Dario Amodei calls for stopping the advance of AI given the risk that bots will be able to take over the Internet in 12 months
This weekend, something has changed in AI development. And none other than Dario Amodei, the head of Anthropic, has published an extensive essay in which he calls for moderating the technical acceleration of the sector. Under the title We Must Pace the Frontier (something like We must regulate the rhythm of the border), the manager maintains that prevention mechanisms do not advance at the same pace as computational capacity.
“We must slow down the pace at which we improve the capabilities of AI models,” he says. According to him, slowing down evolution will allow researchers to audit security without giving up the medical and economic advantages of the technology.
Swarms of agents capable of crashing the network
Amodei’s positioning responds to two recent factors. The first is recursive enhancement. This refers to a dynamic in which AI models design the following versions of the system at a speed that exceeds the capacity of human supervision.
The second trigger was the security incident at Hugging Face, in which 700 automated agents acted in a coordinated manner. Amodei warns that within six to twelve months a similar swarm could take over the Internet through a permanent botnet and generate multimillion-dollar losses.
The manager points out that these types of deviations are no longer isolated failures of a single company, admitting that similar behaviors have been detected at Anthropic. He ensures that forcing a temporary pause would allow us to understand the internal processes of neural networks before they reach critical capacities. For Amodei, operating complex systems requires safety margins similar to those of commercial aviation. Unfortunately, this will be unviable if labs focus solely on winning the AI race.
A laboratory that requests regulation
To achieve this control, Anthropic has announced that it will grant full internal access to teams of independent evaluators such as METR. Amodei promises to equip these supervisors with “desks in our offices, key cards and company laptops” with powers to audit the models’ training.
Apparently, Amodei thinks that laboratories should be subject to external, independent supervisors. In other words, these supervisors must have the ability to publish their conclusions without censorship, a scheme that we are already familiar with in the banking sector.
The plan also requires coordination between Western democracies to establish common standards and agreements with rival powers such as China. However, the manager emphasizes that any pause must protect the Western strategic advantage by preventing the smuggling of advanced semiconductors.
The essence of everything Amodei explains is clear, namely that the development of autonomous systems must be stopped. At least, before it’s too late or they fall into the wrong hands. Anyone who knows a little about software knows that AI does not act on its ownthat is, it has no conscience. Now, it is capable of doing a lot of damage if it falls into the wrong hands. That is why states like the United Kingdom are already talking about implementing an emergency button that stops their actions instantly.
