Excellent article. Surprised this isn’t more upvoted/commented upon. Trying to rectify the lack of comments and looking forward to the rest of the sequence, especially the mind/brain reverse engineering aspects.
The title is worse than useless, the summary isn’t very interesting, the typos don’t help (about half the “it’s” are wrong), and I’m not sure the long digressions into the glass-cannon section adds all that much. Like many LW posts, there is a definite “this letter is long because I didn’t have the time to make it short” feel which makes people tldr off, assuming they didn’t bounce off the title entirely. (I didn’t even read it when it was posted because I rolled my eyes at the title and went “oh great, what on earth is this, more doula-of-life-and-death and circling bullshit?”, skipped it, and didn’t notice the author until days later when Cannell linked it in a comment.)
I pretty much agree—my other title idea was “Why deep learning?”, and it’s based partly on a hoge-poge of notes. My attempts to spice it up a bit didn’t quite work, and the core isn’t something most on LW would find interesting. But that’s all fine, it fulfills a specific niche and back-link reference purpose.
I think after the last 6 years, people would be much more interested in it than they were in BULM which seemed highly speculative and improbable at the time, and it’s an important enough topic that it deserves good writeups. The world is not done with DL or scaling, not by a long shot, and the topic will be interested in most on LW soon enough.
Thanks—this mostly is background material for understanding what new developments we can expect in DL, and what remains to exceed the brain in bayes-efficiency. The later articles will be more direct and focused.
Excellent article. Surprised this isn’t more upvoted/commented upon. Trying to rectify the lack of comments and looking forward to the rest of the sequence, especially the mind/brain reverse engineering aspects.
The title is worse than useless, the summary isn’t very interesting, the typos don’t help (about half the “it’s” are wrong), and I’m not sure the long digressions into the glass-cannon section adds all that much. Like many LW posts, there is a definite “this letter is long because I didn’t have the time to make it short” feel which makes people tldr off, assuming they didn’t bounce off the title entirely. (I didn’t even read it when it was posted because I rolled my eyes at the title and went “oh great, what on earth is this, more doula-of-life-and-death and circling bullshit?”, skipped it, and didn’t notice the author until days later when Cannell linked it in a comment.)
I pretty much agree—my other title idea was “Why deep learning?”, and it’s based partly on a hoge-poge of notes. My attempts to spice it up a bit didn’t quite work, and the core isn’t something most on LW would find interesting. But that’s all fine, it fulfills a specific niche and back-link reference purpose.
I think after the last 6 years, people would be much more interested in it than they were in BULM which seemed highly speculative and improbable at the time, and it’s an important enough topic that it deserves good writeups. The world is not done with DL or scaling, not by a long shot, and the topic will be interested in most on LW soon enough.
Thanks—this mostly is background material for understanding what new developments we can expect in DL, and what remains to exceed the brain in bayes-efficiency. The later articles will be more direct and focused.