Oriol Vinyals: Deep Learning and Artificial General Intelligence | Lex Fridman Podcast #306

at which point is the neural network at which point is the neural network a being a being a being versus a tool the following is a conversation with the following is a conversation with arielle vinialis his second time in the arielle vinialis his second time in the arielle vinialis his second time in the podcast arielle is the research director podcast arielle is the research director podcast arielle is the research director and deep learning lead at deepmind and and deep learning lead at deepmind and and deep learning lead at deepmind and one of the most brilliant thinkers and one of the most brilliant thinkers and one of the most brilliant thinkers and researchers in the history of artificial researchers in the history of artificial researchers in the history of artificial intelligence intelligence intelligence this is the lex friedman podcast to this is the lex friedman podcast to this is the lex friedman podcast to support it please check out our sponsors support it please check out our sponsors support it please check out our sponsors in the description and now dear friends in the description and now dear friends in the description and now dear friends here's arielle vinnie alice here's arielle vinnie alice here's arielle vinnie alice you are one of the most brilliant you are one of the most brilliant you are one of the most brilliant researchers in the history of ai working researchers in the history of ai working researchers in the history of ai working across all kinds of modalities probably across all kinds of modalities probably across all kinds of modalities probably the one common theme is it's always the one common theme is it's always the one common theme is it's always sequences of data uh so that we're sequences of data uh so that we're sequences of data uh so that we're talking about languages images even talking about languages images even talking about languages images even biology and uh games as we talked about biology and uh games as we talked about biology and uh games as we talked about last time so last time so last time so you're a good person to ask this you're a good person to ask this you're a good person to ask this in your lifetime will we be able to in your lifetime will we be able to in your lifetime will we be able to build an ai system that's able to build an ai system that's able to build an ai system that's able to replace me as the interviewer replace me as the interviewer replace me as the interviewer in this conversation in this conversation in this conversation in terms of ability to ask questions in terms of ability to ask questions in terms of ability to ask questions that are compelling to somebody that are compelling to somebody that are compelling to somebody listening and then listening and then listening and then further question is further question is further question is are we close are we close are we close will we be able to build a system that will we be able to build a system that will we be able to build a system that replaces you replaces you replaces you as the interviewee as the interviewee as the interviewee in order to create a compelling in order to create a compelling in order to create a compelling conversation how far away are we do you conversation how far away are we do you conversation how far away are we do you think it's a good question um i think think it's a good question um i think think it's a good question um i think partly i would say do we want that i partly i would say do we want that i partly i would say do we want that i i really like when we start now with i really like when we start now with i really like when we start now with very powerful models interacting with very powerful models interacting with very powerful models interacting with them them them and thinking of them more closer to us

and thinking of them more closer to us and thinking of them more closer to us the question is if you remove the human the question is if you remove the human the question is if you remove the human side of the conversation is that an side of the conversation is that an side of the conversation is that an interesting you know is that an interesting you know is that an interesting you know is that an interesting artifact and i would say interesting artifact and i would say interesting artifact and i would say probably not i've seen for instance um probably not i've seen for instance um probably not i've seen for instance um last time we spoke like was we were last time we spoke like was we were last time we spoke like was we were talking about starcraft um and creating talking about starcraft um and creating talking about starcraft um and creating you know agents that play games involves you know agents that play games involves you know agents that play games involves self-play but ultimately what people self-play but ultimately what people self-play but ultimately what people care about was care about was care about was how does this agent behave when the how does this agent behave when the how does this agent behave when the opposite side is is a human opposite side is is a human opposite side is is a human so so so without a doubt we will probably be more without a doubt we will probably be more without a doubt we will probably be more empowered by ai um maybe you can empowered by ai um maybe you can empowered by ai um maybe you can source some questions from an ai system source some questions from an ai system source some questions from an ai system i mean that even today i would say it's i mean that even today i would say it's i mean that even today i would say it's quite plausible that with your quite plausible that with your quite plausible that with your creativity you might actually find very creativity you might actually find very creativity you might actually find very interesting questions that you can interesting questions that you can interesting questions that you can filter we call this cherry picking filter we call this cherry picking filter we call this cherry picking sometimes in the field of language um sometimes in the field of language um sometimes in the field of language um and likewise if i had now the tools on and likewise if i had now the tools on and likewise if i had now the tools on my side i could say look you're asking my side i could say look you're asking my side i could say look you're asking this interesting question this interesting question this interesting question from this answer i like the words chosen from this answer i like the words chosen from this answer i like the words chosen by this particular system that created a by this particular system that created a by this particular system that created a few words few words few words completely replacing it feels completely replacing it feels completely replacing it feels not exactly exciting to me um although not exactly exciting to me um although not exactly exciting to me um although in my lifetime i think way i mean given in my lifetime i think way i mean given in my lifetime i think way i mean given the trajectory i think it's possible the trajectory i think it's possible the trajectory i think it's possible that perhaps there could be interesting that perhaps there could be interesting that perhaps there could be interesting um maybe self-play interviews as you um maybe self-play interviews as you um maybe self-play interviews as you you're suggesting that would look look you're suggesting that would look look you're suggesting that would look look or sound kind of quite interesting and or sound kind of quite interesting and or sound kind of quite interesting and probably would advocate or you could probably would advocate or you could probably would advocate or you could learn a topic through listening to one learn a topic through listening to one learn a topic through listening to one of these interviews at a basic level at of these interviews at a basic level at of these interviews at a basic level at least so you said it doesn't seem

least so you said it doesn't seem least so you said it doesn't seem exciting to you but what if exciting is exciting to you but what if exciting is exciting to you but what if exciting is part of the objective function the thing part of the objective function the thing part of the objective function the thing is optimized over so you can there's is optimized over so you can there's is optimized over so you can there's probably a huge amount of data probably a huge amount of data probably a huge amount of data of humans if you look correctly of of humans if you look correctly of of humans if you look correctly of humans communicating online and there's humans communicating online and there's humans communicating online and there's probably ways to measure the degree of probably ways to measure the degree of probably ways to measure the degree of you know as they talk about engagement you know as they talk about engagement you know as they talk about engagement so you can probably optimize the so you can probably optimize the so you can probably optimize the question that's most question that's most question that's most created an engaging conversation in the created an engaging conversation in the created an engaging conversation in the past so actually if you strictly use the past so actually if you strictly use the past so actually if you strictly use the word exciting word exciting word exciting there is probably there is probably there is probably a way to create a optimally exciting a way to create a optimally exciting a way to create a optimally exciting conversations conversations conversations that are involved ai systems at least that are involved ai systems at least that are involved ai systems at least one side is ai yeah that makes sense i one side is ai yeah that makes sense i one side is ai yeah that makes sense i think think think maybe looping back a bit to to games and maybe looping back a bit to to games and maybe looping back a bit to to games and the game industry when you design the game industry when you design the game industry when you design algorithms um you're thinking about algorithms um you're thinking about algorithms um you're thinking about winning as the objective right or the winning as the objective right or the winning as the objective right or the reward function but in fact when we reward function but in fact when we reward function but in fact when we discuss this with blizzard the creators discuss this with blizzard the creators discuss this with blizzard the creators of starcraft in this case i think of starcraft in this case i think of starcraft in this case i think what's exciting fun um if you could what's exciting fun um if you could what's exciting fun um if you could measure that and optimize for that measure that and optimize for that measure that and optimize for that that's probably why we play video games that's probably why we play video games that's probably why we play video games or why we interact or listen or look at or why we interact or listen or look at or why we interact or listen or look at cat videos or whatever on the internet cat videos or whatever on the internet cat videos or whatever on the internet so it's true that modeling reward beyond so it's true that modeling reward beyond so it's true that modeling reward beyond the obvious reward functions we've used the obvious reward functions we've used the obvious reward functions we've used to in reinforcement learning is to in reinforcement learning is to in reinforcement learning is definitely very exciting and again there definitely very exciting and again there definitely very exciting and again there is some progress actually into um a is some progress actually into um a is some progress actually into um a particular aspect of ai which is quite particular aspect of ai which is quite particular aspect of ai which is quite critical which is um for instance is a

critical which is um for instance is a critical which is um for instance is a conversation that or is the information conversation that or is the information conversation that or is the information truthful right so you could start trying truthful right so you could start trying truthful right so you could start trying to evaluate um these from to evaluate um these from to evaluate um these from except from the internet right that has except from the internet right that has except from the internet right that has lots of information and then if you can lots of information and then if you can lots of information and then if you can learn a function automated ideally so learn a function automated ideally so learn a function automated ideally so you can also optimize it more easily you can also optimize it more easily you can also optimize it more easily then you could actually have then you could actually have then you could actually have conversations that optimize for conversations that optimize for conversations that optimize for non-obvious things such as excitement non-obvious things such as excitement non-obvious things such as excitement so yeah that's quite possible and then i so yeah that's quite possible and then i so yeah that's quite possible and then i would say in that case it would would say in that case it would would say in that case it would definitely be fun a fun exercise and definitely be fun a fun exercise and definitely be fun a fun exercise and quite unique to have at least one site quite unique to have at least one site quite unique to have at least one site that is fully driven by an excitement that is fully driven by an excitement that is fully driven by an excitement reward function um but obviously reward function um but obviously reward function um but obviously there would be still quite a lot of there would be still quite a lot of there would be still quite a lot of humanity in the system both from who humanity in the system both from who humanity in the system both from who who is building the system of course and who is building the system of course and who is building the system of course and also also also ultimately if we think of labeling for ultimately if we think of labeling for ultimately if we think of labeling for excitement that those labels must come excitement that those labels must come excitement that those labels must come from us because it's just from us because it's just from us because it's just hard to hard to hard to have a computational measure of have a computational measure of have a computational measure of excitement as far as i understand excitement as far as i understand excitement as far as i understand there's no such thing you mentioned truth also i would you mentioned truth also i would actually actually actually venture to say that excitement is easier venture to say that excitement is easier venture to say that excitement is easier to label than truth to label than truth to label than truth or is perhaps uh has lower consequences or is perhaps uh has lower consequences or is perhaps uh has lower consequences of failure of failure of failure but there is but there is but there is perhaps perhaps perhaps the humanness that you mentioned that's the humanness that you mentioned that's the humanness that you mentioned that's perhaps part of a thing that could be perhaps part of a thing that could be perhaps part of a thing that could be labeled and that could mean labeled and that could mean labeled and that could mean an ai system that's doing dialogue an ai system that's doing dialogue an ai system that's doing dialogue that's doing conversations that's doing conversations that's doing conversations should be

should be should be flawed for example flawed for example flawed for example like that's the thing you optimize for like that's the thing you optimize for like that's the thing you optimize for which is uh have inherent contradictions which is uh have inherent contradictions which is uh have inherent contradictions by design have flaws by design by design have flaws by design by design have flaws by design maybe it also needs to have a strong maybe it also needs to have a strong maybe it also needs to have a strong sense of identity sense of identity sense of identity so it has a backstory it told itself so it has a backstory it told itself so it has a backstory it told itself that it sticks to it has memories that it sticks to it has memories that it sticks to it has memories not in terms of the how the system is not in terms of the how the system is not in terms of the how the system is designed but it's able to tell stories designed but it's able to tell stories designed but it's able to tell stories about its past about its past about its past it's it's it's able to have able to have able to have um mortality and fear of mortality in um mortality and fear of mortality in um mortality and fear of mortality in the following way that it has an the following way that it has an the following way that it has an identity identity identity and like if it says something stupid and and like if it says something stupid and and like if it says something stupid and gets cancelled on twitter that's the end gets cancelled on twitter that's the end gets cancelled on twitter that's the end of that system so it's not like you get of that system so it's not like you get of that system so it's not like you get to rebrand yourself that system is to rebrand yourself that system is to rebrand yourself that system is that's it so maybe that the the that's it so maybe that the the that's it so maybe that the the high-stakes nature of it because like high-stakes nature of it because like high-stakes nature of it because like you can't say anything stupid now or oil you can't say anything stupid now or oil you can't say anything stupid now or oil because because because uh you'll be canceled on twitter and uh you'll be canceled on twitter and uh you'll be canceled on twitter and that there's there's stakes to that and that there's there's stakes to that and that there's there's stakes to that and that i think part of the reason that that i think part of the reason that that i think part of the reason that makes it uh makes it uh makes it uh interesting and then you have a interesting and then you have a interesting and then you have a perspective like you've built up over perspective like you've built up over perspective like you've built up over time that you stick with and then people time that you stick with and then people time that you stick with and then people can disagree with you so holding that can disagree with you so holding that can disagree with you so holding that perspective strongly perspective strongly perspective strongly holding sort of a maybe a controversial holding sort of a maybe a controversial holding sort of a maybe a controversial at least a strong opinion all of those at least a strong opinion all of those at least a strong opinion all of those elements it feels like they can be elements it feels like they can be elements it feels like they can be learned because it feels like there's a learned because it feels like there's a learned because it feels like there's a lot of data lot of data lot of data on the internet of people having an on the internet of people having an on the internet of people having an opinion opinion opinion and then combine that with a metric of and then combine that with a metric of and then combine that with a metric of excitement you can start to create excitement you can start to create excitement you can start to create something that as opposed to trying to something that as opposed to trying to something that as opposed to trying to optimize for uh optimize for uh optimize for uh sort of sort of sort of grammatical clarity and truthfulness

grammatical clarity and truthfulness grammatical clarity and truthfulness the the factual the the factual the the factual consistency over many sentences you're consistency over many sentences you're consistency over many sentences you're optimized for optimized for optimized for the humanness the humanness the humanness and there's obviously data for humanness and there's obviously data for humanness and there's obviously data for humanness on the internet on the internet on the internet so i wonder so i wonder so i wonder i wonder if there's a future where i wonder if there's a future where i wonder if there's a future where that's part that's part that's part or i mean i i i sometimes wonder that or i mean i i i sometimes wonder that or i mean i i i sometimes wonder that about myself i'm a huge fan of podcasts about myself i'm a huge fan of podcasts about myself i'm a huge fan of podcasts and i listen to poc some podcasts and i and i listen to poc some podcasts and i and i listen to poc some podcasts and i think like what is interesting about think like what is interesting about think like what is interesting about this what is compelling this what is compelling this what is compelling uh the same way you watch other games uh the same way you watch other games uh the same way you watch other games like you said watch play starcraft or like you said watch play starcraft or like you said watch play starcraft or have magnus carlsen play chess have magnus carlsen play chess have magnus carlsen play chess so i'm not a chess player so but it's so i'm not a chess player so but it's so i'm not a chess player so but it's still interesting to me and what is that still interesting to me and what is that still interesting to me and what is that that's the that's the that's the uh the stakes of it maybe um the end of uh the stakes of it maybe um the end of uh the stakes of it maybe um the end of a domination of a series of wins i don't a domination of a series of wins i don't a domination of a series of wins i don't know there's all those elements know there's all those elements know there's all those elements somehow connect to a compelling somehow connect to a compelling somehow connect to a compelling conversation and i wonder how hard is conversation and i wonder how hard is conversation and i wonder how hard is that to replace because ultimately all that to replace because ultimately all that to replace because ultimately all of that connects the initial proposition of that connects the initial proposition of that connects the initial proposition of how to test of how to test of how to test whether an ai is intelligent or not with whether an ai is intelligent or not with whether an ai is intelligent or not with the turing test the turing test the turing test which i guess my question comes from a which i guess my question comes from a which i guess my question comes from a place of the spirit of that test place of the spirit of that test place of the spirit of that test yes um i actually recall i was just yes um i actually recall i was just yes um i actually recall i was just listening to our first podcast where we listening to our first podcast where we listening to our first podcast where we discussed turing tests um so discussed turing tests um so discussed turing tests um so i would say i would say i would say from a from a from a neural network you know ai builder neural network you know ai builder neural network you know ai builder perspective um there's perspective um there's perspective um there's you know usually you try to map many of you know usually you try to map many of you know usually you try to map many of these interesting topics you discuss to these interesting topics you discuss to these interesting topics you discuss to to benchmarks and then also to actual to benchmarks and then also to actual to benchmarks and then also to actual architectures on the how these systems

architectures on the how these systems architectures on the how these systems are currently built how they learn what are currently built how they learn what are currently built how they learn what data they learn from what are they data they learn from what are they data they learn from what are they learning right we're talking about learning right we're talking about learning right we're talking about weights of a mathematical function and weights of a mathematical function and weights of a mathematical function and then looking at the current state of the then looking at the current state of the then looking at the current state of the game maybe game maybe game maybe what do we what do we what do we need leaps forward to get to the need leaps forward to get to the need leaps forward to get to the ultimate stage of all these experiences ultimate stage of all these experiences ultimate stage of all these experiences um lifetime experience of fears like um lifetime experience of fears like um lifetime experience of fears like words that currently words that currently words that currently barely we're we're seeing um progress barely we're we're seeing um progress barely we're we're seeing um progress just because what's happening today is just because what's happening today is just because what's happening today is you take you take you take all these human interactions um it's a all these human interactions um it's a all these human interactions um it's a large bust of variety of human large bust of variety of human large bust of variety of human interactions online and then you're interactions online and then you're interactions online and then you're distilling these distilling these distilling these sequences right going back to my passion sequences right going back to my passion sequences right going back to my passion like sequences of words letters um like sequences of words letters um like sequences of words letters um images sound there's more modalities images sound there's more modalities images sound there's more modalities here to be to be at play and then you're here to be to be at play and then you're here to be to be at play and then you're trying to trying to trying to just learn a function that will be happy just learn a function that will be happy just learn a function that will be happy that maximizes the the likelihood of that maximizes the the likelihood of that maximizes the the likelihood of seeing all these um through a neural seeing all these um through a neural seeing all these um through a neural network um now network um now network um now i think there's a few i think there's a few i think there's a few places where the way currently we train places where the way currently we train places where the way currently we train these models would clearly like to be these models would clearly like to be these models would clearly like to be able to develop the kinds of able to develop the kinds of able to develop the kinds of capabilities you save i'll tell you capabilities you save i'll tell you capabilities you save i'll tell you maybe a couple one is maybe a couple one is maybe a couple one is the lifetime of an agent or a model the lifetime of an agent or a model the lifetime of an agent or a model so you so you so you you learn from this data offline right you learn from this data offline right you learn from this data offline right so you're just passively observing and so you're just passively observing and so you're just passively observing and maximizing this you know it's almost maximizing this you know it's almost maximizing this you know it's almost like a mountains like a landscape of like a mountains like a landscape of like a mountains like a landscape of mountains and then everywhere there's mountains and then everywhere there's mountains and then everywhere there's data that humans interacted in this way

data that humans interacted in this way data that humans interacted in this way you're trying to make that higher and you're trying to make that higher and you're trying to make that higher and then you know lower where there's no then you know lower where there's no then you know lower where there's no data and then these models generally data and then these models generally data and then these models generally don't don't don't then experience themselves these they then experience themselves these they then experience themselves these they just are observers right they're passive just are observers right they're passive just are observers right they're passive observers of the data and then we're observers of the data and then we're observers of the data and then we're putting them to then generate data when putting them to then generate data when putting them to then generate data when we interact with them but that's very we interact with them but that's very we interact with them but that's very limiting the experience they actually limiting the experience they actually limiting the experience they actually experience um when they could maybe be experience um when they could maybe be experience um when they could maybe be optimizing or further optimizing the optimizing or further optimizing the optimizing or further optimizing the weights we're not even doing that so to weights we're not even doing that so to weights we're not even doing that so to be clear and again mapping to be clear and again mapping to be clear and again mapping to alphago alpha star we train the model alphago alpha star we train the model alphago alpha star we train the model and when we deploy it um to play against and when we deploy it um to play against and when we deploy it um to play against humans or in this case interact with humans or in this case interact with humans or in this case interact with humans um like language models they humans um like language models they humans um like language models they don't even keep training right they're don't even keep training right they're don't even keep training right they're not learning in the sense of the weights not learning in the sense of the weights not learning in the sense of the weights that you've that you've that you've learned from the data they don't keep learned from the data they don't keep learned from the data they don't keep changing changing changing now there's something a bit more now there's something a bit more now there's something a bit more feels magical but it's understandable if feels magical but it's understandable if feels magical but it's understandable if you're into neural nets which is well you're into neural nets which is well you're into neural nets which is well they might not they might not they might not learn in the strict sense of the words learn in the strict sense of the words learn in the strict sense of the words the way it's changing maybe that's the way it's changing maybe that's the way it's changing maybe that's mapping to how neurons interconnect and mapping to how neurons interconnect and mapping to how neurons interconnect and how we learn over our lifetime but it's how we learn over our lifetime but it's how we learn over our lifetime but it's true that true that true that the context of the conversation that the context of the conversation that the context of the conversation that they they that takes takes place with they they that takes takes place with they they that takes takes place with when you talk to these systems it's held when you talk to these systems it's held when you talk to these systems it's held in their working memory right it's in their working memory right it's in their working memory right it's almost like um you start a computer it almost like um you start a computer it almost like um you start a computer it has a hard drive that has a lot of has a hard drive that has a lot of has a hard drive that has a lot of information you have access to the information you have access to the information you have access to the internet which has probably all the internet which has probably all the internet which has probably all the information but there's also a working information but there's also a working information but there's also a working memory memory memory where the these agents as we call them

where the these agents as we call them where the these agents as we call them or start calling them build upon or start calling them build upon or start calling them build upon now this memory is very limited um i now this memory is very limited um i now this memory is very limited um i mean right now we're talking to be mean right now we're talking to be mean right now we're talking to be concrete about 2 000 words that we hold concrete about 2 000 words that we hold concrete about 2 000 words that we hold and then beyond that we start forgetting and then beyond that we start forgetting and then beyond that we start forgetting what we've seen so you can see that what we've seen so you can see that what we've seen so you can see that there's some short-term coherence there's some short-term coherence there's some short-term coherence already right with when you said i mean already right with when you said i mean already right with when you said i mean it's a very interesting topic um having it's a very interesting topic um having it's a very interesting topic um having sort of a mapping sort of a mapping sort of a mapping um an agent to like have consistency um an agent to like have consistency um an agent to like have consistency then you know if if you say oh what's then you know if if you say oh what's then you know if if you say oh what's your name um it could remember that but your name um it could remember that but your name um it could remember that but then it might forget beyond 2000 words then it might forget beyond 2000 words then it might forget beyond 2000 words which is not which is not which is not that long of context if we think even of that long of context if we think even of that long of context if we think even of these podcast um books are much longer these podcast um books are much longer these podcast um books are much longer so so so technically speaking there's a technically speaking there's a technically speaking there's a limitation there super exciting from limitation there super exciting from limitation there super exciting from people that work on deep learning to be people that work on deep learning to be people that work on deep learning to be working on working on working on but but but i would say we lack maybe benchmarks and i would say we lack maybe benchmarks and i would say we lack maybe benchmarks and the technology to have the technology to have the technology to have this lifetime this lifetime this lifetime like experience of memory that keeps like experience of memory that keeps like experience of memory that keeps building up building up building up um however the way it learns offline is um however the way it learns offline is um however the way it learns offline is clearly very powerful right so clearly very powerful right so clearly very powerful right so i you know you asked me three years ago i you know you asked me three years ago i you know you asked me three years ago i would say oh we're very far i think i would say oh we're very far i think i would say oh we're very far i think we've seen the power of this imitation we've seen the power of this imitation we've seen the power of this imitation again again again on the internet scale that has enabled on the internet scale that has enabled on the internet scale that has enabled this um to this um to this um to feel like at least the knowledge the feel like at least the knowledge the feel like at least the knowledge the basic knowledge about the world now is basic knowledge about the world now is basic knowledge about the world now is incorporated into the weights incorporated into the weights incorporated into the weights but then this but then this but then this experience is lacking and in fact as i experience is lacking and in fact as i experience is lacking and in fact as i said we don't even train them when you said we don't even train them when you said we don't even train them when you know when we're talking to them other

know when we're talking to them other know when we're talking to them other than than than their working memory of course is their working memory of course is their working memory of course is affected so that's the dynamic part but affected so that's the dynamic part but affected so that's the dynamic part but they don't learn in the same way that they don't learn in the same way that they don't learn in the same way that you and i have learned right when you and i have learned right when you and i have learned right when from basically when we were born and from basically when we were born and from basically when we were born and probably before probably before probably before so lots of fascinating interesting so lots of fascinating interesting so lots of fascinating interesting questions you asked there i think um questions you asked there i think um questions you asked there i think um the one i mentioned is this idea of the one i mentioned is this idea of the one i mentioned is this idea of memory and experience versus just kind memory and experience versus just kind memory and experience versus just kind of observe the world and learn its of observe the world and learn its of observe the world and learn its knowledge which i think for that i would knowledge which i think for that i would knowledge which i think for that i would argue lots of recent advancements that argue lots of recent advancements that argue lots of recent advancements that make me very excited about the field make me very excited about the field make me very excited about the field and then the second and then the second and then the second maybe issue that i see is maybe issue that i see is maybe issue that i see is all these models all these models all these models we train them from scratch that's we train them from scratch that's we train them from scratch that's something i would have complained three something i would have complained three something i would have complained three years ago or six years ago or 10 years years ago or six years ago or 10 years years ago or six years ago or 10 years ago ago ago and it feels and it feels and it feels if we take inspiration from how we got if we take inspiration from how we got if we take inspiration from how we got here how the universe evolved us here how the universe evolved us here how the universe evolved us and we keep evolving it feels and we keep evolving it feels and we keep evolving it feels that is a missing piece that we should that is a missing piece that we should that is a missing piece that we should not be training models from scratch um not be training models from scratch um not be training models from scratch um every few months that there should be every few months that there should be every few months that there should be some sort of some sort of some sort of way in which we can grow models um much way in which we can grow models um much way in which we can grow models um much like as a species and many other like as a species and many other like as a species and many other elements in the universe is building elements in the universe is building elements in the universe is building from the previous sort of iterations from the previous sort of iterations from the previous sort of iterations and that from uh just purely neural and that from uh just purely neural and that from uh just purely neural network perspective network perspective network perspective even though we we would like to make it even though we we would like to make it even though we we would like to make it work it's proven very hard to not work it's proven very hard to not work it's proven very hard to not you know throw away the previous weights you know throw away the previous weights you know throw away the previous weights right this landscape we learn from the right this landscape we learn from the right this landscape we learn from the data and you know refresh it with a

data and you know refresh it with a data and you know refresh it with a brand new set of weights um given brand new set of weights um given brand new set of weights um given maybe a maybe a maybe a recent snapshot of these data sets we recent snapshot of these data sets we recent snapshot of these data sets we train on etc or even a new game we're train on etc or even a new game we're train on etc or even a new game we're learning so that's learning so that's learning so that's that feels like something is missing that feels like something is missing that feels like something is missing fundamentally we might find it but it's fundamentally we might find it but it's fundamentally we might find it but it's not very clear how it will look like not very clear how it will look like not very clear how it will look like there's many ideas and it's super there's many ideas and it's super there's many ideas and it's super exciting as well yes just for people who exciting as well yes just for people who exciting as well yes just for people who don't know when you approach a new don't know when you approach a new don't know when you approach a new problem in machine learning problem in machine learning problem in machine learning you're going to come up with an you're going to come up with an you're going to come up with an architecture that has a a bunch of architecture that has a a bunch of architecture that has a a bunch of weights and then you initialize them weights and then you initialize them weights and then you initialize them somehow which somehow which somehow which in most cases is some version of random in most cases is some version of random in most cases is some version of random so that's what you mean by starting from so that's what you mean by starting from so that's what you mean by starting from scratch and it seems like it's a it's a scratch and it seems like it's a it's a scratch and it seems like it's a it's a waste waste waste every time you every time you every time you solve uh the game of go in chess solve uh the game of go in chess solve uh the game of go in chess starcraft starcraft starcraft uh protein folding like surely there's uh protein folding like surely there's uh protein folding like surely there's some way to reuse the weights as we grow some way to reuse the weights as we grow some way to reuse the weights as we grow this giant database of this giant database of this giant database of of of of of neural networks that have solved some of neural networks that have solved some of neural networks that have solved some of the toughest problems in the world of the toughest problems in the world of the toughest problems in the world and so and so and so some of that is um some of that is um some of that is um what is that methods what is that methods what is that methods how to reuse weights how to reuse weights how to reuse weights how to learn extract what's how to learn extract what's how to learn extract what's generalizable or at least has a chance generalizable or at least has a chance generalizable or at least has a chance to be to be to be and throw away the other stuff and throw away the other stuff and throw away the other stuff uh and maybe the neural network itself uh and maybe the neural network itself uh and maybe the neural network itself should be able to tell you that should be able to tell you that should be able to tell you that like what like what like what um yeah how do you what ideas do you um yeah how do you what ideas do you um yeah how do you what ideas do you have for better initialization of have for better initialization of have for better initialization of weights maybe stepping back if we look weights maybe stepping back if we look weights maybe stepping back if we look at the field of machine learning but

at the field of machine learning but at the field of machine learning but especially deep learning especially deep learning especially deep learning right at the core of deep learning right at the core of deep learning right at the core of deep learning there's this beautiful idea that is there's this beautiful idea that is there's this beautiful idea that is a single algorithm can solve any task a single algorithm can solve any task a single algorithm can solve any task right so right so right so it's been proven over and over with more it's been proven over and over with more it's been proven over and over with more increasing set of benchmarks and things increasing set of benchmarks and things increasing set of benchmarks and things that were thought impossible that are that were thought impossible that are that were thought impossible that are being cracked by this basic principle being cracked by this basic principle being cracked by this basic principle that is that is that is you take a neural network of you take a neural network of you take a neural network of uninitialized ways so like a blank uninitialized ways so like a blank uninitialized ways so like a blank computational brain computational brain computational brain um then you give it um then you give it um then you give it in the case of supervised learning a lot in the case of supervised learning a lot in the case of supervised learning a lot ideally of examples of hey here is what ideally of examples of hey here is what ideally of examples of hey here is what the input looks like and the desired the input looks like and the desired the input looks like and the desired output should look like this i mean output should look like this i mean output should look like this i mean image classification is very clear image classification is very clear image classification is very clear example images to maybe one of a example images to maybe one of a example images to maybe one of a thousand categories that's what imagenet thousand categories that's what imagenet thousand categories that's what imagenet is like but many many if not all is like but many many if not all is like but many many if not all problems can be mapped this way problems can be mapped this way problems can be mapped this way and then and then and then there's a generic recipe right that you there's a generic recipe right that you there's a generic recipe right that you can use can use can use um and this recipe with um and this recipe with um and this recipe with very little change and i think that's very little change and i think that's very little change and i think that's the core of deep learning research right the core of deep learning research right the core of deep learning research right that what is the recipe that is that what is the recipe that is that what is the recipe that is universal that for any new given task universal that for any new given task universal that for any new given task i'll be able to use without thinking i'll be able to use without thinking i'll be able to use without thinking without having to work very hard on the without having to work very hard on the without having to work very hard on the problem at stake problem at stake problem at stake we have not found this recipe but we have not found this recipe but we have not found this recipe but i think the field is i think the field is i think the field is excited to find um less tweaks or tricks excited to find um less tweaks or tricks excited to find um less tweaks or tricks that people find when they work on that people find when they work on that people find when they work on important problems specific to those and important problems specific to those and important problems specific to those and more of a general algorithm right so at more of a general algorithm right so at more of a general algorithm right so at an algorithmic level i would say we have an algorithmic level i would say we have an algorithmic level i would say we have something general already which is this

something general already which is this something general already which is this formula of training a very powerful formula of training a very powerful formula of training a very powerful model a neural network on a lot of data model a neural network on a lot of data model a neural network on a lot of data and in many cases and in many cases and in many cases you need some specificity to the actual you need some specificity to the actual you need some specificity to the actual problem you're solving um protein problem you're solving um protein problem you're solving um protein folding being such an important problem folding being such an important problem folding being such an important problem has some basic recipe that is learned has some basic recipe that is learned has some basic recipe that is learned from beyond before right like from beyond before right like from beyond before right like transformer models graph neural networks transformer models graph neural networks transformer models graph neural networks um ideas coming from nlp like uh you um ideas coming from nlp like uh you um ideas coming from nlp like uh you know know know something called birth that is a kind of something called birth that is a kind of something called birth that is a kind of loss that you can in place to help the loss that you can in place to help the loss that you can in place to help the model uh model uh model uh knowledge distillation is another knowledge distillation is another knowledge distillation is another technique right so this is the formula technique right so this is the formula technique right so this is the formula we still had to find some particular we still had to find some particular we still had to find some particular things that were specific to alpha fault things that were specific to alpha fault things that were specific to alpha fault right that's very important because right that's very important because right that's very important because protein folding is such a high value protein folding is such a high value protein folding is such a high value problem that as humans we should solve problem that as humans we should solve problem that as humans we should solve it no matter if we need to be a bit it no matter if we need to be a bit it no matter if we need to be a bit specific and it's possible that some of specific and it's possible that some of specific and it's possible that some of these learnings will apply then to the these learnings will apply then to the these learnings will apply then to the next iteration of this recipe that deep next iteration of this recipe that deep next iteration of this recipe that deep learners are about learners are about learners are about but it is true that so far but it is true that so far but it is true that so far the recipe is what's common but the the recipe is what's common but the the recipe is what's common but the weights you generally throw away which weights you generally throw away which weights you generally throw away which feels very sad um feels very sad um feels very sad um although although although maybe in the last especially in the last maybe in the last especially in the last maybe in the last especially in the last two three years two three years two three years and when we last spoke i mentioned this and when we last spoke i mentioned this and when we last spoke i mentioned this area of metal learning which is the idea area of metal learning which is the idea area of metal learning which is the idea of learning to learn of learning to learn of learning to learn that idea and some progress has been had that idea and some progress has been had that idea and some progress has been had starting i would say mostly from gpt3 on starting i would say mostly from gpt3 on starting i would say mostly from gpt3 on the language domain only in which you the language domain only in which you the language domain only in which you could conceive a model that is trained could conceive a model that is trained could conceive a model that is trained once and then this model is not narrow

once and then this model is not narrow once and then this model is not narrow in that it only knows how to translate a in that it only knows how to translate a in that it only knows how to translate a pair of languages or it only knows how pair of languages or it only knows how pair of languages or it only knows how to assign sentiment to a sentence these to assign sentiment to a sentence these to assign sentiment to a sentence these these actually these actually these actually you could teach it by a prompting is you could teach it by a prompting is you could teach it by a prompting is called and this prompting is essentially called and this prompting is essentially called and this prompting is essentially just showing it a few more examples um just showing it a few more examples um just showing it a few more examples um almost like you do show examples input almost like you do show examples input almost like you do show examples input output examples algorithmically speaking output examples algorithmically speaking output examples algorithmically speaking to the process of creating this model to the process of creating this model to the process of creating this model but now you're doing it through language but now you're doing it through language but now you're doing it through language which is very natural way for us to which is very natural way for us to which is very natural way for us to learn from one another i tell you hey learn from one another i tell you hey learn from one another i tell you hey you should do this new task i'll tell you should do this new task i'll tell you should do this new task i'll tell you a bit more maybe you asked me some you a bit more maybe you asked me some you a bit more maybe you asked me some questions and now you know the task questions and now you know the task questions and now you know the task right you didn't need to retrain it from right you didn't need to retrain it from right you didn't need to retrain it from scratch and we've seen these magical scratch and we've seen these magical scratch and we've seen these magical moments almost um in this way to do moments almost um in this way to do moments almost um in this way to do fuchsia prompting through language on fuchsia prompting through language on fuchsia prompting through language on language only domain and then in the language only domain and then in the language only domain and then in the last last last two years we've seen these expanded to two years we've seen these expanded to two years we've seen these expanded to beyond language beyond language beyond language adding vision adding actions and games adding vision adding actions and games adding vision adding actions and games lots of progress to be had but this is lots of progress to be had but this is lots of progress to be had but this is maybe if you ask me like about how are maybe if you ask me like about how are maybe if you ask me like about how are we going to crack this problem this is we going to crack this problem this is we going to crack this problem this is perhaps one way in which you have a perhaps one way in which you have a perhaps one way in which you have a single model single model single model the problem of this model is it's hard the problem of this model is it's hard the problem of this model is it's hard to grow to grow to grow in weights or capacity but the model is in weights or capacity but the model is in weights or capacity but the model is certainly so powerful that you can teach certainly so powerful that you can teach certainly so powerful that you can teach it some tasks right in this way that i it some tasks right in this way that i it some tasks right in this way that i teach you i could teach you a new task teach you i could teach you a new task teach you i could teach you a new task now if we were oh let's a text a now if we were oh let's a text a now if we were oh let's a text a text-based task or a classification a text-based task or a classification a text-based task or a classification a vision style task vision style task vision style task but it still feels like more but it still feels like more but it still feels like more breakthroughs should be hot but it's a breakthroughs should be hot but it's a breakthroughs should be hot but it's a great beginning right we have a good

great beginning right we have a good great beginning right we have a good baseline we have an idea that this maybe baseline we have an idea that this maybe baseline we have an idea that this maybe is the way we want to benchmark progress is the way we want to benchmark progress is the way we want to benchmark progress towards agi and i think in my view towards agi and i think in my view towards agi and i think in my view that's critical to always have a way to that's critical to always have a way to that's critical to always have a way to benchmark the community sort of benchmark the community sort of benchmark the community sort of converging to this overall which is good converging to this overall which is good converging to this overall which is good to see to see to see and then this is actually and then this is actually and then this is actually what excites me in terms of also next what excites me in terms of also next what excites me in terms of also next steps um for deep learning is how to steps um for deep learning is how to steps um for deep learning is how to make these models more powerful how do make these models more powerful how do make these models more powerful how do you train them how to grow them if they you train them how to grow them if they you train them how to grow them if they must grow should they change their must grow should they change their must grow should they change their weights as you teach it the task or not weights as you teach it the task or not weights as you teach it the task or not there's some interesting questions many there's some interesting questions many there's some interesting questions many to be answered yeah you've opened the to be answered yeah you've opened the to be answered yeah you've opened the door about door about door about to a bunch of questions i want to ask to a bunch of questions i want to ask to a bunch of questions i want to ask but let's first return to the but let's first return to the but let's first return to the uh to your tweet and read it like a uh to your tweet and read it like a uh to your tweet and read it like a shakespeare you wrote gato is not the shakespeare you wrote gato is not the shakespeare you wrote gato is not the end it's the beginning and then he wrote end it's the beginning and then he wrote end it's the beginning and then he wrote meow and then an emoji of a cat meow and then an emoji of a cat meow and then an emoji of a cat uh so first two questions first can you uh so first two questions first can you uh so first two questions first can you explain the meow and the cat emoji and explain the meow and the cat emoji and explain the meow and the cat emoji and second can you explain what gatto is and second can you explain what gatto is and second can you explain what gatto is and how it works right indeed i mean thanks how it works right indeed i mean thanks how it works right indeed i mean thanks thanks for reminding me that we're all thanks for reminding me that we're all thanks for reminding me that we're all exposing on twitter and exposing on twitter and exposing on twitter and permanently there yes permanently one of permanently there yes permanently one of permanently there yes permanently one of the greatest ai researchers of all time the greatest ai researchers of all time the greatest ai researchers of all time meow and cat emoji yes there you go meow and cat emoji yes there you go meow and cat emoji yes there you go right so can you imagine like touring uh right so can you imagine like touring uh right so can you imagine like touring uh tweeting tweeting tweeting meow and cat probably he would probably meow and cat probably he would probably meow and cat probably he would probably would probably so yeah the tweet would probably so yeah the tweet would probably so yeah the tweet is important actually um you know i put is important actually um you know i put is important actually um you know i put thought on the tweets i hope people thought on the tweets i hope people thought on the tweets i hope people which part you think okay which part you think okay which part you think okay so there's three sentences

so there's three sentences so there's three sentences gato is not the end gato is not the end gato is not the end gato is the beginning gato is the beginning gato is the beginning meow cat emoji okay which is the meow cat emoji okay which is the meow cat emoji okay which is the important part the meow no no important part the meow no no important part the meow no no definitely um that it is the beginning i definitely um that it is the beginning i definitely um that it is the beginning i mean i i probably was just explaining um mean i i probably was just explaining um mean i i probably was just explaining um a bit a bit a bit where the field is going but um let me where the field is going but um let me where the field is going but um let me tell you about gato so tell you about gato so tell you about gato so first the name gato first the name gato first the name gato comes from maybe a sequence of releases comes from maybe a sequence of releases comes from maybe a sequence of releases that deepmind had that that deepmind had that that deepmind had that named uh like used animal names to name named uh like used animal names to name named uh like used animal names to name some of their models that are based on some of their models that are based on some of their models that are based on this idea of large sequence models this idea of large sequence models this idea of large sequence models initially their only language but we're initially their only language but we're initially their only language but we're expanding to other modalities so we had expanding to other modalities so we had expanding to other modalities so we had a you know we had a you know we had a you know we had gopher gopher gopher chinchilla these were language only and chinchilla these were language only and chinchilla these were language only and then more recently we released flamingo then more recently we released flamingo then more recently we released flamingo which adds vision to the equation and which adds vision to the equation and which adds vision to the equation and then gato which then gato which then gato which adds vision and then also actions in the adds vision and then also actions in the adds vision and then also actions in the mix right um as we discuss actually mix right um as we discuss actually mix right um as we discuss actually actions um especially discrete actions actions um especially discrete actions actions um especially discrete actions like up down left right like up down left right like up down left right i just told you the actions but they're i just told you the actions but they're i just told you the actions but they're words so you can kind of see how actions words so you can kind of see how actions words so you can kind of see how actions naturally map to sequence modeling of naturally map to sequence modeling of naturally map to sequence modeling of words which these models are very words which these models are very words which these models are very powerful powerful powerful so so so gato was named after i believe i can gato was named after i believe i can gato was named after i believe i can only from memory right this you know only from memory right this you know only from memory right this you know these things always happen with an these things always happen with an these things always happen with an amazing team of researchers behind so amazing team of researchers behind so amazing team of researchers behind so before the release yeah um we had a before the release yeah um we had a before the release yeah um we had a discussion about which animal would we discussion about which animal would we discussion about which animal would we pick right and i think because of the

pick right and i think because of the pick right and i think because of the word general agent right and and this is word general agent right and and this is word general agent right and and this is a property quite unique to gato um we we a property quite unique to gato um we we a property quite unique to gato um we we kind of were playing with the ga words kind of were playing with the ga words kind of were playing with the ga words and then you know gato and rice of cat and then you know gato and rice of cat and then you know gato and rice of cat yes um and gato is obviously a spanish yes um and gato is obviously a spanish yes um and gato is obviously a spanish version of cat i had nothing to do with version of cat i had nothing to do with version of cat i had nothing to do with it although i'm from spain it although i'm from spain it although i'm from spain wait sorry how do you say cat in spanish wait sorry how do you say cat in spanish wait sorry how do you say cat in spanish gato oh god okay yeah no okay okay i see gato oh god okay yeah no okay okay i see gato oh god okay yeah no okay okay i see i see i see you now it all makes sense i see i see you now it all makes sense i see i see you now it all makes sense okay how do you say meow in spanish no okay how do you say meow in spanish no okay how do you say meow in spanish no that's that's that's i think you you say it the same way i think you you say it the same way i think you you say it the same way but you write it uh is but you write it uh is but you write it uh is m-i-a-u okay it's universal yeah all m-i-a-u okay it's universal yeah all m-i-a-u okay it's universal yeah all right so then how does the thing work so right so then how does the thing work so right so then how does the thing work so you said general is you said general is you said general is so you said uh language so you said uh language so you said uh language vision vision vision and action action and action action and action action how does this how does this how does this can you explain what kind of neural can you explain what kind of neural can you explain what kind of neural networks are involved what does the networks are involved what does the networks are involved what does the training look like training look like training look like maybe um maybe um maybe um what you are some beautiful ideas within what you are some beautiful ideas within what you are some beautiful ideas within the system yeah so the system yeah so the system yeah so maybe the basics of gato are not that maybe the basics of gato are not that maybe the basics of gato are not that dissimilar from many many work that dissimilar from many many work that dissimilar from many many work that comes so here is where the the sort of comes so here is where the the sort of comes so here is where the the sort of the recipe i mean hasn't changed too the recipe i mean hasn't changed too the recipe i mean hasn't changed too much there is a transformer model that's much there is a transformer model that's much there is a transformer model that's just the kind of recurrent neural just the kind of recurrent neural just the kind of recurrent neural network network network that essentially takes a sequence of that essentially takes a sequence of that essentially takes a sequence of modalities observations that could be modalities observations that could be modalities observations that could be words could be vision or could be words could be vision or could be words could be vision or could be actions and then actions and then actions and then its own objective that you train it to its own objective that you train it to its own objective that you train it to do when you train it is to predict what do when you train it is to predict what do when you train it is to predict what the next the next the next anything is and anything means what's

anything is and anything means what's anything is and anything means what's the next action if this sequence that the next action if this sequence that the next action if this sequence that i'm showing you to train is a sequence i'm showing you to train is a sequence i'm showing you to train is a sequence of actions and observations then you're of actions and observations then you're of actions and observations then you're predicting what's the next action and predicting what's the next action and predicting what's the next action and the next observation right so you you the next observation right so you you the next observation right so you you think of of this really as a sequence of think of of this really as a sequence of think of of this really as a sequence of bytes right so take any sequence um of bytes right so take any sequence um of bytes right so take any sequence um of words a sequence of interleaved words words a sequence of interleaved words words a sequence of interleaved words and images a sequence of um maybe um and images a sequence of um maybe um and images a sequence of um maybe um observations that are images and moves observations that are images and moves observations that are images and moves in atari up down left right and these in atari up down left right and these in atari up down left right and these you just you just you just think of them as bytes and you're think of them as bytes and you're think of them as bytes and you're modeling what's the next byte gonna be modeling what's the next byte gonna be modeling what's the next byte gonna be like and you might interpret that as an like and you might interpret that as an like and you might interpret that as an action as an action and then play it in action as an action and then play it in action as an action and then play it in a game or you could interpret it as a a game or you could interpret it as a a game or you could interpret it as a word and then write it down if you're word and then write it down if you're word and then write it down if you're chatting with the system and so on um chatting with the system and so on um chatting with the system and so on um so gato basically so gato basically so gato basically can be but can be thought as inputs can be but can be thought as inputs can be but can be thought as inputs images images images text text text video video video actions actions actions it also actually inputs some sort of it also actually inputs some sort of it also actually inputs some sort of proprioception sensors from robotics proprioception sensors from robotics proprioception sensors from robotics because robotics is one of the tasks because robotics is one of the tasks because robotics is one of the tasks that it's been trained to do and then at that it's been trained to do and then at that it's been trained to do and then at the output similarly it outputs words the output similarly it outputs words the output similarly it outputs words actions it does not output images um actions it does not output images um actions it does not output images um that's just by design we decided not to that's just by design we decided not to that's just by design we decided not to go that way for now um that's also in go that way for now um that's also in go that way for now um that's also in part why it's the beginning because part why it's the beginning because part why it's the beginning because there's more to do clearly there's more to do clearly there's more to do clearly but that's kind of what the gato is is but that's kind of what the gato is is but that's kind of what the gato is is this brain that essentially you give it this brain that essentially you give it this brain that essentially you give it any sequence of these observations and any sequence of these observations and any sequence of these observations and and modalities and it outputs the next and modalities and it outputs the next and modalities and it outputs the next step and then you off you go you fit the step and then you off you go you fit the step and then you off you go you fit the next the next step into and predict the

next the next step into and predict the next the next step into and predict the next one and so on now next one and so on now next one and so on now it is it is it is more than a language model because even more than a language model because even more than a language model because even though you can chat with gato like you though you can chat with gato like you though you can chat with gato like you can chat with chinchilla or flamingo um can chat with chinchilla or flamingo um can chat with chinchilla or flamingo um it also it also it also is an agent right so that's is an agent right so that's is an agent right so that's why we call it a of gato like the the why we call it a of gato like the the why we call it a of gato like the the word uh the letter a and also it's word uh the letter a and also it's word uh the letter a and also it's general um it's not an agent that's been general um it's not an agent that's been general um it's not an agent that's been trained to be good at only starcraft or trained to be good at only starcraft or trained to be good at only starcraft or only atari or only go it's been trained only atari or only go it's been trained only atari or only go it's been trained on a vast variety of data sets so on a vast variety of data sets so on a vast variety of data sets so what makes an agent if i may interrupt what makes an agent if i may interrupt what makes an agent if i may interrupt the fact that it can generate actions the fact that it can generate actions the fact that it can generate actions yes so yes so yes so when we call it i mean it's a it's a when we call it i mean it's a it's a when we call it i mean it's a it's a good question right what why when do we good question right what why when do we good question right what why when do we call a model i mean everything is a call a model i mean everything is a call a model i mean everything is a model but what is an agent in my view is model but what is an agent in my view is model but what is an agent in my view is indeed the capacity to take actions in indeed the capacity to take actions in indeed the capacity to take actions in an environment that you then send to it an environment that you then send to it an environment that you then send to it and then the environment might return and then the environment might return and then the environment might return with a new observation um and then you with a new observation um and then you with a new observation um and then you generate the next action this this generate the next action this this generate the next action this this actually this reminds me of the question actually this reminds me of the question actually this reminds me of the question from the side of biology what is life from the side of biology what is life from the side of biology what is life which is actually a very difficult which is actually a very difficult which is actually a very difficult question as well what is living question as well what is living question as well what is living what is living when you think about life what is living when you think about life what is living when you think about life here on this planet earth and a question here on this planet earth and a question here on this planet earth and a question interesting to me about aliens what is interesting to me about aliens what is interesting to me about aliens what is life when we visit another planet would life when we visit another planet would life when we visit another planet would we be able to recognize it and this we be able to recognize it and this we be able to recognize it and this feels like it sounds perhaps silly but i feels like it sounds perhaps silly but i feels like it sounds perhaps silly but i don't think it is at which point is the don't think it is at which point is the don't think it is at which point is the neural network neural network neural network a being a being a being versus a tool versus a tool versus a tool and it feels like action ability to

and it feels like action ability to and it feels like action ability to modify its environment as that modify its environment as that modify its environment as that fundamental leap fundamental leap fundamental leap yeah i think it's it certainly feels yeah i think it's it certainly feels yeah i think it's it certainly feels like action is a necessary condition to like action is a necessary condition to like action is a necessary condition to to be to be to be more alive but probably not sufficient more alive but probably not sufficient more alive but probably not sufficient either um yeah so sadly consciousness either um yeah so sadly consciousness either um yeah so sadly consciousness thing whatever yeah yeah we can get back thing whatever yeah yeah we can get back thing whatever yeah yeah we can get back to that later but anyways going back to to that later but anyways going back to to that later but anyways going back to the meow and the legato right so the meow and the legato right so the meow and the legato right so um um um one of the one of the one of the leaps forward and what took the team a leaps forward and what took the team a leaps forward and what took the team a lot of effort and time was lot of effort and time was lot of effort and time was um as you were asking um as you were asking um as you were asking how has gato been trained so i told you how has gato been trained so i told you how has gato been trained so i told you gato is this transformer neural network gato is this transformer neural network gato is this transformer neural network models actions um models actions um models actions um sequences of actions words etc sequences of actions words etc sequences of actions words etc and then the way we train it is by and then the way we train it is by and then the way we train it is by essentially pulling data sets essentially pulling data sets essentially pulling data sets of of of um observations right so it's a massive um observations right so it's a massive um observations right so it's a massive imitation learning algorithm that it it imitation learning algorithm that it it imitation learning algorithm that it it imitates obviously to imitates obviously to imitates obviously to what is the next word that comes next what is the next word that comes next what is the next word that comes next from the usual data sets we used before from the usual data sets we used before from the usual data sets we used before right so these these are these web scale right so these these are these web scale right so these these are these web scale style data sets of people um writing you style data sets of people um writing you style data sets of people um writing you know know know on on webs or chatting or whatnot right on on webs or chatting or whatnot right on on webs or chatting or whatnot right so that's an obvious source that we use so that's an obvious source that we use so that's an obvious source that we use on all language work but then on all language work but then on all language work but then we also took a lot of agents that we we also took a lot of agents that we we also took a lot of agents that we have at deepmind i mean as you know have at deepmind i mean as you know have at deepmind i mean as you know deepmind we're quite um deepmind we're quite um deepmind we're quite um you know we're quite interested in you know we're quite interested in you know we're quite interested in learning um reinforcement learning and learning um reinforcement learning and learning um reinforcement learning and learning agents that play in different learning agents that play in different learning agents that play in different environments so we kind of created a environments so we kind of created a environments so we kind of created a data set of these trajectories as we

data set of these trajectories as we data set of these trajectories as we call them or asian experiences so in a call them or asian experiences so in a call them or asian experiences so in a way there are other agents we train for way there are other agents we train for way there are other agents we train for a single mind purpose to let's say um a single mind purpose to let's say um a single mind purpose to let's say um you know control a 3d game environment you know control a 3d game environment you know control a 3d game environment and navigate a maze so we had all the and navigate a maze so we had all the and navigate a maze so we had all the experience that was created through the experience that was created through the experience that was created through the one agent interacting with that one agent interacting with that one agent interacting with that environment and we added this to the environment and we added this to the environment and we added this to the data set right and as i said we just see data set right and as i said we just see data set right and as i said we just see all the data all these sequences of all the data all these sequences of all the data all these sequences of words or sequences of this agent words or sequences of this agent words or sequences of this agent interacting with that environment interacting with that environment interacting with that environment or you know agents playing atari and so or you know agents playing atari and so or you know agents playing atari and so on we see this as the same kind of data on we see this as the same kind of data on we see this as the same kind of data and so we mix these data sets together and so we mix these data sets together and so we mix these data sets together and we train gato and we train gato and we train gato that's the g part right it's general that's the g part right it's general that's the g part right it's general because it really has mixed it it because it really has mixed it it because it really has mixed it it doesn't have different brains for each doesn't have different brains for each doesn't have different brains for each modality or each narrow task it has a modality or each narrow task it has a modality or each narrow task it has a single brain it's not that big of a single brain it's not that big of a single brain it's not that big of a brain compared to most of the neural brain compared to most of the neural brain compared to most of the neural networks we see these days it has one networks we see these days it has one networks we see these days it has one billion parameters billion parameters billion parameters some models we're seeing get in the some models we're seeing get in the some models we're seeing get in the trillions these days and certainly 100 trillions these days and certainly 100 trillions these days and certainly 100 billion feels like um billion feels like um billion feels like um a size that is very common from from a size that is very common from from a size that is very common from from when you train this this job so the when you train this this job so the when you train this this job so the actual actual actual agent is relatively small but it's been agent is relatively small but it's been agent is relatively small but it's been trained on on a very challenging diverse trained on on a very challenging diverse trained on on a very challenging diverse data set not only containing all of data set not only containing all of data set not only containing all of internet but containing all these asian internet but containing all these asian internet but containing all these asian experience playing very different experience playing very different experience playing very different distinct environments distinct environments distinct environments so this so this so this brings us to the part of the tweet of brings us to the part of the tweet of brings us to the part of the tweet of this is not the end is the beginning it this is not the end is the beginning it this is not the end is the beginning it it feels very cool to see gato

it feels very cool to see gato it feels very cool to see gato in principle is able to control in principle is able to control in principle is able to control any sort of environments um that any sort of environments um that any sort of environments um that especially the ones that it's been especially the ones that it's been especially the ones that it's been trained to do these 3d games atari games trained to do these 3d games atari games trained to do these 3d games atari games and all sorts of robotics tasks and so and all sorts of robotics tasks and so and all sorts of robotics tasks and so on on on um but um but um but obviously it's not as proficient as the obviously it's not as proficient as the obviously it's not as proficient as the teachers it learned from on these teachers it learned from on these teachers it learned from on these environments not obvious environments not obvious environments not obvious it's not obvious that it wouldn't be it's not obvious that it wouldn't be it's not obvious that it wouldn't be more proficient more proficient more proficient it's just the current beginning part it's just the current beginning part it's just the current beginning part right is that right is that right is that the performance is such that it's not as the performance is such that it's not as the performance is such that it's not as good as if it's specialized to that task good as if it's specialized to that task good as if it's specialized to that task right so right so right so it's not as good although i would argue it's not as good although i would argue it's not as good although i would argue size matters here so the fact that i size matters here so the fact that i size matters here so the fact that i would argue always size always matter would argue always size always matter would argue always size always matter yeah that's a different color yeah that's a different color yeah that's a different color but but for neural networks certainly but but for neural networks certainly but but for neural networks certainly size does matter so um it's the size does matter so um it's the size does matter so um it's the beginning because it's relatively small beginning because it's relatively small beginning because it's relatively small so obviously scaling this idea up um so obviously scaling this idea up um so obviously scaling this idea up um might make might make might make the the the connections that exist between connections that exist between connections that exist between you know text on the internet and you know text on the internet and you know text on the internet and playing atari and so on more playing atari and so on more playing atari and so on more synergistic with one another and you synergistic with one another and you synergistic with one another and you might gain and that moment we didn't might gain and that moment we didn't might gain and that moment we didn't quite see but obviously that's why it's quite see but obviously that's why it's quite see but obviously that's why it's the beginning that synergy might emerge the beginning that synergy might emerge the beginning that synergy might emerge with scale right my emerge with scale with scale right my emerge with scale with scale right my emerge with scale and also i believe there's some new and also i believe there's some new and also i believe there's some new research or ways in which you prepare research or ways in which you prepare research or ways in which you prepare the data um that might you might need to the data um that might you might need to the data um that might you might need to sort of make it more clear to the model sort of make it more clear to the model sort of make it more clear to the model that that that you're not only playing atari and it's you're not only playing atari and it's you're not only playing atari and it's just you start from a screen and here is just you start from a screen and here is just you start from a screen and here is up and a screen and down maybe you can up and a screen and down maybe you can up and a screen and down maybe you can think of playing atari as there's some think of playing atari as there's some think of playing atari as there's some sort of context that is needed for the

sort of context that is needed for the sort of context that is needed for the agent before it starts seeing oh i'm agent before it starts seeing oh i'm agent before it starts seeing oh i'm this is an entire screen i'm going to this is an entire screen i'm going to this is an entire screen i'm going to start playing start playing start playing um you might require for instance to to um you might require for instance to to um you might require for instance to to be told in words be told in words be told in words hey this is the in this in this sequence hey this is the in this in this sequence hey this is the in this in this sequence that i'm showing you're going to be that i'm showing you're going to be that i'm showing you're going to be playing an entire game playing an entire game playing an entire game so text might actually be a good driver so text might actually be a good driver so text might actually be a good driver to to to enhance the data right so then these enhance the data right so then these enhance the data right so then these connections might be made more easily connections might be made more easily connections might be made more easily right that's that's an idea that we right that's that's an idea that we right that's that's an idea that we start seeing start seeing start seeing in language but you know obviously in language but you know obviously in language but you know obviously beyond this is going to be effective beyond this is going to be effective beyond this is going to be effective right it's not like i don't show you a right it's not like i don't show you a right it's not like i don't show you a screen and and you from from scratch you screen and and you from from scratch you screen and and you from from scratch you you're supposed to learn a game there is you're supposed to learn a game there is you're supposed to learn a game there is a lot of context we might set so there a lot of context we might set so there a lot of context we might set so there are there might be some work needed as are there might be some work needed as are there might be some work needed as well to set that context um but well to set that context um but well to set that context um but anyways there's a lot of work yeah so anyways there's a lot of work yeah so anyways there's a lot of work yeah so that context puts all the different that context puts all the different that context puts all the different modalities on the same level ground modalities on the same level ground modalities on the same level ground exactly provide the context best so exactly provide the context best so exactly provide the context best so maybe on that point uh so there's this maybe on that point uh so there's this maybe on that point uh so there's this task which task which task which may not seem trivial of may not seem trivial of may not seem trivial of tokenizing the data of converting the tokenizing the data of converting the tokenizing the data of converting the data into data into data into pieces into basic atomic elements pieces into basic atomic elements pieces into basic atomic elements that then could uh that then could uh that then could uh cross modalities somehow so what's cross modalities somehow so what's cross modalities somehow so what's tokenization tokenization tokenization how do you tokenize text how do you how do you tokenize text how do you how do you tokenize text how do you tokenize images how do you tokenize tokenize images how do you tokenize tokenize images how do you tokenize games and actions and robotics games and actions and robotics games and actions and robotics tasks yeah that's a great question so tasks yeah that's a great question so tasks yeah that's a great question so tokenization is tokenization is tokenization is the entry point to actually make all the

the entry point to actually make all the the entry point to actually make all the data look like a sequence because tokens data look like a sequence because tokens data look like a sequence because tokens then are just kind of these little then are just kind of these little then are just kind of these little puzzle pieces we break down anything puzzle pieces we break down anything puzzle pieces we break down anything into these puzzle pieces and then we into these puzzle pieces and then we into these puzzle pieces and then we just model what's the what's this puzzle just model what's the what's this puzzle just model what's the what's this puzzle look like right when you make it you look like right when you make it you look like right when you make it you know lay down in a line so to speak in a know lay down in a line so to speak in a know lay down in a line so to speak in a sequence sequence sequence so so so in gato um in gato um in gato um the text there's a lot of work you the text there's a lot of work you the text there's a lot of work you tokenize text usually by looking at tokenize text usually by looking at tokenize text usually by looking at common commonly used sub strings right common commonly used sub strings right common commonly used sub strings right so there's you know ing in english is a so there's you know ing in english is a so there's you know ing in english is a very common substring so that becomes a very common substring so that becomes a very common substring so that becomes a token um there's quite well studied token um there's quite well studied token um there's quite well studied problem on tokenizing text text and gato problem on tokenizing text text and gato problem on tokenizing text text and gato just used the standard techniques that just used the standard techniques that just used the standard techniques that have been developed from many years even have been developed from many years even have been developed from many years even starting from engram models in the 1950s starting from engram models in the 1950s starting from engram models in the 1950s and so on just for context how many and so on just for context how many and so on just for context how many tokens like what order magnitude number tokens like what order magnitude number tokens like what order magnitude number of tokens is required for a word of tokens is required for a word of tokens is required for a word yeah actually what are we talking about yeah actually what are we talking about yeah actually what are we talking about yeah for a word in in english right i yeah for a word in in english right i yeah for a word in in english right i mean every language is very different um mean every language is very different um mean every language is very different um the current level or granularity of the current level or granularity of the current level or granularity of tokenization generally tokenization generally tokenization generally means is maybe means is maybe means is maybe two to five i mean i i don't know the two to five i mean i i don't know the two to five i mean i i don't know the statistics exactly but to give you an statistics exactly but to give you an statistics exactly but to give you an idea um we don't tokenize at the level idea um we don't tokenize at the level idea um we don't tokenize at the level of letters then it would probably be of letters then it would probably be of letters then it would probably be like i don't know what the average like i don't know what the average like i don't know what the average length of of a word is in english but length of of a word is in english but length of of a word is in english but that would be you know the the minimum that would be you know the the minimum that would be you know the the minimum set of tokens you could use was bigger set of tokens you could use was bigger set of tokens you could use was bigger than letters smaller than words yes yes than letters smaller than words yes yes than letters smaller than words yes yes and you could think of very very common and you could think of very very common and you could think of very very common words like v i mean that would be a words like v i mean that would be a words like v i mean that would be a single token but very quickly you you're single token but very quickly you you're single token but very quickly you you're talking two three four four tokens have talking two three four four tokens have talking two three four four tokens have you ever tried to tokenize emojis emojis

you ever tried to tokenize emojis emojis you ever tried to tokenize emojis emojis are actually just are actually just are actually just um um um sequences of letters so maybe to you but sequences of letters so maybe to you but sequences of letters so maybe to you but to me they mean so much more yeah you to me they mean so much more yeah you to me they mean so much more yeah you can render the emoji but you you might can render the emoji but you you might can render the emoji but you you might if you actually just yeah this is a if you actually just yeah this is a if you actually just yeah this is a philosophical question is emojis an philosophical question is emojis an philosophical question is emojis an image or a text image or a text image or a text the way the way the way we do these things these things is we do these things these things is we do these things these things is they're actually mapped to seek small they're actually mapped to seek small they're actually mapped to seek small sequences of characters yeah so sequences of characters yeah so sequences of characters yeah so you can actually play with these models you can actually play with these models you can actually play with these models and input emojis it will output emojis and input emojis it will output emojis and input emojis it will output emojis back um which is actually quite a fun back um which is actually quite a fun back um which is actually quite a fun exercise you probably can find other exercise you probably can find other exercise you probably can find other tweets about this um out there um but tweets about this um out there um but tweets about this um out there um but yeah so anyways tex there's like it's yeah so anyways tex there's like it's yeah so anyways tex there's like it's very clear how this is done and then in very clear how this is done and then in very clear how this is done and then in gato gato gato what we did for images is we map images what we did for images is we map images what we did for images is we map images to essentially we compressed images so to essentially we compressed images so to essentially we compressed images so to speak into something that looks more to speak into something that looks more to speak into something that looks more like um like um like um less like every pixel with every less like every pixel with every less like every pixel with every intensity that would mean we have a very intensity that would mean we have a very intensity that would mean we have a very long sequence right like if we were long sequence right like if we were long sequence right like if we were talking about 100 by 100 pixel images talking about 100 by 100 pixel images talking about 100 by 100 pixel images that would make the sequences far too that would make the sequences far too that would make the sequences far too long so what was done there is you just long so what was done there is you just long so what was done there is you just use a technique that essentially use a technique that essentially use a technique that essentially compresses an image into maybe 16 by 16 compresses an image into maybe 16 by 16 compresses an image into maybe 16 by 16 patches of pixels and then that is map patches of pixels and then that is map patches of pixels and then that is map again tokenize you just essentially again tokenize you just essentially again tokenize you just essentially quantize this space into quantize this space into quantize this space into a special word that actually maps to a special word that actually maps to a special word that actually maps to these little sequence of pixels and then these little sequence of pixels and then these little sequence of pixels and then you put the pixels together in some you put the pixels together in some you put the pixels together in some raster order and then that's how you get

raster order and then that's how you get raster order and then that's how you get out um or in or in the image that your out um or in or in the image that your out um or in or in the image that your your your process but there's no your your process but there's no your your process but there's no semantic semantic semantic aspect to that so you're doing some kind aspect to that so you're doing some kind aspect to that so you're doing some kind of you don't need to understand anything of you don't need to understand anything of you don't need to understand anything about the image in order to tokenize it about the image in order to tokenize it about the image in order to tokenize it currently no you you're only using this currently no you you're only using this currently no you you're only using this notion of compression so you're trying notion of compression so you're trying notion of compression so you're trying to to to find common it's like jpg or all these find common it's like jpg or all these find common it's like jpg or all these algorithms it's actually very similar at algorithms it's actually very similar at algorithms it's actually very similar at the tokenization level all we're doing the tokenization level all we're doing the tokenization level all we're doing is finding common patterns and then is finding common patterns and then is finding common patterns and then making sure making sure making sure in a lossy way we compress these images in a lossy way we compress these images in a lossy way we compress these images given the statistics of the images that given the statistics of the images that given the statistics of the images that are contained in all the data we deal are contained in all the data we deal are contained in all the data we deal with although you could probably argue with although you could probably argue with although you could probably argue that jpeg that jpeg that jpeg does have some understanding of images does have some understanding of images does have some understanding of images like uh like uh like uh because visual information because visual information because visual information maybe color maybe color maybe color compressing based crudely based on color compressing based crudely based on color compressing based crudely based on color does capture some does capture some does capture some something important about an image something important about an image something important about an image that's about its meaning not just about that's about its meaning not just about that's about its meaning not just about some statistics some statistics some statistics yeah i mean jp as i said is very the yeah i mean jp as i said is very the yeah i mean jp as i said is very the algorithms look actually very similar to algorithms look actually very similar to algorithms look actually very similar to they use this the the they use this the the they use this the the cosine transform in jpg cosine transform in jpg cosine transform in jpg um the the approach we usually do in um the the approach we usually do in um the the approach we usually do in machine learning when we deal with machine learning when we deal with machine learning when we deal with images and we do this quantization step images and we do this quantization step images and we do this quantization step is a bit more data driven so rather than is a bit more data driven so rather than is a bit more data driven so rather than have some sort of fourier basis for how have some sort of fourier basis for how have some sort of fourier basis for how you know frequencies appear in natural you know frequencies appear in natural you know frequencies appear in natural in the natural world in the natural world in the natural world we actually just use we actually just use we actually just use the statistics of the images and then the statistics of the images and then the statistics of the images and then quantize them based on the statistics

quantize them based on the statistics quantize them based on the statistics much like you do in words right so much like you do in words right so much like you do in words right so common subscript sub strings are common subscript sub strings are common subscript sub strings are allocated a token um and images is very allocated a token um and images is very allocated a token um and images is very similar but there's no similar but there's no similar but there's no connection connection connection the token space if you think of oh like the token space if you think of oh like the token space if you think of oh like the tokens are an integer and in the end the tokens are an integer and in the end the tokens are an integer and in the end of the day so now like we work on of the day so now like we work on of the day so now like we work on maybe we have about let's say i don't maybe we have about let's say i don't maybe we have about let's say i don't know the exact numbers but let's say 10 know the exact numbers but let's say 10 know the exact numbers but let's say 10 000 tokens for text right certainly more 000 tokens for text right certainly more 000 tokens for text right certainly more than characters because we have groups than characters because we have groups than characters because we have groups of characters and so on so from one to of characters and so on so from one to of characters and so on so from one to ten thousand those are representing all ten thousand those are representing all ten thousand those are representing all the language and the words we'll see and the language and the words we'll see and the language and the words we'll see and then images occupy the next set of then images occupy the next set of then images occupy the next set of integers so they're completely integers so they're completely integers so they're completely independent right so from ten thousand independent right so from ten thousand independent right so from ten thousand one to twenty thousand those are the one to twenty thousand those are the one to twenty thousand those are the tokens that represent these other tokens that represent these other tokens that represent these other modality images modality images modality images and and and that is an interesting that is an interesting that is an interesting aspect that makes it orthogonal so what aspect that makes it orthogonal so what aspect that makes it orthogonal so what connects these concepts is the data connects these concepts is the data connects these concepts is the data right once you have a data set for right once you have a data set for right once you have a data set for instance that captions images that tells instance that captions images that tells instance that captions images that tells you oh this is someone playing a frisbee you oh this is someone playing a frisbee you oh this is someone playing a frisbee on on a green field now on on a green field now on on a green field now the model will need to predict the the model will need to predict the the model will need to predict the tokens from the text green field to then tokens from the text green field to then tokens from the text green field to then the pixels and that will start making the pixels and that will start making the pixels and that will start making the connections between the tokens so the connections between the tokens so the connections between the tokens so these connections happen as the these connections happen as the these connections happen as the algorithm learns and then the last if we algorithm learns and then the last if we algorithm learns and then the last if we think of these integers the first few think of these integers the first few think of these integers the first few are words the next few are images in are words the next few are images in are words the next few are images in gato we also allocated gato we also allocated gato we also allocated the the highest the the highest the the highest order of integers to actions right which order of integers to actions right which order of integers to actions right which we discretize and actions

we discretize and actions we discretize and actions are very diverse right in atari there's are very diverse right in atari there's are very diverse right in atari there's i don't know if 17 discrete actions in i don't know if 17 discrete actions in i don't know if 17 discrete actions in robotics um actions might be torques and robotics um actions might be torques and robotics um actions might be torques and forces that we apply so we just use kind forces that we apply so we just use kind forces that we apply so we just use kind of similar ideas to compress these of similar ideas to compress these of similar ideas to compress these actions into tokens and then actions into tokens and then actions into tokens and then we just we just we just that's how we map now all the space to that's how we map now all the space to that's how we map now all the space to this sequence of integers but they this sequence of integers but they this sequence of integers but they occupy different space and what connects occupy different space and what connects occupy different space and what connects them is then the learning algorithm them is then the learning algorithm them is then the learning algorithm that's where the magic happens so the that's where the magic happens so the that's where the magic happens so the modalities are modalities are modalities are orthogonal to each other in token space orthogonal to each other in token space orthogonal to each other in token space right so in the input right so in the input right so in the input everything you add you add extra tokens everything you add you add extra tokens everything you add you add extra tokens right and then right and then right and then you're shoving all of that into one you're shoving all of that into one you're shoving all of that into one place yes the transformer and that place yes the transformer and that place yes the transformer and that transformer that transformer transformer that transformer transformer that transformer tries tries tries to look at this gigantic token space and to look at this gigantic token space and to look at this gigantic token space and tries to form some kind of tries to form some kind of tries to form some kind of representation some kind of representation some kind of representation some kind of unique um unique um unique um wisdom wisdom wisdom about all of these different modalities about all of these different modalities about all of these different modalities how's that how's that how's that possible are they do if you were to sort possible are they do if you were to sort possible are they do if you were to sort of like put your of like put your of like put your psychoanalysis hat on and try to psychoanalysis hat on and try to psychoanalysis hat on and try to psychoanalyze this neural network psychoanalyze this neural network psychoanalyze this neural network is it schizophrenic is it schizophrenic is it schizophrenic does it try to given this very few does it try to given this very few does it try to given this very few weights weights weights represent multiple disjoint things and represent multiple disjoint things and represent multiple disjoint things and somehow somehow somehow have them not interfere with each other have them not interfere with each other have them not interfere with each other or is this about building on the or is this about building on the or is this about building on the um um um on the joint strength on whatever is

on the joint strength on whatever is on the joint strength on whatever is common to all the different modalities common to all the different modalities common to all the different modalities like what like what like what if you were to ask questions is it if you were to ask questions is it if you were to ask questions is it schizophrenic or is it uh does it is it schizophrenic or is it uh does it is it schizophrenic or is it uh does it is it of one mind of one mind of one mind i mean it is it is one mind um and it's i mean it is it is one mind um and it's i mean it is it is one mind um and it's actually the very the simplest algorithm actually the very the simplest algorithm actually the very the simplest algorithm which um that's kind of in a way how it which um that's kind of in a way how it which um that's kind of in a way how it feels like the field feels like the field feels like the field hasn't changed since back propagation hasn't changed since back propagation hasn't changed since back propagation and gradient descent was purpose for and gradient descent was purpose for and gradient descent was purpose for learning neural networks so learning neural networks so learning neural networks so there is obviously details on the there is obviously details on the there is obviously details on the architecture this has evolved the architecture this has evolved the architecture this has evolved the current iteration current iteration current iteration is still the transformer which is is still the transformer which is is still the transformer which is a powerful a powerful a powerful sequence modeling architecture but then sequence modeling architecture but then sequence modeling architecture but then the goal of this the goal of this the goal of this you know setting these weights to you know setting these weights to you know setting these weights to predict the data is essentially the same predict the data is essentially the same predict the data is essentially the same as basically i could describe i mean we as basically i could describe i mean we as basically i could describe i mean we described a few years ago alpha star described a few years ago alpha star described a few years ago alpha star language modeling and so on right we we language modeling and so on right we we language modeling and so on right we we take let's say an atari game um we map take let's say an atari game um we map take let's say an atari game um we map it to a string of numbers that will all it to a string of numbers that will all it to a string of numbers that will all be probably image space and action space be probably image space and action space be probably image space and action space interleaved and all we're gonna do is interleaved and all we're gonna do is interleaved and all we're gonna do is say okay say okay say okay given the numbers you know ten thousand given the numbers you know ten thousand given the numbers you know ten thousand one ten thousand four ten thousand five one ten thousand four ten thousand five one ten thousand four ten thousand five the next number that comes is twenty the next number that comes is twenty the next number that comes is twenty thousand six which is in the action thousand six which is in the action thousand six which is in the action space space space and you're just and you're just and you're just optimizing these weights be a very optimizing these weights be a very optimizing these weights be a very simple simple simple gradient like you know mathematical is gradient like you know mathematical is gradient like you know mathematical is almost the most boring algorithm you almost the most boring algorithm you almost the most boring algorithm you could imagine we settle the weights so could imagine we settle the weights so could imagine we settle the weights so that given this particular instance

that given this particular instance that given this particular instance these weights are set to maximize the these weights are set to maximize the these weights are set to maximize the probability of having seen this probability of having seen this probability of having seen this particular sequence of integers for this particular sequence of integers for this particular sequence of integers for this particular game particular game particular game and then and then and then the algorithm does this for many many the algorithm does this for many many the algorithm does this for many many many iterations um looking at different many iterations um looking at different many iterations um looking at different modalities different games right that's modalities different games right that's modalities different games right that's the mixture of the data set we discuss the mixture of the data set we discuss the mixture of the data set we discuss so in a way it's a very simple algorithm so in a way it's a very simple algorithm so in a way it's a very simple algorithm and and and the weights right they're all shared the weights right they're all shared the weights right they're all shared right so in terms of is it focusing on right so in terms of is it focusing on right so in terms of is it focusing on one modality or not the intermediate one modality or not the intermediate one modality or not the intermediate weights that are converting from these weights that are converting from these weights that are converting from these input of integers to the target integer input of integers to the target integer input of integers to the target integer you're predicting next those weights you're predicting next those weights you're predicting next those weights certainly are common and then the way certainly are common and then the way certainly are common and then the way the tokenization happens there is there the tokenization happens there is there the tokenization happens there is there is a special place in the neural network is a special place in the neural network is a special place in the neural network which is we map this integer like number which is we map this integer like number which is we map this integer like number 1001 to a vector of real numbers like 1001 to a vector of real numbers like 1001 to a vector of real numbers like real numbers um we can optimize them real numbers um we can optimize them real numbers um we can optimize them with gradient descent right the the with gradient descent right the the with gradient descent right the the functions we learn are actually um functions we learn are actually um functions we learn are actually um surprisingly differentiable that's why surprisingly differentiable that's why surprisingly differentiable that's why we compute gradients so this this step we compute gradients so this this step we compute gradients so this this step is the only one that this orthogonality is the only one that this orthogonality is the only one that this orthogonality you mentioned applies so you mentioned applies so you mentioned applies so mapping mapping mapping a certain token for text or image or a certain token for text or image or a certain token for text or image or actions this actions this actions this each of these tokens gets its own little each of these tokens gets its own little each of these tokens gets its own little vector of real numbers that represents vector of real numbers that represents vector of real numbers that represents this if you look at the field back many this if you look at the field back many this if you look at the field back many years ago people were talking about word years ago people were talking about word years ago people were talking about word vectors or word embeddings vectors or word embeddings vectors or word embeddings these are the same we have word vectors these are the same we have word vectors these are the same we have word vectors or embeddings we have image vector or or embeddings we have image vector or or embeddings we have image vector or embeddings and action vector of embeddings and action vector of embeddings and action vector of embeddings and the beauty here is that

embeddings and the beauty here is that embeddings and the beauty here is that as you train this model if you visualize as you train this model if you visualize as you train this model if you visualize these little vectors um it might be that these little vectors um it might be that these little vectors um it might be that they start aligning even though they start aligning even though they start aligning even though they're independent parameters there they're independent parameters there they're independent parameters there there could be anything but then it there could be anything but then it there could be anything but then it might be that you take the word gato or might be that you take the word gato or might be that you take the word gato or cat which maybe is common enough that cat which maybe is common enough that cat which maybe is common enough that actually has its own token and then you actually has its own token and then you actually has its own token and then you take pixels that have a cat and you take pixels that have a cat and you take pixels that have a cat and you might start seeing might start seeing might start seeing that these vectors look like they align that these vectors look like they align that these vectors look like they align right so by learning from this vast right so by learning from this vast right so by learning from this vast amount of data amount of data amount of data the model the model the model is realizing the potential connections is realizing the potential connections is realizing the potential connections between these modalities now i will say between these modalities now i will say between these modalities now i will say there would be another way at least in there would be another way at least in there would be another way at least in part to not have these part to not have these part to not have these different different different vectors for each different modality vectors for each different modality vectors for each different modality for instance when i tell you about for instance when i tell you about for instance when i tell you about actions in certain space actions in certain space actions in certain space i'm defining actions by words right so i'm defining actions by words right so i'm defining actions by words right so you could imagine a world in which i'm you could imagine a world in which i'm you could imagine a world in which i'm not learning not learning not learning that the action app in atari is its own that the action app in atari is its own that the action app in atari is its own number number number the action app in atari maybe is the action app in atari maybe is the action app in atari maybe is literally the word or the sentence app literally the word or the sentence app literally the word or the sentence app in atari right and that would mean we in atari right and that would mean we in atari right and that would mean we now leverage much more from the language now leverage much more from the language now leverage much more from the language this is not what we did here but this is not what we did here but this is not what we did here but certainly it might make these certainly it might make these certainly it might make these connections much easier to learn and connections much easier to learn and connections much easier to learn and also to teach the model to correct its also to teach the model to correct its also to teach the model to correct its own actions and so on right so all these own actions and so on right so all these own actions and so on right so all these to to say that gato is indeed the to to say that gato is indeed the to to say that gato is indeed the beginning that it is it is a radical beginning that it is it is a radical beginning that it is it is a radical idea to do this this way but there's idea to do this this way but there's idea to do this this way but there's probably a lot more to be done and the

probably a lot more to be done and the probably a lot more to be done and the results to be more impressive not only results to be more impressive not only results to be more impressive not only through scale but also through some through scale but also through some through scale but also through some new research that will come hopefully in new research that will come hopefully in new research that will come hopefully in the years to come so just to elaborate the years to come so just to elaborate the years to come so just to elaborate quickly you mean quickly you mean quickly you mean one possible one possible one possible next step next step next step or or or one of the paths that you might take one of the paths that you might take one of the paths that you might take next is doing the tokenization fundamentally as doing the tokenization fundamentally as a kind of uh a kind of uh a kind of uh linguistic communication so like you linguistic communication so like you linguistic communication so like you convert even images into language so convert even images into language so convert even images into language so doing something like a crude doing something like a crude doing something like a crude semantic segmentation semantic segmentation semantic segmentation trying to just assign a bunch of words trying to just assign a bunch of words trying to just assign a bunch of words to an image that to an image that to an image that like have like have like have almost like a dumb entity explaining as almost like a dumb entity explaining as almost like a dumb entity explaining as much as you can about the the image and much as you can about the the image and much as you can about the the image and so you convert that into words and then so you convert that into words and then so you convert that into words and then you convert games into words and and you you convert games into words and and you you convert games into words and and you provide the context and words and all of provide the context and words and all of provide the context and words and all of it it it and eventually and eventually and eventually getting to a point where everybody getting to a point where everybody getting to a point where everybody agrees with noam chomsky that language agrees with noam chomsky that language agrees with noam chomsky that language is actually at the core of everything is actually at the core of everything is actually at the core of everything that's it's the base layer of that's it's the base layer of that's it's the base layer of intelligence and consciousness and all intelligence and consciousness and all intelligence and consciousness and all that kind of stuff okay that kind of stuff okay that kind of stuff okay uh you mentioned early on like psy it's uh you mentioned early on like psy it's uh you mentioned early on like psy it's hard to grow what did you mean by that hard to grow what did you mean by that hard to grow what did you mean by that because we're talking about scale might because we're talking about scale might because we're talking about scale might change change change uh there might be and we'll talk about uh there might be and we'll talk about uh there might be and we'll talk about this too like there's a this too like there's a this too like there's a emergent emergent emergent there's certain things about these there's certain things about these there's certain things about these neural networks that are emerging so neural networks that are emerging so neural networks that are emerging so certain like performance we can see only certain like performance we can see only certain like performance we can see only with scale and there's some kind of with scale and there's some kind of with scale and there's some kind of threshold of scale so it threshold of scale so it threshold of scale so it why is it hard to grow something like

why is it hard to grow something like why is it hard to grow something like this meow network this meow network this meow network so so so the meow network the meow network the meow network is is not it's not hard to grow if you is is not it's not hard to grow if you is is not it's not hard to grow if you retrain it yeah what's hard is well we retrain it yeah what's hard is well we retrain it yeah what's hard is well we have now one billion parameters um we have now one billion parameters um we have now one billion parameters um we train them for a while we we spend some train them for a while we we spend some train them for a while we we spend some amount of work towards building these amount of work towards building these amount of work towards building these these weights that are an amazing these weights that are an amazing these weights that are an amazing initial brain for doing this kind of initial brain for doing this kind of initial brain for doing this kind of tasks we care about tasks we care about tasks we care about could we reuse the weights could we reuse the weights could we reuse the weights and expand to a larger brain and that is and expand to a larger brain and that is and expand to a larger brain and that is extraordinarily hard but also extraordinarily hard but also extraordinarily hard but also exciting from a research perspective and exciting from a research perspective and exciting from a research perspective and a practical perspective point of view a practical perspective point of view a practical perspective point of view right so right so right so there's this notion of there's this notion of there's this notion of modularity in software engineering and modularity in software engineering and modularity in software engineering and we're starting to see some examples and we're starting to see some examples and we're starting to see some examples and work that leverages modularity in fact work that leverages modularity in fact work that leverages modularity in fact if we go back one step from gato to a if we go back one step from gato to a if we go back one step from gato to a work that i would say train much larger work that i would say train much larger work that i would say train much larger much more capable network called much more capable network called much more capable network called flamingo flamingo did not deal with flamingo flamingo did not deal with flamingo flamingo did not deal with actions but it definitely dealt with actions but it definitely dealt with actions but it definitely dealt with images in in a in an interesting way images in in a in an interesting way images in in a in an interesting way kind of akin to what agato did but kind of akin to what agato did but kind of akin to what agato did but slightly different technique for slightly different technique for slightly different technique for tokenizing but we don't need to go into tokenizing but we don't need to go into tokenizing but we don't need to go into that detail but that detail but that detail but what flamingo also did which gato didn't what flamingo also did which gato didn't what flamingo also did which gato didn't do and that just happens because these do and that just happens because these do and that just happens because these projects you know they're they're projects you know they're they're projects you know they're they're they're different you know it's a bit of they're different you know it's a bit of they're different you know it's a bit of like the exploratory nature of research like the exploratory nature of research like the exploratory nature of research which is great the research behind these which is great the research behind these which is great the research behind these projects is also modular yes exactly um projects is also modular yes exactly um projects is also modular yes exactly um and it has to be right we need we need and it has to be right we need we need and it has to be right we need we need to have creativity um and sometimes you

to have creativity um and sometimes you to have creativity um and sometimes you need to protect pockets of you know need to protect pockets of you know need to protect pockets of you know people researchers and so on but we people researchers and so on but we people researchers and so on but we believe in humans yes okay and also in believe in humans yes okay and also in believe in humans yes okay and also in particular researchers and maybe even particular researchers and maybe even particular researchers and maybe even further you know deep mine or or other further you know deep mine or or other further you know deep mine or or other such labs and then they act the neural such labs and then they act the neural such labs and then they act the neural networks themselves so it's modularity networks themselves so it's modularity networks themselves so it's modularity all the way down okay all the way down all the way down okay all the way down all the way down okay all the way down so the way that we did modularity very so the way that we did modularity very so the way that we did modularity very beautifully in flamingo is we took beautifully in flamingo is we took beautifully in flamingo is we took chinchilla which is a language only chinchilla which is a language only chinchilla which is a language only model not an agent if we think of model not an agent if we think of model not an agent if we think of actions being necessary for agency so we actions being necessary for agency so we actions being necessary for agency so we took chinchilla we took the weights of took chinchilla we took the weights of took chinchilla we took the weights of chinchilla chinchilla chinchilla and then we froze them we said these and then we froze them we said these and then we froze them we said these don't change we train them to be very don't change we train them to be very don't change we train them to be very good at predicting the next word is a good at predicting the next word is a good at predicting the next word is a very good language model state of the very good language model state of the very good language model state of the art at the time you release it etc etc art at the time you release it etc etc art at the time you release it etc etc going to add a capability to c right we going to add a capability to c right we going to add a capability to c right we are going to add the ability to see to are going to add the ability to see to are going to add the ability to see to this language model so we're going to this language model so we're going to this language model so we're going to attach attach attach um small pieces of neural networks at um small pieces of neural networks at um small pieces of neural networks at the right places in the model it's the right places in the model it's the right places in the model it's almost like almost like almost like injecting injecting injecting the network with some weights and some the network with some weights and some the network with some weights and some substructures substructures substructures in the ways in a good way right so you in the ways in a good way right so you in the ways in a good way right so you need the research to say what is need the research to say what is need the research to say what is effective how do you add this capability effective how do you add this capability effective how do you add this capability without destroying others etc so we without destroying others etc so we without destroying others etc so we created a small sub network created a small sub network created a small sub network initialized not from random but actually initialized not from random but actually initialized not from random but actually from um self-supervised learning that from um self-supervised learning that from um self-supervised learning that you know a model that understands vision you know a model that understands vision you know a model that understands vision um in general and then um in general and then um in general and then we took data sets that connect the two we took data sets that connect the two we took data sets that connect the two modalities vision and language and then

modalities vision and language and then modalities vision and language and then we froze the main part the largest we froze the main part the largest we froze the main part the largest portion of the network which was portion of the network which was portion of the network which was chinchilla that is 70 billion parameters chinchilla that is 70 billion parameters chinchilla that is 70 billion parameters and then we added a few more parameters and then we added a few more parameters and then we added a few more parameters on top train from scratch on top train from scratch on top train from scratch and then some others that were and then some others that were and then some others that were pre-trained from like from with the pre-trained from like from with the pre-trained from like from with the capacity to see like it was a it was not capacity to see like it was a it was not capacity to see like it was a it was not tokenization in the way i described tokenization in the way i described tokenization in the way i described forgato but it's a similar idea forgato but it's a similar idea forgato but it's a similar idea and then we train the whole system parts and then we train the whole system parts and then we train the whole system parts of it were frozen parts of it were new of it were frozen parts of it were new of it were frozen parts of it were new and all of a sudden we developed and all of a sudden we developed and all of a sudden we developed flamingo which is an amazing model that flamingo which is an amazing model that flamingo which is an amazing model that is essentially i mean describing it is is essentially i mean describing it is is essentially i mean describing it is a chat bot where you can also upload a chat bot where you can also upload a chat bot where you can also upload images and start conversing about images images and start conversing about images images and start conversing about images um but it's also kind of a dialogue um but it's also kind of a dialogue um but it's also kind of a dialogue style um style um style um uh chatbot so the input is images and uh chatbot so the input is images and uh chatbot so the input is images and text and the output is text exactly text and the output is text exactly text and the output is text exactly um and how many parameters you said 70 um and how many parameters you said 70 um and how many parameters you said 70 billion 70 billion for chinchilla yeah billion 70 billion for chinchilla yeah billion 70 billion for chinchilla yeah chinchilla is 70 billion and then the chinchilla is 70 billion and then the chinchilla is 70 billion and then the ones we add on top which kind of almost ones we add on top which kind of almost ones we add on top which kind of almost is almost like um a way to overwrite its is almost like um a way to overwrite its is almost like um a way to overwrite its its little activations so that when it its little activations so that when it its little activations so that when it sees vision it does kind of a correct sees vision it does kind of a correct sees vision it does kind of a correct computation of what it's seeing mapping computation of what it's seeing mapping computation of what it's seeing mapping it back towards so to speak um that adds it back towards so to speak um that adds it back towards so to speak um that adds an extra 10 billion parameters right so an extra 10 billion parameters right so an extra 10 billion parameters right so it's total 80 billion the largest one we it's total 80 billion the largest one we it's total 80 billion the largest one we released and then released and then released and then you train it on you train it on you train it on a few data sets that contain vision and a few data sets that contain vision and a few data sets that contain vision and language and once you interact with the language and once you interact with the language and once you interact with the model you start seeing that you can model you start seeing that you can model you start seeing that you can upload an image and start sort of having upload an image and start sort of having upload an image and start sort of having a dialogue about the image um which is a dialogue about the image um which is a dialogue about the image um which is actually not something it's it's very

actually not something it's it's very actually not something it's it's very similar and akin to what we saw in similar and akin to what we saw in similar and akin to what we saw in language only this prompting abilities language only this prompting abilities language only this prompting abilities that it has you can teach it a new a new that it has you can teach it a new a new that it has you can teach it a new a new vision task right it does things beyond vision task right it does things beyond vision task right it does things beyond the capabilities that in theory the data the capabilities that in theory the data the capabilities that in theory the data sets um provided in themselves but sets um provided in themselves but sets um provided in themselves but because it leverages a lot of the because it leverages a lot of the because it leverages a lot of the language knowledge acquired from language knowledge acquired from language knowledge acquired from chinchilla it actually has this few shot chinchilla it actually has this few shot chinchilla it actually has this few shot learning ability and these emerging learning ability and these emerging learning ability and these emerging abilities that we didn't even measure abilities that we didn't even measure abilities that we didn't even measure once we were developing the model but once we were developing the model but once we were developing the model but once developed then once developed then once developed then as you play with the interface you can as you play with the interface you can as you play with the interface you can start seeing wow okay yeah it's cool we start seeing wow okay yeah it's cool we start seeing wow okay yeah it's cool we can we can upload i think one of the can we can upload i think one of the can we can upload i think one of the tweets talking about twitter was this tweets talking about twitter was this tweets talking about twitter was this image from obama that is image from obama that is image from obama that is placing a weight and and someone is kind placing a weight and and someone is kind placing a weight and and someone is kind of waiting themselves and and it's kind of waiting themselves and and it's kind of waiting themselves and and it's kind of a joke style image and it's notable of a joke style image and it's notable of a joke style image and it's notable because i think andriy carpati a few because i think andriy carpati a few because i think andriy carpati a few years ago said years ago said years ago said no computer vision system can can no computer vision system can can no computer vision system can can understand the subtlety of this joke in understand the subtlety of this joke in understand the subtlety of this joke in this image all the things that go on and this image all the things that go on and this image all the things that go on and so what we try to do and it's very so what we try to do and it's very so what we try to do and it's very anecdotally i mean this is not a proof anecdotally i mean this is not a proof anecdotally i mean this is not a proof that we solved this issue but that we solved this issue but that we solved this issue but it just shows that you can upload now it just shows that you can upload now it just shows that you can upload now this image and start conversing with the this image and start conversing with the this image and start conversing with the model trying to make out if it if it model trying to make out if it if it model trying to make out if it if it gets that there's a joke um because the gets that there's a joke um because the gets that there's a joke um because the person waiting themselves don't see that person waiting themselves don't see that person waiting themselves don't see that doesn't see that someone behind is doesn't see that someone behind is doesn't see that someone behind is making the weight higher and so on and making the weight higher and so on and making the weight higher and so on and so forth so it's a fascinating so forth so it's a fascinating so forth so it's a fascinating capability capability capability um and it comes from this key idea of um and it comes from this key idea of um and it comes from this key idea of modularity where we took a frozen brain modularity where we took a frozen brain modularity where we took a frozen brain and we just and we just and we just added a new capability so added a new capability so added a new capability so the question is the question is the question is should we so in a way you can see even

should we so in a way you can see even should we so in a way you can see even from deepmind we have flamingo that this from deepmind we have flamingo that this from deepmind we have flamingo that this this moderate approach um and thus could this moderate approach um and thus could this moderate approach um and thus could leverage the scale a bit more reasonably leverage the scale a bit more reasonably leverage the scale a bit more reasonably because we didn't need to retrain a because we didn't need to retrain a because we didn't need to retrain a system from scratch and the other on the system from scratch and the other on the system from scratch and the other on the other hand we had gato which used the other hand we had gato which used the other hand we had gato which used the same data sets but then it trained it same data sets but then it trained it same data sets but then it trained it from scratch right and so i guess from scratch right and so i guess from scratch right and so i guess big question for the community is big question for the community is big question for the community is should we train from scratch or should should we train from scratch or should should we train from scratch or should we embrace modularity and this lies like we embrace modularity and this lies like we embrace modularity and this lies like this goes back to this goes back to this goes back to modularity as a way to grow but reuse modularity as a way to grow but reuse modularity as a way to grow but reuse seems like natural and it was very seems like natural and it was very seems like natural and it was very effective certainly the next question is effective certainly the next question is effective certainly the next question is if you go the way of modularity if you go the way of modularity if you go the way of modularity is there a systematic way is there a systematic way is there a systematic way of freezing weights and joining of freezing weights and joining of freezing weights and joining different modalities different modalities different modalities across across across you know not just two or three or four you know not just two or three or four you know not just two or three or four networks but hundreds of networks from networks but hundreds of networks from networks but hundreds of networks from all different kinds of places maybe open all different kinds of places maybe open all different kinds of places maybe open source network that looks at weather source network that looks at weather source network that looks at weather patterns patterns patterns and you shove that in somehow and then and you shove that in somehow and then and you shove that in somehow and then you have networks that uh i don't know you have networks that uh i don't know you have networks that uh i don't know do all kinds of to play starcraft and do all kinds of to play starcraft and do all kinds of to play starcraft and play all the other video games and they play all the other video games and they play all the other video games and they you can keep adding them in you can keep adding them in you can keep adding them in without significant effort without significant effort without significant effort like that maybe the effort scales like that maybe the effort scales like that maybe the effort scales linearly or something like that as linearly or something like that as linearly or something like that as opposed to like the more network you add opposed to like the more network you add opposed to like the more network you add the more you have to worry about the the more you have to worry about the the more you have to worry about the instabilities created yeah so that that instabilities created yeah so that that instabilities created yeah so that that vision is beautiful i think vision is beautiful i think vision is beautiful i think um there's still the question about um there's still the question about um there's still the question about within single modalities like chinchilla within single modalities like chinchilla within single modalities like chinchilla was reused but now if we train a next was reused but now if we train a next was reused but now if we train a next iteration of language models are we iteration of language models are we iteration of language models are we going to use chinchilla or not yeah how going to use chinchilla or not yeah how going to use chinchilla or not yeah how do you swap out chinch right so

do you swap out chinch right so do you swap out chinch right so there's there's still big questions but there's there's still big questions but there's there's still big questions but that idea is is actually really akin to that idea is is actually really akin to that idea is is actually really akin to software engineering which we're not software engineering which we're not software engineering which we're not re-implementing you know libraries from re-implementing you know libraries from re-implementing you know libraries from scratch we're reusing and then building scratch we're reusing and then building scratch we're reusing and then building ever more amazing things including ever more amazing things including ever more amazing things including neural networks with software that we're neural networks with software that we're neural networks with software that we're using so i think this idea of modularity using so i think this idea of modularity using so i think this idea of modularity i like it i think it's here to stay and i like it i think it's here to stay and i like it i think it's here to stay and that's also why i mentioned it's just that's also why i mentioned it's just that's also why i mentioned it's just the beginning not the end the beginning not the end the beginning not the end you mentioned metal learning so given you mentioned metal learning so given you mentioned metal learning so given this promise of gatto this promise of gatto this promise of gatto can we try to redefine this term can we try to redefine this term can we try to redefine this term that's almost akin to consciousness that's almost akin to consciousness that's almost akin to consciousness because it means different things to because it means different things to because it means different things to different people throughout the history different people throughout the history different people throughout the history of artificial intelligence but what do of artificial intelligence but what do of artificial intelligence but what do you think meta-learning you think meta-learning you think meta-learning is and looks like now in the five years is and looks like now in the five years is and looks like now in the five years 10 years will it look like system i 10 years will it look like system i 10 years will it look like system i gotta but scaled gotta but scaled gotta but scaled what's your sense of what is what's your sense of what is what's your sense of what is what what does meta learning look like what what does meta learning look like what what does meta learning look like do you think great with all the wisdom do you think great with all the wisdom do you think great with all the wisdom we've learned so far yeah great great we've learned so far yeah great great we've learned so far yeah great great question maybe it's good to give question maybe it's good to give question maybe it's good to give another data point looking backwards another data point looking backwards another data point looking backwards rather than forward so rather than forward so rather than forward so when when we talk when when we talk when when we talk um in 2019 um in 2019 um in 2019 uh uh uh meta learning meta learning meta learning meant something that has changed mostly meant something that has changed mostly meant something that has changed mostly through the through the through the revolution of gpt3 and beyond so what revolution of gpt3 and beyond so what revolution of gpt3 and beyond so what meta-learning meant at the time um meta-learning meant at the time um meta-learning meant at the time um was driven by what benchmarks people was driven by what benchmarks people was driven by what benchmarks people care about in metal learning and the care about in metal learning and the care about in metal learning and the benchmarks were about benchmarks were about benchmarks were about a capability to learn about object a capability to learn about object a capability to learn about object identities so it was very much over

identities so it was very much over identities so it was very much over fitted to vision and object fitted to vision and object fitted to vision and object classification and the part that was met classification and the part that was met classification and the part that was met about that was that oh we're not just about that was that oh we're not just about that was that oh we're not just learning a thousand categories that learning a thousand categories that learning a thousand categories that imagenet tells us to learn we're gonna imagenet tells us to learn we're gonna imagenet tells us to learn we're gonna learn learn learn object categories that can be defined object categories that can be defined object categories that can be defined when we interact with the model so when we interact with the model so when we interact with the model so it's interesting to see the evolution it's interesting to see the evolution it's interesting to see the evolution right the way the way this started was right the way the way this started was right the way the way this started was we have a special language that was a we have a special language that was a we have a special language that was a data set a small data set that we data set a small data set that we data set a small data set that we prompted the model with saying hey here prompted the model with saying hey here prompted the model with saying hey here is a new classification task is a new classification task is a new classification task i'll give you one image and the name i'll give you one image and the name i'll give you one image and the name which was an integer at the time of the which was an integer at the time of the which was an integer at the time of the image and a different image and so on so image and a different image and so on so image and a different image and so on so you have a small prompt in the form of a you have a small prompt in the form of a you have a small prompt in the form of a data set a machine learning data set and data set a machine learning data set and data set a machine learning data set and then you got then a system that could then you got then a system that could then you got then a system that could then predict or classify these objects then predict or classify these objects then predict or classify these objects that you just defined kind of on the fly that you just defined kind of on the fly that you just defined kind of on the fly so so so fast forward fast forward fast forward it was it was it was revealed that revealed that revealed that language models are future learners language models are future learners language models are future learners that's the title of the paper so very that's the title of the paper so very that's the title of the paper so very good title sometimes titles are really good title sometimes titles are really good title sometimes titles are really good so this one is really really good good so this one is really really good good so this one is really really good because that's that's the point of gpt3 because that's that's the point of gpt3 because that's that's the point of gpt3 that showed that look that showed that look that showed that look sure we can we can focus on object sure we can we can focus on object sure we can we can focus on object classification and how what meta classification and how what meta classification and how what meta learning means within the space of learning means within the space of learning means within the space of learning object categories this goes learning object categories this goes learning object categories this goes beyond or before rather to also omniglot beyond or before rather to also omniglot beyond or before rather to also omniglot before imagenet and so on so there's a before imagenet and so on so there's a before imagenet and so on so there's a few benchmarks to now all of a sudden few benchmarks to now all of a sudden few benchmarks to now all of a sudden we're a bit unlocked from benchmarks and we're a bit unlocked from benchmarks and we're a bit unlocked from benchmarks and through language we can define tasks

through language we can define tasks through language we can define tasks right so we're literally telling the right so we're literally telling the right so we're literally telling the model some logical task or little thing model some logical task or little thing model some logical task or little thing that we wanted to do that we wanted to do that we wanted to do we prompted much like we did before but we prompted much like we did before but we prompted much like we did before but now we prompt it through natural now we prompt it through natural now we prompt it through natural language and then language and then language and then not perfectly i mean these models have not perfectly i mean these models have not perfectly i mean these models have failure modes and that's fine but failure modes and that's fine but failure modes and that's fine but no but these models then are now doing a no but these models then are now doing a no but these models then are now doing a new task right so they met to learn um new task right so they met to learn um new task right so they met to learn um these new capabilities now these new capabilities now these new capabilities now now that's where we are now uh flamingo now that's where we are now uh flamingo now that's where we are now uh flamingo expanded this to visual and language but expanded this to visual and language but expanded this to visual and language but it basically has the same abilities you it basically has the same abilities you it basically has the same abilities you can teach it for instance can teach it for instance can teach it for instance an emergent property was that you can an emergent property was that you can an emergent property was that you can take pictures of numbers and then do do take pictures of numbers and then do do take pictures of numbers and then do do arithmetic with the numbers just by arithmetic with the numbers just by arithmetic with the numbers just by teaching it oh that's i mean when when i teaching it oh that's i mean when when i teaching it oh that's i mean when when i show you three plus six you know i want show you three plus six you know i want show you three plus six you know i want you to output nine and and you show it a you to output nine and and you show it a you to output nine and and you show it a few examples and now it does that so it few examples and now it does that so it few examples and now it does that so it went way beyond the oh this image net went way beyond the oh this image net went way beyond the oh this image net sort of category categorization of sort of category categorization of sort of category categorization of images that we were a bit stuck maybe images that we were a bit stuck maybe images that we were a bit stuck maybe before um this before um this before um this revelation moment that happened uh in revelation moment that happened uh in revelation moment that happened uh in 2000 i believe it was 19 but it was 2000 i believe it was 19 but it was 2000 i believe it was 19 but it was after we chat and that way it has solved after we chat and that way it has solved after we chat and that way it has solved metal learning as was previously defined metal learning as was previously defined metal learning as was previously defined yes it expanded what it meant so that's yes it expanded what it meant so that's yes it expanded what it meant so that's what you say what does it mean so it's what you say what does it mean so it's what you say what does it mean so it's an evolving term um but here is an evolving term um but here is an evolving term um but here is maybe now looking forward looking at maybe now looking forward looking at maybe now looking forward looking at what's happening um you know obviously what's happening um you know obviously what's happening um you know obviously in the community with more modalities um in the community with more modalities um in the community with more modalities um what we can expect and i would certainly what we can expect and i would certainly what we can expect and i would certainly hope to see the following and this is a hope to see the following and this is a hope to see the following and this is a pretty drastic pretty drastic pretty drastic hope but in five years maybe we chat

hope but in five years maybe we chat hope but in five years maybe we chat again again again and and and we have a system right a set of weights we have a system right a set of weights we have a system right a set of weights that that that we can teach it to play starcraft we can teach it to play starcraft we can teach it to play starcraft maybe not at the level of alpha star but maybe not at the level of alpha star but maybe not at the level of alpha star but play starcraft a complex game we teach play starcraft a complex game we teach play starcraft a complex game we teach it through interactions to prompting you it through interactions to prompting you it through interactions to prompting you can certainly prompt a system that's can certainly prompt a system that's can certainly prompt a system that's what gato shows to play some simple what gato shows to play some simple what gato shows to play some simple atari games so imagine if you start atari games so imagine if you start atari games so imagine if you start talking to a system teaching it a new talking to a system teaching it a new talking to a system teaching it a new game showing it examples of you know in game showing it examples of you know in game showing it examples of you know in this in this particular game this in this particular game this in this particular game this user did something good maybe the this user did something good maybe the this user did something good maybe the system can even play and ask you system can even play and ask you system can even play and ask you questions say hey i played this game i questions say hey i played this game i questions say hey i played this game i just played this game did i do well can just played this game did i do well can just played this game did i do well can you teach me more so you teach me more so you teach me more so five maybe to ten years these five maybe to ten years these five maybe to ten years these capabilities capabilities capabilities or what meta learning means will be much or what meta learning means will be much or what meta learning means will be much more interactive much more rich and more interactive much more rich and more interactive much more rich and through domains that we were through domains that we were through domains that we were specializing right so you see the specializing right so you see the specializing right so you see the difference right we built alpha star difference right we built alpha star difference right we built alpha star specialized specialized specialized to play starcraft the algorithms were to play starcraft the algorithms were to play starcraft the algorithms were general but the weights were specialized general but the weights were specialized general but the weights were specialized and what what we're hoping is that we and what what we're hoping is that we and what what we're hoping is that we can teach a network to play games to can teach a network to play games to can teach a network to play games to play any game just using games as an play any game just using games as an play any game just using games as an example through interacting with it example through interacting with it example through interacting with it teaching it uploading the wikipedia page teaching it uploading the wikipedia page teaching it uploading the wikipedia page of starcraft like this is of starcraft like this is of starcraft like this is in the horizon and obviously their in the horizon and obviously their in the horizon and obviously their details need to be to be filled and details need to be to be filled and details need to be to be filled and research need to be done but that's how research need to be done but that's how research need to be done but that's how i see metal learning above which is i see metal learning above which is i see metal learning above which is gonna be beyond prompting it's gonna be gonna be beyond prompting it's gonna be gonna be beyond prompting it's gonna be a bit more interactive it's gonna you a bit more interactive it's gonna you a bit more interactive it's gonna you know the system might tell us to give it know the system might tell us to give it know the system might tell us to give it feedback after it maybe makes mistakes

feedback after it maybe makes mistakes feedback after it maybe makes mistakes or it loses a game um but it's or it loses a game um but it's or it loses a game um but it's nonetheless very exciting because if you nonetheless very exciting because if you nonetheless very exciting because if you think about this this way the benchmarks think about this this way the benchmarks think about this this way the benchmarks are already there we just repurpose them are already there we just repurpose them are already there we just repurpose them the benchmarks right so in a way the benchmarks right so in a way the benchmarks right so in a way i like to map the space of i like to map the space of i like to map the space of what maybe agi means to say okay like what maybe agi means to say okay like what maybe agi means to say okay like we went 101 performance we went 101 performance we went 101 performance in go in chess in starcraft the next in go in chess in starcraft the next in go in chess in starcraft the next iteration might be iteration might be iteration might be 20 performance across quote unquote all 20 performance across quote unquote all 20 performance across quote unquote all tasks tasks tasks right and even if it's not as good it's right and even if it's not as good it's right and even if it's not as good it's fine we we actually we have ways to also fine we we actually we have ways to also fine we we actually we have ways to also measure progress because we have those measure progress because we have those measure progress because we have those special agents specialized agents um and special agents specialized agents um and special agents specialized agents um and so on so this is to me very exciting and so on so this is to me very exciting and so on so this is to me very exciting and these next iteration models are these next iteration models are these next iteration models are definitely hinting at that direction of definitely hinting at that direction of definitely hinting at that direction of progress um which hopefully we can have progress um which hopefully we can have progress um which hopefully we can have there are obviously some things that there are obviously some things that there are obviously some things that could go wrong in terms of we might not could go wrong in terms of we might not could go wrong in terms of we might not have the tools maybe transformers are have the tools maybe transformers are have the tools maybe transformers are not enough then we must there's some not enough then we must there's some not enough then we must there's some breakthroughs to come which makes the breakthroughs to come which makes the breakthroughs to come which makes the field more exciting to people like me as field more exciting to people like me as field more exciting to people like me as well of course well of course well of course but that's if i if you ask me five to but that's if i if you ask me five to but that's if i if you ask me five to ten years you might see these models ten years you might see these models ten years you might see these models that start to look more like that start to look more like that start to look more like weights that are already trained and weights that are already trained and weights that are already trained and then then then it's more about teaching are or make it's more about teaching are or make it's more about teaching are or make they're meant to learn what you're they're meant to learn what you're they're meant to learn what you're you're trying um you're trying um you're trying um uh you're trying to to induce in terms uh you're trying to to induce in terms uh you're trying to to induce in terms of tasks and so on well beyond the of tasks and so on well beyond the of tasks and so on well beyond the simple now tasks we're starting to see simple now tasks we're starting to see simple now tasks we're starting to see emerge like you know small arithmetic

emerge like you know small arithmetic emerge like you know small arithmetic tasks and so on so a few questions tasks and so on so a few questions tasks and so on so a few questions around that this is fascinating uh so around that this is fascinating uh so around that this is fascinating uh so that kind of teaching interactive not so that kind of teaching interactive not so that kind of teaching interactive not so it's beyond prompting says interacting it's beyond prompting says interacting it's beyond prompting says interacting with the neural network with the neural network with the neural network that's different than the training that's different than the training that's different than the training process process process so it's different than the optimization over differentiable optimization over differentiable uh functions this is already trained and uh functions this is already trained and uh functions this is already trained and now you're teaching now you're teaching now you're teaching i mean um i mean um i mean um it's almost like akin to the brain then it's almost like akin to the brain then it's almost like akin to the brain then the the neurons already set with their the the neurons already set with their the the neurons already set with their connections on top of that you know connections on top of that you know connections on top of that you know using that infrastructure to build up using that infrastructure to build up using that infrastructure to build up further knowledge further knowledge further knowledge okay okay okay so so so that's a really interesting distinction that's a really interesting distinction that's a really interesting distinction that's actually not obvious from a that's actually not obvious from a that's actually not obvious from a software engineering perspective that software engineering perspective that software engineering perspective that there's a line to be drawn there's a line to be drawn there's a line to be drawn because you always think for a neural because you always think for a neural because you always think for a neural network to learn it has to be retrained network to learn it has to be retrained network to learn it has to be retrained trained and retrained trained and retrained trained and retrained but maybe but maybe but maybe and prompting is a way of and prompting is a way of and prompting is a way of teaching and you'll now work a little teaching and you'll now work a little teaching and you'll now work a little bit of context about whatever the heck bit of context about whatever the heck bit of context about whatever the heck you're trying it to do so you can maybe you're trying it to do so you can maybe you're trying it to do so you can maybe expand this prompting capability by expand this prompting capability by expand this prompting capability by um making it interact that's really um making it interact that's really um making it interact that's really really yeah by the way this is not if really yeah by the way this is not if really yeah by the way this is not if you look at way back um at different you look at way back um at different you look at way back um at different ways to tackle even classification tasks ways to tackle even classification tasks ways to tackle even classification tasks so this this is this comes from from so this this is this comes from from so this this is this comes from from like long-standing literature in machine like long-standing literature in machine like long-standing literature in machine learning um what i'm suggesting could learning um what i'm suggesting could learning um what i'm suggesting could sound to some like a bit like um nearest

sound to some like a bit like um nearest sound to some like a bit like um nearest neighbor so nida's neighbor is almost neighbor so nida's neighbor is almost neighbor so nida's neighbor is almost the simplest algorithm the simplest algorithm the simplest algorithm uh that you can uh that you can uh that you can that does not require learning so it has that does not require learning so it has that does not require learning so it has this interesting like you don't need to this interesting like you don't need to this interesting like you don't need to compute gradients and what nearest compute gradients and what nearest compute gradients and what nearest neighbor does is you quote unquote have neighbor does is you quote unquote have neighbor does is you quote unquote have a data set or upload a data set and then a data set or upload a data set and then a data set or upload a data set and then all you need to do is a way to measure all you need to do is a way to measure all you need to do is a way to measure distance between points and then to distance between points and then to distance between points and then to classify a new point you're just simply classify a new point you're just simply classify a new point you're just simply computing what's the closest point in computing what's the closest point in computing what's the closest point in this massive amount of data and that's this massive amount of data and that's this massive amount of data and that's my answer so you can think of prompting my answer so you can think of prompting my answer so you can think of prompting in a way as you're uploading not not in a way as you're uploading not not in a way as you're uploading not not just simple points and and you know the just simple points and and you know the just simple points and and you know the metric is not the distance between the metric is not the distance between the metric is not the distance between the images or something simple it's images or something simple it's images or something simple it's something that you compute that's much something that you compute that's much something that you compute that's much more advanced but in a way it's very more advanced but in a way it's very more advanced but in a way it's very similar right you you simply are similar right you you simply are similar right you you simply are uploading some uploading some uploading some knowledge to this pre-trained system in knowledge to this pre-trained system in knowledge to this pre-trained system in nearest neighbor maybe the metric is nearest neighbor maybe the metric is nearest neighbor maybe the metric is learned or not but you don't need to learned or not but you don't need to learned or not but you don't need to further train it and then now you further train it and then now you further train it and then now you immediately get a classifier immediately get a classifier immediately get a classifier out of this right now it's just an out of this right now it's just an out of this right now it's just an evolution of that concept very classical evolution of that concept very classical evolution of that concept very classical concept in machine learning which is um concept in machine learning which is um concept in machine learning which is um yeah just learning through what's the yeah just learning through what's the yeah just learning through what's the closest point closes by some distance closest point closes by some distance closest point closes by some distance and that's it yeah it's an evolution of and that's it yeah it's an evolution of and that's it yeah it's an evolution of that and i will say that and i will say that and i will say how how i saw metal learning when we how how i saw metal learning when we how how i saw metal learning when we worked um on a few ideas in in 2016 was worked um on a few ideas in in 2016 was worked um on a few ideas in in 2016 was precisely through the lens of precisely through the lens of precisely through the lens of nearest neighbor which is very common in nearest neighbor which is very common in nearest neighbor which is very common in computer vision community right there's computer vision community right there's computer vision community right there's a very active area of research about how

a very active area of research about how a very active area of research about how do you compute the distance between two do you compute the distance between two do you compute the distance between two images but if you have a good distance images but if you have a good distance images but if you have a good distance metric you you also have a good metric you you also have a good metric you you also have a good classifier right all i'm saying is now classifier right all i'm saying is now classifier right all i'm saying is now these distances and and the points are these distances and and the points are these distances and and the points are not just images they're like not just images they're like not just images they're like words or sequences of words or sequences of words or sequences of words and images and actions that teach words and images and actions that teach words and images and actions that teach you something new but you something new but you something new but it might be that technique wise it might be that technique wise it might be that technique wise those come back and i will say that it's those come back and i will say that it's those come back and i will say that it's not necessarily true that you might not not necessarily true that you might not not necessarily true that you might not ever train the weights a bit further ever train the weights a bit further ever train the weights a bit further some aspect of metal learning some some aspect of metal learning some some aspect of metal learning some techniques in metal learning techniques in metal learning techniques in metal learning do actually do a bit of fine tuning as do actually do a bit of fine tuning as do actually do a bit of fine tuning as it's called right they train the weights it's called right they train the weights it's called right they train the weights a little bit when they get a new task so a little bit when they get a new task so a little bit when they get a new task so as i call the how or or how we're gonna as i call the how or or how we're gonna as i call the how or or how we're gonna achieve this achieve this achieve this um as a deep learner i'm very skeptic um as a deep learner i'm very skeptic um as a deep learner i'm very skeptic we're gonna try a few things whether we're gonna try a few things whether we're gonna try a few things whether it's a bit of training adding a few it's a bit of training adding a few it's a bit of training adding a few parameters thinking of this as nearest parameters thinking of this as nearest parameters thinking of this as nearest neighbor or just neighbor or just neighbor or just simply thinking of there's a sequence of simply thinking of there's a sequence of simply thinking of there's a sequence of words it's a prefix words it's a prefix words it's a prefix and that's the new classifier and that's the new classifier and that's the new classifier we'll see right there's there's the we'll see right there's there's the we'll see right there's there's the beauty of research but um but what's beauty of research but um but what's beauty of research but um but what's what's important is that is a good goal what's important is that is a good goal what's important is that is a good goal in itself that i see as very worthwhile in itself that i see as very worthwhile in itself that i see as very worthwhile pursuing for the next stages of pursuing for the next stages of pursuing for the next stages of not only meta learning i think this is not only meta learning i think this is not only meta learning i think this is basically basically basically what's exciting about machine learning what's exciting about machine learning what's exciting about machine learning period to me well the and then the period to me well the and then the period to me well the and then the interactive aspect of that is also very interactive aspect of that is also very interactive aspect of that is also very interesting the interactive version of interesting the interactive version of interesting the interactive version of nearest neighbor nearest neighbor nearest neighbor yeah to help you uh yeah to help you uh yeah to help you uh pull out the pull out the pull out the classifier from this giant thing

classifier from this giant thing classifier from this giant thing okay okay okay uh is is this the way we can go in five uh is is this the way we can go in five uh is is this the way we can go in five ten ten ten plus years plus years plus years uh from any task so sorry from many uh from any task so sorry from many uh from any task so sorry from many tasks to any task tasks to any task tasks to any task so and what does that mean like what so and what does that mean like what so and what does that mean like what does it need to be actually trained on does it need to be actually trained on does it need to be actually trained on which point is the network had enough which point is the network had enough which point is the network had enough so what um so what um so what um what does a network need to learn about what does a network need to learn about what does a network need to learn about this world in order to be able to this world in order to be able to this world in order to be able to perform any task is it just as simple as perform any task is it just as simple as perform any task is it just as simple as language language language image and action or do you need image and action or do you need image and action or do you need some set of representative images some set of representative images some set of representative images uh like uh like uh like if you only see land images will you if you only see land images will you if you only see land images will you know anything about underwater is that know anything about underwater is that know anything about underwater is that somehow fundamentally different i don't somehow fundamentally different i don't somehow fundamentally different i don't know those i mean those are upward know those i mean those are upward know those i mean those are upward questions i would say i mean the way you questions i would say i mean the way you questions i would say i mean the way you put let me maybe further your example put let me maybe further your example put let me maybe further your example right if if all you see is land images right if if all you see is land images right if if all you see is land images but you're reading all about land and but you're reading all about land and but you're reading all about land and water worlds but in books right imagine water worlds but in books right imagine water worlds but in books right imagine like like like would that be enough i mean good would that be enough i mean good would that be enough i mean good question we don't know but i guess maybe question we don't know but i guess maybe question we don't know but i guess maybe you can you can you can join us if you you can you can you can join us if you you can you can you can join us if you want in our quest to find this that's want in our quest to find this that's want in our quest to find this that's that's precisely water world yeah yes that's precisely water world yeah yes that's precisely water world yeah yes that's precisely i mean the beauty of that's precisely i mean the beauty of that's precisely i mean the beauty of research and and research and and research and and that's the the the the that's the the the the that's the the the the the research business we're in i guess the research business we're in i guess the research business we're in i guess is to figure this out and ask the right is to figure this out and ask the right is to figure this out and ask the right questions and then iterate with with the questions and then iterate with with the questions and then iterate with with the whole community whole community whole community um publishing like um publishing like um publishing like findings and so on uh but yeah these are findings and so on uh but yeah these are findings and so on uh but yeah these are this is a question it's not the only

this is a question it's not the only this is a question it's not the only question but it's certainly as you ask question but it's certainly as you ask question but it's certainly as you ask is is on my mind constantly right and so is is on my mind constantly right and so is is on my mind constantly right and so we'll we'll need to wait for we'll we'll need to wait for we'll we'll need to wait for maybe the let's say five years let's maybe the let's say five years let's maybe the let's say five years let's hope it's it's not 10 to to see what hope it's it's not 10 to to see what hope it's it's not 10 to to see what what are the answers um what are the answers um what are the answers um some people will largely believe in some people will largely believe in some people will largely believe in unsupervised or self-supervised learning unsupervised or self-supervised learning unsupervised or self-supervised learning of single modalities of single modalities of single modalities and then crossing them and then crossing them and then crossing them some people might think end-to-end some people might think end-to-end some people might think end-to-end learning is the answer um modularity is learning is the answer um modularity is learning is the answer um modularity is maybe the answer so we don't know but maybe the answer so we don't know but maybe the answer so we don't know but we're just definitely excited to find we're just definitely excited to find we're just definitely excited to find out but it feels like this is the right out but it feels like this is the right out but it feels like this is the right time and we're at the beginning of this time and we're at the beginning of this time and we're at the beginning of this yeah we're finally ready to do these yeah we're finally ready to do these yeah we're finally ready to do these kind of general kind of general kind of general big models and agents big models and agents big models and agents what do you sort of specific technical what do you sort of specific technical what do you sort of specific technical thing about gato flamingo thing about gato flamingo thing about gato flamingo chinchilla chinchilla chinchilla gopher any of these that is especially gopher any of these that is especially gopher any of these that is especially beautiful that was surprising beautiful that was surprising beautiful that was surprising maybe is there something that just jumps maybe is there something that just jumps maybe is there something that just jumps out at you out at you out at you of course there's the general thing of of course there's the general thing of of course there's the general thing of like you didn't think it was possible like you didn't think it was possible like you didn't think it was possible and then and then and then you realize it's possible in terms of you realize it's possible in terms of you realize it's possible in terms of the generalizability across modalities the generalizability across modalities the generalizability across modalities and all that kind of stuff or maybe the and all that kind of stuff or maybe the and all that kind of stuff or maybe the how small of a network relatively how small of a network relatively how small of a network relatively speaking god was all that kind of stuff speaking god was all that kind of stuff speaking god was all that kind of stuff but is there some weird but is there some weird but is there some weird little things that were surprising little things that were surprising little things that were surprising look i look i look i i'll give you an answer that's very i'll give you an answer that's very i'll give you an answer that's very important because important because important because maybe people maybe people maybe people don't quite realize this but don't quite realize this but don't quite realize this but the teams behind these efforts the

the teams behind these efforts the the teams behind these efforts the actual humans yeah that's maybe the actual humans yeah that's maybe the actual humans yeah that's maybe the surprising surprising surprising um you know obviously positive way so um you know obviously positive way so um you know obviously positive way so anytime you see these these anytime you see these these anytime you see these these breakthroughs i mean it's easy to map it breakthroughs i mean it's easy to map it breakthroughs i mean it's easy to map it to a few people there's people that are to a few people there's people that are to a few people there's people that are great at explaining things and so on great at explaining things and so on great at explaining things and so on that's very nice but that's very nice but that's very nice but maybe the the the learnings or the maybe the the the learnings or the maybe the the the learnings or the method learnings that i get as a human method learnings that i get as a human method learnings that i get as a human about this is um sure we can move about this is um sure we can move about this is um sure we can move forward um forward um forward um and but the surprising bit is and but the surprising bit is and but the surprising bit is how how how how important are all the pieces of of how important are all the pieces of of how important are all the pieces of of these projects these projects these projects how do they come together so i'll give how do they come together so i'll give how do they come together so i'll give you you you uh maybe some of the ingredients of uh maybe some of the ingredients of uh maybe some of the ingredients of success that are common across these um success that are common across these um success that are common across these um but not the obvious ones and machine but not the obvious ones and machine but not the obvious ones and machine learning i i can always always also give learning i i can always always also give learning i i can always always also give you those but you those but you those but basically basically basically there is there is there is engineering is critical so so very good engineering is critical so so very good engineering is critical so so very good engineering uh because ultimately we're engineering uh because ultimately we're engineering uh because ultimately we're collecting um data sets right so the the collecting um data sets right so the the collecting um data sets right so the the engineering of data and then of engineering of data and then of engineering of data and then of deploying the models at scale um into deploying the models at scale um into deploying the models at scale um into some compute cluster that cannot go some compute cluster that cannot go some compute cluster that cannot go understated that is a huge factor of understated that is a huge factor of understated that is a huge factor of success success success and it's hard to believe that and it's hard to believe that and it's hard to believe that details matter so much details matter so much details matter so much we would like to believe that it's true we would like to believe that it's true we would like to believe that it's true that there is more and more of a that there is more and more of a that there is more and more of a standard formula as i was saying like standard formula as i was saying like standard formula as i was saying like this recipe that works for everything this recipe that works for everything this recipe that works for everything but then when you zoom into this each of but then when you zoom into this each of but then when you zoom into this each of these projects then you realize the the

these projects then you realize the the these projects then you realize the the the devil is indeed in the details and the devil is indeed in the details and the devil is indeed in the details and then the teams then the teams then the teams have to work kind of together towards have to work kind of together towards have to work kind of together towards these goals um so engineering of data these goals um so engineering of data these goals um so engineering of data and obviously clusters and large scale and obviously clusters and large scale and obviously clusters and large scale is very important and then is very important and then is very important and then one that is often one that is often one that is often not maybe nowadays it is more more clear not maybe nowadays it is more more clear not maybe nowadays it is more more clear is is is benchmark progress right so we're benchmark progress right so we're benchmark progress right so we're talking here about multiple months of talking here about multiple months of talking here about multiple months of you know tens of researchers um and and you know tens of researchers um and and you know tens of researchers um and and and people that are trying to organize and people that are trying to organize and people that are trying to organize the research and so on working together the research and so on working together the research and so on working together and and and you don't know that you can get there i you don't know that you can get there i you don't know that you can get there i mean it is this this is this is the mean it is this this is this is the mean it is this this is this is the beauty like if you're not risking to beauty like if you're not risking to beauty like if you're not risking to trying to do something that feels trying to do something that feels trying to do something that feels impossible you're not gonna get there um impossible you're not gonna get there um impossible you're not gonna get there um but you need the way to measure progress but you need the way to measure progress but you need the way to measure progress so the benchmarks that you build are so the benchmarks that you build are so the benchmarks that you build are critical um i've seen this beautifully critical um i've seen this beautifully critical um i've seen this beautifully play out in many projects i mean play out in many projects i mean play out in many projects i mean maybe the one i've seen it more maybe the one i've seen it more maybe the one i've seen it more consistently consistently consistently which means we we established the metric which means we we established the metric which means we we established the metric actually the community did and then we actually the community did and then we actually the community did and then we leverage that massively is alpha fault leverage that massively is alpha fault leverage that massively is alpha fault this is a project where this is a project where this is a project where the data the metrics were all there and the data the metrics were all there and the data the metrics were all there and all it took was and it's easier said all it took was and it's easier said all it took was and it's easier said than done an amazing team working than done an amazing team working than done an amazing team working not to try to find some incremental not to try to find some incremental not to try to find some incremental improvement and publish which which is improvement and publish which which is improvement and publish which which is one way to do research that is valid but one way to do research that is valid but one way to do research that is valid but aim very high and work literally for aim very high and work literally for aim very high and work literally for years years years to iterate over that process and working to iterate over that process and working to iterate over that process and working for years with the team i mean

for years with the team i mean for years with the team i mean it is it is tricky that also happened it is it is tricky that also happened it is it is tricky that also happened happened to happen partly during a happened to happen partly during a happened to happen partly during a pandemic and so on um so i think my meta pandemic and so on um so i think my meta pandemic and so on um so i think my meta learning from all these is learning from all these is learning from all these is the teams are critical to the success the teams are critical to the success the teams are critical to the success and then if now going to the machine and then if now going to the machine and then if now going to the machine learning the part that's surprising learning the part that's surprising learning the part that's surprising is is is um um um so we like architectures like neural so we like architectures like neural so we like architectures like neural networks um and networks um and networks um and i would say this was a very rapidly i would say this was a very rapidly i would say this was a very rapidly evolving field until the transformer evolving field until the transformer evolving field until the transformer came so attention might indeed be all came so attention might indeed be all came so attention might indeed be all unique which is the title also a good unique which is the title also a good unique which is the title also a good title although title although title although in hindsight is good i don't think at in hindsight is good i don't think at in hindsight is good i don't think at the time i thought this is a great title the time i thought this is a great title the time i thought this is a great title for a paper but for a paper but for a paper but that that architecture is proving that that that architecture is proving that that that architecture is proving that the dream of modeling sequences of any the dream of modeling sequences of any the dream of modeling sequences of any bites bites bites there is something there that will stick there is something there that will stick there is something there that will stick and and i think these these advance in and and i think these these advance in and and i think these these advance in architectures in in kind of how neural architectures in in kind of how neural architectures in in kind of how neural networks are architecture to do what networks are architecture to do what networks are architecture to do what they do they do they do um it's been hard to find one that has um it's been hard to find one that has um it's been hard to find one that has been so stable and relatively has been so stable and relatively has been so stable and relatively has changed very little changed very little changed very little since it was invented since it was invented since it was invented five or so years ago so that is a five or so years ago so that is a five or so years ago so that is a surprising keeps is a surprise that surprising keeps is a surprise that surprising keeps is a surprise that keeps recurring into other projects try keeps recurring into other projects try keeps recurring into other projects try to to to on a philosophical or technical level on a philosophical or technical level on a philosophical or technical level introspect what is the magic of introspect what is the magic of introspect what is the magic of attention attention attention what is what is the tension what is what is the tension what is what is the tension that's attention in people that study that's attention in people that study that's attention in people that study cognition so human attention i think cognition so human attention i think cognition so human attention i think there's giant wars over what attention there's giant wars over what attention there's giant wars over what attention means means means how it works in the human mind so what

how it works in the human mind so what how it works in the human mind so what this very simple looks at what attention this very simple looks at what attention this very simple looks at what attention is in your network is in your network is in your network from the days of attention is all you from the days of attention is all you from the days of attention is all you need but broad do you think there's a need but broad do you think there's a need but broad do you think there's a general principle that's that's really general principle that's that's really general principle that's that's really powerful here yeah so a distinction powerful here yeah so a distinction powerful here yeah so a distinction between transformers and lstms which between transformers and lstms which between transformers and lstms which were what came before and and you know were what came before and and you know were what came before and and you know there was a transitional period where there was a transitional period where there was a transitional period where you could you could use both in fact you could you could use both in fact you could you could use both in fact when we talked about alpha star we used when we talked about alpha star we used when we talked about alpha star we used transformers and lstms so it was still transformers and lstms so it was still transformers and lstms so it was still the beginning of transformers they were the beginning of transformers they were the beginning of transformers they were very powerful but lstms were still very very powerful but lstms were still very very powerful but lstms were still very also very powerful sequence models also very powerful sequence models also very powerful sequence models so so so the power of the transformer the power of the transformer the power of the transformer is that it has built in is that it has built in is that it has built in what we call an inductive bias of what we call an inductive bias of what we call an inductive bias of attention attention attention that makes the model when when you think that makes the model when when you think that makes the model when when you think of a sequence of integers right like we of a sequence of integers right like we of a sequence of integers right like we discussed this before right this is the discussed this before right this is the discussed this before right this is the sequence of words sequence of words sequence of words um um um when you when you have to do very hard when you when you have to do very hard when you when you have to do very hard tasks over these words this could be tasks over these words this could be tasks over these words this could be we're gonna translate a whole paragraph we're gonna translate a whole paragraph we're gonna translate a whole paragraph or we're gonna predict the next or we're gonna predict the next or we're gonna predict the next paragraph given ten paragraphs before there's some there's some loose loose loose intuition from how we do it as a human intuition from how we do it as a human intuition from how we do it as a human that is very that is very that is very nicely mimicked and re like replicated nicely mimicked and re like replicated nicely mimicked and re like replicated structurally speaking in the transformer structurally speaking in the transformer structurally speaking in the transformer which is this idea of which is this idea of which is this idea of you're looking for something you're looking for something you're looking for something right so you're sort of when you're right so you're sort of when you're right so you're sort of when you're you you just read a piece of text now you you just read a piece of text now you you just read a piece of text now you're thinking what comes next you're thinking what comes next you're thinking what comes next you might want to re-look at the text

you might want to re-look at the text you might want to re-look at the text or look it from scratch i mean or look it from scratch i mean or look it from scratch i mean literally is is because there's no literally is is because there's no literally is is because there's no recurrence you're just thinking what recurrence you're just thinking what recurrence you're just thinking what comes next and comes next and comes next and it's almost hypothesis driven right so it's almost hypothesis driven right so it's almost hypothesis driven right so if if i'm thinking the next word that if if i'm thinking the next word that if if i'm thinking the next word that i'll write is cat or dog okay um i'll write is cat or dog okay um i'll write is cat or dog okay um the way the transformer works the way the transformer works the way the transformer works almost philosophically is it almost philosophically is it almost philosophically is it has these two hypotheses is it is it has these two hypotheses is it is it has these two hypotheses is it is it gonna be cat or is it gonna be dark and gonna be cat or is it gonna be dark and gonna be cat or is it gonna be dark and then then then it says okay if it's cat i'm gonna look it says okay if it's cat i'm gonna look it says okay if it's cat i'm gonna look for certain words not necessarily cat for certain words not necessarily cat for certain words not necessarily cat although cud is an obvious word you although cud is an obvious word you although cud is an obvious word you would look in the past to see whether it would look in the past to see whether it would look in the past to see whether it makes more sense to output cut or dog makes more sense to output cut or dog makes more sense to output cut or dog and then it does some very and then it does some very and then it does some very deep computation over the words and deep computation over the words and deep computation over the words and beyond right so it combines the words beyond right so it combines the words beyond right so it combines the words and and and but but but but it has the query as we call it that but it has the query as we call it that but it has the query as we call it that is cat is cat is cat and then similarly for doc right and so and then similarly for doc right and so and then similarly for doc right and so it's it's very it's a very computational it's it's very it's a very computational it's it's very it's a very computational way to think about way to think about way to think about look if i'm if i'm thinking deeply about look if i'm if i'm thinking deeply about look if i'm if i'm thinking deeply about text i need to go back to to look at all text i need to go back to to look at all text i need to go back to to look at all the texts attend over it but it's not the texts attend over it but it's not the texts attend over it but it's not just attention like what what is guiding just attention like what what is guiding just attention like what what is guiding the attention and that was the key the attention and that was the key the attention and that was the key insight from an earlier paper is not insight from an earlier paper is not insight from an earlier paper is not how far away is it i mean how far away how far away is it i mean how far away how far away is it i mean how far away is it is important what what what did i is it is important what what what did i is it is important what what what did i just write about that's critical but just write about that's critical but just write about that's critical but what you wrote about what you wrote about what you wrote about 10 pages ago might also be critical 10 pages ago might also be critical 10 pages ago might also be critical so so so you're looking not positionally but you're looking not positionally but you're looking not positionally but content-wise right and you transformers content-wise right and you transformers content-wise right and you transformers have this beautiful way to query for have this beautiful way to query for have this beautiful way to query for certain content and pull it out com in a certain content and pull it out com in a certain content and pull it out com in a compressed way so then you can make a

compressed way so then you can make a compressed way so then you can make a more informed decision i mean that's one more informed decision i mean that's one more informed decision i mean that's one way to explain transformers um but i way to explain transformers um but i way to explain transformers um but i think it's it's very it's a very think it's it's very it's a very think it's it's very it's a very powerful inductive bias powerful inductive bias powerful inductive bias there might be some details that might there might be some details that might there might be some details that might change over time but change over time but change over time but i think that is i think that is i think that is what makes transformers so much more what makes transformers so much more what makes transformers so much more powerful than the recurrent networks powerful than the recurrent networks powerful than the recurrent networks that were more recently biased based that were more recently biased based that were more recently biased based which obviously works in some tasks but which obviously works in some tasks but which obviously works in some tasks but it has major flaws transformer itself it has major flaws transformer itself it has major flaws transformer itself has flaws has flaws has flaws and i think the main one the main and i think the main one the main and i think the main one the main challenge is these prompts that we we challenge is these prompts that we we challenge is these prompts that we we just were talking about just were talking about just were talking about they can be a thousand words long but if they can be a thousand words long but if they can be a thousand words long but if i'm teaching you starcraft i mean i'll i'm teaching you starcraft i mean i'll i'm teaching you starcraft i mean i'll have to show you videos i have to i have have to show you videos i have to i have have to show you videos i have to i have to point you to whole wikipedia articles to point you to whole wikipedia articles to point you to whole wikipedia articles about the game um we'll have to interact about the game um we'll have to interact about the game um we'll have to interact probably as you play you'll ask me probably as you play you'll ask me probably as you play you'll ask me questions the context require for us to questions the context require for us to questions the context require for us to achieve achieve achieve me being a good teacher to you on the me being a good teacher to you on the me being a good teacher to you on the game as you would want to do it with a game as you would want to do it with a game as you would want to do it with a model model model what i think goes well beyond the what i think goes well beyond the what i think goes well beyond the current capabilities um so the question current capabilities um so the question current capabilities um so the question is how do we benchmark this and is how do we benchmark this and is how do we benchmark this and then how do we change the structure of then how do we change the structure of then how do we change the structure of the architectures i think there's ideas the architectures i think there's ideas the architectures i think there's ideas on both sides but on both sides but on both sides but we'll have to see empirically right we'll have to see empirically right we'll have to see empirically right obviously what ends up working obviously what ends up working obviously what ends up working and as as you talked about some of the and as as you talked about some of the and as as you talked about some of the ideas could be you know keeping the ideas could be you know keeping the ideas could be you know keeping the constraint of that length in place but constraint of that length in place but constraint of that length in place but then forming like hierarchical then forming like hierarchical then forming like hierarchical representations representations representations to where you can start being much clever to where you can start being much clever to where you can start being much clever in how you use those thousand tokens yeah that's really interesting but it

yeah that's really interesting but it also is possible that this attention also is possible that this attention also is possible that this attention mechanism where you basically you don't mechanism where you basically you don't mechanism where you basically you don't have a recency bias but you you you look have a recency bias but you you you look have a recency bias but you you you look more generally you you make it learnable more generally you you make it learnable more generally you you make it learnable the mechanism in which way you look back the mechanism in which way you look back the mechanism in which way you look back into the past you make that learnable into the past you make that learnable into the past you make that learnable it's also possible where at the very it's also possible where at the very it's also possible where at the very beginning of that beginning of that beginning of that because because because that that that you might become smarter and smarter in you might become smarter and smarter in you might become smarter and smarter in the way the way the way you query the past you query the past you query the past so recent past and distant past and so recent past and distant past and so recent past and distant past and maybe very very distant path so almost maybe very very distant path so almost maybe very very distant path so almost like the attention mechanism like the attention mechanism like the attention mechanism will have to will have to will have to improve and evolve as good as the improve and evolve as good as the improve and evolve as good as the uh the uh the uh the tokenization mechanism where so you can tokenization mechanism where so you can tokenization mechanism where so you can represent long-term memory somehow yes represent long-term memory somehow yes represent long-term memory somehow yes and i mean hierarchies are are very i and i mean hierarchies are are very i and i mean hierarchies are are very i mean it's a very nice word that sounds mean it's a very nice word that sounds mean it's a very nice word that sounds appealing um there's lots of work adding appealing um there's lots of work adding appealing um there's lots of work adding hierarchy to the memories um in practice hierarchy to the memories um in practice hierarchy to the memories um in practice it does seem like we keep coming back to it does seem like we keep coming back to it does seem like we keep coming back to the main formula or main the main formula or main the main formula or main architecture architecture architecture that sometimes tells us something that sometimes tells us something that sometimes tells us something there's such a sentence that a friend of there's such a sentence that a friend of there's such a sentence that a friend of mine told me like whether it wants to mine told me like whether it wants to mine told me like whether it wants to work or not so transformer was clearly work or not so transformer was clearly work or not so transformer was clearly an idea that wanted to work and then an idea that wanted to work and then an idea that wanted to work and then i think there's some principles we i think there's some principles we i think there's some principles we believe will be needed but finding the believe will be needed but finding the believe will be needed but finding the exact details details matter so much exact details details matter so much exact details details matter so much right that's gonna be tricky i love the right that's gonna be tricky i love the right that's gonna be tricky i love the idea that there's like idea that there's like idea that there's like you as a you as a you as a human being you want you want some ideas human being you want you want some ideas human being you want you want some ideas to work and then there's the model that

to work and then there's the model that to work and then there's the model that wants some ideas to work and you get to wants some ideas to work and you get to wants some ideas to work and you get to have a conversation to see which have a conversation to see which have a conversation to see which more likely the model will win in the more likely the model will win in the more likely the model will win in the end end end because it's the one you don't have to because it's the one you don't have to because it's the one you don't have to do any work the model is the one that do any work the model is the one that do any work the model is the one that has to do the work so you should listen has to do the work so you should listen has to do the work so you should listen to the model and i really love this idea to the model and i really love this idea to the model and i really love this idea that you talked about the humans in this that you talked about the humans in this that you talked about the humans in this picture if i could just briefly ask um picture if i could just briefly ask um picture if i could just briefly ask um one is you're saying one is you're saying one is you're saying the benchmarks the benchmarks the benchmarks about the modular humans working on this about the modular humans working on this about the modular humans working on this uh the benchmarks providing a sturdy uh the benchmarks providing a sturdy uh the benchmarks providing a sturdy ground of a wish to do these things that ground of a wish to do these things that ground of a wish to do these things that seem impossible seem impossible seem impossible they they give you they they give you they they give you in the darkest of times give you hope in the darkest of times give you hope in the darkest of times give you hope because little signs of improvement you because little signs of improvement you because little signs of improvement you get you could yes like you're not you're get you could yes like you're not you're get you could yes like you're not you're somehow you're not lost if you have somehow you're not lost if you have somehow you're not lost if you have metrics to measure your your improvement metrics to measure your your improvement metrics to measure your your improvement and then there's other aspect and then there's other aspect and then there's other aspect you said elsewhere and here today like you said elsewhere and here today like you said elsewhere and here today like titles matter titles matter titles matter i wonder how much humans matter in the evolution how much humans matter in the evolution of all this of all this of all this meaning individual humans meaning individual humans meaning individual humans you know something about their you know something about their you know something about their interaction something about their ideas interaction something about their ideas interaction something about their ideas how much they how much they how much they change the direction of all this like if change the direction of all this like if change the direction of all this like if you change the humans in this picture you change the humans in this picture you change the humans in this picture like is is it that the model is sitting like is is it that the model is sitting like is is it that the model is sitting there there there and it wants you it wants some idea to and it wants you it wants some idea to and it wants you it wants some idea to work or is it the humans or maybe the work or is it the humans or maybe the work or is it the humans or maybe the model is providing you 20 ideas that model is providing you 20 ideas that model is providing you 20 ideas that could work and depending on the humans could work and depending on the humans could work and depending on the humans you pick they're they're going to be you pick they're they're going to be you pick they're they're going to be able to hear some of those ideas like in able to hear some of those ideas like in able to hear some of those ideas like in in all the because you're now directing

in all the because you're now directing in all the because you're now directing all of deep learning at deepmind you get all of deep learning at deepmind you get all of deep learning at deepmind you get to interact with a lot of projects a lot to interact with a lot of projects a lot to interact with a lot of projects a lot of brilliant researchers of brilliant researchers of brilliant researchers um um um how much variability is created by the how much variability is created by the how much variability is created by the humans in all of this yeah i mean you i humans in all of this yeah i mean you i humans in all of this yeah i mean you i do believe humans matter a lot at the do believe humans matter a lot at the do believe humans matter a lot at the very least at the very least at the very least at the you know time scale of years you know time scale of years you know time scale of years on when things are happening and what's on when things are happening and what's on when things are happening and what's the sequencing of it right so you get to the sequencing of it right so you get to the sequencing of it right so you get to interact with interact with interact with people that i mean you mentioned this um people that i mean you mentioned this um people that i mean you mentioned this um some people really some people really some people really want some idea to work and they'll want some idea to work and they'll want some idea to work and they'll persist um and then some other people persist um and then some other people persist um and then some other people might be more practical like i don't might be more practical like i don't might be more practical like i don't care care care what idea works i care about you know what idea works i care about you know what idea works i care about you know cracking protein folding yes um and cracking protein folding yes um and cracking protein folding yes um and these at least these two kind of seem these at least these two kind of seem these at least these two kind of seem opposite sides we need both and we've opposite sides we need both and we've opposite sides we need both and we've clearly had both clearly had both clearly had both um historically and that made certain um historically and that made certain um historically and that made certain things happen earlier or later so things happen earlier or later so things happen earlier or later so definitely humans involved in all of definitely humans involved in all of definitely humans involved in all of this endeavor have had this endeavor have had this endeavor have had i would say years of change or of i would say years of change or of i would say years of change or of ordering how how things have happened ordering how how things have happened ordering how how things have happened which breakthroughs came before which which breakthroughs came before which which breakthroughs came before which other breakthroughs and so on so other breakthroughs and so on so other breakthroughs and so on so certainly that does happen certainly that does happen certainly that does happen and so and so and so one other maybe one other axis of one other maybe one other axis of one other maybe one other axis of distinction is distinction is distinction is what i called and this is most commonly what i called and this is most commonly what i called and this is most commonly used in reinforcement learning is the used in reinforcement learning is the used in reinforcement learning is the exploration exploitation trade-off as exploration exploitation trade-off as exploration exploitation trade-off as well it's not exactly what i meant well it's not exactly what i meant well it's not exactly what i meant although quite related so although quite related so although quite related so when you when you when you start

start start trying to help others right like you trying to help others right like you trying to help others right like you you're you're you know you're you become you're you're you know you're you become you're you're you know you're you become a bit more of a mentor to a large group a bit more of a mentor to a large group a bit more of a mentor to a large group of people beat a project or the deep of people beat a project or the deep of people beat a project or the deep learning team or something um or even in learning team or something um or even in learning team or something um or even in the community when you interact with the community when you interact with the community when you interact with people in conferences and so on people in conferences and so on people in conferences and so on um you're identifying um you're identifying um you're identifying quickly right um some some things that quickly right um some some things that quickly right um some some things that are explorative or exploitative and are explorative or exploitative and are explorative or exploitative and it's tempting to try to guide people it's tempting to try to guide people it's tempting to try to guide people obviously i mean that's what makes like obviously i mean that's what makes like obviously i mean that's what makes like our experience we bring it and we try to our experience we bring it and we try to our experience we bring it and we try to shape things um sometimes wrongly and shape things um sometimes wrongly and shape things um sometimes wrongly and there's many times that i've been wrong there's many times that i've been wrong there's many times that i've been wrong in the past that's great in the past that's great in the past that's great but but but it would be wrong to it would be wrong to it would be wrong to dismiss any sort of dismiss any sort of dismiss any sort of of the research styles that i'm of the research styles that i'm of the research styles that i'm observing um and i often get asked well observing um and i often get asked well observing um and i often get asked well you're in industry right so we do have you're in industry right so we do have you're in industry right so we do have access to large compute scale and so on access to large compute scale and so on access to large compute scale and so on so there's certain kinds of research i so there's certain kinds of research i so there's certain kinds of research i almost feel like we need to do almost feel like we need to do almost feel like we need to do responsibly and so on but it is kind of responsibly and so on but it is kind of responsibly and so on but it is kind of we have the particle accelerator here so we have the particle accelerator here so we have the particle accelerator here so to speak in physics so we need to use it to speak in physics so we need to use it to speak in physics so we need to use it we need to answer the questions that we we need to answer the questions that we we need to answer the questions that we should be answering right now for the should be answering right now for the should be answering right now for the scientific progress but then at the same scientific progress but then at the same scientific progress but then at the same time i look at many advances including time i look at many advances including time i look at many advances including attention which was discovered attention which was discovered attention which was discovered in montreal initially because of lack of in montreal initially because of lack of in montreal initially because of lack of compute right so we were working on compute right so we were working on compute right so we were working on sequence to sequence um with with my sequence to sequence um with with my sequence to sequence um with with my friends over at google brain at the time friends over at google brain at the time friends over at google brain at the time and we were using i think eight gpus and we were using i think eight gpus and we were using i think eight gpus which was somehow a lot at the time and which was somehow a lot at the time and which was somehow a lot at the time and then i think montreal was a bit more then i think montreal was a bit more then i think montreal was a bit more limited in the scale but then they

limited in the scale but then they limited in the scale but then they discovered this content-based attention discovered this content-based attention discovered this content-based attention concept that then has obviously concept that then has obviously concept that then has obviously triggered things like transformer not triggered things like transformer not triggered things like transformer not everything obviously starts transformer everything obviously starts transformer everything obviously starts transformer there's there's always a history that is there's there's always a history that is there's there's always a history that is is important to recognize because then is important to recognize because then is important to recognize because then you can make sure that then those who you can make sure that then those who you can make sure that then those who might feel now well we don't have so might feel now well we don't have so might feel now well we don't have so much compute much compute much compute you need to then you need to then you need to then help them help them help them optimize optimize optimize that the kind of research that might that the kind of research that might that the kind of research that might actually produce amazing change perhaps actually produce amazing change perhaps actually produce amazing change perhaps it's not it's not it's not as short term as some of these as short term as some of these as short term as some of these advancements or perhaps it's a different advancements or perhaps it's a different advancements or perhaps it's a different time scale but um the people and the time scale but um the people and the time scale but um the people and the diversity of the field is quite critical diversity of the field is quite critical diversity of the field is quite critical to that we maintain it and at times to that we maintain it and at times to that we maintain it and at times especially mixed a bit with hype or especially mixed a bit with hype or especially mixed a bit with hype or other things it's it's a bit tricky to other things it's it's a bit tricky to other things it's it's a bit tricky to be observing um maybe too much of the be observing um maybe too much of the be observing um maybe too much of the same thinking across the board um but same thinking across the board um but same thinking across the board um but the humans definitely are critical and i the humans definitely are critical and i the humans definitely are critical and i can think of yeah quite a few personal can think of yeah quite a few personal can think of yeah quite a few personal examples where also examples where also examples where also someone told me something that had a someone told me something that had a someone told me something that had a huge you know huge effect on on to some huge you know huge effect on on to some huge you know huge effect on on to some idea and then that's why i'm saying at idea and then that's why i'm saying at idea and then that's why i'm saying at least at the temp in terms of years least at the temp in terms of years least at the temp in terms of years probably some things do happen yeah and probably some things do happen yeah and probably some things do happen yeah and it's also fascinating how constraints it's also fascinating how constraints it's also fascinating how constraints somehow are essential for innovation somehow are essential for innovation somehow are essential for innovation um um um and the other thing you mentioned about and the other thing you mentioned about and the other thing you mentioned about engineering i have a sneaking suspicion engineering i have a sneaking suspicion engineering i have a sneaking suspicion maybe i maybe i maybe i over over over you know my love is with engineering so you know my love is with engineering so you know my love is with engineering so i have a sneaking suspicion that all the i have a sneaking suspicion that all the i have a sneaking suspicion that all the genius genius genius a large percentage of the genius is in

a large percentage of the genius is in a large percentage of the genius is in the tiny details of engineering the tiny details of engineering the tiny details of engineering so like i think so like i think so like i think we like to think our genius our the we like to think our genius our the we like to think our genius our the genius is in the big ideas genius is in the big ideas genius is in the big ideas there's i have a sneaking suspicion that there's i have a sneaking suspicion that there's i have a sneaking suspicion that like because i've seen the genius of like because i've seen the genius of like because i've seen the genius of details of engineering details details of engineering details details of engineering details make uh make uh make uh like the make the night and day like the make the night and day like the make the night and day difference and i wonder if those kind of difference and i wonder if those kind of difference and i wonder if those kind of have a ripple effect over time have a ripple effect over time have a ripple effect over time so that that too so that's that's sort so that that too so that's that's sort so that that too so that's that's sort of the taking the engineering of the taking the engineering of the taking the engineering perspective that sometimes that quiet perspective that sometimes that quiet perspective that sometimes that quiet innovation at the level of an individual innovation at the level of an individual innovation at the level of an individual engineer or maybe at the small scale of engineer or maybe at the small scale of engineer or maybe at the small scale of a few engineers can make all the a few engineers can make all the a few engineers can make all the difference that scales difference that scales difference that scales because we're doing because we're doing because we're doing we're working on computers that are we're working on computers that are we're working on computers that are scaled across large groups scaled across large groups scaled across large groups that one engineering decision can lead that one engineering decision can lead that one engineering decision can lead to ripple effects yes it's interesting to ripple effects yes it's interesting to ripple effects yes it's interesting to think about yeah i mean engineering to think about yeah i mean engineering to think about yeah i mean engineering there's also there's also there's also kind of a historical kind of a historical kind of a historical it might be a bit random it might be a bit random it might be a bit random because if you think of the history of because if you think of the history of because if you think of the history of how especially deep learning and neural how especially deep learning and neural how especially deep learning and neural networks took off feels like networks took off feels like networks took off feels like a bit random because gpus happened to be a bit random because gpus happened to be a bit random because gpus happened to be there at the right time for a different there at the right time for a different there at the right time for a different purpose which was to play video games so purpose which was to play video games so purpose which was to play video games so even the engineering that goes into the even the engineering that goes into the even the engineering that goes into the hardware hardware hardware and it might have a time like the time and it might have a time like the time and it might have a time like the time frame might be very different i mean frame might be very different i mean frame might be very different i mean these the gpus were evolved throughout these the gpus were evolved throughout these the gpus were evolved throughout many years where we didn't even were many years where we didn't even were many years where we didn't even were looking at that right so even at that looking at that right so even at that looking at that right so even at that level right that revolution so to speak

level right that revolution so to speak level right that revolution so to speak um um um the ripples are like like the ripples are like like the ripples are like like we'll see when they stop right but in we'll see when they stop right but in we'll see when they stop right but in terms of thinking of why is this terms of thinking of why is this terms of thinking of why is this happening right there's there's i think happening right there's there's i think happening right there's there's i think that when i try to categorize it in sort that when i try to categorize it in sort that when i try to categorize it in sort of of of things that might not be so obvious i things that might not be so obvious i things that might not be so obvious i mean clearly there's a hardware mean clearly there's a hardware mean clearly there's a hardware revolution we are revolution we are revolution we are surfing thanks to that um data centers surfing thanks to that um data centers surfing thanks to that um data centers as well i mean data centers as well i mean data centers as well i mean data centers are where like i mean at google for are where like i mean at google for are where like i mean at google for instance obviously they're serving instance obviously they're serving instance obviously they're serving google but there's also now thanks to google but there's also now thanks to google but there's also now thanks to that and to have built such amazing data that and to have built such amazing data that and to have built such amazing data centers we can train these models um centers we can train these models um centers we can train these models um software is an important one i think software is an important one i think software is an important one i think if i look at the state of how i had to if i look at the state of how i had to if i look at the state of how i had to implement things to implement my ideas implement things to implement my ideas implement things to implement my ideas how i discarded ideas because they were how i discarded ideas because they were how i discarded ideas because they were too hard to implement too hard to implement too hard to implement yeah clearly the chat the times have yeah clearly the chat the times have yeah clearly the chat the times have changed and thankfully we are in a much changed and thankfully we are in a much changed and thankfully we are in a much better software position as well better software position as well better software position as well and then and then and then i mean obviously there's research that i mean obviously there's research that i mean obviously there's research that happens at scale and more people enter happens at scale and more people enter happens at scale and more people enter the field that's great to see but it's the field that's great to see but it's the field that's great to see but it's almost enabled by these other things and almost enabled by these other things and almost enabled by these other things and last but not least is also data right last but not least is also data right last but not least is also data right curating data sets labeling data sets curating data sets labeling data sets curating data sets labeling data sets these benchmarks we think about maybe these benchmarks we think about maybe these benchmarks we think about maybe we'll we'll want to have all the we'll we'll want to have all the we'll we'll want to have all the benchmarks in one system but it's still benchmarks in one system but it's still benchmarks in one system but it's still very valuable that someone put the very valuable that someone put the very valuable that someone put the thought and the time and the vision to thought and the time and the vision to thought and the time and the vision to build certain benchmarks we've we've build certain benchmarks we've we've build certain benchmarks we've we've seen progress thanks to but seen progress thanks to but seen progress thanks to but we're gonna repurpose the benchmarks we're gonna repurpose the benchmarks we're gonna repurpose the benchmarks that's the beauty of atari that's the beauty of atari that's the beauty of atari is like is like is like we solved it in a way but we solved it in a way but we solved it in a way but we use it in gato it was critical and we use it in gato it was critical and we use it in gato it was critical and i'm sure it's there's there's still a

i'm sure it's there's there's still a i'm sure it's there's there's still a lot more to do thanks to that amazing lot more to do thanks to that amazing lot more to do thanks to that amazing benchmark that someone took the time to benchmark that someone took the time to benchmark that someone took the time to put even though at the time maybe put even though at the time maybe put even though at the time maybe oh you have to think what's the next oh you have to think what's the next oh you have to think what's the next you know iteration of architectures you know iteration of architectures you know iteration of architectures that's what maybe the field recognizes that's what maybe the field recognizes that's what maybe the field recognizes but we need to that's another thing we but we need to that's another thing we but we need to that's another thing we need to balance in terms of humans need to balance in terms of humans need to balance in terms of humans behind we need to recognize all these behind we need to recognize all these behind we need to recognize all these aspects because they're all critical and aspects because they're all critical and aspects because they're all critical and we tend to we tend to we tend to yeah we tend to think of the genius the yeah we tend to think of the genius the yeah we tend to think of the genius the scientists and so on but i'm i'm glad scientists and so on but i'm i'm glad scientists and so on but i'm i'm glad you're i know you have a strong engineer you're i know you have a strong engineer you're i know you have a strong engineer and background so but also i'm a date and background so but also i'm a date and background so but also i'm a date i'm a lover of data and because it's a i'm a lover of data and because it's a i'm a lover of data and because it's a pushback on the engineering comment pushback on the engineering comment pushback on the engineering comment ultimately could be the the creators of ultimately could be the the creators of ultimately could be the the creators of benchmarks who have the most impact benchmarks who have the most impact benchmarks who have the most impact andre capati who you mentioned has andre capati who you mentioned has andre capati who you mentioned has recently been talking a lot of trash recently been talking a lot of trash recently been talking a lot of trash about imagenet which he has the right to about imagenet which he has the right to about imagenet which he has the right to do because of how critical he is about do because of how critical he is about do because of how critical he is about him him him how essential he is to the development how essential he is to the development how essential he is to the development and the success of deep learning around and the success of deep learning around and the success of deep learning around uh imagenet and you're saying that uh imagenet and you're saying that uh imagenet and you're saying that that's actually that benchmark is that's actually that benchmark is that's actually that benchmark is holding back the field holding back the field holding back the field because i mean especially in his context because i mean especially in his context because i mean especially in his context on tesla autopilot that's looking at on tesla autopilot that's looking at on tesla autopilot that's looking at real world behavior of a system real world behavior of a system real world behavior of a system it's it's it's you you there's something fundamentally you you there's something fundamentally you you there's something fundamentally missing about imagenet that doesn't missing about imagenet that doesn't missing about imagenet that doesn't capture the real worldness of things capture the real worldness of things capture the real worldness of things that we need to have the datasets that we need to have the datasets that we need to have the datasets benchmarks that benchmarks that benchmarks that have the impressive unpredictability the have the impressive unpredictability the have the impressive unpredictability the edge cases the whatever the heck it is edge cases the whatever the heck it is edge cases the whatever the heck it is that makes the real world so comp so that makes the real world so comp so that makes the real world so comp so difficult to operate in we need to have difficult to operate in we need to have difficult to operate in we need to have benchmarks with that so benchmarks with that so benchmarks with that so but but but just to think about the impact of just to think about the impact of just to think about the impact of imagenet as a benchmark

imagenet as a benchmark imagenet as a benchmark and and and that really puts a lot of emphasis on that really puts a lot of emphasis on that really puts a lot of emphasis on the importance of a benchmark both sort the importance of a benchmark both sort the importance of a benchmark both sort of internally a deep mind and as a of internally a deep mind and as a of internally a deep mind and as a community so community so community so one is coming in from within like one is coming in from within like one is coming in from within like how do i create a benchmark for me how do i create a benchmark for me how do i create a benchmark for me to to to mark and make progress and how do i make mark and make progress and how do i make mark and make progress and how do i make benchmark for the community to mark benchmark for the community to mark benchmark for the community to mark and uh push um and uh push um and uh push um progress you you uh you have this progress you you uh you have this progress you you uh you have this amazing paper you co-authored a survey amazing paper you co-authored a survey amazing paper you co-authored a survey paper called emergent abilities of large paper called emergent abilities of large paper called emergent abilities of large language models language models language models has again the philosophy here that i'd has again the philosophy here that i'd has again the philosophy here that i'd love to ask you about love to ask you about love to ask you about what's the intuition about the phenomena what's the intuition about the phenomena what's the intuition about the phenomena of emergence in neural networks of emergence in neural networks of emergence in neural networks transform is language models transform is language models transform is language models is there a magic is there a magic is there a magic threshold beyond which we start to see threshold beyond which we start to see threshold beyond which we start to see certain performance certain performance certain performance and is that different from task to task and is that different from task to task and is that different from task to task is that us humans just being poetic and is that us humans just being poetic and is that us humans just being poetic and romantic or is there literally some romantic or is there literally some romantic or is there literally some level of which we start to see level of which we start to see level of which we start to see breakthrough performance breakthrough performance breakthrough performance yeah i mean this is a property that we yeah i mean this is a property that we yeah i mean this is a property that we start seeing start seeing start seeing um in systems that actually um in systems that actually um in systems that actually tend to be tend to be tend to be so in machine learning traditionally so in machine learning traditionally so in machine learning traditionally again going to benchmarks i mean if if again going to benchmarks i mean if if again going to benchmarks i mean if if you have a some input outputs right like you have a some input outputs right like you have a some input outputs right like that is that is that is just a single input and a single output just a single input and a single output just a single input and a single output you generally you generally you generally um when you train these systems you see um when you train these systems you see um when you train these systems you see reasonably smooth reasonably smooth reasonably smooth curves when you analyze how curves when you analyze how curves when you analyze how how much the data how much the data how much the data set size affects the performance or how

set size affects the performance or how set size affects the performance or how the model size affect the performance or the model size affect the performance or the model size affect the performance or how much you long train you how how long how much you long train you how how long how much you long train you how how long you train the system for you train the system for you train the system for affects the performance right so affects the performance right so affects the performance right so you know if we think of imagenet like you know if we think of imagenet like you know if we think of imagenet like the train curves look the train curves look the train curves look fairly smooth and predictable in a way fairly smooth and predictable in a way fairly smooth and predictable in a way um um um and and and i would say that's probably because of i would say that's probably because of i would say that's probably because of the the the it's kind of a one it's kind of a one it's kind of a one a one hop a one hop a one hop um um um reasoning task right it's like here is reasoning task right it's like here is reasoning task right it's like here is an input and you think for a few an input and you think for a few an input and you think for a few milliseconds or 100 milliseconds 300 as milliseconds or 100 milliseconds 300 as milliseconds or 100 milliseconds 300 as a human and then you tell me yeah a human and then you tell me yeah a human and then you tell me yeah there's there's there's there's an alpaca in this image there's an alpaca in this image there's an alpaca in this image so so so in language in language in language we are seeing benchmarks that require we are seeing benchmarks that require we are seeing benchmarks that require more pondering and more more pondering and more more pondering and more thought in a way right this is just kind thought in a way right this is just kind thought in a way right this is just kind of you you you need to look for some of you you you need to look for some of you you you need to look for some subtleties subtleties subtleties that it involves that it involves that it involves inputs that you you might think of or if inputs that you you might think of or if inputs that you you might think of or if even if the input is a sentence even if the input is a sentence even if the input is a sentence describing a mathematical problem um describing a mathematical problem um describing a mathematical problem um there is there is a bit more processing there is there is a bit more processing there is there is a bit more processing required as a human and more required as a human and more required as a human and more introspection so introspection so introspection so i think i think i think the the the how these benchmarks work how these benchmarks work how these benchmarks work means that there is actually a threshold means that there is actually a threshold means that there is actually a threshold um um um just going back to how transformers work just going back to how transformers work just going back to how transformers work in this way of querying for the right in this way of querying for the right in this way of querying for the right questions to get the right answers that questions to get the right answers that questions to get the right answers that might mean that might mean that might mean that performance becomes random performance becomes random performance becomes random until the right question is asked by the until the right question is asked by the until the right question is asked by the querying system of a transformer or of a querying system of a transformer or of a querying system of a transformer or of a language model like a transformer and

language model like a transformer and language model like a transformer and then then then only only then you might start seeing only only then you might start seeing only only then you might start seeing performance going from random to performance going from random to performance going from random to non-random non-random non-random and and and this is more empirical there's there's this is more empirical there's there's this is more empirical there's there's no formalism or theory behind this yet no formalism or theory behind this yet no formalism or theory behind this yet although it might be quite important but although it might be quite important but although it might be quite important but we're seeing these phase transitions of we're seeing these phase transitions of we're seeing these phase transitions of random performance and until some let's random performance and until some let's random performance and until some let's say scale of a model and then it goes say scale of a model and then it goes say scale of a model and then it goes beyond that and it might be that beyond that and it might be that beyond that and it might be that you need to fit you need to fit you need to fit a few a few a few low order bits of thought low order bits of thought low order bits of thought before you can make progress on the before you can make progress on the before you can make progress on the whole task and if you could measure whole task and if you could measure whole task and if you could measure actually actually actually those breakdown of the task maybe you those breakdown of the task maybe you those breakdown of the task maybe you would see more smooth oh like yeah these would see more smooth oh like yeah these would see more smooth oh like yeah these you know once once you get these and you know once once you get these and you know once once you get these and these and these and this and these then these and these and this and these then these and these and this and these then you start making progress in the task you start making progress in the task you start making progress in the task but it's somehow but it's somehow but it's somehow um a bit annoying because then um a bit annoying because then um a bit annoying because then it means that certain it means that certain it means that certain questions we might ask about questions we might ask about questions we might ask about architectures architectures architectures possibly cannot only be done at certain possibly cannot only be done at certain possibly cannot only be done at certain scale and scale and scale and one thing that one thing that one thing that conversely i've seen great progress on conversely i've seen great progress on conversely i've seen great progress on in the last couple years is this notion in the last couple years is this notion in the last couple years is this notion of science of deep learning of science of deep learning of science of deep learning and science of scale in particular right and science of scale in particular right and science of scale in particular right so so so on the negative is that there's some on the negative is that there's some on the negative is that there's some benchmarks for which progress might need benchmarks for which progress might need benchmarks for which progress might need to be measured at at minimum at a to be measured at at minimum at a to be measured at at minimum at a certain scale until you see then what certain scale until you see then what certain scale until you see then what details of the model matter to make that details of the model matter to make that details of the model matter to make that performance better right so that's a bit performance better right so that's a bit performance better right so that's a bit of a con but

of a con but of a con but what we've also seen is that you can what we've also seen is that you can what we've also seen is that you can you can sort of empirically analyze you can sort of empirically analyze you can sort of empirically analyze behavior of models at scales that are behavior of models at scales that are behavior of models at scales that are smaller right so let's say to put an smaller right so let's say to put an smaller right so let's say to put an example um we had this chinchilla paper example um we had this chinchilla paper example um we had this chinchilla paper that revised the so-called scaling laws that revised the so-called scaling laws that revised the so-called scaling laws of models and that whole study is done of models and that whole study is done of models and that whole study is done at a reasonably small scale right maybe at a reasonably small scale right maybe at a reasonably small scale right maybe hundreds of millions up to one billion hundreds of millions up to one billion hundreds of millions up to one billion parameters and then the cool thing is parameters and then the cool thing is parameters and then the cool thing is that you create some loss right some that you create some loss right some that you create some loss right some loss that some trends right you you loss that some trends right you you loss that some trends right you you extract trends from data that you see extract trends from data that you see extract trends from data that you see okay like it looks like the amount of okay like it looks like the amount of okay like it looks like the amount of data required to train now a 10x larger data required to train now a 10x larger data required to train now a 10x larger model would be this and these laws so model would be this and these laws so model would be this and these laws so far these extrapolations have helped us far these extrapolations have helped us far these extrapolations have helped us save compute and just get to a better save compute and just get to a better save compute and just get to a better place in terms of the science of place in terms of the science of place in terms of the science of how should we run these models at scale how should we run these models at scale how should we run these models at scale how much data how much depth and all how much data how much depth and all how much data how much depth and all sorts of questions we start asking sorts of questions we start asking sorts of questions we start asking extrapolating from small scale but then extrapolating from small scale but then extrapolating from small scale but then this emergence is sadly that not this emergence is sadly that not this emergence is sadly that not everything can be extrapolated from everything can be extrapolated from everything can be extrapolated from scale depending on the benchmark and scale depending on the benchmark and scale depending on the benchmark and maybe the harder benchmarks are not so maybe the harder benchmarks are not so maybe the harder benchmarks are not so good for extracting these laws but we good for extracting these laws but we good for extracting these laws but we have a variety of benchmarks at least so have a variety of benchmarks at least so have a variety of benchmarks at least so i wonder i wonder i wonder to which degree to which degree to which degree the threshold the phase shift the threshold the phase shift the threshold the phase shift scale is a function of the benchmark scale is a function of the benchmark scale is a function of the benchmark some some of that some of the science some some of that some of the science some some of that some of the science the scale might be the scale might be the scale might be engineering benchmarks engineering benchmarks engineering benchmarks where that threshold is low where that threshold is low where that threshold is low sort of taking sort of taking sort of taking a main benchmark

a main benchmark a main benchmark and uh reducing it somehow or the and uh reducing it somehow or the and uh reducing it somehow or the essential difficulties left but the essential difficulties left but the essential difficulties left but the emergent the scale at which the emergent the scale at which the emergent the scale at which the emergence happens is lower just for the emergence happens is lower just for the emergence happens is lower just for the science aspect of it versus the actual science aspect of it versus the actual science aspect of it versus the actual real world aspect yeah so luckily we real world aspect yeah so luckily we real world aspect yeah so luckily we have quite a few benchmarks some of have quite a few benchmarks some of have quite a few benchmarks some of which are simpler or maybe they're more which are simpler or maybe they're more which are simpler or maybe they're more like i think people might call this like i think people might call this like i think people might call this systems one versus systems2 style um so systems one versus systems2 style um so systems one versus systems2 style um so i think what we're not seeing luckily is i think what we're not seeing luckily is i think what we're not seeing luckily is that that that extrapolations from maybe slightly more extrapolations from maybe slightly more extrapolations from maybe slightly more smooth or simpler benchmarks are smooth or simpler benchmarks are smooth or simpler benchmarks are translating to the harder harder ones translating to the harder harder ones translating to the harder harder ones but that is not to say that this but that is not to say that this but that is not to say that this extrapolation will hit its limits and extrapolation will hit its limits and extrapolation will hit its limits and when it does when it does when it does then then then how much we scale or how we scale will how much we scale or how we scale will how much we scale or how we scale will sadly be a bit suboptimal until we find sadly be a bit suboptimal until we find sadly be a bit suboptimal until we find better loss right um and these laws better loss right um and these laws better loss right um and these laws again are very empirical loss they're again are very empirical loss they're again are very empirical loss they're not like physical loss of models not like physical loss of models not like physical loss of models although i wish although i wish although i wish there would be better theory about these there would be better theory about these there would be better theory about these things as well but so far i would say things as well but so far i would say things as well but so far i would say empirical theory as i call it is way empirical theory as i call it is way empirical theory as i call it is way ahead than actual theory of machine ahead than actual theory of machine ahead than actual theory of machine learning learning learning let me ask you let me ask you let me ask you almost for fun so this is not auriel as almost for fun so this is not auriel as almost for fun so this is not auriel as a as a deep mind person or anything to a as a deep mind person or anything to a as a deep mind person or anything to do with deep mind or google just as a do with deep mind or google just as a do with deep mind or google just as a human being and looking at these news of human being and looking at these news of human being and looking at these news of a google engineer who claimed a google engineer who claimed a google engineer who claimed uh uh uh that that that uh i guess the lambda language model was uh i guess the lambda language model was uh i guess the lambda language model was sentient or sentient or sentient or had the i still need to look into the had the i still need to look into the had the i still need to look into the details of this

details of this details of this but but but sort of sort of sort of making an official report making an official report making an official report and the claim that he believes there's and the claim that he believes there's and the claim that he believes there's evidence evidence evidence that this system is has achieved that this system is has achieved that this system is has achieved sentience and i think sentience and i think sentience and i think this is a really interesting case this is a really interesting case this is a really interesting case on a human level and a psychological on a human level and a psychological on a human level and a psychological level on a level on a level on a technical machine learning level of how technical machine learning level of how technical machine learning level of how language models transform our world and language models transform our world and language models transform our world and also just philosophical level of the also just philosophical level of the also just philosophical level of the role of ai systems role of ai systems role of ai systems in um in a human world so in um in a human world so in um in a human world so what did you what do you find what did you what do you find what did you what do you find interesting interesting interesting what's your take on all of this as what's your take on all of this as what's your take on all of this as a machine learning engineer and a a machine learning engineer and a a machine learning engineer and a researcher and also as a human being researcher and also as a human being researcher and also as a human being yeah i mean a few reactions yeah i mean a few reactions yeah i mean a few reactions um quite a few actually have you ever um quite a few actually have you ever um quite a few actually have you ever briefly thought is this thing sanctuary briefly thought is this thing sanctuary briefly thought is this thing sanctuary right so never absolutely like even with right so never absolutely like even with right so never absolutely like even with like alpha star wait a minute what uh like alpha star wait a minute what uh like alpha star wait a minute what uh sadly though i think yeah sadly i i have sadly though i think yeah sadly i i have sadly though i think yeah sadly i i have not um yeah i think i think the current not um yeah i think i think the current not um yeah i think i think the current any of the current models although very any of the current models although very any of the current models although very useful and very good um useful and very good um useful and very good um yeah i think we're quite far from that yeah i think we're quite far from that yeah i think we're quite far from that and there's kind of a converse and there's kind of a converse and there's kind of a converse side story so one of one of the my side story so one of one of the my side story so one of one of the my passions is about science in general and passions is about science in general and passions is about science in general and i think i think i think i feel i'm a bit of like a failed i feel i'm a bit of like a failed i feel i'm a bit of like a failed scientist that's why i came to machine scientist that's why i came to machine scientist that's why i came to machine learning because you always feel and you learning because you always feel and you learning because you always feel and you start seeing this that machine learning start seeing this that machine learning start seeing this that machine learning is maybe is maybe is maybe the science that can help other sciences

the science that can help other sciences the science that can help other sciences as we've seen right like you you know as we've seen right like you you know as we've seen right like you you know it's such a powerful tool um so it's such a powerful tool um so it's such a powerful tool um so thanks to that angle right that okay i thanks to that angle right that okay i thanks to that angle right that okay i love science i love i mean i love love science i love i mean i love love science i love i mean i love astronomy i love biology but i'm not an astronomy i love biology but i'm not an astronomy i love biology but i'm not an expert and i decided well the thing i expert and i decided well the thing i expert and i decided well the thing i can do better at these computers but can do better at these computers but can do better at these computers but having especially with when i was a bit having especially with when i was a bit having especially with when i was a bit more involved in alpha fault learning a more involved in alpha fault learning a more involved in alpha fault learning a bit about proteins and about biology and bit about proteins and about biology and bit about proteins and about biology and about about about life life life um um um the complexity the complexity the complexity it feels like it really is like i mean it feels like it really is like i mean it feels like it really is like i mean if you start looking at if you start looking at if you start looking at the things that are going on um the things that are going on um the things that are going on um at you know at that atomic level um at you know at that atomic level um at you know at that atomic level um and and also i mean there's there's and and also i mean there's there's and and also i mean there's there's obviously that obviously that obviously that we are maybe inclined to try to think of we are maybe inclined to try to think of we are maybe inclined to try to think of neural networks as like the brain but neural networks as like the brain but neural networks as like the brain but the complexities the complexities the complexities and the amount of magic that it feels and the amount of magic that it feels and the amount of magic that it feels when i mean i don't i'm not an expert so when i mean i don't i'm not an expert so when i mean i don't i'm not an expert so it naturally feels more magic but it naturally feels more magic but it naturally feels more magic but looking at biological systems as opposed looking at biological systems as opposed looking at biological systems as opposed to these computer to these computer to these computer computational brains computational brains computational brains just makes me like wow this there's such just makes me like wow this there's such just makes me like wow this there's such level of complexity different still level of complexity different still level of complexity different still right like orders of magnitude right like orders of magnitude right like orders of magnitude complexity that um complexity that um complexity that um sure these weights i mean we train them sure these weights i mean we train them sure these weights i mean we train them and they do nice things but they're not and they do nice things but they're not and they do nice things but they're not at the level at the level at the level of biological of biological of biological entities brains entities brains entities brains cells cells cells it just feels like it's just not it just feels like it's just not it just feels like it's just not possible to achieve the same level of possible to achieve the same level of possible to achieve the same level of complexity complexity complexity behavior and but my belief when i talk

behavior and but my belief when i talk behavior and but my belief when i talk to other beings is certainly shaped by to other beings is certainly shaped by to other beings is certainly shaped by this amazement of biology that maybe this amazement of biology that maybe this amazement of biology that maybe because i know too much i don't have because i know too much i don't have because i know too much i don't have about machine learning but i certainly about machine learning but i certainly about machine learning but i certainly feel it's very far feel it's very far feel it's very far fetched and far in the future to be fetched and far in the future to be fetched and far in the future to be calling um calling um calling um or to be thinking well this this this or to be thinking well this this this or to be thinking well this this this mathematical function that is mathematical function that is mathematical function that is differentiable is is um is in fact differentiable is is um is in fact differentiable is is um is in fact sentient and so on so there's something sentient and so on so there's something sentient and so on so there's something on that point it's very interesting so on that point it's very interesting so on that point it's very interesting so you know enough you know enough you know enough about machines and enough about biology about machines and enough about biology about machines and enough about biology to know that there's many orders of to know that there's many orders of to know that there's many orders of magnitude of difference in magnitude of difference in magnitude of difference in complexity but complexity but complexity but you know how machine learning works you know how machine learning works you know how machine learning works so the interesting question from human so the interesting question from human so the interesting question from human beings that are interacting with the beings that are interacting with the beings that are interacting with the system that don't know about the system that don't know about the system that don't know about the underlying complexity underlying complexity underlying complexity and i've seen people probably including and i've seen people probably including and i've seen people probably including myself that have fallen in love with myself that have fallen in love with myself that have fallen in love with things that are quite simple things that are quite simple things that are quite simple yeah so and and so maybe the complexity yeah so and and so maybe the complexity yeah so and and so maybe the complexity is one part of the picture but maybe is one part of the picture but maybe is one part of the picture but maybe that's not a necessary that's not a necessary that's not a necessary um um um that's not a necessary condition for that's not a necessary condition for that's not a necessary condition for sentience for um perception sentience for um perception sentience for um perception uh or emulation of sentience right so i uh or emulation of sentience right so i uh or emulation of sentience right so i mean i guess the other side of this is mean i guess the other side of this is mean i guess the other side of this is that's how i feel personally i mean you that's how i feel personally i mean you that's how i feel personally i mean you asked me about the person right um now asked me about the person right um now asked me about the person right um now it's very interesting to see how other it's very interesting to see how other it's very interesting to see how other humans feel about things right this is humans feel about things right this is humans feel about things right this is this we are like um again like i'm i'm this we are like um again like i'm i'm this we are like um again like i'm i'm not as amazed about things that i feel not as amazed about things that i feel not as amazed about things that i feel like this is not as magical as this like this is not as magical as this like this is not as magical as this other thing because of maybe yeah how i

other thing because of maybe yeah how i other thing because of maybe yeah how i got to learn about it and how i see the got to learn about it and how i see the got to learn about it and how i see the curve a bit more smooth because i you curve a bit more smooth because i you curve a bit more smooth because i you know like just seen the progress of know like just seen the progress of know like just seen the progress of language models since shannon in the 50s language models since shannon in the 50s language models since shannon in the 50s and and and actually looking at that time scale actually looking at that time scale actually looking at that time scale we're not that fast progress right i we're not that fast progress right i we're not that fast progress right i mean it's what what we were thinking at mean it's what what we were thinking at mean it's what what we were thinking at the time like almost 100 years ago the time like almost 100 years ago the time like almost 100 years ago is not that dissimilar to what we're is not that dissimilar to what we're is not that dissimilar to what we're doing now but at the same time yeah doing now but at the same time yeah doing now but at the same time yeah obviously others my experience right obviously others my experience right obviously others my experience right that the personal experience that the personal experience that the personal experience i think no one should um you know i i think no one should um you know i i think no one should um you know i think no one should think no one should think no one should should should tell others how they should should tell others how they should should tell others how they should feel i mean the feelings are very should feel i mean the feelings are very should feel i mean the feelings are very personal right so how others might feel personal right so how others might feel personal right so how others might feel about the models and so on that's one about the models and so on that's one about the models and so on that's one part of the story that is important to part of the story that is important to part of the story that is important to understand for me personally as a understand for me personally as a understand for me personally as a researcher and then researcher and then researcher and then when i maybe disagree or i don't when i maybe disagree or i don't when i maybe disagree or i don't understand or see that yeah maybe this understand or see that yeah maybe this understand or see that yeah maybe this this is not something i think right now this is not something i think right now this is not something i think right now is reasonable knowing all that i know is reasonable knowing all that i know is reasonable knowing all that i know one of the other things and perhaps one of the other things and perhaps one of the other things and perhaps partly why it's great to be talking to partly why it's great to be talking to partly why it's great to be talking to you and reaching out to the world about you and reaching out to the world about you and reaching out to the world about machine learning is hey machine learning is hey machine learning is hey let's make let's demystify a bit the let's make let's demystify a bit the let's make let's demystify a bit the magic and try to see a bit more of the magic and try to see a bit more of the magic and try to see a bit more of the math and the fact that literally to math and the fact that literally to math and the fact that literally to create these models if we had the right create these models if we had the right create these models if we had the right software it would be 10 lines of code um software it would be 10 lines of code um software it would be 10 lines of code um and then just a dump of the internet so and then just a dump of the internet so and then just a dump of the internet so versus like then the complexity of like versus like then the complexity of like versus like then the complexity of like the the the creation of humans um from from their creation of humans um from from their creation of humans um from from their inception right and also the complexity inception right and also the complexity inception right and also the complexity of evolution of the whole universe to

of evolution of the whole universe to of evolution of the whole universe to where we are um that is feels orders of where we are um that is feels orders of where we are um that is feels orders of magnitude more complex and fascinating magnitude more complex and fascinating magnitude more complex and fascinating to me so i think to me so i think to me so i think yeah maybe part of the only thing i'm yeah maybe part of the only thing i'm yeah maybe part of the only thing i'm thinking about thinking about thinking about trying to tell you is yeah i i think trying to tell you is yeah i i think trying to tell you is yeah i i think explaining a bit of the magic there is a explaining a bit of the magic there is a explaining a bit of the magic there is a bit of magic it's good to be in love bit of magic it's good to be in love bit of magic it's good to be in love obviously with what you do at work and obviously with what you do at work and obviously with what you do at work and i'm certainly fascinated and surprised i'm certainly fascinated and surprised i'm certainly fascinated and surprised quite quite often as well but i think quite quite often as well but i think quite quite often as well but i think hopefully as hopefully as hopefully as experts in biology hopefully will tell experts in biology hopefully will tell experts in biology hopefully will tell me this is not as magic and i'm happy to me this is not as magic and i'm happy to me this is not as magic and i'm happy to learn that um through through learn that um through through learn that um through through interactions with the larger community interactions with the larger community interactions with the larger community we can we can we can also have a certain level of education also have a certain level of education also have a certain level of education that that that in practice also will matter because i in practice also will matter because i in practice also will matter because i mean one question is how you feel about mean one question is how you feel about mean one question is how you feel about this but then the other very important this but then the other very important this but then the other very important is is is you starting to interact with this in you starting to interact with this in you starting to interact with this in products and so on um it's good to products and so on um it's good to products and so on um it's good to understand a bit what's going on what's understand a bit what's going on what's understand a bit what's going on what's not going on and what's safe what's not not going on and what's safe what's not not going on and what's safe what's not safe and so on right otherwise um the safe and so on right otherwise um the safe and so on right otherwise um the technology will not be used properly for technology will not be used properly for technology will not be used properly for good which is obviously the goal of all good which is obviously the goal of all good which is obviously the goal of all of us i hope of us i hope of us i hope so let me then ask the next question do so let me then ask the next question do so let me then ask the next question do you think in order to solve intelligence you think in order to solve intelligence you think in order to solve intelligence or to do to replace the lex bot that or to do to replace the lex bot that or to do to replace the lex bot that does interviews as we started this does interviews as we started this does interviews as we started this conversation with do you think conversation with do you think conversation with do you think the system needs to be the system needs to be the system needs to be sentient do you think he needs to sentient do you think he needs to sentient do you think he needs to achieve something achieve something achieve something like consciousness and do you think like consciousness and do you think like consciousness and do you think about what consciousness is in the human about what consciousness is in the human about what consciousness is in the human mind mind mind that could be instructive for creating that could be instructive for creating that could be instructive for creating ai systems ai systems ai systems yeah yeah yeah honestly i think probably not

honestly i think probably not honestly i think probably not to to the degree of intelligence that to to the degree of intelligence that to to the degree of intelligence that there's there's there's this brain that this brain that this brain that can learn can be extremely useful can can learn can be extremely useful can can learn can be extremely useful can challenge you can teach you challenge you can teach you challenge you can teach you um converse you can teach um converse you can teach um converse you can teach it to do things i'm not sure it's it to do things i'm not sure it's it to do things i'm not sure it's necessary personally speaking necessary personally speaking necessary personally speaking but but but if consciousness or any other biological if consciousness or any other biological if consciousness or any other biological or evolutionary or evolutionary or evolutionary lesson lesson lesson can be can be can be repurposed to repurposed to repurposed to then influence our next set of then influence our next set of then influence our next set of algorithms that is a great that is a algorithms that is a great that is a algorithms that is a great that is a great way to actually make progress great way to actually make progress great way to actually make progress right and the same way i try to explain right and the same way i try to explain right and the same way i try to explain transformers a bit how it feels we transformers a bit how it feels we transformers a bit how it feels we operate when we look at text operate when we look at text operate when we look at text specifically specifically specifically these insights these insights these insights are very important right so there's a are very important right so there's a are very important right so there's a distinction between distinction between distinction between um um um details of how the brain might be doing details of how the brain might be doing details of how the brain might be doing computation um i think computation um i think computation um i think my understanding is sure there's neurons my understanding is sure there's neurons my understanding is sure there's neurons and there's some resemblance to neural and there's some resemblance to neural and there's some resemblance to neural networks but we don't quite understand networks but we don't quite understand networks but we don't quite understand enough of the brain in detail right to enough of the brain in detail right to enough of the brain in detail right to to be able to replicate it but then to be able to replicate it but then to be able to replicate it but then more more more if you if you zoom out a bit how we then if you if you zoom out a bit how we then if you if you zoom out a bit how we then our thought process how memory works um our thought process how memory works um our thought process how memory works um maybe even how evolution got us here maybe even how evolution got us here maybe even how evolution got us here what's exploration exploitation like all what's exploration exploitation like all what's exploration exploitation like all the how these things happen i think this the how these things happen i think this the how these things happen i think this clearly can inform algorithmic level clearly can inform algorithmic level clearly can inform algorithmic level research and i've seen some examples um research and i've seen some examples um research and i've seen some examples um of these being quite useful to then of these being quite useful to then of these being quite useful to then guide the research even it might be for guide the research even it might be for guide the research even it might be for the wrong reasons right so i think

the wrong reasons right so i think the wrong reasons right so i think um um um biology and what we know about ourselves biology and what we know about ourselves biology and what we know about ourselves can help a can help a can help a whole lot to build um essentially like whole lot to build um essentially like whole lot to build um essentially like what we call agi this this general um what we call agi this this general um what we call agi this this general um the real gato right the the last step of the real gato right the the last step of the real gato right the the last step of the chain hopefully but the chain hopefully but the chain hopefully but but consciousness in particular i don't but consciousness in particular i don't but consciousness in particular i don't i don't myself at least think too hard i don't myself at least think too hard i don't myself at least think too hard about about about how to add that to to the system but how to add that to to the system but how to add that to to the system but maybe maybe my understanding is also maybe maybe my understanding is also maybe maybe my understanding is also very personal about what it means right very personal about what it means right very personal about what it means right i think this even even that in itself is i think this even even that in itself is i think this even even that in itself is a long debate that i know people uh a long debate that i know people uh a long debate that i know people uh people have often people have often people have often and maybe i should learn more about this and maybe i should learn more about this and maybe i should learn more about this yeah and i personally yeah and i personally yeah and i personally i notice the magic often on a personal i notice the magic often on a personal i notice the magic often on a personal level especially with physical systems level especially with physical systems level especially with physical systems like robots i have a lot of uh like robots i have a lot of uh like robots i have a lot of uh legged robots now in austin that i play legged robots now in austin that i play legged robots now in austin that i play with and even when you program them when with and even when you program them when with and even when you program them when they do things you didn't expect they do things you didn't expect they do things you didn't expect there's an immediate there's an immediate there's an immediate anthropomorphization anthropomorphization anthropomorphization and you notice the magic and you start and you notice the magic and you start and you notice the magic and you start to think about things like sanctions to think about things like sanctions to think about things like sanctions that has to do more with effective that has to do more with effective that has to do more with effective communication and less with any of these communication and less with any of these communication and less with any of these kind of dramatic things kind of dramatic things kind of dramatic things it um it um it um it seems like a useful part of it seems like a useful part of it seems like a useful part of communication communication communication having the perception having the perception having the perception of consciousness of consciousness of consciousness seems like useful for us humans we we seems like useful for us humans we we seems like useful for us humans we we treat each other more seriously we are treat each other more seriously we are treat each other more seriously we are able to uh do a nearest neighbor able to uh do a nearest neighbor able to uh do a nearest neighbor shoving of that entity into your memory shoving of that entity into your memory shoving of that entity into your memory correctly all that kind of stuff seems correctly all that kind of stuff seems correctly all that kind of stuff seems useful at least to fake it even if you useful at least to fake it even if you useful at least to fake it even if you never make it so maybe like yeah

never make it so maybe like yeah never make it so maybe like yeah mirroring the question mirroring the question mirroring the question and since you talk to a few people do and since you talk to a few people do and since you talk to a few people do you then you do think that you then you do think that you then you do think that we'll need to figure something out we'll need to figure something out we'll need to figure something out in order to achieve in order to achieve in order to achieve intelligence in a grander sense of the intelligence in a grander sense of the intelligence in a grander sense of the world yeah i i personally believe yes world yeah i i personally believe yes world yeah i i personally believe yes but i don't even think it'll be like a but i don't even think it'll be like a but i don't even think it'll be like a separate island we'll have to travel to separate island we'll have to travel to separate island we'll have to travel to i think it will emerge very quite i think it will emerge very quite i think it will emerge very quite naturally okay that's easier than for us naturally okay that's easier than for us naturally okay that's easier than for us then thank you but the reason i think then thank you but the reason i think then thank you but the reason i think it's important to think about is you it's important to think about is you it's important to think about is you will start i believe like with this will start i believe like with this will start i believe like with this google engineer you'll start seeing this google engineer you'll start seeing this google engineer you'll start seeing this a lot more especially when you have ai a lot more especially when you have ai a lot more especially when you have ai systems that are actually interacting systems that are actually interacting systems that are actually interacting with human beings that don't have an with human beings that don't have an with human beings that don't have an engineering background engineering background engineering background and we have to prepare for that and we have to prepare for that and we have to prepare for that because there will be i do believe there because there will be i do believe there because there will be i do believe there will be a civil rights movement for will be a civil rights movement for will be a civil rights movement for robots as silly as as it is to say robots as silly as as it is to say robots as silly as as it is to say there's going to be a large number of there's going to be a large number of there's going to be a large number of people that realize there's these people that realize there's these people that realize there's these intelligent entities with whom i have a intelligent entities with whom i have a intelligent entities with whom i have a deep relationship and i don't want to deep relationship and i don't want to deep relationship and i don't want to lose them they've come to be a part of lose them they've come to be a part of lose them they've come to be a part of my life and they mean a lot they have a my life and they mean a lot they have a my life and they mean a lot they have a name they have a story they have a name they have a story they have a name they have a story they have a memory and we start to ask questions memory and we start to ask questions memory and we start to ask questions about ourselves well about ourselves well about ourselves well what uh this thing sure seems like it's what uh this thing sure seems like it's what uh this thing sure seems like it's capable of suffering capable of suffering capable of suffering because it tells all these stories of because it tells all these stories of because it tells all these stories of suffering it doesn't want to die and all suffering it doesn't want to die and all suffering it doesn't want to die and all those kinds of things and we have to those kinds of things and we have to those kinds of things and we have to start to ask ourselves questions what is start to ask ourselves questions what is start to ask ourselves questions what is the difference between a human being in the difference between a human being in the difference between a human being in this thing and wait so when you engineer this thing and wait so when you engineer this thing and wait so when you engineer i believe i believe i believe from an engineering perspective like a from an engineering perspective like a from an engineering perspective like a deep mind or anybody that builds systems

deep mind or anybody that builds systems deep mind or anybody that builds systems there might be laws in the future where there might be laws in the future where there might be laws in the future where you're not allowed to engineer systems you're not allowed to engineer systems you're not allowed to engineer systems with with with displays of sentience displays of sentience displays of sentience unless unless unless they're they're they're explicitly designed to be that unless explicitly designed to be that unless explicitly designed to be that unless it's a pet so if you if you have a it's a pet so if you if you have a it's a pet so if you if you have a system that's just doing customer system that's just doing customer system that's just doing customer support support support you're legally not allowed to display you're legally not allowed to display you're legally not allowed to display sentience we'll start to like ask sentience we'll start to like ask sentience we'll start to like ask ourselves that question ourselves that question ourselves that question and then so that that's that's going to and then so that that's that's going to and then so that that's that's going to be part of the software engineering be part of the software engineering be part of the software engineering process do we do we which features do we process do we do we which features do we process do we do we which features do we have in one of them as have in one of them as have in one of them as communications essentials but it's communications essentials but it's communications essentials but it's important to start thinking about that important to start thinking about that important to start thinking about that stuff especially how much it captivates stuff especially how much it captivates stuff especially how much it captivates public attention public attention public attention yeah absolutely absolutely it's a it's yeah absolutely absolutely it's a it's yeah absolutely absolutely it's a it's definitely a topic that definitely a topic that definitely a topic that is important we is important we is important we think about and i think in a way i i think about and i think in a way i i think about and i think in a way i i always see not not i mean not not every always see not not i mean not not every always see not not i mean not not every movie is is is equally movie is is is equally movie is is is equally on point with certain things but on point with certain things but on point with certain things but certainly science fiction in this sense certainly science fiction in this sense certainly science fiction in this sense at least has prepared society to to at least has prepared society to to at least has prepared society to to start thinking about certain topics that start thinking about certain topics that start thinking about certain topics that even if it's too early to talk about as even if it's too early to talk about as even if it's too early to talk about as long as we are like reasonable um it's long as we are like reasonable um it's long as we are like reasonable um it's certainly gonna prepare us for for both certainly gonna prepare us for for both certainly gonna prepare us for for both um the research to come and how to i um the research to come and how to i um the research to come and how to i mean there's many important challenges mean there's many important challenges mean there's many important challenges and and topics that um come with with and and topics that um come with with and and topics that um come with with building an intelligent system many of building an intelligent system many of building an intelligent system many of which you just mentioned right so which you just mentioned right so which you just mentioned right so i think i think i think being we're never going to be being we're never going to be being we're never going to be fully ready unless we talk about this fully ready unless we talk about this fully ready unless we talk about this and we start also and we start also and we start also as i said just kind of expanding

as i said just kind of expanding as i said just kind of expanding the the the the the the the people we talk to to not include the people we talk to to not include the people we talk to to not include only our our own researchers and so on only our our own researchers and so on only our our own researchers and so on and in fact places like deepmind but and in fact places like deepmind but and in fact places like deepmind but elsewhere elsewhere elsewhere there's more interdisciplinary there's more interdisciplinary there's more interdisciplinary groups forming up to start asking and groups forming up to start asking and groups forming up to start asking and really working with us on these really working with us on these really working with us on these questions um because obviously this is questions um because obviously this is questions um because obviously this is not initially what your passion is when not initially what your passion is when not initially what your passion is when you do your phd but certainly it is you do your phd but certainly it is you do your phd but certainly it is coming right so it's it's fascinating coming right so it's it's fascinating coming right so it's it's fascinating kind of it's it's the the thing that kind of it's it's the the thing that kind of it's it's the the thing that brings me to brings me to brings me to one of my passions that is learning so one of my passions that is learning so one of my passions that is learning so the in this sense this is kind of a new the in this sense this is kind of a new the in this sense this is kind of a new area that area that area that as a learning system myself i want to as a learning system myself i want to as a learning system myself i want to keep exploring and i think it's it's keep exploring and i think it's it's keep exploring and i think it's it's great that um to see you know parts of great that um to see you know parts of great that um to see you know parts of the debate and and even i seen a level the debate and and even i seen a level the debate and and even i seen a level of maturity in the conferences that deal of maturity in the conferences that deal of maturity in the conferences that deal with ai if you look five years ago um with ai if you look five years ago um with ai if you look five years ago um to now just the amount of workshops and to now just the amount of workshops and to now just the amount of workshops and so on has changed so much is is so on has changed so much is is so on has changed so much is is impressive to see how much topics of um impressive to see how much topics of um impressive to see how much topics of um you know safety ethics and so on come to you know safety ethics and so on come to you know safety ethics and so on come to to the surface which is great and if you to the surface which is great and if you to the surface which is great and if you were too early clearly it's fine i mean were too early clearly it's fine i mean were too early clearly it's fine i mean it's a big field and there's lots of it's a big field and there's lots of it's a big field and there's lots of people um with lots of um interest that people um with lots of um interest that people um with lots of um interest that will do progress or make progress um and will do progress or make progress um and will do progress or make progress um and obviously i don't believe we're too late obviously i don't believe we're too late obviously i don't believe we're too late so in that sense like i think it's great so in that sense like i think it's great so in that sense like i think it's great that we're doing this already it's that we're doing this already it's that we're doing this already it's better be too early than yeah too late better be too early than yeah too late better be too early than yeah too late when it comes to super intelligent when it comes to super intelligent when it comes to super intelligent systems let me ask speaking of sentient systems let me ask speaking of sentient systems let me ask speaking of sentient ais you gave props to your friend alias

ais you gave props to your friend alias ais you gave props to your friend alias giver giver giver for being elected uh the fellow of the for being elected uh the fellow of the for being elected uh the fellow of the world society so just as a shout out to world society so just as a shout out to world society so just as a shout out to a fellow researcher and a friend what's a fellow researcher and a friend what's a fellow researcher and a friend what's the secret to the genius of elias the secret to the genius of elias the secret to the genius of elias discover discover discover and also do you believe that his tweets and also do you believe that his tweets and also do you believe that his tweets of as youth hypothesized and andre of as youth hypothesized and andre of as youth hypothesized and andre kapathi did as well are generated by a kapathi did as well are generated by a kapathi did as well are generated by a language model uh yeah language model uh yeah language model uh yeah so so so i i i i strongly believe i ilia is going to i strongly believe i ilia is going to i strongly believe i ilia is going to visit in a few weeks actually so i'll visit in a few weeks actually so i'll visit in a few weeks actually so i'll ask him in person um ask him in person um ask him in person um but will he tell you the truth yes of but will he tell you the truth yes of but will he tell you the truth yes of course yeah absolutely i mean we're you course yeah absolutely i mean we're you course yeah absolutely i mean we're you know ultimately we we all have share know ultimately we we all have share know ultimately we we all have share paths and and there's friendships that paths and and there's friendships that paths and and there's friendships that go beyond obviously institutional go beyond obviously institutional go beyond obviously institutional institutions and and so on so i hope he institutions and and so on so i hope he institutions and and so on so i hope he tells me the truth well maybe the ai tells me the truth well maybe the ai tells me the truth well maybe the ai system is holding him hostage somehow system is holding him hostage somehow system is holding him hostage somehow maybe he has some videos about he maybe he has some videos about he maybe he has some videos about he doesn't want to release so maybe doesn't want to release so maybe doesn't want to release so maybe he it has taken control over him so he he it has taken control over him so he he it has taken control over him so he well i if i see him in person then he well i if i see him in person then he well i if i see him in person then he will he will know yeah will he will know yeah will he will know yeah but but i think the but but i think the but but i think the um i think it's a good i think elia's um i think it's a good i think elia's um i think it's a good i think elia's personality just knowing him for a while personality just knowing him for a while personality just knowing him for a while um um um yeah he's yeah he's yeah he's he's everyone in twitter i guess gets a he's everyone in twitter i guess gets a he's everyone in twitter i guess gets a different persona and and i think elias different persona and and i think elias different persona and and i think elias one one one um um um does not surprise me right so i think does not surprise me right so i think does not surprise me right so i think knowing ilia from before social media knowing ilia from before social media knowing ilia from before social media and before ai was so prevalent i and before ai was so prevalent i and before ai was so prevalent i recognized a lot of his characters so recognized a lot of his characters so recognized a lot of his characters so that's something for me that i feel good that's something for me that i feel good that's something for me that i feel good about a friend that hasn't changed or about a friend that hasn't changed or about a friend that hasn't changed or like is still true to himself right um like is still true to himself right um like is still true to himself right um obviously there is there is though a

obviously there is there is though a obviously there is there is though a fact that fact that fact that your field becomes more popular and he your field becomes more popular and he your field becomes more popular and he is obviously one of the main figures in is obviously one of the main figures in is obviously one of the main figures in the field having done a lot of the field having done a lot of the field having done a lot of advancement so i think that the tricky advancement so i think that the tricky advancement so i think that the tricky bit here is how to balance your true bit here is how to balance your true bit here is how to balance your true self with the responsibility that your self with the responsibility that your self with the responsibility that your words carry so words carry so words carry so in this sense i think yeah like i i i in this sense i think yeah like i i i in this sense i think yeah like i i i appreciate the style and i understand it appreciate the style and i understand it appreciate the style and i understand it but um but um but um it created debates on like some some of it created debates on like some some of it created debates on like some some of his tweets right that maybe it's good we his tweets right that maybe it's good we his tweets right that maybe it's good we have them early anyways right but um but have them early anyways right but um but have them early anyways right but um but yeah it's it's then the reactions are yeah it's it's then the reactions are yeah it's it's then the reactions are usually polarizing i think we're just usually polarizing i think we're just usually polarizing i think we're just seeing kind of the reality of social seeing kind of the reality of social seeing kind of the reality of social media a bit there as well reflected on media a bit there as well reflected on media a bit there as well reflected on on that on that particular topic or set on that on that particular topic or set on that on that particular topic or set of topics he's tweeting about yeah i of topics he's tweeting about yeah i of topics he's tweeting about yeah i mean it's funny that he speak to this mean it's funny that he speak to this mean it's funny that he speak to this tension he was one of the early tension he was one of the early tension he was one of the early seminal figures in the field of deep seminal figures in the field of deep seminal figures in the field of deep learning and so there's a responsibility learning and so there's a responsibility learning and so there's a responsibility with that but he's also with that but he's also with that but he's also from having interacted with him quite a from having interacted with him quite a from having interacted with him quite a bit bit bit he's just a brilliant thinker about he's just a brilliant thinker about he's just a brilliant thinker about ideas ideas ideas and um and um and um which which which as as are you and that there's a tension as as are you and that there's a tension as as are you and that there's a tension between becoming the manager versus like between becoming the manager versus like between becoming the manager versus like the actual thinking through very novel the actual thinking through very novel the actual thinking through very novel ideas ideas ideas the the the yeah the the scientist versus the yeah the the scientist versus the yeah the the scientist versus the manager manager manager and he's and he's and he's uh he's one of the great scientists of uh he's one of the great scientists of uh he's one of the great scientists of our time this was quite interesting and our time this was quite interesting and our time this was quite interesting and also people tell me quite silly which i also people tell me quite silly which i also people tell me quite silly which i haven't quite detected yet but um in haven't quite detected yet but um in haven't quite detected yet but um in private we'll have to see about that private we'll have to see about that private we'll have to see about that yeah yeah yeah yeah i mean just just on the point of i

yeah i mean just just on the point of i yeah i mean just just on the point of i mean ilia has been a mean ilia has been a mean ilia has been a inspiration inspiration inspiration um i mean quite a few colleagues i can um i mean quite a few colleagues i can um i mean quite a few colleagues i can think shaped you know the person you are think shaped you know the person you are think shaped you know the person you are like ilia certainly like ilia certainly like ilia certainly gets probably the top spot if not close gets probably the top spot if not close gets probably the top spot if not close to the top and to the top and to the top and if we go back to the question about if we go back to the question about if we go back to the question about people in the fields like how the role people in the fields like how the role people in the fields like how the role would have changed the field or not i would have changed the field or not i would have changed the field or not i think ilia's case is interesting because think ilia's case is interesting because think ilia's case is interesting because he really has a deep belief in the he really has a deep belief in the he really has a deep belief in the scaling up of neural networks there was scaling up of neural networks there was scaling up of neural networks there was a talk that that that is still famous to a talk that that that is still famous to a talk that that that is still famous to this day um from the sequence to this day um from the sequence to this day um from the sequence to sequence paper um where where he was sequence paper um where where he was sequence paper um where where he was just claiming just give me just claiming just give me just claiming just give me supervised data and large neural network supervised data and large neural network supervised data and large neural network and then you know you'll solve basically and then you know you'll solve basically and then you know you'll solve basically all the problems right that that that all the problems right that that that all the problems right that that that vision right was already already there vision right was already already there vision right was already already there many years ago so it's it's good to see many years ago so it's it's good to see many years ago so it's it's good to see like someone who's in this case very like someone who's in this case very like someone who's in this case very deeply deeply deeply into this style of research um and into this style of research um and into this style of research um and clearly has had a tremendous clearly has had a tremendous clearly has had a tremendous track record of successes and so on um track record of successes and so on um track record of successes and so on um the funny bit about that talk is that we the funny bit about that talk is that we the funny bit about that talk is that we rehearsed the talk in a hotel room rehearsed the talk in a hotel room rehearsed the talk in a hotel room before and before and before and the original version of that talk would the original version of that talk would the original version of that talk would have been even more controversial so have been even more controversial so have been even more controversial so maybe i'm i'm the only person that has maybe i'm i'm the only person that has maybe i'm i'm the only person that has seen the unfiltered version of the talk seen the unfiltered version of the talk seen the unfiltered version of the talk um and you know maybe when the time um and you know maybe when the time um and you know maybe when the time comes maybe we should revisit some of comes maybe we should revisit some of comes maybe we should revisit some of the the the the skip slides from from the from from the skip slides from from the from from the skip slides from from the from from the talk from emilia but i really think the talk from emilia but i really think the talk from emilia but i really think um

um um the deep belief into some certain style the deep belief into some certain style the deep belief into some certain style of research pays out right is is is good of research pays out right is is is good of research pays out right is is is good to be practical sometimes and i actually to be practical sometimes and i actually to be practical sometimes and i actually think ilya and myself are like practical think ilya and myself are like practical think ilya and myself are like practical but it's also good there's some sort of but it's also good there's some sort of but it's also good there's some sort of long-term long-term long-term belief and trajectory um obviously belief and trajectory um obviously belief and trajectory um obviously there's a bit of luck involved but it there's a bit of luck involved but it there's a bit of luck involved but it might be that that's the right path then might be that that's the right path then might be that that's the right path then you clearly are ahead and and hugely you clearly are ahead and and hugely you clearly are ahead and and hugely influential to the field as he has been influential to the field as he has been influential to the field as he has been do you agree with that intuition that do you agree with that intuition that do you agree with that intuition that maybe uh maybe uh maybe uh was was was written about by rich sutton in written about by rich sutton in written about by rich sutton in the the bitter lesson that the biggest the the bitter lesson that the biggest the the bitter lesson that the biggest lesson that can be read from 70 years of lesson that can be read from 70 years of lesson that can be read from 70 years of ai research is that general methods that ai research is that general methods that ai research is that general methods that leverage computation are ultimately the leverage computation are ultimately the leverage computation are ultimately the most effective most effective most effective do you think do you think do you think that intuition is ultimately correct that intuition is ultimately correct that intuition is ultimately correct general methods general methods general methods leverage computation leverage computation leverage computation allowing the scaling of computation to allowing the scaling of computation to allowing the scaling of computation to do a lot of the work do a lot of the work do a lot of the work and so you the basic task of us humans and so you the basic task of us humans and so you the basic task of us humans is to design methods that are more and is to design methods that are more and is to design methods that are more and more general versus more and more more general versus more and more more general versus more and more specific to the tasks at hand specific to the tasks at hand specific to the tasks at hand i i certainly think this i i certainly think this i i certainly think this essentially mimics a bit of the deep essentially mimics a bit of the deep essentially mimics a bit of the deep learning learning learning um research um research um research um um um almost like philosophy almost like philosophy almost like philosophy that that that on the one hand we want to be data on the one hand we want to be data on the one hand we want to be data agnostic we don't want to pre-process agnostic we don't want to pre-process agnostic we don't want to pre-process data sets we want to see the bytes right data sets we want to see the bytes right data sets we want to see the bytes right like the true data as it is and then like the true data as it is and then like the true data as it is and then learn everything on top so learn everything on top so learn everything on top so very much agree with that very much agree with that very much agree with that and i think scaling up feels at the very

and i think scaling up feels at the very and i think scaling up feels at the very least again necessary for least again necessary for least again necessary for building incredible complex systems um building incredible complex systems um building incredible complex systems um it's possibly not sufficient it's possibly not sufficient it's possibly not sufficient bearing that we need a couple of bearing that we need a couple of bearing that we need a couple of breakthroughs um i think rich saturn breakthroughs um i think rich saturn breakthroughs um i think rich saturn mentioned mentioned mentioned search being part of the equation of search being part of the equation of search being part of the equation of skill skill and search i think search skill skill and search i think search skill skill and search i think search i've seen it i've seen it i've seen it that's been more mixed in my experience that's been more mixed in my experience that's been more mixed in my experience so from that lesson in particular search so from that lesson in particular search so from that lesson in particular search is a bit more tricky because is a bit more tricky because is a bit more tricky because it is very appealing to search in it is very appealing to search in it is very appealing to search in domains like go where you have a clear domains like go where you have a clear domains like go where you have a clear reward function that you can then reward function that you can then reward function that you can then discard some search traces discard some search traces discard some search traces but then but then but then in some other tasks it's not very clear in some other tasks it's not very clear in some other tasks it's not very clear how you would do that although recently how you would do that although recently how you would do that although recently one of our one of our one of our recent works which actually was mostly recent works which actually was mostly recent works which actually was mostly mimicking mimicking mimicking or a continuation and even the team and or a continuation and even the team and or a continuation and even the team and the people involved were pretty much uh the people involved were pretty much uh the people involved were pretty much uh very like intersecting with alpha star very like intersecting with alpha star very like intersecting with alpha star was alpha code in which we actually saw was alpha code in which we actually saw was alpha code in which we actually saw the bitter lesson how scale of the the bitter lesson how scale of the the bitter lesson how scale of the models and then a massive amount of models and then a massive amount of models and then a massive amount of search yielded this kind of very search yielded this kind of very search yielded this kind of very interesting result of being able to interesting result of being able to interesting result of being able to have human level code competition so have human level code competition so have human level code competition so i've seen examples of it being literally i've seen examples of it being literally i've seen examples of it being literally mapped to search and scale um i'm not so mapped to search and scale um i'm not so mapped to search and scale um i'm not so convinced about the search bit but convinced about the search bit but convinced about the search bit but certainly i'm convinced skill will be certainly i'm convinced skill will be certainly i'm convinced skill will be needed so we need general methods we needed so we need general methods we needed so we need general methods we need to test them and maybe we need to need to test them and maybe we need to need to test them and maybe we need to make sure that we can scale them given make sure that we can scale them given make sure that we can scale them given the hardware that we have in practice the hardware that we have in practice the hardware that we have in practice but then maybe we should also shape how but then maybe we should also shape how but then maybe we should also shape how the hardware looks like um based on

the hardware looks like um based on the hardware looks like um based on which methods might be needed to scale which methods might be needed to scale which methods might be needed to scale and that's an interesting and that's an interesting and that's an interesting and an interesting contrast of these gpu and an interesting contrast of these gpu and an interesting contrast of these gpu comments that is we got it for free comments that is we got it for free comments that is we got it for free almost because games were using this but almost because games were using this but almost because games were using this but maybe now if sparsity is required maybe now if sparsity is required maybe now if sparsity is required we don't have the hardware although in we don't have the hardware although in we don't have the hardware although in theory i mean many people are building theory i mean many people are building theory i mean many people are building different kinds of hardware these days different kinds of hardware these days different kinds of hardware these days but there's a bit of this notion of but there's a bit of this notion of but there's a bit of this notion of hardware lottery for scale that might hardware lottery for scale that might hardware lottery for scale that might actually actually actually have an impact at least on the year have an impact at least on the year have an impact at least on the year again scale of years on how fast we'll again scale of years on how fast we'll again scale of years on how fast we'll make progress to to maybe a version of make progress to to maybe a version of make progress to to maybe a version of neural nets or or whatever comes next neural nets or or whatever comes next neural nets or or whatever comes next that that that might enable might enable might enable truly intelligent agents truly intelligent agents truly intelligent agents do you think in your lifetime we will do you think in your lifetime we will do you think in your lifetime we will build an agi system build an agi system build an agi system that that that would would would undeniably be a thing that achieves undeniably be a thing that achieves undeniably be a thing that achieves human level intelligence and goes far human level intelligence and goes far human level intelligence and goes far beyond beyond beyond i definitely think it's possible i definitely think it's possible i definitely think it's possible um um um that it will go far beyond but i'm that it will go far beyond but i'm that it will go far beyond but i'm definitely convinced that it will be definitely convinced that it will be definitely convinced that it will be human-level intelligence human-level intelligence human-level intelligence um and i'm i'm hypothesizing about the um and i'm i'm hypothesizing about the um and i'm i'm hypothesizing about the beyond because beyond because beyond because the beyond beat the beyond beat the beyond beat is a bit tricky to define is a bit tricky to define is a bit tricky to define especially when we look at the current especially when we look at the current especially when we look at the current formula of formula of formula of starting from this imitation learning starting from this imitation learning starting from this imitation learning standpoint right so we can certainly standpoint right so we can certainly standpoint right so we can certainly imitate imitate imitate humans humans humans um at language and beyond um at language and beyond um at language and beyond so getting at human level through so getting at human level through so getting at human level through imitation feels very possible

imitation feels very possible imitation feels very possible going beyond going beyond going beyond will require reinforcement learning and will require reinforcement learning and will require reinforcement learning and other things and i think in some areas other things and i think in some areas other things and i think in some areas that certainly already has paid out i that certainly already has paid out i that certainly already has paid out i mean go being an example that's my mean go being an example that's my mean go being an example that's my favorite so far in terms of going beyond favorite so far in terms of going beyond favorite so far in terms of going beyond human capabilities but in general human capabilities but in general human capabilities but in general i'm not sure we can define reward i'm not sure we can define reward i'm not sure we can define reward functions functions functions that from a seat of imitating human that from a seat of imitating human that from a seat of imitating human level intelligence that is general and level intelligence that is general and level intelligence that is general and then going beyond um that that bit is then going beyond um that that bit is then going beyond um that that bit is not so clear in my lifetime but not so clear in my lifetime but not so clear in my lifetime but certainly certainly certainly um human level yes and i mean that in um human level yes and i mean that in um human level yes and i mean that in itself is already quite powerful i think itself is already quite powerful i think itself is already quite powerful i think so um going beyond i think it's so um going beyond i think it's so um going beyond i think it's obviously not we're not gonna not try obviously not we're not gonna not try obviously not we're not gonna not try that if if if it then we get to that if if if it then we get to that if if if it then we get to superhuman superhuman superhuman and discovery and advancing the world and discovery and advancing the world and discovery and advancing the world but um but at least human level is also but um but at least human level is also but um but at least human level is also in general is also very very powerful in general is also very very powerful in general is also very very powerful well especially if human level or well especially if human level or well especially if human level or slightly beyond is integrated deeply slightly beyond is integrated deeply slightly beyond is integrated deeply with human society and there's billions with human society and there's billions with human society and there's billions of agents like that of agents like that of agents like that uh do you think there's a singularity uh do you think there's a singularity uh do you think there's a singularity moment beyond which moment beyond which moment beyond which our world will be just our world will be just our world will be just very deeply transformed by these kinds very deeply transformed by these kinds very deeply transformed by these kinds of systems because now you're talking of systems because now you're talking of systems because now you're talking about intelligence systems that are about intelligence systems that are about intelligence systems that are just i mean this is no longer just a just i mean this is no longer just a just i mean this is no longer just a going from going from going from horse and buggy to to the car horse and buggy to to the car horse and buggy to to the car it feels like a very different kind of it feels like a very different kind of it feels like a very different kind of shift shift shift and what it means to be a living entity and what it means to be a living entity and what it means to be a living entity on earth on earth on earth are you afraid are you excited of this

are you afraid are you excited of this are you afraid are you excited of this world i'm afraid if there's a lot more world i'm afraid if there's a lot more world i'm afraid if there's a lot more so i think maybe we'll need to think so i think maybe we'll need to think so i think maybe we'll need to think about about about if we truly get there if we truly get there if we truly get there just just just thinking of thinking of thinking of limited resources like you know humanity limited resources like you know humanity limited resources like you know humanity clearly hits some limits and then clearly hits some limits and then clearly hits some limits and then there's some balance hopefully that there's some balance hopefully that there's some balance hopefully that biologically um the planet is imposing biologically um the planet is imposing biologically um the planet is imposing and we we should actually try to get and we we should actually try to get and we we should actually try to get better at this as we know there's better at this as we know there's better at this as we know there's there's quite a few there's quite a few there's quite a few you know issues with having too many you know issues with having too many you know issues with having too many people um coexisting in a resource people um coexisting in a resource people um coexisting in a resource limited way so for digital entities it's limited way so for digital entities it's limited way so for digital entities it's an interesting question i think such a an interesting question i think such a an interesting question i think such a limit maybe should exist limit maybe should exist limit maybe should exist but maybe it's going to be imposed by but maybe it's going to be imposed by but maybe it's going to be imposed by energy energy energy availability because this also consumes availability because this also consumes availability because this also consumes energy in fact energy in fact energy in fact most systems are most systems are most systems are more inefficient than we are in terms of more inefficient than we are in terms of more inefficient than we are in terms of energy required energy required energy required but definitely i think as a society but definitely i think as a society but definitely i think as a society we'll need to we'll need to we'll need to just work together to find just work together to find just work together to find what would be reasonable in terms of what would be reasonable in terms of what would be reasonable in terms of growth or how we coexist if that is growth or how we coexist if that is growth or how we coexist if that is to happen to happen to happen i am very excited about i am very excited about i am very excited about obviously the aspects of automation that obviously the aspects of automation that obviously the aspects of automation that make people that obviously don't have make people that obviously don't have make people that obviously don't have access to certain resources or knowledge access to certain resources or knowledge access to certain resources or knowledge um um um for them to have those that access i for them to have those that access i for them to have those that access i think those are the applications in a think those are the applications in a think those are the applications in a way that i'm most exciting to see way that i'm most exciting to see way that i'm most exciting to see um and to personally work towards yeah um and to personally work towards yeah um and to personally work towards yeah there's going to be significant there's going to be significant there's going to be significant improvements in productivity and the improvements in productivity and the improvements in productivity and the quality of life across the whole quality of life across the whole quality of life across the whole population which is very interesting but

population which is very interesting but population which is very interesting but i'm looking even far beyond i'm looking even far beyond i'm looking even far beyond us becoming a multi-planetary species us becoming a multi-planetary species us becoming a multi-planetary species and uh just as a quick bet last question and uh just as a quick bet last question and uh just as a quick bet last question do you think do you think do you think as humans become multi-planetary species as humans become multi-planetary species as humans become multi-planetary species go outside our solar system go outside our solar system go outside our solar system all that kind of stuff do you think all that kind of stuff do you think all that kind of stuff do you think there will be more humans or more robots there will be more humans or more robots there will be more humans or more robots in that future world so will humans be in that future world so will humans be in that future world so will humans be the quirky the quirky the quirky uh intelligent being of the past or is uh intelligent being of the past or is uh intelligent being of the past or is there something deeply fundamental to there something deeply fundamental to there something deeply fundamental to human intelligence that's truly special human intelligence that's truly special human intelligence that's truly special where we we will be part of those other where we we will be part of those other where we we will be part of those other planets not just ai systems planets not just ai systems planets not just ai systems i think we'll we're all excited to i think we'll we're all excited to i think we'll we're all excited to build build build agi to agi to agi to empower empower empower or make us more powerful as human or make us more powerful as human or make us more powerful as human species not to say there might be some species not to say there might be some species not to say there might be some hybridization i mean this is obviously hybridization i mean this is obviously hybridization i mean this is obviously speculation but there are companies also speculation but there are companies also speculation but there are companies also trying to trying to trying to um the same way medicine is is making us um the same way medicine is is making us um the same way medicine is is making us better maybe there are other other better maybe there are other other better maybe there are other other things that are yet to happen on that things that are yet to happen on that things that are yet to happen on that but but but if the ratio is not at most one to one i if the ratio is not at most one to one i if the ratio is not at most one to one i would not be happy so i would hope that would not be happy so i would hope that would not be happy so i would hope that we are part of the equation um but maybe we are part of the equation um but maybe we are part of the equation um but maybe there's there's there's maybe a one-to-one ratio feels like maybe a one-to-one ratio feels like maybe a one-to-one ratio feels like possible um constructive and so on but possible um constructive and so on but possible um constructive and so on but it would not be good to have a it would not be good to have a it would not be good to have a misbalance at least from my core beliefs misbalance at least from my core beliefs misbalance at least from my core beliefs and and and the why i'm doing what i'm doing when i the why i'm doing what i'm doing when i the why i'm doing what i'm doing when i go to work and i research what i go to work and i research what i go to work and i research what i research research research well this is how i know you're human and

well this is how i know you're human and well this is how i know you're human and this is how you've passed the turing this is how you've passed the turing this is how you've passed the turing test test test and you are one of the special humans or and you are one of the special humans or and you are one of the special humans or it's a huge honor that you have talked it's a huge honor that you have talked it's a huge honor that you have talked with me and i hope we get the chance to with me and i hope we get the chance to with me and i hope we get the chance to speak again maybe once before the speak again maybe once before the speak again maybe once before the singularity once after and see how our singularity once after and see how our singularity once after and see how our view of the world changes thank you view of the world changes thank you view of the world changes thank you again for talking today thank you for again for talking today thank you for again for talking today thank you for the amazing work you do you're a the amazing work you do you're a the amazing work you do you're a shining example of a researcher and a shining example of a researcher and a shining example of a researcher and a human being in this community thanks a human being in this community thanks a human being in this community thanks a lot lex yeah looking forward to before lot lex yeah looking forward to before lot lex yeah looking forward to before the singularity certainly the singularity certainly the singularity certainly and maybe after and maybe after and maybe after thanks for listening to this thanks for listening to this thanks for listening to this conversation with arielle vignealis to conversation with arielle vignealis to conversation with arielle vignealis to support this podcast please check out support this podcast please check out support this podcast please check out our sponsors in the description and now our sponsors in the description and now our sponsors in the description and now let me leave you with some words from let me leave you with some words from let me leave you with some words from alan turing alan turing alan turing those who can imagine anything can those who can imagine anything can those who can imagine anything can create the impossible create the impossible create the impossible thank you for listening and hope to see thank you for listening and hope to see thank you for listening and hope to see you next time