Anthropic CEO says it’s time to pump the brakes on AI

Amodei’s call to "pump the brakes" lands at a curious moment in the AI arms race. Anthropic has positioned itself as the safety-first lab, but safety rhetoric has historically been easier to articulate than to operationalize. Opening models to third-party evaluators like METR is a meaningful gesture — it trades some proprietary secrecy for credibility — but it also begs the question: who audits the auditors, and what happens when an evaluation finds something uncomfortable? The essay’s three-step plan is heavy on framing and light on binding commitments, which is typical for a CEO trying to shape norms without ceding competitive ground. The broader context is the classic collective-action problem. Any single lab that slows down unilaterally risks losing talent, funding, and market position to rivals who don’t share the same scruples. Amodei’s proposal is essentially an attempt to build a coalition of the willing, but the history of tech self-regulation suggests such coalitions tend to collapse when the pressure to ship meets the next funding round. His language about "pace" and "frontier" is carefully chosen to sound deliberate, yet the underlying tension remains: safety is a public good, and public goods are chronically underfunded in competitive markets. What makes this notable is not the essay itself — similar calls have come from OpenAI, DeepMind, and various open letters — but the specific mechanism of third-party access. That is a concrete, testable commitment, not just a vibe. If METR and similar groups are given real visibility into model behavior before deployment, we might get something closer to a pre-market safety review. Still, the devil is in the details: what counts as a "safety practice," who defines the thresholds, and what happens when a model fails? Without teeth, this is another promise; with teeth, it could be a template for the industry. The source article (https://www.theverge.com/ai-artificial-intelligence/994337/anthropic-ceo-slow-down-ai-development) captures the announcement but leaves the hard questions hanging. Amodei is right that the pace of frontier AI development has outstripped our ability to understand its consequences, but slowing down is not a technical fix — it’s an economic and political one. Until the incentives change, calls to pump the brakes will remain more symbolic than structural, no matter how many evaluators get invited to the table.