Article Thesis

Working Thesis

AI progress in mathematics and science is shaped by feedback cost. Where outputs can be checked cheaply, systems can search, specialize, and improve quickly. Where the hard part is choosing the question, digesting the answer, training judgment, or maintaining a community that knows how to use the literature, progress does not scale in the same way.

The post should argue that this split matters more than the usual question of whether AI is “good at math” or whether “AGI” is near. The real partition is between work with cheap checks and work where the check is slow, expensive, or missing.

Stronger Thesis

AI can accelerate current mathematical work and change which work looks valuable. Formal proofs, benchmarked tasks, and search-heavy problems will become easier to produce and easier to reward. Question selection, proof digestion, training, synthesis, and community knowledge may become more important while remaining harder to measure.

Use this stronger thesis only if the article has already established the feedback-cost mechanism and the institutional mechanism.

Short Version

AI scales where feedback is cheap. Mathematics and science contain many such pockets, but their deepest bottlenecks are often elsewhere.

What The Post Is Not

  • It is not a prediction of a hard AI plateau.
  • It is not an argument that AI cannot do serious mathematics.
  • It is not a defense of human labor for its own sake.
  • It is not a survey of every AI-for-math paper.
  • It is not a guide to using AI tools.
  • It is not a moral panic about students cheating.

Reader Takeaway

After reading the post, the reader should be able to say:

AI does best where there is a cheap feedback loop. That explains why proof search and benchmarked work can move quickly, while question selection, digestion, education, and community knowledge remain bottlenecks. The institutional danger is that the cheap-to-check part of science starts to define what gets rewarded.

Current Candidate Title Directions

These are working labels, not final titles.

  • “AI Scales Where Feedback Is Cheap”
  • “Proof Is Not The Whole Bottleneck”
  • “The Part Of Mathematics AI Does Not Check”
  • “The Feedback Cost Map Of AI For Science”