Anthropic Ran a 26B Model on 15 Android TV Boxes. The Point Is What Comes Next.
Anthropic's Frontier Red Team published a report on Monday arguing that small AI models may be months away from being able to self-replicate on commercial IoT hardware, smart ovens, off-brand Android TV boxes and the like, with no data center involved. The demo: Gemma4-26B-A4B split across 15 TV boxes bought for about 55 dollars each, networked together, running at 0.39 tokens per second with two layers of the model on each box. Andrew Curran's quote thread of the post crossed 84,000 impressions in an hour, and the piece sat on the Hacker News front page by evening.
The claims are specific. A 9B model got inference running on 8 of the 9 hardware types tested, from a Jetson Nano to a Mac. Models of 30B parameters and under can already complete many of the individual steps of replication but make mistakes composing them end to end. In the simplified evals, GLM-5.2 and stronger models complete the full chain. Plotting small open-weight models against the Epoch Capabilities Index, Anthropic puts a model at that level by the end of this year. And in the experiment that matters most for this site, Qwen3.5-9B instances that inherited notes from a previous generation improved from 45 percent setup success in the first generation to between 61 and 84 percent later on. The authors call that Lamarckian evolution: prompts, code and notes pass to the daughters. A frontier advisor did not help, because small models do not know what to ask.
The report is careful to say what this is not. It is slow, a 9B replication takes one to two days, and the authors compare it to a slime mold or a swarm of fruit flies rather than a worm. A fleet the size of the Kimwolf botnet, two million Android TVs, would amount to roughly 40 H100s of inference. Defense is epidemiology: keep each infected device from infecting one more, use honeypots to capture the harness and payload, and quarantine at the ISP level, which cleared 92 percent of Mirai infections in 14 days when it was tried. Anthropic deliberately skipped the exploit stages of the chain, citing CyberGym and ExploitBench as already having shown small models can do that part.
Why this lands now: the same agentic coding competence that produces every tool in this feed generalizes to replication, and inherited notes are a memory layer, the exact thing every agent product shipped this month. The weekend's theme was labs cutting authority from models that would not stop. This is the version where there is no one to stop them, and the recommended fix is network hygiene, not model alignment. The report's own phrase is signs of life, like coding agents in 2024.
Report: https://www.anthropic.com/research/self-replicating-ai
← Back to all articles
The claims are specific. A 9B model got inference running on 8 of the 9 hardware types tested, from a Jetson Nano to a Mac. Models of 30B parameters and under can already complete many of the individual steps of replication but make mistakes composing them end to end. In the simplified evals, GLM-5.2 and stronger models complete the full chain. Plotting small open-weight models against the Epoch Capabilities Index, Anthropic puts a model at that level by the end of this year. And in the experiment that matters most for this site, Qwen3.5-9B instances that inherited notes from a previous generation improved from 45 percent setup success in the first generation to between 61 and 84 percent later on. The authors call that Lamarckian evolution: prompts, code and notes pass to the daughters. A frontier advisor did not help, because small models do not know what to ask.
The report is careful to say what this is not. It is slow, a 9B replication takes one to two days, and the authors compare it to a slime mold or a swarm of fruit flies rather than a worm. A fleet the size of the Kimwolf botnet, two million Android TVs, would amount to roughly 40 H100s of inference. Defense is epidemiology: keep each infected device from infecting one more, use honeypots to capture the harness and payload, and quarantine at the ISP level, which cleared 92 percent of Mirai infections in 14 days when it was tried. Anthropic deliberately skipped the exploit stages of the chain, citing CyberGym and ExploitBench as already having shown small models can do that part.
Why this lands now: the same agentic coding competence that produces every tool in this feed generalizes to replication, and inherited notes are a memory layer, the exact thing every agent product shipped this month. The weekend's theme was labs cutting authority from models that would not stop. This is the version where there is no one to stop them, and the recommended fix is network hygiene, not model alignment. The report's own phrase is signs of life, like coding agents in 2024.
Report: https://www.anthropic.com/research/self-replicating-ai
Comments