The Director of the NSF Institute for Artificial Intelligence and Fundamental Interactions discusses the fruitful intersection of AI and physics.
Jesse Thaler is a theoretical particle physicist at the Massachusetts Institute of Technology (MIT), who seeks to address outstanding questions in fundamental physics by bringing together techniques from quantum field theory and machine learning. He is also the Director of the NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI).
Launched in 2020 with five years of funding from the US National Science Foundation (NSF), IAIFI was envisioned as a way to bring together the critical mass of people in the Boston area who were thinking about AI issues, physics issues, and their intersection.
FirstPrinciples recently interviewed Thaler about the challenges and opportunities at the intersection of AI and fundamental physics and the work being done by IAIFI researchers. He shared that, not so long ago, his students had to convince him to “go from a machine learning curmudgeon to a machine learning evangelist.” He even joked that follow-up questions from this interview could be directed to ChatJesseT, a playful spin on ChatGPT that his students and postdocs created as an April Fool’s joke.
The interview has been condensed and edited for clarity.
FirstPrinciples: Can you talk a little about IAIFI’s founding and how you became involved?
Jesse Thaler: I'm trained as a theoretical physicist. I was at an event in Zurich with a bunch of friends in theoretical particle physics who were starting to use machine learning. And I joked, “Deep learning? We’re theoretical physicists. We should be doing deep thinking!” That joke ultimately became the tagline for IAIFI: deep learning + deep thinking = deeper understanding. AI and physics is really a two-way street: AI’s influence on how we research new physics phenomena, but also applying physics thinking to the way that AI systems operate.
I made that joke in July 2016. That same September, two spectacular graduate students came to my office with a paper that they had written as master’s students, attempting to convince me that there truly was a synergy between the type of first-principles studies that I do and these methods of advanced computation, statistical reasoning, computer science, and artificial intelligence.
I basically told them, in no uncertain terms, that this was not the way that I thought physics research should be going. And they agreed with me, actually, about the off-the-shelf use of machine learning. Their PhD research was really about injecting physics principles into AI systems, and teaching a machine to think like a physicist.
In some sense, IAIFI was built for those students, Patrick Komiske and Eric Metodiev, upon realizing that there weren't really career opportunities at this intersection between AI and physics. It was (and continues to be) an emerging field. Where would someone live within the academic ecosystem if they wanted to dive into physics topics, but also computation and statistics? That was the motivation for submitting a grant to the NSF to start this institute.
FP: What do you see as the exciting opportunities at this intersection?
JT: If you think about AI solely as it applies to fundamental physics research, we have massive challenges in our field and we’re trying to understand some of the deepest questions in nature.
It's maybe not so well known that there were a large number of machine learning algorithms behind the Higgs boson discovery in 2012. The Large Hadron Collider produces a massive amount of data. And to sift through that data, machine learning is now a very standard part of data analysis. That machine learning is slowly converting from shallow learning to deep learning. We’re even starting to see the growth of things like generative AI influencing how we think about physics analyses.
Cosmology is another area with massive amounts of data and massive computational costs to run simulations of the universe. Without something like machine learning, we simply wouldn't be able to tackle those problems.
Neutrino experiments provide another example. The US is making a heavy investment in a detector technology called liquid argon time projection chambers. This is almost like going back to the bubble chamber days, where you would have humans look at individual images of what was going on with particle interactions – only they are coming fast and furious with all the spills of the beam. You can't actually have humans view each of those pictures one by one. But you need human-style reasoning to figure out what is going on. That's a case where the technology wouldn't even be able to operate without machine learning helping us sift through that information – where AI for fundamental physics seems to be absolutely essential.
In terms of physics for AI, it's maybe not so obvious. Why is it that physics can be used to understand AI systems? Of course, with the recent Nobel Prize, it’s maybe easier to explain that connection. But there are at least four ways that we're seeing physics influence AI.
The first is the incorporation of physics into existing AI systems. For example, if you want a robot to navigate a landscape, you'd like to teach it a little bit about physical systems, three-dimensional space, symmetries of rotations – things like that. Even some of the powerful computer graphics that can be done with AI rely on a form of ray tracing that is based on physics principles about how light propagates.
Next, there is physics for AI, where there are concepts from physics that can help you build better AI tools, even if you aren’t directly studying physical systems. You can build machine learning architectures that have physics-style reasoning embedded and those concepts turn out to be quite powerful.
Then, there's the physics of AI, where you think about AI as if it were an actual physical system. For example, phase transitions are a very standard concept for physicists. Well, AI is not just one thing. AI can exhibit different phases, since you can adjust hyperparameters in a machine learning algorithm to change its behavior. As you adjust those hyperparameters, you actually go through different phases where learning happens in different ways. That's something that IAIFI researchers are figuring out, thinking about using tools of physics analysis to understand AI.
Lastly, there's physics pushing AI to do things that I don't even think AI experts thought their algorithms could do. Cosmology is this multi-scale problem, where you have dynamics happening at the level of the universe as a whole, and then zooming in, and zooming in, down to individual galaxies. How do you deal with phenomena happening on different scales? This is something that's really pushing AI algorithms to the limit because a lot of them were designed for text processing or for image recognition, where this multi-scale nature is not so apparent.
FP: IAIFI scientists work in an area that is changing at a breakneck pace compared to some subfields of physics. Are there any unique challenges or opportunities that come along with that pace of change?
JT: Well, it definitely requires us to think differently about our problems. As a theoretical physicist, my primary toolbox is quantum field theory. And you might ask, “How could quantum field theory possibly benefit from AI?” Quantum field theory is based on rigorous calculations, whereas the lore about AI is that, “Oh, AI can do things like hallucinate.” How could you possibly use AI in a context where you want to do something rigorous?
You have to think creatively about how to do that. One of my IAIFI colleagues is working on simplifying theoretical expressions. Famously, the traditional way to compute quantum scattering amplitudes requires reams and reams of paper that, if you simplify them, can compress down to a single line. But actually figuring out that compression is very challenging because you have special functions that have very complicated properties.
With large language models, we know how to deal with summarization of text. It's not so difficult to think about using that kind of text summary to do equation summary. Because you know what the rules are, you can ask the AI to spit out the rules it uses to do a simplification and then just verify if it actually worked. Machine learning gives us a heuristic that allows us to get reasonable answers in finite computational time.
In my own research, I've been trying to synthesize different types of theory calculations. This is like the story of the blind men and the elephant, except that the elephant is a sought-after fundamental physics calculation. You maybe can't calculate it directly, but you can calculate it in certain limits – you can calculate the trunk, you can calculate the tail – and then you’d like to find a way to synthesize those things together.
Can I turn that synthesis process into an optimization problem? If so, and if I can re-express my problem in that language, then machine learning offers a solution. But the onus is on me to figure out what that solution means. It requires me to think like a machine. But if I can do that, then I can accomplish things that I couldn't have done using my pencil-and-paper calculations.
FP: Creativity isn't what people often associate with AI.
JT: Right. We usually think of creativity as something special about the human brain. But I think we learned from ChatGPT that there's a type of creativity that you can get from just exhaustive search. I don't really understand how it works. But maybe, even just as a thought experiment, we need to think about past discoveries, including something as foundational as Einstein's theory of general relativity. How could we have discovered that as a type of exhaustive search over the space of theoretical possibilities? Is there a framing of that problem that didn't require this leap of insight?
I don't think anyone has really succeeded in doing that yet. But we have hints that already say that numerical and textual data are on a closer footing than you might think. You can imagine us having this conversation – it's not clear whether it's one year or ten years from now – where people are coming up with new conceptual breakthroughs by framing things in this AI language. And that would be really exciting.
FP: You mentioned ChatGPT. Generative AI has become a pervasive topic in the public consciousness. Does this spike in interest make the job of IAIFI researchers easier or more difficult?
JT: Because the work that we're doing is related to the more fundamental curiosity of research, we're not faced (at least directly) with some of the more societal concerns about AI, though we're quite cognizant of the fact that techniques we develop for fundamental physics applications might find their way into societal cases.
There is an algorithmic aspect to ethics where decisions are made using computational tools. An experimental colleague of mine was trying to design an algorithm that would be unbiased with respect to collecting data at the Large Hadron Collider. Of course, that same kind of de-biasing is relevant for more societal applications. He actually applied his tool to some benchmark medical imaging and incarceration data sets, and found that it had better performance than other approaches.
I think the hardest thing to explain to people is that these algorithms are probabilistic. Every time you put in a prompt into ChatGPT, you get a different answer. As a physicist, this is very common. I mean, this is quantum mechanics. Statistical reasoning is the backbone of what we do, at least in modern physics.
Oftentimes, the biggest barriers come from our own community. Many physicists don't embrace AI, partly because they don't understand what it can do or only associate it with deepfake videos. Off-the-shelf AI is not good for physics utilization. But with appropriate tweaks, you can actually have AI systems that are just as robust as traditional methods.
In terms of convincing people, the best way to do it is the same way that I was convinced. You have a junior person come into your office and show what AI can do – that it can satisfy the same standards of scientific rigor that we're used to, and that it can also answer questions you never thought you'd be able to answer using traditional methods. We really have to win over our skeptical colleagues one interaction at a time.
I just don't see AI going away. As scientists, we can’t be ostriches with our heads in the sand. We need to start to understand how these systems work and, in particular, design systems for our own community. The off-the-shelf AI doesn't really operate in a way that we can use for scientific discovery. But it's close. And we can have a nice synergy with the computer science community to drive discovery in both fields.
FP: Can you speak a little about IAIFI's connections to government and industry, and why those are important?
JT: IAIFI researchers are approaching primarily curiosity-driven problems. But there's a continuum between curiosity-driven and application-driven science, and ideas that start on the blackboard eventually can make it into consumers’ hands.
Right now, the AI conversation is being driven by consumer applications or industry applications, but there's a lot to be gained by talking to domain experts in other fields – not just physics, but philosophy, history, cosmology, chemistry, biology, and earth and planetary sciences, to name a few. Each of those fields has a different framing of what their data is, and have insights to gain if they're part of the conversation.
I went to Capitol Hill to talk to congressional staffers. Everyone needs to be literate in AI, but also needs to understand that AI looks different in different communities. At minimum, all of us are going to have to think about it for education. I have a 12-year-old son, so I'm very well aware that he could just fire up a chatbot to do his homework. Having different domain perspectives is quite important. We’re trying to get voices from the physics community in the room, voices beyond just the big companies.
FP: We're roughly four years into IAIFI’s initial five-year investment. What would you say the institute has accomplished so far? And what do you envision for its future?
I'm really proud that we've established “AI plus physics” as a domain that people view as being something. Our IAIFI postdoctoral fellowship has been enormously successful. Our first round of fellows – and, in fact, some of our second round – are going on to jobs in industry and academia. And the jobs that they are taking are jobs that didn't really exist before. In two cases, fellows with physics backgrounds were hired in computer science departments. That's really exciting to see. I'm hoping that eventually physics departments might hire someone with a more computer science background, too. It needs to be a two-way street.
Practically speaking, we're trying to get IAIFI renewed by the NSF. More broadly, the vision is to take what we've done with “AI plus physics” and broaden it to “AI plus science,” emphasizing that the way we do scientific discovery will change and that every field has something to contribute to that.
Exactly what that looks like? We will see whether it's continued government support, whether it comes from foundation support or private philanthropy. But my hope is that, five years from now, you're going to see more positions (and more breakthroughs) in this interdisciplinary space.
