Andrej Karpathy spent a Saturday morning building a tool that tells 143 million American workers how replaceable they are. Then he deleted the source code.
The OpenAI co-founder and former Tesla AI director published an interactive treemap on March 15 that scores 342 occupations from the Bureau of Labor Statistics on a 0-to-10 AI exposure scale. Rectangle size shows how many people hold each job. Color shows how vulnerable it is. The whole thing looks like a heat map of professional anxiety, and it went viral within hours.
The heuristic is blunt
Karpathy's core logic boils down to one question: can you do this job entirely from a home office, on a screen? If yes, your score is at least a 7. Medical transcriptionists got a perfect 10. Software developers landed at 8 to 9. Data analysts, paralegals, copywriters, all in that same range. Roofers and janitors scored 0 to 1, not because the work is easy, but because a large language model can't crawl onto your roof.
The weighted average across all 342 occupations came in at 4.9, though some outlets have reported 5.3 using unweighted figures. Either way, it places the median American worker squarely in the zone Karpathy labels "moderate exposure," which is the kind of phrase that sounds reassuring until you realize it means half the workforce scored higher.
Your salary is the problem
The uncomfortable finding, buried in the data rather than in any headline Karpathy wrote, is the salary correlation. Professions earning more than $100,000 a year averaged a 6.7 exposure score. Those under $35,000 averaged 3.4. The pattern is almost too clean: the more you earn sitting at a desk, the more an LLM can theoretically do your job. According to one breakdown shared widely on X, jobs scoring 7 or above represent roughly 59.9 million workers and about $3.7 trillion in annual wages.
That $3.7 trillion number has been bouncing around social media as if it represents money about to vanish. It doesn't. It's the total payroll of workers in high-exposure roles, which is a very different thing. Most of those people will see their jobs change, not disappear. But the figure makes for a good panic metric, and panic metrics travel fast.
An LLM grading LLMs
The methodology deserves more scrutiny than it's getting. Karpathy fed each BLS occupation description into a large language model with a structured scoring rubric, and the model returned an exposure rating. The entire pipeline was open source: scraper, scorer, visualization. No black box, which is refreshing. But using an LLM to evaluate how replaceable jobs are by LLMs introduces an obvious circularity. The model has no particular insight into its own capabilities or limitations. It's pattern-matching against job descriptions, not measuring actual automation feasibility.
Karpathy was upfront about this on his project page, calling it "not a report, a paper, or a serious economic publication" but rather "a development tool for exploring BLS data visually." The caveat is right there. It just didn't travel as far as the treemap did.
"This was a saturday morning 2 hour vibe coded project inspired by a book I'm reading," he wrote on X after the post went viral. He then removed the GitHub repository, though the website remains live. The repo deletion feels like a researcher realizing his weekend sketch got mistaken for a blueprint.
Musk chimed in, obviously
Elon Musk responded on X by declaring that "all jobs will be optional" and predicting "universal high income." This is roughly what Musk says every time the topic comes up, so it's hard to read it as a meaningful data point. But the exchange did push Karpathy's visualization further into the mainstream, well past the AI-curious audience it was built for.
The timing is worth noting against a different data point. Citadel Securities published a report earlier this year showing Indeed job postings for software engineers are up 11% year over year in 2026. Karpathy gives software developers an 8-to-9 exposure score. Both things can be true simultaneously: AI is reshaping software work while demand for it keeps growing. A high exposure score doesn't predict job loss. It predicts job transformation, which is a less viral but more accurate reading.
What it actually shows
The most interesting feature of Karpathy's tool isn't the AI exposure layer. It's the toggle. The treemap lets you switch between BLS projected growth outlook, median pay, education requirements, and digital AI exposure. Salary and education views are drawn from government data, not LLM guesses. The AI layer is one lens among four, and probably the least reliable one, built to demonstrate what's possible with LLM-powered data coloring rather than to serve as workforce policy.
Karpathy's project page even suggests users could swap in their own prompts: exposure to humanoid robotics, offshoring risk, climate impact. The point was the pipeline, not the conclusions. A distinction his audience largely ignored.
Anthropic published a related labor market study earlier this year using anonymized Claude conversation data, and it arrived at a similar top-line finding: theoretical AI capability far outstrips actual workplace adoption. The gap between what AI can do and what employers are deploying remains wide. Karpathy's treemap maps the theoretical ceiling. The floor is still being negotiated.
