Math/AI
Ideas, predictions, and hope for the future of research mathematics.
What's wrong with proving frontier models?
AI companies are using open conjectures to prove the intelligence of their models. We don't like it, but it'll probably keep happening for a little while. And even if the AI companies don't do it, it'll happen anyway as the newer models are released to the public. But we could turn it around - AI can push us to find harder and more interesting conjectures that the frontier models can't solve. And maybe that's good for everyone.
I've become more optimistic in the last month, and so this will read pretty differently than the last post. We'll see how things are in another month. I'm sure volatility will persist for quite some time.Where we go from here
After some conversations and the Navier-Stokes solution, I was gripped by anxiety about the future of my career and of research mathematics generally. What followed was then my first attempt at some idea of how research math could continue. The main idea is that of "human compression", which is kind of a riff off of what Terry Tao has been saying about canonicalization.
This post was highly influenced by Tao's writings, and might just be a rehashing of his stuff. I claim it's distinct, but in hindsight I'm not as sure.