Been liking Alex O’Connor’s ChatGPT explanation videos and ChatGPT related experiments.
Alex O’Conner makes content related to philosophy and religion but I particularly enjoyed, in addition to this video, one where he gaslights ChatGPT using moral dilemmas.
In this video he tells you the reason why it is so hard to get ChatGPT to do this. Short Answer: most images you find of wine are either empty glasses or partially full because who fills their wine to the top?
This video was nice and simple.
It really drives home the point that chat bots aren’t actually creative and, in simple terms, just spit out averages and probabilities.
Sounds like Francis Galtin’s Ox
“The classic wisdom-of-the-crowds finding involves point estimation of a continuous quantity. At a 1906 country fair in Plymouth, 800 people participated in a contest to estimate the weight of a slaughtered and dressed ox. Statistician Francis Galton observed that the median guess, 1207 pounds, was accurate within 1% of the true weight of 1198 pounds.This has contributed to the insight in cognitive science that a crowd’s individual judgments can be modeled as a probability distribution of responses with the median centered near the true value of the quantity to be estimated.”
It is just an over-powered autocorrect, that same concept applies to images, just repeating patterns of pixels.
Image generation uses DallE and is not Baked into the model.
All it does is give it a prompt. Generates and shows you the results for that prompt.
You can click the pictures to see that prompt and you will see that it verbosely requested it overflowing but dalle does not always interpret that prompt the same, actually i found llms rather suck at prompting image generation models because they behave very strongly on certain words.
A similar experiment was done with “street with no lanterns” always resulting in a lantern.
If I ask you to imagine a street with no lanterns, are you imagining lanterns or no lanterns?
Lanterns of course.
Would take an image generation model at least 3 steps which it doesn’t have right now.
A review step to see if the output matches the prompt.
A identification step to detect elements that don’t match
A redo step to mix that area in the background image (remove) or regenerates an improvement.
Right now you cant iterate on images. Every minor tweak is a completely new image. At least not with dalle because you cant control the seed.
It might be more accurate to use something like Leonardo.AI rather than ChatGPT because you can edit existing images as needed and set the seed. You can even keep a consistent ‘character’ and reuse it in many pictures. Its dreamshaper model is based on SD. I have had the most accurate results with Leonardo. I don’t use ChatGPT/Dall-E for images, it uses too much on a free plan.
And this is a prime example of why these trained models will never be AGIs. It only knows what it’s been trained on and can’t make inferences or extrapolations. It’s not really generating an image’s much as really quickly photoshopping and merging images it already knows about.
It’s just patterns of pixels. It recognizes an apple as just a bunch of reddish pixels etc, then when given an image of a similar colored red ball, or something, it is corrected until it ceases to recognize something not an apple as an apple. It really does not know what an apple looks like to begin with. It’s like declaring a variable. The computer does not know what the variable really means just what to equate it to.
This makes me ponder the assertions that these are exotic compression algorithms.
I’m down with compression being the mechanism asserted for legal reasons.
