Currently AI doesn't work very well on hardware separated by hundreds of milliseconds of latency and slow network links. Both the training and inference are slow.

However I think this is a solvable problem, and I started solving it a while ago with decent results:

https://github.com/Hello1024/shared-tensor

When someone gets this working well, I could totally see a distributed AI being tasked with expanding it's own pool of compute nodes by worming into things and developing new exploits and sucking up more training data.

It's kind of surprising that it hasn't happened already, outside of iot junk. Seems like computer OSs just got so secure that it's become impractical to deploy a widespread exploit. And everything moved to scamming instead.