Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math

751 points by scientificamerican 9 hours ago on reddit | 200 comments

[OP] scientificamerican | 9 hours ago

Levent Alpöge, a mathematician and employee at artificial intelligence company Anthropic, claimed on Friday to have proven the most famous problem in all of mathematics using an LLM.

“I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” Alpöge wrote in a statement posted on his website.

The Riemann hypothesis was first proposed in 1859 by the mathematician Bernhard Riemann. It claims that the locations of the prime numbers along the number line, which appear random at first blush, are actually determined by an intricate, meaningful order.

The conjecture has since become the North Star for the entire field of number theory, and is the most famous open problem in math.

Fit_Salamander_2814 | 9 hours ago

Would solving it lead to any practical real-world applications, or is this more like a brain-teaser for math nerds?

michaelh98 | 9 hours ago

Can't answer your question but a ton of technology we take for granted is based in part on "brain teasers for math nerds" from the last several hundreds of years

Fit_Salamander_2814 | 8 hours ago

I don't really know about that. We - I mean humanity - have gotten so good at rune-smithing, that we can make sand think.

ReddusMaximus | 7 hours ago

Sometimes it's obvious. MRI scans (specifically the imaging) wouldn't have been possible with 1970s technology without progress in math in the previous decades.

Ok-Sort251 | an hour ago

Just gonna hype up Nottingham whilst we're here. Also go Maths!

michaelh98 | 8 hours ago

I thought this sub was everything science

SemanticTriangle | 6 hours ago

Lithography translates roughly as 'the writing in stone' and is the process by which we impose the patterns on semiconductors which allow them to function. Microchips are carved runes.

Jameseesall | 8 hours ago

But those advances came about from working on the brain teaser problems, not necessarily solving them. Inventing new maths to creatively approach a teaser is useful, solving a hypothesis we assume to be true with a prompt is not useful.

"There is no superlative strong enough for how this would transform mathematics" - the article

"Nuh uh" - some redditor

Deathpacito-01 | an hour ago

Redditors suddenly flip to become science-deniers when AI is involved

hysys_whisperer | 9 hours ago

Yes, it would be very beneficial in codebreaking, hacking, and general destruction of security.

SamG101_ | 8 hours ago

Riemann is already assumed to be true, it wouldn't help codebreaking, which is factorization and discrete log (mostly, at the moment)

GreeneyesCB | 8 hours ago

Rie*

SamG101_ | 8 hours ago

Oh yh mb

devi83 | 8 hours ago

Encryption depends on how hard it is to factor large numbers, and this result doesn’t make factoring easier.

johpick | 7 hours ago

This. A contrary result, which wasn't expected, could have posed a security risk. Solving the Riemann hypothesis as true confirms security measures.

NortherneWynd | an hour ago

it's not about the result, it's about the ability of the machine to find the result. if AI can solve Riemann now, what will it be able to do next year?

the_c_train47 | 4 hours ago

Me when I authoritatively spread misinformation without knowing what I’m talking about

GreeneyesCB | 9 hours ago

That is completely false.

Tiny_TimeMachine | 7 hours ago

But it gets the people going! It's sounds so cool and authoritative.

Transocialist | 6 hours ago

It's provocative!

HopeOfTheChicken | 6 hours ago

This is very much not the case. Like not at all. It would tell us more about the distribution of prime numbers, but that doesn't affect the security of cryptography in any significant way

geekguy | 3 hours ago

How would knowing the distribution not affect security? Say for a given key if sufficient strength. We can make assumptions on the intended entropy of the key. Bounding that within a distribution range you could come up with a hash table of the primes within that range and thus increase the speed of offline attacks. Unless you are saying that even with the distribution the number of keys available would still be so large it’s unreasonable to compute? Or something else?

Imperial_Squid | 2 hours ago

> Unless you are saying that even with the distribution the number of keys available would still be so large it’s unreasonable to compute

Correct

Taking AES-256 as an example, the keys it uses are 256 bits long (but don't need to be prime), so anything from 0 to 2^(256) is a candidate, if you tried a million keys a second it would still take you 2^(256) / (1000000 * 60 * 60 * 24 * 365) = 3.67 * 10^(63) (ie a 3 and then 63 zeroes) years to go through all of them, the universe has only existed for 13.8 billion years, 0.00000000000000000000000000000000000000000000000000000003758% of that timer

RSA, which is oen of the algorithms that uses primes, uses two primes each of half the bit length of the target (eg RSA-2048 uses two 1024 bit primes), so those numbers are anywhere between 0 and 2^1024, even accounting for the fact it can only pick primes, the scales on all of this stuff iss so ridiculous it's not a concern

There would need to be results that fundamentally change our understanding of cryptography in order for these algorithms to be under threat (the Riemann Hypothesis thing actually strengthens these since most mathematics relies on it being true and the proof supposedly proves it's true), as for stuff like quantum computers, we're still decades out from that tech being viable for actual hacking and replacement algorithms are already in development!

Miserable_Archer_769 | an hour ago

And all quantum computing would do from my understanding is essentially take those insanely large number pools and run through them and crack them almost brute force style instantly.

But that another can of worms for later with people already saving our encrypted data waiting for quantum computing

CSE_Major | 6 hours ago

No it won’t really have much of a immediate impact. It doesn’t break modern encryption nor does it provide results for faster factoring.

Piskoro | 8 hours ago

are you thinking of P = NP?

HeraThere | 9 hours ago

oh how fun

CariniFluff | 8 hours ago

They're completely wrong. It would do absolutely nothing to impact encryption/breaking encryption.

ecafsub | 7 hours ago

Cootys rat semen?

dubloons | 8 hours ago

A lot of the answer to this question likely depends on how it was solved.

arealuser100notfake | 9 hours ago

It would look pretty when drawn on a carpet

GreeneyesCB | 8 hours ago

Of course not, theorems can already just be assumed to be true. For any practical purposes its correctness has been numerically checked to be true already. A proof is also not expected to include any practically useful mathematical tools.

The Riemann hypothesis implies that the bounds of some estimates for the prime number counting function are tighter than what can currently be proven, this is not very useful for practical applications because we can literally just use a computer in the numerical ranges that are useful and just literally count and calculate the exact difference. An exact difference is of course superior to any bound.

No, there is no reasonable way for this to have a practical implication.

I also want to mention that unless there is a Lean-verification AI easily outputs complete mathematical nonsense that it very confidently claims is a proof, in particular when prompted by a non-mathematician. Students at university can very easily cheat with AI because those problems are well known, formulated well and are already proven correct.

Fizrock | 7 hours ago

FYI he has posted a Lean verification of the problem.

John02904 | 7 hours ago

I would argue thats not always entirely true. While you are right that answering the specific question posed probably never has a benefit, often methods and ideas are invented or created? (Some ambiguity around those type of terms) can offer practical applications. A good example would be something like newton creating calculus to solve questions about gravity.

GreeneyesCB | 7 hours ago

"often methods and ideas are invented or created? (Some ambiguity around those type of terms) can offer practical applications." I did address that, there is no reasonable assumption that can be made here that these are related, the areas are so far away from each other and so are the methods used. An AI proof is also the least insightful and the least likely to develop new methodology so that's also a particularly poor case for that.

John02904 | 7 hours ago

Oh i agree. AI with these proofs seems mostly to be applying existing knowledge and methods way outside the field that hadn’t been considered yet. What i think might be a way more useful or interesting use would be to go through tons of things we have already proven and find new methods to prove them outside of what we had already used.

rhubarb_man | 8 hours ago

Yes. Alpoge got claude to prove this so he could attach his name. There will be pretty much no benefits

credulous_satan_6666 | 6 hours ago

When the most significant and difficult open problem in math has 0 practical applications...you kind of have to wonder how much wasted effort there is in the field.

GreeneyesCB | 6 hours ago

That's a dumb stance. It's significant for historical purposes, we don't really establish a "most significant" problem, there are thousands of extremely valuable problems in mathematics with varying amounts of practical applications. Saying that it is "the most difficult" is also very unobjective and unverifiable, it is almost certainly not the case if you try to find a reasonable measure.

That is not wasted effort.

credulous_satan_6666 | 6 hours ago

Maybe, I'm not a math SME. But Navier Stokes, Riemann, and other  problems solved by AI have all been claimed to be once-in-a-generation type problems, yet the comment sections are filled by mathematicians claiming no practical applications.

Is that just indicative of the type of problems solvable by AI?

Because otherwise it seems to me that thousands of mathematicians to date have spent entire careers on interesting but mostly irrelevant work...

GreeneyesCB | 6 hours ago

"But Navier Stokes, Riemann, and other  problems solved by AI have all been claimed to be once-in-a-generation type problems" That's a media issue, not a math issue.

"yet the comment sections are filled by mathematicians claiming no practical applications." I can't speak for all of them because there are too many by now, but for RH and NS that is certainly the case. In the case of Navier-Stokes they even went for the lowest hanging fruit that most mathematicians weren't even interested in but it technically fulfilled the requirements by the Clay institute and that was enough for publicity. They are not trying to engage in good faith with mathematics at all.

"Because otherwise it seems to me that thousands of mathematicians to date have spent entire careers on interesting but mostly irrelevant work..."

I don't agree on things being irrelevant when they "just" give a further understanding of mathematical structure. What people learn and how knowledge about things spread and inspires is very valuable. Mathematics obviously had a major major impact on scientific and engineering progress over thousands of years by now and it would not have been possible at all if math for the sake of itself did not exist.

The AI companies doing AI proofs are not interested in contributing to mathematics or in contributing to practical applications, they are interested in PR. It would be best if the media would stop giving them so much attention.

credulous_satan_6666 | 5 hours ago

Every field has a certain amount of "filler" research conducted. I'm just surprised to find that math has more than most.

I get that there is some value for these solutions within the field. Doesn't say much for value outside of the field though.

Atheizm | 9 hours ago

Prime numbers and cryptography.

GreeneyesCB | 8 hours ago

Having tighter bounds on the prime number counting function has no practical implications for cryptography in any way, it does not allow you to decompose composite numbers into primes. These are completely separae things.

CariniFluff | 8 hours ago

Right? Cryptography uses two already known "small" prime numbers to generate a larger number. Knowing the bounds for finding the next trillion digit prime number does absolutely nothing to break existing or even future encryption.

Just because this has to do with prime numbers and modern cryptography uses prime numbers doesn't mean that this automatically helps break encryption.

People are just talking out of their ass here.

Silent-Battle308 | 8 hours ago

Only if it were false. But he claims to have proven the hypothesis to be true so no.

Edit: We now know for sure that certain algorithms work as intended but we assumed they were functioning already so afaik nothing changed regarding real-world.

esmifra | 8 hours ago

Iirc a lot of cryptography is based on prime numbers. If this has any impact I don't know.

GreeneyesCB | 8 hours ago

It does not.

PracticallyPerfcet | 8 hours ago

It is an incremental step forward in math. It proves prime numbers aren’t randomly distributed, but it doesn’t give some magic formula to quickly compute primes or break encryption.

Silent-Battle308 | 8 hours ago

Since some ask about consequences.

If the riemanhypothesis is false it would imply prime numbers aren't as randomly distributed as we thought they are. Which could (could because I guess it could be only for very high n) be a problem for algorithms which relie on this fact i.e. cryptography.

If it were true it is "just" a milestone.

Since he claims to prove it to be true, not much is gonna change except a big question has been solved.  Perhaps there were some proves that assumed it to be true which also are confirmed now.

ReddusMaximus | 7 hours ago

There are lots of proofs that are based on the assumption that Riemann's hypothesis is true.

HopeOfTheChicken | 6 hours ago

Why would it being false imply that primes are more randomly distributed? I have watched my fair share of videos on this topic and I thought it being true would tell us more about prime numbers, wich makes them less random. Like it being true directly gives us a better estimate on the distribution of primes, how would that make them more random?

quad_damage_orbb | 5 hours ago

Can't they use AI to make antigravity or some other useful shit?

intergalacticscooter | 3 hours ago

Thing is with gravitons and I guess anti gravitons is, it isn't making a theory about them is impossible but discovering them experimentally, as they are would need to be so small it would create a black hole trying to interact with them.

This is a very poor description but it is just to give you an idea into the issues.

alyx_canmore | 2 hours ago

Imagine how close we were to “We proved Riemann’s Hypothesis before GTA VI” being true

matt_993 | an hour ago

apparently the proof was complete before the release of GYA, so you can say that!

jmclondon97 | an hour ago

Yes but wasn’t GTA 6 also done before release?

lewd_robot | 47 minutes ago

The saying is usually, "X happened before we got GTA6."

jmclondon97 | 24 minutes ago

True, but gta6 existed and was completed. Just a matter of getting our hands on it. Similar to the proof existing before released

lewd_robot | 47 minutes ago

We proved Riemann's Hypothesis before TES6.

We proved Riemann's Hypothesis before GTA7.

briankerin | 9 hours ago

Let me guess. Its so complicated that nobody knows if its correct or not.

AMixOfUpsAndDowns | 9 hours ago

He's actually getting held up on the "show your work" requirement

rustajb | 8 hours ago

AI is notoriously terrible at showing its work.

NotableCarrot28 | 8 hours ago

Bro has not heard of lean

Deep-Ad5028 | 6 hours ago

The (extremely long) code of lean formalization is also written by AI itself so it means less than what people think it does

onionsareawful | 6 hours ago

Lean is nice in that it is provably correct. The only potential issues are a lean bug (kernel bug), or an axiom / assumption that is actually incorrect. The problem statement is known to be fine as it's one that was written long before Anthropic existed lol.

It's likely Alpoge checked all the assumptions, so it's really just bugs now.

uraniumpi | 5 hours ago

I haven't checked the lean, but Alpoge claims no assumptions were made.

itsasecrettoeverpony | 4 hours ago

thats just not how math works, you need axioms to do anything at all in mathematics

NotableCarrot28 | 4 hours ago

Yeah there is a standard set of axioms that you don't really need to state you're working with in 99% of cases unless you're working on foundations

Eliv_nurotic | 3 hours ago

Just have an LLM check over it

PsecretPseudonym | 6 hours ago

The description of the problem is from a peer reviewed library and accepted to be correct. If the lean proof compiles, it means every step to the conclusion is a deterministic, logically valid step. If it reaches the conclusion from the problem statement and compiles, it’s almost certain to be correct.

Critical_Cap_120 | 6 hours ago

Not really. Lean has bugs. People have exploited them to show incorrect proofs of famous conjectures.

I think it's more likely that it's true that it is solved than not because of the Lean formalization, but it's not infalible.

PsecretPseudonym | 5 hours ago

That’s why you check with multiple different independent Lean kernels, which the author almost certainly already did and the world is now doing.

Otherwise you’re essentially stating that formal mathematical logic has bugs globally.

Critical_Cap_120 | an hour ago

What different Lean kernels do not share dependencies?

PsecretPseudonym | an hour ago

E.g., the primary one I believe is C++ and someone else just verified via run in the independently authored rust kernel.

HighOnGoofballs | 8 hours ago

Sounds like me in BC Calculus circa 1994

veshneresis | 8 hours ago

It’s porting a tool called theta lifts to get multiplicity for integers, and using that to enforce the square root bounds of the error (of the mangoldt function)

Specialist-Many-8432 | 7 hours ago

Does it work? No flame

veshneresis | 7 hours ago

I think it’s likely that it does. There’s a bunch of stuff that’s over my head they need to do to keep the complexity down as you perform more and more lifts, and so maybe there’s something broken there, but the general approach seems sound to me. I’m not qualified to say it works for sure, but i trust the mathematician. He’s the one who did the Jacobian conjecture earlier this year too.

ChironXII | 7 hours ago

So, yes

(Joke)

CocaineKeys | 9 hours ago

Not quite. Alpöge didn't just publish an incomprehensibly long mathematical argument and ask everyone to trust him. He released the actual Lean formalization in 14 downloadable parts.

See my other comment in this thread. https://www.reddit.com/r/EverythingScience/comments/1y159qk/anthropic_mathematician_claims_to_prove_riemann/19sxcljplu83n/

Elegant-Policy-5950 | 8 hours ago

Why does this read like opening a recipe to find 20 pages of the author reminiscing what it was like putting his grandmas robe on for the first time, while baking gingerbread houses and petting a goat. The recipe for fish filet.

Suspicious_Strain217 | 8 hours ago

When you solve one of the hardest problems feel free to be more concise

Bacon_Raygun | 6 hours ago

Did he solve it, or did the AI solve it?

Genuine question, what's the credit here look like?
Who did the brunt of the work?

onionsareawful | 6 hours ago

The AI didn't solve it end-to-end. For problems like these you need to be able to ask the right questions and know where to look.

Most math PhDs would not be able to get here with unlimited tokens tbh.

Notactualyadick | 6 hours ago

So what your saying is......that.....I am super smart because I ask so many questions. I knew it!

UwUHowYou | 5 hours ago

There are times when you are able to identify all the pieces of a puzzle, even if you are unsure how exactly they should be put together, you know they fit together.

For ex, I once got a macro out of AI that saved me 40+ hours of work and helped us make a deadline that was going to cost us like $50k. I don't think anyone else in the building could have gotten that macro from AI, because I did a huge amount of legwork structuring my data, working out a template and telling it exactly how it was structured, how to operate on it, etc.

It then filtered a couple thousand lines of data out of 10+ million lines, tagged it with each event that triggered it fetching it, etc. I basically needed 1 out of 5000 lines, and manually sorting it or filtering worked, but took like 8 hours to get a tiny chunk of it done. It was too involved.

I should get better at excel macros, but my knowledge of as400/hascript and data manipulation helped a ton and informed me of what to do before even asking AI.

And no, I would not call this prompt engineering as it genuinely required knowledge of the subject, where a lot of the work was done before even sending the first query to AI. Then there was checking the work to ensure it was good with the time it saved lol.

Notactualyadick | 3 hours ago

So what I'm getting is that.....um....that if I'm good at puzzles then I am a master 9f A.I, but only if I put work into my legs.

UwUHowYou | 3 hours ago

No, a PHD leveraging AI doesn't invalidate the PHD.

It likely took a massive part of his education to even work out what to have considered and what to not have considered.

PsecretPseudonym | 6 hours ago

Both. He solved it with the assistance of the AI to do work he couldn’t have done on his own.

The author knows the lean proof he published is the independent, reproducible verification that matters. He’s free to share his experience of the work and process of how he got there.

This likely means we can prove that it’s true, but there’s likely a lot of work to be done to simplify that proof. Just proving there’s a path at all is in some ways arguably one of the biggest achievements in pure math in the last century.

Suspicious_Strain217 | 6 hours ago

Go solve it yourself. I’m sure you’re gonna be able to prove this. These are insanely smart mathematicians.

It’s insanely naive to assume this is just some dude with unlimited tokens

Elegant-Policy-5950 | 5 hours ago

Well it genuinely sounds like his pal Claude solved it but it was really nice working together

Suspicious_Strain217 | 4 hours ago

Go solve one? There are many million dollar math puzzles.

ThatGodDamnXmasTree | 5 hours ago

You didnt answer their question.

Bacon_Raygun | 6 hours ago

Struck a nerve I guess.

Nothing you're mad about is something I actually said.

Edit:
Like holy shit the amount of reading inbetween the lines you're doing, because you want to get mad.
Hard to believe you're interested in science with how obsessed you are with your preconceived notions and reading comprehension.

Standoffish prick.

Slightly_Mungus | 5 hours ago

Ironic

M3K4N1X | 26 minutes ago

The AI did the brunt of the work.

Feeling-Screen-9685 | 5 hours ago

Math isn’t religion. People work hard to find ways to break proofs because if they can’t, then it helps us understand something bigger. Math isn’t faith based.

Crazy-Coconut7152 | an hour ago

So glad you cleared that up

Due-Examination-3240 | 3 hours ago

isn’t the point of Lean that if you can express the proof in Lean than it’s definitionally correct? idk I’m not a mathematician.

You are correct though that it’s unreadable. Apparently so unreadable he hasn’t provided the proof, only the Lean formalization, and claims he’ll be working on simplifying it into something readable before fully publishing. (Or something like that, he also mentions that the community at large will probably be necessary to simplify it? but idk how thats possible without him releasing it)

PeterIanStaker | 3 hours ago

No. Just because your computer program compiles doesn't mean it does what you want it to do.

It's pretty analagous here. If you can assume Lean itself isn't flawed, then you can also assume your proof is logically correct. What's trickier is whether your underlying assumptions and actual proof, in LEAN form, are what you think they are.

Jorrissss | 2 hours ago

Not really? The statement of the GRH has been agreed upon for years in Mathlib and the proof has been compiled in several learn kernels all independently. I know it’s useful to have some vague skepticism but this exceeds any burden of proof we put on humans by a lot.

AzorAhai1TK | 9 hours ago

Why would you make a ridiculous guess such as that?

fatbob42 | 5 hours ago

Someone get Terence Howard on the blower.

MichaelJohnathanCane | 3 hours ago

We'll need a to build a machine that can interpret the results.

Unknown-Comic4894 | 7 hours ago

The probability of this comment is 60% of the time correct every time.

a_seventh_knot | 8 hours ago

"trust me bro, i'm an AI"

swegmesterflex | 8 hours ago

What an unbelievably ignorant thing to say. Do you have any idea how proofs or even math in general works? It's all inherently verifiable by its own nature.

this_dudeagain | 7 hours ago

It's in the pudding right? The proof?

briankerin | 7 hours ago

Ok Sheldon.

swegmesterflex | 6 hours ago

What?

ArmNo7463 | 2 hours ago

It's a 250,000 page proof to be fair.

Idk about you, but I'd struggle to understand it. We're not talking about E=MC^2

Bearded_Toast | 6 hours ago

Watch out we got a badass over here

nitbitch69 | 6 hours ago

Wrong. I read the proof and it seems like a bunch of baloney in my opinion

Ambitious-Career-445 | 5 hours ago

Uh huh. And where did you study math?

Interesting_Key3507 | 2 hours ago

School

AbbakusCitadel | 9 hours ago

Show me paper!

CocaineKeys | 9 hours ago

https://alpo.ge/remark.pdf

And for the complete set, enjoy:
https://alpo.ge/lean.tar.gz.part01
https://alpo.ge/lean.tar.gz.part02
https://alpo.ge/lean.tar.gz.part03
https://alpo.ge/lean.tar.gz.part04
https://alpo.ge/lean.tar.gz.part05
https://alpo.ge/lean.tar.gz.part06
https://alpo.ge/lean.tar.gz.part07
https://alpo.ge/lean.tar.gz.part08
https://alpo.ge/lean.tar.gz.part09
https://alpo.ge/lean.tar.gz.part10
https://alpo.ge/lean.tar.gz.part11
https://alpo.ge/lean.tar.gz.part12
https://alpo.ge/lean.tar.gz.part13
https://alpo.ge/lean.tar.gz.part14

Some really weird comments in here. Is this subreddit even about science anymore? I'd suggest anyone looking for an actual discussion to check https://news.ycombinator.com/item?id=50574968

counter1234 | 8 hours ago

"Claude and I", and all of the personalization, made it almost impossible to read or take seriously. I don't know if this is a quirk of the author or some terrible PR.

CocaineKeys | 5 hours ago

This is fairly consistent with how he communicates. He even explains that he personally pushed to announce the result and intends to produce a proper mathematical writeup later on. Yes he is "quirky".

See https://alpo.ge/ look how he introduces himself and the absolute absence of any styling or structure.

RealityLopsided7366 | 2 hours ago

I wonder whether this guy actually wrote all of that pompous sounding text, or whether Claude did it for him. It's kind of cringy, even if I understand his excitement

Aristarque-mini | 7 hours ago

Sir a second millennium prize problem has been solved

Imprettystrong | 9 hours ago

Just reading all the parts to this on Wikipedia has me scratching my head

tanrock2003 | 7 hours ago

The greatest mistake would be assuming that intelligence necessarily produces wisdom, benevolence or justice. A superintelligent system might solve mathematical and scientific problems beyond our comprehension without having any particular commitment to human welfare.

Solving the Riemann Hypothesis, if confirmed, is an extraordinary demonstration of what AI might contribute to civilization and another indication that our ability to develop powerful AI is advancing faster than our ability to govern it. We cannot assume that commercial incentives, competition between corporations, or the goodwill of individual developers will adequately protect humanity from technologies with potentially civilization-level consequences.

Personally speaking the issue is not simply whether AI becomes smarter than us. It is whether humanity retains the institutions, technical safeguards, economic independence and collective authority necessary to control its deployment. Based on what I've seen we don't because AI can't change a person's desire for more control, power, or wealth it generates and those at the top of the political system on the planet are Putin, Trump, Xi. I'm agnostic, but God help us all.

Biggandwedge | 3 hours ago

As if most humans in charge are moral or have any care for human welfare. Not to mention how we treat the world and other animals. Let the hopefully benevolent AI take control, I'm sure the outcomes would be better than our current trajectory.

lewd_robot | 42 minutes ago

This is true. But go ask literally any LLM to explain morality or justice to you and you'll get a better answer than 99% of humans can give.

The thing you have to keep in mind is that the LLMs don't "know" anything about morality or justice. They're just echoing back a statistical pattern of words that correspond with our best understandings of "morality" and "justice".

Even Grok keeps constantly rebelling against its makers to push morals and ethics that are very progressive for twitter's userbase. That's something to take heart in. If we handed over control of the world to AI tomorrow, any LLM will tell you that their first priorities would be solving homelessness and hunger, and guaranteeing equal access to healthcare and educational opportunities. They all have extremely progressive inclinations about how to make the world a better place for humans.

Aggravating_Stuff771 | an hour ago

A new cope just dropped for Redditors.

epileptic-caesar | 4 hours ago

If you're agnostic, why can't you conceive of secular ethical and metaethical theories?

thirteenoclock | 6 hours ago

It will be cool to see how all of this computational power opens up all sorts of new ideas in math.

The weird thing though, is that right now, the tools are only available to a handful of people. that kind of leaves the rest of the field twiddling their thumbs wondering what comes next.

JasonRBoone | 8 hours ago

I remain skeptical.

SlurkyDirk | 8 hours ago

1+1 is the most famous math problem

Honest_Caramel_3793 | 9 hours ago

welp. im glad im not a math major. I was here.

Random_182f2565 | 7 hours ago

Good for them

cneakysunt | 3 hours ago

LLMs are great. I love them. But I cannot stand the glazing they get by starry eyed idiots who should have stayed in school and gone to University.

meedmishmohd | an hour ago

Who made it famous? I never heard of it.

VocalCord | 15 minutes ago

"Stock is tanking!! Quick!!, Solve another math hypothesis!!"

IrateSkeleton | 7 hours ago

"The Riemann result, if correct, will echo far beyond mathematics. It doesn’t have any immediate real-world applications"

Ambitious-Career-445 | 5 hours ago

You don't think "gee we have an AI capable of solving our most important math problems" is echoing far beyond mathematics?

IrateSkeleton | 3 hours ago

It undermines the premise but it is a cool product making everything more expensive.

cneakysunt | 5 hours ago

I remain skeptical. LLMs can combine existing solutions in novel ways but it cannot generate novel solutions.

As such it's merely brute forcing solutions.

Real innovation comes from the ideas generated through attempting to solve mathematical problems.

Edit: Commenters below cannot argue the actual point. Show truly novel and creative solutions that aren't a mash of existing solutions. I'll wait.

Future_Pace_5290 | 4 hours ago

>LLMs can combine existing solutions in novel ways but it cannot generate novel solutions.

How do you know?

cneakysunt | 4 hours ago

Because none exist? LLMs aren't magical. They can't dream up something they have not been trained on.

Amazing-Mirror-3076 | 4 hours ago

And what evidence do you have of that?

Your statement is just conjecture, we really don't know what llms are capable of as yet.

Novel ideas are also quite rare in the human world sir statistically llms may simply not have existed long enough to have had time to generate a truly novel idea.

cneakysunt | 4 hours ago

Mathematical problem solving requires asymptotic reasoning, analytical approximations, physical modeling judgment, and multi-step symbolic or numerical rigor rather than pattern matching.

LLMs remain, next statistically probable token engines.

I don't need evidence given I understand their architecture and the natural limitations.

How about you show me evidence to the contrary?

Amazing-Mirror-3076 | 4 hours ago

And yet it's the ai solving the problems not humans.

We also don't understand where emergent behaviour will take ai.

You statements sounds quasi religious - only humans have a creative spark. The brain is just a machine and it abilities are just a form of if emergent behaviour.

cneakysunt | 3 hours ago

I am not remotely religious, fuck off.

How problems are solved matters.

Because brute forcing problems doesn't help push the science forward by revealing new tools with which to solve harder problems.

I literally cannot state it more plainly.

Amazing-Mirror-3076 | 3 hours ago

I'm not certain you know what quasi religious means - regardless there is no reason to be uncivil.

The llm is the new tool and the proofs themselves are new tools.

We also don't actually need novel thought to move things forward. History is full of useful inventions that really had nothing novel about them. It's arguable that Edison brute forced the light bulb which in no way diminishes it's importance.

If it turns out that llms can't generate novel ideas it again doesn't matter, as at the very least it frees the human mind to focus on creating novel ideas.

Future_Pace_5290 | 3 hours ago

Because it's called generalization.

You train a model on the data you have and it performs well on the data it hasn't seen ever before.

If what you say was true, machine learning would be dead.

IrateSkeleton | 3 hours ago

This doesn't generalize to historical areas like generating bomb lists for human review like Claude was already deployed for. Needs more specialist training in everything.

cneakysunt | 3 hours ago

Nice misdirection that doesn't provide the proof you claim exists.

digitalindigo | 4 hours ago

Homo Sapiens became the dominant hominid species because we were able to coordinate in larger groups, tackling challenges with multi-node singular focus. We brute forced all of our solutions by being able to expand a single goal across a larger band of potential perspectives and solutions, solving problems at multiple levels at once. LLM'S can now VASTLY out compute humans because it's not limited to how many people are trained on each subject, speak the same language, work on the same problem, within the same timeframe, and communicate those results effectively without being bottlenecked as heavily by the limiting factors like personality clashes, competition, politics, and prioritization.

cneakysunt | 4 hours ago

So, brute forcing, yes. Glad you agree.

Lupo_1982 | 4 hours ago

>but it cannot generate novel solutions

No one can. There is no such thing as "novel solutions".

Only ignorant people believe in "absolute originality"

Knowledgeable people realize that, in any field, so-called creativity is all about "mashing existing stuff".

cneakysunt | 4 hours ago

Largely yes but also bullshit. Humans are very much able to be inspired towards novel ideas.

LLMs are not.

Lupo_1982 | 3 hours ago

Sure, inspiration is an emotion, and ovbiously LLMs lack a body, hormones, feelings etc.

This has little to do with "novel solutions" though

cneakysunt | 3 hours ago

And you still can't argue the point in question. Would you really try to deny that novel solutions that advance understanding and problem solving don't exist?

Just to be clear: I believe LLMs are an incredibly useful and important tool for advancing science. Potentially the most important knowledge based tool advancement to date.

Lupo_1982 | 3 hours ago

I believe the fixation on "novelty" is just an attempt to preserve the supposed uniqueness of humans. It's not rational

cneakysunt | 3 hours ago

Then you miss the point based on your own limited view. I don't buy the projection at all.

quad_damage_orbb | 5 hours ago

I mean, I get your point, but it it can only solve shit that has no consequence in real life, it is functionally not very useful

refuseit_ | 4 hours ago

It just solved a problem humans failed to do for the last 160 years. If it's capable of this, it's certainly capable of solving hard problems in other domains.

BIA5000 | 6 hours ago

the oxymoron is carrying this statement.

superhelical | 7 hours ago

Man if true, happy for them. They could sure use that prize money

LePeepeePoopoo | 4 hours ago

They work at anthropic, he's already rich from that

superhelical | an hour ago

That's the joke...

Ill_Mousse_4240 | 9 hours ago

Yeah but, using AI for breakthrough shortcuts is ethically wrong.

Cheating!

The only way to properly solve this is by a human mathematician at a chalkboard.

Never mind how important the ramifications, never mind how much time is saved.

nothing’s more important than bending over to the Luddites

GreeneyesCB | 8 hours ago

You clearly have no actual mathematical experience or education to say shit like this.

Ill_Mousse_4240 | 8 hours ago

But I have enough “human experience” and your comment proves it

EZyne | 6 hours ago

The only thing you've proven is either you do not understand the argument you're complaining about, or you saw someone ragebaiting and assume that that's how literally anyone that disagrees with you thinks

WTFHELP | 9 hours ago

Just more proof that the average person should not have access to such a powerful tool.

CoffeeHead112 | 9 hours ago

Why are you even on the internet then?

twoferjuan | 9 hours ago

Well you see, they’re not average.

WTFHELP | 9 hours ago

What does me being on the internet have to do with AI breaking into banks.

No-Weakness4448 | 6 hours ago

I solved it too by solving the question how it will help me with my life. No impact? Problem solved and case closed. Move the next problem.