AI comes up with bizarre physics experiments, but they work
quantamagazine.org
quantamagazine.org
It looks like all the results were driven by optimization algorithms, and yet the writing describes AI 'using' concepts and "tricks". This type of language is entirely inappropriate and misleading when describing these more classical (if advanced) optimization algorithms.
Looking at the paper in the first example, they used an advanced gradient descent based optimization algorithm, yet the article describes "that the AI was probably using some esoteric theoretical principles that Russian physicists had identified decades ago to reduce quantum mechanical noise."
Ridiculous, and highly misleading. There is no conceptual manipulation or intuition being used by the AI algorithm! It's an optimization algorithm searching a human coded space using a human coded simulator.
The article is so dumbed down that it's not clear if there is even any ML involved or if this is just an evaluation of combinatorial experimental setups.
> The outputs that the thing was giving us were really not comprehensible by people,
> Adhikari’s team realized that the AI was probably using some esoteric theoretical principles that Russian physicists had identified decades ago to reduce quantum mechanical noise.
I'll chalk this one up to the Russians, not "AI".
I don't really get your rabid dismissal. Why does it matter that they are using optimisation models and not LLMs? Nobody in the article is claiming to have used LLMs. In fact the only mention of it is lower down where someone says they hope it will lead to advances in automatic hypothesis generation. Like, fair enough?
If it was an LLM based model this could be a correct statement, and it would suggest a groundbreaking achievement: the AI collated esoteric research, interpreting it correctly and used that conceptual understanding to suggest a novel experiment. This might sound far fetched, but we already have LLM based systems doing similar... Their written statement is plausible given the current state of hype (and also a plausible, though ground breaking, given the current state of research).
In reality, the statement is incorrect. The models did not 'use' any concepts (and the only way to know that the article is wrong is to actually bother to consult the original paper, which I did).
The distinction matters: they implied something ground breaking, when the reality is cool, but by no means unprecedented.
Tldr: using concepts is not something classic ML algorithms do. They thus directly erroneously imply (a groundbreaking) foundation model based (or similar) approach. I care because I don't like people being mislead.
Again, the authors never said anything about language models. That's entirely on you.
Philosophical discussions aside, it is entirely possible for current AI to use concepts (but the research they are describing does not employ that kind of AI).
I also think most lay people seeing the term AI are likely to think of something like ChatGPT.
It is a) literally incorrect what they write, and b) highly misleading to a lay person (who will likely think of something like ChatGPT when they read the term AI). Why are you defending their poor writing?
Because that's how language works - it's inherently ambiguous, and we interpret things in the way that makes the most sense to us. Your interpretation makes no sense to me, and requires a whole host of assumptions that aren't present in the article at all (and are otherwise very unlikely, like an AI that can literally work at the level of concepts).
> Why are you defending their poor writing?
I'm defending them because I don't think it's poor writing.
A: a grammatically incorrect statement, saying that "the AI used theory", when they mean that "the AI's design can be understood using theory" (or more sloppy "that the design uses the theory").
B: a grammatically valid if contentious-to-you statement about an LLM or knowledge graph based system (e.g., something like the AI Scientist paper) parsing theory and that parsing being used to create the experiment design.
As I have explained, B is a perfectly valid interpretation, given the current state of the art. It is also valid historically, as knowledge graph based systems have been around for a long time. It is also the likely interpretation of a lay person, who is mainly exposed to hype and AI systems like chatGPT.
Regardless, they a) introduce needless ambiguity that is likely to mislead a large proportion of readers. And b) if they are not actively misleading then they have written something grammatically incorrect.
Both findings mean that the article is a sloppy and bad piece of writing.
This particular sentence is also only a particular example of how the article is likely to mislead.
Anyway, I don't think it's reasonable for me to ask you for evidence here, so let's just agree to disagree.
Out of curiosity, are you familiar with work like "the AI Scientist"? Having an LLM based AI suggest experiments based on parsing scientific literature is not outlandish.
Examples of people being confused:
rlt: The discovering itself doesn’t seem like the interesting part. If the discovery wasn’t in the training data then it’s a sign AI can produce novel scientific research / experiments.
wizzwizz4 in reply: It's not that kind of AI. We know that these algorithms can produce novel solutions. See https://arxiv.org/abs/2312.04258, specifically "Urania".
About a quarter of the comments here I just have to assume what definition of AI they're talking about, which changes the meaning and context significantly.
However, scientifically, I think there's a real challenge to clearly delineate from the standpoint of 2025 what all should fall under the concept of AI -- we really lose something if "AI" comes to mean only LLMs. Everyone can agree that numeric methods in general should not be classed as AI, but it's also true that the scientific-intellectual lineage that leads to modern AI is for many decades indistinguishable from what would appear to be simply optimization problems or the history of statistics (see especially the early work of Paul Werbos where backpropagation is developed almost directly from Bellman's Dynamic Programming [1]). The classical definition would be that AI pursues goals under uncertainty with at least some learned or search‑based policy (paradigmatically but not exclusively gradient-descent of loss function), which is correct but perhaps fails to register the qualitative leap achieved in recent years.
Regardless -- and while still affirming that the OP itself makes serious errors -- I think it's hard to find a definition of AI that is not simply "LLMs" under which the methods of the actual paper cited [2] would not fall.
[1] His dissertation was re-published as The Roots of Backpropagation. Especially in the Soviet Union, important not least for Kolmogorov and Vapnik, AI was indistinguishable from an approach to optimization problems. It was only in the west where "AI" was taken to be a question of symbolic reasoning etc, which turned out to have been an unsuccessful research trajectory (cf the "AI winter").
- methods that were devised with domain knowledge (= numerical methods)
- generic methods that rely on numerical brute forcing to interpolate general behaviour (= AI)
The qualitative leap is that numerical brute forcing is at a stage where it can be applied to useful enough generic models.
There's a fundamental difference between any ML based method and, say, classic optimization. Let's take a simple gradient descent. This solves a very specific (if general) class of problems: min_x f(x) where f is differentiable. Since f is differentiable, someone had the (straightforward) idea of using its gradient to figure out where to go. The gradient is the direction of greatest ascent, so -grad(f) comes as a good guess of where to go to decrease f. But this is local information, only valid at (or rather in the vicinity of) a point. Hence, short of improving the descent direction (which other methods do, like quasi-Newton methods, which allow a "larger vicinity" of descent direction pertinence), the best you can do is iterate along x - h grad(f) at various h and find one that is optimal in some sense. How this is optimal is all worked out by hand: it should provide sufficient decrease, while still giving you some room for progression (not too low a gradient), in the case of the Wolfe-Armijo rules, for example.
These are all unimportant details, the point is the algorithms are devised by carefully examining the objects at play (here, differentiable functions), and how best to exploit their behaviour. These algorithms are quite specific; some assume the function is twice differentiable, others that it is Lipschitzian and you know the constant, in others you don't know the constant, or the function is convex...
Now in AI, generally speaking, you define a parametric function family (the parameters are called weights) and you fit that family of functions so that it maps inputs to desired ouputs (called training). This is really meta-algorithmics, in a sense. No domain knowledge required to devise an algorithm that solves, say, the heat equation (though it will do so badly) or can reproduce some probability distribution. Under the assumption that your parametric function family is large enough that it can interpolate the behaviour you're looking after, of course. (correct me on this paragraph if I'm wrong)
To summarize, in my (classic numerics trained) mind, classic numerics is devising methods that apply to specific cases and require knowledge of the objects at play, and AI is devising general interpolators that can fit to varied behaviour given enough CPU (or GPU as it were) time.
So, this article is clearly not describing AI as people usually mean it in academia, at least. I'll bet you a $100 the authors of the software they used don't describe it as AI.
This description reminds me of NASA’s evolved antennae from a couple of decades ago. It was created by genetic algorithms:
> Just nothing that a human being would make, because it had no sense of symmetry, beauty, anything. It was just a mess.
NASA describing their antenna:
> It has an unusual organic looking structure, one that expert antenna designers would not likely produce.
— https://ntrs.nasa.gov/citations/20060024675
The parallel seems obvious to me.
This comes up a lot and always strikes me as rather anti-science, even anti-rationality in general. To speed run the typical progression of this argument, someone says alchemy and astrology occasionally "work" too if you're determined to ignore the failures. This point is then shot down by a recap about the success of QM despite Einstein's objections, success of the standard model even with lots of quasi-empiricism etc, etc.
Structurally though.. if you want to claim that the universe is fundamentally weird and unknowable, it's very easy to argue this, because you can always ignore the success of past theory and formalisms by saying that "it was nice while it lasted but we've squeezed all the juice out of that and are in a new regime now". Next you challenge your detractors to go ahead and produce a clean beautiful symmetric theory of everything to prove you wrong. That's just rhetoric though, and arguments from model/information/complexity theory etc about fundamental limits on what's computable and decidable and compressible would be much more satisfying and convincing. When does finding a complicated thing that works actually rule out a simpler model that you've missed? https://en.wikipedia.org/wiki/Minimum_description_length#MDL...
I've had a few of those, I think they're usually a symptom of stack corruption.
Timing matters.
The design seemed alien and somewhat organic, but I can’t seem to find it now.
Looking at things like bicycles designed this way leaves me suspicious that it doesn’t actually have the power to derive interesting insights about material properties. I suspect future versions may end up starting to look more mechanical as it discovers that, for example, something under tension should be a straight line.
Why would that be the case, though? It's conceivable that optimal shapes be very different from what our intuitions suggest.
But I recommend logseq for anyone new! All the best features, none of the weird energy of roam, where i worry the founder is training ML models on data.
https://www.damninteresting.com/on-the-origin-of-circuits/
They used genetic algorithms to evolve digital circuits directly on FPGAs. The resulting design exploited things like electromagnetic interference to end up with a circuit much more efficient than a human could've created.
In my mind this brings some interesting consequences for 'AI apocalypse' theories. If the AI understand everything, even an air gap might not be enough to contain it, since it might be able to repurpose some of its hardware for wireless communication in ways that we can't even imagine.
It's a throwaway mechanic in the comic, but it seems plausible.
In certain places the power companies are/were passing time information throughout the whole grid - https://www.nist.gov/publications/time-and-frequency-electri...
Whatever AI comes up with by 2030 is going to be much more clever and unexpected.
I'll have a read of the paper, seems like it's similar in concept
Note for example TEMPEST surveillance, or using a distant laser to pickup speech in a room based on window vibrations. Air-gap traversal is easily done by exploiting human weaknesses (e.g. curiousity to pick up a USB drive to see what's on it), and was successfully done by Stuxnet.
But as someone said earlier, the real interesting part is when/if they start figuring out novel concepts we as humans haven't even considered.
Or data exfiltration through fan noise (60 bits/min): https://www.sciencedirect.com/science/article/abs/pii/S01674...
Or data transfer between computers using only speakers: https://arxiv.org/abs/1803.03422
The list goes on.
> Five individual logic cells were functionally disconnected from the rest — with no pathways that would allow them to influence the output — yet when the researcher disabled any one of them the chip lost its ability to discriminate the tones. Furthermore, the final program did not work reliably when it was loaded onto other FPGAs of the same type.
The bias toward familiarity is detrimental to edge research, but on the other hand if no one smooth the baseline, most advanced knownledge will remain just that and will never reach their full utility to humans. Finding the proper set of concepts that makes it click can be very complicated. Finding a communicable simple thought framework to let other also enjoy it and leverage on it to go further can be at least as hard.
But I'm not paid by the click, so different incentives.
[1] https://en.wikipedia.org/wiki/Random_forest
[2] https://en.wikipedia.org/wiki/Naive_Bayes_classifier#Trainin...
[3] https://en.wikipedia.org/wiki/K-nearest_neighbors_algorithm
AI for attempts at general intelligence. (Not just LLMs, which already have a name … “LLM”.)
ML for any iterative inductive design of heuristical or approximate relationships, from data.
AI would fall under ML, as the most ambitious/general problems. And likely best be treated as time (year) relative, i.e. a moving target, as the quality of general models to continue improve in breadth and depth.
crypto must now be named cryptography and AI must now be named ML to avoid giving the scammers and hypers good press.
I think image and video generation that aren't based on LLMs can also use the term AI without causing confusion.
You just made a lot of 20th century AI researchers cry.
The real problem is not people using the term incorrectly, it's papers and marketing material using the term incorrectly.
You can have your own definition of words but it makes it harder to communicate.
What? I literally don't know a single person anymore who doesn't know what chatGPT is. In this I include several elderly people, a number of older children and a whole bunch of adults with exactly zero tech-related background at all. Far from it being only known to some, unless you're living in a place with essentially no internet access to begin with, chances are most people around you know about chatGPT at least.
For OpenAI, different story, but it's hardly little-known. Let's not grossly understate the basic ability of most people to adapt to technology. This site seems to take that to nearly pathological levels.
I don't question people's ability to adapt, people are adaptible. But if you've never even heard of it, what is there to adapt to?
Example: I live in a country that still has a great deal of deep poverty, it's what's called a "developing economy" (sort of an odd phrase since aren't all economies always still developing at all times? but I digress) and even in all but the most deeply poor rural places here, most people frequently use the internet. And I know nobody who doesn't at least know of chatGPT or about how AI can now talk to you like a person would and answer all kinds of questions, let alone not knowing about things like Google and so forth.
For me, when someone says, "I'm working on AI", it's almost meaningless. What are you doing, actually?
https://github.com/artificial-scientist-lab/GWDetectorZoo/
Nothing remotely LLM-ish, but I'm glad they used the term AI here.
> We develop Urania, a highly parallelized hybrid local-global optimization algorithm, sketched in Fig. 2(a). It starts from a pool of thousands of initial conditions of the UIFO, which are either entirely random initializations or augmented with solutions from different frequency ranges. Urania starts 1000 parallel local optimizations that minimize the objective function using an adapted version of the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm. BFGS is a highly efficient gradient-descent optimizer that approximates the inverse Hessian matrix. For each local optimization, Urania chooses a target from the pool according to a Boltzmann distribution, which weights better-performing setups in the pool higher and adds a small noise to escape local minima.
https://journals.aps.org/prx/abstract/10.1103/PhysRevX.15.02...
> For each local optimization, Urania chooses a target from the pool according to a Boltzmann distribution, which weights better-performing setups in the pool higher and adds a small noise to escape local minima. These choices add a global character to the exploration. When one of the local optimizations of Urania finds a better parameter setting for a setup in the pool, it replaces the old solution with the superior one. Upon convergence, Urania repeats and chooses a new target from the pool. In parallel, Urania simplifies solutions from the pool by probabilistically removing elements whose removal does not impact the overall sensitivity.
> Starting from a dense or fully connected graph, PyTheus uses gradient descent combined with topological optimization to find minimal graphs corresponding to some target quantum experiment
There are a few things like that where we can throw AI at a problem is generating something better, even if we don't know why exactly it's better yet.
Isn't that a delay line? The benefit being that when the undelayed and delayed signals are mixed, the phase shift you're looking for is amplified.
I wonder how many times these designes were dismissed because humans who think out of the box too much are dismissed. It seems that students are encouraged NOT to do so, severely limiting how far out they can explore.
AI is going through the wild phase were people are allowing it to test, as soon as the limits are understood the framework of limitations and the rational system built around will inevitably happen.
I'm guessing AI gets the benefit of the doubt here because its ideas will be interesting and publishable no matter the outcome
* Semmelweis - medicine. Demonstrated textbook obstetric technique at Vienna General Hospital, then produced statistically impeccable data showing that hand-washing slashed puerperal fever mortality. Colleagues drove him out of the profession, and he died in an asylum.
* Barbara McClintock - genetics. Member of the National Academy, meticulous corn geneticist; her discovery of “jumping genes” (transposons) was ignored for 30 years and derided as “mysticism.”
* Georg Cantor - mathematics. Earned a Ph.D. and published dozens of orthodox papers before writing on transfinite numbers; was then declared “a corrupter of youth”. Career was blocked, contributing to a breakdown.
* Douglas Engelbart - computer science. Published conventional reports for years. When he presented the mouse, hypertext, and videoconferencing in “The Mother of All Demos” (1968), ARPA funding was slashed and he was professionally sidelined for the next twenty years.
Then you've got Stravinsky, Van Gogh, Caravaggio, James Joyce; all who displayed perfect 'classical' techniques before doing their own thing.
In economics you've got Joan Robinson and Elinor Ostrom.
And let's not forget Galileo. I'd even put Assange in this list.
So, "following the rules" before attempting to revolutionize your field doesn't seem to actually help all that much. This is a major problem, consistent across many centuries and cultures, which ought to be recognized more.
... which is AI. AI existed long before GPTs were invented, and when neural networks were left unexplored as the necessary compute power wasn't there.
"AI comes up with bizarre ___________________, but it works!"
Imagine these headlines mutating slowly into “all software engineering performed by AI at certain company” and we will just dismiss it as generic because being employed and programming with keyboards is old fashioned. Give it twenty years and I bet this is the future.
Of course today call something "AI" and suddenly interest, and presumably grant opportunities, increase by a few orders of magnitude.
OTOH, AI is very much a search in multidimensional spaces, it is so into it, that it would probably make sense to say that gradient descent is an AI tool. Not because it is used to train neural networks, but because the specialty of AI is a search in multidimensional spaces. People probably wouldn't agree, like they don't agree that Fundamental Theorem of Algebra is not of algebra (and not fundamental btw). But the disagreement is not about the deep meaning of the theorem or gradient descent, but about tradition and "we always did it this way".
The researchers in this article didn't do that. They used gradient descent to choose from a set of experiments. The choice of experiment was the end result and the direct output of the optimisation. Nothing was "learned" or "trained".
Gradient descent and other optimisation tools are used in machine learning, but long predate machine learning and are used in many other fields. Taking "AI" to include "anything that uses gradient descent" would just render an already heavily abused term almost entirely meaningless.
It is a simple iterative algorithm that goes from one point to the next. It doesn't even have memory of previous steps (caveat, the authors used BFGS which approximates the Hessian with previous gradient iterates, but this is still not AI). There is no finding weights or any such thing.
If every for loop is AI, then we might as well call everything AI. Can you pass me the AI, please?
This conflation of everything with AI is precisely why people say things like "gradient descent is most often used in ML" without evidence, and this likely being wrong. No, instead it is 1) ML is the currently most prominent (to the public) use of mathematical optimization and 2) everything else is called AI to the public so they conflate that with ML even when it isn't.
Take a random employee at an engineering or applied sciences (non experimental) lab, ask them if they ever use mathematical optimization, chances are a majority will tell you they do. The vast majority of these are not using or devising ML algorithms.
This matters because of what is clear from this thread. Some people devise a classic algorithm that requires intimate knowledge of the problem at hand, the press calls it AI, the public thinks it's AI, registers one more case of "AI as the tool to replace all others". The Zeitgeist becomes that everything else can go to the bin, and AI (by the more restrictive definition) receives disproportionate attention and funds. Note that funding AI research would not fund the people in the headline, unless they do like the minority of bandits that rebrand their non-AI work with AI keywords.
When the equation is too complex for analytic methods but good enough for gradient descent that equation is overwhelmingly the majority of the time characterized as AI.
A great deal of applied mathematics is related to finding a minimum or maximum quantity of something. There are not always constructive methods, sometimes (often) there's no better way than to step through a generic optimization method.
Some quick examples clearly unrelated to ML, and very common as they relate to CAD (everywhere from in silico studies to manufacturing) and computer vision:
- projecting a point on a surface
- fitting a parametric surface through a point cloud
Another example is non-linear PDEs. Some notable cases are Navier-Stoke's equations, non-linear elasticity, or reaction-diffusion. These are used in many industries. To solve non-linear PDEs, a residual is minimized using, typically, quasi-Newton methods (gradient descent's buff cousin). This is because numerical schemes only exist for linear equations, so you must first recast the problem as something linear (or a succession of those, as it were).
By the way, I might add that most PDEs can be equivalently recast as optimization problems.
Yet another is inverse problems: imaging (medical, non destructive testing...), parameter estimation (subsoil imaging), or even shape optimization. Similarly, optimal control. (similar in that it is minimizing a quantity under PDE constraints)
To summarize, almost every time you seek to solve a non-linear equation of any kind (of which there are many completely unrelated to ML), numerical optimization is right around the corner. And when you seek to find "the best" or "the least" or "the most" of something, optimization. Clearly, this is all the time.
I think I've provided a broad enough set of fields with ubiquitous applications, that it is clear optimization is omnipresent and used considerably more often than ML is. As you see, there is no association from optimization to ML or AI, although there is one the other way around. (much like a bird is not a chicken).
People have been posing examples of similar "weird non-human design" results throughout here that are more than a decade old.
The AI rediscovered an interferometer technique the Russian's found decades ago, optimized a graph in an unusual way and came up with a formula to better fit a dark matter plot.
It's like seeing things in clouds or tea leaves.
At least, that's the thinking.
How would a realistic Turing test for "Shakespeare-ness" look like?
Big experts on Shakespeare likely remember (at least vaguely) all his sonnets, so they cannot be part of a blinded study ("Did Shakespeare write this or no?"), because they would realize that they have never seen those particular lines, and answer based on their knowledge.
Maybe asking more general English Lit teachers could work.
IIRC Scott Alexander is doing something similar with his "AI draws nontrivial prompts" bet, and the difference to last year's results was striking.
Also, this really needs blinding, otherwise the temptation to show off one's sophistication and subtlety is big. Remember how oenologists consistently fail to distinguish between a USD 20 and a USD 2000 wine bottle when blinded.
Unless your response is AI generated. In that case, I would tip my hat.
Yes the AI "resurfaced" the work, but it also incorporated the Russian's theory into the practical design. At least enough to say "hey make sure you look at this" - this means the system produced a workable-something w/ X% improvement, or some benefit that the researchers took it seriously and investigated. Obviously, that yielded an actual design with 10-15% improvement and a "wish we had this earlier" statement.
No one was paying attention to the work before.
If I've understood it right, calling this AI is a stretch and arguably even misleading. Gradient descent is the primary tool of machine learning, but this isn't really using it the way machine learning uses it. It's more just an application of gradient descent to an optimisation problem.
The article and headline make it sound like they asked an LLM to make an experiment and it used some obscure Russian technique to make a really cool one. That isn't true at all. The algorithm they used had no awareness of the Russian research, or of language, or experimental design. It wasn't "trained" in any sense. It was just a gradient descent program. It's the researchers that recognised the Russian technique when analyzing the experiment the optimiser chose.
https://journals.aps.org/prx/abstract/10.1103/PhysRevX.15.02...
Reminds me of the square packing problem, with the absurdly looking solution for packing the 17 squares.
It also reminds me of edge cases in software engineering. When I let an LLM write code, I'm often confused how it starts out, thinking, I would have done it more elegantly. However, I quickly notice, that the AI handled a few edge cases I only would habe caught in testing.
Guess, we should take a hint!
sounds kinda like the chip designer AI
(It was a gradient descent optimizer, so probably unconstrained optimization rather than a Constraint Satisfaction Optimization Problem, but it might have had constraints.)
"Modern" AI is just fuzzy logic, connecting massive probabilities to find patterns.
We’ve been doing that for decades, it’s just more recently that it’s come with so much more funding.