Sakana AI is betting on a different path to frontier progress
Sakana AI has launched a dedicated research group, the Sakana AI RSI Lab, to study recursive self-improvement, or RSI. The concept is ambitious but straightforward in outline: build AI systems that can help redesign, optimize, and extend the technical foundations of future AI systems, creating a compounding loop of improvement.
The startup’s argument is also strategic. Instead of treating progress as mainly a matter of training ever larger models with ever larger compute budgets, Sakana is positioning recursive self-improvement as a route to greater efficiency and wider accessibility. In that framing, the next step for advanced AI may not simply be more scale. It may be better systems for improving systems.
That makes the lab launch more than a branding move. It is a statement about where one of the most closely watched AI startups thinks a meaningful alternative to the current frontier race could emerge.
From theory to controlled experiments
Recursive self-improvement has long carried a speculative aura, often discussed as a distant possibility rather than a practical research program. Sakana is trying to narrow that gap by grounding the idea in specific prior projects. In announcing the RSI Lab, the company pointed to several efforts from the past two years that it sees as stepping stones.
Among them is LLM-Squared, which explores how language models can devise improved training methods for other language models. Another is the Darwin Godel Machine, described as a system that generates, tests, and iterates on variants of its own codebase. Sakana also highlighted ShinkaEvolve, ALE-Agent, and The AI Scientist, all projects tied to evolutionary optimization, trial-and-error strategy discovery, or the automation of parts of scientific research.
One of the stronger signals in that set is the company’s claim that a later version of The AI Scientist wrote a paper that passed peer review, with the underlying research published in Nature in March 2026. Taken together, Sakana presents these projects as evidence that recursive self-improvement has moved from pure thought experiment into controlled, incremental demonstrations.
The four-phase roadmap
Sakana says its roadmap unfolds in four phases, starting with systems designed not as chatbots but as tools for open-ended technical work. The broad idea is to move from human-led optimization toward AI agents that can increasingly contribute to their own underlying architecture, code, and design choices.
The roadmap matters because it implies a measured transition rather than an abrupt leap. Sakana is not saying today’s systems can autonomously reinvent themselves at full scale. It is instead laying out a sequence in which narrowly scoped optimization, experimentation, and code generation gradually become more central to the development loop.
That is a notable contrast with the dominant commercial narrative around AI, where capability gains are often described mainly through model size, training runs, and infrastructure spending. Sakana is emphasizing search, adaptation, and evolution as alternative levers.
Why this approach matters in the current AI market
The launch lands in a period when compute concentration has become one of the defining features of the AI industry. Frontier training runs are expensive, hardware access is uneven, and the gap between the largest labs and everyone else can look structural. Sakana’s thesis is that self-improving systems could soften that advantage by making progress less dependent on brute force scaling alone.
That is not the same as rejecting compute. It is a claim about where marginal gains may come from. If AI systems can discover better training methods, optimize their own code, and automate parts of research and experimentation, then the balance between algorithmic improvement and raw infrastructure could shift.
Even if the strongest form of recursive self-improvement remains far away, intermediate gains could still be valuable. Better tooling for experimental design, code iteration, or automated model improvement would matter to labs far beyond Sakana itself.
The real test ahead
The challenge for the new lab is not conceptual visibility. Recursive self-improvement is already one of the most discussed long-term ideas in AI. The challenge is demonstrating durable, measurable progress without overstating what current systems can actually do.
For now, Sakana’s announcement is best read as an attempt to formalize a research direction it believes has already produced early evidence. The company is arguing that recursive self-improvement should now be treated as an engineering agenda, not merely a philosophical possibility.
If that case holds, the implications would be significant. AI progress could become less about who can spend the most on compute and more about who can build systems that improve the research process itself. That is a harder path to prove, but potentially a more consequential one.
Key points
- Sakana AI has created the RSI Lab to study recursive self-improvement.
- The company says RSI could offer an alternative to the current compute arms race.
- Sakana tied the lab to prior projects including LLM-Squared, Darwin Godel Machine, and The AI Scientist.
- The company outlined a four-phase path toward systems that can increasingly improve their own technical foundations.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
