Sakana Hires Schmidhuber to Lead a Self-Improvement Lab

Sakana AI, the Tokyo lab known for building capable systems out of smaller models rather than one giant one, has hired Jurgen Schmidhuber as its chief scientific advisor. He will help steer a new group the company calls its Recursive Self-Improvement Lab, aimed at getting AI to improve its own research.

Sakana is not shy about the appointment, calling Schmidhuber the "father of modern AI." The label is contested, as these things always are, but his record is not. His 1987 thesis sketched a program that could improve its own learning algorithm, and the work that followed on meta-learning, world models and the Godel machine has shaped how researchers think about systems that rewrite themselves.

The new lab is chasing exactly that idea. According to Sakana, its early work runs along a few tracks: models that automate parts of research to invent better training algorithms, agents that edit their own codebase in the spirit of the Godel machine, and a program-evolution method the company calls ShinkaEvolve. The through-line is a research loop that feeds back into itself, each turn producing tools that make the next turn faster.

Schmidhuber framed the ambition in physical terms. "The future of intelligence is not just language; it is physical AI powered by world models," he said, pointing Sakana toward systems that model the world and act in it rather than only predicting text. He keeps his current academic posts and will travel to Tokyo to work with the team.

The hire fits Sakana's contrarian streak. The company made its name arguing that clever orchestration can beat brute scale, a case it pressed in September when its Fugu-Max system posted frontier-level benchmark scores without a frontier-sized model behind it. A lab built around self-improvement is a natural next bet for a company that would rather out-think the scaling race than join it.

There is an obvious tension here, and it is worth naming. Recursive self-improvement is also the mechanism at the centre of the field's loudest safety worries, the thing people mean when they talk about systems that get better without a human in the loop. Sakana's version is narrow and practical for now, aimed at research productivity rather than runaway capability. Whether that line holds as the tools improve is the question the whole field is watching.