on doing Real work

a working theory

Real as in the opposite of bullshit, as in the opposite of slop, as in the opposite of hype or rhetoric or froth. Real as in true, real as in honest.

---

I’ve been in a somewhat constant low-to-medium-grade crashout all year about what I’ve been calling the Circus. The circus is startups run by new grads and college dropouts claiming to be at the cutting edge of “research”; the circus is Eliezer’s book, and AI safety people defending and disavowing it in turn; the circus is breathless predictions of a 24-month timeline to AI-driven nuclear war, an NYT writeup, an update a few months later that, actually, the headline claim was not what a reader should take from it, and never had been; the circus is so many — so many — “straight lines on a graph.” The circus is SF, the circus is Twitter, the circus is AI, the circus is a shit ton of money.

(If you don’t recognize the circus, I envy you.)

I don’t want to write or think too much about the circus, but the circus is real, in the sense that it involves a lot of people with a lot of power, even if its elements are mostly not, so how can I write about what is real without at least mentioning it?

Another dimension of real that I’ve been forced to think about recently: podcasters without math backgrounds writing blog posts “contributing” to RL theory, a GDM director of engineering claiming to have solved Navier-Stokes, both “assisted” by an LLM. The work itself, of course, is not real; their delusions of having done real work, on the other hand, are.

Expertise is real. Skill and taste and patience — research aptitude — cultivated by the structures of a graduate program are real. But I digress: my pro-academia manifesto will have to wait.

I remember in my first year trying to understand what could possibly motivate anyone to spend their lives studying problems that vanishingly small numbers of people were even capable of understanding, much less care about. I remember very clearly one night coming back from Simons, after some seminar or another; it was very cold, and the twilight was warm and blue as I came down the Li Ka Shing steps to Oxford. I remember feeling struck with an epiphany on those steps. Research is this: you make up a problem, and then you solve it. I remember immediately feeling very silly, for both the thought itself and the feeling of epiphany, but there it was, ringing in my head like a bell.

You can skip the first step by picking a problem someone else has made up, but somewhere along the way, someone has, in fact, made it up. There may be fundamental truths to the world, but the questions we choose to ask about them are of our own construction, and the ways we choose to answer them can be better or worse. Just because a problem is made up doesn’t mean it can’t also be real. Make up a problem, and then solve it.

---

Actually, let me disambiguate further. There’s real as a descriptor of actuality, as in a thing that exists in the world. There’s also real as a descriptor of quality, as in rigorous or substantive or internally consistent. Of course these two reals are related. Work should be both, as in, describe or create something that really exists, and to do it well. In fact, academia is pretty good at, or at least has some language for, what makes something real in those ways; in some sense, that is the entire premise of the academic institution, to train people and filter work product so that, collectively, we find some real phenomenon about the universe or about people. I think, though, that I’m in search of a secret, third real.

---

In November Ben tells me this — “realness” — is something I have been preoccupied with since the fall of my first year, which surprises me because I don’t remember that, or at least putting it in that language. Then I start to remember that first disorienting semester of one-on-ones, when I was just beginning to understand how to have a conversation (let alone working relationship) with him. One week, he suggested that I should compile a spreadsheet of known, large effect sizes. “Smoking causes cancer” — that’s real. “Seatbelts work” — that’s real, but the effect size is smaller (odds ratio of around 2, compared to smoking’s 9ish), so does that make it less real? Any Cochrane review that requires more than two seconds to parse — probably not real.

I didn’t complete this task in part because I was simply lazy and it seemed both boring and underspecified, but I also remember being skeptical that this exercise would illuminate anything to me about what kind of work I should be doing, how to make my work real or how to use my work to draw clearer boundaries around what was real, or not. Or, concretely, how it might ever turn into something that looks like a machine learning conference paper.

What I really wanted to figure out was how to do work that meant something, how to write papers not for the sake of writing papers but for — what?

I killed two projects this year, both of which I had thought about for many months (if not actively worked on), and were already more or less at the stage where I felt like I knew what the shape of the paper would have looked like. While the details differ slightly between the projects, the essential issue with both of them was that I felt that they were not going to shape up to be real, or at least not real enough.

I still think the basic questions that I was asking in these projects are conceptually strong — to what extent was [exciting technique trending among stats-y ML people] actually useful? What was the right objective for [method that recently got some attention], and what were the consequences of choosing such an objective?

They aren’t questions that lend themselves to nice quantitative answers (as in smoking and seatbelts), so when I say that I don’t think they were “real” it’s not that I can necessarily tell you anything about degree. What I mean is that, after sketching out candidate answers for those questions, I could not convince myself that they were worth answering. The stories for each sounded good and important, but the realness of my questions rested on premises that seemed increasingly shaky. Forget [technique], what were people even using predictions for, in the first place? Never mind the objective, when was [method] actually even feasible to run?

Let me revise again: in trying to answer my questions, I lost confidence that their existences were justified at all, and I could not convince myself to massage my negative results into paper-shaped “contributions.”

I know for a fact that we could have resolved those questions with high-quality conference papers (and in fact, variants of them actually have been, by other people). There was even some hard math that we could have done, if we had wanted to. I just didn’t really see the point.

---

The thing is, I am fundamentally a critic at heart. For almost any idea, I can tell you not just its limitations on its own terms, but also any number of reasons why it’s existentially pointless — why fixing those limitations are little more than (what in high school debate we called) rearranging chairs on the deck of the Titanic. I can tell you this for entire research areas and subfields, of course including and especially my own. One typical reason is simply the yawning chasm between knowledge production and (sorry) the real world. Analysis comes far more easily to me than action, or perhaps it’s that critiquing all the possible actions comes more easily to me than committing to one of them.

In October Brandon Taylor visits Berkeley as part of his book tour for Minor Black Figures, which came out this year.¹ One thing that feels important to mention is that he was, once, a biology Ph.D. student, and the subjects of his first books were, in order, biology Ph.D. students, dancers, MFA students. A second thing to know is that I speedran all three of those books the summer between second and third year, hot and sticky.

This new book is about a painter named Wyeth. Wyeth’s big issues are — though you’re not really supposed to say this — uncannily similar to what one might imagine some of Brandon Taylor’s big issues to be. Wyeth, who is black, has been stuck on what it means to make “black art” as opposed to just making art as an artist who happens to be black. Some of his early work had been picked up in the rush of mid-2020 “elevation of black voices”, which he feels ambivalent about, because that’s not what he had meant by the painting at all. He hasn’t painted in ages. Oh, yeah, and he’s also really obsessed with doing real work. Whatever that means.²

Meanwhile, his contemporaries produce work that Wyeth finds empty and soulless and visually uninteresting. He’s fixated on a group called MangoWave, a collective that seeks to “explore the Southeast Asian diasporic visual idiom.” He thinks not only that their art is bad but that they are sellouts, that they’re trading their skin color for art world cachet. And yet it works: they get writeups in arts journals and the New York Times, sold-out shows, flashy parties with sweaty crudites. He’s not not envious of them, but he’s also terrified of becoming like them, which in fact is the direct source of his work paralysis.

It’s not that I’m navigating anything so fraught as representation politics (though, yes! I am tired of writers in the style of MangoWave!). But when I go to that event I find myself doing the Leo DiCaprio meme. Brandon Taylor says that he had originally been planning to write a takedown of artists like MangoWave, but that it turns out that you can’t write a book that’s just snark. He says that being a hater, even if you’re actually right, is fundamentally uninteresting, and just being mean about other people’s work, even if you do it privately, doesn’t make you smarter or better. He says that Wyeth’s essential problem is that he doesn’t just have a cop in his head, he has every imagined audience and every response they would have and every response he would have to their responses in his head. He’s so hemmed in by the “discourse” of it all that he’s completely lost track of what he actually wants for himself.

After I get home from the event I look up all the reviews and interviews and the other bits of the song and dance that happens after a somewhat-buzzy book gets published. In Bookforum Brandon Taylor recounts the experience of seeing a critic saying about one of his prior book announcements, “Do we need another one of these books, eww.” He’s talking about novels but it might as well be about research papers:

It planted the question in my psyche. What is the reason a book should exist? What is the motivation to publish a book? If someone else were to ask me whether the fact that they wanted to publish was a good enough reason, I would say, yeah, of course. For myself, I was like, sure, it’s a good enough reason, but do I want to publish it for the right reasons?

Of course I read the book and what I find out is that Wyeth lives life. Wyeth doesn’t sell out. Wyeth makes art for himself. Wyeth wanted so bad to do something real that he started taking photographs of Union Square protests to use as references, because he thought that those were real, but ultimately couldn’t bring himself to paint them because he knew that the painting would have been simulating something that could have been real, rather than actually being real. I don’t think it’s much of a spoiler for me to say that, by the end of the novel, Wyeth’s triumph is that he finally does succeed in painting something real, and the one person he’s painting for recognizes it as such.

---

My theory of real is that a creative work is real if it has a good reason to exist. If its creator can look themselves in the eye and say with full honesty that they know why they’re making it, and that they like why they’re making it.

It’s a generous standard, if I do say so myself. There are a lot of possible good reasons for something to exist, and everyone has their own relationship to their work; far be it from me to declare those boundaries. Some reasons, though, that I personally find obviously bad: because it will make me look smart. Because this will gain a lot of citations. Because I need to show some output for this period of time. Because someone else has decided that this is an important problem.

The best reasons: I am writing this because I want X to be possible. Because I have a different perspective on Y. Or, even more basic: Because Z should be better-understood. I care about Z.

My prescriptive theory of real when it comes to paper writing, therefore, is simply that researchers should have a point of view, and the work should reflect that. That’s my hack for how to write real papers because I think that working towards your point of view is almost always a good reason to write a paper. It’s not to say that it needs to be overtly political or stated explicitly (in fact for most technical work it probably shouldn’t be either of those things). But if you are spending the time and energy on a paper about X or Y or Z, I sure hope that you care about it, that you believe in it.

When I think about the work I find compelling, vs. the work that I find annoying and makes me think Do we need another one of these [papers], eww, I think more often than not it’s because I can’t tell what the authors think should happen, I can’t tell what they cared about. It’s often possible to get something published that is real only in the first two senses, the existential and the evaluative, and indeed such papers are often good by any academic measure. They’re necessary — something that is real in my secret third sense, but not in the other more quantifiable ones, can’t be good work — but I think we can be more ambitious. Being real in all three ways can’t guarantee success, but I think it’s still a target worth aiming for.

There will always be the artists or researchers or whomever, who are making work you find questionable and maybe a little grifty, who nevertheless seem to be rewarded in ways that you are not, and Brandon Taylor says let them, worry about yourself, how they feel about their work is none of your business. As for the quagmire of, well, the critics and the circus and, crucially, the versions of them that live in your head, Brandon Taylor says you can’t please them so might as well just do what you want to do.

---

I’ve written these every year for the last three years, but everything feels weird and different this time around. At first I didn’t even especially want to write at all, which is mildly alarming given that this is the only really creative writing I feel entirely justified in doing all year. Maybe it’s that damn phone, the Instagram reels frying my already-threadbare attention span to a crisp; maybe it’s that usually by this time, I’ve been able to ski a day or five already, had my neurons temporarily rewired by the cold and the wet and the possibility of seriously injuring myself, but Crystal Mountain has only just opened for the season and my brother is on call all week.

But it’s also just been a weird and different year. I met more VCs and founders than researchers at NeurIPS this month (the circus is earnest anti-imposter-syndrome LinkedIn posts for startup people and VCs at NeurIPS — iykyk). For almost everyone in my BAIR cohort, this will be their last full year of being Ph.D. students (I still plan to graduate in 2027, extinction notwithstanding). It’s not actually possible to think and act and work like a first-year Ph.D. student forever, as I had so sweetly and optimistically thought in December 2022. Time is real, age is real.³

---

the moon is real also

---

Notable books

• Headshot by Rita Bullwinkel. The first book I read of 2025, and also the best. REAL. When it became a Pulitzer finalist (kind of a surprise nomination if my pulse on the litfic scene is accurate, which it really might not be) I was so happy!!!!!

• It Lasts Forever and Then It’s Over by Anne de Marcken. I read this twice in a row, back to back, around the time of Berkeley Ph.D. visit days. (Real!)

• I feel like I read a lot of books (4 of 12…) that were kind of bad this year :( I won’t name them here but it makes me sad :(

Art highlights

• Alysa Liu 2025 world champion!!!!

• SF Symphony: Carmina Burana, Gershwin & Ellington with Helene Grimaud, Crouching Tiger Hidden Dragon got my ass. Also Vienna Phil New World Symphony

• I made Kevin drive 4 hours to Fresno to see Hayato Sumino perform for the California music teachers’ association because his Stanford show was completely sold out. That guy can freakin play the piano!!

• Dance: SF Ballet Manon; Chloe Misseldine in Giselle; this crazy gala at the Liceu in Barcelona, where I saw Chloe again (made for Odile) but also all these other people I wouldn’t normally have seen — Sangeun Lee (a very tall and otherworldly Odette), Yuriko Kajiya (jumping in for Manon at the last minute on top of Mme Butterfly), Maria Khoreva (she did her first 16 Kitri fouettes without arms, wth??). Parsons for the 4th time, duh.

Art lowlights

• Bad jazz again, how is this possible — bassists cannot just pick up a cello the way violinists can swap to viola omg, they’re completely different instruments.

• “Red Carpet” by Hofesh Shechter, performed by the Paris Opera Ballet. What a waste of the POB dancers!!! The girl behind me saw my notes app crashout over my shoulder and validated me, thank god I wasn’t the only one.⁴

---

1. Of all places, the event is at the David Brower center, which I most strongly associate with a yearly gathering of economists/ statisticians/ computer scientists who actually want to talk to people from the other disciplines.

2. Also, the word “real” appears in the book way more than I would think is normal. Also, the title of Brandon Taylor’s first book was Real Life.

3. So many other things about Real I wanted to write, like actually I think there are really interesting questions about what makes something real 1 or real 2 in the academic sense and it’s not nearly as clean and tidy as I’ve made it out to be here, and also I don’t know if computer science is even a real science, and also maybe there is another way in which art can be real, as in Real Life and Headshot and oil paintings of fruit or maybe it’s just real 1, and also I feel like the ocean is real and the sky is real and being outside is real! (Also, completely fair to say that none of this laptop job bs is “real work.”) But this thing ended up being so obscenely long already.

Also, to the people who ended up on this list expecting AI takes or actual research updates sorry if you have made it down here, you can find me on twitter or Reboot for that. Feel free to unsubscribe lol

4. More detailed crashout/review on this story highlight if you tap past the musikverein ones which also, lol