The timing is remarkable. One day after Anthropic warned that Claude now writes more than 80 percent of its own code, another lab turns exactly that topic into its mission: Sakana AI from Tokyo has officially launched its RSI Lab — a lab for recursive self-improvement.
What Sakana is after
Co-founder and CEO David Ha (known on X as hardmaru) put it this way: ‘Today, we are officially launching the Sakana AI RSI Lab in Tokyo to build open-ended, adaptive AI systems that collectively self-improve.’ The core idea: don’t just optimize the model itself — redesign the process of building AI using AI.
With this, Sakana takes a different path than the big frontier labs. Instead of training ever-larger models with ever-more compute, the team bets on nature-inspired approaches — collective intelligence, evolution, swarms. The stated goal: to offer a counterweight to the compute arms race of the major labs.
Two years of groundwork
The RSI Lab didn’t come out of nowhere. It builds on a whole series of projects Sakana has shipped over the past two years: The AI Scientist, a system for fully automated research — from idea to experiment to finished paper — made it into a Nature publication. ALE-Agent took first place out of 804 human participants in an AtCoder contest. And ShinkaEvolve showed how new algorithms can be evolved with just 150 samples.
My take
Recursive self-improvement is the topic of the week — and that’s no accident. While Anthropic raises a warning finger and cautions against losing control, Sakana sees an opportunity. Both sides have a point: it’s the most exciting and, at the same time, the trickiest direction AI can take right now.
On Hacker News, Sakana was promptly accused of just chasing the next hype. The defenders’ counter: this is exactly what they’ve been doing since day one — others are only now following suit. Who’s right remains to be seen. Either way, it stays interesting.
Sources: Hacker News: Sakana AI’s Recursive Self-Improvement (RSI) Lab, Sakana AI: Introducing the RSI Lab