Sakana AI hires Jürgen Schmidhuber, inventor of deep learning, world models, and your next ChatGPT update

What happened

Tokyo-based Sakana AI has appointed Jürgen Schmidhuber, a researcher widely credited with foundational work on deep learning and world models, as its Chief Scientific Advisor. Sakana has described him as a father figure of modern AI research. In his new role, Schmidhuber will help lead the company's newly formed RSI Lab, a research unit focused on recursive self-improvement — systems designed to iteratively refine their own capabilities over time.

According to Sakana AI, Schmidhuber's research from the 1990s has already informed existing company projects, including its Darwin Gödel Machine initiative, which explores self-modifying AI architectures.

Why it matters

This is fundamentally a technology and talent story: Sakana AI is signalling where it believes the next meaningful AI capability frontier lies — not in larger static models, but in systems that can autonomously improve their own architecture and performance. Recursive self-improvement, if it matures, could change how organisations think about AI deployment, shifting from periodically retrained models to systems that evolve continuously in production.

For leaders tracking AI strategy, the hire underscores that foundational research talent is being actively recruited to translate decades-old theoretical concepts into deployable systems. Organisations evaluating AI vendors or partners should watch whether self-improving architectures move from research labs into commercial products, since this could alter assumptions about model lifecycle, governance and oversight.

By the numbers

• 1990s — the decade from which Schmidhuber's foundational research on self-improving systems originates, now informing Sakana's current projects.

The Renascence take

Headline-grabbing hires like this often get framed purely as a talent coup, but the more interesting signal is strategic intent: Sakana is betting that the next competitive edge in AI won't come from scale alone, but from systems that can rewrite their own rules of improvement.

Recursive self-improvement sounds abstract until you consider what it implies for service and experience design: AI systems that adapt their own behaviour over time, rather than waiting for the next human-led retraining cycle. The operational discipline this demands isn't technical alone — it's governance. Any organisation exploring self-improving AI needs clear guardrails on what "improvement" means in a customer-facing context, because a model optimising for its own defined success metric won't automatically optimise for trust, fairness or customer outcomes. The lesson here isn't "hire a famous scientist" — it's "decide now how you'll audit a system that keeps changing itself."