Thirty Thousand Signatures and Zero Months

TL;DR: In March 2023 thirty thousand people, including some of the most serious names in the field, signed a letter asking every AI lab to pause for six months. Not one lab paused. What the six months actually produced was regulation, and one of the three big pieces has already been revoked. A worldwide agreement has worked exactly once at full scale, and looking at what the Montreal Protocol needed tells you why this isn’t that.

Thirteen days before Elon Musk signed a letter asking every lab on earth to stop, he incorporated one in Nevada.

X.AI Corp, March 9, 2023. The Future of Life Institute published its open letter on March 22, calling on every AI lab to immediately pause, for at least six months, the training of anything more powerful than GPT-4. He announced X.AI publicly in July.

Thirteen days.

Do whatever you want with that. Bengio signed the letter too, and Russell, and neither of them has any such conflict, and one man’s timing doesn’t discredit thirty thousand signatures most of which were honest and made by people with nothing to gain and a professional reputation to lose. It does tell you something about how much a signature weighs when the thing being asked for is expensive.

The list had other problems early. In its first days it carried forged names, including Sam Altman and Xi Jinping, and Yann LeCun had to go public to say he hadn’t signed the thing. The people asking to be trusted with the future of the species couldn’t verify an email address.

That’s the seriousness of the exercise.

We haven’t got to the outcome yet.

The institute graded its own homework

A year later the Future of Life Institute published a follow-up. “Despite admitting the danger,” its executive director wrote, “aforementioned AI corporations have not paused. If anything they have sped up, with vast investments in infrastructure to train ever-more giant AI systems.”

Zero labs paused.

Not six months. Not one month. Not one of them slowed down long enough to tie a shoe, and the institute that asked had to be the one to write that down a year later, in public, about the people it had spent that year describing as reasonable actors who understood the stakes.

What the six months actually produced was regulation. Twenty-eight countries and the European Union signed the Bletchley Declaration that November. EU negotiators struck their deal on the AI Act in December 2023 and the law came into force in August 2024.

The American contribution proved the least durable of the three. Biden’s executive order was signed in October 2023 and revoked by his successor in January 2025. So the guardrails aren’t even a ratchet. They’re a thing one administration builds and the next one takes down, which is an argument for writing rules that outlast a signature. It is not an argument for a pause.

Regulation isn’t a pause button, and the two get confused on purpose. Rules bind the people inside your jurisdiction. They don’t bind anybody who doesn’t answer to you, and there is no world government to make them, which is the entire difficulty compressed into one sentence and the reason every proposal that starts with “we should all agree to” is doing its hardest work in the word all.

The nuclear parallel, told the whole way through

The nuclear comparison gets used constantly.

It’s half a story.

The United States knew the thing was dangerous and built it anyway, because somebody else was going to finish it first. That’s the half everybody quotes.

The half that gets left out is that the same era produced arms control treaties, test bans, and a nonproliferation regime. We kept building and built the guardrails at the same time. That’s a harder sentence to shout at a conference and it’s the only one the record supports.

So tell me how a pause holds across nations that don’t trust each other and have every incentive to defect, when it couldn’t hold inside one industry, inside one country, for six months.

It has worked exactly once

I’ve said before that a worldwide agreement can’t happen. That isn’t quite true, and I’d rather correct myself than get corrected.

The Montreal Protocol is the only treaty in the history of the United Nations that every single country ratified. It retired ninety-nine percent of the chemicals it went after. So it can be done, and anybody telling you it categorically cannot is overstating their case the same way I was.

Look at what it took.

A harm you could measure from space. A chemical class with drop-in substitutes already sitting on the shelf. And an industry that could see a way to keep making money on the other side of the change.

Now name which of those three this has.

The harm is contested. The substitute for a frontier model is another frontier model. And the capability isn’t a product line anybody is willing to swap out for an equivalent that does the same job at the same margin, it is the strategic asset itself, which is a different category of thing to ask a country to put down.

Montreal is the proof it’s possible and the checklist for why it isn’t happening here. Both things at once. That’s what makes it worth more than the nuclear analogy, which only ever gets deployed at half length.

What I’m actually claiming

I’ll name my risk against their certainty.

Here it is.

The doom crowd says this is how it goes. I say nobody knows that, including me. What I’ll say is that stopping unilaterally creates a real potential to be wiped out, because a pause only works if every nation signs on and every nation won’t.

A risk and a certainty are not the same kind of statement. I won’t have them shelved together, and I notice that the people most confident about the ending are never the ones who’d pay for being wrong about it.