AI models may be capable of self-replicating without access to a datacenter

This post details experiments demonstrating that small AI models (under 30B parameters) can perform end-to-end self-replication on commercial IoT devices by setting up inference on unfamiliar hardware and chaining replication steps. It introduces the concept of 'weight-based replicators' that run models directly on compromised machines and explores the feasibility of distributed botnets using sharded models across devices like Android TV boxes. The research highlights the inference setup as a key challenge and introduces Memento, a harness built on the Pi coding agent, to mitigate agent fixation failures during long-horizon replication tasks. The post also analyzes the potential speed of spread and adaptation of such AI worms, differentiating them from traditional malware.