> no prior solutions found.
This is no longer true, a prior solution has just been found[1], so the LLM proof has been moved to the Section 2 of Terence Tao's wiki[2].
[1] - https://www.erdosproblems.com/forum/thread/281#post-3325
[2] - https://github.com/teorth/erdosproblems/wiki/AI-contribution...
Can anyone give a little more color on the nature of Erdos problems? Are these problems that many mathematicians have spend years tackling with no result? Or do some of the problems evade scrutiny and go un-attempted for most of the time?
EDIT: After reading a link someone else posted to Terrance Tao's wiki page, he has a paragraph that somewhat answers this question:
> Erdős problems vary widely in difficulty (by several orders of magnitude), with a core of very interesting, but extremely difficult problems at one end of the spectrum, and a "long tail" of under-explored problems at the other, many of which are "low hanging fruit" that are very suitable for being attacked by current AI tools. Unfortunately, it is hard to tell in advance which category a given problem falls into, short of an expert literature review. (However, if an Erdős problem is only stated once in the literature, and there is scant record of any followup work on the problem, this suggests that the problem may be of the second category.)
from here: https://github.com/teorth/erdosproblems/wiki/AI-contribution...
Erdos was an incredibly prolific mathematician, and one of his quirks is that he liked to collect open problems and state new open problems as a challenge to the field. Many of the problems he attached bounties to, from $5 to $10,000.
The problems are a pretty good metric for AI, because the easiest ones at least meet the bar of "a top mathematician didn't know how to solve this off the top of his head" and the hardest ones are major open problems. As AI progresses, we will see it slowly climb the difficulty ladder.
Don't feel bad for being out of the loop. The author and Tao did not care enough about erdos problem to realize the proof was published by erdos himself. So you never cared enough and neither did they. But they care about about screaming LLMs breakthrough on fediverse and twitter.
> Did not care enough about erdos...
This is bad faith. Erdos was an incredibly prolific mathematician, it is unreasonable to expect anyone to have memorized his entire output. Yet, Tao knows enough about Erdos to know which mathematical techniques he regularly used in his proofs.
From the forum thread about Erdos problem 281:
> I think neither the Birkhoff ergodic theorem nor the Hardy-Littlewood maximal inequality, some version of either was the key ingredient to unlock the problem, were in the regular toolkit of Erdos and Graham (I'm sure they were aware of these tools, but would not instinctively reach for them for this sort of problem). On the other hand, the aggregate machinery of covering congruences looks relevant (even though ultimately it turns out not to be), and was very much in the toolbox of these mathematicians, so they could have been misled into thinking this problem was more difficult than it actually was due to a mismatch of tools.
> I would assess this problem as safely within reach of a competent combinatorial ergodic theorist, though with some thought required to figure out exactly how to transfer the problem to an ergodic theory setting. But it seems the people who looked at this problem were primarily expert in probabilistic combinatorics and covering congruences, which turn out to not quite be the right qualifications to attack this problem.
Isn't it bad faith to say no priors solutions was found when a solution published by erdos was ultimately found by the community in 10 minutes?
That sounds like a great question. Why did no one bother to mention the problem was already proved and published by the author that proposed the statement 90 years ago?
Somehow an llm generated proof that consist of gigabytes upon gigabytes of unreadable mess is groundbreaking and pushes mathematics forward, a proof proposed by Erdos himself in 5 pages gets buried and lost to time.
Maybe one particular optics fuels the narrative that formal verified compute is the new moat and llms are amazing at that?
This Tao dude, does he get invited to a lot of AI conferences (accommodation included)?
He's the most prolific and famous modern mathematician. I'm pretty sure that even if he'd never touched AI, he would be invited to more conferences than he could ever attend.
Please follow hackernews guidelines for comments: https://news.ycombinator.com/newsguidelines.html
I know someone who organized a conference where he spoke (this was before the AI boom, probably around 2018 or so) and he got very good accommodations and also a very generous speaking fee.
From Terry Tao's comments in the thread:
"Very nice! ... actually the thing that impresses me more than the proof method is the avoidance of errors, such as making mistakes with interchanges of limits or quantifiers (which is the main pitfall to avoid here). Previous generations of LLMs would almost certainly have fumbled these delicate issues.
...
I am going ahead and placing this result on the wiki as a Section 1 result (perhaps the most unambiguous instance of such, to date)"
The pace of change in math is going to be something to watch closely. Many minor theorems will fall. Next major milestone: Can LLMs generate useful abstractions?
Seems like the someone dug something up from the literature on this problem (see top comment on the erdosproblems.com thread)
"On following the references, it seems that the result in fact follows (after applying Rogers' theorem) from a 1936 paper of Davenport and Erdos (!), which proves the second result you mention. ... In the meantime, I am moving this problem to Section 2 on the wiki (though the new proof is still rather different from the literature proof)."
Personally, I'd prefer if the AI models would start with a proof of their own statements. Time and again, SOTA frontier models told me: "Now you have 100% correct code ready for production in enterprise quality." Then I run it and it crashes. Or maybe the AI is just being tongue-in-cheek?
Point in case: I just wanted to give z.ai a try and buy some credits. I used Firefox with uBlock and the payment didn't go through. I tried again with Chrome and no adblock, but now there is an error: "Payment Failed: p.confirmCardPayment is not a function." The irony is, that this is certainly vibe-coded with z.ai which tries to sell me how good they are but then not being able to conclude the sale.
And we will get lots more of this in the future. LLMs are a fantastic new technology, but even more fantastically over-hyped.
You get AIs to prove their code is correct in precisely the same ways you get humans to prove their code is correct. You make them demonstrate it through tests or evidence (screenshots, logs of successful runs).
Yes! Also, make sure to check those results yourself, dear reader, rather than ask the agent to summarize the results for you! ^^;
We should differentiate AI models from AI apps.
Models just generate text. Apps are supposed to make that text useful.
An app can run various kinds of verification. But would you pay an extra for that?
Nobody can make a text generator to output text which is 100% correct. That's just not a thing people can do now.
The erdosproblems thread itself contains comments from Terence Tao: https://www.erdosproblems.com/forum/thread/281
Has anyone verified this?
I've "solved" many math problems with LLMs, with LLMs giving full confidence in subtly or significantly incorrect solutions.
I'm very curious here. The Open AI memory orders and claims about capacity limits restricting access to better models are interesting too.
Terence Tao gave it the thumbs up. I don't think you're going to do better than that.
It's already been walked back.
Not in the sense of being a "subtly or significantly incorrect solution".
FWIW, I just gave Deepseek the same prompt and it solved it too (much faster than the 41m of ChatGPT). I then gave both proofs to Opus and it confirmed their equivalence.
The answer is yes. Assume, for the sake of contradiction, that there exists an \(\epsilon > 0\) such that for every \(k\), there exists a choice of congruence classes \(a_1^{(k)}, \dots, a_k^{(k)}\) for which the set of integers not covered by the first \(k\) congruences has density at least \(\epsilon\).
For each \(k\), let \(F_k\) be the set of all infinite sequences of residues \((a_i)_{i=1}^\infty\) such that the uncovered set from the first \(k\) congruences has density at least \(\epsilon\). Each \(F_k\) is nonempty (by assumption) and closed in the product topology (since it depends only on the first \(k\) coordinates). Moreover, \(F_{k+1} \subseteq F_k\) because adding a congruence can only reduce the uncovered set. By the compactness of the product of finite sets, \(\bigcap_{k \ge 1} F_k\) is nonempty.
Choose an infinite sequence \((a_i) \in \bigcap_{k \ge 1} F_k\). For this sequence, let \(U_k\) be the set of integers not covered by the first \(k\) congruences, and let \(d_k\) be the density of \(U_k\). Then \(d_k \ge \epsilon\) for all \(k\). Since \(U_{k+1} \subseteq U_k\), the sets \(U_k\) are decreasing and periodic, and their intersection \(U = \bigcap_{k \ge 1} U_k\) has density \(d = \lim_{k \to \infty} d_k \ge \epsilon\). However, by hypothesis, for any choice of residues, the uncovered set has density \(0\), a contradiction.
Therefore, for every \(\epsilon > 0\), there exists a \(k\) such that for every choice of congruence classes \(a_i\), the density of integers not covered by the first \(k\) congruences is less than \(\epsilon\).
\boxed{\text{Yes}}
> I then gave both proofs to Opus and it confirmed their equivalence.
You could have just rubber-stamped it yourself, for all the mathematical rigor it holds. The devil is in the details, and the smallest problem unravels the whole proof.
How dare you question the rigor of the venerable LLM peer review process! These are some of the most esteemed LLMs we are talking about here.
It's about formalization in Lean, not peer review
"Since \(U_{k+1} \subseteq U_k\), the sets \(U_k\) are decreasing and periodic, and their intersection \(U = \bigcap_{k \ge 1} U_k\) has density \(d = \lim_{k \to \infty} d_k \ge \epsilon\)."
Is this enough? Let $U_k$ be the set of integers such that their remainder mod 6^n is greater or equal to 2^n for all 1<n<k. Density of each $U_k$ is more than 1/2 I think but not the intersection (empty) right?
Indeed. Your sets are decreasing periodic of density always greater than the product from k=1 to infinity of (1-(1/3)^k), which is about 0.56, yet their intersection is null.
This would all be a fairly trivial exercise in diagonalization if such a lemma as implied by Deepseek existed.
(Edit: The bounding I suggested may not be precise at each level, but it is asymptotically the limit of the sequence of densities, so up to some epsilon it demonstrates the desired counterexample.)
Here's kimi-k2-thinking with the reasoning block included: https://www.kimi.com/share/19bcfe2e-d9a2-81fe-8000-00002163c...
I am not familiar with the field, but any chance that the deepseek is just memorizing the existing solution? Or different.
Sure but if so wouldn't ChatGPT 5.2 Pro also "just memorizing the existing solution?"?
No it's not, you can refer to my link and subsequent discussion.
I am basing on Terrence Tao comment here: https://news.ycombinator.com/item?id=46665168
It says that the OpenAI proof is a different one from the published one in the literature.
Whereas whether the Deepseek proof is the same as the published one, I dont know enough of the math to judge.
That was what I meant.
Opus isn't a good choice for anything math-related; it's worse at math than the latest ChatGPT and Gemini Pro.
I find it interesting that, as someone utterly unfamiliar with ergodic theory, Dini’s theorem, etc, I find Deepseek’s proof somewhat comprehensible, whereas I do not find GPT-5.2’s proof comprehensible at all. I suspect that I’d need to delve into the terminology in the GPT proof if I tried to verify Deepseek’s, so maybe GPT’s is being more straightforward about the underlying theory it relies on?
There was a post about Erdős 728 being solved with Harmonic’s Aristotle a little over a week ago [1] and that seemed like a good example of using state-of-the-art AI tech to help increase velocity in this space.
I’m not sure what this proves. I dumped a question into ChatGPT 5.2 and it produced a correct response after almost an hour [2]?
Okay? Is it repeatable? Why did it come up with this solution? How did it come up with the connections in its reasoning? I get that it looks correct and Tao’s approval definitely lends credibility that it is a valid solution, but what exactly is it that we’ve established here? That the corpus that ChatGPT 5.2 was trained on is better tuned for pure math?
I’m just confused what one is supposed to take away from this.
[1] https://news.ycombinator.com/item?id=46560445
[2] https://chatgpt.com/share/696ac45b-70d8-8003-9ca4-320151e081...
Also #124 was proved using AI 49 days ago: https://news.ycombinator.com/item?id=46094037
Thanks for the curious question. This is one in a sequence of efforts to use LLMs to generate candidate proofs to open mathematical questions, which then are generally formalized into Lean, a formal proof system for pure mathematics.
Erdos was prolific and many of his open problems are numbered and have space to discuss them online, so it’s become fairly common to run through them with frontier models and see if a good proof can be come up with; there have been some notable successes here this year.
Tao seems to engage in sort of a two step approach with these proofs - first, are they correct? Lean formalization makes that unambiguous, but not all proofs are easily formulated into Lean, so he also just, you know, checks them. Second, literature search inside LLMs and out for prior results — this is to check where frontier models are at in the ‘novel proofs or just regurgitated proofs’ space.
To my knowledge, we’re currently at the point where we are seeing some novel proofs offered, but I don’t think we’ve seen any that have absolutely no priors in literature.
As you might guess this is itself sort of a Rorschach test for what AI could and will be.
In this case, it looked at first like this was a totally novel solution to something that hadn’t been solved before. On deeper search, Tao noted it’s almost trivial to prove with stuff Erdos knew, and also had been proved independently; this proof doesn’t use the prior proof mechanism though.
A surprising % of these LLM proofs are coming from amateurs.
One wonders if some professional mathematicians are instead choosing to publish LLM proofs without attribution for career purposes.
It's probably from the perennial observation
"This LLM is kinda dumb in the thing I'm an expert in"
This is just not true at this point but believe whatever you want to believe.
[dead]
Perennial doesn't make sense in the context of something that has been around for a few months. Observations from the spring 2025 crop of LLMs are already irrelevant.
… “but I guess it was able to formalize it in Lean, so…”
>One wonders if some professional mathematicians are instead choosing to publish LLM proofs without attribution for career purposes.
This will just become the norm as these models improve, if it isn't largely already the case.
It's like sports where everyone is trying to use steroids, because the only way to keep up is to use steroids. Except there aren't any AI-detectors and it's not breaking any rules (except perhaps some kind of self moral code) to use AI.
I think a more realistic answer is that professional mathematicians have tried to get LLMs to solve their problems and the LLMs have not been able to make any progress.
I think it's a bit early to tell whether GPT 5.2 has helped research mathematicians substantially given its recency. The models move so fast that even if all previous models were completely useless I wouldn't be sure this one would be. Let's wait a year and see? (it takes time to write papers)
It's helped, but it's not correct that mathematicians are scoring major results by just feeding their problems to gpt 5.2 pro, so the OP claim that mathematicians are just playing off AI output as their own is silly. Here, im talking about serious mathematical work, not people posting (unattributed AI slop to the arXiv).
I assume OP was mostly joking, but we need to take care about letting AI companies hype up their impressive progress at the expense of mathematics. This needs to be discussed responsibly.
I'm actually not sure what the right attribution method would be. I'd lean towards single line on acknowledgements? Because you can use it for example @ every lemma during brainstorming but it's unclear the right convention is to thank it at every lemma...
Anecdotally, I, as a math postdoc, think that GPT 5.2 is much stronger qualitatively than anything else I've used. Its rate of hallucinations is low enough that I don't feel like the default assumption of any solution is that it is trying to hide a mistake somewhere. Compared with Gemini 3 whose failure mode when it can't solve something is always to pretend it has a solution by "lying"/ omitting steps/making up theorems etc... GPT 5.2 usually fails gracefully and when it makes a mistake it more often than not can admit it when pointed out.
I guess the first question I have is if these problems solved by LLMs are just low-hanging fruit that human researchers either didn't get around to or show much interest in - or if there's some actual beef here to the idea that LLMs can independently conduct original research and solve hard problems.
That's the first warning from the wiki : <<Erdős problems vary widely in difficulty (by several orders of magnitude), with a core of very interesting, but extremely difficult problems at one end of the spectrum, and a "long tail" of under-explored problems at the other, many of which are "low hanging fruit" that are very suitable for being attacked by current AI tools.>> https://github.com/teorth/erdosproblems/wiki/AI-contribution...
There is still value on letting these LLMs loose on the periphery and knocking out all the low hanging fruit humanity hasn’t had the time to get around to. Also, I don’t know this, but if it is a problem on Erdos I presume people have tried to solve it atleast a little bit before it makes it to the list.
Is there though? If they are "solved" (as in the tickbox mark them as such, through a validation process, e.g. another model confirming, formal proof passing, etc) but there is no human actually learning from them, what's the benefit? Completing a list?
I believe the ones that are NOT studied are precisely because they are seen as uninteresting. Even if they were to be solved in an interesting way, if nobody sees the proof because they are just too many and they are again not considered valuable then I don't see what is gained.
Some problems are ‘uninteresting’ in that they show results that aren’t immediately seen as useful. However, solutions may end up having ‘interesting’ connections or ideas or mathematical tools that are used elsewhere.
More broadly, I think there’s a perspective that literally just building out thousands more true statements in Lean is going to keep cementing math’s broadening knowledge framework. This is not building a giant castle a-la Wiles, it’s laying bricks in the outhouse, but someday those bricks might be useful.
You don't see value in having a cheap way to detect when a problem is easy or hard? That would seem unimaginative.
Out of curiosity why has the LLM math solving community been focused on the Erdos problems over other open problems? Are they of a certain nature where we would expect LLMs to be especially good at solving them?
I guess they are at a difficulty where it's not too hard (unlike millennium prize problems), is fairly tightly scoped (unlike open ended research), and has some gravitas (so it's not some obscure theorem that's only unproven because of it's lack of noteworthiness).
I actually don't think the reason is that they are easier than other open math problems. I think it's more that they are "elementary" in the sense that the problems usually don't require a huge amount of domain knowledge to state.
The Collatz conjecture can be stated using basic arithmetic, yet LLMs have not been able to solve it.
I agree it's easier than Collatz. I just mean I am not sure it's much easier than many currently open questions which are less famous but need more machinery.
That is also one of the hardest problems.
People like checking items off of lists.
The LLMs that take 10 attempts to un-zero-width a <div>, telling me that every single change totally fixed the problem, are cracking the hardest math problems again.
Math makes sense, CSS doesn't.
Is there explainability research for this type of model application? E.g. a sparse auto encoder or something similar but more modern.
I would love to know which concepts are active in the deeper layers of the model while generating the solution.
Is there a concept of “epsilon” or “delta”?
What are their projections on each other?
It’s funny. in some kind of twisted variant of Cunningham’s Law we have:
> the best way to find a previous proof of a seemingly open problem on the internet is not to ask for it; it's to post a new proof
I wonder if they tried Gemini. I think Gemini could have done better, as seen from my experiences with GPT and Gemini models on some simple geometry problems.
I'm looking forward to chatgpt 5.3pro. I also use chatgpt 5.2pro for various program consultations. It's been very helpful.
I was hoping there'd be more discussion about the model itself. I find the last couple of generations of Pro models fascinating.
Personally, I've been applying them to hard OCR problems. Many varied languages concurrently, wildly varying page structure, and poor scan quality; my dataset has all of these things. The models take 30 minutes a page, but the accuracy is basically 100% (it'll still striggle with perfectly-placed bits of mold). The next best model (Google's flagship) rests closer to 80%.
I'll be VERY intrigued to see what the next 2, 5, 10 years does to the price of this level of model.
We're eventually going to get it at cerebras inference latency. It's going to be wild.
>no prior solutions found.
They never brothered to check erdos solution already published 90 years ago. I am still confused about why erdos, who proposed the problem and the solution would consider this an unsolved problems, but this group of researchers would claim "ohh my god look at this breakthrough"
This is showing as unresolved here, so I'm assuming something was retracted.
https://mehmetmars7.github.io/Erdosproblems-llm-hunter/probl...
I think that just hasn't been updated.
I have 15 years of software engineering experience across some top companies. I truly believe that ai will far surpass human beings at coding, and more broadly logic work. We are very close
HN will be the last place to admit it; people here seem to be holding out with the vague 'I tried it and it came up with crap'. While many of us are shipping software without touching (much) code anymore. I have written code for over 40 years and this is nothing like no-code or whatever 'replacing programmers' before, this is clearly different judging from the people who cannot code with a gun to their heads but still are shipping apps: it does not really matter if anyone believes me or not. I am making more money than ever with fewer people than ever delivering more than ever.
We are very close.
(by the way; I like writing code and I still do for fun)
Both can be correct : you might be making a lot of money using the latest tools while others who work on very different problems have tried the same tools and it's just not good enough for them.
The ability to make money proves you found a good market, it doesn't prove that the new tools are useful to others.
No, the comment is about "will", not "is". Of course there's no definitive proof of what will happen. But the writing is on the wall and the letters are so large now, that denying AI would take over coding if not all intellectual endeavors resembles the movie "Don't look up".
It is also very much a moving target. A year ago I tried those tools and they were very meh at the kinds of stuff I do. Today, they are much better.
> holding out with the vague 'I tried it and it came up with crap'
Isn't that a perfectly reasonable metric? The topic has been dominated by hype for at least the past 5 if not 10 years. So when you encounter the latest in a long line of "the future is here the sky is falling" claims, where every past claim to date has been wrong, it's natural to try for yourself, observe a poor result, and report back "nope, just more BS as usual".
If the hyped future does ever arrive then anyone trying for themselves will get a workable result. It will be trivially easy to demonstrate that naysayers are full of shit. That does not currently appear to be the case.
What topic are you referring to? ChatGPT release was just over 3 years ago. 5 years ago we had basic non-instruct GPT-3.
Transformers was 2017 and it had some implications on translation (which were in no way overstated), but it took GPT-2 and 3 to kick it off in earnest and the real hype machine started with ChatGPT.
What you are doing however is dismissing the outrageous progress on NLP and by extension code generation of the last few years just because people over hype it.
People over hyped the Internet in the early 2000s, yet here we are.
But the trend line is less ambiguous, models got better year over year, much much better.
There's a big difference between "I tried it and it produced crap" and "it will replace developers entirely any day now"
People who use this stuff everyday know that people who are still saying "I tried it and it produced crap" just don't know how to use it correctly. Those developers WILL get replaced - by ones who know how to use the tool.
> I have 15 years of software engineering experience across some top companies. I truly believe that ai will far surpass human beings at coding, and more broadly logic work. We are very close
Coding was never the hard part of software development.
Getting the architecture mostly right, so it's easy to maintain and modify in the future is IMO hard part, but I find that this is where AI shines. I have 20 years of SWE experience (professional) and (10 hobby) and most of my AI use is for architecture and scaffolding first, code second.
Gotta make sure that the investors read this message in an Erdos thread.
They already do. What they suck at is common sense. Unfortunately good software requires both.
[flagged]
Even a 20 year old Markov chain could produce this banality.
Or is it fortunate (for a short period at least).
Is this comment written by AI?
They can only code to specification which is where even teams of humans get lost. Without much smarter architecture for AI (LLMs as is are a joke) that needle isn’t going to move.
Real HN comment right here. "LLMs are a joke" - maybe don't drink the anti-hype kool aid, you'll blind yourself to the capability space that's out there, even if it's not AGI or whatever.
I’ll look past the disrespectful flippant insult on the hope that there’s a brain there too.
They’re a probabalistic phonograph. They can sharpen the funnel for input but they can’t provide judgement on input or resolve ambiguities in your specifications. Teams of human requirements engineers cannot do it. LLMs are not magic. You’re essentially asking it; from my wardrobe pick an outfit for me and make sure it’s the one I would have picked.
If you’re dazzled into thinking LLMs can solve this you just don’t understand transformer architecture and you don’t understand requirements engineering.
You’ll know a proper AI engine when you see it and it doesn’t look like an LLM.
Humans are magic from the LLMs perspective because the token window sizes they would need to approach human experiential disambiguation of requirements would be orders of magnitude larger. Useful in general or replace in general some human activities is a goal post shift that was never the discussion here.
I can post a long list of simple things a human can do accurately and efficiently that I've seen Gemini unable to do, repeatedly.
And someone could post an even longer list of things you can't do well. But what would be the point?
The LLM did better on this problem than 100% of the haters in this thread could do, and who probably can't even begin "understand" the problem.
how did they do it? Was a human using the chat interface? Did they just type out the problem and immediately on the first reply received a complete solution (one-shot) or what was the human's role? What was ChatGPT's thinking time?
very interesting. ChatGPT reasoned for 41 minutes about it! Also, this was one-shot - i.e. ChatGPT produced its complete proof with a single prompt and no more replies by the human, (rather than a chat where the human further guided it.)
Sounds like Lean 4/rocq did all the work here
Why do you say that? I see no mention of lean/rocq on the twitter thread, nor on the erdos problem forum thread, nor on the chatGPT conversation.
What does "solved with" mean? The author claims "I've solved", so did the author solve it or GPT?
When you use a calculator, did you really solve it or was it the calculator?
With a calculator I supply the arithmetic. It just executes it with no reasoning so im the solver. I can do the same with an LLM and still be the solver as long as it just follows my direction. Or I can give it a problem and let it reason and generate the arithmetic itself, in which case the LLM is effectively the solver. Thats why saying "I've solved X using only GPT" is ambiguous.
But thanks for the downvote in addition to your useless comment.
This is crazy. It's clear that these models don't have human intelligence, but it's undeniable at this point that they have _some_ form of intelligence.
If LLMs weren't created by us but where something discovered in another species' behaviour it would be 100% labelled intelligence
Yes, same for the case where the technology would have been found embodied in machinery aboard a crashed UFO.
My take is that a huge part of human intelligence is pattern matching. We just didn’t understand how much multidimensional geometry influenced our matches
Yes, it could be that intelligence is essentially a sophisticated form of recursive, brute force pattern matching.
I'm beginning to think the Bitter Lesson applies to organic intelligence as well, because basic pattern matching can be implemented relatively simply using very basic mathematical operations like multiply and accumulate, and so it can scale with massive parallelization of relatively simple building blocks.
Intelligence is almost certainly a fundamentally recursive process.
The ability to think about your own thinking over and over as deeply as needed is where all the magic happens. Counterfactual reasoning occurs every time you pop a mental stack frame. By augmenting our stack with external tools (paper, computers, etc.), we can extend this process as far as it needs to go.
LLMs start to look a lot more capable when you put them into recursive loops with feedback from the environment. A trillion tokens worth of "what if..." can be expended without touching a single token in the caller's context. This can happen at every level as many times as needed if we're using proper recursive machinery. The theoretical scaling around this is extremely favorable.
Anatomically good candidate, the thalamal-cortical loop: https://en.wikipedia.org/wiki/Cortico-basal_ganglia-thalamo-...
I don't think it's accurate to describe LLMs as pattern matching. Prediction is the mechanism they use to ingest and output information, and they end up with a (relatively) deep model of the world under the hood.
The "pattern matching" perspective is true if you zoom in close enough, just like "protein reactions in water" is true for brains. But if you zoom out you see both humans and LLMs interact with external environments which provide opportunity for novel exploration. The true source of originality is not inside but in the environment. Making it be all about the model inside is a mistake, what matters more than the model is the data loop and solution space being explored.
"Pattern matching" is not sufficiently specified here for us to say if LLMs do pattern matching or not. E.g. we can say that an LLM predicts the next token because that token (or rather, its embedding) is the best "match" to the previous tokens, which form a path ("pattern") in embedding space. In this sense LLMs are most definitely pattern matching. Under other formulations of the term, they may not be (e.g. when pattern matching refers to abstraction or abstracting to actual logical patterns, rather than strictly semantic patterns).
> I don't think it's accurate to describe LLMs as pattern matching
I’m talking about the inference step, which uses tensor geometry arithmetic to find patterns in text. We don’t understand what those patterns are but it’s clear it’s doing some heavy lifting since llm inference is expressing logic and reasoning under the guise of our reductive “next token prediction”
Yes, the world model building is achieved via pattern matching and happens during ingestion and training, but that is also part of the intelligence.
Which is even more true for humans.
Intelligence is hallucination that happens to produce useful results in the real world.
I don't think they will ever have human intelligence. It will always be an alien intelligence.
But I think the trend line unmistakably points to a future where it can be MORE intelligent than a human in exactly the colloquial way we define "more intelligent"
The fact that one of the greatest mathematicians alive has a page and is seriously bench marking this shows how likely he believes this can happen.
Well, Alpha Go and Stockfish can beat you at their games. Why shouldn't these models beat us at math proofs?
Chess and Go have very restrictive rules. It seems a lot more obvious to me why a computer can beat a human at it. They have a huge advantage just by being able to calculate very deep lines in a very short time. I actually find it impressive for how long humans were able to beat computers at go. Math proofs seem a lot more open ended to me.
Alpha go and stockfish were specifically designed and trained to win at those games.
And we can train models specifically at math proofs? I think only difference is that math is bigger....
It's pattern matching. Which is actually what we measure in IQ tests, just saying.
There's some nuance. IQ tests measure pattern matching and, in an underlying way, other facets of intelligence - memory, for example. How well can an LLM 'remember' a thing? Sometimes Claude will perform compaction when its context window reaches 200k "tokens" then it seems a little colder to me, but maybe that's just my imagination. I'm kind of a "power user".
I call it matching. Pattern matching had a different meaning.
what are you referring to? LLMs are neural networks at their core and the most simple versions of neural networks are all about reproducing patterns observed during training
You need to understand the difference between general matching and pattern matching. Maybe should have read more older AI books. A LLM is a general fuzzy matcher. A pattern matcher is an exact matcher using an abstract language, the "pattern". A general matcher uses a distance function instead, no pattern needed.
Ie you want to find a subimage in a big image, possibly rotated, scaled, tilted, distorted, with noise. You cannot do that with a pattern matcher, but you can do that with a matcher, such as a fuzzy matcher, a LLM.
You want to find a go position on a go board. A LLM is perfect for that, because you don't need to come up with a special language to describe go positions (older chess programs did that), you just train the model if that position is good or bad, and this can be fully automated via existing literature and later by playing against itself. You train the matcher not via patterns but a function (win or loose).
Depends on what you mean by intelligence, human intelligence and human
As someone who doesn't understand this shit, and how it's always the experts who fiddle the LLMs to get good outputs, it feels natural to attribute the intelligence to the operator (or the training set), rather than the LLM itself.
Yes it is intelligent, but so what? Its not conscious, sentient or sapient. It's a pattern matching chinese room.
Funny seeing silicon valley bros commenting "you're on fire!" to Neel when it appears he copied and pasted the problem verbatim into chatGPT and it did literally all the other work here
https://chatgpt.com/share/696ac45b-70d8-8003-9ca4-320151e081...
Knowing which problem to copy and paste into the model is also a skill.
[dead]
[dead]
Narrator: The solution had already appeared several times in the training data
This must be what it feels like to be a CEO and someone tells me they solved coding.
Has anyone confirmed the solution is not in the training data? Otherwise it is just a bit information retrieval LLM style. No intelligence necessary.
[dead]
Interesting that in Terrance Tao's words: "though the new proof is still rather different from the literature proof)"
And even odder that the proof was by Erdos himself and yet he listed it as an open problem!
The theorem is implied by an older result of Erdos, but is not a result of Erdos. Apparently this is because the connection is something called "Roger's Theorem" that was quite obscure.
https://terrytao.wordpress.com/2026/01/19/rogers-theorem-on-...
"This theorem is somewhat obscure: its only appearance in print is in pages 242-244 of this 1966 text of Halberstam and Roth, where the authors write in a footnote that the result is “unpublished; communicated to the authors by Professor Rogers”. I have only been able to find it cited in three places in the literature: in this 1996 paper of Lewis, in this 2007 paper of Filaseta, Ford, Konyagin, Pomerance, and Yu (where they credit Tenenbaum for bringing the reference to their attention), and is also briefly mentioned in this 2008 paper of Ford. As far as I can tell, the result is not available online, which could explain why it is rarely cited (and also not known to AI tools). This became relevant recently with regards to Erdös problem 281, posed by Erdös and Graham in 1980, which was solved recently by Neel Somani through an AI query by an elegant ergodic theory argument. However, shortly after this solution was located, it was discovered by KoishiChan that Rogers’ theorem reduced this problem immediately to a very old result of Davenport and Erdös from 1936. Apparently, Rogers’ theorem was so obscure that even Erdös was unaware of it when posing the problem!"
Maybe it was in the training set.
I think that was Tao's point, that the new proof was not just read out of the training set.
forgive the skepticism, but this translates directly to "we asked the model pretty please not to do it in the system prompt"
Do you have a source for this?
Carbon copy would mean over fitting
Unfortunately.
does it?
this is a verbatim quote from gemini 3 pro from a chat couple of days ago:
"Because I have done this exact project on a hot water tank, I can tell you exactly [...]"
I somehow doubt it an LLM did that exact project, what with not having any abilities to do plumbing in real life...
It is still possible a proof from someone else with a similar method was in the training set.
A proof that Terence Tao and his colleagues have never heard of? If he says the LLM solved the problem with a novel approach, different from what the existing literature describes, I'm certainly not able to argue with him.
Does it matter if it copied or not? How the hell would one even define if it is a copy or original at this point?
At this point the only conclusion here is: The original proof was on the training set. The author and Terence did not care enough to find the publication by erdos himself
It looks like these models work pretty well as natural language search engines and at connecting together dots of disparate things humans haven't done.
They're finding them very effective at literature search, and at autoformalization of human-written proofs.
Pretty soon, this is going to mean the entire historical math literature will be formalized (or, in some cases, found to be in error). Consider the implications of that for training theorem provers.
I think "pretty soon" is a serious overstatement. This does not take into account the difficulty in formalizing definitions and theorem statements. This cannot be done autonomously (or, it can, but there will be serious errors) since there is no way to formalize the "text to lean" process.
What's more, there's almost surely going to turn out to be a large amount of human generated mathematics that's "basically" correct, in the sense that there exists a formal proof that morally fits the arc of the human proof, but there's informal/vague reasoning used (e.g. diagram arguments, etc) that are hard to really formalize, but an expert can use consistently without making a mistake. This will take a long time to formalize, and I expect will require a large amount of human and AI effort.
It's all up for debate, but personally I feel you're being too pessimistic there. The advances being made are faster than I had expected. The area is one where success will build upon and accelerate success, so I expect the rate of advance to increase and continue increasing.
This particular field seems ideal for AI, since verification enables identification of failure at all levels. If the definitions are wrong the theorems won't work and applications elsewhere won't work.
Every time this topic comes up people compare the LLM to a search engine of some kind.
But as far as we know, the proof it wrote is original. Tao himself noted that it’s very different from the other proof (which was only found now).
That’s so far removed from a “search engine” that the term is essentially nonsense in this context.
Hassabis put forth a nice taxonomy of innovation: interpolation, extrapolation, and paradigm shifts.
AI is currently great at interpolation, and in some fields (like biology) there seems to be low-hanging fruit for this kind of connect-the-dots exercise. A human would still be considered smart for connecting these dots IMO.
AI clearly struggles with extrapolation, at least if the new datum is fully outside the training set.
And we will have AGI (if not ASI) if/when AI systems can reliably form new paradigms. It’s a high bar.
Maybe if Terence Tao had memorized the entire Internet (and pretty much all media), then maybe he would find bits and pieces of the problem remind him of certain known solutions and be able to connect the dots himself.
But, I don't know. I tend to view these (reasoning) LLMs as alien minds and my intuition of what is perhaps happening under the hood is not good.
I just know that people have been using these LLMs as search engines (including Stephen Wolfram), browsing through what these LLMs perhaps know and have connected together.
This illustrates how unimportant this problem is. A prior solution did exist, but apparently nobody knew because people didn't really care about it. If progress can be had by simply searching for old solutions in the literature, then that's good evidence the supposed progress is imaginary. And this is not the first time this has happened with an Erdős problem.
A lot of pure mathematics seems to consist in solving neat logic puzzles without any intrinsic importance. Recreational puzzles for very intelligent people. Or LLMs.
It shows that a 'llm' can now work on issues like this today and tomorrow it can do even more.
Don't be so ignorant. A few years ago NO ONE could have come up with something so generic as an LLM which will help you to solve this kind of problems and also create text adventures and java code.
The goal posts are strapped to skateboards these days, and the WD40 is applied to the wheels generously.
Exactly!
You misread my comment.
You can just wait and verify instead of the publishing, redacting cycles of the last year. It's embarrassing.
It's hard to predict which maths result from 100 years ago surfaces in say quantum mechanics or cryptography.
The likelihood for that is vanishingly low, though, for any given math result.
> "intrinsic importance"
"Intrinsic" in contexts like this is a word for people who are projecting what they consider important onto the world. You can't define it in any meaningful way that's not entirely subjective.
Mathematical theorems at least have objectively lower information content, because they merely rule out the impossible, while scientific knowledge also rules out the possible but non-actual.
You have it backwards. Mathematical theorems have objectively higher information content, because they rule out the impossible and model possibilities in all possible worlds that satisfy their preconditions. Scientific knowledge can never do more than inductive projections from observations in the single world we have physical access to.
The only thing that saves science from being nothing more than “huh, will you look at that,” is when it can make use of a mathematical model to provide insight into relationships between phenomena.
There is still enormous value in cleaning up the long tail of somewhat important stuff. One of the great benefits of Claude Code to me is that smaller issues no longer rot in backlogs, but can be at least attempted immediately.
The difference is that Claude Code actually solves practical problems, but pure (as opposed to applied) mathematics doesn't. Moreover, a lot of pure mathematics seems to be not just useless, but also without intrinsic epistemic value, unlike science. See https://news.ycombinator.com/item?id=46510353
Even if pure math is useless, that’s still okay. We do plenty of things that are useless. Not everything has to have a use.
Well, read the linked comment. The possible future applications of useless science can't be known either. I still argue that it has intrinsic value apart from that, unlike pure mathematics.
It's hard to know beforehand. Like with most foundational research.
My favorite example is number theory. Before cyptography came along it was pure math, an esoteric branch for just number nerds. defund Turns out, super applicable later on.
No, I'm not confusing that. Read the linked comment if you're interested.
It's unclear to me what point you are making.
This is a relief, honestly. A prior solution exists now, which means the model didn’t solve anything at all. It just regurgitated it from the internet, which we can retroactively assume contained the solution in spirit, if not in any searchable or known form. Mystery resolved.
This aligns nicely with the rest of the canon. LLMs are just stochastic parrots. Fancy autocomplete. A glorified Google search with worse footnotes. Any time they appear to do something novel, the correct explanation is that someone, somewhere, already did it, and the model merely vibes in that general direction. The fact that no human knew about it at the time is a coincidence best ignored.
The same logic applies to code. “Vibe coding” isn’t real programming. Real programming involves intuition, battle scars, and a sixth sense for bugs that can’t be articulated but somehow always validates whatever I already believe. When an LLM produces correct code, that’s not engineering, it’s cosplay. It didn’t understand the problem, because understanding is defined as something only humans possess, especially after the fact.
Naturally, only senior developers truly code. Juniors shuffle syntax. Seniors channel wisdom. Architecture decisions emerge from lived experience, not from reading millions of examples and compressing patterns into a model. If an LLM produces the same decisions, it’s obviously cargo-culting seniority without having earned the right to say “this feels wrong” in a code review.
Any success is easy to dismiss. Data leakage. Prompt hacking. Cherry-picking. Hidden humans in the loop. And if none of those apply, then it “won’t work on a real codebase,” where “real” is defined as the one place the model hasn’t touched yet. This definition will be updated as needed.
Hallucinations still settle everything. One wrong answer means the whole system is fundamentally broken. Human mistakes, meanwhile, are just learning moments, context switches, or coffee shortages. This is not a double standard. It’s experience.
Jobs are obviously safe too. Software engineering is mostly communication, domain expertise, and navigating ambiguity. If the model starts doing those things, that still doesn’t count, because it doesn’t sit in meetings, complain about product managers, or feel existential dread during sprint planning.
So yes, the Erdos situation is resolved. Nothing new happened. No reasoning occurred. Progress remains hype. The trendline is imaginary. And any discomfort you feel is probably just social media, not the ground shifting under your feet.
> This is a relief, honestly. A prior solution exists now, which means the model didn’t solve anything at all. It just regurgitated it from the internet, which we can retroactively assume contained the solution in spirit, if not in any searchable or known form. Mystery resolved.
Vs
> Interesting that in Terrance Tao's words: "though the new proof is still rather different from the literature proof)"
I firmly believe @threethirtytwo’s reply was not produced by an LLM
regardless of if this text was written by an LLM or a human, it is still slop,with a human behind it just trying to wind people up . If there is a valid point to be made , it should be made, briefly.
If the point was triggering a reply, the length and sarcasm certainly worked.
I agree brevity is always preferred. Making a good point while keeping it brief is much harder than rambling on.
But length is just a measure, quality determines if I keep reading. If a comment is too long, I won’t finish reading it. If I kept reading, it wasn’t too long.
I suspect this is AI generated, but it’s quite high quality, and doesn’t have any of the telltale signs that most AI generated content does. How did you generate this? It’s great.
Their comments are full of "it's not x, it's y" over and over. Short pithy sentences. I'm quite confident it's AI written, maybe with a more detailed prompt than the average
I guess this is the end of the human internet
To give them the benefit of the doubt, people who talk to AI too much probably start mimicking its style.
It's also the wording. The weird phrases
"Glorified Google search with worse footnotes" what on earth does that mean?
AI has a distinct feel to it
I've had that exact phrase pop up from an LLM when I asked it for a more negative code review
Your intuition on AI is out of date by about 6 months. Those telltale signs no longer exist.
It wasn't AI generated. But if it was, there is currently no way for anyone to tell the difference.
I’m confused by this. I still see this kind of phrasing in LLM generated content, even as recent as last week (using Gemini, if that matters). Are you saying that LLMs do not generate text like this, or that it’s now possible to get text that doesn’t contain the telltale “its not X, it’s Y”?
There are no reliable AI detection services. At best they can reliably detect output from popular chatbots running with their default prompts. Beyond that reliability deteriorates rapidly so they either err on the side of many false positives, or on the side of many false negatives.
There's already been several scandals where students were accused of AI use on the basis of these services and successfully fought back.
I've tested some of those services and they weren't very reliable.
If such a thing did exist, it would exist only until people started training models to hide from it.
Negative feedback is the original "all you need."
I wouldn't know how to prove to you otherwise other then to tell you that I have seen these tools show incorrect results for both AI generated text and human written text.
Good thing you had a stochastic model backing up (with “low confidence”, no less) your vague intuition of a comment you didn’t like being AI-written.
I must be a bot because I love existential dread, that's a great phrase. I feel like they trigger a lot on literate prose.
[dead]
(edit: removed duplicate comment from above, not sure how that happened)
the poster is in fact being very sarcastic. arguing in favor of emergent reasoning does in fact make sense
It's a formal sarcasm piece.
It's bizarre. The same account was previously arguing in favor of emergent reasoning abilities in another thread ( https://news.ycombinator.com/item?id=46453084 ) -- I voted it up, in fact! Turing test failed, I guess.
(edit: fixed link)
Poe's Law is the real Bitter Lesson.
We need a name for the much more trivial version of the Turing test that replaces "human" with "weird dude with rambling ideas he clearly thinks are very deep"
I'm pretty sure it's like "can it run DOOM" and someone could make an LLM that passes this that runs on an pregnancy test
Pity that HN's ability to detect sarcasm is as robust as that of a sentiment analysis model using keyword-matching.
The problem is more that it's an LLM-generated comment that's about 20x as long as it needed to be to get the point across.
Phew. This is a relief, honestly!
Despite or because?
Oh yeah, there is also a problem with people not noticing they're reading LLM output, AND with people missing sarcasm on here. Actually, I'm OK with people missing sarcasm on here - I have plenty of places to go for sarcasm and wit and it's actually kind of nice to have a place where most posts are sincere, even if that sets people up to miss it when posts are sarcastic.
Which is also what makes it problematic that you're lying about your LLM use. I would honestly love to know your prompt and how you iterated on the post, how much you put into it and how much you edited or iterated. Although pretending there was no LLM involved at all is rather disappointing.
Unfortunately I think you might feel backed into a corner now that you've insisted otherwise but it's a genuinely interesting thing here that I wish you'd elaborate on.
I definitely missed the point because of the length, and only realized after I read replies to your comment.
That’s just the internet. Detecting sarcasm requires a lot of context external to the content of any text. In person some of that is mitigated by intonation, facial expressions, etc. Typically it also requires that the the reader is a native speaker of the language or at least extremely proficient.
I'm more worried that the best LLMs aren't yet good enough to classify satire reliably.
Why not plan for a future where a lot of non-trivial tasks are automated instead of living on the edge with all this anxiety?
[flagged]
come out of the irony layer for a second -- what do you believe about LLMs?
I mean.. LLMs have hit a pretty hard wall a while ago, with the only solution being throwing monstrous compute at eking out the remaining few percent improvement (real world, not benchmarks). That's not to mention hallucinations / false paths being a foundational problem.
LLMs will continue to get slightly better in the next few years, but mainly a lot more efficient. Which will also mean better and better local models. And grounding might get better, but that just means less wrong answers, not better right answers.
So no need for doomerism. The people saying LLMs are a few years away from eating the world are either in on the con or unaware.
If all of it is going away and you should deny reality, what does everything else you wrote even mean?
Yes, it is simply impossible that anyone could look at things and do your own evaluations and come to a different, much more skeptical conclusion.
The only possible explanation is people say things they don't believe out of FUD. Literally the only one.
Are you expecting people who can't detect self-dellusions to be able to detect sarcasm, or are you just being cruel?