Yoav Goldberg: 880K GPU-Hours vs a Few Hundred Prompted LLM Hours for Same Math Result

yoavgo · x · 2026-09-10

Yoav Goldberg comments on the recent drama around using massive compute to solve a famous open math problem: one effort ran frontier AI for 880,000+ hours at a cost of millions of dollars, while two people who knew what they were doing achieved the same result with the same frontier models in just a couple hundred LLM hours.

His pointed conclusion: expert prompting versus brute-force compute may differ by orders of magnitude in cost-effectiveness — exposing the hidden waste in compute-stacking AI research.

Related event: Yoav Goldberg questions AI math proof hype(2 posts)→

Original post →

More from Fun

Fun channel →