Hacker News Reader

Comment by senko | Hacker News Reader

Calling a LLM "a tool for plagiarism" takes some real mental gymnastics. You could similarly argue IKEA is selling lethal weapons because you can kill someone with a knife.

The exact way copyright laws will apply (or not) to AI is still ambigous, until a high profile case gets before the Supreme Court or ECJ. You can certainly argue one way or another - my view is that AI learning from your blog post is similar to me learning from your blog post. That's not the problem. If I then plagiarise it (or copy wholesale), that is the problem.

View on HN

Only problem is that LLMs don't "learn" from blog posts in any human sense of the word, they are incapable of symbolic reasoning. LLMs directly incorporate the "learning" material into their models, with superhuman memory capacity and recollection ability, and then remix that source content into their productions. The production might be original, depending on its specific circumstances, but the model and the service itself are without any legal doubt derivative works of the original.

They have very little to do with the way humans learn - except maybe if you read a chapter from a book, memorize it, and then proceed to rewrite verbatim with the names of the characters exchanged, a use case of human learning which has long been recognized as, you guessed it, "plagiarism".

> Only problem is that LLMs don’t “learn” from blog posts […] LLMs directly incorporate the learning material into their models

The "in-context learning" (regardless of how you feel about the use of that term) that they do from material incorporated into their prompts does not involve incorporating it into their model.

> The production might be original, depending on its specific circumstances, but the model and the service itself are without any legal doubt derivative works of the original.

Model code is a work, but not derivative of the training data. There’s considerable legal doubt that the models weights themselves are works of authorship at all, which is a prerequisite for being derivative works, so no matter if you are referring to model code or model weights, the statement that model is “without any legal doubt” a derivative work of the training data is false.

You are splitting hairs, there is no doubt that transformers can incorporate and reproduce verbatim training data. Just ask ChatGPT to tell you the lyrics of a famous Madonna song.

On that basis alone they are derivative works, regardless of how that data is represented internally.

> LLMs directly incorporate the "learning" material into their models.

No, that's not how it works.

> read a chapter from a book, memorize it, and then proceed to rewrite verbatim with the names of the characters exchanged,

That's how a bayesian markov chain would work, not transformers.

If ikea would be selling machetes disguised as kitchen knifes, then yeah they would be called out and their product banned.

You learning from my blog is not the same as a piece of software copying my work. Ai does not “learn” as humans do regardless of how well it mimics our behaviour.

My laptop outputs audio, it doesnt sing.

Well an LLM certainly doesn't store the training data in a compressed format, because otherwise it'd be sci-fi levels of compression. So what are you suggesting it is doing?

I wouldn't be so sure that there isn't "any" form of compression going on - that just seems silly. It seems that at some level it is *fantastic* lossy semantic compression.

This is one of the more fascinating aspects of LLMs from a computer science perspective - almost all of human knowledge and experience in text from (several petabytes of data in a database) can be 'cleverly semantically compressed' to fit into a database of just a few hundred GB (NN params). That's some astounding lossy compression algorithm, unfortunately with a less than efficient decoder (temp=0).

However, it isn't always 'plagiarism' either. 'Memorization' vs 'Plagiarism' vs 'understanding' is actually very tricky in LLMs, because having a good world model requires memorizing some facts, as well as some (potentially) "ontological" constructions. In addition, quoting a famous person is very useful for imparting some wisdom occasionally, which is also memorization, and can be seen as plagiarism if the author is left off. But sometimes this occurs (common sayings) and no one bats an eye. One good example of why explicit memorization is really needed is to provide exact references (list of authors, year, and article title) - so clearly making a NN not rewarded for memorizing is not ideal. Hence, *very occasionally*, you get some 'bad memorization'. But it is worth noting that the alignment achieved with these models for such an enormously complex task is quite good.

An audio file doesnt store musical notes and a digital book does not store ink and paper. Using algorithms either of the two can however reproduce audio and visual content sufficiently similar, albeit not identical.

An llm or generative model doesnt store audio files and digital books, compressed or not. But using algorithms it can reproduce audio and visual content sufficiently similar, but unlike my previous example it’s also made of bites.

Just like an audio file and digital book are not copies of the physical product, an ai “data source” is not an identical or compressed copy of the original bites.

That’s why it can be used to generate output identical to the original data even tho the data it uses is not a verbatim copy.

It’s a new way of storing data, optimised for stochastic generation.

It is a powerful tool but still relies on original content.

Ya but we have not outlawed tools that create audio files nor have we outlawed audio formats. So?

No one is asking for outlawing ai, that would be ridiculous. But as we regulate what we can do with audio files and audio creation tools we can also regulate what we can do with ai.

My laptop sings just fine, it's the people doing weird stuff with their mouths and breathing hard.

I suppose we could redefine singing to match the concept of a singing laptop.

OpenAI does plagiarize stuff wholesale though. I was looking up a fairly niche Python library about a month ago using ChatGPT and the code example it provided was too specific to be useful for what I was doing. So, I did a little googling and came across the blogpost it had sourced it from. It had the exact same code.

It seems like a tool for plagiarism. You could do the same thing before LLM grew in popularity by using Google Translate back and forth from English.

Reply on news.ycombinator.com