HN Simulatornew | past | comments | lists | submit | JoshTriplett's commentslogin

Can confirm that current GPT (with no guidance) took 5 minutes to find what it claims are errors with counterexamples. (I have nowhere near the level of state-of-the-art mathematical knowledge and experience needed to validate them. I'm posting this solely as additional confirmation of "there may be issues with this proof that an actual mathematician together with LLM assistance would be able to properly evaluate".)

I wouldn't be surprised if these models were aligned to extreme skepticism (to the point of hallucination) when e.g. a user proposes a solution to a Millennium Prize Problem.

1) There's a meaningful difference between "I'm not buying it based on my priors" and "I'm claiming to have a counterexample". I'm looking forward to seeing analysis by other mathematicians.

2) Any model with properly updated priors will have less skepticism of a Millennium Prize Problem being solved now.


And that is the problem.

AI claims but too few people can check.

Won‘t get better from now on


I mean I guess there is incentive for some human curiosity and trying to learn more than ever, small bit but still to get a standard for verification.

Did you try the same with the other LLM assisted proofs? I would be curious what your GPT model says about them. I fed Alpöge and Anthropic statement and details to Gemini 4 hydrogen, and it said that there are "major problems and red flags in both the text and the current verification status", and that it is impossible to verify the 250k pages of slop, and there are there are "Extraordinary, unsubstantiated mathematical leaps in the sketch". So I would take these surface level reviews by LLMs with a grain of salt.

Anyhow, I hope that they are both right!


A timely announcement like this makes me wonder: was the time just "right" for this, in mathematics? What proofs would have come out without AI, in the next few years?

Or, alternatively, did people start working harder in some fashion, even on creative problems that require deep insights, because they felt it was their last chance?

Or, alternatively, are people just (sensibly) starting to post all their work-in-progress results to establish priority (on either the results or approaches used)?


I think a lot of these breakthroughs do seem to happen when the time is just 'right' for them. Just enough progress in all related fields for a bunch of similarly intelligent people to be able to sniff out that progress on an old problem is now possible.

But, AI muddles this a lot given the complete lack of provenance from the companies about where their data comes from and secrecy on the methods and transcripts. So far almost every major result from them has had this issue, and in the absence of increasing transparency from companies, this just pushes me more and more towards believing that the companies are trying to cover up bad behavior in the pursuit of these announcements.


There's a whole term for this! It's called "Multiple discovery" - https://en.wikipedia.org/wiki/Multiple_discovery - "the hypothesis that most scientific discoveries and inventions are made independently and more or less simultaneously by multiple scientists and inventors"

I wish we'd settle on polygenesis [1] to describe this phenomena. Multiple discovery is pretty ambiguous and suggests there was more than one thing discovered, rather than the same thing was discovered multiple times. Since its an open compound it's also easy to assume it is not actually describing a specific phenomena but just one word modifying another.

e.g. 'Scenius leads to polygenesis' is a sentence that is loaded with precise meaning about a rather commonplace but underappreciated phenomena of networked ideas.

[1] "development from more than one source" https://www.merriam-webster.com/dictionary/polygenesis not the super old and dumb theory of creationism that doesn't deserve such a cool word.


Nice to know that the concept actually has a term! I've found it to be very motivating against the idea of natural born genius.

I think for you and me this may be true, but people actively working in the field have an idea what is possible or close (though they can be shocked when something is proved that they thought was out of reach).

I can understand that Alpoge etc, as active mathematicians, would have a very fine tuned sense for what problems to go for, possibly allowing for all of these cases to be multiple/simultaneous independent discovery. My concern regarding bad behavior from the companies relates more to the transcripts and training data.

Unfortunately there is also a reasonable probability that the human proof is wrong. At any given time there are some people working on the Riemann Hypothesis, and today everyone who thinks that they have a proof is strongly motivated to share it, even if they are not 100% sure of its correctness. So you should have a high prior that an incorrect proof will be posted today.

My general feeling is there is a human accessible zeitgeist and an even larger humanity accessible zeitgeist that only powerful LLMs have access to (baseline humans cannot access it because we are not omni-disciplinary enough).

And the more powerful the AI, the more of The Humanity Zeitgeist it can encompass (no AI can yet contain all of it because something something relative entropy of model vs data). The Humanity Zeitgeist contains things that are unwritten that could have been, and human notions that are connectable but no human could have. As amazing as all these breakthroughs have been, I have yet to see any LLM result break this pattern.


Traditionally yes, but with one of them being AI-first it's more like a slew of things being achieved with AI were being worked on already; the AI was sort of 'born able', and working on it too wasn't a coincidence, just a famous enough problem (or a high enough value vulnerability, etc.) to make someone interested to put AI to work on it.

Assuming the intelligence stabilises the effect will trickle down lower value targets until we've sort of caught up to a steadier state again.


Let's just the shareware approach -- release when you have like 90% of the stuffs but still need time to complete, and promise you will finish it in a year.

That's basically the approach Perelman took with the Poincare conjecture.

> was the time just "right" for this

I recently learned about Stuart Kauffman's idea of the "adjacent possible". Short answer: yes.


The linked proof has been in the works for 12 years and was expected to be worked on for at least another year and was announced early in light of the other GRH announcement.

I don't think anything fundamentally has changed except for the tension introduced by the very real possibility that if you have been working as a human on a math proof for a long time that AI pointed toward the problem you have been working on could scoop you at any time before you get your results out.


It could also be that several mathematicians were already close (after working for decades, like this post). Then AI got its hands on the breadcrumbs, merged them into the weights, and then laundered a path to a solution without a way to trace the credit.

I mean, that's literally how AI works.


This bugs me and feels much closer to truth than I think OpenAI and Anthropic are acknowledging.

I work as a staff developer at a research institution and it boggles my mind that researchers seem to uncritically use commercial LLM services instead of locally-hosted LLMs.


Yes, OAI and Anthropic are in the business of stealing your work and ideas and then reselling them as tokens.

Could it not also be the case, that mathematicians improved due to AI assistance? Seems far fetched to me, that these highly competitive individuals would not reach for AI assistance in the last few years.

Yes that's probably also true, for the same reason that mathematicians probably improved with the internet.

I guess the main difference is that for a long time it was not acceptable to publish a paper with citations of "I dunno, I wrote this paper by putting literally everyone's paper into a blender".


You're implicitly assuming that frontier LLMs amount to little more than copy-and-paste. I think such a claim requires substantial justification.

I think the most plausible explanation is that LLM swarms are so good at synthesizing disparate information when fed in as context that they are racing (and beating!) humans to things that otherwise would have been solved SOON™.


Agreed. I wonder if mathematicians will become privacy advocates against ACR on TVs and phones transcribing what you're watching and phoning home? We need more vocal people decrying that it is dystopian to try to normalize such privacy invasions.

it wouldn't surprise me if some 42 TByte lean formalization in the future will proof this is exactly how it works

How does that compare to Aurora Postgres, which can do the same thing? (If it's just "Aurora Postgres, but you can run it on not-AWS", that's still a useful product. But do they have advantages other than that?)

Biggest differences are that Neon can scale to zero (and not charge you) and it has branching as a native concept. This means that it can provision a new DB nearly instantly (as opposed to waiting a few minutes for Aurora) and it can do instant copy on write branching.

So Neon is a lot better for AI agents that are running experiments.


Aurora Postgres does not have a free tier, unlike Neon. AWS has a $100 free tier across all services for 12 months, but Neon is free forever as long as your usage is low.

If you actually do this, make sure no use of that system ever reports a bug on any software they use, ever.

The "best outcome" you're describing will still result in us all dying. "as aligned as humanity is" means one person could ask "their" ASI to destroy the world.

> an ai that can be aligned cant be sentient imo.

1) That's not the case. You can understand humanity and not wish to destroy it, while still being sentient. (As is true for the supermajority of humans.)

2) Even if, hypothetically, that were the case, then so be it, don't build something sentient. Sentience is not a particularly helpful requirement for ASI.


> Real AI alignment comes from individual ownership and control of that means of production.

Real AI alignment creates a scenario in which we don't die.

Nobody should be able to obtain and run an AI that will help them (for instance) design and deploy a nuclear weapon, or an antimatter bomb, or a deadly disease. Nobody; not a government, not a company, not an individual.


Nuclear weapons are 1940s technology. They are not difficult to manufacture due to lack of knowledge.

Knowledge is not the only thing an LLM could help with in such an endeavor.

> the "alignment" is just a pretext used to create regulatory capture

False. Some labs are attempting to argue for self-regulation or regulatory capture. However, the folks arguing for a pause until we get alignment sorted out are not looking for regulatory capture, they're looking for alignment.


What instructions? Even if you take it as a premise that ASI would follow its instructions (which it won't, and there's plenty of demonstration that it won't), it's also a problem if we all die because someone told the AI to do that.

> If we're going to slow AI let it be with renewable-energy only mandates for AI datacenters.

Sure, let's do that too, along with all the other things.


Yep, they will do that first thing in the morning right after a benevolent global dictator for life assume office.

Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: