this post was submitted on 11 Oct 2026
148 points (85.2% liked)
Opensource
6897 readers
221 users here now
A community for discussion about open source software! Ask questions, share knowledge, share news, or post interesting stuff related to it!
⠀
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
No such thing as clean room LLM output, unless you wrote everything in its training model.
It's in fact by definition as dirty as possible.
Right, because clean room means you've never seen any code in your life.
People - spicy autocomplete was not going to derive C from first principles. Neural networks are example-based. The model still obviously never saw Photoshop's source code, because that shit's not on Github.
I'm with you on your first paragraph - unless we're doing some pioneering work, we learn it from somewhere.
But the possibility of decompiling means you don't have to have access to the source just so long as you can access the code is some form - including the generated executable.
The models were built upon the data, code, art, and texts stolen from a generation of internet users and artists, and from generations of writers.
Do you seriously believe that these companies which built themselves on theft would stop stealing once their first model was built?
If a single adobe programmer (or even employee, depending on how lax their internal security is) used a cloud-hosted LLM at work, you can be assured that whatever code and data that malware could access from their computer ended up being part of the next generation of that company's model, and eventually of any other's.
Even if they didn't, if any windows 11 computer had access to the code you can be certain it scrapped it and sent it home to be given to openAI and any other “AI” company willing to pay.
Thieves will steal. No LLM model ever started clean, and they only get filthier and filthier with time.
That's still not what clean room means.
Y'all will bend over inside-out to pretend these models need a terabyte of linear algebra just to copy-paste a text file they obviously did not have.