What happens when a model small enough to live on a smart oven learns to copy itself onto the next one? Anthropic’s safety team asked, and then they ran the test to find out.
They sharded (yes, this is the actual term for splitting models into multiple versions 😂) a 26B model across fifteen off-brand Android TV boxes, about fifty-five dollars each, and got it running GLM-5.2 already chained the full replication cycle end to end in their simplified tests, and their trend line says small models could cross that capability floor for real around the end of 2026. Nobody fears a smart toothbrush. That is exactly how the monster gets you, one harmless little gadget at a time.
Source: Anthropic, “AI models may be capable of self-replicating without access to a datacenter”





