Karpathy shares ways to make AI outputs easier to understand, from controlled English to explainer videos

OpenAI founding member and former Tesla AI director Andrej Karpathy said on Oct. 2 that people will spend more time trying to understand the outputs of language models, and he outlined several ways to make those outputs easier to consume. In a post on X, he said one useful writing trick is to ask a model to explain something in ASD-STE100, a controlled language standard originally created for aerospace maintenance documents. He said the format’s strict style constraints often make the result much clearer, though he sometimes relaxes the instruction and asks for something that is only "80% like ASD-STE100." Karpathy also pointed to charts, HTML pages, and custom explainer videos as stronger output formats than plain text. He said models are getting better at front-end work and can produce polished interactive web pages and animations. The format he is most optimistic about is tailored explainer video generation, including narration through ElevenLabs or local free alternatives. He added that as language models become more capable, more human work will move toward supervision and interpretation. On the same day, he also shared a separate evaluation in which a model was asked 16,200 times whether given coordinates were on land or in the ocean, producing a world map.