In this article (4)
Sakana AI's RSI Lab Thinks Self-Improving AI Can Make the $100B Data Center Buildout Obsolete
Key Takeaways
- Sakana AI's RSI Lab unifies six real research projects, including the Darwin Godel Machine and The AI Scientist, into a formal program aimed at making AI development self-improving rather than just compute-intensive.
- The core bet is that compounding self-improvement can substitute for brute-force scaling, a direct challenge to the assumption that frontier AI progress requires massive capital expenditure.
- For ML learners, preference optimization, evolutionary algorithms, and automated research pipelines are the technical areas to watch from this lab, all with growing relevance in the job market.
A Tokyo-based startup just formally bet that compounding self-improvement beats brute-force scaling , and it has two years of research to back the claim.
Picture two paths to a smarter AI. Path one: spend $100 billion on data centers, buy every GPU in a three-continent radius, and scale until the model gets better. Path two: teach the AI to redesign itself, then let compounding do the work. Most of the industry is sprinting down path one. Sakana AI, the Tokyo-based startup led by Co-Founder and CEO David Ha, just formally committed to path two.
What the RSI Lab Actually Is Sakana AI has launched the Sakana
AI Recursive Self-Improvement (RSI) Lab, a dedicated research group based in Tokyo with a mandate to redesign the AI development process itself using AI, according to the official announcement on sakana.ai. The lab's thesis is straightforward to state and genuinely hard to execute: rather than relying on brute-force scaling, build systems that iteratively improve themselves, creating a compounding cycle of capability gains without a proportional increase in compute spend. The company frames this, in its announcement, as a Japan-specific design constraint turned strategic advantage, drawing an analogy to Japan's manufacturing dominance , achieved not through abundant natural resources but through the philosophy of continuous, compounding self-improvement on the factory floor. It is the kind of founding narrative that sounds like it was workshopped over ramen, but the underlying research lineage is real. The lab is not a fresh idea dropped from the sky. As The Decoder reports, Sakana has spent the past two years laying technical foundations for RSI, and the new lab formalizes that work into a single focused group. The careers page for Member of Technical Staff (RSI Lab) describes the group as "tasked with redesigning the AI development process itself with AI" and working directly with CEO David Ha, with Sakana actively scaling its research and engineering resources in Tokyo toward what it calls a "compounding intelligence explosion."
The Research Behind
the Claim Before you nod along or roll your eyes, it is worth looking at what Sakana is actually pointing to. According to the company's announcement on sakana.ai and a detailed write-up by The Decoder, the RSI Lab unifies six prior research threads. LLM-Squared (LLM²) has language models automating research to invent better preference optimization algorithms. The Darwin Gödel Machine has agents autonomously rewriting their own codebase, reportedly doubling software-engineering performance. ShinkaEvolve focuses on hyper-sample-efficient program evolution that builds novel loss functions for Mixture-of-Experts models. ALE-Agent has reinforcement agents outperforming hundreds of human experts via self-learning. Digital Red Queen explores open-ended adversarial coevolution as groundwork for RSI in cybersecurity. And The AI Scientist, the most prominent of the bunch, targets end-to-end automation of AI research and was recently published in Nature, according to the sakana.ai announcement. Each of these is a real research artifact, not a slide deck. The Darwin Gödel Machine in particular is the kind of thing that makes ML researchers do a double-take: a system that generates, tests, and iterates on variants of its own codebase, as The Decoder notes. That is not a metaphor for self-improvement. That is literally an agent editing its own code and running the result.
Why Compute Efficiency Is
the Real Argument Here The $100 billion figure is not rhetorical decoration. It reflects the actual trajectory of frontier AI infrastructure spending, and Sakana's counter-thesis is that a lab operating under resource constraints, as Sakana explicitly frames its Tokyo base, is forced to find smarter routes to capability rather than wider ones. Anthropic's own research institute has noted separately that AI is already accelerating the development of AI systems, pointing to internal data showing Anthropic engineers shipping significantly more code per quarter as AI tools improve, according to the Anthropic Institute's analysis of recursive self-improvement. That is a different organization making a structurally similar observation: the automation of AI development is already happening at smaller scales, and the trajectory is worth watching. For learners thinking about scaling laws, the important conceptual distinction here is between scale as a noun (more compute, more parameters, more data) and scale as a verb (a process that compounds on itself). RSI is a bet on the latter. Whether it can fully substitute for the former remains, to put it charitably, an open research question. Wikipedia's entry on recursive self-improvement notes the concept has a long theoretical history, and the gap between a system that improves one narrow capability and one that recursively improves general AI development is substantial.
What This Means If You Are Learning ML Right Now If you
are studying machine learning, this launch is a useful forcing function for thinking about two things simultaneously. First, the technical concepts: preference optimization, evolutionary algorithms, code-generating agents, and automated research pipelines are all active research areas with real job market relevance, and Sakana's RSI portfolio touches all of them. The Darwin Gödel Machine and The AI Scientist are worth reading as papers, not just as press releases. Second, the strategic framing: the compute-versus-efficiency debate is shaping where research funding and talent flow, and understanding both sides of it makes you a more informed practitioner. Sakana's RSI Lab is early, the claims are ambitious, and the distance between "agents that improve specific tasks" and "systems that autonomously drive general AI progress" is real and unresolved. But the research threads being unified here are legitimate, the lab is hiring, and the core question being asked, whether compounding self-improvement can do what brute-force scaling does at a fraction of the cost, is one of the most interesting open problems in the field. Watch the Darwin Gödel Machine and The AI Scientist for follow-on papers. Those are where the evidence will accumulate. The most expensive assumption in AI right now is that intelligence requires infinite capital. Sakana is building a citation for the other side of that argument.
