Hi all, I would say there is a lot that is cognitively distinguishing LLM reasoning from human mathematical reasoning. Namely: caring about the question. AI will silently replace your question with another one, offer a miraculous solution that later turns out to be "natural except on XYZ", have no regard for elegance, etc. It follows evals, and has none of its own aesthetics. I love using AI for math—don't get me wrong, it helps me tremendously—but it is very cognitively different. Indeed, I claim that Turing machines don't care <https://topos.institute/blog/2026-09-22-who-cares-about-a-computation/>, and that caring is cognitively important, because it leads to different solutions than non-caring work does. As Robert Pirsig said, *quality* is the external face of what internally is experienced as *care*. This explains why even their proofs have the quality of "slop". Best, David On Fri, Nov 13, 2026 at 1:53 PM Corey Lionis via Categories < categories-list@categories.org.au> wrote:
Dear All,
As a young mathematician I undoubtedly lack experience influencing the trajectory of this conversation, but I would like to distinguish a related line of questioning: if there is nothing cognitively distinguishing LLM reasoning from human mathematical reasoning, what reasons might we have to prefer mathematics produced from human reasoning?
A strong point-of-view, to me, is the question of taste. If we view creative mathematics as an act of complex optimisation under subjective constraints, then it is clear that LLMs have some advantages and some disadvantages: the clearest advantage is the access to greater compute and a greater knowledge space, which creates the *dis*advantage that LLMs have no need for the highly selective judgements people are good at. In a world where artificial intelligence shares more of our reasoning constraints, it is still likely that the aesthetics of LLM-produced mathematics pull in directions which might not be significant to us in the same way.
If we are producing work for the edification of other humans, then even a perfect simulation of human reasoning from a simulated human does not accomplish all that we ask mathematics to be. For if mathematics is about human-to-human knowledge and skill transferal, then mathematics divorced, in any way, from the human experience no longer performs the developmental and communal role.
In other words, I raise that the question we should be asking is not 'what can AI do', but 'who is mathematics for?'
Corey
P.s. I encourage all to remember that after asking whether we do/do not adopt a tool, there are still many things within our control: Who provides the tool? What functions do we use the tool for? What restrictions and practices produce the most meaningful alignment? To be active stakeholders in the decision-making process, we need to think about these problems quickly, carefully, and independently (preserve our individual stake as well as those we represent).
On Thu, 12 Nov 2026 at 19:01, Henning Basold via Categories < categories-list@categories.org.au> wrote:
Dear Michael,
I concur with your thoughts and it goes back to the fundamental question: what is creativity, what is the process that leads to new ideas and is it something that machines can do? The current conversation around AI is really about LLMs and their interaction as "agents". As such, these are good at absorbing vast amounts of human-produced (and now fed-back machine-produced) writing. However, this comes with the caveat that "new ideas" are applications of patterns that the model has been trained on. Thus, is creativity then mere pattern-recognition or is there more to it? I think that many of us want to believe that there is, but this will not prevent any profit-oriented company from saying otherwise (thank you Ando for rectifying the use of Marx here!).
This leads to a, may I say, spiritual question that is lurking here: is maths a game that can be won by machines that solve problems with no end, and may even ask questions with no end, at a pace that no human can keep up with so that maths should no longer be a human endeavour? There are efforts of many mathematicians (and those that essentially do maths in other fields) to address the issues arising from LLMs being capable of solving maths problems (https://leidendeclaration.ai/, https://mathandai.org/). However, just as (financial) capitalism is a system set on self-preservation, no matter the costs, A"I" will be reinforced by companies pushing for it and researchers forced to by competition for funding and position. I see difficult times for our fields and the people in it ahead, let alone the general societal and environmental repercussions and damage this has and will only increasingly have.
My apologies for this sceptical reply, but hope that we are not blindly endorsing a technological solution to the academic issues that we have and keep technology what it is: a tool to solve a particular problem.
Best,
Henning
PS: I am reminded of mathematics as a mental process that was used by the intuitionist school in order to develop things like Kripke semantics and the philosophy in general. I wonder what the people from that time would have to say.
PPS: For those not on the types mailing list, our esteemed colleague James McKinna made the good point that this discussion cannot be separated from the political questions and suggested some nice reading material:
* Simon Schaffer's speech <https://erasmusprijs.org/en/laureates/simon-schaffer/acceptance-speech/>
* and U. Chicago Press "Working Knowledge: A Simon Schaffer Reader"
* Randy Pollack's "How to believe a machine-checked proof"
On 10/11/2026 21:21, Michael Barr, Prof. via Categories wrote:
I've given some thought to this and I just don't see what AI is going to train on.
For all my career, I have counted on face to face meeting with people. I wanted to talk to them and the conversations were never recorded. I remember running accidentally into Jon Beck in the 42^nd Street Times Square subway station during the winter break 1964-65 and he started telling my about acyclic models. He had just visited Harry Appelgate who had written his PhD on the categorification of that concept. Jon and I eventually modified it further and we were off. We had spent the entire fall semester trying to relate Eilenberg-Mac Lane group cohomology to the free group triple and this solved it instantly.
For about 15 years, John Kennison would spend 6 weeks in Montreal every spring and he and Bob Raphael and I would meet for 4 or 5 hours every day and talk, often going to the board to explain, but mainly just talking. Practically every year we wrote a joint paper as a result. This ended with the pandemic and John canceled in 2020. We have not met since. We tried zooming together but that has not worked. We enjoy the conversations but no mathematics results.
As far as I am concerned, nothing can replace these face to face sessions and I cannot see training AI on that.
Michael ------------------------------------------------------------------------ *From:* Dusko Pavlovic via Categories < categories-list@categories.org.au> *Sent:* Monday, November 9, 2026 3:58 AM *To:* Michael Barr, Prof. via Categories <categories-list@categories.org.au> *Cc:* Dusko Pavlovic <duskgoo@gmail.com> *Subject:* [categories] Re: Comments on AI and math thank you for interesting examples, mike :)
while they will probably be echoing in our heads in the coming days, i think it is safe to say at once that AI bots can in principle be post-trained to be as good at asking questions as they are at answering them --- PROVIDED that samples of the process of question generation can be found, something like your and kleisli's conversations. say if they were recorded as emails or videos.
a PRE-trained language model does just one thing: it generates streams of tokens extending given contexts. given a question, a merely PRE-trained model does not try to answer it, but guesses a continuation of the question itself. to recognize a question as a question, and to try to guess an answer, it must be POST-trained on samples of question-answer protocols.
the models solving all those math problems have been POST-trained on humongous corpuses of processes of solving math problems. answering questions. that is why they are good at answering questions. to also be good and questioning answers, they need to be POST-trained on samples of that.
if we want them to help us ask questions, we need to be able to teach them on LOTS of examples of question generation --- like those that you just provided. or maybe not just examples of questioning, but of the *attitude* that generated kleisli's and freyd's questions and ideas: how they "shake up" each construction to find paths beyond it. if we could find enough such examples, we could not only learn from them ourselves, as we do from the examples you just posted, but we could also process them on scale and post-train some helpers...
i am sure there will be lots of people who will now say that we should keep our question-generation processes to ourselves, maybe also all other thinking processes. and maybe they are right. or maybe we should consider models of our thinking processes as a part of ourselves, like we consider our writings as a part of our identities...
2c, -- dusko
On Sun, Nov 8, 2026 at 3:32 PM Michael Barr, Prof. via Categories <categories-list@categories.org.au> wrote:
While this is not about categories specifically I have some comments
I would like to make public and this would seem to be my outlet. So please bear with me.
The main thing I want to say about math and AI is that AI appears to
be excellent at answering questions, but I see no evidence that it is capable of asking them. And the latter, I feel, is the main thing a research mathematician contributes to the advancement of mathematics. I am going to illustrate this point by two examples from my own career.
I spent a month in Fribourg in 1997 and during that month, Heinrich
Kleisli and I wrote a paper called Topological Balls. The way this paper came about was that every day Heinrich and I would meet in his office and discuss the topic and he would ask a question. I would go back to the apartment I was occupying and answer his question. The next day I would explain my answer and we would have a further discussion and he would raise a new question. And I would answer that in the even and so on and so on till at the end of my stay we had a paper. In my opinion, Heinrich contributed more to it than I did. No questions, no paper.
During the year 1969-70, my second at McGill, I was working on
regular and exact categories, quite consciously trying to adapt the ideas of abelian categories to non-additive ones. After finding what still seem like the correct definitions, I wanted to adapt theorem that a small abelian category has a faithful functor to Ab. By that time Saul Lubkin's proof and a quite different one from Peter Freyd were known. There was also Barry Mitchell's full embedding into a category of modules over a ring. I succeeded in adapting Lubkin's proof to show that a small regular category had a faithful functor to Set, provided that the terminal object had no proper subobject, except possibly an empty object, meaning one that was not the codomain of any map other its own identity (or any other isomorphism to be sure). In the meantime Kleisli had invited me to spend the year 1970-71 in Fribourg and had gotten a grant from the Swiss Research Council to do so.
So, I took an unpaid leave and headed to Fribourg, On the way there,
I stopped in Philly to visit my family and looked in on Peter Freyd at Penn. I explained what I had done and Peter asked me whether it was possible to get a full embedding theorem out of the method by replacing Set by M-Set for a suitable monoid M (which turned out to be the monoid of endomorphisms of the original embedding). To my surprise, it was! And not particularly hard to prove. That means that Lubkin could have had a full embedding theorem too, had he realized. Anyway, on to Fribourg where I showed Heinrich what I had done and he asked whether it would be possible to drop the condition on subobjects of 1 and replace M-Set by a category of the form Set^C for some small category C. It was! And not too hard either. The point was that, unlike me, he thought of a monoid as a small category.
So two questions, one from Peter and one from Heinrich, had elevated
what was probably a minor result to one of the major accomplishments of my career. And this is why I think that questions are more important than answers.
Michael
_______________________________________________ Categories mailing list -- categories-list@categories.org.au To unsubscribe send an email to
categories-list-leave@categories.org.au _______________________________________________ Categories mailing list -- categories-list@categories.org.au To unsubscribe send an email to categories-list-leave@categories.org.au
_______________________________________________ Categories mailing list -- categories-list@categories.org.au To unsubscribe send an email to categories-list-leave@categories.org.au
Categories mailing list -- categories-list@categories.org.au To unsubscribe send an email to categories-list-leave@categories.org.au
_______________________________________________ Categories mailing list -- categories-list@categories.org.au To unsubscribe send an email to categories-list-leave@categories.org.au