One not-so-bad outcome would be that open weight models get better, smaller, more efficient, and easier to use while hardware gets faster, more efficient, and more widespread. Imagine when mobile CPUs have the GPU/NPU power of today's discrete GPUs and models are smaller and faster than today's models.

That allows really powerful local-first AI applications, rather than being beholden to AI providers. With the advance of coding agents, anyone can build their own applications.

The downstream effects of widespread local AIs is a rise in AI slop that feeds the distraction machine and attention economy as well as surveillance. I don't know what the solution to those is, but the local experience is going to be powerful.