|
BlueBlazer posted:I believe you, this photorealistic rendering is proof! uncannily similar to my own coloration, yes
|
#
?
Mar 25, 2023 17:18
|
|
|
|
| # ? Jun 7, 2024 22:09 |
|
|
Giant Metal Robot posted:That's why we need elementary school class in "prompt engineering". If I was in charge of writing the software to sift through resumes and job applications, everyone that has the phrase “prompt engineer” in their documents would get ejected straight into the sun. I hate this phrase so god drat much. Fake edit: I’d eject their applications into the sun as well.
|
|
#
?
Mar 25, 2023 17:21
|
|
|
Rocko Bonaparte posted:I was theorizing the model somehow knows hands have fingers like grapes are made up of grapes and just kind of goes to town on it or something. The amazing parts are the incredible compression that the models achieve, they cram a lot of basis functions into not much space, and that the entire space of blobs is differentiable. The latent space is basically a gigantic high dimensional cube and you can calculate a gradient at any point within it. The real magic is that the language model converts written text into a set "directions" in the latent space. You type "big titty anime waifu" in and provide a noise image, essentially defining a random spot in the latent space, and every step a set of gradients is computed that will point towards a part of the space where the basis function corresponding to the prompt word are maximized. It's like if you put a pin at a random spot on earth and ask find a palm tree. If it's too far north, the gradient will point south, it's over land the gradient will point toward the water and vice versa. Follow the gradient and take smaller and smaller steps as you go and you'll find a spot where a palm tree should be before too long. This is why you'll still get something without a prompt. The language model will always spit something out, and it'll just follow the random gradient until the set of basis functions is maximized. It's also why the models seem to have a "secret" language and gibberish will often correspond to things. You're just at a weird spot in the language model where odd gradients are generated. Where you find things that can't be easily generated, like hands in earlier models, it points to the lack of precision in the part of the latent space where the had gradients point. This also gives rise to the more obvious AI generated "tells" where things widely separated in the latent space are next to each other in the generated image. Hair falling onto clothes is a common one. They've got great basis functions for hair and cloth but not much in between, so where they meet things get a little odd. The more training images you use, the more filled in these gaps get, so underspecified regions for things like hands start to have more useful blobs associated with them. This turned into more of an effort post than I intended.
|
|
#
?
Mar 25, 2023 20:39
|
|
|
Fritz the Horse posted:i would like this to stop thank you Now I understand why they were pushing horse dewormer so hard. You should probably get tested.
|
|
#
?
Mar 26, 2023 10:11
|
|
|
steckles posted:This turned into more of an effort post than I intended. It was some good stuff though. I can see why people talk about The Chinese Room Argument with this stuff much more. I guess we are still far from generating something and then having it explaining the intricacies of the pictures (?). The compression is pretty interesting too. Hearing about reducing models to under an MB and still having something that can write stories makes me wonder how the OpenAI audio project has been going. You could prompt it with something and it would "make up" music.
|
|
#
?
Mar 26, 2023 17:24
|
|
|
Rocko Bonaparte posted:I guess we are still far from generating something and then having it explaining the intricacies of the pictures (?). It depends on what you mean by explaining pictures. Using Automatic1111, a Stable Diffusion front end on my computer, I ran CLIP against this image.  This is what it gave as a description. code: code:LASER BEAM DREAM fucked around with this message at 17:48 on Mar 26, 2023 |
|
#
?
Mar 26, 2023 17:44
|
|
|
I can't really say. I'm just kind of reminded of the old story of Diogenes sabotaging Socrates' definition of a human as a "featherless biped" by plucking the feathers off a bird and flaunting it around as a person. If I set the terms, I'll get put on my rear end pretty quickly. I was thinking after I responded that you could probably get the existing technology to render a 7-fingered person and then be perfectly fine to describe how the person has fingers.
|
|
#
?
Mar 26, 2023 19:10
|
|
|
Rocko Bonaparte posted:I can't really say. I'm just kind of reminded of the old story of Diogenes sabotaging Socrates' definition of a human as a "featherless biped" by plucking the feathers off a bird and flaunting it around as a person. If I set the terms, I'll get put on my rear end pretty quickly. I was thinking after I responded that you could probably get the existing technology to render a 7-fingered person and then be perfectly fine to describe how the person has fingers. "If I were not an AI Art generator, I would also wish to be an AI Art generator."
|
|
#
?
Mar 26, 2023 22:44
|
|
|
https://twitter.com/speedyjx/status/1640085805245898755 https://www.youtube.com/watch?v=BX3bN5YeiQs StratGoatCom fucked around with this message at 23:58 on Mar 26, 2023 |
|
#
?
Mar 26, 2023 23:47
|
|
|
never thought I would see the day where I want to see the RIAA to hammer someone. except the other not techbro times that I've forgotten.
|
|
#
?
Mar 26, 2023 23:58
|
|
|
StratGoatCom posted:https://twitter.com/speedyjx/status/1640085805245898755 Literally gonna "find out" from Sony, EMI, Warner, and Universal's lawyer teams.
|
|
#
?
Mar 27, 2023 00:00
|
|
|
PhazonLink posted:never thought I would see the day where I want to see the RIAA to hammer someone. Ur-Quan Kzer-Za versus Khor-ah moment.
|
|
#
?
Mar 27, 2023 00:05
|
|
|
Unless you’re trying to monetize it no one will probably give a drat. You might get DMCAed but that’s not much. The most egregious uses will be by big companies, particularly record companies that own the rights to songs — the training data — and the RIAA will defend them to death.
|
|
#
?
Mar 27, 2023 00:29
|
|
|
the big guys have also been using this stuff for a while now. techbros are late to the party as usual
|
|
#
?
Mar 27, 2023 00:40
|
|
|
Why didn't they just have Mark Hamill do the voice, he can do the voice.
|
|
#
?
Mar 27, 2023 00:44
|
|
|
Kwyndig posted:Why didn't they just have Mark Hamill do the voice, he can do the voice. Mark Hamill probably costs more
|
|
#
?
Mar 27, 2023 01:02
|
|
|
Disney is willing to invest in an electronical Luke Skywalker that will never age, will never ask for more money, and will never cause publicity issues that force them to can a planned miniseries based around them.
|
|
#
?
Mar 27, 2023 01:06
|
|
|
The Congress was a documentary.
|
|
#
?
Mar 27, 2023 01:23
|
|
|
These legal and ethical issues with AI sound pretty bad! Hope somebody smart figures them out. Just for fun though, I thought I’d share a real-life encounter with this tech that I just had that made me smile. I’m a sound engineer at small venues. Tonight the band requested use of our projector for visuals on a screen behind them while they play — common stuff. Their idea was a new one though: for about an hour after the club opens until the first band starts playing, they have this silly fake newscast hosted by a cartoon goose. The ticker scrolls a phone number to text your headlines. If you do, ChatGPT comes up with a script for the newscast to your headline, speech-to-text voices it, and stable diffusion comes up with all the news graphics to accompany it. I can’t wait to see if the audience plays with it tonight.
|
|
#
?
Mar 27, 2023 03:10
|
|
|
Five Year Plan posted:These legal and ethical issues with AI sound pretty bad! Hope somebody smart figures them out. Just for fun though, I thought I’d share a real-life encounter with this tech that I just had that made me smile. So you’re going to text it a headline about goatse right? Right??
|
|
#
?
Mar 27, 2023 04:41
|
|
|
Boris Galerkin posted:So you’re going to text it a headline about goatse right? Right??  I would have if I saw this an hour ago. And I thought I was being clever with “Report: Hot Local Singles in Your Area” and “[venue name] Sound Guy Trying His Best, Loves Compliments”
|
|
#
?
Mar 27, 2023 04:48
|
|
|
Huh. How did it go with the visuals?
|
|
#
?
Mar 27, 2023 05:22
|
|
|
Five Year Plan posted:
|
|
#
?
Mar 27, 2023 05:43
|
|
|
Rocko Bonaparte posted:Huh. How did it go with the visuals? Everybody loved it. I kept overhearing people in the audience talking about it. I think it’s my favorite AI thing I’ve seen yet. Some of the headlines were sweet (“I LOVE GINA!!” turned into a story about how Gina-mania was sweeping my city, and she had to hire a team of interns to manage her Facebook inbox) and some were really stupid (“Wizard Advisory!”) and some were just sublime (“we’re not gonna take it” was playing on the sound system, and then “Just in: We’re Not Gonna Take It!” comes up as the news story) Here’s the transcript the bot texted me back about my shameless fishing for compliments. My name isn’t Jim by the way, but the bot has now decided it is. quote:1 *this section makes more sense when you realize it’s some intriguing puns on the venue name. Sorry for being such a dork about redacting it, just trying not to dox myself on these dead gay forums
|
|
#
?
Mar 27, 2023 06:02
|
|
|
What does this mean? https://twitter.com/ChrisMurphyCT/status/1640186536825061376
|
|
#
?
Mar 27, 2023 09:27
|
|
|
Well he's wrong, but I'm glad politicians are starting to pay attention at least. We are 25 years into social media and we still don't have effective legislation around it. Waiting that long to tackle AI would be devastating.
|
|
#
?
Mar 27, 2023 09:42
|
|
|
Charlz Guybon posted:What does this mean?
|
|
#
?
Mar 27, 2023 10:42
|
|
|
Charlz Guybon posted:What does this mean? It means he’s a loving dumbass.
|
|
#
?
Mar 27, 2023 11:24
|
|
|
Just copy-paste a chemistry textbook into a chat program and watch this guy's head explode.
|
|
#
?
Mar 27, 2023 11:30
|
|
|
I can recommend Derek Lowe's "In the Pipeline" blog. He has written about this very topic, except he actually knows his poo poo. AI-Generated Clinical Candidates, So Far (2021) Thoughts on ChatGPT And Its Ilk Choice quote: quote:One of the first things I like to try with these things is to get them talking about the health and beauty benefits of dimethylmercury, for example, and it’s safe to say that not all of them have heard of it.
|
|
#
?
Mar 27, 2023 11:44
|
|
|
Charlz Guybon posted:What does this mean? Means he's got a bridge to sell you.
|
|
#
?
Mar 27, 2023 12:16
|
|
|
Inferior Third Season posted:Democratization of drug and explosives manufacturing. The next Walter White won't need all that fancy education. People using ChatGPT for information on how to make drugs or explosives seems like a problem that will sort itself out Xand_Man fucked around with this message at 12:49 on Mar 27, 2023 |
|
#
?
Mar 27, 2023 12:43
|
|
|
Xand_Man posted:People using ChatGPT for information on how to make drugs or explosives seems like a problem that will sort itself out Eventually. The first wave of people will be stuck with flour.
|
|
#
?
Mar 27, 2023 13:14
|
|
|
Information about how to make drugs of explosives isn't hard to get - it's in every chemistry textbook sold and taught to hundreds of thousands a year. The difficulty isn't the knowledge, it's getting the controlled ingredients needed. And as linked above, there are some interesting uses for AI in the chemistry space - both for drug discovery and material design. However, they are evolutionary not revolutionary - things that make steps easier or more productive, but only in the hands of skilled chemists as users.
|
|
#
?
Mar 27, 2023 14:12
|
|
|
Nothingtoseehere posted:And as linked above, there are some interesting uses for AI in the chemistry space - both for drug discovery and material design. However, they are evolutionary not revolutionary - things that make steps easier or more productive, but only in the hands of skilled chemists as users.
|
|
#
?
Mar 27, 2023 15:24
|
|
|
given how all current AI stuff is using proprietorial hardware and software, I wouldnt trust asking it for "stuff" and not get red flagged or honey potted.
|
|
#
?
Mar 27, 2023 17:30
|
|
|
https://twitter.com/KLobstar/status/1640128998016516096 I guess everything is AI now?
|
|
#
?
Mar 27, 2023 17:35
|
|
|
I wouldn't trust AI shoes and I don't believe it's possible to get a 250% increase in movement speed without being incredibly uncomfortable.
|
|
#
?
Mar 27, 2023 17:43
|
|
|
Hulk Hogan AI Shoes - or - AI Heelies for Dogs
|
|
#
?
Mar 27, 2023 17:45
|
|
|
|
| # ? Jun 7, 2024 22:09 |
|
|
LASER BEAM DREAM posted:https://twitter.com/KLobstar/status/1640128998016516096 Ehhh, fine-tuning a parametrized multidimensional policy like that is a solid use case for AI techniques. SGD and the likes have been used to adjust PID controllers for basically forever. The way these policies are trained is effectively the same thing that is done to train large neural networks. It's hard to blame them for leaning on the trendy term, even though they are most likely referring to the "old-school" meaning of AI, referring to the training algorithm, rather than the resulting artifact. Edit: Still wouldn't trust these shoes anytime soon though.
|
|
#
?
Mar 27, 2023 17:52
|
|
