Don't Force Your LLM to Write Terse [Q/Kdb] Code: An Information Theory Argument
7 days ago
- #LLM
- #Information Theory
- #q/kdb+
- The article argues against forcing LLMs to write terse q/kdb+ code, advocating for verbosity to improve LLM accuracy.
- Terseness in q/kdb+ code increases perplexity, making it harder for LLMs to generate reliable outputs.
- Information theory suggests that shorter, more complex code has higher surprisal per token, reducing LLM performance.
- Empirical evidence shows verbose code (e.g., 'i = i + 1') has lower perplexity than terse alternatives (e.g., 'i += 1').
- The article concludes that in the LLM era, developers should prioritize performance over aesthetic preferences for terseness.