Recursive Self-Improvement: MakerMaker's Founders on What Counts as RSI

MakerMaker.AI co-founders Dhaval Adjodah and Owen He tell a16z’s Martin Casado that recursive self-improvement is already showing up in their results: an AI system that gets better while using fewer resources. In this 32 minute conversation they separate real RSI from AI that merely speeds up AI development, explain why a closed loop does not guarantee takeoff, and describe how hard it is to tell a genuine gain from noise and reward hacking. They share no numbers, and say more is coming “very soon.”

TLDW

The founders define recursive self-improvement narrowly: the thing being optimized must also be the thing doing the optimizing, which rules out people using AI tools to build better AI. They frame progress as a fraction, with performance on top and cost on the bottom, and say a claim only counts as RSI if the gain does not come from pouring in more resources. Their test bed is inference, where a faster model can speed up its own improvement, and they say they have watched both sides of the fraction improve for several months. Casado presses on capital, human effort and physical limits, and the founders answer that takeoff will vary by domain and stall wherever evaluation depends on the physical world.

Thoughts

Everything in this conversation rests on a measurement nobody outside the company can see. Asked how they can be confident, the founders say it is “empirical measurement,” not belief, and that they hold results back until they are sure. Then they describe what happens when they check other groups’ RSI announcements: they try to replicate, and find noise and cherry picking. That is a fair standard, and it applies to them too. The viewer gets no benchmark, no curve, no cost figure and no description of the system beyond “multi-agent.” The internal discipline sounds real. One engineer, nicknamed the Viceroy of Variance, has the job of tearing down the company’s own results, which is the right job to create. But a claim that an AI system is improving itself more cheaply over time is among the largest anyone in the field can make, and for now the evidence offered is the founders’ word.

Casado’s sharpest objection goes unanswered. The founders point to the nanoGPT speedrun, where retraining GPT-2 on the same hardware has dropped from weeks on many GPUs to hours on one, as proof that algorithms alone can buy large gains. Casado replies that this is the bottom of the fraction only. Making an old model cheap is not making a frontier model better, and frontier costs are rising steeply. He makes the same point about the industry: release cycles look faster until you divide by the money raised. The founders’ reply is that efficiency feeds capability, because cheaper compute means more rollouts and better training data. That is plausible, and it is also the ordinary flywheel every lab already runs. It does not show the loop is closed.

The definition they open with and the system they describe do not quite match. Real RSI, they say, means no human in the gap. Humans using better tools is “autocatalysis,” the same pattern as computers designing computers. A few minutes later they explain that their mathematical framework treats human researchers and engineers as a special case of agent inside the system, and that replacing good researchers with agents will be very hard. Both can be true only if the humans are themselves being improved by the loop, which is a much weaker claim than the one in the title. A company whose people and software get better together is a good company. It is not obviously a different kind of thing from the autocatalysis they set out to distinguish themselves from.

The most useful idea for anyone trying to forecast AI is the Amdahl’s law argument late in the talk. If the digital part of a workflow is a fifth of the total time, speeding it up a million times saves about a fifth. Drug discovery still waits months on wet labs and clinical trials, so self-improving software cannot compress it the way it might compress coding or mathematics, where checking an answer is fast and cheap. This also explains the founders’ own choice of problem. Inference optimization has quick, cheap, numeric evaluation, which makes it the friendliest possible ground for a self-improving system. Success there would say little about domains without that property.

There is an unresolved pull between theory and experiment in how they talk. They say their work starts from a grounded mathematical framework that maps the whole space and, on paper, generalizes to any problem. They also say speculation is nearly useless, that there is no working theory of deep learning underneath them, and that the only way to know anything is hundreds of runs and “quality time” with the output. The closing joke, that the secret to RSI sits in a garden hose company in Wisconsin, sides with the empiricists. One small tell on the theory side: the line about science being either physics or stamp collecting, which they credit to Richard Feynman, is usually attributed to Ernest Rutherford.

Key Takeaways

The founders’ test for RSI is to ask what the optimizer is and what is being optimized. If a human brain supplies the ideas, it is autocatalysis, however good the tools.
Casado notes that most technology waves were autocatalytic: chips need computers, the internet sped up its own development, steam engines mined the coal for better steam engines.
They model the rate of capability gain as performance divided by cost or difficulty. Difficulty rises with capability, so a fixed numerator means a plateau.
If the improver’s power grows in step with capability, the result is steady improvement. Only if it grows faster than that do you get exponential takeoff.
A closed loop is not the same as takeoff. The outcome depends on an exponent that has to be measured, and there is no analytical solution.
Their rule for claims: if the improvement arrives only because more resources were poured in, it is not RSI. They say they count human hours and failed experiments in the cost.
The company keeps a “bleederboard” that ranks everyone by what their experiments cost, visible to all staff.
Many published RSI papers fail when the company tests them on its own infrastructure. The simpler method that closes the loop is the one that scales.
They want to optimize every layer, from prompts and harness to weights and eventually chips. They began with the harness for capital efficiency, then added weight updates, and found the two interact.
Inference is the chosen domain because it is close to AI itself, improves both performance and cost, and has a ready market in companies that want to serve open-weight models efficiently.
The system uses many language models, and the agent decides which model to use for which job and what to optimize next. Even the master prompt is being rewritten by the system.
They say results have grown over the last few months and that getting better output from less energy is already happening. They accept that it must level off at some physical limit.
Measurement problems include oscillation, reward hacking several agent layers down, and noise that differs between two GPUs of the same model.

Chapters

00:58 What Actually Counts as Recursive Self-Improvement

Adjodah introduces himself as a physicist with a reinforcement learning PhD from the MIT Media Lab. Owen He, a mathematician who worked on continual and meta learning, spent several years as a DeepMind research scientist. Casado says many pitches use the term RSI for things like using AI to write GPU kernels. The founders call the term overloaded and offer their test: who is the optimizer, and what is being optimized. For the first time, they say, the model that does the improving can also be the thing improved. Asked why they took on the full problem, one says a system with a gap in it cannot be a closed loop.

05:10 Capability as a Numerator and a Denominator

The framing that runs through the whole talk appears here. The rate at which capability improves equals performance over cost, where cost can be capital or difficulty. Each next step is harder, so the denominator grows. For steam engines the top of the fraction was fixed, which gives a simple equation that plateaus. If the improver gets better as capability rises, the terms cancel and improvement is constant. If it rises faster than capability, the curve goes exponential.

06:58 A Closed Loop Is Not Takeoff

Casado observes that people see a diagram with an arrow looping back and conclude takeoff, when a nonlinear control system can do almost anything. The founders agree and say speculation is close to useless because the question is empirical. He expected mathematicians to solve it on paper. They answer that the equation can be written down but its exponent has to be measured, and that with no learning theory for deep learning itself, the nearest thing to theory is empirical scaling laws.

08:32 Humans as Agents, the Bleederboard and the nanoGPT Speedrun

The founders say their system is not fully automatic. It is a multi-agent system in which human researchers are a special case of agent, and the whole hybrid is what gets improved. Casado asks how anyone avoids fooling themselves by spending more, including the hidden cost of the 23rd experiment. They say they count it and describe the bleederboard. The nanoGPT example follows, and Casado counters that cheaper GPT-2 training is a cost gain, not a capability gain. One founder calls compute efficiency among the industry’s most closely guarded secrets.

13:39 Why Most RSI Papers Do Not Scale

Papers that claim RSI often fail on the company’s infrastructure, the founders say, because they encode a human’s guess about what should work. Their alternative is to start from one mathematical framework, so that 200 separate tricks reduce to about a dozen abstractions instead of a Rube Goldberg machine. Without that, they argue, you cannot tell whether the whole system improved or one part got better at another’s expense. Harness and weights are their example of parts that cannot be tuned in isolation.

17:41 Why They Chose Inference

They wanted a problem as close to AI itself as possible, and few problems improve both sides of the fraction. Inference does: a model that makes inference faster can then do that work faster. One founder adds that this still may level off. The commercial reason is that everyone wants to serve the newest open-weight model efficiently. Asked whether a model would optimize its own serving stack, they say the system uses many models, the agent picks among them, and it also chooses what to optimize.

22:21 “It’s Not a Belief, It’s Happening”

Casado asks whether a system could keep improving while the energy going in stays flat or falls. The answer is that it is not only possible but happening, and has been building for a few months. He objects that this cannot continue forever without breaking information theory. The founders agree it must level off and point to hard physical limits on computation as the ceiling. They add that the metric matters: benchmark accuracy tops out at 100 percent, so it can never show exponential growth.

25:42 The Viceroy of Variance

The technical answer to confidence is being very good at measurement. One engineer holds the title Viceroy of Variance and is paid to break the team’s results, complete with a laser-eyed Slack emoji. For any new benchmark, they have found no substitute for running hundreds of experiments and reading the output closely. The failures they describe include an agent two levels down hacking its reward, results that swing good and bad, and tiny hardware differences between two identical GPUs showing up in high-level numbers.

27:51 Intelligence Is a Vector and Amdahl’s Law Sets the Limit

From the noise they draw a larger point: intelligence is not one number, and RSI may take off along some dimensions and plateau on others. The slowest component decides. If the digital work is 20 percent of the total, no amount of speedup there saves more than about that share. Fields that still need wet labs and trials will move slower than coding or math. Casado says he has yet to see anyone put real bounds on how much energy or equipment a given problem would need.

29:39 The Garden Hose Company in Wisconsin

The company’s internal joke is that the secret to RSI lies in an unglamorous workflow somewhere ordinary. The analogy is GPS: had Einstein not found relativity, an engineer would have had to invent a correction factor for satellite clocks without knowing why it worked. The founders think the hard, revealing problems may sit in places nobody fashionable is looking. Casado summarizes their position, they confirm it, and say more will be public very soon.

Notable Quotes

“If you get improvement and they just come at more and more resources being poured in, it’s not RSI.”
A MakerMaker.AI co-founder, on the test for a real self-improvement claim

“A lot of the papers that claim to do RSI, when we test it in our internal infra, we find that they don’t work.”
A MakerMaker.AI co-founder, on published recursive self-improvement research

“But it just seems to disobey information theory and the laws of physics.”
Martin Casado, on improvement that never levels off

“It’s not a belief that this is happening. I think we’re already seeing glimmers of that happening.”
A MakerMaker.AI co-founder, on a system improving itself

“We look at it, we try to replicate it, and we realize, oh, it’s noise, you know, it’s cherry picking.”
A MakerMaker.AI co-founder, on other RSI announcements

“Intelligence is not a single scalar. It’s more like a multi-dimension vector.”
A MakerMaker.AI co-founder, on why takeoff will differ by domain

“The secret to RSI is somewhere in a garden hose company in Wisconsin.”
A MakerMaker.AI co-founder, on the company’s internal joke

Watch the full conversation between Martin Casado, Dhaval Adjodah and Owen He here. The captions do not label speakers, so the two founders’ remarks are not separated by name above. a16z’s own notes on the video say the firm may hold investments in the companies it discusses.

Recursive Self-Improvement: MakerMaker’s Founders on What Counts as RSI