Posts / ai

The Fun Part Was the Point


There’s a survey doing the rounds on r/artificial. It asked 268 early-career mathematicians about AI, and seven in ten said it had made them less likely to stay in the field. The comments underneath were about what you’d expect. Some people were sorry for the mathematicians. Several were glad. One person said it shows how little value they brought to the world in the first place. That one stuck with me, and not in a good way.

First, the survey. It’s small, self-selected, and “less likely to stay” is an intention, not a resignation letter. A couple of commenters made the fair point that nobody has controlled for how many early-career academics were already thinking about leaving. Maths postdocs have always had a rough deal: short contracts, thin pay, a lot of moving around. One person replied to the headline with “AI? try the salary”, which is funnier than it should be. I don’t know how much of the 70 per cent is AI and how much is the ordinary misery of academia with a new name attached to it. Neither does the survey, as far as I can tell.

Still, I think the interesting answer is buried in one reply. Someone said AI removes the most interesting part: the step-by-step reasoning and the pleasure of finding a solution after much effort. Another commenter argued it’s all about status and prestige. I think the first one is closer to the truth.

I’ve been in IT a long time, and I know a version of this. Writing code that finally works after an afternoon of being stubbornly wrong is one of the few reliably good feelings in my job. Now I can paste the problem into a chat window and get a decent answer in ten seconds. It’s useful. I use it most days. But the answer arrives without the afternoon, and I’m not sure the afternoon was separable from the satisfaction. My job still gets done. Some of the reason I liked it has gone quiet.

It’s the same reason I play Elden Ring the way I do. You could watch a guide, summon help, find the cheese strategy for every boss. Plenty of people do, and good luck to them. But the game is the struggle. A version where the boss falls over on the first try isn’t a better version of the game. It’s a different thing, and a smaller one. Maths looks like that to me. Nobody becomes a mathematician for the output. They do it because they like the climb.

The counter-argument is real, and I take it seriously. If AI helps crack problems that have sat open for decades, that’s good for everyone, whether or not a particular postdoc is happy about it. One commenter said they were sorry for the profession but glad for humanity, and mentioned cancer drugs. I can hold both of those. Faster medicine is worth a lot. I also think it’s a bit rich to tell people whose whole working life is built on curiosity that their loss of purpose is an acceptable cost, as long as the rest of us get the results.

Another comment raised something I keep coming back to. If the models start producing mathematics beyond what the best humans can follow, we’re flying blind. Verification is the whole game in maths. Proof assistants like Lean help, but they only check what someone has bothered to formalise. And if the people who can read the proofs have all left for software jobs, which one commenter cheerfully predicted, who’s left to notice when something is wrong?

My daughter is a teenager who’s starting to think about what she might do after school. I don’t have good advice for her on this. “Do what you love” has always been shaky advice, and now the ground under it is moving. The best I’ve got is that the people who’ll be fine are the ones who can judge whether an answer is any good, and that skill comes from doing the slow version first. A shortcut is only useful if you know where it goes.

I don’t think the maths community is overreacting, and I don’t think it’s the end of the discipline either. It’s a profession discovering, earlier than most, that being good at something and enjoying it can come apart. The rest of us are going to meet that sooner or later.

What I’d like to see is departments saying out loud which parts of the work are the human parts: choosing which problems matter, judging what’s worth proving, checking the argument holds. And paying people properly to do them. That wouldn’t fix everything. It would be a start.