Jürgen Schmidhuber is joining Sakana AI. Sakana announced on September 24 that the researcher behind LSTM and some of the earliest work on meta-learning and world models is "officially joining Sakana AI as Chief Scientific Advisor" to its "newly formed RSI Lab," a Tokyo group aiming "to trigger a compounding cycle of scientific discovery aimed at improving machine intelligence." For a lab that has built its identity on self-improving AI, it's a major hire: Schmidhuber's 1987 thesis described a program that improves its own learning algorithm, and Sakana's Darwin Gödel Machine descends directly from his Gödel machine idea. See the RSI Lab page.
What he'll actually do
The role is advisory. The lab page says he "will advise the lab" and "be in Tokyo regularly to meet with the team," and he keeps his position directing the AI Initiative at KAUST. The relationship isn't new: Sakana CEO David Ha co-authored the 2018 "World Models" paper with him. The RSI Lab launched in early June, so it's about three and a half months old.
Sakana's announcement calls him "universally recognized as the father of modern AI." The work is real, but "universally" glosses over his long-running priority disputes with Hinton, LeCun and Bengio (an 88-page report, reportedly listing 17 specific claims).
What the lab has shown, and what it promises
The page lists Sakana's lineage: LLM² (an LLM-discovered preference-optimization algorithm), the Darwin Gödel Machine, ShinkaEvolve, ALE-Agent, Digital Red Queen and The AI Scientist. The Darwin Gödel Machine is the closest to "self-improvement": it raised a coding agent from 20% to 50% on SWE-bench (the page's "30 percentage point absolute improvement") by rewriting its own scaffolding around a frozen foundation model. The roadmap's third phase is the leap beyond that, "AI agents actively write, benchmark, and verify the code of their own underlying foundation architectures," stated as a goal, not a result. The Xiaomi launch drew the same criticism here for treating "the RSI path" as a claim without a hedge.
The compute claim contradicts itself
The page's core pitch is that RSI is achievable "on modest, sample-efficient compute," and that Sakana is "building not the most compute-hungry self-improvement engine, but the most sample-efficient one." Its evidence is ShinkaEvolve ("only 150 samples") and ALE-Agent, which the page says "outperformed 804 human heuristics specialists... not by burning more inference." Yet ALE-Agent's own entry, higher on the same page, says it worked by "leveraging massive inference-time scaling and a self-learning mechanism." Both can't be the summary. And 150 samples on a specific optimization problem shows efficient program search, not that improving a foundation model's own training loop is cheap.
Safety: a real section, but a posture, not a protocol
The page has a "Toward Responsible RSI" section, and it's better than most. It names failure modes Sakana says it has seen directly: "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," calls them "the central engineering problem," and promises to "publish openly, including negative results, and design our self-improvement loops with verifiable safeguards from the start."
There are no specifics: no sandbox description (the Darwin Gödel Machine paper had one), no stop conditions, no named oversight, no definition of "verifiable." Two tensions sit beside it:
- Cyber. Digital Red Queen is presented as "the foundation for applying RSI to cybersecurity, modeling how autonomous agents can continuously co-evolve to discover, exploit, and patch vulnerabilities," and hiring text wants scientists applying evolutionary dynamics to "cybersecurity and automated red-teaming." Fugu-Cyber kept a human in the loop before anything was called real; this framing doesn't say the same.
- Diffusion. Phase four is "democratized AI," self-improvement as "a public good rather than a winner-take-all asset," on compute anyone can afford. The safety section covers failure modes of Sakana's own loops, not what happens once a cheap RSI recipe is published. The sovereignty logic is the same one this blog read in Namazu: a genuine, compute-constrained strategy for Japan, with the open question being whether the constraint yields efficiency or just smaller ambitions.