What does AI actually do to 143 million jobs?

Hypothesis

The AI-and-jobs conversation runs on vibes. If you put the actual Bureau of Labor Statistics data on screen — every occupation, sized by real employment — and let an LLM write a grounded deep-dive per occupation, does the conversation get calmer and more useful?

What we built

Two linked views, built on Andrej Karpathy's open jobs visualizer. First, a treemap of 342 occupations covering 143 million US jobs, sized by employment and colored by growth outlook, pay, education, or AI exposure — click any occupation for an LLM-generated deep-dive on how AI reshapes it. Second, an AI Adoption view: a dot grid that morphs between global AI usage (8.1B people, Feb 2026) and those same 143M jobs by AI exposure — surfacing the onboarding gap, the tens of millions whose work is changing faster than their tools, with a profile behind every band.

Learnings

• Grounding the deep-dives in per-occupation BLS numbers keeps the LLM honest — the same prompt without the data produces generic futurism.
• The "onboarding gap" framing landed harder than the exposure map itself: people recognise themselves in "work changing faster than your tools".
• Forking a well-built open project (karpathy/jobs) and adding one strong idea is a far better experiment shape than building a visualizer from scratch.

Log

• 2026-06-10 — AI Adoption dot-grid view added; published at /experiments/jobs-market.
• 2026-06-01 — Fork of karpathy/jobs; treemap + LLM deep-dives working.