LOL, I love kbin’s public downvote records. I quoted a bunch of different sources demonstrating that you’re wrong, and rather than own up to it and apologize for preaching from atop Mt. Dunning-Kruger, you downvoted me and ran off.
I advise you to step out of whatever echo chamber you’ve holed yourself up in and learn a bit about AI before opining on it further.
Unfortunately, you pretty much have to specify a specific time and place for it to be actionable. These guys are very familiar with how those laws work and know exactly how to avoid getting caught by them.
You need to do your own homework. I’m not doing it for you. What I will do is lay this to rest:
https://en.wikipedia.org/wiki/Stable_Diffusion
Stable Diffusion is a latent diffusion model, a kind of deep generative artificial neural network. Its code and model weights have been released publicly […]
https://jalammar.github.io/illustrated-stable-diffusion/
The image information creator works completely in the image information space (or latent space). We’ll talk more about what that means later in the post. This property makes it faster than previous diffusion models that worked in pixel space. In technical terms, this component is made up of a UNet neural network and a scheduling algorithm.
[…]
With this we come to see the three main components (each with its own neural network) that make up Stable Diffusion:
- […]
https://stable-diffusion-art.com/how-stable-diffusion-work/
The idea of reverse diffusion is undoubtedly clever and elegant. But the million-dollar question is, “How can it be done?”
To reverse the diffusion, we need to know how much noise is added to an image. The answer is teaching a neural network model to predict the noise added. It is called the noise predictor in Stable Diffusion. It is a U-Net model. The training goes as follows.
[…]
It is done using a technique called the variational autoencoder. Yes, that’s precisely what the VAE files are, but I will make it crystal clear later.
The Variational Autoencoder (VAE) neural network has two parts: (1) an encoder and (2) a decoder. The encoder compresses an image to a lower dimensional representation in the latent space. The decoder restores the image from the latent space.
https://www.pcguide.com/apps/how-does-stable-diffusion-work/
Stable Diffusion is a generative model that uses deep learning to create images from text. The model is based on a neural network architecture that can learn to map text descriptions to image features. This means it can create an image matching the input text description.
Forward diffusion process is the process where more and more noise is added to the picture. Therefore, the image is taken and the noise is added in t different temporal steps where in the point T, the whole image is just the noise. Backward diffusion is a reversed process when compared to forward diffusion process where the noise from the temporal step t is iteratively removed in temporal step t-1. This process is repeated until the entire noise has been removed from the image using U-Net convolutional neural network which is, besides all of its applications in machine and deep learning, also trained to estimate the amount of noise on the image.
So, I’ll have to give you that you’re trivially right that Stable Diffusion does use a Markov Chain, but as it turns out, I had the same misconception as you did, that that was some sort of mathematical equation. A markov chain is actually just a process where each step depends only on the step immediately before it, and it most certainly doesn’t mean that you’re right about Stable Diffusion not using a neural network. Stable Diffusion works by feeding the prompt and partly denoised image into the neural network over some given number of steps (it can do it in a single step, although the results are usually pretty messy). That in and of itself is a Markov chain. However, the piece that’s actually doing the real work (that essentially does a Rorschach test over and over) is a neural network.
That word has been around since at least the 1980s.
You will never move a boat with nuclear,
I assume you haven’t heard of aircraft carriers and nuclear submarines.
Also, nuclear power can be stored in batteries and capacitors and then used to move electric vehicles (including boats, planes, and tractors), so I don’t know what the hell you’re even talking about.
Eat less meat! How hard is it to compute! So turn off your stupid AI and eat less meat. Do it now, stop eating meat.
I’ve actually cut my meat consumption way down.
That being said, a person using AI consumes an absolutely minuscule amount of power compared to a person eating a steak. One steak (~20kwh) is equivalent to about 60 hours of full time AI usage (300W for an nvidia A100 at max capacity), and most of the time a person spends using an AI is spent idling while they type and read, so realistically it’s a lot longer than that.
Again, your hypothetical data center smashers are going after AI because they hate AI, not because they care about the environment. There are better targets for ecoterrorism. Like my car’s tires, internet tough guy.
So to clarify, are you making the claim that nothing that’s simulated with vector mathematics can have emergent properties? And that AIs like GPT and Stable Diffusion don’t contain simulated neurons?
Except an AI is not taking inspiration, it’s compiling information to determine mathematical averages.
The AIs we’re talking about are neural networks. They don’t do statistics, they don’t have databases, and they don’t take mathematical averages. They simulate neurons, and their ability to learn concepts is emergent from that, the same way the human brain is. Nothing about an artificial neuron ever takes an average of anything, reads any database, or does any statistical calculations. If an artificial neural network can be said to be doing those things, then so is the human brain.
There is nothing magical about how human neurons work. Researchers are already growing small networks out of animal neurons and using them the same way that we use artificial neural networks.
There are a lot of “how AI works” articles in there that put things in layman’s terms (and use phrases like “statistical analysis” and “mathematical averages”, and unfortunately people (including many very smart people) extrapolate from the incorrect information in those articles and end up making bad assumptions about how AI actually works.
A human being is paid for the work they do, an AI program’s creator is paid for the work it did. And if that creator used copyrighted work, then he should be having to get permission to use it, because he’s profitting off this AI program.
If an artist uses a copyrighted work on their mood board or as inspiration, then they should pay for that, because they’re making a profit from that copyrighted work. Human beings should, as you said, be paid for the work they do. Right? If an artist goes to art school, they should pay all of the artists whose work they learned from, right? If a teacher teaches children in a class, that teacher should be paid a royalty each time those children make use of the knowledge they were taught, right? (I sense a sidetrack – yes, teachers are horribly underpaid and we desperately need to fix that, so please don’t misconstrue that previous sentence.)
There’s a reason we don’t copyright facts, styles, and concepts.
Oh, and if you want to talk about something that stores an actual database of scraped data, makes mathematical and statistical inferences, and reproduces things exactly, look no further than Google. It’s already been determined in court that what Google does is fair use.
As the technology improves, data centers that run AI will require significantly less cooling. GPUs aren’t very power-efficient for doing AI stuff because they have to move a lot of data around from their memory to their processor cores. There are AI-specific cards being worked on that will allow the huge matrix multiplications to happen in place without that movement happening, which will mean drastically lower power and cooling requirements.
Also, these kinds of protestors are the same general group of people who stopped nuclear power from becoming a bigger player back in the 1960s and 70s. If we’d gone nuclear and replaced coal, we almost certainly wouldn’t be sitting here at the beginning of what looks to be a major global warming event that’s unlike anything we’ve ever seen before. It wouldn’t have completely solved the problem, but it would have bought us time. An AI may be able to help us develop ideas to mitigate global warming, and it seems ridiculous to me to go all luddite and smash the machines over what will be a minuscule overall contribution to it given the possibility that it could help us solve the problem.
But let’s be real here; these hypothetical people smashing the machines are doing it because they’ve bought into AI panic, not because they’re afraid of global warming. If they really want to commit acts of ecoterrorism, there are much bigger targets.
So clearly we do agree on most of this stuff, but I did want to point out a possibility you may not have considered.
If we’re just talking about what you can do, then these laws aren’t going to matter because you can just pirate whatever training material you want.
This depends on the penalty and how strictly it’s enforced. If it’s enforced like normal copyright law, then you’re right; your chances of getting in serious trouble just for downloading stuff are essentially nil – the worst thing that will happen to you is your ISP will three-strikes you and you’ll lose internet access. On the other hand, there’s a lot of panic surrounding AI, and the government might use that as an excuse to pass laws that would give people prison time for possessing one, and then fund strict enforcement. I hope that doesn’t happen, but with rumblings of insane laws that would give people prison time for using a VPN to watch a TV show outside of the country, I’m a bit concerned.
As for the parent comment’s motivations, it’s hard to say for sure with any particular individual, but I have noticed a pattern among neoliberals where they say things like “well, the rich are already powerful and we can’t do anything about it, so why try” or “having universal health care, which the rest of the first world has implemented successfully, is unrealistic, so why try” and so on. It often boils down to giving lip service to progressive social values while steadfastly refusing to do anything that might actually make a difference. It’s economic conservatism dressed as progressivism. Even if that’s not what they meant (and it would be unwise of me to just assume that), I feel like that general attitude needs to be confronted.
AI is more than just ChatGPT.
When we talk about reinterpreting copyright law in a way that makes AI training essentially illegal for anything useful, it also restricts smaller and potentially more focused networks. They’re discovering that smaller networks can perform very well (not at the level of GPT-4, but well enough to be useful) if they’re trained in a specific way where reasoning steps are spelled out in the training.
Also, there are used nvidia cards currently selling on Amazon for under $300 with 24 gigs of ram and AI performance almost equal to a 3090, which puts group-of-experts models like a smaller version of GPT-4 within reach of people who aren’t ultra-wealthy.
There’s also the fact that there are plenty of companies currently working on hardware that will make AI significantly cheaper and more accessible to home users. Systems like ChatGPT aren’t always going to be restricted to giant data centers, unless (as some people really want) laws are passed to prevent that hardware from being sold to regular people.
Losing their life because an AI has been improperly placed in a decision making position because it was sold as having more capabilities than it actually has.
I would tend to agree with you on this one, although we don’t need bad copyright legislation to deal with it, since laws can deal with it more directly. I would personally put in place an organization that requires rigorous proof that AI in those roles is significantly safer than a human, like the FDA does for medication.
As for the average person who has the computer hardware and time to train an AI (bear in mind Google Bard and Open AI use human contractors to correct misinformation in the answers as well as scanning), there is a ton of public domain writing out there.
Corporations would love if regular people were only allowed to train their AIs on things that are 75 years out of date. Creative interpretations of copyright law aren’t going to stop billion- and trillion-dollar companies from licensing things to train AI on, either by paying a tiny percentage of their war chests or just ignoring the law altogether the way Meta always does, and getting a customary slap on the wrist. What will end up happening is that Meta, Alphabet, Microsoft, Elon Musk and his companies, government organizations, etc. will all have access to AIs that know current, useful, and relevant things, and the rest of us will not, or we’ll have to pay monthly for the privilege of access to a limited version of that knowledge, further enriching those groups.
Furthermore, if they’re using people’s creativity to make a product, it’s just WRONG not to have permission or to not credit them.
Let’s talk about Stable Diffusion for a moment. Stable Diffusion models can be compressed down to about 2 gigabytes and still produce art. Stable Diffusion was trained on 5 billion images and finetuned on a subset of 600 million images, which means that the average image contributes 2B/600M, or a little bit over three bytes, to the final dataset. With the exception of a few mostly public domain images that appeared in the dataset hundreds of times, Stable Diffusion learned broad concepts from large numbers of images, similarly to how a human artist would learn art concepts. If people need permission to learn a teeny bit of information from each image (3 bytes of information isn’t copyrightable, btw), then artists should have to get permission for every single image they put on their mood boards or use for inspiration, because they’re taking orders of magnitude more than three bytes of information from each image they use for inspiration on a given work.
Why do you think people will build data centers in Europe when they can build them elsewhere?
For something to be a fact, it needs to actually be true. AI is currently accessible to everyone.
Wow, you have this all planned out, don’t you?
If that’s what Europe is like, they’ll build their data centers somewhere else. Like the corrupt USA. Again, you’ll be taking away your access to AI, not theirs.
I’m not sure what you’re getting at with this. It will only be a privilege for these groups of we choose to artificially make it that way. And why would you want to do that?
Do you want to give AI exclusively to the rich? If so, why?
Is kbin a place where we just call everyone we don’t like “techbros”?
Also, working in open source means having a proper understanding of licensing and ownership. Open source doesn’t mean “free this and free that” – in fact, many AI based code assistance tools are actually hurting the open source initiative by not properly respecting the license of the code base it’s studying from.
Don’t be patronizing. I’ve been involved in open source for 20+ years, and I know plenty about licensing.
What you’re talking about is changing copyright law so that you’ll have to license content in order for an AI to learn concepts from that content (in other words, to be able to summarize it, learn facts from it, learn an art style, and so on). This isn’t how copyright law currently works, and I hope to god it stays that way.
For example, if you don’t own the right of the original copy of Star Wars, you obviously wouldn’t own any rights over the output of an upscaled Star Wars. Same goes for writing or other “transformative” media and it has been this way for a long time (see: audio sampling)
That’s not the same thing as training and AI on Star Wars. If you feed Star Wars into an upscaling AI, the AI is processing each frame and creating an output that’s a derivative work on that frame, and result of that isn’t something you would be allowed to release without a license. If you train it on Star Wars, the AI would learn general concepts from Star Wars, and not be able to produce an upscaled version of the movie verbatim (although depending on the AI, it may be able to produce images in the general style of Star Wars or summarize the movie).
An appropriate analogy for what’s going on here would be reading a book and then talking about the facts I learned from that book, which is in no way a violation of copyright law. If I started quoting long sections of that book verbatim, I would need a license from the author, but that’s not how AI works. It’s not learning the sentences those people type verbatim, it’s picking up concepts and facts from them. Even if I were to memorize the book from cover to cover, I would be in the clear as long as I didn’t actually start reproducing the book in some way. Neural networks are learning machines, not databases. Their purpose isn’t to reproduce information verbatim.
If you’re still not clear on the difference between training on data and processing it, let me know and I’ll try to clarify further.
Any vegan with half a brain knows that you need more than just fruit to be healthy. Assuming her death by infection is a result of her diet (which is possible, but we don’t know that), she died of being an idiot, not a vegan.