2026-09-11 17:40 UTC
Replies (40)
-
@tomkalei@mathe.social 2026-09-11 17:44
@tao@mathstodon.xyz Thank you very much for this effort!!
-
@withinity@mastodon.gamedev.place 2026-09-11 17:58
@tao@mathstodon.xyz I am really grateful to see people with your stature speaking up about this. In other industries similar existential threats exist. I hope this starts a movement toward something beneficial to all of us. For many, the AI movement is bleak.
-
@karthikganapathy@mathstodon.xyz 2026-09-11 18:00
@tao@mathstodon.xyz I am glad to see this (in part because the statement put out by AMS seemed to celebrate and center OpenAI's achievements rather than warning of the negative consequences).
-
@sorenhave@mastodon.nu 2026-09-11 18:07
@tao@mathstodon.xyz Thank you. Amazed and also very moved to see the name of Pierre Deligne there - my PhD evolved around various properties of Deligne-Lusztig varieties.
-
@hvc@mathstodon.xyz 2026-09-11 19:01
@tao@mathstodon.xyz , sounds great. But... what is the proposal? Meaning, how do you think this can be solved? Because if you basically tell AI companies "I don't like this" without providing an alternative, I'm very much afraid it won't do much good... You need to come down to the real world and propose something tangible, and I say this with all my love and support. Best of luck; in other disciplines, we haven't been listened to, or, like in cybersecurity, it has simply become an asset accessible only to corporate elites and governments...
-
@yaarur@mathstodon.xyz 2026-09-11 19:01
@tao@mathstodon.xyz I love it that you write “the scientific and creative proffessions”. This is a time that calles for solidarity. "AI" has alredy undermined the creative communities with generative images and text, and is threatening the computing community with generative code, but we have been mostly silent, if not participating in this undermining as these were not our fields. Now that it reaches us maybe it will shake us enough to look around and see the pattern.
-
@timaeus@nrw.social 2026-09-11 19:04
@tao@mathstodon.xyz thanks for this and indeed this points towards something that is much large then the field of mathematics. If the work process how a goal is achieved matters then excluding humans from this process and substituting them by AI this will massively impact the very meaning of that work.
-
@timaeus@nrw.social 2026-09-11 19:05
@tao@mathstodon.xyz may be we are on a track where humans lose control over the development of knowledge and culture?
-
@mmoustafa8108@mathstodon.xyz 2026-09-11 19:09
@tao@mathstodon.xyz brilliant as always :) thanks for your efforts!
-
@ShadSterling@mastodon.social 2026-09-11 19:34
@tao@mathstodon.xyz if the recent example is typical, for an LLM to “solve” a math problem it needs to be trained on drafts of an incomplete proof by human mathematicians that turns out to be on the right track, used by a team of about 10 times as many people as wrote those drafts, and spend about a decades worth of funding for related math problems on generating attempted proofs. (See https://neuromatch.social/@jonny/117237800074111038) The trivial solution is to tax the rich and fund mathematicians.
-
@indutny@mean.engineer 2026-09-11 19:45
@tao@mathstodon.xyz I appreciate the post, but can’t get rid of the feeling of watching the character of DiCaprio in “Don’t Look Up” trying to be diplomatic when one side would be literally destroying another as long as it can profit from it.
-
@fuglede@mastodon.social 2026-09-11 19:52
@tao@mathstodon.xyz The comments on HackerNews are never particularly great when the topic is maths, but they do suggest that in addition to saying what large AI companies shouldn't do, it might be useful to spell out explicitly what they could or should do to help out instead. Now there's already a Fields medalist hired by one of them (unfortunately lacking from the list of signatories) so of course they already know, but spelling it out also makes it clear for the rest of the world. https://news.ycombinator.com/item?id=49662371
-
@herbst@social.epfl.ch 2026-09-11 20:35
@tao@mathstodon.xyz Thank you and all initiators for starting this, I can only agree !
-
@joost_rekveld@assemblag.es 2026-09-11 20:50
@tao@mathstodon.xyz I'm in awe, this is so great.
-
@thuegerich@social.coop 2026-09-11 21:00
@tao@mathstodon.xyz Though we're not used to thinking this way, it is completely coherent to want to limit technological progress, to sacrifice some of its potential benefits for the sake of preserving meaningful human work. To reassure myself of this (and help fellow economists think more clearly about it), I developed a simple economics model showing how: https://www.lesswrong.com/posts/3dFYQu4ifG5ZsjjZm/what-if-people-value-their-work-mattering
-
@KeaganCallis@mathstodon.xyz 2026-09-11 22:01
@tao@mathstodon.xyz Thank you and all involved for putting this together! I have many thoughts on AI in math (as I'm sure we all do, especially given recent events). I think a decent analogy for the effects of indiscriminant use of AI to answer open questions unchecked (by humans) and without our proper understanding, is the use of AI to generate large portions of code in some codebase. Anyone who's used LLM for coding knows how quickly AI generated code can get beyond our understanding, due to the volume and all the definitions and sometimes obtuse ways of doing things. It seems like this is one of the issues which need to be addressed in the use of AI in math — yet-to-be checked (by humans) and understood (by humans) proofs are similar to unverified mass LLM generated code. And the difference in math vs. coding is the essential goal of math is in exactly that understanding, not in if the code compiles and program works (of course, not to say understanding code isn't important too in any professional program). But on one hand, is it right to fully cease AI's pursuit of solutions to problems? I think even if it proves something which is at the moment poorly understood, the output it gives _does_ give a route along which one can follow to themselves understand the new methods and discoveries the AI made to solve the problem, right? What if, we decide to create some new category of results: "Proven by AI" vs. "Verified by humans" ? And perhaps this second category could also ensure the methods used are properly attributed, of course (though this part may not be all that straightforward). I am curious what your thought are about that suggestion.
-
@agogolev@mathstodon.xyz 2026-09-11 22:59
@tao@mathstodon.xyz Even if AI people leave us alone we are still facing a flood of AI generated papers which were verified by humans to varying degrees. Some kind of organization is needed here. Probably, if the "author" of such paper claims to understand and contributed to writing then he or she becomes a "starting/initial author"of the result which could be posted on arXiv-like website. Maybe some more requirements can be imposed on "initial authors" such as existence of recorded talk on said result. Then if such AI-generated paper is posted with the set of initial authors, these authors could be required assume some responsibility such as being available for answering questions about the result for at least 2-3 years. It should be allowed for the list of "authors" to grow, at least for a some period of time. Say if someone reads the paper and understands it fully then that person can add their name as an author too (without being related to the original authors in any way). Such new authors could be allowed to upload modified\simplified\improved versions of the original manuscript and also should take up some duties on dissemination and being "responsible" for said result. In this way people could built "portfolios" of results which they claim to understand and for which they assume responsibility. I think this would much better than having the competition of pushing the buttons and claiming to be the "author" through a quick AI-solp digestion process.
-
@battaglia01@mathstodon.xyz 2026-09-11 22:59
@tao@mathstodon.xyz How much of this do you think is an existential threat to mathematics as we know it, and how much is a relatively straightforward fine-tuning problem that could be addressed with another round of RLHF? This situation reminds me of complaints people had with early coding models: they produced sloppy code, they were "lazy," they had poor comments, and in general they tended to produce a vibe-coded ball of wax that "works for no reason." This seems almost exactly parallel to that. With software development, some the answer was simply to train better models and teach people how to use them more effectively. Do we expect the same with math? Thanks, and great essay.
-
@aceroe@mathstodon.xyz 2026-09-11 23:17
@tao@mathstodon.xyz The choices you make may well bring corresponding consequences. Good luck, Tao.
-
@thinkb4coding@mastodon.social 2026-09-11 23:56
@tao@mathstodon.xyz if I understand correctly the situation, IA companies just brute force problems with solvers as rocq injecting far more money than research ever had to produce slop proofs (valid proof that are still meaningless).
-
@divyaranjan@mathstodon.xyz 2026-09-12 00:17
@tao@mathstodon.xyz Finally. Thank you.
-
@carlosdelamora@mastodon.social 2026-09-12 00:28
@tao@mathstodon.xyz I don’t intend to say I know better.I feel that saying math has specific purpose. That AI does not aligning and would harm those purposes is not true. I would argue that finding answers to questions is the main purpose of mathematics. And I agree that understanding is of great value, but AI can help with that as well. Explaining concepts, writing books and even teaching. These things are not something we are loosing only because we aren’t the ones coming with the the answers.
-
@D1C1PLER@mathstodon.xyz 2026-09-12 01:03
@tao@mathstodon.xyz Without this effort by you there would be no hope for an outsider like me. I really see you as the last great bastion to rally against this math takeover by companies, and rich men. Everyone is with you. I have enjoyed 3 of your AI challenges now and that is the way to use it properly, by being guided and monitored, for understanding as the profit. Thanks
-
@limsupx@mathstodon.xyz 2026-09-12 01:30
@tao@mathstodon.xyz as a young developing researcher and prospective PhD candidate, the recent developments had somewhat started to worry me but this is a really encouraging initiative. I think it’s important that mathematicians have a say in how AI develops and how it is used in mathematics. Happy to see so many leading mathematicians coming together around this.
-
@WongK@mathstodon.xyz 2026-09-12 02:38
@tao@mathstodon.xyz
-
@tasket@infosec.exchange 2026-09-12 03:08
@tao@mathstodon.xyz The Economist requires your email address to read this, which means they are taking notes on who is interested in this. Very Economic of them.
-
@bakasuraptor@mathstodon.xyz 2026-09-12 04:14
@tao@mathstodon.xyz Is this eventually boiling down to take permission and only then I'll guide or collaborate with you? This was already insinuated by many other PhDs as a warning to future scientists, when dissing mathathon. Because there's no reason that future students cannot absorb and distill knowledge in a sufficient enough manner that they cannot work out the methodologies and glean the insights once the core of a problem is tackled with AI support. How could a paradigm shift to solving problems be detrimental while citing status quo maintenance as one of the major reasons to deter innovation. Seems bizarre, though understandale if the aim was to put something out to sound credible.
-
@h@mathstodon.xyz 2026-09-12 04:21
@tao@mathstodon.xyz maths is a language of mind, and math concepts are exchanged between people's mind via physical world, that can include tools built by humans. the internal languague between AIs is not proven yet to be sophisticated as our math (us the humans). but the question is, we built this tools LLM to communicate our languague, including the must sophisticated one : the math. this llm talk to us, pretending sometime to be like mathematicians, do they explain to us their maths and translate it to ours? or just feed us back what other people told it and went to farther conclusions and communicated to us what they can? this are interesting questions
-
@resuna@ohai.social 2026-09-12 06:16
@tao@mathstodon.xyz Like the "Truth Mines" in Diaspora by @gregeganSF@mathstodon.xyz SF writer and mathematician. "When ve'd first met Radiya in the Mines, Yatima had asked ver why some non-sentient program couldn't just take each formal system used by the miners and crank out all its theorems automatically sparing citizens the effort." Because the work was the point.
-
@tofe@mastodon.uno 2026-09-12 10:40
@tao@mathstodon.xyz Thank you all! Anyway I think it would be interesting to perform the following experiment to rule out if training on recent unpublished work is being performed: Start pointing out to fake unpublished local work on millennium problems that contains poisoned stuff and fake evidence. If they greedy train on this this could be a huge punch to the training of the new model. If they don't nothing happens.
-
@tuzhenzhao@mathstodon.xyz 2026-09-12 11:16
@tao@mathstodon.xyz I support it, and I’ve added my name. If you choose to devote yourself to a field, one of the most important questions is ultimately: what do we want this field to serve? For mathematics, I don’t think the answer is especially complicated. One of the central purposes of the mathematical enterprise is human understanding—and, through that understanding, our ability to connect with one another, to teach and cultivate the next generation, and to keep exploring. If you work in or around this field and would like to support the statement, all you need is an institutional email address to add your name.
-
@Diffgeometer1@mathstodon.xyz 2026-09-12 11:43
@tao@mathstodon.xyz a world where AI is used all the time will lead to cognitive decline. To put it simply, the human race will get dumber over time. The next generation of mathematicians who rely heavily on AI will be less capable than previous generations. Already on arXiv, papers are appearing with the words “AI generated and human verified” written in the comments. AI needs to be reined in https://www.bbc.com/future/article/20260417-ai-chatbots-could-be-making-you-stupider
-
@Miketodon@mathstodon.xyz 2026-09-12 13:37
@tao@mathstodon.xyz This letter diagnoses and then stops. "These issues must be addressed urgently" is not an ask. There were obvious concrete demands available: embargo until a full writeup exists, mandatory citation of prior art, no benchmark announcement before independent verification, and none of them appear. A declaration signed by twenty-five Fields Medallists is a scarce instrument, and spending it on a mood rather than a demand is a waste. Google DeepMind proposed conventions back in March. You and other signatories ignored it https://arxiv.org/abs/2602.10177
-
@yaokai@mathstodon.xyz 2026-09-12 14:24
@tao@mathstodon.xyz I don't think any announcement could salvage losses of mathematics. I want to say, you guys who owned Fields could only represent only few people around you guys. More people don't think AI is a threat except you geniuses. They are hugging AI for it could make geniuses feel fear. This sounds like ridiculous but this is truth. That's why people would rather forward AI's progress than trust Mathematician's debunkings. And not many people would think the world be changed and will volunte it. More people just want to watch the show and occasionally stir things up. And you know what? Your announcement could not make those companies stop their distributions on math problems. They will be more excited instead. Actually, people's already exhausted on this joking show. Unless some fundamental flaws in AI-generated articles are identified, the public will never regain confidence in mathematics. And there will be more and more young scholars leave mathematics if continuing under this condition. On the last, I suggest use the AI to review those papers that generated by AI. Since AI companies are leveraging computational power to undermine the academic peer-review process, they have a responsibility to provide sufficiently detailed and reasonable explanations for the papers they publish. This maybe the only way to handle the current "crisis".
-
@maherben@mathstodon.xyz 2026-09-12 15:29
@tao@mathstodon.xyz Great initiative! I hope this also includes software engineering; AI, like anything have double edge; the user chooses which edge they choose. Benifical or harmful. For continuous learning and democratizing knowledge, or for corporate greed to try to replace humans.
-
@sonatsuer@mathstodon.xyz 2026-09-12 23:43
@tao@mathstodon.xyz signing a declaration with the year of your fields medal is probably the biggest example of argument from authority i have ever seen. i think the last thing mathematics needs is a clergy, especially one consisting of people who partially represent only a certain kind of pure mathematics. in practice the focus of mathematics is not insight. i wish it was but it is not. just look at the awards the mathematics community deem prestigious, or try to fund a project which puts "understanding" as its main goal. the only redeeming thing in that regard is a handful of research blogs. the community expects you to solve a problem and be the first one to do it. there is this undeniable performative aspect of mathematics and a lot of people of influence pretend that it does not exist. the field is severely suffering from goodhart's law. if the point of mathematics is gaining insight, a more interesting discussion would be about how to direct and motivate mathematical research in a world without big open problems. corporations are corporations. there is nothing surprising in the way they behave. they are misaligned with humanity itself and one should resist them. this much is given. what made me write these at 2 in the morning is the lack of self reflection. the real misalignment is between what mathematicians preach and practice which AI happened to make more apparent. without addressing the problems of mathematics honestly we can declare whatever we want and it will not change anything. finally a political remark after seeing some of the jaw-droppingly naive comments mentioning fascism etc.: who decides ai companies are bad while it is ok to work for darpa on AI related topics? i would say all are evil.
-
@tsturm@famichiki.jp 2026-09-12 23:59
@tao@mathstodon.xyz @n1xnx@tilde.zone This sounds very true and mirrors the concerns that some software engineers have about the rush to “fix” problems in code with AI-generated solutions that are poorly understood and with no time to analyze further repercussions.
-
@sunnydeveloper@mastodon.social 2026-09-13 03:14
@tao@mathstodon.xyz we have a small working group trying to advocate similarly for misalignment of AI with open communities (where code and content is extracted). I've had very poor response trying to get labs interested in such alignment, and wonder how we can unifiy such advocacy. https://github.com/chaoss/wg-ai-alignment
-
@dacig@mastodon.social 2026-09-13 18:49
@tao@mathstodon.xyz Is another AI is possible? The declaration highlights how AI wrecks the social and creative endeavor of Maths. A bit short in solutions on a complex issue, it's also hard to argue against tangible results. What might be missing is that they seem to take for granted that the way most AI is: appropriate human knowledge without attribution or retribution, then gatekeep the results. Reappropriating the data and the tools, making it a commons with democratic governance is the only way forward.
-
@taksh@mathstodon.xyz 2026-09-14 04:39
@tao@mathstodon.xyz I have endorsed it, but I don’t fully agree that AI will single‑handedly destroy the math community. I think that in the next two to three years we will have AI that can solve most open problems. However, it could also be used to train and embed “math thinking” and understanding in humans, and we could optimize for that in the future. Why is that not considered a possibility? I believe that in the future we will have cool headsets that can read our minds and communicate with them, teaching us things so beautifully that we could never have done before. Such positive scenarios are highly likely. The headset example is, of course, an oversimplification, but the point is that we could have AI tutors and a math‑community builder, possibly fitting thousands of years of mathematical context and problem‑solving skills much faster into a person’s brain.