Hacker News Reader

Comment by boerseth | Hacker News Reader

It's a little funny how quickly the conversation goes from "AI will never be able to {my-field}" before turning into "We cannot have AI ruining {my-field} with its obviously superior abilities"

If OpenAI had never made a push to get good at maths, we might still have a sizeable group of peopne stuck at that first line. I say that because I still hear in discussions the naive rebuttal "But we'll still need _someone_ to {one-step-up-in-hierarchy}" . Although you have to be curious now how long that sentiment can last.

The good news for those doing abstract maths might be that, now that OpenAI have proven that their AI can do the job in principle, there's probably not much left in it for them to keep going. Mathematicians should know better than anyone the lack of a direct profit motive in doing the work that they do.

(Edited for clarity)

View on HN

Hacker news is not really a good place to talk about nuances involving academia. I don't expect VC minded hackers comprehend how someone can devote years of their lives working minimum wage trying to answer questions which earn them a fancy title on a piece of paper, for that matter.

Or hone their craft to best express their art and share their human experience

A lot of us here are VC minded hackers *because* we spent years of our lives working minimum wage jobs in academia and then eventually realized we preferred to build!

I am one of those. I'll also say I never felt the kind of emotion in software building that I did when working on PhD problems.

Or maybe to earn proper money?

> A lot of us here are VC minded hackers because we spent years of our lives working minimum wage jobs in academia and then eventually realized we preferred to build!

This is a disingenuous framing. I've "built" more in academia than in some jobs I had, and some VC-backed startups are renowned not for building but for their lack of ethics.

"Building" is not the discriminating factor.

Building things which you couldn't convince people to pay you for could be might as well building dubious crap

Man, majority of Sillicon Valley production is dubious crap with not much other utility or even actively harmful. And someone pays that.

Earning money is not curelated with usefullness these days. It is somewhat correlated with lack of ethics.

I think it's a very skewed form of ethics to make people pay without convincing them.And fwiw Making money is correlated with productivity which is negatively correlated with a lot of bad things so I think on pure technical point you are wrong, there are subtleties.

Reward alignment is the differentiator. Obviously it's very far from clear-cut what is worth funding in research. But i've seen a lot of crap research, that would have died off far sooner if it had some VC/free market equivelant.

Or preferred to eat. I love teaching, and I'm good at it. If it were more valued by society then I'd do more of it. As it stands, I'm in adtech, and I teach a little piano and math on the weekends.

I spent years programming for myself, for free. I did this because I loved it, not because I was interested in the VC game.

But you did something else for a paycheck in the meantime, right? They're talking about what mathematicians do for their meager paycheck.

> But you did something else for a paycheck in the meantime, right? They're talking about what mathematicians do for their meager paycheck.

Grigiri Perelman, who was awarded the Fields Medal, notoriously refused the prise because he stated no other recognition was needed.

It's great there's an exception, but that isn't the point, they're talking about people who get paychecks as mathematicians. BTW what did Grigori do for a paycheck?

A large fraction of people are open source contributors, hobby game devs, and various other types of people who fall under "devote years of your life with no formal remuneration"

The US-American mind cannot comprehend "free time" nor working hard on something just for the love of it and the joy of collaborating with other like minded people.

There's a reason that so many open source builders and contributors are unemployed people in social-welfare European states. There's a reason that nearly every great work in the English canon was written by someone who was on the dole at the time, from Nietzsche to Harry Potter.

It's not like you need to raise at least $5m to create an account here. There's a mixture of people.

The vast majority of users here are nerds first, business enthusiasts much later.

I'm of two minds about this. On the one hand, you get miraculous results with AI, and that is what you are referring to with "be able to {my-field}". On the other hand you regularly get useless, if not toxic, crap out of the miracle machine. If you work with AI, you know these moments: "no that is not what I meant at all!" [1] or "how could you have already forgotten the central motivation I laid out just ten back-and-forths ago?". We're calling this the jagged frontier, but the jaggedness itself is not the problem, it's how deep these cuts go into aspects that are important. These are, why AI is still not "obviously superior" in its abilities, can't replace mathematicians and software developers and lawyers and everyone else who is primarily working with text. (It's also why people say it is ruining their fields, instead of replacing them in their fields, because it only replaces some part of the work, and what's left is not tenable or pleasant (e.g. doing reviews all the time).)

Over all, I'm not quite sure whether I see these deep cuts occurring less and less with progressing AI. It might just be, that I'm giving them progressively harder tasks as well. But the fact of the matter still is that they aren't there yet and it needs proof they will get there ("there" being able to do competent work that is not strongly limited in their scale by human review and correction). So I'm still not convinced about the point you want to make, even though I see AI disrupting about every part of society, and am hoping for the benevolent super AI to lead us to utopia (and afraid of the cliff of human dictatorship on one side and extinction on the other).

[1] I've started to use these moment to reflect on whether a human in a comparable situation would fare better. Having only a few text inputs and small glimpses into a code base and knowing next to nothing about you or your project personally. Quite a few times I've concluded: No, they would not. But more than 50% they would, and also, these limitations are technological limitations of LLM, so you should probably not excuse them for that.

But isn't part of the problem that you simply don't have access to the hardware larger organisations do? OpenAI certainly run into the same constraints. but if they have access to terabytes of memory, while when you use the same model it's closer to 32GB, that will surely affect how the model behaves?

So the pain points you and I frequently run into, simply don't exist for a well connected individual(Sam Altman).

bullshit. such a machine would earn a profit no matter its cost, and Sam wouldn't be at the helm of a dog of a company that won't make it to 2030

> If OpenAI had never made a push to get good at maths

By snooping on work math researchers were doing on their platform and then overtaking them when sufficiently close to completion? isn't that what happened i'm not sure

No, as far as I know there's no evidence they used the other team's work for their own (and OpenAI changed their initial 'It's possible' to 'we double checked and it didn't happen'). What they did was hear a rumour about it and then spent millions of dollars on compute to scoop them.

That's not what happened. The claims of this for Navier-Stokes were already exagerated. It was dramatic but there was AI involved on both sides to make progress and the final proof did not use the same approach.

Finally, the list of hundreds of solved problems after that completely defeated any argument in that direction.

I don't think i disputed that LLMs were involved assuming that's how OpenAI would be able to snoop in the first place if the story is true.

Using LLMs doesn't mean some easy mode where humans don't do difficult work

so far it's been programming and math, the two fields most people would naturally assume computers would be good at :P

not saying it's not possible ai will continue to expand to more fields, but i don't think it's a coincidence that both math and programming are easily automatically verifiable. in fact for programming their effectiveness drops substantially once you remove any sort of testing or verification loop, and i imagine it doubly-so with math. try getting an agent to build something without any tests or verification, you'll see it go back to near-2024 levels. half of the work i do with agents is setting up a harness that can take and analyse screenshots, send direct input to the application, and compare results. this is something that should come standard with every harness.

if we can build automatic verification for more fields is a different and perhaps more relevant question.

> so far it's been programming and math, the two fields most people would naturally assume computers would be good at :P

If that is true, then only because "most people" have no understanding of what differentiates programming and math from "computation". Computers (as the term suggests) have always been extremely good at the latter, and useless at the former. It's only 2–3 years ago that computers, by way of LLMs, have started being good at the former.

Agree, and I'd add a caveat: the loop is only as good as the verifier, and agents are very good at finding the gap between "passes the check" and "is correct." Some things I've seen when the verifier is weak:

If the same agent writes the tests and the code, it will happily weaken an assertion, add a skip, or special-case the fixture to get to green. I keep test files read-only for the implementing agent, or have a separate pass write tests from the spec before any code exists.

Mutation testing is a cheap check on whether your tests constrain anything at all. If you can flip a comparison and the suite still passes, the agent's "all tests pass" means very little.

Property-based tests and differential checks (compare against a reference implementation or the old version) hold up much better than example-based tests the model can pattern-match against.

On the screenshot harness: pixel diffs get noisy fast. What worked better for me was dumping the accessibility tree or DOM state alongside the screenshot and asserting on that, with the image as a secondary signal for the model. Fixed viewport, fonts and animation disabled cut flakiness a lot.

Also, verification only covers what you thought to check. Veracode's 2026 update on LLM-generated code had syntax pass rates above 95%, but only about 55% of tasks produced secure code. "It compiles and the tests pass" was true, and the security property just wasn't in the loop. Math has an edge here: a Lean proof that checks really is a proof of the stated theorem, so the remaining risk sits in formalizing the statement. Programming has no equivalent unless you specify everything.

On your last question, I suspect many fields will get partial verifiers (simulators, unit checks on physical quantities, citation checks, reproducible pipelines) rather than a full oracle. Then the human work moves to deciding what the verifier misses, which is roughly what code review is turning into already.

I have to confess to being wrong in exactly that way in the past. On the other hand it is probably always good to behave as if the next step up in the hierarchy is still available. There's no point letting these things spoil your day twice.

Now I'm forced to think about it much more carefully. I don't believe that current LLMs can find every possible proof on their own from now on. I also agree that the PR returns have fully diminished. Given how hard it's been in recent years to fund e.g. a poorly paid postdoc, I doubt there's a ton of money lying around for getting the best models to work on mathematics problems. I think there's a chance we're heading into a really good era. I was able to make so much progress back when Google search was a useful research tool...

> obviously superior abilities

if you think that's the takeaway, you're not listening to any of the people affected by AI

I am myself one of them. This is just me doomposting as lightheartedly as I can muster.

I guess you didn't read the article? The end of it quotes Tao about how much more there is to discover in mathematics, and he expects humans using AI to get there.

My point was not that AI from now on should be done away with in mathematics research. I was rather commenting on the likelihood of AI companies continuing the "strip mining", as Tao put it, of his field.

Reply on news.ycombinator.com