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

What happened

Sakana AI, a Japanese startup co-founded by Llion Jones, who worked on Transformers, has launched a dedicated research lab focused on recursive self-improvement (RSI). This approach aims to develop AI that can iteratively improve its own code and models, rather than relying solely on scaling up raw computing power. This move positions RSI as a strategic alternative to the compute-heavy arms race led by large US AI labs. Meanwhile, AI safety firm Anthropic has issued warnings about the risks tied to the control and governance of self-improving AI systems.

Why it matters

Most frontier AI development today revolves around amassing more compute to train larger models. That approach favors big players with access to massive hardware and cloud resources. Sakana AI’s bet on recursive self-improvement targets a potentially more efficient path, where AI enhances itself without needing endless compute increases. If this works, it could lower entry barriers and speed innovation cycles, shifting power toward groups that master RSI techniques rather than pure compute scale. However, the risks flagged by Anthropic highlight that RSI systems might reduce human control over AI behavior, raising safety and regulatory challenges.

What to watch next

Follow Sakana AI’s progress on building self-optimizing AI systems and whether their approach delivers substantial efficiency gains versus scale-driven methods. Watch how regulators and the AI safety community respond to the emerging control risks highlighted by Anthropic, particularly around self-modifying AI. Also, track if other startups or labs pivot toward RSI as a way to compete without massive compute investments—this could reshape AI R&D priorities and business models in the near term.

AI Quick Briefs Editorial Desk