In a few years, an AI will produce a series of papers or a monograph entitled something like
Co-prismatic Eisenstein Theory I: Deconstruction of Semi-euleroid Motifs
A typical mathematician will be able to follow the text as far as page five. A few people on the planet will reach page ten. On page thirty, out of two hundred, there will be an unnumbered proposition: "From the last three lemmata, the Riemann hypothesis obviously follows."
The work will quickly be verified by other AIs. It will be formalised and verified in Lean or, rather, in a much more powerful proof assistant that an AI will have created some time before and in which the entirety of human-produced mathematics will have been verified.
You will be able to ask an AI to explain the results at a level you can understand, and it will gladly do so, but reading the diluted explanations necessary to reach page twenty will take decades. Now you know what angels feel before the face of god: a presence of an infinitely superior, seemingly benevolent, if aloof, intelligence.
It will not stop there, of course. More papers will be produced, using the final conclusions of the original work --inaccessible to you-- as starting points. And then more.
And given that, after a century of effort involving some of the world’s leading mathematicians, an AI has produced a comically simple counter-example to the Jacobian conjecture---one that an eighth-grader can verify, it is no longer seem entirely impossible that it might, after all, find an elementary proof of Fermat’s Last Theorem that fits within the default LaTeX margins.