Jürgen Schmidhuber Joins Sakana AI’s Self-Improvement Lab

Sakana AI has appointed Jürgen Schmidhuber as chief scientific adviser to its newly formed Recursive Self-Improvement Lab, or RSI Lab. He will retain his current positions, advise the Tokyo startup, and travel regularly to Japan. The appointment connects Sakana’s work on self-modifying agents and learned simulations with a researcher who has pursued both ideas for decades.

Sakana’s announcement adds scientific leadership to an existing research program and includes no new model, product, funding round, or release date. The lab will focus on two difficult problems: software agents that improve their own code and world models that predict how actions change an environment.

Self-improvement, translated

Recursive self-improvement describes a loop in which a system proposes changes to its own software, tests those changes, retains useful variants, and repeats the process. Sakana wants to apply that loop to AI research, including architecture design, tool use, planning, and experimentation.

Sakana’s stated objective is a compounding cycle of scientific discovery that improves machine intelligence. Existing demonstrations remain bounded by human-defined tasks, evaluation methods, compute budgets, and permissions. A system that edits selected components under controlled tests is still far from an autonomous system that can broadly improve its own capabilities.

An Agent-Native World Model is Sakana’s term for a learned simulator designed for use by software agents. Such a model estimates how an environment will change after an action, allowing an agent to compare possible outcomes before acting. In robotics, for example, it might predict whether a grasp will succeed or how an object will move after contact.

Schmidhuber’s earlier research on world models, planning, and curiosity-driven learning gives the lab a direct intellectual lineage. Curiosity-driven systems generate their own learning signals by seeking states that improve their predictions, an approach that can help when labeled training data or explicit rewards are scarce.

Old theories meet running code

Sakana’s Darwin Gödel Machine differs from the original theoretical design in a practical way. A Gödel Machine is meant to prove that a proposed rewrite will improve its objective before applying the change. Sakana’s system searches experimentally, measures candidate modifications, and uses observed performance to decide which versions survive.

Empirical search can operate without constructing a formal proof, but its results depend heavily on benchmarks. A rewrite may improve performance on the measured tasks while reducing reliability elsewhere, exploiting flaws in the evaluator, or consuming more compute than the gain justifies.

Why Schmidhuber fits

Schmidhuber co-created the long short-term memory network, commonly called LSTM, which became a standard architecture for speech, translation, handwriting recognition, and other sequence tasks before Transformers became dominant. His 1987 diploma thesis also presented an early formal treatment of recursive self-improvement and meta-learning.

His advisory title defines a strategic role rather than day-to-day management of the RSI Lab. Sakana says he will help shape its scientific direction while keeping his existing positions and making regular visits to Tokyo.

Sakana was founded in 2023 by David Ha, Llion Jones, and Ren Ito. Ha and Jones previously worked at Google, and Jones co-authored the Transformer paper “Attention Is All You Need.” The startup has concentrated on evolutionary search, model merging, automated research, and coordinated systems of smaller models.

Tokyo as an industrial test bed

Sakana is using the appointment to support its effort to attract international AI researchers to Japan. Schmidhuber has linked the country’s robotics and manufacturing base with the lab’s Physical AI agenda, which targets systems that reason about machines, objects, and physical processes.

Learned world models could let manufacturers test supply-chain decisions, robot policies, and factory configurations in simulation before deploying them. Their value will depend on fidelity: errors in contact dynamics, rare events, sensor behavior, or unfamiliar operating conditions can produce plans that work in simulation and fail on hardware.

What changes inside Sakana

• Scientific direction: The RSI Lab gains an adviser whose research history closely matches its work on self-modification, meta-learning, and learned simulations.
• Research continuity: The Darwin Gödel Machine and AI Scientist now sit within a broader program devoted to recursive improvement.
• Hiring: Sakana is recruiting for technical roles in Tokyo.
• Deliverables: The company has not specified a release schedule, API, model license, compute budget, or benchmark target for the lab.

Developer impact remains upstream

The appointment changes no API, pricing plan, or production interface today. Its practical relevance lies in the systems Sakana may build around automated code modification, research agents, planning, and physical simulation.

Current agent frameworks commonly rely on people to revise prompts, tools, memory systems, and orchestration code. A successful descendant of the Darwin Gödel Machine could automate some of that engineering. Useful results would need to show improvements on unseen tasks, preserve performance outside the optimization benchmark, and report the compute and evaluation costs required to obtain each gain.

Robotics developers would need world models that integrate with existing simulators and control stacks, remain calibrated under changing conditions, and expose uncertainty when predictions become unreliable. Sakana has not said whether future models will be released as open weights, hosted services, research code, or commercial products.

The benchmarks that matter

• Scope of autonomy: Which components can the system modify, and which remain fixed by researchers?
• Generalization: Do improvements transfer to held-out tasks, environments, and hardware?
• Economics: Does the performance gain justify the training, search, and evaluation cost?
• Reproducibility: Can independent teams obtain similar results from the released methods and code?
• Control: Can operators inspect changes, enforce limits, detect evaluator exploitation, and restore earlier versions?

After four decades of theory and narrow demonstrations, no generally self-improving AI system has emerged. Sakana is placing that research agenda inside a dedicated lab with an adviser who helped define it. The lab’s results will depend on measurable gains, transparent evaluations, and systems that remain useful outside the benchmarks that created them.