After OpenAI's 722 math papers, why AI will crack verifiable fields first but not bug-free code

ziv_ravid · x · 2026-10-08

Reacting to OpenAI publishing 722 math papers overnight, the author lays out a framework for how AI automation will spread:

The result is impressive, though error rates and proof reusability remain unclear.
Core claim: automation speed and quality are inversely related to the entropy of the solution space and how easily outputs can be verified.
Predicted order: mathematics (structured, verifiable) → exact sciences (strict laws, noisy measurements) → financial markets → medicine → law → and only last, art, culture, and soft traits like leadership.
On programming: it's only half-solved, unlike the consensus. Measurable parts—performance optimization, output-preserving rewrites—are at or near superhuman level, since you can just let the model try many things and keep what works.
But bug-free software is not coming soon: with real users in the loop, entropy is the real bottleneck—humans can't be formalized in Lean. User-facing coding sits closer to medicine and law than math on the difficulty ladder.

More from AGI Musings