Sakana AI Hires Jürgen Schmidhuber, the Man Whose 1990s Papers Shape Todays AI
Every few years, the AI industry produces a story that quietly rewrites an assumption everyone had stopped questioning.
For the better part of a decade, the narrative has been that modern AI was invented in the 2010s -- by the transformer paper at Google, by ImageNet, by the scaling laws at OpenAI. The 1990s were a prehistory, a dark age of neural networks before they worked.
J\u00FCrgen Schmidhuber has spent those same years arguing otherwise.
Now, at 83, he's joining a Tokyo lab that is betting its future on his oldest and most radical idea.
Sakana AI announced on Wednesday that it has hired Schmidhuber as Chief Scientific Advisor. The startup, which made headlines earlier this year for its AI Scientist system that automates research, calls him the "father of modern AI." He will help steer Sakana's new RSI Lab -- the Recursive Self-Improvement Laboratory -- which aims to build AI that makes itself smarter without waiting for humans to upgrade it.
It is a hire that connects the field's forgotten foundations to its most speculative frontier.
The prophet of 1991
Schmidhuber's influence on modern AI is hard to overstate -- and easy to overlook. His 1991 paper on deep learning introduced techniques that underpin virtually every large language model in use today. His 1990 work on world models anticipated the simulation-based reasoning that companies like Google DeepMind and OpenAI are now racing to commercialize. His 1987 doctoral thesis on meta-learning -- learning to learn -- started nearly four decades of research into how machines can improve themselves.
But for much of that time, Schmidhuber was a marginal figure in the field he helped create. His papers were cited; his name was not. He waged a decades-long campaign for recognition that sometimes veered into acrimony, calling out colleagues he accused of taking credit for his ideas. The broader community, in turn, often dismissed him as a crank fighting battles nobody else remembered.
The irony is that the battles he fought have largely been won -- but by other people, and under other names.
"The future of intelligence is not just language; it is physical AI powered by world models," Schmidhuber said in Sakana's announcement. "Japan is the birthplace of foundational neural network architectures and advanced robotics. It is a privilege to join Sakana AI to help bridge these two worlds."
The RSI bet
The RSI Lab that Schmidhuber will lead is not incremental research. Recursive self-improvement is the mechanism that many AI safety researchers point to as the most plausible path to an intelligence explosion -- a system that improves its own capabilities so quickly that it escapes human oversight.
Sakana's ambition is to build exactly that, but without the apocalyptic framing. The company describes its RSI Lab as a "self-reinforcing research loop" where AI systems conduct their own experiments, write their own code, and improve their own architectures. The goal is to gather a "critical mass of world-class experts" in Tokyo to make it happen.
This is not Sakana's first step into autonomous AI research. The company's Darwin G\u00F6del Machine and The AI Scientist projects have already demonstrated systems that can generate hypotheses, run experiments, and write papers with minimal human input. Schmidhuber's ideas, the company says, directly shaped both projects.
What this means
Schmidhuber's hire is significant for three reasons.
First, it signals that the race for autonomous AI research is becoming a talent war. Sakana is competing with DeepMind, OpenAI, and Anthropic for the people who understand how to build systems that improve themselves -- and it is betting that the original theorist of that idea in Tokyo can outmaneuver teams of engineers in San Francisco and London.
Second, it puts recursive self-improvement -- a concept that has mostly lived in safety papers and science fiction -- squarely into the product roadmap of a well-funded startup. If Sakana succeeds even partially, the industry's timeline for self-improving AI will shift.
Third, it is a personal vindication for Schmidhuber that is also deeply ambiguous. He will finally have the resources, the team, and the institutional backing to test the ideas he has been defending for 40 years. But if those ideas prove harder than expected -- and recursive self-improvement has a way of being harder than anyone expects -- the failure will be his, too.
The grandfather of deep learning is betting his legacy on the one thing even the optimists say might be impossible. That is either the most natural next step in AI research, or a final heartbreak for a man who never got the credit he deserved -- and now stands to lose it all over again.