AI-assisted research in mathematics II

Now everyone is talking about AI and its potential to reshape mathematics. Without adding my own opinion to this, I want to contribute by sharing how I am actually using it right now and what has changed since April. I have a lot of opinions on the different models, but I want to keep it brief.

  • I still have a Gemini subscription, but the new version of Gemini (Gemini Flash) is not a frontier model. It is just really fast and the subscription API tokens are cheap.
  • I pivoted to Anthropic. Fable is great, but too expensive to run. The new Opus 5 is my mathematical workhorse for more than just routine tasks.
  • arXiv digest: everyday I run a little AI-written script that fetches new preprints into a nice looking html page.
  • VSCode: I switched LaTeX editor to VSCode and have an AI agent running on the side.
  • Live-TeXing: I started live-TeXing with an AI agent, most recently for my lecture notes from Lausanne. In between, I tell it to ‘fix compile’. You can even be really sloppy and also use macros that do not yet exist. It can also on the spot generate something like this, a pair-of-pants cobordism from $M$ to $N_1 \sqcup N_2$:

I am leaving this in for being too confident.

  • Counterexamples: We are now all convinced that AI is good at that, in particular in fields with a large corpus of literature. Very helpful to check edge cases of your arguments.
  • some convenient things that you can do now: Let AI generate preamble for your TeX documents; a polished preamble (and .bib file) are very pleasing. There is an endless number of ways to facilitate typing a TeX document with AI.