In the essay, published on September 12, Amodei asked frontier AI companies to pace how quickly their models gain new capabilities, so that safety measures have time to catch up. Without that, he warned, AI could within six to twelve months be capable of leading a swarm of agents that takes over the internet.
The proposal calls for embedded independent evaluators inside leading AI companies and governments, coordination among companies in democratic countries, and eventually an international agreement that includes China.
Amodei’s said they should look at their own record and assume Anthropic is not perfect; organise the rest of the industry around common standards rather than going it alone; and add an international component, on the view that the problem cannot be solved within one country’s borders.
When asked by Benioff on what had surprised him most over a decade in the field, Amodei pointed not to the technology but to the pace of capability gains and of the economic pull drawing the tools into everyday use. He tied it to the scaling principle he helped formalise, that models grow more capable as more compute and data are poured into them, and said what he had underestimated was how quickly that would translate into commercial adoption of newer offerings such as MCP, Cowork and Claude skills.
Salesforce is both a major Anthropic customer and an investor.
The writer is in San Francisco at Salesforce’s invitation.