I definitely agree with the final sentiment here - we need, more than ever, more mathematicians and human experts! And the good news is, it's easier than ever before to become such an expert. You can use AI to explain concepts, talk through proofs, etc. But... 1/
Tony Feng
@tonylfeng
Sep 15
I like this essay. But I'd like emphasize that the premise is hypothetical: even after the Navier-Stokes solution, I do not believe that current AI models are robustly superhuman yet. It seems plausible that AI could reach the capability to solve any precisely formulated problem faster than any human could. Asking questions is also an important aspect of the profession, and while AI models are getting better at asking interesting questions, I'm not sure what it would mean for them to be "robustly" superhuman at it -- surely they will not ask every question that a human could ask? Still, I think it would be prudent for academia to start moving towards the future this essay depicts. A first concrete step could be to make Ph.D. defenses, which are typically rubber stamped (in my experience), more rigorous. I would also advocate for the public value of mathematicians in such a future. There are many wrong/confused/misinformed mathematical claims on this website, which could have been rectified by spending a few minutes with ChatGPT, but were not. You can't filter every sentence you read through an LLM, and you can't filter every sentence you write through an LLM. The math discourse on this site is proof that even in the presence of superintelligent AI, you need dedicated experts to serve as stewards of the truth.

Sep 15, 2026 · 9:35 PM UTC

4
8
102
16,994
You still need to do the work. Think things through for yourself, make sure you actually understand. Everyone has had access to the internet and all the world's knowledge free for decades - yet this did not make everyone an expert because more is required. 2/
1
49
1,985
AI makes the path to understanding these concepts easier than ever before - but it is still not easy. If these maths headlines have got you curious then fantastic- go, learn, explore, think! 3/
1
33
1,519
But it is such a waste to have that interest, and the endless learning opportunities afforded by AI (and many other resources available), but instead just copy-paste a prompt and stare at a screen, hoping for it to return an answer you can parrot without understanding. 4/
1
25
1,384
Use your brain. It's brilliant. 5/5
47
1,205
Sort replies: Relevant Recent Liked
Replying to @thomasfbloom
„Everyone has had access to the internet and all the world's knowledge free for decades” I respectfully disagree. I’m a software engineer, working in industry for 20 years, so I’m definately not a Sunday internet user. Yet I can’t even compare the experience of researching something now with a help of LLM to how was it before. Ease of access is as important as source availability. I have a job, family and very little time for a hobby. So if getting the source, such as specific academic paper, takes days or weeks of research, I will not get there. Probably ever. If however it takes one prompt which results in multiple sources I pointed to, cited, summarized and even explained how and in which way they are relevat - this changes everything. Does this on its own make me an expert? Of course not. But it opens a door for people willing to walk through. LLM changes access to knowledge globally and irreversibly. It is it’s biggest value.
1
347
I'm not sure what you disagree with - I agree with what you say! Of course LLMs can make access to knowledge, and learning about it, much easier. My point was that much of this knowledge was already available - but yes, available does not mean easily available.
1
5
260
more replies
Replying to @thomasfbloom
Cool how LLMs seem to empower (i) motivated people who want to learn a ton of math and (ii) people who don’t want to learn any math
2
503
Replying to @thomasfbloom
We actually went deeper on the debate surrounding AI-generated proofs here:
AGTP
@AGTPinsights
Sep 13
Mathematicians are now debating whether OpenAI's own team understands the Navier-Stokes proof it published. Here's what you need to know. Mathematician Scott Armstrong posted on September 12 that the OpenAI construction is a huge contribution to PDE analysis, but said it will take analysts longer than a week, maybe less than a month, to understand it well. He argued human understanding will now typically lag behind machine-generated proofs instead of coming first. The post drew pushback and debate across X. Daniel Litt said Armstrong was too uncharitable toward concerns about training, incentives, and lab conduct. Sasha Gusev compared it to biology, where results get replicated before they're explained. Noah Snyder said he's more bothered that OpenAI staff don't seem to care about understanding the proof at all, not just that they don't understand it yet. Iosif Lazaridis doubted understanding will keep pace, and Dmitry Krachun asked whether it still would if labs spent $1B to clear 99.99% of the roughly 100,000 conjectures on arXiv in a week. This follows OpenAI's September 9 announcement that roughly 10,000 agents, using a next-generation model more capable than GPT-6 Astra, produced the 166-page proof, titled 'Finite Time Blowup for Navier-Stokes,' in 88 hours, resolving two of the four statements in the problem. Key numbers: - 10,000 agents used - 88 hours to produce the proof - 166 pages in the paper - 2 of 4 statements resolved The proof includes a formal Lean writeup and has not yet been peer-reviewed.
756