Karpathy Deletes Repository in Emergency! AI Job Apocalypse Chart Goes Viral, 60 Million White-Collar Jobs at Risk

60 Million White-Collar Jobs in the U.S., Really in Danger!

This project went viral across the internet, yet within minutes of going live, Karpathy deleted the post, and it's now a GitHub 404.

Fortunately, AI influencer Josh Kale cloned the entire repository before it was taken down.

As can be seen, the far-left column of the project homepage lists all key metrics, including Exposure and Salary.

For all 342 U.S. occupations, comprising 143 million jobs, scored by Gemini Flash, the average exposure level for all occupations is as high as 4.9 points.

Portal: https://joshkale.github.io/jobs/

Among them, the most affected positions (6-10) account for 42%, or 59.9 million; the least affected (0-1) account for only 4%, with only 6.2 million positions.

Positions with annual salaries exceeding $100,000 (scoring 6.7) are more likely to be replaced by AI; while those with salaries below $35,000 are the least affected (3.4 points).

Moreover, occupations requiring a bachelor's degree are the most vulnerable to AI impact.

The overall trend is that AI is precisely targeting jobs along the axis of "information processing density."

Those white-collar clerical positions reliant on word processing, data analysis, code writing, and standardized processes, regardless of salary, are collectively "flashing red."

Conversely, positions involving physical operation, complex interpersonal interaction, or on-the-spot judgment remain in the safe zone.

White-Collar Job Massacre

In the interactive area on the right side of the homepage, professions with similar natures are closely clustered together.

Let's first count the positions with an AI exposure index exceeding 6 points.

In the lower-left area, mainly office and administrative roles, all scoring above 7, including clerks, receptionists, etc.

Furthermore, their median annual salaries generally hover around $43,000, and the educational requirement is typically a high school diploma.

For example, Office Clerk positions (9/10), median annual salary: $43,630, job count: 2.6 million.

Financial Clerks (9/10), median annual salary: $48,650, job count: 1.2 million.

The core responsibilities of these roles are mostly routine, data entry, and document formatting tasks, which are almost entirely digitized and routine, making them extremely susceptible to AI automation.

The细分d "Business and Financial Operations" roles in the upper-right corner are almost entirely in the red.

These positions have median annual salaries between $50,000 and $100,000 and require a bachelor's degree.

For example, Financial Analysts (9/10), median annual salary: $101,910, job count: 429,000.

The work content is almost "fully digital," involving processing large datasets, trend analysis, and report generation—precisely AI's forte.

Of course, computer-related jobs are also significantly impacted by AI. After all, Dario Amodei once predicted that AI would replace software engineers in the next 6-12 months.

As evident in the chart below, Software Developers (9/10), Computer Systems Analysts (8/10), and Computer Support Specialists (8/10) are all in the high-risk range.

They hold median salaries as high as $130,000 but are among the most replaceable groups.

Additionally, positions like Lawyers (8/10), Data Scientists (9/10), Graphic Designers (9/10), and Cashiers (7/10) all face high risks of being replaced by AI.

<极端的mg style="max-width:100%;overflow:hidden;" src="https://d1x7dwosqaosdj.cloudfront.net/images/2026-03/a7222106d1f64d479a1426d7b极端的e9d480.png" alt="">

It is worth mentioning that Medical Transcriptionists have the highest risk among all positions.

Go Be a Plumber

Now, the safest professions truly只剩下 those involving "manual interaction with physical entities."

In the interactive chart, it's clear that the large areas colored green are基本上 related to complex on-site environments and hands-on operational jobs.

As shown below, construction and specialized trades have average exposure indices between 1 and 3; these physical tasks must be performed by humans.

Take Plumbers, Pipefitters, and Steamfitters, for example: they only require a high school diploma, have a median salary of $62,970, and are among the least likely to be淘汰d.

Their core work属于 "heavy physical labor," requiring not only dexterity and strength but also the ability to solve various突发 situations in real-time within complex and changing environments like narrow crawl spaces or construction sites.

The core hands-on installation and repair work is still something AI cannot do.

Similarly, food service occupations, including chefs, waiters, bartenders, and food preparation workers, are also in the safe zone.

Furthermore, barbers, animal caretakers, cleaners, healthcare personal care aides, and transportation material movers are less impacted by AI.

In short, the value of Hinton's statement continues to rise.

The Internet Explodes, Karpathy Responds

Last night, this chart quickly went viral online, with many predicting that white-collar workers are about to suffer.

Two weeks ago, Anthropic also released a report titled "The Impact of AI on the Labor Market: New Metrics and Early Evidence."

Similar to Karpathy's data, the report指出 that tasks for computer programmers currently have an AI coverage rate as high as 75%.

Following closely are customer service representatives, data entry keyers, and medical records specialists—these are the "disaster zones" hit hardest by AI.

In contrast, about 30% of occupations are largely unaffected, such as cooks, lifeguards, and dishwashers, because these jobs require significant human physical collaboration.

However, the current actual adoption rate of AI represents only a small fraction of its theoretically feasible capabilities.

Precisely because this chart caused massive panic on social media, Karpathy subsequently urgently deleted the data.

He explained, "This was just a fun project I coded up 'by feel' over a weekend in 2 hours, and it's being over-interpreted by everyone."

Harvard Confirms: AI Isn't Just "Killing" Jobs

The panic is real, but panic is not the whole picture.

Harvard Business School professor Suraj Srinivasan,联合 researchers from Hong Kong University of Science and Technology and Ohio State University, published a heavyweight working paper, "Substitute or Complement? The Impact of Generative AI on the Labor Market," providing a more robust and complex answer.

Paper address: https://www.hbs.edu/ris/Publication%20Files/25-039_05fbec84-1f23-459b-8410-e3cd7ab6c88a.pdf

The research team directly pulled a dataset covering almost all online job postings in the U.S., tracking real changes in job supply and demand from 2019 to March 2025, entry by entry.

First, the substitution side.

After ChatGPT's release, hiring for the batch of positions with the highest automation potential (top 25%) decreased by an average of 95 per company per quarter, a drop of 17%.

The finance and tech industries were the first hit. "Screen搬砖" type jobs like文书 clerks, payroll clerks, medical transcriptionists, and telemarketers are being systematically phased out by AI.

Now, the enhancement side.

During the same period, hiring for the batch of positions with the highest enhancement potential (top 25%) increased by an average of 80 per company per quarter, a rise of 22%.

Microbiologists, financial analysts, clinical neuropsychologists—these professions share a common characteristic: part of the work can be accelerated by AI, while another part must be驾驭d by human experience, intuition, and social skills.

Behind these two sets of numbers lies a precise quantification method.

The research team used GPT-4o to evaluate 19,000 specific tasks across 900+ occupations one by one, categorizing them into four levels—"No Exposure," "Direct Exposure," "Application Exposure," "Image Exposure"—based on whether AI could reduce task completion time by more than half. They then combined the importance weight of each task within the occupation to calculate an "Automation Score" and an "Augmentation Score" for each profession.

The分化 at the skill level is even more striking.

In high-automation positions, demand for AI-related skills plummeted by 24%, total skill requirements also contracted, and the frequency of new skills appearing continued to decline.

These positions are being "hollowed out." As AI takes over most structured tasks, the remaining work becomes simpler and more standardized, and companies require less from people.

In high-augmentation potential positions, the trend is completely reversed. Demand for AI-related skills grew by 15%, and total skill requirements and the number of new skills are rising.

These positions have become more complex. Employees not only need to know how to use AI tools but also need the ability to supervise AI output and integrate human-AI collaboration processes. Taking the financial industry as an example, investment managers and analysts use AI to process vast amounts of market data, but the final judgment and decision-making still rest with humans.

AI is not wielding a broad axe against all white-collar workers equally. It's more like a "occupational reorganization." Pure information handlers are being淘汰d, while those who can collaborate with AI are becoming more valuable.

How Much Time is Left in the Window?

Karpathy deleted the post, but the data remains. Harvard's paper is calmer, but its conclusions are equally merciless.

Whether you look at Gemini Flash's score sheet or the empirical study covering the entire U.S. job market, they point to the same fact. The reorganization of white-collar jobs by AI is already happening.

It's just not a one-size-fits-all massacre; it's a分化.

What's being cut are those positions whose work content can be fully described and whose processes can be standardized and broken down.

What remains, and even becomes more valuable, are those positions that require judgment in gray areas, building trust between people, and making final decisions based on AI output.

This分化 brings a cruel consequence.

In the past, the first step on the white-collar career ladder was often standardized entry-level work: data entry, report writing, junior coding, basic analysis.

Young people started here, doing repetitive tasks, slowly accumulating experience and judgment, and eventually growing into irreplaceable individuals.

Now, AI is pulling out this first step.

The entrance has narrowed, but the reward at the end has反而 become greater.

For everyone still in the workforce, there's really only one question to answer.

What percentage of your work can AI *not* do?

If the answer makes you uneasy, then the time to act is not tomorrow, it's now.