Tokyo-based Sakana AI has formally established an RSI Lab, a dedicated research group for recursive self-improvement: AI systems that conduct the research to improve themselves.
The lab inherits an unusually concrete track record. Sakana's Darwin Godel Machine maintained an evolving lineage of agents rewriting their own code and more than doubled its baseline SWE-bench performance, a 30 percentage point absolute gain. ShinkaEvolve solved optimisation problems in 150 samples that brute-force search treats as intractable, ALE-Agent beat all 804 human entrants in an AtCoder heuristics contest, and the AI Scientist reached Nature in March.
The distinctive angle is economic. Where American labs frame self-improvement as a way to absorb ever more compute, Sakana argues Japan's constraints force the opposite bet: the lab is chasing the most sample-efficient self-improvement engine it can build, with advances that "compound on national, rather than hyperscale, compute budgets". Whether sovereign-scale RSI can keep pace with hyperscale RSI is now a testable question, and Sakana has volunteered to run the experiment.