A free preview, generated by MasteryLoopHQ.
Two announcements this year put the question in front of a lot of people. On August 1, 2026, OpenAI shared ten results from an internal version of its Astra model, which was unreleased at the time, that it said resolve or make substantial progress on long-standing open problems. It published a manuscript of about 250 pages and said the token cost of the successful runs was roughly $2,000 at its API rates. On October 8, 2026, Google DeepMind's AlphaProof Nexus appeared in Science, reporting proofs for 9 of 353 open Erdős problems it targeted and 44 of 492 open conjectures from the OEIS, the online encyclopedia of integer sequences. Both are research-level results that a machine produced, which is new. Neither is the same as a machine replacing mathematicians.
Both efforts use Lean, a proof assistant. Lean has a small trusted core that checks every step, and its verdict is binary: the proof compiles or it doesn't. AlphaProof Nexus works directly in Lean, looping between a language model that drafts proofs and Lean that checks them, and its final proofs must contain no unfinished placeholders. OpenAI's model reasoned in ordinary language first, and then formalized each argument in Lean afterward, with the Lean files published under an open license so anyone can run them. That is a strong kind of checking, because Lean's acceptance does not depend on believing the AI's reasoning. It only covers the statement as it was written into Lean.
A person still has to confirm that the statement Lean checked is the problem someone meant to solve. In the AlphaProof Nexus paper, experts validated each formal statement after each solve, and the meaning of density had to be corrected in the statements of two problems. The paper also says its process has a bias toward problems that are easy to express in Lean. The nine Erdős results count sub-questions and variants, so they are not nine famous problems solved in full, and Erdős Problem 12, open since 1970, had two of its questions resolved, not the whole problem. In failed attempts the model leaned on lemmas it claimed were established results that turned out not to exist, and Lean refused to accept those proofs. On the OpenAI side, the picture has moved since August 1. OpenAI released GPT-6 Astra on September 3, and on September 8 it claimed a solution to the Navier-Stokes problem, one of the Millennium Prize Problems, using a newer internal model. Mathematicians are still examining that claim, and one preprint argues that its Lean version does not match the written proof.
On October 16, 2026, the YouTube channel 3Blue1Brown published a video titled "There's never been a better time to go into math." It has passed 2 million views. Its chapters cover AI as a proof machine, a conjecture machine and a theory machine, and the possibility that AI gets better than people at everything. The host, Grant Sanderson, argues that the answer is yes: the value of math as training, as teaching and as joy survives even if machines out-prove people. The Hacker News thread about it had 552 points and 796 comments as of the afternoon of October 20, and many of the arguments were about meaning and money more than mathematics. One widely shared YouTube comment, quoted in the thread, compared mathematicians and programmers to craftspeople whose craft is being automated. Another commenter asked why anyone would spend decades on a hard problem if a future model could make the effort moot. A third said the value is in the journey and the questions, and that AI strips away the prestige and the feeling of time well spent.
Here is a reasoned point, built on the limits above. A proof that Lean accepts still has to be matched to the question someone meant to ask. That takes a person who understands the problem well enough to see whether the formal statement says what it should, and who can notice when a definition is slightly off, as happened with density. Terence Tao, quoted by TechCrunch on October 8, criticized a pattern where problems are solved by AI prompters "who have no interest in the broader field itself once their initial target is 'solved', and do not understand the AI output well enough to answer questions on the result." The article does not claim this settles whether to study math as a career, since neither the announcements nor the thread tested that.
The best-controlled evidence is about learning, not careers. A field experiment published in the Proceedings of the National Academy of Sciences in June 2025 gave nearly 1,000 high school math students access to GPT-4 while practicing. One group had an ordinary ChatGPT-style assistant. Another had a tutor instructed to give hints without directly giving the answer, with teacher-written solutions and common mistakes in its prompt. While they had access, grades on practice problems rose 48% for the first group and 127% for the second. When access was taken away, students who had used the ordinary assistant scored 17% worse than students who had never had it. The tutor with safeguards largely removed that harm, though its students did no better on their own than students who never had AI. The study took place at one high school in Turkey.
A classic result in memory research points the same way. In a 2006 study in Psychological Science, students studied prose passages and then either took recall tests or restudied the material. Five minutes later, restudying produced better recall. After two days and after a week, the students who had been tested remembered substantially more, even though restudying had made them feel more confident. The finding is that pulling an answer out of your own memory builds retention in a way that rereading does not. By the same logic, an assistant that supplies the solution takes away exactly that effort, though that is an inference, not something the study tested.
None of the studies above tests a specific routine, so this is a reasonable rule, not a proven method. First, attempt the problem or write your own explanation before you ask for help. Second, ask the assistant for a hint or a check on your reasoning, not for the answer, which is the design of the tutor that avoided harm in the PNAS experiment, though it did not improve scores there. Third, close the chat and explain the idea back in your own words, without looking, which draws on the testing result. Fourth, come back to it days later and try it again. That last step is an added habit, since neither study tested revisiting. The aim of these steps is to avoid the crutch use that left students worse off on their own. The open question the research does not answer is whether any person should plan a career around math, and that depends on more than what a model can prove.
AI systems have now produced real research-level proofs, and Lean makes those proofs checkable without trusting the machine. But the claims are narrower than the headlines, a human still has to confirm that a proof answers the question that was asked, and the strongest evidence we have about learning says that unrestricted AI help during practice can leave people worse off on their own. Whether to go into math as a career is a different question that this evidence does not settle. What it does support is a way of studying: try first, get checked not told, explain it back, and return to it later.
In your own words: AlphaProof Nexus proved 9 of 353 open Erdős problems it targeted (counting sub-questions and variants), and experts had to confirm each formal statement matched the original. Explain why a proof that Lean accepts can still fail to answer the original question, and what that means for whether a person needs to understand the math.
Reading a good explanation feels like understanding it. Usually it isn't the same thing — and you don't find out which one you've got until someone asks you to explain it back.