Andrej Karpathy thinks that reading and understanding what an LLM produces may start to take more time. So instead of reading an answer's content at length as it is, he proposed ways to turn it into an output format that is easier to understand. The key is to start with short sentences and expand to visuals and interactive explanations.
Concise sentences in the style of ASD-STE100
The first method is to ask the LLM to explain in simple technical English close to ASD-STE100. ASD-STE100 is the Simplified Technical English specification created to write aerospace maintenance documents clearly. It uses words with consistent meanings and focuses on conveying one idea per sentence, briefly and clearly.
Karpathy adds that if the standard is hard to follow strictly, applying about 80% of it is enough. What matters is keeping sentences short and cutting ambiguous phrasing so the reader grasps the point quickly. You can see the standard's principles in the official ASD-STE100 guide and the official FAQ.
On a related note, our blog's custom instructions for AI answers based on ASD-STE100 also covers how to apply this in a prompt.
See relationships with diagrams and images
The second is to ask for a diagram or image. In a complicated explanation, turning the relationships and flow between concepts into a visual can reduce the burden of following long sentences. The idea is to ask for material that shows structure and connections at a glance, rather than a picture that merely decorates what the text says.
Explore with an interactive HTML page
The third format is an HTML web page. If you ask the model to build a page that explains a topic, you can go a step beyond a static answer and add interaction or animation. The reader can click the parts they need or follow the motion to understand the concept.
Explainer videos tailored to a topic
The format Karpathy is especially excited about is a custom explainer video for a topic you want. His example prompt is a request to "make a 3Blue1Brown-style explainer video about X and use an ElevenLabs API key for the voice." It bundles visualizations and narration suited to the concept into a single video.
Karpathy says this way of making things is now actually starting to work. As one example of a tool for making custom explainer videos yourself, you can look at how to make a code-based explainer video from a topic with anything2explainer.
Source: X @karpathy
One-off custom software becomes natural
These four formats lead to a broader view. As intelligence and code get more abundant, the value grows of the custom outputs you build to understand a particular question or handle a single job. If you can make software tailored to the purpose, like a web app or an explainer video, at the moment you need it, you can naturally choose outputs you wouldn't have made in the past because of the cost of making them.
Karpathy's suggestion doesn't stop at reading more of an LLM's answer. It points toward choosing the format that suits understanding, whether sentences, images, web pages, or video, and, when needed, even building a one-off tool for that purpose.