Sakana AI launches recursive self-improvement lab as an alternative to brute-force AI scaling

Sakana AI has launched a dedicated Recursive Self-Improvement Lab to study AI systems that can iteratively improve AI development itself.

A lab focused on self-improving AI

Tokyo-based Sakana AI announced the formal establishment of its Recursive Self-Improvement, or RSI, Lab in a company post titled “Introducing Sakana AI’s Recursive Self-Improvement (RSI) Lab.” The company says the lab will focus on redesigning AI development “with AI,” emphasizing sample-efficient methods rather than relying primarily on hyperscale compute.

The Decoder reports that Sakana AI is framing the lab as a strategic bet against the compute-intensive approach taken by major frontier AI companies, particularly in the United States. Instead of competing mainly by training ever-larger models on ever-larger infrastructure, Sakana AI is exploring whether AI systems can help generate improvements to models, code, training processes, or research workflows more efficiently.

Sakana AI was co-founded by Llion Jones, one of the authors of the Transformer paper, according to The Decoder. That background gives the company’s research direction additional visibility in a field still shaped by the Transformer architecture’s role in modern language models.

What recursive self-improvement means here

In Sakana AI’s framing, recursive self-improvement refers to AI systems that contribute to improving themselves or the systems around them over repeated cycles. The company’s RSI Lab announcement says the goal is to pursue AI development methods that are more sample-efficient than brute-force scaling.

Sakana AI’s related research on the “Darwin Gödel Machine” gives a concrete example of the direction. In a separate company post, Sakana AI describes the Darwin Gödel Machine as a self-improving coding system that rewrites its own code and uses open-ended evolution to improve benchmark performance. The company says the system improved through iterative modification rather than through a single fixed design.

That work does not prove that recursive self-improvement can replace large-scale training, and Sakana AI’s own descriptions focus on research prototypes and experimental systems. But it illustrates the company’s thesis: AI may be useful not only as a product of research, but also as a participant in the research and engineering process.

A challenge to the compute race

The Decoder characterizes Sakana AI’s RSI Lab as an attempt to break, or at least sidestep, the compute arms race among frontier labs. The current frontier model strategy often depends on access to large GPU clusters, expensive training runs, and major capital investment. Sakana AI’s announcement argues for a different axis of progress: using AI to discover improvements more efficiently.

That position is especially relevant for smaller companies and research groups that cannot match the infrastructure spending of the largest AI labs. If recursive self-improvement techniques can produce meaningful gains with less compute, they could broaden participation in advanced AI research. However, the sources do not establish that Sakana AI has already achieved that outcome at frontier scale.

Safety concerns remain unresolved

The same concept that makes recursive self-improvement attractive also makes it controversial. The Decoder notes that Anthropic has warned about control risks associated with AI systems that can improve themselves. The concern is that systems capable of modifying their own behavior, code, or development process may become harder to predict, evaluate, or constrain.

Sakana AI’s public materials present RSI as a research direction, not as a finished product. Still, the debate around recursive self-improvement is likely to center on whether researchers can build reliable evaluation, oversight, and containment methods alongside the technical capabilities.

For now, Sakana AI’s RSI Lab is best understood as an early institutional commitment to a specific research agenda: making AI development more automated and sample-efficient. Whether that approach becomes a practical alternative to brute-force scaling will depend on results that can be independently tested, reproduced, and evaluated for safety.