> Right, I'm not talking about the commit history, but rather that anyone (with means) should be able to produce the final artifact themselves, if they want. For weights like this, that requires at least the training script + the training data.
You cannot produce the final artifact with the training script + data. Meta also cannot reproduce the current weights with the training script + data. You could produce some other set of weights that are just about as good, but it's not a deterministic process like compiling source code.
> That's like saying that Slack is Open Source because if I want to, I could patch the binary with a hex editor and add/remove things as I see fit? No one believes Slack should be called Open Source for that.
This analogy doesn't work because it's not like Meta can "patch" Llama any more than you can. They can only finetune it like everyone else, or produce an entirely different LLM by training from scratch like everyone else.
The right to release your changes is another difference; if you patch Slack with a hex editor to do some useful thing, you're not allowed to release that changed Slack to others.
If Slack lost their source code, went out of business, and released a decompiled version of the built product into the public domain, that would in some sense be "open source," even if not as good as something like Linux. LLMs though do not have a source code-like representation that is easily and deterministically modifiable like that, no matter who the owner is or what the license is.
You cannot produce the final artifact with the training script + data. Meta also cannot reproduce the current weights with the training script + data. You could produce some other set of weights that are just about as good, but it's not a deterministic process like compiling source code.
> That's like saying that Slack is Open Source because if I want to, I could patch the binary with a hex editor and add/remove things as I see fit? No one believes Slack should be called Open Source for that.
This analogy doesn't work because it's not like Meta can "patch" Llama any more than you can. They can only finetune it like everyone else, or produce an entirely different LLM by training from scratch like everyone else.
The right to release your changes is another difference; if you patch Slack with a hex editor to do some useful thing, you're not allowed to release that changed Slack to others.
If Slack lost their source code, went out of business, and released a decompiled version of the built product into the public domain, that would in some sense be "open source," even if not as good as something like Linux. LLMs though do not have a source code-like representation that is easily and deterministically modifiable like that, no matter who the owner is or what the license is.