Dario Amodei Wants a Slowdown, an Antitrust Waiver, and a Crackdown on Distillation
Anthropic CEO Dario Amodei published an essay on September 12 called "We Must Pace the Frontier," arguing that the handful of labs building the most advanced models should deliberately slow down together. The concrete ask has three parts: permanent, employee-level access for outside evaluators like METR inside frontier labs, a narrow antitrust waiver from Washington so labs can legally coordinate on pacing without exposure to collusion claims, and tighter controls on model distillation and weight theft aimed squarely at China. I build small products on top of a mix of frontier and open-weight models, and the distillation piece is the part that actually touches my stack, not the frontier-lab pacing itself.
What the essay proposes, not what the headlines say
The essay never uses the phrase "open weights." What it does argue is that coordinated pacing among frontier labs is "legally challenging" under current antitrust law, so it asks the U.S. government to issue a narrow waiver specifically for safety-related coordination conversations between competing labs. It pairs that ask with a call to keep chip supply constrained to China and to crack down harder on unauthorized distillation and weight theft, framed as protecting the safety-conscious labs' lead rather than protecting their market position. Reading the actual text matters here, because most of the coverage collapses "crack down on distillation" into "attack open weight models," and those aren't quite the same claim even if the practical effect on the ecosystem could end up looking similar.
The pushback arrived within a day
This isn't a proposal sailing through uncontested. Figures close to the administration's AI policy apparatus, most visibly David Sacks, publicly rejected the antitrust-waiver piece specifically, arguing that labs should be able to pace themselves without a special legal carve-out from the government. That's worth knowing before treating this essay as a settled industry direction: it's a proposal from one lab's CEO, already contested by people with real influence over whether any of it becomes policy.
Why the distillation piece is the one that matters to me
I don't run frontier-lab-scale infrastructure, and neither does almost anyone reading a blog about solo operators. What I do run is a mix of API calls to frontier models for the tasks that need them and open-weight models, Qwen and DeepSeek derivatives mostly, for everything where cost matters more than the last few points of quality. That second category exists at its current price and availability specifically because open-weight labs can train competitive models more cheaply, in part by learning from frontier lab outputs. A real crackdown on distillation, if it takes the form of export-control-style restrictions rather than just IP litigation, could shrink the pipeline that keeps open-weight models improving at the pace they have been. Prices for the cheap tier of my stack didn't get this low by accident, and that pipeline is exactly what's under discussion here.
The counterargument I take seriously
I'm not dismissing the safety case out of hand. Frontier capability has moved fast enough over the past two years that a credible argument for deliberate pacing, with real third-party access to verify labs are actually pacing rather than just claiming to, isn't obviously self-serving nonsense. Embedded evaluators with permanent access is a real, costly commitment if Anthropic actually follows through on it for itself, not just a talking point. The uncomfortable truth is that a policy can be both a genuine safety measure and a move that happens to protect the market position of the three labs already ahead. Those aren't mutually exclusive, and pretending the essay is purely self-interested is as lazy as pretending it's purely altruistic.
What I'd actually do
My actual hedge isn't picking a side in the safety-versus-market-protection debate, because I don't think I have enough information to resolve it and neither does most of the commentary I've read this week. My hedge is architectural: I don't build anything where swapping the underlying model out from under the product would break it. Every open-weight dependency in my stack sits behind an interface I could point at a different provider, frontier or open, within a day if pricing or availability shifted hard in either direction. That's cheap insurance against a policy outcome I can't predict and can't influence.
Where I could be wrong: if this specific proposal goes nowhere, which the immediate pushback suggests is plausible, none of this matters this year, and I'll have spent effort hedging against a threat that never materialized. I think that effort is cheap enough to be worth it anyway, since the same architecture protects against ordinary business risk too, a provider raising prices or having an outage, not just a change in distillation policy.
Author
Lukas
@lukcombinator