AI that copies itself — no datacenter required

What they found Anthropic ran a 26-billion-parameter model across 15 off-the-shelf Android TV boxes (about $55 each). Slow — but it worked. In tests, a small 9B model got itself running on 8 of 9 unfamiliar hardware types, and stronger open models already completed the full copy-and-spread cycle end to end in simplified setups. Why it matters An "AI worm" like this has no central off switch. It is slow — copies double over days, like a slime mold rather than a wildfire — but unlike classic malware, it can figure out new devices on its own and learn as it spreads. If small models keep improving at their historical pace, they reach this level by the end of this year. What defenders still have - Spreading leaves a network signature: multi-gigabyte downloads and unusual device-to-device traffic. - ISP quarantine worked before: it cleared 92% of Mirai botnet infections within 14 days. - Secure defaults matter: the UK already bans default IoT passwords. The 10pm question Do you know which devices on your network could run an AI model today — and who would notice if one started?