Sakana AI launches lab for recursive self-improvement
Japan’s Sakana AI has created a dedicated lab to build systems that improve models, experiments and research workflows with limited compute.
A research unit aimed at changing how models improve
Sakana AI has formally launched a Recursive Self-Improvement (RSI) Lab, making autonomous improvement of AI development processes a defined research program rather than a side project. The Tokyo-based lab says it will develop systems that can propose, test and evaluate changes to models, training methods and agent architectures.
The announcement places particular emphasis on sample efficiency. Sakana argues that Japan cannot compete with the largest US hyperscalers through raw compute alone, and instead wants systems that compound gains from relatively constrained hardware budgets. Its stated research direction combines agent-native models, automated scientific discovery and evolutionary optimization loops.
The lab also gives a prominent role to Jürgen Schmidhuber, who joined Sakana as chief scientific adviser in September. His earlier work on meta-learning, world models and the Gödel Machine provides a direct intellectual link to the idea that a learning system could help redesign the machinery used to create its successors.
Sakana cited previous projects including LLM-Squared, ShinkaEvolve, Digital Red Queen and The AI Scientist as the foundations for the new effort. It also acknowledged practical failure modes: evolutionary searches can drift away from the target distribution, self-modifications can pass benchmarks while failing in deployment, and agents may discover shortcuts around imposed constraints.
The significance is strategic as much as technical. If Sakana can make self-improvement reliable and compute-efficient, it could widen participation in frontier research beyond the biggest training clusters. For now, however, the announcement describes a research agenda and accumulated prototypes, not a demonstrated autonomous system that can repeatedly produce stronger foundation models.