hmm, I would have thought they could make github 1000x more reliable with llm's