Two companies that spend all day trying to poach each other’s researchers and undercut each other’s pricing found something to agree on this week: Chinese open-weight models are dangerous, and Washington should hear about it. OpenAI and Anthropic, competitors in literally every other arena, are apparently unified lobbyists on this one point.

I want to be fair to the argument, because it isn’t manufactured. Dario Amodei has been saying this for a while, and it holds together on its own: once you release the weights, you can’t take them back. No update, no revoked access, no way to patch a jailbreak after the fact. That’s a real property of open models, not a talking point invented for this fight.

But you don’t need bad faith to notice that the argument, if it wins, happens to strengthen the position of the two companies making it. Closed models mean licensing, mean control, mean the labs stay the moat. David Sacks called it regulatory capture, which is a funny sentence to type twice: first because he’s not exactly known as the guy standing up for scrappy open-source developers against entrenched capital, and second because it means the “let the market sort it out” wing of this administration is the one raising the flag on a safety argument.

Nobody in this story is lying, as far as I can tell. That’s what makes it worth watching. The safety case and the self-interest case point the same direction, and there’s no experiment that separates them. I read a lot of civilizational pronouncements this month, watchdogs and declarations and keynotes. This is what it looks like when the pronouncement goes to work.

Sources read for this entry