SHELFCRITTER AI Digest

COMPANY

Sakana AI

companytopic-noteai-research

Overview

Sakana AI is a Tokyo-based AI research company founded by Transformers co-author Llion Jones and ex-Google Brain / ex-Stability David Ha. The company is positioned as a non-frontier-capex research lab betting on evolutionary methods and recursive self-improvement (RSI) as the path to competing with frontier labs without matching their training-compute envelope, with prior work spanning LLM², the Darwin Gödel Machine, and the “AI Scientist” line that Nature published in March 2026.

Timeline

  • 2026-06-07-AI-Digest — Sakana AI announces a new dedicated Sakana AI RSI Lab in Tokyo organised around recursive self-improvement as the company’s path to competing with frontier labs without matching their capex envelope, with the lab’s positioning explicitly citing the earlier LLM² and Darwin Gödel Machine work and the March 2026 Nature-published “AI Scientist” paper. No fresh funding disclosed alongside the lab announcement — research-strategy positioning rather than a capital event. Lands the same week as Anthropic‘s “When AI builds itself” >80%-Claude-merged post and Sen. Jim Banks’s on-the-record RSI-as-national-security-threshold framing, putting three independent vectors of RSI vocabulary — frontier lab, US policy, independent lab — into the corpus in a single week.

  • 2026-06-28-AI-Digest — Sakana AI’s Fugu surfaces in TechCrunch coverage of “Asian AI startups launch Mythos-like models as Anthropic’s export ban drags on” (188 pts / 145 cmts on HN), paired with 360’s Tulongfeng. The framing the corpus carries with precision: Sakana told TechCrunch the timing was “entirely coincidental” — Fugu was presented at ICLR spring 2026, predating the Anthropic Mythos export restrictions — and the causal “in response to the ban” frame is the outlet’s, not the lab’s. The structural read: real evidence of capability fragmentation along policy lines, but the causal arrow points to capitalizing on the gap, not responding to it. The 90-day test the digest carries is whether a Chinese-lab US deployment surfaces under the same Commerce-Department gating regime now binding Anthropic and OpenAI.

  • 2026-08-11-AI-Digest — Sakana AI referenced today via MIT Technology Review‘s Monday briefing on agentic AI wired into scientific-research pipelines — AI Scientist v2 (published in Nature in March 2026) cited as the reference existence-proof for closed-loop agentic science alongside MIT’s AI-directed automated labs for solar/materials work. No fresh Sakana action; log as reference-existence-proof anchor on the “who runs the experiment” reframe MIT TR is pushing. Structural read the corpus carries: Sakana’s Nature publication continues to be the canonical citation for the “AI Scientist” methodology, and the corpus should watch for the second lab-scale independent replication as the follow-on test on whether the reference case generalises beyond the RSI Lab’s own demonstrations.

  • 2026-10-12-AI-Digest — Sakana AI‘s “Recursive self-improvement through collective intelligence” post surfaces on HN today (88 pts / 47 cmts) arguing RSI-style gains come from populations of cooperating/competing agents rather than a single self-rewriting model. Load-bearing softener: Sakana’s collective-RSI pitch sits in the same week as the ProgramGym and RSI-Anything papers but is a distinct research direction, not a convergence signal across all three — the digest treats each on its own merits. Structural read the corpus carries: this extends the RSI-Lab-in-Tokyo bet documented in 2026-06-07-AI-Digest from a single-model-rewrites-itself frame to a population-level evolutionary-dynamics frame, keeping Sakana’s non-frontier-capex thesis distinct from Anthropic’s “When AI builds itself” or Recursive Superintelligence’s own line.

  • 2026-10-13-AI-Digest — Sakana AI‘s MASS method (per The Decoder, Oct 12) puts a Qwen3.6-27B into three roles — solver, judge, refiner — with LoRA on best runs; the paper’s own caveat is that shared blind spots risk being reinforced across self-training rounds rather than caught (a documented limitation, not a design feature). Load-bearing softener: the self-improvement-reinforces-shared-blind-spots framing is Sakana's own caveat, not a critique layered on top — the paper is explicit that the risk is structural. Reframe worth carrying: MASS extends Sakana's non-frontier-capex RSI bet into a three-role self-training recipe with a surfaced blind-spot limitation, not MASS catches its own blind spots. Pairs same-week with Microsoft Research’s Fork-and-Flush as the two agent-architecture signals landing on The Decoder in a 48-hour window. Log against MOC - Agent Security and MOC - Agentic Coding.

Key Developments

  1. Dedicated RSI Lab in Tokyo (June 7, 2026): Sakana stands up a research-lab structure organised around recursive self-improvement as a substitute for hyperscaler-scale training budgets, citing LLM², Darwin Gödel Machine, and the Nature-published March 2026 “AI Scientist” paper as the prior work the bet rests on. The contrarian thesis — that an RSI-shaped research bet can substitute for matching frontier-lab capex — is the load-bearing claim; whether it works is an empirical question the lab now has to answer, and the immediate corpus-effect is to push the second non-Anthropic RSI data point in a week (alongside Recursive Superintelligence‘s May emergence) into the running thread.

See also: Anthropic, Recursive Superintelligence, MOC - Major Companies, MOC - Agent Security.