Karpathy Shares Four Techniques for Understanding LLM Output
Karpathy posted a long piece arguing that in the future we will spend far more time trying to understand LLM output, and that understanding deserves its own systematic approach. He offered a layered set of practical techniques whose core idea is to change the presentation medium rather than asking AI to explain itself in words over and over. Multiple accounts (@danielmac8, @APPSO, @机器之心) relayed the content.
Confirmed
- Writing: Have the LLM write explanations in ASD-STE100, a controlled-English standard originally developed for aerospace maintenance documentation. Per @机器之心, its rules include keeping one instruction within 20 words, one action per sentence, favoring active voice, and using one consistent name for a concept throughout. If too strict, ask for "80% compliance" to relax the constraints.
- Presentation escalation: Posts describe four progressive moves — controlled-language writing, diagrams, web pages, and finally custom explainer videos.
- Community practice: User @blixt shared that simply prompting "follow ASD-STE100" makes LLM output noticeably cleaner and more controllable.
Why it matters
- The approach turns "understanding AI output" into an engineering problem, lowering comprehension cost via style constraints and medium changes instead of relying on the model re-explaining itself. Notably, one of the most effective tools is a 40-year-old aviation writing standard.
- User @AryHHAry responded with a framing: AI hasn't eliminated thinking but changed which parts remain human — in the AI era production becomes cheap while verification is the core step left to humans.
2026-10-02 ~ 2026-10-03 · 7 related posts
Primary sources
- [source] Karpathy: From controlled-language docs to bespoke explainer videos, better ways to read LLM output — karpathy · 2026-10-02
- Karpathy shares tips for reading LLM outputs, using the ASD-STE100 controlled language spec — daniel_mac8 · 2026-10-02
- Karpathy: we'll spend far more time understanding LLM outputs — and shares ASD-STE100 writing trick — DavideCrapis · 2026-10-02
- AI isn't killing thinking — production is cheap, verification isn't — AryHHAry · 2026-10-02
- Karpathy shares 4 tricks for understanding AI output, from controlled language to custom explainer videos — APPSO · 2026-10-02
- [source] Prompting LLMs with ASD-STE100 aerospace spec yields cleaner output, devs share tips — blixt · 2026-10-03
- [source] Karpathy's trick: use a 40-year-old aviation writing standard to make LLMs clearer — 机器之心 · 2026-10-03