Proofs are where concepts come from. Recent developments in mathematics have come as both a shock and a source of excitement for people in the field. If you are an expert in the field and are excited, you realise you can now do more proofs with this incredible tool. However, if you're shocked, you realise a lot. A part of the field is gone. Prof. Tao calls it the end of the romantic part of mathematics. The days of hero celebrations are kind of coming to an end. I believe Perelman will be glad about that side, but the AI part, he might have a different opinion on it; I think so.
I'm writing to talk about the problems that I see as a mathematics student. I may present my writing in a mathematical structure of a sort.
Main conjecture: the basic and important stone for mathematics is proofs, i.e finding the truth of a result. Why this conjecture? I am going to talk about the four personalities in the field to shorten the essay and maybe out of interest. Their contributions gave rise to my developments in the field.
There's a saying that mathematicians write proofs, good mathematicians bring out conjectures, and great mathematicians write definitions. The statements may suggest three different working mathematicians, but in fact, it is the development cycle of a mathematician. It always starts with proofs.
Prof. Grothendieck and Prof. Weil. I will mention a few of these cycles happening. Weil counted points on curves and diagonal hypersurfaces over finite fields. Grothendieck worked through Serre's sheaf cohomology and Riemann-Roch. This is the proof stages for these mathematicians. What happened? You realise that point counts behave as if a cohomology theory lay behind them. For Grothendieck, the maps between varieties mattered more than the varieties themselves. These patterns were emerging from working out proofs for other results. This journey gave rise to the Weil Conjectures. The development of Abelian categories, K-theory, schemes, étale cohomology and topoi. The theory that contains these definitions became modern algebraic geometry.
Our next two personalities are Prof. Eilenberg and Prof. Mac Lane. Mac Lane has been working on group extensions, and Eilenberg on Steenrod's homology of solenoids. The two compared notes and found the same group, ext, appeared in both. By working out the link gave them a precise form of the universal coefficient theorem. Once again, the proof stages for the two mathematicians. What did they realise after doing proofs? Their computations kept producing isomorphisms that depended on no choices. An example is a finite abelian group and its double dual. The group is isomorphic to its dual only by a choice of isomorphism, but to its double dual canonically. "Natural" sounded more intuitive. This led them to define what we call "Natural Transformations". From proofs to definitions. Again, we start from proofs. What are the theory we got? We had category theory itself - the goldmine of mathematics. Eilenberg-Steenrod axioms for homology and, with Cartan, homological algebra.
Why these statements and examples? Others perceive proof as just an end goal. I disagree. Coming up with a proof requires clear expositions, intuition-building, and understanding of the concepts. The rigour and standards of mathematics is of itself the ability to make logical sense which is to do proofs. The results of these proofs birth definitions, theory building and conjectures. How? The journey to finding proof has no clear path or direction. I will use this analogy. Once again in the style of mathematics.
Definition: Let \(\mathcal{F}\) be a forest ( a math topic), \(S\) be the starting hypotheses, and \(T\) be the target theorem. Let \(P\) denote a valid proof path from \(S\) to \(T\), and let \(D\)denote the set of all dead ends (failed attempts and counterexamples) surrounding \(P\).
Theorem 1 (Asymmetry of Terrain Knowledge). Suppose Person A constructs \(P\) manually, while Person B queries an AI oracle \(\mathcal{O}: S \mapsto P\). At completion, both posses \(T\), yet their internal maps \(M\) differ:
\(M(A)=P\cup D\quad \text{whereas}\quad M(B)=P\text{\ only.}\)
Theorem 2(The illusion of competence). Because \(\mathcal{O}\) outputs only the polished path \(P\), the failed set \(D\) is erased. Consequently, the justification for why a specific method \(f\) succeeds where the alternative \(g \in D\) fails remains locked in \(\mathcal{O}\).
Corollary(Inability to Generalise or Explain). True intuition and exposition require knowing \(D\) as well as \(P\). Therefore:
1. Recitation is not equal to understanding: Person B can read \(P\) around, creating an illusion of competence, without knowing how \(P\) was discovered.
2. Fragile: if the target shifts by \(\epsilon\) i.e \(T + \epsilon\) and \(\mathcal{O}\) is removed, Person A roam using \(D\), while Person B is lost.
We are seeing these effects play out in the blow-ups of the Navier-Stokes results. We know it but don't understand it. We know it is true, but the \(D\), which gives the reasons for the next steps, is lost. We now consult to know why. Do we learn? I believe we don't.
The excitement: now we can prove more results than ever. The above theorem may have less effect on established mathematics. They have passed through the training. To them, the AI is complementing them because they already have the judgment and intuition that is built from this journey, the Proof Journey. This is the reason why I disagree with Sanderson(3blue1brown), but proof is the starting point. A clear example is Levent Alpöge, famous for the Jacobian conjectures and, more recently, the Riemann Hypothesis. While he's making advances in the field, the path to developing the next Jacobian and Riemann is narrowing.
Now to myself and the young mathematician incoming, we have a different story to tell the established ones who are getting excited that the path is dangerous. The trajectory kills apprenticeship. We need to maintain the traditional route. Celebrate more people if results are achieved without AI. I mean finding a different proof to AI. Build an AI resistance until expertise is reached. The bar to become an expert is now far too high that I can't see it. While Anthropic and OpenAI will continue to churn out proofs, we must be guided to see it as certainty measure that the results is true or false, and that the human development and the dead ends is crucial to the field. If we continue with this end, AI might give recommendations, and we'll be clueless about what to do. It is not the romantic part of mathematics that is dying; it might be the vital part of it that is dying.
