Can you explain how I'd train my own version of this, reproducing the final deliverable that runs on the gpu?
I don't understand what gripe people have with open weight models and wanting it to be purely open source. the training dataset is only going to be a copyrighted set of contents you don't want to touch with a 10 feet pole. let alone have a publicly traded company host it for you, even if they internally are training on it.
Sure, there might be currently contraints, but I think it is quite possible training will get optimized over time or crowsourced training can be organized.
Without fully end-to-end open source models you are still at the mercy of the model provider to keep providing updates, you have no idea what garbage they trained the model on & can't fix that, not to mention might end up getting sued for using the open weigth model once all those "AI stole my data" lawsuits are finally decided.
I love that Apple made OSX open source.
The more open source software out there the better. And the more open weights or even over source AI stuff the better too right?open weight models are better than open-assembly binaries (as you put it) because models are grown like plants, you can shape the open-weight model in a direction you want by feeding it more data and compute (aka finetuning).
which is something that is impossible in a binary.
embrace the new paradigm and it's tradeoffs. without scoffing at semantics and criticising from an armchair.
In comparison to cloud services, having access to the assembly code can still be quite useful.
This is especially true in the modern era, where you can use AI to much more easily decompile the assemble and reconstruct the source code even.
"Open-assembly" code may be more useful than you might previously have thought, given the ease of recreating the source code or making changes with AI these days.