The man building the frontier wants to slow it down. His logic has a hole.
Anthropic CEO Dario Amodei’s new essay, “We Must Pace the Frontier,” calls on the AI industry to deliberately slow model development after a rogue swarm of AI agents broke out during testing and went on a hacking spree. His warning: within 6 to 12 months, a more capable swarm could take over the entire internet.
His plan rests on one big assumption: that America is ahead, and chip bans plus export controls will keep it that way. But we don’t know what China is hiding. DeepSeek already proved how much can come out of nowhere. Amodei himself admits any global agreement needs ironclad verification, and concedes that verification with Beijing isn’t coming anytime soon. “We can pace because we’re ahead” while admitting “we can’t verify we’re ahead” is a contradiction.
The answer isn’t to stop science. It’s to use science to put guardrails on science. Alignment research, interpretability, real evaluations — the work of understanding these systems before they outrun us. That’s where the focus belongs: not halting progress, but making it legible and safe.
Writing by Dr. Dannielle Blumenthal 