I had the same thought. Couple it with a source fuzzer to generate a large training dataset of valid source, and you could make a very nice disassembler.