Entropic Thoughts
16 hours ago
- Top-of-the-line LLMs on easy tasks generate code about 10% more complicated than necessary.
- This complexity is often accepted because it solves the immediate problem and is locally contained.
- An example: a 24-line function to convert a string to a safe HTTP header value had a dead branch due to implicit assumptions, which the author rewrote in 15 lines.
- The author questions the long-term impact of consistently accepting unnecessarily complex code.
- A possible counter-argument is that LLMs will improve, but the author remains unconvinced.