logoalt Hacker News

crote • yesterday at 6:49 PM • 1 reply • view on HN

> The question remains if LLM-assisted design is actually ‘clean’ or not

I genuinely don't understand how one could possibly make that argument.

The entire point of clean-room design is to kill every possibility of copyright infringement before it even starts, by completely separating the source and destination. Everyone who has ever read the copyrighted source is considered "tainted" and may not ever contribute any code to the "clean" destination. The readers and writers are two completely distinct teams, with only the known-clean specification passing between them.

LLMs, on the other hand, are fundamentally nothing more than a "code blender". You mix together giant piles of "tainted" code, and then statistically pull some combined fragments out of it. Following the clean-room legal argument, it should be treated as if everything coming out of it is tainted. There is, by definition, absolutely zero separation between input and output - and the early models verbatim reproducing popular F/LOSS code made that pretty obvious. LLMs are the literal opposite of a clean-room design!

Granted, going for the clean-room approach is overkill - and intentionally so. It is the nuclear option chosen when you believe the risk of getting sued is substantial. It is perfectly possible to write non-infringing code without using a clean-room approach, it is just a lot harder to defend in court when you get sued - be it justifiably or not. But "LLMs are non-clean-room non-infringing" is quite a different argument from "LLMs are clean-room, therefore non-infringing".


Replies

xoa • yesterday at 7:43 PM

I agree with you overall, but it's an interesting question to turn over and contemplate.

>I genuinely don't understand how one could possibly make that argument.

With perhaps a bit of genuine irony inasmuch as it's a real reversal from most of computing history, highly locked down proprietary software might actually be a place where a case could be made even with current LLMs. After all, have any LLMs (or anyone else who hasn't signed an NDA for that matter) ever seen the source code of iOS 27 or other Apple crown jewels for example? So if someone used one LLM to produce the factual spec of a piece of software whose code is not part of its training data, humans to verify it was indeed just an uncopyrightable factual spec, and then a separate isolated freshly initialized LLM to produce something exclusively from said spec, it might well work with respect to that specific target? Like yes, the LLM would be dirty as hell for other stuff, but only the copyright holder can sue. Apple couldn't sue the output on behalf of F/LOSS projects or copyright infringement in general, only with respect to its own.

In principle, I suppose this could also be extended with enough compute power to any arbitrary software. Build an LLM perfectly from scratch with a signed verified training set (which you save) that includes nothing of what you want to CRE, and you should be able to prove that it's "virgin" on that front. Of course in practice that'd mean jumping back in time and be super expensive right now, no current tainted LLM could be a base point. But maybe eventually compute will drop back down in price enough again, and/or meta breakthroughs in training, that for valuable enough CRE targets, someone might create a business around creating "CRE LLMs certifiably and legally/technically provably clean of any copyrighted material related to your target" which would still leave most of the corpus of human works. That could really disrupt a lot of the current regime.

F/LOSS I assume would actually be the most resistant to this in principle due to being insanely hard to not have be part of a core training set in any way.