> The code is open source but the model is proprietary.

Are you saying that because the Open RAIL M license doesn't conform to the OSI definition of "open source"?

I agree that it's not "open source", but I don't think that means it's "proprietary".

Yes, the OpenRAIL licences are deliberately not open source. Being open source or not is binary, and proprietary is the antithesis of open source.

There are of course additional ways to describe specific proprietary things, like distributable in this case. Distributable but non-free, because the software license is morally judgemental, and limits 'fields of endeavour'.

I'm with you on "not open source", but I don't think "proprietary" works as the opposite of "open source" here.

The dictionary definition of "proprietary" tends towards "one that possesses, owns, or holds exclusive right to something" - that doesn't quite fit here, because models licensed under OpenRAIL aren't exclusively restricted to their owners.

They have terms of how they can used that are more restrictive than open source licenses, but they don't enforce "exclusive" usage by their owners.

Source available is the fitting term here.

Model weights can't really be described as source code though. The equivalence isn't exact, but I'd describe the weights more as the compiled binary, with the training data & schedule being the source (which is sort of under an open source license, with the complication of LAION's "it's just links"). The fact it costs $1 million to "compile" isn't relevant.

This isn't to defend Stability particularly though - they've been getting increasing slow and restrained in their model releases. Charitably because they're attracting a lot of heat from political and anti-AI aligned groups. Uncharitably because they've taken a lot of funding now.

(Edit: typo)

> Model weights can't really be described as source code though. The equivalence isn't exact, but I'd describe the weights more as the compiled binary, with the training data & schedule being the source

I think this is a really interesting discussion! I see where you're coming from, but I'm minded to disagree in part.

For one, I think it's possible to release model weights under a liberal licence, yet train on proprietary data. (ChatGPT is trained on oodles of proprietary data, but that doesn't limit what OpenAI do with the model). Normally, obviously, the binary is a derivative work of the source.

Also, the GPL defines source code as 'the preferred form for modification'. I don't disagree that model weights are a black box. But we've seen loads of fine tuning of LLaMA, so we don't always need to train models from scratch.

Ideally, of course, having both unencumbered training data and model weights would be perfect. But in the interim, given I don't have that million dollars, I'll settle for the latter.

Yeah, neither view is a perfect fit. Another example is vision transformer backbones, where a common generic base weight is used to fine-tune all sorts of different processes (segmentation, image to text, etc). The terminology (and licenses) haven't really kept up.

A properly unencumbered model would be my preference too. The community generally seems a bit laissez-faire with license compliance though, so the restrictions currently don't generate much push back. (Plus it's not totally clear that you can copyright model weights at all, given they're the output of an automatic process).