Replies (1)

• @djsumdog@djsumdog.com 2026-04-22 01:41

  How would one make a truly "open source" LLM anyway? Not the bullshit OSI open-weights definition, but a system where you could possess all the source material. You'd have to be limited to only public domain or Creative Common works, and only open source code. You could include GPL, but the documentation would have to warn anything generated could be GPL or CC-NC/SA tainted. You'd need a few terabytes of the published "source," the exact setup to train the model and the final safesenors fine. And if regular people could even afford the thousands (sometimes millions) of dollars worth of compute to run the training, they'd end up with different models depending on how they did the reinforced feedback learning.

  Open ##1507719