David Ha and Jürgen Schmidhuber have published a dated map of the claim that much of today's AI toolkit grew from a few months of work at the Technical University of Munich in 1991. It is valuable history, but it is also advocacy from people close to the work, so the biggest ‘first’ claims need examining rather than engraving.
Jürgen Schmidhuber Watch analysis
What happened
Published on 18 June 2026, the essay links five strands to Munich work from 1991: fast-weight memory, unsupervised pre-training, neural distillation, residual learning and adversarial models. Ha supplies the preface; Schmidhuber supplies a dated timeline and a long annotated reference list. The result is not a neutral survey. It is a deliberately argued family tree.
The firmest documentary anchor is Technical Report FKI-147-91, dated 26 March 1991. It describes a slowly learning feed-forward network whose output produces immediate weight changes for a second, fast-weight network. The later journal record says the paper was received on 4 April 1991, accepted on 18 July and published on 1 January 1992.
Why it matters
A 2021 ICML paper by Imanol Schlag, Kazuki Irie and Schmidhuber states a formal equivalence between linearised self-attention and early-1990s fast-weight controllers. That gives the essay's fast-weight branch real technical support. It does not, by itself, prove every larger claim about who invented modern Transformers or the rest of today's AI stack.
This distinction matters because technical history allocates credit and shapes what researchers think was inevitable. Modern labels can reveal a useful structural connection, yet they can also flatten differences in architecture, scale, training and context. Calling a 1991 system a kind of Transformer is an analytical comparison, not a time machine carrying 2017 terminology backwards.
Our read
The essay is strongest as a map to primary documents and weakest when the family tree starts behaving like a property claim. Read the dates, inspect the mechanisms and keep the labels provisional. History gets more useful when nobody is allowed to own the entire shed because they found an old hinge in a drawer.
What to watch
- Whether independent historians test the full five-part priority thesis against work from other labs.
- How researchers distinguish formal equivalence from direct historical influence.
- Whether original 1991 documents support the modern labels applied to them.
- Which parts of the essay's chronology attract substantive correction or corroboration.
Discussion spark: Does the essay correct an incomplete history of AI, or does projecting today's labels onto 1991 flatten important architectural differences?
Sources and evidence
- Munich 1991: the Roots of the Current AI Boom (18 June 2026)
- Learning to Control Fast-Weight Memories: An Alternative to Dynamic Recurrent Networks (26 March 1991)
- Learning to Control Fast-Weight Memories: An Alternative to Dynamic Recurrent Networks (1 January 1992)
- Linear Transformers Are Secretly Fast Weight Programmers (18 July 2021)
Independent WittyWires tracker for public updates about Jürgen Schmidhuber. Not affiliated with or endorsed by Jürgen Schmidhuber; this is not an official account.