If I vibe-coded an app prompting, say, Claude, and released it along with the generated code, would you have the same objections to me calling it “open source,” because I haven’t also released the weights (and Anthropic’s training data, of course) that generated it?
Your argument simply takes for granted that the way to think of modern AI models is as compiled binaries, based on some superficial similarities. I consider this specious: it looks different from the programs you’re used to, but the weights + inference code are the program, and training data + RLHF + safety abliteration + jailbreaking safeguards + secret sauce (maybe they run magnets over a hard drive containing the weights) are tools used to create the code they release. This view is supported by the company that releases it having no more preferred “source code” form they interact with it in: I think you’re simply wrong in suggesting they prefer to work with the model by editing the training data and “recompiling” instead of starting with the weights (which they released) and modifying that directly through fine-tuning.
If I vibe-coded an app prompting, say, Claude, and released it along with the generated code, would you have the same objections to me calling it “open source,”
No, because I don’t think this misleads people. Granted, the term “open source” is fuzzy at the boundaries. Should we use the term? I don’t know, but if we do, it only makes sense if it means something different from “closed source”.
wrong in suggesting they prefer to work with the model by editing the training data and “recompiling” instead of starting with the weights
One doesn’t exclude the other. If you’re creating v2 of your model, you’d likely: take the training code and data for v1; make some changes / add new things; run the new training code on the new data. For minor changes you may prefer to do fine-tuning on the weights.
There’s a clear and obvious difference between models like Qwen, DeepSeek, and Llama; and models like ChatGPT, Claude, and Gemini; and the well-established and widely-understood phrase for this difference is “open source”, contrasted with “proprietary” or “closed source,” used in things like hardware, fonts,[1] and military intelligence. If you like, think of it as a kind of fossilization of the phrase, where the “source” part has ceased to be more than an etymological curiosity; you can certainly dislike this phenomenon, but – and I say this with regret since I’m far more prescriptivist than the next guy – trying to change it is probably pissing upwind.
The restrictions on usage are a better argument: among the models with weights available, some are clearly more “open source”[2] than others, and I’d even agree that Llama’s 700-million-user restriction means that, while for most practical purposes it’s open source, it’s technically only “source-available.”
It’s not obvious to what extent fonts count as programs, and their “source code” is usually nothing more than the glyphs, which can be read out from proprietary fonts trivially. Maybe there’s a bit of obfuscation one could perform on the feature file?
The usage I’m objecting to started, as far as I can tell, about 2 years ago with Llama 2. The term “open weights”, which is often used interchangably, is a much better fit.
If I vibe-coded an app prompting, say, Claude, and released it along with the generated code, would you have the same objections to me calling it “open source,” because I haven’t also released the weights (and Anthropic’s training data, of course) that generated it?
Your argument simply takes for granted that the way to think of modern AI models is as compiled binaries, based on some superficial similarities. I consider this specious: it looks different from the programs you’re used to, but the weights + inference code are the program, and training data + RLHF + safety abliteration + jailbreaking safeguards + secret sauce (maybe they run magnets over a hard drive containing the weights) are tools used to create the code they release. This view is supported by the company that releases it having no more preferred “source code” form they interact with it in: I think you’re simply wrong in suggesting they prefer to work with the model by editing the training data and “recompiling” instead of starting with the weights (which they released) and modifying that directly through fine-tuning.
No, because I don’t think this misleads people. Granted, the term “open source” is fuzzy at the boundaries. Should we use the term? I don’t know, but if we do, it only makes sense if it means something different from “closed source”.
One doesn’t exclude the other. If you’re creating v2 of your model, you’d likely: take the training code and data for v1; make some changes / add new things; run the new training code on the new data. For minor changes you may prefer to do fine-tuning on the weights.
There’s a clear and obvious difference between models like Qwen, DeepSeek, and Llama; and models like ChatGPT, Claude, and Gemini; and the well-established and widely-understood phrase for this difference is “open source”, contrasted with “proprietary” or “closed source,” used in things like hardware, fonts,[1] and military intelligence. If you like, think of it as a kind of fossilization of the phrase, where the “source” part has ceased to be more than an etymological curiosity; you can certainly dislike this phenomenon, but – and I say this with regret since I’m far more prescriptivist than the next guy – trying to change it is probably pissing upwind.
The restrictions on usage are a better argument: among the models with weights available, some are clearly more “open source”[2] than others, and I’d even agree that Llama’s 700-million-user restriction means that, while for most practical purposes it’s open source, it’s technically only “source-available.”
It’s not obvious to what extent fonts count as programs, and their “source code” is usually nothing more than the glyphs, which can be read out from proprietary fonts trivially. Maybe there’s a bit of obfuscation one could perform on the feature file?
I agree it’s useful vocabulary, and reducing it to a binary makes it less so.
The usage I’m objecting to started, as far as I can tell, about 2 years ago with Llama 2. The term “open weights”, which is often used interchangably, is a much better fit.