TopDigg

The "HTML Hacker Era" of AI Output Bandwidth: Karpathy's Tweet Laid Bare the Interface Bottleneck

· TopDigg · AI / Human-AI Interaction / HTML / Karpathy / Multimodal AI

On May 11, 2026, Andrej Karpathy reshared an article by Anthropic engineer Thariq on X, casually dropping a practical tip: "Append 'structure your response as HTML' to your query, then open the generated HTML file in your browser." The numbers are stark: as of this writing, Karpathy's post had 903,949 views, 9,568 likes, and 905 reposts; Thariq's original post (May 8) had 10.84 million views and 14,773 likes.

This isn't just a UI trick. Karpathy reframed the issue at the human-AI bandwidth level: humans prefer audio on the input side, AI is better suited for visuals (images/animations/video) on the output side. He laid out an evolution path:

  1. Plain text (hard to read)
  2. Markdown (current default — bold, lists, tables, slightly better)
  3. HTML (code-as-process, but far more flexible with graphics, layout, and interactivity) ← forming the new default
    ...
    n. Interactive neural video/simulation (diffusion nets generating directly — not quite ready yet)

He also pulled in hard neuroscience data: roughly one-third of the human cerebral cortex is dedicated to visual processing — a "10-lane information highway." This isn't hyperbole: the visual cortex occupies ~30% of the cerebral cortex, compared to 8% for touch and 3% for hearing. Some studies suggest up to two-thirds of brainwave activity during wakefulness is vision-related, with the retina itself being an extension of the brain.

The Real Cost Data of the Current "HTML Hacker" Approach

(Benchmarked against 2026 mainstream LLM APIs and industry reports)

Token Cost Comparison (Same Content)

  • Markdown vs HTML: Markdown typically saves 20-30% tokens; extreme cases (Cloudflare analysis) show up to 80% fewer tokens
  • Developer实测 in threads: HTML output token volume is often ~3x that of Markdown. For a product handling 100 million conversations daily, output format directly determines unit economics

API Cost Impact

(Using typical pricing where output tokens cost 3-10x more than input tokens)

  • Output tokens carry higher per-unit cost (OpenAI and similar models charge 3-10x more for output). HTML bloat can increase total cost per million tokens by 20-80%
  • High-frequency use cases (reports, code review, planning): HTML is generated once, then rendered by the browser with near-zero additional latency; but in bulk chat products, the "token killer" effect is pronounced

Real-World Examples

(Thariq's GitHub + community)

  • Thariq published 20+ HTML examples: PR reviews with SVG charts, draggable triage boards, interactive reports. Developer feedback: "information density went through the roof"
  • Similar tools are already popular in Cursor, Claude Code, and Obsidian plugins: shifting from plain Markdown summaries → HTML dashboards/slides, cutting reading time from 30 seconds to 2 seconds (self-reported by developers)

The Input-Side Asymmetry

Karpathy noted that audio/text/video still aren't enough — we're still missing "gestures pointing at the screen." This isn't sci-fi: by 2026, multimodal models can process video in real time, but input bandwidth is still bottlenecked on "typing" or "speech-to-text." Meanwhile, output is desperately chasing the visual highway.

Multimodal Trend Data

(2026 industry reports)

  • Market: Multimodal AI was $1.2 billion in 2023, with 2024-2032 CAGR exceeding 30%
  • Model progress: Diffusion models (Sora-class) already convert text/images → video; by 2026, "any-to-any" is mainstream, but truly interactive neural simulation remains in the "open questions" stage (Karpathy's exact words)
  • Viral case: The link Karpathy referenced — the recent breakout hit — points toward the direction of diffusion-net-generated interactive simulations

Putting All the Numbers Together

The logic is brutally clear:

  1. Brain's visual cortex at 30% vs. AI output defaulting to Markdown (extremely low information density)
  2. HTML as a transitional bridge, hacker-style bandwidth boost, but token costs genuinely rising 20-80%
  3. Downstream developers/products are already adopting it (Thariq's 10M+ views prove demand), while upstream models are still optimized around "output token billing"
  4. The n-th step is diffusion neural video — meaning output is no longer "text + formatting" but direct neural mesh rendering, with exponentially greater bandwidth potential

Conclusion

In 2026, AI model intelligence is no longer the tightest bottleneck — the human-machine interface's bandwidth asymmetry is. Karpathy and Thariq quantified the problem with a single HTML trick: we're still running 20th-century text pipes into a brain with 30% visual cortex.

The data proves HTML works, but also exposes its inefficiency — every additional interactive element burns more tokens, more money. This isn't UI optimization; it's a signal from the entire AI economic model: when output shifts from "writing text" to "painting neural video," inference compute demand will overtake training, and the current "token economy" may need to be rewritten.

The real question isn't "Will AI get too smart?" — it's "Are we willing to widen the output channel to match the brain's highway?" HTML is just a temporary bridge. When diffusion models make interactive video the default, AI will no longer be a chatbot — it'll be a cognitive partner that can look at the screen with you and modify things in real time. At that point, the $2 trillion compute flywheel (referencing prior OpenAI/Anthropic commitments) won't be burning money on training — it'll be burning it on real-time visual rendering.

The data is on the table: the brain has a 10-lane highway ready, and we're still driving on a single lane. Karpathy's hot tip isn't teaching you a shortcut — it's reminding us that the interface revolution has only just begun.

This article is based on publicly available data as of May 2026.