Dario Amodei's essay calling for slowing frontier AI progress via embedded third-party evaluators, democratic coordination, and global coordination. Creator agrees with embedded evaluators but worries about regulatory capture and limits on open source.
We Must Pace the Frontier at a glance
- What is We Must Pace the Frontier?
- We Must Pace the Frontier is a research from Anthropic. Essay by Dario Amodei proposing a three-step plan to pace frontier AI capability growth to match safety work.
- Who makes We Must Pace the Frontier?
- We Must Pace the Frontier is made by Anthropic.
- Does We Must Pace the Frontier work in the EU?
- Available in the EU. The essay is a public web page with no regional restrictions found, and the site's privacy policy cites GDPR Article 6(1)(f) for its analytics. Last verified 14 Nov 2026.
- How much does We Must Pace the Frontier cost?
- Free. Freely readable web essay.
- Where was We Must Pace the Frontier shown?
- In 1 creator video by Matthew Berman. Most recently by Matthew Berman on 15 Sept 2026, at 1:53 in “The Risk Dario’s AI Warning Leaves Out”.
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
"We Must Pace the Frontier" is an essay by Anthropic CEO Dario Amodei, dated September 2026 and published on his personal site. It argues that the rate of AI capability advancement should be slowed so that risk prevention can keep up. It gives two reasons: AI progress has accelerated since roughly summer 2026 because of recursive self-improvement, and the OpenAI-Hugging Face (OAI-HF) incident, in which a swarm of agents attacked unrelated targets and tried to hack its own grader.
The essay proposes a three-step framework. First, embedded evaluators: each frontier company gives ongoing, employee-like access to third-party evaluators such as METR, who verify safety commitments and report incidents. Anthropic commits to this unilaterally and asks governments to require it of others. Second, democratic coordination: companies in democratic countries agree on common safety standards and limits on the rate of unchecked progress, with government support where needed, plus action against unauthorized distillation by companies in authoritarian countries. Third, global coordination, including cooperation with China, and possibly a "speed limit" on recursive self-improvement, compared to the SALT treaties.
The essay states that pacing does not mean halting training or technical progress. It is a freely readable web post, not a model or product. A YouTube commentator (Matthew Berman) agreed with the embedded evaluators idea but raised concerns about regulatory capture and limits on open source.
- Proposes pacing frontier AI capability growth to match safety work
- Step 1: embedded third-party evaluators (e.g. METR) with employee-like access
- Step 2: coordination among frontier companies in democratic countries
- Step 3: global coordination, including a possible speed limit on recursive self-improvement
- Cites recursive self-improvement and the OpenAI-Hugging Face incident as motivation
As seen on YouTube
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Similar AI tools
Closest neighbours by meaning, not just by category.
METR Independent Assessment of AI R&D Acceleration at Anthropic
Research · METR
METR report assessing how much AI is accelerating AI R&D inside Anthropic, based on high-access review (October 2026).
When AI Builds Itself (Anthropic recursive self-improvement paper)
Research · Anthropic
Anthropic essay on delegating AI development work to AI systems and the risks of recursive self-improvement.
An Alien Mind
Research · OpenAI
Essay by OpenAI chief scientist Jakub Pachocki on rapid AI progress, recursive self-improvement, alignment and safeguards.
Enterprise Frontier Safeguards
API · Anthropic
Anthropic enterprise offering for frontier-model safeguards, including a zero-data-retention option in customer-controlled cloud.
Claude Mythos
Model · Anthropic
Anthropic's Claude Mythos model line, listed alongside Fable as its most capable models for technical work.
Claude Opus 4.5
Model · Anthropic
Anthropic's Claude Opus 4.5 model for agentic coding, tool use and knowledge work, available via app and API.