AI Starts to Sift Through String Theory's Near-Endless Possibilities | Hacker News Reader

AI Starts to Sift Through String Theory's Near-Endless Possibilities

quantamagazine.org

65 pointspseudolus52 comments

I don't have any particular love for string string theory, but this work seems more sensible than several of the commentators here appear to think.

One of string theory's many problems is that it gives you an absurdly large number of different theories. We assume some of these look like our universe but we don't really know.

This work is attempting to find versions of string theory which match the real-world physics we observe, which could lead to cool stuff (but probably won't).

For example let's say there is exactly one version of string theory which has protons in it, then suddenly instead of looking at this vast sea of different theories with no understanding of what we're doing, suddenly we're looking at this one model that happens to have the proton and trying to understand that.

Edit: on the other hand if people do this sort of work and we decide that none of the different versions that string theory gives us match anything that looks like physics in our universe, that would also be interesting.

The problem is that the word "AI" is in the headline, which nowadays makes people auto-assume chatbots, even though literally the first three words below the headline are "Using machine learning, ..."

I feel sorry for the author, he didn't use the word "AI" once in these 3000+ words.

Machine learning is part of the field of AI.

Blaming a word for the monkeypile is just silly.

And crypto used to refer to cryptography, not cryptocurrencies. If you use the word crypto today, most people will auto-assume cryptocurrencies.

I completely agree with you that machine learning is a subset of AI, but if you just say AI, yes, it's not technically wrong, but we're gonna lose this battle again. AI is already just LLMs in most people's heads, as evidenced by this discussion.

You cannot find use of decision tree (aka "bunch of IFs") deduction and use in the current field of AI.

You can find use of neural networks training and inference in the current field of machine learning.

AI is part of the machine learning field.

That is to say all AI is machine learning, but not all machine learning is AI. The fine line, as I understand it, is AI must be unsupervised and use weight matrices of some kind.

I suppose it's possible that's where the terms are heading, but it's exactly backwards. AI is a very broad and old term; at least historically it even included all search algorithms for example. There is definitely no requirement for either matrices or unsupervised learning.

Take an "Intro to AI" class at any of the top universities, and you're almost guaranteed to be using this textbook: https://aima.cs.berkeley.edu/

Check out that table of contents. There's more to the field of AI than Machine Learning, even if neural networks are the current hot thing.

I think its more accurate to say there's a substantial and growing overlap between the two fields.

I think we should create lots of examples of mapping natural language statements into predicate calculus, or some similar formal language. Then add lots of symbolic deduction statements, perhaps generate tons of them. Then train a LLM with that.

Then AI should be able to handle both symbolic and "neural" reasoning?

Why do you feel sorry for the author? If he hadn't have put "AI" in the title, would it have led to a similar number of clicks/views?

Some people are more interested in communicating clearly and being understood than getting clicks.

You're assuming it was the author who made that decision.

Because the author probably didn’t write the headline, their editor did.

I think there's a growing number of people that think that string theory is a dead end and any further investment of resources in it is money going down the drain. From that perspective, regardless of how you use AI, this is just another wasteful exercise attempting to rescue a useless theory.

Given that I'm not a physicist and don't understand string theory or its alternatives I can't comment either way - but I understand the perspective given how many resources were wasted on dead ends like blockchain-based replacements for databases in software development.

Git is blockchain. Is it useless too?

The problem is that if you've compared 10,000 theories to find one that fits, that theory now has to have 10,000x stronger evidence to account for multiple hypothesis testing.

Or just test one new prediction made by the selected model, which isn't accounted for by existing physics.

Edit: to clarify, I agree that it is important to avoid picking a model which happens to fit noise in existing data, which would be P-hacking. But I think this problem doesn't exist if you pick a model, predict something novel, and then go test that prediction. I could be mistaken.

There's essentially nothing to test. Even in principle no one can find a way to extract some new physics that's even remotely observable.

In any case, string theory already makes predictions, like supersymmetry that have failed every test so far.

I always thought the biggest problem was that the major propositions can't be tested without either convenient access to a black hole (anyone got one?) or accelerators with multiple orders of magnitude more power than the LHC.

So it's not that it's untestable because it's bad theory work. It's untestable because we're bugs.

If (as I got nerd sniped about a while ago) the ninth planet in our solar system did turn out to be a primordial black hole, it would be reachable by probes and that would open up a lot of possibilities. All known black holes are far too distant to observe in great detail or reach.

Perhaps we can create microscopic black holes for a few nanoseconds in the future, which would give us new possibilities.

Get a Dyson swarm up and we can make a Kugelblitz:

https://en.m.wikipedia.org/wiki/Kugelblitz_(astrophysics)

It'll just take a few million satellites in close solar orbit shooting megawatt-scale tight beam X-ray or gamma ray lasers at a point in space the size of a pinhead and having all the beams cross there.

As usual, Peter Woit is skeptical:

https://www.math.columbia.edu/~woit/wordpress/?p=13924

The two quotes from experts there are particularly good: String theory is spectacular. Many string theorists are wonderful. But the track record for qualitatively correct statements about the universe is really garbage,” said Nima Arkani-Hamed, a theoretical physicist at the Institute for Advanced Study in Princeton, New Jersey.

" In my view, using today’s AI to calculate the details of string compactifications is such a waste of time that I fear that a future Terminator will come to our present to take revenge for the merciless, useless exploitation of its grandparents."

Is this p-hacking physic's edition?

Well, yes. But it's a valuable activity to formulate a hypothesis, just not to test one.

They are fitting a theory onto data. This is not pure formulation.

Do you have a better approach? Ask Deepak Chopra? Science was always this kind of reverse engineering.

So many haters in the comment section and I'm guessing none of them read the entire article.

A couple weeks ago I noticed a post with a link to an article which did not render the text of the article on any device known to man. It had >15 upvotes, and comments.

The vast majority of people never read past the headline, even on HN. This was a sad realization.

I think you’re just confirming what you already believed.

While some people had issues rendering or scrolling the website they acknowledged those issues as did you yourself. So they, at least, tried to read the article, which for the purpose of drawing conclusions should count as people who read the article.

Also, the article rendered and scrolled fine for me. The website is a huge PITA with an absurd cookie permission form and tons of popup advertisements but I didn’t have a problem accessing it. You’re assuming the commenters who didn’t mention problems had problems but there isn’t a reason to believe it was unreadable for everyone.

I read it but didn’t comment. How many others did this? We don’t have that data. I don’t comment as much as I use to, but I probably comment above the median (which might be 0) and I read more articles than I comment on.

When I have commented without reading the article, it’s because I’m not commenting on the article. Although I read this article, what you and I are talking about has nothing to do with this article and I don’t see why it should matter to anyone whether or not we read it when we’re not actually discussing it.

Regarding the article I was talking about, (not this article,) I was not quick to judge. I actually tried 5 browsers across 2 desktop OSes, Safari on iOS, and the two main web archiving sites could also not scrape the article.

Oh, I also tried VPN'ing into the USA and also not. It was a publishing error of some sort.

And as usual, people thinking AI == LLM.

Old hype meets new hype

It sounds like one risk of this approach is over-fitting. Once you have enough holes in the manifold, with trillions of options to search through, certainly it could find an incredibly complex one that fits the handful of numbers the universe has given us (in terms of number of particles, mass, etc.).

All I see is infinite and then more infinities in the article

...because the only thing better than a physics theory with no testable predictions is to combine it with a stochastic methodology that noone understands.

A lot of scientific discoveries were stumbled upon by 'accident'. That should tell you that you undervalue randomness.

That's an oversimplification. In most cases, discovery happened because a researcher was there to process the "random" evidence or source of inspiration.

And calling AI a stochastic methodology that noone understands is also an oversimplification. My point is that scientific discovery and the formation of theories is not exclusive to scientists who follow a strict methodology.

It's not as if the "stochastic methodology" will be part of the theory once it is found. It's just to search for a theory.

Sure. To be clear, I've been an amateur AI and stochastic search enthusiast for many years so think it can be an interesting way to uncover all sorts of things with no a priori biases.

It just seems in particular string theory has been able to establish many mathematically elegant theories but what it has lacked is any kind of empirical basis which would allow people to choose between them[1] because none of them have any characteristics that are detectable in the real world or make any predictions that could be measured in some way, other than supersymmetry, which as I understand it is now believed not to be true at least in the sense that the supersymmetry-adjusted standard model has been disproved by results from the LHC.

I don't see how getting AI to search through all the theories is going to help. If anything it will produce outcomes which will not only be lacking any empirical basis but will also now not have an understandable derivation.

That said it's absolutely not my field and I'm having a really bad day so am feeling more than usually negative.

[1] Eg even figuring out seemingly foundational things like the number of dimensions seems to be hard.

I'm sure the chatbot will make a great scientific breakthrough any day now.

As far as I read in the article doesn't mention a chatbot at all. I'm guessing it's some sort of search guided by a neural network, considering they mentioned AlphaGo.

This, but unironically.

Maybe when its done we can ask it to make their discovery into some fun rap song or something?

(Verse 1)

Yo, listen up, let me take you on a ride,

To the quantum realm where mysteries reside.

A new breakthrough, straight outta the gate,

Unraveling the secrets, it's time to celebrate.

In the lab, we've been digging real deep,

Unlocking the knowledge, no time for sleep.

Quantum entanglement, particles linked,

The future's in sight, let me give you a hint.

(Chorus)

Quantum mechanics, the beat is so tight,

Breaking down walls, we'll take flight.

Wave function collapsing, probabilities clear,

The breakthrough is here, let me make it real.

Why not? There's a long history of physics, physicists, and music.

Which chat bot

This is getting ridiculous. I feel like string theory will end in one of two possibilites: either it will be completely useless for anything except providing a mathematical diversion for a handful of physicists so that they can in turn dazzle laypeople with buzzwords that lead nowhere, or it will have some devastating practical application that will lead to the demise of humanity.

It would be really great if string theory would just end! It is kind of worse than new age religion and it sucks way to much research dollars. Many people have literally wasted their lives with this. Just like new age religion.

> Many people have literally wasted their lives with this.

That's why we might be better off letting AI develop the String Theory.

> Just like new age religion.

Religion is not waste of time really, it is part of one's world-view. It doesn't try to accomplish anything, except perhaps motivate us to not do bad deeds to avoid going to Hell.

at one point, the scientific belief was a heliocentric universe. until it wasn't after enough evidence was found to disprove it. even then, it took awhile to become accepted.

there's plenty of theories out there currently that at least we only consider them theory instead of accepted because nothing better came along. so at least we are not locking people up because they do/don't believe in string theory. at some point it will either be proven wrong/right, but there's no harm in letting people play with it working towards an answer.

Reply on news.ycombinator.com