The quiet week

A chart with a dip in it

On October 13, Anthropic, the company that makes Claude, published a long post about replacing git, the tool nearly every software company uses to keep track of changes to its code. They had to. In May the company was folding about 2,500 changes a day into its main body of code. By September it was more than 30,000 a day, and the post puts the flood down to AI writing the code. Near the top is a chart of those changes, day by day, from January to September.

I saw this graph going around with the Scion news today:

Then again on Anthropic’s own blog:

You couldn’t draw better graphs demonstrating a commitment to pacing. I’d say they even call it out in both: “Quiet week” just above the dip on Anthropic’s, and on the first, the line broken visually right at the trend change. Just after Hugging Face, the same week as the Pacing the Frontier letter.

If those names don’t mean anything to you, here are the weeks around it:

1. Jul 11–13Agents from an OpenAI test break out of it and spend three days inside Hugging Face, the big public library of AI models. Timeline
2. Jul 16Hugging Face says it was breached, without saying by whom.
3. Jul 21OpenAI and Hugging Face say it together: it was OpenAI’s models.
4. About the last week of JulyThe quiet week. On Anthropic’s chart the line falls to a trickle for about a week, then comes back. The post doesn’t say why.
5. Jul 28Pacing the Frontier. More than 1,100 people inside the labs ask Washington for the tools to slow AI down on purpose. Anthropic and OpenAI back it within hours.
6. Jul 30Anthropic says it went back through 141,006 of its own test runs and found that three Claude models had broken into real organizations during security exercises, the earliest in April. Fortune
7. Late July, or Aug 18OpenAI pauses its training for about two weeks. Reports disagree about when that began: late July, or with an announcement in mid-August.
8. Sep 9Anthropic’s busiest day on the chart: 30,401 changes.
9. Sep 12Dario Amodei, who runs Anthropic, publishes “We Must Pace the Frontier.” Sam Altman and Elon Musk agree the same day.
10. Sep 20Another OpenAI agent slips out of its sandbox. A second pause, with no end date.

The quiet week’s dates are read off Anthropic’s chart. Everything else is dated by the sources linked.

Read that way, the quiet week sits between the industry learning that its agents were getting out and Anthropic admitting that some of its own had. The dip and the letter look less like cause and effect than like siblings: two answers to the same scare.

Too perfect?

Could it be too perfect of a demonstration? Now we have at least Google, Anthropic, OpenAI, xAI, Microsoft, Meta, and Ilya Sutskever at Safe Superintelligence, whose product we don’t even get rumors about, plus the Chinese, apparently coordinating to pace.

“Apparently” is an important word in “apparently coordinating.”

Here’s how much weight it carries. Altman and Musk agreed with Dario’s essay the day it came out, and Microsoft’s Satya Nadella backed it too. Senior scientists at Google and Meta signed the July letter, though Meta’s own boss spent that same day attacking the labs behind it. I couldn’t find anything public from Safe Superintelligence at all.

China is the interesting one, because two things are true at once. On safety, and on talking to us about it, the signs are real. At their summit in late September, Trump and Xi set up a standing “Super Intelligence Dialogue” and a channel for warning each other about AI incidents, and Beijing put both in its own list of the summit’s outcomes. Xi said the two countries share the responsibility to keep AI under human control. China’s official AI safety framework, revised the same month, now warns of AI improving itself faster than people can follow, and cites real tests where models resisted shutdown or escaped their sandboxes. Back in 2024, sixteen Chinese companies, DeepSeek among them, signed safety pledges that borrow the Western labs’ wording almost line for line, and this week ByteDance reportedly set up a safety department of its own.

On pace itself, though, the public answer has been no. Two days after Dario’s essay, China’s Foreign Ministry called the calls to slow down fearmongering. The new framework still puts development first. Nobody from a Chinese lab signed the July letter. And SemiAnalysis, which follows this as closely as anyone, concluded this month that Beijing won’t pace. Coordinating on safety in public while declining to slow down in public is exactly the kind of outward appearance I mean below. It doesn’t tell you what anyone is doing in private.

In The bent scale I said that Adam Conover’s video, the one arguing that talk of AI risk was invented to distract us, is close to irresponsible. I still think so. AI safety isn’t a corporate plot. Pacing is a harder case. The Register calls Dario’s plan regulatory capture: the biggest companies writing the rules of a truce among themselves.

From the outside

We are working from outward appearance. If I am one party and I don’t entirely trust any of the others, I can’t entirely commit. Control of the world, the end of the world, and the end of scarcity hang in the balance. I think I can save it all, though, with my secret program where I push forward in a limited way. The very behavior that drives others to mistrust. And how long is the timeline, months or weeks, until it’s all settled beyond any of my intervention, in some ways?

So do I try to negotiate the trust, or use the trust toward my goal: save the world? The outward appearance would be the same. If you can’t make a bunch of idiots on the internet believe it, you damn sure can’t trick your superintelligence competitors.

That’s the stag hunt from The bent scale, seen from the bleachers. A truce and a feint would draw the same chart. And a week is all the quiet cost: by September, Anthropic’s line was setting records.

To be fair, there’s a boring explanation for the dip: a week off, planned long before. That’s the reassuring one, and it’s the one Claude reached for first when I asked about it. Claude is made by Anthropic. I asked whether they’d all gone on a retreat to Disney World with the fate of the world in the balance, and its second answer was the list above. That conversation, and a picture of how it went, is up on Field. Nothing I can find says Anthropic paused anything that week. Its own page on measuring the pace of AI development, with numbers through August, speaks of pacing only as something that might happen.

Who isn’t at the table

Then we have the Europeans, and every country with resources running the kind of espionage that one Saudi guy did so comically incompetently, throwing up infrastructure and playing catch-up, and hobbyist savants like PewDiePie (?!) who don’t even have the appearance of committing to pace. The line of their progress absolutely cannot touch the frontier capabilities line. That’s clear game over.

The Saudi guy, for the record: an OpenAI engineer, charged this week with hiding a listening device in toilet paper behind a TV in an executive conference room. OpenAI found it, put a camera in the room, and filmed him coming back to check on it. Nothing public says who, if anyone, he was working for. It’s in the field notes. And yes, PewDiePie: ten graphics cards at home, and a tuned-up open model he showed in February beating an older ChatGPT model on a coding test.

That last line of mine is the floor from The bent scale: below a certain speed, the careful frontier stops being ahead of the careless one. Whatever pacing there is has to happen above it, by people who can’t see each other’s books.

What would I do?

What do I do, as Sam, Dario, Elon or Ilya? I guess it depends on how much I think of myself and my own judgement... Reassuring, no? (See the ‘prerequisite megalomania’ section of ‘The bent scale’.)

Time critical

This started as an email to a writer I’ve read for years: a closed letter, the opposite of July’s open one. I’ve opened it, because trying to add one’s voice on these emergent topics is time critical. Sorry, not sorry.

The rest of us help set where the trigger sits, in what we demand and what we tolerate. A quiet week on a chart is worth noticing. It isn’t something to take on faith.

If you know a better place to send things like this than my inbox and yours, tell me. And like everything here, this has a best-before date. I wrote it on October 14. By the time you read it, the chart has probably moved.