Hacker News Reader

Comment by inglor_cz | Hacker News Reader

I think the concerning problem is when the LLM reproduces some copyrighted code verbatim, and the user doesn't even stand a chance to know it.

View on HN

Yes, but that's not what the grandparent comment was talking about.

If I’m the grandparent comment, it was a big part of what I mean. Stolen/Unknown content goes in for training, verbatim or very close “inspired by” code comes out and there is no way to verify the source - “violation as a service”.

Verbatim dumping is one thing but otherwise this seems closer to the issue of plagiarism than copyright. If someone studies the Linux kernel and then builds a new kernel that follows some of the design decisions and idioms that's not really copyright infringement.

The bigger issue (spiritually anyway) seems to be the need to develop free software LLM tools the same way FSF needed to develop free compilers. That's what's going to keep users from being able to adapt and control their machines. The issue is more ecological that programmers equipped with LLM are likely much more productive at creating and modifying code.

Some of the rest seems more like saying that anyone who studies GCC internals is forever tainted and must write copyleft code for life which seems laughable to me. Again this is more a topic of plagiarism than copyright which are fairly similar but actually different and not as clear cut.

> more a topic of plagiarism than copyright

You’re right, in the context of a technical legal interpretation they’re different. In the context of right or wrong, they amount to the same.

> anyone who studies GCC internals

LLMs are not a someone, they’re more like … the indigo printout of some text or design, you then use to make a scrapbook to be mass produced for profit. Very different situation.

When the AI bubble pops, I hope we will have some equalisation back to something more ethical.

LLMs are not a someone

This is in line with my disagreement over the fair use rulings. Most people who published works that have been used to train AI systems, created those works and published them for other people to consume and benefit from, not for proprietary software systems to consume and benefit from. The existing licenses and laws did not account for this; nobody was anticipating it.

I don't know... there's a pretty clear difference between copyright and a say a utility patent or trade secret. The right and wrong in FSF isn't about labor, it's about control over machines and ability to modify. Free software has never tried to control the community using patents and trade secrets and in general are rather hostile to them. In fact its fairly contemptuous of copyright and uses it from a purely utilitarian perspective. And frankly FSF is not opposed to commercial software. They're opposed to users being unable to modify machines and software that they are using. That's the core of the ethics. See the origins in that damn printer firmware RMS did battle with.

But I also disagree in general about LLMs. LLMs are statistical text models but the general concept of what if there were an "AI" that wasn't a LLM and was trained on open source software is the same at the end of the day. I think whether or not LLM are intelligent or equivalent to humans is a red herring. There's no reason to not consider the implications of machines that are indistinguishable or even superior to human programmers. Particularly if we're discussing ethics getting lost in implementation details seems like a distraction and then all the derived ethics gets thrown out after the next innovation.

When that happens, it's because the code was trivial enough to be compressed to a minuscule handful of bits... either because it literally is trivial, or because it's common enough to have become part of our shared lexicon.

As a society, we don't benefit from copyright maximalism, despite how trendy it is around here all of a sudden. See also Oracle v. Google.

Quake's sqrt approximation is not trivial and is not common.

[1] https://www.reddit.com/r/programming/comments/oc9qj1/copilot...

(Shrug) It's a math trick, documented by Abrash among others and very heavily discussed on forums such as this one. And it didn't originate in the Quake codebase. Like much IEEE754 hackery, it goes back to the father of IEEE754 himself, William Kahan.

Nobody benefits from a law that says that LLMs can't regurgitate the Quake sqrt() approximation. If that's what the law actually says, which it isn't.

Reply on news.ycombinator.com