AI, Work and the Economy

Latest/Story

AI doing original research

Four items, from 7 Oct 2026 to 20 Nov 2026, oldest first.

See also: Point 1 of our view: AI does more office work, cheaper

AI and technologyNew data

OpenAI (6 Oct): internal model produced 722 maths manuscripts on open problems; 185 main results checked in Lean

OpenAI released 722 manuscripts in 372 result families from an unreleased internal model, posed about 4,000 problems and spent roughly three hours of ChatGPT Pro compute per result; 185 main results are machine-checked in Lean. It claims advances on named open problems (e.g. Hilbert's tenth problem over Q, Goldfeld's conjecture); the advisory group says this is 'not an endorsement' and much is still under verification.

Why it matters. An AI system produced hundreds of research papers on unsolved maths problems at about three hours of computing each, a sign that AI is reaching the most specialised research work, though experts have yet to check most of the claims.

Next. Mathematicians' verification of the headline results, and whether OpenAI releases the model.

Sources: openai.com, github.com

AI and technologyNew dataUpdate

OpenAI withdraws 3 of its AI-produced maths papers (7 Oct) after a sign error; 14 others get proof repairs

OpenAI's maths repository records on 7 Oct the withdrawal of three manuscripts on the Hodge conjecture for K3 surfaces and related results, because one sign error invalidated an argument two of them depended on; 14 further manuscripts received repaired proofs or corrected statements and 13 more updated citations. This follows its 6 Oct report of 722 manuscripts produced by an internal model.

Why it matters. OpenAI has withdrawn three of the mathematics papers its model produced, after a single sign error undermined a proof two of them relied on, and repaired fourteen others. Machine-written research at this volume may need checking capacity that grows as fast as the output; errors that spread through dependent results are one way such work could mislead.

Next. Independent review of the remaining results by mathematicians over the coming weeks.

Sources: github.com, news.ycombinator.com, mathstodon.xyz · Point 1 of our view

AI and technologyReportUpdate

OpenAI (26 Oct): an internal model claims a disproof of the Exponential Time Hypothesis, stated in Lean; open review set

OpenAI published a preprint dated 25 Oct claiming a deterministic subexponential-time algorithm for 3-SAT, which would refute the Exponential Time Hypothesis, a central conjecture of complexity theory since 1999. It released claimed results on the Kakeya conjecture and on quantum chromodynamics in the same post. Russell Impagliazzo and Ramamohan Paturi, two of the conjecture's authors, and Ryan Williams are to lead an open examination. Commenters call the algorithm 'galactic', with no practical use.

Why it matters. OpenAI says one of its internal models has disproved a widely believed conjecture about how hard certain computing problems are, and three of the field's leading researchers, two of them its original authors, are to check it. If the proof holds, it would be among the first cases of a machine settling a central open question in a scientific field largely on its own, a step from assisting research toward doing it.

Next. The open review led by Impagliazzo, Paturi and Williams; OpenAI withdrew three earlier AI-written maths papers on 7 October after an error was found.

Sources: github.com, openai.com, news.ycombinator.com · Point 1 of our view

AI and technologyNew dataUpdate

Anthropic mathematician Alpöge claims a Lean-formalised proof of the Generalized Riemann Hypothesis made with Claude (20 Nov)

Levent Alpöge published a note claiming that, working with Claude, he has proved the Riemann Hypothesis and its generalisation, presenting only a Lean formalisation (an informal draft runs past 250,000 pages). Terence Tao the same day lists it beside Euler and Navier-Stokes blow-up proofs by Alpöge, Buckmaster and Coiculescu made with heavy AI assistance, which are human-readable and Lean-formalised; none is yet independently verified.

Why it matters. A researcher at Anthropic says he and its model Claude have proved the Generalized Riemann Hypothesis, among the most famous open problems in mathematics, and has released a machine-checkable version; on the same day, two long-open fluid-flow problems were reported solved with heavy AI help. None is yet confirmed by others. Should the proof check, it would be strong evidence that AI can now carry expert research far beyond routine work.

Next. Independent checks of the formal proof by mathematicians and Lean users over the coming days and weeks.

Sources: alpo.ge, terrytao.wordpress.com, news.ycombinator.com · Point 1 of our view