I think GP's point is we could get much better results if we fed the LLMs with more than the input text. It could still be text, or at least bytestreams: progr

I think GP's point is we could get much better results if we fed the LLMs with more than the input text. It could still be text, or at least bytestreams: program traces, OpenTelemetry logs, objdump of the compiled binaries, LLVM IR dumps, compiler errors, syntax highlighting markers, etc.