Karpathy US Job Market Visualizer Scored 342 Occupations for AI Exposure — Then Got Deleted Within Hours

Andrej Karpathy, OpenAI co-founder and former Tesla AI director, published an interactive visualization on March 15 that scored every major U.S. occupation for AI exposure. Within hours, it went viral. Elon Musk amplified it with a bold claim: “All jobs will be optional.” Fortune ran a feature story. Hacker News lit up with 363 points and hundreds of comments. And then Karpathy deleted the GitHub repository.

The Karpathy US Job Market Visualizer remains live as an interactive treemap, but the source code is gone. What started as a self-described “saturday morning 2 hour vibe coded project” turned into one of the most widely discussed AI labor analyses of the year — and a case study in how quickly exploratory data work can be taken out of context.

What the Karpathy US Job Market Visualizer Actually Does

The tool pulls occupation data from the Bureau of Labor Statistics Occupational Outlook Handbook, covering 342 occupations and 143 million jobs across the U.S. economy. Each occupation appears as a rectangle in a treemap layout, sized by total employment. Users can toggle the color layer between BLS projected growth outlook, median pay, education requirements, and — the one that made headlines — “Digital AI Exposure.”

For the AI exposure layer, an LLM (Gemini Flash) reads each occupation’s BLS description and assigns a score from 0 to 10 based on how much of the work is “purely digital and screen-based.” The scoring framework uses clear anchors: 0–1 for physical, unpredictable environments (think roofers and landscapers), 8–10 for routine digital tasks (data entry, transcription, financial analysis).

The methodology is transparent, which is part of what makes the results both interesting and easy to over-interpret. Karpathy built a quick, visual exploration tool — not a peer-reviewed economic study.

The Numbers That Sparked the Firestorm

The headline figures from the Karpathy US Job Market Visualizer are striking:

Weighted average AI exposure across all jobs: 4.9 out of 10
42% of jobs scored 7 or higher, representing 59.9 million workers
Those high-exposure roles account for roughly $3.7 trillion in annual wages

But the most provocative finding is the income-exposure correlation. Workers earning over $100,000 per year averaged an AI exposure score of 6.7, while those earning under $35,000 averaged just 3.4. In other words, the more you earn, the more your work overlaps with what current AI can theoretically handle.

Specific occupation scores paint a vivid picture:

Occupation	AI Exposure Score	Employment
Medical transcriptionists	10/10	~53,000
Accountants & financial analysts	9/10	~1.5M
Lawyers	9/10	~1.6M
Software developers	8–9/10	~1.8M
Customer service representatives	8/10	~2.8M
Registered nurses	4–5/10	—
Electricians & plumbers	2–3/10	—
Roofers	0/10	—

There is also an education paradox buried in the data. Doctoral degree holders averaged 5.7 out of 10 on exposure, while occupations requiring no formal degree averaged 2.7. The implication: the more education your job requires, the more overlap it has with LLM capabilities — at least on paper.

Why Karpathy Deleted the Repo — and What That Says About AI Discourse

The deletion happened fast. Karpathy published the visualization on Saturday, March 15. By Sunday morning, the GitHub repository returned a 404. The interactive website stayed up, but the code was gone.

His explanation on X was brief: “It’s been wildly misinterpreted (which I should have anticipated even despite the readme docs) so I took it down.”

He didn’t specify what the misinterpretation was, but the context makes it fairly obvious. Within hours of publication, the tool was being cited as definitive proof that AI will eliminate tens of millions of jobs. Elon Musk’s repost — declaring “all jobs will be optional” and predicting “universal high income” — transformed a weekend data exploration into ammunition for sweeping economic claims.

The original GitHub README was explicit that this was “not a report, a paper, or a serious economic publication” but “a development tool for exploring BLS data visually.” That framing was lost almost immediately as the viral cycle took hold.

This pattern is increasingly common in AI discourse: an exploratory analysis or prototype gets amplified far beyond its intended scope, and the creator is left choosing between letting the misinterpretation spread or pulling the work entirely.

How This Compares to Other AI Job Impact Studies

The Karpathy US Job Market Visualizer is not the first attempt to quantify AI’s impact on employment, but it stands apart in approach and accessibility.

OpenAI’s own research (2023) estimated that roughly 80% of the U.S. workforce could have at least 10% of their tasks affected by GPTs, and about 19% could see 50% or more of their tasks impacted. That study used a more formal methodology with human annotators and GPT-4, and it focused on task-level exposure rather than whole-occupation scores.

The Brookings Institution has published multiple analyses mapping AI exposure across metro areas and demographics, typically finding that higher-paid, more-educated workers face greater theoretical exposure — consistent with Karpathy’s results.

McKinsey’s 2023 report projected that generative AI could automate activities absorbing 60–70% of employees’ time, with the biggest impact on knowledge work, customer operations, and sales.

Where Karpathy’s tool differs is in its immediacy and visual clarity. A static PDF report from a consulting firm doesn’t get 363 Hacker News upvotes. An interactive treemap where you can hover over your own occupation and see a score from 0 to 10 does. That accessibility is both its strength and its vulnerability — it makes the data feel more personal and more definitive than it actually is.

The methodological gap is also worth noting. Using an LLM to score jobs for LLM replaceability introduces a self-referential bias. The model is essentially evaluating which jobs it thinks it could do, which is a useful signal but not the same as measuring actual workplace deployment or economic displacement.

The “Exposure Is Not Elimination” Problem

One critical nuance that got lost in the viral spread: a high AI exposure score does not mean a job is about to disappear. Karpathy himself emphasized that high exposure signals transformation, not necessarily elimination.

Software developers scored 8–9 out of 10, yet demand for developers has not collapsed — if anything, AI tools are increasing developer productivity, which may expand the market for software rather than shrink it. The BLS still projects software development as one of the fastest-growing occupations through 2032.

The same logic applies to lawyers, financial analysts, and accountants. AI can draft contracts, summarize case law, and run financial models. But the judgment, client relationships, and regulatory navigation that these roles require are not captured by a single exposure score.

This is the fundamental limitation of any 0-to-10 scoring system for something as complex as labor market dynamics. It compresses multidimensional reality into a single number, which is useful for visualization but dangerous for policy conclusions.

What Happens Next

The Karpathy US Job Market Visualizer will likely be referenced for months in AI labor discussions, even with the repo deleted. The interactive site remains live, and screenshots are already embedded in dozens of articles and social media threads.

For anyone exploring this space, the tool works best as what it was originally intended to be: a starting point for visual exploration, not a conclusion. Toggle between the different data layers. Notice where AI exposure diverges from employment growth projections. Look at which high-exposure jobs the BLS still expects to grow.

The real value is not in the scores themselves but in the questions they prompt: If AI can theoretically handle 80% of a financial analyst’s screen-based tasks, what happens to the role? Does it shrink, transform, or become more productive? Those questions don’t have 0-to-10 answers.

FAQ

Is the Karpathy US Job Market Visualizer free to use?

Yes. The interactive treemap is accessible at no cost. The GitHub repository with source code was briefly available before Karpathy deleted it, but the visualization itself remains live.

How accurate are the AI exposure scores?

The scores are generated by an LLM (Gemini Flash) reading BLS occupation descriptions and estimating digital task overlap. They reflect theoretical exposure to current AI capabilities, not actual job displacement rates. Karpathy explicitly described this as a visual exploration tool, not a formal economic analysis.

Which jobs have the highest AI exposure according to the tool?

Medical transcriptionists scored 10/10, followed by accountants, lawyers, financial analysts, and software developers in the 8–9 range. Generally, occupations that are primarily screen-based and involve processing, analyzing, or generating text and data scored highest.

Which jobs are least exposed to AI?

Physical labor roles scored lowest. Roofers received 0/10, with construction workers, janitors, landscapers, and truck drivers ranging 1–2/10. Jobs requiring hands-on work in unpredictable physical environments showed minimal overlap with current AI capabilities.

How does this compare to other AI job impact studies?

The findings broadly align with research from OpenAI, Brookings, and McKinsey — all of which find that higher-paid, knowledge-work occupations face greater theoretical AI exposure. The key difference is format: Karpathy’s interactive visualization makes the data more accessible but also more prone to oversimplification compared to detailed academic or consulting reports.

Top AI Product