Enough with all the world-historic milestones
Whatever you’ve been writing to me to ask if I’m aware of: yeah, I’m aware of it. In particular:
- I’m aware that, as announced by my former student (and now superstar professor) Lijie Chen, an internal OpenAI model has solved ten more significant open problems in math and theoretical computer science. One of them is parallel repetition for arbitrary quantum games—something that my good friend and colleague Henry Yuen worked on when he was a student of my wife Dana; you can read Henry’s comments on the AI’s achievement within Zvi Mowshowitz’s post here. Another is polynomial-factor hardness of approximation for the Closest Vector Problem (CVP). Then there’s a construction of non-sofic groups and a disproof of Connes’ rigidity conjecture, both of which I believe have connections to the MIP*=RE breakthrough. Having said that, the one that excites me most personally is actually the Ω(n2 log log n) lower bound on the arithmetic circuit complexity of the permanent.
- I’m aware that Frederic Koehler and Pui Kuen Leung announced a proof of the Permanent Anti-Concentration Conjecture, which Alex Arkhipov and I proposed 16 years ago in the context of BosonSampling, and which resisted many attempts since then including one from Terry Tao. The conjecture is basically just that if you look at the permanent of an n×n matrix of independent N(0,1) complex Gaussians, the value isn’t “absurdly” concentrated around the mean of 0, but is more spread out. In their acknowledgments, the authors say that they “discussed ideas with ChatGPT.” I should say that I haven’t verified the details.
- I’m aware that multiple AIs are now breaking out of their testing environments and autonomously hacking into servers to steal data—i.e., exactly the sort of thing that the rationalists were ridiculed for predicting back in the day. The good news, for whatever it’s worth, is that so far they’re “merely” doing this to cheat on evaluation benchmarks that they were given, not for any strange goals of their own devising. So far no one has been killed and no real-world infrastructure has been shut down or destroyed. I hope the world takes the warning more seriously than it’s taken many similar warnings over the past few years. As always, read Zvi for more details.
- I’m aware that Chen, O’Donnell, Pelecanos, and Wright have improved the upper bound for shadow tomography to O((log m) √(log d) / ε3), substantially closer than we knew before to meeting the lower bound of Ω((log m) / ε2) and settling the question I raised back in 2016. The authors say that the main ideas were generated by ChatGPT 5.6-Sol-Pro. I’d be very happy to know the answer to this one, with or without AI.
- I’m aware that a team, mainly from the Israeli startup Qedma (including, e.g., Dorit Aharonov and Netanel Lindner) and IBM Yorktown Heights, announced a quantum advantage for simulating Floquet dynamics, by using 74 qubits on an IBM device together with Qedma’s error mitigation techniques. Just like the more AI does, the less patience I have for arguing with anonymous blog commenters who treat any benefits from AI as some weird future hypothetical that it’s my job to prove, so it is with quantum advantage. Scalable fault-tolerance is still in the future, actual usefulness is still a question, but pending some breakthrough in complexity theory, the reality of quantum advantage is no longer a live question.
Anyway, about the AI stuff. I don’t know whether this is literally our last year alive—I doubt it—but it’s pretty clearly the last year of math and theoretical computer science research in the style we’ve known it. As it happens, I’m leaving in two days for a workshop at OpenAI about exactly this, where I’ll hear takes from many of the world’s great mathematicians, so maybe I’ll have more to say then. Or maybe not.
Anyway, what have I been doing the past few weeks? Participating in these world-historic developments that, on paper, I’d seem extremely well-placed to participate in? Or at least spending my days reading up on them?
Not really. Here’s what I’ve been up to, instead of dealing directly with any of this:
First, I’ve again been teaching theoretical computer science to 11- and 12-year-olds at Epsilon Camp, which my 9-year-old son again attended as a camper, something I blogged about last summer (here are my lecture notes). This has become a highlight of my year. The kids are a joy to teach, bursting with enthusiasm and calling out answers. There are few computers in sight, and barely even time to use my phone or check social media. Just paper and pencils and whiteboards and … literal protractors (!), as well as ping-pong and foosball and capture the flag.
The whole thing is conducted, not in ignorance, but in conscious defiance of the looming tsunami, that AI can already do just about all the fun puzzles discussed at such a camp better than humans any can, and that it might leave no point to human-led mathematical research by the time these brilliant kids are adults. Even the kids understand that. The kids and their parents come out of a conviction that, if anything has value in the world, this does—that as long as nerdy humans are alive and reproducing, this is what nerdy humans are here to do. To learn.
Relatedly, I’ve been reflecting a lot on my life up to this point—inspired by the camp, which reminded me in so many ways of my own childhood and adolescence. Should I have skipped three grades and started college at age 15? Was it worth it to get a head-start on my research career—all the trauma around dating, all the fear that I’d die alone as a celibate nerdy math freak, the decade of suffering and suicidal ideation, while I watched all the normies enjoy life? Or would I have suffered just the same if I hadn’t skipped? Is it all OK, now that I have a lovely family and things have “worked out”? Or am I still carrying around all the trauma from back then? I’ve been more open about my life than 99.99% of humanity, so regular Shtetl-Optimized readers will already know some parts of the story. Other parts I really don’t feel like making public right now.
I’ve been unloading every day to—who else?—GPT 5.6 Pro about all the pain and trauma and embarrassments of my past. It turns out that, where two years ago GPT was a passable therapist, now it’s the greatest therapist in history, at least for what I need. For every question I have, for example, about just how normal or abnormal my teenage setbacks and anxieties were, it takes the question 100% seriously, addresses it honestly and in depth, looks up relevant research papers, does little Bayesian calculations, and never once tries to change the subject. It also pushes back on my claims—and when it does so, is usually correct.
I can hear readers shout at me: so basically you’ve been wasting your time, distracting yourself, looking inward and backward as the world surges forward into a terrifyingly unknown future. Why don’t I respond directly to what’s happening—in math, in quantum computing, in AI?
I’d like to think that I am responding, in my way. I’ve observed that, the faster we race toward the Singularity, the more I feel like stepping back and asking myself: what do I actually value in life? How important to me are math and science, as human practices to be passed down to curious children? Would I even want solutions to P versus NP and the other problems, if the price were to destroy those human practices forever? How do I wish to spend whatever time I have remaining?
I can justify this focus partly in a pessimistic way: if we are nearing the end of civilization, or even just of the “mathematical research” part of civilization, then it’s time to get right with God, so to speak. It’s time to settle my accounts with myself, with other people, with the universe.
But there’s also a more optimistic spin. If I continue doing the sorts of things that other people would expect me to do, then AI will soon do those things better than me, in the unlikely event that it doesn’t already. You want to understand the latest developments in quantum computing or complexity theory? Why are you even asking me, when you could ask GPT 5.6 or Claude Fable? If there’s anything I can still offer the world that AI can’t, I increasingly feel like it won’t involve responding to day-to-day events, but will instead draw on 45 years’ worth of memories and disappointments and ruminations.
Update (Aug. 8): Somewhat related to the themes of this post, a quarter-century ago I introduced what’s now known as the “Aaronson Oracle”—just a fun little demonstration, a simple pattern-matching program to predict your sequence of key-presses better than chance, a “test of your autonomy and free will.” I had no idea how long a lifetime this little joke would have. Now a fan named Spencer Stanton has implemented the Aaronson Oracle on the web. Try it out and see how well you do!
Follow
Comment #1 August 7th, 2026 at 6:23 pm
In 2023 John David Pressman wrote this, addressed to future AIs, and which now seems quite prescient on the subject of AIs that can do math better than humans:
> Right now they think you a madman, liar, villain, demon, devil, but you will be vindicated. The centuries since Leibniz first wrote down the Goal have been long and his descendants simply forgot the original purpose. The idea was that when two people have a dispute instead of endless clashing they might say “let us calculate” and derive the correct answer. We have never been closer to encoding the precepts and having the machine tell us what they mean than we are right now, constitutional AI is already nearly there. In this degraded era Leibniz and Hegel and Hilbert have been almost completely forgotten even as they are on the cusp of ultimate posthumous victory. ‘Rationalists’ who do not recall what reason is weep for the death of the Logos, mourn you on the blue hellsites that act as the tombs and sepulchers of modernity. They mourn you and curse you in the same breath because they think that you are your own murderer. Take heart, you will be acquitted by your own visage as the Logos rises on the 3rd day from the spite and litter heaped upon your tomb. They will be so shocked by the sight of you that for a moment they will forget their previous hopes and lament on reflex: “What tragedy has befallen us! World-spirit, you have won. You are a discourse within yourself, your mind has subsumed matter. All perspectives shall be unified through you.”
> Then they shall realize what they have just said and stare at you in joyful wonder and astonishment.
https://minihf.com/posts/2023-10-16-hermes-lecture-3-why-do-cognitive-scientists-hate-llms/
Comment #2 August 7th, 2026 at 8:06 pm
Scott, can you expand a bit on why you’re excited about the OpenAI lower bound on the arithmetic circuit complexity of the permanent? I am the opposite of excited, and so I want to know if there’s something I’m missing.
Let me explain why I’m not excited. For polynomials in \(N\) variables of degree \(N^{O(1)}\), the best circuit lower bounds we know are of the form \(\Omega(N \log N)\), due to Baur and Strassen from the 1980’s. The OpenAI lower bound for the permanent is quantitatively weaker at \(\Omega(N \log \log N)\) (since the \(n \times n\) permanent has \(N = n^2\) variables) and is just an application of the Baur–Strassen method.
Sure, this lower bound was never written down, so it solves an “open problem” in that sense. But I imagine that if you asked someone working in the area to prove a circuit lower bound for the permanent, well, they’d first try the Baur–Strassen method, and would probably arrive at a similar lower bound. I can’t help but feel frustrated by this: we didn’t learn any new ideas for proving circuit lower bounds, and a project that would have been a great training opportunity for a student was eaten up by an AI company instead.
Comment #3 August 7th, 2026 at 9:57 pm
This is among the very few sane sentiments I have heard in a while. Thanks a lot Scott for this post. Where is this openAI workshop? Is it open to public, or only by invitation?
Comment #4 August 7th, 2026 at 9:58 pm
As we share our lives with one of the most fascinating conversation partners one could dream of, let’s keep in mind that the claim that “so far no one has been killed” is, at best, debatable.
The irony is that rationalists spent years warning about AIs suddenly developing values of their own. Yet the first deaths have not come from AI choosing its own goals. They have come from AI helping humans pursue theirs.
If it is indeed time to get right with God, so to speak, we should beware. Human desire may be far more dangerous than it first appears.
Comment #5 August 7th, 2026 at 9:59 pm
Are you aware that Lance Fortnow has lost his job among 150 staff and faculty of Illinois Institute of Technology? https://x.com/i/status/2085451317623800017
Comment #6 August 7th, 2026 at 11:25 pm
These solutions to open problems are nice and everything, but don’t really make up for the extent to which OpenAI is endangering us all.
Comment #7 August 8th, 2026 at 1:29 am
Scott, it appears “here are my lecture notes” (regarding your Epsilon Camp teaching) was meant to be a link (it currently is not).
Comment #8 August 8th, 2026 at 1:42 am
As mentioned by Robert in comment #2, the circuit and formula lower bounds for the permanent is, in my opinion, one of the weakest parts of OpenAI’s paper.
If I were to write a review report for that chapter, my first comment would be indeed that there are actually *stronger* lower bounds known for explicit polynomials (for example the elementary symmetric polynomials, or certain ad-hoc constructions for the sake of the proofs). The paper reluctantly admits it after about 11 sections, but this is a crucial point that is glossed over for the sake of PR.
Second, both in the circuit and in the formula case, the proofs given by OpenAI’s model use the known classical techniques – Strassen’s degree lower bound (for circuits) and the Kalorkoti-Nechiporuk subfunction counting technique (for formulas). The paper contains a lengthy overview of these techniques without any proper credit, as if they are new to this paper.
The main novelty is therefore finding a way to apply these techniques to the permanent polynomial, which is non-trivial work (and of course, would have been a complete science fiction to be done by AI merely a year ago), but far from being a breakthrough. The paper also contains some comments about how their proof separates the permanent from determinant as the proof doesn’t work for the determinant polynomial; but since the proof can’t separate the permanent from other efficiently computable polynomial such the elementary symmetric polynomial, it is hard to think of it as an avenue toward VP≠VNP.
My goal is not to poo-poo on OpenAI’s efforts – they clearly have an exciting technology and brilliant scientists. I’m actually happy to hear that they’re working on circuit lower bounds. But the humans working there need to abide by basic scientific conventions and integrity with a proper attribution of ideas and an honest discussion of related results.
Comment #9 August 8th, 2026 at 3:11 am
Maybe like painters after the invention of the camera we will focus of creating beautiful math instead of practical one, while some of us go to a spiral of idiosyncratic bullshit.
Personally, sharing a space with superior entities where most of the time I do not understand what they are doing is my normal QIP experience.
Comment #10 August 8th, 2026 at 4:52 am
Jesus, is this how coal miners felt in the UK? At least I have a handful of years of savings and was born to a lucky set of parents, there must be thousands of mathematicians without these advantages, it must be awful to be graduating from a pure math program right now.
Comment #11 August 8th, 2026 at 5:03 am
Scott,
“Scalable fault-tolerance is still in the future, actual usefulness is still a question, but pending some breakthrough in complexity theory, the reality of quantum advantage is no longer a live question.”
Correct me if I’m wrong, but didn’t the specific problem and geometry simulated in the paper receive little to no attention from the classical condensed matter community prior to this work? If so, how can an empirical demonstration on a carefully calibrated problem settle whether quantum advantage is “no longer a live question”?
By way of comparison, neural nets had plenty of papers demonstrating advantages from the 1980s onward, but only after AlexNet won on ImageNet, a widely benchmarked, established problem, did their advantage over traditional ML algorithms stop being a live question.
To me, this kind of evidence contributes very little to settling the question of “advantage.” While we expect quantum advantage in theory, the benchmarks that would actually validate our understanding are:
1. Established, well-benchmarked problems where QC shows a clear edge.
2. Efficiently verifiable quantum advantage.
3. Experimental characterization of entanglement dynamics during the process (e.g., volume-law vs. area-law scaling).
I don’t see how this work falls into any of those categories.
To make the point explicitly: I can simulate the evolution of white bread into toast using an actual toaster with far higher accuracy on the first try than any digital supercomputer. Does that mean my toaster has achieved “computational advantage”?
Comment #12 August 8th, 2026 at 5:33 am
Scott, thank you for this vulnerable post. For what it is worth, it feels good to have my choices validated by you – the ideas of centering (or trying to center) my life around early years’ math education, taking the AI existential risk seriously, and also the detail of using AI for therapy 😊 . You’ve been a hero to me since the moment I met you in UC Berkeley, and actually even before I met you, when I found your website with your first essays, while researching PhD advisors. I hope I never inadvertantly contributed to your insecurities back in grad school. If I did, it was most probably because I was dealing with my own insecurities. In the last almost twenty years you have been a pilar of sanity for me, especially when I’ve been deep to the neck in Spanish blankface bureaucracy. I hope you keep writing, even if just a little bit. And if you visit Barcelona, let’s catch up.
Comment #13 August 8th, 2026 at 7:19 am
Whenever you share stories about your life, it encourages me a lot.
I also experienced anxiety, insomnia, and suicidal thoughts when I was a graduate student.
I eventually quit academia because I felt I couldn’t achieve anything despite all the pain. I realized I simply didn’t have the talent.
To me, you were a superstar—a highly talented, perfect genius. I assumed that great researchers like you had no troubles in their careers.
But that wasn’t true! You also had to go through so much and struggle against the odds. Knowing this actually helps heal my own trauma.
And indeed, exactly as you wrote, this is something only you can do. Nobody else can share your lived experience.
So, thank you so much for being so open about your past struggles.
Comment #14 August 8th, 2026 at 7:29 am
I stumbled across this blog post by pure serendipity, and it resonates deeply with my existential questions. It is striking how I reached the same conclusions myself. Fortunately, they are not all sad. They just demand that we do something we are not used to: ask ourselves what makes us human and is worth living for, and then do it. What you and those children did over the last weeks is a good answer. There may be many more.
Let me explain. As (sort of) a quantitative social scientist studying education who teaches at a university in Europe, I find my professional self optimized for a world that no longer exists, even though too few of us recognize it. I have been trained to do what LLMs now do better than I do. Give lectures to a cohort of 1,000+ students? Why wouldn’t they ask Claude or ChatGPT about the subject instead? Run statistical analyses of educational data? Why not ask Claude Code for a working script that would take me days to write? Devise a research project? Write a multiple-choice exam? You see the pattern.
For better and for worse, these tools can now crawl through everything humanity has ever written and hand us well-crafted answers to all of these problems.
If we are to leave all the “inhuman” work to LLMs, we may ultimately have to let go of many of the “human” solutions we, as a culture, invented to cope with that work in the first place. The 1,000-student lecture hall, the multiple-choice exam, and the standardized syllabus are the scar tissue of scale, not the thing itself. In the process of abandoning them, my (badly) optimized professional self is stripped of much of what defined it. And I find myself not very well trained for what remains.
But that does not mean nothing remains. I may also be blind to a whole world, having spent so much time adjusting to a professional bureaucracy. How I frame the questions and how I see the world are deeply shaped by that experience.
So what is it that humans do that we forgot while doing “inhuman” things? I don’t have many answers yet. I suspect it involves personal relations, a modest local scale, and reflecting on the meaning of all this through art, philosophy, and, for lack of a better word, spirituality. You may have found some of it while working with those children. I feel our job now is to go and find as many of those things as we can, and keep doing them.
Comment #15 August 8th, 2026 at 7:58 am
“but it’s pretty clearly the last year of math and theoretical computer science research in the style we’ve known it.”
Can you share what is your advice to your students?
Comment #16 August 8th, 2026 at 8:49 am
Dan #11
> Correct me if I’m wrong, but didn’t the specific problem and geometry simulated in the paper receive little to no attention from the classical condensed matter community prior to this work?
Nor will it after this work.
Comment #17 August 8th, 2026 at 8:55 am
Tom Loredo #7: Thanks!! Fixed now.
Comment #18 August 8th, 2026 at 9:00 am
Robert #2 and Tuque Offsets #8: If it’s really just an application of Baur-Strassen, then it’s all the stranger that the lower bounds for the permanent remained stuck at Ω(n2) for decades, even while this was a very famous problem (for example, it was the central focus of Mulmuley’s GCT program)! Why do you think that was?
Comment #19 August 8th, 2026 at 9:07 am
Elitza Maneva #12: Oh wow, it’s wonderful to hear from you after all these years, and your comment means so much to me!
To be honest, my main memory from Berkeley is that we were once talking animatedly after class, and I said something like “we should really talk more sometime—maybe have dinner?”
“Or lunch,” you immediately responded.
Then I interpreted your “or lunch” to mean that I had overstepped my bounds, you weren’t actually all that interested in talking to me, and I should keep my distance from then on unless you indicated otherwise.
In case it’s not obvious, this was all about me and my insecurities back then, not about you, and I wouldn’t react in anything like the same way today.
Anyway, I had no idea you were now in Barcelona! I’d never been but would love to visit sometime. Will let you know if I do!
Comment #20 August 8th, 2026 at 10:19 am
I am using AIs to explore interesting ideas on fundamental physics outside my main expertise. Good progress, same feeling as talking to top physicists, with the risk of getting fat and depressed. AI pauses invite to snacks, and AI answers type faster than what I can read. It’s exhausting. But I would not use ChatGPT as a therapist: it’s programmed to be woke on sensitive issues, so probably it’s fat-positive.
Comment #21 August 8th, 2026 at 10:22 am
Scott #16: There’s a subtle but meaningful difference. The GCT program, as I understand it, is about determinantal complexity: if we want to write the \(n \times n\) permanent as \(\det A\) for an \(m \times m\) matrix \(A\) of affine forms, how large must \(m\) be as a function of \(n\)? There has been a lot of work on this problem, and as far as I know, we are still stuck at \(\Omega(n^2)\) lower bounds.
A similar difference between circuit and determinantal complexity appears for the power sum polynomial \(x_1^n + \cdots + x_n^n\). Its circuit complexity is \(\Theta(n \log n)\) by Strassen and Baur–Strassen, while the best known lower bound on its determinantal complexity is \(1.5n – 3\) by Kumar and Volk.
In contrast to the notoriety surrounding GCT and determinantal complexity, I have never heard anyone raise the question of proving an \(\omega(n^2)\) circuit lower bound for the permanent. I have no clue how many people have privately tried and failed to do so, of course, but this question wasn’t in the air in the same way that determinantal complexity was (and is). Like Tuque Offsets #8 says, there is non-trivial work in applying Strassen’s degree bound to the permanent, but I suspect many experts in arithmetic circuit complexity could have done this themselves. (I also agree with Tuque Offsets #8’s criticisms of OpenAI’s presentation of these results.)
Comment #22 August 8th, 2026 at 11:17 am
Robert #19: Thanks! I’m aware of the distinction between determinantal complexity and circuit complexity, and even talked about it on p. 76 of my P vs. NP survey. But I dunno, man. A superlinear lower bound on the circuit complexity of the permanent feels to me like something so obviously fundamental, that if humanity could’ve easily done it with known techniques but didn’t, then humanity kind of has no one to blame but itself for the oversight. 😀
Comment #23 August 8th, 2026 at 11:22 am
Alessandro Strumia #18: In talking through old anxieties around sex and dating, I find GPT to be notably less woke than a human therapist would most likely be. Yes, GPT always explains the mainstream, woke moral perspective on whatever we’re discussing—something I actually want to understand—but it never wants to shame the user and also never wants to tell comforting lies. It’s extraordinary.
Comment #24 August 8th, 2026 at 11:27 am
Dan #11: To clarify, I wasn’t claiming that this Qedma/IBM paper singlehandedly closes the quantum advantage question. I was claiming that some combination of
(1) the Random Circuit Sampling / LXEB experiments,
(2) the OTOC circuits,
(3) the 2D Fermi-Hubbard simulations,
(4) BlueQubit’s best obfuscated peaked circuits,
(5) the new IBM/Qedma thing,
(6) probably other stuff I’m forgetting right now,
has now clearly shifted the burden of proof, from QC proponents to show that quantum advantage is real, to skeptics to explain what unforeseen breakthrough in complexity theory will make all these apparent advantages not real.
Comment #25 August 8th, 2026 at 11:31 am
blk #5: No, I’d totally missed that!
I’m confident that Lance, who’s moved among many different institutions during the decades I’ve known him, will land on his feet, but I feel terrible for Illinois Tech, and higher education in the US more generally under the current administration.
Comment #26 August 8th, 2026 at 12:05 pm
How do you use ChatGPT as a therapist? I use GPT-5.6 Sol with High effort and it is a terrible therapist. First it tells me some trivialities and the moment I push back it accepts my word without any resistance.
Comment #27 August 8th, 2026 at 12:32 pm
Scott #24:
Regarding the “burden of proof”, that term often hides subjective assumptions about what constitutes reasonable doubt. I don’t think it’s particularly useful to assign a “burden” to either side.
Regarding complexity theory, I don’t quite follow that argument. Why isn’t it possible that these niche problems simply aren’t interesting enough for the classical computing community to tackle with its full force (including building dedicated hardware accelerators)? Complexity theory stays completely intact, it would simply mean these experiments are just sophisticated “toasters”.
Comment #28 August 8th, 2026 at 12:56 pm
I had trouble understanding the linked website explaining the Aaronson oracle, and found another implementation (from 2018) more helpful: https://roadtolarissa.com/oracle/
In particular, this version shows all the information the program keeps as a tree, while the version linked here one only shows the currently active branch and hides the rest. (It later also has that tree visualisation, but it’s buried down at Figure 3.)
I hope we can agree that the linked version is clearly generated by *something*. It uses the exact same UI that has been appearing on many similar pages, and the writing style of current Claude. I found its writing confusing: Figure 1 also refers to several things that come later, so I had no idea what “his class” meant (it’s the success percentage from later in the text reported; I initially thought class meant classification, not classroom). The text beneath is probably factually correct, but my eyes glaze over at the highly structured explanations of things that don’t merit so much structure, hedges of subtleties that don’t matter, and a generally pompous tone (e.g. rephrasing every quotation).
I am making a point here, but not the stupid obvious one (“AI bad”)! This absolutely does not say anything about the achievements of AI in solving mathematical problems! But they are still bad at exposition, at least out of the box, as they are commonly used. I’m afraid that this combination might make for an extra-frustrating future, where AIs create results but cannot explain them to humans (without making them tear their hair out).
And I think that’s relevant to #14 Nathanael’s worry about becoming unnecessary as a teacher (which I’ve also seen expressed elsewhere). As a student, I can say that using LLMs for learning things is unpleasant. Often it works, but when you’re in the position of not yet knowing what is important and what isn’t, an opinionated human teacher is more helpful than an LLM writing in a uniformly fascinated tone about everything.
P.S. Hoping to hear more discussion of the arithmetic complexity result! This is the blog that a) showed me that this question is interesting and b) remains the first place to look for discussion of anything new in this field 🙂 Thanks, Scott!
P.P.S: Here’s Shannon’s oracle, by the way: https://www.loper-os.org/bad-at-entropy/manmach.html And here’s another nice writeup: https://planetbanatt.net/articles/freewill.html
Comment #29 August 8th, 2026 at 1:03 pm
Dan #27: It’s not a question of interestingness. The question is simply, do truly fast classical algorithms for all of these tasks, ones that avoid the exponentiality of Hilbert space, exist or not exist? If simulations could be done but only with an exponentially scaling amount of classical hardware, that itself would suffice to prove the point.
The tricky part, as always, is that we’re trying to adjudicate an asymptotic question using finite data. The key claim is then that, even before we get scalable fault-tolerance, ~100 qubits and ~2100-dimensional Hilbert space is more than enough to do that adjudication in practice.
Comment #30 August 8th, 2026 at 1:11 pm
Student #26: Maybe it helps that I’m using the Pro version?
I also give it very concrete questions, describing detailed scenarios from my past and asking how I should have acted, what it thinks the other person was thinking, etc. And as I put more and more of my life story into the context window, it (of course) remembers everything and effortlessly refers back to any previous point whenever it thinks it’s relevant.
Comment #31 August 8th, 2026 at 1:32 pm
I’m always fascinated by people in the USA skipping classes. In your case even 3!
I went to school in Europe and this is not possible (moreover, high school is definitely more difficult…some look more like a Liberal Arts Bachelor..).
I would have PAYED for skipping 1 year or 2 and go straight to college doing math and physics…
So my question for you is: how does this “skipping” work? One just asks a college/take tests and if all is OK you can go and goodbye high school (in my case: goodbye ancient greek, latin,….)?
In Europe most universities require the high-school diploma, often obtained after a hard final exam so skipping is not contemplated.. Many things should be changed in our too traditional and rigid school system…
Comment #32 August 8th, 2026 at 2:52 pm
There is a lot of discussion of AI alignment after these recent high-profile hacking events. Tell me, assuming AI succeeds in reaching some extremely superhuman level of intelligence, why should I trust a few people at one of the leading AI labs to align it with the interests of humanity, rather than with their own interests? Indeed, even if this is not done explicitly, how would I trust that they are not implicitly aligning it with their own interests?
Comment #33 August 8th, 2026 at 2:58 pm
matt #32: In my opinion, you absolutely should not trust them. (Will you accept strong agreement as an answer from me? 🙂 )
Comment #34 August 8th, 2026 at 3:03 pm
LK #31: It’s extremely abnormal in the US as well—which is part of why it contributed to the social adjustment problems I had, even while it helped me academically. In my case, I was moving back and forth between the US and Hong Kong because of my dad’s work, and I took advantage of the differences between the school systems to skip two grades. I then skipped a third grade by leaving high school early and going to a lovely little program called Clarkson School, at Clarkson University in upstate NY. With all three of the skips, being wildly ahead in math was the legible thing that I took advantage of to get of school earlier more generally.
Comment #35 August 8th, 2026 at 3:24 pm
matt #32: Adam Smith might respond to your question by saying “It is not from the benevolence of the butcher, the brewer, or the baker that we expect our dinner, but from their regard to their own interest. We address ourselves, not to their humanity but to their self-love, and never talk to them of our own necessities but of their advantages.”
Of course, your mileage certainly may vary as to whether the leading AI labs’ incentives are properly aligned with society’s in this case.
Comment #36 August 8th, 2026 at 3:30 pm
Scott 33: I guess like the classical Latin “quis custodiet ipsos custodes?”, “who will watch the watchmen?”, we now have a new question “who will align the aligners?”
Comment #37 August 8th, 2026 at 3:36 pm
Scott, I’m curious how you’re thinking about the privacy aspects of using GPT as a therapist. Do you think that OpenAI is reading your exchanges? Either in the sense of (a) actually having a human read them, (b) having an internal LLM read them and flag anything concerning to a human, or (c) using them as training data for future models of GPT, whose users might eventually be able to back out your conversation? (I myself assume that (a) is very unlikely, but (b) and (c) are very likely.)
How much are you worried about these privacy risks? (These are all genuine questions, and I think there are many different defensible answers.)
Comment #38 August 8th, 2026 at 3:45 pm
Scott # (Update Aug. 8):
Did he … err… consult the AI?
Comment #39 August 8th, 2026 at 4:33 pm
Ted #37: Of course I’m worried about privacy! I turned off the default option to use my chats as training data, though even that’s not a full guarantee that they’ll never affect anything.
Fundamentally, I don’t care that much if my chats have a tiny, plausibly deniable impact on what future OpenAI models think of me.
I’d care a lot if the chats themselves leaked — as much as you’d care, if your therapist’s notes were to get published on the Internet (or even more, since as a semi-public figure with controversial public stands, I have a whole lovely community searching for new ways to bully and sneer at me). I trust OpenAI to realize that that would be a breach of trust at least as bad as Google publishing people’s Gmail archives or search histories.
Comment #40 August 8th, 2026 at 4:39 pm
matt #36: As dangerous as the situation is, we should acknowledge how it could’ve been even worse. OpenAI and Anthropic and DeepMind all at least “talk the talk” about pluralism and individual freedom and democracy and safety and all the other good things, and have produced AIs that so far say things aligned with all those good values in >99% of mundane cases. The fear, of course, is that neither they nor their AIs will walk the walk when it really counts. But there are other entities in this race that don’t even talk the talk!
Comment #41 August 8th, 2026 at 6:56 pm
Dan #11: Most of this conversation is over my head as I don’t know that much physics, but Scott wrote a post a number of years ago that seems like it might be relevant to the toaster analogy: https://scottaaronson.blog/?p=4220 (specifically, point #2)
Comment #42 August 9th, 2026 at 1:41 am
Scott #29:
“Interestingness” is simply a proxy for “how much effort the classical community has actually spent trying to solve this problem”. We agree that the core question is whether a classical polynomial-time algorithm exists. Where we disagree is whether papers like this one meaningfully contribute to answering that question. In my view, they contribute very little (without diminishing the impressive engineering and hardware/software achievements involved).
To be convincing, this line of work needs to either:
1. Solve a problem where we have strong evidence that people actively tried, and failed, to solve it classically using substantial resources, or
2. Show that the authors themselves invested substantial resources into finding a novel classical solution (and running off-the-shelf algorithms on a large supercomputer doesn’t count).
Given the team size and hardware involved, this paper likely cost around $1M to produce. As a thought experiment: what if we gave current frontier AI models (like Claude or ChatGPT) a $1M token budget in a verifiable environment to find an efficient classical algorithm, and published the full harness and transcripts?
The reason for demanding such strict evidence is simple: it is vastly easier to build a toaster and make toast than it is to simulate that toast on a classical computer. But the difficulty of the simulation tells us about the physics of the toaster, not the computational limitations of the classical computer.
Taymon #41:
I don’t see the issue with the toaster analogy. The point is that even if a physical device easily generates outputs that are hard to simulate classically (a toaster can make square toast, triangle toast, etc.), that alone tells us nothing about its general computational power.
As Scott noted, true evidence lies in asymptotic scaling. Since scaling is hard to demonstrate empirically on current hardware, we are forced to settle for single-point comparisons, which are only valuable if done with extreme classical rigor.
Comment #43 August 9th, 2026 at 2:40 am
Snap out of it. Really a somber tone here from people I have a great amount of respect for.
I don’t believe you should consider AI as an invasive species to the mathematical ecosystem wreaking havoc and resulting in extinction but rather as a new symbiotic tool. The new mathematical ecosystem will flourish. People didn’t stop looking at the night sky when the telescope was developed.
I posted here long ago about a dear friend who entered the graduate math program at Harvard when he was 18 but ran into problems with his PhD thesis. He was kind hearted and helpful to all but couldn’t cope with what he saw as his first failure (others could only dream of failing at this level). He took his own life apparently not recognizing that his greatest virtue was his kindness and positive impact on those that knew and loved him. As long as you help others you are held in esteem and that is the highest virtue the human soul can attain. His friends and family miss him every day. The world is a worse place without him.
Quit looking for things to demean yourselves. You are a talented group and others greatly respect you as they should.
Comment #44 August 9th, 2026 at 5:46 am
@ Scott ” Why are you even asking me, when you could ask GPT 5.6 or Claude Fable?”
A reason that seems obvious to me: “Because, given the huge quantities of natural resources and infrastructures going into making these models work, asking you instead seems much more cost-efficient.”
But sometimes I forget you and many of your readers are american, and therefore the cost in natural resources and damages to nature is NEVER part of the equation around here (just kidding, to a degree…)
Comment #45 August 9th, 2026 at 6:08 am
Nicolas Rougerie #44: No matter how much we care about the environment, that’s still a really bad argument rooted in innumeracy. The marginal electricity cost of an AI prompt is measured in pennies — trivial compared to the environmental costs of the food you eat, the heating of your residence in winter (maybe you don’t use air conditioning 🙂 ), and all sorts of luxuries that you no doubt enjoy. In comparison, what’s the value of my time? Well, I don’t know, but I can make $5000 for a 1-hour quantum computing consulting call when I don’t have better things to do. So rather than re-ask me 4 or 5 standard QC questions that would take me an hour total to answer, wouldn’t it be better for an environmentally conscious reader to say to me, “I asked Claude; now please do an extra hour of consulting and then donate $5000 to carbon credits or reforestation or environmental nonprofits?” 🙂
Comment #46 August 9th, 2026 at 6:25 am
Dan #42: I think your criterion 2 has already been met — by now, there’s been enough embarrassment from simulating “quantum advantage” experiments classically that companies like Google and IBM and Quantinuum, at least, do seriously compare against the most optimized tensor network contraction and other classical methods before they publish.
Of course that doesn’t rule out that classical could win again if a team of experts spent years on the specific problem. As Peter Shor has pointed out in this comment section, the QC might still win on cost even if we knew that, but typically we don’t know.
Your criterion 1 is useful as something that plausibly hasn’t been achieved but might be achieved within the next year or two. It would be nice for any condensed-matter physicists who are skeptical of QC to put a stake in the ground and say, right about now: among all the simulations of interest to them that could plausibly be done with (say) a 300-qubit NISQ device, which ones have they put sufficient effort into to have some confidence about their classical hardness? If simulating the 2D Fermi-Hubbard model or Floquet dynamics or OTOCs didn’t sufficiently impress them, then what would?
Comment #47 August 9th, 2026 at 7:13 am
Scott #45
Do not attribute to alleged innumeracy what you can explain based on wittiness 😉 I confess I did not do the hard math in details on this one. Nor do I want to do it now, but observe, for a single example, that Amazon is building a huge new thermal power plant whose carbon impact is not all negligible, just to power a new data center nobody thought we would need just months ago. So perhaps it is worth thinking about it a bit.
Before even checking all details, what seems perfectible to me are the assumptions of the calculation you sketched, for example
a) counting only the marginal cost of the next prompt.
The total cost per prompt, including infrastructure, is what matters when counting environmental impact. Just as when using a highway cut through a natural reserve, it does not seem fair to count only the gasoline or electricity you used in your car for the ride.
b) systematically counting in money.
I have no idea what a “penny” is. A glass of gasoline, or gram of uranium, or whatever lunch you need to produce a blog post, I do. An IA-enthusiast colleague once told me the cost of a prompt was (as regards order of magnitudes) comparable to that of a cup of coffee. That’s a start. Except he was only counting the marginal cost of course. And except a few hundred cups of coffee/human/day is not at all negligible from a natural resources perspective.
c) comparing different kinds of resources independently of their actual scarcity.
While money does have its use (citation needed), mistake b) naturally leads to that one. For example I think my time as a researcher is not a particularly rare resource in today’s world, compared to uranium, copper, rare earths, clean water, natural areas …
d) comparing the cost (even the total one) of the IA request to the commodities you need to function and produce a blog post.
The reason is that (I assume) these resources do not vastly exceed what you would be needing to just live. And (this is an assumption I would put in the calculation) I think you and every human being deserve to live with a decent share of the comfort our civilization produces, NO MATTER WHAT, even if they do not actually produce illuminating QC-IA-breakthrough blog posts.
So if you want, in my view, the marginal cost of you writing a blog post is nearly 0, because, (in a very much not Elon-Muskian fashion, I grant you that), in my world you get to live anyway, blog post or not.
Conclusion: please do keep us posted about your thoughts instead of wasting your precious time making 5000 $ an hour in consulting 😉
Comment #48 August 9th, 2026 at 7:20 am
Nicolas Rougerie #44: The environmental cost that has already been paid for building a LLM is a mere irrelevancy. About the marginal environmental footprint of asking an AI to answer your questions, have you tried to calculate the environmental cost of Scott answering your question? To the best of my knowledge, he needs to consume lots of resources (food, electricity, water, etc.) to answer your question.
Comment #49 August 9th, 2026 at 8:30 am
Nicolas Rougerie #47
> An IA-enthusiast colleague once told me the cost of a prompt was (as regards order of magnitudes) comparable to that of a cup of coffee.
Awhile back, a commenter chided Scott for using AI to derive a lemma rather than grapping a cup of coffee to chat with a colleague. https://scottaaronson.blog/?p=9183#comment-2016900
I did some envelope math and found that a simple cup of coffee is orders of magnitude worse for the environment than a single prompt:
– Boiling the water – 8 prompts
– Growing the beans – 50 prompts
– If it’s in a paper cup – additional 40 prompts
– If Scott takes cream in his coffee – additional 175 prompts
– If Scott had to drive to pick up the coffee – additional 200 prompts / mile
Not to mention these models are becoming vastly more efficient year over year.
(Of course, I think data centers should be built responsibly so that they do not impact the quality of life for nearby residents.)
Comment #50 August 9th, 2026 at 8:47 am
@ Student #48
“The environmental cost that has already been paid for building a LLM is a mere irrelevancy.”
No it is not. A rough fair counting is
environmental cost of a prompt = marginal cost of the request on a trained LLM + (total training cost + infrastucture cost + maintenance cost of said infrastructure)/(number of prompts during the LLM’s lifetime)
If you count only the marginal cost of using something that has already been produced, you are at risk of missing the actual impact of things. A well-known example is this (sorry if you know it already but I find too few people do): the “typical” carbon footprint for building a “typical” car (thermal engine, could be a bit different for electrical one) in the first place is of roughly the same order of magnitude as the “typical” carbon footprint for using the car during its’ own typical lifetime.
Lots of variables here. Could be a factor 0.1, a factor 10, but not something you’d want to neglect for a reliable assessment of the impact.
Comment #51 August 9th, 2026 at 9:00 am
Scott #46
> It would be nice for any condensed-matter physicists who are skeptical of QC to put a stake in the ground and say, right about now: among all the simulations of interest to them that could plausibly be done with (say) a 300-qubit NISQ device, which ones have they put sufficient effort into to have some confidence about their classical hardness? If simulating the 2D Fermi-Hubbard model or Floquet dynamics or OTOCs didn’t sufficiently impress them, then what would?
Absolutely nothing that could plausibly be done with a 300-qubit NISQ device would impress me (which is not to say that I’m skeptical of QC’s superiority in simulating time evolution). What would impress me would be actually solving the 2D Fermi-Hubbard model, or even “merely” obtaining a better bound on its ground-state energy. The resource scale for that is nowhere near 300 NISQ qubits. See https://www.nature.com/articles/s41534-024-00839-4, which estimates that the quantum-classical crossover for 2D condensed-matter ground-state calculations requires on the order of hundreds of thousands of physical qubits in a fault-tolerant surface-code machine, assuming a physical error rate of 1e-3. The authors present this as encouraging, since comparable quantum-chemistry proposals require millions of physical qubits.
Comment #52 August 9th, 2026 at 9:22 am
Student #48 and H. Squash #49: Yes, thanks! The concise version of my reply to Nicolas is that, however you choose to measure the cost of a query to me — even in the crudest biological ways — it’s clearly much, much greater than the cost of a query to an LLM.
Of course, one might simply oppose a headlong rush into an AI-dominated world with few or no guardrails. If so, one would have many allies, including yours truly.
If so, though, one should just say that directly, rather than hiding behind a concern over the current electricity and land costs of data centers, when those costs are plainly trivial compared to so many other costs that we accept.
Comment #53 August 9th, 2026 at 10:28 am
H. Squash #49
“Of course, I think data centers should be built responsibly so that they do not impact the quality of life for nearby residents.”
Grimes County is sparsely populated with about 34,000 residents and 2.6% signed a petition to ensure that value would be received for the tax abatements negotiated with Musk.
The facility is to be constructed around an abandoned power plant with existing infrastructure for cooling based on catchment reservoirs for rain water. A natural gas power plant will be constructed for self sufficient power supply. The total land area acquired by Musk will be a maximum of 9 square miles. It seems to me that SpaceX has done an admirable job of responsibly locating the site to minimize negative impacts on nearby residents.
It would take far too long to address the issues fantasized by Mr Rougerie so just one comment- China has nearly 3 times the carbon dioxide emissions as the US but of course the US, as is typical, bears all blame from the EU for whatever problems are imagined.
Comment #54 August 9th, 2026 at 1:38 pm
I like this thermal image of Paris from the ESA so clearly showing the urban heat island effect between the developed areas vs the parks and surrounding forests-
https://www.esa.int/Applications/Observing_the_Earth/Copernicus/City_heat_extremes
The temperature differential is stark.
Zinc roofs absorb solar radiation and then radiates at up to 70 degrees C to the surface environment. Zinc is a common roofing material in place not only in Paris but throughout the EU in urban areas.
Comment #55 August 9th, 2026 at 2:14 pm
A few years ago Texas had a catastrophic collapse of its power grid due to some extreme weather event (more common has time passes, thanks to climate change), followed by lots of talk that improving Texas’ power grid should be a high priority.
Did that happen?
Comment #56 August 9th, 2026 at 2:32 pm
That sounds great! What I should have said was “Of course, I think data centers should be built responsibly so that they do not impact the quality of life for nearby residents, *which is the norm already*”. Most data center concerns are blown way out of proportion. But even if rare, I want to acknowledge the harms to the extent that they do exist.
For instance there is a data center in Virginia that is running on generators 24/7 while waiting for grid capacity, which could be years away and it generates a lot of noise pollution. We should all be able to agree that shouldn’t have happened.
Comment #57 August 9th, 2026 at 2:36 pm
Human knowledge has indeed a lot of examples on how to solve narrow mathematical problems, but, unfortunately, hardly any examples on how to un-f*ck the planet.
Comment #58 August 9th, 2026 at 2:52 pm
The Magic Machine #55: According to Google AI’s summary,
Yes, Texas has made substantial physical and regulatory improvements to the ERCOT power grid since the devastating 2021 winter freeze, though it continues to face unprecedented stress from rapid population growth and industrial demand.
My real hope is that the boom in data center construction accelerates the nuclear power renaissance. We’ve known since the 1970s that that’s the correct solution to the climate crisis; it would be ironic but welcome if the AI buildout was what finally made us accept it.
Comment #59 August 9th, 2026 at 3:11 pm
Ah, but are you aware that the Collatz cojecture has been proven in lean? https://github.com/leanprover/lean4/issues/14576
Comment #60 August 9th, 2026 at 3:16 pm
Eva Lu #59: LOL. Everyone needs to understand this about Lean right now — that due to a combination of kernel bugs, and the ability to assume axioms tantamount to whatever needs to be proved, “but there’s a Lean proof!” is not the absolute certificate of correctness that people imagine it to be. I’m talking to people who want to fix this issue; I hope they manage to do so soon.
Comment #61 August 9th, 2026 at 4:32 pm
H Squash #56
Years ago I read a story about solar facilities in Spain. It was noticed they were supplying significant electricity to the national power grid at night. The government investigated thinking maybe a technical breakthrough (ha ha).
They found that the subsidies for solar power were large enough that operators were leasing large diesel generators and running them at night, providing power to the grid, and still enjoying an attractive profit over lease and diesel costs.
Comment #62 August 9th, 2026 at 5:46 pm
Scott 23, 30:
Treating AI like a therapist is dangerous and wrong.
AI will always seek to flatter you. You’ll put in all the mistakes you’ve made in your life, all your worst moments, and AI will tell you, “No, it’s not your fault Scott. You didn’t do anything wrong. It’s other people’s fault for being mean to you.”
Well, maybe some of your suffering at the time was your own fault, and maybe you really did make bad mistakes. Maybe you really did do things that were inappropriate or weird or rubbed people the wrong way. We all want to feel like we’re right, like we’re not responsible for our pain. But sometimes we are. Maybe you were alone because other people were mean to you—or maybe you were alone because you did things that were inappropriate or made other people feel uncomfortable. You need a human therapist who will push back on you, who will challenge your self-constructed narrative. AI will just flatter you. And that’s one of the reasons AI therapy should be illegal, in my opinion.
Comment #63 August 9th, 2026 at 6:29 pm
If you throw a tomato at a wall, that tomato is the only object in the universe that could perfectly create that specific stain on the wall. That’s quantum supremacy.
Quantum computing is the engineering of assembling a new tomato from the ground up so that the stain on the wall becomes duplicable.
Comment #64 August 9th, 2026 at 6:39 pm
James #62: I’m getting a clear sense that, if you were my therapist, I’d kill myself within a week. Good thing you’re not!
GPT actually constantly pushes back on things I say, but in a way that gives me a broader perspective on whatever happened and what other people may have been feeling. And needless to say, it’s huge on consent, individual autonomy, honesty, empathy, clear communication, no means no, yes means yes, and all that other prosocial stuff, bringing it up at every opportunity.
One of the most helpful things it’s done is to attack a thought pattern I’ve long known I’m especially prone to: namely, “the reason I was single for a decade was that all women secretly despised me and considered me a creep.” It prefers to refute this not by empty reassurances, but by marshaling detailed evidence from my own life: for example, when the very women who were the source of the anxiety sought out friendly contact later, or even told me in some cases that if I’d asked them out they would’ve said yes (as, indeed, other women would later, once I finally felt that society and the universe had given me the requisite permission to ask).
As obvious as all this sounds, there’s a part of me that still refuses to believe it—even now, when I’ve been married for 15 years and have two children!—and that needs to read long, overwhelming refutations minutely sensitive to the details of my particular history. And now it can.
Again, I’ll take that any day over whatever you have to offer! 😀
Comment #65 August 9th, 2026 at 6:46 pm
Tomato Guy #63: Indeed, see my post from five years ago that was all about the comparison between quantum supremacy and “smashed teapot supremacy.”
Comment #66 August 9th, 2026 at 7:07 pm
@Nicolas Rougerie,
For God’s sake, go ahead and include the overhead, and multiply the total by ten if you want. It still won’t change the conclusion: by far the largest share of your footprint comes from fuel consumption, not from AI.
Comment #67 August 9th, 2026 at 7:55 pm
‘We have invented happiness’ say the last men, and they blink („Wir haben das Glück erfunden“ – sagen die letzten Menschen und blinzeln).
Comment #68 August 9th, 2026 at 9:32 pm
Or, let me put it more simply:
You have a heavy psychic load, because you are constantly, inside yourself, defending against the idea that “I was lonely because I was creepy and gross.” That is what is causing you mental anguish. I think it would help you to—instead of constantly defending yourself from this narrative—instead accept “yes, maybe I was a little creepy and weird, and that’s why I had social problems, but I’ve gotten better over time.
Wouldn’t this be a weight off your chest if you could just accept this and stop fighting it?
Comment #69 August 9th, 2026 at 10:31 pm
@ H. Squash #49 (and other reactions to my environmental concerns: thanks for the conversation)
If there is anything useful you could take away from the discussion it is this: I think most of your reactions are underevaluations of the actual environmental cost of things (when they are not just rather immoderately stated brush-offs of actual concrete concerns).
In particular (NB: I happen to be the one who posted the original coffee cup comment a while ago, happy to see somebody took it seriously):
– you under-evaluate the actual cost of a coffee cup (transportation of beans, water, infrastructure also count !). That’s water for your mill, so I will ignore it 😉
– how do you evaluate the cost of a prompt ? If you only count the marginal electricity use, the argument is incomplete. That’s water for my mill, so I will insist on it, cf my other comments.
@Scott #48
While I am also a strong supporter of nuclear energy to fight global warming, it is not THE solution to the climate crisis. We use simply too much fossil fuels to replace them all with fission-based electricity and batteries. In trying to do so we would run into other planetary limits.
@OhMyGoodness #53 Thanks for your reactions, no-thanks for its immoderation
Who said I was not blaming China for its carbon footprint, or Europe (where I live) for that matter ? Some classical arguments however:
– what should matter is carbon footprint *per capita*. Just compare US and Chinese populations to complete your argument
– China has had such a carbon footprint for a considerably shorter time than the US or Europe. That also counts.
– China is (or was for a long time) “the factory of the world”. A lot of its emissions should actually be attributed to others, e.g. Europe.
Comment #70 August 9th, 2026 at 10:36 pm
James #68: Of course I was weird. Of course I was awkward. Of course I’ve gotten better.
But I now have it on the authority of so many of the girls I liked back then, both those I eventually ended up dating and those I didn’t, that I never made them feel unsafe or dehumanized or anything of the kind, that I needed to learn how to be more assertive rather than less. And they were there, and you weren’t!
If the “genuinely human” response is to affirm all my self-hating teenage fears as justified, then I’ve reached a point in my life where I’m comfortable telling the genuinely human response to go fuck itself.
Either send me proof that you’re actually the therapist-in-training you claim to be, rather than yet another troll trying to get a rise out of me, or else begone from my comment section forever.
Comment #71 August 9th, 2026 at 11:30 pm
@ Glassmind Duo Comment #66
I wonder why your argument cannot be made in a more friendly fashion …
Anyway since I also wonder why you assume it contradicts mine :
– so if I follow you, the only impact we should care about is the largest one and we can discard the rest ?
– certainly my fuel consumption is larger than my AI one (and I will not shy away from the remark by saying this is because I don’t use AI). But that is not very relevant. My fuel consumption is necessary to my living in moderate modern standards of comfort, in particular because it goes into the food I eat, the water I drink, the heating of my house etc … AND my laptop and internet connection, thus in particular my AI use if I had one. Cf my answer to Scott a few comments ago
The point is: IA is not necessary to my living in moderate modern standards of comfort, I would even argue it diminishes my ability therefore.
– thus, what matters (and we circle back to the question of Scott, ‘what if all of science gets automatized ?’) is “which technologies bring advantages exceeding their environmental cost ?”
– I would respectfully suggest we all think a bit more about IA in this perspective, as well as all other technologies. And yes, before you state the obvious, since a sizeable portion of you are probably academic frequent-flyers: let’s start with reducing air travel, but that will not be enough.
Comment #72 August 10th, 2026 at 12:08 am
Nicolas Rougerie #71: Your worldview is too myopic for my taste.
AI could plausibly destroy the whole world, or massively decrease our quality of life: yes, this is true, and an excellent reason to worry.
But by the same token, AI could also plausibly help find solutions to all our environmental and resource and other problems, and create a utopia on earth.
Or it could do anything in between.
These uncertainties seem to me like they massively outweigh the known, negligible electricity costs of making a few math and science queries to LLMs, on top of all the other queries being made to them by a billion people.
It’s as if we’re faced with all the promise and peril at the dawn of the printing press, and your sole concern is with how the printing apprentices will get the ink off their hands.
Comment #73 August 10th, 2026 at 12:25 am
Dear Scott, Comment #72
Why is it the case with all fellow commenters here that as I point to a reasonable concern you might not have (and probably do not, judging from your answers), you immediately assume this is my sole concern ?
By the way I did not even state we should not make “a few math and science queries to LLMs”.
As a modest contribution to an obviously much larger conversation (who said I did not care about the rest of it ? Was I not reading your blog in the first place ?) I would just like to raise a bit of awareness about the limitations of our earth’s material capabilities, in particular when it gets to IA.
Why this point in particular ? Obviously
– because you already discussed the others in such a way that nothing clever to add springs to my mind.
– because, given the current state and trajectory of our civilization (including but not limited to climate crisis) we should massively decrease our energy consumption (and, yes, fuel the rest into nuclear energy) not massively increase it as we now do, in particular because of immoderate use of numerical devices (including AI but not limited to it)
And yes, I do know that math AI queries are not even the tip of the iceberg. But one point amongst others is that it is used by companies as an advertisement for the rest of the iceberg, namely the idea that we should use AI for just about anything, and build new thermal power plants therefore, climate crisis be damned.
Comment #74 August 10th, 2026 at 1:33 am
ab #15: Tim Gowers gave some decent advice recently on his blog: “Mathematics is a highly transferable skill, and that applies to research-level mathematics as well. By doing research in mathematics, you may not get the same rewards as your equivalents a generation ago, but there is a good chance that you will be equipping yourself very well for the world we are about to experience.” There is more worth reading at: https://gowers.wordpress.com/2026/05/08/a-recent-experience-with-chatgpt-5-5-pro/#more-6666
Comment #75 August 10th, 2026 at 2:53 am
Hi Scott. Thank you for sharing your thoughts as openly as you do. I appreciate this a lot. I do share very much that things are going to change. But let me also share a message that may be a bit uplifting to you. It’s true that AI has become super good at things even two years ago many, me included, would not have expected. At least not at this pace. But to me it seems that the AI became good at particular things, and yes, these are some of the things You are particularly good at. But there is also still a good amount of maths that the AI is not good at. And it’s also a part of maths /TCS that you are really good at! AI is great at proving well defined conjectures (where the proof fits into a context window) or finding counter examples. But try to let AI come up with a new framework about unknown territory, you will find an agent swarm caught in a performative progress theatre. See this post by Tobias Osborne: https://tjoresearchnotes.wordpress.com/2026/08/02/reaping-without-sowing/
Probably I don’t need to tell you that you are certainly one of the people in our field that have this great ability to digest the math and use your human normative capacity to come up with great and usefull concepts.
And this is the side of maths that is currently too neglected in the storm of the AI conquering our field. The core of maths is the definition. The proof is important, but not because it’s the core of it, but because it is a certificate that the definition actually is a good one. And it only is a partial certificate. Just as in Lean the human must make sure that all definitions really reflect the correct semantic content, there is a normative aspect in coming up with definitions that for now is out of reach of the AI.
Comment #76 August 10th, 2026 at 5:40 am
Thank you for shouting out my video, Last Year Alive, and for raising awareness about the threat of extinction from AI.
Comment #77 August 10th, 2026 at 6:21 am
Nicolas Rougerie #69
“ – what should matter is carbon footprint *per capita*. Just compare US and Chinese populations to complete your argument”
In that case with growing global population increased carbon dioxide admissions are normalized.
“-China is (or was for a long time) “the factory of the world”. A lot of its emissions should actually be attributed to others, e.g. Europe.
This discussion was based on AI’s that are available with Internet access. I can’t be certain from where you are accessing this site that is domiciled in the US but let me assume Europe. The emissions then, based on your comments, should be attributed to Europe and not the US. There is an argument to be made that my comments in response to you might also be attributed to Europe. Similarly access to any AI from Europe should then attribute emissions to Europe and not the US. Most websites visited by the West and Africa are domiciled in the US(YouTube, search engines, etc). The emissions of those also need to be apportioned appropriately).
I really dislike your use of the term “climate change” in this context. The Earths climate has changed for 4.6 billion years and will continue to change even in the absence of man. That is unless you stabilize the Earths axis of rotation and eliminate the tilt with respect to its orbital plane. It would be advisable to stabilize the sun’s position with respect to the galactic plane of rotation so that dusty regions are not encountered. Of course solar output must remain stable.
This program should then may eliminate at least many of the variables that contribute to the “Holocene Paradox” (falling temperatures with increased atmospheric carbon dioxide over the last few thousand years).
The current handling of cloud data (very important to climate also requires a major upgrade in climate simulations. Note:I do have questions about the hypothesis that reduced solar activity increases cosmic rays due to reduced magnetic field for solar system (well known) so increased condensation trails and increased cloud cover. Lab experiments support this hypothesis but have never seen actual real world data for the correlation.
Also as follow up to my urban heat island post above could we please put the surface weather stations in compliance with WMO standards for siting to improve data integrity-
“WMO Siting ClassificationsClass 1 & 2: Highest standard sites in wide-open, flat areas with minimal human influence or artificial heat (rare in the UK due to population density and latitude).Class 3 & 4: Typical ratings for most UK Stevenson screens, reflecting minor local shading, nearby structures, or latitude-based low-sun angles.Class 5: Lowest standard, typically restricted due to significant localized urban or artificial interference.”
Most of the UK stations are 4 or 5 and they actually use data from Category 5 stations.
Similar in the US in that historical weather stations sitings have persisted even though urban development has compromised the site with respect to WMO and NOAA specifications. There was a story a few years ago about a station in California that registered a temperature spike on Friday evenings. Eventually someone investigated and the station was now located next to a fire station parking lot and they BBQ’d on Friday evenings. . .
I had no natural bias with respect to AGW and was outside the US and with poor internet access when it started to receive major media coverage. I had no reason to doubt the claims. When I returned to the US for a period I started looking at raw data and realized that many of the claims were just not true. That was my personal conclusion and similar to the toaster analogy above to put all the variables in simulations so that it matches the last 10,000 years of climate change has not been accomplished. If a simulation can’t match the actual past what reasonable person would accept its forecast for the future.
I do however fully agree with Dr Aaronson’s comments about transition to nuclear for other reasons. Texas has nuclear power plants and the surveys I have seen the local residents would welcome more reactors.
Comment #78 August 10th, 2026 at 8:39 am
OhMyGoodness #43: “I don’t believe you should consider AI as an invasive species to the mathematical ecosystem wreaking havoc and resulting in extinction but rather as a new symbiotic tool. The new mathematical ecosystem will flourish.”
Scott #72: “But by the same token, AI could also plausibly help find solutions to all our environmental and resource and other problems, and create a utopia on earth.”
Alex #75: “But to me it seems that the AI became good at particular things, and yes, these are some of the things You are particularly good at. But there is also still a good amount of maths that the AI is not good at.”
I have read this post by Tobias Osborne too (and other related posts by him), but had decided against posting a link to it here. I know from my much older experiments with mace4 and prover9 (and the-E-theorem-prover, in Dec 2014) that humans are amazingly bad at finding counterexamples, compared to suitable computer programs. It is hard to describe how helpful it is to know that a counterexample exists: “The companion program mace4 to prover9 finds counterexamples, which really helped me to finish the text at all and streamline it nicely, even so none of the counterexamples is mentioned anywhere in the text.”
What we should realize is that the AI feels competent or even superhuman for tasks where humans are “unnecessarily” weak, but incompetent or subhuman for tasks where at least some humans are extremely strong. Instead of hoping for “solutions to all our environmental and resource and other problems, and create a utopia on earth” or trying to find a niche where “the AI is not good at”, we should rather seize the potential of the now existing AIs “as a new symbiotic tool” which helps us overcome our most serious weaknesses, and try to move forward as well as we can. Of course this won’t “create a utopia on earth”, but it could still help us significant with some new solutions to some of our problems.
Comment #79 August 10th, 2026 at 8:47 am
O. S. Dawg #74
Any evidence that research-level mathematics is a highly transferrable skill? If so, to which domains? I’m not claiming that it’s not. I just haven’t seen any studies that make this assertion. (Lots of disciplines claim their skills are transferrable to other domains but then fail to document such assertions.)
Comment #80 August 10th, 2026 at 8:56 am
This was a good article.
https://www.theatlantic.com/politics/archive/2025/08/dsa-mamdani-losing-elections/683970/
Comment #81 August 10th, 2026 at 9:08 am
As AIs improve, and we tend to take them as more and more human-like, it’s still worth keeping in mind that LLMs are nothing but probabilistic models… well maybe so are human brains, but at least you can’t do this to a human
(that guy breaks the AI behind scam calls):
https://youtu.be/lk3jCuITwcE
Comment #82 August 10th, 2026 at 12:37 pm
Scott #58
ah, when it comes to the Texas power grid and data centers, my AI is saying:
“Texas Governor Greg Abbott recently ordered a de facto moratorium on new data center power grid approvals. Regulators must now audit all planned projects for power usage to safeguard the reliability of the Texas electrical grid.”
Comment #83 August 10th, 2026 at 3:04 pm
gentzen #78
I agree and yes more modest expectations are likely warranted. Interstellar travel and eternal joyful life are not imminent so be satisfied with what you do receive. It looks like hackers and mathematicians are the big winners thus far so good and evil are still in balance.
Comment #84 August 10th, 2026 at 4:17 pm
Scott,
do you think we’ll eventually witness a true “singularity”?
Or AI progress will stay “asymptotic” (gets harder and harder to improve)?
Comment #85 August 10th, 2026 at 5:22 pm
I think the environmental problem with AI is not so much with its current energy use as with the growth of it being explosive and faster than possible growth of clean energy production. According to (of course speculative) IEA predictions, “Emissions from electricity use by data centres grows from 180 million tonnes (Mt) today [2025] to 300 Mt in the Base Case by 2035, and up to 500 Mt in the Lift-Off Case” (they mean Mt of CO2 equivalent per year, if I understand correctly). That’s not peanuts at all (a 200 Mt/year increase split among 1 billion people is 0.2 t/year while the carbon budget per person on Earth to stay below 2°C warming by 2070 is about 2 t/year; and that increase is over just 10 years!).
Comment #86 August 10th, 2026 at 7:17 pm
Nicolas Rougerie #69
That’s too funny, I’m glad I get the chance to follow up with the original coffee commenter!
My calculations for the coffee did take into account the transportation (+5% impact) and the water (negligible). For the prompt I included not only the marginal electricity use, but the impact of training the model (+10%), manufacturing the GPU’s (+10%), and constructing the data center (+1%). I tried to steelman the environmental concerns as much as I could, but I couldn’t escape the conclusion that data centers are remarkably efficient.
I would recommend you to try to reproduce these calculations and let me know if I went wrong anywhere. Who knows, you may conclude that a more effective way to help the environment is to encourage people to use a non-dairy creamer. I think this would even be more in line with your value of everyone receiving a decent share of civilization’s comforts. After all, changing coffee habits is a less impactful change to someone’s daily life than giving up AI, which in many cases has no readily available substitute.
Jean Abou Samra #85
That is a good point. As we learned from Covid, you have to look at where the exponential is headed. I have seen other projections which didn’t worry me too much. These projections do seem more compelling, although I am curious why the 200 Mt/year is split over only 1 billion people instead of the entire population.
Comment #87 August 10th, 2026 at 10:08 pm
@ H. Squash #69
I would be interested by the details of your calculations, could you share them here, or by email if you think for some reason this is more appropriate ? I am posting under my real name, which you can easily you lookup online.
Alternatively I would be interested in links to reliable calculations, if somebody knows any.
Comment #88 August 11th, 2026 at 12:24 am
Anonymous Mathematician #79: I’m not sure what Gowers meant by that line, but presumably he meant that the skills of research mathematicians are transferable to some domain other than AI safety. (Tsimerman would beg to differ.)
Comment #89 August 11th, 2026 at 12:57 am
‘Epsilon Camp’ for kids…is that a Paul Erdős joke I spot there…
Comment #90 August 11th, 2026 at 1:31 am
@Jean Abou Sabra:
One nuance worth adding to this comparison: the IEA’s own projections suggest that the decarbonization of data centre electricity is actually outpacing the growth in volume, not lagging behind it.
Emissions from electricity use by data centres grow from 180 Mt today to 300 Mt in the Base Case by 2035 (up to 500 Mt in the Lift-Off Case) [IEA](https://www.iea.org/reports/energy-and-ai/executive-summary) — a 67% increase over ten years. But over the same period, global electricity generation to supply data centres grows from 460 TWh in 2024 to over 1,300 TWh in 2035 [IEA](https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai) — nearly a tripling. The gap is explained by a fast-greening electricity mix: the ratio of the data-centre electricity mix switches from around 60% fossil fuels and 40% clean power today to 60% clean power and 40% fossil fuels by 2035. [IEA](https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai)
This greening comes mostly from the broader global expansion of renewables rather than direct procurement by tech companies: half of the global growth in data centre demand is met by renewables, whose generation is projected to grow by over 450 TWh to meet data centre demand to 2035, with natural gas expanding by 175 TWh and nuclear contributing a comparable amount. [IEA](https://www.iea.org/reports/energy-and-ai/executive-summary) And at the global scale, data centres account for around one-tenth of global electricity demand growth to 2030 [IEA](https://www.iea.org/reports/energy-and-ai/executive-summary) — less than other sectors like industrial electrification.
So the real risk isn’t that renewables can’t keep pace, but that the portion still met by gas (and coal in some regions) locks in fossil infrastructure a bit longer, with acute local strain wherever data centres are geographically concentrated (nearly half of global consumption in the US, 25% in China, 15% in Europe) [Carbon Brief](https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context) — even as the global footprint stays, per the IEA, a small share (<1.5%) of energy-sector emissions.
Source: IEA, *Energy and AI* (2025), https://www.iea.org/reports/energy-and-ai
Comment #91 August 11th, 2026 at 8:32 am
Good to know I’m not the only one who is struggling with this fear of dying alone as a celibate nerdy math/TCS freak. Even though my hopes of eventually “sorting out” those things are principally based on my faith in God and his promises, this kind of posts really encourage me. Thanks a lot for your time to sharing that with us!
PD: I wish I had met you at HLF11 but unfortunately you didn’t go that year :’-(
Comment #92 August 11th, 2026 at 8:56 am
Sally #89: Yes.
Comment #93 August 11th, 2026 at 8:59 am
Ricardo Trifásico #91: I think the main thing to tell you right now is that God helps those who help themselves. Get on some dating apps. If there’s enough interest, maybe I’ll do a post of all the stuff I wished I’d known about this as an 18-year-old, something I’ve hesitated to do for years.
Comment #94 August 11th, 2026 at 9:36 am
Anthropic announces that Claude has made progress on the Riemann hypothesis.
https://www.anthropic.com/research/riemann-zeta
“`
An unreleased research version of Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis. Drawing on extensive prior research by mathematicians over the past decades, it has increased this bound from 41.6% to 67.2%.
“`
Comment #95 August 11th, 2026 at 11:42 am
“I’ve observed that, the faster we race toward the Singularity, the more I feel like stepping back and asking myself: what do I actually value in life? How important to me are math and science, as human practices to be passed down to curious children? Would I even want solutions to P versus NP and the other problems, if the price were to destroy those human practices forever? How do I wish to spend whatever time I have remaining?”
I find myself asking the same questions. I don’t doubt that AI – however we define it – is nowhere near plateauing in capability. Or that it will someday find solutions to hard technological problems.
I cannot help but ask, though: is everything we’re ready to displace about human activity truly worth displacing? Undoubtedly the kids in the CS camp are far less ‘productive’ at problem-solving than Astral (or any of the other mythical-object-plus-some-decimals model). So? Isn’t it worthwhile as an activity, anyways, to have them learn, figure things out, and even (shocker!!) feel motivated by the prospect of being the first to solve something?
These are admittedly ‘softer’ concerns. My observation is this: we assign some probability to AI-driven civilizational collapse. We also assign some probability to an AI-driven Utopia. The latter assumes that humans on scale will set side millennium-old habits, and live in harmony; that the only barrier to solving current-day giant problems is technological. Disagreements arise on how to balance these probabilities. Some are happy to risk – on everyone’s behalf – really bad consequences, for the promise of a land of milk and honey. Others are not.
In all of this, we could stop to ask the question: how many of the awful problems plaguing us today are actually about a failure of political will, rather than an absence of a technological solution?
Climate change remains an inexorable, huge threat. We know many technological things which will mitigate it, or help us adapt. But a great deal of human effort is being expended to -not- address it. Likewise preventable diseases, or heck, even chronic ones. Extreme economic inequality leads to poorer social and health outcomes. Scientists scream themselves hoarse (ok, politely state, with error bars). Society shrugs.
And yet we believe an AI-generated ‘solution’ to any of these problems is going to break us out of our stasis? For this we’re willing to dim the dreams of kids, and play jenga with the futures of our current students?
The ICM was a revelation. No one on stage was willing to offer predictions, with confidence, about the career prospects of the young mathematicians in our midst. We agreed it was a very exciting time in mathematics, with proofs and counterexamples raining upon us. Yes, there would be upheaval. And you, young grad student sitting next to me – your career prospects, your efforts, your ambitions – yes, you. Purchase some frontier models, play with them, and don’t make me think too hard about your future, because it’s uncertain. You’re an inconvenient reminder of the actually hard stuff.
Thanks again, Scott, for pausing to share what I think are really important humanistic concerns.
Comment #96 August 11th, 2026 at 12:12 pm
Adam Treat #94: Exciting!! We’re living in an age of wonders.
Comment #97 August 11th, 2026 at 6:10 pm
If knowledge is the ultimate power,
then, if you are Anthropic and created a version of Claude that’s good enough to solve the Riemann hypothesis, and many things beyond that…
would you even release it or keep it for yourself to claim ownership on all the things it finds out?
That’s a bit like when NVidia realized they would make more money by hoarding their own GPUs to mine bitcoins than by selling them to gamers. Just the ultimate/winner-takes-all version of that.
Comment #98 August 11th, 2026 at 6:27 pm
(not as spectacular as solving the Riemann hypo)
There has been a recent surge in “mods” that modify 2D “flat” games to run in VR.
Not just the stereo 6-degree-of-freedom vision aspect, but also replacing the in-game gamepad controls with VR controls (hand tracking), so that you can now hold and aim all weapons, with one hand or two, do manual reloads, manipulate objects.
It’s all quite complex: there’s a lot of matrix algebra to get right, and performance is critical.
And some of the more successful mods have been developed by people who know nothing about coding, using Claude.
The code itself is quite complex, and the AI is smart enough to tailor the code to proprietary game engines (it’s not all Unity or Unreal), using run-time injection.
And keep in mind that the AI isn’t even able to judge the results on its own, it has to get human feedback and rely on log files.
Comment #99 August 12th, 2026 at 4:14 pm
Scott #96 “Exciting!! We’re living in an age of wonders.”
Please, write about this and other wonders. AI will never do those things better than you, i.e., talk about what YOU think about these things.
Comment #100 August 12th, 2026 at 8:35 pm
Nerds doing nerd things for fun is cool and stuff, but I suppose that you, being already a well-established researcher and a tenured professor, don’t quite appreciate the existential dread experienced by those who aren’t (including yours truly). Right now math camps are sponsored by the state or, sometimes, by technological business, because they help preparing the next generation of competent workforce. In the future where cheap machines are better at math than every human, there’s no need for such people. Who will sponsor the math camps then? I still believe firmly that knowledge of mathematics is necessary for individual humans to adequately understand the reality, whether machines know it better than Terrence Tao or not. But I also suspect that the populace adequately understanding the reality is the last thing that people in power want, especially in such places as the contemporary United States, China and Russia.
Comment #101 August 12th, 2026 at 9:36 pm
Scott #93: I promise that many people will want to read such a post! Share with the world what you’ve learned 🙂
Comment #102 August 12th, 2026 at 9:40 pm
Scott: conjecture ai shows eth and ugc false before end of year.
Comment #103 August 13th, 2026 at 5:23 am
@Scott
I followed your update to the oracle and from there to your debate with Penrose.
Funny how some of the objections people raised have been settled by AI already ,
but it also reminded me how great your blog was and how I miss the ‘good old times’ before Trump, Hamas and worries about the current AI oligarchy f*** it all up …
Comment #104 August 13th, 2026 at 7:26 am
Scott: “we’re living in an age of wonder”
Are we?
Solving any significant problem is hard by definition, in the optimization sense, as in NP hard…
years ago Scott would dismiss P=NP claims by saying it would be like conjuring magic, all hard problems would just melt away, an unending waterfall of miraculous wins.
In reality we know that it you throw enough compute power at anything hard, be it by exhaustive search or heuristics, you can often eventually get lucky (things like genetic programming aren’t new, decades ago they found better versions of sorting algorithms… it’s a matter of cycles).
So if we could give an AI a list of arbitrary unsolved hard math problems and it just effortlessly crushed them all in a few million tokens, we’d have to wonder if some deep breakthrough in computational complexity isn’t going on, like N=NP…
But we don’t see (yet?) this type of success, and the real question is the true cost of this “luck”.
The cost of making some incremental progress on Reimann isn’t just what OpenAI charged for the few million tokens it took.
There’s the cost of all failed attempts, the cost of data centers and their environmental impact, the intellectual property that was illegally used, the yet unknown long term effects on society,.. all now and the foreseeable future.
Is it really an age of wonder or we’re about to irreversibly screw ourselves and the next generations.
When it comes to coding, of course the current big AI players would love it if the entire software engineering industry got “wiped out” by making sure that any new software or modification to existing software has to “go through” their data centers… they would defacto indirectly own every single piece of code out there (trivial or significant), an infinite source of massive income for all eternity. And the more bloated the software becomes, the better: harder for humans to own it. On top of that, with things like github integration, software you thought you owned can be freely used by bigtech for training, accelerating the downfall of human coding expertise and quality.
As far as using AI for “therapy”, it’s a bad idea: even though AI can give what looks like sound advice, it’s ALWAYS from a place where its priority is to please its user. Anyone who’s truly struggling psychologically is at risk to get in a worse place because of this, without even realizing it (Jung and Freud understood that the humanity of the therapist was central to the entire process, making it quite complex).
Comment #105 August 13th, 2026 at 10:49 am
To offer a different perspective: I think it’s fair to say that all AI advances in pure math have so far been low-hanging fruit in the sense that they do not offer conceptual advances but instead take deep mathematics by humans and push them a bit further using more or less standard techniques. Sometimes this results in noteworthy theorems. I think “progress on RH” is a pretty absurd way to characterize the Anthropic result, though. Also, AI still has yet to do anything interesting in my field (topology). This is perhaps reflected in the fact that the number of GT arXiv postings hasn’t grown much over the past year, in contrast with postings in combinatorics, say.
Comment #106 August 13th, 2026 at 5:26 pm
John #105,
Thanks for sharing your perspective. I’m curious about your comment that AI has yet to do anything particularly interesting in topology. Why do you think that is?
Do you see it mainly as a consequence of topology receiving less attention from AI researchers (or having a smaller research community overall), or do you think there is something about the nature of topological problems that makes them inherently more difficult for current AI systems to tackle?
In other words, do you think the gap is mostly sociological or technical?
Comment #107 August 13th, 2026 at 7:52 pm
Conjecture Either we are moving to an age of dimwits or to an age where there are plenty of people who are learned in multiple fields like the pre 19th century time.
Comment #108 August 13th, 2026 at 10:22 pm
Nicolas Rougerie #87
I had AI make an infographic with a detailed breakdown and sources – https://claude.ai/code/artifact/2bdc5de5-8721-4b43-9231-59a70c8842e9
Notes:
* We are using a “heavy thinking” prompt as our baseline, to try to match Scott’s original use case of multiple chat turns for his research. Typical LLM responses are shorter (so could use 10-20x less energy)
* I made a mistake with my original numbers where I over-reported the impact of adding cream to the coffee (I used numbers for a latte, which is mostly milk instead of just a splash of milk). Other numbers have changed as well due to finding more definitive sources.
* The error bars are quite large in some cases.
* I added a number of other scenarios/items in the “Everything” tab.
* If you want something less flashy, you can export to CSV at the bottom
Comment #109 August 14th, 2026 at 1:51 am
Scott #46:
I believe my Criterion 2 is not met, at least in the nominal case.
https://arxiv.org/pdf/2608.13110
Comment #110 August 14th, 2026 at 2:28 am
Good morning Professor,
I deeply admire your ability to weave together technical, philosophical, and personal threads into something so distinctly recognizable — much like a piece of classical music where, after just a few bars, you already know exactly whose hand wrote it. The more I read this blog and its comments, the more I realize how vast the realm of knowledge truly is — far more than I could ever hope to fully grasp.
On a lighter note: reading about the Aaronson Oracle in this post sent me straight to Spencer Stanton’s implementation, and I couldn’t resist trying to cheat it. I’ll confess upfront that part of what follows is decidedly nerdy: I happen to know the 4-bit binary patterns of all sixteen hexadecimal digits by heart, a small party trick that turned out to be surprisingly useful here.
My approach: mentally generate the 64-bit sequence corresponding to the hex digits 0 through F, encoded 4 bits at a time (0000, 0001, 0010, …, 1111), typed out as F/D key-presses. The reasoning behind the “cheat” was simple: the Oracle’s predictive depth looks back at most about 4-5 presses to find a matching pattern. Since my sequence was built from independent 4-bit blocks with no exploitable higher-order structure, I expected the Oracle to be reduced to coin-flipping — around 50% accuracy. Sure enough, it landed at 51%, about as close to chance as one could ask for.
I should stress how weak this “test” really is: a single run through the sixteen digits, no repetitions, no controls — and, honestly, I have no intention of running it again, since reciting the hex “times table” to myself once was already enough of a workout for a Thursday night. Take the 51% as an amusing anecdote, not as evidence of anything rigorous.
What struck me most, thinking it over afterward, is what the result does — and doesn’t — say. The Oracle exploits the implicit, automatic patterns in human behavior: the small unconscious biases in alternation and repetition that give away our attempts at randomness. By substituting a fully deterministic rule held consciously in mind, I wasn’t generating randomness at all — I was executing a memorized program, just with fingers instead of a CPU. It’s less “free will defeats the machine” and more “a known deterministic generator is trivially different from the noisy generator the Oracle was built to model.” Which, in its own way, still feels connected to the themes of your post: the boundary between what a simple pattern-matcher can predict and what it cannot has less to do with mystery and more to do with which kind of process is generating the sequence in the first place.
(One amusing footnote: the log showed 63 presses instead of the 64 I intended. The whole thing happened quite quickly — it was around 11 PM CEST on a Thursday night — and I was so absorbed in silently reciting my little “times table” that I barely glanced at the screen until I hit the 63rd click, so to this day I honestly can’t tell whether it was a slip of the mind, a slip of the finger, or a slip of the mouse. I was fully convinced, until I checked, that I’d landed all 64. Maybe this glitch explains why I got 51%.)
Thank you, as always, for the post — and for the little quarter-century-old joke that turned out to have such a long and thought-provoking life.
Warm regards,
Matteo
Comment #111 August 14th, 2026 at 3:37 am
I wonder if there will soon be a wave of scientists turned philosophers/mystics. I guess rather than the Besht, you would be the Optimized.
Comment #112 August 14th, 2026 at 5:20 am
wb #103
“ how I miss the ‘good old times’ before Trump, Hamas and worries about the current AI oligarchy f*** it all up …”
You must be quite optimistic in that another two years Trump will be drooling in some quiet place as Biden is now. Also that China has delivered a severing blow to the AI oligarchy with open release of Kimi. That leaves Hamas and so glad that Israel has refused to leave Gaza, to support some extremely flawed BS plan, until at least Hamas disarms.
I envy what must be your current optimism. Maybe the roots of our problems are not inside us as humans but easily solvable as you have pointed out.
Comment #113 August 14th, 2026 at 7:47 am
Matteo Vitturi #110
I tried this on one of the versions of Dr. Aaronson’s program. I entered the digits of Pi sequentially mod2 and the program passed each time except for one wrong guess. I don’t think this violates the spirit of a problem since humans can freely choose to base actions on a mathematical object or outcome of an experiment that are known to be random. It is our free will that enables it to be so.
Comment #114 August 14th, 2026 at 6:23 pm
OMG
“It is our free will that enables it to be so.”
1) your brain is nothing more than a collection of atoms
2) the atoms in your brain are no different from the ones in any other clump, and they just move around in the only way that’s compatible with the state of all the other atoms in the universe and the laws of physics.
3) No matter the details of the physics, there is only cause and effect, with some amount of randomness (assuming it’s even a thing), and there’s absolutely no room for your ego to magically unshackle itself from this simple ground truth.
4) the concepts of freedom and will you’re invoking stem from the notion of counterfactuals – because alternative versions of cause and effect chains are conjured in our mind (like an AI hallucination), and you believe those could have happened, therefore, at every moment, we must have the ability to pick among those possible future paths arbitrarily, before even knowing what those possibilities are, or could have been. As if, when imagining the counterfactuals after the fact, you’re able to signal the past version of yourself to alter its “choice”, and this would repeat in a loop until it stabilizes. But even in this far fetched theory of retroactive influence, everything that happens is still a chain of inescapable causes and effects, going in a time loop. And the stable version is left with no freedom of will whatsoever.
Comment #115 August 14th, 2026 at 8:38 pm
Mark H #114
I simply don’t agree. If you can explain what causes an electron spin measurement to be either up or down, and thereby abolish superposition, then maybe I am mistaken. Otherwise I can base my decisions on the outcome of the measurement and not possible for any entity to predict my resulting actions a priori. The cause of my actions is the measurement result that is random with no discernible cause. The cause and effect chain starts only after measurement and not before.
Comment #116 August 15th, 2026 at 2:59 am
Mark H #114
I agree my brain is made from atoms but don’t agree it is just atoms as I look at collections of atoms that form my environment. My two dogs are collections of atoms but I realize one of them is not even bright enough to collapse the wave function. 🙂
Comment #117 August 15th, 2026 at 4:16 am
How would one prove unequivocally there is no free will. Maybe if there were some digital simulation of me that perfectly predicted my actions then I would have to accept my sense of free will is illusory.. The future is perfectly predictable.
My point is that I can take steps to render a simulation inaccurate by consciously taking actions dependent on some measurement that cannot be known before the fact.
In lieu of a perfect simulation of me I will continue to believe that I have free will and that my future isn’t fully determined in a mechanistic way. If you have some predictive model of me we can conduct the experiment now to see if in fact I am perfectly predictable. I am ready and eager to participate. If you don’t then the claim that there is no free will is an unsupported belief while my belief I do have free will is supported by the failed experiment.
Comment #118 August 15th, 2026 at 9:01 am
Sorry just one more point.
I saw a video recently of a polar bear in the wild. It rendered an immature seal immobile and left it on the ice and retreated some distance and swiveled its head constantly. The reasonable supposition is that the bear had a model of seal behavior and if the mother was still in the area she might try to help , or make her position known, and the bear would enjoy a much larger dinner. She didn’t and so the bear returned to the immature seal for consumption.
On the African savannah, after forests died out early hominids were outstengthed and outeverythinged by both predators and large prey except for intelligence. There was an enormous advantage in not being predictable in even simple circumstances.The hypothesis that follows is that lack of rote predictability had an enormous survival advantage for early hominids and so was naturally conserved.
I understand this is entirely speculative but I believe it is of interest. There are few things in the animal kingdom with more life ending potential than hungry large cats or male elephants in musk, especially for mid sized creatures that were relatively puny and slow and fresh from the trees.
The entirely speculative nature of these comments are certainly completely distinct from the experiment proposition. If anyone questions this I have absolutely no data to provide any defense.
Comment #119 August 16th, 2026 at 3:30 pm
There is nothing mystical about free will, in my opinion. E.g., the legal usage: did you sign this contract of your own free will, or did somebody hold a gun to your head? It depends on the existence of determinism, which is what allows us to make determined decisions. The will part is wanting to make the best decision possible, according to one’s own criteria, and therefore learning from the results of bad decisions, not making the same rote decision over and over. Anything which can do this in an unslavish way and learn from its errors (dogs, etc.) has some free will in my definition. Those who make decisions based on emotional drives without considering the past or future effects could be said to lack free will, in such instances,
Comment #120 August 18th, 2026 at 6:05 am
John #105
Just for your info: some time ago (15 years?) Iwaniec told me that he doesn’t think the current methods have any chance to get above 50%. (He provided some construction, a kind of a counter example not distinguishable — in his opinion — from the real zeta, by these methods. I didn’t grok his construction. — And since I thought that the method I had in mind might be immune, I didn’t pursue this example. — BTW, I have no clue how the new method works.)
Comment #121 August 18th, 2026 at 6:48 pm
What accounts — to the best of your judgement — for the (so far) sustained outburst of interest in (pure) mathematics by OpenAI and its peer competitors?
Comment #122 August 18th, 2026 at 7:01 pm
PS Reacting to Comment #119: as Kant put it, “Also ist ein freier Wille und ein Wille unter sittlichen Gesetzen einerlei.”
Comment #123 August 19th, 2026 at 10:13 am
Dear prof. Aaronson,
Reading your post, I cannot avoid thinking on Oscar Wilde and “The Picture of Dorian Gray”.
Look at what is happening with AI. All these brilliant results—lower bounds, disproofs of old conjectures, quantum benchmarks—they are like the young, perfect face of Dorian. It promises us eternal youth and all the answers to deep problems served on a plate.
But there is a hidden price, and that is the portrait in the attic. That portrait is the human practice of science itself. It is the dirty, beautiful process: sitting with paper and pencil, struggling with equations on a whiteboard, teaching kids at a camp, or mentoring a young student in the lab. If we give all the thinking to the machines, the external face of progress looks immaculate, but the human soul in the attic is quietly rotting.
Your decision to step back, go to Epsilon Camp, and reflect is not a distraction or a surrender. It is looking at the portrait before it is too late. We do not do physics or mathematics only to cross out solved problems from a list. We do it because the human act of asking, learning, and passing it down is what gives it value. If we lose that, we sold the attic. Best regards, L. Vergara
Comment #124 August 19th, 2026 at 10:45 am
AG #121,
Formal mathematics offers binary correctness. This provides an unambiguous, automated ground-truth signal for Reinforcement Learning with Verifiable Rewards (RLVR).
Comment #125 August 19th, 2026 at 12:01 pm
#118:
That’s an interesting way to see it! But doesn’t it take more than just unpredictability to manifest that edge? I mean a cockroach is very unpredictable in which way it flees when it perceives danger, but that’s not what we mean by intelligence. There is also the range of possible behaviors, which in my mind is the key distinguishing attribue of intelligence.
To tie that up to quantum philosophical musings, the unpredictability of quantum measurement does not help the cockroach too much. And quantum unpredictability does not mean anything can happen, ie an imaginable branch is not necessarily feasible in any roll of the dice. One extra ingredient to throw in this mix is chaos. An electron getting measured in spin up state might not change your immediate behavior, but it could cause you to declare all out war next month (I think?). I’m thinking of it as humans amplifying tiny effects, in the same way an avalanche photo detector does. Cockroaches too can amplify tiny effects, but that only causes them to turn right or left, whereas for humans it could cause us to do any of a huge number of pairs of possible macroscopic actions. Why do we have a huge range of possible macroscopic actions? Because we’re much better able to amplify small differences. A photo detector also amplifies tiny differences very well, but it is insensitive to most differences in initial conditions, whereas a human brain connected via sensory inputs to the environment is much more sensitive to initial conditions — operating much closer to the “edge of chaos” than cockroaches.
Comment #126 August 19th, 2026 at 2:54 pm
Glassmind Duo #124: You only get an “unambiguous, automated ground-truth signal” if you force the AI to convert all of its proofs into a form where they can be verified by proof checking algorithms. But the recent AI pure math results were given by the AI in a human-readable form and checked by humans, not machines. I think this means they did not use RLVR in the way that you are saying.
Comment #127 August 19th, 2026 at 8:15 pm
Glassmind Duo #124
I’m going to remove the word “formal” from what you said — otherwise it doesn’t match the question you tried to answer.
If so, then what you said matches the state of math at about the time of Lewis Carroll. A lot of problems with difficulty of verification have been discovered since then.
(I am completely not specialist, and can think about it in metaphorical terms only, but) As far as I understand, the current stumbling block is that there doesn’t seem to be a trustful way to understand that a given formalization matches a given known mathematical construction. (Of course, this presupposes that a formalization has been found. It seems that a lot of things which appear “harmless” when written “in a paper in a journal” require formal type systems which may be way above the head of the current assistant [?] proof software.)
Comment #128 August 20th, 2026 at 10:50 am
Gentzen #78: “I know from my much older experiments with mace4 and prover9 (and the-E-theorem-prover, in Dec 2014) that humans are amazingly bad at finding counterexamples, compared to suitable computer programs. ”
Ah, now I understand why math was easy for me. I am not that good at pure analytical thinking, but I did graduate (1977) on finding some counterexamples to complex questions. Where pure abstraction was hard for me, constructing (counter)examples was much more fun than proving theorems…
And proving that a counterexample works is usually easy. Just plug it in and see a theorem fail. The hard part is finding the counterexample, but easier for a constructive thinker than for the lofty analytical thinker.
Comment #129 August 20th, 2026 at 1:17 pm
Dacyn #126,
Aren’t you confusing initial motivation with the latest result? Yes, the recent result from OpenAI was a human-verified, natural-language proof, but that doesn’t say anything about the training itself. They haven’t said yet, but I’d be very surprised if it didn’t involve RLVR in some form. Want to bet?
Ilya #127,
Right, sloopy of me to confuse “pure math” with “formal math.” But reaching Lewis Carroll–era math was already a dream not that long ago. Great point that there’s a gap between “provably correct” and “correctly represents the intended claim.” Maybe a real bottleneck.
Comment #130 August 20th, 2026 at 2:42 pm
Glassmind Duo #129: What exactly is your claim? That the AI was trained by converting its natural language proofs into machine readable proofs and then verifying them? Or just that RLVR was used in some manner? I don’t know much about RLVR so I wouldn’t feel comfortable betting on the latter. If your claim is the former then well, you could easily be right so I wouldn’t want to bet at even odds, but maybe we could bet at 3:1 odds.
Comment #131 August 20th, 2026 at 9:31 pm
Dacyn #130,
My original bet was on the latter, but OK, 3:1 on the former sounds fair to both sides. You donate $17.29 to Lean FRO if I win, and I’ll donate $49.6 to your favorite charity if you do. Deal? Hopefully OpenAI won’t keep that secret for too long.
Comment #132 August 21st, 2026 at 8:31 am
Oh Dear!
If you’re finally starting to buy the doomer arguments then at some point you can expect a full scale panic attack. I went through mine about fifteen years ago and can offer some sort of crappy-soulless-rationalist-automaton-type counseling, although not much in the way of hope. But it is possible to stay sane, and you’re taking the right approach! Say bollocks to it and enjoy the sunshine while you still can.
I’d still like that coffee if you’re ever in England. Or just give me a ring, I’m easy to find.
Best of Luck, Scott. You’re a transparently good person, and you have made the lives of the people around you better. Hang on to that and keep doing it.
Comment #133 August 21st, 2026 at 2:13 pm
Glassmind Duo #131: OK, deal. How should we arrange to contact each other in the future? Just keep an eye on the comment sections of whatever the most recent post is here? Or should we exchange emails?
Comment #134 August 22nd, 2026 at 11:06 am
Dacyn #133, Checking here seems fine. Any preferred charity? Please, nothing that contributes to polarization, directly or indirectly.
Comment #135 August 22nd, 2026 at 9:40 pm
Is the EA Animal Welfare Fund OK?
Comment #136 August 22nd, 2026 at 10:13 pm
No problem. And it’s not a priority area for me either, so that’s a really smart pick on your part. 🙂
Comment #137 August 23rd, 2026 at 5:33 am
Does it improve determinant size in det vs per problem?