4 ms·

I heard a great quip about the way people misreason about AI capabilities. Announcement: "AI can play chess!" Public: "People can play chess. And AI can play

by ethbr0 4y ago

I heard a great quip about the way people misreason about AI capabilities. Announcement: "AI can play chess!" Public: "People can play chess. And AI can play chess. People can also X. Therefore AI can also X." ... Ignoring the underlying nature of the chess problem and how it was different than other problems' structures. Hands are the same. You can image-bash together faces from examples and get something mostly-right simply through pattern copying. You cannot do the same with hands, because rendering them plausibly requires at least intuition and approximation of inverse kinematics -- something the recent set of image generative AI didn't include. Which isn't to say it can't, simply that the "hands problem" is unlike "the face problem."

ElFitz 4y ago
For some reason, the phrasing, the structure, the rhythm, they all very strongly remind me of Alan Watts’. Quite amusing and surprising.
muyuu 4y ago
yep, drawing hands well requires an understanding of the underlying anatomy that is more apparent than most other salient features hands are obvious to most people, but there would be many features that an AI would require a vast training set to completely capture, but that humans would also miss most of the time for instance look at Michelangelo's Moses, it was sculpted with models and with a very thorough knowledge of anatomy by the artist, and includes details like the muscle of the forearm that contracts when someone lifts the pinky finger: https://i.imgur.com/0vjAOnR.png https://i.imgur.com/0vjAOnR.png what are the chances that, for instance, the average person would notice that detail missing without being told about it? let alone reproduce it generatively, representing an imaginary person
fwlr 4y ago
“ You can image-bash together faces from examples and get something mostly-right simply through pattern copying. … You cannot do the same with hands, because rendering them plausibly requires at least intuition and approximation of inverse kinematics” Not sure I agree. I think hands probably just require ~1 OOM more image-bashing than faces, and training sets had ~1 OOM less samples of hands than faces. E.g. faces need 1 trillion cumulative Tflops while hands need 10 trillion cumulative Tflops, and because there were more faces than hands in the training set, by the time we reached 1T on faces we had reached 0.1T on hands. (numbers made up) Appeals to needing understanding of deeper or underlying principles like chess rules or inverse kinematics compel me to bring up the bitter lesson http://www.incompleteideas.net/IncIdeas/BitterLesson.html http://www.incompleteideas.net/IncIdeas/BitterLesson.html

by ethbr0 4y ago

ElFitz 4y ago

For some reason, the phrasing, the structure, the rhythm, they all very strongly remind me of Alan Watts’. Quite amusing and surprising.

muyuu 4y ago

yep, drawing hands well requires an understanding of the underlying anatomy that is more apparent than most other salient features hands are obvious to most people, but there would be many features that an AI would require a vast training set to completely capture, but that humans would also miss most of the time for instance look at Michelangelo's Moses, it was sculpted with models and with a very thorough knowledge of anatomy by the artist, and includes details like the muscle of the forearm that contracts when someone lifts the pinky finger: https://i.imgur.com/0vjAOnR.png https://i.imgur.com/0vjAOnR.png what are the chances that, for instance, the average person would notice that detail missing without being told about it? let alone reproduce it generatively, representing an imaginary person

fwlr 4y ago

“ You can image-bash together faces from examples and get something mostly-right simply through pattern copying. … You cannot do the same with hands, because rendering them plausibly requires at least intuition and approximation of inverse kinematics” Not sure I agree. I think hands probably just require ~1 OOM more image-bashing than faces, and training sets had ~1 OOM less samples of hands than faces. E.g. faces need 1 trillion cumulative Tflops while hands need 10 trillion cumulative Tflops, and because there were more faces than hands in the training set, by the time we reached 1T on faces we had reached 0.1T on hands. (numbers made up) Appeals to needing understanding of deeper or underlying principles like chess rules or inverse kinematics compel me to bring up the bitter lesson http://www.incompleteideas.net/IncIdeas/BitterLesson.html http://www.incompleteideas.net/IncIdeas/BitterLesson.html