in Sociology 619 Words

A Nobel prize in physics for something that is not physics

When I wrote about Nobel prizes a little while back, I did not expect to return to the subject. I assumed the prize this year would be awarded for some meritorious advance in laboratory physics, like last year’s prize “for experimental methods that generate attosecond pulses of light for the study of electron dynamics in matter.” Instead, we find that the 2024 prize has been awarded to John Hopfield and Geoffrey Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.” This is the Nobel prize in physics we’re talking about.

One small issue: that’s not physics.

I’ve been concerned for a long time with the interface between astronomy and physics – where they are distinct fields and where they overlap. One of the reasons I left physics as a grad student was because the string theorists were taking over. They were talking about phenomena that were tens of orders of magnitude beyond any conceivable experimental test. That sort of theoretical speculation is often fun, sometimes important, and very rarely relevant to physical reality. Lacking exposure to experimental tests or observational consequences, to my mind it was just that: speculation, not physics.

Nearly forty years on, my concerns about string theory have not been misplaced. And while, in the strictest sense, I don’t think it qualifies as physics – it’s more of a physics-adjacent branch of mathematics – it is at least attempting to be physical theory. But machine learning is not physics. It’s computer science. Computers are a useful tool, to be sure. But programming them is no more physics than teaching a horse to count.

I’m not sure we should even consider machine learning to be meritorious. It can be useful, but it is also a gateway drug to artificial intelligence (AI). I remember the more earnest proponents of early AI propounding on the virtues of LISP and how it would bring us AI – in the 1980s. All it brought us then was dystopian fantasies about killer robots nuking the world. Despite the current hype, we have not now developed intelligent machines – what we’re calling AI is certainly artificial but not at all intelligent. It uses machine “learning” to reprocess existing information into repackaged forms. There is zero original thought, nothing resembling intelligence. Modern AI is, in essence, a bullshit generator. Now, we can all think of people who qualify as organic bullshit generators, but that begs the question:

Why is the Nobel prize in physics being awarded for something that is clearly not physics?

Maybe it has something to do with the hype around AI. I don’t know what the decision process was, but I do know that I am not the only scientist to have this reaction.

Myself, I’m not mad, just disappointed. I’m not unique in feeling that physics has lost its way. This just emphasizes how far it has strayed.

Apparently the Nobel committee is sensing the blow-back, as this poll currently appears on the award page:

I… don’t think this helps their case. Did you know that molecules are made of atoms? Ergo all of chemistry is just applied atomic physics. I mean, it is a long-standing trope that physicists think every other science is just a lesser, applied form of physics. At the level of being based on the equations of physics, that’s almost kinda if not really true. So asserting that machine learning models are based on physics equations comes nowhere near to making machine learning into physics. It’s fancy programming, not physics.

Well, there will be complaints about this one for a while, so I won’t pile on more. I guess if you give out 118 prizes since 1901, one of them has to rank 118th.

42 thoughts on “A Nobel prize in physics for something that is not physics”

  1. Hi Stacy,

    Thank you for this post on the 2024 Nobel prize for Physics. As you mentioned previously, the wording for the Nobel prize in physics includes “the most important discovery or invention within the field of physics”. I take it that “discovery” relates to observation and “invention” relates to equipment & theory. The 1927 Nobel prize went to Charles Wilson for inventing the cloud chamber. So I think a good case can be made for awarding the prize for complementary work. To me it is interesting that the Nobel committee did not find any physicists more worthy than Hopfield and Hinton.

    Separately: despite some previous comments, most of us really appreciate the time & care you put into writing this blog.

    1. I’m a big advocate of the importance of software tools, for example, in the operation of large instruments on modern telescopes. It’s almost malpractice at this point to build a big, fancy imaging camera or spectrograph without a software pipeline to handle the output. (The attitude for a long time was that observers would figure that part out on their own.) Even so, I think it is a stretch to call machine learning an invention (the computers they run on certainly are!) but it is even more of a stretch to call it physics. It isn’t really in the same category of a physical object like a cloud chamber that is experimentally useful.

  3. The Nobel Prize for Chemistry was also given to AI researchers this year, so this isn’t just restricted to physics.

  4. In 1994, the Nobel Prize for Economics had been awarded to a mathematician,

    https://web.archive.org/web/20090630015308/http://www.samuelbrittan.co.uk/text172_p.html

    I do happen to strongly agree with your sentiment.

    Sadly, “publish or perish” leads people to re-invent their subjects with popular vocabulary, and, the subject in which I have personal interest is deeply infected with the contagion of information technology. A recently posted discussion of set theory at the n-category cafe includes:

    “The data to which the axioms apply…”

    in its explanation. I own approximately 400 mathematics texts. The sentence fragment above is not a standard vocabulary. It is symptomatic of confusing constructive Herbrand semantics with non-constructve first-order semantics. It is symptomic of a generation trying to redefine “mathematics” for sociological reasons.

    Meanwhile the FOM mailing list has devolved into little more than announcements of seminars and “cool stuff” being done with proof assistants (or Wolfram’s new science).

    Dependence on computers transforms people. You may expect to start hearing about how physics has become too difficult for physicists. A Fields medalist has already declared that mathematics is too hard for mathematicians.

  5. This year’s Nobel in Chemistry was described to me by colleagues who work on proteins as “not a matter of if but when” – they view the changes brought by the work of AlphaFold as truly transformative for biological chemistry. And the subject of the award is very much chemistry – the “protein folding problem” has long been viewed as a grand challenge.
    I haven’t heard anyone from the Physics community share similar views about the Physics prize this year.

    1. I am not qualified to have an opinion on the chemistry prize. Certainly protein folding is important and challenging, so this could well be a worthy advance.

      I’ve certainly heard plenty of complaints about the physics prize, but I’ve also heard it defended. So far, these boil down to “the committee wanted to recognize Hinton” which sounds like an argument that there is a star chamber that gives awards to the in-crowd, and being recognized as a member of the in-crowd suffices with the stated reason for the award being secondary.

      1. If I was going to be pedantic I would have said that it should have won the Nobel Prize for Medicine or Physiology on the grounds that protein-folding is principally important for living creatures as it is the basis of enzymes.

  6. Hinton received a Turing Award in 2018 for neural nets, commonly referred to as the “Nobel Prize for Computer Science”. Hinton joined U of T in 1987, a year after I graduated from Computer Science, otherwise I probably would have had him as a prof. Not sure why he needs (or should have) both a Turing and a Nobel for essentially the same thing. I believe I read he was shocked as well.

    I think this means that the Nobel has officially jumped the shark. I personally believe the whole AI space has been grossly over promoted, and I still don’t believe machine learning is intelligence, although to be fair, I’m also not sure what “intelligence” is. I still can’t help thinking of current AI as “Cat, not a cat”. It is incredibly powerful and useful, particularly in pattern recognition, but mainly as a tool for other pursuits, not something unto itself. Essentially what I believe about the role of the computer as a whole.

    The “gold standard” test of AI used to be the Turing Test; whether a human found human and computer response indistinguishable. Chat-GPT recently passed the Turing Test, the first “AI” to do this, but I think this may say more about the current gullibility and lack of curiosity of humans rather than representing a true singularity.

  7. Machine learning are statistical models; they have nothing to do with intelligence except that they can associatively recognize patterns. And I agree, it has little to do with physics.

    I do think they can be really useful, however their output is completely dependent of their training input. My definition of intelligence would be that it can output really original surprising ideas and word those in an intelligent fashion. Now it can only give “original” content where pattern matching and language processing matter a lot.

  8. Well, Going back Dalen got one for work on the lighthouse and Marconi for radio. But I mostly agree. I wanted to respond to your last ‘Nobel’ post and nominate Bob Dicke as person perhaps most deserving that never won one.

    1. Yes, Dicke certainly should have won. I think the well-deserved prize to Peebles belatedly acknowledges this, without explicitly doing so since it was too late, and then we have to already know Dicke should have been in on it in order for us to even notice.

  9. Machine learning is certainly not physics; it does bear some resemblance to mathematics. I think of it as automated trial and error and interpolation; and I personally think those processes are the basis of what we call intelligence. It is at the mercy of its training data, but aren’t we all? In cases where its training data is complete, ie.g., the rules of Go and Chess, it outperforms humans. Treatises have been written on Go for at least 2400 years. In AlphaGo’s tournament against the world champion, all the watching experts were amazed by one of AlphaGo’s moves, and the world champion later said, “I never thought a machine could beat me, but when I saw that move, I knew I would lose the tournament.”

  10. “One of the reasons I left physics as a grad student was because the string theorists were taking over. ”

    in the 80s-00s and even today were string theorists guaranteed to get hired and tenure faculty member by top physics departments like Harvard University and Princeton universities ?

    is astronomy phd easier to get a hired by universities ?

    1. No, not at all. In the ’80s, the joke was that string theorists were training to be taxi drivers. While that stereotype didn’t persist, there is not now, nor has there ever been a time in my lifetime when it was easy to get hired as a faculty member anywhere. Quite the opposite: it is an extraordinarily hard road, lots of perfectly well qualified people don’t succeed, and many of us who do make tremendous sacrifices to do so, including myself: http://astroweb.case.edu/ssm/ronin.html

      1. @Dr. McGaugh

        Your observation that merit becomes meaningless in tight job markets struck a chord with me (not to diminish the stress placed upon your family during those years).

        Grothendieck spoke at CERN in 1972,

        “By dint of producing highly qualified people we’ve really produced too many since the great boom in the production of young scientists—since Sputnik some fifteen years ago—and there’s more and more unemployment in scientific careers. The problem is becoming increasingly acute for a growing number of young people, especially young scientists. In the USA we must produce something like 1000 or 1500 theses a year in mathematics alone and the number of job openings is about a third of that.”

        https://github.com/Lapin0t/grothendieck-cern

        1. Sadly, this is true, and has continued to be true. There have been a few ups & downs in the job market, but that’s the difference between “tight” and “insufferable.” It’s all demographics, compounded by the Sputnik splurge Grothendieck notes, which clogged up the job market with people of a certain age like a python swallowing an elephant.

      2. but I thought only the very smart students who maybe get phD from elite programs like Edward Witten at Princeton get hired

        surely studying under and a letter of recommendation from a professor like Edward Witten could open the door of jobs

        so what fields of between astronomy and physics – where graduate programs and students who has career aspirations to get hired as a faculty member ?

        1. First, there are no fields where this aspiration is easy. Some are easier than others, but which those are can change in the time it takes to do a PhD.

          Certainly it helps to have a letter from a Famous Person. That is a sociological effect; it needn’t correlate with actual achievement, and often doesn’t. From what I’ve seen, which is a lot, this is a good way to encourage brown-nosing to become part of the in-crowd, which strongly discourages people from following the scientific method. It is more important to stay on the good side of the perceived powers that be than to speak up for what is right.

  11. Since career aspirations have been brought into the conversation, note that the application of the scientific method to solving the problem of finding satisfying work is appropriate. Of course, what is ‘satisfying work’ will vary with each individual and will especially depend on the individual situation and frame of mind. Finding success in academia or beyond will always involve lots of strenuous effort. Nevertheless, assistive resources for looking beyond a person’s current narrow field are more available today than in the century past.

  12. have you seen

    arXiv:2410.02612 (astro-ph)
    [Submitted on 3 Oct 2024]
    On the nature of the missing mass of galaxy clusters in MOND: the view from gravitational lensing
    Benoit Famaey, Lorenzo Pizzuti, Ippocratis D. Saltas

    Modified Newtonian Dynamics (MOND) has long been known to fail in galaxy clusters, implying a residual missing mass problem for clusters in this context. …Clusters with lower observed gas mass display larger and more scattered values for both the density ratio and cut-off radius. These lensing results can in principle serve as a crucible for relativistic theories of MOND in galaxy clusters, or for any other tentative hypothesis regarding the nature of the clusters residual missing mass in the MOND context.

    1. That’s really cool. There is a missing baryon problem in the cores of rich clusters – this is true in LCDM as well as in MOND, we just don’t care in the former case. That there is some systematic in the scaling is intriguing. Maybe it is a clue about the parent theory of MOND or maybe it is a hint about the nature of the missing mass – a big ball of massive neutrinos might do the trick.

      1. a big ball of massive neutrinos might do the trick.

        aren’t massive neutrinos the same as sterile neutrinos and a type of dark matter

        are you suggesting the combination of MOND + sterile neutrinos +missing baryon in the cores of rich clusters to explain galaxy clusters

        1. I’m saying these are all possibilities, on their own or in combination. I don’t find any of these possibilities particularly appealing: clusters don’t make sense in any paradigm – https://tritonstation.com/2024/02/06/clusters-of-galaxies-ruin-everything/ – which is why Famaey’s paper is interesting: there’s a clue there that we haven’t yet used to inform our ideas.

          I’m not sure what you mean by “a type of dark matter.” Anything that we can’t easily see could be considered a type of dark matter – by that standard, there are lots of as-yet undetected baryons in the universe that could be considered by dark matter. But they’re not anything fundamentally new like WIMPs any more than rocks in caves that you can’t see because it is dark inside caves. Neutrinos are known particles that are dark in the sense that they don’t radiate, but they are also particles that are known to exist with some (as yet unknown) mass, so these are more like rocks in this context than they are to novel dark matter. Sterile neutrinos are related to normal neutrinos but also not quite the same beast, so would be something new. Then there is the infinite number of possibilities for completely novel dark matter.

          These things are meaningfully different. The term “dark matter” can encompass them all, but using it that way obscures the need for fundamentally new physics. It also gives the false impression that we already know “dark matter” exists when all we really know is that there are rocks and neutrinos that don’t radiate when these have exactly nothing to do with novel dark matter particles.

          1. to clarify

            are massive neutrinos standard model neutrinos that have more mass than ordinary neutrinos say from radioactive or a new type of neutrino?

            1. Standard Model neutrinos have a tiny but finite mass; this is irrelevant on all scales smaller than clusters, but might matter there depending on what their mass turns out to be.

              1. instead of massive neutrinos could more Standard Model neutrinos have a tiny but finite mass also work based on arithmetic multiplication, perhaps “slowed” by some force like gravity of black hole

              2. Would the inertial mass of neutrino’s be the cause, or should we use their relativistic mass?

                The amount of mass lost (that’s the max of captured nonluminous mass/energy) by all stars in a cluster integrated over the estimated period of time might become important on the scale of clusters.

              3. @Maarten
                Rest mass is a good first approximation. Just to double rest mass requires v = 0.866c and that’s far too fast to be bound in a galactic cluster. We would see some pretty big gravitational red-shifts from galaxies near the centre of clusters by the time the gravitational potential well of a cluster got deep enough to bind neutrinos from beta decay arising from stars.

                My assumption is that any neutrino contribution comes from thermonuclear reactions in the Big Bang. Those would be slowed by the expansion of the universe and might be slow enough to be captured by galactic clusters. We also know fairly well how many of them there are because there are two neutrinos per helium atom.

      2. A big ball of massive neutrinos is my favorite thing.
        (I’m in the middle of reading the paper.)

    1. I don’t find this phrasing too problematic, as it is just the way the field talks about it. That is to say, the underlying presumption of dark matter is so widespread that it is just how we talk about observed discrepancies.

  14. > It’s fancy programming, not physics.

    It’s funny you say this. For me, as a programmer, it’s a fancy statistical physics, not programming. Essentially, neural nets are not programmed, they are evolved.

    1. Neural nets are trained on examples. That’s not the same as biological evolution. In either case, I don’t see how it counts as either a discovery or an invention.

  15. I don’t think we humans are that intelligent really.
    We can communicate and work together very well. But the typical individual human being is not really intelligent.
    If you like: See also the video from at least minute 1:50.

  16. Two of my best colleagues are protein folders, one experimental and one theoretical. I dabbled in the experimental part with some success. Its clear all of us that if we had to bet on the field in which AI makes its first really big success, protein folding will be it.

    Chemistry of course is just physics. But its not just particle (quantum) physics.
    Its also statistical thermodynamics and to some extent statistical mechanics and in the real world, kinetics..
    And there really is now lots and lots of training data. In protein folding the thermodynamics and kinetics rule, but thermodynamics is ruled by statistics.

  19. One of the reasons I left physics as a grad student was because the string theorists were taking over. They were talking about phenomena that were tens of orders of magnitude beyond any conceivable experimental test.

    in the 80s to 90s and 00’s string theorists Greene and Kaku et al – They were talking about supersymmetry at the LHC …

    and supersymmetry at a future 100 tev collider

    1. Supersymmetry is not the same as string theory, though string theories can contain supersymmetry. So what you’re referring to is not the same as what I was referring to.

      The achievable, golden test of supersymmetry was the decay of the B_s meson. It flunked.

      The LHC found no evidence for SUSY and there is no reason to expect a 100 TeV collider would do so. The LHC did find the expected Higgs particle, but it is “too normal” (i.e., consistent with the Standard Model) to provide a clear window to any new physics.

  20. @Augustin

    Well, setting quips and appeals to authority aside, computer science is precisely about computers.

    Prior to the investigations on effective computability leading to the Church-Turing thesis, computing had been a profession in which people brought home paychecks to feed their families and keep roofs over their heads. Turing’s landmark work had been based upon how human computers manipulated (finitely many) syntactic symbols on a (metrically bounded) sheet of paper. Significantly, the context of calculation is rule based.

    The significance of the rule-based manipulation of symbols had been previously analyzed by Augustus de Morgan. The contemporary account of mathematics being “formal” (contrary to what de Morgan actually concluded) in terms of language signatures generating manipulable expressions arises from de Morgan’s work.

    The relationship between symbolic methods and non-constructive mathematics is introduced as information theory by Claude Shannon. Shannon had been fully aware of symbolic methods, but, chose to use logarithms “for engineering purposes.” Apparently, if Warren Weaver is to be believed, Claude Shannon had been enamored with the work of Boltzmann, whence the analogy with entropy is introduced into the subject.

    In turn, the introduction of non-constructive mathematics admits a topological analysis of symbolic truth tables in terms of convex hulls in the sense of combinatorial topology. You can find the mathematics of this without Turing’s delusions of machine intelligence or Weiner’s analogies with biological systems in Hu’s book, “Threshold Logic,”

    https://www.amazon.com/Threshold-Logic-Sze-Tsen-Hu/dp/0520369084

    Dr. McGaugh’s blog is motivated, in part, by the frustration of answering to the sociology of academia as infused with logical fallacy.

    Computer science is founded upon the imitation of human beings manipulating symbols on pieces of paper.

    That only changes when one permits history to be replaced with Orwellian Newspeak.

Comments are closed.