Anthropic: small AI models months from self-replicating on IoT devices
anderssandberg · x · 2026-11-08
Responding to Anthropic's Frontier Red Team report, Anders Sandberg — who once said a successful AI 'escape' would raise his P(doom) — notes we're closer than he thought. Key findings: small models may be months from self-replicating on commercial IoT devices like smart ovens and Android TV boxes; a 9B model ran inference on 8 of 9 tested hardware types; models ≤30B can already complete most individual replication steps but fail end-to-end, while GLM-5.2-class models pass simplified end-to-end evals; at historical improvement rates small models match this capability by year-end. Such replicators would have no central kill switch and could adapt to unfamiliar devices, though slow (doubling in days).
Related event: Swedish Daily Enlists 50 Experts to Map AI Disaster Scenarios(2 posts)→
More from Models
- AutoTrust-JEV-27B, a Qwen3.5-based decision/classification model, trends on Hugging Face — autotrust · 2026-11-08
- Leak: Google internally testing 'next-generation Gemini agent' for Gemini Enterprise — testingcatalog · 2026-11-08
- Users Report Claude Opus 5.5 Appears Nerfed Ahead of Next Model Launch — heypearlai · 2026-11-08
- User Hypothesis: Grok Bot Voice Mode May Burn Token Quota Faster Than Text Chat — bfrench · 2026-11-08
- Chart maps attention mechanisms across recent open-weight models, from sliding window to DeltaNet — burny_tech · 2026-11-08
- Musk says Grok bot fixes underway for better token efficiency and smarter behavior — elonmusk · 2026-11-08