On Saturday 12 September, Dario Amodei published a 3,800-word essay asking the AI industry to move more slowly. Hours later, Sam Altman wrote on X: "I agree with Dario that we need to pace the frontier". Elon Musk summed it up in three words: "Dario is right". Demis Hassabis joined the next day.
The companies pouring hundreds of billions into winning this race are now saying everyone should run a little slower. That almost never happens.
Forty-eight hours later, governments said no.
What they actually signed up to
It is not a pause. Amodei spells it out: pacing "does not mean halting model training or technical progress", but making sure companies take enough time to align and safeguard what they ship. And he adds a line that changes the whole reading: progress will still seem fast.
Of the three steps in the plan, only the first is under way: third-party evaluators working inside the companies, with employee-level access and the right to publish without editorial control. Anthropic is committing unilaterally and OpenAI said it will do the same. Coordination among democratic countries, and then globally, is intent rather than commitment.
The agreement is also thinner than it looks. Hassabis backed the direction but proposes something different: an industry-funded, government-supervised standards body along the lines of FINRA, examining models 30 days before release. Amodei wants oversight inside the company and during development; Hassabis wants it outside and before launch. They agree on the headline and disagree on who pays for supervision, who carries it out and when.
What really worries Amodei is recursive self-improvement: models helping to build the next generation of models, accelerating the cycle from within. The essay admits this is already happening across the industry, Anthropic included. The goal is to buy a year or two of margin.
How governments answered
Trump rejected it that same weekend with a line that sums up his position: "whoever wins AI, wins". His argument is that the United States is ahead and that slowing down would hand over the lead.
China replied on Monday. Foreign ministry spokesperson Guo Jiakun called the framing fearmongering: "fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance and serve the interests of no one".
It is worth understanding why Beijing reacted that way, because the coverage stopped at the headline. Amodei did not only call for slowing down. In the same essay he called for keeping restrictions on advanced chip sales to China and warned that a Chinese lead in AI would be "a grave danger for the United States and the world". Everyone slows down, and China stays blocked on top of that.
That is the knot. The technical proposal and the geopolitical position travel in the same text, which makes them very hard to separate. Trump and Xi meet on 24 September in Washington with AI governance on the agenda.
Meanwhile, one gesture carries more weight than any statement: OpenAI has ruled out going public this year, citing the safety work ahead. Altman called it an ill-advised moment for a listing. Postponing one of the decade's largest IPOs is not a tweet.
The year we argued about whether we had already arrived
In March, Jensen Huang said a line on the Lex Fridman podcast that still echoes: "I think it's now. I think we've achieved AGI". He was not alone. Altman had spent months saying OpenAI knew how to build it, before later admitting the term had become "very imprecise".
The technical response was close to unanimous. Google DeepMind described an uneven cognitive profile: models beat humans at mathematics and factual recall, and lag behind on learning from experience and social understanding. Hendrycks and Bengio scored the most capable model below 60%. And on ARC-AGI-3, François Chollet's abstract reasoning test, frontier models score under 1% where a person is around 100%.
We are not at AGI. But one curve really has accelerated, and it is the one that matters.
METR measures how long a task an agent can complete on its own with reasonable reliability. That horizon used to double every seven months. In the most recent stretch, every three or four. In May, an agent sustained more than two hours of expert human work. Meanwhile success on OSWorld, which measures operating a computer the way a person would, went from 12% to around 66%.
Agents stopped being a two-minute demo and started chaining hundreds of steps across hours. That is the real leap of 2026, more than any single model.
Why we think this is good news
Our view has nothing to do with safety or geopolitics. It is about the gap between what the technology allows and what companies actually do with it. That gap keeps growing.
In six months everything changed at once. Agents working alone for hours. Image and video at genuine delivery quality. Long context, computer use, protocols for connecting tools. Each release able to unseat the last one overnight.
And something perverse happens: when the tool changes every four weeks, you never master any of them. The new thing gets tested before the previous one has been squeezed. Workflows pile up on top of versions that no longer exist. And two different things get confused, having access to the latest tool and knowing how to get everything out of it.
Most organisations are nowhere near using what AI already allows today, and not for lack of models. What is missing is time to settle the method, embed it in real processes and learn where it breaks.
That is what a calmer pace buys. Not less capability, more depth on the capability that already exists. Time for judgement to catch up with power.
The trap would be reading this as permission to wait. With governments saying no, the real slowdown will be far softer than the headlines suggest. Progress will still seem fast.
When everyone has the same models for the price of a subscription, what separates a good result from a mediocre one is not the engine. It is knowing what to ask of it and recognising when it got things wrong. That does not update itself every February.
At GROS we work exactly there: visual production with AI, directed by people. From brief to master. Days, not weeks.
More on our approach to AI audiovisual production.
Companies want to slow down and governments do not. While that gets resolved, what models already allow today remains far ahead of what almost anyone is using.
Sources
- Dario Amodei, We Must Pace the Frontier: https://darioamodei.com/post/we-must-pace-the-frontier
- TechCrunch, Anthropic CEO outlines plan to pace the frontier: https://techcrunch.com/2026/09/12/anthropic-ceo-outlines-plan-to-pace-the-frontier/
- NPR, Beijing hits back at Anthropic CEO's call to curb China's AI development: https://www.npr.org/2026/09/14/nx-s1-5968456/china-hits-back-ai-development
- Infobae, Los cuatro jefes de la IA ya coinciden en frenar: ninguno propone lo mismo: https://www.infobae.com/estados-unidos/2026/09/13/los-cuatro-jefes-de-la-ia-ya-coinciden-en-frenar-ninguno-propone-lo-mismo/
- Fortune, Sam Altman confirms OpenAI won't go public this year: https://fortune.com/2026/09/12/sam-altman-openai-ipo-delay-ill-advised-moment-safety-concerns/
- Fox News, Trump says Xi Jinping will visit the US on Sept 24 to discuss AI: https://www.foxnews.com/politics/trump-says-chinese-president-xi-jinping-visit-us-sept-24
- Fortune, Nvidia's Jensen Huang says 'we've achieved AGI': https://fortune.com/2026/03/30/agi-definition-jensen-huang-lex-fridman-deepmind-turing-text-cognitive-taxonomy/
- METR, Measuring AI Ability to Complete Long Tasks: https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
- Stanford AI Index 2026 (OSWorld, computer use): https://aiindex.stanford.edu/