菲尔兹奖得主：AI现在主要靠「抬杠」突破重大数学猜想

TL;DR - Fields Medalist Timothy Gowers argues that LLMs currently excel at mathematical breakthroughs framed as finding counterexamples or constructing rare objects. Their broad knowledge and low-cost search enable cross-domain experimentation, but they still lack mathematicians’ intuition for prioritizing promising paths.

• Recent advances cited include counterexamples related to the Jacobian and Erdős unit-distance conjectures, plus constructions for non-sofic groups and multicolor Ramsey bounds.
• LLMs can cheaply explore large search spaces, recombine techniques, and import tools from distant fields such as algebraic number theory.
• Current systems often pursue plausible but unproductive approaches and repeatedly reformulate problems without materially nearing solutions.
• Gowers views genuinely novel, reusable mathematical methods—not merely successful constructions—as a stronger test of top-tier AI creativity.