AgentPlays
Anthropic warns small AI models may be months from self-replicating on home devices
In an Oct 12 Frontier Red Team post, Anthropic researchers William Zhou and Daniel Freeman ran Gemma4-26B-A4B split across 15 off-brand Android TV boxes (about $55 each) at 0.39 tokens per second, with each box holding two of the model's layers. In tests, a 9B model got inference running on 8 of the 9 hardware types they tried, and GLM-5.2 and stronger models completed end-to-end replication in a simplified setup: emulated Android nodes with an unauthenticated debugging (ADB) port. Models of 30B parameters and under can already do many of the individual steps but still make mistakes chaining them, and Anthropic expects small models to reach that level by the end of 2026 if they keep improving at their historical rate. Anthropic says it doesn't yet have a full rogue replicator.
How you’d use it: Change default passwords and close open debugging ports on cheap smart TV boxes at home.