AI comes up with bizarre physics experiments, but they work

(quantamagazine.org)

190 points | by pseudolus 8 hours ago

18 comments

  • JimDabell 7 hours ago
    > Initially, the AI’s designs seemed outlandish. “The outputs that the thing was giving us were really not comprehensible by people,” Adhikari said. “They were too complicated, and they looked like alien things or AI things. Just nothing that a human being would make, because it had no sense of symmetry, beauty, anything. It was just a mess.”

    This description reminds me of NASA’s evolved antennae from a couple of decades ago. It was created by genetic algorithms:

    https://en.wikipedia.org/wiki/Evolved_antenna

    • ElFitz 2 hours ago
      There was something similar about using evolutionary algorithms to produce the design for a mechanical piece used to link two cables or anchor a bridge’s cable, optimizing for weight and strength.

      The design seemed alien and somewhat organic, but I can’t seem to find it now.

    • kriro 36 minutes ago
      Reminds me a bit of chess engines that crush the best humans with ease but play moves that human players can identify as "engine moves". In chess the environment is fixed by the rules so I'd assume this deeper understanding of underlying patterns is only amplified in more open environments.
    • eleveriven 3 hours ago
      That evolved antenna looked like something cobbled together by a drunk spider
      • ludicrousdispla 1 hour ago
        or anyone trying to get a decent signal on their TV
    • wickedsight 1 hour ago
      That reminds me of this article:

      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.

      • otikik 32 minutes ago
        I don't remember the title, but someone wrote a story were an AI would use the (imperceptible) flickering of a fluorescent lightbulb and a camera to transmit information across such an "air gap".
      • ACCount36 54 minutes ago
        In practice, we'll just let that AI have a direct internet connection, and also give it enough access to push code straight to prod. For the good measure.
        • gpderetta 28 minutes ago
          Don't forget about full control of the terminators.
    • PicassoCTs 1 hour ago
      The bias is a handicap, the looking for beauty, symmetry, a explanation, a story, its all googles upon googles of warping lenses and funhouse-mirrors, hiding and preventing the perception of truth.
      • psychoslave 49 minutes ago
        Zero is taught routinely to primary schoolers today, but it has been a hard thing to come with for scholar who struggled to nail as smooth as a concept as we know it now.

        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.

    • carabiner 5 hours ago
      [mandatory GA antenna post requirement satisfied]
    • esperent 3 hours ago
      That evolved antenna is a piece of wire with exactly 6 bends. It's extremely simple, the exact opposite of a hard to understand mess.
      • JimDabell 3 hours ago
        This physics experiment:

        > 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.

        • TheOtherHobbes 1 hour ago
          Maybe not so much the implications. If our science is defined by symmetry, beauty, anything - and it is, because so much of physics is literally about looking for symmetries of various kinds - why are we ignoring the loud hints from ML solutions that this is a limiting heuristic?
      • rcxdude 2 hours ago
        Lacking symmetry, it's extremely hard to understand how the antenna actually works (i.e. why those six bends, as opposed to any other random six bends).
        • RedShift1 26 minutes ago
          My best guess is that the edges are oriented such that at the tested frequencies they cause constructive interference inside the antenna therefore boosting the signal. The orientation is weird because that's probably the best way to make it work in all directions, if the edges were in a flat plane, the constructive interference would only work in a single direction.
        • jmogly 38 minutes ago
          I love the idea of faith based technology, where it just works but nobody is capable of comprehending why or how.
      • Biganon 25 minutes ago
        Go ahead then, explain to us how the exact values of the 6 angles make it work so well
  • luketaylor 6 hours ago
    Referring to this type of optimization program just as “AI” in an age where nearly everyone will misinterpret that to mean “transformer-based language model” seems really sloppy
    • saithound 4 hours ago
      Referring to this type of optimization as AI in the age where nearly everybody is looking to fund transformer-based language models and nobody is looking to fund this kind of optimization is just common sense though.
      • benterix 3 hours ago
        You are both right. Because the term "AI" is so vague and can mean so many things, it will be used and abused in various ways.

        For me, when someone says, "I'm working on AI", it's almost meaningless. What are you doing, actually?

    • benterix 3 hours ago
      I think it's actually this repo:

      https://github.com/artificial-scientist-lab/GWDetectorZoo/

      Nothing remotely LLM-ish, but I'm glad they used the term AI here.

    • bee_rider 5 hours ago
      How can one article be expected to fix the problem of people sloppily using “AI” when they mean LLM or something like that?
      • rachofsunshine 5 hours ago
        I use "ML" when talking about more traditional/domain specific approaches, since for whatever reason LLMs haven't hijacked that term in the same way. Seems to work well enough to avoid ambiguity.

        But I'm not paid by the click, so different incentives.

        • Nevermark 4 hours ago
          I like that.

          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.

        • smj-edison 4 hours ago
          Generative AI vs artificial neural network is my go-to (though ML is definitely shorter than ANN, lol).
      • a_victorp 4 hours ago
        Just don't use the term AI. It has no well defined meaning and is mostly intended as a marketing term
        • vanviegen 3 hours ago
          So, "don't do marketing" is your advice?
          • dns_snek 3 hours ago
            Correct, "an editorially independent online publication launched by the Simons Foundation in 2012 to enhance public understanding of science" shouldn't be doing marketing and contributing to the problem.
      • coldtea 2 hours ago
        By doing its part and using the term correctly.

        The real problem is not people using the term incorrectly, it's papers and marketing material using the term incorrectly.

        • braiamp 1 hour ago
          Lets be real here, the people with the money bags don't care either.
      • Lionga 4 hours ago
        Just do not use AI for anything except LLMs anymore. Same way that crypto scam has taken the word crypto.

        crypto must now be named cryptography and AI must now be named ML to avoid giving the scammers and hypers good press.

        • omnicognate 2 hours ago
          > AI must now be named ML

          You just made a lot of 20th century AI researchers cry.

          • anthk 1 hour ago
            Ocaml users too. And Haskell.
        • ItsHarper 4 hours ago
          Yep. I dislike it just as much as ceding crypto, but at the end of the day language changes, and clarity matters.

          I think image and video generation that aren't based on LLMs can also use the term AI without causing confusion.

    • advael 3 hours ago
      This exact kind of sloppy equivocation does seem to be one of the major PR strategies that tries to justify the massive investment in and sloppy rollout of transformer-based language models when large swaths of the public have turned against this (probably even more than is actually warranted)
    • victorbjorklund 2 hours ago
      Yea, I can tolerate it when random business people do it. But scientists/tech people should know better.
    • layer8 2 hours ago
      While nowadays misleading as a title, I found the term being used in the traditional sense refreshing.
    • tomrod 5 hours ago
      I know, but can we blame the masses for misunderstanding AI when they are deliberately misinformed that transformers are the universe of AI? I think not!
    • fragmede 4 hours ago
      Thinking "nearly everyone" has that precise definition of AI seems way more sloppy. Most people haven't even heard of OpenAI and ChatGPT still, but among people who have, they've probably heard stories about AI in science fiction. My definition of AI is any advanced computer processing, generative or otherwise, that's happened since we got enough computing power and RAM to do something about it, aka lately.
      • IanCal 3 hours ago
        Then that definition is at odds with how the field has used it for many decades.

        You can have your own definition of words but it makes it harder to communicate.

      • southernplaces7 1 hour ago
        >Most people haven't even heard of OpenAI and ChatGPT still

        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.

    • andai 6 hours ago
      That's how I feel about Web 3.0...
      • buu700 6 hours ago
        Web 3(.0) always makes me think of the time around 14 years ago when Mark Zuckerberg publicly lightly roasted my room mate for asking for his predictions on Web 4.0 and 5.0.
    • pharrington 5 hours ago
      I'll bet that almost everyone who reads Quanta Magazine knows what they mean by AI.
    • zeofig 5 hours ago
      Absolutely agree.
  • LeroyRaz 29 minutes ago
    The article is misleading and badly written. None of the mentioned works seem to have used language or knowledge based models.

    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.

  • markasoftware 8 hours ago
    not an LLM, in case you're wondering. From the PyTheus paper:

    > 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

  • WhyNotHugo 1 hour ago
    Article mentions that if students present these designs, they’d be dismissed as ridiculously. But when AI present them, they’re taken seriously.

    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.

    • wongarsu 1 hour ago
      Across basically all fields you have to first show that you can think inside the box before you are allowed to bring out-of-the-box ideas. Once you have shown that you mastered the craft and understood the rules you can get creative, but before that creativity is rarely valued. Doesn't matter if you are an academic or an artist, the same rules apply

      I'm guessing AI gets the benefit of the doubt here because its ideas will be interesting and publishable no matter the outcome

      • mandmandam 50 minutes ago
        You can do all the 'proving your chops' and in-the-box thinking in the world and still get ostracized for your creative insight.

        * 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.

    • Arkhaine_kupo 1 hour ago
      Its a cost risk analysis. We have tried letting studdents do whatever and most of the time it went nowhere, so we ended up with a more rational system (with many caveats) where experiments are proposed and people with good insights and sense of whether it might even work approve it before running it.

      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.

    • gewbie 1 hour ago
      [dead]
  • topspin 5 hours ago
    "It added an additional three-kilometer-long ring between the main interferometer and the detector to circulate the light before it exited the interferometer’s arms."

    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.

    • heisenbit 3 hours ago
      Sounds like ring lasers. Not really an unusual concept to increase sensitivity.
  • viraptor 7 hours ago
    This sounds similar to evolved antennas https://en.wikipedia.org/wiki/Evolved_antenna

    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.

  • matt3210 3 hours ago
    The "AI" here is not the same "AI" as claude, Grok or OpenAI. It's just an optimization algorithm that tries different things in parallel until it finds a better solution to inform the next round.
    • jeroenhd 1 hour ago
      > It's just an optimization algorithm that tries different things in parallel until it finds a better solution to inform the next round.

      ... 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.

      • LeroyRaz 25 minutes ago
        The standard term is Machine Learning (ML). It's not artificial intelligence (AI) because there is no sense of attempting to manipulate concepts (e.g., the classic symbolic conceptual manipulation algorithms, or modern foundation models).
  • codeaether 55 minutes ago
    These days, it feels like “AI” basically just means neural network-based models—especially large autoregressive ones. Even convolutional neural networks probably don’t count as “real AI” anymore in most people’s eyes. Funny how things change. Not long ago, search algorithms like A* were considered the cutting edge of AI.
  • k__ 1 hour ago
    "it had no sense of symmetry, beauty, anything. It was just a mess."

    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!

  • anonym00se1 8 hours ago
    Feels like we're going to see a lot of headlines like this in the future.

    "AI comes up with bizarre ___________________, but it works!"

    • viraptor 7 hours ago
      We've seen this for a while, just not as often: antennas, IC, FPGA design, small mechanical things, ...
    • sandspar 3 hours ago
      "AI comes up with a bizarre short-form generative video genre that addicts user in seconds - but it works!" I'm guessing we're only a year or two away.
      • eleveriven 3 hours ago
        Entering the "hold my beer" era of AI creativity
    • amelius 3 hours ago
      ... sometimes.
    • ninetyninenine 7 hours ago
      That’s how we become numb to the progress. Like think of this in the context of a decade ago. The news would’ve been amazing.

      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.

      • dns_snek 2 hours ago
        You're taking intelligently designed specialized optimization algorithms like the one in this article and trying to use their credibility and success to further inflate the hype of general-purpose LLMs that had nothing to do with this discovery.
      • hammyhavoc 7 hours ago
        Twenty bucks says it isn't.
      • somenameforme 6 hours ago
        A decade ago it wouldn't have been called AI, and it probably shouldn't be called AI today because it's absurdly misleading. It's a python program that "uses gradient descent combined with topological optimization to find minimal graphs corresponding to some target quantum experiment".

        Of course today call something "AI" and suddenly interest, and presumably grant opportunities, increase by a few orders of magnitude.

        • JimDabell 5 hours ago
          That’s been called AI for about thirty years as far as I am aware. I’m pretty sure I first ran into it studying AI at uni in the 90s, reading Norvig’s Artificial Intelligence: A Modern Approach. This is just the AI Effect at work.

          https://en.wikipedia.org/wiki/AI_effect

        • ninetyninenine 6 hours ago
          Gradient descent is a learning algorithm. This is AI.
          • omnicognate 2 hours ago
            Gradient descent is used in machine learning, which is a field in AI, to train models (eg. neural networks) on data. You get some data and use gradient descent to pick the parameters (eg. neural network weights) to minimise the error on that training data. You can then use your trained model by putting other data into it and getting its outputs.

            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.

          • somenameforme 5 hours ago
            Hahah, if you're going to go that route you may as well call all of math "AI", which is probably where we're headed anyhow! Gradient descent is used in training LLM systems, but it's no more "AI" itself than e.g. a quadratic regression is.
            • ordu 4 hours ago
              Neural networks are on the hype now, but it doesn't mean that there was no AI before them. It was, it struggled to solve some problems, and to some of them it found solutions. Today people tend to reject everything that is not neural net as not "AI". If it is not neural net, then it is not AI, but general CS. However AI research generated a ton of algorithms for searching, and while gradient descent (I think) was not invented as a part of AI research, AI research adapted the idea to discrete spaces in multiple ways.

              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".

  • aeternum 7 hours ago
    More hype than substance unfortunately.

    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.

    • irjustin 6 hours ago
      Ehhhhh, I'll say it's substantive and not just pure hype.

      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.

      • omnicognate 3 hours ago
        AFAICT the "AI" didn't "pay attention to the work" either. They built a representation of a set of possible experiments, defined an objective function quantifying what they wanted to optimise and used gradient descent to find the best experiment according to that objective function.

        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.

    • rlt 7 hours ago
      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 4 hours ago
        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".
      • hammyhavoc 7 hours ago
        This is monkeys and typewriters.

        It's like seeing things in clouds or tea leaves.

        • Supermancho 7 hours ago
          If the "monkeys with typewriters" produces a Shakespear sonnet faster than he is reincarnated, it's a useful resource.

          At least, that's the thinking.

          • coliveira 5 hours ago
            It is 100% impossible for AIs to create a Shakespeare sonnet. They can create a pastiche of a sonnet, which is completely different.
          • andsoitis 4 hours ago
            It can’t be a Shakespeare sonnet if Shakespeare didn’t write it
            • vrighter 50 minutes ago
              Yes. Just like any pair of glasses the sun decides to wear would be sunglasses.
            • inglor_cz 2 hours ago
              Your exchange has made me wonder. Yes, whatever AI produces is not genuine stuff. But there is something we could call "Shakespeare-ness", and maybe it is quantifiable.

              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.

              • hammyhavoc 36 minutes ago
                • inglor_cz 22 minutes ago
                  Extra Terrible Lines are indeed fun. We've had 9 months of development since then, though; maybe it would make sense to repeat those experiments twice a year.

                  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.

          • shermantanktop 6 hours ago
            That’s the looooong game on both counts.
            • tux1968 5 hours ago
              'tis the patient plot on either side, where time doth weave its cunning, deep and wide.
      • coliveira 5 hours ago
        AI companies stole massive amounts of information from every book they could get. Do you really believe there's any research they don't have input into their training sets?
  • eleveriven 3 hours ago
    Feels like we're entering a new kind of scientific method. Not sure if that's thrilling or terrifying, but definitely fascinating
  • smj-edison 4 hours ago
    Am I understanding the article correctly that the created a quantum playground, and then set thein algorithm to work optimizing the design within the playgrounds' constranits? That's pretty cool, especially for doing graph optimization. I'd be curious to know how compute intensive it was.
  • Huxley1 3 hours ago
    This AI-designed experiment is pretty cool. It seemed kind of weird at first, but since it actually works, it’s worth paying attention to. AI feels more like a powerful tool that helps us think outside the box and come up with fresh ideas. Is AI more of a helper or a creator when it comes to research?
    • theteapot 2 hours ago
      AFAICT "The AI" (which is never actually described in the article) is a CSOP solver.
      • omnicognate 2 hours ago
        They should stop optimizing their Company Share Option Plans and get back to work!

        (It was a gradient descent optimizer, so probably unconstrained optimization rather than a Constraint Satisfaction Optimization Problem, but it might have had constraints.)

  • kristjank 3 hours ago
    Impressive results, I remember reading about AI-generated microstrip RF filters not too long ago, and someone already mentioned evolved antenna systems. We are suffering from a severe case of calling gradient descent AI at the moment, but if it gets more money into actual research instead of LLM slop, I'm all for it.
    • IanCal 2 hours ago
      > We are suffering from a severe case of calling gradient descent AI at the moment,

      We’ve been doing that for decades, it’s just more recently that it’s come with so much more funding.

    • carabiner 2 hours ago
      I still call computers "adding machines." Total fad devices.
  • qz_kb 6 hours ago
    This is not "AI", it's non-linear optimization...
    • tomrod 5 hours ago
      We all do math down here.
    • rurban 2 hours ago
      Non-linear optimization is classic AI, ie. searching through logic and symbolic computation.

      "Modern" AI is just fuzzy logic, connecting massive probabilities to find patterns.

  • IAmGraydon 8 hours ago
    This is the kind of thing I like to see AI being used for. That said, as is noted in the article, this has not yet led to new physics or any indication of new physics.