An AI Solved Ten Open Problems for Two Grand

A lab's model reportedly solved ten open math problems and proved each answer formally. "Formally" is the whole story. Everything else is commentary.

The report going around this week: a lab's internal model took on ten previously unsolved problems in math and theory, solved them, and published formally verified proofs. The compute cost was reportedly low, something on the order of a couple thousand dollars. Ten open problems. Verified. Cheap.

If that holds up, it is a genuine milestone. Let me say why, and then let me say where the hype is getting ahead of it, because both things are true.

why "verified" is the word

A formally verified proof is not a vibe. It is a proof that a machine has checked, step by step, against the rules of the system. It is either right or it is not, and there is no arguing about it afterward. That is the difference between this and every demo that ever wowed a room.

Most AI output lives in the land of "looks right." An essay, a summary, a plan. You read it, it sounds good, and maybe it is. Maybe it is not. A verified proof does not ask for your trust. It brings receipts that a computer checked. That is a different category of result, and it is why this story deserves more than the usual hype-cycle shrug.

It also matters that the target was math. Mathematics is the friendliest possible terrain for this kind of work: closed rules, checkable answers, no messy human context. If machine reasoning was going to prove itself anywhere first, it was going to be here.

where the hype outruns it

Ten problems is not ten fields. It is ten problems, in areas close enough to the model's training that the leap was crossable. That is still impressive. It is not "mathematics is solved now," and anyone telling you that is selling something.

The low cost cuts both ways too. Yes, it suggests the capability is getting cheaper, which is the trend that matters most. But a headline number like "two grand" invites people to imagine that all hard problems now cost two grand to solve. They do not. These were the right problems, attacked the right way, with the right verification machinery around them. Change any of those and the price changes with it.

And "internal model" is doing quiet work in that sentence. We are reading a report about a system we cannot touch, evaluated by the people who built it. I am not calling it false. I am saying the verification that matters, the proof checking, is public and checkable, while the claims around it deserve the usual discount you apply to self-reported results.

the honest read

Here is what I actually take from this. The direction is real: machine-checked reasoning is getting stronger and cheaper, and that combination will eventually matter far beyond math. Proofs are the beachhead. The mainland is every field where "show your work" can be formalized.

But a beachhead is not the mainland. Celebrate the milestone, discount the victory lap, and watch what happens when this machinery points at problems where the rules are not written down yet. That is the test that actually matters, and it has not happened.