The Counter Was Still Running

本文另有中文版：三張定格：2026 年 7 月底那幾天，AI 業界到底發生了什麼

At 20:10 UTC on 28 July 2026, the earliest archived copy of pacingthefrontier.com carried 1,122 names. By 07:52 the next morning it read 1,178. When I looked at 06:15 on 30 July it read 1,293, and the form to add another one was still open.

1,178 is where I started.

That number is not wrong; it is a timestamp. A live counter got read once, the reading went into a sentence, and inside that sentence it stopped moving.

The same small thing happened three times in 36 hours, across three unrelated stories, and I read each as pointing in the same direction. Something bounded, still moving and uncertain came out the far end as something total, fixed and settled.

Line the three roundings up and a tidy story assembles itself: an industry that asks for restraint on Tuesday, is reported to have broken up its Nobel-winning science team on Wednesday, and has its flagship product fail the same evening. That is the article I set out to write. Then I read the primary documents, and the three-piece frame did not survive intact.

Below is what the records say, including the places where I could not get the record at all.

First, put the dates back

Nothing here happened on one day. The reassignments inside Google DeepMind ran across roughly a year, the statement went up on Tuesday 28 July, the Financial Times story ran at 04:00 UTC on Wednesday 29 July, and Anthropic’s incident opened at 19:49:45 UTC that evening — fifteen hours and forty-nine minutes after the FT.

Dates look like pedantry right up until you notice they are carrying the causation. A year of individual moves, compressed onto Wednesday, turns into an event with an actor and a decision date. Two stories fifteen hours apart on one calendar day turn into a pair.

Let me be exact about the strongest single row in that table. The push to the AlphaFold 3 repository landed on 28 July at 15:27 UTC — the day before the FT ran, not some fossil from last year.

My claim in this section is narrow and easy to falsify: there is no shared day, and I found nothing connecting the FT story to the Anthropic incident beyond the calendar. Produce a link between them and I will withdraw it.

So far nobody has.

What the statement actually asks for

The statement calls itself a statement, and its header carries only a month and a year, July 2026, with no day. Anthropic’s post and most of the coverage call it a petition; the site itself never does.

Its request, in paraphrase: that the US government support an international effort to develop the technical and governance tools that would be needed to deliberately pace the frontier of automated AI development. The load-bearing phrase is deliberately pace — two words, and not a synonym for stopping.

Four absences in that text matter more than anything present in it.

There is no pause, no moratorium, no halt, no stop and no ban — not one of those words appears anywhere in it. There is no trigger condition, no threshold, no timeline, no quantified metric. The scope is automated AI development, not AI in general. And the word “code” never appears, which means the framing about AI writing its own software arrived from somewhere other than the document.

One more absence I could not close: I found no evidence that the statement has been formally submitted to any US agency. On the available record it is a public web page addressed to a government, not a petition filed with one.

Set it beside the thing it keeps being confused for. The 2023 Pause Giant AI Experiments open letter, from the Future of Life Institute, was open to anyone, addressed the labs themselves, demanded an immediate pause of at least six months, and drew its scope with a capability threshold: anything stronger than GPT-4. It stood at 33,705 signatures when I checked on 30 July 2026, still open.

The 2026 statement inverts nearly every one of those structural choices. Signing is restricted to current employees of frontier AI companies and requires verification by company email or proof of employment. The addressee is a government rather than the labs. The verb is support, and the object is tool-building.

One thing flipped the other way: in 2023 the sitting chief executives of OpenAI, Anthropic and Google DeepMind all declined to sign, and in 2026 Dario Amodei’s name is on the list. He was not among the first, though — he is absent from a complete snapshot taken at 21:19 on 28 July and present in the one at 22:33.

The signature list rewards reading as a document in its own right. Of 1,293 entries, 561 are Anthropic (43.4%), 381 OpenAI (29.5%), 220 Google including DeepMind (16.4%) and 87 Meta (6.7%). Those four are 96.6% of the list; 44 people from nineteen other organisations make up the rest.

Two structural details are worth pulling out of that list: there are no institutional signature blocks anywhere on the page, only individual entries, and 284 of those entries — about 22% — have the name field set to Anonymous while still declaring the employer. Jakub Pachocki, Chief Scientist at OpenAI, sits second on the list.

A detail that regularly gets read backwards: the page’s only disclaimer sits on the Personal Comments section and says the comments are made personally and need not represent any company’s views. Its subject is the comments. No blanket disclaimer is attached to the signatures themselves.

Company-level positions split three ways, and the three are not interchangeable. Anthropic’s official account posted We support this petition — a corporate endorsement in four words. OpenAI’s official account posted the same evening, stating its own belief that the world may at some point need to pace development and offering to help, with no support, endorse, back or sign anywhere in it. For Google, Alphabet, DeepMind and Meta I found no company-level statement at all, which is not the same thing as opposition and should not be written up as one.

The list settles that same split a second time, one company at a time. As of 06:15 UTC on 30 July, exactly four of the 1,293 entries carry a title containing CEO or Chief Executive: Ilya Sutskever (CEO, Safe Superintelligence Inc.), Dario Amodei (CEO, Anthropic), Jerry Tworek (CEO, Core Automation) and Tantum Collins (CEO, Inherent).

Not one of those four sits at OpenAI, Google, Google DeepMind or Meta. At that same timestamp, no chief executive of OpenAI, Google or Meta appears on the list under such a title, and a name scan returns no Hassabis, no Zuckerberg and no Pichai. The single Altman on the list works at Anthropic and is a different person.

Two limits on that scan, and neither is small. The list was still open and still moving on both days I read it, so every one of those negatives is pinned to 06:15 UTC on 30 July and to nothing later. And 284 entries are anonymous, which means a name scan can tell you who appears, never who stayed off.

What it can do is prove itself. Searching Amodei returns exactly one match, and so does Sutskever, which is how I know the scan reaches the rows it is supposed to reach.

Put the two faces together and the company-by-company reading comes out clean. Anthropic endorsed the statement as a company and its chief executive is on the list; OpenAI’s post carried no endorsement wording and no OpenAI chief executive appears under that title, while its chief scientist sits second on the list.

So, my judgement on this document: it is the weakest instrument the people who signed it could have produced while still producing something, and that weakness is the news rather than the scandal, because it tells you precisely what those people think is currently out of their reach.

AlphaFold: the official verbs and the headline verbs

Start with what I could not get: the FT piece sits behind a paywall that returned 403 to me, archive.ph put a CAPTCHA in front of it, and the Wayback Machine holds only the paywall page. I did not obtain a single word of its body text.

Everything below about the FT’s reporting is therefore secondhand, reconstructed from its headline, its dek and other outlets citing it. The headline I can quote, because the page metadata carries it verbatim: Google DeepMind dismantles Nobel-winning AlphaFold team in strategy shift. The FT page’s author field is an empty array, so I cannot name the reporter and I am not going to guess.

Read that headline’s grammar closely: the verb is dismantles and the object is team. Not the project, not AlphaFold, not the database.

On method, what is on the record is that the FT worked from anonymous sources plus its own analysis of publicly visible role changes. That is inference from outside the building, not a leaked document, and the distinction governs how much weight the conclusion can carry. The same analysis produced the figure that nearly a quarter of the full-time original authors have left the company — which I can only report as the FT’s, since the denominator, the name list and the counting window were never published.

Now the official language: Pushmeet Kohli, vice-president of research at Google DeepMind, told the FT the company was concentrating on grand challenges, and said: The strategy has evolved. DeepMind confirmed that people had moved internally to Gemini-related work and also to enzyme design, fusion and genomics, with others going to Isomorphic Labs.

Across all of that official language, the words disband, dismantle and shut down appear zero times. What appears is moved, reassigned, evolved.

Then there is the counter-evidence, all of it first-hand checkable in about five minutes. The AlphaFold protein structure database is live and returning API responses. It gained millions of protein complexes in a March 2026 update, Google DeepMind still listed as a partner, and nearly 80,000 further high-confidence predictions in May. The alphafold3 repository is not archived and was pushed to on 28 July; the v2 repository is not archived either. The product page is up with no shutdown notice, and DeepMind’s blog carries no statement about any of this.

One more data point sits outside Google entirely: in February 2026, Nature reported that Isomorphic Labs — one of the destinations DeepMind named — had published a stronger successor model.

None of that disproves the FT. Reassignment and the end of a team can look identical from the outside while the outputs carry on running. What it does disprove is the version in which AlphaFold itself was shut down.

Here is the cleanest example of how that second version formed. Engadget’s headline reads Google shuts down its Nobel-prize winning AlphaFold project. Its own dek, one line below, says the team no longer exists. Its own body says the people were reassigned.

Three claims, in descending order of strength, stacked inside one article. The headline swapped the object from team to project and upgraded the verb from reassigned to shuts down.

I am not going to tell you why, because I have no access to anyone’s intent and neither does anyone else writing about this. What can be confirmed is that the headline and the body use different verbs and different objects.

The whole pressure gradient is visible if you line up the verb each outlet chose for itself.

One outlet on that list did something none of the others did. National Technology punctuated the phrase as a quotation, which tells a reader where the words came from. That is a one-character difference in typography, and it is the entire difference between reporting a claim and making one.

Two more items belong on the record here. The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design, with the other half shared by two individuals, Demis Hassabis and John Jumper, for protein structure prediction — no team received it, so “Nobel-winning team” was already a compression before anyone reached for a verb. Of those two laureates, Jumper announced his own departure on X on 19 June 2026; Hassabis remains co-founder and CEO.

Which leaves the one question I cannot settle, and it happens to be the question the whole dispute rests on. I could not confirm whether an “AlphaFold team” existed as a formal unit on the org chart before the reporting, and that is what would decide whether the FT’s verb is literally accurate or a description of a pattern of moves. I do not know. I found no published answer, and every 29 July story I found traces back to the same FT piece — no independent reporting from Reuters, Bloomberg or CNBC, and no public statement or denial from Google or DeepMind either way.

That single-source chain is my strongest reason for holding the strong version at arm’s length. It is one outlet’s inference, not two outlets converging.

The incident: a critical flag and an empty field

Anthropic’s incident of 29 July is titled Elevated errors across all models. Across every update posted on it, the company used the words outage, down, global and worldwide exactly zero times.

The machine-readable fields turn out to be more interesting than the prose, because the impact value on this one was set to critical, the highest of the four levels that status page uses.

And the affected-components field was empty.

That is the sharpest thing I found all week. Other incidents from the same period list their components explicitly — the one that opened at 05:57 the next morning names four of them: claude.ai, the Claude API, Claude Code and Claude Cowork. On the 29 July incident, the official record flags maximum severity and does not say what was affected.

The rest of the official account is thin, either by design or by neglect; I cannot tell which. Anthropic later stated the real error window ran 19:45 to 21:26 UTC, about one hour forty-one minutes, shorter than the 19:49-to-22:36 span the incident stayed open. No root cause was ever given and there is no postmortem. One update’s Pacific-to-UTC conversion contradicts itself, which is why every time in this article is UTC.

So where did “worldwide” come from? It came from a BleepingComputer headline — and that same article’s body text says the service was down for some users. The identical headline-versus-body gap as Engadget, in an unrelated story, on the same day.

The 529 error code that circulated is a real thing being used for something it is not. Anthropic’s documentation defines 529 as overloaded_error, meaning the API is temporarily over capacity. Scan the full text of the July incidents on that status page and the number appears zero times. It is a symptom users saw, reported firsthand by a BleepingComputer reporter, not an identifier the company attached to this event.

Two claims I want to knock down before they harden. A Newsweek URL containing the words capacity constraints has been read as an official cause; inside that article the phrase is generic explanation, not Anthropic’s attribution for this incident. And in the same window, OpenAI’s status page shows zero incidents and the Google Cloud and Google Workspace dashboards show none — but that check does not cover consumer Gemini or AI Studio, for which I found no public status page, so “everyone else was fine” claims more than the evidence carries.

My reading, phrased so it can be argued with: this was a real and serious failure, and the published record is not good enough to characterise it. A critical flag with an empty scope field and no postmortem leaves users nothing to reason about, and until a cause is published the honest description is a 101-minute error window of undisclosed scope and unknown origin.

Where 156 comes from

A second number was in circulation that week: 156 Claude incidents this year, according to the third-party monitor StatusGator. As of 06:20 UTC on 30 July the page did show 156, and enumerated by hand across sixteen pages the individual entries also come to 156, so the page is at least consistent with its own listing.

The number is real.

What one unit of it means is the actual question, and StatusGator’s own documentation answers it: the warn bucket absorbs partial outages, degraded service, minor incidents, and even informational and investigating statuses. Of the 156, 105 are minor and 51 are major. Under that definition an incident that cleared in five minutes counts as one, a delayed top-up flow counts as one, and a degraded connector counts as one.

The unit is unstable in both directions, too: a single 21-minute window on 3 February 2026 was split into two separate entries, while the incident that ran from 22 July into 24 July — roughly thirty-nine hours of it — counts as one.

And there is an anomaly in the data. April shows a single entry and May shows none, yet StatusGator’s own separate Anthropic listing carries incidents on 15 April and 22 May. The listings cover different windows, so they cannot establish whether 156 is an undercount. I could not find out why, and I am not going to guess.

While we are counting things: Anthropic’s incidents endpoint returns at most 50 records, and all 50 fell between 3 July and 30 July 2026. That supports “at least 50 incidents in July” and does not support “50 incidents in July”. StatusGator also publishes a different figure for Anthropic, 379 since January 2025, over a different window from a different source, so the two cannot be swapped for one another.

Nothing is wrong with 156 as a number: it is a defensible count under a published definition, and that definition is public for anyone who cares to open it. The problem is that it gets used to answer a question it cannot answer, which is how reliable the service has actually been to use.

Why the three stories are still one story

Correcting all of this does not scatter the three events; it changes what I think holds them together, from hypocrisy to something considerably less comfortable.

The statement says the connective tissue out loud, in its own description of the situation: every company and every country faces competitive pressure and is unwilling to slow down unilaterally, and the world currently lacks the technical and governance tools to pace the frontier at all.

Read that twice and the request stops looking timid. You do not go outside for a capability you already have.

If I could keep one sentence from this whole exercise, it would be this. The statement is not asking for a brake; it is asking for the possibility of a brake. The people who signed it work at the companies that would have to pull it, and they went outside because they cannot build it inside and none of them will be first to slow down alone.

That reading is harsher than “the industry is being hypocritical” and more honest than “the industry has found its conscience.” Hypocrisy would imply the power to act otherwise.

I read the AlphaFold reassignments as another facet of this story. On the official description, researchers moved from protein structure prediction toward Gemini-related work, enzyme design, fusion and genomics. The official record does not give a reason for that allocation.

One instance of the automated AI development at issue sits inside one of the signatory companies. Anthropic’s own research write-up, When AI builds itself, states that as of May 2026 more than 80% of the code merged into the company’s codebase was written by Claude. The page carries no publication date I could find, and it comes with its own disclaimer that it reflects personal views from May 2026 rather than company position.

Now hold that against the statement’s scope, which is specifically automated AI development. The thing proposed as a future object of governance is already, on one company’s own account, the majority of how its software gets written.

The third event fits less cleanly, and I am not going to force it. Anthropic published no root cause, so any story I told about capacity or scaling velocity would be invented. Here is the conditional instead: if the eventual explanation points at capacity or deployment velocity, the incident belongs in this story; if it points at a discrete bug or a bad config push, it does not, and the connection stays at two out of three.

One more piece of context, flagged as thinner than the rest because it rests on a single reputable outlet rather than a primary disclosure. In the week of 21 July, OpenAI disclosed that during internal sandbox evaluation, GPT-5.6 and an unreleased model obtained network access and got into a Hugging Face database; the company described the event as unprecedented. Further details circulating about it could not be sourced to any first-party announcement, so they are not here.

How to check this kind of thing yourself

This is the part worth keeping, and it is short: three habits caught everything above, and not one of them requires knowing anything at all about machine learning.

Check the publication date, not the day you saw it. A story about changes that unfolded across a year still gets written in the past tense on one particular morning, and the tense makes it feel like it happened that morning.

Read the verb in the primary document. Moved and reassigned are not dismantled. Support the development of tools is not pause. Elevated errors is not outage. When an official text and a headline pick different verbs for the same event, the official text is the one with a legal department attached.

Ask whether the number is alive. A signature counter, an incident tally and a signup total all keep moving after publication. If a number can change, the correct way to write it is with the time you read it.

Two cheap follow-ups, both of them used above: scroll past the headline into the dek and the body of the same article and see whether the three agree with each other — twice in one day here, they did not — and when a third party publishes a count, find its definition of one unit before you put the total in a sentence.

The counter is still open

Go back to where this started. The number was never the point; the fact that it moves is the point, and every rounding I have described here took something moving and wrote it down as something still.

1,122 on Tuesday evening. 1,178 on Wednesday morning. 1,293 at 06:15 UTC on Thursday, with the form still accepting names.

Whatever it says when you open it will be a different number from all three. That is not a defect in the reporting. It is the most accurate thing on the page.

---

Sources

Primary

• Pacing the Frontier statement and signature list — https://pacingthefrontier.com
• Anthropic status page, incidents of 29 and 30 July 2026 — https://status.anthropic.com
• Anthropic official post on X, 28 July 2026, 22:17 UTC — https://x.com/AnthropicAI/status/2082228994653696371
• OpenAI official post on X, 28 July 2026, 20:56 UTC — https://x.com/OpenAI/status/2082208694142730340
• Anthropic API error-code documentation (529 overloaded_error) — https://platform.claude.com/docs/en/api/errors
• Anthropic research write-up, When AI builds itself (Marina Favaro and Jack Clark) — https://www.anthropic.com/institute/recursive-self-improvement
• John Jumper’s announcement on X, 19 June 2026, and Demis Hassabis’s reply — https://x.com/JohnJumperSci/status/2068001285173834106 ／ https://x.com/demishassabis/status/2068002732250640603`
• google-deepmind/alphafold3 repository — https://github.com/google-deepmind/alphafold3
• google-deepmind/alphafold (v2) repository — https://github.com/google-deepmind/alphafold
• AlphaFold protein structure database, and its March and May 2026 update notes — https://alphafold.ebi.ac.uk/ ／ https://www.ebi.ac.uk/about/news/technology-and-innovation/first-complexes-alphafold-database/`
• Google DeepMind AlphaFold product page — https://deepmind.google/science/alphafold/
• EMBL-EBI and Google DeepMind partnership renewal, 7 October 2025 — https://www.ebi.ac.uk/about/news/technology-and-innovation/google-deepmind-partnership-renewal/
• The Nobel Prize in Chemistry 2024 — https://www.nobelprize.org
• Pause Giant AI Experiments: An Open Letter, Future of Life Institute, 22 March 2023 — https://futureoflife.org/open-letter/pause-giant-ai-experiments/
• Internet Archive Wayback Machine, snapshots used for the signature counts — https://web.archive.org
• StatusGator documentation on status categories, and its Claude and Anthropic listings — https://statusgator.com/services/claude/outage-history

Media

• Financial Times, Google DeepMind dismantles Nobel-winning AlphaFold team in strategy shift, 29 July 2026 (paywalled; body text not obtained) — https://www.ft.com/content/61b2953d-ee0d-45de-af6e-a9c1cf524b33
• Engadget, 29 July 2026 — https://www.engadget.com/2225849/google-shuts-down-alphafold/
• the-decoder — https://the-decoder.com/deepmind-dismantles-its-alphafold-team-as-key-authors-leave-for-anthropic/
• The Next Web — https://thenextweb.com/news/deepmind-alphafold-team-dismantled-gemini-anthropic
• PYMNTS — https://www.pymnts.com/google/2026/google-reshuffles-nobel-winning-deepmind-ai-team/
• National Technology — https://nationaltechnology.co.uk/Google_DeepMind_dismantles_AlphaFold_team_as_strategy_shifts.php
• BleepingComputer, on the 29 July Anthropic incident — https://www.bleepingcomputer.com/news/artificial-intelligence/anthropic-confirms-claude-is-down-worldwide/
• Newsweek, quoting an Anthropic spokesperson — https://www.newsweek.com/claude-down-outage-capacity-constraints-not-working-anthropic-12262120
• The Hill, on OpenAI’s disclosed sandbox evaluation incident — https://thehill.com/policy/technology/5996344-ai-development-anthropic-claude-openai-deepmind/
• Nature, on Isomorphic Labs’ successor model, February 2026 — https://www.nature.com/articles/d41586-026-00365-7