Anthropic Ran AI on $55 TV Boxes and Tested Device-to-Device Replication
Fifteen cheap TV boxes were chained together to run an AI model, and in an environment sealed off from the outside, researchers tested moving a model onto other devices. Anthropic said that in a simplified experiment, GLM-5.2 completed the replication process from start to finish. The ability of a small model to replicate itself on a target device, however, was not demonstrated.
On October 12 local time, Anthropic published research on running small AI models on Internet of Things (IoT) devices. The work examines whether AI can generate responses on cheap hardware outside the data center and spread from one device to another. The researchers ran two sets of tests: one that chained together multiple low-cost devices to run a model, and another that moved a model onto other devices inside an experimental environment isolated from the outside world.
Can AI run on a TV box?
When a model is split across several devices, AI can run even on hardware with very little memory. The researchers linked 15 Android TV boxes, each costing roughly $55, to run Gemma4-26B-A4B, distributing the model’s layers across the devices. It was extremely slow, producing about 0.39 tokens per second.
The team also tested, separately, whether an AI model could autonomously set up an AI runtime environment on each device. A 9-billion-parameter model managed to build a working environment on 8 of the 9 hardware types tested. The researchers broke the process into stages: finding and gaining access to a target device, delivering the model and getting it running, and then spreading to the next device.
How far did the replication experiment get?
The replication tests were conducted on eight virtual Android devices created inside Google Cloud and isolated from the outside world. GLM-5.2 delivered Llama3.1-8B to each device and configured the runtime environment. In four of five trials it spread the model to all eight devices; in the remaining trial it reached seven. Importantly, the GLM-5.2 instance directing the work ran on a separate server, not on the devices themselves.
The access conditions were simpler than what compromising real-world devices would involve. The researchers used an Android diagnostic channel (the ADB port) that could be connected to without authentication. Anthropic said that in this simplified setup, GLM-5.2 completed the replication process from start to finish. That said, the ability of a small model to replicate itself on a target device was not demonstrated.
What distinguishes this from conventional malware is that the model prepares the runtime environment to fit each device. Mirai, the IoT malware that corrals infected devices into a botnet, spread using a fixed list of usernames and passwords. The approach Anthropic examined is closer to an AI working out each device’s environment on its own and adapting it so the model will run.
So is this something to worry about now?
This is research that used actual hardware to check two things: whether AI can run on the kind of cheap devices people use every day, and whether AI can handle the installation work needed to adapt to a given device environment. But the replication experiment took place in a virtual environment sealed off from the outside, and the AI directing delivery and execution sat on a separate server.
The thing to watch going forward is how much of that separate server’s role a small model can take over. The researchers estimate that if performance gains continue at their current pace, small models could reach the required capability level by the end of 2026. Whether the chain from access to execution to further spread holds up on real devices, and under what conditions each stage succeeds, still needs more verification.