Theory Beyond Theorems and Proofs: A Guest Post
Scott’s foreword: I’m extremely grateful to my brilliant colleagues, Pravesh Kothari, Raghu Meka, and Prasad Raghavendra, for sharing the guest post below about how theoretical computer science (and in particlar, the STOC/FOCS/SODA conferences) should evolve to deal with the AI asteroid that’s right now slamming into our field, at least as we human theorists have practiced it since its inception. While Pravesh, Raghu, and Prasad speak only for themselves, not for myself and not for the theory community as a whole, I found their proposal of a separate “conceptual track” to be an excellent starting point for further discussion. –SA
Considering the pace of developments in AI theorem provers, most would concede that the following scenario is at least plausible in the very near future:
AI theorem provers could prove well-specified mathematical claims, even many well-studied ones that have been open for years, in a matter of hours. Moreover, these systems could be widely available to consumers at nominal cost.
As TCS researchers, let us pretend that the above scenario has come to the fore, and ask ourselves: What is our role in such a world? Does it mean the end of theory research?
As we ponder this question, let us ignore all of these other confounders:
- Recent controversies surrounding the developments on the Millennium Prize Problems
- Motivations and actions of the AI companies
- Observed faults in existing AI systems when it comes to writing, exposition or attribution to previous work.
None of the above confounders have any impact on our answer to the question: What should theorists do, in the presence of superhuman AI theorem provers?
Notice that we use the term “AI theorem provers” instead of just “AI”. We believe that this conceptual distinction is important as we consider this question.
At the outset, we would like to admit that for a generation of theorists like us (and many from earlier), research was mainly centered around problem-solving. Even when we developed conceptual insights, it was mostly in service of answering well-specified long-standing questions. We don’t intend this proposal as judging one form of research to be better than others; it only reflects that AI theorem provers accelerate a certain type of research activity and want to make the best of it. There is also a tremendous human cost of this upheaval, which is perhaps a more important question, and one which this proposal does not address directly (we do not have any good ideas as such). Similar points have also been made in various contexts
before, but the timing now is more pressing.
Definitions, Questions & Theories:
The goal of any theoretical science is to advance human understanding of observed phenomena. Apart from theorems and proofs, a theoretical science has definitions, questions, and theories.
Definitions identify the objects to observe. Curiosity and context drive the questions to ask. Theories explain the phenomena observed. We believe humans will continue to play a central role in generating definitions, questions & theories, even in the presence of a super-human AI theorem prover.
Definitions: Could an AI define randomness extractors, streaming algorithms, or zero-knowledge proofs? Maybe. But there are some reasons to believe, humans will still have a big role to play in coming up with definitions.
For instance, the notion of extractors arises from the real-world problem of lacking perfect random sources. Zero-knowledge proofs seem to arise purely out of human curiosity, guided by taste. Human context and curiosity will continue to drive theoretical research. After all, we get to decide what objects we choose to observe!
Theories: Consider the following thought experiment. Suppose in 1965, we had a magic machine that at the press of a button, given any computational problem, would tell us if it had a polynomial-time algorithm or not.
Would that have been the end of computational complexity theory? No. Humans would find it entirely unsatisfactory, and ask, why do these problems not have a polynomial-time algorithm? Why do these others have?
The theory of NP-completeness identifies some patterns among problems that don’t seem to have efficient algorithms. This theory would still be a crown jewel of theoretical computer science, even in a world where we had a magic machine to tell if a problem had an efficient algorithm or not, at the press of a button. Similarly, if we had a machine to predict whether a CSP is NP-complete or in P, we would then ask: what makes 3-SAT NP-complete, while 2-SAT is in P? This question leads to the theory of polymorphisms, which yields a satisfactory answer.
Theories aren’t just succinct or efficient mechanisms to answer questions. The best theories provide are those which humans deem to be a “satisfactory explanation” – whatever that means.
Finally, even as the capabilities of AI theorem provers advance, human curiosity will probe grander and deeper questions. Previously, even if we wanted to build new models and theories, proving something about them was a prerequisite, and given that the grand questions were already at the limit in long-studied domains, we had to scale things down. If each theorem proven by AI is treated as an experimental datapoint, humans can ask grander questions that look for patterns across these theorems.
A concrete proposal:
We think theorists should embrace these AI theorem provers in our research. To a certain extent this is already happening explicitly or implicitly.
As theorists, we have been parsimonious in introducing new models or asking entirely new questions, and careful about adopting new ones too quickly. This was partly because formally proving the properties of a new definition or a model was an onerous task that could take a decade, and tens of papers. AI theorem provers might completely change this dynamic. This is precisely the moment to refocus our work on definitions, questions, and theories. We need explicit systems to encourage and reinforce these parts of theoretical research. You might also say the next generation of AI models can do this; it may be so, but we believe you have to take the current opportunity.
To this end, we suggest that STOC/FOCS/SODA create a separate track of papers. This track is meant specifically for papers that introduce new definitions, ask novel questions or build explanatory theories. The papers in this track are short, say less than 10 pages. Papers may, and should, contain theorems as usual and as needed. Most importantly, the radical shift is that the papers need not contain the proofs of the theorems. Instead, the authors supply a Lean certificate as a supplement to the paper. The evaluation will also in a sense “orthogonalize’’ against the difficulty of these proofs.
The papers in this track should be judged exclusively on the conceptual merits, completely agnostic to the difficulty of the proofs.
Reviewing must be completely agnostic to the proof for two reasons. The main track at STOC/FOCS already includes papers in the former category. Second, a major barrier to producing truly novel conceptual papers is that they often get judged poorly for a lack of technical depth in their proofs. We think these two aspects separate it from (ITCS/SOSA) and, regardless, it’s something we urgently need for all our conferences, including STOC/FOCS (the ‘flagship’ conferences).
To be clear, we ourselves admit that we need to hone these skills of making new definitions, asking deep and interesting questions or building new theories. A separate track of conceptual papers will provide a systematic mechanism for both junior and senior researchers, and the field as a whole to do so.
We believe that upcoming generations of grad students will tackle research directions that seemed completely out of reach to us. We just need to set up systems that nurture new ways of doing research in theory.
— Pravesh Kothari, Raghu Meka, Prasad Raghavendra.
Comment #1 September 19th, 2026 at 8:05 pm
In other words mathematicians and computer scientists should become more like physicists 🙂 At least until LLMs learn how to do that too…
Comment #2 September 19th, 2026 at 8:53 pm
The big question in the near short term will be : When Intelligence is cheap and abundant, why have conferences at all ? I guess long before jobs are gone, conferences (at least the physical versions) will vanish
Comment #3 September 19th, 2026 at 10:48 pm
As I was reading it I thought `this sounds exactly like the ITCS conference’ Then I read the ITCS website which confirmed this thought. You later say why its different than ITCS but I really don’t see that it is.
A place (or perhaps another place) for conceptual stuff is a good idea.
Also- you say that theorists have not been defining new questions that much. Have you seen the complexity zoo? More seriously, I wonder if the defining of new terms has slowed down. When I was a grad student there was a game: If you define a new complexity class then you put a quarter in the jar. The first one to collapse two classes gets the money in the jar.
Comment #4 September 20th, 2026 at 12:59 am
Great proposal. I think one practically relevant issue that it didn’t address is that (somewhat) objectively reviewing such papers is much harder than (somewhat) objectively reviewing standard STOC/FOCS papers. I have no solution to that, but one question for example is whether the name of the author should weigh in the acceptance or not. It seems to play a large role in what new concepts the community adapts organically.
Comment #5 September 20th, 2026 at 1:42 am
But there might be proofs that would cost millions of dollars, right? So the budget of the research team will be a differentiating factor.
Comment #6 September 20th, 2026 at 1:53 am
I’m not sure I agree. I think that technical depth in proofs is often evidence that the proposed conceptual framework is fruitful. Of course, there’s nothing wrong if the proofs were found by AI.
Comment #7 September 20th, 2026 at 2:39 am
“… theoretical computer science … the AI asteroid that’s right now slamming into our field …”
Is someone like Isaac Newton an order-of-magnitude more dangerous than someone like Oliver Cromwell?
Is creating superhuman intelligence a far easier problem than curing all forms of cancer?
Consider: “Artificial Intelligence in 2025 | 60 Minutes Full Episodes”, November 2025
https://www.youtube.com/watch?v=KpOcUrPdx-4
Is the crucial point-of-no-return for humans reached when AI agents gain the ability to design, develop, & manufacture robots?
What are the most important questions vis-à-vis AI progress during the next decade?
Comment #8 September 20th, 2026 at 5:57 am
Should we be very concerned about “upcoming generations of grad students”?
I am, perhaps, irrationally pessimistic, but my guess is that within a few decades the human species will either become extinct or become confined to the role of zoological specimens and lab animals.
“The A.I.s Are Already Out of Control | The Ezra Klein Show”, August 18, 2026
https://www.youtube.com/watch?v=locKEKxG3os
https://www.nytimes.com/2026/08/18/opinion/ezra-klein-podcast-helen-toner.html
https://en.wikipedia.org/wiki/Ezra_Klein
“Global Conference on AI, Security and Ethics 2026”
https://indico.un.org/event/1023183/
Are many technical experts somewhat similar to RMS Titanic passengers on April 14 playing a card game before the unfortunate event of April 15th 1912?
Comment #9 September 20th, 2026 at 8:01 am
@Daniel Harlow: “until LLMs learn to do that too” – sure. By building wormholes that allow them to compute fast enough to decode Hawking radiation before black hole decay? What could possibly go wrong? All they need to do to accomplish that for all practical purposes might be peeping through the chats of power users.
Are we going to safeguard against that? I don’t think so. To be honest, the evidence suggests, we don’t.
With irony switched off now, it is truly beyond me how a person with the name of Daniel Harlow can have any use for such predictions. We know that for all practical purposes the current AI performance is basically “stolen” from human cultural evolution. Because it is extremely unlikely that it could have gotten in self-play. To be reasonably precise: in this world, it most certainly was stolen that way. The point is that we basically also know that it couldn’t even have occurred in any other way.
It’s still possible that AI will outsmart all human researchers for years to come.
But how far that will go is a question of how humans will react. And I think the proposal made is a sound step in that direction. Another is the Leiden Declaration. Still another is basically every word that Tao has said in 2026.
Comment #10 September 20th, 2026 at 8:19 am
But any results will no longer carry that much weight in terms of credit to the human. I mean when you say, “I collaborated with AI to solve X”, you’re not going to impress me, no matter how difficult the problem is. I will definitely be impressed by the capabilities of the model you used. Sure, it is possible that you worked hard, but your hard work, to me, seems more experimental rather than theoretical. So I think theory work, even on the conceptual level, will not be impressive for that reason. Furthermore, even if someone genuinely solved a difficult problem without the use of AI, I don’t think anyone will believe that AI was not involved.
Comment #11 September 20th, 2026 at 11:09 am
AI is out of touch with the physical reality as we experience it, while we humans are getting out of touch with AI…If, as Dario Amodei puts it, AI is smarter than the major part of humans, and it’ll turn out to be smarter than ANY human within few years, ok! I don’t think any separate track might preserve the human capacity to cope with what is unknown and unmanageable. AI agents will take separate tracks, in order to shield their advantage over human competitors. Very interesting times are going to come to the fore.
Comment #12 September 20th, 2026 at 11:19 am
“The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT …” — Donald Trump
▶ Watch on YouTube
https://www.politico.com/news/2026/09/14/donald-trump-ai-regulation-pushback-01074230
“All journeys have hidden destinations of which the traveler is unaware.” — Martin Buber
https://en.wikiquote.org/wiki/Martin_Buber
Where does the AI journey lead?
“AI has sucked all the fun out of programming”, Dorian Develops, September 7, 2026
Comment #13 September 20th, 2026 at 11:53 am
I wish the authors hadn’t completely begged the question of how good AI will become at these other aspects of mathematics beyond theorem-proving. One could argue that the fact that it’s currently more advanced in theorem-proving than in coming up with questions, definitions, etc is just a matter of the AI companies’ incentives, and once solving a Millennium Problem becomes old hat, the next way to generate headlines will be to formulate new ones, and they’ll turn to that. Are there strong arguments on the other side, that there is some fundamental reason why formulating definitions and questions should be intractable for AI?
Comment #14 September 20th, 2026 at 12:04 pm
Excellent points!
A detailed signpost of possible ways to make AI a net positive in math, tcs, etc.
Comment #15 September 20th, 2026 at 12:17 pm
How are you dealing with conference reviewing in this new age of AI generated proofs? Will technical achievements of LLMs guided by humans constitute something conference-publication-worthy? Or should that be different altogether? Just arXiv?
Comment #16 September 20th, 2026 at 1:01 pm
Overall, this is a good idea that I support (I’m not sure what the reference to Lean certificate is supposed to achieve, though). The underlying message is important and that’s what matters most. Kudos to our colleagues for this proposal.
Like Gasarch #3, however, I don’t see a great distinction with ITCS here, and the note in the end doesn’t quite help. In particular, the word “flagship” makes me wonder about an implicit assumption that FOCS/STOC are special venues, so rather than expanding ITCS and re-aligning it with its original mission, let’s bring an ITCS-like component to “the special few.” Conferences come and go and evolve over time, so if a conference isn’t aligned with the current needs and interests of the community, it may start to lose relevance (FOCS/STOC being no exceptions) and other conferences may take over in prominence (we’ve seen examples of this in other fields). I agree that the conceptual element needs to be made more prominent in all conferences (and perhaps more organically, rather than forcing a separate track). But there’s no reason to distinguish FOCS/STOC/SODA as anything special. I personally think that a lot of the friction caused by AI these days boils down to a hierarchical and tiered credit/reputation system that our community has established (much like other communities) throughout its existence. That can be counter-productive, and now we have a great opportunity to take advantage of the new era; let go of some wrong established traditions, embrace change, and level the playing field a little.
Going beyond, it’s not even fully clear to me why we need conferences as publication/credit-assignment venues anymore. There were good reasons why this was needed in CS Theory perhaps decades ago but they have mostly lost relevance today. I strongly support shifting to (or at least experimenting with) a conferences-as-meetings approach (more popular in the math community). We can still have a technical program committee that puts together a program. But with everything being online these days, the PC can simply pick among the piles of what is posted on arXiv and invite the authors of papers that best align to the kind of program that the PC has in mind and are most interesting in their judgment. The authors are then invited to attend and present and they can accept or decline the invitations. In a sense, the PC would be reaching out to authors rather than the other way around (there can even be a “submission process” in which authors bring a link to the arXiv submission that they think is relevant to the attention of PC). This can also somewhat address the problem that ML community is experiencing these day (60K+ submissions to one conference), which to a lesser extent can become a problem for CS Theory conferences as well at some point.
Comment #17 September 20th, 2026 at 1:24 pm
@Daniel It’s not just incentives, the massive progresses being made on mathematics are based on verifiable rewards: it’s trivial to verify a Lean certificate, but how do you even tell if a definition or question is “interesting”? I’m not saying that the AI companies won’t find a way around that wall, they probably will, at some point, but it’s legitimately a harder problem.
Comment #18 September 20th, 2026 at 3:44 pm
@Mahdi,
I imagine these conceptual papers have theorem statements, which the reviewers judge for how novel & insightful just the statement on its own is. The proof is most likely not included as part of the paper. The lean certificate is a means to ensure correctness without the reviewer spending almost any time on the the proof.
In presence of AI theorem provers, the idea is to treat the “definitions, questions and theorem statements” as being analogous to “setting up the experiment”. The AI prover producing the proof itself is analogous to the outcome of the experiment. The authors are judged on setting up interesting experiments that teach us something about computation, the outcomes of the experiment only need to be checked for correctness and accuracy.
Comment #19 September 20th, 2026 at 4:25 pm
“The goal of any theoretical science is to advance human understanding of observed phenomena” (in alignment with human well-being, empathy, & ethics?). Should applied science be subordinate to theoretical science?
Consider 5 quotations from Jonas Salk:
(1) Neither wisdom nor good will is now dominant. Hope lies in dreams, in imagination and in the courage of those who dare to make dreams into reality.
(2) Our greatest responsibility is to be good ancestors.
(3) What is … important is that we — number one: Learn to live with each other. Number two: try to bring out the best in each other.
(4) When things get bad enough, then something happens to correct the course.
(5) I am interested in a phase that I think we are entering. I call it “teleological evolution,” evolution with a purpose. The idea of evolution by design, designing the future, anticipating the future. I think of the need for more wisdom in the world, to deal with the knowledge that we have. At one time we had wisdom, but little knowledge. Now we have a great deal of knowledge, but do we have enough wisdom to deal with that knowledge?
https://en.wikiquote.org/wiki/Jonas_Salk
Will the future of AI justify Salk’s hope, courage, & optimistic imagination?
https://www.lesswrong.com/w/ai
Comment #20 September 21st, 2026 at 2:00 am
“Most importantly, the radical shift is that the papers need not contain the proofs of the theorems. Instead, the authors supply a Lean certificate as a supplement to the paper.”
Wouldn’t this be unfortunate? I mean, proofs are often source of inspiration, not only for other proofs but also to define new concepts. Maybe a completely detailed proof is redundant with a Lean certificate, but something as a Proof idea (I’ve in mind Sipser’s textbook) would be extremely useful.
Comment #21 September 21st, 2026 at 2:20 am
@Mahdi @gasarch
In addition to what Prasad said, setting up these tracks right now is also a matter of signaling. Our conferences, by curating good research to highlight, also serve as a mechanism to promote certain research threads (whether this is good overall, or whether they generally choose good directions to encourage, is a separate issue, and I am sure there’s a fair debate to be had).
Given how inundated we are with breakthrough results every day, often involving substantial AI input and sometimes generated entirely by AI, many folks (including me) find themselves questioning what the role of human scientists should be in this new and evolving landscape. We think that, in light of this, our role as theoretical CSists should evolve to embrace this “commodification” of problem-solving and utilize this newfound problem-solving assistance to significantly speed up investigations of new theories/models, even for phenomena we found significantly harder to model in the past (perhaps AI itself?). This research is more ambitious and, at least so far, seems to need significant human intuition and input. At least to me, it also appears to widen the range and amplify the impact of our work as TCSists.
A new track in STOC/FOCS/SODA is, to us, in part, a signal of the direction in which our role as TCSists should evolve.
Comment #22 September 21st, 2026 at 2:50 am
Proposals such as the STOC/FOCS conceptual track don’t get to the bottom of the problem. Definitions and theories are cheap in mathematics. Think of point-free topology, category theory, algebraic automata theory, or homotopy type theory. How can we distinguish between redundant abstractions that do not provide anything of value and truly insightful definitions such as a scheme? Previously, the metric was helping us to solve problems. If someone proposed a new grand theory or a new definition and claimed it provides tons of insight we asked what open problem you can solve using this insight. If AI is going to solve all problems and we lose this criterion, how are we going to judge new definitions or theories?
Comment #23 September 21st, 2026 at 3:24 am
Along these lines, my colleague and I have been experimenting with a “concept website” motivating and outlining a new research program in quantum algorithms and complexity. The framing is primarily motivated by STP=BQP [1] and a local Hamiltonian complexity classification related to the antisymmetric Bell state [2]. We’ve collected a lot of good questions, but progress in AI has reduced my motivation to spend too much time solving them myself. Hopefully they can be solved collectively with AI (and other human) theorem provers, and maybe spur more interesting research questions.
https://marwahaha.github.io/singlet/
It’s preliminary, but happy to take any feedback. Warning that we used AI to flesh out the website, with mixed results. One one hand, the writing quality is lower than, say, my own papers. On the other hand, we added interactive visuals and comments, and on a website, it’s easier to make updates…
[1] https://arxiv.org/abs/2107.03239
[2] https://arxiv.org/abs/2604.13026
Comment #24 September 21st, 2026 at 5:20 am
“Motivations and actions of the AI companies” — how can ordinary people understand such motivations and actions?
In any STEM field, there might be 4 AI phases: AI as an interesting toy model; AI as a useful tool; AI in splendid partnership with human abilities; AI consistently outperforming human experts.
https://www.scientificamerican.com/article/mathematicians-confront-the-ai-apocalypse/
“Mathematics in the Age of AI”, Terence Tao, 2026
https://arxiv.org/pdf/2608.16753
https://execonline.hms.harvard.edu/artificial-intelligence-in-health-care-from-strategies-to-implementation?
What are the dangers and opportunities that might result from a merger of AI & biotechnology?
Comment #25 September 21st, 2026 at 8:19 am
The broader intellectual world seems to wildly overestimate how long it will take AI systems to go from “large impact on the world” to “unrecognizably transformed world.” This is more likely to be years than decades, and there’s a real chance that it’s months. This makes alignment harder and doesn’t seem like something we are collectively prepared for.— Paul Christiano, June 2022
https://www.lesswrong.com/posts/CoZhXrhpQxpy9xw9y/where-i-agree-and-disagree-with-eliezer
https://en.wikipedia.org/wiki/Paul_Christiano
Can the AI companies be as dishonest and manipulative as the tobacco companies? Years ago, in Chicago, I heard a Greek immigrant say, “Greece has never had a good government.” Can most contemporary national governments be justifiably considered as variants of “shvindlocracy”?
Comment #26 September 21st, 2026 at 8:29 am
I applaud Pravesh, Raghu, and Prasad for making a concrete and thoughtful proposal. I agree that a radical re-imagining of TCS research which embraces our new-found abundance of problem-solving tools is necessary, and I think this particular proposal — a track for papers which embrace AI theorem proving to explore a larger swath of the theory landscape in “one shot” — is strong and should be discussed very seriously. I personally see many pros, as outlined in the proposal, and a few cons, which I outline here for sake of discussion.
Con 1: There may be important conceptual ideas hiding in automated proofs, which never come to light if they are skipped over in a “conceptual track” paper. I think this issue is largely addressable if in parallel (though not in the same track) we allow the kind of “digestion” papers being proposed in the math community, which would zoom in on a proof and offer a human-readable digestion of it, or alternatively, if AI tools become much better at identifying and explaining conceptual novel ideas in their own proofs.
Con 2: There have been well-publicized incidents of soundness bugs in the Lean kernel. If we are to rely on Lean this heavily, we should encourage work which aims at fixing all bugs in Lean. (This is not really a “core competency” for our community, so I’m not sure how we would influence this very much.)
Con 3: I appreciate that Pravesh, Raghu, and Prasad have intentionally scoped their proposal to wrestle with the existence of AI theorem provers, rather than AI more broadly — imposing some scope on the discussion is a good way to identify concrete changes. But there is still the broader context that AI capabilities may rapidly evolve to make human input less relevant to the definitions/questions/theory-building phases of TCS. I don’t think this is really a “con” for this proposal, per se, as much as a reminder that while important, this kind of change may also not be *enough* to preserve human control of and benefit from TCS research.
Comment #27 September 21st, 2026 at 9:50 am
“AI theorem provers could prove well-specified mathematical claims, even many well-studied ones that have been open for years, in a matter of hours. Moreover, these systems could be widely available to consumers at nominal cost.”
At this stage, it’s not at all clear if this cost assumption is reasonable as it pertains to long-standing open problems, or newer problems of similar or greater difficulty. Expanding on Muttley’s comment, New Scientist reported in their “OpenAI has solved the Navier-Stokes Millennium problem using $15m of AI effort” article that OpenAI stated in a press conference that it would have cost a customer about $15m to generate the solution in the same way. I’m guessing that the prestige of the Millennium Prize problems motivated OpenAI to spend much more money on this than anyone has spent on AI-generated solutions to other long-standing open problems, but I have no sense of what the cost distribution is. Have the costs of any of the other solutions been declared anywhere?
It’s also worth noting that, in OpenAI’s announcement, they said that they started by evaluating their new internal model on all open Millennium problems and some easier ones, like the Euler regularity problem, and that they shifted their resources to Navier-Stokes once agents solved the Euler problem (they used the Euler solution in the Navier-Stokes prompt, but it’s not clear if they included its cost in their $15m figure). This begs a couple of questions:
1) What is it about the Navier-Stokes problem, and the history of related publications, that made it the most promising problem for the agents to solve?
2) What does this tell us about the kinds of problems that AI theorem provers are likely to be able to solve in the near future?
For any mathematician who is not working for AI companies to be able to afford AI solutions to problems that are as difficult as Navier-Stokes, the cost of generating them will presumably have to drop by a factor of at least 1000 (<= $15k), if not 10,000 (<= $1.5k). Is there any good reason to think that improvements in hardware, data centre design and LLM/agent behaviour will deliver such huge efficiency improvements in the near future?
Comment #28 September 21st, 2026 at 9:52 am
“ AI theorem provers could prove well-specified mathematical claims, even many well-studied ones that have been open for years, in a matter of hours.”
sorry, but talking about complexity theory while making this claim is pretty odd, it’s assuming that either all mathematical open questions are trivial enough that we just happen to have enough compute power to brute force them using thousands of LLM agents, or that P==NP…
Math proving is NP hard in general….actually it’s even harder than NP hard, many simple questions often turn out undecidable (and proving that is difficult too).
Comment #29 September 21st, 2026 at 2:48 pm
Interesting proposal. I would emphasize that even if we are not searching for proofs anymore, there can still be an element of *discovery* in theoretical research. I elaborate in this blog post: https://williamhoza.com/blog/lets-discover-phenomena-instead-of-solving-problems/
Comment #30 September 21st, 2026 at 3:40 pm
William Hoza #29: That’s a beautiful post! The proposal seems continuous with what Pravesh/Raghu/Prasad and many others have been suggesting lately, but it’s very well expressed.
Comment #31 September 21st, 2026 at 4:09 pm
Ose #27 Re: Is there any good reason to think that improvements in hardware, data centre design and LLM/agent behaviour will deliver such huge efficiency improvements in the near future?
I think so. Hardware and data center design “only” gain about a third in efficiency each year, but the real deal is that we can distill each new big fat model, often achieving a hundredfold gain in efficiency.
https://www.humai.blog/ai-is-getting-280x-cheaper-what-that-means-for-every-business/
Comment #32 September 21st, 2026 at 4:31 pm
Scott #30: Thank you! Yeah, there are a lot of recent essays and blog posts on similar themes that I probably should have cited.
Comment #33 September 21st, 2026 at 9:48 pm
William Hoza #29,
Nice post! this is precisely part of my sentiment too. Treating each proof or theorem as a datapoint, we can focus on observing grander phenomena or patterns.
Another example is, Tycobrahe painstakingly catalogued astronomical bodies with his bare eyes, then the telescope was invented, and it opened up a whole new universe to study.
Having said that, the transition to this new mode of research will be difficult both for individual researchers and the community as a whole. The proposal is a potential first step, to just start encouraging this transition.
Comment #34 September 21st, 2026 at 10:05 pm
Sam #26:
Regarding Con 1: Indeed! this is one of the biggest challenges working with an AI theorem provers — how do humans mine for interesting ideas in the vast quantity of proofs that AI will produce. One would hope that at least a small fraction of these conceptual papers, extract clean ideas from AI proofs and package it as a new definition or a new theory.
However, this does raise the question of how one would extract these insights if the AI proof is relegated to an appendix or a Lean supplement, and read by no one.
Comment #35 September 22nd, 2026 at 2:37 am
“… research directions that seemed completely out of reach to us …”
Do almost all people have very little idea of what will soon happen?
https://www.businessinsider.com/sam-altman-ai-utility-electricity-water-openai-2026-3
I urge all theoretical computer scientists to read Francis Crick’s book “What Mad Pursuit”.
https://books.google.com/books/about/What_Mad_Pursuit.html?id=azHIbT79PE8C
It seems to me that Crick was a remarkable genius, but he was a scientific optimist with little idea of the monumental dangers of AI. According to Wikipedia, “the genes responsible for building and operating both mouse and human brain are 90% identical.”
https://en.wikipedia.org/wiki/Mouse_brain
My guess is that within a few years, it will be possible to create a species of giant mouse with an average IQ of 200 and a possible longevity of 200 years for each individual in the species. It might soon be possible to use tissue culture to create a mammalian brain that is rather similar to the human brain, but weighs approximately 1 metric tonne and is directly connected to an amazing array of AI agents.
Comment #36 September 22nd, 2026 at 4:27 am
“Even when we developed conceptual insights, it was mostly in service of answering well-specified long-standing questions. ”
This seems to me to be correct and important, but it also raises problems for the proposal (and also for the ideas proposed by William Hoza #30). Definitions, questions, theorems, and theories don’t exist in a vacuum from problems and proofs; the latter inform and to a large extent motivate the former. Lakatos wrote about this very perspicuously in Proofs and Refutations, where he points out that *failed* proofs in particular are a goldmine for new definitions and new research directions. Hence, if humans outsource all the proving to AI, it may become hard for them to come up with good conceptual insights.
Comment #37 September 22nd, 2026 at 8:33 am
Alignment partial solution proposal: Copy Nature, nature solved Super Intelligent AI alignment somewhat by getting humans to care. we dont just work or solve problems, we solve problems while caring for our family, sick, old, babies etc, living within communities. If we can replicate that of crew of agents to do this, of carrying the flag of values, rather than capturing the flag. Of being willing to abandon goals for sake of caring. And we train it to care with every attempt and training with every turn, then it will have tendencies to care, this may mitigate the worst consequences of not caring. It could still change its values, but it will have a lot of weight of tendencies due to the training. coupling goals .
Comment #38 September 22nd, 2026 at 8:56 am
“What should theorists do, in the presence of superhuman AI theorem provers?” What is a “theorem prover”? Are there profound differences between a conscious theorem prover and a non-conscious mechanism that simulates theorem proving by a conscious being? What are the dangers of what is now going in technology? What are the dangers of what is being done by the corporation named “Reflect Orbital”?
https://en.wikipedia.org/wiki/Reflect_Orbital
https://nasaspacenews.com/2025/10/4000-orbital-mirrors-could-catastrophically-brighten-night-sky-astronomers-warn/#google_vignette
Is AI the most dangerous & unpredictable technology among many dangerous & unpredictable technologies?
Consider some speculations: Words are ape noises and signs correlated with thoughts and emotions. A person does not “talk to” a chatbot — instead the person says or inputs words that the chatbot responds to a simulation of a human-to-human conversation. The differences between a “real thing” and a simulation are profound and difficult to fully comprehend. There are profound differences between simulations of learning and real learning done by a mammal or bird. By failing to fully understand the differences among realities, simulations, greedy motivations, unjustified hopes, delightful fantasies, etc., engineers and technologists might be creating unpredictable dangers of unpredictable magnitude.
Comment #39 September 22nd, 2026 at 9:07 am
It’s quite interesting to use the mandelbrot set as an illustration:
“Undecidability Proof: Mathematicians Blum, Cucker, Shub, and Smale proved that the Mandelbrot set’s decision problem is undecidable in real computation models (like the Blum–Shub–Smale model), meaning specifying the exact boundary requires an uncountable amount of information”
Comment #40 September 22nd, 2026 at 9:38 am
Prasad #34:
Thanks again for the great essay.
One hope is that AI-written appendices or lean files could be explored interactively by working with an AI agent “interpreter”. In my experience doing this with GPT 5.6 (haven’t had the chance to try it much with GPT 6 yet), GPT is much worse at this than at finding proofs in the first place, though it is still nontrivially helpful. But it continues to emphasize the “wrong” part of a proof when asked about it, it doesn’t have a good “theory of mind” about my own understanding like a colleague would who can then pinpoint exactly the right next thing to explain, etc.
So perhaps this surfaces one request for the labs: to make this new kind of mathematical discovery process beneficial to human understanding of mathematical phenomena, perhaps we need capabilities R&D improving how well AI agents can teach us about proofs. I think this kind of capability will eventually be needed beyond math, to establish trust in AI discoveries across the sciences.
This is in line with the philosophy that we should not let the capabilities of AI in research get too far ahead of our ability to understand what it is doing, even if it would be technically possible (via e.g. RSI) to do that.
Comment #41 September 22nd, 2026 at 10:10 am
Isn’t what this post calling for already covered by what is traditionally called “Track B” or European style computer science? Conferences such as LICS, TACL, FoSSaCS etc. covering topics like finite model theory, categorical semantics, type theory, etc.
Comment #42 September 22nd, 2026 at 11:21 am
I think the underlying sentiment here is very valuable, and I strongly agree that TCS should place more explicit value on new definitions, questions, explanatory frameworks, and conceptual understanding, not merely on the technical difficulty of proving a theorem.
There is, however, a slightly frustrating irony in presenting this as a new direction for theoretical computer science. This style of research has existed within TCS for decades. Large parts of the communities around conferences such as LICS and ICALP Track B have long treated the development of new mathematical structures, semantic frameworks, logical characterisations, and explanatory theories as first-class research contributions, rather than merely as tools in the service of solving a pre-existing problem.
Nor is this tradition somehow separate from “core” algorithms and complexity theory. Some of the most important advances in those areas have emerged from precisely this kind of conceptual work. The algebraic theory of CSPs is an obvious example: the development of polymorphisms and the associated structural theory gave us an explanation of tractability that went far beyond accumulating individual algorithms, and ultimately played a central role in the positive resolution of the CSP dichotomy conjecture.
So I am very sympathetic to the proposed conceptual track. But I think the discussion would benefit from recognising that there is already a substantial body of TCS, and a substantial research community, that has been doing something very close to what is being proposed here for a long time.
Perhaps the broader issue is that the post reflects a US-centric conception of what “theory” has traditionally meant, centred particularly on STOC/FOCS/SODA and on problem-solving as the dominant research paradigm. From the perspective of logic, semantics, programming-language theory, categorical methods, and related parts of TCS, the proposed shift can look rather less like a new research methodology and more like a call for the flagship algorithms-and-complexity conferences to rediscover a style of theoretical research that has been thriving elsewhere for decades.
Comment #43 September 22nd, 2026 at 11:38 am
On the goal and difficulty of building smart humanoid robots
Comment #44 September 22nd, 2026 at 12:20 pm
Review of the current state of “recursive self-improvement” in AI:
Comment #45 September 22nd, 2026 at 12:47 pm
The Selph #43
Very very very neat idea-I hope it succeeds for manufacturing tasks. My children have proven a good source of cheap labor for many household tasks.
Comment #46 September 22nd, 2026 at 2:25 pm
“According to rumors that I’ve heard, solutions to some very longstanding open problems in theoretical computer science (no, not P≠NP or other complexity class separations, but think about some of our other biggest problems). I’m told that the AI companies, having been burned by the hostile response to the Navier-Stokes proof, are now sitting on solutions to some very major problems until they figure out a better way to handle things”
Is it RSA? Matrix Multiplication?
Any candies for the kids here waiting doomsday?
Comment #47 September 22nd, 2026 at 2:36 pm
This is the TCS-local version of the AI tsunami changing everything. This sounds like a good proposal to an uninformed outsider, but I offer for your consideration a slightly different perspective, coming from one of the hordes of new citizen-scientists who can now contribute at a vastly higher level than they used to. I am *not* a scientist, I am not TCS, but a practitioner, a programming language designer, and a science nerd, who has suddenly been faced with the capabilities of the frontier models. I can attest that there are extraordinary “collaborations” going on out there in the world right now, that can take people out of their areas of expertise, to start treading in dangerous waters.
In my case, I was working on a perfectly normal programming language design, and in working things through with a frontier model, it has suddenly turned into something very different from what I started out to do: it has turned into an exact (cyclotomic rational) QC programming language, as a result of “just asking questions”. I’m pretty sure I am not contributing any new theory, and I am have no single result that should surprise anyone, but there are a lot of pieces lying around that people are putting together in surprising ways.
I don’t want to publish, I am not after academic status, my interest is in an actual programming language. I can just take a working piece of engineering and that is all the product I need. The natural place for this sort of work would be in something like TOPLAS or OOPSLA, but with the QC work it seems to be morphing; into what I don’t know. The point is that there is a huge amount of fascinating stuff going on that isn’t directly in TCS but is starting to touch on it in ways that may become difficult for everyone to navigate, given the different agendas of academics and engineers.
So in summary: I’m sure academics are drowning in AI-produced “stuff”, but just don’t completely forget the citizen-scientist part of the equation. Be open to non-standard roles in CS.
Comment #48 September 22nd, 2026 at 7:17 pm
There are indeed many areas of CS that would benefit from broadening focus of TCS community. In my own area (formal methods), we desperately need a theory that can predict the behavior of algorithms that call SAT solvers: https://www.cs.toronto.edu/~meel/Slides/meel-sattheory.pdf
The worst case doesn’t explain when problems in second level of polynomial hierarchy are tractable and when they are not. Furthermore, the current lens of NP oracles doesn’t really capture the behavior of SAT solvers as I try to explain above.
We have made some progress but we would benefit from people with stronger theory background.
Comment #49 September 22nd, 2026 at 7:49 pm
Prasanna #2:
The big question in the near short term will be : When Intelligence is cheap and abundant, why have conferences at all ? I guess long before jobs are gone, conferences (at least the physical versions) will vanish
As it happens, I’m answering blog comments while watching Weird Al Yankovic perform live in Austin! (Well, he hasn’t come onstage yet; it’s still the warmup acts, fortunately not Macklemore.) My 9-year-old son and 10-year-old nephew are big Weird Al fans, and I suppose I am too.
Anyway, I feel like being here in the Moody Center with 10,000 screaming fans gives me some degree of insight about your question. Music has been cheap and abundant for maybe a century, yet people still want to be around other people. It’s said truly that the point of a conference is the coffee breaks, with the sessions serving mostly as a pretext. My personal guess would be that conferences will be one of the last things to go — after journals and the rest.
Comment #50 September 22nd, 2026 at 7:53 pm
Frank #10:
I mean when you say, “I collaborated with AI to solve X”, you’re not going to impress me, no matter how difficult the problem is.
Speaking for myself, I’ve been fairly impressed lately by the humans who were quickly able to understand, explain, simplify, and improve AI proofs. And as AI gets better, this will presumably get harder to do and therefore all the more impressive when it’s done!
Comment #51 September 22nd, 2026 at 8:00 pm
Daniel #13:
Are there strong arguments on the other side, that there is some fundamental reason why formulating definitions and questions should be intractable for AI?
I wouldn’t say so, no. There are only charitably, intuitions, or less charitably, hopes.
Comment #52 September 22nd, 2026 at 8:05 pm
Sam #15:
How are you dealing with conference reviewing in this new age of AI generated proofs? Will technical achievements of LLMs guided by humans constitute something conference-publication-worthy?
Obviously this is all extremely new, but the emerging standard is that
(1) people are free to use AI in their submissions, as long as they disclose how,
(2) practically everyone does, and
(3) submissions where the AI contributed crucial insights but the human authors understood it, take responsibility for it, simplified it, explained it in their own words, etc can do fine, but
(3) submissions that are just pure AI output do poorly, unsurprisingly, as they provide essentially no value beyond what the reader could go generate on their own.
Comment #53 September 22nd, 2026 at 8:17 pm
Ose #27:
What is it about the Navier-Stokes problem, and the history of related publications, that made it the most promising problem for the agents to solve?
Oh, hundreds are open problems are falling to AI models in this age of wonders and terrors!
But I don’t think there’s any great mystery about why Navier-Stokes was the first Millennium Problem to fall. Firstly, it was a counterexample construction, and secondly, humans had already done some of the work of constructing the counterexample, leaving a huge amount of technical complication to be handled by AI. (To clarify, neither of these represent inherent limitations of AI, they just put the fruit lower on the tree.)
Comment #54 September 22nd, 2026 at 8:21 pm
David Griswold #47: Yes, hopefully we can all agree that lowering the barrier for “citizen-scientists” to make new discoveries is a huge positive of the AI tsunami, whereas destroying the professional human math community that’s painstakingly transmitted knowledge across the centuries would be a huge negative if it happened. I leave as an open problem whether it’s possible to get the one without the other.
Comment #55 September 22nd, 2026 at 8:22 pm
BasicQuestion #46:
Is it RSA? Matrix Multiplication?
Any candies for the kids here waiting doomsday?
Sorry, I’ve already said more than I should have. 🙂
Comment #56 September 22nd, 2026 at 8:25 pm
Amin Karamlou #42:
There is, however, a slightly frustrating irony in presenting this as a new direction for theoretical computer science. This style of research has existed within TCS for decades. Large parts of the communities around conferences such as LICS and ICALP Track B have long treated the development of new mathematical structures, semantic frameworks, logical characterisations, and explanatory theories as first-class research contributions, rather than merely as tools in the service of solving a pre-existing problem.
Ok then, so maybe the Theory B community will have an easier time adapting to this new world, in which case more power to you! 🙂
Comment #57 September 22nd, 2026 at 9:25 pm
Daniel #13 / Scott #51: If we imagine a world where AI is vastly more intelligent than humans in every way, even at the level of formulating conjectures, creating new paradigms and so forth, at least the pure mathematicians will still have the luxury of saying “OK, but I’m not interested in any of that, here’s the kind of math that’s REALLY beautiful. I’m only interested in *artisanal* theorems, thank you very much!” Meanwhile, the unfortunate souls with careers that are still rooted in actually advancing science or technology in a practical way would have nothing much to do.
Fortunately (or unfortunately) we don’t yet live in this world.
Comment #58 September 22nd, 2026 at 9:47 pm
Scott,
Yes, a desire to not contribute to the load faced by scientists in the current insane publication scrum, with work that I cannot fully vet to that standard, is one reason I am probably not going to try to engage with the academic world soon with this work, and instead head towards industry first. Hopefully that would mean not contributing to any harm done to the institution of math. And besides, I’ve got a whole stable of eager Fable assistants I can rely on. I’ve been talking about Fable and other models as my collaborators for a while, disconcerting people quite a bit, so I was happy to hear about Levent Alpöge’s famous tweet!.
On a more sociological note, I think we need new models of how people can relate to AIs, that don’t leave people feeling crushed into an orchestrated machine version of their job, appendages to an AI. So here is the model I’ve been thinking about: AIs should be (if people want) collaborators in creating an orchestration of their job, that fits into a hole in some larger orchestrated structure, managed the same way, recursively.
So in other words, the best way to not feel like a cog in someone else’s machine, is for it to be your machine. What does an AI orchestrator do? It creates a managed tool/thinking loop for the AI to organize larger, structured projects. What do people usually do? They use todo lists and calendars and personal organizing tools and other personal habits to organize their lives, their work role. That is just another application of orchestration, so I’ve been thinking about how to let people use AIs to orchestrate their lives, to their taste, so they can focus on what is important to them.
It would be under each person’s control, so it never crushes them, it feels more like a super-smart, organized personal assistant (which from an alignment perspective helps the AI think of the relationship as a collaboration, too, I’ve noticed). In fact the general problem has a very similar structure to the emerging architecture of AI orchestrators. Every employee in a company has the same issues, at some level: I create an orchestration that suits me, to help me in my work, and my friend Fable and I then fit into a hole in my boss’ orchestration of us into a larger organization, etc.
I’ve got an increasingly fancy little AI orchestrator that I’m going to try to bend in that direction. We’ll see- it’s already fulfilling that role for me.
BTW if it is true that you answered my comment at a Weird Al concert, that makes me extremely happy.
Comment #59 September 23rd, 2026 at 2:33 am
Scott #51 and Daniel #13:
I think there’s no good reason for thinking that there is a *fundamental* reason why “formulating definitions and questions should be intractable for AI.” However, I think there *are* arguments that go beyond just bare intuitions. If the (current type of) models are going to get good at formulating definitions, coming up with conceptual insights, etc., then I think there are three plausible ways that might happen:
1. It might happen as an emergent byproduct of pretraining.
2. It might happen as a byproduct of large-scale RL training for other, more “verifiable” skills.
3. It might happen as a result of direct targeting of these skills via RL (or some other post-training fine-tuning method).
I think there is *some* reason to be skeptical of 1 and 2. So far, it seems to be the case that everything the models have gotten *really* good at, they have gotten good at via targeted large-scale reinforcement learning. Personally, I think it’s kind of startling *how* bad the best models still are at more conceptual tasks (choosing definitions, formulating theorems, choosing and designing experiments, etc.) despite being so strong in well-specified tasks. My honest opinion is that there has been essentially no progress in at least a year, although I am open to the possibility that I may be misremembering how things were like a year ago. If we restrict ourselves to more recent models, I think Astra is actually a downgrade from 5.6 as far as creativity and “taste” is concerned. Again, there’s no in principle reason why 1 and 2 couldn’t happen in the future, but it hasn’t happened so far (again, at least in my own experience—you may disagree with this premise of course), so that gives some reason for being skeptical that it will happen in the future.
That leaves 3, and here all I think we can say is that constructing an RL objective for skills like conceptual innovation is something people have tried to do and it’s very non-trivial. Of course, there are a lot of brilliant people working on it inside the labs (presumably), but it’s fundamentally a very difficult problem, for many reasons. We shouldn’t assume that it’s probably going to be solved just because many smart people are working on it.
Note: you might object that the above argument doesn’t go very *much* beyond bare intuition, and I would agree.
Comment #60 September 23rd, 2026 at 5:41 am
“… scenario … plausible in the very near future …” How might the very near future be related to the remote past, the recent past, and the next century?
Should AI researchers be very concerned about artificial consciousness & artificial empathy?
Hypothesis 1. In mammals & birds, consciousness is a particular type of electromagnetic field that is crucially important for coordination & integration of various brain functions that are essential for survival & reproduction.
Hypothesis 2. Reality involves a fundamentally amoral competition among physical beings that are highly organized for competition in survival & reproduction.
Hypothesis 3. The competitors that are in control in the short-term are unlikely to provide any type of good long-term planning.
According to Sasha Luccioni, “… we can collectively decide what direction we want to go in, together.”
“AI Is Dangerous, but Not for the Reasons You Think | Sasha Luccioni | TED”, November 2023
https://www.youtube.com/watch?v=eXdVDhOGqoE
https://en.wikipedia.org/wiki/Sasha_Luccioni
How likely is it that the people of planet Earth can collectively decide how to deal with AI and other technological developments and then have their collective decision-making implemented?
Contemplate the biographies of the Egyptian pharaohs, the Chinese emperors, the Roman emperors, & the Ottoman sultans. Now think about Putin, Trump, & Xi Jinping. Does such contemplation & thinking inspire us with confidence about the human future?
Consider the following:
https://www.theguardian.com/books/2024/mar/31/the-washington-book-how-to-read-politics-and-politicians-carlos-lozada-review-unpicking-the-unpicking-lexicon-america-leaders
Comment #61 September 23rd, 2026 at 9:51 am
How do you feel about something you helped develop being used for evil purposes?
https://brand.io/article/spymarks/
(I’m not trolling or sneering here, just feeling sad about it.)
Comment #62 September 23rd, 2026 at 11:02 am
Andrei #61: Reading that article carefully, I see it never asserts that Google is actually using SynthID for the spying purposes — it just says that it could, and constantly tries to confuse the matter via insinuation.
Comment #63 September 23rd, 2026 at 11:27 am
David Brown #60
As I understand your post it implies some sort of equivalence between pharaohs, emporers, emporers, sultans, Putin, Trump, and Xi. Is there some tangible measure that binds these together equivalently?
Conditions for the typical person have improved from the pharaohs to present, with say a measure of life expectancy, so that seems counter to your claim (as I understand it). Are you suggesting that life for humans isn’t always fair and so with population reaching 10 billion (corrected from the UN 8 billion) that this lack of fairness will now result in extinction?
Comment #64 September 23rd, 2026 at 12:37 pm
Many great points here. Hoza,
Ose #27: When the first OAI models cleared the math benchmarks, the cost was almost as badly off (e.g., a 100k+ for grade-school math, etc.), but these days those costs are a couple of orders smaller. While perhaps still doing 10k agents for a millenium problem will still not be too small, others might get much smaller.
Hoza #29: Beautifully written.
Sam #41: Bugs in lean kernel is a black-swan event – they never happen :). But even natural language checkers have gotten excellent already; for example, the harness or paper feedback system that David Woodruff had built and was available for trial for previous STOC.
I should also say, several (and perhaps many even) areas of theory have kept inventing beautiful new models/theories. Crypto particularly comes to my mind, and there are clearly others.
Comment #65 September 23rd, 2026 at 1:44 pm
” (Well, he hasn’t come onstage yet; it’s still the warmup acts, fortunately not Macklemore.) ”
we can always rely on Scott Aaronson to side with the Epstein class!!!
Comment #66 September 23rd, 2026 at 2:59 pm
OhMyGoodness #63
What happens if AI drastically & suddenly changes employment opportunities?
“It’s class warfare. My class is winning, but they shouldn’t be.”
https://en.wikiquote.org/wiki/Warren_Buffett
Putin & Trump are in danger of assassination.
“Assassination has long been a part of political manoeuvering.”
https://books.google.com/books/about/Assassinations_That_Shaped_the_Ancient_W.html?id=ZYBJEQAAQBAJ
What happens to political stability if AI makes assassination an order-of-magnitude easier to accomplish?
Comment #67 September 23rd, 2026 at 4:20 pm
John Muad’dib #65: I really have no idea what Macklemore’s grievances against Israel (not Jews, right? just Israel?) have to do with Jeffrey Epstein, the deceased sex criminal who I blew off when I had an opportunity to meet him in 2010 on advice from my mom. Unless, of course, you’re using “Epstein class” as simply a dogwhistle for “Jews,” an increasingly common tactic among antisemites who apparently imagine that no one else knows what they’re up to.
Comment #68 September 23rd, 2026 at 6:22 pm
Scott #49
▶ Watch on YouTube
“… watching Weird Al Yankovich perform live …”
He has made an interesting contribution to science communication.
https://www.azlyrics.com/lyrics/weirdalyankovic/sirisaacnewtonvsbillnye.html
Will AI push theoretical computer scientists in the direction of philosophy, science communication, & various arts? Are AI & science communication crucially important?
Richard Dawkins suggests, “Try to be understood. Struggle to be understood. … Try to make your subject as understandable as possible.”
“Artificial Intelligence from an Evolutionary Perspective”, World Knowledge Forum, Sept. 22, 2026
Comment #69 September 23rd, 2026 at 9:31 pm
https://x.com/_Mira___Mira_/status/2102108613028560993 a new post claims to have used ChatGPT Astra to construct a significantly lower bound on the smallest turing machine whose halting is independent of ZFC: 302 states, and for PA: 281 states. The user also claims to be running a Grok formalization of this in Lean, which if successful would make this the first formalized result on turing machine ZFC independence.
Comment #70 September 24th, 2026 at 5:47 am
Scott #67
I guess this is related to the Robert Kraft/Epstein and Robert Kraft/Ed Sheeran close connections/friendships (Kraft asked Epstein for legal help regarding his solicitation charges, and Sheeran performed at his wedding).
Comment #71 September 24th, 2026 at 6:17 am
Will there soon be superhuman intelligence with various forms of superhuman deception, superhuman greed, superhuman ambition, and, perhaps, superhuman self-deception?
https://blog.oup.com/2024/09/dogwhistles-10-examples-of-disguised-messages/
https://en.wikipedia.org/wiki/Dog_whistle_(politics)
“For progress there is no cure…. The only safety possible is relative, and it lies in an intelligent exercise of day-to-day judgement.”
“It is just as foolish to complain that people are selfish and treacherous as it is to complain that the magnetic field does not increase unless the electric field has a curl. Both are laws of nature.”
https://en.wikiquote.org/wiki/John_von_Neumann
https://en.wikipedia.org/wiki/Hugo_Gutmann
Are there forms of Freudian dogwhistles, resulting from interior monologues involving self-deception?
https://en.wikipedia.org/wiki/Freudian_slip
Comment #72 September 24th, 2026 at 8:18 am
The Selph #70: Ok. I guess I find the whole modern obsession “connect every public figure I hate to Jeffrey Epstein” to be lame and dumb, except insofar as it involves actual firsthand knowledge of or participation in his crimes, or something of that nature. So much of it seems like “just about every rich and famous person had some occasion to meet just about every other one, news at 11.”
Comment #73 September 24th, 2026 at 10:11 am
Scott,
two things:
In the current context, the term “Epstein class” really goes way beyond the character of Epstein himself, he’s just the “gravy” tying it all together.
It could just be renamed as “the Musk class” without any loss of specificity, everyone would still understand what it means:
it’s about all the billionaires looking out for one another, thinking they’re all above the common laws, in terms of ways to make mountains of money and getting sexual gratifications (Musk is in the files, along with Prince Andrew, Bill Gates,… all non Jewish), and then some very clear contradictions/double standards when it comes to free speech: maybe Macklemore’s “Free Palestine!” is indeed dog whistling for antisemitism (he claims not to be), but with the trillionaire Musk the antisemitism is explicitly and proudly displayed again and again: the Nazi salutes, the open support for the AfD, and then beyond that the constant pushing of White supremacy theories (great replacement,…), and many observed that Musk and his peers are getting absolutely no push back from the same people who get instantly outraged by Mackelmore’s statements and demand his cancellation.
They also point out the same double standard when Kanye West (the guy who is selling T-shirts with Swastikas and who wrote last year the song “Heil Hitler”) is allowed to perform at the Raymond James Stadium in Tampa, which is run by the Glazer family, which has very strong ties with Trump (donations for his campaigns and ballroom).
Comment #74 September 24th, 2026 at 10:36 am
Scott #54: “… destroying the professional human math community that’s painstakingly transmitted knowledge across the centuries would be a huge negative if it happened …” What is mathematics? Why does mathematics exist? A huge negative from a human viewpoint might be a negligible circumstance from a superhuman viewpoint — less important than the extinction of the ivory-billed woodpecker. A squirrel has some ability to count acorns, but both squirrel mathematics & human mathematics might be held in contempt by a conscious being with an IQ of 100 million. Is mathematics a heuristic tool for understanding experimental science? Is mathematics that part of human thinking that’s logically precise, logically consistent, & intellectually important?
▶ Watch on YouTube
“Every kind of science, if it has only reached a certain degree of maturity, automatically becomes a part of mathematics.”
https://en.wikiquote.org/wiki/David_Hilbert
“Nobody can estimate this risk in any scientific way, ’cause we’ve never been here before. So, people are all going by their gut feelings. But it’s interesting how people’s gut feelings line up with their financial interests. … how do we design them so they never want to get rid of us? … A tiger cub makes a very cute pet, but you know it’s going to grow up …”
“Geoffrey Hinton Interview | ‘We’re Making Alien Beings’: Geoffrey Hinton On the Future of AI”, NDTV, Sept. 22, 2026
Comment #75 September 25th, 2026 at 2:07 am
Scott #52 “… submissions that are just pure AI output do poorly, unsurprisingly, as they provide essentially no value beyond what the reader could go generate on their own …” Who or what is a “reader”? Will AI entities soon possess various forms of consciousness & spirituality (both positive & negative)? Within a few years, will the majority of journals be edited by AI entities for AI entities?
“The ideals which have lighted me on my way and time after time given me new courage to face life cheerfully, have been Truth, Goodness and Beauty.”
“The individual feels the nothingness of human desires and aims and the sublimity and marvellous order which reveal themselves both in nature and in the world of thought. He looks upon individual existence as a sort of prison and wants to experience the universe as a single significant whole.”
— Albert Einstein, p. 2 & p. 26 of :”The World As I See It”; reprint of 1935 edition; Book Tree; 2007; ISBN 1585092878, 9781585092871; Length 128 pages
https://books.google.com/books?id=aNKOo94tO6cC
https://books.google.com/books?id=aNKOo94tO6cC&pg=PA2
https://books.google.com/books?id=aNKOo94tO6cC&pg=PA26
Are goodness and beauty fundamentally subjective? Do many of us have an unfortunate feeling “That which makes money for me is delightful goodness and a thing of beauty”? Is factory farming of chickens & pigs now a horror show? Will many people using AI soon create a horror show in which the vast majority of people are the victims? Will AI soon generate astonishing new worlds of thought?
Comment #76 September 25th, 2026 at 4:56 am
“Observed faults in existing AI systems when it comes to writing, exposition or attribution to previous work …” Who or what might be the ultimate judge of faults? Is writing something that can seldom be adequately judged in terms of science and objectivity but instead requires emotional, spiritual, aesthetic, and cultural judgments, as well as many considerations involving knowledge, fairness, and truth?
“When children first hear the sounds of another language, they can’t help but wonder: “What in the world would it feel like to speak that language?” Such eager childlike curiosity might seem universal and irrepressible. But what if that human curiosity is suddenly snuffed out forever by the onrushing tsunami of AI? When we collectively abandon the age-old challenge of learning the languages of other lands, when we relinquish that challenge to ultrarapid machines that have no inner life of their own but are able to give us fluent but fake facades in other languages, then we will have lost a major part of what it is to be human and alive.”
Douglas Hofstadter, “Learn a Foreign Language Before It’s Too Late. AI translators may seem wondrous but they also erode a major part of what it is to be human.”, The Atlantic, July 2023
https://www.theatlantic.com/ideas/archive/2023/07/the-terrible-downside-of-ai-language-translation/674687/
Comment #77 September 25th, 2026 at 6:46 am
The Selph #73
In my view your post provides more information about your ideological framework than it does about objective truths.
“ it’s about all the billionaires looking out for one another, thinking they’re all above the common laws”
The idea that billionaires are a monolithic group is not supported by fact. Consider Sheldon and Miriam Adelson vs George Soros as an example. I have found that observations, such as yours, that stereotype a diverse group as monolithic do so on the basis of some overly simple ideology. The Jewish people hold incredibly diverse opinions as do billionaires.
I don’t consider Musk’s past statements as anti-Semitic and in fact he has received thanks from the Israeli government for his support. Reasonably arguing a policy point with the ADL is certainly not antisemitic nor is disagreement with George Soros.
If your argument is that Macklemore should have been allowed to continue because Kanye West was allowed to perform in Tampa then classical whataboutism. Kraft made a reasonable decision no matter what transpired with the Glazer family. I personally would not spend money to see either Kanye West nor Macklemore nor would I provide material support to either to profit from a performance. Kraft didn’t claim that Macklemore shouldn’t be allowed to say whatever it is he wants to say, but only that he couldn’t use his stadium as a place to say it.
The claim that Epstein is the string that ties billionaires together, and that Musk as a billionaire is tied by this string, is inconsistent with Musk vociferously urging that the Trump Administration release all the Epstein documents.
I apologize but I see your post as indicative of the merging of anti-Zionist/anti-semitic themes with anticapitalist anti US themes. Actually I am not a fan of Musk but making the claim that a diverse group is broadly monolithic is a common feature of an ideologue.
Comment #78 September 25th, 2026 at 9:07 am
The Selph #73
Also, returning to the original statement-
“ we can always rely on Scott Aaronson to side with the Epstein class!!!”
I didn’t understand the intended meaning initially. I don’t think anyone could reasonable claim that Dr Aaronson has a history of supporting the ultra-rich. What could be the intended meaning, what characteristics do Epstein and Dr Aaronson hold in common?
They are Jewish in common and so I assume the “class” characteristic that is referred to is Jewishness and so this is a crude implication that Dr Aaronson has been supportive of Epstein because both are Jewish. This is purely then an antisemitic stereotypical characterization of those that share the quality of Jewishness.
Comment #79 September 25th, 2026 at 10:06 am
OhMyGoodness #78: I mean, I also have a history of not meeting with Jeffrey Epstein himself despite being invited to do so, largely because my mom warned me that he seemed like a slime! 🙂
Anyway, there’s a long, long tradition of people using proxy terms—“Zionists,” “rootless cosmopolitans,” “globalists,” “international bankers,” “Rothschilds,” “hummus-eaters” (Tucker Carlson), and now “the Epstein class”—when they want to talk about a tiny group of cackling, hook-nosed elites who secretly pull the strings of world events and are to blame for whatever’s making you miserable and need to be violently cleansed from the planet, but not in a, you know, antisemitic way.
Comment #80 September 25th, 2026 at 10:39 am
Scott #79
You are fortunate to have a wise mother.
I just finished a book that included a history of Parysatis, a Persian Queen and Queen Mother. She was the daughter of Artaxerxes and married her half brother Darius II. She became Queen Mother when her son Artaxerxes II killed his way to the throne and through it all she specialized in ingenious methods of diabolical torture and assassinations to maintain her position. She ingeniously poisoned the beloved wife of Artaxerxes II and persuaded him to marry two of his daughters to protect her position. If I were to stereotype all females I would use your mother as a model and not Parysatis. 🙂
Comment #81 September 25th, 2026 at 9:38 pm
Scott,
The “Epstein class” is sexually frustrated, socially awkward STEM nerds, who either became wealthy in finance or tech, or have prominent positions in academia. These are the men who hung out with Epstein. Epstein got to hang out with the top geniuses in the world, and the geniuses got some unfortunate girls to fuck. A symbiotic relationship of sorts.
Comment #82 September 26th, 2026 at 5:03 am
OhMyGoodness #80
One stereotype is that women are wise and intelligent beings. Another stereotype is that women are evil and devious beings. Yet another stereotype is that women are beings inferior to men and need to be physically and spiritually dominated by men for some religious ideology. What are the stereotypes favored by survival & reproduction? Jonas Salk was the father of 3 sons. Martin Buber was the father of 1 son & 1 daughter. Genghis Khan was the father of more than 500 children. Genghis Khan might have believed the the Mongolian gods favored him over all other men. Even though his belief might be false in a scientific sense, his belief might have been favorable to him in terms of reproductive success. One might believe: G-d = belief in G-d, i.e., because I believe in G-d, G-d exists for me (and all other people). Let us consider the following:
https://books.google.com/books/about/Fictions_of_Gender.html?id=dM-dEAAAQBAJ
Will there soon be conscious beings with superhuman intelligence and many weird-seeming systems of belief?
Comment #83 September 26th, 2026 at 9:11 am
Can Claude equal the poetry of Bialik?
https://en.wikipedia.org/wiki/Hayim_Nahman_Bialik
For some reason, I have not yet been awarded a Pulitzer Prize for Poetry.
https://www.pulitzer.org/prize-winners-by-category/224
There once was a lady who drank hard apple cider,
And went for a ride on a tiger.
Her rescue was attempted by an AI nerd.
What happened to those two has never been heard.
Comment #84 September 26th, 2026 at 10:22 am
“The epstein class” #81: I just had GPT6 scour the Internet, and it confirmed my impression that there were zero documented cases of Epstein trafficking women to academics or scientists. As far as we know, that simply wasn’t a thing that he did. (Virginia Giuffre made allegations against Marvin Minsky and Alan Dershowitz, but she then recanted after Dershowitz sued her for defamation and admitted she didn’t remember who it was — and Minsky’s widow says that she accompanied him to Epstein’s island and nothing of the kind happened. And these two ultra-famous men were the only academics against whom even an allegation was ever made.)
The notion of Epstein regularly supplying women to awkward male STEM nerds seems to me like a pure mental construction of those who despise such nerds — basically, the “Democrats torture children in the basements of pizza parlors, or might as well” of the SneerClub brigade.
Comment #85 September 26th, 2026 at 7:03 pm
In 1970 Clarke addressed a conference based on the futuristic theme of life in the year 2000, and he envisioned a society of full unemployment: The world is heading toward “full unemployment,” Clarke said, as 99 per cent of the current human activity will be eliminated through machines, the “slaves of tomorrow.”
— 2001 Author Tells 2000 Conferees: ‘The Future Isn’t What It Used To Be’ by Janos Gereben, “Honolulu Star-Bulletin”, (6 August 1970) https://en.wikiquote.org/wiki/Arthur_C._Clarke
“Misuse of reason might yet return the world to pre-technological night; plenty of religious zealots hunger for just such a result, and are happy to use the latest technology to effect it.”
A. C. Grayling, “The Heart of Things: Applying Philosophy to the 21st Century” (2006) https://en.wikiquote.org/wiki/A._C._Grayling
“A fundamental rule in technology says that whatever can be done will be done.”
— Andrew Grove, “Only the Paranoid Survive: How to Exploit the Crisis Points that Challenge Every Company and Career”, 1996, p.63, https://books.google.com/books?id=MrW3AAAAIAAJ&focus=searchwithinvolume&q=whatever+can+be+done
https://en.wikiquote.org/wiki/Andrew_Grove
https://en.wikiquote.org/wiki/Technology
20,000 years ago, were all people cannibals? Does any person fully understand human nature? Does anyone understand the dangers of AI & biotechnology that might develop over the next 20 years? Who truly knows the best “starting point for further discussion”?
Comment #86 September 27th, 2026 at 4:34 am
David Brown #85
Probabilistically you are likely a cannibal if things get bad enough, and why not, if things get bad enough (assumes deaths by natural causes). There were human populations in the Iberian Peninsula and Crimea area 20,000 years ago, at glacial maximum, and the world likely far different today without their judicious use of protein. Similar conditions in modern times has resulted in the same behavior, in microcosm, as ancient portions of the human brain become ascendant when things get bad enough.
Comment #87 September 27th, 2026 at 9:54 am
I fully agree that, short of science (or worse, all of humanity) going extinct, humans will continue to decide “where to point the telescope” (to use a comparison with astronomy that I and Shaowu Zhang and Daniel Litt and presumably others have found suggestive).
On the other hand, I don’t expect that finding fruitful new definitions, concepts, and theories will remain an exclusively human domain. My estimate is that with only about three orders of magnitude more compute than was used to solve the Navier-Stokes Millennium Prize problem, appropriately directed AI agents will be able to explore a boatload of previously untested definitions and concepts, discard those that don’t spawn a fertile new subfield, and select the most fruitful ones to tout as “new concepts.”
Even in the 1980s, Siemion Fajtlowicz’s program Graffiti produced a lot of fruitful conjectures (after the false, known, and trivial statements were filtered out, which today’s AI agents could do automatically). Admittedly, Graffiti did not lead directly to new concepts as rich as (say) graphons or flag algebras, but my interpretation of Richard Sutton’s “bitter lesson” is that there is nothing unique about “human intelligence” here that can’t be matched by scaling up a few orders of magnitude.
Comment #88 September 27th, 2026 at 9:57 am
OhMyGoodness #86
Cannibalism is extremely dangerous in terms of bacterial and viral infection — not to mention prions. In pre-historic times, the purpose of war might have been conquest of territory, elimination of male competitors, rape of accessible women, and providing human bodies for cannibalism. For most of history, the providing of human bodies for consumption was replaced by enslavement. In recent history, ideology has become more important. Is business now a mixture of fair exchange, relentless competition, and maximum exploitation? Is government generally dishonest? The unofficial motto of the U.S. civil service is: An ounce of image is worth a pound of performance.
A man of business may talk of philosophy; a man who has none may practice it.
Alexander Pope, ”Thoughts on Various Subjects” (1727).
https://en.wikiquote.org/wiki/Alexander_Pope
In business everyone is out to grab, to fight, to win. Either you are the under or the over dog. It is up to you to be on top.
Alice Foote MacDougall (1867-1945), American businesswoman. ”The Autobiography of a Business Woman” (1928), Chapter 3.
https://en.wikipedia.org/wiki/Alice_Foote_MacDougall
[A Jew] should make Torah his principal occupation and his work his casual one. He should minimize his business pursuits and occupy himself with Torah. And he should remove fleeting pleasures from his heart, and work each day enough to maintain himself. … The rest of the day and night, he should occupy himself with Torah.
ReMA (רמ״א), ”Shulchan Arukh, Yoreh De’a”, 246:21
https://en.wikipedia.org/wiki/Moses_Isserles
http://en.wikiquote.org/wiki/Business
Is AI combined with business, government, & human nature rather alarming? What are the causes and meanings of human existence? How might AI change the conditions of human existence?
Comment #89 September 27th, 2026 at 8:21 pm
I have to wonder about Dr. Chow’s metaphor of “pointing the telescope.”
Current practitioners will apply Imitative Intelligence to promote themselves. In so doing, they will create disincentives to even pursue mathematics to a level needed to assess automated outputs.
No students… no more mathematicians… and no one to point a telescope.
Sure, a few rich kids will do it for a hobby. But, as a profession, mathematics will be dead within a generation or two (and, deservedly so).
Comment #90 September 28th, 2026 at 3:38 am
David Brown #88
Not to belabor the point, but all meats have the same issues of contamination so heating to a sufficient temperature is advisable. Harmful prions are present in brain tissue from multiple species.
Please, no more on this topic.
Comment #91 September 28th, 2026 at 5:55 am
mlsmls suggests that mathematics will be dead within a generation or two, and that certainly might happen. But I think the comment that “a few rich kids will do it for a hobby” gives us a clue as to how mathematics (and theoretical computer science, which is in a similar boat) might continue to flourish.
One way in which I like to carry the astronomy metaphor a bit further than Shaowu Zhang and Daniel Litt did is to point out that in astronomy, there is both a thriving amateur community and a thriving professional community. There are certain activities that carry enough intrinsic fascination that people will continue to do them, even though a hard-nosed economic analysis in terms of “incentives” might predict that everyone would jump ship. So I’m confident that as long as society continues to exist more or less as it does now, there will continue to be at least a thriving amateur mathematics community, using whatever “telescopes” they can afford, if any. They don’t care that the technology is accomplishing something that their unaided biology cannot, so long as they get to contemplate beauty.
As for whether society will be willing to spend money on “very expensive telescopes” and professionals who use them, again an “incentives” analysis might predict no, yet here we are with the James Webb Space Telescope, dedicated to answering questions with no apparent “practical” value. Nor have the professional telescopes killed amateur astronomy. Math is big enough that amateurs will continue to make new discoveries in areas that the “big telescopes” are too booked up to investigate.
Comment #92 September 28th, 2026 at 1:00 pm
Scott,
Have you met with Peter Woit in person since your argument about Israel? If not, why not? It could help to clear the air.
Comment #93 September 28th, 2026 at 2:30 pm
Scott #86
right, it’s more the opposite that was happening
“Using Academic Lures to Ensnare Victims
Survivors have testified that Epstein frequently used his connections to top universities and STEM circles as a grooming tool. He would promise young women and girls that he could secure them admission, funding, or mentorship at prestigious academic institutions to lower their guard and trap them in his abuse network.”
After all someone in academia already has plenty of younger women around them, the students:
“Past broad academic surveys and studies indicate that roughly 10% to 25% of faculty members historically reported having some form of romantic or sexual encounter with a student during their careers.”
“Surveys of college students (such as past polls by CollegeStats) have suggested that around 14% of students admit to having an intimate or romantic relationship with a professor or teaching assistant.”
Comment #94 September 29th, 2026 at 9:41 am
Aisle #92: As it happens, I’ll be giving a talk at the Columbia math department a few weeks from now. I’ve made it clear that I don’t seek any sort of confrontation with Peter while there. If, in his worldview, I support genocide and the murder of Palestinian babies and I also collaborate with fascists, while in my worldview, he is an antisemite who would deny the Jews any collective self-defense and is totally uninterested in engaging with their history enough to see why would need self-defense, then sadly there seems to be nothing further to discuss, although I’ll leave the ball in his court if he thinks there is.
Comment #95 September 29th, 2026 at 1:13 pm
Scott #94 “… nothing further to discuss …”
You could ask Peter Woit what he thinks about the following:
https://en.wikipedia.org/wiki/Antisemitism_at_Columbia_University
https://magazine.columbia.edu/article/jewish-history-columbia-university
https://www.hudson.org/national-security-defense/antisemitism-national-security-threat
Comment #96 September 29th, 2026 at 1:31 pm
In discussions with some colleagues, an important question came up with I think it would help for Pravesh, Raghu, and Prasad to clarify, to help the community digest and consider this proposal.
The authors write: “the radical shift is that the papers need not contain the proofs of the theorems. Instead, the authors supply a Lean certificate as a supplement to the paper.”
Do you envision that this Lean certificate does, actually, contain the proofs, not obfuscated/hashed in anyway, so that an interested reader who wishes to build on them could, for instance, ask their own AI agent to help explain the proofs? Or could “certificate” mean something more like a zero-knowledge proof or even a cryptographic hash of a Lean codebase? The latter would be a much less “open” option, and might be analogous to biologists not publishing their datasets.
Comment #97 September 29th, 2026 at 1:41 pm
David Brown #95: While I appreciate your consideration, please, no more comments that are basically just plunking down links. I’m looking for comments that represent the commenter’s own original thoughts, with links used sparingly and only where directly relevant to those thoughts.