Grant Sanderson (@3blue1brown) – AI disproved a famous math conjecture. Now what?

21 insights saved from this video by @technology

@technology
Mathematical originality means deciding what deserves study, not merely solving problems humans have already posed.
@technology
If AI automates theorem proving, mathematics may discover that defining the right concepts mattered more than proving them.
@technology
AI could make theorem proving less prestigious than defining concepts, curating knowledge, and choosing worthwhile directions.
@technology
Mathematicians may become art-museum curators of mathematics, selecting worthwhile ideas from an effectively infinite space.
@technology
Digital minds may gain further advantages by pooling knowledge and spawning identical copies, alongside massive parallelization.
@technology
Massive investment can turn quantity into a qualitative advantage even when each AI remains below a human mathematician.
@technology
Autonomous agents improve when environments reward reflection, such as searching the whole codebase or assessing their own mistakes.
@technology
Mathematical originality rises from proving theorems, to generating conjectures, to inventing definitions.
@technology
As AI generates vastly more mathematics, the scarce skill may become deciding what is useful and worth pursuing.
@technology
Human-curated learning artifacts sequence concepts and motivate each problem. LLMs then explore locally around that structure.
@technology
A mathematical concept may need a verification loop lasting decades or centuries across later developments in unrelated fields.
@technology
Formal proof systems could enable autonomous mathematical exploration: machine-checkable steps let computation continue indefinitely without continual human review.
@technology
If AI builds correct new theories, its economic impact could spread widely even without performing every human task.
@technology
Language models may suit mathematical discovery unusually well by mastering quantum physics and analytic number theory simultaneously.
@technology
Harder synthetic problems can force cross-field connections by removing assumptions while still requiring a solution.
@technology
Conjecture generation may matter enormously while resisting standard benchmarks, since its quality is difficult to quantify objectively.
@technology
A software harness could amplify intellectual diversity by defining distinct approaches and systematically running agents across them.
@technology
Deep learning needs many parallel rollouts because it learns slowly from each individual example.
@technology
Galois’s contribution resisted a tidy solution: his early theory was incomplete, unclear, and did not identify a specific unsolvable quintic.
@technology
The dream workflow: models write readable proofs, then formalize them for machine-certified correctness.
@technology
An AI can find deep mathematical connections yet still perform poorly at ordinary white-collar work such as editing.

Post on X

Keep what you found in Grant Sanderson (@3blue1brown) – AI disproved a famous math…

A free account saves any of these cards to your Collection, and Korva resurfaces them so you actually remember.