Hacker News Reader

Comment by riedel | Hacker News Reader

This a hard stance if you talk just about the code (not the model weights). The llama community licence is a bit weird and probably not an OSI compliant licence, but close. Regarding weights this is different, but to me it is actually difficult to understand still now to aplly copyright law here. Having said that one nicht und erstand why certain stupid looking clauses went into the code licence. If we do not understand copyright of model weights and do not have court rulings in the use oft training data und er different copyright regimes (US and EU), I would not care too much. We are still in the Wild West.

View on HN

Model Weights are not the Source. Why can't that be obviously like a binary isn't source code - a binary is compiled from source. You can open-license the data in a binary so it can be reverse-engineered / modded but that doesn't make it open source.

“Close” is not good enough for using a term with a very specific meaning. OSI = open source, everything else is source-available (which its arguable that either even applies, because the source of the weights, the dataset, is not available).

I agree that for Llama, things are weird and they want to cover their bases, and that its better than nothing, but the specific use of “open source” is a long-running corporate dilution of what open source really means and I am tired of it.

Funny thing is that what they implemented in that licence is kind of a copy left thingy. Because you can use llama to train other models, but you would have to declare them derivative and thus keep the licence terms. (Not saying this is OSI compliant, but at least a IMHO a new kind of beast)

I feel the arbitrary split between code and weights makes little sense when discussing if these models are "open source" in the copyleft meaning of the term. If the average user can't meaningfully use your product without agreeing to non-free terms then it's morally closed source.

Anything else and you're just open-source-washing your proprietary technology.

I tend to see weights as nothing more than data, data which may not even be copyrightable. But Meta keeps calling their data "open source" when they clearly do not release the model under an open source license, and that is terrible, awful and misleading.

Having access to a weight doesn't make it open. Else you can make the argument that Microsoft Word is open source because you have access to the binary.

Word is a good analogy here.

The model is a static data file like a word doc.

Meta open sourced the code to run inference on the model (ie the code for Microsoft word reading a doc file).

They also open sourced the code to train/fine tune the model. (Ie the code for Microsoft word writing a doc file)

Then they released a special doc (the llama 3 model), but didn’t include the script they used to create that doc.

Indeed, weights are literally binary data. Not human-readable!

The access to modify (uptrain/finetune) these weights is the same between Meta & others, unlike with Word (where Microsoft has an advantage because they have code and can recompile it). I think this is the only thing which matters in practical terms.

Lots of binary executables and libraries can be customised too. That doesn't make them open-source.

This neither makes modification easy, or that’s how owner of the code does modification themselves, and that’s where the difference is.

Reply on news.ycombinator.com