In the quest to make artificial intelligence more reliable and easier to understand, one of the industry’s leading minds is looking to an unexpected source of inspiration: the highly regulated world of aviation maintenance.
Andrej Karpathy, a founding member of OpenAI who joined Anthropic’s pretraining team in May, recently shared a novel approach to prompting large language models (LLMs). He suggests asking these systems to explain complex topics using ASD-STE100—a controlled form of English originally developed to make aircraft maintenance manuals foolproof.
The Power of Controlled English
Maintained by the AeroSpace and Defence Industries Association of Europe (ASD), ASD-STE100—or Simplified Technical English (STE)—is an international writing standard designed to eliminate ambiguity. It features a strict set of grammatical rules and a core dictionary of approximately 900 approved words, most of which are restricted to a single, precise meaning.
According to Karpathy, modern LLMs possess a deep understanding of this standard. He noted that the “heavy constraints on clean writing style” enforced by ASD-STE100 result in AI-generated text that is “a lot more readable” than standard prose. Because the official standard is exceptionally stringent, Karpathy occasionally soft-pedals the instruction, prompting the model to write “80% of the way to ASD-STE100.”
The technique is already gaining traction in the developer community. An open-source tool on GitHub now applies STE rules to text read by AI agents, though its creators note the system is explicitly not intended for marketing copy.
Climbing the Abstraction Ladder
For Karpathy, prompting in controlled English is just the first step in a broader shift in how humans interact with generative AI. He outlines a four-tier progression of model outputs, with each step offering a superior way to digest information:
- Controlled Writing: Using standards like ASD-STE100 for ultra-clear, unambiguous text.
- Diagrams: Visual representations that are “a lot easier to process, parse, and understand” than dense paragraphs.
- Interactive HTML Pages: Generating complete web pages rather than static Markdown. This aligns with recent practices from Anthropic’s Claude Code team, who favor HTML for richer, interactive interfaces.
- Custom Explainer Videos: Generating tailored video content on demand. Karpathy expressed immense enthusiasm for this format, suggesting prompts that ask models to generate “3b1b style” explainer videos (referencing the popular 3Blue1Brown math animations) complete with narration via voice APIs like ElevenLabs or free, locally run alternatives.
A Shift Toward Oversight
This evolution highlights a fundamental change in the nature of knowledge work. Karpathy argues that as AI models take over the foundational legwork of coding, writing, and animating, human professionals will shift their focus. “A lot more of our work will rise up the abstractions into oversight and understanding,” he wrote.
Instead of manually crafting individual documents, users are encouraged to treat LLMs as engines for creating “large, custom, discardable software artifacts”—highly complex, single-use outputs that would have been entirely impractical to produce before the rise of generative AI.