This is Jessica. Last week the 2024 Turing award winners were announced. It went to pioneers in reinforcement learning Andrew Barto and Richard Sutton.
Academic computer scientists get excited about Turing awards. On the positive side, people like to see hard work getting recognized and celebrate their peers or mentors. On the cringier side, award announcements give people an excuse to casually boast about their education on the awarded topic or their proximity to the winners. And suddenly everyone is very interested in hearing any pearls of wisdom the winners have to share about their specific area but also the broader research landscape that their contributions intersect.
In this case, some bad advice once given by one of the winners trickled up to social media. It seems Sutton once gave some truly horrible advice in a talk back in 2022. There was a slide about how to go about “keeping one’s eyes on the prize of understanding intelligence” if you want to be ambitious in AI research. To do this, he suggested not being distracted by applications. Or domain knowledge. Or unnecessary problem variations. Or AI privacy, explainability, or safety.
This advice is questionable for many reasons. Do we really want to encourage the already palpable preoccupation with “artificial general intelligence” or proving progress on benchmarks of questionable relation to real world tasks they supposedly embody? How does one not worry at all about domain knowledge, given that many real world decision pipelines where AI is deployed still depend on human oversight? And how do you reconcile an inventor of algorithms that are vulnerable to unintended behavior through “reward hacking” and the like telling people not to waste their time on topics like AI risk management or making these models interpretable to humans?
It’s probably not what we should be telling the younger generation. But at the same time, it lines up with a certain preoccupation with conceptual purity that I see often among computer scientists, where the closer your work is to math, the more brilliant you must be. There’s a certain glamour associated with solving an old math problem (e.g., he cracked the sunflower lemma!) Not surprisingly, I see this most in CS theorists, but I don’t think it’s at all specific to theory. Conceptual purity also seems to carry weight in areas like programming languages, privacy and security, and ML. More generally, it leads to vague but recurring questions that come up in faculty hiring or grant proposal review, like “How is this advancing ‘core CS’?”
It’s interesting how sharply this aesthetic contrasts with ideals from some of the more applied areas of computer science. In systems, it’s more about conquering messes at scale. If your code somehow tames something ugly and complex into submission, then you’ve earned a sort of valor, not unlike a firefighter putting out some raging blaze or a lawn crew bringing out the power tools to hack away a bunch of dense brush. This is the impression I get often when listening to my systems colleagues discuss faculty candidates or research in their area. In human computer interaction, there can be a preoccupation with context, with very specific things and what can’t be formalized. If you want to write a paper applying anthropological methods to understand the role of mobile interfaces in witchcraft in some small rural community, go for it. I guess on the bright side, when you combine a bunch of individually perverse aesthetics in a single discipline, things feel more balanced.
In this case, it’s unfortunate Sutton once advocated for ignoring applications and safety and the like, because he has often spoken out against exactly the kind of arrogance that tends to accompany this kind of preoccupation with conceptual purity. But I guess the purity fetish can run deep. I know I have been guilty of it myself sometimes, e.g., when I find myself pointing to examples from the more applied areas of CS when I want to make a point about research with questionable generalizability. I’ve seen faculty use the term “pure CS” and then catch themself and ask whether it’s the right ideal. Part of this is probably what we’re conditioned to appreciate in computer science education (e.g., clever generic solutions over practical or contextually-specific ones). But perhaps we’ve selected into it as well.
I think it is also a reaction to the perceived threat of interdisciplinarity. Computer science appears to have subsumed many other fields, but this leads to a kind of threat from within that whatever it was historically will disappear. Focusing on the so-called conceptual “core” can seem to protect academic CS from being subsumed by the rapidly bleeding edges. On the flip side, the undeniable relevance of computer science to the world perhaps gives contributions to “pure CS” some kind of penultimate value that requires no rationalization as they might in other fields. One can focus on conceptual purity without the risk of being judged irrelevant to the world.
I don’t know if it is still like this, but there used to be some CS departments in a university’s college of arts and science and some CS departments in engineering. This distinction seems related to the purity divide.
At Northwestern where I am there is an increasing demand for computer science majors among students in the college of arts and science, and for AI or data science degrees to be grantable by other units on campus (like the business school or statistics). The demand definitely makes people nervous that forces outside the computer science department will step in and provide some kind of lesser substitute for the “real” computer science.
Nothing new.
Early on, many CS departments (like mine at Penn State, College of Science) grew from math departments (, at least as many grew from EE,and some schools have gone back and forth between {EE, CS} and EECS. My PSU department later moved to Engineering & reorganized several times.
I think Pitt’s came from Library Science, and I recall U Michigan had 3 CS departments from math, EE and business.
Some evolution seems akin to what happens when engineering colleges decide service course taught by statistics departments don’t match their needs, so they create their own.
In 1990s, I sat in on a terrific introductory CS course at Princeton taught in operations research dept, said to be more populat than the equivalent taught by CS. So, that’s life in dynamic interdisciplinary fields.
I do wonder whether the tendency to distil everything down to this kind of deterministic process is the common factor among a lot of bad behaviours in CS when applied to real-world settings. You can hide behind a veneer of “this is just what the data say” or “this is the best model” and ignore the consequences of implementation even when your work can directly lead to harm. You could argue that much of the current AI landscape is driven by this kind of thinking.
Is this saying that researchers in universities (outside China, of course) should vet their proposed research for political and social acceptability, and only pursue research which is politically acceptable, in their eyes? An alternative would be for some researchers to specialise in investigating the possibilities of technologies, for researchers specialised in the relevant areas to investigate the possible social conequences of the various options revealed, and for governments to decide which options to promote, which to discourage, and which to disallow. I note that various disturbances and following hearings appear to show at least some topics on which the prevailing political opinions in Ivy League universities are not identical to those in the electorate at large.
Or maybe only a few will make big progress in AI research, whereas anyone can do the “AI privacy, explainability, or safety” stuff. Indeed, the latter seem seems to be the favored domain of those who want an easy way to capitalize on the AI craze.
So while it might be bad advice if applied to all AI researchers, it’s perfectly fine if applied to the few. It’s a big world and there’s plenty of room in it for people who concentrate on their core contribution as well as those who don’t.
Or it’s just a rational thing to tell your students if you understand what gets respect in computer science.
You say that like “respect” is divorced from anything important in the real world.
Most of computer science, like most of statistics, is extraordinarily ephemeral. The window in which anyone will care about it is measured in years if you’re very lucky. For most efforts it’s measured in seconds.
Occasionally though, something transcends the throwaway nature of computer science (and statistics) to become a permanent contribution. People will still care, for example, about a proof of P=NP centuries from now.
Some people want to concentrate all their effort on that sort of thing. Since the other stuff can and regularly is done by almost anyone, nothing is lost thereby.
There is, of course, a near-equivalent but much older divide in statistics between those who think it math and those who think it is applied math. In math itself, the purists won long ago, and the offspring “Applied Math” still lives, but as a different discipline, usually joint with engineering departments.
Stephen Senn: “A theoretical statistician knows all about measure theory but has never seen a measurement whereas the actual use of measure theory by the applied statistician is a set of measure zero.”
That’s a great example. It may be that the dividing line in statistics articles is the simple word “σ-algebra.” If the article has it, it’s really just math.
Grumble. It’s cutesy, but “measure theory” isn’t about (what we’d call) “measurement” at all. It’s about functions on abstract spaces that have certain properties that relate to _distances between things_ not _magnitudes of measured values_.
Measure theory is actually pretty neat. (I watched the first half of the first lecture in an MIT course thereon.) Functions that at first glance couldn’t possibly act as distance measurements between points in such spaces actually do. Surprisingly fun.
I didn’t say measure theory wasn’t cool. I like a Lebesgue integral as much as the next guy. But if we were to survey the world’s applied statisticians and ask them to define one, much less how they might use one, you’d have a set of measure 0. (That’s a terrible joke of course, but it converges to a funny joke eventually.)
I’m not sure I agree with this interpretation of the slide, although I agree it’s bad advice. In fact I just think this post understates how bad and strange the advice is.
For example he writes “don’t be distracted by optimality, regret, or variations thereof”. This is saying ignore basically all theoretical CS as it relates to AI. (So, make sure to never get distracted by learning anything about multi-armed bandits. Also, you should make sure to tear out the pages from Sutton’s book where he proves convergence of various RL algorithms.)
Likewise don’t be distracted by inverse RL or multi-agent systems etc etc — ruling out whole subfields, which aren’t particularly interdisciplinary or focused on applications.
Of course he does also say ignore domain knowledge (his well known “Bitter Lesson”) and ignore social impact and ignore applications. So if you’re interested in recommender systems or robots because you want to make money, this is a waste of time…
He’s basically saying the entire mainstream of AI research, including all theory and almost all applications, including both very conceptually pure CS and interdisciplinary stuff, is a distraction. I guess I assume it must just be intended as a provocation.
Oops, I replied but its below, not in this thread.
>I guess I assume it must just be intended as a provocation.
I agree. Like many statements that are intended primarily to be provocative, it ends up making little sense. This one reminded me of other faux pas I’ve witnessesd when senior researchers have been invited to give talks. It’s like the pressure to say something visionary often brings out weird, offensive claims instead.
Jessica:
Yes, as I wrote once in another context, if you’re gonna be provocative, you shouldn’t be surprised if people get provoked.
He’s also decided to become a heavy poster on social media. It seems like sometimes for academics with big ideas and big personalities, this can tip them over the edge.
“it ends up making little sense. ”
I didn’t read it all that carefully, but at fist glance, I thought he was basically banging his own drum, namely that reinforcement learning is the best way to make progress towards “AI”. As a dyed-in-the-wool believer in symbolic processing* being the core of intellegence (the very first thing I learned in Minsky’s AI seminar in 1972 was that Skinner’s main intellectual contribution was in that his students made great circus animal trainers), I disagree, of course, but “AI” has turned into a hodge-podge of folks with essentially religious beliefs that some particular form of mindless processing will lead to “intelligence”, without any understanding of, or even explication of the very basics of what we ought to mean by, “intelligence”.
I tried to write a response or two to this post, but found it hard. If you look at, for example, the computational complexity subset of computer science, those blokes are quite rightly up to their ears in mathematical rigor on top of mathematical rigor. There’s nothing wrong whatsoever with what those guys are doing, and the mathematical rigor is absolutely required. The programming language blokes have figured out all sorts of kewl stuff, but programming languages in the real world tend to be social phenomena with the “community” (or originators) of a particular language taking the bit in their teeth and running with it. I love and use Python, but it’s a horrifically uninformed reinvention of Lisp, with enough of the good things (an interpreter + REPL, symbol tables) to make it useful, but also reinventing and retaining most of the mistakes that Sussman and Steele figured out and fixed. Whatever, Except for AI, it seems to me that Computer Science is doing just fine…
Whatever, keep up giving everyone hell. We all deserve it at some level.
*: There’s an overview article on conciousness as something not just human in the 21 February 2025 Science (“Evaluaing animal consciousless”), and, grumble, there is no hint that symbolic reasoning may or may not have something to do with conciousness. How can Ms. Pipa (a neighborhood cat who comes by for treats and snacks) be “concious” of David and Masako without a symbol for “human”???
I’ve been conciously and concientiously avoiding scientific discussions of conciousness on the grounds that reading it would result in my going balistic way to often, but I really need to organize my thoughts thereon based on what they’re trying to do. I get the impression that they are finally realizing that simple stimulu/response circuits isn’t conciousness…
In the linked slide, the computer science professor characterizes applications, domain knowledge, optimality, privacy, sustainability, and safety as “incidental things that could absorb all your energy” if you want to “be ambitious in your AI research.”
I wonder if he was forgetting about the division of labor. Different people can contribute in different ways. It could be that this guy’s contributions were such that applications, domain knowledge, etc., were unnecessary for him to do his research. But for others that will not be the case.
I’m ambitious in my statistics research, and applications, domain knowledge, etc., have been very important to my work–not “distractions” at all! But other people have made equally important contributions of their own without getting deep into applications. It would be fine for me to give advice of the form, “For me, it would be a disaster to drift too far from applications and domain knowledge,” but it would me a mistake for me to aim this advice at everybody.
This guy could still be a great adviser, though, because students self-select, so I’d hope that students with strong applied interests would find other supervisors to work with.