Recursive Self-Improvement and the Future of AI
Martin CasadoDhaval AdjodahOwen He
- The founders of an RSI startup—Dhaval Adjodah, who did a PhD in reinforcement learning at MIT Media Lab, and Owen He, formerly a DeepMind research scientist—say they are seeing empirical glimmers of recursive self-improvement: performance improving while the resource denominator stays fixed or falls. Asked whether this is happening, Owen says, “I don’t think it’s just possible. I think it’s happening”; they say results have grown over the last few months and they have held back releases until they are sure the results are correct.
- Their definitional test distinguishes RSI from “AI creating AI”: ask what the optimizer and the optimizee are. Humans using better AI tools is “autocatalysis”—the ideas and improvement still come from human brains. RSI, they argue, would put the language model itself in the optimizee role, closing the loop; in their framing, marginal gains could then stay constant or grow.
- Casado frames capability improvement in terms of performance over cost, but a closed loop alone doesn’t establish takeoff. A fixed numerator leads to a plateau; if it grows with capability, improvement can be constant, and if it grows as a higher power, the model implies exponential takeoff. A founder says the relevant exponent must be measured empirically; they lack a learning theory of deep learning.
- The denominator matters when evaluating RSI claims. Casado notes that model release times are falling, but says the trend looks different relative to capital raised. The founders track experiment costs—including human effort—on a “bleederboard.” They cite nanoGPT: GPT-2, once taking weeks on multiple GPUs, can now be retrained in hours on the same GPU for about $100, using the same data. Casado’s rejoinder is that he has yet to see the equivalent for GPT-5; the founders add that efficiency can buy more rollouts and better training trajectories.
- Their commercial wedge is inference efficiency, an AI problem they believe can improve both performance and cost. Faster inference also lets a system do more inference; the founders say their multi-LLM agent can choose what to optimize, and even its master prompt has been improving. They aim to optimize across prompts, harnesses, weights and chips, rather than isolated components.
- They say many published RSI ideas don’t scale in their tests, and practical bounds remain unclear. Their framework includes human researchers as agents and is intended to generalize across problem domains, though physical processes and the slowest component constrain progress. They contrast fast-evaluation areas like coding and math with drug discovery’s wet-lab and clinical-trial bottlenecks. Owen speculates that the universe’s computation sets an ultimate ceiling; he says he has yet to see a task-specific bound on the energy needed to solve a problem.
1. RSI is not “AI building AI” — it’s the optimizer becoming the optimizee
- Their definitional test for pitches Casado hears, such as “we use AI to create GPU kernels,” is to ask “what is the optimizer and what is the optimizee.” Humans using ever-better AI tools is “autocatalysis”: the ideas and improvement still come from human brains. In their view, RSI puts the language model itself in the optimizee role, reinvesting cognitive improvement into itself.
- Casado points out that every technical epoch had autocatalytic effects—computers helping build computers, for example. The founder’s distinction, via control theory, is that autocatalysis still needs humans to fill a gap, while a closed loop could make marginal gains constant or larger over time.
2. The capability-improvement model separates plateau, constant improvement and takeoff
- Casado frames the rate of capability improvement in terms of performance (the numerator) over cost or difficulty (the denominator). With a fixed numerator, as in his steam-engine example, the system plateaus. If the numerator correlates with capability, the result can be constant improvement; if it correlates with a higher power of capability, the model yields exponential takeoff.
- Casado cautions that a loop diagram alone doesn’t establish takeoff. A founder says it is hard to know what a nonlinear system will do and that the exponent must be measured empirically. They lack a good theory of deep learning; scaling laws are among the closest available guides, and are empirical too.
3. The denominator matters — count experiment costs and human effort
- Casado observes that model release times are decreasing, but says that trend no longer holds when considered relative to the capital companies have raised. The founders say improvement that requires pouring in more resources is not RSI, and that they include human effort and experimentation in the denominator. They track experiment costs on an internal “bleederboard.”
- They cite nanoGPT: GPT-2, which originally took weeks to train on multiple GPUs, can now be retrained in hours on the same GPU hardware, with the same data, for about $100. The improvements include algorithmic, data and kernel techniques; many are auto-research results. Casado’s rejoinder is that this is not yet demonstrated for GPT-5. The founders add that improving efficiency can enable more rollouts, better trajectories and smarter models.
4. A unified framework and the full stack matter
- The founders say many papers claiming RSI fail when tested in their internal infrastructure at scale on benchmarks and real economic problems. They also say that results announced in the field can prove to be noise or cherry-picking when they try to replicate them. Their view is that the simpler system that closes the loop is what scales.
- Casado invokes Feynman’s distinction between “stamp collecting and physics.” A founder argues that without a unified mathematical framework, researchers can’t tell whether the overall system is improving or only one component; optimizing one part may affect another.
- RSI can span prompts, harnesses, weights and chips. The founders started with harness improvements, then added weight updates, and expect to co-optimize multiple parts. Their agent chooses what to work on, and even the master prompt has been improving.
- They chose inference efficiency because it is close to AI itself and can improve both performance and cost. Faster inference means the system can do more inference; the founders also describe efficiency gains as enabling more rollouts and better training trajectories. Serving the latest open-weight models, such as Kimi K3, is also an economically important problem.
5. “It’s not a belief — it’s empirical measurement”: glimmers, variance and limits
- Asked whether performance can keep rising while the denominator stays fixed or falls, Owen says, “I don’t think it’s just possible. I think it’s happening.” He says they have seen results grow over the last few months, but have not released results until they are sure they are correct. He also notes that benchmark accuracy is capped at 100%, so it may not be the right metric for detecting exponential progress.
- Their measurement culture includes an engineer nicknamed the “Viceroy of Variance,” whose job is to try to break their results. Owen says there is no substitute for “quality time”: running hundreds of experiments to distinguish real effects from variance and reward hacking, including cases where an agent two levels down has hacked the process. Dhaval describes stochastic noise, oscillations and even different behavior between two B200 GPUs.
- Owen says intelligence is not a single scalar but a multidimensional vector: some areas may take off while others plateau, with progress constrained by the slowest component. Dhaval applies Amdahl’s law: speeding up digital work by a thousandfold or a millionfold has limited impact if it is only part of the total process. Drug discovery remains bottlenecked by wet-lab work and clinical trials, unlike coding and math, where evaluations are faster and cheaper.
- Owen says he has yet to see a quantitative bound on the energy needed to improve a system by a given amount or solve a given problem. He speculates that physical limits ultimately bound computation, citing roughly 10⁸⁰ particles and the speed of light as inputs to an upper bound. Dhaval’s “garden hose company in Wisconsin” joke suggests that important RSI applications may be hidden in unfashionable workflows; he compares this to engineers treating relativistic effects as an empirical correction before understanding them as physics. When Casado asks when to expect more, Dhaval says, “Very soon.”
Full transcript
The fundamental idea here is to build AI that improves itself, and I think we’re getting there.
For the first time, we could swap the optimizer and the optimizee. This leads to a compounding reinvestment of cognitive improvement into itself. You can think of RSI at many different levels of the stack: prompts, harnesses, weights, and all the way to chips. Our goal is to optimize all the levels of the stack. Our approach to RSI is a multi-agent system working together that’s been recursively improved.
1. It's not a belief, it's happening
I’m so impressed by your belief in this idea of a model improving itself.
But it’s not a belief.
It is a dramatic statement. I think it’s the first time I’ve heard you say it.
We’ve been so thorough that we haven’t released things until we’re sure they’re correct.
Do you envision a system where the amount of energy that goes into it somehow decreases over time relative to the improvement?
2. What actually counts as RSI
We can’t share too much, but one of the very important things we’ve tried to do—
Welcome, both of you. We’re going to be talking about everybody’s favorite topic: RSI, or recursive self-improvement. Maybe to start, it’d be great if you could introduce yourselves and your backgrounds. Then we’ll talk a little about what you’re working on now, and then we’ll dig into the content.
I think of myself as a physicist of some kind. I studied physics, then did a PhD in reinforcement learning at the MIT Media Lab. I really like building things, and I’m excited to be building a cool startup with Owen and Grace.
My name is Owen. I first studied math. One day, I realized that all mathematicians are proving theorems. I was taking a class called Formal Language and Logic, and they were showing how you can build a theorem prover. I thought, “Oh, that would be cool. I can just build a machine that proves theorems for me.” That’s how I got into AI.
Amazing.
For my PhD, I worked on continual learning and meta-learning. Then I worked for a few years at DeepMind as a research scientist. Now I’m working on RSI.
I see a lot of pitches come in, and a lot of them use the term RSI. I’ll ask, “Oh, so what do you do?” They’ll say, “We use AI to create AI.” That sounds very interesting. So what do you do? “We use AI to create GPU kernels.” That doesn’t quite match the definition of what I would consider RSI. So maybe we can start by you defining what you think you would include in RSI, definitionally?
Yeah. RSI is definitely an overloaded term today. Many people mean different things when they say RSI. The important thing when you’re differentiating them is to ask: What is the optimizer, and what is the optimizee? When you’re self-improving, who is improving?
There are people who say, “Okay, we’re doing RSI. We have humans using AI to develop better AI.” But to us, that’s just autocatalysis. The improver is still a human with tools, even though those tools are getting better and better. The ideas and the improvement still come from human brains.
Right, and just—I don’t mean to interrupt, but to be clear, most technical epochs have had autocatalytic effects, right? There’s no way you can create a computer chip without a computer. You could argue that the development of the internet was sped up because people who weren’t geographically located could coordinate. The steam engine—you could build better steam engines because you could get coal faster, et cetera. So this notion of using the thing to build the thing faster isn’t new. You’re saying something different.
Yes. Throughout the history of civilization, we’ve had autocatalysis—computers using computers to build better computers—but the idea for this change, for this improvement, comes from the human brain. That human brain was never the thing being improved, never the optimizee. Whereas now, for the first time, we could swap the optimizer and the optimizee. The brain of the agents, which is the language model, can be the thing that you’re optimizing. This leads to a compounding reinvestment of cognitive improvement into itself.
But there has to be a reason you decided to take on this hard problem, as opposed to what anybody else would do, which is chew off pieces of the problem.
Yeah. Before I answer your more important question, the reason I did this is that I think it’s a hard problem, and it’s fun working on the hardest problem.
But do you have an intuition that this is the actual right answer, as opposed to something more pragmatic?
Yeah. I think a pragmatic thing would be, from a control theory perspective, that if you have a gap in the system, you can’t have a closed loop. Autocatalysis has no closed loop, right? Whatever you improve, there’s still a gap. Humans have to come in. You have to pour in other resources to make it work.
3. Capability as a numerator and denominator
But if you have a closed loop, you still have to pour in resources, but the marginal gains ideally are either constant or larger over time. That’s the fundamental difference. What’s really different here can be explained pretty easily.
So if you think of capability—
Yeah.
What we’re trying to do here is improve the rate of improvement of capability, right? The first derivative of capability with respect to time is what everybody is trying to improve.
Yeah.
And now you can put that as being driven by two things: the numerator and the denominator. The numerator would be performance.
Yeah.
Which is what everybody is trying to improve. And the denominator is some kind of cost, let’s say capital, right? This one is pretty set because the higher the capability, the higher the difficulty of the next step. Right now, the numerator is the interesting part. For things like steam engines and whatnot, that was fixed.
Yeah.
Right. So then you end up with a very simple differential equation where you basically plateau.
Yeah.
Right. Now, if you have a system where the thing being improved is also the thing that’s doing the improvement—
You now have a new way to think about it. The simplest case would be that the numerator is also correlated with capability, which makes sense, right? Because the better you are, the better you are at improving.
And if you do that, the math cancels out, and you basically get constant improvement over time. That could be one level of RSI, right? And then there’s one more case—
Now the numerator could be correlated with a higher power of capability.
Yeah.
And then you get exponential takeoff.
4. Closed loop vs takeoff
Yeah. Can we dig in a little more on the difference between RSI having closed the loop and actual convergence properties? For some reason, when people see a graph where there’s a circle with an arrow to another circle, and it goes back, they look at that closed control system and say, “Takeoff!”
Right, which—
I mean, if this is a nonlinear control system—
It’s very difficult to know what the hell happens, by the way. Most people don’t seem to understand this distinction.
Because we very quickly go to takeoff and FOOM and all that stuff when we talk about these things.
Well, as a network guy, you know it depends on the exponent of your function. That’s self-referencing. What we’ve found is that it’s almost not useful to speculate on what we think could happen, because this is really an empirical problem. The real way to know what’s happening is to test things. Testing things is really hard; reward hacking is one of the many problems we have to deal with. The question you’re asking is very empirical.
Interesting. I’m a networker, so that’s how I would have answered. I would have thought math people would have been like, “We’ve got this differential equation,” and taken out a pencil and written down the answer.
I mean, it is possible to write down the differential equation, but you still have to empirically measure the variable, the exponent.
So there’s no analytical solution?
Honestly, we don’t even have a learning theory of deep learning. We don’t fundamentally know why it works.
Right. And now you’re five or six layers on top of deep learning.
So we don’t have a good theory. I think the closest things we have are scaling laws, and those are empirical laws. One misunderstanding people have about our approach to RSI is that they think the entire system has to be fully automatic and digital. But if you look at our mathematical foundation, we have a multi-agent system, not just one digital agent. In this multi-agent system, human researchers and engineers are actually a special case of these agents. It’s a hybrid system of multiple agents working together that’s been recursively improved.
I see. We don’t want to leave out any component and fail to optimize it. If it’s possible to optimize it in the model, why wouldn’t we do that? You’ll have more gains if you include those. Anything you don’t optimize will later become your bottleneck.
One observation we’ve made is that there are claims that speed is improving, which is actually the case, right? If you look at model release time, it’s decreasing.
And so, on one hand, you can say, “Clearly, we’re in some sort of rapidly improving phase.” But if you divide by the actual amount of capital these companies have raised, that’s no longer the case. Maybe they just raised a whole bunch of money, and AI allows them to apply that money. This isn’t for you personally, but how would you encourage people who are evaluating these claims, or even working on them, not to fool themselves by throwing more money at the problem? How can they show that they’re making meaningful improvement, accounting for the denominator?
For us, it’s pretty straightforward. If you get improvement only by pouring in more and more resources, it’s not RSI. You have to be able to back up your claim about how much improvement you’re getting.
But you do include all the man-hours and experimentation that went into it, right? They’ll talk about the actual FLOPs and money that go in, but the reality is that this is the 23rd experiment, and a human has burned a lot of mental capital to get to that final configuration. Do you actually include that in the denominator?
We do, and I think it’s good company discipline to do that.
Yeah. We’ve tried from the beginning to be very intentional about having a culture of tracking where things go: cost, time, all that kind of stuff. I’ll share a funny thing internally. We have this thing people talk about, leaderboards, which is like a score—
You know what I’m about to say: we have something called a bleederboard.
Where we keep track of the cost of experiments for everybody who’s running them.
That’s awesome.
Everybody can look any day at who’s high up on the bleederboard. I can also give an example from AI. I don’t know if you know the nanoGPT benchmark they’re running. It’s basically asking you to retrain GPT-2, which was a very expensive model to train when it first came out and took weeks on multiple GPUs.
And right now, people can train it on a single GPU within hours, and that time has been shrinking constantly.
Yeah. But that’s also totally separate, right? That’s just Moore’s law or something.
No, no, it’s the same GPU. It’s the same hardware.
Using the same hardware, and then you improve the techniques of pre-training?
Yeah.
Oh, interesting.
What’s improved is algorithmic techniques, data techniques, kernels, all of that kind of stuff.
Do you know how much it cost to train GPT-2?
I think it’s a few hundred dollars.
A hundred dollars.
Oh, wow. So, the same GPU?
Same GPU, same data.
And this is just algorithmic. Do those improvements come from humans or agents?
A lot of them are auto-research results.
But that’s definitely the denominator, not the numerator. What you want to see is that you can do that with GPT-5, which we’re not seeing. The cost of these things is going up dramatically.
I think one of the most closely guarded secrets in the industry is exactly this: how much compute you can get out of this.
Yeah.
Because, in some sense, even if all you were improving was your denominator, you could get faster training because now you can get more rollouts.
Right, yeah, of course.
And that means you can sample the manifold of language more extensively, get better trajectories, and then get smarter models when you use that to train.
For sure, improving the denominator is good.
Yeah.
For sure, improving the numerator is good.
5. Why most RSI papers don't scale
I’ve yet to see that, though it’s great that you have glimmers of it, where you can improve both the numerator and the denominator. That’s a pretty meaningful statement, which is awesome.
What we found is that a lot of the papers that claim to do RSI don’t work when we test them in our internal infrastructure. A lot of them are cool ideas that humans try to introspect and say, “Oh, this is what I think should happen.” They’re complicated and sound cool; they get you a good paper, but when we test them at scale on benchmarks and real economic problems, they don’t scale. What we’ve found is that the simpler thing that closes the loop is what scales.
Yeah.
One thing we do that’s probably unique is that we always start with a very grounded mathematical framework that lays out the map of the whole space.
Yeah.
The reason is that this space is so large that if you don’t do that, and don’t have a principled way of doing it, you get lost very quickly in the complexity. The other reason is that with a really complex system, there are 2 extremes in how you can build it. You can build 200 different systems and hope that the whole Rube Goldberg machine works out.
Yes.
Or you can spend some time thinking about those 200 things and realize they’re actually just 12 abstractions that generalize to all 200. We’ve tried to create a culture of really thinking from first principles: looking at the math, identifying the assumptions we make to be practical, and then seeing what works. When you do that, you can see that multiple papers or ideas are actually subclasses of the same general framework.
Yeah.
If you do that well, you can scale. I think that’s something I see missing. It’s hard to do, and it takes a very unusual group of people to make it work.
I think Richard Feynman said once that there are 2 kinds of science: stamp collecting and physics.
Without a unified mathematical framework, all this research is just stamp collecting. A lot of the papers we’re reading are ideas inspired by some human behavior, which they try to mimic. Then they say, “Look, we have RSI.” But if you don’t put it into a unified system, you don’t know whether the overall system is actually improving or whether you’re just locally applying an idea that’s getting better. We’re putting all the components that you could possibly recursively improve into a unified framework and looking at them holistically, so improving one part doesn’t negatively impact the other.
Right. I see, I see.
For example, some people are doing recursive improvement just on the harness, and others are only doing it on the weights. But weights and harnesses aren’t isolated components; they have an impact on each other. You can think about RSI at many different levels of the stack: prompts, harness, weights, and all the way down to chips.
Our goal is to optimize all levels of the stack. We started with the harness, and we’ve always tried to be very capital-efficient because we care so much about the denominator. Then we started doing weight updates, and we found that you get more gains, but you still have to do harness work. You can keep going down the stack. My guess is that you’ll probably need to co-optimize many different parts of the system to make it work.
6. Why they chose inference
Can you talk a little bit about the problem domain you’re going after, just to prove that this works?
We’re trying very hard to pick areas of AI that are as close as possible to AI itself. Again, we try to be recursive. There aren’t that many problems that are super close to AI and give you improvement on both the numerator and the denominator. One of those is the inference problem.
Right, because the more you can get improvements and efficiency out of inference, the more everything unlocks.
And it’s not just for technical reasons. It’s also the area with the most promise for improving both the numerator and the denominator.
It kind of breaks the mind. Just for everybody listening, there are 2 improvements going on here. One is the model’s ability to make inferences faster. The second is that, now that inference is faster, it can do more of it, faster.
And again, that doesn’t necessarily mean takeoff. That can still happen at some point.
Yeah.
But it’s wild to think of this kind of compounding impact.
Yeah. We’re a startup, not a research lab. One of the reasons we chose this is that everybody in the world right now wants to serve their own model.
Yeah.
Everybody wants to use the latest open-weight model—for example, Kimi K3 or whatever’s just coming out—and serve it as efficiently as possible. It’s an economically extremely important problem right now.
Just for the sake of discussion, if I wanted to make inference for Kimi K3 faster, would I use Kimi K3 recursively, or would I use a separate model and system to do that for Kimi K3?
To improve the serving stack of Kimi K3.
Well, our system isn’t just using one LLM. We have many different LLMs doing many different things.
Oh, so you’re optimizing inference for all of them?
I mean, we pick whichever is most effective, and it’s being used most of the time. But the interesting thing is that the choice of which LLM to use for which task isn’t always up to us. We let the model choose. To be honest, today it’s not really us deciding what to optimize; it’s our agent, because it’s recursively self-improving. We’re trying not to limit it to just one component, so it chooses what to work on. We try to be as hands-off as possible and let it do its thing.
Do you have a single master prompt? Is it just “improve yourself”?
Even that master prompt has been improving.
Really? It improves the master prompt?
Yeah. And it’s not one RSI system. You can think of it as having one thing that does the improvement, but you also need something that knows what to improve. Those are both dynamic problems, and they’re very coupled.
How general do you think this ends up being? Right now, you’re focused on speeding up inference in a specific domain. Do you think the approaches you’re coming up with will generalize, or will there be a different one for each target solution space?
Mathematically, we’ve defined it so that it’s one system that generalizes to everything. I think the one part of the system—
And by “everything,” you mean any problem domain you can point AI at?
Correct. But there are physical limits, and one big limit is the human element. I think a lot of AI companies are thinking, “Let’s use AI to replace humans.” But we’ve found there’s so much that good researchers and engineers can bring to the table that will be very hard to replace with agents. From the very beginning, we’ve tried to think of our entire company as a multi-agent system, where humans are a kind of agent cohabitating with other AI agents.
Can we at least derive bounds? For example, do you envision a system where the amount of energy that goes into it somehow decreases over time relative to the improvement?
We can’t share too much, but one of the very important things we’ve tried to do in our company is work not just on the numerator, which is performance, but also on the denominator: how much you can get out of the same amount of physics, money, or whatever you want to call it.
Right, but do you think it’s possible to have a system where the numerator continues to increase at the same rate while the denominator stays fixed or decreases?
I don’t think it’s just possible. I think it’s happening.
Really?
Yeah.
So it takes less energy to get better results?
Yeah, over time.
Do you think there’s a universe in which that continues without asymptoting? Because if that were the case, you’d actually have takeoff.
I mean—
But it just seems to disobey information theory and the laws of physics.
Yeah. At some point, you’re going to reach the Planck constant, Planck length, Planck energy, Planck time. It will asymptote at some point, but I think—
Oh, I see. So you think that asymptote is basically the computation of the universe?
At least, yeah. How many particles are there? 10⁸⁰. We know the speed of light. We know what it is. You multiply those two and you basically get an upper bound on computation. Given this fixed amount of computation, we know we have a limit.
I’m so impressed by your belief in this idea of a model improving itself.
But it’s not a belief.
No, you’ve got to tackle that. That’s amazing. I mean—
It’s also a question of what metric you use to view progress. If you use benchmark accuracy, for example, it’s upward-bounded by 100%. You’re never going to see exponential improvement on that, because eventually you bottleneck at 100%. For a long time, we had this question: How do we even know if something has taken off if we don’t have the right metric? It’s not a belief that this is happening. I think we’re already seeing glimmers of it.
But it is a dramatic statement. I think it’s the first time I’ve heard you say it. Has this been a continuous result, or is it relatively recent that you’re starting to see glimmers of this?
We’ve been seeing these results grow and grow over the last few months.
That’s amazing. How do you have confidence in the results, given that it’s not working yet and you don’t know?
7. The Viceroy of Variance
No, it’s not a belief. It’s empirical measurement. But as a company, there’s a very funny thing happening: because we’ve been so thorough, we haven’t released things until we’re super sure they’re correct. Every few months, there’s an announcement of something RSI-related. We look at it and think, “We should have released our results or something.” Then we try to replicate it and realize it’s noise, or cherry-picking. It’s very hard to know, in this sea of AI—everybody’s trying to do AI and RSI stuff—what’s legitimate and what isn’t. The technical answer is that you have to be really good at measuring things. For example, we named one of our engineers the “Viceroy of Variance.” That’s a real thing: his job is to try to destroy our results and find what’s wrong. We even have Slack emojis for it, including one with red lasers coming out of his eyes when he finds bad variance in something.
One thing we’ve found is that, when you’re dealing with a new benchmark or a new problem of economic value that you want to apply this to, we haven’t found a replacement for what we call quality time: literally running hundreds of experiments, looking at them, developing an intuition, and figuring out what’s wrong, what isn’t, and what’s just variance. You have to work out what’s going on when something weird shows a kind of reward hacking two levels down: the agent that built the agent that built the agent did hack something, and now you have these weird results.
Right. It’s not easy. In nonlinear control systems, and in networking, there’s a lot of degenerate behavior you have to deal with. It’s not just convergence and divergence; you have things like oscillations.
And this is a very multidimensional problem. Do you have to deal with all of that, like potential oscillations?
Absolutely.
Meaning it’ll get good, then get bad, then get good?
Absolutely. And on top of that, the noise is fundamentally very stochastic. For a specific agent on a specific benchmark, there are levels of noise that are sometimes unpredictable. Some of it is extremely weird: you can get different behavior from one GPU and another GPU. They’re both B200s, but they have slightly different behaviors on the physics, and you can see some of that replicated at the high level. That’s why it has to be empirical.
From that, you can already see that intelligence isn’t a single scalar. It’s more like a multidimensional vector. There will be domains and fields where RSI gives you takeoff, and there will be dimensions where you might plateau or see relatively slow improvement. I think the defining factor is the slowest component in the process.
So you’re an optimization person. You know about Amdahl’s law: the speedup you can get by optimizing a component is—
You just take it to zero. Nothing else dominates.
Exactly. The digital part could be optimized and accelerated a thousand times or a million times. But if it’s only 20% of your overall time, you’re only going to get a 20% speedup. For domains where you still need to move atoms that interact with the physical world, that’s really slow, so you’re still going to be bottlenecked by it. In drug discovery, for example, if every evaluation requires wet-lab experiments and then clinical trials that take months, every new idea has to go through a months-long or multiple-month evaluation. Those fields probably won’t be accelerated as fast as coding or mathematics, where evaluations are very fast and cheap.
8. The garden hose company in Wisconsin
What I’ve yet to see is anybody put a bound on that—to say, in order for me to improve this by X or solve this problem, I need this much net-new energy to do it. If you had that, you could predict how big a warehouse I need to solve a drug discovery problem, or how many robots I would need. Do you think we’re getting any closer to having bounds? I think this is such a critical question for all of us to answer.
One of the internal jokes we have in the company is that the secret to RSI is somewhere in a garden hose company in Wisconsin. The joke here is that if Einstein hadn’t discovered relativity—
Yeah.
—he wouldn’t have these laws that tell you how speed and the contraction of time work.
Yeah.
If some unfortunate engineer had to design the GPS system, they would have had to find a correction for time. They probably would have said, “No, no, it’s clearly not the fabric of spacetime that’s contracting,” right? But I don’t know why; this is the formula I’ve devised empirically.
Some fudge.
9. Humans as agents in the system
Exactly right. And then people later would have been like, “Holy, this is spacetime contracting,” and we have this theory. RSI right now is being thought of as, “Let’s solve all these super-hard, fashionable problems.” But what we think is that the complexity of the world is such that the real challenges probably exist in spaces that aren’t being applied to. That’s the joke: there’s some workflow somewhere in a garden hose company in Wisconsin—
—if we solve it, we get, you know—
I love that. I just want to quickly summarize: You’re going after RSI. In your case, this is not autocatalytic effects; this is actual RSI, where the system improves the full system. For you, it’s an actual system that includes the harness and the weights.
If you think about the problem, you’ve got a numerator and a denominator. The numerator is how good the system gets, and the denominator is how much energy goes into it. You believe you’re seeing glimmers of a sign that both improve at the same time using this type of approach. Is that correct?
Yeah.
When can we expect to hear more from you?
Very soon.
Amazing. Thanks so much, both of you, for joining.
Yeah. Thank you.