Can sample efficiency replace compute scale in recursive self-improvement?

Sakana AI announces the formal establishment of its RSI Lab in Tokyo, with Jürgen Schmidhuber joining as Chief Scientific Advisor, and states the lab's defining choice explicitly: "we are building not the most compute-hungry self-improvement engine, but the most sample-efficient one. Its advances should compound on national, rather than hyperscale, compute budgets." The lab's own portfolio is offered as evidence this is already the house style — ShinkaEvolve "solved complex optimization problems using only 150 samples," and ALE-Agent "secured 1st place out of 804 human participants" by "extracting structured lessons from its own failures, not by burning more inference."

Sakana AI's stated reason for the choice is geographic, not merely technical: "Frontier RSI is being attempted, almost exclusively, inside the world's two largest compute clusters," while Japan offers "deep scientific talent, strong engineering culture, and a compute envelope that is large by global standards but modest next to the hyperscalers." Under that constraint, Sakana AI argues, sample efficiency is "not a preference but a structural necessity," and the resulting techniques are "exactly the ones most likely to generalize beyond the two countries currently sprinting on raw scale." The lab maps its own trajectory as four phases — Agent-Native Models, the AI Scientist, Recursive Self-Improvement, and what it calls Democratized AI, "the point at which exponential self-improvement becomes a public good rather than a winner-take-all asset." Sakana AI also names its failure modes directly as "the central engineering problem": "evolutionary loops that drift off-distribution, self-modifications that pass benchmarks but fail in deployment, agents that find shortcuts around the constraints they were given," and commits to publishing negative results and building "verifiable safeguards from the start."

This gives a concrete answer to a question Can recursive self-improvement speed up the research process itself? leaves open — what variable actually raises the fixed efficiency of the research process. That paper treats sustained self-improvement as the remedy in the abstract; Sakana AI names sample efficiency, not raw compute, as the specific lever, and ties the choice to a resource constraint (national versus hyperscale budgets) rather than a purely algorithmic one. The announcement also leans on Can AI systems improve themselves through trial and error? as part of its own lineage — DGM's self-rewriting evolutionary archive is listed as the 2025 milestone the RSI Lab builds on, so the two notes describe the same technique from the lab's internal roadmap and the paper's own account. Sakana AI's promise to publish negative results and treat shortcut-finding as a central engineering problem also stands against Do AI researchers view automating AI research as a severe risk?: this is a lab publicly naming the failure modes that the cited survey found academic participants giving limited consideration to.

As a lab's own announcement, the excerpt establishes what Sakana AI is claiming and planning, not what has been independently verified. The portfolio list itself is contested elsewhere in the library: Does Sakana's AI Scientist deliver autonomous research without human help? checked one of the exact systems named here — the AI Scientist — and found failed code and misjudged novelty where this announcement reports a landmark, globally recognized result. Nothing here quantifies what "verifiable safeguards" will mean in practice, or sets a timeline for the "critical inflection point" phase. The sample-efficiency framing and the democratization promise should be read as a stated strategic bet and research agenda, not as a measured outcome.