I have mixed feelings about this definition. It’s at least better than the status quo, where you see companies both big and small haphazardly throwing model weights out in the world and calling them “Open Source”, despite them not being… source code. I’ve also seen the term “open weights”, but as far as I can tell, this isn’t rigorously defined anywhere, so usually it just means “you can download the weights” but anything beyond that is a coin toss

Then, there’s the fact that it doesn’t require the data to be provided, just described in detail such that someone could reproduce the results. This is practically sensible, since as best as I can tell, no one is releasing the full data sets for their state-of-the-art models. But it still feels like a very weak stance (personally, I feel like training data sources is one of the biggest ethical problems for LLMs / AI models right now, and my impressions from the sidelines is that it’s an open secret that basically all the competitive players in the space scrape data that they don’t really have the right to train on). The FAQ provides a justification for this, but I’m still not swayed

score 4 · canonical

Kudos to them for starting from a consistent outcome goal (“access to the preferred form to make modifications”) and making an honest effort to apply it to AI models, as opposed to just coming up with some rationalization for using the words “open source”, as so many AI companies have been doing.

Nobody seems to be willing to explore the idea that there is actually no practical way large AI models can be open source, and we just need to come up with a different, accurate term. I would rather have that than confuse the meaning of the existing term by stretching it to fit.

score 3 · canonical