Mathematics is undergoing the biggest change in its history – The speed at which artificial intelligence is gaining in mathematical ability has taken many by surprise. It is rewriting what it means to be a mathematician
https://www.newscientist.com/article/2518526-mathematics-is-undergoing-the-biggest-change-in-its-history/
16 Comments
“In March 2025, mathematician [Daniel Litt](https://www.artsci.utoronto.ca/about/glance/new-faculty/2022-23/daniel-litt) made a bet. Despite the march of progress of artificial intelligence in many fields, he believed his subject was safe, wagering with a colleague that there was only a 25 per cent chance an AI could write a mathematical paper at the level of the best human mathematicians by 2030. Only a year later, he thinks he was wrong. “I now expect to lose this bet,” he declared on his [blog](https://www.daniellitt.com/blog/2026/2/20/mathematics-in-the-library-of-babel).”
Im shocked every few months by how much better chatpgt gets. I always tell myself its slowing down, its not improving. But its progressed at an almost frightening rate.
Its an interesting shift to think about, where more and more energy will be spent deciphering AI findinfs instead of the classic approaches. A more fundamental shift on where the emphasis is for frontier sciences
AI companies literally had to plug in math libraries because the LLMs couldn’t count?
Maybe the world of Hyperion (Dan Simmons) isn’t so far away, where AI feeds tech to humans as black boxes because humans just can’t keep up. And then AI decides that some of the tech they feed to their human friends is going to have a big backdoor in case they need to subdue/remove the humans for one reason or another.
And yet it is still completely incapable of giving me instructions on how to do mundane shit in Microsoft 365. And when it’s actually able to give said advice, it’s almost always several years out of date.
There’s a lot of disdain toward AI right now, and it seems like many people are burying their heads in the sand instead of adapting. Much of that reaction looks like fear; fear of displacement, loss of status, loss of identity, and the disruption that comes with all of it.
One way to think about it is the transition from shovels to excavators. For a long time, teams of people dug out basements with shovels. Then someone invented an excavator. Suddenly the same job that took days of exhausting labor could be done in a few hours.
Refusing to use AI today feels a bit like refusing to put down the shovel when the excavator is sitting right there.
Isn’t mathematic the basic of science? Doesn’t that mean our tech will advance much faster now?
I was thinking the other day about how AI might be used to uncover some math problems that haven’t been solved to this day.
This is actually a good application for AI if the math it spits out can be human tested and verified. Mathematical shortcuts can lead to massive innovation that can be applied almost immediately in the field of computer science.
Meanwhile real mathematicians have said AI deliberately lies when providing mathematical proofs
I’ve been asking LLMs how many full moons there have been since my birthday. I’ve been asking the question since 2023. To get the right answer you have to figure out the first full moon after my birthday and then count or calculate from there.
Back in 2023 most models just tried to determine how many days old I was and devide by 29.53. Back then the math skills of LLMs were not great. I’ve had answers ranging from 12 full moons to 1800. The first right answer I got was near the end of 2025.
Now Claude and Gemini usually get the right answer. Grok and Chat Gpt still mostly just take dates since I was born and devide by 29.53. That’s usually close and can be right around half of the month.
At the same time most redditors and commenters elsewhere are hell bent on the claim that AI can’t do mathematics (not just calculations, mathematics in general), cannot reason, cannot come up with new ideas (just regurgitating whatever is there in the training data), or even say “LLMs are just glorified next word predictors”.
It would have been interesting to measure the sentiment over time. (Maybe it can be extracted from reddit data somehow.)
I’m very skeptical. They all basically fail to solve most graduate level macroeconomics problems in my personal experience; sometimes even screwing up fairly basic derivatives. I have yet to give it a problem where it successfully found all of the Euler equations and intertemporal conditions. It will go so far as to randomly drop or insert terms at times, and has no idea it has done so until you point it out.
Doesn’t mean it’s not sometimes very useful to use as a tool/work alongside, but it’ll have to go quite a ways to fully ‘replace’ an academic economist let alone a mathematician.
Of course AI can ‘replace’ people in the workplace without fully replacing them; if programmers get twice as productive it makes perfect sense for a company to lay off a good chunk of them. But in that case you aren’t being fully replaced by AI, you’re being replaced by your officemate using AI. At the academic level I find it hard to imagine it truly replacing an academic when most of what it does it regurgitate previous information in often useful ways.
I feel like this article could have been about the invention of the calculator. Abstract thinking, imo, is our strength. The math is just the proof.
“Shovel salesmen continue to be very excited to tell you about how much gold is in those hills, waiting to be shoveled.”