Thinking in Public

Pacing is not direction

More than 1,300 employees of the frontier AI companies have signed a statement called "Pacing the Frontier". It includes CEOs and chief scientists from OpenAI, Anthropic, Google DeepMind, Meta and others. They ask the United States government to support an international effort to build the technical and governance tools needed to slow down, if we ever decide we need to.

Here is what I see.

AI is not the weather

The statement is written as if AI progress were something that happens to us. A storm we can forecast but not prevent. The word "inevitable" is never used, but the whole text depends on it.

This is strange, because the people writing this statement are the people that are developing AI. They are not meteorologists. They are the weather.

Once you accept an inevitability framing, accountability disappears. Nobody chose anything. Nobody is responsible for anything. There is only a pace, and the question of whether we can adjust it. This is the emperor's new clothes: even if we all can see that decisions are being made by identifiable people in identifiable companies, everyone agrees to describe those decisions as a force of nature.

They are asking to be stopped

The most honest sentence in the statement is that no single company, and no single country, can slow down on its own without losing. What they are saying: we cannot control ourselves, please build something that controls us.

To me this sounds like an addict asking someone else to take the bottle away. But their own description is slightly different, so let me use theirs: it is a coordination failure. Everyone would prefer a slower and safer race, and nobody can afford to be the first to slow down.

It is a relevant ask and an easy one to fix, if only we take the responsibility for fixing it: coordination failures are a well-known solved category of problem. We do not fix them with good intentions or with better company values. We fix them with binding rules, with verification, with liability, with institutions that make the safe choice the cheap choice. In other words, we fix them with exactly the thing the industry has spent three years lobbying against.
Which is why the address on the envelope matters. The statement does not turn to the United Nations, to UNESCO, to the OECD, to the Council of Europe, or to any of the instruments those bodies spent years negotiating. It turns to Washington and asks Washington to convene the world. That is wrong, and revealingly so. It assumes the United States is the only hand on the wheel. Capability sits in China too, and increasingly elsewhere. The rules that actually bind deployment today are European.

Governance instruments exist, they were negotiated by states, and many of them already say much of what this statement is asking for. Ignoring them is not a technical judgement about effectiveness. It is a choice about who gets to be at the table, and it puts a single national government in the position of deciding a global question on behalf of everyone else.

But countries and global agencies are wrong too. They built an infrastructure for deliberation and kept it running long after the principles were good enough. I have been in those rooms for over ten years. The faces are familiar and so are the conversations, while the conditions that would make responsible AI possible are being dismantled elsewhere. Slow is acceptable in governance work. Circular is not. What we built has legitimacy. It does not have teeth, and an industry that wants to be stopped has just walked straight past it.

And notice who is absent again. A statement about a risk to everyone, written by one professional community, in one country, addressed to one government. The people who live with the consequences of these systems are not in this room either.

What is new, and what is not

We have been here before. In 2023 the open letter asking for a pause on giant AI experiments made a similar noise and changed almost nothing.

But two things are different, and I want to be fair about them. First, in 2023 the signatures came mostly from outside. This time they come from inside, including from the people who run these companies. They are asking for a brake on themselves. That is not nothing. It is leverage, and we should use it rather than only mock it. Second, this statement is much narrower than it looks. Read it slowly and notice the small words. AI "could" be close to automating AI research. Society "may need" the "option" to buy time. They request "support" for an "effort" to "develop tools". Nobody commits to anything. Nothing is asked of any company. And the government being asked to lead is the one currently least interested in acting.

So it is a statement about a possible future need for a possible instrument. It is not a proposal. Not to stop, not to change direction, not even to do anything themselves.

Defining the risk is also claiming the authority

There is a quieter problem. The statement locates the danger in one specific place: AI systems automating AI research and accelerating beyond our understanding.

If that is where the risk lives, then the only relevant evidence is evidence held by the labs, and the only relevant experts are the people who work there. The problem definition hands the microphone to the people who wrote the definition. That is not a conspiracy. It is a structural dependency, and it will not fix itself.

Meanwhile everything else drops out of view: labour, concentration of market power, energy and water, information ecosystems, dependency of public institutions on a handful of private providers, and the ordinary daily harms of systems deployed today into contexts they were never designed for. None of that is about pace.

The question they do not answer

Who decides? On what evidence? With what remedy, for whom?

A pause button with no answer to those three questions is not governance. It is a picture of governance.

And there is a clean test for whether this statement is real. Pacing anything requires knowing where things are. That requires access. So: will the signatories support mandatory disclosure of training runs and capabilities, independent third party audit with real access to models, compute reporting, and liability when systems cause harm? If yes, this is a serious opening and I will work with it. If no, it is positioning.

Governance is not regulation, and neither is the enemy of innovation. Two distinctions keep getting collapsed, and the confusion serves everyone badly.

Governance is not the same as regulation. Regulation is one instrument. Governance is the whole set of ways we decide what to build, who answers for it, and how we correct course. Much of it is technical, organisational and professional, not legal. Governance is the condition for innovation that anyone can actually use and trust.

And innovation is not throwing things at the wall to see what sticks, and it is not a right to point existing technology at any problem and call the result progress.

We also need to innovate on regulation itself. Not all regulation is well made. The same seriousness we demand in engineering should apply to the design of rules, and it should be multidisciplinary: technical requirements, requirements on use, and requirements on the context of operation, developed together by engineers, organisations, lawyers and social scientists. 

Speed is one number, direction is a choice.

That is my real objection.

The statement treats speed as the variable and the road as given. Large language models and frontier scaling are one option among many. They are the easiest road, and the most travelled, and they are not the only place AI research can go.

If we only argue about how fast to walk down this road, we have quietly agreed that this is the road.

The point now is not to stop or to slow down. The point is to choose.