Dario Amodei has called on frontier AI companies to slow the rate at which they improve model capabilities, rather than halt development altogether, to create time for alignment work and safeguards. The proposal has drawn conspicuous public support: Reuters reported that OpenAI CEO Sam Altman agreed with the call to “pace the frontier,” while Elon Musk responded on X, “Dario is right.”

The alignment matters because it places a shared word at the center of a high-stakes argument among competing AI leaders. But “pacing” is presently a direction, not a demonstrated industry mechanism. Public endorsement of caution is not the same thing as a jointly administered slowdown.

Amodei’s reported framework gives the term more shape than a general appeal for responsibility. It calls for independent evaluators with access comparable to employees; coordination among frontier companies around safety standards and limits; and international cooperation on AI risks. Those are substantial categories of governance. They point toward a model in which labs accept scrutiny from outsiders, agree on conditions for proceeding, and seek coordination beyond any one company.

Each category also contains the questions that determine whether the proposal can operate outside an essay or social-media exchange. Who chooses the evaluators? What access do they receive, and what can they disclose? Which safety standards matter, who defines them, and what happens when a company believes it has met them while others disagree? International cooperation introduces another layer of unanswered design: agreement on risk is not yet agreement on a process for acting on it.

Altman’s reported statement offers the most specific company-level commitment in public reporting. Reuters reported that he said OpenAI intended to adopt independent evaluators with employee-like access. That would be a meaningful shift if carried out: evaluation cannot be fully independent if the people conducting it lack sufficient proximity to the systems under examination. But the statement is an intention, not evidence that such evaluator access has already been implemented. Reuters’ public account quotes that statement only in part, so it should not be read as proof the access is already in place.

That is not proof that no additional commitments exist. Public reporting cannot conclusively rule out unpublished arrangements, later negotiations, or separate commitments among companies. It does mean that the public case for an industry-wide slowdown remains incomplete. There is a difference between leaders converging on the language of restraint and institutions accepting enforceable limits on their behavior.

The current proposal has one concrete virtue: it moves the debate beyond the false choice between unrestricted acceleration and a full halt. Slowing capability improvements to make room for safeguards is a recognizable policy position. Yet its credibility will depend on whether its advocates can convert that position into commitments that outsiders can inspect.

Critics have suggested that public safety language can also serve reputational or commercial purposes. That is a plausible interpretive critique, not established evidence of any company’s intent. The more immediate problem is simpler and more testable. A pact to pace development needs terms: evaluators, access, thresholds, limits, monitoring, and consequences. Until those terms are publicly defined, the industry’s apparent consensus is best understood as a shared aspiration rather than a concrete slowdown.

Read the full piece: Don't Worry About the Vase →

AIUC has raised a $40 million Series A led by Ribbit Capital, with First Harmonic also participating, bringing its total funding to $55 million. The company is building a third-party audit and certification layer for enterprise AI agents, centered on its AIUC-1 standard.

Read more: TechCrunch →

OpenAI has reportedly acquired smartphone-camera software company Glass Imaging for more than $300 million, per multiple reports tracing to an initial Wall Street Journal account. Neither company has confirmed the deal or said how the technology will be used.

Read more: SiliconANGLE →

Microsoft has published a draft Humanist AI Code of Conduct for its MAI models, prioritizing human control — barring deception and resistance to shutdown. The document is still in development, not yet used in training, and open for a six-week public comment period before a year-end revision.

Read more: Microsoft AI →