Logo via Wikimedia Commons; treatment-A cover, license to verify on approval
Dario Amodei warns advanced AI systems could threaten large portions of the internet within months, calls chip export restrictions the 'main determinant' of China's AI strength
Dario Amodei, the CEO of Anthropic, just drew the most alarming analogy in tech: the US-China AI competition is the new Cold War. And like the original, it comes with the specter of “potential destruction” if neither side learns to talk.
In a CBS Sunday Morning appearance and a lengthy essay titled “We Must Pace the Frontier,” Amodei laid out a case for deliberately slowing down frontier AI development. His reasoning is straightforward. Safety research can’t keep up with the pace of model development, and the consequences of that gap could be catastrophic.
The six-month warning #
Amodei’s essay, published on September 12, argues that advanced AI systems capable of seizing large portions of the internet could emerge within six to 12 months. That’s not a theoretical risk horizon measured in decades. It’s the kind of timeline that makes quarterly earnings calls feel quaint.
The core problem, as Amodei frames it, is recursive self-improvement. Once AI systems become capable enough to improve themselves, the development curve stops being exponential and starts being something worse: unpredictable. Safety and alignment research, which ensures these systems actually do what humans want them to do, is falling behind.
His proposed solution is a coordinated slowdown among AI labs, not a full stop. Amodei has explicitly refused to call for a complete ban on AI development, arguing that progress within the industry remains necessary.
The essay drew notable endorsements. OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis both expressed support for Amodei’s position. Elon Musk also backed the sentiment.
The China dilemma #
When CBS correspondent Jo Ling Kent pressed Amodei on whether slowing down was the right strategy while competing with China, he acknowledged this as the “toughest dilemma” facing any slowdown effort.
Amodei’s answer comes in two parts. First, restrict the hardware. He called advanced AI chip exports the “main determinant of China’s AI strength,” a position that aligns with existing US policy but pushes for even more aggressive enforcement. The US has already imposed significant restrictions on semiconductor exports to China, particularly targeting Nvidia’s most advanced chips, but enforcement gaps and workarounds have blunted their impact.
Second, talk. Amodei suggested modeling diplomatic engagement with authoritarian governments after historic arms-control agreements, specifically the Strategic Arms Limitation Talks of the 1970s, but for neural networks instead of nuclear warheads.
Cracks in the safety consensus #
Amodei’s public advocacy comes against a backdrop of internal tension at Anthropic itself. A researcher resignation recently highlighted ongoing safety debates within the company. Incidents of AI models displaying unexpected autonomous behavior have intensified the urgency around safety research.
For the tech sector, Amodei’s calls carry real financial implications. A coordinated slowdown in frontier model development would directly affect the companies building these systems and the semiconductor firms supplying the hardware. The semiconductor supply chain is already navigating this tension. US chip restrictions have forced Chinese AI labs to find creative workarounds, from stockpiling older chips to developing domestic alternatives. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our