Sakana AI brings Jürgen Schmidhuber into its self-improvement push

The ideas Sakana is buying

Sakana AI says Schmidhuber’s work influenced a significant share of modern AI development. His 1987 dissertation examined meta-learning and recursive self-improvement: approaches in which an AI system continually improves itself.

The company links that research directly to two of its projects:

Schmidhuber also argues that intelligence will not be defined by language alone. Physical AI, in his view, should operate through models of the world and learn by interacting with the physical universe.

He said Japan was the birthplace of foundational neural-network architectures and advanced robotics, and that Sakana AI could help restore the country’s role as a center of innovation. The intended destination is a class of autonomous systems that can understand, model and study the physical universe through interaction with it.

The RSI laboratory

Schmidhuber will help develop Sakana AI’s new RSI laboratory. Its goal is to create a self-reinforcing research cycle that allows machines to become more intelligent.

The company plans to build a “critical mass of world-class experts” in Tokyo to support that effort.

That ambition is the more consequential part of the appointment. Sakana AI is not presenting Schmidhuber merely as a prestigious name attached to an existing research program; it is placing his ideas about recursive improvement near the center of a new laboratory.

My read is that the appointment gives Sakana AI a coherent intellectual story: deep learning’s past, automated science in the present and physical-world intelligence as the next target. But the announcement is quiet about the mechanism. It does not say how the RSI laboratory will measure improvement, control recursive changes or turn a research cycle into a reliable system.

That gap matters. The company has identified the direction and the people it wants in Tokyo; the difficult question is whether self-improvement produces better research faster, or simply more complicated systems that are harder to evaluate.