I really love this writeup, and I think the approach of posting the lean proof and leaving the long process of digestation for future work is a perfectly acceptable approach. It's the difference between closing fields of research (openai) and opening them.
Thread for to follow along with verification: https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
(login seems to be required)
> The precise timing is horrible. If I could have, I’d have finished this all much earlier. Hell, strictly before this age would have sure been nice.
Is he over 40, and referring to being ineligible for the Fields Medal? It would be strange if you could get a fields medal for work done by an AI, but I suppose that is how many Fields Medals will be won from now on. I hadn't thought of that until just now.
If true, the infinite agent theorem needs to supersede the infinite monkey theorem. Just launch 1,000,000 agents with 10hr windows at one open math problem per day and see what happens. I know OpenAI is sort of trying this already for marketing purposes but I'm thinking a more formal, BOINC approach to problems.
https://en.wikipedia.org/wiki/Infinite_monkey_theorem#Histor...
This (https://github.com/tzu-chen/slim) can be used to navigate the lean source and search for key statements without building the whole release, which can take hours.
What are some of the practical applications of this discovery?
I'm sure people have done it, but I haven't seen any news about it so I guess I'll ask here... has anyone asked thrown Inter-universal Teichmuller Theory at Claude?
> The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail, in my view of my path I have just taken a step. I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly.
Next step, make the proof intelligible for humans.
I wonder how specifically he used Claude to do this.
Was it a case of having more dollars to spend by virtue of being an employee and just pointing it in the right direction, or did he provide some insight into the context that allowed it to produce superior results?
It must be the second - but if so, how? Was it as simple as spelling it out in chat and analysing output?
If true, it seems like Goldbach's conjecture may fall shortly after, see this recent proof conditional on Generalized Riemann Hypothesis [0]. It has other conditions too though.
> Does [Alpoge's proof] give anything about square roots mod pq (p,q prime)?
> Claude Notice: Fable 5.2's safeguards flagged this message.
Yikes
I've downloaded the Lean formalisation. It's 175,187 Lean files, 17,374,215 lines of Lean code, and comes in at 0.96 GB of code. On my 20 core i7 box, it's projected to take about a day to build...
How tricky is it to translate the statement of the Riemann Hypothesis into Lean?
If we're confident in the accuracy of the Lean RH statement (which must be part of this proof), and if Lean is sufficiently bug-free, then we can trust the proof is correct.
If proven then we can say with confidence that we now share the world with super intelligence, albeit in certain areas and indeed one that was endowed by humanity. I can only hope that in time the benefits are many and accrued to all of us.
We have been conditioned to be impressed. Should we update that conditioning now?
If this is true it would be astounding. Most Physicist's placed Navier–Stokes at the top of the list of what AI could quickly solve and the Riemann Hypothesis last.
An unreadable proof is basically none at all. Suggest you have a model produce a concise summary and work edge in from that. Apparently none of these problems has been verified and anybody working with models would be justly skeptical of these helium filled claims which are almost certainly not going to pan out.
Also apparently the Riemann claim is based on it compiling in Lean. This is classic because just the differences between an actual verification of a proof and it compiling in Lean are the standard model failure tropes, lack of judgement, confabulation, etc., to which this adds shortcomings of Lean.
>>The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail
How do we know the Lean formalization is an accurate representation of the incredibly long and unreadable argument?
This is such a wild ride. The conjecture itself is not so hard to understand, but approximately _all_ of the maths here are beyond me. The paper is written up to suggest "we knew which direction to go, but really had no idea how far the journey would be" which certainly makes it seem like the human expert was useful in this process. Perhaps that's just wishful thinking, but it's consistent with what I see elsewhere that AI is having an effect. The biggest effects accrue to the most effective humans. Random folks prompting is better than random folks toiling alone, but experts prompting is a different thing.
This is an amazing result but I can't help but think about how much the world has changed since the controversy around the four color theorem's computer assisted proof.
There are certain math problems where the only guidance has been the prompt "you can do it". Now that we know it's possible, this should be a new benchmark. Probably too expensive for the next few years.
Formalized P!=NP proof by end of year?
My hunch is that stuff like GRH and other problems we have will be childrens' toys in the future and truly novel, advanced math is not yet even within our grasp yet. We are struggling to walk here, we need machines to help us with this. We felt good about ourselves for a while because everyone was equally bad at it, but now reality is hitting us. Hard.
Out of interest, how do they use an AI to solve these maths problems? I assume they're not pointing the chat at the wikipedia page and saying "Go solve that for me."
Are they using code to do it? Is it a special maths AI harness>
Software development and mathematics were two professions people used to say were the ones most stable for future .... I guess they couldn't have been more wrong :(
For example, a Field Medal-level Claude, if made available to the general public, could potentially enable people to design and develop something as technologically complex as SpaceX's Starship in a matter of months rather than years. What took hundreds of brilliant engineers at SpaceX years to accomplish could potentially be achieved in a fraction of the time, if the current pace of AI advancement continues.
Math is becoming the measure of intelligence.
I wonder if the money open ai and anthropic have spent just on solving these problems was spent on human mathematicians and resources would we have gotten these solved? Sooner?
The shift now needs to be towards statements that can't be posed as a singular Lean statement (or, even, a collection of Lean or similar statements that is understandable by a human). Mathlib/NumberTheory/LSeries/RiemannZeta.lean already had the statement formalized in a way that us mortals can understand. The great open problems of the future will be the ones which cannot so easily be interpretably formalized in the first place, I fear.
Seems like a great point to stop AI development at, internationally
The timeline given by Claude at the end of the note seems to jump back and forth in time. Are these just typos or what is going on there?
If the laws of AI scaling continue to be true, does this mean that in a year or two we would be seeing frequent revolutions in mathematics and physics?
I kind of expected this to happen by EoY.
Combining the brightest mathematicians we have with intelligence-in-a-box, it was just a matter of time before the hardest math challenges fall.
Congratulations to Levant, but also to us as a species. We have managed to create machines made out of pure math that contain insurmountable knowledge and are helping us solve the mysteries of the universe.
At some point I was looking at ways to try and find counter-examples computationally. I'm glad I didn't spend too much time on that...
Man this is so fucking cool, I cannot wait to see what other problems fall in the future (and what new ones this opens).
250, 000k pages is bonkers, I hope they will be able to condense it into something a bit more grokable.
I think this is a pretty clear indication that the hybrid approach (e.g amazing mathematician and great models with a huge token budget) can accomplish absolutely incredible things.
Math is the verifiable domain; I wonder if this ratio still holds in less verifiable domains as well, or it shifts more AI/more human (or how it changes in future!)
The most important thing now is to find ways to visualize LEAN proves so that we can understand them faster by factors.
Whoa. If they also solve P=NP I will really be scared.
Has anyone set up a GitHub repo with the Lean code a comparator and verified this?
Looking forward to the release of the approach and also someone to "distill" it down to the most intuitive and still-correct version.
So... its like paying technical debt in math, considering a lot of stuff assumed it was true, without a proof -- until today?
I really hope the major AI labs steer away from mathematics as a means to prove capability of their models, and start putting their tokens to real-world use cases.
At this rate do you think the frontier labs have a result on P vs NP?
I must say I like this approach to the announcement way better than OpenAI's "dump"
In about an hour or so, I'll have a rough automatic informalization of this up on https://digestivo.dev. I've been working on a way to digest these proofs faster via parsing lean and translating it into natural language and this is as good as time as any to finally launch this thing.
> It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
> But I hope in having the experience of seeing all of number theory’s beauty brought together I am just one of the first, and those who love the mathematics will all be able to enjoy such transcendence in the future.
Bus meme irl. I guess I'm lucky that I have this experience even when working on boring things like data work. But these types of attitudes can really make a difference in how we feel and our end results.
Alright, what's the over/under of the Collatz conjecture getting solved by the end of the year?
Thank god for Grigori Perelman, who isolated himself for a decade to prove the Poincaré Conjecture, and we can at least claim that one of the millennium problems was not solved by a machine.
The write-up is beautiful and poetic, but light on the details of how this was achieved from a practical standpoint. How many human and agentic hours were necessary? Was a breakthrough only possible after Fable 5.2 or another internal model?