On October 6, OpenAI published findings on 377 open problems in mathematics, including what it described as resolutions of the Kakeya conjecture and progress on the Riemann, Hodge, and Birch and Swinnerton-Dyer conjectures. If you saw the headlines on your feed that morning, you might have concluded that AI can now basically do your math degree for you.

It can’t. And understanding exactly why is more useful to you than the results themselves.

What the result actually is

These are research-level conjectures — problems that have been open for decades and that working mathematicians build entire careers around. Solving them is a genuinely significant claim, and it’s the kind of multi-step, long-horizon reasoning that models from even two years ago couldn’t attempt. OpenAI’s own framing is that its newest models are built for exactly this sort of extended work, not for one-shot answers.

Two honest caveats before the headlines settle in. First, “OpenAI says” is doing a lot of work in that sentence — the claims are the company’s, and the mathematical community’s verification is the part that actually counts, and that process runs on a much slower clock than a press release. Second, and this is the part that matters for you: a model capable of attacking a decades-old conjecture is not the same thing as a model that reliably handles the routine algebra on your problem set.

Why the gap between “solves conjectures” and “helps with homework” is real

The two tasks stress completely different failure modes. A research conjecture is one hard problem you can afford to spend enormous compute on, where being interesting is the goal. Your homework is twenty routine problems where being reliably correct is the goal — and routine correctness is where models have historically been shakiest, because they pattern-match their way through familiar-looking steps and can produce a confident, clean, wrong answer.

We’ve covered this exact failure before on AIStackHub: AI math tutors lose accuracy the moment a problem is reworded, because the model recognizes the shape of a problem type rather than truly following the logic. Nothing in the October 6 announcement changes that. A model that makes a genuine leap on the hardest problem in the field can still stumble on a sign change in a routine derivative, and assuming otherwise is how people lose points.

What do you actually need from AI in math? Understand a concept AI is genuinely useful here Get the answer to tonight's problem set Ask for the why, not the number reword the problem yourself after Verify every step routine algebra is still a weak spot

How to actually use a model this capable

The useful mental shift: treat the model as a tutor who is brilliant but has no idea whether they’re right this time, not as an answer machine. That means:

The frontier genuinely moved on October 6 — that’s not hype. But the thing that will move your grade is still whether you can do the routine work reliably, and the fastest way to get there is using AI to understand, not to skip.

VERDICT

The 377-problems result is a real milestone in long-horizon reasoning, and it's worth knowing about — but it changes essentially nothing about how you should tackle a due-Thursday problem set. Keep using AI as a reasoning tutor: ask for the approach, reword the problem to test your understanding, and check every routine step yourself, because routine arithmetic is still where confident models quietly get it wrong. Understanding the difference between "impressive" and "reliable" is the actual skill here.

Sources