▶ YouTube Summaries

← All summaries
t3dotgg#▶ Watch

On "The Unreasonable Effectiveness of HTML"

2026-05-13 Wed ⏱ 36 min @t3dotgg

Theo reacts to Thoric's article "The unreasonable effectiveness of HTML" (Claude Code team) and Karpathy's tweet endorsing HTML output from LLMs. Thesis: Markdown is the dominant agent output format but is expressively weak; HTML lets agents emit richer artifacts — diagrams, mockups, interactive playgrounds, hierarchical PR reviews, shareable reports. Theo agrees with most of it but warns about token cost, novelty inflation, version-control friction, and the limitations of Anthropic's Artifacts implementation.

Why Markdown is wearing thin

  • Agents now emit large specs/plans; >100-line markdown files rarely get read end-to-end.
  • ASCII tables / unicode color swatches / mermaid diagrams break or misalign frequently.
  • Diff rendering in markdown is poor.
  • Users edit these files via the agent rather than by hand, removing one of markdown's main benefits.

Thoric's case for HTML

  • Higher information density: tables, CSS, SVG diagrams, code blocks, interactive widgets, images, spatial layouts in one file.
  • Better readability and ease of sharing (upload to S3, send link).
  • Interaction: sliders/knobs to tune designs; export-as-prompt buttons to round-trip back into Claude Code.
  • Data ingestion: Claude Code can pull from filesystem, git, MCPs (Slack, Linear, etc.) and produce a rich HTML synthesis.
  • Engagement: more fun → developer stays in the loop instead of rubber-stamping plans.

Theo's pushback / nuance

  • Some HTML "value" is novelty — once everyone spams HTML the signal/noise drops back.
  • Token cost: HTML is heavier than markdown; Thoric waves it off (1M-context Opus 4.7), Theo finds the dismissal glib.
  • Image generation in HTML reports often hallucinates b64 / URLs.
  • Markdown viewers (VS Code) already handle code-block syntax highlighting and headings well; many "HTML wins" are just the reader not knowing markdown features.
  • Mobile-responsive HTML adds bloat — the shared examples weren't mobile-friendly.
  • Version control: HTML diffs are noisy.
  • Anthropic's Artifacts implementation is "garbage" — slow load, broken UX, undermines the format's promise on Anthropic's own site.

Strong use cases Theo endorses

  • Exploration / fan-out: ask for N distinct design directions in one HTML grid → much more variety than serial follow-ups.
  • Custom PR review pages: re-organize diffs by importance, group related changes, annotate hot spots. Mentions devinreview.com as a working example.
  • Reports for stakeholders: nice HTML weekly status from Jira / git history. "Career-growth hack" — managers feel informed.
  • Throwaway editors: build single-purpose HTML UIs to manipulate one piece of data, with an export-as-prompt button. ~70% of his generated code is one-shot, thrown away after a single run.
  • Research/learning artifacts: tell the model you're a novice so it produces audience-facing explainers, not insider notes.

Karpathy's broader frame

  • Audio is the preferred input to LLMs; vision (images, animations, video) is the preferred output.
  • Progression: raw text → markdown → HTML → eventually diffusion-generated interactive video / simulations.
  • Open question: how procedural HTML/JS artifacts will weave together with neural-rendered UI.

Theo's closing thought

  • HTML is a stepping stone, not the endpoint. He bets MDX (HTML + React components) is the under-explored sweet spot — "all roads lead to React."
  • Hints at building an HTML-output mode and a custom skill into T3 Code.