Sakana AI Challenges the Arms Race with Self-Improving AI

The Japanese startup Sakana AI recently launched the Sakana AI RSI Lab, a research laboratory dedicated to recursive self-improvement (RSI) of artificial intelligence systems. This initiative aims to explore how AI can self-improve, thereby optimizing its own development without requiring massive computing resources.

Unlike the current trend that favors training increasingly large models, Sakana AI focuses on evolutionary optimization. The company envisions a future where AI agents could not only design but also enhance their own technical architectures. This vision is outlined in a four-phase roadmap, starting with models designed for open agent tasks and evolving towards systems capable of self-evaluating and rewriting their own code.

Since its founding in 2023, Sakana AI has developed several innovative projects, such as LLM-Squared, which improves language model training methods, and the Darwin Gödel Machine, which generates and tests variations of its own code. The projects ShinkaEvolve and ALE-Agent also illustrate work on optimizing evolutionary programs and agents deriving new strategies from trial-and-error loops.

Sakana AI sees RSI as a potential solution to the computing arms race. The idea is that recursive self-improvement could operate with moderate computing power and rely less on the massive GPU clusters that large American AI labs and cloud providers currently exploit. However, this innovative approach is not without risks. Anthropic, another entity in the AI field, has recently expressed concerns regarding the safety of recursive self-improvement. They fear that these systems may advance faster than the regulatory capabilities of institutions, suggesting a global pause in the development of cutting-edge AI.

Founded by former Google researchers, including Llion Jones and David Ha, Sakana AI stands out for its unique approach. Rather than focusing on the augmentation of individual models, the company prioritizes scalable and adaptive AI systems. The name "Sakana," meaning "fish" in Japanese, reflects this philosophy, inspired by collective intelligence and swarm behavior.