Andrej Karpathy, the OpenAI co-founder who popularized the concept of "vibe coding," is throwing his weight behind the Claude Code team's push to replace Markdown with HTML for AI output. He's not just endorsing the change — he's sketching out a full evolution roadmap for how AI interfaces will look, predicting the ultimate destination will be "interactive neural video."
https://twitter.com/karpathy/status/2053872850101285137
Karpathy sees AI output formats evolving from early, barely readable plain text to Markdown, and now to HTML, which offers far greater layout flexibility. But that's not the end. He envisions several intermediate forms (4, 5, 6, etc.) before reaching the final form (n): interactive neural videos generated directly by diffusion models. As a concrete example, he points to Flipbook, a no-code pixel-level rendering prototype recently released by a former OpenAI researcher.
This progression is driven by the physical bandwidth of the human brain. Karpathy notes that roughly one-third of our brain is a parallel processor dedicated to visual signals — a "ten-lane highway" for input. That logic dictates the optimal human-AI interaction: the best way for humans to give instructions (input) is voice — efficient for communication — while the best way for AI to respond (output) is high-bandwidth visuals: images, animations, or video.
Still, Karpathy points out a remaining pain point on the input side: relying solely on voice or text isn't enough. What's missing is spatial pointing — the ability to gesture at a specific area on a screen, like two people standing side by side looking at a computer. As a quick workaround for now, he strongly recommends adding "Structure the reply as HTML" at the end of your prompts.