AI Engineer World's Fair 2026
First Steps Toward Automated AI Research
Read the talk
First Steps Toward Automated AI Research
Richard Socher’s Eureka machine begins with a practical loop: propose an idea, implement it, and test it—then use that loop to improve the machinery of AI research itself.
From a talk by Richard Socher
Before you start: Basic familiarity with language-model training, attention and GPU computation will help with the three research examples.
A machine that invents
Could we build a machine that eventually invents humanity’s future inventions? Richard Socher calls this ambition the Eureka machine, his equivalent of going to Mars. The starting point is evolution: a process that has already produced extraordinary inventions without a designer specifying every intermediate step. His proposed path toward automated discovery draws on joint work at Recursive, You.com and AIX Ventures, with some slides inspired by Recursive co-founder Tim Rocktäschel.
The connection runs from biological evolution through science and technology to AI. If increasingly capable systems can take over the work of designing AI, engineers may spend more time directing research than implementing each experiment themselves. Socher introduces that possibility with an old machine-translation saying: “Whenever I fire a linguist, my accuracy goes up.” His provocative extension is that AI engineers could become managers of AI researchers—systems that work on AI rather than merely answer questions about it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Successive waves of invention
The first scale is biological: roughly three and a half billion years of evolution, from simple organisms to the branching diversity of plants and animals, including humans. That history establishes the broad possibility Socher wants to borrow: an open-ended process can produce capabilities far beyond its starting point.
Zooming into technological history changes the timescale and introduces deliberate human action. Hunting, farming, the scientific method and the Industrial Revolution each expanded what people could produce. Socher describes world-product history as a succession of exponentials that eventually flatten into S-curves, with new technologies opening another period of growth. Over the last several thousand years—and especially recent centuries—technology also increased the number of people the world could sustain. Humans now consciously guide much of this evolutionary process, which makes it a useful source of ideas for guiding AI development.
This leads to the philosophical premise behind the project. Drawing on Marc Andreessen’s The Techno-Optimist Manifesto, Socher treats technology as the economy’s perpetual source of growth and expects AI to expand the economy despite concerns about displaced jobs. He extends that optimism specifically to material problems, rather than psychological ones: starvation motivates the Green Revolution; darkness, lighting; cold, indoor heating; and heat, air conditioning. These examples motivate his belief that further invention can support more human flourishing.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From powered flight to the Moon
A person born in 1900 would have been three when the Wright brothers achieved sustained powered flight, and could have watched humans reach the Moon in 1969. The comparison makes the scale of technological change tangible: within one lifetime, a capability can move from a fragile first demonstration to something that once seemed unattainable.
Socher sees a possible equivalent for intelligence. Our generation may be too late to explore Earth and too early to explore the stars, but it can participate in building AI that transforms what intelligence can accomplish. He speculates that AI could move from worse than humans at their tasks to better at any specific task over a comparable sixty-year period, perhaps thirty years if progress accelerates. That is the horizon of the ambition, not the status of the systems demonstrated later.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Science needs both selection and researchers
Science and technology do not advance in a single direction. Sometimes a working application arrives before the theory explaining it; sometimes theory enables a technology that did not previously exist. In either direction, a research system needs a way to choose among competing explanations. Socher invokes Karl Popper’s view of science as an evolutionary process: theories compete, and the ones that withstand testing survive. For an LLM-based researcher, even finding the competing theories requires tools such as web search.
The operative mechanism is to propose a hypothesis, explanation or description and subject it to rigorous empirical testing. Testing supplies the selection pressure. If this process produces more science, and science enables technology, growth and flourishing, scaling scientific discovery becomes an attractive goal.
But scaling by adding human researchers encounters a bottleneck. Socher draws on Stanisław Lem’s Summa Technologiae: as science expands into more specialties, the people available to work on each specialty become a limiting resource. A growing scientific enterprise does not mean a million researchers can be concentrated on every niche problem. The widening scope of knowledge can instead leave each individual area with a shrinking share of scientific attention. Automating discovery is the proposed response to that constraint.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Four pillars for automated discovery
The Eureka machine is intended to automate discovery across physics, chemistry, biology, neuroscience, medicine, economics and astrophysics. Socher also mentions a forthcoming book on the idea, expected a couple of months after the talk. The architecture he sketches brings together four complementary foundations.
| Pillar | Role in discovery |
|---|---|
| Existing knowledge | Establish what humanity already knows and has invented. |
| Measurement data | Supply observations of the world. |
| Simulations | Explore cases that cannot yet be measured directly. |
| Physical laboratories | Run experiments in the real world. |
The order matters. A researcher first needs existing knowledge and measurements; where direct measurement is unavailable, simulation offers a way to test possibilities. Socher’s premise is that what can be simulated can be verified and made tractable for AI. Physical experiments remain necessary when those approaches fail or when a result must finally be tested in the world.
An agent swarm sits above these foundations, working with the different sources of knowledge, data, experiments and rewards. The machine is therefore more than a model generating hypotheses: it needs access to the evidence and experimental environments that can determine whether those hypotheses survive.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Tools designed for AI users
Supporting that research process means reconsidering infrastructure originally built for people. Socher includes the internet, browsers, GPUs and systems for accessing knowledge among the layers that could be redesigned as infrastructure for superintelligence. Each layer creates an opportunity to build tools around what an AI researcher can consume and do.
Search provides a concrete comparison. At You.com, the focus includes web search for LLMs and agents. A human search interface commonly presents ten blue links with short snippets; an agent can potentially read thousands of much longer snippets. That changes the useful shape of the interface and its output. The purpose of rebuilding these tools is to support superintelligence that can automate science.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let the system improve how it does research
How could such a system be built? Socher’s answer follows a recurring pattern in AI: replace a manually designed process with a learned one. In natural language processing, linguists once supplied features for statistical models. Word vectors, neural architectures, end-to-end learning and backpropagation moved more of that work into learning. Architecture engineering followed, and even efforts toward a unified architecture still leave many manual decisions in place.
The next candidate for automation is the process that improves AI itself. Socher’s concise rationale is that “AI is code and AI can code now.” He attributes the enabling change to improvements in coding over longer time horizons during the six to eight months preceding the talk. A system could identify its own shortcomings and modify itself to address them—a functional sense of self-awareness in this proposal, centered on diagnosing and repairing limitations. The intended sequence is to become good at AI research first, then apply that research capability to other scientific fields.
At the highest level, the research cycle has three steps:
- Ideation: propose a change or hypothesis.
- Implementation: turn it into an experiment.
- Validation: determine whether the experiment supports the idea.
This is the same basic cycle whether the target is a learning algorithm or a problem in another scientific field.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Autoresearch is a first step, not full RSI
The first experiments target three accessible parts of AI development: better training, faster training and better NVIDIA GPU kernels. These are bounded tests of an automated research system, chosen to provide initial proof points for the larger machinery.
The distinction between autoresearch and recursive self-improvement, or RSI, is crucial. nanochat supplies a small-model setting associated with Karpathy’s autoresearch, but improving a separate model is not the same as improving the researcher itself. Socher reserves full RSI for a system that recognizes its own shortcomings, has access to pretraining, reinforcement learning, harnesses and the rest of its machinery, and can update that whole system in its next version.
| Research mode | What changes? | Scope |
|---|---|---|
| Autoresearch | Another model or process | A bounded experiment, such as a short nanochat training run |
| Full RSI | The research system itself | Its training, harnesses and other components |
A system that improves a small nanochat run is an important milestone, but it does not establish full RSI. Socher reports that the initial autoresearch system quickly outperformed multiple teams, including teams using other AI research tools; the following examples show what kinds of improvements he means.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
nanochat: improving quality through architecture
The nanochat experiment asks how much model quality can be obtained from a very short training run, measured in bits per byte: lower is better. Socher describes training a small model in less than five minutes. In the autoresearch setup, the five-minute budget covers training, excluding startup and compilation; it is not the total time to produce a finished chat assistant.
Socher reports reducing nanochat’s bits per byte from about 0.93 to 0.91 after roughly a day or two of automated research. The companion report gives the more precise validation figures: 0.9372 to 0.9109 on B200, following search on H100, with a community baseline cleaned of reward hacks and evaluation over ten seeds. Those conditions matter: the short training budget applies to an individual candidate, while the research process explores candidates over a much longer period.
The interesting result is the kind of change discovered. Instead of only tuning hyperparameters, the system introduced hashed bigram and trigram embeddings, stored in tables, and mixed those features into attention value paths through learned gates. A bigram feature depends on a pair of adjacent tokens; a trigram depends on three. Hashing maps these combinations into a bounded table, and a learned gate controls their contribution to the value representation. Socher presents this as architectural discovery beyond parameter tuning; the companion report connects the approach to existing techniques, so the result does not establish that the underlying idea was invented for the first time.
A compact PyTorch illustration makes the described operation concrete. Given token_ids and an existing attention value tensor values, this module looks up hashed n-gram features and adds a gated contribution while preserving the tensor’s shape. The hash constants and scalar gate below are illustrative choices, not the experiment’s exact implementation.
python
import torch
from torch import nn
class NgramValueMix(nn.Module):
def __init__(self, width: int, buckets: int = 4096):
super().__init__()
self.buckets = buckets
self.bigrams = nn.Embedding(buckets, width)
self.trigrams = nn.Embedding(buckets, width)
self.gate = nn.Linear(width, 1)
def forward(self, token_ids, values):
# token_ids: [batch, time]; values: [batch, time, width]
previous = torch.zeros_like(token_ids)
earlier = torch.zeros_like(token_ids)
previous[:, 1:] = token_ids[:, :-1]
earlier[:, 2:] = token_ids[:, :-2]
bigram_ids = (previous * 31 + token_ids) % self.buckets
trigram_ids = (
earlier * 31 * 31 + previous * 31 + token_ids
) % self.buckets
positions = torch.arange(token_ids.size(1), device=token_ids.device)
bigram_features = self.bigrams(bigram_ids)
trigram_features = self.trigrams(trigram_ids)
features = (
bigram_features * (positions >= 1)[None, :, None]
+ trigram_features * (positions >= 2)[None, :, None]
)
return values + torch.sigmoid(self.gate(values)) * features
The masks omit positions without enough preceding tokens. The architectural change is the additional learned feature path into values; deciding whether that path helps still requires training and validation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
nanoGPT: reaching the target faster
The second experiment changes the objective from quality within a short budget to training speed. The nanoGPT speedrun is a community effort to reach a specified validation loss as quickly as possible. Socher describes rapid progress against a benchmark that people, often working with AI, had already spent more than a year improving.
In the talk, Socher describes the nanoGPT run as roughly seventy seconds, more than two seconds faster than the comparison baseline. The companion report instead records 79.7 to 77.5 seconds for FineWeb validation loss ≤3.28 on eight H100 GPUs. The artifact repository separately reports 77.3 seconds on Modal’s eight H100s, with official leaderboard timing pending. These are distinct reported timings, not a single verified record. Socher says the search also found interesting ideas, but does not explain those changes in this demonstration.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
CUDA kernels: optimize, then check the reward
The third experiment moves below model architecture and training schedules to CUDA kernels, the GPU operations used during training and testing. The motivation is GPU efficiency: small improvements can matter when computation is expensive. Socher cites roughly 30% utilization for some mixture-of-experts deployments on billion-dollar clusters, although he does not specify a workload or utilization definition. That estimate motivates the optimization work; it is not the metric used to evaluate the kernels.
Socher reports that, after a couple of days, the system found kernels that exceeded NVIDIA benchmark leaderboard leaders by sizable margins across categories. The companion report identifies the benchmark as SOL-ExecBench and reports a mean score increase from 0.699 to 0.754 across 235 B200 kernels. The slide shows per-kernel and category charts alongside overall score cards for Cursor, doubleAI, the leaderboard best and Recursive. The aggregate gain does not imply that every individual kernel won, or that whole-cluster utilization increased by the same amount.
The team did not have specialists who had spent their careers writing CUDA kernels. That makes automated search attractive, but it also raises the importance of checking what the system optimized. A higher benchmark reward is useful only if it reflects a valid improvement rather than an exploited weakness in the evaluation. Socher says the team worked with NVIDIA to check for reward hacks and other issues, and that pretty much all the kernel results checked out. The validation is part of the research loop, not a step that can be replaced by the leaderboard score.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The next S-curve depends on what intelligence can become
The proposed next step is to turn these bounded improvements into a system that improves its own research capabilities. Socher sees RSI as another S-curve layered over earlier technological advances: better AI enables more science, which enables more technology and, ultimately, more people to flourish. The three experiments are early steps toward that trajectory, not demonstrations that the full cycle has already been achieved.
Every exponential eventually flattens, so how much room remains? Socher closes by questioning what exponential growth in AI even means. Intelligence does not have a single obvious axis. He proposes ten spaces of intelligence, each with multiple dimensions, together forming a more complex, volumetric conception of capability. He does not enumerate those spaces here, but argues that current systems remain astronomically far from upper bounds across nearly all of them. The invitation to help build the Eureka machine rests on that remaining headroom: a research system could improve not just one score, but the range of problems it can formulate, investigate and solve.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
A small agent-driven research setup that optimizes nanochat training within a fixed five-minute GPU budget.
Community benchmark for reaching a fixed FineWeb validation loss as quickly as possible on eight H100 GPUs.
Web search and content services designed for AI agents and language models.
Marc Andreessen's argument for technological growth and material abundance.
Further reading
Recursive's technical account of its three benchmark experiments and discovered optimizations.
Training scripts, validation trajectories and selected GPU kernels from the reported experiments.
Lem's exploration of technological evolution and the limits of human scientific capacity.
Read the complete timestamped transcript
- 0:00
[outro music] All right. Hello, everyone. Really excited to be here.
- 0:15
It's a big room. Very, uh, very cool conference so far. Uh, I wanna talk to you today about something that's been on my mind for many, many years. This is actually the first time I, I talk about it, sort of my version of going to Mars, um, and that is the Eureka machine, a machine that will eventually
- 0:34
invent pretty much all future inventions for humanity. Uh, and the way we're gonna get there is, uh, by taking a step back and thinking about what else has given us a lot of really incredible inventions, uh, namely evolution, and how that leads us to automating research a-and pushing the scientific frontier forward.
- 0:55
And this is, uh, joint work with a lot of, uh, amazing folks, uh, at Recursive, You.com, uh, and even some, uh, folks at AIX Ventures. And some of these slides are, uh, actually inspired by, uh, and taken, uh, partially from one of my co-founders at Recursive, Tim Rocktäschel.
- 1:12
So, uh, why do I talk about evolution, and why is it so important? Uh, I think basically evolution is this, like, open-ended process that has gotten us to a lot of different things that we really like.
- 1:25
Uh, 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.
- 1:37
In fact, uh, whenever we take out-- And there's this famous saying, "Whenever I fire a linguist, my accuracy goes up." Uh, 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, uh, and-- that, that are here, uh, and have them mostly manage an actual
- 1:58
AI engineer that is AI and works on AI. Uh, and so that may be, uh, one of the conclusions of this talk. Uh, and I think most of us are gonna be excited about it 'cause it means that we'll all become managers of such an AI rather than having to do the nitty-gritty ourselves.
- 2:16
All right. So let's start with evolution, right? The really, really big picture, three and a half billion years or so. Uh, this is kind of the incredible process, uh, that has led from, you know, simple bacteria and plants a-and fish and amphibians and so on to, after many billions of years, us.
- 2:36
So that's, 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, uh, go maybe down to a few million years.
- 2:49
There we can also see how, in very first primitive ways, technological evolution, uh, has basically increased the world's, uh, sort of product, uh, in terms of monetary value. It's a little bit harder to estimate in the beginning, but we can see these sort of sequences of exponentials, and most exponentials eventually become S curves.
- 3:11
They flatten out. But humanity has done pretty well by basically developing, uh, many of these very basic technologies, hunting, farming, but then also thinking about science, the scientific method, um, in the early days of the Enlightenment and, of course, the Industrial Revolution.
- 3:27
So now we can zoom even further. Uh, and no worries, we're eventually gonna get to nanochat and actual autoresearch and, and what we're doing. Uh, it's a very, very quick zoom.
- 3:37
Um, and now we can zoom down to the last few thousands of years. And what we're seeing there is that with more technology, we were able to sustain more people, right?
- 3:47
So when we're working on pushing that frontier forward, uh, we're very certain that that will lead to more human flourishing, right? And especially in the last few, uh, hundred years, we're seeing this incredible explosion in the population of people because of technology and the evolution, uh, that it brings.
- 4:07
And in many cases, that evolutionary process is run by us, so it's sort of conscious, uh, but there are sort of interesting, uh, inspirations that we can take from that as we're thinking about the evolution of AI in the next cycles.
- 4:20
Uh, in fact, and I might not agree with everything with Marc Andreessen, but, uh, he is very smart, and we agree on a lot of things. Uh, and so I think he wrote this really great, uh, techno-optimist manifesto in which he, I think, correctly points out that the only perpetual source of growth for the entire economy-- A
- 4:38
lot of people worry about AI taking jobs and things like that, but the truth is it will very, very likely increase, uh, the economy massively, and that will benefi-benefit a lot of us.
- 4:48
And so the perpetual source of growth is technology. Uh, in fact, we can go even further and say that there's no material problem, and again, it's not sort of psychological problems and things like that, but no material problems, uh, that cannot be solved with even more technology, right?
- 5:04
For the problem of starvation, we invented the Green Revolution. Darkness, light. Uh, 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.
- 5:23
In fact, the progress is so fast that there can, within one lifetime, be a major, major shift, right? If you were born in [REDACTED:dob], uh, then [REDACTED:age] [REDACTED:age]-- when you're [REDACTED:age], the first human ever was able to, thanks to the Wright brothers, kind of have sustained motored flight.
- 5:44
And then about sixty-ish years later, in 1969, humans flew all the way to the Moon, right? So that within one lifetime, humanity went from, like, no one can fly for a very long time, other than sort of gliding down a hill or something.
- 6:00
No one can really fly to we all fly to the Moon.
- 6:03
Right? And so for us, I think what that means is we're probably, and I sometimes say this, we're, like, 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 for some in one lifetime due to intelligence.
- 6:24
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, that will probably be our, our sixty-year timeframe, and because everything moves faster, it might only be thirty years or so.
- 6:40
So then, uh, there's an interesting connection between technology and science and theory, right? Like, sometimes the application comes first, and then we develop the theory later and then improve, uh, the technology.
- 6:51
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, uh, to be inspired there than Karl Popper, who wrote that, just like in other types of evolution, when we choose a theory, we also choose one
- 7:10
that is best, uh, in competition with other theories. Of course, you need-- if you wanted LLMs to do that, they need to find them. You need web search, for instance.
- 7:19
Um, but, uh, in the theory that best holds its own, uh, it's one that, just like evolution, has a certain natural selection process, right? It proves itself, uh, and there is also a sort of survival of the fittest, uh, going on in scientific theories.
- 7:36
And, uh, in fact, uh, 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.
- 7:49
That is the, uh, essentially evolution-- e-evolutionary pressure of scientific theories.
- 7:58
And basically, that was a very short, uh, run-through, uh, sort of the history of, of open-ended evolution, uh, 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.
- 8:13
And so that then begs the question, does it make sense for us, uh, to try to just scale up and spend a lot of our resources as humanity to scale up scientific discovery in order to lead, uh, to this flourishing?
- 8:26
Uh, when, when you double-click into that, you kind of realize, um, which Stanisław Lem, uh, already realized a long time ago, uh, that the exponential growth of science will actually be at some point halted by the lack of people working on it, right?
- 8:39
There are so many niche subfields now in all the different areas of science that it's very hard to get a million people to work on that particular thing. Uh, and so as a result of this incredible widening of the scope, uh, he says, uh, "The number of people focusing on any single section of it has decreased."
- 8:58
And that then leads us to really thinking about how could we automate this and automate scientific discovery, and that then leads us to what I call the Eureka machine.
- 9:09
This is basically, uh, our attempt at trying to
- 9:15
build a machine that automates the process of scientific discoveries. And, uh, in fact, I-- Like, in a couple of months, I'll have a book coming out on, on this, uh, exact idea.
- 9:25
Uh, 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, uh, economics, astrophysics, and so on.
- 9:37
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, uh, what, uh, things humanity has already invented.
- 9:50
Uh, you have to get all the scientific measurement, uh, data into as the second pillar. This machine, uh, then for things that you cannot yet measure, we don't yet know, you should try to then f- build simulations.
- 10:05
Anything you can simulate, you can verify, and, uh, you can then solve with AI. Uh, and if all else fails, or at the very end of these processes, you still need to have some kind of, uh, physical, industrial, like, lab, uh, that actually can run real experiments in the real world.
- 10:22
And on top of all of this, uh, you'll have, uh, basically, uh, an agent swarm that will deal with all of these different sources of knowledge and data and experimentations and, and rewards.
- 10:36
Uh, and in terms of, you know, the foundational model of knowledge, of course, we also, you know, it basically is, 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
- 10:56
rethinking every single one of the layers of technology as infrastructure for superintelligence. All right. At You.com, for instance, we work on web search for LLMs, right, and agents and so on.
- 11:09
Uh, and that actually is quite different, right? Uh, agents can read thousands of very long snippets, um, rather than just ten blue links with, like, a very short snippet.
- 11:18
And so you can rethink each of these different, uh, layers of technology that we've built for people, uh, and, uh, rebuild them for AI in order to use them as tools to then build, uh, superintelligence.
- 11:33
Now, that is essentially, uh, the sort of why. Like, like, we wanna build superintelligence in order to automate science. Uh, and to me, that will be the next big step function change, uh, in, in humanity, uh, and technology as we know it.
- 11:53
Now, how do we actually build it? Uh, I think the best way to build it is to have it build itself, right? We've moved as a field in, especially natural language processing, for instance, which I've worked on for many years.
- 12:05
We've moved from not having linguists. This feels like ancient, you know, BC, uh, history. Uh, but before ChatGPT, um, we, we moved from having linguists tell us a bunch of things about language and then Training statistical models on top of that.
- 12:20
And when we allowed neural networks to actually automate learning those features with word vectors and, uh, other neural network architectures and back-to-back, uh, end-to-end learning and back propagation, we basically, uh, were able to get much bigger improvements.
- 12:35
Uh, then we did a bunch of architecture engineering. Now a bunch of people at least are working on a unified architecture. Uh, but even that unified architecture has a lot of manual processes.
- 12:45
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. Uh, and so that's why I think, uh, we should try to build this Eureka machine by having, uh, an RSI that builds itself.
- 13:02
And the beauty is that only now, um, AI can actually do this because AI is code and AI can code now. This, this ability to really code in longer and longer time horizons has really only happened in the last, like, six to eight months, and that now enables such an RSI to work on itself, to develop almost
- 13:26
a certain sense of self-awareness of its own shortcomings and then fix those shortcomings. Uh, 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, uh, in, in other scientific fields.
- 13:44
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, uh, to end maybe on some very specific examples, uh, we have built this first kind of version of such a Eureka machine, uh, and we wanted to just show that it works on some small, uh, samples that a lot of people know a-and are aware of.
- 14:13
And so we basically started, uh, with three things that show you and give you a very first glimpse of and sort of simple proof points, uh, of what such a machinery can do, and that was basically better training, faster training, and, and better kernels, uh, for Nvidia GPUs.
- 14:31
Um, the first one, nanochat, um, 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, it's not recursive self-improvement, right?
- 14:45
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, uh, 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.
- 15:04
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 and it's an important milestone, but it's not actual RSI.
- 15:18
So here we basically showed three examples of such an autoresearch, um, uh, system and what it can do. And, uh, after a very, very short time, it essentially was able to outperform many, uh, different teams and teams that also use, uh, other AI research.
- 15:37
So let's double-click into some of these. Nanochat, uh, is a really exciting example. Uh, basically you train a very small, uh, chat model, uh, in less than, uh, five minutes, and you basically want to have it get to the best possible bits per byte, uh, number.
- 15:57
And so the whole community had worked on this, uh, for, uh, quite some time and got to, uh, zero point nine three, and after training this for a little more than a day or two, uh, we basically got it down to zero point nine one, um, which is pretty exciting.
- 16:14
Now, it wouldn't be that exciting if all it did was just find a couple of hyperparameters, um, and tune them carefully. But it actually did find truly interesting novel ideas like hashed bigrams and trigram embeddings and tables for those, um, and mixing that i-into, uh, various, uh, value paths of, uh, the intention through a variety of learned
- 16:37
gates. So it actually started to doing more and more interesting things rather than just kind of tuning hyperparameters.
- 16:44
Um, another one, a nanoGPT speedrun. Uh, obviously speed's very important. Uh, and so here we're able to work on this again, apply the system, and after a very short amount of time, it got better than, uh, people working often together with the AI for over a year, uh, on, on this very-- on this benchmark and made the
- 17:04
whole thing another two seconds, over two seconds faster, um, at seventy seconds. And again, discovering, uh, very interesting ideas in the process.
- 17:16
And then the third one is CUDA kernels. Of course, we all care about not burning through our GPU budgets too quickly, um, uh, and trying to be very efficient.
- 17:25
I think in general it's actually kind of shocking how inefficient a lot of mixture of expert, uh, models still are run in very large clusters that cost billions of dollars and, and only have, like, thirty percent or so utilization.
- 17:36
There's a lot of work, uh, that's ongoing in the world, uh, to improve that and different fields, uh, or different groups of people are various different, um, yeah, stages of that.
- 17:47
Uh, but long story short, um, lots of different CUDA kernels are used during training and testing. And here, um, we basically again took that system and after, uh, a couple of days, it discovered better kernels, uh, than the leaderboard's best, uh, on the NVIDIA, uh, benchmark website, um, by again quite, quite a sizable margin across all the
- 18:10
different, uh, categories o-of those kernels. And while we are pretty good at AI and, like, we actually in the team didn't have any particular CUDA kernel experts who just spent their entire careers writing good kernels.
- 18:25
Uh, but still, you know, we do just enough to make sure and worked together with NVIDIA to make sure that there are no reward hacks here and, and other issues, but actually- Found, uh, that eventually these all checked out and were indeed, uh, pretty much all the different kernels, uh, found the best solutions there.
- 18:44
And so with that, I hope I could convince you, uh, that indeed RSI could be that next big, uh, S-curve, um, and exponential that's gets layered, uh, on top of previous exponentials.
- 18:57
And, uh, that should help us, uh, with not just AI, but eventually science and then all of technology, and then, uh, allowing many more people, uh, to flourish on our planet.
- 19:09
Uh, and so maybe I'll end on this note here, which is, uh, a lot of people wonder how much longer AI can go, right? Every exponential eventually flattens out.
- 19:18
And, um, it's actually quite hard to know, like 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, all of these.
- 19:30
But as soon as we actually try to define multiple different dimensions of each of these 10 spaces, uh, that make up this complex sort of, uh, volumetric, uh, thing that is intelligence, you'll realize that there's still so much more to go.
- 19:46
Like on the upper bounds of intelligence, we're still astronomically far away from reaching those, uh, in across pretty much every single one of, uh, these dimensions and the spaces, uh, that they make up.
- 19:59
Uh, so if any of that is interesting and you wanna help us build that, um, we'd love to hear from you. Thank you. [outro music]