I've seen several relevant quotes posted today, so I used GPT-Sol 6.1 to gather expert reactions to the results released on October 6 and checked each quote against its original source.
On “invent a new concept that does real work,” the nearest candidate here is the new code Moshkovitz describes. Ahle describes a different route to a bound, built on Strassen’s existing asymptotic-spectrum framework. These are early reactions to OpenAI’s October 6 mathematics release; the last two discuss the same plane-colouring proof.
Dana Moshkovitz: UT Austin professor specializing in probabilistically checkable proofs and pseudorandomness.
“Upon further inspection, the proof constructs a completely new tree-based code that somehow has a noise test with properties similar to that of the short code.”
Her addendum, posted by her husband Scott Aaronson in comment #2, October 7. In her own comment #5, she emphasizes the existing frameworks and results the AI built on, while calling the results “paradigm shifts.” She also criticizes the exposition and citations.
Thomas Dybdahl Ahle: theoretical computer scientist whose research includes algorithms and tensor sketching; Head of Machine Learning at Normal Computing.
“the new FMM paper is surprisingly elegant. It uses a very different approach than previous results. Instead of powering the CW tensor and trimming it, [they] define a potential function on tensors, and show the whole result by contradiction by studying the simple convolution tensor.”
Original X post, October 7. “CW” means Coppersmith–Winograd; “[they]” corrects a typo. The paper describes a profile, separation construction and growth argument; “potential function” is Ahle’s description.
Simon Pepin Lehalleur: University of Duisburg-Essen; algebraic and arithmetic geometry, particularly motives and motivic homotopy theory.
Of the OpenAI proofs he had skimmed, he singled out:
“The most original … is the algebraicity of the Kuga-Satake correspondence for K3 surfaces, using techniques from mirror symmetry.”
Later that day, on a very superficial reading, he judged the Bloch-conjecture proof “not based on deep new insights into codimension 2 cycles in general,” though “very substantial progress on cycles on (powers of) surfaces.”
Joshua Zelinsky: published number theorist with a mathematics PhD from Boston University.
“I've only had time to start looking at the chromatic plane proof, and it doesn't look like the method is a direction that the prior lit used to my knowledge. This is far beyond merely building on existing methods or seeing connections between different problems.”
Bluesky reply, October 6. The paper credits earlier invariant randomization and calls its averaging standard; the proof combines a transfer to weak measurable colourings with a geometric obstruction to five colours.
Dmitry Rybin: cofounder and chief scientist of a Shenzhen startup working on AI for mathematics; trained in pure mathematics at HSE; CUHK machine-learning PhD, 2026.
“Somehow the model found that any K-coloring <=> "weakly measurable" K-coloring, which seems out of nowhere”
Original X post, October 7. His follow-up says it used “ridiculous human-developed machinery”: amenable groups, Pontryagin duality and Furstenberg–Zimmer towers.
Since I'm being quoted here about the chromatic plane/Hadwiger Nelson I do think I should note a clarfiication here:
First, there's been some people who have erroneously concluded from the paper that "weakly measurable coloring" which the AI defines with not citations is novel. There's a lot of very similar ideas to it in the literature which are not cited, and if I were refereeing the paper I would insist they cite. In particular, there is a 2014 paper of Bachoc, DeCorte, de Oliveira Filho, and Vallentin. (I poked around with Claude and it indicated some other papers that are also precursors here that are relevant.)
The interesting new idea is there Theorem 1.3 which says that proer colorings correspond to weakly measurable colorings. That's a really key ingredient and doesn't seem (to me at least) to have been in the literature.
As a mathematician I don't think there's some kind of qualitatively different "creativity" other than being able to notice and track high-order patterns across different kinds of data and different levels of abstraction.
I've always emphasized the importance of sheer persistence in mathematics. I had an olympiad coach who told us, "the easiest way to build an edge over the competition is to be able to stomach twice as many variables on the page." There are some famous notebooks of Gauss containing pages of tens of thousands of prime numbers calculated by hand in ranges around a million, looking for patterns. When I visited Waterloo a couple years ago, I was given a tour of a dusty archive where one can find Tutte's (one of the founder's of graph theory) old binders, there was a stack of probably half a dozen binders where hundreds of individual cases of some theorem (maybe the four-color theorem?) were proved by hand.
Many a mathematician's career is built on pushing a method farther, in more technical detail, than anyone else. In some sense it's not surprising that computers with at least human level math understanding and infinite knowledge and persistence can do what they do.
As a mathematician I don't think there's some kind of qualitatively different "creativity" other than being able to notice and track high-order patterns across different kinds of data and different levels of abstraction.
As a theoretical physicist I also think this is basically true but that the models seems to remain much worse at this than humans. With a particular emphasis on "high-order".
However they do have much more data to work from and, as you say, massive persistence. As far as I can tell the discoveries that have come out of them so far have come from fairly low-order pattern matching on a huge knowledge corpus, and trying many ad hoc moves. Low conceptual intuition, high compute. Time will tell if this applies to the latest batch as well. I suspect so.
I think it's not an ASI crux whether LLMs are significantly creative, because being creative is by itself not a path towards fast accumulation of technological culture (accessing far future technologies and theoretical disciplines; as opposed to looking ahead some fixed distance, even if a somewhat greater distance). Instead, 2026 demonstrates that scaling LLMs 10x compared to 2025 was sufficient to reach high levels of common sense (the kind of thing RL only elicits but doesn't create), which is key to applying technical skills in complicated engineering projects, and to setting priorities (including for developing new skills). That same amount of scaling also allowed LLMs to go from doing well at IMO to solving Millennium Prize problems. Since there's another 100x of scaling until 2032, it's almost certain that all engineering that humans can do is going to be automated just with the current methods, including creation of RL tasks/envs to plug LLM capability gaps as they get discovered (which enables LLMs to automatically learn the skills for all the other engineering things in short order).
This tells me that 2026 establishes a likely 2040-2045 upper bound on timing of ASI (fast architecture-rewriting strong RSI), because automation of engineering enables a robot-industrial explosion. It's also more likely that scaling to 2028-2032 LLMs (after 1-2 more leaps like the 2025-2026 one) makes them capable enough that continual learning in a strong sense (sufficient to get fast cultural accumulation started) gets within the lookahead distance accessible with the kind of engineering-type capabilities we are seeing this year. But that depends more on the intrinsic difficulty of continual learning, which could remain inaccessible to LLMs if it requires more conceptual breakthroughs (than they can make in one step), in which case the whole 2022-2026 LLM boom is not informative about the timing of when those breakthroughs will occur.
That is, even if LLMs are capable of conceptual breakthroughs (within some lookahead distance of the current technological-cultural frontier), they also need to be capable of chaining them together (which is what I mean by unbounded/fast cultural accumulation). This requires learning the previous ideas in order to become capable of inventing the next ideas, with fast invent-learn-invent cycles that build on each other. Currently, this could only happen when new models get trained (even when the training is also automated by LLMs), which might be too slow to matter for the kinds of inventions where the chaining of inventions (that are only accessible in sequence, not in parallel) is the bottleneck, at least until an industrial explosion speeds up even the next-model training step a lot.
Nice, yeah I’d also be curious about this.
Some of this will happen organically as experts digest the results. I’ve found a bit of relevant information already, but will refrain from commenting, everything will be clearer in a few weeks.
Many people expect the AI paradigm to approximately AGI complete, and expect it to smoothly scale to AI that can do innovative science. i.e. the kind of science you'd need to do to develop a plague from scratch that actually kills all humans, with very limited opportunity to experiment … A periodic thread of disagreement has been TsviBT, Steven Byrnes and some others arguing that, yes, the AIs are getting better at verifiable tasks, but they still seem to be missing some basic sauce that lets them actually form new concepts on the fly and apply them …
(For the record , I’ve long been a strong advocate of (something in the vicinity of) “LLMs can’t form and use novel concepts”, but am somewhat more agnostic on “…and therefore we needn’t worry about LLMs killing everyone / taking over”. I’m optimistic about the latter, but wouldn’t make a much stronger statement than that. Cf. recent discussion here.)
One of the things is in §1 of my post “Sharp Left Turn” discourse: An opinionated review, or my linear algebra thought experiment. Like, if an LLM comes up with some new mathematical construction, then that construction exists only in its context window, not its weights, which is fine if it’s just one new construction but rapidly hits limits if you want to build giant new edifices of new concepts built on top of new concepts etc. The context window can’t handle all that, rather the “knowledge” has to somehow get into the weights.
So then people will say “Well humans can’t build giant new edifices of new concepts built on top of new concepts either”
and then I’ll say “Of course they can; the most magnificent example is that billions of humans over thousands of years built all knowledge from scratch”
and then they’ll say “Well the context window is just gonna keep growing”
and then I’ll say “It’s not about length, it’s about serial complexity, again see that linear algebra thought experiment thread”
and then they’ll say “OK fine, we’ll set up the LLM to produce math proofs and then automatically fine-tune on them”
and then I’ll say “I think that would go off the rails because it will make a few mistakes and then when we fine-tune on those mistakes it will get dumber and make more mistakes in a vicious cycle”
and then they’ll say “we’ll Lean-verify before we feed a proof back in”
and then I’ll say “sure that might work in principle, but isn’t it interesting that humans don’t need anything like that to do math? And also, that works in math but not in lots of other domains.”
Thanks for the reply! Two points.
In your original comment you said “LLMs can’t form and use novel concepts”. However, in your explanation you seem to be saying something different. I would summarise what you say there as: "Sure, LLMs can form and use new concepts. However, the degree to which they can generate new ideas and then continue to develop new ideas on top of those (serial complexity) is limited by their context window because they can't do continual learning." Is that a fair summary of what you are saying?
The second point is that even if you agree that continual learning is an important bottleneck, the question remains how far you can get without it. You say that a current LLM "rapidly hits limits if you want to build giant new edifices of new concepts built on top of new concepts etc." My question is, how do you view things like OpenAI solving Navier-Stokes by having 10,000 agents cooperate, work on different tasks, communicate etc.? It seems to me like they managed to produce a result with a pretty high degree of serial complexity by doing so. (Like you say, this doesn't mean that they then were able to integrate that into their weights a la continual learning, but the output certainly seems to have had a high degree of serial complexity)
My underlying argument is that "form and use novel concepts" or 'serial complexity' might not be the right metrics to think about. Also LLMs already seem to be able to do both. I think that the thing people care about when they ask if LLMs can be creative is something different, more what you were gesturing at with your linear algebra experiment. My hypothesis is that often when people say things like LLMs are 'not creative' or 'can't form novel concepts' or 'can't achieve high serial complexity' they have something else in mind, but putting that thing into words is actually difficult. One framework that seems to come closer to putting into words what people mean is Margaret Boden's three levels of creativity.
Is this what you were getting at with your linear algebra experiment?
Also curious about this. I've just been working on an article draft provisionally titled "Can LLMs develop novel concepts" and before the most recent dump, mathematician comments seemed to be pointing me towards "no", but it's a bit problematic to define what exactly "novel" means here and with this many results, it'd be odd if there weren't at least some results that met that criteria. Here's what I had in my earlier draft (that hasn't yet incorporated any discussion of these most recent results):
LLMs keep doing new things every day, and by now claims of “they don't really reason” seem hollow. They can find decade-old vulnerabilities in extensively audited codebases, solve open math problems, and create new kinds of mathematical objects. Prompt a recent model with “come up with a novel concept in the domain of X”, and it will certainly come up with something.
Still, there is a difference between “finding instances of a known concept” versus “coming up with the concept in the first place”, and “applying a known procedure to create something” versus “inventing the procedure from scratch”. As very basic examples:
To be clear, being only good at the former can still have an immense practical impact. If you could magically learn about every single buffer overflow in any software ever, you could cause an immense amount of damage. Just being superhuman at applying existing human concepts would still be enough to drastically change society.
At the same time, a security researcher who only knew about buffer overflows and wasn’t able to find out about any other types of vulnerability would remain permanently capped in terms of what they could do. [...]
For mathematical proofs in general: I’m not a mathematician, so it’s hard for me to evaluate this. But reading commentaries from mathematicians and Fable on various AI-driven math results, the sense I’ve been getting is that the proofs tend to be successful for the same reason why AI is good at finding cybersecurity vulnerabilities: AI knows a lot of results that act as building blocks it can try combining together, it has the patience to grind through a lot of fiddly stuff, and it’s willing to try approaches humans have rejected as unsuitable.
For example, in this X thread from March 2025, Daniel Litt reflects on the way o3-mini-high shocked him by solving a FrontierMath problem that he had authored. He recounts that he thought the problem would be hard because it required a lot of background, specifically knowing four different math facts that most mathematicians don’t know. “Putting those facts together, we win.” He concludes that probably 90% of math research is “routine” in the sense of looking up or knowing lots of facts and putting them together, and that the FrontierMath benchmark then ends up measuring something like “(a) how much known (if possibly obscure) mathematical knowledge does the LLM have, and (b) can it match a problem to the knowledge it has memorized”.
Later, in May 2026, Thomas Bloom’s commentary on OpenAI’s disproof of the unit distance conjecture goes along similar lines:
While I was still very surprised to hear of the this result, this was dampened slightly when I learnt it was a construction of a counterexample, and still further when I learnt that nature of the construction, being (with the benefit of hindsight) a natural, albeit highly non-trivial, generalisation of the original lattice-based construction of Erdős.
On examining the construction, it becomes more clear how people had missed this before – it requires the confluence of several different unlikely events: that a good mathematician is
(1) spending significant time in thinking about the unit distance conjecture in the first place;
(2) seriously trying to disprove it, despite the oft-repeated belief of Erdős that it is true;
(3) believes that there is mileage in generalising the original construction to other number fields, and so is willing to expend significant time in exploring such constructions
; and
(4) sufficiently familiar with the relevant parts of class field theory to recognise that the appropriately phrased question about infinite towers of number fields with appropriate parameters can be solved using existing theory.
The AI met all of these criteria, and its success here echoes previous achievements: it often produces the most surprising results by persevering down paths that a human may have dismissed as not worth their time to explore, combining superhuman levels of patience with familiarity with a vast array of technical machinery. [...]
On the other hand, perhaps some in the area will be a little disappointed with how little this tells us: it does not introduce any powerful new geometric tools, or hitherto unsuspected structural results, that a proof of the unit distance conjecture would likely have called for. Still, while perhaps not the proof of a conjecture that we had hoped for, no doubt this construction and the ideas involved will have a major impact in discrete geometry.
From following some mathematicians on social media and asking Fable to analyze various recent math results, there at least doesn’t seem to be a clear consensus disagreement with the claim that AIs would mostly be doing “routine” math very, very well.
One admittedly cherry-picked example more is Fable's commentary on the AI proof of Feige's conjecture:
… it combines the exact Dirichlet calibration theorem of Vlassis and Thomas (2026) — itself a resolution of Gaffke's conjecture in statistics — with results from convex geometry, including Grünbaum's centroid theorem from 1960 and its 2024 generalization by Letwin and Yaskin. Read that ingredient list again: a statistics result from this year, a sixty-six-year-old centroid theorem about convex bodies, and a 2024 refinement, welded together to settle a probability conjecture from theoretical computer science. This is proof-by-synthesis — the proof effectively existed, distributed across three literatures that don't read each other, and the contribution was noticing the pieces fit.
As an obligatory comment on the recent Navier-Stokes result, the situation looks a bit messy - for instance, there was a previous human strategy that Terry Tao had previously highlighted as one possibly leading to a solution, but that paper was by itself AI-assisted, and it’s unclear how much of the contribution came from humans and how much from AI. After some discussion, Fable endorsed the following analysis as a fair description of its interpretation of the events:
The strategic idea was human and predates the AI wave; the route was flagged as promising by the field's leading commentator; every rung built on it in the past year involved AI in escalating roles — assistant, then co-author-in-effect, then autonomous builder; and the final rung's machinery was, per the paper's own account, invented by the swarm. "Invention" here means new fusion and execution at least as much as new gadgetry — the components existed; running them together in the smooth setting, with error terms vanishing to all orders at the singular time, did not.
For mathematical objects: Possibly the most significant AI-driven novel construction so far, besides the “final rung’s machinery” mentioned above, is Alpöge and Claude’s complex structure on S^6. This one is totally over my head, but I asked Claude Fable to look for expert discussions of the result, and it said that the construction is based on standard tools, that the novelty mostly comes from the choice of an unusual approach, and that there was a previous paper already suggesting this kind of an approach. So similar to the other math results.
The main question here is how to define whether something is a novel idea or not. I think that ideas that seem very creative or novel are still in some way constructed from ideas or concepts that were available before. (In some trivial sense this has to be true, if there were no connection to any previous thoughts the idea would be incomprehensible.)
You make a useful distinction that I would like to respond to:
“applying a known procedure to create something” versus “inventing the procedure from scratch”.
It seems to me like 'inventing the procedure from scratch' is a potentially misleading way to put it. For example you mention mathematical induction, but is it true that a human actually invented that procedure from scratch? What I understand is that induction as a method of argumentation was already used by the ancient Greeks (and probably others), then mathematicians started using it as well but at first in a more handwavy informal way, then later its usage became rigorous and conceived as a step by step procedure that allowed for a definite conclusion, but there wasn't really a single point where mathematical induction was 'invented from scratch'. I think you could probably tell a similar story about your other examples.
Just to point to one salient example: the highly creative idea of evolution was invented by Darwin and Wallace in connection to reading 'An Essay on the Principle of Population' by Malthus. Similar points could be made about the invention of calculus and a number of other concepts and inventions.
So how should we judge how creative an idea is? I don't have a good answer to that, at least not a formal one. It seems like several dimensions contribute, like inferential distance to prior thought, level of abstraction and generativity (how much new work can be done after generating this new idea). Regarding LLMs, I would say that they are already capable of generating novel ideas and that the degree to which their ideas seem creative and novel on such metrics will probably only go up over time.
Are OpenAI's math results "creative" in an important way?
My gut says yes.
I believe as these results gain more readable expositions, it will become obvious that the models, as applied to math, are creative in a different way to humans.
I predict some people, including some mathematicians, will use these differences to claim "not creative", and I believe this will be cope.
@DaemonicSigil linked https://scottaaronson.blog/?p=10169, which states:
For those who are wondering: apparently, the AI model that produced all these wonders was not bespoke contraption of 10,000 agents burning millions of dollars worth of compute, as was used for example to construct a finite-time blowup for the Navier-Stokes equations. Instead, it was simply the latest internal OpenAI model—one that might be released to paying ChatGPT customers within the next couple of months, depending on the recommendations of OpenAI’s safety board! (My 9-year-old son: “Oh they definitely shouldn’t release that. If it could solve all those math problems, it can’t possibly be safe.”) Apparently they used about 3 hours of GPT-Pro level compute on average per problem solved.
Can someone confirm this? That's kinda surprising to me. (Though "average" is vague about the time for the more impressive ones.)
From https://github.com/openai/math:
The vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model. On average, each result used three hours of ChatGPT Pro thinking compute with that model. Over the course of the evaluation, the model was posed approximately 4,000 problems. Aggregating the output into result families and manuscripts and requiring an appropriate level of significance led to the catalog outlined above.
Exceptions to this fixed procedure include work on a zero-free region for the Riemann zeta function and proof of the Hodge Conjecture for CM abelian varieties. Additionally, the writeup for the Re(s) > 11/12 zero-free region for the Riemann zeta function was human edited for readability.
This would be interesting. (There is discussion of this among mathematicians, e.g. Daniel Litt and Elliot Glazer, but I think more would be better. Terry Tao did some digestion of the Jacobian conjecture counterexample.)
Do any of the proofs have the vibe of AlphaGo's Move 37?
I think AlphaGo's Move 37 logically must have the vibe of AlphaGo's Move 37, and I think is somewhat widely agreed to not be indicative of deep general creativity? I think Navier-Stokes and other things like that are already Move 37 in the sense of "wow we are glimpsing something beyond us, in this area" and the attendant social narrative shifts. The question isn't whether something is somehow above you (obviously current AI is that in many areas!), but as you say, on a level above you.
On guy reacted "examples?" to the "I've seen the demonstrate good creativity and judgment in increasingly many domains."
Examples include my personal eval of designing innovative games (going from "always saying 'time manipulation'" to generating some actual interesting new ideas), and choosing animations to achieve particular emotional effects when I have them make 3D music videos.
Here's a message I sent to X last week recounting Claude getting good creativity and judgement in engineering:
Claude has gotten as smart as a smart human. I was using it as my intern, using it for tedious calculations, shopping for parts, summarizing research, and checking for obvious oversights. But it recently had a genuinely good idea. I've been working on a design for a new cyber-optical gizmo since last November. I've spent hundreds of hours thinking about the design, and I missed a simple and important improvement. But Claude thought of it. It's my own personal move 37.
Do you have any examples of your innovative game design benchmark that you'd be willing to share via link, DM or similar? I'm curious to see what you saw there; I'd be interested even if it's just proof-of-concept snippets.
(I'm also doing similar small evals for myself, and noticed a jump with Fable and now Opus 5.5 showing noticeably better judgement on game design and video animation.)
I've now posted them here:
Thanks! Since I don't have access to any of the fancy models, I was under the impression that most current LLMs, even if they're capable of something, don't have good taste/judgement.
Or, plausibly, the models reasoning traces explain the insights which motivated the proofs, but we don't get to see those.
Not really. They have PDFs for a total of 11 of the problems, and those are only summaries: https://github.com/openai/math/blob/main/reasoning_traces/
I am not a number theorist (I'm a topologist), but it seems like the BSD formula is quite conditional. For instance:
1) the statement needs some "corank 0/1 condition";
2) density 1 is only for the special case of quadratic twists of elliptic curves (and density 1 does not imply full BSD anyways; this statement doesn't treat higher rank ECs at all);
3) in the general case, ChatGPT (5.6 extra-high) says the current results give BSD for 83.75% of elliptic curves ordered by naive height.
Anyways this seems like very significant progress but not close to the full BSD. Notably, talking to some friends in number theory and algebraic geometry, my impression was that the two largest results in these fields are 1) quasi-RH and 2) rational Hodge for CM abelian varieties, and the BSD result is not nearly as significant as these two.
Mathematicians I spoke with today say the proofs have deep insights that they don't understand; the proofs are clear but the understanding that led to the proof is not. The analogy one used was having a proof pick a very specific coefficient in an equation, showing that it works, but giving no explanation of why or how it was found.
But there's still a kind of creativity that a) would have way higher efficiency and b) is maybe necessary for doing some kinds of breakthrough work.
For the sake of argument (as of October 2026), I could believe that there is some special kind of creativity that AIs are still mostly missing. They're smart enough to write sophisticated code, but they do much better with strong tests and metrics. They're smart enough to solve hundreds of famous open problems, but there's a weird grindy quality to some of it. (Or they rely on superhuman knowledge of obscure techniques in other subfields.)
However, none of this reassures me, because of the risks of RSI. We are rapidly reaching a point where we could ask:
Then, once we have a dozen different hypotheses about what's missing, and we have metrics to measure our progress, then we could just fire up a few thousand internal frontier agents and let them try to optimize those metrics.
I strongly suspect that it will take one more paradigm shit to actually build SkyNet/The Blight/your favorite horror. But I fear the current paradigm will soon be enough to aggressively grind enough experiments to find those breakthroughs. However, I don't think the paradigm will go "LLMs → AGI → ASI." I suspect the paradigm will go "LLMs → theoretical breakthrough → broken ASI → engineering polishing →true ASI." [2]
I see two possible paths forward, both worth fighting for:
(I don't hold out much hope for alignment, for reasons I discuss elsewhere. But I really need to write this all up better.)
I have a number of hypotheses about what is missing in current AI, and most of them could potentially be boiled down to metrics an AI could iterate on. No, I will not post specific capabilities-related proposals on LessWrong, even though the labs probably have better proposals than I do. ↩︎
I actually suspect that Fable and Astra are weakly superhuman intelligences by many metrics, but that they fall well below humans in certain critical areas. Once those areas are addressed, I think we'll almost immediately get a superhuman intelligence that's still somewhat broken for prosaic engineering reasons. But at that point, applying the remaining engineering tweaks will be inevitable. My fear is that the earlier breakthrough may be simple, something that could be written up in a nice journal paper or two. ↩︎
To the extent LLMs are limited in what they can do (especially fast learning), they might fail to find one of the key improvements on the path to ASI even in RSI mode (or at least find it quickly, maybe even compared to humanity). And reaching ASI (the kind that can have oceans bloom with physical compute, at doubling times measured in days) probably requires some amount of architecture-rewriting strong RSI progress without such limitations (not just scaling of some breakthrough), after the initial key improvements were all already found, which LLMs aren't necessarily capable of doing quickly (meaning within months/years rather than decades, and without making use of an industrial explosion). So LLM-driven RSI using current methods also doesn't seem like an ASI crux (for the same reasons that LLM creativity doesn't seem to be an ASI crux). It is however a robot-industrial explosion crux (which enables the brute-force path).
The reason it's plausible LLMs will be able to eventually brute-force hill-climb without much creativity all the way to AGIs that don't have more basic cognitive limitations than humans (which then leads to ASI) is that evolution found humans, being local in its search. But this could still take time, LLM-driven RSI or not, because the invent-learn-invent loops are throttled at the crucial step of learning deep skills, which with the current methods proceeds at next-model building speed, not at next-token generation speed.
Yes.
Overwhelmingly so.
This has been apparent far before the recent spate of results. LLMs were already capable of minor mathematical results in late 2024-early 2025 with some minor handholding, where minor is minor in the grand scheme of things but would be pretty impressive for most math grad students.
I'm surprised to see so much questioning about this on this forum [as opposed to the wider world]. There is no secret sauce to being creative. LLMs do it and do it in spades.
If you are still seriously questioning whether LLM can do 'truly creative work' you haven't been updating.
Could you give some examples, ideally pretty concretely, about what creative steps they took in recent proofs? I can't evaluate any of this directly without spending considerable effort, so I've mostly been trying to get a sense of it from reading various mathematician's takes which afaict are pretty mixed. Some of what I find difficult to assess for myself is how much creativity I should assign to things like "the model combined two known techniques in an interesting way," or "it took a direction no one had tried before." I can see the case for that being anywhere from extremely impressive to something more like brute force ++.
I agree there is no secret sauce to creativity in principle, but that doesn't necessarily mean AI systems have it. I also don't think it's obvious that the impressiveness of the results mean the cognition was commensurately impressive. E.g. the results of Deep Blue were clearly impressive, while its way of producing them was far less so. Ofc all else equal the ability to do a ton of problems that are very hard for us is suggestive of using cognition we'd consider impressive, but it's not obvious to me that it necessarily implies this as opposed to to e.g. the relative cognitive advantages LLMs have. But I'd like to know!
This may not be a question that can be answered for a while.
I do a lot of LLM coding where Claude Code can produce impressive artifacts like games quickly and then it takes me a while to understand the code and figure out how it works and appreciate it.
It seems like the situation in math is similar.
As Davidmanheim said in another comment:
Mathematicians I spoke with today say the proofs have deep insights that they don't understand; the proofs are clear but the understanding that led to the proof is not. The analogy one used was having a proof pick a very specific coefficient in an equation, showing that it works, but giving no explanation of why or how it was found.
So I think what will happen is:
Daniel Litt did, in fact, tweet "lmao holy...", in response to the counterexample to the Jacobian conjecture. (Not the original Levent Alpoge tweet, which was optimized for mechanical verfication but a tweet by Andy Jiang who I think replicated it on GPT and read the transcript of how it found the solution.)
It appears to me that lots of people asserting it is slop while refusing to even look at the output.
It still seems there is a strong verifiable/non-verifiable component though, we aren't seeing similar progress in materials science, biology, .
To be clear if you had asked me a year ago I also would have said LLM's won't be capable of solving unsolved mathematics in '26/'27 so I have already been wrong in an important way.
Right now it seems likely that doing huge amounts of RL on a specific verifiable field (such as math) generalizes well to insights within that field. (RLing on hard math problems results in broad math solving ability). But it does not necessarily transfer or translate into success within other fields. I think of these kind of on this axis of possibilities:
I thought we were in world 2, now I think we are in 3.
Why do you think it's only in-field transfer, with no cross-field transfer?
I'm not sure what exactly firms do RL on, so maybe they're just training on ~all fields, but despite capabilities being jagged, I see some general improvements across the board, including in practically less useful areas. To me this seems in line with my expectation of modest cross-field transfer.
(My reason for expecting some cross-field transfer is that success in each field depends not exclusively on highly specialized techniques. From my human experience, I noticed that some of the techniques and ways of thinking I developed during my PhD research were actually usefully transferable for my later work in an unrelated industry job.)
Math may catch up with the rest of science.
"We need a budget to do an experimental run on a particular problem, we're hoping to get a useful result but we may just get crap, plz fund" where instead of a particle accelerator, bioreactor, or combinatorial chemistry plant, it's a blob of GPUs running some model.
Individual work remains important in the sense that one breakthrough by a human could avert the need for six or seven figure LLM runs.
Costs will continue to drop, no?
https://en.wikipedia.org/wiki/Jevons_paradox may come into play.
I guess it's possible that "all of math" is finite, and once a certain level of intelligence is achieved, there will be no more math to solve. Thus, there will be no math problem worth throwing millions of dollars of compute against.
I expect the opposite: the most capable AI systems will identify interesting problems that they cannot solve with their present capability.
Current frontier AIs can solve math problems at a few hundred dollars apiece (Astra, Gemini 3.1 Pro, the solved problems with costs here, etc), make that a few thousand dollars adjusting for failure rates. At ~13x per year cost drop we're on track to this being much more affordable pretty soon. Are you saying these cheap-ish results are low-hanging fruit that will quickly be picked so the average cost per result will keep rising, and that Jevons paradox will make it rise faster than the 13x per year cost drop? If so I don't buy it: there's much more to math than just solving problems; some problems can be arbitrarily scaled up with no ceiling (e.g. Ramsey numbers, busy beaver, etc) and so on. I also suspect there's more math out there (in some interesting sense) than there are computational operations available in the entire universe for cognition to run on (only 1.35×10^120 operations, which isn't really that high if you're used to things like the game-tree complexity of board games like go, to say nothing of googology...).
https://openai.com/index/navier-stokes-solution/ claimed to use 300 billion output tokens. So I can say that more math tasks exist that are both expensive to solve with present tools and worthwhile. I expect this trend to intensify as the frontier advances.
Yes, I think the recent cheap results are low hanging fruit. I think we are in an overhang where there are a lot of cheap and easy things that can (and should) be addressed by the new tools. I do not know how long this overhang lasts, it could absolutely be long enough to provide full careers to people coming out of school today.
I am asserting that it will be possible to do more with more. As the cost of compute comes down, and humans get better at math, mathematicians will hopefully identify challenges that merit massive investments of resources.
I view that as an improvement on the status quo, 'big science' in the form of supercomputers, space telescopes and particle accelerators has delivered a lot, at high cost. I hope math, as a field, is gaining a tool where it is possible to scale in with massive investment to obtain massive rewards.
The irrationality exponent ofis 2 (same as most other real numbers).
This and other errors, looks like none of the math symbols in the quoted comment were copied properly.
Many people expect the AI paradigm to approximately AGI complete, and expect it to smoothly scale to AI that can do innovative science. i.e. the kind of science you'd need to do to develop a plague from scratch that actually kills all humans, with very limited opportunity to experiment. Or, the kind of science you'd need to solve alignment, so you could build a more powerful successor AI that shares your values.
A periodic thread of disagreement has been @TsviBT, @Steven Byrnes and some others arguing that, yes, the AIs are getting better at verifiable tasks, but they still seem to be missing some basic sauce that lets them actually form new concepts on the fly and apply them.
The AIs are getting better at "brute force creativity", where you try every single idea ever devised by Man in parallel and maybe one of them works. But there's still a kind of creativity that a) would have way higher efficiency and b) is maybe necessary for doing some kinds of breakthrough work.
On one hand, my subjective experience has been seeing AIs be able to demonstrate increasingly interesting judgement and taste in increasingly many domains (even open-ended ones). But, I do indeed still run into AIs that get confused and "just don't get it", in a way that is pretty suggestive that there is still something significant missing.
OpenAI has lately been shipping some major results in mathematics, for problems that people had previously agreed were important and hard. But, I'd heard for previous results that the proofs turned out sort of "not actually interesting", compared to how sometimes makes you go "Holy shit that's beautiful/elegant/surprising" and conveys someone is a level above you."
Recently they shipped a lot of new proofs on important problems.
I'm not a math guy. But, seems fairly important for some people with enough context to wade through them and figure out "What sort of cognition was involved here?". Do any of the proofs have the vibe of AlphaGo's Move 37? Does it invent a new concept that does real work?
I'm guessing answering this is a nontrivial task that'll involve some... archaeology? Might be cool for a competent math theorist who doesn't have a higher priority project to organize the labor. (I'm guessing the broader math community will do this on their own, but not necessarily organize the information in a way that's optimized for figuring out "is this the kind of mind that could solve its own alignment or quickly develop radically more powerful weapons than we have available?").
I'm shamelessly copying in @DaemonicSigil's quick take about this, to give some hooks for people to start thinking through.