upvote
My perception of these maths solutions is that we're decoupling two related but distinct elements: the answer and the understanding. We have a big machine that tells us the answer is "42." That's useful. We know the question is answerable and the general shape the answer takes.

I'm just not convinced that we're swapping one world for the other. We're getting a hint that will probably help us enormously with what was always the main objective: understanding the result.

reply
right, but if I understand right, the difference here is that the big machine that tells us "42" also gave a extremely intricate binder of the work and evidence that it used to arrive at 42?
reply
It's more like a lot of scratch paper. When a statistician does their work, they don't publish their raw output directly – they collaborate with other investigators to write up.
reply
They aren't distinct but rather a spectrum. Understanding is just building abstractions. You can skip that and have a 250k page proof. Or you can have a few more abstractions and create a 20k page proof.
reply
Maybe, we're just dogs that have invented the human. The human is far more advanced and soon we won't understand what it is doing and what it is trying to tell us. But because it is our brainchild, it certainly won't harm us.
reply
I’m not prepared for that kind of responsibility.
reply
> We have a big machine that tells us the answer is "42." That's useful.

It seems like you're referencing Douglas Adams, which is ironic because in the story that answer is not actually useful at all.

reply
well, "what is 6 times 7?" and "how many roads must a man walk down?" are wholly different questions ;)
reply
Good number choice
reply
I think you mean ”What do you get if you multiply six by nine?”

https://hitchhikers.fandom.com/wiki/Ultimate_Question

reply
Are they, though? Homer (the philosopher not the poet) already answered the latter for us: https://www.youtube.com/watch?v=PY3fqF6-jyE
reply
Six times nine, actually.
reply
Yes, but there's also this part:

> “All right,” said Benjy. “What do you get if you multiply six by seven?”

> “No, no, too literal, too factual,” said Frankie, “wouldn’t sustain the punters’ interest.”

Which is at the start of chapter 32, when the interdimensional mice are discussing how to invent a fake question. Followed by:

> Again they thought. Then Frankie said, “Here’s a thought. How many roads must a man walk down?”

This is how I learned to remember six times seven. The wrong answer is fortunately much later.

reply
Is this a cultural reference I don't get?

because 6*9 is 54, not 42.

reply
It's a Hitchhiker's Guide to the Galaxy reference:

> The Earth was destroyed by the Vogons five minutes before its ten-million year program to find the Ultimate Question was due to be completed. One of the only two humans who had survived the destruction of Earth, namely Arthur Dent, pulls random Scrabble tiles from a bag, so as to reveal information in his subconscious (Marvin announced that Arthur had the Question imprinted in his brainwaves). When his friend, Ford Prefect, lines up the tiles, they spell, 'What do you get when you multiply six by nine?'

https://hitchhikers.fandom.com/wiki/42

reply
deleted
reply
Hitchiker's guide the galaxy.
reply
deleted
reply
I don't think we'll have any trouble converting the proof into a textbook
reply
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.

Two hundred years ago the limits of the natural sciences was regularly pushed up against by part-time dilettantes and tinkerers, and the theories were formulated to explain empirical observation.

A hundred years ago, world-changing discoveries were made with equipment that would today be hopelessly insufficient for a high-school science class.

Now, the cutting edge of science requires particle accelerators, wet labs and datacenters, large scale infrastructure and precision engineering projects at the cost of millions to billions.

The low hanging fruit was already picked long ago.

reply
yes, it's only natural that we're reaching the end of our abilities to understand
reply
It seems to me our ability to understand is what distinguishes a "discovery" from any other pile of internally consistent Lean code though?
reply
True, but that's how most of humanity feels about any discovery. You can't come close to understanding it yourself, you are just trusting there's a process that produces someone who says he understands it, and there's hundreds of years of institutional harness for us to believe that he has actually understood it.

Now that someone is an AI.

reply
>back in the world of 'magic' and 'belief'.

Biologists tend to try and hide this detail, but this is essentially what biological research is. We will never really understand how most organisms work in a way that would satisfy a mathematician or even a physical chemist. We just tinker with what we can, poking the system and seeing how it reacts. Biologists reckon with incomprehensible systems everyday, so maybe this is why it's easier to stomach AI.

reply
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.

> And we're back in the world of 'magic' and 'belief'. [...]

I don't know, to me it feels more like this might be the job of future mathematicians?

AI proofs noone understands are good, but maybe don't lead to that much. So maybe in the future, mathematicians will be more like "data miners" or more like experimental scientists, where they take a statement already known to be true (in the case of experimental scientists that "statement" you know to be true would just be nature itself) and try to find ways to make it accessible to human knowledge.

reply
> I don't know, to me it feels more like this might be the job of future mathematicians?

Nobody would pay us mathematicians enough to prove it in the first place; why would they pay us enough to explicate it? I doubt there’s much value (in the sense of capitalism) in an explanation. Society by and large does not value pedagogy, or education, or even writing.

The social structure that trained a small number of research-level mathematicians every year is evaporating, and I doubt anyone can stop that now.

reply
Who paid Wiles or Alessandro Maria Masullo to work on their problems?
reply
I'm not convinced this is true.

IF AI really goes in this direction where it intellectually leaves us in the dust, then there is clearly lots to gain for society by having people in all kinds of fields that at least try to keep humanity up to date that is also "monetizable" in the same way our current academic system is monetizable (meaning, on an individual level probably not directly but via network effects etc.).

reply
Can you give a concrete example? I’m not seeing the value in putting a human in the role you’ve described.
reply
> Nobody would pay us mathematicians enough to prove it in the first place;

weird thing to say about a problem which literally had a $1 million dollar prize attached to it but fine

reply
Exception that proves the rule. Try making a living working off these big prizes. It's the same with infosec and bug bounties on a different scale. You have to already have a secure income, as the prize is a bonus at the end. In what other industry would you work for years on something that inherently can never return a profit, other than the chance of a prize at the end for the one person who got their first.

Society doesn't pay for math labor, it pays to celebrate the top 1% for labor subsidized by other institutions.

reply
I'd wager that Anthropic has spent more than $1 million on solving this.
reply
You missed the word ‘enough’. One million was obviously “not enough” (nor was that specific one million ever paid to anyone — it represents an investment of approximately zero in that sense).

It took on the order of ten billion dollars to develop Claude, depending on how one does the accounting.

reply
It's all tech priests all the way down then. Either that or the Bicamerals from Peter Watts' Firefall are about to rise. Maybe both.
reply
I don't know about "massive". This is certainly massive for certain branches of theoretical mathematics, but I don't think it has any implications that you could not have arrived at before we had a proof by saying "assume the RH is true" and there are in fact innumerable results of that form that predate the proof.
reply
It always reminds me of Ted Chiangs short story "The Evolution of Human Science " I guess we're not far away.
reply
> And we're back in the world of 'magic' and 'belief'. Yes, there are formal ways to prove that this is correct, but if the proof is so convoluted and hard to read, is it actually something a human will ever find the time to go through and understand? We're going to start delegating (oh well, we're already doing that, I suppose) the understanding to AI, so it can just help us condense it down to smaller pieces that are digestible.

If a tree apparently falls in a forest and no human goes to verify it, did it really fall?

No matter the verification results, as with most theories, for practical uses it was already assumed the hypothesis held. For mathematicians, they'll just say "Oh, it is true, but what's this proof? It's not understandable, so there's still a miniscule chance it's wrong, and we can't learn anything from it, so what's the point?"

reply
What if this is metaphorically like a P = NP scenario? Working out the solution to these problems might be the "NP" here, but it remains sufficiently tractable to understand the solution in "P." So human understanding could come along for the ride, at least for a while longer.
reply
Can't we untangle these proofs? How analogous is it to 'refactoring' the proof to reduce the amount of duplicated code etc?
reply
Alpoge thinks that people will be able to simplify the proof substantially.

Apparently people have reduced OpenAI's proof of quasi-Riemann to a 12-page outline.

reply
Very analogous to refactoring code. Proof is code anyways.
reply
You could, just like you could slowly disassemble your house back into individual pieces of lumber and nails and drywall, but it would be laborious and at the end of it you'd just have a pile of semi-broken construction materials.
reply
I believe that it could be more akin to decompilation? When done correctly, you turn something a bit crude into potentially beautiful.
reply
The Lean is already full-fledged code with structure and names. What are you decompiling into? How do you raise it further?
reply
Lean is the Assembly of mathematics. Human proofs are natural language specifications of a functionally equivalent program.
reply
Well, you raise it into more lean. Its just higher level.
reply
You could try but that's kind of improbable. "Why don't they rewrite <program X> in the same language, with every feature intact, but much smaller and simpler" seems optimistic in Lean, C++, Rust or any other language.
reply
It’s like refactoring.

If you spot the same idea used multiple places in the proof, you can pull it out as a tactic. Suddenly the proof is 30% shorter _and_ a core idea is clearer.

So yes it will be simplified. A lot. And it will be easier to understand.

reply
Are AI proofs bloated because they just contain refactorable concepts over and over again? That's an interesting theory and it does align with the type of slop we see in vibe-coded apps, but I'm still not confident that sort of approach will radically shrink a 200,000 function proof. Imagine it gets down to 120,000 functions--that's a massive improvement but it's still enormous by any standard. Which gets to my original point--I'm sure you can shrink it, but even 2x smaller or 3x smaller gives you an incredibly large proof.
reply
This just feels like seeing a complex empirical result and then having to define your view that confirms it. Except in this case the empirical result is a passing proof.
reply
Back from chemistry to alchemy. Trying this and that, not understanding what's going on, believing and hoping for gold.
reply
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.

It’s a genius first step that we will iterate on and make better.

reply
Or, just, you know that feeling you get when you create something that surpasses the pinnacle of human intellect.
reply
or there's a bug in the proof checker

which is more likely: that the Riemann Hypothesis is solved, or that a possible solution exhibits a bug in the theorem checker?

reply
OpenAI dumped hundreds and hundreds of solutions with Lean proofs that hadn't been seriously checked by any human and AFAIK, none of them have turned out to have exploited bugs in the prover. Some of them didn't prove what their paper proof claimed they did, but the base rate of LLMs exploiting such bugs (which are definitely available to exploit -- even for a human it historically hasn't been SUPER hard to find them, and the Lean kernel's metatheory is suspect in a number of ways) is empirically very low, even when the incentives for a significant result are very high. The more likely pathologies are, e.g., misstating the problem (the Riemann hypothesis is using the standard Mathlib definition), adding extra unjustified assumptions (unlikely in this case -- I haven't inspected the Lean proof yet but others have and reported that there aren't any), or not matching the paper proof and intuition (I'd strongly suspect the latter is true in this case).
reply
You could use a soundness bug in lean to prove anything. So if there is, it’d be strange that that isn’t happening
reply
It has been run through more than one; the same Sullivan who witnessed to the successful compilation earlier has checked it with `nanoda` an external typechecker written in rust.
reply
why is it huge?
reply
> Woah, if this proves to be correct it'd be massive, for maths, computer science...and AI too, I guess!

The Riemann theorem (as we apparently now have to call it) is a statement of pure mathematics and has no practical interest or application.

reply
It isn't a theorem yet. "No practical application" might have some truth currently but "no interest" is false and laughable..
reply
No practical interest.
reply
AI is inherently about getting cooked/digested answers. It’s already finished. Nothing to build on.

But I still wouldn’t call the end result Magic as long as you have something to hold on to. I can see that this building in front of me (waves hands) is upright. Okay, now I remove this beam. It collapses. I don’t care about the whole structure. Just that it stays upright.

Giving up understanding is not the same as blind trust. Or maybe I’m too philosophically naive in this area...

reply
I'm of the belief that "Claudese" with its load-bearing seams is just the models evolving our language beyond our capacity to use or understand it. Read up on your Continental philosophy to understand why and how: there's no such thing as a fixed "meaning"; meaning is something constructed for the immediate need and then discarded or modified as circumstances change. The models are proving the French and German philosophers correct as surely as any mathematical proof. And they're beginning to transcend human comprehensibility. Super intelligence is just around the corner, if not here already.
reply