The real bottleneck for AI discovery beyond math may be verification

Afinetheorem · x · 2026-11-22

Replying to Matt Clancy's thread on why AI-driven cross-silo discovery is happening in math but not elsewhere, @Afinetheorem argues that "we haven't tried" and "even if AI produced the right answer, we wouldn't know" are more plausible explanations than "AI was just RL-trained on math." He says he has ideas for measuring this, discussed previously in Berkeley.

Original post →

More from AGI Musings

AGI Musings channel →