Tao did call it the biggest advance in 50 years. The clip stops before the bound

The line is real. "OpenAI did make a major breakthrough, the biggest advance on the subject in 50 years, I'd say." Terence Tao said it on camera to Numberphile, in a twenty-minute interview published on 15 October, and by Saturday it was a screenshot with his name above it and no link below. The post that carried it has drawn roughly 2,900 likes. The attribution survives scrutiny. What does not is the sentence's surroundings, which include Tao explaining, seconds later, how far the advance goes and where the method stops.

Where the sentence actually lives

The passage does not appear in any Tao blog post, paper or Mastodon thread that I could find. The primary is a video: "No More Heroes", Brady Haran's interview for Numberphile, published on 15 October. At about sixteen minutes the interviewer asks whether the Riemann hypothesis will fall soon, and Tao answers with the paragraph now circulating. The auto-generated transcript has no punctuation and a few mangled words; the X post tidies both without changing the substance.

The relay adds one editorial mark. The bracket in "[on quasi-Riemann Hypothesis]" belongs to the poster and narrows Tao's referent: the question put to him was about the Riemann hypothesis, not the weaker version, and he says "the subject" before drawing the distinction himself.

Digg's summary is more careful than what it summarises: "A post attributes to Terence Tao the claim..." That hedge is correct. Its "4 Sources" panel is not: two of the four are retweets of a single post by @MehmetMars7, a third is commentary, and the fourth is the post itself. One artifact cited four times.

Twenty seconds of caveats

What follows "I will give them that". Tao puts the result in scale: the zeros are now known to lie to the left of 7/8 where previously they were only known to lie a little to the left of 1, so "we're 1/4 of the way to the Riemann hypothesis". Then he bounds the method, saying it "cannot get beyond" a certain point "without another extremely clever idea". His summary is that the hypothesis "hasn't fallen" but "has been dented quite a bit", and his forecast is that "something beyond current AI is needed". He also volunteers that he has read "about 10 of the 300 odd papers".

None of that contradicts the compliment: the biggest advance in fifty years and a method with a visible ceiling answer different questions, and Tao gave both inside ninety seconds. The clip is not a fabrication. It is a crop, and what it removes is the only part that says how much of the problem is left.

A weaker claim, correctly named

The Riemann hypothesis says every non-trivial zero of the zeta function has real part one half, as Wikipedia states it. The quasi-Riemann hypothesis says something looser: that there is a fixed line Re(s) = θ, with θ below 1, to the right of which there are no zeros at all. The first implies the second; the second does not imply the first, and proving it leaves the critical line untouched.

Unconditionally, the best general zero-free regions are strips that hug the line σ = 1 and thin out as the imaginary part grows; none gives a fixed half-plane. That makes 7/8 a serious statement rather than a bookkeeping one, and also, in Tao's arithmetic, a quarter of the distance.

OpenAI's own scope note is precise: the 7/8 bound for the zeta function and every Dirichlet L-function, uniformly over all moduli and characters, plus finite-order Hecke L-functions over Q(sqrt(-3)), with the pole at s = 1 excluded. The note adds: "The paper's later applications are not included."

What the company put its name to

OpenAI announced the release in a blog post on 6 October, describing a broad range of results from an internal frontier model, Lean formalisations for "many of the proofs", and an average cost of roughly three hours of ChatGPT Pro thinking per result. The post does not single out the quasi-Riemann hypothesis; the claim lives in a paper title inside the repository.

Scientific American counted more than 700 manuscripts covering 372 open problems, and reported that some had already been retracted after outside scientists found mistakes, that OpenAI published no full prompts or run-time details, and that it promised better citations in future releases. No retraction is reported as touching this paper. The gap that matters is simpler: nobody has refereed it.

The first ten days of other people's arithmetic

On 8 October, Baiying Liu posted a paper that takes OpenAI's 7/8 as its starting point, varies three parameters, improves the constant to 0.874957019..., and proves it optimal in that scheme, with a Lean formalisation. The gain sits in the fourth decimal place; the signal is that an outsider could re-run the argument within 48 hours.

On 14 October, Matt Visser posted the opposite kind of paper: assume the claim and see what follows. He derives, among other things, a prime between n^8 and n^8 + n^7 for every positive integer n, and calls the exercise "contingent on the validity of the claimed 7/8 version". That is what early verification looks like: strangers building on the result with the caveat attached.

In Scientific American, Hector Pasten of the Pontifical Catholic University of Chile supplied the fairest account of the stakes: many applications of the Riemann hypothesis never needed the full thing, and the quasi version "is way more than enough for many applications". On whether it is progress toward the real problem: "Some people claim it is progress toward Riemann; other people say it's not."

The blog post that is not about this

Tao's own blog carries something that looks, at a glance, like a rebuttal. On 12 October it published "This is normal bad work", a guest post by Sophie Huiberts, converted with AI and signed off by Tao with a single initial. It calls OpenAI's PDFs "exhaust fumes", says they "actively set back the state of the literature", and calls them "easy rejects". It is not about the quasi-Riemann hypothesis and does not dispute the interview. It is about the packaging, and both judgements can hold: the best advance in fifty years, delivered in papers his blog calls poor.

What would settle it

The attribution question is closed. The mathematical one is not, and the difference is the story. What would move it: mathematicians reading the argument rather than the abstract; a check that the Lean statement matches the informal theorem, since a kernel verifies steps and not intent; and a referee report. Scott Aaronson made the related point in a comment on his blog: a certificate can mislead through kernel bugs, an unsound formal system, or a formal statement that does not match the informal one.

Until then the defensible reading is the one Tao gave and the crop removed: the biggest advance in fifty years, by a method whose limits were already visible, on a problem that did not fall. The quote is accurate. The theorem is a claim.