Comment by dathinab | Hacker News Reader

Comment by dathinab | Hacker News Reader

Parent Full thread

the author speaks about code which is syntactically completely different but semantically does the same

i.e. a re-implementation

which can either

- be still derived work, i.e. seen as you just obfuscating a copyright violation

- be a new work doing the same

nothing prevents an AI from producing a spec based on a API, API documentation and API usage/fuzzing and then resetting the AI and using that spec to produce a rewrite

I mean "doing the same" is NOT copyright protection, you need patent law for that. Except even with patent law you need to innovations/concepts not the exact implementation details. Which means that even if there are software patents (theoretically,1) most things done in software wouldn't be patentable (as they are just implementation details, not inventions)

(1): I say theoretically because there is a very long track record of a lot of patents being granted which really should never be granted. This combined with the high cost of invalidating patents has caused a ton of economical damages.

No, that depends on whether or not the AI work product rests on key contributions to its training set without which it would not be able to the the work, see other comment. In that case it looks like 'a new work doing the same' but it still a derived work.

Ted Nelson was years ahead of the future where we really needed his Xanadu to keep track of fractional copyright. Likely if we had such a mechanism, and AI authors respected it then we would be able to say that your work is derived from 3000 other original works and that you added 6 lines of new code.

No, training and inference are two separate processes. Training data is never redistributed, only obtained and analyzed. What matters is what data is put into context during inference. This is controlled by the user.

AI/ML is complex, so as a simpler analogy: If I watch The Simpsons, and I create an amusing infographic of how often Homer says "D'oh!" over time, my infographic would be an original work. AI training follows the same principle.

> my infographic would be an original work.

> AI training follows the same principle.

If you really believe that then we can't have a meaningful conversation about this, that's not even ELIF territory, that's just disconnected. You should be asking questions, not telling people how it works.

How exactly is it different? All the model itself is is a probability distribution for next token given input, fitted to a giant corpus. i.e. a description of statistical properties. On its own it doesn't even "do" anything, but even if you wrap that in a text generator and feed it literal gcc source code fragments as input context, it will quickly diverge. Because it's not a copy of gcc. It doesn't contain a copy of gcc. It's a description of what language is common in code in general.

In fact we could make this concrete: use the model as the prediction stage in a compressor, and compress gcc with it. The residual is the extent to which it doesn't contain gcc.

There already have been multiple documented cases of LLMs spitting out fairly large chunks of the input corpus. There have been some experiments to get it to replicate the entirety of 'Moby Dick' with some success for one model but less success with others most likely due to output filtering to prevent the generation of such texts, but that doesn't mean they're not in there in some form. And how could they not be, it is just a lossy compression mechanism, the degree of loss is not really all that relevant to the discussion.

Are you referring to this?

https://osyuksel.github.io/blog/reconstructing-moby-dick-llm...

I see a test where one model managed to 85% reproduce a paragraph given 3 input paragraphs under 50% of the time.

So it can't even produce 1 paragraph given 3 as input, and it can't even get close half the time.

"Contains Moby Dick" would be something like you give it the first paragraph and it produces the rest of the book. What we have here instead is a statistical model that when given passages can do an okay job at predicting a sentence or two, but otherwise quickly diverges.

I'm no longer certain what point you're trying to make.

Getting close less than half the time given three paragraphs as input still sounds like red-handed copyright infringement to me.

If I sample a copyrighted song in my new track, clip it, slow it down, and decimate the bit rate, a court would not let me off the hook.

It doesn't matter how much context you push into these things. If I feed them 50% of Moby Dick and they produce the next word, and I can repeatedly do that to produce the entire book (I'm pretty sure the number of attempts is wholly irrelevant: we're impossibly far from monkeys on typewriters) then we can prove the statistical model encodes the book. The further we are from that (and the more we can generate with less) then the stronger the case is. It's a pretty strong case!

That's... not how this works.

> If I feed them 50% of Moby Dick and they produce the next word and I can repeatedly do that to produce the entire book... then we can prove the statistical model encodes the book.

It can't because it doesn't. That's what it means to say it diverges.

The "number of attempts" is you cheating. You're giving it the book when you let it try again word by word until it gets the correct answer, and then claiming it produced the book. That's exactly the residual that I said characterizes the extent to which it doesn't know the book. Trivially, no matter how bad the model is, if you give it the residual, it can losslessly compress anything at all.

If you had a simple model that just predicts next word given current word (trained on word pair frequency across all English text, or even all text excluding Moby Dick), and then give it retries until it gets the current word right, it will also quickly produce the book. Because it was your retry policy that encoded the book, not the model. Without that policy, it will get it wrong within a few words, just like these models do.

The point is that the AI's themselves and their backers are on the record as saying that the AI could reproduce copyrighted works in their entirety but that there are countermeasures in place to stop them from doing so.

I wonder what the results would be if I spent time to train a model up from scratch without any such constraints. But I'm much too busy with other stuff right now, but that would be an interesting challenge.

Yeah just like a star could appear inside of Earth from quantum pair production at any given moment. But realistically, it can't. And you can't even show a test where any model can get more than a few tokens in a row correct.

These companies just don't want to deal with people complaining that it reproduces something when they don't understand that they're literally giving it the answer.

You do realize you are now arguing against your own case don't you?

But it does encode it! Each subsequent token's probability space encodes the next word(s) of the book with a non-zero probability that is significantly higher than random noise.

If you had access to a model's top p selection then I'd bet the book is in there consistently for every token. Is it statistically significant? Might be!

I'm not cheating because the number of attempts is so low it's irrelevant.

If I were to take a copyrighted work and chunk it up into 1000 pieces and encrypt each piece with a unique key, and give you all the pieces and keys, would it still be the copyrighted work? What if I shave off the last bit of each key before I give them to you, so you have a 50% chance of guessing the correct key for each piece? What if I shave two bits? What if it's a million pieces? When does it become transformative or no longer infringing for me to distribute?

The answer might surprise you.

Your test is more like the following:

Consider a password consisting of random words each chosen from a 4k dictionary. Say you choose 10 words. Then your password has log_2(4k)*10 entropy.

Now consider a validator that tells you when you gets a word right. Then you can guess one word at a time, and your password strength is log_2(4k*10). Exponentially weaker.

You're constructing the second scenario and pretending it's the first.

Also in your 50% probability scenario, each word is 1 bit, and even 50-100 bits is unguessable. A 1000 word key where each word provides 1 bit would be absurdly strong.

You're still missing the point. The numbers don't matter because it's copyright infringement as long as I can get the book out. As long as I know the key, or the seed, I can get the book out. In court, how would you prove it's not infringement?

Because you put the book in. Again, this is measurable. Compress the book with a model as the predictor. The residual is you having to give it the answer. It's literally you telling it the book.

For an infographic, perhaps you claim claim fair use. I think it makes a lot of sense, but IANAL.

For a fan fiction episode that is different from all official episodes, you may cross your fingers.

For a remake of one of the episodes with a different camera angle and similar dialog, I expect that you will get in problems.

Is the claim that these models can 1 shot a Simpsons episode remake with different camera angle and similar dialog from a prompt like "produce Simpsons episode S01E04"? Or are we falling into the "the user doesn't notice that they told the model the answer, and the model in fact did not memorize the thing" trap?

Reply on news.ycombinator.com