Schmidhuber Joins Sakana AI to Guide Its RSI Lab

Jürgen Schmidhuber has spent nearly four decades publishing ideas that other people's companies later turned into products. This week, one of those companies brought him inside directly. Sakana AI announced on September 24 that Schmidhuber, the researcher it calls "the father of modern AI," is joining as Chief Scientific Advisor, guiding a new Tokyo research group called the RSI Lab. RSI stands for recursive self-improvement: AI systems that help build better versions of themselves, and then use those better versions to build the next ones after that.

Table of contents

• What Schmidhuber is actually taking on
• Why his name carries weight here
• What the RSI Lab has already shipped
• The physical-world bet
• A partnership built around geography
• What "credibility" doesn't prove
• Frequently asked questions

What Schmidhuber is actually taking on

Schmidhuber's role is an advisory one, not a full-time move. Superpower Daily's coverage confirms he keeps his existing positions and will travel to Tokyo regularly rather than relocate, guiding the lab's scientific direction while Sakana builds out its research team there. Nikkei Asia's report, dated the day of the announcement, frames it plainly: a well-known AI pioneer joining a startup as science adviser, without describing a change in his primary affiliations.

Sakana's own announcement quotes Schmidhuber connecting the move to Japan specifically, saying the country produced both foundational neural-network architectures and advanced robotics, and calling it a privilege to help bridge the two. That framing, bridging two research traditions rather than starting a new one, matches the advisory scope of the role: guidance and direction, not day-to-day management of the lab.

Why his name carries weight here

The Decoder's profile of the appointment traces Schmidhuber's influence back further than most researchers active in AI today. Sakana credits him with early work on world models around 1990 and with deep learning techniques from 1991 that underpin much of the current AI boom. The company also points to his contributions to ideas now standard across the field: unsupervised pretraining, distillation, residual learning and early forms of Transformer-like architectures.

None of that history is unique to this announcement. Schmidhuber has spent years publicly arguing that his lab's early papers anticipated techniques later popularized elsewhere, a claim that has made him a somewhat divisive figure in AI research circles even as his technical contributions are widely acknowledged. What is new is a company building its research program explicitly around him rather than simply citing his papers.

What the RSI Lab has already shipped

Sakana did not build the RSI Lab around Schmidhuber's arrival. The lab's own site describes it as a formal expansion of research Sakana has run for roughly two years, and lists concrete projects that predate this week's announcement. The Darwin Gödel Machine and The AI Scientist, described as a system capable of generating research ideas, running experiments, writing papers and executing peer reviews end to end, are both cited as work that shaped the lab's direction, and Schmidhuber's own ideas are credited as an influence on both.

Sakana's site also describes an adversarial sandbox applying recursive self-improvement to cybersecurity, where autonomous agents continuously co-evolve to find, exploit and patch vulnerabilities. That is a research environment, not a deployed security product, and Sakana's own framing treats it as a foundation for future work rather than a finished result.

The physical-world bet

KuCoin's coverage of the announcement highlights the specific technical direction Sakana is pointing the lab toward: agent-native world models that simulate physical outcomes before a robot or system acts, aimed at supply chains and robotic deployment rather than only text or image generation. Schmidhuber's own quote in Sakana's announcement makes the same point: "the future of intelligence is not just language; it is physical AI powered by world models."

That bet is not without competition. Several major labs are pursuing versions of world-model research aimed at robotics and physical reasoning, and Sakana's own materials do not claim a unique technical approach so much as a research philosophy, described on the lab's page as progress through ideas rather than through compute alone. Whether that philosophy produces results competitive with better-funded labs pursuing similar goals is an open question the appointment itself does not answer.

A partnership built around geography

Global Sources' coverage frames Schmidhuber's role as carrying a geographical message as much as a scientific one: positioning Japan as a place where foundational AI research and advanced robotics meet, rather than treating the two as separate industries. Sakana's own announcement leans into this directly, describing Japan as a country that "did not merely participate in the AI revolution" but "sparked it," and saying the company is assembling a "critical mass of world-class experts" in Tokyo.

That framing gives Sakana a recruiting pitch beyond salary and equity: a chance to work alongside a researcher of Schmidhuber's standing, in a country positioning itself as a serious AI research hub rather than primarily a market for AI products built elsewhere. Whether that pitch translates into the specific hires Sakana needs is something only time, not the announcement, will show.

What "credibility" doesn't prove

A different lab's numbers give some sense of how fast automation of AI research itself can move: our report on Anthropic's R&D automation index found Claude went from leading under 1% of internal research tasks to 26% in six months, though that measured automating existing research work rather than a system redesigning its own architecture the way recursive self-improvement aims to. The Decoder's own framing is worth sitting with: the appointment gives Sakana's RSI research "a heavyweight scientific voice," and that phrase is accurate as far as it goes. An advisory hire, however well-known the adviser, is not itself evidence that recursive self-improvement research is close to a breakthrough, and Sakana's public materials do not claim otherwise. The Darwin Gödel Machine and The AI Scientist are real, published projects, but they are also two years of prior work Schmidhuber is now formally connected to rather than results produced under his direct guidance.

What the hire does establish is intent: a mid-sized lab making a public, high-profile bet that recursive self-improvement and physical-world modeling, rather than simply scaling larger language models, is where it wants to compete. Whether that bet pays off is a multi-year question. This week's announcement is the starting gun, not the result.

Frequently asked questions