Karpathy backs HTML over Markdown for AI outputs, hints at neural video future

Andrej Karpathy, former OpenAI co-founder and AI researcher, has voiced strong support for using HTML over Markdown as the output format for large language models. He explicitly aligned with Thariq Shihipar, an engineer on Anthropic's Claude Code team, who previously advocated for structured markup formats. Karpathy's argument centers on HTML's far greater flexibility in layout, rich graphics, and embedded interactive components compared to Markdown's simplicity. He notes that as models become more capable, restricting them to Markdown limits the expressive potential of generated content, especially for UI-heavy applications.

Karpathy further outlined an evolution roadmap for AI-generated outputs. The current practice involves models producing text or code that humans then render. The next step is structured output formats like HTML. But the 'final state,' according to Karpathy, is when diffusion models—capable of generating images and videos—directly produce interactive 'neural videos' or real-time simulations tailored to human perception. This would bypass traditional display methods entirely, allowing AI to create bespoke visual and interactive experiences on the fly. The vision suggests a future where AI doesn't just write content but directly generates immersive, interactive environments.