TL;DR
The short version
Dario Amodei's essay We Must Pace the Frontier argues labs should slow capability gains enough for safety work to keep up — without a full halt — starting with embedded third-party evaluators. Unusual same-weekend support from rival leaders made the proposal feel like the start of industry negotiations, not another lone safety post.
This page works from The AI Daily Brief's breakdown of Amodei's plan and the reactions around it, checked against the primary essay and contemporaneous press.
Built on The AI Daily Brief with Nathaniel Whittemore, with claims checked against Amodei's essay.
Pacing is not a pause
On 12 September 2026, Anthropic CEO Dario Amodei published We Must Pace the Frontier — about 3,800 words arguing that early recursive self-improvement and incidents like the OpenAI–Hugging Face agent swarm mean capability gains must slow enough for safety work to catch up. Pacing, he stresses, does not mean stopping training. It means taking time to align and safeguard models, and letting third parties confirm that work.
Pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models and for third party evaluators to confirm this.
Dario Amodei
What would the extra time buy? Amodei names four buckets: operational excellence, alignment, interpretability, and testing/evals. Whittemore notes that even if you discard "alignment," three of the four are broadly uncontroversial.
Three steps — only one binds today
The plan climbs in difficulty:
- Embedded evaluators — third parties (METR-style) with ongoing, employee-like access to verify practices, report incidents, and assess training pipelines, not just finished models. Anthropic commits unilaterally.
- Democratic coordination — common safety standards and limits on unchecked progress across frontier firms in democratic countries, with antitrust overhang and talk of government waivers.
- Global coordination — including attempts to work with authoritarian governments, with hard verification problems.
Only step one is a near-term commitment. Sam Altman said OpenAI would match independent evaluators with employee-like access. Elon Musk wrote "Dario is right." Demis Hassabis said the direction is correct. Satya Nadella welcomed embedded evaluators and deliberate pacing while stressing open and closed ecosystems. Meta's Alexander Wang stayed more cautious, treating alignment as a possible gating factor rather than a full pact.
Pushback is the negotiation
Critics arrived immediately. Yann LeCun cast it as ladder-pulling. Investors framed pacing as concentrating power or dressing up cost/IPO pressure. Legal voices warned industry-wide slowdown talk can look like an invitation to collude. Holly Elmore and others challenged METR's independence. China's Global Times framed the essay as technological barriers and US hegemony — a "silent AI Cold War."
Whittemore's frame: specificity made this discourse more productive than last week's researcher warnings. Labs still set the frontier. They no longer get to set it alone.
What to watch if you build or buy AI
Treat "pace the frontier" as a negotiation frame, not finished policy. Watch who actually embeds independent evaluators, what they can publish without lab editorial control, and whether coordination stays voluntary or becomes regulation. Prefer mechanisms — access, incident reporting, pipeline evals — over slogan fights about pause versus accelerate.
Keep present-tense operational risk in the same room: sandbox hygiene, agent containment, and customer trust after incidents. Geopolitics will dominate headlines; reliability failures will dominate your week.
