Andrej Karpathy has published an interactive visualization scoring AI exposure across all 342 US occupations tracked by the Bureau of Labor Statistics. Available at karpathy.ai/jobs, the tool rates each job on a 0–10 scale and renders them as bubbles where area encodes total employment and color encodes exposure level. Source code is on GitHub.
The methodology matters. Rather than treating all 342 occupations equally, Karpathy weights each score against actual BLS employment figures, so high-headcount jobs drive the overall average rather than niche roles with small workforces. That makes the weighted mean a genuinely useful number instead of an artifact of how the government categorizes work.
A "wages exposed" metric adds economic specificity: it totals annual wages concentrated in occupations scoring 7 or above. The tool also breaks down exposure by pay level and education level, letting users check whether high-credential or high-wage workers <a href="/news/2026-03-14-tech-layoffs-45000-march-ai-attributed">face more risk</a> than the aggregate suggests. The BLS grounding makes the analysis reproducible.
The model choice will get attention. Karpathy used Google's Gemini Flash to score all 342 occupations, not an OpenAI model. With no current lab affiliation, he had no institutional reason to favor either provider. Gemini Flash's large context window suits batch processing of occupational descriptions in fewer API calls, and its pricing scales well for classification workloads at this volume. That Karpathy chose it over OpenAI's models is a revealed preference the AI community is likely to read carefully.
Since founding Eureka Labs in 2024, Karpathy has drawn on multiple model providers across his public work rather than defaulting to one. His selection of Gemini Flash here, on a project where context window size and API cost were real constraints, is the kind of real-world endorsement that carries more weight than any benchmark.