Sakana AI opens Tokyo RSI lab, bets on compute-light path

A Tokyo lab betting that the next real gains in AI won't come from bigger training runs is a story worth paying attention to, especially when the lab has receipts. Sakana AI announced a dedicated Recursive Self-Improvement Lab in Tokyo, framing it as an attempt to redesign the AI development process itself with AI, rather than out-scaling the frontier labs on parameters and clusters.

The lab is not standing on a whiteboard. Sakana's writeup points to a stack of work already published: the Darwin Gödel Machine, which the company says more than doubled its baseline software-engineering performance on SWE-bench, driving a 30 percentage point absolute improvement by letting agents rewrite their own codebases; ShinkaEvolve, an open-source program-evolution framework that only needed 150 samples on problems the writeup treats as intractable for brute-force search; ALE-Agent, which secured 1st place out of 804 human participants in the AtCoder Heuristic Contest 058; and The AI Scientist, an end-to-end automated research system whose work culminated in a publication in Nature on March 26, 2026.

The pitch to a working engineer is that these gains came from technique rather than compute. Sakana's positioning line, from the English announcement, is progress through ideas, not just compute, and the writeup explicitly reads Japan's actual position in the global compute landscape as a design constraint rather than a disadvantage, arguing that the country's sovereign AI infrastructure strategy provides institutional support. According to Stack Futures, which places the announcement on June 5, 2026, the Tokyo-based lab was founded by ex-Google Brain researchers including David Ha, and RSI Lab is now the dedicated home for that agenda.

The honest caveats are the ones you'd expect. Benchmark gains, especially on SWE-bench and heuristic coding contests, have been the AI subfield most prone to looking transformative in a paper and modest in production, so take the specifics as reported, not settled. What the reporting doesn't give you is a headcount for the new lab, a funding figure, or a clear answer on whether Sakana intends to keep training its own frontier-scale base models alongside the RSI work.

What's worth watching is whether smaller teams and academic collaborators pick up the open-sourced pieces of this stack. If ShinkaEvolve-style optimization and Darwin Gödel Machine-style self-editing scaffolds prove useful outside a benchmark harness, the RSI Lab's real contribution won't be a leaderboard number, it will be a working recipe for teams that were never going to compete on cluster size.