In the 1980’s, there were no freely available C compilers (I guess Borland was inexpensive, but not free). The GNU C Compiler was the first such. The project focused on being bootstrappable, reproducibility from source with a fixed dependency chain.
Corporates took advantage of GCC, and continue to, but have also been happy that LLVM came along with its permissive licensing. LLVM was a free lunch for GPU deployment, letting corporates customize it to compile shaders without having to release the sources.
In the LLM era of mostly closed weights, closed and pirated datasets and inaccessible compute requirements, what would, or does, an open alternative look like? Or should something else be done?
A giant, properly ethical sourced dataset of training material, along with an AI architecture that comes without “hallucination” and an affordable GPU which can train such a model in a few weeks?