Karpathy just published an AI job exposure map for 342 occupations It scored 0–10 by an LLM, visualized as a treemap where:
- size = employment volume
- color = exposure level
Jobs with irreducible physical presence. Roofers, janitors, construction laborers, plumbers, electricians, groundskeepers. You can’t reroof a house from behind a computer, so that’s the entire point of Karpathy’s scoring rubric
Nurses and home health aides sit in the 2–4 range despite being partially digitizable. The physical caregiving component anchors them. Same logic applies to dental hygienists, surgical techs, and EMTs
ii) Least safe (high exposure, 8–10):
Medical transcriptionists top out at 10. Pure digital input, pure digital output, zero physical component. Software developers, data analysts, paralegals, and financial analysts cluster at 8–9. Accountants land squarely in this zone too
iii) The ‘uncomfortable middle’ grey area is where it gets interesting. Secretaries, general office clerks, customer service reps, and cashiers are large employment blocks sitting at 5–7. These are the roles where AI potentially replaces some headcount; one person with AI tooling does what three did before
iv) The key takeaway from the visualization:
the biggest rectangles (highest employment) tend to sit in the moderate-to-high exposure range
The safest jobs are physically demanding but employ fewer people in aggregate
Food for thought if you’re thinking about reskilling to future proof your income - although once robots become more capable even some of the more manual tasks may be at risk
Perhaps plumbers will soon have their “LeBron James” of pipe laying
But even then if everyone retrains to become a plumber, won’t it become crowded too?