

The thread we’re commenting on is about reimplementations of open source code libraries to skirt their licensing requirements; so I think the fact that most (probably all) of the AIs capable of something like that have all trained heavily on what is likely the very code they’re being asked to recreate “from scratch” is kind of relevant to this.










We weren’t discussing proprietary code in this thread. You can’t use LLMs to make “clean room reimplementations” of OSS libraries to skirt licensing because they trained on those libraries is what I was getting at.
The rest of my original comment was just me trying to establish that the only way to get a clean room style reimplementation from an LLM is to vibe code something it wasn’t already trained on. Which is legal but a lot less useful, given all the issues with vibe coding, and likely impossible for any OSS project.