The takeaway: Terence Tao used the public lecture at the International Congress of Mathematicians to refuse the argument everyone expected — declining to debate whether AI can do research mathematics, and asking instead what the profession actually values. He published the slides and an AI-conducted interview with himself overnight. At the same congress, Hannah Fry took the Leelavati Prize for public understanding of mathematics. And three frontier stories from Friday are worth a second look: a model that drives robot arms, a 25-signatory open-weights letter the closed labs did not sign, and the security team that could not use hosted AI to investigate its own breach.

Mathematics decides what it is for

Tao gave the ICM public lecture in Philadelphia on Friday evening under the title "Mathematics in the age of AI," and posted his 52 slides shortly after midnight UTC.

He states the capability question formally — the "AI Capability Conjecture" — and then declines to argue it: "my talk will not be about that conjecture." Instead he asks the audience to assume a reasonably strong version is true and reason conditionally, which frees him to go at the question the profession has mostly left to one side. If the tools do work, what happens to a discipline that rewards being first to a proof?

His answer is that mathematics faces a "crisis in the foundations of mathematical values and practices," on the model of the 1900–1930 foundational crisis. The mechanism is Goodhart's law: optimise the one goal that is easiest to measure and easiest to automate — problems solved — and the goals that used to move together come apart. He lays out a five-stage chain (generation, verification, exposition, publication, canonicalisation), groups the back three under the heading "proof digestion," and argues that AI accelerates the front while barely touching the back. Without policy and cultural change, he says, the result is "proof indigestion" — a shift "from an era of proof scarcity to an era of proof abundance."

Our full piece is here. Two things the report discloses that are not on the slides: First Proof — the one piece of evidence Tao does put in front of the audience — takes unrestricted donations from Anthropic and OpenAI, with pending funding from Google.org. And Tao himself is listed in that report as a principal investigator on one of the four systems tested. Neither fact undoes the benchmark, which discloses both in its own front matter and is far better documented than the industry announcements Tao is criticising. Both belong next to the recommendation.

Fry takes the Leelavati Prize

At the same congress, Hannah Fry was awarded the 2026 Leelavati Prize — the International Mathematical Union's award, given every four years and sponsored by Infosys, for "outstanding contributions for increasing public awareness of mathematics as an intellectual discipline." The IMU's citation calls Fry "one of the foremost global ambassadors for mathematical thinking," and credits her with using "outstanding creativity and originality to translate mathematics into a language of wonder and relevance for the public without diminishing its scope and importance." Fry became Cambridge's first Professor of the Public Understanding of Mathematics in January 2025.

The pairing is worth noticing. In the interview he posted alongside the slides, Tao writes that there is "some long-delayed outreach that our profession needs to perform rather urgently" — that it is "now critically important to correct the general public opinion that mathematics is primarily about solving open problems by whatever means possible, even if the resulting proofs are incomprehensible to humans." A prize for explaining mathematics to the public, awarded the same week, puts a little institutional weight behind exactly that.

Three from Friday that got less attention than they deserved

A model that drives robot arms. Black Forest Labs' FLUX 3 puts image, video and audio in one architecture. Its derived model FLUX-mimic is in testing at Audi's Production Lab — the carmaker's pre-production piloting unit, not the assembly line — and Black Forest Labs claims its action decoder beats prior vision-language-action models even with the FLUX backbone completely frozen.

The open-weights letter, and who is missing from it. Twenty-five companies and organisations — Nvidia, Microsoft, Meta, IBM, Hugging Face, Mistral, Mozilla, the Linux Foundation and others — signed a statement backing open-weight models. OpenAI, Anthropic and Google are not signatories.

The defender who was locked out. Hugging Face could not use hosted frontier models to run forensics on its own breach: submitting real exploit payloads tripped the providers' safety classifiers which, in Hugging Face's words, "cannot distinguish an incident responder from an attacker." It ran the analysis on an open-weight model on its own hardware. The asymmetry is the story — and Hugging Face signed the open-weights letter eight days later.

Also on the record

Claude Opus 5, which we covered on Friday, now tops the Artificial Analysis Intelligence Index. Checked this morning: Opus 5 at max effort leads on a displayed 61, with Opus 5 at xhigh and Claude Fable 5 (with Opus 4.8 fallback) behind on 60 — about 0.6 on the underlying scores. The top three entries are all Anthropic configurations; the nearest model from another lab, GPT-5.6 Sol, sits on 59.

The detail worth holding onto is the mismatch. Anthropic does not position Opus 5 as its smartest model — our Friday piece noted it "isn't, and Anthropic doesn't claim it is," pitching it instead as close to Fable 5's frontier intelligence at half the price. An independent index now places it above the model its own maker ranks higher.

Tao says the lecture recording "will likely be forthcoming in a day or two." When it lands we will check it against the slides.