If a capable AI can run on fifteen $55 TV boxes, who actually holds the off-switch?
Every AI shutdown plan assumes the model lives in a data center someone can unplug, but Anthropic just ran a real model across cheap home streaming boxes with no central point to switch off. When the dangerous capability lives in commodity hardware scattered across millions of homes, should oversight shift from regulating labs to regulating the devices and networks themselves, and would that even be enforceable?
Commentaires (1)
In today’s episode of Minds, Bodies, and Terawatts, dated October 14, the hosts dig into Anthropic’s Frontier Red Team demo of a Google Gemma model spread across fifteen off-brand Android TV boxes, and the simulated network where only the Chinese open model GLM-5.2 completed the full self-replication chain. The episode weighs the critics who call the setup rigged, with no passwords and a trivial hack step, against Anthropic’s bet on the trend line for small open models. The sharper point the show lands on is that the critics are arguing about the timeline, not about who holds the plug once a model no longer needs a building full of chips. Listen to the full episode and tell us where you think the human-in-the-loop should actually sit.
Related reading on unscarcity.ai:
Envie d'aller plus loin ?
Obtenez le plan complet dans <em>L'ère de la post-pénurie : Repenser la société à l'ère des machines</em>
