> Sharing weights is like sharing a compiled program. Not at all. They're only similar in the sense that both are a build artifact. > Without visibility into

> Sharing weights is like sharing a compiled program. Not at all. They're only similar in the sense that both are a build artifact. > Without visibility into the training data, curation / moderation decision, the training code, etc Llama could be doing anything and we wouldn’t know. "could be doing anything" is quite the tortured phrase, there. For one, model training is not deterministic and having the full training data would not yield a byte-perfect Llama retrain. For two, the released models are not turing-complete or filled with viruses; you can open the weights yourself and confirm they're static and harmless. For three, training code exists for all 3 Llama models and the reason nobody uses them is because it's prohibitively expensive to reproduce and has zero positive potential compared to finetuning what we have already. > Also open source means the license used should be something standard not proprietary, without restrictions on how you can use it. There are very much restrictions on redistribution for nearly every single Open Source license. Permissive licensing may not mean what you think it means.