Richard Socher at AI Engineer: The Eureka Machine and the Path to Recursive Self-Improvement

Summary

In this talk at AI Engineer, Richard Socher presents the Eureka Machine, a system meant to automate scientific discovery, and explains why he thinks the way to build it is to have AI research and improve itself. He starts from evolution and the history of technology, then argues that science is limited by the number of people working in each narrow subfield.

He lays out four pillars for the machine, covering existing knowledge, scientific measurements, simulation, and real-world labs, with an agent swarm on top. He then shares early results from Recursive's first version: improvements on the nanochat and nanoGPT benchmarks and on CUDA kernels for NVIDIA GPUs. He closes by saying we are still astronomically far from the upper bounds of intelligence.

Key highlights

• Evolution as inspiration. Socher frames evolution as an open-ended process that moved from biology to science and technology, and argues it can inspire better AI systems.
• Technology and growth. He cites Marc Andreessen's Techno-Optimist Manifesto and the span from the Wright brothers to the moon landing to argue that technology solves material problems, and that AI could do for intelligence what flight did in one lifetime.
• A people bottleneck in science. Science is split into so many niche subfields that it is hard to put enough people on any one of them, which is the case for automating discovery.
• Four pillars plus a swarm. The Eureka Machine combines existing knowledge, scientific measurement data, simulation, and physical labs, with an agent swarm working across them.
• Rebuild infrastructure for agents. He says agents can read thousands of long snippets rather than ten blue links, so search, browsers, and other layers can be rethought for AI.
• Replace manual steps with learned ones. His pattern from NLP is that removing a manual process and replacing it with a learned system tends to bring improvements, and AI can now attempt this because it can code over long time horizons.
• Auto research vs. recursive self-improvement. He says tuning another model is a weak form, while true recursive self-improvement means an AI with access to its whole stack that updates itself in the next version.
• Early results. On nanochat, the system brought bits-per-byte from 0.93 to 0.91. On the nanoGPT speedrun, it made the benchmark more than two seconds faster, to 70 seconds. It also found CUDA kernels that beat the leaderboard's best on an NVIDIA benchmark, and the team worked with NVIDIA to check for reward hacks.

Watch the talk

https://www.youtube.com/watch?v=pWXUkLP9uWM

Transcript

Lightly edited for readability: timestamps follow the source transcript, and obvious transcription errors and name spellings are corrected. Section headings come from the video's chapters.

[0:12] All right. Hello everyone. Really excited to be here. It's a big room. Very, very cool conference so far. I want to talk to you today about something that's been on my mind for many, many years. This is actually the first time I talk about it. Sort of my version of going to Mars. And that is the Eureka Machine, a machine that will eventually invent pretty much all future inventions for humanity. And the way we're going to get there is by taking a step back and thinking about what else has given us a lot of really incredible inventions, namely evolution, and how that leads us to automating research and pushing the scientific frontier forward. And this is joint work with a lot of amazing folks at recursive.com, and even some folks at AIX Ventures. And some of these slides are actually inspired by, and taken partially from, one of my co-founders at Recursive, Tim Rocktäschel.

[1:12] So why do I talk about evolution and why is it so important? I think basically evolution is this open-ended process that has gotten us to a lot of different things that we really like. It started in biology. It's moving to science, technology, and eventually AI, and I think it can inspire us in a lot of different ways to build better AI systems as well. In fact, whenever we take out...

Why this matters now

[1:38] ...and there's this famous saying, whenever I fire a linguist, my accuracy goes up. I think that's true for machine translation back in the day, and it may be true that we should fire all the AI engineers that are here and have them mostly manage an actual AI engineer that is AI and works on AI. And so that may be one of the conclusions of this talk. And I think most of us are going to be excited about it, because it means that we'll all become managers of such an AI rather than having to do the nitty-gritty ourselves.

Compressing the timeline of progress

[2:15] All right, so let's start with evolution, right? The really, really big picture, three and a half billion years or so. This is kind of the incredible process that has led from simple bacteria and plants and fish and amphibians and so on, after many billions of years, to us, right? That's a good starting point that gives us some indication that evolutionary processes can do pretty amazing things, right? But now let's zoom in and go maybe down to a few million years. There we can also see how, in the very first primitive ways, technological evolution has basically increased the world's product, in terms of monetary value. It's a little bit harder to estimate in the beginning, but we can see these sequences of exponentials, and most exponentials eventually become S-curves. They flatten out. But humanity has done pretty well by developing many of these very basic technologies, hunting, farming, but then also thinking about science, the scientific method in the early days of the Enlightenment, and of course the Industrial Revolution. So now we can zoom in even further, and no worries, we're eventually going to get to nanochat and actual auto research and what we're doing. It's a very, very quick zoom. And now we can zoom down to the last few thousand years, and what we're seeing there is that with more technology we were able to sustain more people, right? So when we're working on pushing that frontier forward, we're very certain that that will lead to more human flourishing, right? And especially in the last few hundred years we're seeing this incredible explosion in the population of people because of technology and the evolution that it brings. And in many cases that evolutionary process is run by us, so it's sort of conscious, but there are sort of interesting inspirations that we can take from that as we're thinking about the evolution of AI in the next cycles.

Technoptimism and material limits

[4:19] In fact, and I might not agree with everything with Marc Andreessen, but he is very smart and we agree on a lot of things. And so I think he wrote this really great Techno-Optimist Manifesto in which he, I think correctly, points out that technology is the only perpetual source of growth for the entire economy. A lot of people worry about AI taking jobs and things like that, but the truth is it will very, very likely increase the economy massively and that will benefit a lot of us. And so the perpetual source of growth is technology. In fact, we can go even further and say that there's no material problem, and again, it's not psychological problems and things like that, but no material problems that cannot be solved with even more technology, right? We had a problem of starvation. We invented the Green Revolution. Darkness, light. Cold, indoor heating. Heat, air conditioning. And the list goes on. So I think we can kind of realize that this evolutionary process has been going on for a very long time and continues to make a huge amount of progress. In fact, the progress is so fast that within one lifetime there can be a major, major shift, right? If you're born in 1900, then when you're three years old, the first human ever, thanks to the Wright brothers, was able to have sustained motored flight. And then about 60-ish years later, in 1969, humans flew all the way to the moon, right? So within one lifetime, humanity went from, no one can fly, other than sort of gliding down a hill or something, to we all fly to the moon, right? And so for us, I think what that means is, and I sometimes say this, we're too late to explore Earth. We're too early to explore the stars, but we're right on time to build an AI that could actually do what flying did in one lifetime, due to intelligence. We can build and move from AI being worse at everything that we do to possibly being better at any specific task that we do, right? And that will probably be our 60-year time frame. And because everything moves faster, it might only be 30 years or so.

Popper and open-ended evolution

[6:42] So then there's an interesting connection between technology and science and theory, right? Sometimes the application comes first and then we develop the theory later and then improve the technology. Sometimes the theory comes first and from that we can build new kinds of technologies. And so it's very helpful to think a little bit about the philosophy of science, and no better to be inspired there than Popper, who wrote that, just like in other types of evolution, when we choose a theory we also choose one that is best in competition with other theories. Of course, if you wanted LLMs to do that, they need to find them. You need web search, for instance. But the theory that best holds its own is one that, just like evolution has a certain natural selection process, right, it proves itself. And there is also a sort of survival of the fittest going on in scientific theories. And in fact a lot of science, according to Popper, is basically us proposing a new theory, hypothesis, or explanation or description and then subjecting it to rigorous empirical testing. That is essentially the evolutionary pressure of scientific theories.

[7:58] And basically that was a very short run through sort of the history of open-ended evolution, which hopefully makes us all realize that more science will lead to more technology, which will lead to more growth, which will lead to more human flourishing. And so that then begs the question: does it make sense for us to try to just scale up and spend a lot of our resources as humanity to scale up scientific discovery in order to lead to this flourishing? When you double-click into that, you kind of realize, which [unclear] already realized a long time ago, that the exponential growth of science will actually at some point be halted by the lack of people working on it, right? There are so many niche subfields now in all the different areas of science that it is very hard to get a million people to work on that particular thing. And so as a result of this incredible widening of the scope, he says, the number of people focusing on any single section of it has decreased. And that then leads us to really thinking about how could we automate this, and automate scientific discovery.

The Eureka machine

[9:07] And that then leads to what I call the Eureka Machine. This is basically our attempt at trying to build a machine that automates the process of scientific discoveries. And in a couple of months I'll have a book coming out on this exact idea. And so I'll just give you a super high-level highlight of how such a Eureka Machine could be built for basically everything from physics, chemistry, biology, neuroscience, medicine, economics, astrophysics and so on. And there are essentially four pillars that are all extremely important to this machine. One is, of course, you have to understand what knowledge is already out there, what things humanity has already invented. You have to get all the scientific measurement data into, as the second pillar, this machine. Then, for things that you cannot yet measure, that we don't yet know, you should try to build simulations. Anything you can simulate, you can verify, and you can then solve with AI. And if all else fails, or at the very end of these processes, you still need to have some kind of physical, industrial-like lab that actually can run real experiments in the real world. And on top of all of this you'll have basically an agent swarm that will deal with all of these different sources of knowledge and data and experimentation and rewards.

Rethinking search, browsers, and GPUs

[10:37] And in terms of the foundational model of knowledge, of course, it basically is a good example of how every single technology we've built so far, especially in AI but also before that, the internet, browsers, GPUs and so on, we can rethink, and there are a lot of startups possible in rethinking every single one of the layers of technology as infrastructure for superintelligence, right? At You.com, for instance, we work on web search for LLMs and agents and so on, and that actually is quite different, right? Agents can read thousands of very long snippets rather than just 10 blue links with a very short snippet. And so you can rethink each of these different layers of technology that we've built for people and rebuild them for AI in order to use them as tools to then build superintelligence. Now that is essentially the why: we want to build superintelligence in order to automate science. And to me that will be the next big step-function change in humanity and technology as we know it.

[11:53] Now, how do we actually build it? I think the best way to build it is to have it build itself. We moved as a field, and especially in natural language processing, for instance, which I've worked on for many years, in the ancient, BC days before ChatGPT, from having linguists tell us a bunch of things about language and then training statistical models on top of that, to allowing neural networks to actually automate learning those features, with word vectors and other neural network architectures and end-to-end learning and backpropagation, and we basically were able to get much bigger improvements. Then we did a bunch of architecture engineering. Now a bunch of people at least are working on a unified architecture. But even that unified architecture has a lot of manual processes. And so it's clear over and over again in AI that when we take out a manual process and we replace it with a learned system, improvements will follow. And so that's why I think we should try to build a Eureka Machine by having an RSI that builds itself.

Recursive self-improvement

[13:02] And the beauty is that only now can AI actually do this, because AI is code and AI can code. This ability to really code on longer and longer time horizons has really only happened in the last six to eight months, and that now enables such an RSI to work on itself, to develop almost a certain sense of self-awareness of its own shortcomings and then fix those shortcomings. And then once we have that machine that has gotten really, really good at doing research in AI itself, we can then use it to do AI research for a lot of other things in other scientific fields. And so at a high level it's quite easy, right? We have three steps: ideation, implementation, and validation of ideas. That's true for basically almost every scientific field.

[13:56] And so, to end, maybe on some very specific examples: we have built this first version of such a Eureka Machine, and we wanted to just show that it works on some small samples that a lot of people know and are aware of. And so we started with three things that give you a very first glimpse of, and sort of simple proof points of, what such a machine can do. And that was basically better training, faster training, and better kernels for NVIDIA GPUs. The first one, nanochat. I'm sure many of you have heard of it. A lot of people think that's already recursive self-improvement, and it is kind of a weak form, in the sense that usually when you do auto research, it's not recursive self-improvement, right?

Proof point: improving a model

[14:46] True recursive self-improvement is when you have an AI that has a sense of self-awareness of its own shortcomings, full access over everything in its arsenal, from pre-training to RL training and harnesses and everything, and then actually updates that entire system in the next version of itself. Now, you can also take such a system and just ask it to improve some other process, some other AI, like a small nanochat run where you can train something in five minutes, and that is really exciting. It's an important milestone, but it's not actual RSI. So here I basically showed three examples of such an auto research system and what it can do, and after a very, very short time it was essentially able to outperform many different teams, and teams that also use other AI research. So let's double-click into some of these. Nanochat is a really exciting example. Basically you train a very small chat model in less than five minutes and you want to have it get to the best possible bits-per-byte number. And so the whole community had worked on this for quite some time and got to 0.93. And after training this for a little more than a day or two, we basically got it down to 0.91, which is pretty exciting. Now, it wouldn't be that exciting if all it did was just find a couple of hyperparameters and tune them carefully, but it actually did find truly interesting novel ideas, like hash bigrams and trigram embeddings and tables for those, and mixing that into various value paths of the attention through a variety of learned gates. So it actually started doing more and more interesting things rather than just kind of tuning hyperparameters.

Proof point: architecture search

[16:44] Another one, a nanoGPT speedrun. Obviously speed is very important. And so here we were able to apply the system again and after a very short amount of time it got better than people working, often together with the AI, for over a year on this benchmark, and made the whole thing over two seconds faster, at 70 seconds, again discovering very interesting ideas in the process.

Proof point: CUDA kernels

[17:16] And then the third one is CUDA kernels. Of course, we all care about not burning through our GPU budgets too quickly and trying to be very efficient. I think in general it's actually kind of shocking how inefficient a lot of mixture-of-experts models still are, run in very large clusters that cost billions of dollars and then only have like 30% or so utilization. There's a lot of work that's ongoing in the world to improve that, in different fields or different groups of people or various different stages of that. But long story short, lots of different CUDA kernels are used during training and testing, and here we basically again took that system and after a couple of days it discovered better kernels than the leaderboard's best on the NVIDIA benchmark website, by again quite a sizable margin across all the different categories of those kernels. And while we are pretty good at AI, we actually in the team didn't have any particular CUDA kernel experts who just spent their entire careers writing good kernels. But still, we do just enough to make sure, and worked together with NVIDIA to make sure, that there are no reward hacks here and other issues, but actually found that eventually these all checked out and were indeed pretty much all the different kernels, found the best solutions there. And so with that I hope I could convince you that indeed RSI could be that next big S-curve, an exponential that gets layered on top of previous exponentials, and that should help us with not just AI but eventually science and then all of technology, and then allowing many more people to flourish on our planet.

How far we still are, and an invitation

[19:08] And so maybe I'll end on this note here, which is a lot of people wonder how much longer AI can go, right? Every exponential eventually flattens out, and it's actually quite hard to know, when we even talk about exponential growth in AI, what does that even mean? There are many different, I call them, spaces of intelligence, and we won't have time to go into all of these, but as soon as you actually try to define multiple different dimensions of each of these 10 spaces that make up this complex, sort of [unclear] thing that is intelligence, you'll realize that there's still so much more to go on the upper bounds of intelligence. We're still astronomically far away from reaching those across pretty much every single one of these dimensions and the spaces that they make up. So if any of that is interesting and you want to help us build that, would love to hear from you. Thank you.