The Request to Slow Down AI Development

In March 2023 an open letter asked the AI laboratories to pause for six months. More than a thousand people signed it. Nothing paused.

In September 2026 the request came from inside the buildings. Anthropic's chief executive published an essay, We Must Pace the Frontier, arguing that the companies must slow the rate at which they improve model capability. Within a day OpenAI's chief executive agreed publicly. Within four days the rest of the industry had publicly disagreed with each other about how.

This film is about that disagreement, and about the thing underneath it that almost nobody is discussing: every proposal on the table requires somebody to be able to measure how fast this is going, in units, to a precision that would survive a dispute.

First, what the essay actually proposes: embedded third-party evaluators with what it calls ongoing, employee-like access; coordination on standards and rate limits among frontier companies in democracies; and coordination with authoritarian governments, which the essay itself does not present as easy. Its four stated quantities — 5 to 10 years to cure most major diseases, 6 to 12 months for a rogue-agent swarm to threaten the internet, a 3 to 5 year geopolitical window, 1 to 2 years of available alignment progress — are shown as ranges, because all four are ranges in the original, and are labelled forecasts rather than measurements.

Then the split, as reported by the Associated Press on 16 September 2026. Demis Hassabis proposing a standards body modelled on FINRA. Satya Nadella on human control. Mustafa Suleyman on embedded evaluators and broader representation. Elon Musk on competitors reviewing each other's models. And against a coordinated slowdown: Mark Zuckerberg on individual liability, Jensen Huang on market forces — both stated at full strength, because these are arguments about mechanism rather than about whether harm matters.

The criticism that the essay is a business strategy rather than a safety argument, made by Chamath Palihapitiya, is in the film unanswered by the narrator.

Then the measurement. METR's Time Horizon 1.1, published 29 January 2026, puts the highest measured 50% time horizon at 320 minutes for Claude Opus 4.5 — with a confidence interval of 170 to 729 minutes, a top end more than four times the bottom. Doubling times of 196.5 days across the whole trend, 130.8 days since 2023 and 88.6 days since 2024: three windows, three answers, published together. Of the 31 longest tasks in the 228-task suite, five have a measured human baseline.

You cannot enforce an agreement to halve a rate when compliance and violation sit inside the same error bar. That is not an argument against pacing. It is a description of what the first year of any serious pacing regime would actually consist of, which is not slowing anything down — it is building the instrument.

And the one rule already written against a number: the EU AI Act's 10^25 FLOP threshold, set against Epoch AI's projection (30 May 2025) of models above 10^26 FLOP — about 1 in 2025, about 10 in 2026, 10 to 80 in 2027, and 61 to 528 by 2030. A fixed threshold in a world where the quantity compounds names one system in its first year and dozens within three, without a word of it changing.

Disclosure: Anthropic's chief executive wrote the essay this film is about, Anthropic makes the model used to produce this film, and an Anthropic model is the one named in the METR figures quoted. That is stated on screen, and the criticism of the essay is stated alongside it.

This film reaches no verdict on whether AI development should be slowed. It asks a question that can be checked instead: what would each proposal have to be able to measure, and can anybody measure that yet.

Educational documentary. Not financial or investment advice.