Comment by wolframhempel | Hacker News Reader

Comment by wolframhempel | Hacker News Reader

I think the underlying question is one of "degrees of derivation".

There's a famous Carl Sagan quote: “If you wish to make an apple pie from scratch, you must first invent the universe” which hints at the problem: Nothing is created in a vacuum.

Let's compare what Stable Diffusion does with what Franz von Holzhausen, head of design at Tesla, does. Franz didn't come into existence out of nothing and knew how to design cars. Instead he studied transportation design and worked at Volkswagen, General Motors and Mazda before joining Tesla. All these steps trained his (actual) neural network with inputs from copyrighted car designs.

Based on these inputs he was able to create the designs for the Model S, Model 3, Model X and others. Does this mean that Mazda can now levy a copyright lawsuit against Tesla? It could, based on the reasoning employed in some of the Stable Diffusion suits, but it won't based on the lack of similarity between the cars of both brands.

I believe that the law around AI will come to a similar conclusion. AI learning is neither fundamentally add odds with copyrighted material, nor confirming it. It will be a matter of degrees of derivation - how different is the output of the AI from its copyrighted inputs.

I can't help but feel like you're slightly anthropomorphising an algorithm. It's a really damn cool and powerful algorithm, don't get me wrong, but it's still not a person. At the end of the day, it's also not the algorithm "benefiting" from it, but the corporation using the algorithm.

It's also a bit hypocritical, because if you did the same thing to them as a human (in this example let's say be a Tesla copycat) you'd likely be sued into the ground because there's probably a patent somewhere in there.

The anthropomorphisation of the ML baffles me. This is not AI but a large ML model trained on often proprietary data. The whole discussion is ridiculous. There should be a lineage of model data so we should know what's the original source (at least for the "core" of the answer) in form of citations. I assume that the model loose that information during the training and that would be quite hard as everything is mixed together.

Is it really that baffeling? Take ChatGPT, a model specifically fine tuned on dialog interactions to seem more human like. Of course people are going to anthropomorphise it. And the whole thing about it not being AI but an ML model, I think we can let that one go. It didn't stop the term cloud computing ("it's just someone else's server") and it won't stop the term AI from being used for this tech.

It's baffling because it's being done by people who would otherwise insist on technical correctness and rebut any argument otherwise with "well, how do humans think?"

I think the parent poster has the right idea. This is not just anthropomorphization; it's an analogy.

They take away is contained in the first and last sentences.

Derivation is key.

Copyright protects original expressions, and copying means to reproduce (read and write) something. The analogy OP made is focused on the reading and writing done by humans and the reading and writing done by an algorithm.

Algos like stable diffusion are reading, and their user is controlling what is written.

If the user produces a work that is unique, but uses the style of a particular artist, that seems like it should be valid, since style is simply a process. It's how to create art, but it is not art, and processes are not subject to copyright based on copyright.gov.

With all that said legality isn't morality, and I sympathize with the artist.

I guess. Imagine you're an artist, and you've created your career based around your unique art style. It gets you jobs as a concept artist, or whatnot (idk). Now, imagine there's a tool someone "trained" in a few hours to do what you you do, but withi seconds instead of hour or days. Is there any reason to pay you, ever again? Your hourly rate is surely going to be higher than their electricity bill for running the algorithm.

So, how do we protect you (as the artist) from this? Copyright, even if flawed, currently protects you from that

No. You cannot copyright an abstract style. Process is not copyrightable.

Sorry, I realise how my comment was a bit confusing. I kinda butchered my point.

I guess my point is more ethical than legal.

Sure, another artist could copy your style, but they would still need to study it for a long time (composition, palette, perspective, grading, line work, etc) to get an accurate understanding of it. They also needed to have spent years training their hand-eye coordination, as well as art theory to achieve it. Whether it's ethical is debatable, but they've earned their skills. And they'd still take a while to produce it, so they earned their money.

If someone without training just asked a computer program to produce "dragon in style of X" then this means people can sidestep artists to create works.

One could be smug and say "work smart not hard" here, but it creates a tricky situation.

What if people start pulling down their works or not posting them for fears a megacorp will crawl their works without permission?

What if someone asks for the same piece (by describing it as a prompt and saying "in the style of X"); is that plagiarism or copytheft?

What will happen to the models in this case? Do they become "inbred" over time?

What becomes of artists? It will no longer be a viable career (especially digital artists) for them. I know most artist enjoy the process more than selling the art, but the process won't put food on the table.

I guess it's a lot of what ifs too here, but they're not unlikely scenarios.

It's not an analogy because it doesn't analogize "learning" which is the thing that's being used as a defense. It hides any criticism of learning by implying that it's the same learning that humans do. That's not an analogy, it's just bullshit. It's bullshit word salad. A really tiny salad, here.

The key point is: an AI calls itself an intelligence but this so far is quite a bit of marketing. A human is considered an intelligent being. Where a process of creating new things happen. They are based on true learning and not just reproduction. And where they have been just reproduction, of course it went to the courts.

I'm really unsure if there is a qualitative difference between a human looking at lots of images, deriving patterns and recompiling them into a new image or a computer doing the same.

Currently there is definitely a wide gap because the current "AI" are completely incapable of a real intellectual process. Wheras humans can develop a true understanding and consequently it is an actual learning and not memorizing process. Of course there are overlaps and consequently from time to time there are law suits about copyright infringements by artists.

The pure fact that Stable Diffusion tends to produce 3 legged humans shows the complete lack of understanding of its doings.

I'm familiar with this line of reasoning, but I'm always struggling to understand the exact thing that humans can do that computers don't. Usually, the differentiation is that humans have e.g.

"true understanding"

I assume you mean the process of looking at an image and not just deriving patterns, but seeing that you are looking at a cat, that a cat is an "animal" which has "four legs and a tail" and that cats can be friendly towards you or aggressive, depending on your own behavior and theirs.

Neural Nets are certainly capable of the first two: classification and creating taxonomies. The last one I admit is tricky as it requires the Neural Net to be an entity within the observed world

"intellectual process"

the intellectual process is arguably exactly the process input->categorize and analyze->compile->produce output loop that we've modelled AI based upon

"creativity"

is the ability to create something truly new. This one seems obvious as Neural Nets only can derive patterns (plus maybe a random input) - but I would posit the question if any human ever created something truly new in the "apple pie from scratch" sense or if we've only ever created higher level works derived from existent things.

"consciousness"

this one is hard to grasp. I would argue that consciousness is the realization that one exists (in the descartian sense) - coupled with the desire to continue to do so. It is a quality that wouldn't make much sense for an output focused neural net like the one behind Stable Diffusion - but it might be a desirable trait in a decision making focused deep learning setup - similar to a self-healing cloud deployment.

"love/emotion"

This builds on the previous consciousness example. Not to sound like Rick Sanchez /some other cynic - but aren't these at their core adjustment mechanisms that help us further evolutionary goals like survival and continuation of our lineage. Wouldn't a decision making focused deep learning setup be more stable/have a higher uptime if it would facilitate its goal of "staying on" through a strong drive of survival/expansion?

The last two examples are where my point falls apart a bit. But I still stand by my general thesis: We are way too certain that our particular human way of processing information and "thinking" has some divine quality to it that isn't replicable in neural networks. Against that, I would argue that neural networks are largely the same mechanism we employ in our thinking and that they are just a couple of millenia in evolution behind, but are catching up at a multiple of the speed it took us to get to where we are now intellectually.

Regression =/= memorization and is exactly how humans synthesize information as well.

Do you think there is there a qualitative difference between the human and the computer?

I guess the question is, what exactly is "true learning" and why is that impossible for a machine to achieve?

Keep in mind, we don't know how humans learn either (on a neurological level). It might end up being that we stumbled onto the same general idea, using matrices and linear algebra instead of neurones, synapses and neurotransmitters.

>There's a famous Carl Sagan quote: “If you wish to make an apple pie from scratch, you must first invent the universe” which hints at the problem: Nothing is created in a vacuum.

See also the story of Trurl's Electropoet (from the Cyberiad). Trurl first had to simulate a universe to get the poet to work.

Reply on news.ycombinator.com