本文由 AI 分析生成
建立時間： 2026-05-23 來源： https://x.com/karpathy/status/2053872850101285137
Summary
Karpathy proposes that asking LLMs to output HTML (rather than markdown) is a practical near-term improvement for consuming AI-generated content, leveraging the brain’s visual processing bandwidth. He sketches a progression from raw text through markdown and HTML toward eventual interactive neural video/simulations, while noting that input modalities also need improvement (pointing, gesture, spatial reference).
Karpathy 建議請 LLM 以 HTML 格式輸出，利用大腦視覺處理的高頻寬通道。他描繪從純文字、Markdown、HTML 到互動式神經影片的輸出演進路徑，並指出輸入端（指向、手勢）也需要改進才能真正實現人機心智融合。
Key Points
- Asking LLMs to “structure your response as HTML” and viewing in browser is practical today
- ~1/3 of human brain is visual cortex — vision is a “10-lane superhighway” into the brain
- Progression: raw text → markdown → HTML → … → interactive neural videos/simulations
- Audio is the preferred human input to AI; visual output is the preferred AI-to-human channel
- Input still needs gesture/pointing capabilities for full human-AI collaboration
Insights
The asymmetry (audio input, visual output) reflects genuine human cognitive architecture. The trajectory toward “diffusion-generated interactive simulations” as the end state of AI output is a bold claim but consistent with the trend of neural rendering. The practical tip (ask for HTML today) is immediately actionable and underutilized. This framing also implies that markdown-centric interfaces like most chat UIs are a temporary compromise, not the natural end state.
Connections
Raw Excerpt
audio is the human-preferred input to AIs but vision (images/animations/video) is the preferred output from them. Around a ~third of our brains are a massively parallel processor dedicated to vision, it is the 10-lane superhighway of information into brain.