Sakana AI Hires Jürgen Schmidhuber to Advise New Self-Improvement Lab

Sakana AI has brought Jürgen Schmidhuber in as chief scientific adviser to a new Tokyo lab pursuing an ambitious loop: AI systems that help improve their own code and predict the physical consequences of their actions. The appointment adds guidance to the effort, but not a demonstrated self-improving system.

The research loop

The Recursive Self-Improvement Lab aims to create a cycle in which AI helps conduct research that improves machine intelligence. Sakana’s longer-term goal includes systems that research, discover and rewrite their own code. Improving one component would be a step toward that goal; the harder test is whether successive changes produce useful gains.

Schmidhuber’s connection to the idea goes back to his 1987 thesis on recursive self-improvement and meta-learning, or improving how a system learns. Sakana says his work influenced two of its existing projects, the Darwin Gödel Machine and The AI Scientist. Those projects give the new lab a research lineage, rather than evidence that its proposed cycle already works.

The future of intelligence is not just language; it is physical AI powered by World Models.

Jürgen Schmidhuber, in Sakana AI’s announcement

Rehearsing an action before taking it

The lab’s physical-world goal is easier to picture. Sakana says it is developing its first agent-native world models: systems intended to simulate what might happen after an action, before that action is taken. An AI agent could use such a model to consider a move’s consequences rather than work only with text or code.

Sakana proposes using the models to explore supply-chain choices and robotic deployments without relying entirely on costly real-world trial and error. That would make predictive accuracy crucial: a rehearsal helps only if its results remain useful when a machine acts. The announcement offers no manufacturing or robotics test of the models’ predictions.

An adviser, not a new system

Schmidhuber will retain his current positions and visit Tokyo regularly to work with Sakana’s team. He will help guide the lab’s scientific direction; Sakana is also assembling researchers in the city. The company frames the appointment as part of its effort to attract international AI talent to Japan.

His role connects two research ambitions, each with a different test. Code-rewriting agents must produce improvements that matter beyond the act of changing code. World models must make predictions useful beyond a simulation. Sakana has described both as goals for its new lab; the appointment itself supplies expertise to pursue them, not results from either effort.