Hacker News Reader

Comment by AlienRobot | Hacker News Reader

If I remember correctly, Stallman's whole point about FLOSS was that consumers were beholden to developers who monopolized the means to produce binaries.

If I can't reproduce the model, I'm beholden to whoever trained it.

>"If you're explaining, you're losing."

That is an interesting point, but isn't this the same organization that makes "open source" vs. "source available" a topic? e.g. why Winamp wouldn't be open source?

I don't think you can even call a trained AI model "source available." To me the "source" is the training data. The model is as much of a binary as machine code. It doesn't even feel right to have it GPL licensed like code. I think it should get the same license you would give to a fractal art released to the public, e.g. CC.

View on HN

It's not clear that copyright applies to model weights at all, given they are generated by a computer and isn't really a creative work. It is closer to a quantitative description of the underlying data, like a dictionary or word frequency list.

That's interesting. I wonder what will protect these models then, if anything? NDAs? Or maybe the model can be a trade secret or patented?

I think dictionaries are copyrightable, however?

I think this makes the most sense. The only meaningful part of the term is whether or not you can hack on it, without permission from (or even coordination with) owners, founders or creators.

Heck, a regular old binary is much less opaque than “open” weights. You can at least run it through a disassembler and slowly, dreadfully, figure out how it works. Just look at the game emulator community.

For open weight AI models, is there anything close to that?

It's a bit impressive that AI managed to produce something blobbier than a binary blob. AI is the blobbiest blob, so blobby that it's a black box to even its own authors.

I wonder how could anyone be an open source enthusiast, distrusting source code they can't verify, and yet a LLM enthusiast, trusting a huge configuration file that can't be debugged.

Granted I don't have a lot of knowledge about LLMs. From what I know, there are some tools that can tell you the confidence/stickiness of certain parts of the generate output, e.g. "for a prompt like this, this word WILL appear almost every time, while this other word will almost never appear." I think there was something similar for image generation that could tell what areas of an image stemmed from what terms in the prompt. I have no idea how this information is derived, but it doesn't feel like there are many end-user tools for this. Maybe the AI researchers have access to more powerful tooling.

For source code I can just open a file in notepad.exe to inspect it. I think that's the standard.

If, for example, a computer program was programmed using an esoteric language that read used image files instead of text files as source code, I don't think you would be able to consider that program "open source" unless the image format it used was also open source, e.g. PNG. If it was some proprietary format, people can't create tools for it, so they can't actually do anything the image blob, which restricts their freedoms.

Reply on news.ycombinator.com