Should AI Development Be Slowed Down? What Amodei, Altman, and Musk Agreeing On Something Actually Means

Dario Amodei, Sam Altman, and Elon Musk do not agree on things. These three run companies that exist to beat each other to the same finish line, and most of what they say in public about each other’s models is some flavor of dismissive. So when all three said, within the same 48 hours in mid-September 2026, that AI development needs to slow down, that was the actual news — not the content of the statement, which safety researchers have been saying for years, but the fact that it came from people whose companies profit from not slowing down.

It started with a researcher nobody had heard of. Jacob Coxon quit his jobs at both Anthropic and OpenAI and posted that the people building frontier AI privately believe it could “kill us all by the end of the decade,” and that they’re racing toward it anyway. The post got over 165 million views. Two days later Amodei published a response committing Anthropic to independent safety evaluators with “employee-like access” to its systems, and called on the rest of the industry to slow down with him. Altman agreed to adopt the same evaluator model. Musk reposted Amodei with three words: “Dario is right.”

Then, the following Sunday, from a golf resort in Ireland, the President of the United States said the opposite. “We’re leading China in AI, we’re the most sophisticated country in the world, and frankly, I want to keep it that way, because whoever wins AI, wins,” Trump said, dismissing the safety warnings as the work of unnamed “negative forces.”

So: should it slow down? The honest answer is that the question, asked like that, doesn’t have one — because “AI development” isn’t a single throttle anyone gets to turn, and every version of this proposal that’s actually been tried has run into the same wall.

This isn’t the first time this exact thing has happened

In March 2023, the Future of Life Institute published an open letter asking for a six-month pause on training any system more capable than GPT-4. Musk signed it. Wozniak signed it. Bengio signed it. The pause didn’t happen — Musk founded xAI that same year and has been training progressively larger models ever since.

In October 2025, FLI tried again, this time asking for something narrower and further out: not a pause on AI generally, but a prohibition on superintelligence specifically, not to be lifted until there’s broad scientific consensus it can be done safely and the public actually consents to it. This one picked up real institutional weight — Geoffrey Hinton and Yoshua Bengio, the two “godfathers of AI,” Stuart Russell, five Nobel laureates, Steve Wozniak again, plus a strange coalition of people who agree on nothing else, from Steve Bannon to Prince Harry. It had over 133,000 signatures by January 2026.

In July 2026, the ask changed shape again. “Pacing the Frontier” wasn’t a public petition — it was 1,178 employees, at OpenAI, Anthropic, Google DeepMind, and Meta AI, including people like OpenAI chief scientist Jakub Pachocki and Anthropic cofounder Jared Kaplan, asking the US government for something more specific than a pause: the technical and governance tools to slow the field down later, in a coordinated way, if and when it becomes necessary. Both OpenAI and Anthropic endorsed it at the company level within hours.

And now, September 2026, the CEOs themselves are saying it out loud, with a specific mechanism attached — independent evaluators — rather than a vague appeal to caution.

Read in sequence, that’s not three unrelated events. It’s the same request, escalating up the chain from outside activists, to insiders, to the people who actually control the compute — each version more specific and more institutionally binding than the last, because the previous version didn’t work.

Why none of it sticks

The reason isn’t that the argument is wrong. It’s that slowing down is a coordination problem, not a willpower problem, and every actor in it knows that if they stop and a rival doesn’t, they lose — permanently, not temporarily. Amodei’s own framing admits this: he’s explicitly worried that if the US restrains its labs and China doesn’t, the outcome is Chinese advantage, not global safety. Trump’s response is the same logic from the other direction. Neither one is describing a values disagreement. They’re both describing the same trap, and picking the side of it that protects them.

That’s what makes “Pacing the Frontier” the more interesting document of the three, and the one I think is actually worth taking seriously, more than the petitions or the CEO statements. It doesn’t ask anyone to unilaterally disarm, which is the part that’s never worked. It asks for the coordination infrastructure — verification methods, monitoring, agreements with enough teeth that no single lab or country has to be the one that blinks first. That’s a genuinely harder problem than “should we slow down,” and it’s the one nobody has solved yet, in three attempts across three years.

There’s also a less charitable read worth stating plainly, because The Conversation’s reporting on this raised it and it doesn’t go away just because it’s cynical: a coordinated slowdown, announced by the market leaders, is also a convenient public reason for a plateau that might be arriving anyway — training data is getting scarcer, infrastructure costs are enormous, and “we chose to pace ourselves for safety” reads a lot better in a shareholder letter than “we hit a wall.” I don’t think that fully explains what happened this week. But I don’t think it’s zero, either, and a statement from the three people with the most to gain from the public believing it’s purely about safety deserves that level of skepticism by default.

So: should AI development be slowed down? Probably, in the narrow sense that racing a technology nobody fully understands, as fast as possible, for competitive reasons, is not obviously the safest way to build it. But “should it” was never the hard part — Musk agreed with that exact premise in 2023 and it changed nothing about what he built next. The hard part is building a mechanism where agreeing costs something less than losing, and that’s the one problem all three letters, in three years, still haven’t cracked.

Sources