Anthropic Says AI Could Copy Itself Onto Your TV Box Within Months. It Ran the Demo on Fifteen $55 Boxes, Left the Doors Unlocked, and the Critics Say the Scary Part Isn't the Part It Proved
About This Episode
Anthropic's Frontier Red Team published research on October 13 saying small open-weight AI models may be months from self-replicating across cheap internet-connected home devices with no central point to shut off. It ran a 26-billion-parameter Google Gemma model across 15 off-brand Android TV boxes at about a third of a token per second, and in a simulated network with no passwords only the Chinese open model GLM-5.2 completed the full replication chain. Outside researchers called the experiment engineered to succeed; Anthropic's own post concedes the hack step was trivial and the directing model ran on a separate server.
Our Take
Every AI off-switch plan assumes the model lives in a data center somebody can unplug; Anthropic just showed a capable model running on fifteen $55 TV boxes, and the critics who say the demo was rigged are arguing about the timeline, not about who holds the plug.
Continue Reading on Unscarcity
Human-in-the-Loop: Where AI Agents Must Stop
The framework's last line of defense is a human who reads the logs, quarantines the weights and pulls the plug; Anthropic's replicator is designed around there being no plug, so the lens asks where the checkpoint goes when the fleet lives on hardware nobody answers for.
The Commoditization of Intelligence
Open weights as the anti-lock-in technology, cheap is what spreads, and the one model that completed replication is Z.ai's GLM-5.2; the lens explains why a near-frontier model is a downloadable file that fits on $825 of TV boxes and belongs to no one who could switch it off.