SILO: Funsearch: a novel approach to using large language models as part of mathematical practice

I’ll report on a collaborative project with researchers at DeepMind, published here in which we find novel examples in extremal combinatorics, not by optimizing a function for an objective in a traditional way, but by using an LLM “one level up” to search the space of function-generating programs.  I’ll talk about both the promise and limitations of this approach in its current form.  Overall message: I think this has potential to be a very useful tool in mathematical practice (and I will try to pitch people in the room on working with me to build out tools of this kind) and a step towards interpretable machine learning (whatever that means.)

Bio:

Jordan Ellenberg is a math professor and frequent SILO-attender at UW-Madison.