Pasting a broken script into ChatGPT and getting a working correction back felt like cheating. I was quite happy to spend less time fixing a bad script.
So I kept using it.
After a few more useful experiences, I started to understand the scale of what I was looking at. I had welcomed help with one tedious part of programming. There was no reason its capabilities would stop at the parts I wanted to hand over.
That was a shock. This was reaching into my whole understanding of programming, and into the career I had built around it.
Finding AI useful changed how I worked. It did not settle what I wanted to preserve about programming. I think that is a reasonable distinction for mathematicians to make too.
Reading mathematicians discuss AI, I recognize something in their wariness. I cannot speak for every mathematician. I find it difficult to dismiss a reaction I have felt myself.
The help arrives before the larger question
The first decision was small. I wanted the script fixed, and I could check whether the answer worked.
The larger question arrived through repeated use. Each useful result gave me another reason to come back, and another reason to reconsider what I thought required my expertise.
I still use coding agents. I replaced ESLint with Biome because Claude Code or Codex could finish a patch before the linter finished checking it. By then, I was making the rest of my workflow keep up with the agent.
That saves me time. It also means I am handing over more of the work through which I learned to think.
A correct result still needs people who understand it
On October 6, OpenAI released mathematical results produced by an internal model. Its repository includes preprints and formalized proofs. I have not independently checked the mathematics.
Jeremy Avigad’s guest essay on Terence Tao’s blog describes a profession reconsidering its daily work as results that once supported a respectable publication become easier to generate with AI. He also describes mathematics more broadly as a culture of rigorous reasoning and communication.
Solving a problem gives people something useful. Working on it can also be how they develop skill and find collaborators. Getting the result faster changes how much of that work remains for them to do.
In his October 6 thread, Tao describes what traditionally follows a breakthrough: talks, workshops, questions, new connections, and proofs becoming easier to understand and teach. Producing the answer starts a much larger exchange.
Those exchanges let someone else question a step, learn the method, or apply it elsewhere. If the process stops at the result, those opportunities do not arrive automatically with it. AI could help with the explanation too; that still requires someone to make understanding part of the work.
Programming has its own version of that exchange. I can read an implementation, ask why it was designed that way, and learn something I can use elsewhere. Accepting the patch gets the change shipped. Understanding the choices helps me maintain it.
The comparison has limits. Software tests and mathematical proofs do different jobs. I am drawing a connection between people’s relationships with their work, rather than claiming the fields will change identically.
Using it does not settle what I want to keep
I do not think someone has to stop using AI to take that change seriously. I also do not think finding the tool useful makes every reservation disappear.
Use it, and you may feel that you are participating in a change that unsettles you. Avoid it, and you give up useful help while the field may continue changing around you. That is the damned-if-you-do, damned-if-you-don’t feeling I recognize.
Someone using an agent may be trying to finish necessary work or reclaim an evening. Someone hesitating may be worried about how they will teach the work once fewer people do it themselves. Finding the tool useful does not answer that concern.
I prefer getting the script fixed. I would still rather ask mathematicians what they want to preserve and what the tools could help them explore than assume they have not tried a good enough model yet.
What happens after the answer?
Tao’s Math 2.0 proposal gives explanation, community building, and new directions more weight alongside raw problem solving. He sees AI contributing positively to those activities too.
OpenAI’s October 15 update also describes support for a Simons Foundation initiative to help mathematicians understand and build on AI-produced results. That is a concrete response to the question of what happens after a result appears.
In my own work, I can ask an agent to compare approaches, explain an unfamiliar design, or help test a prediction. I can also accept a patch and move on. The patch may ship either way. What I learn from it depends on what I stop to understand.
I do not know whether mathematics will follow programming along the same path. Repeated usefulness changed how I worked before I understood how much it might change the work itself.
I still want the help. I do not think mathematicians are wrong to ask what comes with it.
Was this useful?
Thanks for the feedback.