AI News

OpenAI model proves that Landau-Siegel zeros do not exist; academia discusses its distance from P≠NP

36Kr · AIUpdated Nov 9 20:20

AI summary

OpenAI's internal model produced a proof that "Landau-Siegel zeros do not exist"; the original proof was 8 pages, and mathematicians subsequently compressed it to 5 pages. On October 6, OpenAI released 722 manuscripts on GitHub (withdrawing 3 the next day), grouped into 372 result families; the internal model attempted about 4,000 open problems, averaging about 3 hours of ChatGPT Pro compute per problem. Lance Fortnow pointed out that AI's solve rate for given problems is less than 10%, none of these results brought academia any closer to solving P vs NP, and only about 42% came with Lean formal proofs.

Why read this

AI score 74

The original article outlines the scale of math manuscripts mass-produced by OpenAI's internal model and the academic community's skepticism, so readers can judge the real boundaries of progress in AI conquering open problems.

Topics

Tags

Body · AI translation

A number theory mystery that has stumped humanity for decades—did AI crack it in just 8 pages?

Over the past two days, the entire mathematics community and the tech world have been in shock.

On October 18, Stanford number theorist Jared Duker Lichtman set the internet ablaze with a post on X—

"The existence of such a proof strongly suggests that P≠NP."

The trigger for this statement was the astonishing answer just delivered by an internal OpenAI model—a proof that "no Landau-Siegel zero exists."

This key problem in analytic number theory has been stuck for decades, and the academic community widely regards it as extremely difficult.

As it turned out, not only did OpenAI solve it—the original proof was only 8 pages long!

And the excitement didn't stop there. Top mathematicians quickly followed up:

Bill Banks of the University of Missouri first wrote a detailed 19-page breakdown;

then Bryce Kerr, using biquadratic extensions

and Masser's classical results, compressed the proof to 6 pages;

after that, University of Oxford professor Ben Green trimmed it further, directly producing a 5-page ultimate version!

From decades of hopeless deadlock to a mere 5 pages. Abhishek Saha, a professor at Queen Mary University of London, posed a soul-searching question:

If this problem can be proved in just 5 pages, why did mathematicians miss it for decades?

Finding is like climbing to heaven, verifying is like drinking water

Why would a 5-page number theory proof make a Stanford professor immediately think of the holy grail of computer science—P vs NP?

The reason is simple.

A 5-page proof can be read and verified by any expert in a few hours; but for those 5 pages, humanity's brightest minds struggled for decades.

"Finding is extremely hard, verifying is extremely easy." These eight words are precisely the underlying intuition of P≠NP!

In the comments, Anthropic researcher Joshua Batson pointed out incisively—

Behind the joke lies an extremely serious question: when AI delivers over 700 mathematics manuscripts within weeks, is the millennium prize problem P vs NP also coming soon?

Three high walls have locked humanity in for fifty years

To explain P vs NP in one sentence: can every problem whose answer can be verified quickly also have its answer computed quickly?

P is the set of problems that "can be solved quickly," and NP is the set of problems that "can be verified quickly."

Sudoku, scheduling, and codebreaking all belong to NP: give you the answer, and you can check whether it's right in a second; but starting from scratch, you can only do something close to exhaustive search.

If P=NP, modern cryptography collapses instantly, global finance is left exposed, but all complex problems would gain an ultimate shortcut;

if P≠NP, it means the universe contains problems that are inherently impossible to shortcut.

This problem was posed in 1971, with a $1 million prize. The vast majority of scientists believe P≠NP, but for over fifty years no one has been able to prove it.

Why is it so hard? Because humanity has not only failed to find the key—it has proved that existing lock-picking tools are doomed to fail.

These are the famous three "barriers":

1. The relativization barrier (1975): classical methods based on simulation and diagonalization simply cannot solve it;

2. The natural proofs barrier (1994): most proof approaches for circuit lower bounds, if they could prove P≠NP, would incidentally break pseudorandom generators, leading to self-contradiction;

3. The algebrization barrier (2008): upgraded methods incorporating algebraic techniques were also proved to be insufficiently powerful.

In June this year, complexity theorist Lance Fortnow pessimistically stated: at present, humanity does not even have a viable line of attack.

He suggested that challengers first try to prove a much simpler problem such as "SAT has no quadratic-time algorithm," then added the finishing blow: "You won't succeed there either."

On October 6, OpenAI released 722 manuscripts at once on GitHub (withdrawing 3 the next day), grouped into 372 result families.

The internal model attempted about 4,000 open problems, averaging only about 3 hours of ChatGPT Pro compute per problem.

Theoretical computer science accounted for 40 families. Although it did not directly take down P vs NP, several papers landed precisely on the main road leading to it.

In June this year, complexity theorist Lance Fortnow wrote: apart from simulation, we have no way to handle general polynomial-time algorithms; circuit complexity, proof complexity, algebraic geometry, and other routes are all stuck far below the full power of polynomial time. His conclusion was that at present there is not even one viable line of attack. Coincidentally, OpenAI's six results this time each planted a flag on one of these routes.

Family No. 112 (depth-3 circuit lower bounds): the most favored route to proving P vs NP is to prove that some NP problem requires "super-polynomial-size circuits." Humanity has long been stuck on explicit lower bounds for depth-3 circuits at

Family No. 102 (Unique Games Conjecture, UGC): proved this core conjecture proposed by Subhash Khot in 2002 (including partial Lean verification). It is closely adjacent to P vs NP and greatly strengthens the inapproximability limits of many NP-hard optimization problems.

Foundational algorithms and derandomization: achieved full derandomization of randomized logarithmic space (L=RL=BPL); compressed the matrix multiplication exponent to

; integer multiplication and DFT fell below

time.

Around the 102nd "Unique Games Conjecture," a particularly stinging episode also unfolded.

Dana Moshkovitz, the wife of the renowned scholar Scott Aaronson, devoted her entire career to this conjecture. On the night of the release, their 9-year-old son ran over to mock his mother:

"I heard a robot solved the problem you've been researching your whole life!"

Quanta Magazine also disclosed that MIT's Dor Minzer team, after hearing rumors about OpenAI on September 11, rushed to release their own proof of a related variant before OpenAI's release so that a machine wouldn't scoop them.

Lance Fortnow wrote the next day that AI's solve rate on the given problems was under 10%, and that none of these results brought us any closer to resolving P vs NP.

Although Scott Aaronson called October 6 "one of the most important days in the history of mathematics," he also pointed out that the list contained not only no P≠NP, but not even "smaller" hard problems like P=BPP—clearly not because no one had tried.

He revealed on a podcast that OpenAI and Anthropic are racing to crack P vs NP using their strongest internal models, but neither has succeeded yet.

Viewed calmly, there are still four major obstacles before a breakthrough:

1. Verification is far from complete: only about 42% come with Lean formal proofs; the rest rely entirely on humans. Aaronson admitted that almost no human has truly read any of the long proofs.

2. Readability is extremely poor: Moshkovitz described the AI's proof of the Unique Games Conjecture as written "so badly that you can't understand it without AI."

3. Lean is not omnipotent: machines can only guarantee that the reasoning is correct; they cannot guarantee that the formal definitions match the original problem exactly, nor can they judge whether the result is important.

4. The core route remains blank: the AI breakthroughs are in depth-3 circuits and algebraic circuits, while the "general Boolean circuit lower bounds" that could imply P≠NP remain unwritten.

First, the "hardness" of hard problems has been repriced. The 5-page proof of the Siegel zero shows that some dead ends lack the one step no one thought of. How many conjectures are still just one layer of window paper away? No one is sure.

Third, the scale and speed are not on a human order of magnitude. 4,000 problems, averaging 3 hours each. When major labs aim their next-generation strongest compute squarely at P vs NP, quantitative accumulation is highly likely to trigger qualitative change.

As for the final outcome, an interesting divergence in bets has even emerged in academia.

Aaronson still bets heavily on P≠NP; but some scholars believe that if AI really can break through, the answer is more likely to be P=NP—because having AI search out a clever algorithm is far easier to imagine than using existing mathematics to prove "impossible."

However, the most intriguing humor of the whole affair is this—

The more powerful AI becomes, the more it itself looks like living evidence for P≠NP.

It expends massive compute to "search for" proofs, while humans spend a few hours to "verify" the results.

If one day AI really burns through its compute and proves P≠NP, that would be the most perfect self-proof of this millennium problem.

This article comes from the WeChat public account "Xinzhiyuan," author: ASI Revelation, published by 36Kr with authorization.

Source

36Kr · AI