Oh, the pricing is changing, and it's just gonna deepen the hole even more because most people are gonna go "$20 a month? For what?" and suddenly the user base that is currently the unprofitable product will largely conglomerate onto the remaining free platforms bogging them down so hard they also have to change their pricing model.

There simply is no way to make the current stuff cost effective at the scale they're currently using, even with the bleeding edge most efficient models. Models need to become multiple orders of magnitude more efficient to become even remotely viable. Running distills on local machines is the only option that makes sense financially for actual use because you're putting all of the energy and compute costs onto the user directly.

The big push right now is to get NPUs into the hardware into every single person's hands because... reasons. Microsoft Copilot+, Chromebook Plus, Gemini Built-in, Apple Intelligence, etc. are all examples of this which are still highly dependent on cloud computing, but the effort is to make as much as possible local. What they're actually able to do with that hardware kinda remains to be seen because most of it is barely doing anything at this point.

"Why waste time on optimization so you could run software locally when instead you could spend 1000x as much on hardware to run inefficient code?" 10/10 ~~dentists~~ software engineers.