There's literally a list of opcodes to be executed in the model. There's a whole lot of data too, but that's part of the build just as much as anything in a .data section.
Forgive my ignorance, I haven't studied the AI tooling landscape yet. Are you saying these models have a structured binary format and "running" them is just a matter of having a "player" with the right "codec"?
They are basically a series of intermediate bytecodes to be compiled to the hardware they actually run on, in addition to the large tables of weights that bytecode references.
Just a few billion of them that no person created or understands, which are based on the their input data and not design decisions, which include random initialization. I see no reason to treat them any differently than software in a conversation about OSS.
Model inputs are definitely design decisions, what data to use, how to fine tune them, what methods were used in weighting things. This is LITERALLY the source of the model. The model is a binary like an EXE
reply