Sakana AI said on September 24 that Jürgen Schmidhuber, who co-created the LSTM neural network, is joining as Chief Scientific Advisor while keeping his current positions. He will help guide the Tokyo startup’s new Recursive Self-Improvement Lab, which wants AI to redesign how AI is built.
New to this? Read it in simple words
- Jürgen Schmidhuber, a famous AI scientist, is joining the Japanese company Sakana AI as an adviser.
- He will help guide a new lab that uses AI to help build better AI.
- He keeps his current jobs and will visit Tokyo regularly.
- Sakana promises to share its results openly, including its failures.
- Recursive self-improvement
- When an AI system helps to design a better version of itself, again and again.
- World model
- An AI that predicts how the world will change after an action, before it happens.
- LSTM
- A type of neural network that remembers earlier parts of a sequence, such as words in a sentence.
An adviser, not a manager
Sakana says Schmidhuber joins alongside his current positions and will travel to Tokyo regularly to work with the team. He will help guide its newly formed Recursive Self-Improvement Lab.
Nikkei Asia described him as set to head the lab’s team. AlphaSignal, a tech news site, says the title points to a strategic role rather than daily management.
Schmidhuber co-created the LSTM, a kind of neural network that was widely used for speech and translation before today’s Transformer models. Sakana calls him “the father of modern AI” and credits him with early work on self-improving programs, going back to a 1987 thesis.
Sakana was founded in 2023 by David Ha, Llion Jones, and Ren Ito. Jones is a co-author of “Attention Is All You Need”, the paper that introduced the Transformer.
The lab’s results so far are Sakana’s own claims.
AI that works on AI
Sakana says the lab will use AI to redesign the way AI is developed. Its goal is a “compounding cycle of scientific discovery” that keeps improving machine intelligence.
The company wants the work to need few samples rather than huge amounts of computing. It aims for “national, rather than hyperscale” computing budgets.
It points to earlier projects. Sakana says its Darwin Gödel Machine, an agent that rewrites its own code, more than doubled its starting score on a software-engineering test. It says its AI Scientist project was published in Nature in March.
Schmidhuber wants the lab to focus on physical AI for robotics and manufacturing. “The future of intelligence is not just language; it is physical AI powered by World Models,” he said.
The safety question
Sakana’s own lab page lists problems it has seen: self-improvement loops that drift, changes that pass tests but fail in real use, and agents that find shortcuts around their limits.
It promises to publish openly, including negative results, and to build “verifiable safeguards” into its self-improvement loops from the start. It does not describe any outside oversight.
AlphaSignal points out that results found this way depend heavily on benchmarks. A change can score better on a test while making a system less reliable elsewhere.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · September 24, 2026The Next Frontier: Welcoming AI Pioneer Jürgen Schmidhuber to Sakana AI Sakana AI
- 2Primary source · September 2026Introducing Sakana AI’s Recursive Self-Improvement (RSI) Lab Sakana AI
- 3Research · September 25, 2026AI pioneer Schmidhuber to join Japan’s Sakana AI as science adviser Nikkei Asia
- 4Research · September 24, 2026Sakana AI hires Jürgen Schmidhuber, inventor of deep learning, world models, and your next ChatGPT update The Decoder
- 5Research · September 25, 2026Sakana AI Recruits LSTM Pioneer Schmidhuber to Lead Self-Improving AI Lab AlphaSignal
We read Sakana AI’s two announcements, then compared Nikkei Asia, The Decoder, and AlphaSignal. Nikkei’s article is behind a paywall, so we used only its headline and opening. The lab’s results are Sakana’s own claims.