@SuperDicq @kaia

> These local models are still proprietary and you have no control over them.

Simply not true. Many many models out there (and not just academic, irrelevant models, I mean models that people actually use) are licensed freely (Apache 2.0 or MIT are popular licenses) and that includes the parameters, i.e. all those numbers in the gigabytes-large .safetensors file that you download and then stuff into your GPU's VRAM. The software used for inferencing and (additional) training with these models is also freely licensed.

What's true is that, more often (but not always!) the training data used to produce those parameters is proprietary and/or undisclosed. The computing power required to train a model *completely from scratch* is also prohibitively time-consuming and/or expensive for any single user with consumer hardware.

So it's easy to assume that the big blob file used for inferencing is completely opaque and immutable, regardless of licensing, and effectively people are still using a black box, but that's not true - just like an executable binary program isn't really a black box. It's on your computer, you can look at the binary, you can patch it, you can disassemble it, you can run it in a debugger, you can run it in a sandbox, you can inject your own code into it, etc.

You can do a lot of similar things with models on your computer, like training Low-Rank Adaptations for these models using small datasets on consumer hardware to adapt and modify the output of the model, or modify the parameters directly to modify the output of a model. People routinely do both of these things to teach models new tricks (example: Generate images of things that the base model does not know about) or to make the model do things the original creator trained it not to do (example: Do ERP or use politically incorrect language).