Karpathy shares tips for reading LLM outputs: ASD-STE100 writing and diagrams
dair_ai · x · 2026-10-05
Karpathy argues we'll spend much more time understanding LLM outputs and shares tricks: ask models to write in ASD-STE100, an aerospace-controlled language spec whose strict constraints yield cleaner, more readable prose (or request '80% of the way to ASD-STE100'); and prefer asking for diagrams over text since they're easier to parse.
dair-ai's Omar Sanseviero notes most of these practices are already common, but says it raises a real question: is there a universal interface/language for seamless human-agent collaboration? He has been experimenting with one for months (LLM wikis, artifacts, AI-generated visual explainers) without finding the perfect solution.
More from coding & agent
- Supermemory launches smfs: filesystem ergonomics for agent memory without the token tax — gaganghotra_ · 2026-10-05
- Claude Code spent 6 hours and 33% of a weekly quota making its own reply video — Vivid-Snow-2089 · 2026-10-05
- AI-Generated Match-3 x Flappy Hybrid Level: 5 Minutes and $0.77 to Build — Flomerboy · 2026-10-05
- Agent stuck in a loop cost $12 overnight and nearly $1,200 — how do you build kill switches? — Full_Collar9026 · 2026-10-05
- Open-source plugin pools multiple API keys with auto-rotation to dodge rate limits — ahriad · 2026-10-05
- Matt Pocock: Give Your Coding Agents Tools That Build Code the Way You Like — _AustinCalvert_ · 2026-10-05