Building Pi With Pi
12 hours ago
- Pi, part of Earendil, remains Mario's project, and the author reflects on using Pi to build Pi.
- LLM-generated issue descriptions, often inaccurate but confident, become inputs for Pi sessions, altering the issue tracker's role.
- Issues lacking the human's original voice, filled with plausible but wrong diagnoses, create extra maintenance work.
- The author prefers issues condensed to observed facts (e.g., a stack trace) rather than guesses or LLM-generated slop.
- LLMs over-engineer fixes, adding complexity by locally handling edge cases instead of maintaining global invariants.
- The issue tracker sees high volume: ~3,145 external issues/PRs in 90 days, 26% of issues and 8% of PRs are ultimately addressed.
- Open source faces pressure from increased code/projects, but without more maintainers, leading to fragmented effort and need for stronger collaboration.