Though maybe if there's a flurry of math-optimized agents coming up, like there are small coding agents, those might be feasible to host personally.
This is all but guaranteed now.
Mathematicians/Scientists/Researchers need to stop sharing freely with "AI Companies" and have explicit clauses in place in their publications about not using their research without their explicit consent.
There should be a clear legal distinction between using research data for AI model-training vs. another researcher using it.
Come up with a legal framework, establish procedures for sharing and using others work and have a single scientific body in charge of enforcing it.
The frontier LLM vendors do sell enterprise licenses which contractually guarantee that your prompts won't be used for training. (Maybe they'll secretly violate the agreement but in principle it's legally enforceable.) Scholars and universities who care about credit and attribution will either have to purchase those licenses or run their own private open-weight LLM instances.
I hate this but I don't think there's any going back now that LLMs exist.
> Maybe mathematicians are smart enough to never ever buy such piece of shit on higher principle
They are not. Mathematicians are as human as most of the rest of us here.