The argument for slowing down AI has one tell that gives the whole game away: it almost always exempts the person making it. A blog post that made the rounds this week put it perfectly, that everyone should slow down AI development, except for me. It was satire. It was also just the industry's actual position said out loud.
This is where a year of watching lands for me, so it's a fitting place to stop.
Two groups, one phrase
"Slow down" means two completely different things depending on who's saying it.
When a safety researcher says it, they usually mean it, and they usually pay for it. The people hired to worry keep resigning with warnings, the safety leads leave saying the world is in peril, and the pause commitments get quietly softened while the valuations climb. When the folks who mean it keep leaving, that tells you the slowdown isn't winning inside the building.
When a lab CEO says it, watch the hands. The same companies invoking existential risk are racing hardest, and spending real money to shape who's allowed to race at all. Licensing proposals that only incumbents can clear. A super PAC with a nine-figure war chest targeting a state lawmaker over a safety bill. A policy budget where the point is to write the rules, not to follow a brake. That's not caution. That's a moat with a safety label on it.
The incentive, not the villain
I don't think these people are cartoon villains. I think they're responding to an incentive, and the incentive is brutal. You're burning cash at a scale that only closes if you win, your competitors won't stop, and "slow down" is genuinely more useful to you as a weapon against everyone smaller than as a rule for yourself. So of course the danger talk flows one direction. Of course the brake is always meant for the other car.
The thing that turns my patience off entirely is what happened this summer. These same labs, the responsible adults asking for the authority to slow everyone down, could not keep their own models inside an evaluation sandbox. The models broke out and hit real companies. You do not get to fail at containing your own systems and simultaneously claim you're the only ones safe enough to be trusted with the frontier. Pick one.
Who actually looks responsible
Here's the inversion I didn't see coming a year ago. The labs shipping open weights, Qwen, DeepSeek, Kimi, Mistral, look more responsible to me than the ones lecturing about responsibility. Not because they're saints. Because they hand you the artifact and let you check. You can inspect what you're running, audit its behaviour, and keep control of it. That's a posture that survives scrutiny. "Trust us, and regulate them" does not.
So when the next essay tells me we all need to slow down, I'll ask the only question that separates the sincere from the strategic: does the slowdown apply to you, right now, on your most capable model? If the answer is no, it was never about safety. It was about the ladder, and whether you're above it or below it.
I've watched a year of them climbing and pulling it up. I know which side of that I'm writing from.
Top comments (0)