After the panel on "when are AI systems free/open" at #FOSDEM I struggle to see why we don't just equate ML/training to compilers. They take data (source code for compiler, corpus for AI) and turn them into a decision making program (machine code for compiler, weights for AI model). Clearly a program doesn't become open source by distributing it in binary form/weights, and a whole AI system can only be open source if all of its components are: compiler (ML/training process) and program (model)?