PufferLib author rants about LLM-generated contributor code, semi-automates fixes with a Skill file
jsuarez · x · 2026-09-21
- PufferLib author jsuarez published a ranty piece: while porting 2,000 lines of Python to CUDA C, he found the repo had grown 1,500 extra lines of mostly LLM-generated code from well-meaning contributors, forcing him to spend time "fixing their fucking stupid code."
- He semi-automated the cleanup into a "Skill" file that codifies review/refactor rules for agent-assisted PR handling — while mocking "Skill" as the best pure-marketing naming since LLM folks co-opted "Agent" from RL researchers.
- The piece is a vivid look at open-source maintenance in the LLM era: lower contribution barriers, higher reviewer burden.
More from coding & agent
- Google indexes your trycloudflare tunnel URLs — here's a safer setup — iannuttall · 2026-09-21
- "Don't vibe code if you can't code without AI" sparks skill-dependency debate — prasenx · 2026-09-21
- Model grades itself full marks for a chart that never existed — memokris · 2026-09-21
- AI finds key-independent collisions in six major hash functions — thomasahle · 2026-09-21
- 7 document parsing APIs tested on 11 PDFs: no single winner — latentnoise_ · 2026-09-21
- Building a Claude API chatbot with live web search: 5 open engineering questions — angel-queen123 · 2026-09-21