Karpathy says AI should explain complex topics with custom videos

Why it matters

Karpathy is describing a shift from AI-generated text toward custom-built explanations, while the standard he cites cautions that readable output still needs expert review.

Andrej Karpathy says people will spend more time understanding language-model output, and he expects models to help by turning explanations into diagrams, interactive webpages and custom videos. In an October 2nd post on X, he described a progression from clearer writing toward richer, purpose-built formats, calling bespoke explainer videos the output format he is most bullish on.

Karpathy's first suggestion is to prompt a model to write in ASD-STE100 Simplified Technical English, a controlled language originally developed for aerospace maintenance documentation. He says he has found the approach more readable and sometimes asks for output "80% of the way" to the standard because its rules can be stringent. The standard's maintainers describe it as a set of writing rules and a controlled dictionary intended to make technical documentation easier to understand.

The recommendation comes from a researcher whose work has repeatedly involved making technical ideas teachable. Karpathy's personal site says he designed and taught Stanford's CS231n deep-learning course, was a founding member and research scientist at OpenAI, and later led Tesla's computer-vision team for Autopilot. He also lists educational videos on AI among his current work. In the post, he moves from readable prose to diagrams, then interactive HTML, and finally custom videos that could pair generated narration with visuals.

That sequence is a practical argument about what cheap code generation can make possible. A custom webpage or video may once have required enough design and engineering time to make it impractical for a single explanation. Karpathy's claim is that models can now do enough of that work autonomously to make such artifacts disposable: build one for a specific question, use it to understand the answer, then move on. He expects human work to shift toward oversight and understanding as models handle more of the execution.

The post is advice, not a product announcement or a test of whether these formats improve comprehension. Karpathy offers examples of prompts, including asking for an explainer in the style of 3Blue1Brown and using ElevenLabs for narration, but reports no comparative results or measured improvement. The recommendation to write in a formal controlled language also should not be mistaken for proof that a model follows the standard. ASD's current Issue 9 guidance warns that AI-generated text can appear consistent with Simplified Technical English while failing to apply its rules or vocabulary correctly; the group says expert review remains necessary.

That caution sets a useful boundary around the thesis. A clear-looking diagram, webpage or video can make an explanation easier to inspect, but presentation alone does not establish that the underlying answer is correct. Karpathy's post puts the human task in judging and understanding what models produce, while the standard's maintainers emphasize that technical accuracy still depends on informed review.

Karpathy's proposed formats also broaden the idea of an AI answer. A response need not end as a block of text: it could be an interactive page or a short, tailored lesson assembled for one user. The bet is that abundant code makes that customization cheap enough to use routinely. The post points to a workflow possibility, not evidence that custom videos or generated interfaces have become dependable substitutes for written explanations.