ATAILA Newsroom · Budapest ·
2026-W38
The people building it say the pace is unsafe. That is a procurement fact before it is a philosophical one.
On 8 September a 27-year-old researcher named Jacob Coxon resigned from Anthropic and wrote that neither Anthropic nor OpenAI, the two companies he had worked for, is acting responsibly. Within a week the story had run on every Hungarian news portal we watch. All of them asked the same question: does this end with humanity dying? It is a fair question and we are not qualified to answer it. We want to ask the much smaller one that nobody asked, because it is the one with a decision attached.
What actually happened
Jacob Coxon is a 27-year-old British researcher who worked on pretraining at OpenAI and then at Anthropic. On 8 September he announced his resignation. In his words, quoted by TIME: "I resigned from Anthropic today...Neither company is acting responsibly." He left two months before his equity would have vested.
The reaction inside the industry was not dismissal. Anthropic's CEO Dario Amodei published a 3,800-word essay on 12 September arguing for pacing the frontier. Evan Hubinger, who leads alignment science at Anthropic, publicly agreed with Coxon and put the chance of AI killing all humans within a decade at greater than ten percent.
We are an infrastructure company. We have no standing to adjudicate whether those probabilities are right, and we are not going to pretend otherwise.
153 million
views of Coxon's resignation posts within 36 hours, per TIME, 15 September 2026
3,800 words
Dario Amodei's essay of 12 September calling for pacing the frontier
greater than 10%
chance of AI killing all humans within a decade, per Anthropic alignment lead Evan Hubinger
2 months
of unvested equity Coxon forfeited by resigning when he did
The Hungarian coverage was total, and uniform
This is the part that interested us most. A foreign personnel story does not usually travel. This one did. Inside a week it ran on Portfolio, Telex, Index, HVG, Economx, Forbes.hu, Qubit and Bitport — eight of the portals we read every week, several of them more than once.
And the framing was the same everywhere: extinction, 2030, percentages, whether the industry should slow down. Telex covered Obama's concern and Trump's dismissal. Qubit covered Senator Sanders' bill to ban superintelligence development. Economx covered the split between Zuckerberg on one side and Altman, Musk and Amodei on the other.
Every one of those is a legitimate story. But a Hungarian company that spent this summer wiring its document review, its customer correspondence or its accounting workflow into a US model API read eight versions of are we all going to die and zero versions of what, if anything, you should do differently on Monday morning.
That gap is why we are writing.
Where this argument does not help us
Before the argument, the concession, because it is a real one and leaving it out would be the dishonest move.
Coxon is not making our case. His concern is capability racing — labs pushing toward self-improving systems faster than anyone can verify they are controllable. That is an argument about what is built. Ours is an argument about where data sits and who can reach it. They are different problems, and a private deployment of an open-weight model does nothing whatsoever about the first one. If frontier capability is genuinely dangerous, running a smaller model in a Hungarian rack does not make the world safer.
Taking an extinction-risk story and landing it on a hosting invoice is the kind of move that deserves the eye-roll it gets. So we are not going to make that move.
The narrower point, which we do stand behind
Strip out the existential question entirely and something ordinary is left, and it is a procurement fact.
A supplier whose own senior researchers publicly describe its roadmap as unsafe is a supplier you cannot plan around. Not because the supplier is bad, but because its direction is genuinely contested inside its own walls. In the same fortnight: a researcher resigned over the pace, the alignment lead agreed with him in public, the CEO called for the whole industry to slow down, and a US senator introduced a bill to ban a category of development outright.
None of that tells you the technology is dangerous. All of it tells you the terms, the availability and the direction of that service are less predictable than a normal enterprise dependency, and that the predictability is not in the vendor's sole control either.
The remedy for an unpredictable dependency is not to avoid it. It is to make sure you can leave. Where the model runs, whether the weights are yours to keep, whether the data ever left your jurisdiction, and whether there is a documented way out — those are the questions that change what happens to you if the direction changes again next quarter.
What we would actually ask
If we were sitting with a Hungarian company this week, the questions would be dull and answerable:
- Which processes now have a hard dependency on a single external model API, and what happens to each one if its terms, price or availability change at 30 days' notice?
- For each of those, does the work involve material you would not email to a competitor? If so, whose jurisdiction is it in right now?
- If you had to move to a different model in one quarter, what would that cost, and does anyone own that plan?
Most companies we meet cannot answer the third one. That is the gap worth closing, and it was worth closing before anybody resigned.
Sources, and what we could not read
The primary source for this piece is The AI Tipping Point, Billy Perrigo, Harry Booth and Naomi Nix, TIME, 15 September 2026, which we opened and read in full. The direct quotation, the 153 million figure, the 3,800-word essay, the Hubinger probability and the forfeited equity all come from it.
Hungarian coverage referenced: Portfolio (09-09), Telex (09-10), Index (09-12), HVG (09-13), Economx (09-16), Forbes.hu, Qubit (09-10) and Bitport.
What we could not read, and it is worth saying out loud. CNN covered this story heavily all week, including an Anderson Cooper interview with Coxon. Every CNN URL returns HTTP 451 Unavailable For Legal Reasons to a reader in the EU, through both an automated fetch and an ordinary browser. We therefore quote no CNN sentence anywhere in this article; we know only its headlines and dates. A US news organisation making itself unreadable in Europe is a small, sharp illustration of the thing we write about every week: whose infrastructure you are on determines what you are allowed to see.
Reported view counts for Coxon's posts differ between outlets and by the hour — 145 million, 153 million, later "more than 170 million". We quote TIME's figure with its date rather than the largest one available.
The original article
“The AI Tipping Point”
TIME, 15 September 2026
- Read it live at TIME
- ATAILA archive, captured 18 September 2026
The live page has changed since we captured it.
← Back to the Newsroom Press enquiries: contact us