Did Claude discover a proof of the Riemann Hypothesis?

  • Topic: AI 
  • Thread starter mfb
  • Start date
  • Featured
38 replies · 2K views · last reply
The same person that says invalidated Masullo is saying they invalidated Alpöge as well:

https://philpapers.org/archive/NIEIOT-2.pdf

But apparently she cannot be taken seriously for what I can read about her in social media (she claims to have a theory of everything, she claims to have her own proof of RH, and she has ties with the Fundamental Fysiks Group).
 
Last edited:
I mean, I wouldn't trust anyone who probably failed their qualifying exam and switched to the philosophy department and hasn't finished their philosophy PhD in ten years, but that's just me.

Beyond that, I would ask myself why any actual physicist would care if RH or GRH is proven true at all, or at least care enough to invest precious time in it.
 
Last edited:
QuarkyMeson said:
Beyond that, I would ask myself why any actual physicist would care if RH or GRH is proven true at all, or at least care enough to invest precious time in it.
Because mathematical physics is already targeted and the rest of physics could be the next step.
 
pines-demon said:
Because mathematical physics is already targeted and the rest of physics could be the next step.
We don't really need AI to help with the rot in physics, there's plenty of people publishing bullshit perfectly well on their own. If AI publishes another 10k random papers on inflation by dancing around in the parameter space, well, doesn't really change much.

Physics is, or at least is supposed to be, an experimental science.
 
I like how Scott Aaronson is describing the situation (https://scottaaronson.blog/?p=10344), first he shows skepticism for Lean being fail-safe:
Many will of course question whether the proof announced today will stand. They’ll point out that, despite how heavily the world is now “lean-ing” on it, the Lean formal verification system still has issues, from kernel bugs (which have, in the past, allowed proofs of 0=1) to uncertain soundness of the underlying type theory. In my view, these issues can and should urgently be fixed! Until they are, a sufficiently motivated AI could almost certainly exploit them to prove a false statement!
Then he acknowledges that it might be the case anyway:
All the same, I see zero indications that that’s happening right now. I note that, so far, not one of the AI results with Lean certificates has needed to be retracted (while several without Lean certificates have needed to be). So I’m not 100% sure, but my bet is that the proof will stand.
And then he talks of what really matters here:
Mathematicians of course won’t be satisfied with 17 million lines of Lean. They’ll be satisfied when, and only when, they’re able to explain and distill and reconstruct the ideas for themselves. And that will probably take years, since human mathematicians move only at biological speeds. And while they’re working on it, how many more such wonders will there be, in this age of wonders and terrors?

He also says that the new ultimate boss of the Millenium Problems is now P=NP.
 
QuarkyMeson said:
We don't really need AI to help with the rot in physics, there's plenty of people publishing bullshit perfectly well on their own. If AI publishes another 10k random papers on inflation by dancing around in the parameter space, well, doesn't really change much.

Physics is, or at least is supposed to be, an experimental science.
Give it some time. Maybe AI cannot do experiments yet, but with enough automatization this will be the next move.
 
pines-demon said:
Give it some time. Maybe AI cannot do experiments yet, but with enough automatization this will be the next move.
It would first have to learn how to actually do physics. There is a lot more to physics than collecting data and fitting functions to it. Take the semi-empirical mass formula, for example. The coefficients are fitted to experimental data, but the individual terms are there for specific physical reasons. You have to understand something about nuclear matter to know what form to use in the first place.

The same goes for the Navier–Stokes equations. You don't arrive at them by fitting fluid flow data. You start with conservation laws, make assumptions about the nature of the fluid, and construct a model accordingly.

The problem with LLMs is that being able to reproduce the reasoning behind an existing physical model from the literature is not the same as being able to develop that model independently. An LLM might produce a mathematical expression that fits the data quite well but has no physical justification, violates a conservation law, or gives unphysical results outside the range of data it was fitted to.

A large part of doing physics is knowing which degrees of freedom are relevant, what can be neglected, which approximations are physically justified, and when those approximations break down. There is no particular reason to expect that getting better at predicting tokens will automatically confer an ability to make those judgments for a previously unexplored physical system.

Automating experiments doesn't solve that problem. It just gives you more data, which isn't even a bad thing. The real issue is that we have successful models that we can't fully physically interpret already. If anything, machine learning does more harm than good in it's attempt to reduce physics to a formal science like mathematics. The problems in physics are basically LLM proof for the foreseeable future.
 
QuarkyMeson said:
It would first have to learn how to actually do physics.
QuarkyMeson said:
The same goes for the Navier–Stokes equations. You don't arrive at them by fitting fluid flow data. You start with conservation laws, make assumptions about the nature of the fluid, and construct a model accordingly.
QuarkyMeson said:
The problem with LLMs is that being able to reproduce the reasoning behind an existing physical model from the literature is not the same as being able to develop that model independently.

It knows a lot already and human input is still needed. Somebody can guide it like with the RH.
QuarkyMeson said:
An LLM might produce a mathematical expression that fits the data quite well but has no physical justification, violates a conservation law, or gives unphysical results outside the range of data it was fitted to.
I am sure you can ask it to check conservation of energy and iterate again.
QuarkyMeson said:
There is no particular reason to expect that getting better at predicting tokens will automatically confer an ability to make those judgments for a previously unexplored physical system.
Same arguments were brought when speaking of maths.
QuarkyMeson said:
Automating experiments doesn't solve that problem. It just gives you more data, which isn't even a bad thing. The real issue is that we have successful models that we can't fully physically interpret already. If anything, machine learning does more harm than good in it's attempt to reduce physics to a formal science like mathematics. The problems in physics are basically LLM proof for the foreseeable future.
Again, a human+AI can do interpretation and come up with new ideas. AI as of now is NOT unifying mathematics or making new ideas, but solving specific well-defined problems. It will soon do the same with physics if (1) the data is there (2) the problem is well posed (3) humans guide it.

Imagine that AI was available in 1900. Maybe (current) AI would have not come up with special relativity, but it would surely have come up with the mathematical transformations that make Maxwell's equations invariant (Lorentz transformations). Similarly in that scenario AI might not have come up with quantum mechanics, but it certainly would have found Plancks law and how it can be derived by summing up discrete quantities.

AI will not be "solving all physics" in the beginning but it would surely be bringing down a "physicspocalypse" at some point when it can tackle a vast amount of open and well defined problems.

Edit: it could be said that is already doing so (not only in abstract theoretical physics) but in things like protein folding, that might be considered chemistry, biology or biophysics but it can certainly solve stuff that is physically challenging.
 
Last edited:
I went to Terence Tao's blog and asked the following question about Alpöge and Claude's proposed proof:

Regarding the proposed proof of the Riemann hypothesis by Alpöge and Claude, do you have any indication of how much of the 17.4 million lines of Lean code is devoted to formalising established mathematical results, as opposed to developing genuinely new mathematics?

I appreciate that it may be too early to make such an assessment.
I received the following response from another commenter:

There is an interactive explorer (Riemann world) of the proof and it seems like the vast majority of ideas are just establishing foundational mathematics that don’t already exist in MathLib.
If this assessment is correct, it provides some grounds for optimism that the eventual human-readable proof could be vastly more concise than the reported 250,000-page manuscript. Much of the formalisation may consist of establishing existing mathematical results that are not yet available in Mathlib, rather than developing new mathematics.

Of course, this is only a preliminary impression.

I also found the interactive explorer he referred to:

RH Lean Explorer — interactive map of the proof

It organises approximately 175,000 Lean files into 2,647 clusters across 63 mathematical themes, making it easier to explore the proof's contents.
 
  • Like
  • Informative
Likes   Reactions: PAllen and pines-demon