Karpathy: Ask Your AI to Respond in HTML
Original: Karpathy: Ask Your AI to Respond in HTML View original →
A One-Line Addition That Changes Everything
On May 11, 2026, Andrej Karpathy published a widely-shared observation: adding structure your response as HTML to the end of an LLM query and opening the result in a browser produces dramatically richer output than markdown. He also noted success asking the model to present output as slideshows and other visual formats.
A Framework for Human-AI Interface Evolution
Beyond the practical tip, Karpathy outlined a progression of human-AI communication formats:
- Stage 1: Raw text
- Stage 2: Markdown
- Stage 3: HTML with interactive elements
- Stage 4: Interactive neural videos and simulations
Why Visual Output Matters
Karpathy grounded his argument in neuroscience: approximately one-third of human brainpower is dedicated to visual processing, making visual outputs far more efficient than text for information transfer. Audio is humanity preferred input to AI; images, animations, and video are AI preferred output to humans.
Plenty of Runway Before Neural Interfaces
Significant advances in human-AI interaction design remain possible before brain-computer interfaces become necessary. The tweet resonated widely with 16,000+ likes as a rare combination of a concrete, immediately actionable tip and a compelling long-range research direction for the field.
Related Articles
AI researcher Andrej Karpathy argues that LLMs fundamentally change software constraints, excelling at code translation. He predicts large fractions of all software ever written will be rewritten many times over as AI reshapes the programming landscape.
AI researcher Andrej Karpathy argues that programming has fundamentally changed over the last two months, particularly since December when coding agents started actually working. Developers are shifting from writing code to directing and managing AI agents in parallel.
Cloudflare tested Anthropic's security-specialized Mythos Preview model against their own infrastructure under Project Glasswing. Mythos can chain low-severity bugs into working exploits, demonstrating reasoning comparable to senior security researchers — but with inconsistent safeguards and significant triage overhead.