Why is Anthropic's own CEO calling for a pause?

Dario Amodei says he didn't fully grasp what exponential growth would actually feel like once it arrived. Now he's telling that it's a warning sign the industry can't ignore.

The Anthropic CEO sat down with Jo Ling Kent at the company's San Francisco headquarters for an interview airing on "CBS News Sunday Morning".

He described AI's trajectory in blunt terms, doubling from one to two, four, eight, sixteen, and thirty-two, and said the curve is now visibly bending upward. "It doesn't mean we need to panic today," he said. "But I would say it's a warning sign that we need to slow down."

Amodei's comments follow directly from former Anthropic researcher Jacob Coxon's resignation earlier in the week, in which Coxon accused Anthropic and OpenAI of "gambling with our lives" in the race toward self-improving systems.

Coxon told CBS separately that the risk "doesn't look that different from... 'Terminator'", arguing a sufficiently advanced intelligence would be capable of killing humans outright.

Amodei didn't dismiss the concern, but he also drew a clear line: slowing down doesn't mean Anthropic stops releasing models. It means every release gets tested more rigorously than the pace has allowed so far.

Amodei has provided a specific solution that involves giving independent inspectors “employee-like” access to Anthropic’s models in the way that a food inspector can confirm that the firm is fulfilling its safety obligations.

The AI expert understands that his proposal does not constitute a complete solution on its own. He believes in the need for federal regulations while opposing any kind of outright prohibition of AI technology development.

The reason is that the potential of the technology is real in the field of medicine and that authoritarians will continue to develop it regardless of what democracies do.

The question Amodei can't fully answer

When asked if the United States should slow down even if China does not, Amodei pointed out that it was the hardest part of the whole conundrum, citing agreements on the control of nuclear weapons during the Cold War era as an imperfect example of possible future collaboration, despite having low hopes of achieving it.

His views on responsibility were even more far-reaching than those usually expressed by CEOs in such a position: he admitted that he himself felt uneasy about artificial intelligence research being conducted in a privately-owned corporation, suggesting some kind of joint governance by democratic nations instead.