Recursive Self-Improvement and the Future of AI

Summary · 4 min read instead of 32 min

Dhaval Adjodah trained as a physicist and did a PhD in reinforcement learning at the MIT Media Lab. Owen He studied math, did a PhD on continual and meta learning, and was a research scientist at DeepMind. Together with Grace, they cofounded MakerMaker.AI, which is building recursively self-improving AI systems.

Adjodah and He separate true recursive self-improvement (RSI), where the AI being improved is also the improver, from autocatalysis, where humans use better AI tools. They frame progress as capability gains divided by the resources spent, and argue that real RSI must improve both sides rather than simply absorb more capital. They say a closed loop does not guarantee takeoff, so the outcome has to be measured, and measurement is hard because of noise, variance and reward hacking. They apply their system to inference efficiency and claim results over recent months show better performance from fewer resources. Physical bottlenecks such as wet-lab experiments will slow RSI in some domains.

Minute by minute

1. 1:43

   Most RSI pitches are autocatalysis. Martin Casado says many pitches call AI-built GPU kernels RSI. Owen He says the key question is who the optimizer is. When humans use AI tools to build better AI, the ideas still come from human brains, which he calls autocatalysis. In true RSI, the model is both the optimizer and the thing being optimized, so cognitive improvement gets reinvested into itself.

2. 4:08

   Closed loops change the improvement math. The founders argue from control theory that autocatalysis leaves a gap that humans must fill, while a closed loop can keep marginal gains constant or growing. Adjodah frames the rate of capability improvement as a ratio of a performance term and a cost term. If the improving term rises with capability, improvement stays constant. If it rises with a higher power of capability, the result is exponential takeoff.

3. 8:31

   Humans are agents in the system. He says people wrongly assume RSI must be fully automatic and digital. Their mathematical framework is a multi-agent system in which human researchers and engineers are a special case of agents. He argues every component should be optimized, because anything left unoptimized later becomes the bottleneck.

4. 9:12

   Count the denominator, including human hours. Casado observes that model release times are shrinking, but not once divided by the capital raised. The founders say gains that come from pouring in more resources are not RSI. They track human hours, cost and time, and keep a "bleederboard" that ranks staff by experiment spend. He cites the nanoGPT benchmark, where retraining GPT-2 on the same GPU fell from weeks on multiple GPUs to hours, at about a hundred dollars.

5. 13:26

   Most RSI papers fail at scale. The founders say many published RSI ideas don't work when tested on their internal infrastructure, because they are complex, human-intuited designs. The simpler approach that closes the loop is what scales. Adjodah describes starting from a unified mathematical framework, such as finding 12 abstractions behind 200 systems. He cites Richard Feynman to call research without such a framework "stamp collections".

6. 16:32

   Optimize every level of the stack. He says harnesses and weights affect each other, so improving them in isolation misleads. The founders describe RSI across prompts, harness, weights and even chips. They started with the harness for efficiency, moved to weight updates for more gains, and expect co-optimization of many parts to be needed.

7. 17:28

   Inference is the chosen proving ground. Adjodah says they pick problems as close to AI as possible, and inference can improve both numerator and denominator at once. It is also economically important, since everyone wants to serve open-weight models such as Kimi K3 efficiently. Their system uses many LLMs, lets the agent choose what to optimize and which model to use, and even its master prompt has been improved by the system.

8. 20:58

   General in theory, limited by humans. The founders say their system is defined mathematically to generalize to any domain, but there are physical limits, the human element being a big one. Instead of replacing people, they treat the whole company as a multi-agent system where humans cohabit with AI agents, since good researchers and engineers bring things agents can't easily replace.

9. 22:12

   Less energy for better results. Asked whether a system could keep improving with fixed or falling energy, a founder says "It's not just possible. I think it's happening." They accept it must eventually asymptote at physical limits, and estimate an upper bound on computation from about 10 to the 80 particles and the speed of light. They say results have grown over the last few months and rest on empirical measurement.

10. 24:50

    Proving gains requires adversarial measurement. Adjodah says they hold results until certain, and that many announced RSI results turn out to be noise or cherry picking when they try to replicate them. An engineer they call the "Viceroy of Variance" tries to destroy internal results. They rely on quality time with hundreds of experiments to catch variance and reward hacking, including hacks by a builder agent two levels down. Even two B200 GPUs can behave slightly differently.

11. 27:32

    Slow physical steps cap acceleration. He argues intelligence is a multidimensional vector, so RSI may produce takeoff in some domains and plateaus in others. Citing Amdahl's law, he says accelerating a digital part that is only 20% of the time yields limited speedup. Drug discovery, with wet labs and clinical trials taking months, will move slower than coding or math, which have fast, cheap evaluations.

12. 28:56

    The secret may sit in unglamorous work. Casado asks whether anyone can predict the energy or resources needed to solve a problem. Adjodah shares the company joke that the secret to RSI lies in a garden hose company in Wisconsin. By analogy to an engineer finding a GPS time correction without relativity, he argues the real challenges likely sit in overlooked workflows rather than fashionable problems.

Published Oct 12, 2026.