I expect that if you call a release "open source", it's, you know, source. That their engineers used to build the release.
> The "recipe" is missing, yes, but this is very different from what you're saying.
The "recipe" is what we generally call source.
> So if there's miscommunication then let's communicate better instead of getting upset at one another.
Who's getting upset here? I'm simply calling for not diluting a term. A free, permissive, binary release is great. It's just not open source.
> Because 14G of a bunch of algorithmically constructed numbers and a few text tiles is definitely all you need to use, edit, and/or modify the work.
Just like my Windows install ISO from when they were giving windows licenses away from free.
Not really. At least not in normal software. The recipe is honestly only really interesting to researchers (like me). But for building and production stuff, you have everything you need.
> Just like my Windows install ISO from when they were giving windows licenses away from free.
I repeat, a free windows ISO doesn't have an Apache license attached. This is an inane comparison.
Yes, in normal software, the 'recipe' used to create the build artifacts is the source.
> The recipe is honestly only really interesting to researchers (like me). But for building and production stuff, you have everything you need.
A lot of excuses about why it's not useful to a lot of people and that you don't actually need it to use it in production is exactly the argument made for why not open sourcing is ok. That doesn't mean this is a source release, open or otherwise.
> I repeat, a free windows ISO doesn't have an Apache license attached. This is an inane comparison.
Even if the Windows binary iso was released with an Apache license, it wouldn't be open source since no actual source was released. That's the point of that line of argument.
> Source code for a work means the preferred form of the work for making modifications to it.
I'm sorry, but what do you expect? Literally all code is "a bunch of numbers" when you get down to it. Realistically we're just talking about if the code/data is 1) able to be read through common tools and common formats and 2) can we edit, explore, and investigate it. The answer to both these questions is yes. Any parametric mathematical model is defined by its weights as well as its computational graph. They certainly provide both of these.
What are we missing? The only thing that is missing here is the training data. That means of course that you could not reproduce the results were you to also have tens of thousands to millions of dollars to do so. Which if you're complaining about that then I agree, but this is very different from what you've said above. They shouldn't be providing the dataset, but they should be at least telling us what they used and how they used it. I would agree that it's not full "open source" when the the datasets are unknown and/or unavailable (for all intents and purposes, identical). The "recipe" is missing, yes, but this is very different from what you're saying. So if there's miscommunication then let's communicate better instead of getting upset at one another. Because 14G of a bunch of algorithmically constructed numbers and a few text tiles is definitely all you need to use, edit, and/or modify the work.
Edit: I should also add that they don't provide any training details. This model is __difficult__ to reproduce. Not impossible, but definitely would be difficult. (within some epsilon, because models are not trained in deterministic manners, so training something in identical ways twice usually ends up with different results)