My vague prediction right now is that in five years LLMs will be heavily used by universities in grant-funded math research but nobody else will be able to afford it, much like supercomputer clusters 25 years ago.

Well, if progress in LLMs will steadily continue over next 5 years, then models will be so powerful that there will be no longer place for (most of) human researchers in math (remember that 5 years ago there was no chatgpt!). But I think it's more likely that progress will stall and then open models will catch up to frontier models and almost everyone will be able to afford them.

Seems way too binary a statement. I am guessing you mean "frontier LLMs". Small models keep getting better and better and if you make domain specific ones, it will likely be even smaller. Companies renting smaller LLMs or using enterprise models might very well remain in the future. Consumers getting LLMs whose performance dont improve (think gpt 6 forever on premium or gpt 4.x on a cheap tier) might well become a thing.

Sounds very good for regular joe software dev, almost too good to be true