If the LLM has ‘seen’ the GPL-licensed code, it is debatable whether a reimplementation by that LLM is from scratch.

It doesn’t help the “yes it is; the LLM learned from reading that code, but didn’t copy it” camp that there have been cases where an LLM output blocks of code that are very similar to parts of GPL-licensed source code.

Question for copyright to apply then is whether those blocks are copyrightable and if so, whether the ‘copied’ text is similar enough to the source text.

▲

theturtletalks a day ago | parent | next [-]

But a lot of these reverse engineered projects claim “clean lab.” If that LLM was only given documents and didn’t have web search or access to the copy-left repo, cloning it is reverse engineering according to the courts.

Even if that model has been trained on that copy left code.

	▲	Someone a day ago \| parent [-]
		> If that LLM […] didn’t have web search or access to the copy-left repo, […] > Even if that model has been trained on that copy left code How can an LLM be trained on code that it doesn’t have access to? > cloning it is reverse engineering according to the courts. What courts? Has there been any court case on this w.r.t. LLMs? AFAIK, learning from copyrighted material is fine, copying copyrighted material (with some exceptions) isn’t. Where one ends and the other starts is a gray area, but at the far end of that scale lies learning by head and then reproducing verbatim. That definitely is copyright infringement. Also, claiming “clean lab” isn’t worth much. Question is whether that holds up in court. I would guess/hope/think one need not go as far as Phoenix technologies went to make a BIOS (https://en.wikipedia.org/wiki/Clean-room_design), but I can see why they did things that way to be on the safe side.

▲

snvzz a day ago | parent | prev [-]

By that logic, anybody who's learned operating systems development via the book Operating Systems: Design and Implementation (Tanenbaum and Woodhull) would be legally unable to write an operating system, ever. That book had Minix source code, which at the time was, nevermind copyleft, proprietary.

AIUI this includes Linus Torvalds, and would have affected Linux. Which, by the way, is heavily written by LLM these days, including by Linus himself using Anthropic's tools.

An LLM reading some code and learning from it (at any point in time, in training or as used through agentic programming tools) is in this regard no different from an human being doing the same thing.

Nothing stops me from reading Asahi's source code and then writing my own from what I learned. My agent can do the same thing faster and better. It would also be less likely to resemble Asahi's, as an LLM has seen and learned from a far larger body of code than I have in my entire 30+ years of programming.

If anything, code is becoming no longer a product but a commodity, which brings us closer to the sort of freedom GNU and the FSF aimed for; Something to celebrate.

No, there's no way the genie is going back into the bottle. The past is not coming back. Reality is best accepted and worked with, not fought.

▲

jasomill 9 hours ago | parent [-]

Much of this is far from settled law. All it takes is one law declaring code produced by LLMs to be a derivative work of training data for the entire house of cards to collapse.

Not saying this is likely, but it's possible, and the combined market power of many industries dwarfs the market power of the fledgeling LLM industry, so it's not necessarily unlikely, either.

In other words, it's a significant risk.

	▲	snvzz 7 hours ago \| parent [-]
		>Much of this is far from settled law. All it takes is one law declaring code produced by LLMs to be a derivative work of training data for the entire house of cards to collapse. This is, realistically, not happening. We aren't going back to the stone age because of some dumb judge. Especially not worldwide. Even if some country is left behind momentarily, this won't prevent other countries from continuing to push the frontier forward.