Sakana AI bets AI that improves itself can break the compute arms race of frontier labs

Sakana AI has launched a research lab focused on recursive self-improvement (RSI), the idea that AI systems can iteratively redesign and improve themselves, creating a compounding cycle of progress. The startup sees RSI as a potential way beyond the compute arms race.

Japanese AI startup Sakana AI has created a new research group exploring how AI can speed up and improve the development of new AI systems. The Sakana AI RSI Lab builds on the company's earlier work. Since its founding in 2023, Sakana has focused on evolutionary, adaptive AI systems and, more recently, on practical steps toward recursive self-improvement.

In its announcement, Sakana points to several research milestones from the past two years. These include LLM-Squared, where language models design better training methods for other language models, and the Darwin Gödel Machine, which generates, tests, and iterates on variants of its own codebase.

With ShinkaEvolve and ALE-Agent, Sakana also highlights work on evolutionary program optimization and agents that derive new strategies from trial-and-error loops. Another key project is The AI Scientist, a system for automating parts of scientific research. A later version wrote a paper that passed peer review, according to Sakana. The underlying research was published in Nature in March 2026.

Taken together, Sakana AI presents these projects as evidence that recursive self-improvement is no longer purely theoretical, but is already being tested in controlled research environments.

Four phases of the road to RSI

In its blog post, Sakana outlines a four-phase transition from conventional, human-led AI optimization to self-improving systems.

The roadmap builds partly on Sakana's own research. It starts with models designed not as chatbots but for open-ended agent tasks from the ground up. With The AI Scientist, Sakana has already shipped a system that applies agent capabilities to automated research, from idea generation and experiments to writing scientific papers.

The next step is recursive self-improvement itself: AI agents that actively work on their own technical foundations by writing, benchmarking, and verifying code for their underlying architectures.

The fourth phase and long-term goal is broader access to frontier AI. Sakana positions RSI as a counter to the dominant scaling paradigm: instead of training ever-larger monolithic models with ever more compute, the company bets on adaptive systems and evolutionary optimization, where AI finds better solutions in as few attempts as possible.

The idea is that recursive self-improvement could work with moderate compute and depend less on the massive GPU clusters that big US AI labs and cloud providers run today. It's a plausible research direction, but still a strategic bet. There's no proof yet that self-improving systems can actually offset the structural advantage of large-scale data centers.

Anthropic warns of RSI risks

The RSI Lab's launch puts a spotlight on an issue that Anthropic recently flagged as a potential safety concern. Sakana focuses on the upside, framing RSI as a path toward more efficient and more widely accessible frontier AI.

Anthropic shares that view but also warns about the risks: full recursive self-improvement hasn't been achieved yet, but once it is, AI systems could drive their own development faster than institutions can keep up. For that scenario, Anthropic has floated the idea of a global pause on frontier AI development as something worth considering.

Sakana AI is one of the more notable AI startups outside the United States. The company was founded by former Google researchers, including Llion Jones, one of the authors of the Transformer paper "Attention Is All You Need," and David Ha, who previously worked at Google Brain and Stability AI.

Unlike many frontier AI labs, Sakana focuses less on scaling up individual large models and more on evolutionary, adaptive, and multi-part AI systems. The name "Sakana" means "fish" in Japanese, a nod to swarm behavior, evolution, and collective intelligence. That's what sets the startup apart: Sakana combines deep roots in Transformer research with a program that explicitly seeks alternatives to the dominant scaling paradigm.