The "vibe coding" proponent and OpenAI co-founder Andrej Karpathy has publicly backed Anthropic Claude Code engineer Thariq Shihipar's assertion that HTML should replace Markdown as the default AI output format. Shihipar's post on X last week, complete with 20 real-world examples, surpassed 4.4 million views. Karpathy not only agreed but expanded the discussion into what the final form of AI output should be.
From Text to HTML, Then to Interactive Neural Video
Karpathy outlined a clear evolution of AI output formats. The first generation is raw plain text, cognitively taxing to read. The second is the current standard Markdown, which reduces load through bold, italics, and tables. The third generation is HTML, offering far greater freedom in layout, graphics, and interactive elements — and it's quickly becoming the new default. He noted there will be several intermediate generations (4th, 5th, 6th) but refrained from defining them prematurely.
The endgame (nth generation) is interactive neural videos or simulations directly generated by diffusion models. AI would no longer produce text for humans to read; instead, it would output dynamic visual content directly into human perception. Karpathy explained that roughly one-third of the human brain is a visual parallel processor, and the visual channel is a "ten-lane highway" for information, with vastly higher bandwidth than reading text. Therefore, high-bandwidth visual output — images, animations, and videos — is the optimal interface.
Flipbook Prototype: Real-Time Pixel-Level Generation Without Any Code
Karpathy specifically cited the Flipbook prototype by former OpenAI researcher Zain Shah's team as a glimpse of the end state. Flipbook uses no HTML or CSS at all. Instead, it's built on the open-source DiT (Diffusion Transformer) video model LTX Studio, streaming 1080p/24fps pixel-level frames via WebSocket in real time. Every "page" a user sees is an AI-generated image; clicking any area triggers the next frame. The entire interface contains zero traditional code.
While Karpathy framed this as a vision for a perceptual interface, he offered a pragmatic immediate tip for LLM users: append "structure your response as HTML" to prompts and open the output in a browser.