Sakana AI opens Tokyo RSI Lab, hiring frontier researchers

Sakana AI has posted its Member of Technical Staff (RSI Lab) role in Tokyo, and the interesting thing about the listing is how openly it stakes out a position against the prevailing playbook. The lab, named for Recursive Self-Improvement, tells applicants it wants people "entirely unsatisfied with the status quo" who are ready to "break away from standard benchmarking and brute-force scaling." That reads as a hiring pitch as much as a research one, aimed at whoever inside the big US labs has been quietly muttering the same thing.

The actual work described sits across four tracks: discovering "fundamental new laws of machine intelligence that bend the scaling curve," building world models that act as "verifiable simulators for agentic reasoning and planning," applying "open-ended evolutionary dynamics to high-stakes algorithmic domains," and building the systems and infrastructure backbone of the RSI loop itself. The associated RSI Lab overview frames this as a phased trajectory from agent-native models through an "AI Scientist" to recursive self-improvement, and points to prior Sakana projects (the Darwin Gödel Machine on SWE-bench, ALE-Agent taking first place at AtCoder Heuristic Contest 058, ShinkaEvolve solving optimization problems with 150 samples) as receipts.

Two tracked researchers in our directory shared the posting shortly after it went up, which fits the pattern for Sakana openings: they circulate among people who left larger labs looking for a smaller, more contrarian bet.

Read the fine print, though. The listing itself doesn't disclose salary, headcount, or who leads the lab day to day, and the RSI framing is famously easier to announce than to deliver. It also requires physical relocation to Tokyo with English-only application materials, at a moment when the US frontier labs are paying more to keep the same shortlist.

If Sakana can convert this hiring round into two or three concrete papers of the "small model, big result" flavor that made its earlier work land, the geographic bet on Japan and on world models becomes genuinely interesting. Right now it is a strongly opinionated job page attached to a lab that has yet to publish under its new banner.