Andrej Karpathy, co-founder of OpenAI and former Director of AI at Tesla, has released an open-source project that systematically scores every US occupation by its exposure to artificial intelligence. The project, available at github.com/karpathy/jobs with a live demo at karpathy.ai/jobs, scrapes all 342 occupations from the Bureau of Labor Statistics Occupational Outlook Handbook, scores each using Google's Gemini Flash model accessed via OpenRouter, and presents results as an interactive treemap visualization. The complete data pipeline — Playwright for scraping, BeautifulSoup for HTML parsing, and LLM-based scoring — is fully reproducible and requires only a free OpenRouter API key to run.
The scoring methodology applies a single AI Exposure axis from 0 to 10, accounting for both direct automation (AI replacing the work entirely) and indirect displacement (AI multiplying worker productivity so that fewer workers are needed). Karpathy's key heuristic is whether a job is fundamentally digital: occupations performable entirely from a computer carry the highest exposure, while jobs requiring physical presence, manual dexterity, or real-time human interaction have natural barriers. The dataset yields an average exposure of 5.3 out of 10 across all 342 occupations. Medical transcriptionists score a maximum 10, while software developers, paralegals, data analysts, and editors cluster at 8 to 9. At the low end, roofers, janitors, and construction laborers score 0 to 1, with electricians, plumbers, and firefighters in the 2 to 3 range. Registered nurses and physicians land in the moderate 4 to 5 range.
The visualization renders each occupation as a rectangle sized proportionally to total employment and colored on a green-to-red scale by exposure score, grouped by BLS category. This design makes aggregate sector risk immediately legible rather than requiring users to scan ranked lists. Karpathy also packaged all data — statistics, tier breakdowns, exposure by pay and education level, growth projections, and all 342 scored occupations with LLM-generated rationales — into a single roughly 45,000-token prompt.md file, enabling anyone to conduct a data-grounded LLM conversation about labor market disruption without running any code.
Karpathy left OpenAI in 2024 and now runs Eureka Labs, an AI education startup, giving him no obvious institutional stake in where the scores land. His choice of Gemini Flash via OpenRouter rather than any OpenAI model suggests he wasn't trying to flatter his former employer. The scoring rubric tracks closely with an analytical frame visible across his public writing: fully digital jobs that produce verifiable outputs — code, documents, transcripts — score highest; anything requiring physical presence or real-time human judgment scores lowest. The 342-occupation dataset is now freely available as both raw JSON and the prompt.md file. Anyone who disputes a score can rerun the pipeline with a different model or a modified prompt and publish their own version.