Does anybody have a description of what type of system was involved in Alpoge’s work on the Riemann hypothesis?
Did it use tools? A harness? Code interpreters? Loops? Conditionals? Symbolic theorem provers? Was it just a scaled up version of o1?
A lot of people are writing as if they know the answers to these questions. (Or as if they don’t understand the questions).
Do we even have a flowchart of what system was involved?
Or is everyone just making stuff up?
Please send a link if there is one.
Why do people listen to this guy in 2026?
He is not active, right?
His radiologist prediction failed.
> Geoffrey Hinton: The argument that LLMs are just predicting the next word falls apart when you test their ability to understand something they've never seen before.
We will write about 2026 like we have 1543, 1666, 1905. But what's unthinkable until now in human history is that we could have a year that would be an annus mirabilis for mathematics while something of an annus horribilis for mathematicIANs. Future reconfigured in real time.
Q: “But Gary, why can’t you admit that the Riemann results are impressive?”
A: I actually did, multiple times! I even called them “brilliant”! What more do you need?
And why you are guys so needy, anyway, that you need my constant validation? What is up with that?
If the Claude-Anthropic drop (via proxy Alpoge) really depends on this, how the hell the Lean people did not see this and quickly "certified" the drop. Would they be that unserious?
“We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct.” -- Lean community confirms the *biggest* achievement in AI for math!
However, two points are worth noting:
(1) This proof is 17M lines of code!!! It speaks to the genius of Riemann that he could show how the zeros of the zeta function govern the distribution of primes and formulate his hypothesis 167 years ago, with nothing but pen and paper.
(2) One of the main reasons proving GRH mattered so much was the prospect of a much deeper understanding of mathematics, perhaps even the discovery of new subfields. An AI-generated proof we cannot yet understand doesn’t, by itself, deliver that understanding. It may still take a while to extract the ideas that mathematicians can build on.