|
TIP posted:if it makes you feel better this is true of everyone It's definitely going to be the second, it's made by a for profit corporation whose explicit goal is to make general ai as fast as possible so that they can steer the future of it.
|
#
¿
Apr 22, 2022 22:20
|
|
|
|
| # ¿ May 15, 2024 13:35 |
|
|
Nuts and Gum posted:There will always be a need for programmers, but they won’t be writing code, they’ll be shaving off the rough corners of code produced by other software, and then implementing the last mile portions. I suspect the same will be true for other disciplines that are affected by advancements in machine learning. I've always been more skeptical of this than the other stuff tbh. Like we can forgive a lot of the rough edges for the images we're seeing itt, but I don't think we'll be as ready to forgive the edges in a computer program. I have this feeling that if you asked an ai to make a computer program for you to do a task, you'll just incur a lot of "technical debt" and not really get out ahead in the end. There's that and I just don't trust programs written by humans to begin with, but I feel like we'll just get even more black boxy if we ask an ai to program poo poo for us. There are a lot of domains where it must be verifiable and audited. Like maybe you could have it build a website for you but idk if you can ask it to write most of excel 2.0. People make real-life decisions with software, but they don't do that with illustrations. Nuts and Gum posted:This is an interesting example because humans are weird and have irrational opinions, like wanting to talk to a person when calling customer service, even if an automated solution could solve the issue. I wonder how many of these hang ups simply go away as new generations grow up with these technologies. My guess would be that yes, younger generations will be quicker to do that kind of stuff. Like my wife's nephew was very excited to go to one of those Amazon grocery stores. But I refuse to enter an Amazon grocery store. Well that's mostly because I think Amazon is a lovely company we'd be better without, but the creepy panopticon aspect of Amazon doesn't really help. But if you're just used to Amazon listening and recording every waking minute of your life then well why not go there?
|
|
#
¿
Apr 23, 2022 04:31
|
|
|
I wonder if this is going to just produce endless amounts of white people when photorealistic pictures are requested. Like that mariners game scene was looking pretty monochrome. I wonder if it would give you different things if you explicitly ask it for, for example, a latino xxx doing yyy. I mean I hope they mitigated that flaw, but otoh they’re tech bros so lol.
|
|
#
¿
Apr 23, 2022 06:30
|
|
|
The Protagonist posted:it's doing deliberate face warping for liability issues as i understand it The dall-e mini one (not from the dall-e-2 creators) is just not trained as long and a smaller model, so it just doesn't have the capability to do faces. They have a short limitation and biases section on their writeup, where they say that "Faces and people in general are not generated properly" and "Animals are usually unrealistic". Their version was trained for only like a couple days, and is reportedly 27 times smaller than the dall-e-1 model. They also address the fact that their model predominantly generates white people. It's interesting what things the model can and can't do well. I found that the watercolor style can give good results. I tried radish lightbulb and watercolor radish lightbulb.  vs  another prompt: bird with flower feathers vs watercolor bird with flower feathers.  vs  The watercolor prompt seems to keep it more focused or something. It will also tend to give a single object in the center with a white/plain background which maybe helps it too.
|
|
#
¿
May 6, 2022 16:01
|
|
|
It's weird to me how a lot of those style mashups were rendered as if they were on canvas. But they're also all from the same angle? It's just an oddly specific thing for it to produce, I wonder what part of the learning set it came from. Anyway, it can do passable legos 
|
|
#
¿
May 9, 2022 15:10
|
|
|
Yeah that aspect of it is pretty cool, but otoh you're also limited to whatever the model can produce, and all the biases and assumptions that it has learned. Like it's good at imagining the future of the golden gate bridge, but can it show you the future of Lagos with a similar verisimilitude? Anyway, tried to get dall-e mini to give me a goatse stargate. Prompt: "the stargate gripped by two hands". So close.  
|
|
#
¿
May 18, 2022 15:17
|
|
|
|
| # ¿ May 15, 2024 13:35 |
|
|
Brutal Garcon posted:There's already people calling bullshit on this, but it's interesting in a "hosed up if true" way I was thinking about similar things too. I know when the big image recognition models from google were first getting good there were some papers that showed that you can get a 99% prediction of something being a turtle, when it's an image of random noise. Or like you can subtly distort a picture of a panda and make it so the model says that 99% certain it's a dog. Given that I was wondering if it would ever produce just totally garbled or noise images that it still thinks match very well to the prompt. But I didn't think about the other way, you can send it totally garbled words and have it really think that you mean birds. IDK it's a bit weird, like the way we perceive and identify things we don't get random hits like this.
|
|
#
¿
Jun 2, 2022 04:19
|
|