Karpathy shares ways to understand AI answers: ASD-STE100, diagrams, HTML and video
Karpathy shares ways to understand AI answers: ASD-STE100, diagrams, HTML and video
Andrej Karpathy described prompting techniques that help him better understand AI responses, including the simplified aviation standard ASD-STE100, diagrams, interactive HTML pages and explanatory videos.
Andrej Karpathy shared several techniques that help him make sense of AI answers. The first is to ask the AI to write according to ASD-STE100, a simplified English standard used in aircraft maintenance manuals.
ASD-STE100 was created for aviation mechanics for whom English is not their native language. Its vocabulary contains about 900 words; each instruction is at most 20 words long, and each sentence contains one instruction. Language models already know this standard well, so explanations become short, with one idea per sentence, and are easy to read. If the constraints seem too strict, you can ask the model to get “80% of the way to ASD-STE100.” Someone turned this into an open-source skill on GitHub.
If the text still does not help, try another format: first a diagram, then an interactive HTML page with animations and visualizations. The most advanced option, which Karpathy considers promising, is an explanatory video. For example, you can ask a model to create a video in the style of 3Blue1Brown with narration generated via ElevenLabs.
Videos that explain exactly what you want to understand could change education. As an example, Karpathy mentions an explanation of how Shazam works; the original post did not include the link.
Why it matters
AnalysisThis approach shows that the quality of AI answers can be improved through formats and language constraints, not only through more complex prompts. ASD-STE100, visualizations and explanatory videos make difficult topics easier to understand.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Loop Engineering: Architecture of Autonomous Agentic Loops and the Post-Prompting Era
A deep dive into Loop Engineering: self-sustaining execution loops for Claude Code and autonomous agents, the DOER/CHECKER topology, deterministic verification, the /goal primitive, and reward hacking mitigation.
ChatGPT: How to Choose a Model, Plan, and Write Effective Prompts
1. The Operating System in One Page...
Plan-and-Solve Prompting
A two-stage agent architecture that separates the strategic decomposition of a task into a global plan from its sequential tactical execution with dynamic replanning.
ReAct Pattern (Reasoning + Acting)
A fundamental algorithmic pattern for autonomous agents that alternates between internal reasoning steps (Thought), executing external tools (Action), and analyzing the resulting output (Observation).