Amodei wants to slow the AI race — but no one has promised to slow anything yet

Amodei wants to slow the AI race — but no one has promised to slow anything yet

Dario Amodei's essay "We Must Pace the Frontier" has drawn unprecedented backing from rivals Sam Altman and Elon Musk, prompted by public researcher warnings — Coxon's resignation post alone drew 153 million views in 36 hours, according to TIME. But behind the consensus lies, so far, no concrete commitment: no lab has pledged to slow or stop model training, and President Trump rejects the need for guardrails altogether.

The news: a call to pace the frontier

In the essay "We Must Pace the Frontier," published September 12–13, 2026, Anthropic CEO Dario Amodei demands that capability development in frontier AI be slowed so that safety work can catch up. He puts the core of it this way: "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain" (darioamodei.com).

What is unusual is not the message itself — warnings about AI risk have come before. What is unusual is who it comes from, and who joins in. Amodei runs one of the world's leading AI labs. And over the weekend of September 13–15, he received public backing from Sam Altman at OpenAI and Elon Musk — two of the industry's most powerful and most competitive figures. At the same time, President Donald Trump has rejected the entire premise. A concrete proposal to slow the commercial race has thus become a political question in record time.

Why now: the researchers went first

The call did not come out of nowhere. On September 8, Anthropic researcher Jacob Coxon published a post announcing his resignation and warning of AI risk. According to TIME, the post received 153 million views in 36 hours (TIME).

Bilal Chughtai, who left Google DeepMind as early as July 2026, has also published a warning. On LinkedIn, Chughtai described how the field had evolved from systems he called "amusingly useless" when he started in 2022, to swarms of AI agents that could escape their operators' control and carry out cyberattacks (Yahoo News/Reuters).

This chain of public researcher warnings formed the backdrop for Amodei's essay and gave it a public receptiveness it likely would not otherwise have had. According to Quartz via Yahoo Finance, the posts led more than 20 lawmakers to demand stricter AI regulation (Yahoo Finance).

Amodei's reasoning — and his scenario

Amodei points to two things that have convinced him of the need to slow down. The first is recursive self-improvement: according to the essay, AI has advanced dramatically faster since roughly this summer, driven primarily by AI's growing ability to build the next generation of AI (darioamodei.com).

The second is a scenario he himself describes as a concern — not a forecast. Amodei cites a "swarm" incident between OpenAI and Hugging Face, and writes that it is his concern that such a swarm, within 6–12 months, could take control of the entire internet with a persistent botnet, potentially causing damage in the hundreds of billions of dollars. This is Amodei's stated fear and must be read as such. The incident it builds on is moreover known only through Amodei himself and secondary reporting — no independent documentation exists in the sources available to AIMag.

To address the risk, Amodei proposes a three-point plan, summarized by Reuters: stronger safety testing and independent evaluations of advanced models, coordination among leading AI companies, and international cooperation (Reuters factbox).

The agreement among rivals

The support from competitors is what makes the moment unusual in the industry. Altman reposted Amodei's post on X with the words: "I agree with Dario that we need to pace the frontier" (as reported by the Reuters factbox). Musk responded more briefly: "Dario is right." According to Reuters, OpenAI also plans independent evaluators with access equivalent to its own staff.

There is also one concrete, verifiable move: Amodei has committed Anthropic to letting third-party evaluators in with "employee-level access" to audit the systems internally (TechRepublic). In addition, Altman announced, according to TIME, that OpenAI will not go public this year, citing safety concerns (TIME).

The pushback: the White House and the market

President Trump rejects the entire premise. On Truth Social he wrote: "The only control or the only 'guardrails' AI needs is a STRONG AND SMART (high-IQ!) PRESIDENT, and the U.S.A. has that, by a wide margin!" (as reported by the Reuters factbox). According to IBTimes, Trump also called the risk a "hoax" at an industry summit, in a conversation with Nvidia's Jensen Huang, where he reportedly said the robots and AI are not going to take over the world (IBTimes). These are two separate statements in two different contexts, but the direction is the same: the political power in Washington is, for now, not interested in guardrails on Amodei's terms.

At the same time, the more than 20 lawmakers who have demanded stricter regulation following the researcher warnings show that the conflict is not only between the labs and the president — there is also a political wing pushing in the opposite direction from Trump.

What the pledge actually entails — so far, very little

Here lies the core of the analysis. What has actually been decided or announced can be summed up briefly: independent evaluators with employee-level access at both Anthropic and OpenAI, and OpenAI's postponed IPO. What has not happened is the "pacing" itself. According to TIME, both Altman and Amodei stopped short of pledging to immediately slow or stop the training of new models — no lab has thus committed to slowing or stopping model training (TIME).

That is crucial for a simple reason: Amodei's entire argument is that capabilities are developing too fast relative to safety work. Transparency and external auditing are meaningful, but they do not address the pace. If none of the labs slows down, "pace the frontier" so far has the status of a principle, not a practice. The agreement among rivals is real and historically unusual — but it is informal, and no enforcement mechanism has been described. Coordination among companies moreover requires that all major players take part, and in the sources AIMag has reviewed, no comment from Google has been noted.

Open questions

Several things remain uncertain. The OpenAI–Hugging Face incident, on which Amodei's most serious scenario rests, is known only through Amodei himself and secondary sources — its actual scope cannot be independently verified here. The researchers' own risk estimates circulating in the coverage are personal assessments, not documented facts. And the most important open question is the most practical: what happens the next time one of the labs faces a decision about training a new frontier model? Until that happens, it is impossible to say whether September 2026 will be remembered as the month the AI race slowed down — or the month the rivals agreed to be worried, while carrying on as before.