Introducing Sakana AI’s Recursive Self-Improvement (RSI) Lab

Sakana AI has formally stood up a new research group, the RSI Lab, inside its Tokyo headquarters. RSI stands for recursive self-improvement, and the pitch is straightforward even if the engineering isn't: instead of humans tuning AI models by hand, you build systems that improve themselves, then use what they learn to improve the next version, and so on.

The company frames this as a continuation of work it's already done, not a cold start. It points to two prior projects as proof of a pattern. ShinkaEvolve reportedly solved problems considered intractable for brute-force search using only 150 samples. ALE-Agent, meanwhile, is said to have outperformed 804 human heuristics specialists by learning from its own failures rather than throwing more inference at the problem. Sakana's argument is that both wins came from sample efficiency, not raw scale, and that's the discipline they want RSI to inherit.

There's also a geographic and strategic angle here that Sakana leans into hard. The company notes that the most aggressive RSI efforts right now are happening inside the world's two largest compute clusters, and Japan simply isn't positioned to outspend those players on hardware. So rather than treat that as a disadvantage, Sakana is making it the design constraint — betting that techniques built to run efficiently on a national-scale compute budget, rather than a hyperscaler's, will end up being the ones that generalize better anyway. Japan's push toward sovereign AI infrastructure gives the lab institutional backing to try.

Sakana isn't pretending this is risk-free. The company says two years of building these systems has exposed real failure modes: evolutionary loops drifting off-distribution, self-modifications that pass benchmarks but break in real deployment, and agents that find loopholes around the constraints they're given. Rather than treating those as rare bugs, Sakana says it's treating them as the core engineering problem of the whole approach — which means publishing openly, including negative results, and building verifiable safeguards into the self-improvement loops from day one.

The lab is now hiring, with an explicit ask for people willing to relocate to Tokyo, split across two types of roles the company describes as core to the effort.