India, Oct. 13 -- Anthropic warned Monday that increasingly capable AI models could be only months away from autonomously replicating across internet-connected devices, potentially creating AI-powered botnets that could evade shutdown attempts and seek more computing resources.
The company said experiments showed that AI agents can combine the computing power of inexpensive connected devices and replicate across a simplified test network.
Anthropic said models with 30 billion parameters or fewer can already perform many of the individual tasks needed for self-replication, although they currently struggle to complete the entire process reliably.
googletag.cmd.push(function() { googletag.display('div-gpt-ad-1568189509843-0'); });
In tests...