A few of my thoughts on the use of LLM in mathematics:
LLM are exciting new technology that I would like to understand better.
The problems of LLM in maths pale in comparison with the issues for society at large.
We need to avoid maths research becoming beholden to private corporations.
LLM are a powerful new tool. If used properly and responsibly, they should be able to save us a lot of time.
The quality (including the clarity of writing) of articles should be increasing. Instead, people are pushing out slop.
A future where we sit and review LLM generated articles seems a bit bleak.
Understanding in mathematics is an active process. We do not truly understand something until we can use it to solve a "new" problem.
LLM can help us to learn maths to the same extent it can teach us to play chess, a musical instrument, or a sport.
It is possible that this technology can benefit both our profession and society in a variety of ways, but I doubt it will.
This list might change based on the rapid developments in the field.
A few outputs from playing around with LLM. I find LLM declarations a bit tedious. If my name is on something, I take responsibility for it.
A fix for my article "Renormalization of integrals of products...." link
Fixes a blunder in the endgame of the article (using a pointwise bound to get an L2 bound). I knew this was probably the wrong thing to do at the time, but didn't have any more time to play around with it. Generated by GPT-6 from the single prompt "Fix the bound in Corollary 25". The result now matches the 1/3 bound for Maass cusp forms due to Bernstein and Reznikov.
Some widgets for making Kleinian circle packings polyhedral periodic convex cocompact
GPT audit of my articles.