Karpathy suggests that large models should be output in HTML format, and that the ultimate form of AI interaction should be "interactive neural video".

According to Beating, Andrej Karpathy, the originator of the "vibe coding" concept and a founding member of OpenAI, published an article today strongly supporting the Claude Code team's advocacy of "replacing Markdown with HTML." He not only strongly agrees with this change but also outlines a roadmap for the evolution of AI interactive interfaces, predicting that after multiple iterations, the ultimate form of output from large models will be "interactive neural video." Karpathy believes that the evolution of AI output formats has progressed from the earliest, extremely difficult-to-read plain text, to today's Markdown, and now to HTML, which is gradually becoming the new standard with extremely high layout flexibility. In the future, it will go through multiple intermediate forms (4, 5, 6, etc.), ultimately reaching the final stage (n): interactive neural video directly generated by a diffusion model. Regarding the specific appearance of this form, he directly named Flipbook, a no-code pixel-level rendering prototype recently released by a former OpenAI researcher. The underlying logic of this evolutionary trend lies in the physical bandwidth of the human brain. Karpathy points out that about one-third of the human brain is dedicated to parallel processors that handle visual signals—a "ten-lane highway" for inputting information. This determines the optimal solution for human-computer interaction: the best way for humans to give instructions (input) to AI is through highly efficient speech, while the best way for AI to provide feedback (output) to humans is through high-bandwidth visual images (images, animations, or videos). Furthermore, he points out that current input methods still have significant limitations; relying solely on voice or text is insufficient, and there is an urgent need to add spatial indication capabilities, such as pointing to a specific area of the screen when two people are looking at a computer side-by-side. As a shortcut to improving the experience at this stage, he strongly suggests that users add "structure your response as HTML" directly to the end of the prompt.