Hacker News Reader

Comment by stefano | Hacker News Reader

How do you guarantee it doesn't copy a GPL-ed function line-by-line?

Yup, this isn't a theoretical concern, but a major practical one. GPT models are known for memorizing their training data: https://towardsdatascience.com/openai-gpt-leaking-your-data-...

Edit: Github mentions the issue here: https://docs.github.com/en/github/copilot/research-recitatio... and here: https://copilot.github.com/#faq-does-github-copilot-recite-c... though they neatly ignore the issue of licensing :)

> GPT models are known for memorizing their training data

Hash each function, store the hashes as a blacklist. Then you can ask the model to regenerate the function until it is copyright safe.

What if it copies only a few lines, but not an entire function? Or the function name is different, but the code inside is the same?

If we could answer those questions definitively, we could also put lawyers out of a job. There’s always going to be a legal gray area around situations like this.

Matching on the abstract syntax tree might be sufficient, but might be complex to implement.

You can probably tokenize the names so they become irrelevant. You can ignore non-functional whitespace, so that code C remains. Maybe one can hash all the training data D such that hash(C) is in hash(D). Some sort of Bloom filter...

That second link says the following:

> We found that about 0.1% of the time, the suggestion may contain some snippets that are verbatim from the training set

That's kind of a useless stat when you consider that the code it generates makes use of your existing variable/class/function names when adapting the code it finds.

I'm not a lawyer, but I'm pretty sure I can't just bypass GPL by renaming some variables.

It's not just about regurgitating training data during a beam search, it's also about being a derivative work, which it clearly is in my opinion.

Surprised not to see more mention of this. It would make sense for an AI to "copy" existing solutions. In the real world, we use clean room to avoid this.

In the AI world, unless all GPL (etc.) code is excluded from the training data, it's inevitable that some will be "copied" into other code.

Where lawyers decide what "copy" means.

It's not just about copying verbatim. They clearly use GPL code during training to create a derivative work.

Then you have the issue of attribution with more permissive licenses.

How do you know that when you write a simplish function for example, it is not identical to some GPL code somewhere? "Line by line" code does not exist anywhere in the neural network. It doesn't store or reference data in that way. Every character of code is in some sense "synthesized". If anything, this exposes the fragility of our concept of "copyright" in the realm of computer programs and source code. It has always been ridiculous. GPL is just another license that leverages the copyright framework (the enforcement of GPL cannot exist outside such a copyright framework after all) so in such weird "edge cases" GPL is bound to look stupid just like any other scheme. Remember that GPL also forbids "derivative" works to be relicensed (with a less "permissive" one). It is safe to say that you are writing code that is close enough to be considered "derivative" to some GPL code somewhere pretty much every day, and you can't possibly prove that you didn't cheat. So the whole framework collapses in the end anyways.

> How do you know that when you write a simplish function for example, it is not identical to some GPL code somewhere?

I don't, but then I didn't go first look at the GPL code, memorize it completely, do some brain math, and then write it out character by character.

I truly don't think they can guarantee that. Which is a massive concern.

Reply on news.ycombinator.com