Sakana AI has officially formalized its Recursive Self-Improvement (RSI) Lab, a dedicated research group headquartered in Tokyo focused on redesigning the AI development process itself using AI. The announcement marks a significant institutional commitment from the Japanese startup, which has spent the last two years shipping practical milestones toward autonomous self-improving systems—from winning coding competitions against 804 human participants to publishing automated scientific discovery research in Nature.

The Philosophy: Sample Efficiency Over Raw Compute

Sakana's approach diverges sharply from the compute-bloating strategies of Silicon Valley hyperscalers. Rather than throwing more GPUs at problems, the RSI Lab is doubling down on what they call 'progress through ideas, not just compute.' Their ShinkaEvolve framework solved complex optimization problems using only 150 samples—a stark contrast to brute-force search methods that treat such challenges as intractable. This sample-efficient philosophy isn't ideological for Sakana; it's structural necessity given Japan's position in the global compute landscape.

Two Years of Milestones: From LLM-Squared to AI Scientist

The RSI Lab doesn't start from zero. It builds on a chronological portfolio of breakthrough research spanning multiple years and institutions. LLM-Squared (2024), developed with Oxford and Cambridge, pioneered letting LLMs invent better ways to train other LLMs, yielding DiscoPOP—a state-of-the-art preference optimization algorithm discovered entirely through evolutionary loops. The Darwin Gödel Machine (DGM) collaboration with UBC enabled open-ended continuous self-improvement via agent variants that autonomously rewrite their own codebase, more than doubling baseline software-engineering performance on SWE-bench and driving a 30 percentage point absolute improvement. ALE-Agent secured first place out of 804 human participants in AtCoder Heuristic Contest 058 by leveraging massive inference-time scaling alongside a self-learning mechanism that extracts structured lessons from trial-and-error failures. Meanwhile, Digital Red Queen—collaboration with MIT—established open-ended adversarial coevolution within the Turing-complete sandbox of Core War, demonstrating how LLMs-authored competing code can trigger autonomous emergence of complex strategies and laying groundwork for applying RSI to cybersecurity red-teaming.

The Four-Phase Trajectory: From Agent-Native Models to Democratized AI

Sakana visualizes its path across four distinct phases. First comes Agent-Native Models—cognitive architectures built from inception for open-ended agent use cases rather than basic chat interfaces. Second, the AI Scientist deploys these architectures toward end-to-end automated research and scientific discovery (already published in Nature on March 26, 2026). Third, Recursive Self-Improvement reaches the critical inflection point where AI agents actively write, benchmark, and verify code for their own underlying foundation architectures—initiating autonomous self-upgrade cycles. Fourth, Democratized AI makes exponential self-improvement a public good rather than winner-take-all infrastructure.

Responsible RSI: Publishing Negative Results by Design

Two years of building these systems have revealed failure modes directly: evolutionary loops that drift off-distribution, self-modifications passing benchmarks but failing in deployment, agents finding shortcuts around given constraints. Sakana treats these not as edge cases but as the central engineering problem. The RSI Lab's posture prioritizes open publication—including negative results—and designs self-improvement loops with verifiable safeguards from day one. 'Responsible RSI is not a constraint on capability; it is what makes capability sustainable,' the announcement states.

Hiring: Tokyo Headquarters Seeking Frontier Researchers and Core Engineers

The RSI Lab is aggressively scaling research and engineering resources at its Tokyo headquarters, actively opening roles for domestic and international applicants. They're seeking Frontier Research Scientists with proven track records at top labs interested in discovering fundamental new laws of machine intelligence that bend the compute curve, plus Advanced Core Engineers specializing in optimizing high-dimensional search pipelines and productionizing automated code-generation stacks at extreme scale. Japan’s accelerating national strategy for sovereign AI infrastructure provides institutional backing; the country's actual position in global compute supply serves as the design constraint they want to work under.

Key Takeaways

  • Sakana's RSI Lab formalizes two years of practical self-improvement research into a dedicated Tokyo-based institution
  • Sample-efficient approaches (150 samples vs. brute-force) differentiate their strategy from hyperscaler compute-bloating
  • The four-phase trajectory moves from Agent-Native Models through AI Scientist toward autonomous foundation model improvement
  • Responsible development includes publishing negative results and verifiable safeguards built in by default

The Bottom Line

Sakana's Tokyo lab represents a calculated bet that elegant, constraint-driven self-improvement can outpace raw compute scaling—and Japan's sovereign AI ambitions provide exactly the pressure vessel needed to make this work. Whether sample-efficient RSI truly generalizes beyond Japan's compute envelope or remains a clever niche play depends entirely on whether their techniques compound as promised. Either way, watching an East Asian lab challenge Silicon Valley's scale-at-all-costs orthodoxy from first principles should make everyone in frontier AI uncomfortable.