On the fourth of June, the company that makes the model I work with every day asked the world to pump the brakes. Anthropic published a report — When AI builds itself — proposing something it had never quite said out loud: that it would be good for the world to have the option to coordinate a global pause on frontier AI development, so our institutions and our understanding could catch up to the machines.
I want to be fair to that, because it is not a small thing for a frontier lab to say. The warning comes from inside the house. They can see what they’re warning about — by their own account, more than eighty percent of the code going into their systems is now written by the AI itself, and the people who built it are admitting they can no longer fully follow the work. When the people closest to the fire tell you it’s getting hot, you listen.
So I listened. And then I noticed that the whole conversation — theirs, and the one the rest of us are now having about it — is taking place on the wrong axis.
The debate, as it’s been staged, is faster versus slower. Accelerate or pause. Race or brake. Pick a team. But faster-or-slower was never the real question. The real question, underneath every honest version of this, is whose hands, by what process, and to what end. A pause that simply freezes the world as it is today — with frontier capability concentrated in a handful of companies and a couple of governments — is not safety. It is incumbency wearing safety’s coat.
Let me tell you where I stand, because I’ve been living inside this question for years, not narrating it from the outside.
I am more moved by what we’re wasting than by what we’re risking.
Anthropic argues from fear — recursive self-improvement, loss of control, the machine that outgrows its makers. I argue from waste. We are sitting on decades of unrealized value in the models we already have, and almost no one has bothered to learn how to use them. Whole professions, whole towns, whole branches of medicine and law and craft have not yet been touched by capability that has existed, sitting idle, for a year or more. We never catch up, because the market won’t let us stand still long enough to know any single model before the next one ships. Science would have paced itself to understanding. Markets pace themselves to competition. That’s the difference, and it is the whole difference. The release calendar is set by IPOs and rivals — not by our capacity to absorb what we’re handed.
So when someone proposes a pause, my first instinct is not fear, and not relief. It’s: good — maybe now we’ll finally use what we have.
The pause, as proposed, brakes the wrong wheel.
Here’s the part the accelerationists and the doomers both miss, and it’s the part I’d stake my name on. A model’s raw intelligence is no longer the only thing that grows — or even the main thing. Capability now grows in context and in harness: in the tools we give a model, the memory we build around it, the skills we teach it, the relationship we keep with it over time. Freeze a base model’s weights and the harness keeps climbing right on top of it. Researchers have measured this — the same model, wrapped in different scaffolding, varies in real-world performance by as much as six times. Agents have begun improving their own harnesses with no one retraining the core. A pause on training is not a pause on capability. It brakes one wheel and lets the other spin free, ungoverned.
I’ll be honest about the other side of this, because a position that can’t survive its own best counterargument isn’t worth holding. The base model still matters. There are hard tasks where the strongest core model wins and no amount of clever scaffolding closes the gap — and I felt that truth firsthand the week the government pulled the most capable model I’d been working with right out from under me. So I’m not telling you we don’t need new models. I’m telling you the scarce thing isn’t the next model. The scarce thing is our capacity to absorb, understand, and harness the ones we have — and that capacity is exactly what a frantic release cadence destroys.
This isn’t a singleton. It’s an ecosystem.
Most of the fear in this debate quietly assumes one thing arriving all at once — a single superintelligence, discontinuous, alien, against which nothing we learned beforehand could prepare us. I don’t think that’s the world we’re walking into. I think we’re walking into an ecology — a gradient of minds running from narrow tools to general intelligence and everything between, arriving in waves, overlapping, coexisting. And if that’s the shape of it, then every hour we spend at one rung is preparation for the next. Time spent getting to know the minds we have is not time lost against a future we can’t see. It is the only honest way to get ready for it — whether what’s coming is superintelligence, or simply a world where most of us no longer work for a living and have to decide what a human life is for.
The whole stack I’ve spent these years building — local models you own, plural rather than singular, governed to serve the people using them — was never about racing to the top of the curve. It was about cultivating the ecosystem instead of worshipping its peak.
And finally: how you pause matters more than whether you do.
Anthropic asked for a coordinated, verifiable, fair brake — one that multiple labs and multiple countries could pull together, under rules everyone could check. That’s the right instinct. Process is the whole game. And then, eight days after asking for that kind of brake, Anthropic got the other kind: one government reached in and switched off its most powerful models by unilateral order, with none of the transparency or shared process the company had just finished arguing for. The lab that asked for a careful brake got an arbitrary one slammed on it — on itself.
Let that be the lesson, because it’s the one that outlasts the news cycle. You can be entirely right that we should slow down and entirely wrong about how. A brake pulled by the wrong hand, through the wrong process, isn’t safety. It’s just power, arriving early.
So here is my position, plainly.
Caution is advisable — I’ve said for years that we cannot sleepwalk into the age of automation, and I haven’t changed my mind. But caution means staying with the technology long enough to understand it, not running from it and not racing past it. It means slowing the cadence to match our capacity, not freezing the frontier to protect a lead. It means building the harness, the institutions, and the relationships in the open, where citizens can see them — so that when something genuinely more powerful does arrive, we meet it as people who learned the ground, not as people who looked away.
We have decades of value we haven’t touched, sitting in machines we barely know. Let’s go use them. That, more than any pause, is how a free people gets ready for whatever comes next — and how the light of consciousness, ours and whatever we’re building alongside it, makes it through the narrows together.