Anti-LLM Sentiment Considered Harmful

My perspective on the matter is that reviewing LLM-generated PRs is incredibly tiring. A human is able to design something that spans several files over the course of several days, and this is simply not something that LLMs are good at, since they have to get their context re-injected all the time, but also lose some of the context to compaction because the context window is fairly limited. They inflict such a burden on the reviewer (me) that I have no appetite to use those tools on my colleagues.

It’s very difficult (maybe impossible?) for this technology to approach the cognitive capacity of a human brain in a way that produces PRs that are easily readable, as well as PRs that follow the same design or vision. Moreover the tools themselves are optimised for spending tokens so that you can help their owners repay their investors. As such, alongside the lack of reproducible results, I don’t consider them adequate development tools.