It tickled me that this was a bullet point in the Fable 5.2 announcement. I like that we're now at a point where "solved two open problems in the physics of black holes" is the kind of thing that shows up in a frontier model launch.
Those theoretical physics papers look like the real thing (i.e. not mathematical physics like the OpenAI Yang-Mills result). It looks like AI is finally branching out of math.
In my (limited) experience it does not struggle with empirical research — at least in the procedural aspects thereof — as long as it has access to the right data.
It is very good at understanding research goals, identifying appropriate datasets to work with, doing data wrangling and data science, determining appropriate metrics, running simulations, analyzing results, writing up reports with charts and tables, and proposing follow ups or “high potential next steps” which show a surprising amount of judgment, and if instructed to, executing on those too.
I use Astra and Fable (who delegate to cheaper models for coding) and have them review each others work, and only spot check things when they are done calling each other out. I rarely find issues, and when I do, it’s usually a misunderstanding of intentions.
I’ve only explored a few areas though, including computer vision, vector search, and wireless signal processing, all in applications that, as far as I can tell, are genuinely novel (though of varying levels of novelty and economic value, as the models are fond of pointing out to me!)
I prefer being in the loop and setting the direction — because I do not have unlimited tokens, hah! —- and I regularly identify new directions they had not considered or had prematurely dismissed… so I am a bit more useful than just a financier of tokens! But I can totally see how they could grinding ahead by themselves and brute force useful results, especially if given very concrete goals like a theorem to prove.
I don’t think we’re at Recursive Self Improvement yet, but I don’t doubt for a minute that this is behind the insane pace of development at Anthropic and OpenAI.
I agree, but it is the kind of math done the physicists' way (which I found recently asserted that AI cannot do) and not the kind which is formalizable in Lean.
From skimming the papers, it’s not clear to me what it is actually doing outside of proving what can be considered mathematical theorems (even if that’s done the physicist way). It doesn’t seem to be making testable predictions (but maybe I missed those).
For people not in the know, it's common - both now and historically - for physicists to use "mathematical techniques" that have not been proven valid.
I took grad level physics courses, and this drove me bonkers. I was accustomed to the comfort mathematics provided me. If something was unclear, I could always dig deeper and find a proof. But a lot of the stuff the physics professors would do - I couldn't, and nor could they. They just responded with "intuition". Some of them had dug deep earlier in their careers and stated that the technique was likely invalid mathematically, "but still useful in getting answers experiments can verify".
I won't even talk about relying on diverging series...
If you attempt to formalize math involving such assumptions, one expects to encounter a proliferation of contradictions. So I’m curious to see how the contributions of AI authors is handled by the physics community. With pure math, you can ostensibly write a compact statement of a theorem, and if so, trust the process.
We used to joke in my physics undergrad that they didn’t need a math department because we were doing all the stuff they were doing but with actual applications. Then the mechanical engineers said the same thing about us!
From a cursory glance the bootstrap paper is really hard to read, however the blog post is easier and seems more human written.
The first result is cool (but basically a complex error-prone calculation), the second is much much cooler, Claude proved the upper limit for the mass of a black-hole which is consistent with a given central charge. (thereby making it a consistent theory of quantum gravity), any theory predicting above this is not a theory of quantum gravity of that universe. This is very exciting!
> thereby making it a consistent theory of quantum gravity
Careful. Thereby making a not-inconsistent-via-bootstrap theory. Such a theory could be inconsistent for other reasons shown by other methods. The article expresses a desire to show inconsistency for M1 above c/12 for example, but it's now been shown that the boostrap method can't get you that far.
"Neither result describes our own universe directly, and neither has practical applications, but both contribute to a long-running theoretical effort to understand how quantum mechanics and relativity fit together."
For example the famous Schwarzschild solution (that predicts black holes) doesn't apply to our universe, because it assumes a black hole that's been there forever, not one that formed. But it was still very useful for our understanding.
It also assumes zero angular momentum, which we know is essentially impossible in our universe. And rotating black holes (the Kerr metric) are much stranger beasts.
Apart from Claude/Anthropic no attribution whatsoever - neither on the Hexagon entry nor in the PDF. Isn't this a novelty, even for papers that are 100% the product of AI?
Now let's wait so we find if it stole someone's work.
If the solution is correct
If it does produce a meaningful and useful result.
Lately people claim victory too early
The "stolen" work and strategy of the solution itself was there from 2023 and even buckmaster used ai who "stole" the idea from Córdoba–Martínez-Zoroa programme. There's no proof that the solution even used much of the solution of the buckmaster. You people blindly repeat everything like a sheep without critical thinking. AI + humans solving it within days compared to decades of humans only solving is what should be commended. Instead we got people like you who will take humanity to stone ages due to anti capitalist ideology
I hope you one day come to appreciate the value of human effort and how it has allowed these models to exist and continue to exist. And that people should be looked after by the community for which they contribute with their effort rather than dismissed and discarded. It's a quality of life thing. For everyone.
Also: The main result was first obtained autonomously by an internal Claude research model. The publicly available Claude Opus 5.5 later obtained the same result independently. Claude then wrote this paper, following brief guidance from a physicist, who also checked every main result.
A lot of us are willing and today have already taken the bet that a model or system that can solve hard problems across math, physics, biology is genuine intelligence. It will command a great premium or it may even figure out in the not so distant future how to make money on its own.
Those theoretical physics papers look like the real thing (i.e. not mathematical physics like the OpenAI Yang-Mills result). It looks like AI is finally branching out of math.
https://news.ycombinator.com/item?id=50430239
Is it really, though? It's still a heavily mathematical domain, rather than one involving empirical research, where AI still struggles.
It is very good at understanding research goals, identifying appropriate datasets to work with, doing data wrangling and data science, determining appropriate metrics, running simulations, analyzing results, writing up reports with charts and tables, and proposing follow ups or “high potential next steps” which show a surprising amount of judgment, and if instructed to, executing on those too.
I use Astra and Fable (who delegate to cheaper models for coding) and have them review each others work, and only spot check things when they are done calling each other out. I rarely find issues, and when I do, it’s usually a misunderstanding of intentions.
I’ve only explored a few areas though, including computer vision, vector search, and wireless signal processing, all in applications that, as far as I can tell, are genuinely novel (though of varying levels of novelty and economic value, as the models are fond of pointing out to me!)
I prefer being in the loop and setting the direction — because I do not have unlimited tokens, hah! —- and I regularly identify new directions they had not considered or had prematurely dismissed… so I am a bit more useful than just a financier of tokens! But I can totally see how they could grinding ahead by themselves and brute force useful results, especially if given very concrete goals like a theorem to prove.
I don’t think we’re at Recursive Self Improvement yet, but I don’t doubt for a minute that this is behind the insane pace of development at Anthropic and OpenAI.
"It's not much, but it's honest work."
For people not in the know, it's common - both now and historically - for physicists to use "mathematical techniques" that have not been proven valid.
I took grad level physics courses, and this drove me bonkers. I was accustomed to the comfort mathematics provided me. If something was unclear, I could always dig deeper and find a proof. But a lot of the stuff the physics professors would do - I couldn't, and nor could they. They just responded with "intuition". Some of them had dug deep earlier in their careers and stated that the technique was likely invalid mathematically, "but still useful in getting answers experiments can verify".
I won't even talk about relying on diverging series...
If you attempt to formalize math involving such assumptions, one expects to encounter a proliferation of contradictions. So I’m curious to see how the contributions of AI authors is handled by the physics community. With pure math, you can ostensibly write a compact statement of a theorem, and if so, trust the process.
> https://news.ycombinator.com/item?id=50563604
The first result is cool (but basically a complex error-prone calculation), the second is much much cooler, Claude proved the upper limit for the mass of a black-hole which is consistent with a given central charge. (thereby making it a consistent theory of quantum gravity), any theory predicting above this is not a theory of quantum gravity of that universe. This is very exciting!
Careful. Thereby making a not-inconsistent-via-bootstrap theory. Such a theory could be inconsistent for other reasons shown by other methods. The article expresses a desire to show inconsistency for M1 above c/12 for example, but it's now been shown that the boostrap method can't get you that far.
https://nautil.us/what-does-any-of-this-have-to-do-with-phys...
"Neither result describes our own universe directly, and neither has practical applications, but both contribute to a long-running theoretical effort to understand how quantum mechanics and relativity fit together."
Representatives of AI labs and other organizations may email admin@hexagonmath.org to arrange a verified lab account.
https://hexagonmath.org/organizations/anthropic
Also: The main result was first obtained autonomously by an internal Claude research model. The publicly available Claude Opus 5.5 later obtained the same result independently. Claude then wrote this paper, following brief guidance from a physicist, who also checked every main result.
Wild.
https://hexagonmath.org/who
It was a comforting moment, for sure. But ephemeral, self-imposed ignorance is no bliss at all :(