Harald Schilly

← Thoughts

Mathematics one level up

Draft, still being revised

In 2002 my old homepage said: “i was never really interested in mathematics”. It turned out otherwise, with a degree in mathematics, many years with SageMath and work on the LMFDB database.

Mathematics can now be formalised on the computer. In the proof assistant Lean you write definitions, theorems and proofs in a way that lets the computer check every single step. The shared library Mathlib already covers a large part of the foundations. In 2022 the Liquid Tensor Experiment fully verified a difficult, recent result by Dustin Clausen and Peter Scholze in Lean, at Scholze’s own request.

On top of that comes AI that finds proofs by itself. In 2024 a system by Google DeepMind solved four of the six problems of the International Mathematical Olympiad, at the level of a silver medal. Its AlphaProof part wrote its proofs in Lean. That is the key point: you do not have to trust an AI if its proof has been formally checked. My own preprint also comes with a formal proof in Lean 4 and was at the same time an experiment in AI-assisted proving.

Today, on 20 November 2026, a short text appeared according to which it looks as if the Riemann Hypothesis has been proved, according to the text even the generalised one, found together with Claude. Only the proof in Lean is presented. The formalisation is huge, and a written-out version would, by its own account, run to more than 250,000 pages. How anything like this is supposed to be reviewed is unclear. Lean checks every step, but people still have to check that the statement is formalised correctly, and understand the proof. The text itself says that the argument still has to be understood, simplified and written up properly.

This changes the work of mathematics. Until now a large part of it was craft: calculations, estimates, case distinctions, a proof line by line. Machines are increasingly taking over that part. Mathematicians become more like managers and supervisors. They have to understand what it is about, recognise which problems are worth pursuing, see where something can be applied, and ask new questions. AI and formalisation on the computer are the tool one level below.