Current Position & Biography
Jürgen Schmidhuber is a professor of computer science and co-chair of the Center of Excellence for Generative AI at King Abdullah University of Science and Technology (KAUST) in Saudi Arabia. In September 2026, Sakana AI announced his appointment as chief scientific advisor alongside his existing positions, with involvement in its Recursive Self-Improvement Lab. His career connects neural-network research in Europe with contemporary AI research programs in Saudi Arabia and Japan.
He earned his doctorate in computer science at the Technical University of Munich in 1991 and completed his habilitation there in 1993. Before KAUST, his institutional career included leadership at the Swiss AI laboratory IDSIA and a professorship at the University of Lugano. His research has addressed learning over time, training deep networks, internal models of environments and systems that improve their own learning procedures.
Notable Contributions
With Sepp Hochreiter, Schmidhuber coauthored the 1997 paper introducing Long Short-Term Memory, or LSTM. The work addressed the difficulty of learning dependencies across long sequences with recurrent neural networks. Its memory-cell architecture provided a method for maintaining useful learning signals over time. The contribution belongs to the named coauthors and the subsequent research community, not to an individual alone.
In 2015, Rupesh Kumar Srivastava, Klaus Greff and Schmidhuber introduced highway networks. Their architecture used learned gates to regulate information passing through a deep network, helping train networks with many layers. This is a specific, documented contribution to deep-network optimization; it does not establish sole authorship of every later architecture that uses skip connections or gating.
With David Ha, Schmidhuber coauthored World Models in 2018. The work studied compact representations of reinforcement-learning environments and controllers that could learn within a model-generated simulation before being evaluated in the underlying environment. The results were demonstrations in defined research tasks, not evidence of a system that perfectly understands physical reality or performs safely in arbitrary settings.
Sakana AI's 2026 appointment announcement connects this line of research to its plans for world models, physical AI and recursive self-improvement. Those plans are an organizational research agenda rather than proof that open-ended self-improving intelligence has been achieved. KAUST describes his current work as spanning research, education and engagement with public and private organizations.
Why Katharos Is Following Them
Schmidhuber's published collaborations provide an important thread through the history of recurrent learning, deep architectures and model-based agents. His current institutional roles also connect research communities across several countries at a time when the distinction between demonstrated capability and projected general intelligence matters.
Areas to Follow
Sequence-learning architectures; learned world models; meta-learning and self-improvement; reproducible evaluations of physical AI; and the translation of research ideas into dependable systems.
Reporting Scope
Coverage anchors attribution in dated papers and named collaborators. It distinguishes documented contributions from broader historical-priority claims, which remain a subject of public debate, and separates institutional ambitions from independently demonstrated results. Predictions about general intelligence are treated as predictions, not settled scientific outcomes.
Schmidhuber’s own account of AI priority disputes is included among the sources as an interested participant’s perspective, not an independent resolution of those debates.
