The labs negotiate their own brake

On September 12 Dario Amodei published “We Must Pace the Frontier,” arguing that the industry should slow the rate at which model capabilities improve. His proposal has three steps: third-party evaluators such as METR embedded inside the labs to verify safety commitments and incident reports, common safety standards among leading firms in democratic countries with the US government mediating the antitrust problem, and international agreements — China included — limiting the most dangerous capabilities. Anthropic committed to the first step unilaterally.

Agreement came the same day. Sam Altman wrote “I agree with Dario that we need to pace the frontier” and said OpenAI would take embedded evaluators too; Elon Musk posted “Dario is right.” By September 18 Anthropic and OpenAI had a joint proposal to give independent evaluators access to their staff.

The obvious objection arrived just as fast. Outside experts published a letter arguing that evaluators chosen, funded or hosted by the labs they assess are not independent, and critics note METR's ties to the same effective-altruist network as Anthropic. It is also the month Anthropic files to go public, which is why the sceptics keep asking whose moat a coordinated slowdown protects.

• Reported Amodei’s three steps: embedded third-party evaluators, common safety standards among democracies, and international limits on the most dangerous capabilities. TechCrunch
• Reported Altman: “I agree with Dario that we need to pace the frontier,” pledging embedded evaluators at OpenAI. Musk posted “Dario is right.” TechCrunch
• Reported Outside experts replied in a public letter that evaluators funded or chosen by the labs are not independent. CNBC