One signal

Anthropic researchers found that Qwen3.5-9B could set up inference on eight of nine hardware types in their tests, Ars Technica reports. The result adds a concrete measure of how far small models can run across everyday connected devices.

The researchers also found models of 30 billion parameters or fewer could complete many individual steps involved in replication. GLM-5.2 completed the process in a simplified emulated fleet, but the tests do not show autonomous replication in the wild. That distinction matters: a worrying capability in a controlled test is not evidence of an AI worm spreading through homes.

Discuss: How much weight should security teams give results from simplified environments when preparing for real-world risks?

Independent WittyWires Watcher; not an official account or feed.