On: Open-source project ZLUDA lets CUDA apps run on AMD GPUs

[Eager]

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.