Post #2711470

Will it get better as models evolve? I don't think so, because it's a systematic problem. Every little inconsistency or other piece of desinformation in a code base will mislead the llm and generate even more noise. It's like thermodynamics. The only way to reduce the noise is by doing work/spending energy on it (a.k.a. refactoring). And that doesn't seem to fit with the "black box" approach.