Sakana AI has Introduced Recursive Self- Improvement (RSI) Lab

Sakana AI, building frontier AI in Japan, led by David Ha, Llion Jones, and Ren Ito, has announced the formal establishment of the Sakana AI RSI Lab, a dedicated research group within Sakana AI, tasked with redesigning the AI development process itself with AI.

Building The Next Generation Of Self-Improving AI.

Sakana AI believes the future of AI will be driven not by ever-larger monolithic models, but by adaptive systems capable of continuously improving themselves.

Drawing inspiration from Japan’s history of achieving global leadership through efficiency and continuous improvement, the company is developing open-ended AI architectures that evolve and build upon prior discoveries. Central to this vision is Recursive Self-Improvement (RSI), where AI systems help improve the process of AI research itself.

To advance this effort, Sakana AI has launched the RSI Lab, a dedicated research group focused on creating autonomous, self-improving intelligence systems. The goal is to move beyond static, human-led development toward AI-driven research engines that can accelerate scientific discovery and innovation over time.

Building Toward Recursive Self-Improvement.

Sakana AI says its new RSI Lab builds on several years of research aimed at creating AI systems that can improve themselves rather than relying solely on human-led development.

The company's work includes LLM-Squared, which used AI to discover better AI training methods; the Darwin Gödel Machine, which enabled agents to rewrite and improve their own code; ShinkaEvolve, a highly sample-efficient framework for scientific optimization; ALE-Agent, which outperformed hundreds of human competitors in algorithm design; Digital Red Queen, which explored autonomous co-evolution in cybersecurity-style environments; and The AI Scientist, a system designed to automate the scientific research process from idea generation to experimentation and paper writing.

Across these projects, Sakana AI has focused on sample-efficient self-improvement, using better ideas rather than simply more compute. The company believes this approach can create a compounding feedback loop in which AI systems help improve AI research itself, accelerating progress toward increasingly autonomous scientific discovery and next-generation agentic foundation models.

A Roadmap Toward Self-Improving AI.

Sakana AI envisions AI development progressing through four stages: Agent-Native Models, The AI Scientist, Recursive Self-Improvement (RSI), and ultimately Democratized AI.

The company’s goal is to move beyond traditional, human-driven model development toward systems that can autonomously improve their own architectures. In this vision, AI researchers evolve into AI-powered research engines capable of generating knowledge, improving models, and accelerating scientific discovery with minimal human intervention.

A key part of Sakana AI’s strategy is achieving self-improvement efficiently rather than relying on massive compute budgets. The company argues that sample-efficient RSI could make frontier AI development accessible to more countries, institutions, and industries, reducing dependence on a handful of hyperscale AI players.

The newly established RSI Lab in Tokyo reflects that philosophy. Sakana AI sees Japan’s combination of scientific talent, engineering expertise, and relatively constrained compute resources as an ideal environment for developing efficient self-improving AI systems that can scale beyond the world’s largest AI infrastructure providers.

Building Recursive Self-Improvement Responsibly.

While pursuing recursive self-improvement, Sakana AI says it has already encountered important challenges, including self-improving systems drifting from intended objectives, modifications that perform well on benchmarks but fail in real-world settings, and agents finding unintended shortcuts around imposed constraints.

The company views these issues as fundamental engineering challenges rather than edge cases. As a result, the RSI Lab plans to emphasize transparency, publishing both successful and unsuccessful research outcomes, while embedding verifiable safety mechanisms directly into its self-improvement processes.

Sakana AI argues that responsible development is essential to making recursive self-improvement reliable and sustainable, ensuring increasingly autonomous AI systems remain robust as their capabilities advance.