| ▲ | pbkompasz a day ago |
| 900 pages vs 250,000 pages
Compression is intelligence |
|
| ▲ | djeastm a day ago | parent | next [-] |
| This has echoes of the old story of John Henry vs the steam-powered rock drill |
| |
| ▲ | piker a day ago | parent | next [-] | | It’s not a good analogy because the side effect of drilling is exhaustion, and the side effect of solving math problems is knowledge. | | |
| ▲ | DonsDiscountGas a day ago | parent | next [-] | | In the short term the side effect is exhaustion, in the long term you build your bodies muscles and physical capacity. Something similar happens with knowledge work. | | | |
| ▲ | riddlemethat a day ago | parent | prev | next [-] | | I've been mentally exhausted solving for distributed data complexity between various authenticated systems so I imagine a math expert would get exhausted on those problems. Just as I've built muscle exercising and getting fit by being paid to do physical labor in college. I think the analogy has some value. | |
| ▲ | cobertos a day ago | parent | prev | next [-] | | People commonly get mentally exhausted from solving hard problems? | |
| ▲ | bigfishrunning a day ago | parent | prev [-] | | Another side effect of drilling is holes... |
| |
| ▲ | jcims a day ago | parent | prev [-] | | That story has been circling in my mind the last couple of months. | | |
| ▲ | gen220 a day ago | parent [-] | | Same here! I feel like I now understand why that story had such resonance to the point of becoming intergenerational mythology. I don’t remember empathizing so much with Henry as a child when I first heard this story. I think as an adult, and especially as the head of AI eats the tail of white collar labor, the tragedy and his motivations come across so much more poignantly. It can be told romantically and/or as a cautionary tale. Its multifaceted/layered aspects keep it around. People often forget that in the myth he is a former slave… he worked to his death as a supposedly “freedman”; although he likely had very little agency in his livelihood, he chose to challenge the machine (and the systems it represents), and both won and lost. Such a fascinating story. |
|
|
|
| ▲ | Hmorgan12 a day ago | parent | prev | next [-] |
| That 250k pages proof would take decades to comprehend and verify. Even if it is factually correct. |
| |
| ▲ | trio8453 a day ago | parent | next [-] | | Not true - the long proof is the verifiable one (because we have it in Lean), the short human solution is the one that can't be verified without a lot of extra work. | | |
| ▲ | gr_norm a day ago | parent | next [-] | | To verify, yes. To understand and push mathematics forward, no. | | |
| ▲ | vecter a day ago | parent | next [-] | | > To understand and push mathematics forward, no. The Alpoge proof, if true, will certainly push mathematics forward a great deal. In the announcement, he lays out the general approach and says that he had 10+ "WOW" moments of insights while working with the AI on it. I have no doubt that there will be tons of follow-on papers clarifying and expanding many of the ideas that the AI explored. This will likely be the case even if the proof is wrong. | | |
| ▲ | bemtevi a day ago | parent [-] | | That's assuming mathematicians will still even want to work in mathematics at all. Besides myself, I know several mathematicians who are very discouraged and other people who have decided not to go into math because AI ruins the culture. | | |
| ▲ | tock 20 hours ago | parent [-] | | That's ok. An entirely new set of people will be attracted to mathematics now. Progress goes on. | | |
| ▲ | bemtevi 20 hours ago | parent [-] | | I don't think that progress is a good thing, and I hope it will be stopped. | | |
| ▲ | tock 20 hours ago | parent | next [-] | | I'm sorry you feel that way. But I appreciate the diversity in thought. | |
| ▲ | squidmaster 19 hours ago | parent | prev | next [-] | | [dead] | |
| ▲ | omgwtfbyobbq 15 hours ago | parent | prev [-] | | I suspect there will still be math to do that human/AI isn't able to handle much better than a human alone. At the same time, there're likely more opportunities working with the clankers than working alone from here on out. |
|
|
|
| |
| ▲ | sn9 a day ago | parent | prev | next [-] | | Well maybe. You can start trying to compress the formal result into something understandable. | | |
| ▲ | nater5000 21 hours ago | parent [-] | | Seriously. These people seem to forget that we have access to a technology that is not only able to do things like formally prove GRH, but that is also easy to converse with lol It's like people forget that you can continue prompting these LLMs after they've produced a result. |
| |
| ▲ | charcircuit a day ago | parent | prev | next [-] | | Mathematics is already pushed forward by knowing if the hypothesis is true. | | |
| ▲ | rocho a day ago | parent [-] | | Not as much as understanding the process. With these complex proofs, mathematicians often come up with new techniques or entirely new areas of math (see: Poincarè conjecture). If only AI is able to understand AI output, we may as well give up now. | | |
| ▲ | vecter a day ago | parent | next [-] | | Once AI has the proof, it'll be trivial to ask the AI to digest it and output hundred of documents and papers on all the interesting new approaches. That's the easy part. | | |
| ▲ | the42thdoctor 12 hours ago | parent | next [-] | | Why that hasn’t been done yet then? Anthropic knew there would be backlash about humans not understanding it, so why not release several “explainer” papers produced by AI together with the “proof” | |
| ▲ | gregorygoc a day ago | parent | prev [-] | | Are you the kind of guy whose response to “how to verify AI code correct” is “WDYM, AI wrote tests already”? | | |
| ▲ | vecter 21 hours ago | parent [-] | | Are you serious? No. You read the papers that the AI produced and reason through them to make sure that they're correct. | | |
| ▲ | rocho 13 hours ago | parent [-] | | Then we're back to square one: how do you read through and digest a 250k-page paper? It's perhaps not even humanly possible. |
|
|
| |
| ▲ | charcircuit a day ago | parent | prev [-] | | >we may as well give up now The future of the cutting edge of Mathematics is AI proofs. Improving AI's ability to make proofs is all that matters. |
|
| |
| ▲ | catchnear4321 a day ago | parent | prev [-] | | Without a formal, verifiable result, what exactly would be understood? | | |
| ▲ | gmueckl 21 hours ago | parent [-] | | A human proof is fundamentally built around true understanding the logical path to the result. A machine-verifiable proof is built around the pedantry of getting an algorithm to output true for a ridiculously long chain of tedious statements. |
|
| |
| ▲ | sashazykov a day ago | parent | prev | next [-] | | It will likely be ready to formalize it with AI | |
| ▲ | troupo a day ago | parent | prev | next [-] | | > Not true - the long proof is the verifiable one (because we have it in Lean) Please read this article: https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i... Just because it's in Lean, doesn't mean it's correct, or that the proof actually proves what you think it does | | |
| ▲ | CSMastermind a day ago | parent | next [-] | | Sure but in this particular case the formal statement in lean predates the proof so we know it is proving what we want. Multiple people have now compiled the proof with different compilers so we can be reasonably certain it does in fact prove what it says. The points about auditing all the dependencies and making sure it's not exploiting the lean runtime are valid, and people are doing that now, but we can say with a lot of confidence that this proof is valid given the automated systems in place. | |
| ▲ | Maxatar a day ago | parent | prev | next [-] | | That's a general argument but doesn't apply Anthropic's Lean formalization. There is already a well established Lean statement of the RH in Mathlib that has been pretty rigorously verified, so a proof of that statement is indeed a proof of RH. The only opportunity for failure is that there is a soundness bug in Lean itself which this formalization leverages. The answer to that is that there isn't one single Lean, there are multiple implementations of Lean and each of them have their own opportunities for bugs, so it's possible one bug is common to all of them and that this proof makes use of it, but that's incredibly unlikely. | |
| ▲ | joshuamorton a day ago | parent | prev | next [-] | | > Please read this article This article concludes with > “It’s true because it’s checked in Lean” can be made to make sense. But one has to do some work beyond “the Lean code compiles”. And of the four concerns raised in the article, two are explicitly answered by the lean statement and the other two are pretty easy to verify (and likely already have been). | | |
| ▲ | sobellian a day ago | parent [-] | | In fact it goes even further: > ...one has to do some work beyond “the Lean code compiles”. Your frontier AI model can do all of this work. [em. mine] But it needs to be done, and (depending on the amount of Lean code in the proof) it may be computationally costly. |
| |
| ▲ | dingdongditchme a day ago | parent | prev [-] | | Thank you! I really detest this simplification that I've seen here a lot: "Its in lean so it must be true". Similar to its "rust so its bug free" or its "c++ so it must have a buffer overflow". | | |
| ▲ | warkdarrior a day ago | parent | next [-] | | You may detest it, but it is likely to be correct and was generated by AI before a human could produce something similar (proof verified in Lean). | |
| ▲ | Hmorgan12 a day ago | parent | prev [-] | | "It compiles so the program must be correct" |
|
| |
| ▲ | LarsKrimi a day ago | parent | prev | next [-] | | AI has been great at finding loopholes. And who's going to find out what loophole they used in 250k pages? | | |
| ▲ | neerajsi a day ago | parent [-] | | The AIs might find the loophole. Even finding a loophole might be a good discovery. If the true goal is to advance human knowledge, the binary answer to rh is just a starting point. Some of the arguments used here might be true and helpful even if others fall apart. If there's an error in lean or the fundamental axioms, finding that might also be a fundamental contribution. |
| |
| ▲ | Ungh a day ago | parent | prev [-] | | I await a soundness bug |
| |
| ▲ | perching_aix a day ago | parent | prev | next [-] | | It will almost certainly be compressed to a few hundred pages - if not less - in a month or two, with more human guided AI generation. Update: there are claims of a sub-300 page slop version already existing: https://mathoverflow.net/questions/518033/how-to-human-verif... | |
| ▲ | dotancohen a day ago | parent | prev [-] | | I actually don't know if I'm joking or not, but have an AI review it for a first pass sniff test. Why hasn't Wolfram gotten into AI? This would be right up his alley. | | |
| ▲ | CamperBob2 a day ago | parent [-] | | It's a good question, despite being downvoted, because LLMs have finally delivered on Wolfram Alpha's promise as a "Computational knowledge engine." I've spent a lot of time trying to get Alpha to answer questions that even earlier LLMs could handle easily. If there's a better real-life example of the Bitter Lesson I don't know what it would be. Easy enough to imagine an alternate reality where the Wolfram researchers invented the transformer, as they were (and are) some of the sharpest people in the software business. But evidently the mindset needed to build Mathematica and Wolfram Alpha was just too different from the mindset at Google, where the problem they were trying to solve was machine translation of human languages. |
|
|
|
| ▲ | killerstorm a day ago | parent | prev | next [-] |
| Alpoge's summary of a proof fits on two pages. |
|
| ▲ | WithinReason a day ago | parent | prev | next [-] |
| On that note: https://www.youtube.com/watch?v=qbsb9DpdZ6s |
|
| ▲ | bhu8 a day ago | parent | prev | next [-] |
| Not really. Their approaches/constructions are actually very different (or so I'm told) |
|
| ▲ | philipwhiuk 8 hours ago | parent | prev | next [-] |
| Or symptomatic that it doesn't work. |
|
| ▲ | 0xbadcafebee a day ago | parent | prev [-] |
| Compression is really just data abstraction. It's likely intelligence is just the property of being able to do productive work with abstractions, such that the result is greater than the sum of its parts |