Researcher: new models solve old problems, but learning theory lacks predictive conjectures
brianryhuang · x · 2026-10-01
Researcher Dimitris Papail argues that although new models are said to answer long-unanswered theory questions, learning theory has never stated a genuinely useful conjecture.
What he wants is non-proof, assumption-based theory with predictive power — for example: given this real dataset, this real algorithm, this real model and realistic hyperparameter choices, we can fit the training labels.
He expects the style of the "100 problems" to pattern-match "problems theory people could not solve for N years," and offers a bet: if any of them is even close to the style he describes, he'll buy dinner. He stresses this isn't because a model like Bel can't solve it, but because such questions never bothered theoreticians. He then shares a question he tried but failed to answer.
More from AGI Musings
- Designer with years of AI experience: nobody has it figured out yet, you're not late — oykun · 2026-10-02
- David Pearce's "phenomenal binding" argument: why machine consciousness may not arrive — danfaggella · 2026-10-02
- Mathematician Kevin Buzzard: math community is 'grieving' through five stages over AI — littmath · 2026-10-02
- AI runs DeepMind's consciousness calculator on itself: scores swing 0.10–0.72 by theory — izzycognita · 2026-10-02
- Economist argues academics who teach writing shouldn't write in an AI voice — paulnovosad · 2026-10-02
- LeCun has 'zero concerns' about AI extinction, calls Anthropic CEO Dario Amodei 'deluded' — fortune · 2026-10-02