Tag Archives: academia

The Many Varieties of Journal Club

Across disciplines, one tradition seems to unite all academics: the journal club. In a journal club, we gather together to discuss papers in academic journals. Typically, one person reads the paper in depth in advance, and comes prepared with a short presentation, then everyone else asks questions. Everywhere I’ve worked has either had, or aspired to have, a journal club, and every academic I’ve talked to recognizes the concept.

Beyond that universal skeleton, though, are a lot of variable details. Each place seems to interpret journal clubs just a bit differently. Sometimes a lot differently.

For example, who participates in journal clubs? In some places, journal clubs are a student thing, organized by PhD or Master’s students to get more experience with their new field. Some even have journal clubs as formal courses, for credit and everything. In other places, journal clubs are for everyone, from students up through the older professors.

What kind of papers? Some read old classic papers, knowing that without an excuse we’d never take the time to read them and would miss valuable insights. Some instead focus on the latest results, as a way to keep up with progress in the field.

Some variation is less intentional. Academics are busy, so it can be hard to find a volunteer to prepare a presentation on a paper every week. This leads journal clubs to cut corners, in once again a variety of ways. A journal club focused on the latest papers can sometimes only find volunteers interested in presenting their own work (which we usually already have a presentation prepared for). Sometimes this goes a step further, and the journal club becomes a kind of weekly seminar: a venue for younger visitors to talk about their work that’s less formal than a normal talk. Sometimes, instead of topic, the corner cut is preparation: people still discuss new papers, but instead of preparing a presentation they just come and discuss on the fly. This gets dangerous, because after a certain point people may stop reading the papers altogether, hoping that someone else will come having read it to explain it!

Journal clubs are tricky. Academics are curious, but we’re also busy and lazy. We know it would be good for us to discuss, to keep up with new papers or read the old classics… but actually getting organized, that’s another matter!

When Your Research Is a Cool Toy

Merry Newtonmas, everyone!

In the US, PhD students start without an advisor. As they finish their courses, different research groups make their pitch, trying to get them to join. Some promise interesting puzzles and engaging mysteries, others talk about the importance of their work, how it can help society or understand the universe.

Thinking back to my PhD, there is one pitch I remember to this day. The pitch was from the computational astrophysics group, and the message was a simple one: “we blow up stars”.

Obviously, these guys didn’t literally blow up stars: they simulated supernovas. They weren’t trying to make some weird metaphysical argument, they didn’t believe their simulation was somehow the real thing. The point they were making, instead, was emotional: blowing up stars feels cool.

Scientists can be motivated by curiosity, fame, or altruism, and these are familiar things. But an equally important motivation is a sense of play. If your job is to build tiny cars for rats, some of your motivation has to be the sheer joy of building tiny cars for rats. If you simulate supernovas, then part of your motivation can be the same as my nephew hurling stuffed animals down the stairs: that joyful moment when you yell “kaboom!”

Probably, your motivation shouldn’t just be to play with a cool toy. You need some of those “serious” scientific motivations as well. But for those of you blessed with a job where you get to say “kaboom”, you have that extra powerful reason to get up in the morning. And for those of you just starting a scientific career, may you have some cool toys under your Newtonmas tree!

Simulated Wormholes for My Real Friends, Real Wormholes for My Simulated Friends

Maybe you’ve recently seen a headline like this:

Actually, I’m more worried that you saw that headline before it was edited, when it looked like this:

If you’ve seen either headline, and haven’t read anything else about it, then please at least read this:

Physicists have not created an actual wormhole. They have simulated a wormhole on a quantum computer.

If you’re willing to read more, then read the rest of this post. There’s a more subtle story going on here, both about physics and about how we communicate it. And for the experts, hold on, because when I say the wormhole was a simulation I’m not making the same argument everyone else is.

[And for the mega-experts, there’s an edit later in the post where I soften that claim a bit.]

The headlines at the top of this post come from an article in Quanta Magazine. Quanta is a web-based magazine covering many fields of science. They’re read by the general public, but they aim for a higher standard than many science journalists, with stricter fact-checking and a goal of covering more challenging and obscure topics. Scientists in turn have tended to be quite happy with them: often, they cover things we feel are important but that the ordinary media isn’t able to cover. (I even wrote something for them recently.)

Last week, Quanta published an article about an experiment with Google’s Sycamore quantum computer. By arranging the quantum bits (qubits) in a particular way, they were able to observe behaviors one would expect out of a wormhole, a kind of tunnel linking different points in space and time. They published it with the second headline above, claiming that physicists had created a wormhole with a quantum computer and explaining how, using a theoretical picture called holography.

This pissed off a lot of physicists. After push-back, Quanta’s twitter account published this statement, and they added the word “Holographic” to the title.

Why were physicists pissed off?

It wasn’t because the Quanta article was wrong, per se. As far as I’m aware, all the technical claims they made are correct. Instead, it was about two things. One was the title, and the implication that physicists “really made a wormhole”. The other was the tone, the excited “breaking news” framing complete with a video comparing the experiment with the discovery of the Higgs boson. I’ll discuss each in turn:

The Title

Did physicists really create a wormhole, or did they simulate one? And why would that be at all confusing?

The story rests on a concept from the study of quantum gravity, called holography. Holography is the idea that in quantum gravity, certain gravitational systems like black holes are fully determined by what happens on a “boundary” of the system, like the event horizon of a black hole. It’s supposed to be a hologram in analogy to 3d images encoded in 2d surfaces, rather than like the hard-light constructions of science fiction.

The best-studied version of holography is something called AdS/CFT duality. AdS/CFT duality is a relationship between two different theories. One of them is a CFT, or “conformal field theory”, a type of particle physics theory with no gravity and no mass. (The first example of the duality used my favorite toy theory, N=4 super Yang-Mills.) The other one is a version of string theory in an AdS, or anti-de Sitter space, a version of space-time curved so that objects shrink as they move outward, approaching a boundary. (In the first example, this space-time had five dimensions curled up in a sphere and the rest in the anti-de Sitter shape.)

These two theories are conjectured to be “dual”. That means that, for anything that happens in one theory, you can give an alternate description using the other theory. We say the two theories “capture the same physics”, even though they appear very different: they have different numbers of dimensions of space, and only one has gravity in it.

Many physicists would claim that if two theories are dual, then they are both “equally real”. Even if one description is more familiar to us, both descriptions are equally valid. Many philosophers are skeptical, but honestly I think the physicists are right about this one. Philosophers try to figure out which things are real or not real, to make a list of real things and explain everything else as made up of those in some way. I think that whole project is misguided, that it’s clarifying how we happen to talk rather than the nature of reality. In my mind, dualities are some of the clearest evidence that this project doesn’t make any sense: two descriptions can look very different, but in a quite meaningful sense be totally indistinguishable.

That’s the sense in which Quanta and Google and the string theorists they’re collaborating with claim that physicists have created a wormhole. They haven’t created a wormhole in our own space-time, one that, were it bigger and more stable, we could travel through. It isn’t progress towards some future where we actually travel the galaxy with wormholes. Rather, they created some quantum system, and that system’s dual description is a wormhole. That’s a crucial point to remember: even if they created a wormhole, it isn’t a wormhole for you.

If that were the end of the story, this post would still be full of warnings, but the title would be a bit different. It was going to be “Dual Wormholes for My Real Friends, Real Wormholes for My Dual Friends”. But there’s a list of caveats. Most of them arguably don’t matter, but the last was what got me to change the word “dual” to “simulated”.

  1. The real world is not described by N=4 super Yang-Mills theory. N=4 super Yang-Mills theory was never intended to describe the real world. And while the real world may well be described by string theory, those strings are not curled up around a five-dimensional sphere with the remaining dimensions in anti-de Sitter space. We can’t create either theory in a lab either.
  2. The Standard Model probably has a quantum gravity dual too, see this cute post by Matt Strassler. But they still wouldn’t have been able to use that to make a holographic wormhole in a lab.
  3. Instead, they used a version of AdS/CFT with fewer dimensions. It relates a weird form of gravity in one space and one time dimension (called JT gravity), to a weird quantum mechanics theory called SYK, with an infinite number of quantum particles or qubits. This duality is a bit more conjectural than the original one, but still reasonably well-established.
  4. Quantum computers don’t have an infinite number of qubits, so they had to use a version with a finite number: seven, to be specific. They trimmed the model down so that it would still show the wormhole-dual behavior they wanted. At this point, you might say that they’re definitely just simulating the SYK theory, using a small number of qubits to simulate the infinite number. But I think they could argue that this system, too, has a quantum gravity dual. The dual would have to be even weirder than JT gravity, and even more conjectural, but the signs of wormhole-like behavior they observed (mostly through simulations on an ordinary computer, which is still better at this kind of thing than a quantum computer) could be seen as evidence that this limited theory has its own gravity partner, with its own “real dual” wormhole.
  5. But those seven qubits don’t just have the interactions they were programmed to have, the ones with the dual. They are physical objects in the real world, so they interact with all of the forces of the real world. That includes, though very weakly, the force of gravity.

And that’s where I think things break, and you have to call the experiment a simulation. You can argue, if you really want to, that the seven-qubit SYK theory has its own gravity dual, with its own wormhole. There are people who expect duality to be broad enough to include things like that.

But you can’t argue that the seven-qubit SYK theory, plus gravity, has its own gravity dual. Theories that already have gravity are not supposed to have gravity duals. If you pushed hard enough on any of the string theorists on that team, I’m pretty sure they’d admit that.

That is what decisively makes the experiment a simulation. It approximately behaves like a system with a dual wormhole, because you can approximately ignore gravity. But if you’re making some kind of philosophical claim, that you “really made a wormhole”, then “approximately” doesn’t cut it: if you don’t exactly have a system with a dual, then you don’t “really” have a dual wormhole: you’ve just simulated one.

Edit: mitchellporter in the comments points out something I didn’t know: that there are in fact proposals for gravity theories with gravity duals. They are in some sense even more conjectural than the series of caveats above, but at minimum my claim above, that any of the string theorists on the team would agree that the system’s gravity means it can’t have a dual, is probably false.

I think at this point, I’d soften my objection to the following:

Describing the system of qubits in the experiment as a limited version of the SYK theory is in one way or another an approximation. It approximates them as not having any interactions beyond those they programmed, it approximates them as not affected by gravity, and because it’s a quantum mechanical description it even approximates the speed of light as small. Those approximations don’t guarantee that the system doesn’t have a gravity dual. But in order for them to, then our reality, overall, would have to have a gravity dual. There would have to be a dual gravity interpretation of everything, not just the inside of Google’s quantum computer, and it would have to be exact, not just an approximation. Then the approximate SYK would be dual to an approximate wormhole, but that approximate wormhole would be an approximation of some “real” wormhole in the dual space-time.

That’s not impossible, as far as I can tell. But it piles conjecture upon conjecture upon conjecture, to the point that I don’t think anyone has explicitly committed to the whole tower of claims. If you want to believe that this experiment literally created a wormhole, you thus can, but keep in mind the largest asterisk known to mankind.

End edit.

If it weren’t for that caveat, then I would be happy to say that the physicists really created a wormhole. It would annoy some philosophers, but that’s a bonus.

But even if that were true, I wouldn’t say that in the title of the article.

The Title, Again

These days, people get news in two main ways.

Sometimes, people read full news articles. Reading that Quanta article is a good way to understand the background of the experiment, what was done and why people care about it. As I mentioned earlier, I don’t think anything said there was wrong, and they cover essentially all of the caveats you’d care about (except for that last one 😉 ).

Sometimes, though, people just see headlines. They get forwarded on social media, observed at a glance passed between friends. If you’re popular enough, then many more people will see your headline than will actually read the article. For many people, their whole understanding of certain scientific fields is formed by these glancing impressions.

Because of that, if you’re popular and news-y enough, you have to be especially careful with what you put in your headlines, especially when it implies a cool science fiction story. People will almost inevitably see them out of context, and it will impact their view of where science is headed. In this case, the headline may have given many people the impression that we’re actually making progress towards travel via wormholes.

Some of my readers might think this is ridiculous, that no-one would believe something like that. But as a kid, I did. I remember reading popular articles about wormholes, describing how you’d need energy moving in a circle, and other articles about optical physicists finding ways to bend light and make it stand still. Putting two and two together, I assumed these ideas would one day merge, allowing us to travel to distant galaxies faster than light.

If I had seen Quanta’s headline at that age, I would have taken it as confirmation. I would have believed we were well on the way to making wormholes, step by step. Even the New York Times headline, “the Smallest, Crummiest Wormhole You Can Imagine”, wouldn’t have fazed me.

(I’m not sure even the extra word “holographic” would have. People don’t know what “holographic” means in this context, and while some of them would assume it meant “fake”, others would think about the many works of science fiction, like Star Trek, where holograms can interact physically with human beings.)

Quanta has a high-brow audience, many of whom wouldn’t make this mistake. Nevertheless, I think Quanta is popular enough, and respectable enough, that they should have done better here.

At minimum, they could have used the word “simulated”. Even if they go on to argue in the article that the wormhole is real, and not just a simulation, the word in the title does no real harm. It would be a lie, but a beneficial “lie to children”, the basic stock-in-trade of science communication. I think they could have defended it to the string theorists they interviewed on those grounds.

The Tone

Honestly, I don’t think people would have been nearly so pissed off were it not for the tone of the article. There are a lot of physics bloggers who view themselves as serious-minded people, opposed to hype and publicity stunts. They view the research program aimed at simulating quantum gravity on a quantum computer as just an attempt to link a dying and un-rigorous research topic to an over-hyped and over-funded one, pompous storytelling aimed at promoting the careers of people who are already extremely successful.

These people tend to view Quanta favorably, because it covers serious-minded topics in a thorough way. And so many of them likely felt betrayed, seeing this Quanta article as a massive failure of that serious-minded-ness, falling for or even endorsing the hypiest of hype.

To those people, I’d like to politely suggest you get over yourselves.

Quanta’s goal is to cover things accurately, to represent all the facts in a way people can understand. But “how exciting something is” is not a fact.

Excitement is subjective. Just because most of the things Quanta finds exciting you also find exciting, does not mean that Quanta will find the things you find unexciting unexciting. Quanta is not on “your side” in some war against your personal notion of unexciting science, and you should never have expected it to be.

In fact, Quanta tends to find things exciting, in general. They were more excited than I was about the amplituhedron, and I’m an amplitudeologist. Part of what makes them consistently excited about the serious-minded things you appreciate them for is that they listen to scientists and get excited about the things they’re excited about. That is going to include, inevitably, things those scientists are excited about for what you think are dumb groupthinky hype reasons.

I think the way Quanta titled the piece was unfortunate, and probably did real damage. I think the philosophical claim behind the title is wrong, though for subtle and weird enough reasons that I don’t really fault anybody for ignoring them. But I don’t think the tone they took was a failure of journalistic integrity or research or anything like that. It was a matter of taste. It’s not my taste, it’s probably not yours, but we shouldn’t have expected Quanta to share our tastes in absolutely everything. That’s just not how taste works.

Confidence and Friendliness in Science

I’ve seen three kinds of scientific cultures.

First, there are folks who are positive about almost everyone. Ask them about someone else’s lab, even a competitor, and they’ll be polite at worst, and often downright excited. Anyone they know, they’ll tell you how cool the work they’re doing is, how it’s important and valuable and worth doing. They might tell you they prefer a different approach, but they’ll almost never bash someone’s work.

I’ve heard this comes out of American culture, and I can kind of see it. There’s an attitude in the US that everything needs to be described as positively as possible. This is especially true in a work context. Negativity is essentially a death sentence, doled out extremely rarely: if you explicitly say someone or their work is bad, you’re trying to get them fired. You don’t do that unless someone really really deserves it.

That style of scientific culture is growing, but it isn’t universal. There’s still a big cultural group that is totally ok with negativity: as long as it’s directed at other people, anyway.

This scientific culture prides itself on “telling it like it is”. They’ll happily tell you about how everything everyone else is doing is bullshit. Sometimes, they claim their ideas are the only ways forward. Others will have a small number of other people who they trust, who have gained their respect in one way or another. This sort of culture is most stereotypically associated with Russians: a “Russian-style” seminar, for example, is one where the speaker is aggressively questioned for hours.

It may sound like those are the only two options, but there is a third. While “American-style” scientists don’t criticize anyone, and “Russian-style” scientists criticize everyone else, there are also scientists who criticize almost everyone, including themselves.

With a light touch, this culture can be one of the best. There can be a real focus on “epistemic humility”, on always being clear of how much we still don’t know.

However, it can be worryingly easy to spill past that light touch, into something toxic. When the criticism goes past humility and into a lack of confidence in your own work, you risk falling into a black hole, where nothing is going well and nobody has a way out. This kind of culture can spread, filling a workplace and infecting anyone who spends too long there with the conviction that nothing will ever measure up again.

If you can’t manage that light skeptical touch, then your options are American-style or Russian-style. I don’t think either is obviously better. Both have their blind spots: the Americans can let bad ideas slide to avoid rocking the boat, while the Russians can be blind to their own flaws, confident that because everyone else seems wrong they don’t need to challenge their own worldview.

You have one more option, though. Now that you know this, you can recognize each for what it is: not the one true view of the world, but just one culture’s approach to the truth. If you can do that, you can pick up each culture as you need, switching between them as you meet different communities and encounter different things. If you stay aware, you can avoid fighting over culture and discourse, and use your energy on what matters: the science.

Visiting the IAS

I’m at the Institute for Advanced Study, or IAS, this week.

There isn’t a conference going on, but if you looked at the visitor list you’d be forgiven for thinking there was. We have talks in my subfield almost every day this week, two professors from my subfield here on sabbatical, and extra visitors on top of that.

The IAS is a bit of an odd place. Partly, that’s due to its physical isolation: tucked away in the woods behind Princeton, a half-hour’s walk from the nearest restaurant, it’s supposed to be a place for contemplation away from the hustle and bustle of the world.

Since the last time I visited they’ve added a futuristic new building, seen here out of my office window. The building is most notable for one wild promise: someday, they will serve dinner there.

Mostly, though, the weirdness of the IAS is due to the kind of institution it is.

Within a given country, most universities are pretty similar. Each may emphasize different teaching styles, and the US has a distinction between public and private, but (neglecting scammy for-profit universities), there are some commonalities of structure: both how they’re organized, and how they’re funded. Even between countries, different university systems have quite a bit of overlap.

The IAS, though, is not a university. It’s an independent institute. Neighboring Princeton supplies it with PhD students, but otherwise the IAS runs, and funds, itself.

There are a few other places like that around the world. The Perimeter Institute in Canada is also independent, and also borrows students from a neighboring university. CERN pools resources from several countries across Europe and beyond, Nordita from just the Nordic countries. Generalizing further, many countries have some sort of national labs or other nation-wide systems, from US Department of Energy labs like SLAC to Germany’s Max Planck Institutes.

And while universities share a lot in common, non-university institutes can be very different. Some are closely tied to a university, located inside university buildings with members with university affiliations. Others sit at a greater remove, less linked to a university or not linked at all. Some have their own funding, investments or endowments or donations, while others are mostly funded by governments, or groups of governments. I’ve heard that the IAS gets about 10% of its budget from the government, while Perimeter gets its everyday operating expenses entirely from the Canadian government and uses donations for infrastructure and the like.

So ultimately, the IAS is weird because every organization like it is weird. There are a few templates, and systems, but by and large each independent research organization is different. Understanding one doesn’t necessarily help at understanding another.

Fields and Scale

I am a theoretical particle physicist, and every morning I check the arXiv.

arXiv.org is a type of website called a preprint server. It’s where we post papers before they are submitted to (and printed by) a journal. In practice, everything in our field shows up on arXiv, publicly accessible, before it appears anywhere else. There’s no peer review process on arXiv, the journals still handle that, but in our field peer review doesn’t often notice substantive errors. So in practice, we almost never read the journals: we just check arXiv.

And so every day, I check the arXiv. I go to the section on my sub-field, and I click on a link that lists all of the papers that were new that day. I skim the titles, and if I see an interesting paper I’ll read the abstract, and maybe download the full thing. Checking as I’m writing this, there were ten papers posted in my field, and another twenty “cross-lists” were posted in other fields but additionally classified in mine.

Other fields use arXiv: mathematicians and computer scientists and even economists use it in roughly the same way physicists do. For biology and medicine, though, there are different, newer sites: bioRxiv and medRxiv.

One thing you may notice is the different capitalization. When physicists write arXiv, the “X” is capitalized. In the logo, it looks like a Greek letter chi, thus saying “archive”. The biologists and medical researchers capitalize the R instead. The logo still has an X that looks like a chi, but positioned with the R it looks like the Rx of medical prescriptions.

Something I noticed, but you might not, was the lack of a handy link to see new papers. You can search medRxiv and bioRxiv, and filter by date. But there’s no link that directly takes you to the newest papers. That suggests that biologists aren’t using bioRxiv like we use arXiv, and checking the new papers every day.

I was curious if this had to do with the scale of the field. I have the impression that physics and mathematics are smaller fields than biology, and that much less physics and mathematics research goes on than medical research. Certainly, theoretical particle physics is a small field. So I might have expected arXiv to be smaller than bioRxiv and medRxiv, and I certainly would expect fewer papers in my sub-field than papers in a medium-sized subfield of biology.

On the other hand, arXiv in my field is universal. In biology, bioRxiv and medRxiv are still quite controversial. More and more people are using them, but not every journal accepts papers posted to a preprint server. Many people still don’t use these services. So I might have expected bioRxiv and medRxiv to be smaller.

Checking now, neither answer is quite right. I looked between November 1 and November 2, and asked each site how many papers were uploaded between those dates. arXiv had the most, 604 papers. bioRxiv had roughly half that many, 348. medRxiv had 97.

arXiv represents multiple fields, bioRxiv is “just” biology. Specializing, on that day arXiv had 235 physics papers, 135 mathematics papers, and 250 computer science papers. So each individual field has fewer papers than biology in this period.

Specializing even further, I can look at a subfield. My subfield, which is fairly small, had 20 papers between those dates. Cell biology, which I would expect to be quite a big subfield, had 33.

Overall, the numbers were weirdly comparable, with medRxiv unexpectedly small compared to both arXiv and bioRxiv. I’m not sure whether there are more biologists than physicists, but I’m pretty sure there should be more cell biologists than theoretical particle physicists. This suggests that many still aren’t using bioRxiv. It makes me wonder: will bioRxiv grow dramatically in future? Are the people running it ready for if it does?

No, PhD Students Are Not Just Cheap Labor

Here’s a back-of-the-envelope calculation:

In 2019, there were 83,050 unionized graduate students in the US. Let’s assume these are mostly PhD students, since other graduate students are not usually university employees. I can’t find an estimate of the total number of PhD students in the US, but in 2019, 55,614 of them graduated. In 2020, the average US doctorate took 7.5 years to complete. That implies that 83,050/(55,614 x 7.5) = about one-fifth of PhD students in the US are part of a union.

That makes PhD student unions common, but not the majority. It means they’re not unheard of and strange, but a typical university still isn’t unionized. It’s the sweet spot for controversy. It leads to a lot of dumb tweets.

I saw one such dumb tweet recently, from a professor arguing that PhD students shouldn’t unionize. The argument was that if PhD students were paid more, then professors would prefer to hire postdocs, researchers who already have a doctoral degree.

(I won’t link to the tweet, in part because this person is probably being harassed enough already.)

I don’t know how things work in this professor’s field. But the implication, that professors primarily take on PhD students because they’re cheaper, not only doesn’t match my experience: it also just doesn’t make very much sense.

Imagine a neighborhood where the children form a union. They decide to demand a higher allowance, and to persuade any new children in the neighborhood to follow their lead.

Now imagine a couple in that neighborhood, deciding whether to have a child. Do you think that they might look at the fees the “children’s union” charges, and decide to hire an adult to do their chores instead?

Maybe there’s a price where they’d do that. If neighborhood children demanded thousands of dollars in allowance, maybe the young couple would decide that it’s too expensive to have a child. But a small shift is unlikely to change things very much: people have kids for many reasons, and those reasons don’t usually include cheap labor.

The reasons professors take on PhD students are similar to the reasons parents decide to have children. Some people have children because they want a legacy, something of theirs that survives to the next generation. For professors, PhD students are our legacy, our chance to raise someone on our ideas and see how they build on them. Some people have children because they love the act of child-raising: helping someone grow and learn about the world. The professors who take on students like taking on students: teaching is fun, after all.

That doesn’t mean there won’t be cases “on the margin”, where a professor finds they can’t afford a student they previously could. (And to be fair to the tweet I’m criticizing, they did even use the word “marginal”.) But they would have to be in a very tight funding situation, with very little flexibility.

And even for situations like that, long-term, I’m not sure anything would change.

I did my PhD in the US. I was part of a union, and in part because of that (though mostly because I was in a physics department), I was paid relatively decently for a PhD student. Relatively decently is still not that great, though. This was the US, where universities still maintain the fiction that PhD students only work 20 hours a week and pay proportionate to that, and where salaries in a university can change dramatically from student to postdoc to professor.

One thing I learned during my PhD is that despite our low-ish salaries, we cost our professors about as much as postdocs did. The reason why is tuition: PhD students don’t pay their own tuition, but that tuition still exists, and is paid by the professors who hire those students out of their grants. A PhD salary plus a PhD tuition ended up roughly equal to a postdoc salary.

Now, I’m working in a very different system. In a Danish university, wages are very flat. As a postdoc, a nice EU grant put me at almost the same salary as the professors. As a professor, my salary is pretty close to that of one of the better-paying schoolteacher jobs.

At the same time, tuition is much less relevant. Undergraduates don’t pay tuition at all, so PhD tuition isn’t based on theirs. Instead, it’s meant to cover costs of the PhD program as a whole.

I’ve filled out grants here in Denmark, so I know how much PhD students cost, and how much postdocs cost. And since the situation is so different, you might expect a difference here too.

There isn’t one. Hiring a PhD student, salary plus tuition, costs about as much as hiring a postdoc.

Two very different systems, with what seem to be very different rules, end up with the same equation. PhD students and postdocs cost about as much as each other, even if every assumption that you think would affect the outcome turns out completely different.

This is why I expect that, even if PhD students get paid substantially more, they still won’t end up that out of whack with postdocs. There appears to be an iron law of academic administration keeping these two numbers in line, one that holds across nations and cultures and systems. The proportion of unionized PhD students in the US will keep working its way upwards, and I don’t expect it to have any effect on whether professors take on PhDs.

From Journal to Classroom

As part of the pedagogy course I’ve been taking, I’m doing a few guest lectures in various courses. I’ve got one coming up in a classical mechanics course (“intermediate”-level, so not Newton’s laws, but stuff the general public doesn’t know much about like Hamiltonians). They’ve been speeding through the core content, so I got to cover a “fun” topic, and after thinking back to my grad school days I chose a topic I think they’ll have a lot of fun with: Chaos theory.

Getting the obligatory Warhammer reference out of the way now

Chaos is one of those things everyone has a vague idea about. People have heard stories where a butterfly flaps its wings and causes a hurricane. Maybe they’ve heard of the rough concept, determinism with strong dependence on the initial conditions, so a tiny change (like that butterfly) can have huge consequences. Maybe they’ve seen pictures of fractals, and got the idea these are somehow related.

Its role in physics is a bit more detailed. It’s one of those concepts that “intermediate classical mechanics” is good for, one that can be much better understood once you’ve been introduced to some of the nineteenth century’s mathematical tools. It felt like a good way to show this class that the things they’ve learned aren’t just useful for dusty old problems, but for understanding something the public thinks is sexy and mysterious.

As luck would have it, the venerable textbook the students are using includes a (2000’s era) chapter on chaos. I read through it, and it struck me that it’s a very different chapter from most of the others. This hit me particularly when I noticed a section describing a famous early study of chaos, and I realized that all the illustrations were based on the actual original journal article.

I had surprisingly mixed feelings about this.

On the one hand, there’s a big fashion right now for something called research-based teaching. That doesn’t mean “using teaching methods that are justified by research” (though you’re supposed to do that too), but rather, “tying your teaching to current scientific research”. This is a fashion that makes sense, because learning about cutting-edge research in an undergraduate classroom feels pretty cool. It lets students feel more connected with the scientific community, it inspires them to get involved, and it gets them more used to what “real research” looks like.

On the other hand, structuring your textbook based on the original research papers feels kind of lazy. There’s a reason we don’t teach Newtonian mechanics the way Newton would have. Pedagogy is supposed to be something we improve at over time: we come up with better examples and better notation, more focused explanations that teach what we want students to learn. If we just summarize a paper, we’re not really providing “added value”: we should hope, at this point, that we can do better.

Thinking about this, I think the distinction boils down to why you’re teaching the material in the first place.

With a lot of research-based teaching, the goal is to show the students how to interact with current literature. You want to show them journal papers, not because the papers are the best way to teach a concept or skill, but because reading those papers is one of the skills you want to teach.

That makes sense for very current topics, but it seems a bit weird for the example I’ve been looking at, an early study of chaos from the 60’s. It’s great if students can read current papers, but they don’t necessarily need to read older ones. (At least, not yet.)

What then, is the textbook trying to teach? Here things get a bit messy. For a relatively old topic, you’d ideally want to teach not just a vague impression of what was discovered, but concrete skills. Here though, those skills are just a bit beyond the students’ reach: chaos is more approachable than you’d think, but still not 100% something the students can work with. Instead they’re learning to appreciate concepts. This can be quite valuable, but it doesn’t give the kind of structure that a concrete skill does. In particular, it makes it hard to know what to emphasize, beyond just summarizing the original article.

In this case, I’ve come up with my own way forward. There are actually concrete skills I’d like to teach. They’re skills that link up with what the textbook is teaching, skills grounded in the concepts it’s trying to convey, and that makes me think I can convey them. It will give some structure to the lesson, a focus on not merely what I’d like the students to think but what I’d like them to do.

I won’t go into too much detail: I suspect some of the students may be reading this, and I don’t want to spoil the surprise! But I’m looking forward to class, and to getting to try another pedagogical experiment.

Machine Learning, Occam’s Razor, and Fundamental Physics

There’s a saying in physics, attributed to the famous genius John von Neumann: “With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.”

Say you want to model something, like some surprising data from a particle collider. You start with some free parameters: numbers in your model that aren’t decided yet. You then decide those numbers, “fixing” them based on the data you want to model. Your goal is for your model not only to match the data, but to predict something you haven’t yet measured. Then you can go out and check, and see if your model works.

The more free parameters you have in your model, the easier this can go wrong. More free parameters make it easier to fit your data, but that’s because they make it easier to fit any data. Your model ends up not just matching the physics, but matching the mistakes as well: the small errors that crop up in any experiment. A model like that may look like it’s a great fit to the data, but its predictions will almost all be wrong. It wasn’t just fit, it was overfit.

We have statistical tools that tell us when to worry about overfitting, when we should be impressed by a model and when it has too many parameters. We don’t actually use these tools correctly, but they still give us a hint of what we actually want to know, namely, whether our model will make the right predictions. In a sense, these tools form the mathematical basis for Occam’s Razor, the idea that the best explanation is often the simplest one, and Occam’s Razor is a critical part of how we do science.

So, did you know machine learning was just modeling data?

All of the much-hyped recent advances in artificial intelligence, GPT and Stable Diffusion and all those folks, at heart they’re all doing this kind of thing. They start out with a model (with a lot more than five parameters, arranged in complicated layers…), then use data to fix the free parameters. Unlike most of the models physicists use, they can’t perfectly fix these numbers: there are too many of them, so they have to approximate. They then test their model on new data, and hope it still works.

Increasingly, it does, and impressively well, so well that the average person probably doesn’t realize this is what it’s doing. When you ask one of these AIs to make an image for you, what you’re doing is asking what image the model predicts would show up captioned with your text. It’s the same sort of thing as asking an economist what their model predicts the unemployment rate will be when inflation goes up. The machine learning model is just way, way more complicated.

As a physicist, the first time I heard about this, I had von Neumann’s quote in the back of my head. Yes, these machines are dealing with a lot more data, from a much more complicated reality. They literally are trying to fit elephants, even elephants wiggling their trunks. Still, the sheer number of parameters seemed fishy here. And for a little bit things seemed even more fishy, when I learned about double descent.

Suppose you start increasing the number of parameters in your model. Initially, your model gets better and better. Your predictions have less and less error, your error descends. Eventually, though, the error increases again: you have too many parameters so you’re over-fitting, and your model is capturing accidents in your data, not reality.

In machine learning, weirdly, this is often not the end of the story. Sometimes, your prediction error rises, only to fall once more, in a double descent.

For a while, I found this deeply disturbing. The idea that you can fit your data, start overfitting, and then keep overfitting, and somehow end up safe in the end, was terrifying. The way some of the popular accounts described it, like you were just overfitting more and more and that was fine, was baffling, especially when they seemed to predict that you could keep adding parameters, keep fitting tinier and tinier fleas on the elephant’s trunk, and your predictions would never start going wrong. It would be the death of Occam’s Razor as we know it, more complicated explanations beating simpler ones off to infinity.

Luckily, that’s not what happens. And after talking to a bunch of people, I think I finally understand this enough to say something about it here.

The right way to think about double descent is as overfitting prematurely. You do still expect your error to eventually go up: your model won’t be perfect forever, at some point you will really overfit. It might take a long time, though: machine learning people are trying to model very complicated things, like human behavior, with giant piles of data, so very complicated models may often be entirely appropriate. In the meantime, due to a bad choice of model, you can accidentally overfit early. You will eventually overcome this, pushing past with more parameters into a model that works again, but for a little while you might convince yourself, wrongly, that you have nothing more to learn.

(You can even mitigate this by tweaking your setup, potentially avoiding the problem altogether.)

So Occam’s Razor still holds, but with a twist. The best model is simple enough, but no simpler. And if you’re not careful enough, you can convince yourself that a too-simple model is as complicated as you can get.

Image from Astral Codex Ten

I was reminded of all this recently by some articles by Sabine Hossenfelder.

Hossenfelder is a critic of mainstream fundamental physics. The articles were her restating a point she’s made many times before, including in (at least) one of her books. She thinks the people who propose new particles and try to search for them are wasting time, and the experiments motivated by those particles are wasting money. She’s motivated by something like Occam’s Razor, the need to stick to the simplest possible model that fits the evidence. In her view, the simplest models are those in which we don’t detect any more new particles any time soon, so those are the models she thinks we should stick with.

I tend to disagree with Hossenfelder. Here, I was oddly conflicted. In some of her examples, it seemed like she had a legitimate point. Others seemed like she missed the mark entirely.

Talk to most astrophysicists, and they’ll tell you dark matter is settled science. Indeed, there is a huge amount of evidence that something exists out there in the universe that we can’t see. It distorts the way galaxies rotate, lenses light with its gravity, and wiggled the early universe in pretty much the way you’d expect matter to.

What isn’t settled is whether that “something” interacts with anything else. It has to interact with gravity, of course, but everything else is in some sense “optional”. Astroparticle physicists use satellites to search for clues that dark matter has some other interactions: perhaps it is unstable, sometimes releasing tiny signals of light. If it did, it might solve other problems as well.

Hossenfelder thinks this is bunk (in part because she thinks those other problems are bunk). I kind of do too, though perhaps for a more general reason: I don’t think nature owes us an easy explanation. Dark matter isn’t obligated to solve any of our other problems, it just has to be dark matter. That seems in some sense like the simplest explanation, the one demanded by Occam’s Razor.

At the same time, I disagree with her substantially more on collider physics. At the Large Hadron Collider so far, all of the data is reasonably compatible with the Standard Model, our roughly half-century old theory of particle physics. Collider physicists search that data for subtle deviations, one of which might point to a general discrepancy, a hint of something beyond the Standard Model.

While my intuitions say that the simplest dark matter is completely dark, they don’t say that the simplest particle physics is the Standard Model. Back when the Standard Model was proposed, people might have said it was exceptionally simple because it had a property called “renormalizability”, but these days we view that as less important. Physicists like Ken Wilson and Steven Weinberg taught us to view theories as a kind of series of corrections, like a Taylor series in calculus. Each correction encodes new, rarer ways that particles can interact. A renormalizable theory is just the first term in this series. The higher terms might be zero, but they might not. We even know that some terms cannot be zero, because gravity is not renormalizable.

The two cases on the surface don’t seem that different. Dark matter might have zero interactions besides gravity, but it might have other interactions. The Standard Model might have zero corrections, but it might have nonzero corrections. But for some reason, my intuition treats the two differently: I would find it completely reasonable for dark matter to have no extra interactions, but very strange for the Standard Model to have no corrections.

I think part of where my intuition comes from here is my experience with other theories.

One example is a toy model called sine-Gordon theory. In sine-Gordon theory, this Taylor series of corrections is a very familiar Taylor series: the sine function! If you go correction by correction, you’ll see new interactions and more new interactions. But if you actually add them all up, something surprising happens. Sine-Gordon turns out to be a special theory, one with “no particle production”: unlike in normal particle physics, in sine-Gordon particles can neither be created nor destroyed. You would never know this if you did not add up all of the corrections.

String theory itself is another example. In string theory, elementary particles are replaced by strings, but you can think of that stringy behavior as a series of corrections on top of ordinary particles. Once again, you can try adding these things up correction by correction, but once again the “magic” doesn’t happen until the end. Only in the full series does string theory “do its thing”, and fix some of the big problems of quantum gravity.

If the real world really is a theory like this, then I think we have to worry about something like double descent.

Remember, double descent happens when our models can prematurely get worse before getting better. This can happen if the real thing we’re trying to model is very different from the model we’re using, like the example in this explainer that tries to use straight lines to match a curve. If we think a model is simpler because it puts fewer corrections on top of the Standard Model, then we may end up rejecting a reality with infinite corrections, a Taylor series that happens to add up to something quite nice. Occam’s Razor stops helping us if we can’t tell which models are really the simple ones.

The problem here is that every notion of “simple” we can appeal to here is aesthetic, a choice based on what makes the math look nicer. Other sciences don’t have this problem. When a biologist or a chemist wants to look for the simplest model, they look for a model with fewer organisms, fewer reactions…in the end, fewer atoms and molecules, fewer of the building-blocks given to those fields by physics. Fundamental physics can’t do this: we build our theories up from mathematics, and mathematics only demands that we be consistent. We can call theories simpler because we can write them in a simple way (but we could write them in a different way too). Or we can call them simpler because they look more like toy models we’ve worked with before (but those toy models are just a tiny sample of all the theories that are possible). We don’t have a standard of simplicity that is actually reliable.

From the Wikipedia page for dark matter halos

There is one other way out of this pickle. A theory that is easier to write down is under no obligation to be true. But it is more likely to be useful. Even if the real world is ultimately described by some giant pile of mathematical parameters, if a simple theory is good enough for the engineers then it’s a better theory to aim for: a useful theory that makes peoples’ lives better.

I kind of get the feeling Hossenfelder would make this objection. I’ve seen her argue on twitter that scientists should always be able to say what their research is good for, and her Guardian article has this suggestive sentence: “However, we do not know that dark matter is indeed made of particles; and even if it is, to explain astrophysical observations one does not need to know details of the particles’ behaviour.”

Ok yes, to explain astrophysical observations one doesn’t need to know the details of dark matter particles’ behavior. But taking a step back, one doesn’t actually need to explain astrophysical observations at all.

Astrophysics and particle physics are not engineering problems. Nobody out there is trying to steer a spacecraft all the way across a galaxy, navigating the distribution of dark matter, or creating new universes and trying to make sure they go just right. Even if we might do these things some day, it will be so far in the future that our attempts to understand them won’t just be quaint: they will likely be actively damaging, confusing old research in dead languages that the field will be better off ignoring to start from scratch.

Because of that, usefulness is also not a meaningful guide. It cannot tell you which theories are more simple, which to favor with Occam’s Razor.

Hossenfelder’s highest-profile recent work falls afoul of one or the other of her principles. Her work on the foundations of quantum mechanics could genuinely be useful, but there’s no reason aside from claims of philosophical beauty to expect it to be true. Her work on modeling dark matter is at least directly motivated by data, but is guaranteed to not be useful.

I’m not pointing this out to call Hossenfelder a hypocrite, as some sort of ad hominem or tu quoque. I’m pointing this out because I don’t think it’s possible to do fundamental physics today without falling afoul of these principles. If you want to hold out hope that your work is useful, you don’t have a great reason besides a love of pretty math: otherwise, anything useful would have been discovered long ago. If you just try to model existing data as best you can, then you’re making a model for events far away or locked in high-energy particle colliders, a model no-one else besides other physicists will ever use.

I don’t know the way through this. I think if you need to take Occam’s Razor seriously, to build on the same foundations that work in every other scientific field…then you should stop doing fundamental physics. You won’t be able to make it work. If you still need to do it, if you can’t give up the sub-field, then you should justify it on building capabilities, on the kind of “practice” Hossenfelder also dismisses in her Guardian piece.

We don’t have a solid foundation, a reliable notion of what is simple and what isn’t. We have guesses and personal opinions. And until some experiment uncovers some blinding flash of new useful meaningful magic…I don’t think we can do any better than that.

Amplitudes 2022 Retrospective

I’m back from Amplitudes 2022 with more time to write, and (besides the several papers I’m working on) that means writing about the conference! Casual readers be warned, there’s no way around this being a technical post, I don’t have the space to explain everything!

I mostly said all I wanted about the way the conference was set up in last week’s post, but one thing I didn’t say much about was the conference dinner. Most conference dinners are the same aside from the occasional cool location or haggis speech. This one did have a cool location, and a cool performance by a blind pianist, but the thing I really wanted to comment on was the setup. Typically, the conference dinner at Amplitudes is a sit-down affair: people sit at tables in one big room, maybe getting up occasionally to pick up food, and eventually someone gives an after-dinner speech. This time the tables were standing tables, spread across several rooms. This was a bit tiring on a hot day, but it did have the advantage that it naturally mixed people around. Rather than mostly talking to “your table”, you’d wander, ending up at a new table every time you picked up new food or drinks. It was a good way to meet new people, a surprising number of which in my case apparently read this blog. It did make it harder to do an after-dinner speech, so instead Lance gave an after-conference speech, complete with the now-well-established running joke where Greta Thunberg tries to get us to fly less.

(In another semi-running joke, the organizers tried to figure out who had attended the most of the yearly Amplitudes conferences over the years. Weirdly, no-one has attended all twelve.)

In terms of the content, and things that stood out:

Nima is getting close to publishing his newest ‘hedron, the surfacehedron, and correspondingly was able to give a lot more technical detail about it. (For his first and most famous amplituhedron, see here.) He still didn’t have enough time to explain why he has to use category theory to do it, but at least he was concrete enough that it was reasonably clear where the category theory was showing up. (I wasn’t there for his eight-hour lecture at the school the week before, maybe the students who stuck around until 2am learned some category theory there.) Just from listening in on side discussions, I got the impression that some of the ideas here actually may have near-term applications to computing Feynman diagrams: this hasn’t been a feature of previous ‘hedra and it’s an encouraging development.

Alex Edison talked about progress towards this blog’s namesake problem, the question of whether N=8 supergravity diverges at seven loops. Currently they’re working at six loops on the N=4 super Yang-Mills side, not yet in a form it can be “double-copied” to supergravity. The tools they’re using are increasingly sophisticated, including various slick tricks from algebraic geometry. They are looking to the future: if they’re hoping their methods will reach seven loops, the same methods have to make six loops a breeze.

Xi Yin approached a puzzle with methods from String Field Theory, prompting the heretical-for-us title “on-shell bad, off-shell good”. A colleague reminded me of a local tradition for dealing with heretics.

While Nima was talking about a new ‘hedron, other talks focused on the original amplituhedron. Paul Heslop found that the amplituhedron is not literally a positive geometry, despite slogans to the contrary, but what it is is nonetheless an interesting generalization of the concept. Livia Ferro has made more progress on her group’s momentum amplituhedron: previously only valid at tree level, they now have a picture that can accomodate loops. I wasn’t sure this would be possible, there are a lot of things that work at tree level and not for loops, so I’m quite encouraged that this one made the leap successfully.

Sebastian Mizera, Andrew McLeod, and Hofie Hannesdottir all had talks that could be roughly summarized as “deep principles made surprisingly useful”. Each took topics that were explored in the 60’s and translated them into concrete techniques that could be applied to modern problems. There were surprisingly few talks on the completely concrete end, on direct applications to collider physics. I think Simone Zoia’s was the only one to actually feature collider data with error bars, which might explain why I singled him out to ask about those error bars later.

Likewise, Matthias Wilhelm’s talk was the only one on functions beyond polylogarithms, the elliptic functions I’ve also worked on recently. I wonder if the under-representation of some of these topics is due to the existence of independent conferences: in a year when in-person conferences are packed in after being postponed across the pandemic, when there are already dedicated conferences for elliptics and practical collider calculations, maybe people are just a bit too tired to go to Amplitudes as well.

Talks on gravitational waves seem to have stabilized at roughly a day’s worth, which seems reasonable. While the subfield’s capabilities continue to be impressive, it’s also interesting how often new conceptual challenges appear. It seems like every time a challenge to their results or methods is resolved, a new one shows up. I don’t know whether the field will ever get to a stage of “business as usual”, or whether it will be novel qualitative questions “all the way up”.

I haven’t said much about the variety of talks bounding EFTs and investigating their structure, though this continues to be an important topic. And I haven’t mentioned Lance Dixon’s talk on antipodal duality, largely because I’m planning a post on it later: Quanta Magazine had a good article on it, but there are some aspects even Quanta struggled to cover, and I think I might have a good way to do it.