@mildsunrise@tech.lgbt @gabboman@gabboman.xyz some people also talk about disassembled code as clean room, proper clean room practices (i.e. codeweavers, intel, apple, etc) are not frequently publicly disclosed or talked about, by the nature of how the clean room operates so the ai has a higher bias towards very liberal interpretations of clean room.
0 Boosts
0 Likes
@ada @gabboman I'm not quite getting the license laundering point, care to explain a bit more? i thought the point of clean room was to distill a protected work into ideas, which are not protected, and then use those to create another work. if so, I'd expect the providence of these ideas wouldn't matter at all to courts
@mildsunrise@tech.lgbt @gabboman@gabboman.xyz because the ai itself is trained on proprietary code, the code it emits can accidentally be licensed, by emitting it into a secondary language that isn't native to the code it isn't direct plagiarism but laundering the code.
If I describe a proprietary patented algorithm to you in vague terms, it's not unlikely that if you had the same background as the algorithm designer you would end up on the same algorithm even if you implemented it with different terms. You yourself would be clean but me as the one describing it or tasking someone to describe a known thing would be laundering the license away.
AI is that concept on steroids. AI clean room can't compensate for that unless the training corpus is highly curated.
@ada @gabboman ah, from the training data. i agree, and I made very much that point here https://tech.lgbt/@mildsunrise/117426701448506658
@mildsunrise@tech.lgbt @gabboman@gabboman.xyz like right now "AI Clean Room" is co-opting Intel's clean room practices where agent A documents something and agent B (different model) implements it in a bespoke language that is exclusively FLOSS code.
but this has 3 problems