I appreciate this idea of territory-expanding research.
While I agree we should be aiming higher and for, say, deep novelty, I'm not convinced higher ambition is positively correlated with failure. The opposite seems true! I see in the screenshot that Ilya authored a winning test-of-time paper *every single year* of the modern era. Einstein published 4 absolutely seminal papers in *7 months* during his famous "miracle year" of 1905 (all of which expanded huge territory in quite different directions). This phenomenon of repeat success has also been noted in other domains[1].
This reminds me of Fei-Fei Li's autobiography, "The Worlds I See":
"'Well, Fei-Fei…' he began, choosing his words carefully. 'Everyone agrees that data has a role to play, of course, but…' He paused for a moment, then continued. 'Frankly, I think you’ve taken this idea way too far.' I took a shallow breath. 'The trick to science is to grow with your field. Not to leap so far ahead of it.'... A frightening idea was beginning to sink in: that I’d taken a bigger risk than I realized, and it was too late to turn back."
I’ve felt this for a really, really long time. 3-4 years ago (halfway through the PhD, and only after I had published in a sufficiently prestigious journal—and in my field, one is enough to graduate) I decided to follow this advice; I went for a risky idea and wrote something up.
I submitted it to NeurIPS and got really terrible scores! Part of it was quality, definitely. But the reviewers’ main issues were that no benchmarks had been surpassed; no obvious domain was going to be incrementally improved.
Plus, I am fairly confident that submitting to other venues won’t change anything—it’s just not “publishable” research, whatever that means these days. Yet I really think it’s more interesting than my prior work.
My only comfort is that I’m a theorist in an experimental field (neuroscience), and it’s fine—not great, but fine—if I don’t publish for a couple years. If I want job security eventually, though, this is still a risky and probably dumb move, and I imagine it must be an order of magnitude worse in ML.
Really makes me wonder how to address the bigger publish-or-perish problem. The social psychologist Adam Mastroianni seems to have an interesting model (he’s got a substack called Experimental History)—curious to see how that will work out, and whether other people have any thoughts.
> Part of it was quality, definitely. But the reviewers’ main issues were that no benchmarks had been surpassed; no obvious domain was going to be incrementally improved.
This might be my biggest gripe with publishing these days. To publish, one must run their own, closed-door benchmarks, where they cherry-pick results to make their method _appear_ more effective. Their method, one they are familiar with, can be specifically optimized for the task at hand, while simultaneously casting the competition in a negative light by presenting it in its worst form.
And its frustrating! Because I feel like I constantly see headlines to the effect of, "_______ was just released, and it beats everything with half the compute 🤯🤯🤯" -- just for me to never hear of it again. Why? Turns out "in practice" it doesn't actually perform all that better, so it dies. Endless up and down hype.
I think sometimes, its ok to publish things that are interesting for the sake of them being interesting, even if they don't beat everything all the time.
How does the idea of “dreaming bigger” compare with serendipitous research? Aren’t early-stage ideas often extremely fragile? For instance, in OpenAI’s early days, there was a seemingly trivial project -- a Reddit chatbot that was just a language model trained on Reddit data. If the team had filtered for only the “most important” or “time-safe” research directions, something like that likely wouldn’t have happened.
Do you think this highlights a deeper difference between what defines great research versus great product work? Research tends to optimize for significance and rigor, while great product work often emerges from small, easy-to-overlook experiments. Is it possible that the latter depends more on exploring fragile ideas before their value becomes obvious?
You might like this quote from Christopher Alexander on aiming higher:
"In my life as an architect, I find that the single thing which inhibits young professionals, new students most severely, is their acceptance of standards that are too low. If I ask a student whether her design is as good as Chartres, she often smiles tolerantly at me as if to say, “Of course not, that isn’t what I am trying to do. . . . I could never do that.”
Then, I express my disagreement, and tell her: “That standard must be our standard. If you are going to be a builder, no other standard is worthwhile. That is what I expect of myself in my own buildings, and it is what I expect of my students.”
Gradually, I show the students that they have a right to ask this of themselves, and must ask this of themselves. Once that level of standard is in their minds, they will be able to figure out, for themselves, how to do better, how to make something that is as profound as that.
Two things emanate from this changed standard. First, the work becomes more fun. It is deeper, it never gets tiresome or boring, because one can never really attain this standard. One’s work becomes a lifelong work, and one keeps trying and trying. So it becomes very fulfilling, to live in the light of a goal like this.
But secondly, it does change what people are trying to do. It takes away from them the everyday, lower-level aspiration that is purely technical in nature, (and which we have come to accept) and replaces it with something deep, which will make a real difference to all of us that inhabit the earth."
I appreciate this idea of territory-expanding research.
While I agree we should be aiming higher and for, say, deep novelty, I'm not convinced higher ambition is positively correlated with failure. The opposite seems true! I see in the screenshot that Ilya authored a winning test-of-time paper *every single year* of the modern era. Einstein published 4 absolutely seminal papers in *7 months* during his famous "miracle year" of 1905 (all of which expanded huge territory in quite different directions). This phenomenon of repeat success has also been noted in other domains[1].
1. https://blakemasters.tumblr.com/post/23435743973/peter-thiels-cs183-startup-class-13-notes
This reminds me of Fei-Fei Li's autobiography, "The Worlds I See":
"'Well, Fei-Fei…' he began, choosing his words carefully. 'Everyone agrees that data has a role to play, of course, but…' He paused for a moment, then continued. 'Frankly, I think you’ve taken this idea way too far.' I took a shallow breath. 'The trick to science is to grow with your field. Not to leap so far ahead of it.'... A frightening idea was beginning to sink in: that I’d taken a bigger risk than I realized, and it was too late to turn back."
Yeah, it’s definitely not just AI, but science in general—maybe.
https://www.nature.com/articles/s41586-022-05543-x
Oops, accidentally hit send before I was done lol
I’ve felt this for a really, really long time. 3-4 years ago (halfway through the PhD, and only after I had published in a sufficiently prestigious journal—and in my field, one is enough to graduate) I decided to follow this advice; I went for a risky idea and wrote something up.
I submitted it to NeurIPS and got really terrible scores! Part of it was quality, definitely. But the reviewers’ main issues were that no benchmarks had been surpassed; no obvious domain was going to be incrementally improved.
Plus, I am fairly confident that submitting to other venues won’t change anything—it’s just not “publishable” research, whatever that means these days. Yet I really think it’s more interesting than my prior work.
My only comfort is that I’m a theorist in an experimental field (neuroscience), and it’s fine—not great, but fine—if I don’t publish for a couple years. If I want job security eventually, though, this is still a risky and probably dumb move, and I imagine it must be an order of magnitude worse in ML.
Really makes me wonder how to address the bigger publish-or-perish problem. The social psychologist Adam Mastroianni seems to have an interesting model (he’s got a substack called Experimental History)—curious to see how that will work out, and whether other people have any thoughts.
> Part of it was quality, definitely. But the reviewers’ main issues were that no benchmarks had been surpassed; no obvious domain was going to be incrementally improved.
This might be my biggest gripe with publishing these days. To publish, one must run their own, closed-door benchmarks, where they cherry-pick results to make their method _appear_ more effective. Their method, one they are familiar with, can be specifically optimized for the task at hand, while simultaneously casting the competition in a negative light by presenting it in its worst form.
And its frustrating! Because I feel like I constantly see headlines to the effect of, "_______ was just released, and it beats everything with half the compute 🤯🤯🤯" -- just for me to never hear of it again. Why? Turns out "in practice" it doesn't actually perform all that better, so it dies. Endless up and down hype.
I think sometimes, its ok to publish things that are interesting for the sake of them being interesting, even if they don't beat everything all the time.
How does the idea of “dreaming bigger” compare with serendipitous research? Aren’t early-stage ideas often extremely fragile? For instance, in OpenAI’s early days, there was a seemingly trivial project -- a Reddit chatbot that was just a language model trained on Reddit data. If the team had filtered for only the “most important” or “time-safe” research directions, something like that likely wouldn’t have happened.
Do you think this highlights a deeper difference between what defines great research versus great product work? Research tends to optimize for significance and rigor, while great product work often emerges from small, easy-to-overlook experiments. Is it possible that the latter depends more on exploring fragile ideas before their value becomes obvious?
Thanks for the post, well said. Not in AI, but I think these ideas apply just as well to biotech and most life science research as well
You might like this quote from Christopher Alexander on aiming higher:
"In my life as an architect, I find that the single thing which inhibits young professionals, new students most severely, is their acceptance of standards that are too low. If I ask a student whether her design is as good as Chartres, she often smiles tolerantly at me as if to say, “Of course not, that isn’t what I am trying to do. . . . I could never do that.”
Then, I express my disagreement, and tell her: “That standard must be our standard. If you are going to be a builder, no other standard is worthwhile. That is what I expect of myself in my own buildings, and it is what I expect of my students.”
Gradually, I show the students that they have a right to ask this of themselves, and must ask this of themselves. Once that level of standard is in their minds, they will be able to figure out, for themselves, how to do better, how to make something that is as profound as that.
Two things emanate from this changed standard. First, the work becomes more fun. It is deeper, it never gets tiresome or boring, because one can never really attain this standard. One’s work becomes a lifelong work, and one keeps trying and trying. So it becomes very fulfilling, to live in the light of a goal like this.
But secondly, it does change what people are trying to do. It takes away from them the everyday, lower-level aspiration that is purely technical in nature, (and which we have come to accept) and replaces it with something deep, which will make a real difference to all of us that inhabit the earth."