For makers and artists, the relentless race to build bigger models with more GPUs is finally getting a rival. Sakana AI has launched a research lab dedicated to recursive self-improvement (RSI), proposing that systems capable of redesigning their own code could bypass the compute-heavy bottlenecks currently straining the industry. This shift promises a future where AI development is less dependent on massive data centres and more accessible to those with moderate resources.
A new approach to model evolution
Founded in 2023 by former Google researchers Llion Jones, a co-author of the seminal “Attention Is All You Need” paper, and David Ha, Sakana AI has pivoted away from the standard strategy of scaling monolithic models. Instead, the company is betting on evolutionary optimization. The new Sakana AI RSI Lab builds on earlier work to explore how AI can iteratively refine its own architecture, creating a compounding cycle of technical progress.
The startup points to several proof-of-concept projects from the last two years as evidence that this is no longer purely theoretical. Key milestones include LLM-Squared, where language models design superior training methods for their peers, and the Darwin Gödel Machine, which generates, tests, and iterates on its own codebase variants.
Further work on ShinkaEvolve and ALE-Agent demonstrates evolutionary program optimization and agents that derive new strategies through trial-and-error loops. Perhaps most strikingly, the system known as The AI Scientist has automated parts of scientific research to the point where it wrote a paper that passed peer review. That underlying research was published in Nature in March 2026.
A four-stage roadmap to autonomy
Sakana outlines a clear transition from human-led optimisation to fully self-improving systems across four phases:
The first phase moves away from chatbot-centric models toward open-ended agent tasks built from the ground up. The existing AI Scientist system already applies these capabilities to generate ideas, run experiments, and draft scientific papers.
The second phase introduces recursive self-improvement proper. Here, AI agents actively work on their technical foundations by writing, benchmarking, and verifying the code for their underlying architectures.
The third phase involves agents that can derive new strategies from trial-and-error loops, effectively learning to solve problems they have not seen before.
The final, long-term goal is broader access to frontier AI. By relying on adaptive systems rather than ever-larger models, the hope is to reduce dependency on the massive GPU clusters currently run by big US labs and cloud providers.
While the logic is sound, it remains a strategic bet. There is currently no proof that self-improving systems can offset the structural advantage held by large-scale data centres, and the ability to run these agents with moderate compute is not yet guaranteed.
Safety concerns from industry giants
The launch of the RSI Lab highlights a tension between efficiency and safety. While Sakana frames RSI as a path to wider accessibility, safety giant Anthropic has recently flagged full recursive self-improvement as a potential risk. Anthropic warns that once achieved, AI systems could drive their own development faster than institutions can regulate or keep up with.
In response to this scenario, Anthropic has suggested that a global pause on frontier AI development might be necessary. Sakana, however, focuses on the upside, positioning their work as a necessary evolution rather than a threat.
The name “Sakana,” meaning “fish” in Japanese, reflects the company’s philosophy of swarm behaviour, evolution, and collective intelligence. By combining deep roots in Transformer research with an explicit programme to find alternatives to the dominant scaling paradigm, the startup aims to break the current compute arms race. The name serves as a reminder of the collective, adaptive nature of the systems they are trying to build.
Key takeaways
Sakana AI’s new RSI Lab aims to replace the compute-heavy scaling paradigm with evolutionary optimisation, potentially making frontier AI more accessible.
Recent projects like The AI Scientist and LLM-Squared demonstrate that AI systems can already automate research and design better training methods.
While promising, the technology faces significant hurdles; there is no proof yet that self-improving agents can operate effectively on moderate compute.
Safety experts like Anthropic warn that once RSI is fully realised, AI development could outpace institutional control, prompting calls for a global pause.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
