Tag Archives: particle physics

About the OpenAI Amplitudes Paper, but Not as Much as You’d Like

I’ve had a bit more time to dig in to the paper I mentioned last week, where OpenAI collaborated with amplitudes researchers, using one of their internal models to find and prove a simplified version of a particle physics formula. I figured I’d say a bit about my own impressions from reading the paper and OpenAI’s press release.

This won’t be a real “deep dive”, though it will be long nonetheless. As it turns out, most of the questions I’d like answers to aren’t answered in the paper or the press release. Getting them will involve actual journalistic work, i.e. blocking off time to interview people, and I haven’t done that yet. What I can do is talk about what I know so far, and what I’m still wondering.

Context:

Scattering amplitudes are formulas used by particle physicists to make predictions. For a while, people would just calculate these when they needed them, writing down pages of mess that you could plug in numbers to to get answers. However, forty years ago two physicists decided they wanted more, writing “we hope to obtain a simplified form for the answer, making our result not only an experimentalist’s, but a theorist’s delight.”

In their next paper, they managed to find that “theorist’s delight”: a simplified, intuitive-looking answer that worked for calculations involving any number of particles, summarizing many different calculations. Ten years later, a few people had started building on it, and ten years after that, the big shots started paying attention. A whole subfield, “amplitudeology”, grew from that seed, finding new forms of “theorists’s delight” in scattering amplitudes.

Each subfield has its own kind of “theory of victory”, its own concept for what kind of research is most likely to yield progress. In amplitudes, it’s these kinds of simplifications. When they work out well, they yield new, more efficient calculation techniques, yielding new messy results which can be simplified once more. To one extent or another, most of the field is chasing after those situations when simplification works out well.

That motivation shapes both the most ambitious projects of senior researchers, and the smallest student projects. Students often spend enormous amounts of time looking for a nice formula for something and figuring out how to generalize it, often on a question suggested by a senior researcher. These projects mostly serve as training, but occasionally manage to uncover something more impressive and useful, an idea others can build around.

I’m mentioning all of this, because as far as I can tell, what ChatGPT and the OpenAI internal model contributed here roughly lines up with the roles students have on amplitudes papers. In fact, it’s not that different from the role one of the authors, Alfredo Guevara, had when I helped mentor him during his Master’s.

Senior researchers noticed something unusual, suggested by prior literature. They decided to work out the implications, did some calculations, and got some messy results. It wasn’t immediately clear how to clean up the results, or generalize them. So they waited, and eventually were contacted by someone eager for a research project, who did the work to get the results into a nice, general form. Then everyone publishes together on a shared paper.

How impressed should you be?

I said, “as far as I can tell” above. What’s annoying is that this paper makes it hard to tell.

If you read through the paper, they mention AI briefly in the introduction, saying they used GPT-5.2 Pro to conjecture formula (39) in the paper, and an OpenAI internal model to prove it. The press release actually goes into more detail, saying that the humans found formulas (29)-(32), and GPT-5.2 Pro found a special case where it could simplify them to formulas (35)-(38), before conjecturing (39). You can get even more detail from an X thread by one of the authors, OpenAI Research Scientist Alex Lupsasca. Alex had done his PhD with another one of the authors, Andrew Strominger, and was excited to apply the tools he was developing at OpenAI to his old research field. So they looked for a problem, and tried out the one that ended up in the paper.

What is missing, from the paper, press release, and X thread, is any real detail about how the AI tools were used. We don’t have the prompts, or the output, or any real way to assess how much input came from humans and how much from the AI.

(We have more for their follow-up paper, where Lupsasca posted a transcript of the chat.)

Contra some commentators, I don’t think the authors are being intentionally vague here. They’re following business as usual. In a theoretical physics paper, you don’t list who did what, or take detailed account of how you came to the results. You clean things up, and create a nice narrative. This goes double if you’re aiming for one of the most prestigious journals, which tend to have length limits.

This business-as-usual approach is ok, if frustrating, for the average physics paper. It is, however, entirely inappropriate for a paper showcasing emerging technologies. For a paper that was going to be highlighted this highly by OpenAI, the question of how they reached their conclusion is much more interesting than the results themselves. And while I wouldn’t ask them to go to the standards of an actual AI paper, with ablation analysis and all that jazz, they could at least have aimed for the level of detail of my final research paper, which gave samples of the AI input and output used in its genetic algorithm.

For the moment, then, I have to guess what input the AI had, and what it actually accomplished.

Let’s focus on the work done by the internal OpenAI model. The descriptions I’ve seen suggest that it started where GPT-5.2 Pro did, with formulas (29)-(32), but with a more specific prompt that guided what it was looking for. It then ran for 12 hours with no additional input, and both conjectured (39) and proved it was correct, providing essentially the proof that follows formula (39) in the paper.

Given that, how impressed should we be?

First, the model needs to decide to go to a specialized region, instead of trying to simplify the formula in full generality. I don’t know whether they prompted their internal model explicitly to do this. It’s not something I’d expect a student to do, because students don’t know what types of results are interesting enough to get published, so they wouldn’t be confident in computing only a limited version of a result without an advisor telling them it was ok. On the other hand, it is actually something I’d expect an LLM to be unusually likely to do, as a result of not managing to consistently stick to the original request! What I don’t know is whether the LLM proposed this for the right reason: that if you have the formula for one region, you can usually find it for other regions.

Second, the model needs to take formulas (29)-(32), write them in the specialized region, and simplify them to formulas (35)-(38). I’ve seen a few people saying you can do this pretty easily with Mathematica. That’s true, though not every senior researcher is comfortable doing that kind of thing, as you need to be a bit smarter than just using the Simplify[] command. Most of the people on this paper strike me as pen-and-paper types who wouldn’t necessarily know how to do that. It’s definitely the kind of thing I’d expect most students to figure out, perhaps after a couple of weeks of flailing around if it’s their first crack at it. The LLM likely would not have used Mathematica, but would have used SymPy, since these “AI scientist” setups usually can write and execute Python code. You shouldn’t think of this as the AI reasoning through the calculation itself, but it at least sounds like it was reasonably quick at coding it up.

Then, the model needs to conjecture formula (39). This gets highlighted in the intro, but as many have pointed out, it’s pretty easy to do. If any non-physicists are still reading at this point, take a look:

Could you guess (39) from (35)-(38)?

After that, the paper goes over the proof that formula (39) is correct. Most of this proof isn’t terribly difficult, but the way it begins is actually unusual in an interesting way. The proof uses ideas from time-ordered perturbation theory, an old-fashioned way to do particle physics calculations. Time-ordered perturbation theory isn’t something any of the authors are known for using with regularity, but it has recently seen a resurgence in another area of amplitudes research, showing up for example in papers by Matthew Schwartz, a colleague of Strominger at Harvard.

If a student of Strominger came up with an idea drawn from time-ordered perturbation theory, that would actually be pretty impressive. It would mean that, rather than just learning from their official mentor, this student was talking to other people in the department and broadening their horizons, showing a kind of initiative that theoretical physicists value a lot.

From an LLM, though, this is not impressive in the same way. The LLM was not trained by Strominger, it did not learn specifically from Strominger’s papers. Its context suggested it was working on an amplitudes paper, and it produced an idea which would be at home in an amplitudes paper, just a different one than the one it was working on.

While not impressive, that capability may be quite useful. Academic subfields can often get very specialized and siloed. A tool that suggests ideas from elsewhere in the field could help some people broaden their horizons.

Overall, it appears that that twelve-hour OpenAI internal model run reproduced roughly what an unusually bright student would be able to contribute over the course of a several-month project. Like most student projects, you could find a senior researcher who could do the project much faster, maybe even faster than the LLM. But it’s unclear whether any of the authors could have: different senior researchers have different skillsets.

A stab at implications:

If we take all this at face-value, it looks like OpenAI’s internal model was able to do a reasonably competent student project with no serious mistakes in twelve hours. If they started selling that capability, what would happen?

If it’s cheap enough, you might wonder if professors would choose to use the OpenAI model instead of hiring students. I don’t think this would happen, though: I think it misunderstands why these kinds of student projects exist in a theoretical field. Professors sometimes use students to get results they care about, but more often, the student’s interest is itself the motivation, with the professor wanting to educate someone, to empire-build, or just to take on their share of the department’s responsibilities. AI is only useful for this insofar as AI companies continue reaching out to these people to generate press releases: once this is routinely possible, the motivation goes away.

More dangerously, if it’s even cheaper, you could imagine students being tempted to use it. The whole point of a student project is to train and acculturate the student, to get them to the point where they have affection for the field and the capability to do more impressive things. You can’t skip that, but people are going to be tempted to.

And of course, there is the broader question of how much farther this technology can go. That’s the hardest to estimate here, since we don’t know the prompts used. So I don’t know if seeing this result tells us anything more about the bigger picture than we knew going in.

Remaining questions:

At the end of the day, there are a lot of things I still want to know. And if I do end up covering this professionally, they’re things I’ll ask.

What was the prompt given to the internal model, and how much did it do based on that prompt?
Was it really done in one shot, no retries or feedback?
How much did running the internal model cost?
Is this result likely to be useful? Are there things people want to calculate that this could make easier? Recursion relations it could seed? Is it useful for SCET somehow?
How easy would it have been for the authors to do what the LLM did? What about other experts in the community?

The Timeline for Replacing Theorists Is Not Technological

9 Replies

Quanta Magazine recently published a reflection by Natalie Wolchover on the state of fundamental particle physics. The discussion covers a lot of ground, but one particular paragraph has gotten the lion’s share of the attention. Wolchover talked to Jared Kaplan, the ex-theoretical physicist turned co-founder of Anthropic, one of the foremost AI companies today.

Kaplan was one of Nima Arkani-Hamed’s PhD students, which adds an extra little punch.

There’s a lot to contest here. Is AI technology anywhere close to generating papers as good as the top physicists, or is that relegated to the sci-fi future? Does Kaplan really believe this, or is he just hyping up his company?

I don’t have any special insight into those questions, about the technology and Kaplan’s motivations. But I think that, even if we trusted him on the claim that AI could be generating Witten- or Nima-level papers in three years, that doesn’t mean it will replace theoretical physicists. That part of the argument isn’t a claim about the technology, but about society.

So let’s take the technological claims as given, and make them a bit more specific. Since we don’t have any objective way of judging the quality of scientific papers, let’s stick to the subjective. Today, there are a lot of people who get excited when Witten posts a new paper. They enjoy reading them, they find the insights inspiring, they love the clarity of the writing and their tendency to clear up murky ideas. They also find them reliable: the papers very rarely have mistakes, and don’t leave important questions unanswered.

Let’s use that as our baseline, then. Suppose that Anthropic had an AI workflow that could reliably write papers that were just as appealing to physicists as Witten’s papers are, for the same reasons. What happens to physicists?

Witten himself is retired, which for an academic means you do pretty much the same thing you were doing before, but now paid out of things like retirement savings and pension funds, not an institute budget. Nobody is going to fire Witten, there’s no salary to fire him from. And unless he finds these developments intensely depressing and demoralizing (possible, but very much depends on how this is presented), he’s not going to stop writing papers. Witten isn’t getting replaced.

More generally, though, I don’t think this directly results in anyone getting fired, or in universities trimming positions. The people making funding decisions aren’t just sitting on a pot of money, trying to maximize research output. They’ve got money to be spent on hires, and different pools of money to be spent on equipment, and the hires get distributed based on what current researchers at the institutes think is promising. Universities want to hire people who can get grants, to help fund the university, and absent rules about AI personhood, the AIs won’t be applying for grants.

Funding cuts might be argued for based on AI, but that will happen long before AI is performing at the Witten level. We already see this happening in other industries or government agencies, where groups that already want to cut funding are getting think tanks and consultants to write estimates that justify cutting positions, without actually caring whether those estimates are performed carefully enough to justify their conclusions. That can happen now, and doesn’t depend on technological progress.

AI could also replace theoretical physicists in another sense: the physicists themselves might use AI to do most of their work. That’s more plausible, but here adoption still heavily depends on social factors. Will people feel like they are being assessed on whether they can produce these Witten-level papers, and that only those who make them get hired, or funded? Maybe. But it will propagate unevenly, from subfield to subfield. Some areas will make their own rules forbidding AI content, there will be battles and scandals and embarrassments aplenty. It won’t be a single switch, the technology alone setting the timeline.

Finally, AI could replace theoretical physicists in another way, by people outside of academia filling the field so much that theoretical physicists have nothing more that they want to do. Some non-physicists are very passionate about physics, and some of those people have a lot of money. I’ve done writing work for one such person, whose foundation is now attempting to build an AI Physicist. If these AI Physicists get to Witten-level quality, they might start writing compelling paper after compelling paper. Those papers, though, will due to their origins be specialized. Much as philanthropists mostly fund the subfields they’ve heard of, philanthropist-funded AI will mostly target topics the people running the AI have heard are important. Much like physicists themselves adopting the technology, there will be uneven progress from subfield to subfield, inch by socially-determined inch.

In a hard-to-quantify area like progress in science, that’s all you can hope for. I suspect Kaplan got a bit of a distorted picture of how progress and merit work in theoretical physics. He studied with Nima Arkani-Hamed, who is undeniably exceptionally brilliant but also undeniably exceptionally charismatic. It must feel to a student of Nima’s that academia simply hires the best people, that it does whatever it takes to accomplish the obviously best research. But the best research is not obvious.

I think some of these people imagine a more direct replacement process, not arranged by topic and tastes, but by goals. They picture AI sweeping in and doing what theoretical physics was always “meant to do”: solve quantum gravity, and proceed to shower us with teleporters and antigravity machines. I don’t think there’s any reason to expect that to happen. If you just asked a machine to come up with the most useful model of the universe for a near-term goal, then in all likelihood it wouldn’t consider theoretical high-energy physics at all. If you see your AI as a tool to navigate between utopia and dystopia, theoretical physics might matter at some point: when your AI has devoured the inner solar system, is about to spread beyond the few light-minutes when it can signal itself in real-time, and has to commit to a strategy. But as long as the inner solar system remains un-devoured, I don’t think you’ll see an obviously successful theory of fundamental physics.

On Theories of Everything and Cures for Cancer

5 Replies

Some people are disappointed in physics. Shocking, I know!

Those people, when careful enough, clarify that they’re disappointed in fundamental physics: not the physics of materials or lasers or chemicals or earthquakes, or even the physics of planets and stars, but the physics that asks big fundamental questions, about the underlying laws of the universe and where they come from.

Some of these people are physicists themselves, or were once upon a time. These often have in mind other directions physicists should have gone. They think that, with attention and funding, their own ideas would have gotten us closer to our goals than the ideas that, in practice, got the attention and the funding.

Most of these people, though, aren’t physicists. They’re members of the general public.

It’s disappointment from the general public, I think, that feels the most unfair to physicists. The general public reads history books, and hears about a series of revolutions: Newton and Maxwell, relativity and quantum mechanics, and finally the Standard Model. They read science fiction books, and see physicists finding “theories of everything”, and making teleporters and antigravity engines. And they wonder what made the revolutions stop, and postponed the science fiction future.

Physicists point out, rightly, that this is an oversimplified picture of how the world works. Something happens between those revolutions, the kind of progress not simple enough to summarize for history class. People tinker away at puzzles, and make progress. And they’re still doing that, even for the big fundamental questions. Physicists know more about even faraway flashy topics like quantum gravity than they did ten years ago. And while physicists and ex-physicists can argue about whether that work is on the right path, it’s certainly farther along its own path than it was. We know things we didn’t know before, progress continues to be made. We aren’t at the “revolution” stage yet, or even all that close. But most progress isn’t revolutionary, and no-one can predict how often revolutions should take place. A revolution is never “due”, and thus can never be “overdue”.

Physicists, in turn, often don’t notice how normal this kind of reaction from the public is. They think people are being stirred up by grifters, or negatively polarized by excess hype, that fundamental physics is facing an unfair reaction only shared by political hot-button topics. But while there are grifters, and people turned off by the hype…this is also just how the public thinks about science.

Have you ever heard the phrase “a cure for cancer”?

Fiction is full of scientists working on a cure for cancer, or who discovered a cure for cancer, or were prevented from finding a cure for cancer. It’s practically a trope. It’s literally a trope.

It’s also a real thing people work on, in a sense. Many scientists work on better treatments for a variety of different cancers. They’re making real progress, even dramatic progress. As many whose loved ones have cancer know, it’s much more likely for someone with cancer to survive than it was, say, twenty years ago.

But those cures don’t meet the threshold for science fiction, or for the history books. They don’t move us, like the polio vaccine did, from a world where you know many people with a disease to a world where you know none. They don’t let doctors give you a magical pill, like in a story or a game, that instantly cures your cancer.

For the vast majority of medical researchers, that kind of goal isn’t realistic, and isn’t worth thinking about. The few that do pursue it work towards extreme long-term solutions, like periodically replacing everyone’s skin with a cloned copy.

So while you will run into plenty of media descriptions of scientists working on cures for cancer, you won’t see the kind of thing the public expects is an actual “cure for cancer”. And people are genuinely disappointed about this! “Where’s my cure for cancer?” is a complaint on the same level as “where’s my hovercar?” There are people who think that medical science has made no progress in fifty years, because after all those news articles, we still don’t have a cure for cancer.

I appreciate that there are real problems in what messages are being delivered to the public about physics, both from hypesters in the physics mainstream and grifters outside it. But put those problems aside, and a deeper issue remains. People understand the world as best they can, as a story. And the world is complicated and detailed, full of many people making incremental progress on many things. Compared to a story, the truth is always at a disadvantage.

For Newtonmas, One Seventeenth of a New Collider

Reminder to Physics Popularizers: “Discover” Is a Technical Term

C. N. Yang, Dead at 103

1 Reply

I don’t usually do obituaries here, but sometimes I have something worth saying.

Chen Ning Yang, a towering figure in particle physics, died last week.

Picture from 1957, when he received his Nobel

I never met him. By the time I started my PhD at Stony Brook, Yang was long-retired, and hadn’t visited the Yang Institute for Theoretical Physics in quite some time.

(Though there was still an office door, tucked behind the institute’s admin staff, that bore his name.)

The Nobel Prize doesn’t always honor the most important theoretical physicists. In order to get a Nobel Prize, you need to discover something that gets confirmed by experiment. Generally, it has to be a very crisp, clear statement about reality. New calculation methods and broader new understandings are on shakier ground, and theorists who propose them tend to be left out, or at best combined together into lists of partial prizes long after the fact.

Yang was lucky. With T. D. Lee, he had made that crisp, clear statement. He claimed that the laws of physics, counter to everyone’s expectations, are not the same when reflected in a mirror. In 1956, Wu confirmed the prediction, and Lee and Yang got the prize the year after.

That’s a huge, fundamental discovery about the natural world. But as a theorist, I don’t think that was Yang’s greatest accomplishment.

Yang contributed to other fields. Practicing theorists have seen his name strewn across concepts, formalisms, and theorems. I didn’t have space to talk about him in my article on integrability for Quanta Magazine, but only just barely: another paragraph or two, and he would have been there.

But his most influential contribution is something even more fundamental. And long-time readers of this blog should already know what it is.

Yang, along with Robert Mills, proposed Yang-Mills Theory.

There isn’t a Nobel prize for Yang-Mills theory. In 1953, when Yang and Mills proposed the theory, it was obviously wrong, a theory that couldn’t explain anything in the natural world, mercilessly mocked by famous bullshit opponent Wolfgang Pauli. Not even an ambitious idea that seemed outlandish (like plate tectonics), it was a theory with such an obvious missing piece that, for someone who prioritized experiment like the Nobel committee does, it seemed pointless to consider.

All it had going for it was that it was a clear generalization, an obvious next step. If there are forces like electromagnetism, with one type of charge going from plus to minus, why not a theory with multiple, interacting types of charge?

Nothing about Yang-Mills theory was impossible, or contradictory. Mathematically, it was fine. It obeyed all the rules of quantum mechanics. It simply didn’t appear to match anything in the real world.

But, as theorists learn, nature doesn’t let a good idea go to waste.

Of the four fundamental forces of nature, as it would happen, half are Yang-Mills theories. Gravity is different, electromagnetism is simpler, and could be understood without Yang and Mills’ insights. But the weak nuclear force, that’s a Yang-Mills theory. It wasn’t obvious in 1953 because it wasn’t clear how the massless, photon-like particles in Yang-Mills theory could have mass, and it wouldn’t become clear until the work of Peter Higgs over a decade later. And the strong nuclear force, that’s also a Yang-Mills theory, missed because of the ability of such a strong force to “confine” charges, hiding them away.

So Yang got a Nobel, not for understanding half of nature’s forces before anyone else had, but from a quirky question of symmetry.

In practice, Yang was known for all of this, and more. He was enormously influential. I’ve heard it claimed that he personally kept China from investing in a new particle collider, the strength of his reputation the most powerful force on that side of the debate, as he argued that a developing country like China should be investing in science with more short-term industrial impact, like condensed matter and atomic physics. I wonder if the debate will shift with his death, and what commitments the next Chinese five-year plan will make.

Ultimately, Yang is an example of what a theorist can be, a mix of solid work, counterintuitive realizations, and the thought-through generalizations that nature always seems to make use of in the end. If you’re not clear on what a theoretical physicist is, or what one can do, let Yang’s story be your guide.

The Rocks in the Ground Era of Fundamental Physics

5 Replies

It’s no secret that the early twentieth century was a great time to make progress in fundamental physics. On one level, it was an era when huge swaths of our understanding of the world were being rewritten, with relativity and quantum mechanics just being explored. It was a time when a bright student could guide the emergence of whole new branches of scholarship, and recently discovered physical laws could influence world events on a massive scale.

Put that way, it sounds like it was a time of low-hanging fruit, the early days of a field when great strides can be made before the easy problems are all solved and only the hard ones are left. And that’s part of it, certainly: the fields sprung from that era have gotten more complex and challenging over time, requiring more specialized knowledge to make any kind of progress. But there is also a physical reason why physicists had such an enormous impact back then.

The early twentieth century was the last time that you could dig up a rock out of the ground, do some chemistry, and end up with a discovery about the fundamental laws of physics.

When scientists like Curie and Becquerel were working with uranium, they didn’t yet understand the nature of atoms. The distinctions between elements were described in qualitative terms, but only just beginning to be physically understood. That meant that a weird object in nature, “a weird rock”, could do quite a lot of interesting things.

And once you find a rock that does something physically unexpected, you can scale up. From the chemistry experiments of a single scientist’s lab, countries can build industrial processes to multiply the effect. Nuclear power and the bomb were such radical changes because they represented the end effect of understanding the nature of atoms, and atoms are something people could build factories to manipulate.

Scientists went on to push that understanding further. They wanted to know what the smallest pieces of matter were composed of, to learn the laws behind the most fundamental laws they knew. And with relativity and quantum mechanics, they could begin to do so systematically.

US particle physics has a nice bit of branding. They talk about three frontiers: the Energy Frontier, the Intensity Frontier, and the Cosmic Frontier.

Some things we can’t yet test in physics are gated by energy. If we haven’t discovered a particle, it may be because it’s unstable, decaying quickly into lighter particles so we can’t observe it in everyday life. If these particles interact appreciably with particles of everyday matter like protons and electrons, then we can try to make them in particle colliders. These end up creating pretty much everything up to a certain mass, due to a combination of the tendency in quantum mechanics for everything that can happen to happen, and relativity’s $E=mc^2$ . In the mid-20th century these particle colliders were serious pieces of machinery, but still small enough to make industrial: now, there are so-called medical accelerators in many hospitals based on their designs. But current particle accelerators are a different beast, massive facilities built by international collaborations. This is the Energy Frontier.

Some things in physics are gated by how rare they are. Some particles interact only very faintly with other particles, so to detect them, physicists have to scan a huge chunk of matter, a giant tank of argon or a kilometer of antarctic ice, looking for deviations from the norm. Over time, these experiments have gotten bigger, looking for more and more subtle effects. A few weird ones still fit on tabletops, but only because they have the tools to measure incredibly small variations. Most are gigantic. This is the Intensity Frontier.

Finally, the Cosmic Frontier looks for the unknown behind both kinds of gates, using the wider universe to look at events with extremely high energy or size.

Pushing these frontiers has meant cleaning up our understanding of the fundamental laws of physics up to these frontiers. It means that whatever is still hiding, it either requires huge amounts of energy to produce, or is an extremely rare, subtle effect.

That means that you shouldn’t expect another nuclear bomb out of fundamental physics. Physics experiments are already working on vast scales, to the extent that a secret government project would have to be smaller than publicly known experiments, in physical size, energy use, and budget. And you shouldn’t expect another nuclear power plant, either: we’ve long passed the kinds of things you could devise a clever industrial process to take advantage of at scale.

Instead, new fundamental physics will only be directly useful once we’re the kind of civilization that operates on a much greater scale than we do today. That means larger than the solar system: there wouldn’t be much advantage, at this point, of putting a particle physics experiment on the edge of the Sun. It means the kind of civilization that tosses galaxies around.

It means that right now, you won’t see militaries or companies pushing the frontiers of fundamental physics, unlike the way they might have wanted to at the dawn of the twentieth century. By the time fundamental physics is useful in that way, all of these actors will likely be radically different: companies, governments, and in all likelihood human beings themselves. Instead, supporting fundamental physics right now is an act of philanthropy, maintaining a practice because it maintains good habits of thought and produces powerful ideas, the same reasons organizations support mathematics or poetry. That’s not nothing, and fundamental physics is still often affordable as philanthropy goes. But it’s not changing the world, not the way physicists did in the early twentieth century.

Some Dumb AI Ideas

8 Replies

Sometimes, when I write a post about AI, I’ve been sitting on an idea for a long time. I’ve talked to experts, I’ve tried to understand the math, I’ve honed my points and cleared away clutter.

This is not one of those times. The ideas in this post almost certainly have something deeply wrong with them. But hopefully they’re interesting food for thought.

My first dumb idea: instruction tuning was a mistake.

I’m drawing the seeds of this one from a tumblr post by nostalgebraist, someone known for making a popular bot trained on his tumblr posts in the early days before GPT became ChatGPT.

AIs like ChatGPT are based on Large Language Models, insanely complicated mathematical formulas that predict, given part of a text, what the rest of that text is likely to look like. In the early days, this was largely how they were used. Loosely described nostalgebraist’s bot, called nostalgebraist-autoresponder, began with a list of tumblr posts and asks and determines what additional posts would best fit in.

If you think about it, though, ChatGPT doesn’t really work like that. ChatGPT has conversations: you send it messages, it sends you responses. The text it creates is a dialogue, with you supplying half the input. But most texts aren’t dialogues, and ChatGPT draws on a lot of non-dialogue texts to make its dialogue-like responses.

The reason it does this is something called instruction tuning. ChatGPT has been intentionally biased, not to give the most likely completion to a task in general, but to give completions that fit this dialogue genre. What I didn’t know until I read nostalgebraist’s post was that this genre was defined artificially: AI researchers made up fake dialogues with AI, cheesy sci-fi conversations imagining how an AI might respond to instructions from a user, and then biased the Large Language Model so that rather than giving the most likely text in general, it gives a text that is more likely to look like these cheesy sci-fi conversations. It’s why ChatGPT sounds kind of like a fictional robot: not because sci-fi writers accurately predicted what AI would sound like, but because AI was created based on sci-fi texts.

For nostalgebraist, this leads into an interesting reflection of how a sci-fi AI should behave, how being warped around a made-up genre without history or depth creates characters which act according to simple narratives and express surprising anxiety.

For myself, though, I can’t help but wonder if the goal of dialogue itself is the problem. Dialogue is clearly important commercially: people use ChatGPT because they can chat with it. But Large Language Models aren’t inherently chatbots: they produce plausible texts, of any sort you could imagine. People seem to want a machine that can, for example, answer scientific questions as part of a conversation. But most competent answers to scientific questions aren’t conversations, they’re papers. If people stuck with the “raw” model, producing excerpts of nonexistent papers rather than imitating a dialogue with a non-existent expert, wouldn’t you expect the answers to be more accurate, with the model no longer biased by an irrelevant goal? Is the need to make a sell-able chatbot making these AIs worse at everything else people are trying to use them for?

I’m imagining a world where, instead of a chatbot, OpenAI built an “alternate universe simulator”. You give it some context, some texts or parts of texts from a universe you made up, and it completes them in a plausible way. By imagining different universes, you can use it to answer different questions. Such a gimmick would get fewer customers, and fewer investors, it would probably do worse. But I have to wonder if the actual technology might have been more useful.

My second idea is dumber, to the point where I mostly know why it doesn’t work. But thinking about it might help clarify how things work for people unused to AI.

I saw someone point out that, unlike something like Wikipedia, AI doesn’t give you context. You shouldn’t trust Wikipedia, or a source you find on Google, blindly. If you want to, you can look through the edit history on Wikipedia, or figure out who wrote a page you found on Google and how. If ChatGPT tells you something, by default you don’t know where that knowledge came from. You can tell it to search, and then you’ll get links, but that’s because it’s using Google or the like behind the scenes anyway. You don’t know where the model is getting its ideas.

Why couldn’t we get that context, though?

Every text produced by a Large Language Model is causally dependent on its training data. Different data, different model, different text. That doesn’t mean that each text draws from one source, or just a few sources: ChatGPT isn’t copying the training data, at least not so literally.

But it does mean that, if ChatGPT says something is true, you should in principle be able to ask which data was most important in making it say that. If you leave a piece of data out of the training, and get similar answers, you can infer that the response you got doesn’t have much to do with that piece of data. But if you leave out a text in training, and now ChatGPT gives totally different responses to the same question…then there’s a pretty meaningful sense that it got the information from that source.

If this were the type of non-AI statistical model people use in physics, this would be straightforward. Researchers do this all the time: take one experiment out of the data, see how their analysis changes, and thereby figure out which experiments are most important to check. One can even sometimes calculate, given a model, where you should look.

Unfortunately, you can’t do this with ChatGPT. The model is just too big. You can’t calculate anything explicitly about it, the giant mathematical formulas behind it are so complicated that the most you can do is get probabilities out case by case, you can’t “unwind” them and see where the numbers come from. And you can’t just take out sources one by one, and train the model again: not when training takes months of expensive computer time.

So unlike with the previous idea, I understand even on a technical level why you can’t do this. But it helped me to be able to think about what I would like to do, if it were possible. Maybe it helps you too!

Did the South Pole Telescope Just Rule Out Neutrino Masses? Not Exactly, Followed by My Speculations

9 Replies

Recently, the South Pole Telescope’s SPT-3G collaboration released new measurements of the cosmic microwave background, the leftover light from the formation of the first atoms. By measuring this light, cosmologists can infer the early universe’s “shape”: how it rippled on different scales as it expanded into the universe we know today. They compare this shape to mathematical models, equations and simulations which tie together everything we know about gravity and matter, and try to see what it implies for those models’ biggest unknowns.

Some of the most interesting such unknowns are neutrino masses. We know that neutrinos have mass because they transform as they move, from one type of neutrino to another. Those transformations let physicists measure the differences between neutrino masses, but but themselves, they don’t say what the actual masses are. All we know from particle physics, at this point, is a minimum: in order for the neutrinos to differ in mass enough to transform in the way they do, the total mass of the three flavors of neutrino must be at least 0.06 electron-Volts.

(Divided by the speed of light squared to get the right units, if you’re picky about that sort of thing. Physicists aren’t.)

Neutrinos also influenced the early universe, shaping it in a noticeably different way than heavier particles that bind together into atoms, like electrons and protons, did. That effect, observed in the cosmic microwave background and in the distribution of galaxies in the universe today, lets cosmologists calculate a maximum: if neutrinos are more massive than a certain threshold, they could not have the effects cosmologists observe.

Over time as measurements improved, this maximum has decreased. Now, the South Pole Telescope has added more data to the pool, and combining it with prior measurements…well, I’ll quote their paper:

Ok, it’s probably pretty hard to understand what that means if you’re not a physicist. To explain:

There are two different hypotheses for how neutrino masses work, called “hierarchies”. In the “normal” hierarchy, the neutrinos go in the same order as the particles they interact with with the weak nuclear force: electron-neutrinos are lighter than muon neutrinos, which are lighter than tau neutrinos. In the “inverted” hierarchy, they come in the opposite order, and the electron neutrino is the heaviest. Both of these are consistent with the particle-physics data.
Confidence is a statistics thing, which could take a lot of unpacking to define correctly. To give a short but likely tortured-sounding explanation: when you rule out a hypothesis with a certain confidence level, you’re saying that, if that hypothesis was true, there would only be a 100%-minus-that-chance chance that you would see what you actually observed.

So, what are the folks at the South Pole Telescope saying? They’re saying that if you put all the evidence together (that’s roughly what that pile of acroynms at the beginning means), then the result would be incredibly uncharacteristic for either hypothesis for neutrino masses. If you had “normal” neutrino masses, you’d only see these cosmological observations 2.1% of the time. And if you had inverted neutrino masses instead, you’d only see these observations 0.01% of the time!

That sure makes it sound like neither hypothesis is correct, right? Does it actually mean that?

I mean, it could! But I don’t think so. Here I’ll start speculating on the possibilities, from least likely in my opinion to most likely. This is mostly my bias talking, and shouldn’t be taken too seriously.

5. Neutrinos are actually massless

This one is really unlikely. The evidence from particle physics isn’t just quantitative, but qualitative. I don’t know if it’s possible to write down a model that reproduces the results of neutrino oscillation experiments without massive neutrinos, and if it is it would be a very bizarre model that would almost certainly break something else. This is essentially a non-starter.

4. This is a sign of interesting new physics

I mean, it would be nice, right? I’m sure there are many proposals at this point, tweaks that add a few extra fields with some hard-to-notice effects to explain the inconsistency. I can’t rule this out, and unlike the last point there isn’t anything about it that seems impossible. But we’ve had a lot of odd observations, and so far this hasn’t happened.

3. Someone did statistics wrong

This happens more often. Any argument like this is a statistical argument, and while physicists keep getting better at statistics, they’re not professional statisticians. Sometimes there’s a genuine misunderstanding that goes in to testing a model, and once it gets resolved the problem goes away.

2. The issue will go away with more data

The problem could also just…go away. 97.9% confidence sounds huge…but in physics, the standards are higher: you need 99.99994% to announce a new discovery. Physicists do a lot of experiments and observations, and sometimes, they see weird things! As the measurements get more precise, we may well see the disagreement melt away, and cosmology and particle physics both point to the same range for neutrino masses. It’s happened to many other measurements before.

1. We’re reaching the limits of our current approach to cosmology

This is probably not actually the most likely possibility, but it’s my list, what are you going to do?

There are basic assumptions behind how most theoretical physicists do cosmology. These assumptions are reasonably plausible, and seem to be needed to do anything at all. But they can be relaxed. Our universe looks like it’s homogeneous on the largest scales: the same density on average, in every direction you look. But the way that gets enforced in the mathematical models is very direct, and it may be that a different, more indirect, approach has more flexibility. I’ll probably be writing about this more in future, hopefully somewhere journalistic. But there are some very cool ideas floating around, gradually getting fleshed out more and more. It may be that the answer to many of the mysteries of cosmology right now is not new physics, but new mathematics: a new approach to modeling the universe.

Bonus Info on the LHC and Beyond

4 Replies

Three of my science journalism pieces went up last week!

(This is a total coincidence. One piece was a general explainer “held in reserve” for a nice slot in the schedule, one was a piece I drafted in February, while the third I worked on in May. In journalism, things take as long as they take.)

The shortest piece, at Quanta Magazine, was an explainer about the two types of particles in physics: bosons, and fermions.

I don’t have a ton of bonus info here, because of how tidy the topic is, so just two quick observations.

First, I have the vague impression that Bose, bosons’ namesake, is “claimed” by both modern-day Bangladesh and India. I had friends in grad school who were proud of their fellow physicist from Bangladesh, but while he did his most famous work in Dhaka, he was born and died in Calcutta. Since both were under British India for most of his life, these things likely get complicated.

Second, at the end of the piece I mention a “world on a wire” where fermions and bosons are the same. One example of such a “wire” is a string, like in string theory. One thing all young string theorists learn is “bosonization”: the idea that, in a 1+1-dimensional world like a string, you can re-write any theory with fermions as a theory with bosons, as well as vice versa. This has important implications for how string theory is set up.

Next, in Ars Technica, I had a piece about how LHC physicists are using machine learning to untangle the implications of quantum interference.

As a journalist, it’s really easy to fall into a trap where you give the main person you interview too much credit: after all, you’re approaching the story from their perspective. I tried to be cautious about this, only to be stymied when literally everyone else I interviewed praised Aishik Ghosh to the skies and credited him with being the core motivating force behind the project. So I shrugged my shoulders and followed suit. My understanding is that he has been appropriately rewarded and will soon be a professor at Georgia Tech.

I didn’t list the inventors of the NSBI method that Ghosh and co. used, but names like Kyle Cranmer and Johann Brehmer tend to get bandied about. It’s a method that was originally explored for a more general goal, trying to characterize what the Standard Model might be missing, while the work I talk about in the piece takes it in a new direction, closer to the typical things the ATLAS collaboration looks for.

I also did not say nearly as much as I was tempted to about how the ATLAS collaboration publishes papers, which was honestly one of the most intriguing parts of the story for me. There is a huge amount of review that goes on inside ATLAS before one of their papers reaches the outside world, way more than there ever is in a journal’s peer review process. This is especially true for “physics papers”, where ATLAS is announcing a new conclusion about the physical world, as ATLAS’s reputation stands on those conclusions being reliable. That means starting with an “internal note” that’s hundreds of pages long (and sometimes over a thousand), an editorial board that manages the editing process, disseminating the paper to the entire collaboration for comment, and getting specific experts and institute groups within the collaboration to read through the paper in detail. The process is a bit less onerous for “technical papers”, which describe a new method, not a new conclusion about the world. Still, it’s cumbersome enough that for those papers, often scientists don’t publish them “within ATLAS” at all, instead releasing them independently. The results I reported on are special because they involved a physics paper and a technical paper, both within the ATLAS collaboration process. Instead of just working with partial or simplified data, they wanted to demonstrate the method on a “full analysis”, with all the computation and human coordination that requires. Normally, ATLAS wouldn’t go through the whole process of publishing a physics paper without basing it on new data, but this was different: the method had the potential to be so powerful that the more precise results would be worth stating as physics results alone.

(Also, for the people in the comments worried about training a model on old data: that’s not what they did. In physics, they don’t try to train a neural network model to predict the results of colliders, such a model wouldn’t tell us anything useful. They run colliders to tell us whether what they see matches the analytic, Standard, model. The neural network is trained to predict not what the experiment will say, but what the Standard Model will say, as we can usually only figure that out through time-consuming simulations. So it’s trained on (new) simulations, not on experimental data.)

Finally, on Friday I had a piece in Physics Today about the European Strategy for Particle Physics (or ESPP), and in particular, plans for the next big collider.

Before I even started working on this piece, I saw a thread by Patrick Koppenburg on some of the 263 documents submitted for the ESPP update. While my piece ended up mostly focused on the big circular collider plan that most of the field is converging on (the future circular collider, or FCC), Koppenburg’s thread was more wide-ranging, meant to illustrate the breadth of ideas under discussion. Some of that discussion is about the LHC’s current plans, like its “high-luminosity” upgrade that will see it gather data at much higher rates up until 2040. Some of it is assessing broader concerns, which it may surprise some of you to learn includes sustainability: yes, there are more or less sustainable ways to build giant colliders.

The most fun part of the discussion, though, concerns all of the other collider proposals.

Some report progress on new technologies. Muon colliders are the most famous of these, but there are other proposals that would specifically help with a linear collider. I never did end up understanding what Cooled Copper Colliders are all about, beyond that they let you get more energy in a smaller machine without super-cooling. If you know about them, chime in in the comments! Meanwhile, plasma wakefield acceleration could accelerate electrons on a wave of plasma. This has the disadvantage that you want to collide electrons and positrons, and if you try to stick a positron in plasma it will happily annihilate with the first electron it meets. So what do you do? You go half-and-half, with the HALHF project: speed up the electron with a plasma wakefield, accelerate the positron normally, and have them meet in the middle.

Others are backup plans, or “budget options”, where CERN could get a bit better measurements on some parameters if they can’t stir up the funding to measure the things they really want. They could put electrons and positrons into the LHC tunnel instead of building a new one, for a weaker machine that could still study the Higgs boson to some extent. They could use a similar experiment to produce Z bosons instead, which could serve as a bridge to a different collider project. Or, they could collider the LHC’s proton beam with an electron beam, for an experiment that mixes advantages and disadvantages of some of the other approaches.

While working on the piece, one resource I found invaluable was this colloquium talk by Tristan du Pree, where he goes through the 263 submissions and digs up a lot of interesting numbers and commentary. Read the slides for quotes from the different national inputs and “solo inputs” with comments from particular senior scientists. I used that talk to get a broad impression of what the community was feeling, and it was interesting how well it was reflected in the people I interviewed. The physicist based in Switzerland felt the most urgency for the FCC plan, while the Dutch sources were more cautious, with other Europeans firmly in the middle.

Going over the FCC report itself, one thing I decided to leave out of the discussion was the cost-benefit analysis. There’s the potential for a cute sound-bite there, “see, the collider is net positive!”, but I’m pretty skeptical of the kind of analysis they’re doing there, even if it is standard practice for government projects. Between the biggest benefits listed being industrial benefits to suppliers and early-career researcher training (is a collider unusually good for either of those things, compared to other ways we spend money?) and the fact that about 10% of the benefit is the science itself (where could one possibly get a number like that?), it feels like whatever reasoning is behind this is probably the kind of thing that makes rigor-minded economists wince. I wasn’t able to track down the full calculation though, so I really don’t know, maybe this makes more sense than it looks.

I think a stronger argument than anything along those lines is a much more basic point, about expertise. Right now, we have a community of people trying to do something that is not merely difficult, but fundamental. This isn’t like sending people to space, where many of the engineering concerns will go away when we can send robots instead. This is fundamental engineering progress in how to manipulate the forces of nature (extremely powerful magnets, high voltages) and process huge streams of data. Pushing those technologies to the limit seems like it’s going to be relevant, almost no matter what we end up doing. That’s still not putting the science first and foremost, but it feels a bit closer to an honest appraisal of what good projects like this do for the world.

4 gravitons

Stories about physics from someone who's been there