Researchers Challenge AI on Math Conjectures
Szegedy and Fleuret reject claims AI will disprove remaining Millennium Prize problems.
Christian Szegedy states that finding counterexamples is not inherently simpler than proving theorems and that their complexities span a wide spectrum. François Fleuret replies very not to the same discussion. Both comments address suggestions in the thread that large language models could soon disprove the remaining Millennium Prize problems. The packet records these as direct quotes from the researchers on X with no independent confirmation of outcomes.
I would challenge this. This opinion is based on the idea that finding counterexamples is inherently simpler than proving general statements. Intuitively, we can expect the complexity of finding counterexamples and proving theorems to span a wide spectrum.
This is almost certainly right If a conjecture survives sustained attacks from AI then it is probably true https://twitter.com/prz_chojecki/status/2083529701788721645
