90
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 08 Oct 2026
90 points (83.1% liked)
Open Source
49520 readers
156 users here now
All about open source! Feel free to ask questions, and share news, and interesting stuff!
Useful Links
- Open Source Initiative
- Free Software Foundation
- Electronic Frontier Foundation
- Software Freedom Conservancy
- It's FOSS
- Android FOSS Apps Megathread
Rules
- Posts must be relevant to the open source ideology
- No NSFW content
- No hate speech, bigotry, etc
Related Communities
- !libre_culture@lemmy.ml
- !libre_software@lemmy.ml
- !libre_hardware@lemmy.ml
- !linux@lemmy.ml
- !technology@lemmy.ml
Community icon from opensource.org, but we are not affiliated with them.
founded 7 years ago
MODERATORS
The difference is the original source code, and similar source code that is incompatibly licensed, was in the LLM training data, making it not a clean room.
Thats not a difference. Other clean room projects, for example ReactOS, have source projects that are incompatibly licenced. Its the entire point of clean room reversing, to bypass copyright and licencing.
The training data is on the dirty room side, it doesnt get passed to the clean side.
The clean room concept is violated if raw source code is in the model generated, but to my knowledge its not. There is no way to pull out a chunk of the model, and say "This is a direct copy of main.c from xyz project". At best, you can ask the LLM to guess what it should be, and it'll regurgitate something statistically close. I just tried asking my local qwen3.8, and did produce a close summary of the nginx main file, but it wasn't even close to a 1 to 1 copy.
The clean room would also be violated if you consider the training, model and inference as a singular entity, but that would be tough to argue when the inference can happen on one server, the resulting model isn't functional in any way, its a big box of data, and the inference can happen on any GPU anywhere else in the world.
A human doing a clean room reimplementation is not supposed to have seen the original source code for something they are reimplementing. If that source code is in the training data for the LLM, that seems like a similar situation.
Except that they are separated. The clean side (model + implementors) doesnt have the original source code. The dirty side is the training room, which does have original source.
The model that you download from huggingface is basically the notebook passed between the two rooms, and as far as im aware, the model doesnt have the actual training data in it, just enough statistics to "hallucinate" something close.