Andrej Karpathy, a founding member of OpenAI, has outlined a vision for the evolution of AI interaction interfaces, predicting a shift towards 'interactive neural video' as the ultimate output form of large models. In a recent post, Karpathy endorsed the Claude Code team's proposal to replace Markdown with HTML, citing HTML's superior typographic flexibility. He anticipates that AI output formats will evolve through several stages, eventually leading to video content generated by diffusion models. Karpathy argues that the human brain's capacity for parallel visual processing makes high-bandwidth visual content the most effective way for AI to communicate with humans. He suggests that while current input methods like voice and text are limited, incorporating spatial referencing capabilities could enhance interaction. As an immediate improvement, he recommends users prompt AI to structure responses in HTML.