3 ms·

There's a bit of a difference between interpolating to fill in spaces between pixels with hints from what's next to them and adding whole new words and sentence

There's a bit of a difference between interpolating to fill in spaces between pixels with hints from what's next to them and adding whole new words and sentences.

jacobr1 6y ago
Right. A better example might be removing every 5th word from a famous work, say one of shakespeare's sonnets. Then run GPT3 on it to interpolate the missing words. I bet a reasonable fraction of would match the correct words. Some would be acceptable substitutions, maybe you lose some fidelity or wit from the original, but it doesn't impact the overall enjoyment or understanding of the work. Then will be some minority of cases where we get something weird or wrong. Maybe a word that matches the general meaning, but is modern slang for example. Or maybe it guessed something that is plausible in context, but is opposite in meaning that changes something substantial in the work. I would expect that last scenario to be very rare as a fraction of total word substitutions, but also quite likely over the course of a whole piece of work. On the other hand, if we were able to render have a new undiscovered play available constructed from some sort of damaged fragment, rather than my hypothetical test ... having something we could read or see performed rather than just analyze academically seems worth the chance of getting some (even important) nuance in the work wrong.
- jacobwilliamroy 6y ago
  That's actually a worse analogy. You're obscuring what's actually going on here. I was being nice when I said it was like "doubling the length of the federalist papers." In a proper analogy, our fake document would be 24 times the length of the federalist papers. 96% of our document would be completely fabricated. FAKE. At that point you're learning more about the computer than whatever document you're viewing and it's the same for video. Scaling up from 480p to 4k requires the computer to fill in 96% of the information. That's really only good for entertainment value, not serious academic research into a historical period. Computers aren't magic time-traveling machines. If you want 4k reproductions of old films, just scan the actual film with better equipment. If you want colorization, that requires actual investigation and to figure out what colors were in use, and how to map them onto the video data. If no such investigation is possible then we're SHIT OUT OF LUCK. It happens. Move on. Stop revising history for entertainment.

There's a bit of a difference between interpolating to fill in spaces between pixels with hints from what's next to them and adding whole new words and sentences.

jacobr1 6y ago

Right. A better example might be removing every 5th word from a famous work, say one of shakespeare's sonnets. Then run GPT3 on it to interpolate the missing words. I bet a reasonable fraction of would match the correct words. Some would be acceptable substitutions, maybe you lose some fidelity or wit from the original, but it doesn't impact the overall enjoyment or understanding of the work. Then will be some minority of cases where we get something weird or wrong. Maybe a word that matches the general meaning, but is modern slang for example. Or maybe it guessed something that is plausible in context, but is opposite in meaning that changes something substantial in the work. I would expect that last scenario to be very rare as a fraction of total word substitutions, but also quite likely over the course of a whole piece of work. On the other hand, if we were able to render have a new undiscovered play available constructed from some sort of damaged fragment, rather than my hypothetical test ... having something we could read or see performed rather than just analyze academically seems worth the chance of getting some (even important) nuance in the work wrong.

jacobwilliamroy 6y ago
That's actually a worse analogy. You're obscuring what's actually going on here. I was being nice when I said it was like "doubling the length of the federalist papers." In a proper analogy, our fake document would be 24 times the length of the federalist papers. 96% of our document would be completely fabricated. FAKE. At that point you're learning more about the computer than whatever document you're viewing and it's the same for video. Scaling up from 480p to 4k requires the computer to fill in 96% of the information. That's really only good for entertainment value, not serious academic research into a historical period. Computers aren't magic time-traveling machines. If you want 4k reproductions of old films, just scan the actual film with better equipment. If you want colorization, that requires actual investigation and to figure out what colors were in use, and how to map them onto the video data. If no such investigation is possible then we're SHIT OUT OF LUCK. It happens. Move on. Stop revising history for entertainment.