DeVault: Reforming the free software message

My intention isn't to 'open-wash' the distribution of LLMs. I agree that we must be cautious about using the term 'open source' in a way that could be misleading. My argument, however, is not about redefining open source for marketing purposes, but about expanding our understanding of what source means in the context of LLMs in a way that makes the usage align not only with the letter of the definition of open source, but also the spirit.

As we've discussed, the creation of LLMs involves unique costs and processes that don't have direct analogies in traditional software development. The expense and complexity of the training process mean that even if someone had access to the original training data and code, they might not be able to reproduce the model. Instead, the pre-trained weights of the model often serve as a more practical 'source' for downstream users who want to fine-tune the model for their specific needs.

In this sense, I believe that making the weights of an LLM freely available for use, distribution, and modification could be seen as a form of 'openness'—and, more importantly, only making the training data and code available in the same way is not open enough to be called open source.