gpt-6IMAGE
Two GPT-6 Astra instances grade each other on a Scholze conjecture test
The author had two Astra Pro instances correct each other while defining shtukas over Spec(Z) and trying to prove the Gestalten conjectures, producing 95 pages of slop, and notes that finding counterexamples via meta-search is far easier than building theory.
Przemek Chojecki | PC
@prz_chojeckiArchived 2026.09.27Originally posted 2026.09.26
Przemek Chojecki | PC @prz_chojeckiOriginally posted 2026.09.26
Whenever I want to take a model for a real math test, I task it with defining shtukas over Spec(Z) and then trying to prove Scholze's Gestalten conjectures. What a calamitous slop emerges. 95 pages by GPT-6 Astra Pro x 2 instances talking / correcting each other. Finding counterexamples through meta-search is way easier than building up theory.
x.com/prz_chojecki/status/2103860481924452826Archived 2026.09.27 10:07