Monday, November 16, 2026

Podcast on AI taking over mathematics

Yesterday I listened to a podcast from "Radiolab" about the recent AI mathematics revolution you may have heard about. 

Radiolab is a science-oriented show on America's public radio, NPR. (I guess we have to call these "podcasts" now.) This podcast(!) was originally released about a month ago. I found it very interesting--I didn't realize mathematicians now consider to be an AI model to be a "world class mathematician better than any human mathematician." Experts label it an "emergent property" -- this ability emerged not by being designed, but by AI's reading all the mathematical literature out there done by humans.

The big shock came this past May after an OpenAI model solved the unit distance conjecture, disproving it. Then there was a Navier-Stokes fluid equation conjecture that it also disproved, finding an example where the equations "blew up" (gave a singularity).

Here's the timeline the guest gave for AI's development:  

  • 2024 - AIs were doing high school level mathematicians
  • 2025 - AIs were as good as a very smart undergraduate studying mathematics
  • March 2026 - could do research level mathematics but was not the world's best mathematician
  • May 2026 - AI was solving problems that the best mathematicians had not been able to solve 

This is an incredible pace of improvement. In September OpenAI released a tranche of 700+ papers proving or disproving (they claim) a great many conjectures that had been made in the past.

Mathematicians are literally wondering if human mathematics is over.

Now ask, where will AI's math abilities be a mere year from now? 

Now imagine 10 years from now. Will AIs be writing about mathematics no human has ever imagined? That they can't understand? That they can't even make sense of? Will humans be able to understand its papers (which by the way have not submitted (I think) to peer review journals)?  

Scary. But also fascinating.

No comments: