The State of Recursive Self-Improvement in AI Research
Real recursive self-improvement means the thing being optimized is also the optimizer, with gains measured against cost, and it must be shown empirically, across the whole stack, and within physical limits.
Watch on YouTube ↗Charlie's summary
- ▶ 1:58
Real RSI means the optimizer and the thing optimized are the same; humans using AI tools is just autocatalysis.
The guests argue RSI is an overloaded term. Using computers to build better computers, or AI to write GPU kernels, is autocatalysis: the ideas still come from human brains. In their view, RSI starts when the agent's own model can be the thing being optimized, which reinvests “cognitive improvement into itself”. Ask of any RSI claim: who is the optimizer, and who is the optimizee?
- ▶ 4:30
A closed loop changes the math: capability improvement can go from plateau to constant to exponential.
The speaker frames improvement in capability as performance divided by cost or difficulty. With a fixed numerator, as with steam engines, you plateau. If the improver's strength rises with capability, the terms cancel and you get constant improvement. If it scales with a higher power of capability, you get exponential takeoff. The exponent can't be solved analytically; it has to be measured empirically, since there is no learning theory of deep learning.
- ▶ 9:37
Count the denominator: improvement bought with more money, compute, or human effort is not RSI.
Faster model release cycles look like acceleration, but divided by capital raised they may not be. The guests say they track cost, time and human experiment effort, even running a bleederboard of experiment spend. They point to the nanoGPT benchmark: GPT-2 training fell from weeks on many GPUs to hours on one, for a few hundred dollars, on the same hardware, thanks to algorithmic gains, many from auto-research. Their test is whether numerator and denominator both improve.
- ▶ 13:14
Start from a unified mathematical framework, because most RSI papers are stamp collections that fail at scale.
When the guests tested papers claiming RSI on their internal infrastructure, many didn't work. They were clever, human-inspired ideas that didn't scale, and the simpler thing that closes the loop is what scales. They begin with a math framework that maps the whole space, hoping to find about 12 abstractions behind 200 apparent systems. Citing Feynman's split between stamp collecting and physics, they say that without unification you can't tell whether the overall system is improving.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.