@lrhodes@merveilles.town at 11/21/2026, 4:46:50 AM

Some fun facts about the Millinneum Prize Problems:

1. A proof must be published in a qualifying journal and independently verified in order to qualify. No AI-generated solution yet meets those criteria.

2. The only solution so far accepted was produced without AI in 2003.

3. Each prize is $1M, a substantial sum for most mathematicians, but likely less than what AI companies spend on the compute required to produce their solutions.

4. GenAI proofs tend to be extremely verbose and inefficient (the Reimann hypothesis proof is reportedly ~250,000 pages long), and will take years, if not decades, to verify or disprove. Mathematicians could plausibly waste entire careers wading through inconclusive slop.

5. The problems aren't particularly urgent, eg. people don't regularly die of Hodge conjecture, but LLMs are better at generating plausible-looking proofs than they are at, say, solving climate change, and AI companies need splashy stories to sustain the narrative that they can solve Big Problems, so they've converged on the Prize Problems as a more attainable way to attract favorable coverage.

If I had to make a prediction, I would say that the full solutions LLMs propose to the Prize Problems will never be independently verified. The sheer amount of work involved is too vast, and most mathematicians will prefer to spend their time working out their own proofs. Human-produced proofs will likely beat most AI proofs to the prizes simply because they'll be easier for their colleagues to verify.

In effect, these corporate sponsored AI-generated proofs are the mathematical equivalent of flooding FOSS projects with vibecoded pull requests. Taking them seriously means committing huge amounts of unpaid time and labor to checking their work, with a disproportionately high risk that they'll turn out to be useless.

Since the last few AI proofs have only narrowly beaten human-written proofs to print, here's a not-at-all hypothetical question: Who wins the prize if the human proof is verified before the AI proof? Does the human win, or is the prize held in abeyance to make sure that the AI didn't solve the problem first? And if the AI proof is never verified, is the prize never awarded?

In some ways, these AI proofs seem tailor-made to break the Prize Problem format.

This, incidentally, sounds like someone in desperate need of a detox. It's the sort of self-disclosure that later gets quoted in biographies as an early warning sign that tragically went unheeded.

Text clipped from Alpogë's announcement of a proof for the Reimann hypothesis: "Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line."

FWIW, I don't really buy the idea that there was any academic or mathematical reason to rush out the Lean version of the proof. "There is an obvious cynical view, to do with the timing," Alpöge writes in the particularly oblique paragraph where he talks about coordinating release strategy directly with Dario Amodei. Yeah, the cynical view is that Anthropic's stock was already down almost 25% on Friday, so a little juice couldn't hurt them going into Monday.

If we want to be even more cynical about it: Alpöge is on staff at Anthropic. Presumably, that makes him one of the employees who became a millionaire when Anthropic's IPO launched. Whether or not his proof wins a prize matters less to his net worth than keeping the company's stock price as high as possible.

Maybe those consideration never entered into his decision. But publishing the way he did certainly casts that shadow over his work.

Counter-intuitive maybe, given all I've written so far, but I wouldn't rule out the possibility that Alpöge's Lean proof is true. The key word being "Alpöge's." It seems plausible to me that a trained mathematician looking for relevance could steer an LLM to produce convincingly relevant output, find useable material in 250,000+ pages of output, and assemble it into a workable proof. A person sufficiently motivated to believe in the capabilities of the LLM may not even realize how much of the sense-making work he's actually done.

Here's a relatively accessible explainer on why Lean proofs on their own don't necessarily solve a problem: youtu.be/yUoezeDBhD0

And here's the paper it references: arxiv.org/abs/2610.08144

Basically: Because LLMs are unreliable translators, there's no guarantee that an autoformalized proof actually corresponds in meaning to the natural language proof it's derived from, and reliable autoformalization may be computationally impossible anyway.

It would follow that the likelihood of mistranslation grows in proportion to the length and complexity of the natural language proof — though mostly as a consequence of the number of opportunities for ambiguity that introduces to a text.