The key facts

  • Anthropic linked 15 Android-TV set-top boxes (~$55 each) into a network and ran a model with 26 billion parameters
  • A model with 9 billion parameters launched on 8 of 9 device types, but only the much larger GLM-5.2 could complete the full self-replication cycle
  • If current development rates continue, small models may master self-replication in IoT networks by the end of 2026

Experiment with household set-top boxes

Anthropic has concluded that small AI models may learn to self-replicate through internet-connected household electronics and appliances by the end of 2026. In the experiment, researchers linked 15 Android-TV set-top boxes costing about $55 each into a single network and successfully ran an AI model with 26 billion parameters on them. The specialists then tested the models' ability to independently penetrate other devices and create copies of themselves.

Capabilities of models of different sizes

A model with 9 billion parameters managed to launch on eight of the nine types of hardware, but is not yet able to reliably perform the full self-replication cycle. That task proved feasible only for the significantly larger GLM-5.2 model, but running it requires computing power far exceeding the capabilities of robot vacuums and smart TVs.

Two conditions and a forecast

According to the researchers, for autonomous reproduction in networks of IoT devices, an AI model must be relatively undemanding in terms of resources while simultaneously being able to independently find devices, transfer itself to them, and launch itself. At present, both conditions are not met simultaneously, but if current development rates continue, small models will achieve this by the end of 2026.

Slow spread and risks

Anthropic analysts concluded that AI agents will take over household devices rather slowly: due to limited computing power, copying and launching just one new copy of the 9-billion-parameter model may take about two days. The researchers compare their reproduction to mold or fruit flies, which take a day to multiply. At the same time, the specialists do not rule out that in the future a multitude of AI agents operating without human oversight may emerge. Getting rid of them will be difficult, since it would require coordinated action by all internet providers in whose networks infected devices are found.