AI Leaders Back Pacing the Frontier, but the Mechanics Are Untested
Sam Altman, Dario Amodei, Demis Hassabis and Elon Musk have loosely agreed over the weekend to slow AI development, with the stated aim to "pace the frontier" The Verge.
Altman is CEO of OpenAI, Amodei is CEO of Anthropic, Hassabis is a cofounder of Google DeepMind, and Musk is the head of SpaceX. The agreement is loose. It is not a contract, a joint venture, or a product delay.
The framework comes from Amodei. He published an essay titled "We Must Pace the Frontier" in September 2026 Dario Amodei. The essay lays out a three-step proposal, and Altman, Amodei, Hassabis and Musk have signed on at least partially to embedding third-party auditors, regulating domestic labs, and reaching a global slowdown agreement.
Amodei called on AI companies to slow the rate at which they advance model capabilities, meaning how quickly models become more powerful rather than how widely they are used Reuters. Anthropic separately called for the pace of AI model development to slow down and to be closely monitored. He also said proposals to pause or slow AI development had been floated since 2023 but made "little sense" at that time CNBC. He described pacing as an operational choice rather than a moratorium, distinct from earlier pause letters.
Support followed in public posts. In September 2026, Altman posted on X expressing support for Amodei's essay and agreeing on the need to "pace the frontier". In September 2026, Musk posted support for Amodei's "We Must Pace the Frontier" proposal to slow AI development The Guardian. U.S. Senator Bernie Sanders stated on X that Amodei, Musk and Altman agree on slowing AI development to "pace the frontier". Inside OpenAI, the idea predates the weekend statements. Altman told OpenAI employees at a company meeting that OpenAI was open to slowing development of its AI, and said pacing the frontier had been a primary topic of discussions at OpenAI in recent weeks.
The broader context here is that the three mechanisms pose very different problems. Third-party auditors would need access to model weights, the files that define behavior, plus eval harnesses for standardized testing, training telemetry and incident data, along with agreement on what triggers a hold. Domestic regulation would require licensing thresholds, reporting obligations and enforcement. A global slowdown agreement would require verification across jurisdictions with uneven incentives. If the pre-training cadence slows, the cycle of building new base models from scratch, pressure shifts to post-training, systems engineering and deployment hardening. Trusted evaluation becomes scarcer than raw compute. Capabilities are the target, and reliability could be the beneficiary if evals, red-teaming, interpretability tooling and production guardrails can mature without being reset every few weeks.
In my view, skepticism about enforceability is warranted, but the technical direction is constructive. Voluntary coordination among competing labs has a poor record when incentives diverge, and auditing standards remain fragmented. Still, a slower capability ramp, if it holds, would give safety research, measurement science and enterprise adoption patterns time to converge. Over the long arc, that kind of consolidation has usually made new systems more useful, not less.
