When I look at the current capacity of AI, it really feels like there's so much potential for frontier models to be used that just aren't being applied yet. Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness, and a good deal more expensive as well.

This is a strange conclusion. AFAIK in mathematics it doesn't work yet like article says: there is a problem, point agent swarm on it, and mark it done. The math results from OpenAI and Anthropic were got by seeding the surface of known problems and randomly trying to grow solutions; while they got some good solutions indeed, it doesn't mean current LLMs will work for arbitrary math problems.

Correct me if I am wrong on this, I am only a student of mathematics and by no means a researcher.

For the programming part, I think author already says that automated solutions need a good harness, and this is indeed correct, but in my experience building good harness is hard and agents can not do this currently without human closely guiding them. And in pre-LLM programming practice we did not need harness at all, or needed only small part of it: so, solving programming problems with LLM involves more work than without LLM, even if big part of this work can be automated.

The rest of the article seem to be a speculation mostly based on the premise that LLM can really replace human in the loop, which seems to be not the case currently.