Someone should have an LLM start generating random valid CUDA programs. Compile each one to get a binary. Train a language model with the source and output bi

Someone should have an LLM start generating random valid CUDA programs. Compile each one to get a binary. Train a language model with the source and output binary. Hey presto, clean room compiler. Edit: Oh wait.. duh.. just train it on the equivalent target source. Presumably you can do this for other targets as well.