I still think/hope/pray the future will be on-device models that don't need constant retraining. That will blow up the existing business model but I think a company could still make good money with a "majority local/remote for the really challenging stuff" model.

The problem is that today's AI companies have taken on so much funding that a reasonable, not crazy profit ratio isn't enough for them.

The future is already pretty much here. Note the recent stories about Chrome adding a local model, not to mention the Googlebook demo (if it works as advertised, there's a 0% chance you could get that kind of latency with a non-local model).

If it continues to be a numbers game - the more resources you throw at it, the better it is - then on-device is always going to be not as good. I guess it might be good enough for some uses?

I kind of loathe the move away from a world where we could control our own computers and run our own software on them.