4 ms·

> It generates the most likely image/text/code in its training set to be found following/around/correlating with the prompt. This is simply false. It's not a s

> It generates the most likely image/text/code in its training set to be found following/around/correlating with the prompt. This is simply false. It's not a search engine that outputs the training item closest to the prompt. In reality, it is "learning" (in some sense) how to correlate text to images, and then generating brand new images in response to input text. If this is legal for humans to do, then it's probably legal for machines to do the same thing.

toss1 4y ago
Perhaps I was not clear enough to prevent ambiguity. >>It's not a search engine that outputs the training item closest to the prompt. Correct, it is not outputting the training ITEM, it is outputting finer-grained slices of many items, more of a mash-up of the training items. Of course it is not taking an entire specific image the closely matches the search term, it is taking averages of component images of "astronaut riding a horse over the moon in style of Rembrandt". That image won't exist in the training set, but astronauts, horses, and Rembrandt-style coloring and shading do exist, and it is assembling those from averages of the components found it's training set, not from some abstract imagination or understanding. The fact that the astronaut suit may not be the exact same as any of it's training images is the same as if I averaged 100 faces in photoshop, not because there is some kind of "learning" or "understanding". Ability to do useful statistical mashups is NOT the same as "learning". This can be shown in a different "AI" engine'd failure to solve a child's puzzle. ChatGPT, when presented with: "Mike's mom had four kids, three are named Lucia, Drake, and Kelly, what is the fourth kid's name?". It said there is insufficient info, and doubled down when told that the answer is in the question. >>how to correlate text to images yes, as I pointed out, "correlating with the prompt." I didn't say it correlated an entire image, but I also failed to specify that it was correlating components. >> If this is legal for humans to do, then it's probably legal for machines to do the same thing. This [0] is I'm quite sure, not legal. Asked for an image of a person named "Ann Graham Lotz", it returned the image in the training set, slightly degraded. First, that is literally the search engine functionality you were deriding. Second, if you asked a human artist to produce the same image, without infringing copyright, they would produce something likely recognizable as the person, but obviously not resembling the training photo. It doesn't matter if they are a portrait painter, sketch artist, Photoshop jockey, or Picasso-like impressionist. So, no, this does not represent learning in any conceptual, creative, or human-like sense. It does represent mashing-up averages of inputs of various components. Feed in enough "astronaut" photos, and it'll be able to select out the humans in the spacesuit as the response to that prompt. Same for "horse", "moon", "riding", and "Rembrandt". and it can mash them together into something useful with good prompts. But give it something very specific, like a person's name, and you get basically a search-engine result, because it doesn't have enough input data variety to abstract out the person 'object' from the background. [0] https://techxplore.com/news/2023-02-ai-based-image-generation-generate-trainer.html https://techxplore.com/news/2023-02-ai-based-image-generatio...
- d110af5ccf 4y ago
  to me the "search engine" case where it reproduces a specific training image seems like a failure mode that's distinct from normal operation > it is assembling those from averages of the components found it's training set, not from some abstract imagination or understanding how exactly are you so certain that the human brain handles abstract concepts any differently? please note that I'm not claiming that I myself know, but rather that you almost certainly do not know and thus are presenting an invalid argument what is human imagination anyway? > assembling those from averages of the components found it's training set > slices of many items, more of a mash-up of the training items > But give it something very specific ... it doesn't have enough input data variety to abstract out the person 'object' from the background so is it abstracting or not? where's the line between that and a mere statistical mashup?
  - toss1 4y ago
    >>how exactly are you so certain that the human brain handles abstract concepts any differently? Good question. At the very least, we have a far deeper understanding of physical reality. Humans would not unintentionally (e.g., for effect) produce images of people with three ears, or of a bikini-clad girl seated on a boat with her head and torso facing us, and also her butt somehow facing us and thighs/knees away... yet I've seen both of these in the last week (sorry, couldn't find the reference, it was a hilarious image, looked great for 2sec until you saw it) I admit that it is possible (tho I think unlikely) that this is a difference in quantity, not in kind. One reason to doubt this is that Stable Diffusion was trained on 2.3 billion images. This is a vastly larger library than any human has seen in their lifetime (considering that viewing 2.3 billion images at one per second would take 72.8 years). Yet even if you count every second of eyesight as 'training', children under 1/10 of that age, who have seen only 10% of those images would not make the same kinds of mistakes. Plus, the neuron/synapse/neurotransmitter and brainstem/midbrain/cerebellum micro & macro-architectures are vastly different than the computer training models. So, I think we can be confident that something different is happening. >>so is it abstracting or not? where's the line between that and a mere statistical mashup? Good question. There is definitely something we might call, or that might resemble abstraction. It's definitely able to associate the cutout images of an astronaut in a spacesuit from the backgrounds. It can evidently assemble those from different angles. But it certainly does not have the abstraction to understand even the correct relationship between the parts of a human. E.g., it seems to keep astronauts' parts in the right relationship, but not bikini-clad-girls' parts (because of the variety of positions in the dataset?). There's no understanding of kinesiology, anatomy, or anything else that an actual artist would have. Could this be trained in? I expect so, but I think it would require multiple engines, not merely six orders of magnitude more training of the same type. Even if 10^6X more training eliminated these error types and even performed better than humans, I'm not sure it would be the same, just different and useful. I'd want to see evidence that it was not merely cut-pasting components of images in useful ways, but generating it from an understanding of the sub-sub components: "the thigh bone connects to the hip bone, the hip can rotate this far but not that far, the center of mass is supported...+++" as an artist builds up their images. Good artists study anatomy. These "AI"s haven't a clue that it exists. >>to me the "search engine" case where it reproduces a specific training image seems like a failure mode that's distinct from normal operation Au contraire, it seems that this merely exposes the normal operation. Insufficient images of that person prevented it from abstracting the person components from the background, so it just returned the whole thing. IDK whether it would take a dozen, hundred, or thousand more images of the same person, to work properly. But, if they all had some object in the background (e.g., a lamp) that was the same, the "AI" would include it in their abstraction. (but I could be wrong).
oldgradstudent 4y ago
> In reality, it is "learning" (in some sense) Only because some people named their field "machine learning" and called it "learning". It has no relation to human learning. If your child accidently confuses a giraffe with some other animal you correct then, you don't add the picture to their training set and show them again thousands of pictures of giraffes hoping that their success rate improves. If you ask Stable Diffusion for a starry night, you get van Gogh's starry night.

toss1 4y ago

Perhaps I was not clear enough to prevent ambiguity. >>It's not a search engine that outputs the training item closest to the prompt. Correct, it is not outputting the training ITEM, it is outputting finer-grained slices of many items, more of a mash-up of the training items. Of course it is not taking an entire specific image the closely matches the search term, it is taking averages of component images of "astronaut riding a horse over the moon in style of Rembrandt". That image won't exist in the training set, but astronauts, horses, and Rembrandt-style coloring and shading do exist, and it is assembling those from averages of the components found it's training set, not from some abstract imagination or understanding. The fact that the astronaut suit may not be the exact same as any of it's training images is the same as if I averaged 100 faces in photoshop, not because there is some kind of "learning" or "understanding". Ability to do useful statistical mashups is NOT the same as "learning". This can be shown in a different "AI" engine'd failure to solve a child's puzzle. ChatGPT, when presented with: "Mike's mom had four kids, three are named Lucia, Drake, and Kelly, what is the fourth kid's name?". It said there is insufficient info, and doubled down when told that the answer is in the question. >>how to correlate text to images yes, as I pointed out, "correlating with the prompt." I didn't say it correlated an entire image, but I also failed to specify that it was correlating components. >> If this is legal for humans to do, then it's probably legal for machines to do the same thing. This [0] is I'm quite sure, not legal. Asked for an image of a person named "Ann Graham Lotz", it returned the image in the training set, slightly degraded. First, that is literally the search engine functionality you were deriding. Second, if you asked a human artist to produce the same image, without infringing copyright, they would produce something likely recognizable as the person, but obviously not resembling the training photo. It doesn't matter if they are a portrait painter, sketch artist, Photoshop jockey, or Picasso-like impressionist. So, no, this does not represent learning in any conceptual, creative, or human-like sense. It does represent mashing-up averages of inputs of various components. Feed in enough "astronaut" photos, and it'll be able to select out the humans in the spacesuit as the response to that prompt. Same for "horse", "moon", "riding", and "Rembrandt". and it can mash them together into something useful with good prompts. But give it something very specific, like a person's name, and you get basically a search-engine result, because it doesn't have enough input data variety to abstract out the person 'object' from the background. [0] https://techxplore.com/news/2023-02-ai-based-image-generation-generate-trainer.html https://techxplore.com/news/2023-02-ai-based-image-generatio...

d110af5ccf 4y ago
to me the "search engine" case where it reproduces a specific training image seems like a failure mode that's distinct from normal operation > it is assembling those from averages of the components found it's training set, not from some abstract imagination or understanding how exactly are you so certain that the human brain handles abstract concepts any differently? please note that I'm not claiming that I myself know, but rather that you almost certainly do not know and thus are presenting an invalid argument what is human imagination anyway? > assembling those from averages of the components found it's training set > slices of many items, more of a mash-up of the training items > But give it something very specific ... it doesn't have enough input data variety to abstract out the person 'object' from the background so is it abstracting or not? where's the line between that and a mere statistical mashup?
- toss1 4y ago
  >>how exactly are you so certain that the human brain handles abstract concepts any differently? Good question. At the very least, we have a far deeper understanding of physical reality. Humans would not unintentionally (e.g., for effect) produce images of people with three ears, or of a bikini-clad girl seated on a boat with her head and torso facing us, and also her butt somehow facing us and thighs/knees away... yet I've seen both of these in the last week (sorry, couldn't find the reference, it was a hilarious image, looked great for 2sec until you saw it) I admit that it is possible (tho I think unlikely) that this is a difference in quantity, not in kind. One reason to doubt this is that Stable Diffusion was trained on 2.3 billion images. This is a vastly larger library than any human has seen in their lifetime (considering that viewing 2.3 billion images at one per second would take 72.8 years). Yet even if you count every second of eyesight as 'training', children under 1/10 of that age, who have seen only 10% of those images would not make the same kinds of mistakes. Plus, the neuron/synapse/neurotransmitter and brainstem/midbrain/cerebellum micro & macro-architectures are vastly different than the computer training models. So, I think we can be confident that something different is happening. >>so is it abstracting or not? where's the line between that and a mere statistical mashup? Good question. There is definitely something we might call, or that might resemble abstraction. It's definitely able to associate the cutout images of an astronaut in a spacesuit from the backgrounds. It can evidently assemble those from different angles. But it certainly does not have the abstraction to understand even the correct relationship between the parts of a human. E.g., it seems to keep astronauts' parts in the right relationship, but not bikini-clad-girls' parts (because of the variety of positions in the dataset?). There's no understanding of kinesiology, anatomy, or anything else that an actual artist would have. Could this be trained in? I expect so, but I think it would require multiple engines, not merely six orders of magnitude more training of the same type. Even if 10^6X more training eliminated these error types and even performed better than humans, I'm not sure it would be the same, just different and useful. I'd want to see evidence that it was not merely cut-pasting components of images in useful ways, but generating it from an understanding of the sub-sub components: "the thigh bone connects to the hip bone, the hip can rotate this far but not that far, the center of mass is supported...+++" as an artist builds up their images. Good artists study anatomy. These "AI"s haven't a clue that it exists. >>to me the "search engine" case where it reproduces a specific training image seems like a failure mode that's distinct from normal operation Au contraire, it seems that this merely exposes the normal operation. Insufficient images of that person prevented it from abstracting the person components from the background, so it just returned the whole thing. IDK whether it would take a dozen, hundred, or thousand more images of the same person, to work properly. But, if they all had some object in the background (e.g., a lamp) that was the same, the "AI" would include it in their abstraction. (but I could be wrong).

oldgradstudent 4y ago

> In reality, it is "learning" (in some sense) Only because some people named their field "machine learning" and called it "learning". It has no relation to human learning. If your child accidently confuses a giraffe with some other animal you correct then, you don't add the picture to their training set and show them again thousands of pictures of giraffes hoping that their success rate improves. If you ask Stable Diffusion for a starry night, you get van Gogh's starry night.