Adaption Labs: Gradient-Free Continual Learning — Sara Hooker, Adaption

>> Amazing. >> Amazing. Um it is Oh, sorry. Pardon me. Um it is so lovely to be here. So, I Um it is so lovely to be here. So, I wanted to share today uh some thoughts wanted to share today uh some thoughts wanted to share today uh some thoughts that I have around who gets to be at the that I have around who gets to be at the that I have around who gets to be at the frontier of discovery. frontier of discovery. frontier of discovery. So, modern computer science as a field So, modern computer science as a field So, modern computer science as a field has only existed for the last 77 years. has only existed for the last 77 years. has only existed for the last 77 years. It's kind of bizarre when you think It's kind of bizarre when you think It's kind of bizarre when you think about it. So, World War II, about it. So, World War II, about it. So, World War II, uh all the transistor technology that uh all the transistor technology that uh all the transistor technology that was developed for radio, we finally had was developed for radio, we finally had was developed for radio, we finally had our first versions of the computer. But, our first versions of the computer. But, our first versions of the computer. But, when you think about it, that's only two when you think about it, that's only two when you think about it, that's only two generations of people working on these generations of people working on these generations of people working on these tools. However, within that time, even tools. However, within that time, even tools. However, within that time, even for computer science, who and what and for computer science, who and what and for computer science, who and what and what topics we work on has dramatically what topics we work on has dramatically what topics we work on has dramatically changed. And I think it's an interesting changed. And I think it's an interesting changed. And I think it's an interesting setting because actually, if you look setting because actually, if you look setting because actually, if you look back across science as a whole, how we back across science as a whole, how we back across science as a whole, how we do discovery has been markedly different do discovery has been markedly different do discovery has been markedly different uh at different points in time. So, when uh at different points in time. So, when uh at different points in time. So, when we started, the whole idea of a we started, the whole idea of a we started, the whole idea of a researcher was what we call like a researcher was what we call like a researcher was what we call like a gentleman scientist, typically someone gentleman scientist, typically someone gentleman scientist, typically someone with adequate wealth to dabble in a with adequate wealth to dabble in a with adequate wealth to dabble in a discovery. And these were all discovery. And these were all discovery. And these were all individuals, independent researchers.

individuals, independent researchers. individuals, independent researchers. Uh when we uh see like the first Uh when we uh see like the first Uh when we uh see like the first associations emerge with the Royal associations emerge with the Royal associations emerge with the Royal Society in the 1600s, Society in the 1600s, Society in the 1600s, the idea of being a scientist as a the idea of being a scientist as a the idea of being a scientist as a full-time job was very special. And full-time job was very special. And full-time job was very special. And these became the predominant spaces for these became the predominant spaces for these became the predominant spaces for discovery. Why am I talking about this? discovery. Why am I talking about this? discovery. Why am I talking about this? Because to be honest, like the Because to be honest, like the Because to be honest, like the professionalization professionalization professionalization of science uh led to what I would call, of science uh led to what I would call, of science uh led to what I would call, and you know, my good friend Rosanne Liu and you know, my good friend Rosanne Liu and you know, my good friend Rosanne Liu calls, the unreasonably narrow path. So, calls, the unreasonably narrow path. So, calls, the unreasonably narrow path. So, I'm an AI researcher. Uh this is Yans' I'm an AI researcher. Uh this is Yans' I'm an AI researcher. Uh this is Yans' actual career. Uh and it's interesting, actual career. Uh and it's interesting, actual career. Uh and it's interesting, if you were to be an AI researcher at if you were to be an AI researcher at if you were to be an AI researcher at the forefront, you basically had to the forefront, you basically had to the forefront, you basically had to follow this exact narrow path. You had follow this exact narrow path. You had follow this exact narrow path. You had to get into the right PhD program. You to get into the right PhD program. You to get into the right PhD program. You had to then go to the right industry had to then go to the right industry had to then go to the right industry lab. You had to do sufficiently lab. You had to do sufficiently lab. You had to do sufficiently interesting work, and then finally you interesting work, and then finally you interesting work, and then finally you got to contribute to the frontier. This got to contribute to the frontier. This got to contribute to the frontier. This was my story, too. So, a lot of my work was my story, too. So, a lot of my work was my story, too. So, a lot of my work has been on efficiency at scale. has been on efficiency at scale. has been on efficiency at scale. Uh I did my PhD and worked at DeepMind Uh I did my PhD and worked at DeepMind Uh I did my PhD and worked at DeepMind and a lot of different frontier labs, and a lot of different frontier labs, and a lot of different frontier labs, but this was in many ways um a very but this was in many ways um a very but this was in many ways um a very aggressively filtered system. If you did aggressively filtered system. If you did aggressively filtered system. If you did not make it or you were not curious not make it or you were not curious not make it or you were not curious about the right problem at the right about the right problem at the right about the right problem at the right time, you didn't have um a place to play time, you didn't have um a place to play time, you didn't have um a place to play at a frontier lab. And this is like the at a frontier lab. And this is like the at a frontier lab. And this is like the standard successful scientist. Like you standard successful scientist. Like you standard successful scientist. Like you have a famous advisor, hopefully.

have a famous advisor, hopefully. have a famous advisor, hopefully. Uh you hopefully get one or two Uh you hopefully get one or two Uh you hopefully get one or two important internships. important internships. important internships. And what's interesting about this is And what's interesting about this is And what's interesting about this is computer science was really about computer science was really about computer science was really about representing the world, and it was like representing the world, and it was like representing the world, and it was like all the tools that do that. But the all the tools that do that. But the all the tools that do that. But the reason why most people got into computer reason why most people got into computer reason why most people got into computer science is the question at the end of science is the question at the end of science is the question at the end of it. Like if you can represent the world, it. Like if you can represent the world, it. Like if you can represent the world, what questions can you answer? So, today what questions can you answer? So, today what questions can you answer? So, today I'm going to talk about what I think is I'm going to talk about what I think is I'm going to talk about what I think is like one of the most profound and like one of the most profound and like one of the most profound and interesting topics, which is like why interesting topics, which is like why interesting topics, which is like why this is so important when computer this is so important when computer this is so important when computer science at this is changing. So, I'll science at this is changing. So, I'll science at this is changing. So, I'll also speak to this. It was double also speak to this. It was double also speak to this. It was double compounded in computer science because compounded in computer science because compounded in computer science because of the need for compute. of the need for compute. of the need for compute. So, this resulted in jokes. Um it's fun So, this resulted in jokes. Um it's fun So, this resulted in jokes. Um it's fun that Merve was here. This is her tweet that Merve was here. This is her tweet that Merve was here. This is her tweet about GPU poor versus GPU rich. Uh it about GPU poor versus GPU rich. Uh it about GPU poor versus GPU rich. Uh it led to barriers of entry on who can led to barriers of entry on who can led to barriers of entry on who can contribute frontier AI. So, I put here contribute frontier AI. So, I put here contribute frontier AI. So, I put here company A, company B, company C. But to company A, company B, company C. But to company A, company B, company C. But to be honest, if we pulled, I think there be honest, if we pulled, I think there be honest, if we pulled, I think there would be significant majority votes would be significant majority votes would be significant majority votes about who those companies are. But about who those companies are. But about who those companies are. But basically a handful of frontier labs basically a handful of frontier labs basically a handful of frontier labs have been able to build the technology have been able to build the technology have been able to build the technology we use. we use. we use. Um it's also determined who gets to Um it's also determined who gets to Um it's also determined who gets to participate in breakthroughs and who participate in breakthroughs and who participate in breakthroughs and who doesn't. So, uh this is a map of like doesn't. So, uh this is a map of like doesn't. So, uh this is a map of like what Stanford calls uh where what Stanford calls uh where what Stanford calls uh where statistically or significant statistically or significant statistically or significant breakthroughs have come through come breakthroughs have come through come breakthroughs have come through come from, and you can see the whole sections from, and you can see the whole sections from, and you can see the whole sections of the world are completely left out.

of the world are completely left out. of the world are completely left out. And so for me, uh this is a very And so for me, uh this is a very And so for me, uh this is a very important question worth answering. Who important question worth answering. Who important question worth answering. Who gets to shape the frontier? Who gets to gets to shape the frontier? Who gets to gets to shape the frontier? Who gets to answer the questions at the end of the answer the questions at the end of the answer the questions at the end of the pursuit? We've seen that the shift has pursuit? We've seen that the shift has pursuit? We've seen that the shift has dramatically changed from academia to dramatically changed from academia to dramatically changed from academia to industry. industry. industry. And it's also meant that we ship the And it's also meant that we ship the And it's also meant that we ship the same model to everyone. Uh same model to everyone. Uh same model to everyone. Uh why that's why that's why that's particularly interesting is that most particularly interesting is that most particularly interesting is that most people intuitively understand that you people intuitively understand that you people intuitively understand that you shouldn't ship the same model to shouldn't ship the same model to shouldn't ship the same model to billions of people. And they also billions of people. And they also billions of people. And they also understand that it's not a particularly understand that it's not a particularly understand that it's not a particularly good use of compute, right? You're good use of compute, right? You're good use of compute, right? You're spending the same amount of compute on spending the same amount of compute on spending the same amount of compute on everything. And some problems are hard everything. And some problems are hard everything. And some problems are hard and some are very easy. and some are very easy. and some are very easy. So where does that leave us? What's my So where does that leave us? What's my So where does that leave us? What's my talk for today? Um I would like to say talk for today? Um I would like to say talk for today? Um I would like to say that we are ripe for a revolution. And that we are ripe for a revolution. And that we are ripe for a revolution. And we are ripe for a revolution in who gets we are ripe for a revolution in who gets we are ripe for a revolution in who gets to participate at the frontier of AI. to participate at the frontier of AI. to participate at the frontier of AI. Uh I'll tell you two reasons why I'm Uh I'll tell you two reasons why I'm Uh I'll tell you two reasons why I'm bullish on this. Um I'll definitely bullish on this. Um I'll definitely bullish on this. Um I'll definitely cover one, and then I'll actually see cover one, and then I'll actually see cover one, and then I'll actually see interest and timing cuz I want to leave interest and timing cuz I want to leave interest and timing cuz I want to leave plenty of time for questions. I think plenty of time for questions. I think plenty of time for questions. I think that's you know, unless you have a few that's you know, unless you have a few that's you know, unless you have a few questions and a bit of banter, these questions and a bit of banter, these questions and a bit of banter, these things can be kind of boring. So we'll things can be kind of boring. So we'll things can be kind of boring. So we'll see, but I'll cover one definitely that see, but I'll cover one definitely that see, but I'll cover one definitely that I'm actively thinking about. And this is I'm actively thinking about. And this is I'm actively thinking about. And this is what if we could allow anyone to build what if we could allow anyone to build what if we could allow anyone to build the same frontier intelligence um as the same frontier intelligence um as the same frontier intelligence um as that in labs. And I've been in a few that in labs. And I've been in a few that in labs. And I've been in a few labs. I've done my tour of duty. And labs. I've done my tour of duty. And labs. I've done my tour of duty. And this is the core question I care about this is the core question I care about this is the core question I care about now is like how do you build now is like how do you build now is like how do you build intelligence that continuously adapts intelligence that continuously adapts intelligence that continuously adapts and that builders everywhere can have and that builders everywhere can have and that builders everywhere can have more control.

more control. more control. So instead of taking years of training So instead of taking years of training So instead of taking years of training to learn how to build the tools, to learn how to build the tools, to learn how to build the tools, scientists just skipped the questions. scientists just skipped the questions. scientists just skipped the questions. So a few weeks ago we released Auto So a few weeks ago we released Auto So a few weeks ago we released Auto Scientist. And Auto Scientist is really Scientist. And Auto Scientist is really Scientist. And Auto Scientist is really about how do you automate the training about how do you automate the training about how do you automate the training of models itself. Um I'll share a few of models itself. Um I'll share a few of models itself. Um I'll share a few things that are really interesting about things that are really interesting about things that are really interesting about this is that one it's this is that one it's this is that one it's um co-optimizes the entire loop. So it's um co-optimizes the entire loop. So it's um co-optimizes the entire loop. So it's like from data to alignment uh and it it like from data to alignment uh and it it like from data to alignment uh and it it chooses and self-evolves based upon the chooses and self-evolves based upon the chooses and self-evolves based upon the domain and the type of data. domain and the type of data. domain and the type of data. Um what's interesting as well is that it Um what's interesting as well is that it Um what's interesting as well is that it actually outperforms research staff. And actually outperforms research staff. And actually outperforms research staff. And mainly because like a lot of our mainly because like a lot of our mainly because like a lot of our research staff has experience with research staff has experience with research staff has experience with certain model types, and we're testing certain model types, and we're testing certain model types, and we're testing it across many different model it across many different model it across many different model architectures, different different size architectures, different different size architectures, different different size models, um as well as like dense and models, um as well as like dense and models, um as well as like dense and mixture of experts. And that search mixture of experts. And that search mixture of experts. And that search space is a lot broader. And so, space is a lot broader. And so, space is a lot broader. And so, exploiting it using like how do you how exploiting it using like how do you how exploiting it using like how do you how do you self-improve from experience and do you self-improve from experience and do you self-improve from experience and scale is very effective. scale is very effective. scale is very effective. Um I think this is very interesting. Um I think this is very interesting. Um I think this is very interesting. This only worked when we co-optimized This only worked when we co-optimized This only worked when we co-optimized the data. Uh so, there's a lot of auto the data. Uh so, there's a lot of auto the data. Uh so, there's a lot of auto research projects right now, which research projects right now, which research projects right now, which basically treat data as like basically treat data as like basically treat data as like um the agent is it decides whether to um the agent is it decides whether to um the agent is it decides whether to create data or not or what to do.

create data or not or what to do. create data or not or what to do. Frankly, we did not get the returns for Frankly, we did not get the returns for Frankly, we did not get the returns for like how much you can squeeze out of like how much you can squeeze out of like how much you can squeeze out of performance until you control for data performance until you control for data performance until you control for data quality. So, we actually co-optimized quality. So, we actually co-optimized quality. So, we actually co-optimized based on all the adaptation we did with based on all the adaptation we did with based on all the adaptation we did with the data exactly what we would do with the data exactly what we would do with the data exactly what we would do with the model. And that was super the model. And that was super the model. And that was super interesting. It speaks to like the need interesting. It speaks to like the need interesting. It speaks to like the need to control the entire flow. Um What's to control the entire flow. Um What's to control the entire flow. Um What's fun about this is like really what auto fun about this is like really what auto fun about this is like really what auto scientists do and it's a combines all scientists do and it's a combines all scientists do and it's a combines all the knowledge it gained from the the knowledge it gained from the the knowledge it gained from the adaptive data component with also the adaptive data component with also the adaptive data component with also the knowledge of the domain and also the knowledge of the domain and also the knowledge of the domain and also the ability to self-improve for a domain and ability to self-improve for a domain and ability to self-improve for a domain and to learn from other components of that to learn from other components of that to learn from other components of that domain. Um What What is a cheeky fact domain. Um What What is a cheeky fact domain. Um What What is a cheeky fact and this is quite fun. You'll notice all and this is quite fun. You'll notice all and this is quite fun. You'll notice all these percentages for win rates are like these percentages for win rates are like these percentages for win rates are like 60 plus. Um that's because we put the 60 plus. Um that's because we put the 60 plus. Um that's because we put the budget stopping it stopping it above 60. budget stopping it stopping it above 60. budget stopping it stopping it above 60. So, like once it was above 60, the our So, like once it was above 60, the our So, like once it was above 60, the our our genetic flow could exit. But we our genetic flow could exit. But we our genetic flow could exit. But we since like removed that barrier and like since like removed that barrier and like since like removed that barrier and like you can see it just go up over time, you can see it just go up over time, you can see it just go up over time, which is super fascinating. which is super fascinating. which is super fascinating. Um and then I think what's interesting Um and then I think what's interesting Um and then I think what's interesting is like it changes a lot of the is like it changes a lot of the is like it changes a lot of the hyperparameters. Typically, that humans hyperparameters. Typically, that humans hyperparameters. Typically, that humans are much more wary about changing all at are much more wary about changing all at are much more wary about changing all at once. And so, you get massive once. And so, you get massive once. And so, you get massive exploitation of the search space. I see exploitation of the search space. I see exploitation of the search space. I see this is crucial for like how do you this is crucial for like how do you this is crucial for like how do you reduce the amount of compute you use for reduce the amount of compute you use for reduce the amount of compute you use for customization because you train with customization because you train with customization because you train with much more predictability. But also, how much more predictability. But also, how much more predictability. But also, how do you um leverage like your domain do you um leverage like your domain do you um leverage like your domain knowledge to really unlock how you build knowledge to really unlock how you build knowledge to really unlock how you build frontier AI. Um and this was fun. We we frontier AI. Um and this was fun. We we frontier AI. Um and this was fun. We we did uh we announced a beta like 4 weeks did uh we announced a beta like 4 weeks did uh we announced a beta like 4 weeks ago. The excitement is most acute for ago. The excitement is most acute for ago. The excitement is most acute for like medical and sciences. And that's like medical and sciences. And that's like medical and sciences. And that's largely I think and legal and and code largely I think and legal and and code largely I think and legal and and code as well. But these are like domains

as well. But these are like domains as well. But these are like domains where typically current models fall where typically current models fall where typically current models fall short. And also domains where um in many short. And also domains where um in many short. And also domains where um in many ways like the degree of last mile ways like the degree of last mile ways like the degree of last mile customization is really acute. Um and customization is really acute. Um and customization is really acute. Um and this is the core point. I you know, I this is the core point. I you know, I this is the core point. I you know, I think this is fun. I think I might even think this is fun. I think I might even think this is fun. I think I might even have time to cover have time to cover have time to cover The other point I wanted to make about The other point I wanted to make about The other point I wanted to make about why now is very important for like why now is very important for like why now is very important for like changing who shapes. But this really the changing who shapes. But this really the changing who shapes. But this really the main factor that this does is it main factor that this does is it main factor that this does is it increases your innovation cycle. Like increases your innovation cycle. Like increases your innovation cycle. Like and it also increases the likelihood and it also increases the likelihood and it also increases the likelihood that when you train and spend compute that when you train and spend compute that when you train and spend compute you'll succeed. And those combined you'll succeed. And those combined you'll succeed. And those combined factors are super interesting. Like one factors are super interesting. Like one factors are super interesting. Like one thing that we're doing next is extending thing that we're doing next is extending thing that we're doing next is extending that. So even your test time compute that. So even your test time compute that. So even your test time compute should be adaptive based on your task. should be adaptive based on your task. should be adaptive based on your task. Um Um Um So this kind of brings me back to where So this kind of brings me back to where So this kind of brings me back to where I started. And like kind of the grumpy I started. And like kind of the grumpy I started. And like kind of the grumpy statement I said, which is like you statement I said, which is like you statement I said, which is like you know, we have this rarefied super narrow know, we have this rarefied super narrow know, we have this rarefied super narrow compounding issue of barriers to entry. compounding issue of barriers to entry. compounding issue of barriers to entry. One is like that you need to do this um One is like that you need to do this um One is like that you need to do this um very narrow funnel of who gets to build very narrow funnel of who gets to build very narrow funnel of who gets to build frontier AI. And the other is that um frontier AI. And the other is that um frontier AI. And the other is that um typically compute and cost really typically compute and cost really typically compute and cost really dominate. Um we want to change that.

dominate. Um we want to change that. dominate. Um we want to change that. Like we decided okay, we're going to Like we decided okay, we're going to Like we decided okay, we're going to cover languages from day one, 242 cover languages from day one, 242 cover languages from day one, 242 languages. And also like a big interest languages. And also like a big interest languages. And also like a big interest for us is actually non-verifiable tasks. for us is actually non-verifiable tasks. for us is actually non-verifiable tasks. I think this is super interesting I think this is super interesting I think this is super interesting because um this is really the bulk of because um this is really the bulk of because um this is really the bulk of like everyday tasks that people do. And like everyday tasks that people do. And like everyday tasks that people do. And um it's really where the meat of like um it's really where the meat of like um it's really where the meat of like what is interesting for progress is what is interesting for progress is what is interesting for progress is going to be over the next year. going to be over the next year. going to be over the next year. Um and this leads me into our mandate. Um and this leads me into our mandate. Um and this leads me into our mandate. We care deeply about how do you We care deeply about how do you We care deeply about how do you accelerate learning in a way that models accelerate learning in a way that models accelerate learning in a way that models should be able to um learn from their should be able to um learn from their should be able to um learn from their environment. So right now we've moved environment. So right now we've moved environment. So right now we've moved from a era of like the the model is from a era of like the the model is from a era of like the the model is monolithic. You know, when I was at um monolithic. You know, when I was at um monolithic. You know, when I was at um different parts of my research career, different parts of my research career, different parts of my research career, basically a whole team would be around basically a whole team would be around basically a whole team would be around building a model. You give it to someone building a model. You give it to someone building a model. You give it to someone else to serve, and like you have someone else to serve, and like you have someone else to serve, and like you have someone else do the front end. And actually now, else do the front end. And actually now, else do the front end. And actually now, like the most important intelligence is like the most important intelligence is like the most important intelligence is a model who interacts. And so, this idea a model who interacts. And so, this idea a model who interacts. And so, this idea of how efficiently are you going to of how efficiently are you going to of how efficiently are you going to interact, how will you continuously interact, how will you continuously interact, how will you continuously learn from the environment, is pretty learn from the environment, is pretty learn from the environment, is pretty core. And um I think about it a lot. So, core. And um I think about it a lot. So, core. And um I think about it a lot. So, um um um let's see. I think I do have time, let's see. I think I do have time, let's see. I think I do have time, right? How are we doing for time? Oh, I right? How are we doing for time? Oh, I right? How are we doing for time? Oh, I do. I have plenty. This is lovely. So, do. I have plenty. This is lovely. So, do. I have plenty. This is lovely. So, we'll have time for questions, and I'll we'll have time for questions, and I'll we'll have time for questions, and I'll share a little bit about what I think um share a little bit about what I think um share a little bit about what I think um the next uh the next uh the next uh component is. I think core to this, so component is. I think core to this, so component is. I think core to this, so if we just did auto scientist, but it if we just did auto scientist, but it if we just did auto scientist, but it still took enormous compute to do still took enormous compute to do still took enormous compute to do frontier AI trainings, I think we'd be frontier AI trainings, I think we'd be frontier AI trainings, I think we'd be in a bit of a pickle, right? Like I'd be in a bit of a pickle, right? Like I'd be in a bit of a pickle, right? Like I'd be saying, "Oh great, you can use this saying, "Oh great, you can use this saying, "Oh great, you can use this agent, but don't worry, just bring your agent, but don't worry, just bring your agent, but don't worry, just bring your uh 10,000 GPUs with you."

uh 10,000 GPUs with you." uh 10,000 GPUs with you." But, I think there's another trend, But, I think there's another trend, But, I think there's another trend, which I think makes this very important which I think makes this very important which I think makes this very important timing, and rooms like this probably timing, and rooms like this probably timing, and rooms like this probably much more optimistic than like have been much more optimistic than like have been much more optimistic than like have been in a few years ago about who can build in a few years ago about who can build in a few years ago about who can build frontier AI. Um and one of that is like frontier AI. Um and one of that is like frontier AI. Um and one of that is like the rules of like where you get rate of the rules of like where you get rate of the rules of like where you get rate of return for computer are totally return for computer are totally return for computer are totally changing. So, um changing. So, um changing. So, um I wrote a very paper about this called I wrote a very paper about this called I wrote a very paper about this called So Death of Scaling. Um but empirically, So Death of Scaling. Um but empirically, So Death of Scaling. Um but empirically, we do now know that um pre-training size we do now know that um pre-training size we do now know that um pre-training size in particular is not your most lucrative in particular is not your most lucrative in particular is not your most lucrative axis of scale. And what does this mean? axis of scale. And what does this mean? axis of scale. And what does this mean? Like if pre-training scale isn't going Like if pre-training scale isn't going Like if pre-training scale isn't going to dominate performance, it actually to dominate performance, it actually to dominate performance, it actually really greatly changes who can create really greatly changes who can create really greatly changes who can create the best recipes for innovation, because the best recipes for innovation, because the best recipes for innovation, because pre-training compute typically has to be pre-training compute typically has to be pre-training compute typically has to be co-located. Um it has to be uh in many co-located. Um it has to be uh in many co-located. Um it has to be uh in many ways large volume to accommodate for ways large volume to accommodate for ways large volume to accommodate for redundancy. Inference compute and other redundancy. Inference compute and other redundancy. Inference compute and other places where you actually apply compute, places where you actually apply compute, places where you actually apply compute, typically you can have much more like typically you can have much more like typically you can have much more like distributed. It's also much more higher distributed. It's also much more higher distributed. It's also much more higher return given the amount of flops. And return given the amount of flops. And return given the amount of flops. And so, it's interesting when we talk about so, it's interesting when we talk about so, it's interesting when we talk about what is the state of pre-training what is the state of pre-training what is the state of pre-training compute, we know it's not giving the compute, we know it's not giving the compute, we know it's not giving the same returns largely because our same returns largely because our same returns largely because our architecture is saturated. So, um we see architecture is saturated. So, um we see architecture is saturated. So, um we see much smaller models outperforming much much smaller models outperforming much much smaller models outperforming much larger ones. Uh this is like the Open larger ones. Uh this is like the Open larger ones. Uh this is like the Open LLM leaderboard, and this is like the LLM leaderboard, and this is like the LLM leaderboard, and this is like the daily submission of like the best small daily submission of like the best small daily submission of like the best small model under 13B versus all the larger model under 13B versus all the larger model under 13B versus all the larger models.

models. models. Um and you can see over time that ratio Um and you can see over time that ratio Um and you can see over time that ratio totally flips. Um and also there's kind totally flips. Um and also there's kind totally flips. Um and also there's kind of the grumpy assessment that most of the grumpy assessment that most of the grumpy assessment that most recent models that have severely um recent models that have severely um recent models that have severely um played with just increasing model size played with just increasing model size played with just increasing model size haven't provided the same stepwise haven't provided the same stepwise haven't provided the same stepwise change um as their predecessors. And a change um as their predecessors. And a change um as their predecessors. And a lot of that is because where the most lot of that is because where the most lot of that is because where the most returns for performance are now are on a returns for performance are now are on a returns for performance are now are on a broader action space. broader action space. broader action space. Um and this is really what I was getting Um and this is really what I was getting Um and this is really what I was getting at when we move from an algorithm to we at when we move from an algorithm to we at when we move from an algorithm to we are expanding optimization space in new are expanding optimization space in new are expanding optimization space in new places. And what's the fun about that is places. And what's the fun about that is places. And what's the fun about that is that these are new places where um the that these are new places where um the that these are new places where um the barriers to entry are much more nimble, barriers to entry are much more nimble, barriers to entry are much more nimble, and where recipe and algorithm and and where recipe and algorithm and and where recipe and algorithm and research matters again. Um and things research matters again. Um and things research matters again. Um and things like how do you automate that discovery? like how do you automate that discovery? like how do you automate that discovery? Um and so this is what I'll state, and I Um and so this is what I'll state, and I Um and so this is what I'll state, and I think then we should open up for think then we should open up for think then we should open up for questions. Um and I would encourage good questions. Um and I would encourage good questions. Um and I would encourage good grumpy questions or fun positions. Let's grumpy questions or fun positions. Let's grumpy questions or fun positions. Let's make the use of the time. I know I was make the use of the time. I know I was make the use of the time. I know I was told earlier that almost no talks have told earlier that almost no talks have told earlier that almost no talks have time for questions. I find that so time for questions. I find that so time for questions. I find that so disappointing, so disappointing, so disappointing, so um so we'll need some brave people to um so we'll need some brave people to um so we'll need some brave people to start the conversation. Um but I will start the conversation. Um but I will start the conversation. Um but I will say this means we're better off, and I say this means we're better off, and I say this means we're better off, and I would say it's a very good time to be would say it's a very good time to be would say it's a very good time to be like working on intelligence because like working on intelligence because like working on intelligence because instead of just a handful of people instead of just a handful of people instead of just a handful of people getting to getting to getting to getting to create it, it's much more now getting to create it, it's much more now getting to create it, it's much more now about the question you want to answer at about the question you want to answer at about the question you want to answer at the end of the day. The reason why the end of the day. The reason why the end of the day. The reason why people did a computer science PhD was to people did a computer science PhD was to people did a computer science PhD was to learn the tools to get to the question,

learn the tools to get to the question, learn the tools to get to the question, and now you can just get to the and now you can just get to the and now you can just get to the question, which is super meaningful. Um question, which is super meaningful. Um question, which is super meaningful. Um okay, let me open up. Where should we okay, let me open up. Where should we okay, let me open up. Where should we start? We um have an abundance. I hear start? We um have an abundance. I hear start? We um have an abundance. I hear there's no microphone. So, if you want there's no microphone. So, if you want there's no microphone. So, if you want to ask a question, you want to make a to ask a question, you want to make a to ask a question, you want to make a statement, I will indulge a statement if statement, I will indulge a statement if statement, I will indulge a statement if it's interesting. it's interesting. it's interesting. And yeah, go for it. Just raise your And yeah, go for it. Just raise your And yeah, go for it. Just raise your hand and I'll I'll repeat it afterwards. hand and I'll I'll repeat it afterwards. hand and I'll I'll repeat it afterwards. Let me just get to the end of Let me just get to the end of Let me just get to the end of this in case people want to reach me this in case people want to reach me this in case people want to reach me afterwards. Nice. Yes, go ahead. afterwards. Nice. Yes, go ahead. afterwards. Nice. Yes, go ahead. Gentleman in the fourth row. Go for it. Gentleman in the fourth row. Go for it. Gentleman in the fourth row. Go for it. >> You mentioned >> You mentioned >> You mentioned uh you know uh you know uh you know uh AI from here uh AI from here uh AI from here will be more democratized. will be more democratized. will be more democratized. Can you point to like how? Can you point to like how? Can you point to like how? >> So, I think how it's twofold. One is >> So, I think how it's twofold. One is >> So, I think how it's twofold. One is there's very few people who know how to there's very few people who know how to there's very few people who know how to train frontier models. I would say train frontier models. I would say train frontier models. I would say realistically probably less than 5,000 realistically probably less than 5,000 realistically probably less than 5,000 in the world at scale. I think that type in the world at scale. I think that type in the world at scale. I think that type of knowledge, that's a very exploitable of knowledge, that's a very exploitable of knowledge, that's a very exploitable search space. And actually as humans, search space. And actually as humans, search space. And actually as humans, all those configurations, we're not all those configurations, we're not all those configurations, we're not particularly good at. It's kind of like particularly good at. It's kind of like particularly good at. It's kind of like secret knowledge we pass as if, you secret knowledge we pass as if, you secret knowledge we pass as if, you know, we're apprentices.

know, we're apprentices. know, we're apprentices. So, that's one. Like once you automate a So, that's one. Like once you automate a So, that's one. Like once you automate a lot of that knowledge, you just lot of that knowledge, you just lot of that knowledge, you just accelerate innovation cycles. Which accelerate innovation cycles. Which accelerate innovation cycles. Which means that you can explore and do more means that you can explore and do more means that you can explore and do more questions. questions. questions. Typically, what people I think often Typically, what people I think often Typically, what people I think often miss is that the cost of miss is that the cost of miss is that the cost of asking something informs what is asked. asking something informs what is asked. asking something informs what is asked. And if you make it cheaper to ask And if you make it cheaper to ask And if you make it cheaper to ask something, you change like the volume of something, you change like the volume of something, you change like the volume of things that are asked, which is super things that are asked, which is super things that are asked, which is super interesting. The other reason though, I interesting. The other reason though, I interesting. The other reason though, I do think it's very much a facet of like do think it's very much a facet of like do think it's very much a facet of like the changing nature of compute. So, the changing nature of compute. So, the changing nature of compute. So, agentic compute, post-training compute agentic compute, post-training compute agentic compute, post-training compute matters a significant amount for matters a significant amount for matters a significant amount for performance. That does not require the performance. That does not require the performance. That does not require the same type of um same type of um same type of um I dare I say hoarding of GPUs. I dare I say hoarding of GPUs. I dare I say hoarding of GPUs. >> [laughter] >> [laughter] >> [laughter] >> But like I think it it's very different >> But like I think it it's very different >> But like I think it it's very different compute purchasing dynamics. And again, compute purchasing dynamics. And again, compute purchasing dynamics. And again, it means that the person with the best it means that the person with the best it means that the person with the best idea has a higher chance of winning. idea has a higher chance of winning. idea has a higher chance of winning. Which is fun. Nice. What else? Who wants Which is fun. Nice. What else? Who wants Which is fun. Nice. What else? Who wants to go? I see to go? I see to go? I see Yeah, we can go up here and then I saw a Yeah, we can go up here and then I saw a Yeah, we can go up here and then I saw a hand back there. Okay, yes, I do see hand back there. Okay, yes, I do see hand back there. Okay, yes, I do see you. The glare is high, but you go first you. The glare is high, but you go first you. The glare is high, but you go first and then we'll come up here.

and then we'll come up here. and then we'll come up here. >> So, Frontier Labs care a lot about >> So, Frontier Labs care a lot about >> So, Frontier Labs care a lot about safety. So, one of the challenges in safety. So, one of the challenges in safety. So, one of the challenges in safety >> Yeah, so the question I'll just repeat >> Yeah, so the question I'll just repeat it so cuz I think there's it so cuz I think there's it so cuz I think there's probably people who want to know who are probably people who want to know who are probably people who want to know who are in the room. So the question was in the room. So the question was in the room. So the question was one of the I guess one of the I guess one of the I guess the counterpoints on some Frontier Labs the counterpoints on some Frontier Labs the counterpoints on some Frontier Labs about not about not about not enabling Frontier AI but outside is that enabling Frontier AI but outside is that enabling Frontier AI but outside is that is a safety question. So I think it is a safety question. So I think it is a safety question. So I think it would be I definitely am not one of would be I definitely am not one of would be I definitely am not one of those people who who says that those people who who says that those people who who says that open source doesn't carry any risk. So open source doesn't carry any risk. So open source doesn't carry any risk. So when you make a tool more readily when you make a tool more readily when you make a tool more readily available, there's a profile of risk available, there's a profile of risk available, there's a profile of risk associated with it. Um Auto Scientists associated with it. Um Auto Scientists associated with it. Um Auto Scientists is to be fair like it's about enabling is to be fair like it's about enabling is to be fair like it's about enabling people to customize their models. You people to customize their models. You people to customize their models. You can think of that as like a slightly can think of that as like a slightly can think of that as like a slightly different question from whether those different question from whether those different question from whether those are open source. It's giving people way are open source. It's giving people way are open source. It's giving people way more control. Whether that's local or more control. Whether that's local or more control. Whether that's local or private or within that company, it's private or within that company, it's private or within that company, it's about like how do they own their own about like how do they own their own about like how do they own their own intelligence? Um intelligence? Um intelligence? Um what do I think broadly about the impact what do I think broadly about the impact what do I think broadly about the impact of open source on safety?

of open source on safety? of open source on safety? You can do so much the dynamic has often You can do so much the dynamic has often You can do so much the dynamic has often conflated like that real risk of like conflated like that real risk of like conflated like that real risk of like wider access with wider access with wider access with slight sense that slight sense that slight sense that that it it it restrains who can actually that it it it restrains who can actually that it it it restrains who can actually participate and I think that's a participate and I think that's a participate and I think that's a delicate balance. And I think you have delicate balance. And I think you have delicate balance. And I think you have to acknowledge risk by also navigating to acknowledge risk by also navigating to acknowledge risk by also navigating that and acknowledging that it limits that and acknowledging that it limits that and acknowledging that it limits who can participate. Yeah, so nuanced who can participate. Yeah, so nuanced who can participate. Yeah, so nuanced answer. So I I guess I should be more answer. So I I guess I should be more answer. So I I guess I should be more bombastic on that one but I I guess I bombastic on that one but I I guess I bombastic on that one but I I guess I have been in this discussion a few times have been in this discussion a few times have been in this discussion a few times and I find the binary views on other and I find the binary views on other and I find the binary views on other sides sides sides I feel like they miss a lot. But I feel like they miss a lot. But I feel like they miss a lot. But anyways, okay, go ahead. Oh, I think for for automating and Oh, I think for for automating and speeding up learning, one of the core speeding up learning, one of the core speeding up learning, one of the core questions is how do you balance what you questions is how do you balance what you questions is how do you balance what you store in the parametric space and the store in the parametric space and the store in the parametric space and the non-parametric space? And actually, one non-parametric space? And actually, one non-parametric space? And actually, one of the most interesting things, I of the most interesting things, I of the most interesting things, I mentioned that this only work because we mentioned that this only work because we mentioned that this only work because we co-optimized data and model, um it will co-optimized data and model, um it will co-optimized data and model, um it will only work to do like an AutoScientist only work to do like an AutoScientist only work to do like an AutoScientist for harnesses if you also co-optimize it for harnesses if you also co-optimize it for harnesses if you also co-optimize it with a model. And so, it's interesting, with a model. And so, it's interesting, with a model. And so, it's interesting, it's actually a long horizon problem, it's actually a long horizon problem, it's actually a long horizon problem, and that's super fascinating to think and that's super fascinating to think and that's super fascinating to think about where you're optimizing the about where you're optimizing the about where you're optimizing the choices for each and co-training, which choices for each and co-training, which choices for each and co-training, which is cool. Nice. Uh I think we have time is cool. Nice. Uh I think we have time is cool. Nice. Uh I think we have time for maybe two more and then we can I'll for maybe two more and then we can I'll for maybe two more and then we can I'll pass on to the next speaker. Nice, go pass on to the next speaker. Nice, go pass on to the next speaker. Nice, go ahead.

ahead. ahead. >> I have a question. >> I have a question. >> I have a question. We talked a little bit about We talked a little bit about We talked a little bit about this this this actually working on the post-training actually working on the post-training actually working on the post-training side of the models is actually cheaper side of the models is actually cheaper side of the models is actually cheaper than doing pre-training with public than doing pre-training with public than doing pre-training with public data. But like data. But like data. But like still, especially talking like really still, especially talking like really still, especially talking like really large models, like reinforcement large models, like reinforcement large models, like reinforcement learning, even fine-tuning is still learning, even fine-tuning is still learning, even fine-tuning is still pretty like GPU-intensive, so like you pretty like GPU-intensive, so like you pretty like GPU-intensive, so like you talked a little bit about like smaller talked a little bit about like smaller talked a little bit about like smaller models. models. models. But I still think that most like But I still think that most like But I still think that most like frontier smaller models still rely on frontier smaller models still rely on frontier smaller models still rely on like the bigger models to distill like the bigger models to distill like the bigger models to distill knowledge like downwards, not knowledge like downwards, not knowledge like downwards, not necessarily to be trained. So, how do necessarily to be trained. So, how do necessarily to be trained. So, how do you see this like you see this like you see this like like for the future, how to work like like for the future, how to work like like for the future, how to work like with these smaller models, how they can with these smaller models, how they can with these smaller models, how they can be be be start working without it being trained start working without it being trained start working without it being trained without depending on these larger without depending on these larger without depending on these larger models? models? models? >> Yeah, actually that's an excellent >> Yeah, actually that's an excellent >> Yeah, actually that's an excellent point. I think the question amounts to point. I think the question amounts to point. I think the question amounts to two points. One, are larger models two points. One, are larger models two points. One, are larger models necessary for distillation benefits? And necessary for distillation benefits? And necessary for distillation benefits? And then second, so frontier models are then second, so frontier models are then second, so frontier models are still pretty large. still pretty large. still pretty large. Um so, I think for the second one, Um so, I think for the second one, Um so, I think for the second one, frontier models are still pretty large, frontier models are still pretty large, frontier models are still pretty large, yes. I don't think I'm arguing that you yes. I don't think I'm arguing that you yes. I don't think I'm arguing that you um I I My argument is slightly um I I My argument is slightly um I I My argument is slightly different, and my argument is in no different, and my argument is in no different, and my argument is in no frontier AI lab is going to four frontier AI lab is going to four frontier AI lab is going to four exercise their model again for exercise their model again for exercise their model again for pre-training. So, it's almost like we pre-training. So, it's almost like we pre-training. So, it's almost like we know we're at an opposite for this know we're at an opposite for this know we're at an opposite for this architecture. If someone comes out with architecture. If someone comes out with architecture. If someone comes out with a new architecture, that's totally a new architecture, that's totally a new architecture, that's totally different, you can different, you can different, you can the architecture determines your the architecture determines your the architecture determines your ceiling, and I'm saying we are probably ceiling, and I'm saying we are probably ceiling, and I'm saying we are probably at the ceiling of size, which means that at the ceiling of size, which means that at the ceiling of size, which means that that's fine because it means okay, it's that's fine because it means okay, it's that's fine because it means okay, it's what you innovate within that. Um so, so what you innovate within that. Um so, so what you innovate within that. Um so, so size does matter. I I that's that's a size does matter. I I that's that's a size does matter. I I that's that's a very good point to bring up.

very good point to bring up. very good point to bring up. Meaning I'm not advocating everyone use Meaning I'm not advocating everyone use Meaning I'm not advocating everyone use a 0.8B, but I am saying that we now have a 0.8B, but I am saying that we now have a 0.8B, but I am saying that we now have a more equal feeling for playing field a more equal feeling for playing field a more equal feeling for playing field at the the top. Second point is at the the top. Second point is at the the top. Second point is interesting, distillation like interesting, distillation like interesting, distillation like the impact certainly. So data quality in the impact certainly. So data quality in the impact certainly. So data quality in general means you use capacity a lot general means you use capacity a lot general means you use capacity a lot more. So you what you will see in more. So you what you will see in more. So you what you will see in pre-training is instead of the size pre-training is instead of the size pre-training is instead of the size people are just moving post-training people are just moving post-training people are just moving post-training further back, which is very fascinating further back, which is very fascinating further back, which is very fascinating and a bigger lever. So I agree and a bigger lever. So I agree and a bigger lever. So I agree distillation is helpful. It's just that distillation is helpful. It's just that distillation is helpful. It's just that again we've hit the ceiling and so it's again we've hit the ceiling and so it's again we've hit the ceiling and so it's almost like no one is going to almost like no one is going to almost like no one is going to supersize their model supersize their model supersize their model or if they do it's not clear it's or if they do it's not clear it's or if they do it's not clear it's beneficial except for small size of the beneficial except for small size of the beneficial except for small size of the distribution, which is distribution, which is distribution, which is very much the long tail and that's kind very much the long tail and that's kind very much the long tail and that's kind of interesting like where that trade-off of interesting like where that trade-off of interesting like where that trade-off is where worth that much pre-training is where worth that much pre-training is where worth that much pre-training compute. So very good question. One more compute. So very good question. One more compute. So very good question. One more and then I think we are done. Yes, go and then I think we are done. Yes, go and then I think we are done. Yes, go ahead. Yeah, it's actually in beta. So you can Yeah, it's actually in beta. So you can I shared here. You can try it in beta.

I shared here. You can try it in beta. I shared here. You can try it in beta. So we actually are offering the GPUs for So we actually are offering the GPUs for So we actually are offering the GPUs for free. Okay, oh that's a nice question. free. Okay, oh that's a nice question. free. Okay, oh that's a nice question. >> [laughter] >> [laughter] >> [laughter] >> I promise I don't know this gentleman. >> I promise I don't know this gentleman. >> I promise I don't know this gentleman. But yes, I think actually we're trying But yes, I think actually we're trying But yes, I think actually we're trying to remove the compute hurdle and I think to remove the compute hurdle and I think to remove the compute hurdle and I think it's quite cool to see. So feel free to it's quite cool to see. So feel free to it's quite cool to see. So feel free to take a look at the beta. Nice, lovely. take a look at the beta. Nice, lovely. take a look at the beta. Nice, lovely. Thank you so much. Really nice. Thank Thank you so much. Really nice. Thank Thank you so much. Really nice. Thank you.