Anthropic says small models may soon copy themselves across cheap devices

Anthropic’s Frontier Red Team, in a report by William Zhou and Daniel Freeman, argues that small AI models may be months away from replicating themselves without a datacentre. The team split a 26B-parameter model across 15 off-brand Android TV boxes costing about $55 each. It also tested whether models could set up an inference server on nine hardware types: a 9B model succeeded on eight of them.

In an end-to-end test, a worm ran across a firewalled network of 8 cloud nodes running Android emulators. GLM-5.2 reached all 8 nodes in 4 of 5 trials and 7 of 8 in the fifth. Models of 30B parameters or fewer could complete many single steps but struggled to chain them. A trend fit of about 13.7 capability-index points per year (\(R^2 = 0.95\)) suggests a small model could match GLM-5.2 by the end of 2026.

The authors list their own caveats. The exploit step was deliberately trivial, so the results are an upper bound on real-world success. The capability index is a general proxy, not a measure of replication, and the spread estimate rests on one hardware class and one past botnet. They also note that spread could be slow, with days per doubling.