First Steps Toward Automated AI Research — Richard Socher, CEO Recursive AI

>> All right. Hello everyone. Really >> All right. Hello everyone. Really excited to be here. excited to be here. excited to be here. It's a big room. It's a big room. It's a big room. Very uh very cool conference so far. Uh Very uh very cool conference so far. Uh Very uh very cool conference so far. Uh I want to talk to you today about I want to talk to you today about I want to talk to you today about something that's been on my mind for something that's been on my mind for something that's been on my mind for many many years. This is actually the many many years. This is actually the many many years. This is actually the first time I I talk about it, sort of my first time I I talk about it, sort of my first time I I talk about it, sort of my version of going to Mars. Um and that is version of going to Mars. Um and that is version of going to Mars. Um and that is the Eureka machine. A machine that will the Eureka machine. A machine that will the Eureka machine. A machine that will eventually invent pretty much all future eventually invent pretty much all future eventually invent pretty much all future inventions for humanity. Uh and the way inventions for humanity. Uh and the way inventions for humanity. Uh and the way we're going to get there is uh by taking we're going to get there is uh by taking we're going to get there is uh by taking a step back and thinking about what else a step back and thinking about what else a step back and thinking about what else has given us a lot of really incredible has given us a lot of really incredible has given us a lot of really incredible inventions, uh namely evolution, and how inventions, uh namely evolution, and how inventions, uh namely evolution, and how that leads us to automating research uh that leads us to automating research uh that leads us to automating research uh and pushing the scientific frontier and pushing the scientific frontier and pushing the scientific frontier forward. forward. forward. And this is a joint work with a lot of And this is a joint work with a lot of And this is a joint work with a lot of uh amazing folks uh at Recursive.com uh amazing folks uh at Recursive.com uh amazing folks uh at Recursive.com uh and even some uh folks at AIX uh and even some uh folks at AIX uh and even some uh folks at AIX Ventures. And some of these slides are Ventures. And some of these slides are Ventures. And some of these slides are uh actually inspired by uh and taken uh uh actually inspired by uh and taken uh uh actually inspired by uh and taken uh partially from one of my co-founders at partially from one of my co-founders at partially from one of my co-founders at Recursive, Tim Rocktäschel. Recursive, Tim Rocktäschel. Recursive, Tim Rocktäschel. So, So, So, uh why do I talk about evolution and why uh why do I talk about evolution and why uh why do I talk about evolution and why is it so important? Uh I think basically is it so important? Uh I think basically is it so important? Uh I think basically evolution is this like open-ended evolution is this like open-ended evolution is this like open-ended process that has gotten us to a lot of process that has gotten us to a lot of process that has gotten us to a lot of different things that we really like. Uh different things that we really like. Uh different things that we really like. Uh it started in biology, it's moving to it started in biology, it's moving to it started in biology, it's moving to science, technology, and eventually AI.

science, technology, and eventually AI. science, technology, and eventually AI. And I think it can inspire us And I think it can inspire us And I think it can inspire us in a lot of different ways to build in a lot of different ways to build in a lot of different ways to build better AI systems as well. better AI systems as well. better AI systems as well. In fact, uh whenever we take out and In fact, uh whenever we take out and In fact, uh whenever we take out and there's this famous saying, "Whenever I there's this famous saying, "Whenever I there's this famous saying, "Whenever I fire a linguist, my accuracy goes up." fire a linguist, my accuracy goes up." fire a linguist, my accuracy goes up." Uh I think that's true for machine Uh I think that's true for machine Uh I think that's true for machine translation back in the day. And it may translation back in the day. And it may translation back in the day. And it may be true that we should fire all the AI be true that we should fire all the AI be true that we should fire all the AI engineers uh and that that are here uh engineers uh and that that are here uh engineers uh and that that are here uh and have them mostly manage an actual AI and have them mostly manage an actual AI and have them mostly manage an actual AI engineer that is AI and works on AI. Uh engineer that is AI and works on AI. Uh engineer that is AI and works on AI. Uh and so that may be uh one of the and so that may be uh one of the and so that may be uh one of the conclusions of this talk. Uh and I think conclusions of this talk. Uh and I think conclusions of this talk. Uh and I think most of us are going to be excited about most of us are going to be excited about most of us are going to be excited about it cuz it means that we'll all become it cuz it means that we'll all become it cuz it means that we'll all become managers of such an AI rather than managers of such an AI rather than managers of such an AI rather than having to do the nitty-gritty ourselves. having to do the nitty-gritty ourselves. having to do the nitty-gritty ourselves. All right. So, let's start with All right. So, let's start with All right. So, let's start with evolution, right? The really, really big evolution, right? The really, really big evolution, right? The really, really big picture, 3 and 1/2 billion years or so. picture, 3 and 1/2 billion years or so. picture, 3 and 1/2 billion years or so. Uh this is kind of the incredible Uh this is kind of the incredible Uh this is kind of the incredible process uh that has led from, you know, process uh that has led from, you know, process uh that has led from, you know, simple bacteria and plants and fish and simple bacteria and plants and fish and simple bacteria and plants and fish and amphibians and so on to, after many amphibians and so on to, after many amphibians and so on to, after many billions of years, us. So, I That's billions of years, us. So, I That's billions of years, us. So, I That's That's a good starting point. That gives That's a good starting point. That gives That's a good starting point. That gives us some indication that evolutionary us some indication that evolutionary us some indication that evolutionary processes can do pretty amazing things, processes can do pretty amazing things, processes can do pretty amazing things, right? But now, let's zoom in and uh go right? But now, let's zoom in and uh go right? But now, let's zoom in and uh go maybe down to a few million years.

maybe down to a few million years. maybe down to a few million years. There, we can also see how in a very There, we can also see how in a very There, we can also see how in a very first primitive ways, technological first primitive ways, technological first primitive ways, technological evolution uh has basically increased the evolution uh has basically increased the evolution uh has basically increased the world's uh sort of product uh in terms world's uh sort of product uh in terms world's uh sort of product uh in terms of monetary value. It's a little bit of monetary value. It's a little bit of monetary value. It's a little bit harder to estimate in the beginning, but harder to estimate in the beginning, but harder to estimate in the beginning, but we can see these sort of sequences of we can see these sort of sequences of we can see these sort of sequences of exponentials. And most exponentials exponentials. And most exponentials exponentials. And most exponentials eventually become S curves. They flatten eventually become S curves. They flatten eventually become S curves. They flatten out, but humanity has done pretty well out, but humanity has done pretty well out, but humanity has done pretty well by basically developing uh many of these by basically developing uh many of these by basically developing uh many of these very basic technologies, hunting, very basic technologies, hunting, very basic technologies, hunting, farming, but then also thinking about farming, but then also thinking about farming, but then also thinking about science, the scientific method um in the science, the scientific method um in the science, the scientific method um in the early days of the Enlightenment, and early days of the Enlightenment, and early days of the Enlightenment, and then of course the Industrial then of course the Industrial then of course the Industrial Revolution. So, now we can zoom even Revolution. So, now we can zoom even Revolution. So, now we can zoom even further, uh and no worries, we're further, uh and no worries, we're further, uh and no worries, we're eventually going to get to Nanochat and eventually going to get to Nanochat and eventually going to get to Nanochat and actual auto research and and what we're actual auto research and and what we're actual auto research and and what we're doing. Uh it's a very, very quick zoom. doing. Uh it's a very, very quick zoom. doing. Uh it's a very, very quick zoom. Um Um Um and now we can zoom down to last few and now we can zoom down to last few and now we can zoom down to last few thousands of years. And what we're thousands of years. And what we're thousands of years. And what we're seeing there is that with more seeing there is that with more seeing there is that with more technology, we were able to sustain more technology, we were able to sustain more technology, we were able to sustain more people, right? So, when we're working on people, right? So, when we're working on people, right? So, when we're working on pushing that frontier forward, uh we're pushing that frontier forward, uh we're pushing that frontier forward, uh we're very certain that that will lead to more very certain that that will lead to more very certain that that will lead to more human flourishing, right? And especially human flourishing, right? And especially human flourishing, right? And especially in the last few in the last few in the last few uh, hundred years, we're seeing this uh, hundred years, we're seeing this uh, hundred years, we're seeing this incredible explosion in the population incredible explosion in the population incredible explosion in the population of people because of technology.

of people because of technology. of people because of technology. And the evolution And the evolution And the evolution uh, that it brings. And in many cases, uh, that it brings. And in many cases, uh, that it brings. And in many cases, that evolutionary process is run by us, that evolutionary process is run by us, that evolutionary process is run by us, so it's sort of conscious. Uh, but there so it's sort of conscious. Uh, but there so it's sort of conscious. Uh, but there are sort of interesting uh, inspirations are sort of interesting uh, inspirations are sort of interesting uh, inspirations that we can take from that as we're that we can take from that as we're that we can take from that as we're thinking about the evolution of AI in thinking about the evolution of AI in thinking about the evolution of AI in the next cycles. the next cycles. the next cycles. Uh, in fact, and I might not agree with Uh, in fact, and I might not agree with Uh, in fact, and I might not agree with everything with Marc Andreessen, but uh, everything with Marc Andreessen, but uh, everything with Marc Andreessen, but uh, he is very smart and we agree on a lot he is very smart and we agree on a lot he is very smart and we agree on a lot of things. of things. of things. Uh, and so I think he wrote this really Uh, and so I think he wrote this really Uh, and so I think he wrote this really great uh, techno-optimist manifesto in great uh, techno-optimist manifesto in great uh, techno-optimist manifesto in which he, I think, correctly points out which he, I think, correctly points out which he, I think, correctly points out that the only perpetual source of growth that the only perpetual source of growth that the only perpetual source of growth for the entire economy, a lot of people for the entire economy, a lot of people for the entire economy, a lot of people worry about AI taking jobs and things worry about AI taking jobs and things worry about AI taking jobs and things like that, but the truth is it will like that, but the truth is it will like that, but the truth is it will very, very likely increase uh, the very, very likely increase uh, the very, very likely increase uh, the economy massively and that will benefit economy massively and that will benefit economy massively and that will benefit uh, benefit a lot of us. And so the uh, benefit a lot of us. And so the uh, benefit a lot of us. And so the perpetual source of growth is perpetual source of growth is perpetual source of growth is technology. Uh, in fact, we can go even technology. Uh, in fact, we can go even technology. Uh, in fact, we can go even further and say that there's no material further and say that there's no material further and say that there's no material problem, and again, it's not sort of problem, and again, it's not sort of problem, and again, it's not sort of psychological problems and things like psychological problems and things like psychological problems and things like that, but no material problems that, but no material problems that, but no material problems uh, that cannot be solved with even more uh, that cannot be solved with even more uh, that cannot be solved with even more technology. Right? For problem of technology. Right? For problem of technology. Right? For problem of starvation, we invented the green starvation, we invented the green starvation, we invented the green revolution, darkness, light, uh, cold, revolution, darkness, light, uh, cold, revolution, darkness, light, uh, cold, indoor heating, heat, air conditioning, indoor heating, heat, air conditioning, indoor heating, heat, air conditioning, and the list goes on.

and the list goes on. and the list goes on. So, I think we can kind of realize that So, I think we can kind of realize that So, I think we can kind of realize that this evolutionary process has been going this evolutionary process has been going this evolutionary process has been going on for a very long time and continues to on for a very long time and continues to on for a very long time and continues to make a huge amount of progress. In fact, make a huge amount of progress. In fact, make a huge amount of progress. In fact, the progress is so fast that there the progress is so fast that there the progress is so fast that there can within one lifetime be a major, can within one lifetime be a major, can within one lifetime be a major, major shift. Right? If you're born in major shift. Right? If you're born in major shift. Right? If you're born in 1900, uh, then 3 years, when you're 3 1900, uh, then 3 years, when you're 3 1900, uh, then 3 years, when you're 3 years old, the first human ever was able years old, the first human ever was able years old, the first human ever was able to, thanks to the Wright brothers, kind to, thanks to the Wright brothers, kind to, thanks to the Wright brothers, kind of have sustained motored flight. of have sustained motored flight. of have sustained motored flight. And then about 60-ish years later, in And then about 60-ish years later, in And then about 60-ish years later, in 1969, 1969, 1969, humans flew all the way to the moon, humans flew all the way to the moon, humans flew all the way to the moon, right? So, that within one lifetime, right? So, that within one lifetime, right? So, that within one lifetime, humanity went from like no one can fly humanity went from like no one can fly humanity went from like no one can fly for a very long time for a very long time for a very long time other than sort of gliding down a hill other than sort of gliding down a hill other than sort of gliding down a hill or something, no one can really fly to or something, no one can really fly to or something, no one can really fly to we all fly to the moon, right? And so, we all fly to the moon, right? And so, we all fly to the moon, right? And so, for us, I think, for us, I think, for us, I think, what that means is we're probably, and I what that means is we're probably, and I what that means is we're probably, and I sometimes say this, we're like too late sometimes say this, we're like too late sometimes say this, we're like too late to explore Earth, we're too early to to explore Earth, we're too early to to explore Earth, we're too early to explore the stars, but we're right on explore the stars, but we're right on explore the stars, but we're right on time to build an AI that could actually time to build an AI that could actually time to build an AI that could actually do what flying did for some in one do what flying did for some in one do what flying did for some in one lifetime due to intelligence. We can lifetime due to intelligence. We can lifetime due to intelligence. We can build and move from AI being worse at build and move from AI being worse at build and move from AI being worse at everything that we do to possibly being everything that we do to possibly being everything that we do to possibly being better at any specific task that we do.

better at any specific task that we do. better at any specific task that we do. Right? And that that will probably be Right? And that that will probably be Right? And that that will probably be our our 60-year time frame, and because our our 60-year time frame, and because our our 60-year time frame, and because everything moves faster, it might only everything moves faster, it might only everything moves faster, it might only be 30 years or so. be 30 years or so. be 30 years or so. So, then, uh there's an interesting So, then, uh there's an interesting So, then, uh there's an interesting connection between technology and connection between technology and connection between technology and science and theory, right? Like science and theory, right? Like science and theory, right? Like sometimes the application comes first, sometimes the application comes first, sometimes the application comes first, and then we develop the theory later, and then we develop the theory later, and then we develop the theory later, and then improve at the technology. and then improve at the technology. and then improve at the technology. Sometimes the theory comes first, and Sometimes the theory comes first, and Sometimes the theory comes first, and from that we can build new kinds of from that we can build new kinds of from that we can build new kinds of technologies. And so, it's very helpful technologies. And so, it's very helpful technologies. And so, it's very helpful to think a little bit about the to think a little bit about the to think a little bit about the philosophy of science, and no better uh philosophy of science, and no better uh philosophy of science, and no better uh to be inspired there than Karl Popper, to be inspired there than Karl Popper, to be inspired there than Karl Popper, wrote that just like in other types of wrote that just like in other types of wrote that just like in other types of evolution, when we choose a theory, we evolution, when we choose a theory, we evolution, when we choose a theory, we also choose one that is best uh in also choose one that is best uh in also choose one that is best uh in competition with other theories. Of competition with other theories. Of competition with other theories. Of course, you need if you wanted LLMs to course, you need if you wanted LLMs to course, you need if you wanted LLMs to do that, they need to find them, you do that, they need to find them, you do that, they need to find them, you need web search, for instance. Um need web search, for instance. Um need web search, for instance. Um but, uh in the theory that best holds but, uh in the theory that best holds but, uh in the theory that best holds its own, uh it's one that, just like its own, uh it's one that, just like its own, uh it's one that, just like evolution, has a certain natural evolution, has a certain natural evolution, has a certain natural selection process, right? It proves selection process, right? It proves selection process, right? It proves itself, uh and there is also a sort of itself, uh and there is also a sort of itself, uh and there is also a sort of survival of the fittest uh going on in survival of the fittest uh going on in survival of the fittest uh going on in scientific theories.

scientific theories. scientific theories. And uh in fact, uh a lot of science, And uh in fact, uh a lot of science, And uh in fact, uh a lot of science, according to Popper, is basically us according to Popper, is basically us according to Popper, is basically us proposing a new theory hypothesis or proposing a new theory hypothesis or proposing a new theory hypothesis or explanation or description and then explanation or description and then explanation or description and then subjecting it to rigorous empirical subjecting it to rigorous empirical subjecting it to rigorous empirical testing. That is the testing. That is the testing. That is the essentially evolution evolutionary essentially evolution evolutionary essentially evolution evolutionary pressure of scientific theories. And basically that was a very short run And basically that was a very short run through so the history of open-ended through so the history of open-ended through so the history of open-ended evolution evolution evolution which hopefully makes us all realize which hopefully makes us all realize which hopefully makes us all realize that more science lead to more that more science lead to more that more science lead to more technology which will lead to more technology which will lead to more technology which will lead to more growth which lead to more human growth which lead to more human growth which lead to more human flourishing. And so that then begs the flourishing. And so that then begs the flourishing. And so that then begs the question does it make sense for us to question does it make sense for us to question does it make sense for us to try to just scale up and spend a lot of try to just scale up and spend a lot of try to just scale up and spend a lot of our resources as humanity to scale up our resources as humanity to scale up our resources as humanity to scale up scientific discovery in order to lead to scientific discovery in order to lead to scientific discovery in order to lead to this flourishing. this flourishing. this flourishing. When when you double click into that you When when you double click into that you When when you double click into that you kind of realize kind of realize kind of realize which Stanisław Lem already realized a which Stanisław Lem already realized a which Stanisław Lem already realized a long time ago long time ago long time ago that the exponential growth of science that the exponential growth of science that the exponential growth of science will actually be at some point halted by will actually be at some point halted by will actually be at some point halted by the lack of people working on it, right?

the lack of people working on it, right? the lack of people working on it, right? There's so many There's so many There's so many niche subfields now in all the different niche subfields now in all the different niche subfields now in all the different areas of science that is very hard to areas of science that is very hard to areas of science that is very hard to get a million people to work on that get a million people to work on that get a million people to work on that particular thing. particular thing. particular thing. And so as a result of this incredible And so as a result of this incredible And so as a result of this incredible widening of the scope he says widening of the scope he says widening of the scope he says the number of people focusing on any the number of people focusing on any the number of people focusing on any single section of it has decreased. single section of it has decreased. single section of it has decreased. And that then leads us to really And that then leads us to really And that then leads us to really thinking about how could we automate thinking about how could we automate thinking about how could we automate this and automate scientific discovery this and automate scientific discovery this and automate scientific discovery and that then leads us to what I call and that then leads us to what I call and that then leads us to what I call the Eureka machine. the Eureka machine. the Eureka machine. This is basically This is basically This is basically our attempt at trying to our attempt at trying to our attempt at trying to build a machine that automates the build a machine that automates the build a machine that automates the process of scientific discoveries. process of scientific discoveries. process of scientific discoveries. And in fact I like in a couple months And in fact I like in a couple months And in fact I like in a couple months I'll have a book coming out on on this I'll have a book coming out on on this I'll have a book coming out on on this exact idea exact idea exact idea and so I'll just give you a super and so I'll just give you a super and so I'll just give you a super high-level highlight of how such a high-level highlight of how such a high-level highlight of how such a Eureka machine could be built for Eureka machine could be built for Eureka machine could be built for basically everything from physics, basically everything from physics, basically everything from physics, chemistry, biology, neuroscience, chemistry, biology, neuroscience, chemistry, biology, neuroscience, medicine, medicine, medicine, economics, astrophysics and so on. And economics, astrophysics and so on. And economics, astrophysics and so on. And there are essentially four pillars that there are essentially four pillars that there are essentially four pillars that are all extremely important to this are all extremely important to this are all extremely important to this machine. One is, of course, you have to machine. One is, of course, you have to machine. One is, of course, you have to understand what knowledge is already out understand what knowledge is already out understand what knowledge is already out there, what things humanity has already there, what things humanity has already there, what things humanity has already invented.

invented. invented. You have to get all the scientific You have to get all the scientific You have to get all the scientific measurement data into, as in the second measurement data into, as in the second measurement data into, as in the second pillar, this machine. Then, for things pillar, this machine. Then, for things pillar, this machine. Then, for things that you cannot yet measure, we don't that you cannot yet measure, we don't that you cannot yet measure, we don't yet know, you should try to then build yet know, you should try to then build yet know, you should try to then build simulations. Anything you can simulate, simulations. Anything you can simulate, simulations. Anything you can simulate, you can verify, and you can then solve you can verify, and you can then solve you can verify, and you can then solve with AI. with AI. with AI. And And And if all else fails, or at the very end of if all else fails, or at the very end of if all else fails, or at the very end of these processes, you still need to have these processes, you still need to have these processes, you still need to have some kind of physical industrial like some kind of physical industrial like some kind of physical industrial like lab that actually can run real lab that actually can run real lab that actually can run real experiments in the real world. experiments in the real world. experiments in the real world. And on top of all of this, you'll have a And on top of all of this, you'll have a And on top of all of this, you'll have a basically basically basically an agent swarm that will an agent swarm that will an agent swarm that will deal with all of these different sources deal with all of these different sources deal with all of these different sources of knowledge and data and of knowledge and data and of knowledge and data and experimentations and and rewards. experimentations and and rewards. experimentations and and rewards. And in terms of, you know, the And in terms of, you know, the And in terms of, you know, the foundational model of knowledge, of foundational model of knowledge, of foundational model of knowledge, of course, we also, you know, course, we also, you know, course, we also, you know, basically is is a good example of how basically is is a good example of how basically is is a good example of how every single technology we've built so every single technology we've built so every single technology we've built so far, especially in AI, but also before far, especially in AI, but also before far, especially in AI, but also before that, the internet, browsers, GPUs, and that, the internet, browsers, GPUs, and that, the internet, browsers, GPUs, and so on, we can rethink, and there are a so on, we can rethink, and there are a so on, we can rethink, and there are a lot of startups possible in rethinking lot of startups possible in rethinking lot of startups possible in rethinking every single one of the layers of every single one of the layers of every single one of the layers of technology as infrastructure for technology as infrastructure for technology as infrastructure for superintelligence.

superintelligence. superintelligence. And at you.com, for instance, we work on And at you.com, for instance, we work on And at you.com, for instance, we work on web search for LLMs, right, and agents web search for LLMs, right, and agents web search for LLMs, right, and agents and so on. and so on. and so on. And that actually is quite different, And that actually is quite different, And that actually is quite different, right? Agents can read thousands of very right? Agents can read thousands of very right? Agents can read thousands of very long snippets, long snippets, long snippets, rather than just 10 blue links with like rather than just 10 blue links with like rather than just 10 blue links with like a very short snippet. And so, you can a very short snippet. And so, you can a very short snippet. And so, you can rethink each of these different layers rethink each of these different layers rethink each of these different layers of technology that we've built for of technology that we've built for of technology that we've built for people, people, people, and rebuild them for AI in order to use and rebuild them for AI in order to use and rebuild them for AI in order to use them as tools to then build a them as tools to then build a them as tools to then build a superintelligence. superintelligence. superintelligence. Now, Now, Now, that is essentially uh the sort of why. that is essentially uh the sort of why. that is essentially uh the sort of why. Like like we want to build Like like we want to build Like like we want to build superintelligence in order to automate superintelligence in order to automate superintelligence in order to automate science. science. science. Uh and to me, that will be the next big Uh and to me, that will be the next big Uh and to me, that will be the next big step function change uh in in humanity step function change uh in in humanity step function change uh in in humanity uh and technology as we know it. uh and technology as we know it. uh and technology as we know it. Now, Now, Now, how do we actually build it? Uh I think how do we actually build it? Uh I think how do we actually build it? Uh I think the best way to build it is to have it the best way to build it is to have it the best way to build it is to have it built itself, right? We moved as a field built itself, right? We moved as a field built itself, right? We moved as a field and especially natural language and especially natural language and especially natural language processing for instance, which I've processing for instance, which I've processing for instance, which I've worked on for many years. We moved from worked on for many years. We moved from worked on for many years. We moved from not having linguists. This feels like not having linguists. This feels like not having linguists. This feels like ancient, you know, BC uh history, uh but ancient, you know, BC uh history, uh but ancient, you know, BC uh history, uh but before ChatGPT, um before ChatGPT, um before ChatGPT, um we we moved from having linguists tell we we moved from having linguists tell we we moved from having linguists tell us a bunch of things about language and us a bunch of things about language and us a bunch of things about language and then training statistical models on top then training statistical models on top then training statistical models on top of that. And when we allowed neural of that. And when we allowed neural of that. And when we allowed neural networks to actually automate learning networks to actually automate learning networks to actually automate learning those features with word vectors and uh those features with word vectors and uh those features with word vectors and uh other neural network architectures and other neural network architectures and other neural network architectures and back-to-back uh and end-to-end learning back-to-back uh and end-to-end learning back-to-back uh and end-to-end learning and backpropagation, we basically uh and backpropagation, we basically uh and backpropagation, we basically uh were able to get much bigger

were able to get much bigger were able to get much bigger improvements. Uh then we did a bunch of improvements. Uh then we did a bunch of improvements. Uh then we did a bunch of architecture engineering. Now a bunch of architecture engineering. Now a bunch of architecture engineering. Now a bunch of people at least are working on a unified people at least are working on a unified people at least are working on a unified architecture, architecture, architecture, uh but even that unified architecture uh but even that unified architecture uh but even that unified architecture has a lot of manual processes. And so, has a lot of manual processes. And so, has a lot of manual processes. And so, it's clear over and over again in AI it's clear over and over again in AI it's clear over and over again in AI that when we take out a manual process that when we take out a manual process that when we take out a manual process and we replace it with a learned system, and we replace it with a learned system, and we replace it with a learned system, improvements will follow. Uh and so, improvements will follow. Uh and so, improvements will follow. Uh and so, that's why I think uh we should try to that's why I think uh we should try to that's why I think uh we should try to build a supermachine by having uh an RSI build a supermachine by having uh an RSI build a supermachine by having uh an RSI that builds itself. And the beauty is that builds itself. And the beauty is that builds itself. And the beauty is that that that only now only now only now um um um AI can actually do this because AI is AI can actually do this because AI is AI can actually do this because AI is code and AI can code now. This this code and AI can code now. This this code and AI can code now. This this ability to really code in longer and ability to really code in longer and ability to really code in longer and longer time horizons has really only longer time horizons has really only longer time horizons has really only happened in the last like 6 to 8 months. happened in the last like 6 to 8 months. happened in the last like 6 to 8 months. And that now enables such an RSI to work And that now enables such an RSI to work And that now enables such an RSI to work on itself, to develop almost a certain on itself, to develop almost a certain on itself, to develop almost a certain sense of self-awareness of its own sense of self-awareness of its own sense of self-awareness of its own shortcomings and then fix those shortcomings and then fix those shortcomings and then fix those shortcomings. shortcomings. shortcomings. Uh and then once we have that machine Uh and then once we have that machine Uh and then once we have that machine that has gotten really, really good at that has gotten really, really good at that has gotten really, really good at doing research in AI itself, we can then doing research in AI itself, we can then doing research in AI itself, we can then use it to do AI research for a lot of use it to do AI research for a lot of use it to do AI research for a lot of other things uh in in other scientific other things uh in in other scientific other things uh in in other scientific fields.

fields. fields. And so at a high level it's quite easy, And so at a high level it's quite easy, And so at a high level it's quite easy, right? We have three steps: ideation, right? We have three steps: ideation, right? We have three steps: ideation, implementation, and validation of ideas. implementation, and validation of ideas. implementation, and validation of ideas. That's true for basically almost every That's true for basically almost every That's true for basically almost every scientific field. scientific field. scientific field. And so uh to end maybe on some very And so uh to end maybe on some very And so uh to end maybe on some very specific examples, uh we have built this specific examples, uh we have built this specific examples, uh we have built this first kind of version of such a Eureka first kind of version of such a Eureka first kind of version of such a Eureka machine uh and we wanted to just show machine uh and we wanted to just show machine uh and we wanted to just show that it works on some small uh samples that it works on some small uh samples that it works on some small uh samples that a lot of people know uh and are that a lot of people know uh and are that a lot of people know uh and are aware of. And so we basically started uh aware of. And so we basically started uh aware of. And so we basically started uh with three things that show you and give with three things that show you and give with three things that show you and give you a very first glimpse of an and sort you a very first glimpse of an and sort you a very first glimpse of an and sort of simple proof points uh of what such a of simple proof points uh of what such a of simple proof points uh of what such a machinery can do. And that was basically machinery can do. And that was basically machinery can do. And that was basically better training, faster training, and better training, faster training, and better training, faster training, and and better kernels uh for for Nvidia and better kernels uh for for Nvidia and better kernels uh for for Nvidia GPUs. Um GPUs. Um GPUs. Um the first one, NanoChat, um I'm sure the first one, NanoChat, um I'm sure the first one, NanoChat, um I'm sure many of you have heard of it. many of you have heard of it. many of you have heard of it. A lot of people think that's already A lot of people think that's already A lot of people think that's already recursive self-improvement and it's kind recursive self-improvement and it's kind recursive self-improvement and it's kind of a weak form in the sense that usually of a weak form in the sense that usually of a weak form in the sense that usually when you do auto research, it's it's not when you do auto research, it's it's not when you do auto research, it's it's not recursive self-improvement, right? True recursive self-improvement, right? True recursive self-improvement, right? True recursive self-improvement is when you recursive self-improvement is when you recursive self-improvement is when you have an AI that has a sense of have an AI that has a sense of have an AI that has a sense of self-awareness of its own shortcomings, self-awareness of its own shortcomings, self-awareness of its own shortcomings, full access over everything uh in its full access over everything uh in its full access over everything uh in its arsenal from pre-training to RL training arsenal from pre-training to RL training arsenal from pre-training to RL training and harnesses and everything, and then and harnesses and everything, and then and harnesses and everything, and then actually updates that entire system in actually updates that entire system in actually updates that entire system in the next version of itself.

the next version of itself. the next version of itself. Now you can also take such a system and Now you can also take such a system and Now you can also take such a system and just ask it to improve some other just ask it to improve some other just ask it to improve some other process, some other AI, like a small process, some other AI, like a small process, some other AI, like a small NanoChat run where you can train NanoChat run where you can train NanoChat run where you can train something in 5 minutes. And that is something in 5 minutes. And that is something in 5 minutes. And that is really exciting and it's an important really exciting and it's an important really exciting and it's an important milestone, but it's not actual RSI. So milestone, but it's not actual RSI. So milestone, but it's not actual RSI. So here here here we basically showed three examples of we basically showed three examples of we basically showed three examples of such an auto research such an auto research such an auto research system and what it can do and after a system and what it can do and after a system and what it can do and after a very very short time it essentially was very very short time it essentially was very very short time it essentially was able to outperform many different teams able to outperform many different teams able to outperform many different teams and teams that also use other AI and teams that also use other AI and teams that also use other AI research. So, let's double click into research. So, let's double click into research. So, let's double click into some of these. Nano chat is really some of these. Nano chat is really some of these. Nano chat is really exciting example. Basically, you train a exciting example. Basically, you train a exciting example. Basically, you train a very small very small very small chat model chat model chat model in less than 5 minutes and you basically in less than 5 minutes and you basically in less than 5 minutes and you basically want to have it get to the best possible want to have it get to the best possible want to have it get to the best possible bits per byte bits per byte bits per byte number. number. number. And so, the whole community had worked And so, the whole community had worked And so, the whole community had worked on this for quite some time and got to on this for quite some time and got to on this for quite some time and got to 0.93 0.93 0.93 and after training this for a little and after training this for a little and after training this for a little more than a day or two, more than a day or two, more than a day or two, we basically got it down to 0.91.

we basically got it down to 0.91. we basically got it down to 0.91. Which is pretty exciting. Now, it Which is pretty exciting. Now, it Which is pretty exciting. Now, it wouldn't be that exciting if all it did wouldn't be that exciting if all it did wouldn't be that exciting if all it did was just find a couple of hyper was just find a couple of hyper was just find a couple of hyper parameters and tune them carefully, but parameters and tune them carefully, but parameters and tune them carefully, but it actually did find truly interesting it actually did find truly interesting it actually did find truly interesting novel ideas like hash bi-grams and novel ideas like hash bi-grams and novel ideas like hash bi-grams and tri-gram embeddings and tables for those tri-gram embeddings and tables for those tri-gram embeddings and tables for those and mixing that into various value paths and mixing that into various value paths and mixing that into various value paths of the intention through variety of of the intention through variety of of the intention through variety of learned gates. So, it actually started learned gates. So, it actually started learned gates. So, it actually started to doing more and more interesting to doing more and more interesting to doing more and more interesting things rather than just kind of tuning things rather than just kind of tuning things rather than just kind of tuning hyper parameters. hyper parameters. hyper parameters. Another one nano GPT speed run. Another one nano GPT speed run. Another one nano GPT speed run. Obviously, speed is very important. So, Obviously, speed is very important. So, Obviously, speed is very important. So, here we're able to here we're able to here we're able to work on this again, apply the system and work on this again, apply the system and work on this again, apply the system and after very short amount of time it got after very short amount of time it got after very short amount of time it got better than people working often better than people working often better than people working often together with the AI for over a year together with the AI for over a year together with the AI for over a year on on this very on this benchmark and on on this very on this benchmark and on on this very on this benchmark and made the whole thing another two seconds made the whole thing another two seconds made the whole thing another two seconds over two seconds faster over two seconds faster over two seconds faster at 70 seconds. And again, discovering at 70 seconds. And again, discovering at 70 seconds. And again, discovering very interesting ideas in the process.

very interesting ideas in the process. very interesting ideas in the process. And then the third one is CUDA kernels. And then the third one is CUDA kernels. And then the third one is CUDA kernels. Of course, we all care about not burning Of course, we all care about not burning Of course, we all care about not burning through our GPU budgets too quickly through our GPU budgets too quickly through our GPU budgets too quickly and trying to be very efficient. I think and trying to be very efficient. I think and trying to be very efficient. I think in general it's actually kind of in general it's actually kind of in general it's actually kind of shocking how inefficient a lot of shocking how inefficient a lot of shocking how inefficient a lot of mixture of expert models still are mixture of expert models still are mixture of expert models still are running very large clusters that cost running very large clusters that cost running very large clusters that cost billions of dollars and then only have billions of dollars and then only have billions of dollars and then only have like 30% or so utilization. So a lot of like 30% or so utilization. So a lot of like 30% or so utilization. So a lot of work that's ongoing in the world to work that's ongoing in the world to work that's ongoing in the world to improve that and different fields or improve that and different fields or improve that and different fields or different groups of people are various different groups of people are various different groups of people are various different different different stages of that. stages of that. stages of that. But long story short, lots of different But long story short, lots of different But long story short, lots of different CUDA kernels are used during training CUDA kernels are used during training CUDA kernels are used during training and testing and here and testing and here and testing and here we basically again took that system and we basically again took that system and we basically again took that system and after after after a couple days it discovered better a couple days it discovered better a couple days it discovered better kernels than the leader boards best on kernels than the leader boards best on kernels than the leader boards best on the Nvidia benchmark website by again the Nvidia benchmark website by again the Nvidia benchmark website by again quite quite a sizable margin quite quite a sizable margin quite quite a sizable margin across all the different categories of across all the different categories of across all the different categories of those kernels. And while we are pretty those kernels. And while we are pretty those kernels. And while we are pretty good at AI and we actually in the team good at AI and we actually in the team good at AI and we actually in the team didn't have any particular CUDA kernel didn't have any particular CUDA kernel didn't have any particular CUDA kernel experts who just spent their entire experts who just spent their entire experts who just spent their entire careers writing good kernels.

careers writing good kernels. careers writing good kernels. But still, you know, we do just enough But still, you know, we do just enough But still, you know, we do just enough to make sure and work together with to make sure and work together with to make sure and work together with Nvidia to make sure that there no reward Nvidia to make sure that there no reward Nvidia to make sure that there no reward hacks here and and other issues. But hacks here and and other issues. But hacks here and and other issues. But actually actually actually found that eventually these all checked found that eventually these all checked found that eventually these all checked out and were indeed pretty much all the out and were indeed pretty much all the out and were indeed pretty much all the different kernels different kernels different kernels found the best solutions there. found the best solutions there. found the best solutions there. And so with that I hope I could convince And so with that I hope I could convince And so with that I hope I could convince you that indeed RSI could be that next you that indeed RSI could be that next you that indeed RSI could be that next big S-curve an exponential that gets big S-curve an exponential that gets big S-curve an exponential that gets layered on top of previous exponentials layered on top of previous exponentials layered on top of previous exponentials and that should help us and that should help us and that should help us with not just AI but eventually science with not just AI but eventually science with not just AI but eventually science and then all of technology and then and then all of technology and then and then all of technology and then allowing many more people to flourish allowing many more people to flourish allowing many more people to flourish on our planet. And so on our planet. And so on our planet. And so maybe I'll end on this note here, which maybe I'll end on this note here, which maybe I'll end on this note here, which is is is a lot of people wonder how much longer I a lot of people wonder how much longer I a lot of people wonder how much longer I can go, right? Every exponential can go, right? Every exponential can go, right? Every exponential eventually flattens out and eventually flattens out and eventually flattens out and it's actually quite hard to know like it's actually quite hard to know like it's actually quite hard to know like when we even talk about exponential when we even talk about exponential when we even talk about exponential growth in the eye, what does that even growth in the eye, what does that even growth in the eye, what does that even mean? There are many different, I call mean? There are many different, I call mean? There are many different, I call them spaces of intelligence and we won't them spaces of intelligence and we won't them spaces of intelligence and we won't have time to go into all of all of have time to go into all of all of have time to go into all of all of these, but as soon as you actually try these, but as soon as you actually try these, but as soon as you actually try to define multiple different dimensions to define multiple different dimensions to define multiple different dimensions of each of these 10 spaces that make up of each of these 10 spaces that make up of each of these 10 spaces that make up this complex sort of volumetric uh this complex sort of volumetric uh this complex sort of volumetric uh thing that is intelligence, you'll thing that is intelligence, you'll thing that is intelligence, you'll realize that there's still so much more realize that there's still so much more realize that there's still so much more to go. Like on the upper bounds of to go. Like on the upper bounds of to go. Like on the upper bounds of intelligence, we're still astronomically intelligence, we're still astronomically intelligence, we're still astronomically far away from reaching those far away from reaching those far away from reaching those and across pretty much every single one and across pretty much every single one and across pretty much every single one of these dimensions and the spaces that

of these dimensions and the spaces that of these dimensions and the spaces that they make up. So, if any of that is they make up. So, if any of that is they make up. So, if any of that is interesting and you want to help us interesting and you want to help us interesting and you want to help us build that, um build that, um build that, um we'd love to hear from you. Thank you.