deerenbung wrote:OpenAI are absolutely taking record of the things you mention but the results of that will only come out in future models. All improvements we've seen to Chat GPT so far have been learning to use it better, or putting a prettier front on it.
The fact is that training models is much harder than using them. This is why a Stable Diffusion model (the only "AI" I've got hands-on experience with) can be a few gigabytes in size and generate images in seconds when according to Wikipedia it took 150,000 GPU hours to train on hundreds of TB of images. Training a model for more specific tasks on top of an existing one is easier, sure, but still much more intensive than generating an image.
At the end of the day it's not smart like a person is at all. "AI" as it currently stands is a giant ball of statistics with no knowledge besides a vague representation of a slice of the internet. So the fact that it can do what it does is pretty crazy! But if we want to make AI that is "smart like a person" we'll need at least one more paradigm shift to produce models that truly learn and grow in real time.
How embarrassing for me. Sorry - I meant Un/Stable Diffusion and and OpenAI. Not ChatGPT. Yeah, ChatGPT definitely isn't what people turn to to create porn.
The updated versions of these platforms are getting to the point where they can create videos in real time. Despite the "wacky hands" problem still not being fully solved, somehow AI can now, very realistically, start creating actual scenes.
On one of the following issues you're half-right/half-wrong - the other you're just wrong. I'm not saying this angrily or in a mean way - just going to provide context of my understanding for that claim.
For the half-right/half-wrong - yes, you're right that the AI programs aren't learning in real time - you're wrong in that it matters. Every single image AI generates - there is coding within it that detail the process it took - the result it got - and how that result was interpreted. So, if I put a picture of a naked lady on there and tell the prompt, "Make her breasts bigger and like X and like Y and whatever else," it will keep track of everything it does, and then it will keep track of how I react to it (whether I download it - whether I save it - whether I enter the same prompt again and ask it to redo it).
Every site that uses Unstable Diffusion, you better believe it, has it within their contract that in order to do so, Unstable Diffusion, or OpenAI, or whatever, gets to keep records of the computations that occured. Not that they care about the people who used it - but the results that were achieved. Okay, so then ...
They take ALL the data from these billions upon billions of creations and they ... you guessed it ... insert all of it into the next version of the program. So, Stable DIffusion 5.1 will, most likely, have 30 billion more inputs than 5.0 did.
So, rather than learning in real time, data gets compiled, and then uploaded en-masse, to create the next iteration. So ... not real-time, but ... still learning from everything. But ... how does it do this?
Thing is - you cannot have any one person sit down and teach AI what was a good result and what wasn't. Like ... how is anybody going to go through billions upon billions of creations? Well, obviously, they can't. But ... AI can.
That's where all the metrics of user-data gets used in conjunction with the images that were generated. AI then has algorithms that determine, exactly, what went right and what went wrong. How does it do this? How can it do this? That would take billions of lines of code, right?
Except that ... it doesn't. Because, what AI did, was create it's own computational programming languages. Just that ... they don't operate like human coding. Human coding is 2-dimensional. It goes up and down.
if X value > 200
then do this
if X value < 200
then do this
AI coding is ... I'm pretty sure it's 4th dimensional at this point. Imagine you had a stack of a million sheets of paper. Now, each of these pieces of paper had a grid of 10,000 by 10,000 squares. AI programming takes all of these sheets, and then, simultaneously, based on which grid spaces are filled with what information, and the relationship and distance all of those grid spaces (with their information) have between all of the other grid spaces .. determines what it's knowledge is. So, by a certain block of information being 2,000 grid spaces to the left, 5,000 above, and 236,890 sheets away from another block of information, determines how that value is interpreted. That interpretation can change based on it's repositioning, or the addition of other blocks of information close to it.
It's ... literally beyond anything that anybody, beyond the absolute smartest people on the planet, can even comprehend anymore as to how it actually works.
This is the machine that is making your porn. And the results are ... as you noted ... completely jaw-dropping within an 8-month time period. AI did this. On it's own. I said it's as smart as a human, but human's have difficulty reading the pull sign on a door as they're pushing it. These AI models are ... starting to do more than we could have ever fathomed.
Does this deserve it's own topic? Maybe. But, like, the only point I had, really, was that this is something that is going to be making exponential leaps and bounds in the upcoming 8 months that will make the last 8 look like it was a pre-game warm-up.
If Analvids could bother with that 3D animation stuff because, as they realized, there are folks who like it - when would be the right time for them to jump on the bandwagon that's pretty much made everyone a porn creator. Why not get the drop on anyone else - by applying all the professional know-how and techniques to this field while it's still young - so that once it blossoms and really comes into it's own - people will be dropping wads of cash for fantasy girls doing BBC TAP at Pornbox.
And not AI-Porn-DAP or DAIPORN dot com.