The research that gets done in any field of science is heavily influenced by the priorities set by those who fund the research. For science in the US in general, and the field of theoretical physics in particular, recent years have seen a reordering of priorities that is becoming ever more pronounced. As a prominent example, recently the NSF announced that their graduate student fellowships (a program that funds a large number of graduate students in all areas of science and mathematics) will now be governed by the following language:
Although NSF will continue to fund outstanding Graduate Research Fellowships in all areas of science and engineering supported by NSF, in FY2021, GRFP will emphasize three high priority research areas in alignment with NSF goals. These areas are Artificial Intelligence, Quantum Information Science, and Computationally Intensive Research. Applications are encouraged in all disciplines supported by NSF that incorporate these high priority research areas.
No one seems to know exactly what this means in practice, but it clearly means that if you want the best chance of getting a good start on a career in science, you really should be going into one of
- Artificial Intelligence
- Quantum Information Science
- Computationally Intensive Research
or, even better, trying to work on some intersection of these topics.
Emphasis on these areas is not new; it has been growing significantly in recent years, but this policy change by the NSF should accelerate ongoing changes. As far as fundamental theoretical physics goes, we’ve already seen that the move to quantum information science has had a significant effect. For example, the IAS PiTP summer program that trains students in the latest hot topics in 2018 was devoted to From Qubits to Spacetime. The impact of this change in funding priorities is increased by the fact that the largest source of private funding for theoretical physics research, the Simons Foundation, shares much the same emphasis. The new Simons-funded Flatiron Institute here in New York has as mission statement
The mission of the Flatiron Institute is to advance scientific research through computational methods, including data analysis, theory, modeling and simulation.
In the latest development on this front, the White House announced today \$1 billion in funding for artificial intelligence and quantum information science research institutes:
“Thanks to the leadership of President Trump, the United States is accomplishing yet another milestone in our efforts to strengthen research in AI and quantum. We are proud to announce that over $1 billion in funding will be geared towards that research, a defining achievement as we continue to shape and prepare this great Nation for excellence in the industries of the future,” said Advisor to the President Ivanka Trump.
This includes an NSF component of \$100 million dollars in new funding for five Artificial Intelligence research institutes. One of these will largely be a fundamental theoretical physics institute, to be called the NSF AI Institute for Artificial Intelligence and Fundamental Interactions (IAIFI). The theory topics the institute will concentrate on will be
- Accelerating Lattice Field Theory with AI
- Exploring the Multiverse with AI
- Classifying Knots with AI
- Astrophysical Simulations with AI
- Towards an AI Physicist
- String Theory Conjectures via AI
As far as trying to get beyond the Standard Model, the IAIFI plan is to
work to understand physics beyond the SM in the frameworks of string and knot theory.
I’m rather mystified by how knot theory is going to give us beyond the SM physics, perhaps the plan is to revive Lord Kelvin’s vortex theory.
Update: Some more here about the knots. No question that you can study knots with a computer, but I’m still mystified by their supposed connection to beyond SM physics.
Looks like Chris Anderson was right when he proclaimed the end of theory.
(There’s a “b” missing in the last sentence.)
Thanks Sabine,
Typo fixed. The Chris Anderson argument she refers to is here
https://www.wired.com/2008/06/pb-theory/
How about
• NSF run by AI
Reply to Chris Anderson – https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2711825/
It is absolutely fascinating, from a Frankenstein monster that will not die perspective, that the multiverse and string theory live on through AI.
I think what this probably really means is that the latest AI bubble is about to burst (or has burst). Certainly an AI winter is coming.
I did not think I’d live to see the day that the direction of physics research was announced by Ivanka Trump.
This sounds like an Onion article. Let’s just have everything run by AI shall we? What about AI run by AI? Then we can all just stay in bed in the morning and avoid all this bullshit.
I suspect a large part of the “White House announcement” is congressionally mandated, in spite of the boilerplate, “Leadership of President Trump” bombast, in part because that’s where the money has to come from. (And way down at the bottom of the announcement is an acknowledgment that “The National Quantum Initiative Act, bipartisan legislation signed by President Trump in 2018, called for the creation of research centers nationwide to accelerate foundational QIS research and development.”)
@Warren Siegel. How do you know that NSF is not run by AI today?
Artificial intelligence is better than none.
David Brahm/Peter Erwin,
While Ivanka Trump was doing the announcing, the decisions that this is the way science needs to go are getting made elsewhere, often with the participation of the physics community. What I wonder more is what sort, if any, of peer review signed off on the multiverse AI business.
All this reminded me that I’ve written about some of this before, see
https://www.math.columbia.edu/~woit/wordpress/?p=10680
I think we are underestimating the imminent role that AI is about to play in expanding our knowledge of the Universe. I’m with those ‘tech geeks’ who see the future discoveries of physics not coming out of Princeton, Harvard or Cambridge but rather out of Apple, Google and their Chinese equivalents. AI will be an existential tool to expand knowledge, but Chris Anderson’s view is too radical. In about twenty or twentyfive years, 200 million teens in their bedrooms are each going to have access to more computing power and tools than the top physics institutions have today.
Quantum computers and AI are ALREADY the future of cosmology/string theory/multiverse/knot theory… whatever field you prefer.
Theoretical physics, as has been done/studied up till now… is at the verge of a massive paradigm shift. I know there are a couple of radicals out there (some are string theorists pretty well known in this blog, especially one of them from Central Europe…) who still believe that men alone will find that final TOE (which they consider to be ST itself!), but… sorry for them. It’s not going to happen as they expect. Not a snowball’s chance in hell!!
The field is screwed without the help of that powerful technology. And contrary to what some top theoretical physicists go about saying that it’s a great time to become a theoretical physics student, well… bollocks.
The current approach to AI is highly data-dependent and statistical in nature. Causal understanding has not been introduced yet. And most active Machine Learning research is not primarily geared towards that.
Given that, as physicists do you think there is even a possibility that theoretical physics topics have any chance of benefitting from AI (without it having causal understanding) in the near future? Are there any planned physics experiments to which AI would add a significant value?
(I am from computer science and engineering background.)
DB,
It’s not just certain string theorists who believe human beings can come up with a successful unified theory, since I happen to also believe this, and would have thought that if anyone’s not a string theorist, it’s me. I might be wrong though about what I am, since I just saw in Katie Mack’s new book the following definition of string theory:
“String theory is a blanket term for theories that try to bring together gravity and particle physics in new ways” and that definition fits what I work on.
Apple and Google might be sources for certain kinds of physics discoveries, but I’d bet these will always be in certain restricted subfields (e.g. those relevant to building quantum computers). Those companies are pretty tightly focused on making money and achieving world domination, not likely to devote much effort to other fields with little practical, exploitable, applications.
The idea that increase in computer power and new computer algorithms will revolutionize physics has been around for at least the last 50 years I’ve been following the subject. Actual progress of this kind has always been much more limited than expected. When I was a grad student in the early 1980s I spent a lot of time doing computer calculations (lattice gauge theory Monte-Carlo simulations). That field has had interesting results, but not revolutionized the subject at all. One of the most likely possible applications of quantum computers would be in that area, but I suspect they will always be rather specialized, not affect most people’s work.
On the experimental side of physics, I’m quite willing to believe that AI will find some very useful applications, from what I hear it’s already being seriously exploited in LHC data analysis, with more to come. There’s no evidence though that this sort of thing will lead to breakthroughs in fundamental experimental issues, for instance helping overcome the barriers to getting information about what’s happening at energies above the TeV scale.
It’s not hard to believe that a large fraction of the stupid things people do can be done just as stupidly, but faster, by artificial intelligence. Given the depressing state of the world, this might mean that most of what human beings do can be done by machines. On the issue of finding a better unified theory, all evidence I’ve seen in that the machines will do a great job of calculating away on bad, failed ideas, none that they’ll help come up with better ones.
Yash Sharma,
The main problem with beyond the SM HEP theory is that there is essentially no relevant data at all (maybe, making a lot of assumptions about astrophysics, a couple numbers about “dark matter”). So, it seems no role for “Big Data”- based approaches.
Back before the LHC turned on, a bunch of people organized an “LHC Olympics” designed to get algorithms ready to sort through and figure out the correct beyond the SM model based on the large number of new (non-SM) physics signals they expected. It always seemed more likely that things would work exactly as they had always worked in the past: intensive effort devoted to finding even one new physics signal. That’s where the field is now: experimentalists will use all the help they can get to find such a signal, but for theorists whether there’s no signal or one signal, Big Data is not what they will ever have.
Thanks, Peter.
About the other discipline set to receive great funding, Quantum Information Science, it seems much more likely to help theoretical physics, guessing from this nice article – https://www.quantamagazine.org/john-preskill-quantum-computing-may-help-us-study-quantum-gravity-20200715/ . Thats about quantum gravity. Do you see any other ways in which quantum computing might help (apart from lattice gauge theory that you mentioned). Specifically, are there any experiment or simulations that quantum physicists cannot perform in labs because they are infeasible or too expensive, but with a scalable quantum computer they might be able to do it?
Yash Sharma,
Theorists have been promoting the idea that quantum information science will give us a theory of quantum gravity for a long time now, with little solid to show for this, and I’m extremely dubious that that will ever change. The only substantive reference in that Preskill article, was to the article discussed here
https://www.math.columbia.edu/~woit/wordpress/?p=11648
While I thought that Preskill article was disappointing, the best answer I know of to your question is a much better article he wrote a while back, see
https://arxiv.org/abs/1811.10085
As a researcher in AI I am quite surprised to see such cranky topics being given respectability.
There is a lot of serious work on bringing optimisation, search and automated theorem proving (all of them techniques employed in research in AI) to help with experimental design, finding proofs and programming quantum computers.
Seeing a topic “exploring the multiverse with AI” – that sounds like something out of a work of fiction such as Neal Stephenson’s “Anathema” – breaks my heart.
Hi Peter and others,
Regarding the AI in theoretical physics, the only potentially useful application I can imagine is the following — unlike any human, the AI could “read” and sift through all scientific papers on the arXiv, with a possible outcome of finding one or two obscure papers which contain some crucial piece of result for some open problem, written by some people not very well known by the research community.
Namely, a human simply cannot read *all* papers, and most of the researchers only pay attention to papers authored by the people they are familiar with. So some crucial insight, put on the arXiv, can go completely unnoticed for a long time, if it was authored by a researcher who has no social contact with leaders in the field. Simply, nobody bothered to read it. On the other hand, in principle if properly trained, an AI could analyze all the arXiv papers, and turn everyone’s attention to that one inconspicuous little paper written by those anonymous couple of guys from that obscure University of Whereever — which actually solves something important, but nobody paid any attention to it because its authors are not recognized as relevant by the main community of researchers.
I think the above could be a useful application of AI in theoretical physics, and IMO this is the best case scenario — I am in fact very skeptical that AI could do even that (perform a semantic analysis of papers on the arXiv), since semantic analysis usually requires actual understanding of the underlying problem, paths to possible solutions, etc. — all that stuff that we do in science that AI usually *cannot* do.
Other than that, I don’t see AI being helpful in theoretical physics in any way whatsoever, except maybe in some highly specific usecases, like sifting through petabytes of LHC data looking for a non-SM signal. But do note that the latter is just a more efficient way of performing a very dull, boring and non-intelligent work what humans (or even monkeys) could eventually do even without AI, given enough time. A machine just does it faster.
Also, I am baffled that some people have an impression that AI can be somehow “smarter” or “more inteligent” than humans (whatever the definitions of those words may be). On the contrary, AI is just an efficient lossy compression algorithm (think jpeg or mp3). It is a mathematical function with a huge number of tunable parameters, and once tuned to give desired output to certain known inputs (process called “training” the AI), one can use it to approximate outputs for unknown inputs, hoping that the resulting outputs would not be “too far off” from the outputs one would expect to get. That is why mp3 playback sounds rather similar to the original song, and why google seems to be able to “predict” what will be the next word you type into its search box. There is nothing “intelligent” about it, it’s just pure (and very stupid) large scale data-fitting.
All this stuff (that makes up the AI) has absolutely nothing whatsoever to do with creative thinking, which is what science is all about, IMO.
Best, 🙂
Marko
Using AI to perform analysis on arxiv papers is possible now. Hopefully we will soon know if this approach works, as they have already released arXiv papers dataset for a large ML community. https://blogs.cornell.edu/arxiv/2020/08/05/leveraging-machine-learning-to-fuel-new-discoveries-with-the-arxiv-dataset/
Understanding intelligence will be an interesting science, important for theoretical physics because it’s plausible that IQ will be scalable. For the moment, “*AI” (adding AI to whatever theorists were doing) is not going to be a breakthrough. But it’s nice that some physicists become apolitical when Ivanka throws money.
Yash Sharma,
Yes, arXiv has opened its database for ML purposes. But training the AI to do something useful with it is a completely different story, and a much harder problem to solve, IMO.
Best, 🙂
Marko
You all have already been fooled. The NSF announcement was written by an IA. (That is why “strings” and “knots” are in the same sentence!) Penrose predicted this. The Emperor has no mind.
“Exploring the Multiverse with AI” – helping sift all that data?
Peter,
there are good reasons not to model particles as knots. One is that in the weak interaction, particles can change (electrons to neutrinos, for example) and this change cannot be described with knots, which are stable structures. Another is that knots do not yield spin 1/2. (Even stronger reasons arise in string theory: in 10/11/12/26 dimensions, knots of one-dimensional objects do not exist at all: every knot is equivalent to the unknot. Maybe the AI system will rediscover this result …) As a consequence, modelling particles as knots has never been successful, despite about a dozen attempts in the literature – starting with Kelvin, as you mentioned.
Peter,
thanks for your detailed reply above.
Preskill’s arxiv paper and the link that you include in your update are on the same track of the idea I’m trying to convey.
No matter what string theorists or non string theorists say and proclaim all over the world, there’s not a f… chance for us humans to discover that final TOE… if it exists.
We will need the help of quantum computers/advanced AI, and, slowly but surely, they will be taking over that task from us.
Does that mean that we will disappear and have no role at all?
No, it just means that evolution will carry on its due course and our role will be diminishing. Of course, it won’t happen overnight. We will probably have a big role to play at least during this century. After that… I’m not betting my money.
Re the world’s situation… I can’t agree with you more. The disaster is absolute. And it doesn’t look better for the near future at least. It just sends massive shudders down my spine to think that in a decade (or less!) we could be in the hands of people like Ivanka or Jared… Not to mention the totalitarian oligarchs that already run many other parts of the world.
Finally, the reason why I don’t believe we will be able to find that TOE, is because, IMO, “Ultimate Reality” is ineffable. It can’t be expressed with words because it is diffused, almost imperceptible and lacks precision. It seems to me that the material Universe comes from the Immaterial. And that sounds like jumping over the physical realm into the META-physical, which I consider totally out of reach.
Oh, and I’m not referring to a god as religious people understand. I mean something so abstract that it can’t be perceived/understood, a noumenon type of idea/concept.
André,
I agree, and am not aware of any non-crackpot ideas to use knot theory to get a viable beyond SM model. That’s why I was surprised to see that there’s a new NSF-funded institute that advertises work on this.
DB,
I don’t see any point in arguing from anyone’s human intuitions about what fundamental physics should look like or not look like at the deepest level, including whether it should be comprehensible to us, or beyond our comprehension. Those who look at the problem and are sure that engaging with it would be like a dog trying to learn GR should do something else. Personally there seem to me several things about the SM and GR which I don’t understand, for which continuing to try and learn more about them may lead to something new, and I see possible ways forward that might work.
I’m in a good position to keep working on what I want to work on, unfortunately young people at this point are going to face overwhelming obstacles if they try and pursue this kind of fundamental theory. Instead they’re encouraged to go to work in an AI institute training them to develop more powerful tools to further the goals of our new overlords. Even if we are all just dogs, better to howl at the moon…
Peter,
the arguments against knots as particles are strong and hold water. But related topological structures, maybe made of open strings, might be possible – of course again only in 3d.
To me it is astonishing that a string theory institution includes a knot theory group. Knot theory is only possible in 3d. Knots and strings are simply incompatible. Or is there a kind of open string theory derivate in 3d?
André,
This isn’t a string theory institution, it’s an “AI and fundamental physics” institution. The theory of knots is very intricate and complicated, and you can easily generate lots of data, so it’s plausible that AI techniques might be useful. The problem is that this has no connection to fundamental physics, and I don’t see one even in the Northeastern PR material specifically about this.
If I had to guess, the supposed connection would be through some sort of extremely complicated constructions of possible “string vacua”, where you manage somehow to use knots as one piece of the construction. I’d guess though that such constructions are of even less interest than the usual complicated constructions, but that doesn’t mean they couldn’t play a role in attracting NSF funding.
As an extragalactic astronomer (optical, radio) I have some experience with AI and how it might help, perhaps even leading to learning something new about dark matter. Yes, it’s certainly “sexy”, but I think we’re still very much learning that AI GIGO is real, and that a lot of work has yet to be done before autonomous cars stop killing jaywalkers (to stretch an analogy). Before AI unleashed on SKA datasets starts telling us marvelous things about dark matter and dark energy, perhaps AI could help with identifying consistencies, inconsistencies, myths, and even reproducibility issues in astrophysics and cosmology?
Jean Tate,
Autonomous vehicles are easy. Autonomous vehicles on the same road with humans, on the other hand, is hard because humans are infinite fonts of novel behavior which simply cannot be anticipated. Because of that, we are facing a couple of decades of random slaughter and mayhem while we catalog and suppress most of the edge cases we can’t foresee. That may seem a steep price to pay but, to put it in perspective, it likely pales compared to the WMD-like prospect of aging baby boomers behind the wheel en masse over the same coming decades. Personally, I’ll take my chances with autonomous vehicles so we better get the AI cracking on that front…
But AI advancing theoretical physics? Autonomous vehicles will be accident-free before that comes to pass.
It seems to me much of this work would be useful in advancing AI research, but not advancing fundamental physics research. With particle physics, we have a large amount of data that can be explained quite efficiently with a relatively small number of abstract rules. Current AI/machine learning research is able to build predictive models from large amounts of data, but isn’t able to produce human comprehensible rules.
If you can understand how to build an AI able to go from raw LHC data to the Standard Model, that would be incredibly valuable from an AI perspective. If you apply current AI models to physics data you’d get a black box model which in the best case scenario would be able to reproduce the physics data, but provide little to no additional understanding.
The comment above by Joseph Healy is an outrage. People have to be killed randomly by driverless cars, until driverless car programmers have anticipated “every situation” that can arise? Why not put in a set of semi-catchall control conditions, so that for example if a driverless car is confused about the situation ahead, or it’s not clearly safe to proceed, and it’s not a fast busy road, then the driverless car will slow down and stop? Why not require prototype driverless cars to go through an appropriate very lengthy period of testing with dual controls, so that a human supervisor inside a prototype driverless car can take control immediately, if the car starts to behave in a dangerous manner?
Chris Austin,
I don’t want to host a discussion of driverless car technology or of the social significance of AI here, but will use this as an excuse to use my control of the blog to editorialize.
Most of the AI advances I hear about don’t seem to me to be doing anything to make the world a better place. In my city, there’s no good reason for most people to own a car. What we need is good public transportation and taxis. I don’t believe AI can anytime soon reliably drive a cab or bus around New York. Even if it could, I don’t see how eliminating the professions of bus driver and cab driver are likely to make New York a better place.
Dear Peter,
I don’t know if this is the reason that knots are mentioned in the announcement along side of strings, but it is worth mentioning that knot theory is fundamental to loop quantum gravity. The basis result of the field, proved by Rovelli and myself, is
that the gauge and spatially diffeomorphism invariant states are in one to one correspondence with diffeo classes of knots and graphs. Thus many basic steps to construct quantum gravity are problems in knot theory. One example of this is that the Jone’s polynomial (Kauffman bracket) is a specification of a physical state of the gravitational field, with a non-zero cosmological constant.
Meanwhile, Will Cunningham at PI has been using machine learning to get unprecedented results on the causal set approach to QG. He and I are also part of a collaboration of computer scientists and physicists developing applications of machine learning to problems in fundamental physics.
Thanks,
Lee Smolin
Knots: It’s not a bad thing if an area of real inquiry with the possibility of definite results gets a splash from the latest mania.
IAIFI: This sounds like science fiction. I could see data processing being useful in astronomy … how is it going to help *theory*?
Thanks for a very nice rep theory book, Dr. Woit. It took me a long time to be ready to appreciate it; now I’m very thankful to you.