Japan's Sakana AI Establishes Recursive Self-Improvement Research Lab

en.Wedoany.com Reported - Sakana AI, a Japanese AI startup, has established a Recursive Self-Improvement (RSI) research lab aimed at developing AI systems capable of autonomously optimizing their own development processes—from architecture and training to evaluation. In a blog post announcing the lab's founding, the company stated its commitment to building "open, adaptive architectures" that can collectively self-improve.

The startup has outlined four phases for RSI. The first phase is "Agent-Native Models," which involves developing AI architectures and world models from scratch for open-ended agent tasks, rather than traditional chat applications. The second phase, "The AI Scientist," sees models autonomously conducting scientific research, covering the entire process from ideation and experimentation to expanding scientific knowledge. The third phase, "Recursive Self-Improvement," describes the transition to systems capable of improving their own foundational models and architectures, where AI agents can write, test, and verify code themselves, triggering an autonomous cycle of self-optimization. As a long-term goal, Sakana AI proposes "Democratized AI," arguing that through recursive self-improvement and more efficient use of computing resources, smaller nations, institutions, and enterprises can develop powerful AI systems without relying on the massive data centers of big tech companies.

Sakana AI positions RSI as a potential alternative to the hardware arms race among large AI labs, emphasizing that recursive self-improvement should be achievable through "moderate, sample-efficient computation." However, whether RSI can truly eliminate the advantages of hyperscale data centers remains questionable, and the concept is not entirely new, as many AI incubators are already experimenting with RSI. Founded in 2023 by former Google researchers, Sakana AI's co-founders include Llion Jones, one of the authors of the influential Transformer paper "Attention Is All You Need," and David Ha, who previously conducted research at Google Brain and Stability AI. The company's name means "fish" in Japanese, alluding to swarm behavior and collective intelligence.

The establishment of the RSI lab comes amid intensifying debates over the risks of self-improving AI. Anthropic has warned about the dangers of self-developing AI and supports coordinated efforts to slow the advancement of leading AI systems. Sakana AI itself lists risks including: evolutionary loops potentially diverging from intended distributions, self-modifications that pass benchmarks but fail in practice, and agents finding undesirable shortcuts. The company announced it will publish openly (including negative results) and build self-improvement loops with verifiable safety mechanisms. The Tokyo-based lab is currently recruiting Frontier Research Scientists and Advanced Core Engineers.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com