> Somehow no one talks about LLM speed.

When I've raised speeds about local inference I've been told 60-75 t/s is perfectly usable. It makes sense that people aren't talking about speed yet since you either already have a response fast enough to wait for, or you go do something else and check back in a few minutes.

I would love to wait for the latter type of tasks though, because those are typically the ones that require the most work from me to verify and I don't want my attention divided with multitasking.