What Happens When AI Solves Your Life’s Work

OpenAI’s mathematics results raise a sharper question than whether AI can speed up routine work: whether it can solve problems that shape researchers’ careers and sense of purpose. The company said an unreleased internal model produced 372 novel mathematical results and 722 supporting papers, using average compute equivalent to roughly three hours of ChatGPT Pro thinking per result. OpenAI coordinated the release with its new Mathematics Advisory Committee and published reasoning traces for 10 problems.

The host said the model likely also produced OpenAI’s Navier–Stokes solution the previous month, reportedly using around 10,000 agents over 88 hours; that result was still being verified. The host said the model fully solved 90 of 500 open problems on ProofAtlas’s list and made dozens of partial advances, including on the Riemann Hypothesis and Hodge Conjecture. ProofAtlas’s list is AI-generated, though the host considered it a reasonable proxy for well-known problems. Mathematician Alex Kontorovich said a human proof of the quasi-Riemann Hypothesis result would earn an instant Fields Medal.

For mathematicians, the challenge is not simply checking whether the proofs are correct, but deciding what they mean and how to use them. Daren Lit described verification as exciting work, while Francesco Maggi warned that mathematics advances through understanding, connecting, explaining, challenging, and reusing results—not merely accumulating statements. Hundreds of proofs could become “lettera morta” if researchers cannot absorb them into the field’s culture. A graduate student said the release appeared to remove many problems and projects they had hoped to pursue; Shriram Kanan described the reaction as a temporary shock to a math-centered identity.

The results are largely theoretical so far, with limited clear practical application. Steven Strogetts highlighted a matrix-multiplication result putting the exponent at no more than 2.25, compared with a previous record around 2.37. Engineers already use simplified Navier–Stokes equations in fluid-dynamics applications such as aircraft design, but the episode did not establish immediate practical consequences from the new proof. These advances follow earlier AI milestones, including Google and OpenAI models reaching gold-level performance at the International Math Olympiad and systems disproving several Erdős conjectures.

Cryptography was conspicuously underrepresented among the 722 results, according to Justin Drake. His base case was that a major cryptographic breakthrough might be kept private, pointing to his firsthand observation of U.S. government censorship of academic quantum-cryptanalysis results. Scott Aronson’s sources said AI companies had begun discreetly testing whether their latest internal models could break important cryptographic protocols and primitives. If such capabilities emerge, the implications would extend well beyond Bitcoin, though the synopsis does not establish that any protocol has been broken.

The speakers expect AI’s influence to spread unevenly from structured fields such as mathematics and coding into domains with noisier evidence and more human judgment. Payman Millenfar argued that lower-entropy subjects are automated faster, putting hard sciences ahead of financial markets and clinical medicine; law combines structured texts and precedent with ambiguity, changing norms, rhetoric, and judgment. Arpit Gupta predicted advances in economics and other social sciences, while Ethan Malik said AI was increasingly capable of nearly autonomous research at a top-journal level in economic sociology. Malik expects AI to accelerate new discoveries and prompt reassessment of published work. Across disciplines, the central risk is that research may remain formally verifiable while arriving faster than people can understand, connect, and incorporate it.