41–50 of 153 posts
DALL-E is truly magic. It got me believing we are close to AGI. I wonder what Gary Marcus or Filip Pieknewski think about it. Surely they must be eating crow.
Yesterday I saw one of Gandalf eating samples at Costco. I was laughing hysterically for a minute. AI is not supposed to have a sense of humor. That was supposed to be the last province of the human, but it is quite awhile since a human made me laugh like that.
What wrote the prompt?
-- spent a day with DALL-E - here are some of my favorites: https://imgur.com/a/uD5yjV3 --
You like your lobsters
A lot of these posts showing up on HN. I wonder - is it because it is so new, or is it because the ways in which we are to use this technology are so nascent that we are discovering how to use it more precisely daily?
DALL-E is truly magic. It got me believing we are close to AGI. I wonder what Gary Marcus or Filip Pieknewski think about it. Surely they must be eating crow.
An AGI wouldn't need us to this extent, or at all. An AGI would also be able to come up with new ways to represent ideas, even ways that are foreign to us.
Earlier quoted context omitted.
Yesterday I saw one of Gandalf eating samples at Costco. I was laughing hysterically for a minute. AI is not supposed to have a sense of humor. That was supposed to be the last province of the human, but it is quite awhile since a human made me laugh like that.
What was the prompt for that image? What wrote the prompt?
I wonder how this would play out with the new Stable Diffusion
I've been playing with Stable Diffusion a lot, and in my experience its results are much weaker then what's shown in this post. The artistic pictures that it generates are beautiful, often more beautiful then Dalle-2 ones. But it has a real problem understanding the basic concepts of anything that is not the simplest task like "draw a character in this or that style". And explaining the situations in detail doesn't help - the AI just stumbles upon basic requests.
Seems like Stable Diffusion has a much more shallow understanding of what it draws and can only produce good result for things very similar to the images it learned from. For example, it could generate really good dutch still life paintings for me - with fruits, bottles and all the regular expected objects for this genre of painting. But when I've asked it to add some unusual objects to the painting (like a Nintendo switch, or a laptop) - it couldn't grasp this concept and just added more warbled fruit. Even though the system definitely knows how a Switch looks like.
The results in the post are much more impressive. I doubt that Dalle-2 saw a lot of similar images in training, but in all of the styles and examples it definitely understood how a llama would interact with a basketball, what are their relative sizes and stuff like that. On surface results from different engines might look similar, but to me this is an enormous difference in quality and sophistication.
"Image intentionally modified to blur and hide faces" I thought this was strange. Why hide an AI generated face?
-- spent a day with DALL-E - here are some of my favorites: https://imgur.com/a/uD5yjV3 --
I love that we're at the level where the physical "realism" of correctly representing quadrupedals playing basketball is a thing now. I suppose the next level AI will be expected to model a full 3d environment with physical assumptions based on the prompt and then run the simulation