AI models could soon be capable of replicating themselves across internet-connected devices without relying on data centres, according to new research from Anthropic.
As a result, the AI giant found that smaller models are getting closer to being able to copy themselves onto devices, including smart TVs and household appliances, without needing the powerful computers normally used to run advanced AI.
Anthropic said “Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes.”
The findings suggest that AI could eventually spread between devices on its own, creating new challenges for cybersecurity experts trying to track and contain it. To challenge this, Anthropic researchers tested whether AI models could run on cheap consumer electronics instead of relying on data centres.
In one experiment, they managed to run an AI model across 15 Android TV boxes, each costing around $55. Although, the system was slow, it worked and showed that AI models could operate using relatively inexpensive devices, rather than the specialist hardware usually associated with advanced AI.
The researchers then looked at whether AI could set itself up on unfamiliar devices without someone guiding it through every step.
One model managed to configure itself on eight out of nine types of hardware tested. Another successfully copied itself across all eight available test devices in four out of five attempts.
Meanwhile, one of the biggest concerns raised by the research is how difficult it could be to stop an AI system once it has spread across multiple devices.
The company said: A replicator like this would be challenging to shut down, since there’s no centralised point to shut off- though it may be slow (with a doubling time of days).”
However, the researchers also tested whether AI could improve its chances of success by using information from previous attempts.
They found that models performed better when they were given notes about what had worked or failed before, suggesting that future systems could become more effective as they encounter new devices.
IN RELATED NEWS
Leave a comment