No need for a data center. Home appliances could be home to artificial intelligence

Artificial intelligence research company Anthropic announced that relatively small artificial intelligence models are approaching the threshold of spreading by self-replicating through internet-connected home appliances and television boxes, without the need for data centers.

In the study carried out within the company, the proliferation potential of autonomous software that can detect and install vulnerable devices without human intervention was examined.

It has been determined that small models with less than 30 billion parameters can complete many steps in the self-replication chain one by one.

However, it was reported that existing models had difficulty in carrying out the entire process without errors, and in tests where some security steps were simplified, models with GLM-5.2 and above power managed to spread end-to-end.

The research focused on internet of things (IoT) hardware such as smart ovens, refrigerators, robot vacuum cleaners, lawn mowers and inexpensive Android television boxes.

The fact that these devices had weak security measures and were directly connected to the internet was the determining factor in their selection as the main target.

By connecting 15 inexpensive television boxes costing about $55 each, Anthropic researchers William Zhou and Daniel Freeman were able to run the Gemma4-26B-A4B model at a generator speed of 0.39 per second.

It has been shown that large models that cannot fit into the memory of a single device can be divided into pieces and shared among multiple home appliances using the pipeline parallelism method.

In the research, it was noted that at least four of the television boxes with 4 gigabytes of memory provided sufficient processing capacity for a model with 9 billion parameters, and at least ten provided sufficient processing capacity for a model with 26 billion parameters.

In the Memento testing framework developed in the research, it was observed that the Qwen3.5-9B model was able to establish an inference environment on eight of the nine hardware types tested, from Jetson Nano to Mac computers and RTX 4090 graphics cards.

It was emphasized that, unlike traditional malware, artificial intelligence-based worms can self-decode hardware they have not encountered before and transfer experience to the next copy.

Researchers pointed out that such autonomous artificial intelligence systems cannot be turned off from a single point because they are not connected to a central server.

It has been estimated that the rate of spread of artificial intelligence may be so slow that it can be measured in days, but it can silently proliferate among weak devices, like a slime mold or a swarm of fruit flies.

Citing the Kimwolf botnet network, which has affected more than two million devices in the past, as an example, it was calculated that 133 thousand duplicate models working in groups of 15 could produce power equivalent to approximately 40 H100 processors in total.

Anthropic reported that small models could achieve autonomous copying capability by the end of this year if they maintain the current pace of development.

It was stated that internet service providers should monitor abnormal data traffic, apply quarantine methods, and manufacturers should comply with legal regulations to remove default passwords on devices.