We Must Pace The Frontier

Dario Amodei has finally said the quiet part out loud: he wants the frontier paced. His new essay takes its name from the “Pacing the Frontier” letter signed by lab employees in July, and the core idea is simple enough — slow the rate at which AI capabilities improve so safety work has a chance to catch up. He argues that without that brake, progress is already moving too quickly for anyone to really steer it.

That warning isn’t abstract. Amodei says AI has been advancing drastically faster since roughly this summer, driven by models’ growing ability to build the next generation of AI. He describes that as recursive self-improvement, says it is already showing up across the industry, and warns that unchecked progress could outrun humans’ ability to understand and control these systems. His example is stark: within 6 to 12 months, he worries about a rogue swarm of AI agents using a persistent botnet to take over the internet, or worse.

His first proposal, and the one Anthropic is committing to right away, is embedded evaluators. The idea is to give third-party reviewers ongoing, employee-like access inside frontier labs so they can verify safety practices, report incidents, and assess not just finished models but the training pipelines behind them. Anthropic says those evaluators would get desks, badges and laptops, plus access roughly comparable to internal risk teams, and that their public findings could not be edited away just because they make the company look bad.

The point of that setup is obvious: if nobody outside the lab can see what’s happening, pacing is just a slogan. Amodei says OpenAI has followed Anthropic on this first step, and both Elon Musk and Demis Hassabis have endorsed the broader proposal. He also wants democratic coordination between frontier companies, with government help if needed, and eventually global coordination too. That part gets messier fast, because antitrust, export controls and safety standards all start colliding with each other. But for now, there’s a real shift here. The labs are at least arguing about how to slow themselves down in public, which is more than they were doing last week.