The no-human-in-the-loop Vashtis have Lemoine syndrome. Things Vashti won't admit: LLMs are not good coders. It's mostly over-engineered dogshit. Just like your noob coder LLMs will propose a deep change with hundreds of diffs, all internally consistent and 'knowledgeable', but where real understanding would have been a one liner. Humans rarely get it right first time either, hence code reviews pre-date LLMs. Inconvenient then for Vashti et al since humans are the bottleneck. We all know what /1
If I get the suspicion that someone is slopping a #curl PR, I ask questions.
If they answer as human and explain their thoughts, this is fine. They used the LLM as a tool, but they are in charge of it. We review, they improve until it‘s accepted.
If their answer is likely from an LLM itself, we are dealing with a meatproxy („Apologies…“ etc). The chances of this PR landing become practically zero.
But I look at the idea behind the PR: was there one? Is it maybe worth doing in another way?