OpenAI founding member and former Tesla AI director Andrej Karpathy said in an Oct. 2 post on X that people will spend much more time trying to understand the outputs of language models. He also shared several methods he uses to make those outputs easier to read and work with, covering text, charts, web pages, and tailored explainer videos.
Using controlled English for writing
Karpathy said one technique that has worked well for him is asking an LLM to explain something in ASD-STE100. He described it as a controlled language specification originally developed for aerospace maintenance documentation. According to him, language models are very familiar with it, and its strict writing constraints make the output much clearer to read.
Because the specification is quite strict, he said he sometimes loosens the instruction and asks the model to write in a style that is only "80% like ASD-STE100."
Charts, HTML pages, and custom explainer videos
Karpathy then laid out three more advanced approaches beyond plain text.
- First, ask the model to draw charts instead of returning text. He said charts are often easier to process and understand.
- Second, ask the model to output HTML and directly generate a polished interactive web page. He said language models are getting stronger at front-end development and can now produce refined interactive experiences and animations.
- Third, and the format he is most optimistic about, is a custom explainer video built for any topic.
He suggested prompts such as asking the model to "make a 3Blue1Brown-style explainer video and use my ElevenLabs API key for voice narration." If no key is available, he said users can also ask the model to find free alternatives that run locally. Karpathy said this approach is "actually starting to work."
Human work shifts toward supervision and understanding
Karpathy said that as language models become more capable, they will handle more basic work on their own, while human work will move upward toward supervision and understanding.
He added that models can help there as well. As intelligence and code become more abundant, people can ask models to create large, customized, disposable software artifacts such as web apps or explainer videos, things that previously were not economical to build.
A separate latitude-longitude test shared the same day
On the same day, Karpathy also shared another evaluation. He gave the model sets of latitude and longitude coordinates in text form and asked whether each point was on land or in the ocean. After 16,200 prompts, he plotted the answers and the result was a world map.
He wrote: "The model knows. This comes from compressing the entire internet."