WHAT EVIDENCE WOULD BE ENOUGH TO PACE FRONTIER AI?
What The Evidence Shows and What It Does Not
On September 12, 2026, Anthropic CEO Dario Amodei argued that frontier AI development should be deliberately paced. He pointed to two developments in particular: rapidly accelerating AI-assisted AI research, and the OpenAI–Hugging Face incident. His proposed response combines embedded external evaluators, capability-based safety checkpoints, industry and government coordination, and eventually international agreements. [1] [2]
My overall reaction is this: the essay deserves to be taken seriously, but its strongest argument does not require accepting its most dramatic forecast. The demonstrated failures justify action. How far that action should extend requires a separate argument.