Embed Notice

@7666 @sun the appeal of testing is that it’s double-entry bookkeeping for code. You have the desired behavior expressed multiple ways that can be mechanically confirmed to agree with each other. That’s all: these different ways of saying what the program should do, they agree, so the code might just do what it’s supposed to do.

Faking that with LLM-generated code, which does not in any real sense know what the program should do, is just going to result in obvious bugs that pass tests.