Hacker News Reader

This is the difference between the law and the spirit of the law. Copyright is there to incentivize creation and innovation, and the large-scale use of AI might reasonably go against this incentive. The law doesn't exist in a vacuum where is it the word of god and We Have To Obey It. It has a purpose, and we can choose to adapt it to better suit our needs.

A note : RNG means it's random. An LLM's output is not random, it's based on previously look-at data (which is the same way humans function, if we're being honest).

>The law doesn't exist in a vacuum where is it the word of god and We Have To Obey It. It has a purpose, and we can choose to adapt it to better suit our needs.

And this is why this agreement is so important for the future of AI even though it is really only governs movies and TV. Hollywood is a major political force. We all know the story about how Disney has repeatedly pushed for extensions on the life of copyrights. This deal will likely give us good indication into how active and in what areas Hollywood will participate in shaping the law around AI.

> An LLM's output is not random, it's based on previously look-at data

Just because the set of tokens that are considered to be output next have weighted probabilities, does not mean the one selected isn't "random".

How is the selected token "random" if it's chosen based on weighted probabilities, especially if you use a temperature of 0?

Because even with weighted probabilities, they're still probabilities. If you have "the cat sat on the", and the next possible tokens are "mat" with a probability of 90%, "floor" with a probability of 9%, and "dog" at 1%, how are you going to pick "mat" 90% of the time without consulting a (P)RNG?

They are only probabilities if you treat them as such. If you always just pick the highest probability option (temperature = 0), it's not random in any way - then it's just a confidence value.

You can use randomness, but it's in no way inherent to the process.

> If you always just pick the highest probability option (temperature = 0), it's not random in any way

Huh. That feels like a recipe for either outputting parts of the training data verbatim, or getting stuck in an output loop.

Although I earlier gave an example with 90/9/1 weights where the obvious choice is significantly higher than the others, I'd have thought that in a different scenario where you have, say, 51/49 or 34/33/32 options to choose from, always picking the one with only a fractionally higher confidence than the reasonable alternatives could lead to a "sameyness" in the outputs. Isn't there any value in having variance in the generated responses?

> Isn't there any value in having variance in the generated responses?

Sure, which is why the technique is used. But that doesn't mean "random" is a useful descriptor. I, as a human, try to add variance to my responses as well - does that mean my mind is "random"?

That's why my point is that there is no inherent randomness in the process. We use a bit of randomness to make it subjectively better, but it's fundamentally not random.

Copyright is a special protection under law to incentivize people to produce creative works. Machines don’t need this incentive and thus don’t need the protection. It’s fairly straightforward.

You still prompt AI.

But your prompt is merely a suggestion to the AI model. You aren't the one making the image.

This is stemming from a conviction about where the difficulty in the creation process is at. That changes with technological advances. Automation often makes practical skills economically worthless.

Art is separate from economics. It has existed before there were markets or even complex societies. Also, the amount of effort is not what determines art. However, I think art requires one to be the actor in the process that generates the work. The big issue I have with AI art is that you could have very clear image in mind that you wish to generate, and spend lots of time tweaking your prompt to the model, but you will never get the image you had in mind, because there is disconnect between you and the machine. Your imagination doesn't extend across the screen.

>The big issue I have with AI art is that you could have very clear image in mind that you wish to generate, and spend lots of time tweaking your prompt to the model, but you will never get the image you had in mind, because there is disconnect between you and the machine. Your imagination doesn't extend across the screen.

Couldnt you say the same about pen and pencil? Or any other medium? Turning imagination into something physical/digital/real always entails loss. This might be a matter of a few more iterations of better usability. If you are able to put the the limits of the machine representation of your imagination into words they might iteratively vanish. You wont connect your imagination to a machine but to words.

Posts were also about creativity and design in the context of copyright thus the focus on economics. Art for its own sake as a personal pursuit doesnt really relate to that perspective. Unless you want to trade in other currencies (fame and number of impressions) in which case its the age old competition of what is able to speak to people on a deeper level. For which the representation is often less important then the idea / perspective and how that is able to resonate with people. For that the medium is just that, a medium.

I place take-out orders as well. Doesn’t make me a chef.

Given sufficient technological advances some skills might become practically meaningless in the economy. The idea might be where its at. Handloomers also thought they brought something worth while to the table.

Most takeout today is also warmed up according to strict recipe. "Sandwich artists" arent chefs either. And even the initial recipe is designed for mass reproducibility and supply chain feasibility over taste. You still get food but a lot fewer people employed as actual chefs doing chef stuff.

edit: Removed unsensible comparison

And prompts can be copyrighted, like computer code. That is not what the conflict over AI is about.

You can copyright the prompt but not the result generated by AI when fed the prompt

> "Next, you plug it into a computer and have the computer use a RNG to determine how to shake it. Still copywritable. Lastly, you have an LLM generate a sequence of angles and velocities with which to shake the brush. No longer copywritable? Why?"

Consider an alternative image generation technique where the same computer is connected to a camera, and that video feed drives the brush strokes. Let's assume it imitates the Pollock style so well that if you place an image of a Pollock painting in front of the camera, the computer will produce what looks like a near-perfect painted replica of the original.

Clearly this technique would have a problem with copyright. Programming a computer to make copies doesn't absolve you of copyright.

In the eyes of the law currently, image diffusion models are seen as something closer to that hypothetical camera input that waves the brush, in the sense that the AI model has been trained directly on the works of specific artists and can recall stylistic aspects of their works on command.

Because that LLM does not exist without tremendous amounts of input created by actual human beings. These oversimplifications always skip this step, and pretend LLMs sprang into being through sheer force of will. They did not. They are built on the output of other people, and using a statistical approach to copyright laundering remains copyright laundering.

That is betting LLMs wont get enough non copyrighted training material to be useful. Especially for text thats likely not the case. Pretty sure its a moot point anyway as you have no way to proof what was in the training set and what wasnt.

I very much expect parallel construction if push comes to shove.

An LLM trained on public-domain or liberal-license material would be a great counterexample ... if anyone built it. The popular models in wide use do disclose what they train on (sometimes under NDA) and they all contained copyrighted work.

There is no need to build it or tell you atm. Once that does change i expect parallel construction.

Once you know what behavior you want from the net, recreating that from non-copyrighted material might turn out to be a lot less difficult.

It was my understanding that the debate was never about whether AI work is copyrightable. The debate is whether the work can be attributed to the AI. In your example, it would be like attributing a work to the brush, rather than to the artist.

The problem of reducing AI to a tool is that your hammer has no intelligence. There are more complex tools, you might consider your computer a tool, once again it doesn't have intelligence albeit it can achieve very complex tasks. I think that your example of "you have an LLM generate a sequence of angles and velocities with which to shake the brush" is copywritable. In that sense you used it as a tool, as you would use a RNG. But if instead the LLM created the whole outcome, then it shouldn't be copywritable.

PS: I use the term "intelligence" loosely here for the sake of the argument

You might be too ahead of your time in this thinking, but I see the logic there.