The viral post on OpenAI and Navier–Stokes: what NIK says, what is true, what is missing

In short: on 8 September 2026, while mathematicians published preprints and OpenAI briefed the press, the story that fixed public perception was written in 11 lines by a pseudonymous account: NIK's post (@ns123abc), 1.4 million views. We checked it line by line against the documents: 4 claims hold, the rest are misleading, dramatized or incomplete — and the most important one remains unanswered.

Last updated: 9 September 2026 · Evidence frozen at 8 September 2026, 23:00 UTC

How to read the evidence labels

OpenAI claim stated by OpenAI, not publicly checkable · Independent evidence public document, preprint or page · Mathematically established independent scrutiny plus coherent formalization · Unverified metric or content not directly inspectable

This article is about a post, not a theorem. The theorems — or rather, the theorems — we reconstructed in our analysis of OpenAI and Navier–Stokes: what was actually proven. What matters here is a question worth as much as the mathematics: how did an 11-line summary, written by a pseudonym, become the official version of events for millions of people before the documents were even read?

The answer is neither "because it tells the truth" nor "because it lies". It does both at once: it wraps real facts in a frame that exceeds them. Taking it apart piece by piece is the only serious way to use it — and an information-hygiene exercise that stays valid for the next viral scientific controversy, whatever it is.

What does the post say? / Cosa dice il post?

Direct answer: NIK's post of 8 September 2026, 11:34 AM, summarizes the mathematicians' version in 11 greentext-style lines, with two attached screenshots. Here it is in full, quoted as a document:

>math professor spends a year on one of the hardest unsolved problems in math
>"Navier-Stokes"
>his drafts for the whole project went through codex sessions
>openai had his logs in codex
>suddenly rumor leaks AI solved it
>he emails openai to ask what's going on
>openai calls within days
>"funny story, our model also solved it"
>using his exact approach
>asks openai: did you train on my sessions?
>no answer

What it says: a professor spends a year on one of the hardest open problems, Navier–Stokes; every draft goes through Codex sessions; OpenAI holds his logs; suddenly rumor leaks the AI solved it; he emails OpenAI asking what is going on; OpenAI calls within days ("funny story, our model solved it too") with his exact approach; he asks whether the model was trained on his sessions; no answer.

Metrics verified in-page at inspection time: 1.4 million views Independent evidence. Likes and reposts are live counters: the reported snapshot was about 33.8k likes and 3.7k reposts, our fetch shows about 35k and 3.9k — normal drift, not a discrepancy. The two attached screenshots were not inspected by us Unverified: nothing in them can enter the verdicts.

Who amplified it? / Chi lo ha amplificato?

Direct answer: three accounts with different roles: a viral aggregator (NIK), a careful technician (konstiwohlwend) and a partisan commentator (rynorhn). Confusing them is the first reading error.

NIK (@ns123abc) is a pseudonym active since 2019, about 76.6k followers at snapshot, specialized in sensationalist BREAKING news on the AI industry ("ITS HAPPENING"). Nothing publicly verifiable exists about the physical person, and it should not be sought: what counts here is the function — a high-reach amplifier with a mixed track record. An auto-generated wiki attributes a past at xAI to the account: a weak source, reported only as an unattributed attribution.

@konstiwohlwend is Konsti Wohlwend, a builder in San Francisco (Hexclave, YC S24; former Google and Jane Street intern; ICPC/IPhO competition background), about 2.5k followers. His "drama summary" is the most accurate of the three and carries the warning the others omit: the two mathematicians do not hold the million-dollar proof, but proofs of related problems. A small voice with enormous amplification: the textbook case of virality rewarding readability over authority.

@rynorhn is Ryan Orhan, an AI commentator (about 3.7k followers, November 2025 account), previously quoted by Techmeme on other topics. His thread is live chronicle fused with explicit invective against OpenAI: read it as advocacy, not as a neutral report — even where the facts it carries are correct.

Line-by-line verdicts / Il verdetto riga per riga

Direct answer: of 9 statements, 4 hold against the documents, 1 is misleading, 1 is dramatization, 1 is incomplete, 1 is ambiguous and 1 remains unresolved. The full matrix, the core of this article:

Post line	Verdict	Documentary basis
A professor, a year on the problem	Supported	Buckmaster statement: collaboration since 19/09/2025, breakthrough 15/08, Lean 22/08
"Navier-Stokes", the hardest problem	Misleading	Published proofs concern Euler/Boussinesq/IPM; on Navier–Stokes only the non-public OpenAI claim exists
Drafts in Codex sessions	Supported as stated	Statement: Codex with GPT-5.6 Sol for the whole project, costs paid from own funds
"OpenAI had his logs"	Ambiguous	As service operator, yes; use of the contents unestablished — see the three channels below
Rumor "AI solved it"	Supported	Curran prediction post plus misreading of Tao's hypothetical 3/09 scenario
Email to OpenAI	Supported	3/09 email quoted in full in the statement
"Funny story, our model also solved it"	Dramatization	The account speaks of a ~100-page claim on 6/09; the sketch tone is NIK's, OpenAI's version differs
"Using his exact approach"	Unresolved	Same narrow route per Buckmaster; OpenAI says different Euler proof (New Scientist); undecidable without the manuscript
"Did you train on my sessions? No answer"	Incomplete	Training question unanswered on record; but lookup denied, human access denied, official denial arrived after the post

The rule applied here too: no amount of optimization can justify a statement stronger than the available evidence.

The three data channels / I tre canali dietro una domanda

Direct answer: "did you train on my sessions?" confuses three different things, with three different answers. It is the real information gap the post leaves behind — and this article's original contribution.

1. Training — unanswered on record

OpenAI policy for individual plans (ChatGPT, Codex) says contents may be used for training unless the user opts out, with separate Codex controls; for Business/API the default is no. Buckmaster paid out of pocket, so the channel was abstractly possible — but whether that internal model used them is established by nobody.

2. Runtime lookup — denied

To whether the model had consulted user data, the reported answer is no. The post never says so: "no answer" flattens a partial answer into a missing one.

3. Human access — denied by OpenAI

At the press conference OpenAI denied human access to Codex, with Mark Chen extending the formula to agents ("that's also our understanding", New Scientist). A corporate statement, not an independent check — but it exists, and the post predates it.

A timing note nobody cites: August 2026 drafts can hardly live inside a model trained months earlier, short of continuous fine-tuning — a possible but undocumented hypothesis. Even suspicion needs a mechanism, not just a motive.

What the post leaves out / Cosa manca nel post

Direct answer: everything that arrived after 11:34 AM on 8 September — the second half of the story.

Hours after the post, the official @OpenAI account published two things: the solution announcement ("We're sharing a solution to the Navier-Stokes Millennium Prize Problem... The proof was produced by a group of agents...") and an explicit denial: congratulations to the two mathematicians plus "did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve..." (text cut by "Show more"; quoted up to the visible part). Bubeck added technical points, starting from the sprint's genesis: viral rumors of two Anthropic-solved Millenniums. And the press conference produced on-record details absent from the post: 1,000 agents for 50 hours on Euler, then 10,000 agents for 11 hours on the extension, estimated customer cost around $15 million (New Scientist).

The rest of the picture is missing too: the "false and inflammatory" denial with promised reply, the Brown–Douglas exchange, Tao's Euler-only validation, Anandkumar's independent PINN result, the two-year Clay path. An 11:34 AM post could not contain them; anyone reading it today must know the story moved on.

Layered timeline: documents, rumors, viral posts, press. The NIK post compresses into 11 lines what takes three lanes here.

19/09/2025 → 22/08/2026 — Documents lane. Collaboration, 15/08 breakthrough, 22/08 Lean, 7–8/09 preprints. Independent evidence
3 → 6/09 — Rumors and calls lane. Curran prediction, 3/09 email, OpenAI sprint since 1/09, two 6/09 calls in conflicting versions.
8/09 11:34 AM — Viral lane. The NIK post freezes the mathematicians' version at 1.4M views; the threads detail and color it.
8/09 afternoon–evening — Press and denials lane. Axios, Scientific American, New Scientist, TechCrunch; official @OpenAI denial; Bubeck promises a reply. The post contains none of it.

Why it exploded / Perché è esploso

Direct answer: because it is a perfect morality tale in a perfect format: David versus Goliath, 11 lines, a corporate villain and an unanswered question for a finale.

The ingredients are measurable: a sympathetic protagonist (the professor paying out of pocket), a structural suspicion (whoever runs the tool sees the data), suspicious timing (days between rumor and counter-announcement), an open ending ("no answer") that invites sharing more than any conclusion. Quote-tweets amplify through indignation, not verification: "rooting for Team Human Professors" is declared fandom, and that is fine — as long as it is recognized as fandom.

The paradox that makes the story genuinely interesting: the most accurate thread (konstiwohlwend, who denies the Millennium framing) carries a fraction of the driest post's reach. Virality rewarded compression, not precision. Exactly the mechanism our first article described as the bottleneck: generating narratives is cheap, checking them stays expensive.

How to read viral science posts / Come leggere i post virali sulla scienza

Direct answer: with five one-minute checks each — the same ones applied here.

Who speaks? Pseudonym or verifiable byline? Track record or view counter? NIK is an amplifier, not a witness.
What is cited? Does the post link documents or only other posts? Here: zero links, two uninspectable screenshots.
What is compressed? "Had his logs" hides three different technical channels; every compression is a verification hypothesis.
Who denies, and when? Always look for the timestamped reply: the @OpenAI denial postdates the post but predates your reading.
What stays open? An unanswered question ("training?") is not proof of guilt: it is an open inquiry, and should be treated as one.

Known vs unknown / Cosa sappiamo e cosa non sappiamo — 9 September 2026

We know

The post text and its headline metrics (1.4M views verified).
Who the three amplifiers are, and in which style.
That 4 of 9 lines hold against the documents.
That Codex policy makes training abstractly possible on individual plans.
That OpenAI denied every access path, in conference and on X.

We do not know

What the two attached screenshots show.
Whether the internal model was trained on the sessions (unanswered question).
Whether OpenAI's route truly matches the mathematicians' (manuscript needed).
Who NIK is in real life — and there is no need to know.
How Bubeck's promised reply will land.

Quantitative audit — 9 September 2026

Post lines checked: 9 — 4 supported, 1 misleading, 1 dramatization, 1 incomplete, 1 ambiguous, 1 unresolved.
Metrics verified in-page: 1.4M views; likes/reposts as drifting snapshots (~33.8k to ~35k; ~3.7k to ~3.9k).
Accounts profiled: 3, all from dated bio and reach snapshots; zero non-public personal data.
Primary documents crossed: statement, preprints, Tao, OpenAI policy (2 help-center articles), @OpenAI and Bubeck posts, 4 outlets (Axios, SciAm, New Scientist, TechCrunch).
Attached screenshots inspected: 0 of 2 — explicitly excluded from the verdicts.

Method and limits: descriptive statistics over this investigation's corpus. X metrics are live counters and drift over time; secondary-thread likes are attributed snapshots, not independently verified.

An unsentimental conclusion. NIK's post did its amplifier job better than any press release: it put a million people in front of the right question — who controls the data science is made of — starting from a partial reconstruction. Our job is the opposite and the complement: keep the question, drop the partiality. Even if OpenAI's denial were complete and sincere, the structural problem the post made visible would stand intact: when the tools of research belong to those who compete in research, every coincidence will demand proof of innocence. And proofs of innocence, in mathematics as in reporting, get published — not announced.

Method and limits: X quotes verified via direct fetch on 8–9 September 2026; "Show more"-truncated texts quoted up to the visible part; metrics as dated snapshots. We will update the verdicts on Bubeck's reply, the OpenAI manuscript or policy clarifications.

Frequently asked questions / FAQ

What does NIK's viral post say?

It condenses the mathematicians' version into 11 lines: a year of Navier–Stokes work with drafts in Codex, rumor of an AI solution on the same approach, an unanswered question on session training. Four lines hold against the documents, the rest is partial or open.

Who is NIK (@ns123abc)?

A pseudonym active since 2019 commenting on AI news, with about 76 thousand followers at snapshot. Nothing publicly verifiable exists about the physical person: evaluate the account as an amplifier, not a witness.

Is the post accurate?

Partly: the year of work, the Codex drafts, the 3 September email and the training question check out; the "solved Navier–Stokes" framing is misleading, and the denials that arrived after publication are missing.

Did OpenAI use Buckmaster's Codex data?

Unestablished. Three channels must be kept apart: training unanswered on record, lookup denied, human access denied by OpenAI. Policy makes training abstractly possible on individual plans unless the user opts out.

What is missing from the post?

OpenAI's official denial, Bubeck's technical points, the press-conference on-record details, Tao's Euler-only validation and the Clay path. The post is frozen at 11:34 AM on 8 September.

Who are konstiwohlwend and rynorhn?

Two secondary amplifiers: the first a builder with a scientific-competition background who signs the most careful summary, the second an AI commentator with an openly partisan tone. Both useful, neither neutral.

Primary and secondary sources / Fonti

Primary and direct evidence — @ns123abc post 08/09 11:34 AM (direct fetch, 1.4M views); X profiles @ns123abc, @konstiwohlwend, @rynorhn, @OpenAI (direct fetches); Buckmaster statement (NYU PDF); OpenAI help center on training and Codex (articles 5722486 and Codex FAQ). Attributed reporting — New Scientist 08/09 (sprint figures and conference denials), Axios, Scientific American, TechCrunch (Codex policy and opt-out), GadgetsNow, OfficeChai, Decoder, 4SAPI. Mathematical context from our analysis: OpenAI and Navier–Stokes: what was actually proven and Italian version of this fact-check.

More on our method: AI for SMEs (in Italian) and how we structure citable content (in Italian).