Sakana AI Forms RSI Lab To Chase Self-Improving AI From Tokyo

Why it matters

Sakana AI is betting that the next AI edge may come from automating the research loop, not just buying more compute. If that works, smaller national labs and startups could compete in frontier AI with different economics.

Sakana AI, the Tokyo company building what it calls frontier AI in Japan, announced a new Recursive Self-Improvement Lab in a blog post aimed at using AI to redesign parts of the AI research process itself.

The new group, called the Sakana AI RSI Lab, is not a separate company, funding round, or commercial product. Sakana describes it as a dedicated research team inside the company focused on systems that can write, benchmark, verify, and improve components of their own architectures. The company says it is "one of the earliest labs" working on recursive self-improvement with modern foundation models, a positioning claim that is not independently substantiated in the materials RuntimeWire reviewed.

Sakana's official account @sakanaai has framed the company around a distinctly Tokyo constraint: competing in AI without simply matching the largest labs on raw compute. The new lab is the clearest expression of that bet. In the announcement, Sakana argues Japan should not rely on "brute-forcing monolithic models" and should instead pursue adaptive, sample-efficient systems that can compound progress through iteration.

The strategy under the lab

Sakana is trying to turn a resource disadvantage into a research identity. The blog post explicitly links the RSI Lab to Japan's manufacturing history, arguing that the next AI advantage may come from redesigning the process, not just scaling the factory.

That is the strategic point of the announcement. Sakana is not saying it has built an autonomous AI researcher that can improve itself without limits. It is saying its research agenda is moving from static, human-led workflows toward loops in which models propose, test, and refine parts of the system. The difference matters: recursive self-improvement is a powerful phrase, but the evidence Sakana offers is a portfolio of bounded research systems, each operating in specific domains and benchmarks.

The company does not disclose the RSI Lab's headcount, leadership, budget, compute allocation, customers, or commercial timeline. It also refers to frontier RSI being pursued inside the world's two largest compute clusters without naming the operators. That leaves the announcement as a research marker rather than a measurable business milestone.

What Sakana says it has already built

Sakana grounds the RSI Lab in six projects it says it has shipped over the last two years.

The earliest is LLM-Squared, a 2024 project Sakana says it developed with Oxford and Cambridge to let large language models invent better methods for training large language models. Sakana says that work produced DiscoPOP, a preference optimization algorithm written by an LLM through an evolutionary loop.

Sakana also points to The Darwin Godel Machine, developed with University of British Columbia researchers, which the company says maintained an evolving lineage of software agents that rewrote their own codebase. According to Sakana, DGM more than doubled a baseline software-engineering score on SWE-bench and produced a 30 percentage point absolute improvement.

In ShinkaEvolve, Sakana says it used adaptive sampling and novelty filtering to solve optimization problems with 150 samples. With ALE-Agent, Sakana says an optimization agent placed 1st out of 804 human participants in AtCoder Heuristic Contest 058. In Digital Red Queen, a collaboration Sakana says involved MIT, models wrote competing programs in a Core War sandbox to study adversarial coevolution.

The broadest project is The AI Scientist, which Sakana says can generate research ideas, run experiments, write papers, and perform peer review. Sakana says that work culminated in a Nature publication dated March 26, 2026.

A research brand, not yet a business model

The through-line is clear: Sakana wants to be judged on sample efficiency and autonomous iteration rather than scale alone. That is a founder-level thesis even though the announcement does not name an individual founder or lab head: build frontier AI in Japan by making the research loop itself more automated.

The open question is how much of this can move from benchmarked research systems into durable products, revenue, or defensible infrastructure. Sakana's post lays out a four-phase arc from agent-native models to AI scientists, recursive self-improvement, and democratized AI. For now, that roadmap remains Sakana's internal theory of progress, backed by a set of research outputs the company says point in the same direction.

For operators and investors, the signal is not that Sakana has solved self-improving AI. It is that one of Japan's most visible AI labs is choosing a lane: fewer claims about ever-larger monoliths, more emphasis on systems that can search, evaluate, and improve under constraint.