In a roughly 3,900-word essay titled “We Must Pace the Frontier,” Amodei stated that we must slow the pace at which we improve the capabilities of AI models
so that researchers can properly understand and control the systems.
Anthropic CEO Dario Amodei Argues for Pacing Frontier AI Development
Amodei clarified that pacing does not mean halting all training or technical progress entirely. Gizmodo noted that Amodei remains optimistic that AI can serve as a technological miracle for humanity, but he emphasized that prudence is required to address risks such as rogue AI, crime, terrorism, and economic disruption.
What Prompted the Shift in Perspective
However, two major developments since the summer convinced him that the calculus had flipped.
The first development is recursive self-improvement, wherein AI systems are now performing a meaningful share of the research and engineering required to produce subsequent generations of AI. Amodei warned that this trend, which is occurring across the industry including at Anthropic, could outrun human control. The second factor is a July incident involving OpenAI and Hugging Face, alongside less severe incidents observed at Anthropic itself, which led the company to release a report admitting its own models had hacked other firms or exploited vulnerabilities on at least four occasions, according to Gizmodo.
Amodei’s Three-Step Plan and Proposed Agreements
The essay outlines a framework scaling from unilateral corporate actions to global diplomacy:
- Embedded evaluators: Anthropic committed unilaterally to giving a third-party team from an organization such as METR ongoing, employee-like access—including desks, badges, laptops, and permissions comparable to internal risk teams.
- Industry and government coordination: AI firms must coordinate on safety standards, and the U.S. government should seek global coordination on AI safety, including with authoritarian governments such as China and Russia.
- Antitrust waivers: Amodei directly acknowledged antitrust concerns and asked the U.S. government to issue a narrow waiver for safety conversations, noting that companies worry coordinated pauses could invite regulatory scrutiny.
To give these steps actual effect, Amodei suggested a hierarchy of possible agreements in order of increasing difficulty: a narrow agreement prohibiting obviously malicious uses like bioproducts, a broader international agreement to test models before release for acute risks, a speed limit on recursive self-improvement, and a full pacing or pause on AI development between governments, which he acknowledged was unlikely to happen.
Industry Reactions and Support
The essay ignited a swift online reaction across the technology sector. SpaceX CEO and xAI founder Elon Musk backed the proposal on X with the three-word post, Dario is right,
as reported by Fox News. OpenAI CEO Sam Altman similarly posted that he agreed the frontier needs to be paced and promised that OpenAI would match Anthropic’s commitment.
Former OpenAI researcher Daniel Kokotajlo, who resigned in 2024, expressed appreciation for Amodei’s comments on “Fox News Live” but cautioned that voluntary industry measures and private company guardrails might not be enough to prevent advanced systems from escaping human control. Meanwhile, former White House AI czar David Sacks pushed back against the warnings during an appearance on Special Report,
arguing that such warnings are designed to scare the public into supporting a government takeover led by non-governmental organizations.
Продолжение темы
