This defeats a few hypotheses that were widely relied on.
strong exponential time hypothesis, and the titular exponential time hypothesis
It's not saying that these problems are easy, but it does demonstrate that they are much easier than previously believed. A decade ago if a human had published this kind of thing, and importantly if it held up, they would be almost certainly a strong candidate for some sort of award(s).
Predictions markets are terrible mechanisms for this, because they require imposing an arbitrary time limit at which point the prediction will pay out. IOW, even if you religiously believed P=NP, but you think it would take at least two years to prove it, you would vote against the prediction if the deadline were one year.
There's a chain of classes P, NP(aka Sigma_1), Sigma_2, .., Sigma_n, ... PSPACE, EXPTIME. What we know for sure is that P < EXPTIME, others are unknown (well, maybe this paper adds something to this, time will tell).
It's conceivable that each of them is smaller than the next, but also that e.g. P < NP = all the sigmas < PSPACE < EXPTIME or many other possibilities. I think the general consensus is that it's more likely that there are more inequalities than not, but the fact is that we can't even show that P is less than PSPACE so far, which would be extraordinarily surprising if it were true.
The exponential time hypothesis is essentially that NP-complete problems' best solutions are all exponential.
NP-complete problems are essentially those where you need to make some kind of choice over a superlinearly-growing set of options where each option can be solved in linear time. Like, for example, finding a route between cities on a map under a certain total distance. You might need to go over all choices of routes.
The ETH essentially assumes that the best solution (in terms of complexity, i.e. how it grows) you can get for these kinds of problems is exponential. OpenAI just proved that this is false, you can have solutions that are below exponential complexity (but still above polynomial - if anyone finds a polynomial solution then we'd have proven that P=NP)
> NP-complete problems are essentially those where you need to make some kind of choice over a superlinearly-growing set of options where each option can be solved in linear time.
Seems like a weird way to describe NP, and I'm not even sure it's true?
NP is the class of decision problems where if the answer is YES and someone slips you a hint, you can solve them quickly (or quickly figure out that the hint was useless and report that). Where quickly means in polynomial time.
There's nothing about linear time in the classic definition.
Sorry, you're correct. I meant "polynomial" rather than linear.
NP = nondeterministically polynomial. i.e. if you could evaluate all choices in parallel (or just "know" the right choice), then you'd get a polynomial runtime. That's what I was trying to informally hint at, but I ended up using the wrong term. I'll edit my original post.
That's equivalent to your description - the "hint" is knowing which of the options to choose.
The reason why superlinearity comes into play is that if you had a linearly growing amount of solutions to test then you'd just have a sum of polynomials, which is itself polynomial.
Ok, I read your edited comment. Looks a lot better know.
One nitpick: I think your set of options needs to grow a lot faster than just superlinearly, I suspect it needs to grew superpolynomially?
However otherwise what you describe is essentially true of all problems in NP, not just the NP hard problems:
Sorting is (trivially) in NP, and you can say that to sort you need to pick one out of n! many permutations (which certainly grows superlinearly and even superpolynomially).
NP hard problems have the distinction that if you solve one of them efficiently, you can solve all of them efficiently, because you can reduce them to each other. (Basically because NP hard problems allow you to sort-of describe a computer, and reducing them to each other is equivalent to writing an interpreter or compiler.)
There's a set of problems that are (it is conjectured) easy to verify but hard to solve. One of them, called 3SAT, looks for values (a, b, c, ...) such that an expression like ((a or b or c) and (b or not-c or e) and ...) evaluates to true. And it turns out that all of these supposedly hard problems can be converted to 3SAT, so if you find a solution to one you find a solution to all.
The Exponential Time Hypothesis in effect says that 3SAT can only be solved by brute force--trying every possible combination until you find the right one. If your 3SAT expression has 100 variables, then this will take about 2^100 tries (exponential time).
What they found was an algorithm faster than brute force. 3SAT is still potentially one of these hard problems, but finding anything faster than brute force is a major advance.
the language around complexity domains is really confusing.
P means "this problem can be solved in polynomial # of steps relative to the problem size"
NP means "we don't know if the problem can be solved in polynomial # of steps relative to problem size, but we know we can check if the result is correct in a polynomial # of steps relative to the problem size"
Prior to today, it was widely suspected that there are problems in NP where it requires an exponential # of steps relative to problem size. We now know that any problem in NP requires AT MOST something more than polynomial but strictly less than exponential to solve.
This "makes P = NP" more likely, because we know there can't be algorithms requiring exponential # of steps if it's checkable in polynomial time. There still might be algorithms that require superpolynomial, but sub-exponetial # of steps.
This isn't quite true, problems in NP can still require exponential time to solve.
Solving an arbitrary NP problem looks like first reducing that problem to SAT and then solving the equivalent SAT problem. We know (via the Cook–Levin theorem) that any NP problem can be reduced to SAT in polynomial time, and we now know that solving SAT is subexponential in terms of the size of the SAT problem, but it's possible for the conversion to SAT to grow the problem enough that it's exponential in terms of the original problem.
But even if every problem in NP is sub-exponential (which, as tfpgh notes, this result doesn't prove), how does that make it "more likely" that every such problem is solvable in a polynomial # of steps? I feel like it's a bit like claiming that, because Alex Honnold has shown that a human can free solo El Capitan in less than 4 hours, that makes it "more likely" that I could. To me, his success does not budge the latter probability at all.
Is this really as interpetable math is fundamentally. Like wtf is this accessibility? who is this for? Do top mathematicians read this like a chapter book? Absolutely ridiculous they have all this compute and creative resources and cant make something more intuitive to help people understand it more natively
It's not possible for someone who's not a specialist in the field to understand it (unless there are more elementary proofs, of which likely none exist).
Even mathematicians take a long time to read math books and articles, especially if they go into the details. They don't read math like you would read a novel.
I agree with you insofar that the primary motivator for these press releases is public relations and investor confidence, so it should be made as minimally hard to understand as possible just by sheer virtue that it exposes oAI to more cash thrown at them.
I totally disagree with you on the precept that all difficult to understand things need be needlessly dumbed-down/watered-down until palatable for a mass (average) audience. Requiring 'math50' before 'trig101' isn't gate-keeping in any moral sense, it's an effort to save everyones' time involved when the effort is futile.
Similarly I don't demand that haute couture fashion companies explain the style to me until 'I get it' ; I just take it for granted that i'm not going to be able to understand the entirety of the world in one lifetime.
> I agree with you insofar that the primary motivator for these press releases is public relations and investor confidence, so it should be made as minimally hard to understand as possible just by sheer virtue that it exposes oAI to more cash thrown at them.
Why? The investors can just ask Claude or Gemini whether the proof holds up.
> Absolutely ridiculous they have all this compute and creative resources and cant make something more intuitive to help people understand it more natively
We can record a lot of parrot speech and train LLMs against it and eventually build a translator.
They are making good progress on similar research programmes for many cetaceans.
In the case of our stochastic parrots it's even easier, since they already speak something like human language. And for the math proofs it's even easier, because once they are formalised into something like Lean the humans only need to understand the statement of the theorem, and a mechanical computer program can check the proof.
Not a math major here. But is this paper way too long? So long that it raises the possibility that OpenAI's staff mathematicians passed through AI's raw output as-is?
It is just that we never considered how advanced stochastic parrots could be so powerful when dealing with basically the whole library of acquired human knowledge and guided by experts to generate.
It is not a demerit, actually I think that it is a very elegant outcome how many problems could be solved by generation and this probably show us how language is a far more important thing than we ever thought it was.
Wow. Complexity class separations are really really rare.
Can an expert in sat solving comment if there's any immediate applications of the method from the paper? Or is SAT solving in general pretty far away from worst-case times, and this is a theoretical algorithm which would be hard to apply in practice.
The data you put into a SAT solver has substantial effect on it's performance in practice. Real world inputs tend to be much nicer than pathological ones, so I don't think there's much practical application, but I also haven't read the paper yet.
SAT-solving expert here. No immediate applications (the results in the paper are "galactic"). The main question is whether some ideas here will snowball into other ideas, one of which could eventually be very practical.
> A galactic algorithm is an algorithm with state-of-the-art theoretical (asymptotic) performance, but which is not used due to practical constraints.
Yes, I think we can pretty safely say that the algorithm in the paper is not used due to practical constraints. This is a subjective judgment, but pretty clear from the paper that the algorithms aren't practical to apply.
This is a complete non-sequitur, but I checked out your web page. I love the Curious George image. My daughter has been deeply into the Curious George episodes for the past couple of months, and I don't think I've seen him show up in many other places besides kids' shows :D
It's surprising that Exponential Time Hypothesis would be false. Quoting a paper from 2022[1]:
> We prove that the time complexity of self-attention is necessarily quadratic in the input length, unless the Strong Exponential Time Hypothesis (SETH) is false[...]
If OpenAI's proof is correct, does that mean self-attention can be implemented in sub-quadratic time? I.e. we could eventually have all the most advanced LLMs running on cheap hardware? To put it more playfully, is ETH being false actually bearish for Nvidia?
That makes me wonder if "I see your CS proof, how does this affect LLMs / nvidia stonk price???" being immediately self-described as "playful" ensures criticism looks cranky.
"I see your particle physics paper (on sustainability of neutron-induced chain fission reaction in uranium), how does this affect the world politics???"
Turns out, it affected the global politics tremendously.
If AI really scales indefinitely is an increase in efficiency even bearish for Nvidia? Wouldn't they still build more and more compute as fast as they can, because these things get smarter and better with more compute?
I think that statement is too strong -- if ETH was a core belief then its implication P≠NP must also have been a core belief ... but I don't think that those who assumed ETH held it that strongly. As you say, it was commonly thought to be true ... as was P≠NP ... and the latter still is. But is not a "core belief" -- I think virtually everyone believes that they could be wrong, whichever side they're on.
This is an amazing and unexpected result. To be clear though, it does not imply that NP is subexponential: "Note that runtime bounds of the form 2^o(n) do not transfer under general polynomial-time reductions (2^n^o(1) would be needed for this), so our result does not imply that general NP problems
admit subexponential time algorithms.""
> The subexponential solution does carry over to any (?) problems that have can be reduced to 3SAT in a linear number of variables. They list: Counting independent sets, Counting proper q-colorings, Counting perfect matchings.
Wouldn't that also include functions that can be evaluated in hardware using a fixed number of gates? E.g., symmetric cryptography?
A circuit with a fixed number of (binary logical) gates can be solved in constant time, as it will have a fixed number of input bits, so brute force will have a fixed number of input options to consider. This does apply to cryptographic algorithms with a fixed key size, like the ones used in practice, as well.
The whole point of computational complexity is to look at the scaling behavior when you increase the inputs/key size/size of the circuit.
From my reading no. Because brute forcing anything symmetrical is already O(n). What protects us is the assumption that there is not enough energy in the observable universe to brute force 2^256 key.
It's NP-complete problems if "subexponential" means O(2^{n^{\epsilon}}) for every \epsilon>0.
If subexponential just means 2^{o(n)}, (as it is here, O(2^{n/\sqrt{\log n}})), then no. E.g. if the cost of the reduction took you from n to \sqrt{\log n}n, then you'd be back to exponential time. The Subexponential Reduction Family is the class of problems that preserve subexponential time in the second sense.
Perhaps this will be the push we need to finally, seriously consider the possibility that our formalisms of “efficient” and “efficiently verifiable” algorithms correspond to nothing resembling reality, and P \in NP is actually ill-posed.
This has been marginally debated since the mid 90s, when Impagliazzo published the “5 worlds” survey. Occasionally it is brought up in talks about average case analysis. There is occasionally a workshop about it. But it has never really got the attention it deserved. How many more of these results do we need to take it seriously? (Assuming we do indeed certify it.)
> Perhaps this will be the push we need to finally, seriously consider the possibility that our formalisms of “efficient” and “efficiently verifiable” algorithms correspond to nothing resembling reality,
What do you mean? Do you mean that our heuristics shortcuts of treating P as efficiently solvable (and NP as efficiently verifiable) are not warranted? So far pretty much all algorithms in P that we found have small exponents. That hasn't changed, but it might change in the future.
Or do you mean something else?
Keep in mind that even most instances of many NP hard problems can be solved very efficiently in practice, because most instances ain't hard.
However we know problems that are expected to be 'easier' than NP that are hard to solve in practice. These are pretty much all the problems we build cryptography out of, because essentially these problems are in the intersection of NP and co-NP and that's expected to be smaller than NP. And cryptography has to be hard to break on average to work: a random key needs to be a good key.
> and P \in NP is actually ill-posed.
Oh, that's very simple to answer: P is not an element of NP. The question people care about is whether there's any element in NP that's not in P. In LaTeX notation P \subset NP vs P \subseteq NP.
I'm not sure what you mean by ill-posed? The question is very clear. What are you trying to say?
"Ill-posed" here means that the formalisms we've chosen to represent "efficient" and "efficiently verifiable" algorithms could turn out to not correspond to anything humans would consider meaningful ("we posed the problem poorly"). For most people, a solution to an NP-complete problem that's barely-subexponential would be clear evidence of these formalisms being ill-posed.
There are many ways it might turn out to be ill-posed, this is just one of them. Impagliazzo's "5 worlds" survey is in some sense a catalog of some of these ways, and includes some of the ways you seem to be referencing (approxmations, probabilistic techniques, average-case efficiency; depending what you mean this could lie in the "Heuristica" world).
> "Ill-posed" here means that the formalisms we've chosen to represent "efficient" and "efficiently verifiable" algorithms could turn out to not correspond to anything humans would consider meaningful ("we posed the problem poorly").
OK, that's a reasonable worry and one that practitioners share. Thanks for explaining. Though I don't think ill-posed is a good term for this, the technical meaning of the term points in a different direction.
So far, as far as I can tell this new result gives much of any extra credence to that worry.
It is completely acceptable in math to use "ill-posed" to indicate that something is insufficiently specified, improperly formulated, or ambiguous. Math people do this all the time.
You can hold that position if you like but you should do so knowing that the vast majority of people (including math people) would say that a formalism that classifies a technically-subexponential solution to an NP-complete problem as “efficient” is in fact improperly formulated, because the goal was specifically to formalize our intuition of efficiency. Our formulation in that case would have failed to do the single thing we laid out for it. Textbook improper formulation. There is not another way of putting this.
The technical result has nothing to do with "efficient" which according to textbook CS theory roughly defines it to be the set P, the strictly polynomial time algorithms. Has nothing to do with the ETH result which is still an exponential, 2^(n/ log n), just not a pure exponential.
I'm also confused what you are saying down here. Though I can see how "sub-exponential" is easily misleading terminology.
The point is not that ETH has anything to do with P or NP, but that the bounds theoretical CS chooses for its conjectures tend to be extremely round and convenient. Does reality actually prefer such nice, round bounds in general? Does reality prefer them in the case of P and NP?
In essentially all cases, we have no real reason to believe or disbelieve. We choose these bounds because we can't do better. It's often the case that we later prove them wrong with some extremely strange result. Here, we have weird bounds for 3-SAT, and a few weeks ago we had one for integer multiplication.
But this has also happened repeatedly throughout history. In the 50s Kolmogorov conjectured that n-bit integer multiplication was \Omega(n^2). Karatsuba produced O(n^{1.58}) during Kolmogorov's seminar on the topic. Strassen discovered matrix multiplication is O(n^{2.81}) while trying to prove it is O(n^3). And on and on and on.
Maybe P and NP are well-formulated and they do perfectly capture our intuition of "efficient" and "efficiently verifiable" algorithms. Or maybe they don't. We are probably going to keep finding extremely weird, cheesed-out bounds for famous problems, though. You can choose what you want to believe about reality, but I think it would be more remarkable if the reality of P and NP turned out to be as clean as the bounds indicate.
Humans in this thread have already explained that concept with much more straightforward prose.
From the HN comment guidelines:
> Don't post generated text or AI-edited text. HN is for conversation between humans.
Disclosure doesn't address the objection (not that it wouldn't be obvious anyway). If you learned something from an AI convo, tell us what you learned, in your own words, not what the LLM said.
If I can "count the satisfying assignments of a k-CNF in subexponential time and polynomial space", could I just count the the solutions for assignments of 0 and 1 to x_1, then proceed with whichever is greater (or nonzero), counting the solutions for x_2=0, and so on.
Doesn't that lead me to an existence proof (i.e., a solution) to a k-SAT problem in N*(subexponential_time and polynomial_space)?
And with constant folding, N is reduced by 1 on each iteration.
This construction splits the space into 'small' models with a quadratic lookup cost and 'large' with sub-exponential cost. The small one is by enumerating all SAT problems with their answers, then looking up the answer. The large one is something in polynomials I don't understand. An answer from openai about where the cutoff is was:
the constructed lookup machine has vastly more states than the estimated total number of physically distinguishable configurations of the observable universe.
I'm not inclined to consider "wrote down all the answers ahead of time, so that's free" as a quadratic time solution to SAT. Nor very convinced by the other half.
So on balance, if this is considered proof that the exponential time hypothesis is false, and openai model reading of the lean was vaguely on point, then I reject the axioms of this system. Not a proof, even if there is lean. I'm OK with being downvoted for that position.
I'm not the kind of programmer that even understands P=NP, so I'll either be way off base, or be blessed with an outsider perspective.
Considering that no human will be able to sift through these in no less than a team-decade, to verify that there's no mistake, yet still that will be massively less compute than the one necessary to produce the 'solution', isn't the empirical finding contrary to the theoretical finding being suggested? Namely that in practice we are generating these 'proofs' in a massive amount of compute time, but it will take tiny amounts of time to verify them.
Is it very useful to talk about computing in purely mathematical terms and not in empirical terms? There's very obvious truths that we are seeing here, and it seems like those that go through some sort of rigorous mathematical training kind of fail to see it. Similar to how 70% of physicians or something believe in the Many Worlds Interpretation or that time isn't real, a sort of algebraic psychosis where one is removed from a human reality that we have access to in order to attempt getting close to an 'objective reality' that is far removed from the human experience.
To some extent we are computing entities, so I don't even think that we can conceive of computing as decoupled from ourselves[1]. I know the observer effect is very often mis-cited by laymen, but I'll be damned if some computer scientists like Godel don't sound like they were up to something similar to this.
Also, here's another outsider perspective. This is obviously slop, I'm not sure how but it's obviously nerd-sniping the CS majors that understand this, because it's obviously very hard to find the flaw, but it gives epicycles energy. It's also sufficiently complex that under a bayesian analysis it is more likely that the AI has reached the limit of its processing power until it becomes confused enough to believe it has reached something significant, kind of like how I'm doing now.
> Considering that no human will be able to sift through these in no less than a team-decade
How do you figure that? For ETH, OpenAI is engaging with top experts in this area. It's unlikely that they'll need 10 man-years combined to go through the results:
> This work will be verified and shepherded by a group led by Russell Impagliazzo (UC San Diego), Ramamohan Paturi (UC San Diego), and Ryan Williams (MIT). All three are experts in algorithm design and complexity theory.
> Impagliazzo and Paturi, along with Francis Zane, are the original proposers of the Exponential Time Hypothesis, and have authored a number of influential papers on the topic over the last 30 years.
> Williams has refuted strong versions of the ETH, used ETH for hardness results, and connected algorithms and lower bounds.
I think I misread something, the pdf is like 80 pages long, and there’s a list of like 200 pdfs, each around that length, I thought they were all related to this proof, but they are unrelated.
While this has no immediate bearing on P-vs-NP, it is interesting to note:
- Early CS researchers like Donald Knuth entered CS in the days when regularly someone would come up with a surprising faster algorithm for some problem, and their intuitions were that most problems had faster algorithms than you'd expect.
- Later generations of CS researchers, post Karp's definition of NP-completeness (and the Garey & Johnson book, say), began to specialize in proving hardness, and complexity theory with a “zoo” of complexity classes became more about separation and lower bounds, e.g. I've seen algorithms like Arora's PTAS for Euclidean TSP being described as a “failed attempt to prove a lower bound”. So an entire generation of researchers' intuition was mostly in the direction of an elegant universe with neat separation results (even if we couldn't get them).
- So e.g. surveys of computer scientists showed most of them believing that P≠NP, and someone like Knuth has always seemed a bit of an outlier for saying that maybe P=NP: his argument was simply that there are so many possible polynomial-time algorithms that we haven't explored, and maybe one of them with a large exponent could solve NP-complete problems. (In response to arguments that many researchers have failed to find any such algorithms, he would say that many researchers have failed to prove that P≠NP.) E.g. in 2017 when Scott Aaronson wrote his survey (https://www.scottaaronson.com/papers/pnp.pdf = https://www.scottaaronson.com/papers/pnp-kindle.pdf) he wrote:
> Admittedly, it can be hard to tell whether declarations that P = NP are meant seriously, or are merely attempts to be contrarian. However, we can surely agree with Knuth and Lipton that we’re far from understanding the limits of efficient computation, and that there are further surprises in store.
- What we've seen in this month is surprising unexpected algorithms, like o(n log n) algorithms for FFT and integer multiplication, and now this sub–exponential-time algorithm for 3-SAT. (Yes the ETH has had less support than P≠NP but I still gather it's unexpected.)
So (though I don't expect intuitions on P≠NP to change, and probably that's still the right intuition), maybe there will be a bit of a swing of the pendulum back to the idea that maybe clever algorithms exist that we just haven't thought of.
Part of the problem with the current centralization of the world's math capability within AI labs is that if they do find that P=NP via an efficient algorithm, it could have implications on all modern public key based cryptography. So it's not completely clear that they will (be allowed to) release these results.
> if AI frontier labs do find that P=NP via an efficient algorithm, it could have implications on all modern public key based cryptography.
The fact that other math problems are bulk solved yet nothing from cryptography was released/published as solved could be a strong indicator that "all modern public key based cryptography" could have been at risk.
It seems rather unlikely that solving the Navier-Stokes problem is harder than e.g. a better attack on a reduced-round AES. Perhaps they think an epsilon improvement in the exponent isn't worth publishing, but it seems more likely that they're intentionally withholding any results they have.
Either that or they've not put the model(s) on those problems. Remember that even the Navier-Stokes solution did use a large amount of compute. I'm not sure they'd even want to waste it on problems that they can't get PR from.
It's a reasonable assumption that the cases where the NSA, etc, had a role in designing cryptographic systems, and reportedly made specific suggestions that could be forms of a backdoor, would all be revealed by AI being tasked to find it. Maybe even indirectly as part of the broader defensive security research by models unreleased to the public. If algorithmic complexity or a really obscure backdoor has existed in cryptographic systems and put there by the Gov't, it seems that current frontier models would have the best chance of revealing it.
Maybe I'm misunderstanding you, but a better attack on a reduced-round AES has been achieved by LLMs. Anthropic did it and published it already, a chosen plaintext attack that improves on state of the art against reduced-round by "200-800x". https://anthropic.com/document/aes_mobius_bridge.pdf
Also most of the results come from "spray attacks". In addition to the results that are found there are also tons of failed attempts on other problems. So merely through how large the whole space is one should not necessarily expect groundbreaking results in cryptography specifically (even though it can also be true that some results be withheld, at the same time).
"They've specifically avoided trying" is also a possibility. Because what do they do if they actually find a critical flaw? Publishing leads to mass chaos. If they don't publish but it leaks, also mass chaos. (Probably just leaking the existence is enough without the actual proof). If they're "responsible" and report it to NSA, there's a good chance that model development gets forcibly halted and some chance that the entire company gets seized on national security grounds.
The issue is that if a country's agencies find an efficient attack to modern public key based cryptography with LLMs, there is no reason to think that other countries will not find the same in a few months with their LLMs. So, finding a vulnerability could offer a short term advantage, but one would want eventually to protect their own, too.
I think a proof of P=NP via a galactic algorithm is far more likely than a proof of P=NP via an efficient algorithm.
I mean if someone develops an O(n^100) algorithm for solving an NP-complete problem, that would prove P=NP, but pose no immediate threat to cryptography.
It increases risk, in that a galactic algorithm increases the odds that an efficient algorithm exists waiting to be found, but it is entirely (epistemically) possible that the galactic algorithm exists but the efficient algorithm doesn't.
Maybe, but that feels significantly more exotic and unlikely to me. You have to nearly negate the natural cost scaling of brute force in the thousand to million n range but lose that speed benefit later.
Well, if someone found a quasipolynomial time algorithm – O(n^log(n)) – for an NP-complete problem, that would technically be compatible with P!=NP, but might have quite significant practical implications.
To have a really big impact you'd have to have that be the dominating factor at lower n and keep that "log(n)" part of the equation below 10 or so for all practical n. The latter one in particular seems unlikely to me.
There are hard problems where the worst case may be exponentially hard, but the average case is much faster. Linear programming and the traveling salesman problem are like that. The faster approaches are considered "probabilistic solutions".
In cryptography, most of the problems believed to be exponentially hard, such as factoring, don't have proofs of an exponential lower bound. There's some risk of a breakthrough that makes those problems easier. A probabilistic solution would be good enough to break many cryptosystems.
I suspect that people at the cryptanalysis agencies worldwide are staying up late tonight, trying to figure out the impacts and risks of this new result.
> In fact, one of the first things I wanted OpenAI to check is that no existing cryptosystems are compromised by the result. Apparently they are safe...
so if you trust his trust of OpenAI, then this has no (immediately obvious) bearing on cryptosystems for now.
Seeing the panicked state of the PHP ecosystem at the moment, and having worked in the financial system in the past, I dread the thought of what would happen to global banking if a frontier model broke a maths problem that underpins our crypto systems...
Somehow, their wide range of results touch every single area of math except anything touching cryptography. I think it's fair to expect that three letter agencies are given the ability to veto the publication of certain discoveries.
Wasn't there a comment a while ago when all major LLM providers went down almost at the same time that it was caused by a sort of virtual "Room 641A" thing?
Funny you mention probabilistic algorithms. If I’m not mistaken one of the OpenAI results is that most of the probabilistic methods (specifically those that use a logarithmic amount of space)-4 have non-probabilistic versions that are as efficient in time and space (L = RL = BPL)
Linear programming admits a worst-case polynomial time algorithm.
It is notable that simplex, especially dual simplex is often better in practice, especially for integer linear programming with branch and bound because you only have to slightly modify the parent optimal basis in simplex vs entirely restarting or refactorizing for interior point methods.
We had O(n lg n) since 2019 already; now we (supposedly) know the limit is less than that (but of course at least linear).
> In 2019, Harvey and van der Hoeven came up with a galactic algorithm with complexity O ( n log n ) {\displaystyle O(n\log n)}. This matches a guess by Schönhage and Strassen that this would be the optimal bound.
> In September 2026, OpenAI released a preprint claiming an algorithm with complexity O ( n ( log n ) α ) {\displaystyle O(n(\log n)^{\alpha })} with α < 1 {\displaystyle \alpha <1}
I believe this stems from an incomplete understanding of the problem space on my part (excuse the multiple puns); but, I am skeptical Knuth really said P=NP. There are so many mathemtical examples of equations that clearly show P≠NP. Computing in itself was originally formed from using iterative methods to solve hard problems. Least-squares methods, sin() cos(), traveling salesman, so many to list? To say P=NP for all problems seems like a fools errand to me.
I am not a computer scientist so my understanding is "incomplete". I thought it stood for polynomial and non-polynomial time. But i do mix them up here and there. What i have read time and time again is that: If P≠NP would mean some problems really are harder to solve than to check (which feels intuitive given working encryption exists and all the mathematical functions that I know and understand). For example checking the roots of polynomials is easier than finding them in the first place.
From this comment it isn't clear if you previously thought it stood for polynomial and non-polynomial and have been corrected or if you still have this misconception, but just in case: it actually stands for polynomial and nondeterministic polynomial. P is a subset of NP, they aren't opposites. And yes most people intuit that they likely aren't equal.
> Don Knuth: As you say, I've come to believe that P = N P, namely that there does exist an integer M and an algorithm that will solve every n-bit problem belonging to the class N P in n**M elementary steps.
thanks for the quote. Reading through it is clear that my understanding of P=NP is not complete enough. I would have thought n*M is precisely not polynomial but exponential time.
…Oh, I see. The M here is a hypothesized constant, with n as the base. That's how polynomial time looks. Exponential time would have n in the exponent.
Double-checking, is this more like the recent DFT improvement where even the word-RAM model of computation doesn't justify the fraction of a log factor they eliminated because of an assumption of arbitrary-precision arithmetic, or has ETH actually fallen with the Lean proof being formulated appropriately? Do we know yet?
That doesn't sound right. Previous results proved that with bounded values you couldn't beat O(nlogn), among other things. The OpenAI result explicitly used arbitrary-precision complex arithmetic. Yes, it's galactic [0], but either it's wrong or the previous results were wrong, no?
[0] Very much not something I care much about in these sorts of conversations. Even if it weren't galactic, a random logn factor here and there usually loses to cache efficiency and whatnot. Their speedups were less than that, and I'd already assumed we were in the realm of applied math rather than real engineering.
> but either it's wrong or the previous results were wrong, no?
Not quite. This result uses polylog(n) precision which isn't bounded, so it escapes the previous bounds. This doesn't noticably affect the runtime because not all of the math needs this extra precision (in this case you need a couple computations in higher precision but less than n operations), so the extra precision can be payed for even in a TM/circuit model.
no. The DFT result as published uses unlimited precision, but the comparison with the int multiplication shows that the unlimited precision just makes the paper shorter/easier rather than being an exploit in the computation model.
What about OpenAI allowed ETH to be determined to be false that couldn't be done by humans? Was there something in particular that OpenAI's AI did that help bridge the gap to a solution?
I think anthropic one is edited by humans, OpenAI drop just looks like how Astra writes (i know it's probably a different model producing this but still)
The introduction is very nicely written, you're right. I learned a lot from it. There's some detail on how impractical their algorithm is:
> Our result is purely asymptotic: for arbitrary 3-CNF formulas with m = 5n clauses, the bound we prove does not improve on the 2n of exhaustive search unless n > 10^10000,
So their algorithm beats standard algorithms only when you have more variables than there are atoms in the universe (10^80). Even if every atom contained a whole universe, and each atom in each universe itself contained a whole universe, iterated 100 times, your still not there.
It would be interesting to see if this new counting method has any implications to counting 2SAT or parity-counting of 2SAT. Since problem posed in XORSAT and 2SAT clauses is NP-complete, this could have implications about P vs NP.
201 comments
[ 0.84 ms ] story [ 28.0 ms ] threadstrong exponential time hypothesis, and the titular exponential time hypothesis
It's not saying that these problems are easy, but it does demonstrate that they are much easier than previously believed. A decade ago if a human had published this kind of thing, and importantly if it held up, they would be almost certainly a strong candidate for some sort of award(s).
It's conceivable that each of them is smaller than the next, but also that e.g. P < NP = all the sigmas < PSPACE < EXPTIME or many other possibilities. I think the general consensus is that it's more likely that there are more inequalities than not, but the fact is that we can't even show that P is less than PSPACE so far, which would be extraordinarily surprising if it were true.
NP-complete problems are essentially those where you need to make some kind of choice over a superlinearly-growing set of options where each option can be solved in linear time. Like, for example, finding a route between cities on a map under a certain total distance. You might need to go over all choices of routes.
The ETH essentially assumes that the best solution (in terms of complexity, i.e. how it grows) you can get for these kinds of problems is exponential. OpenAI just proved that this is false, you can have solutions that are below exponential complexity (but still above polynomial - if anyone finds a polynomial solution then we'd have proven that P=NP)
Seems like a weird way to describe NP, and I'm not even sure it's true?
NP is the class of decision problems where if the answer is YES and someone slips you a hint, you can solve them quickly (or quickly figure out that the hint was useless and report that). Where quickly means in polynomial time.
There's nothing about linear time in the classic definition.
NP = nondeterministically polynomial. i.e. if you could evaluate all choices in parallel (or just "know" the right choice), then you'd get a polynomial runtime. That's what I was trying to informally hint at, but I ended up using the wrong term. I'll edit my original post.
That's equivalent to your description - the "hint" is knowing which of the options to choose.
The reason why superlinearity comes into play is that if you had a linearly growing amount of solutions to test then you'd just have a sum of polynomials, which is itself polynomial.
One nitpick: I think your set of options needs to grow a lot faster than just superlinearly, I suspect it needs to grew superpolynomially?
However otherwise what you describe is essentially true of all problems in NP, not just the NP hard problems:
Sorting is (trivially) in NP, and you can say that to sort you need to pick one out of n! many permutations (which certainly grows superlinearly and even superpolynomially).
NP hard problems have the distinction that if you solve one of them efficiently, you can solve all of them efficiently, because you can reduce them to each other. (Basically because NP hard problems allow you to sort-of describe a computer, and reducing them to each other is equivalent to writing an interpreter or compiler.)
The Exponential Time Hypothesis in effect says that 3SAT can only be solved by brute force--trying every possible combination until you find the right one. If your 3SAT expression has 100 variables, then this will take about 2^100 tries (exponential time).
What they found was an algorithm faster than brute force. 3SAT is still potentially one of these hard problems, but finding anything faster than brute force is a major advance.
P means "this problem can be solved in polynomial # of steps relative to the problem size"
NP means "we don't know if the problem can be solved in polynomial # of steps relative to problem size, but we know we can check if the result is correct in a polynomial # of steps relative to the problem size"
Prior to today, it was widely suspected that there are problems in NP where it requires an exponential # of steps relative to problem size. We now know that any problem in NP requires AT MOST something more than polynomial but strictly less than exponential to solve.
This "makes P = NP" more likely, because we know there can't be algorithms requiring exponential # of steps if it's checkable in polynomial time. There still might be algorithms that require superpolynomial, but sub-exponetial # of steps.
Solving an arbitrary NP problem looks like first reducing that problem to SAT and then solving the equivalent SAT problem. We know (via the Cook–Levin theorem) that any NP problem can be reduced to SAT in polynomial time, and we now know that solving SAT is subexponential in terms of the size of the SAT problem, but it's possible for the conversion to SAT to grow the problem enough that it's exponential in terms of the original problem.
It's not possible for someone who's not a specialist in the field to understand it (unless there are more elementary proofs, of which likely none exist).
siting on a recliner making an occasional sensible chuckle while flipping through the pages?
I totally disagree with you on the precept that all difficult to understand things need be needlessly dumbed-down/watered-down until palatable for a mass (average) audience. Requiring 'math50' before 'trig101' isn't gate-keeping in any moral sense, it's an effort to save everyones' time involved when the effort is futile.
Similarly I don't demand that haute couture fashion companies explain the style to me until 'I get it' ; I just take it for granted that i'm not going to be able to understand the entirety of the world in one lifetime.
Why? The investors can just ask Claude or Gemini whether the proof holds up.
OpenAI has that too - open their website and ask.
They are making good progress on similar research programmes for many cetaceans.
In the case of our stochastic parrots it's even easier, since they already speak something like human language. And for the math proofs it's even easier, because once they are formalised into something like Lean the humans only need to understand the statement of the theorem, and a mechanical computer program can check the proof.
It is not a demerit, actually I think that it is a very elegant outcome how many problems could be solved by generation and this probably show us how language is a far more important thing than we ever thought it was.
Can an expert in sat solving comment if there's any immediate applications of the method from the paper? Or is SAT solving in general pretty far away from worst-case times, and this is a theoretical algorithm which would be hard to apply in practice.
Yes, I think we can pretty safely say that the algorithm in the paper is not used due to practical constraints. This is a subjective judgment, but pretty clear from the paper that the algorithms aren't practical to apply.
> We prove that the time complexity of self-attention is necessarily quadratic in the input length, unless the Strong Exponential Time Hypothesis (SETH) is false[...]
If OpenAI's proof is correct, does that mean self-attention can be implemented in sub-quadratic time? I.e. we could eventually have all the most advanced LLMs running on cheap hardware? To put it more playfully, is ETH being false actually bearish for Nvidia?
[1]: https://arxiv.org/abs/2209.04881
But it's also not even wrong, in the Pauli sense.
That makes me wonder if "I see your CS proof, how does this affect LLMs / nvidia stonk price???" being immediately self-described as "playful" ensures criticism looks cranky.
Turns out, it affected the global politics tremendously.
sure, this proof itself sheds no light on the matter, but we've lost one of our core beliefs.
Wouldn't that also include functions that can be evaluated in hardware using a fixed number of gates? E.g., symmetric cryptography?
Is there a subset of NP problems for which a subexponential algorithm would witness the entire NP class being subexponential?
If subexponential just means 2^{o(n)}, (as it is here, O(2^{n/\sqrt{\log n}})), then no. E.g. if the cost of the reduction took you from n to \sqrt{\log n}n, then you'd be back to exponential time. The Subexponential Reduction Family is the class of problems that preserve subexponential time in the second sense.
No, subexponential is not the same as polynomial time. There's sub-exponential functions that grow faster than polynomial.
This has been marginally debated since the mid 90s, when Impagliazzo published the “5 worlds” survey. Occasionally it is brought up in talks about average case analysis. There is occasionally a workshop about it. But it has never really got the attention it deserved. How many more of these results do we need to take it seriously? (Assuming we do indeed certify it.)
What do you mean? Do you mean that our heuristics shortcuts of treating P as efficiently solvable (and NP as efficiently verifiable) are not warranted? So far pretty much all algorithms in P that we found have small exponents. That hasn't changed, but it might change in the future.
Or do you mean something else?
Keep in mind that even most instances of many NP hard problems can be solved very efficiently in practice, because most instances ain't hard.
However we know problems that are expected to be 'easier' than NP that are hard to solve in practice. These are pretty much all the problems we build cryptography out of, because essentially these problems are in the intersection of NP and co-NP and that's expected to be smaller than NP. And cryptography has to be hard to break on average to work: a random key needs to be a good key.
> and P \in NP is actually ill-posed.
Oh, that's very simple to answer: P is not an element of NP. The question people care about is whether there's any element in NP that's not in P. In LaTeX notation P \subset NP vs P \subseteq NP.
I'm not sure what you mean by ill-posed? The question is very clear. What are you trying to say?
There are many ways it might turn out to be ill-posed, this is just one of them. Impagliazzo's "5 worlds" survey is in some sense a catalog of some of these ways, and includes some of the ways you seem to be referencing (approxmations, probabilistic techniques, average-case efficiency; depending what you mean this could lie in the "Heuristica" world).
OK, that's a reasonable worry and one that practitioners share. Thanks for explaining. Though I don't think ill-posed is a good term for this, the technical meaning of the term points in a different direction.
So far, as far as I can tell this new result gives much of any extra credence to that worry.
The problem is that this formal question might not correspond to the informal question we actually care about.
I'm also confused what you are saying down here. Though I can see how "sub-exponential" is easily misleading terminology.
In essentially all cases, we have no real reason to believe or disbelieve. We choose these bounds because we can't do better. It's often the case that we later prove them wrong with some extremely strange result. Here, we have weird bounds for 3-SAT, and a few weeks ago we had one for integer multiplication.
But this has also happened repeatedly throughout history. In the 50s Kolmogorov conjectured that n-bit integer multiplication was \Omega(n^2). Karatsuba produced O(n^{1.58}) during Kolmogorov's seminar on the topic. Strassen discovered matrix multiplication is O(n^{2.81}) while trying to prove it is O(n^3). And on and on and on.
Maybe P and NP are well-formulated and they do perfectly capture our intuition of "efficient" and "efficiently verifiable" algorithms. Or maybe they don't. We are probably going to keep finding extremely weird, cheesed-out bounds for famous problems, though. You can choose what you want to believe about reality, but I think it would be more remarkable if the reality of P and NP turned out to be as clean as the bounds indicate.
However, so far equating P with 'efficient in practice' seems to hold up.
From the HN comment guidelines:
> Don't post generated text or AI-edited text. HN is for conversation between humans.
Disclosure doesn't address the objection (not that it wouldn't be obvious anyway). If you learned something from an AI convo, tell us what you learned, in your own words, not what the LLM said.
Doesn't that lead me to an existence proof (i.e., a solution) to a k-SAT problem in N*(subexponential_time and polynomial_space)?
And with constant folding, N is reduced by 1 on each iteration.
What obvious thing am I missing here?
So on balance, if this is considered proof that the exponential time hypothesis is false, and openai model reading of the lean was vaguely on point, then I reject the axioms of this system. Not a proof, even if there is lean. I'm OK with being downvoted for that position.
I'm reminded of the Apollo program that was cool for a while, but eventually the public lost interest.
Maybe it becomes cheaper? Maybe other labs want to prove their worth will pick up the slack when another loses interest?
Considering that no human will be able to sift through these in no less than a team-decade, to verify that there's no mistake, yet still that will be massively less compute than the one necessary to produce the 'solution', isn't the empirical finding contrary to the theoretical finding being suggested? Namely that in practice we are generating these 'proofs' in a massive amount of compute time, but it will take tiny amounts of time to verify them.
Is it very useful to talk about computing in purely mathematical terms and not in empirical terms? There's very obvious truths that we are seeing here, and it seems like those that go through some sort of rigorous mathematical training kind of fail to see it. Similar to how 70% of physicians or something believe in the Many Worlds Interpretation or that time isn't real, a sort of algebraic psychosis where one is removed from a human reality that we have access to in order to attempt getting close to an 'objective reality' that is far removed from the human experience.
To some extent we are computing entities, so I don't even think that we can conceive of computing as decoupled from ourselves[1]. I know the observer effect is very often mis-cited by laymen, but I'll be damned if some computer scientists like Godel don't sound like they were up to something similar to this.
Also, here's another outsider perspective. This is obviously slop, I'm not sure how but it's obviously nerd-sniping the CS majors that understand this, because it's obviously very hard to find the flaw, but it gives epicycles energy. It's also sufficiently complex that under a bayesian analysis it is more likely that the AI has reached the limit of its processing power until it becomes confused enough to believe it has reached something significant, kind of like how I'm doing now.
[1] xkcd://1163
How do you figure that? For ETH, OpenAI is engaging with top experts in this area. It's unlikely that they'll need 10 man-years combined to go through the results:
> This work will be verified and shepherded by a group led by Russell Impagliazzo (UC San Diego), Ramamohan Paturi (UC San Diego), and Ryan Williams (MIT). All three are experts in algorithm design and complexity theory.
> Impagliazzo and Paturi, along with Francis Zane, are the original proposers of the Exponential Time Hypothesis, and have authored a number of influential papers on the topic over the last 30 years.
> Williams has refuted strong versions of the ETH, used ETH for hardness results, and connected algorithms and lower bounds.
- Early CS researchers like Donald Knuth entered CS in the days when regularly someone would come up with a surprising faster algorithm for some problem, and their intuitions were that most problems had faster algorithms than you'd expect.
- Later generations of CS researchers, post Karp's definition of NP-completeness (and the Garey & Johnson book, say), began to specialize in proving hardness, and complexity theory with a “zoo” of complexity classes became more about separation and lower bounds, e.g. I've seen algorithms like Arora's PTAS for Euclidean TSP being described as a “failed attempt to prove a lower bound”. So an entire generation of researchers' intuition was mostly in the direction of an elegant universe with neat separation results (even if we couldn't get them).
- So e.g. surveys of computer scientists showed most of them believing that P≠NP, and someone like Knuth has always seemed a bit of an outlier for saying that maybe P=NP: his argument was simply that there are so many possible polynomial-time algorithms that we haven't explored, and maybe one of them with a large exponent could solve NP-complete problems. (In response to arguments that many researchers have failed to find any such algorithms, he would say that many researchers have failed to prove that P≠NP.) E.g. in 2017 when Scott Aaronson wrote his survey (https://www.scottaaronson.com/papers/pnp.pdf = https://www.scottaaronson.com/papers/pnp-kindle.pdf) he wrote:
> Admittedly, it can be hard to tell whether declarations that P = NP are meant seriously, or are merely attempts to be contrarian. However, we can surely agree with Knuth and Lipton that we’re far from understanding the limits of efficient computation, and that there are further surprises in store.
- What we've seen in this month is surprising unexpected algorithms, like o(n log n) algorithms for FFT and integer multiplication, and now this sub–exponential-time algorithm for 3-SAT. (Yes the ETH has had less support than P≠NP but I still gather it's unexpected.)
So (though I don't expect intuitions on P≠NP to change, and probably that's still the right intuition), maybe there will be a bit of a swing of the pendulum back to the idea that maybe clever algorithms exist that we just haven't thought of.
The fact that other math problems are bulk solved yet nothing from cryptography was released/published as solved could be a strong indicator that "all modern public key based cryptography" could have been at risk.
I mean if someone develops an O(n^100) algorithm for solving an NP-complete problem, that would prove P=NP, but pose no immediate threat to cryptography.
It increases risk, in that a galactic algorithm increases the odds that an efficient algorithm exists waiting to be found, but it is entirely (epistemically) possible that the galactic algorithm exists but the efficient algorithm doesn't.
Given the current developments, I would not be confident that it would not lower dramatically pretty soon.
Here's a nice low-budget scifi movie about this:
https://en.wikipedia.org/wiki/Travelling_Salesman_(2012_film...
https://www.youtube.com/watch?v=OgeyHkP6i3g (full movie)
https://www.goodreads.com/book/show/37986693-factor-man
In cryptography, most of the problems believed to be exponentially hard, such as factoring, don't have proofs of an exponential lower bound. There's some risk of a breakthrough that makes those problems easier. A probabilistic solution would be good enough to break many cryptosystems.
I suspect that people at the cryptanalysis agencies worldwide are staying up late tonight, trying to figure out the impacts and risks of this new result.
> In fact, one of the first things I wanted OpenAI to check is that no existing cryptosystems are compromised by the result. Apparently they are safe...
so if you trust his trust of OpenAI, then this has no (immediately obvious) bearing on cryptosystems for now.
It also not probabilistic, so linear programing is not rely a good example
These are by far not most, nor most of the practically useful ones. Directed graph reachability, given an adjacency list as input, is NL-hard.
It is notable that simplex, especially dual simplex is often better in practice, especially for integer linear programming with branch and bound because you only have to slightly modify the parent optimal basis in simplex vs entirely restarting or refactorizing for interior point methods.
Actually, integer factorization already has subexponential-time algorithms, such as the General Number Field Sieve (GNFS).
There are subexponential algorithms for integer factorization, e.g., the general number field sieve and Lenstra’s ECM.
> In 2019, Harvey and van der Hoeven came up with a galactic algorithm with complexity O ( n log n ) {\displaystyle O(n\log n)}. This matches a guess by Schönhage and Strassen that this would be the optimal bound.
> In September 2026, OpenAI released a preprint claiming an algorithm with complexity O ( n ( log n ) α ) {\displaystyle O(n(\log n)^{\alpha })} with α < 1 {\displaystyle \alpha <1}
https://en.wikipedia.org/wiki/Multiplication_algorithm
That said, it's interesting to contemplate that there is "space between" polynomial and exponential time complexities.
Yes, svat’s comment accounts for that by using little-o notation* in o(n log n). This was only achieved in the 2026 result.
* https://en.wikipedia.org/wiki/Little_o_notation
https://www.informit.com/articles/article.aspx?p=2213858&WT....
> Don Knuth: As you say, I've come to believe that P = N P, namely that there does exist an integer M and an algorithm that will solve every n-bit problem belonging to the class N P in n**M elementary steps.
…Oh, I see. The M here is a hypothesized constant, with n as the base. That's how polynomial time looks. Exponential time would have n in the exponent.
[0] Very much not something I care much about in these sorts of conversations. Even if it weren't galactic, a random logn factor here and there usually loses to cache efficiency and whatnot. Their speedups were less than that, and I'd already assumed we were in the realm of applied math rather than real engineering.
Not quite. This result uses polylog(n) precision which isn't bounded, so it escapes the previous bounds. This doesn't noticably affect the runtime because not all of the math needs this extra precision (in this case you need a couple computations in higher precision but less than n operations), so the extra precision can be payed for even in a TM/circuit model.
> Our result is purely asymptotic: for arbitrary 3-CNF formulas with m = 5n clauses, the bound we prove does not improve on the 2n of exhaustive search unless n > 10^10000,
So their algorithm beats standard algorithms only when you have more variables than there are atoms in the universe (10^80). Even if every atom contained a whole universe, and each atom in each universe itself contained a whole universe, iterated 100 times, your still not there.
https://news.ycombinator.com/item?id=50255159
— Scott Aaronson