本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/karpathy/status/2053872850101285137
Summary
Karpathy argues that requesting HTML output from LLMs is underutilized — the browser renders rich, interactive, visually structured information that raw text or markdown can’t match. He sketches a progression from raw text to markdown to HTML toward eventually interactive neural video simulations. His asymmetry claim: audio is the preferred human input to AI, but vision (images, animations) is the preferred output format from AI.
Karpathy 提出人機互動的演進序列：原始文字 → Markdown → HTML → 互動式神經影片。核心觀點是大腦有三分之一的容量專門處理視覺資訊，AI 輸出應利用這條「10線高速公路」。在輸入端，他認為音訊是人類對 AI 最自然的輸入方式，但還需要能夠指向和手勢操作。
Key Points
- Requesting “structure your response as HTML” is a useful prompt trick — browser renders it richly
- Output progression: raw text → markdown → HTML → interactive neural videos
- ~1/3 of brain is visual processing — vision is the high-bandwidth output channel
- Input gap: need pointing/gesture alongside audio for natural human-AI interaction
- Endpoint: neural video/simulations generated directly by diffusion nets
Insights
The asymmetry observation (audio in, vision out) is a useful mental model for designing AI interfaces. Most current tools default to text/markdown output and ignore that the rendering surface could be doing much more work. The practical tip (ask for HTML, view in browser) requires no infrastructure change but meaningfully improves information density.
Connections
Raw Excerpt
“Audio is the human-preferred input to AIs but vision (images/animations/video) is the preferred output from them.”