Hacker News Reader

Comment by diggan | Hacker News Reader

> Btw Sam has tweeted about an open source model. Stay tuned... https://x.com/sama/status/1932573231199707168

Sneaky wording but seems like no, Sam only talked about "open weights" model so far, so most likely not "open source" by any existing definition of the word, but rather a custom "open-but-legal-dept-makes-us-call-it-proprietary" license. Slightly ironic given the whole "most HN posters are confidently wrong" part right before ;)

Although I do agree with you overall, many stories are sensationalized, parts-of-stories always lack a lot of context and large parts of HN users comments about stuff they maybe don't actually know so much about, but put in a way to make it seem so.

View on HN

There are ten measures by which a model can/should be open:

1. The model code (pytorch, whatever)

2. The pre-training code

3. The fine-tuning code

4. The inference code

5. The raw training data (pre-training + fine-tuning)

6. The processed training data (which might vary across various stages of pre-training and fine-tuning)

7. The resultant weights blob

8. The inference inputs and outputs (which also need a license; see also usage limits like O-RAIL)

9. The research paper(s) (hopefully the model is also described in literature!)

10. The patents (or lack thereof)

A good open model will have nearly all of these made available. A fake "open" model might only give you two of ten.

Open weights is unobjectionable. You do get a lot.

It's nice to also know what the training data is, and it's even nicer to be aware of how it's fine-tuned etc., but at least you get the architecture and are able to run it as you like and fine tune it further as you like.

> Open weights is unobjectionable

Yeah? Try me :)

> but at least you get the architecture and are able to run it as you like and fine tune it further as you like.

Sure, that's cool and all, and I welcome that. But it's getting really tiresome of seeing huge companies who probably depend on actual FOSS to constantly get it wrong, which devalues all the other FOSS work going on, since they wanna ride that wave, instead of just being honest with what they're putting out.

If Facebook et al could release compiled binaries from closed source code but still call those binaries "open source", and call the entire Facebook "open source" because of that, they would. But obviously everyone would push back on that, because that's not what we know open source to be.

Btw, you don't get to "run it as you like", give the license + acceptable use a read through, and then compare to what you're "allowed" to do compared to actual FOSS licenses.

Reply on news.ycombinator.com