The White-Collar Reckoning: AI Meta’s Looming Layoffs and Why Electricians Can’t Be Replaced

On the morning of Saturday, March 15, Andrej Karpathy—co-founder of OpenAI and former director of AI at Tesla—published a small website. It was, by his own description, a “two-hour vibe coded project inspired by a book I’m reading.” Within hours, it had redrawn the map of how millions of Americans understood their professional futures.

The tool at karpathy.ai/jobs visualizes 342 occupations drawn from Bureau of Labor Statistics data, covering roughly 143 million jobs across the U.S. economy. Each job receives an AI Exposure score from 0 to 10, generated by a large language model (LLM) scoring how much of a role’s core tasks current digital AI can perform or substantially reshape. The result is an interactive treemap where the size of each block reflects how many people work in a field, and the color tells you how much AI is bearing down on it.

The headline finding: the weighted average exposure across all jobs is 4.9 out of 10. But the distribution is the story. The higher a profession’s salary, the more exposed it is. Workers earning over $100,000 a year average a score of 6.7. Those earning under $35,000 average just 3.4. For three decades, knowledge work meant security. Now, it is the most threatened category of labor.

What the Tool Actually Shows—and Doesn’t

Before the viral spread stripped away its caveats, Karpathy was careful to state what his scores do and don’t mean. The tool’s own description is unambiguous: “This is not a report, a paper, or a serious economic publication—it is a development tool for exploring BLS data visually.” The scores are rough LLM estimates, not rigorous economic predictions. A software developer scoring 9 out of 10 doesn’t mean software developers are going to disappear; it means a large share of the tasks in that role are theoretically performable by current LLMs. Demand for those skills could easily grow as individual developers become dramatically more productive.

Equally important: the model scores jobs only for digital AI exposure. It does not attempt to account for demand elasticity, regulatory barriers, social preferences for human workers, or the rise of physical robotics. Karpathy even included a suggested prompt for users to re-run the scoring pipeline with entirely different criteria—”exposure to humanoid robotics, offshoring risk, climate impact.” The point was the pipeline, not the conclusions.

Elon Musk responded on X, declaring that “all jobs will be optional” and predicting “universal high income”—which promptly amplified the tool far beyond its intended audience. Karpathy subsequently deleted the GitHub repository hosting the raw code and data, while leaving the website itself live. The most plausible reason: he was uncomfortable with how definitively an exploratory, self-described hobby project was being cited as authoritative economic analysis.

The Exposure Tiers: Who Is Actually at Risk?

Using Karpathy’s published scoring rubric—which was built into the tool itself—the occupational landscape breaks into clear tiers. A score of 8–10 indicates work that is “almost entirely digital, knowledge-based,” where current AI can perform the core tasks. A score of 6–7 reflects predominantly knowledge work with some human judgment or physical presence still required. Below 5, physical reality provides meaningful protection for now.

The inversion of assumptions is striking. For decades, people worried about machines replacing assembly-line workers. The actual frontier of AI disruption has opened in offices, law firms, and tech companies—not on factory floors.

Meta and the Real-World Consequences

The same weekend Karpathy’s tool went viral, Reuters reported that Meta’s top executives had instructed senior leaders to begin planning for significant workforce reductions. Three anonymous sources cited by Reuters said the cuts could affect more than 20 percent of the company’s workforce. A Meta spokesperson called the reporting “speculative reporting about theoretical approaches.” No timeline or final headcount has been confirmed.

The scale, if realized, would be significant. Meta employed 78,865 people as of December 31, 2025. A 20 percent reduction would affect roughly 15,800 workers—its largest single workforce reduction since the “Year of Efficiency” cuts of late 2022 and early 2023, when the company eliminated around 22,000 positions. Meta’s stock climbed nearly 3 percent on Monday following the Reuters report.

The context is financial as much as operational. Meta revealed in its fourth-quarter earnings that AI-related capital expenditure for 2026 is projected at $115 billion to $135 billion—roughly double what it spent in 2025. Total company expenses for the year are forecast between $162 billion and $169 billion. CEO Mark Zuckerberg has made his priorities explicit: 2026 is, in his framing, a defining year for AI, anchored by Meta’s new Superintelligence Labs. Workforce cuts in non-AI areas are the logical counterweight to that investment surge.

Meta is not alone. In early 2026, software firm Atlassian announced a 10 percent reduction, citing AI-driven efficiency. So far this year, AI has been cited in over 12,000 job cuts across U.S. companies, according to consulting firm Challenger, Gray & Christmas. Some analysts, including OpenAI CEO Sam Altman, have suggested that AI is being used as cover for corrections to pandemic-era overhiring. The truth is likely both things simultaneously: genuine productivity gains from AI tools, and executives using the moment to achieve headcount reductions they may have pursued regardless.

Anthropic’s Related Finding: The Gap Between Capability and Adoption

Shortly before Karpathy’s tool appeared, Anthropic published its own labor market study using anonymized data from Claude conversations. The headline finding, as reported by Fortune, is that actual AI adoption in workplaces remains a small fraction of what AI tools are technically capable of performing. The gap between theoretical AI capability—what tools like Claude or GPT-4 can do—and what employers are actually deploying at scale remains wide. Karpathy’s treemap maps the theoretical ceiling. The floor is still being negotiated, workplace by workplace, contract by contract.

The Safest Job in the AI Economy: The Electrician

In Karpathy’s rubric, electricians score 2–3 out of 10. The explanation is simple: the core of the job cannot be done from a screen. Electricians work in unpredictable physical environments—crawl spaces, live panels, muddy trenches, and industrial sites where conditions change constantly. A language model cannot pull wire through conduit or troubleshoot a tripped breaker at a data center at 2 a.m.

That low exposure score coincides with a demand curve that is running almost vertically upward. According to the Bureau of Labor Statistics, electrician employment is projected to grow 9 percent over the next decade—more than twice the average rate for all occupations—creating approximately 81,000 job openings annually. The median annual salary sits at around $62,000, with the top 10 percent earning over $104,000. In data center construction, wages regularly exceed $100,000, and some specialized roles top $200,000.

The driver is AI itself. Electrical work accounts for 45 to 70 percent of total data center construction costs, according to the International Brotherhood of Electrical Workers. Google, which gave software developers a 9/10 exposure score on Karpathy’s scale, issued a policy report stating that a lack of electricians “may constrain America’s ability to build the infrastructure needed to support AI.” The company has donated $10 million to the Electrical Training Alliance to train 100,000 existing electricians and 30,000 new apprentices by 2030. Microsoft President Brad Smith has written publicly about electricians having to commute long distances or temporarily relocate because local supply cannot meet project demand.

The IBEW describes the shortage as a “life or death” situation for companies like Amazon, Meta, and Microsoft. Nearly 30 percent of union electricians are between the ages of 50 and 70. About 20,000 are expected to retire annually over the next decade. The Associated Builders and Contractors trade group estimates the construction industry will need 349,000 net new workers in 2026 alone. Against that backdrop, a generation of young people that was told a four-year degree was the only path to stability is increasingly reconsidering.

One Caveat the Original Article Got Wrong

The narrative that “electricians are safe from AI” is largely correct for now—but it carries a hidden asterisk. Karpathy’s tool scores only for digital AI exposure. Physical automation—robotics, autonomous vehicles, warehouse automation—operates on a separate trajectory that his scoring system explicitly does not address. Roofers score 0 on digital AI exposure. They may score differently on a robotics exposure prompt. Electricians score 2–3. The precision physical dexterity required for their work is beyond current robotics, but that ceiling is not fixed.

The honest framing is not “trades are forever safe.” It is that trades are safe right now, at this specific moment in AI development, in a way that screen-based knowledge work is not. That window may be a decade. It may be longer. It may not. Anyone entering the trades today is making a reasonable bet, with excellent near-term prospects. But “AI hasn’t figured out how to turn a wrench yet” is a description of the present, not a guarantee about the future.

The Real Takeaway from Karpathy’s Weekend Project

What Karpathy actually built is not a prediction machine. It is a mirror. It reflects, with unusual clarity, where the current frontier of AI capability overlaps with how Americans actually spend their working lives. That overlap is concentrated at the top of the salary distribution—among precisely the people who were supposed to be most insulated from technological displacement.

A high exposure score does not mean a job disappears. It means the job changes—possibly radically, possibly subtly. A software engineer with AI tools may do the work of three engineers from a few years ago. That is a transformation, not a termination—unless the company decides to employ one engineer instead of three. For the other two, the question of what comes next is not abstract. It is the defining labor question of the moment.

Karpathy built a two-hour Saturday project. Meta is planning cuts that may affect 15,800 people. The construction industry needs 349,000 new workers this year alone. These are not separate stories. They are the same story, unfolding at different speeds across different sectors of the same economy—one in which the most durable job credential right now may not be a computer science degree, but a journeyman electrician’s license.