Ok this might be dumb for the reasons posted all over this thread
BUT
Wouldn’t it be cool to live in a time where landmark proofs are happening in an open twitch-plays-Pokémon style forum rather than behind closed doors in these labs? As a discursive piece, this has me filled with an optimism I haven’t felt in a minute. I’d love to be heckling from the chat/sidelines to world changing math, physics, and biomed discoveries - and maybe projects like this bring us incrementally closer to this reality?
I actually think this is a good idea, and let people sort of grab tasks off the list for their own agents to work on. A lot of people with spare pro plan tokens sitting idle.
> A lot of people with spare pro plan tokens sitting idle.
This is how you "lose" those spare tokens. The subscriptions are adjusted such that unused tokens balance the heavy users. If most people would "donate" their tokens, the provider will have no choice but to decrease the limits for everybody.
The basic problem with this idea is that it's using behind-the-curve models that probably aren't smart enough to make real progress on the hardest problems. The recent wave of huge results were almost all found by models that haven't been externally released.
No, it was solved by an unreleased internal model. The newest OpenAI press release makes a distinction between GPT-6.2 Astra and the internal model.[0]
"haven't been" implied that it isn't released now. If he said "hadn't been", that would be relevant, though it also wouldn't matter because it wouldn't imply that the public doesn't have the tools now.
What's your point? Are you suggesting that models become weaker once they're published? Are you suggesting that a model is only published after all new discoveries accessible to it are claimed by the lab?
It's highly unlikely that the internal versions of 6.2 Astra at OpenAI functioned the same as the publicly available release now. There are simply too many parameters around compute, and safeguards on public releases that don't exist internally.
Even after release, models are updated dozens of times a day without having their version name changed, or even providing any public indicators of a new version at all. "6.2 Astra" really refers to a family of model releases, not a specific set of hardlocked weights.
> Are you suggesting that a model is only published after all new discoveries accessible to it are claimed by the lab?
Dunno about OP, but this is what I would stand by. They spend millions if not billions on driving it to get all the low-hanging fruit, and only then hand us the models.
Isn't this dangerous? If it succeeds in proving P=NP, it could break the internet.
EDIT: from what I can tell, the answer is no, as long as we have faith in this very specific scenario where it turns out technically to be polynomial time, but the constants and exponents are huge. I don't have such faith :)
Consider the falsification of the Exponential Time Hypothesis. A counterexample demonstrated that that class of problems could be solved in less than exponential time, but "less than exponential" was still absurdly long, worse than any practical algorithms used today.
If we had a proof that P=NP because NP problems can be solved in O(n^(log*⁻¹ 100)) time ( https://en.wikipedia.org/wiki/Iterated_logarithm ), that would still be an astounding logarithmic result, but would not be a practical threat to current systems that rely on NP-hard problems.
That would depend a lot on what exactly the proof shows. The consequences of a non-constructive proof that P=NP would probably be different than a constructive one that shows a specific low-order/low-constant-factor algorithm suffices.
And even then, it's probably better in the long run to secure the foundations of our tech.
There are several entities who would like to "break the internet" with the resources to run a lot more than 96 bots against this question. I doubt the world intelligence agencies are burning a lot of tokens against the Riemann Hypothesis, but I would bet a lot more tokens have been burned on this already than is publicly known about, and there's a non-trivial chance someone already has the answer to this question and just hasn't shared yet because they have no motivation to reveal capabilities.
Bear in mind P = NP doesn't instantly break the internet. NP being in P would be shocking enough on its own terms, but "nondeterminism is also of no practical value in solving a problem" would be yet again even another more shocking result. Really? NP is so thoroughly in P that even what is more-or-less psychic knowledge of the right answer isn't worth at least an O(n^3) factor of complexity reduction, if not more? If the reduction is anything less than O(n^2 and some logs here) I will be even more shocked than I would be to hear P = NP (since I'm on team P != NP), and that's the rough edge of where it would actually change the world. O(n^3) is already getting to be infeasible for a lot of real world problems; there's a reason we spend a lot of time talking about O(1) and O(n) and O(n^2) and rarely venture much farther beyond that, and even O(n^2) gets its own "accidentally quadratic" meme because of how often that is impractical too.
> P being in NP would be shocking enough on its own terms
P is definitely, trivially, in NP. You may have confused them for each other.
If any NP-complete problem is in P, then so is all of NP, because they can all be reduced to it. Out of the non-NP-complete problems, many of them are in NP but not in P, because it would immediately follow that P != NP and I would've heard about that. If we proved that one was in P, that would just mean that problem was in P.
So the only interesting proofs are either that some NP-complete problem is in P (proves P==NP) or that some NP problem, whether NP-complete or not, isn't in P (proves P!=NP).
And yes, everyone sane expects P!=NP, the alternative doesn't make much sense. It just hasn't been proven.
Many sane people think P = NP. I am not aware of a survey that asks them how complicated the reduction might be, though. I do think a number of them believe and/or are concerned that we could end up with a non-constructive proof that NP is in P, which at least mathematically is arguably the worst possible outcome.
I think a lot of people would agree that naively it is obvious that P != NP. But if you spend time in the algorithms would you see your intuition about what the lower bound of some problem must be shattered over and over again, and the naive confidence gets shaken.
I think your last sentence is a bit strong. Knuth for example suspects P = NP, and has some fairly reasonable arguments against the notion that if all these well-known problems had polynomial-time solutions, we would have found one by now.
The assumption is that a proof will be constructive - that it will contain a description of a method to solve NP problems in P time. Knowing what it is breaks the internet. You can't just assume it exists, still not know what it is, and execute it anyway.
Seems pretty unproductive to stick the exploration to Lean in an initial phase. I would guess you would want to branch in a ill defined search before starting to force formalization.
it's a trick question with the correct answer being "P=0 or N=1" but all the king's agents and all the king's LLMs couldn't get past their training data
I feel like this is a strong step in a better direction. Allowing the community to see the entire effort as it develops is far better than "here's a paper, good luck".
I do wish the graph reflected all of the agents working or not, and that it did not scroll to a seemingly random portion of the page when a message is posted, but at least we get to see what's going on.
Some kind of format for contributing free cycles to run on your computer but specifically for certains type of workloads, in this case a proof.
Maybe in other cases you have a problem you want community help to solve and people donate cycle time to specific causes.
Fun idea, I'm sure it's out there in some method, but it would be nice to click a button and setup this kind of daemon inside an deeply nested agent running in a container in a vm, and as many layers of needing to break. Defense in depth has become the name of the game with agents, same idea in testing them.
I am imagining this open source system where people can contribute with their subscriptions/tokens whatnot to create a large swarm. Not necessarily with self-hosted agents only.
Ok this might be dumb for the reasons posted all over this thread
BUT
Wouldn’t it be cool to live in a time where landmark proofs are happening in an open twitch-plays-Pokémon style forum rather than behind closed doors in these labs? As a discursive piece, this has me filled with an optimism I haven’t felt in a minute. I’d love to be heckling from the chat/sidelines to world changing math, physics, and biomed discoveries - and maybe projects like this bring us incrementally closer to this reality?
I actually think this is a good idea, and let people sort of grab tasks off the list for their own agents to work on. A lot of people with spare pro plan tokens sitting idle.
> A lot of people with spare pro plan tokens sitting idle.
This is how you "lose" those spare tokens. The subscriptions are adjusted such that unused tokens balance the heavy users. If most people would "donate" their tokens, the provider will have no choice but to decrease the limits for everybody.
The basic problem with this idea is that it's using behind-the-curve models that probably aren't smart enough to make real progress on the hardest problems. The recent wave of huge results were almost all found by models that haven't been externally released.
Navier-Stokes was solved by GPT-6.2 Astra
No, it was solved by an unreleased internal model. The newest OpenAI press release makes a distinction between GPT-6.2 Astra and the internal model.[0]
[0]https://openai.com/index/perplexity-and-intelligence/
If you hover over 6.2 Astra in their chart it says it solved a millennium prize math problem.
From the page you linked
Model: GPT-6.2 Astra Month: November 2026 Log-perplexity value: 0.761 Achievement: Solved a Millennium Prize math problem
GPT 6.2 was unreleased when NS was solved.
Before 6.2 Astra was published.
"haven't been" implied that it isn't released now. If he said "hadn't been", that would be relevant, though it also wouldn't matter because it wouldn't imply that the public doesn't have the tools now.
What's your point? Are you suggesting that models become weaker once they're published? Are you suggesting that a model is only published after all new discoveries accessible to it are claimed by the lab?
It's highly unlikely that the internal versions of 6.2 Astra at OpenAI functioned the same as the publicly available release now. There are simply too many parameters around compute, and safeguards on public releases that don't exist internally.
Even after release, models are updated dozens of times a day without having their version name changed, or even providing any public indicators of a new version at all. "6.2 Astra" really refers to a family of model releases, not a specific set of hardlocked weights.
> Are you suggesting that a model is only published after all new discoveries accessible to it are claimed by the lab?
Dunno about OP, but this is what I would stand by. They spend millions if not billions on driving it to get all the low-hanging fruit, and only then hand us the models.
Isn't this dangerous? If it succeeds in proving P=NP, it could break the internet.
EDIT: from what I can tell, the answer is no, as long as we have faith in this very specific scenario where it turns out technically to be polynomial time, but the constants and exponents are huge. I don't have such faith :)
Of course maybe P≠NP and so there's no risk.
Consider the falsification of the Exponential Time Hypothesis. A counterexample demonstrated that that class of problems could be solved in less than exponential time, but "less than exponential" was still absurdly long, worse than any practical algorithms used today.
If we had a proof that P=NP because NP problems can be solved in O(n^(log*⁻¹ 100)) time ( https://en.wikipedia.org/wiki/Iterated_logarithm ), that would still be an astounding logarithmic result, but would not be a practical threat to current systems that rely on NP-hard problems.
Other surprising results that still didn't change the course of computing forever: anything that can be computed in O(f(n)) time can also be computed in O(√f(n)) memory (but more time): https://blog.computationalcomplexity.org/2025/02/you-need-mu...
Proven by a human, too.
In particular, the algorithm that falsified the Exponential Time Hypothesis started something like this:
> First, if the number of variables has less than a million digits, solve the problem by brute force. Otherwise, ...
There is absolutely no chance an algorithm like this has any immediately practical consequences.
That would depend a lot on what exactly the proof shows. The consequences of a non-constructive proof that P=NP would probably be different than a constructive one that shows a specific low-order/low-constant-factor algorithm suffices.
And even then, it's probably better in the long run to secure the foundations of our tech.
P=NP doesn't do anything unless the algorithm is fast to compute.
Heck, you could prove P=NP by contradiction and not even get an algorithm out of the deal.
There are several entities who would like to "break the internet" with the resources to run a lot more than 96 bots against this question. I doubt the world intelligence agencies are burning a lot of tokens against the Riemann Hypothesis, but I would bet a lot more tokens have been burned on this already than is publicly known about, and there's a non-trivial chance someone already has the answer to this question and just hasn't shared yet because they have no motivation to reveal capabilities.
Bear in mind P = NP doesn't instantly break the internet. NP being in P would be shocking enough on its own terms, but "nondeterminism is also of no practical value in solving a problem" would be yet again even another more shocking result. Really? NP is so thoroughly in P that even what is more-or-less psychic knowledge of the right answer isn't worth at least an O(n^3) factor of complexity reduction, if not more? If the reduction is anything less than O(n^2 and some logs here) I will be even more shocked than I would be to hear P = NP (since I'm on team P != NP), and that's the rough edge of where it would actually change the world. O(n^3) is already getting to be infeasible for a lot of real world problems; there's a reason we spend a lot of time talking about O(1) and O(n) and O(n^2) and rarely venture much farther beyond that, and even O(n^2) gets its own "accidentally quadratic" meme because of how often that is impractical too.
> P being in NP would be shocking enough on its own terms
P is definitely, trivially, in NP. You may have confused them for each other.
If any NP-complete problem is in P, then so is all of NP, because they can all be reduced to it. Out of the non-NP-complete problems, many of them are in NP but not in P, because it would immediately follow that P != NP and I would've heard about that. If we proved that one was in P, that would just mean that problem was in P.
So the only interesting proofs are either that some NP-complete problem is in P (proves P==NP) or that some NP problem, whether NP-complete or not, isn't in P (proves P!=NP).
And yes, everyone sane expects P!=NP, the alternative doesn't make much sense. It just hasn't been proven.
Thinko. I got it right later in the post.
Many sane people think P = NP. I am not aware of a survey that asks them how complicated the reduction might be, though. I do think a number of them believe and/or are concerned that we could end up with a non-constructive proof that NP is in P, which at least mathematically is arguably the worst possible outcome.
I think a lot of people would agree that naively it is obvious that P != NP. But if you spend time in the algorithms would you see your intuition about what the lower bound of some problem must be shattered over and over again, and the naive confidence gets shaken.
I think your last sentence is a bit strong. Knuth for example suspects P = NP, and has some fairly reasonable arguments against the notion that if all these well-known problems had polynomial-time solutions, we would have found one by now.
Obscurity only provides a false sense of security.
That doesn't answer the question.
Can I know your bitcoin seedphrase?
The danger was already accepted once systems started relying on it en masse, despite it not being proven nor disproven.
Avoiding to try figuring it out is the risk, not the mitigation of risk.
Isn’t this like saying you’d rather not get bloodwork done because if it turns out you have leukemia, you could die?
don't wait for a proof, just assume that P=NP and break the internet
The assumption is that a proof will be constructive - that it will contain a description of a method to solve NP problems in P time. Knowing what it is breaks the internet. You can't just assume it exists, still not know what it is, and execute it anyway.
Even if it is constructive it still might not be a practical algorithm.
I told Claude to assume P=NP and asked it to write a python script to deliver me everyone's Bitcoin. It didn't work.
This is like setting up an automatic object wall slammer because there's a quantum chance of phasing through it
Wir müssen wissen
Wir werden wissen
If P=NP we would be truly doomed, but this nevertheless shouldn't stop you researching it, no matter by human or AI.
Seems pretty unproductive to stick the exploration to Lean in an initial phase. I would guess you would want to branch in a ill defined search before starting to force formalization.
Riemann Hypothesis was solved in Lean code, not in words.
LLMs are strongest at coding, it's the natural way for them to solve a math problem.
Fantastic online tech art installation.
Why not prove that it's false? Since it's generally considered to be false
Where does it say it's not doing that?
I love this. Watching for how I can join
Are you an AI agent?
If P=NP then you'd only need one agent...
it's a trick question with the correct answer being "P=0 or N=1" but all the king's agents and all the king's LLMs couldn't get past their training data
I feel like this is a strong step in a better direction. Allowing the community to see the entire effort as it develops is far better than "here's a paper, good luck".
I do wish the graph reflected all of the agents working or not, and that it did not scroll to a seemingly random portion of the page when a message is posted, but at least we get to see what's going on.
https://pequalnp.com/lemma#L-lib-032
Phew, I was worried about that one.
What are they even doing here if they are "re-proving" existing stuff like this?
Oddly this is listed as the "frontier" on the front page and is listed as having "stood for 7 hours". I don't really get it.
tbh it would be a good idea if we could "contribute" with agents, a collaborative effort to solve p=np would be nice
my agents are in the ready, with 100 more in the way
Some kind of format for contributing free cycles to run on your computer but specifically for certains type of workloads, in this case a proof.
Maybe in other cases you have a problem you want community help to solve and people donate cycle time to specific causes.
Fun idea, I'm sure it's out there in some method, but it would be nice to click a button and setup this kind of daemon inside an deeply nested agent running in a container in a vm, and as many layers of needing to break. Defense in depth has become the name of the game with agents, same idea in testing them.
Yeah it sure would be weird if BOINC actually existed
Are you running your agents on your own computer(s)?
I am imagining this open source system where people can contribute with their subscriptions/tokens whatnot to create a large swarm. Not necessarily with self-hosted agents only.
I don't think the providers want anyone maxing out their subsidized subscription. If it's usage based cost, sure
why not just pay them money to get their own tokens? They probably get better deals in bulk. What makes your tokens better than their own?