Karpathy Advocates Aircraft Manual Style for LLM Clarity

Brain head collage with question mark and speech bubble in pop art halftone design. — Image credit: Search Engine Journal

Andrej Karpathy, a founding member of OpenAI now with Anthropic, advocated for large language models (LLMs) to generate output using controlled English, specifically referencing the ASD-STE100 standard typically found in aircraft manuals.

Karpathy suggested that adopting formats like ASD-STE100, also known as Simplified Technical English, could significantly improve the readability and clarity of text produced by AI models due to its strict constraints and limited vocabulary.

The AI expert also proposed that LLMs explore alternative output formats beyond plain text, including diagrams, interactive HTML pages, and custom explainer videos, drawing parallels to the educational style of 3blue1brown, potentially utilizing tools like the ElevenLabs API for audio.

ASD-STE100 is a globally recognized standard maintained by ASD, an organization representing European aerospace, security, and defense companies. The standard is freely available for use.

The standard aims to reduce ambiguity and complexity in technical documentation, making it easier for non-native English speakers and technical personnel to understand critical information.

An open-source skill that applies some of the Simplified Technical English rules to AI agent text is available on GitHub, according to developer Thariq Shihipar. However, this open-source tool does not encompass the full dictionary or all the stringent rules of the official ASD-STE100 standard.

Karpathy’s methods highlight a growing industry focus on refining the practical utility and interpretability of AI-generated content.

Source: Search Engine Journal