alt Hacker News> this is the point of the "clean room" dance
which is the actual relevant part: they didn't do that dance AFIK
AI is a tool, they set it up to make a non-verbatim copy of a program.
Then they feed it the original software (AFIK).
Which makes it a side by side copy, as in the original source was used as reference to create the new program. Which tend to be seen as derived work even if very different.
IMHO They would have to:
1. create a specification of the software _without looking at the source code_, i.e. by behavior observation (and an interface description). I.e. you give the AI access to running the program, but not to looking into the insides of it. I really don't think they did it as even with AI it's a huge pain as you normally can't just brute force all combinations of inputs and instead need to have a scientific model=>test=>refine loop (which AI can do, but can take long and get stuck, so you want it human assisted, and the human can't have inside knowledge about the program).
2. then generate a new program from specification, And only from it. No git history, no original source code access, no program access, no shared AI state or anything like that.
Also for the extra mile of legal risk avoidance do both human assisted and use unrelated 3rd parties without inside knowledge for both steps.
While this does majorly cut cost of a clean room approach, it still isn't cost free. And still is a legal mine field if done by a single person, especially if they have enough familiarity to potentially remember specific peaces of code verbatim.
Replies
> Then they feed it the original software (AFIK).
My understanding is they did do the dance. From the article: "He fed only the API and the test suite to Claude and asked it to reimplement the library from scratch."
One could still make the argument that using the test suite was a critical contributing factor, but it is not a part of the resulting library. So in my uninformed opinion, it seems to me like the clean room argument does apply.
Well sure they didn't do the dance, but you don't have to do the dance. The reason to do it is that it's a good defense in a lawsuit. Like you say, all of this is a legal minefield.
So my understanding was that the original code was specifically not fed into Claude. But was almost certainly part of its training data, which complicates things, but if that's fair use then it's not relevant? If training's not fair use and taints the output, then new-chardet is a derivative of a lot of things, not just old-chardet...
This is all new legal ground. I'm not sure if anyone will go to court over chardet, though, but something that's an actual money-maker or an FSF flagship project like readline, on the other hand, well that's a lot more likely.