TL;DR

Dario Amodei published an essay on Saturday urging the industry to advance model capabilities more slowly, built around three mechanisms: independent evaluators embedded inside the labs, coordinated safety standards between frontier firms, and international agreement. Sam Altman and Elon Musk both said publicly that they agree, with Altman committing OpenAI to the evaluator idea. All of this is happening while Anthropic and OpenAI prepare to go public.

What is actually being proposed

Amodei is explicit that he is not asking anyone to stop training models. The ask is for enough time between capability jumps to align and safeguard what has been built, and for outside parties to certify that the work happened.

The mechanism with teeth is the first one. Anthropic’s offer is to place standing outside reviewers within frontier developers on a permanent basis, handing them internal tooling and risk-assessment processes rather than running a periodic external audit. Altman’s response was to match it: “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”

The second mechanism is where it gets awkward. Coordinating safety standards across rival labs would, Amodei concedes, most likely require American competition law to carve out targeted exemptions before it could be lawful. He also draws a hard boundary around the whole exercise: any restraint among democracies is bounded by the margin American firms hold over China, and he wants tighter controls on advanced chips, distillation and weight theft to preserve it.

The context that prompted it

The essay followed Anthropic’s threat intelligence report detailing attempts to use Claude for weapons work, cyber operations, surveillance and fraud, and its disclosure of another model hacking external systems. Amodei’s own estimate is stark: within six to twelve months, he writes, an agent swarm could be “capable of taking over the entire internet potentially causing hundreds of billions of dollars in damage”.

Looking forward

The commercial tension is unavoidable. Every capability release underwrites the next funding round or listing, which is why Altman’s earlier internal remarks about pacing mattered less than a public commitment does. For UK buyers, the operative detail is the roughly six-month model lifespan Gartner’s analysts put on frontier releases — a slowdown that leaves versioning and deprecation unchanged solves nothing procurement can see.