Karpathy urgently deletes repositories! AI career doomsday chart goes viral, 60 million white-collar jobs at risk

Author: Xin Zhiyuan

The “Workplace Verdict” in the AI Era: Will 60 Million People Lose Their Jobs?

Last night, AI expert Karpathy launched a trending project—karpathy.ai/jobs/—deeply analyzing the extent of AI’s “erosion” of employment.

He extracted 342 occupations from the U.S. Bureau of Labor Statistics (BLS) and assigned each job a risk score of AI replacement (0-10).

The results are shocking: the average exposure score across all industries is as high as 4.9.

Especially for “screen-dependent” jobs, the risk is rapidly increasing—most are within AI’s reach—

Software Developers 9/10

Medical Transcriptionists 10/10

Lawyers 8/10

Office Workers 9/10

Statistics show about 60 million jobs are at high risk, with over 42% scoring above 7, totaling $3.7 trillion in annual salaries.

Which jobs are safest? The answer is cleaners, plumbers, roofers—those involving complex physical labor are the safest havens.

Hinton once suggested: become a plumber.

In response, Elon Musk commented sharply, “All future jobs will become optional.”

Some netizens compiled a video gathering AI leaders’ predictions of unemployment.

All 60 million white-collar jobs in the U.S. are truly at risk!

This project went viral online, but just minutes after launching, Karpathy deleted the post, and it now leads to a 404 on GitHub.

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

On the project homepage, the leftmost section shows all key indicators, including exposure and salary.

Out of 342 occupations nationwide, totaling 143 million jobs, all are scored by Gemini Flash, with an average exposure level of 4.9.

Link:

Among these, jobs with the highest impact (6-10) account for 42%, about 59.9 million; those with the least impact (0-1) make up only 4%, roughly 6.2 million jobs.

Jobs paying over $100,000 annually (score 6.7) are more easily replaced by AI; meanwhile, jobs earning less than $35,000 (score 3.4) are least affected.

Moreover, jobs requiring a bachelor’s degree are most vulnerable to AI disruption.

Overall, AI is targeting jobs based on “information processing density.”

White-collar jobs relying on text processing, data analysis, coding, and standardized workflows—regardless of high salaries—are collectively “lighting up red.”

Conversely, roles involving physical operations, complex human interactions, or on-site real-time judgment remain in the safe zone.

White-Collar Job Massacre

On the right side of the homepage, similar jobs are tightly grouped.

First, let’s look at jobs with exposure scores above 6.

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

Their median annual salaries hover around $43,000, with education requirements generally high school diploma.

For example, office clerks (9/10) have a median salary of $43,630, with 2.6 million positions.

Financial clerks (9/10) have a median salary of $48,650, with 1.2 million positions.

These roles mainly involve routine tasks like data entry and document formatting, which have been largely digitized and standardized, making them highly susceptible to AI automation.

The “Business and Financial Operations” segment in the upper right is almost entirely red.

These jobs have median salaries between $50,000 and $100,000, requiring a bachelor’s degree.

For example, financial analysts (9/10) with a median salary of $101,910 and 429,000 positions.

Their work involves processing large datasets, trend analysis, and report generation—areas where AI excels.

Of course, computer-related roles are also significantly impacted. Dario Amodei predicted that within 6-12 months, AI will replace software engineers.

In the chart below, software engineers (9/10), computer systems analysts (8/10), and support specialists (8/10) are all in high-risk zones.

They earn median salaries up to $130,000 but are among the most easily replaced.

Other high-risk roles include lawyers (8/10), data scientists (9/10), graphic designers (9/10), cashiers (7/10), all facing high AI replacement risk.

Notably, medical transcriptionists are among the highest risk positions.

Become a Plumber

Today, the safest jobs are truly those involving “manual and physical interaction.”

The interactive chart clearly shows that large green areas are mostly related to complex on-site environments and hands-on roles.

For example, construction and specialized trades have an average exposure index between 1-3, requiring physical labor that must be done by humans.

Take plumbers, pipefitters, and steamfitters—they only need a high school diploma, with a median salary of $62,970, and are among the least likely to be eliminated.

Their core work involves heavy physical labor, requiring agility, strength, and the ability to handle unexpected situations in confined spaces or complex construction sites.

AI still cannot perform these core installation and repair tasks.

Similarly, food service roles like chefs, waiters, bartenders, and food processors are in the safe zone.

Other relatively safe jobs include hairdressers, animal caretakers, cleaners, personal healthcare workers, and material handlers.

In summary, Hinton’s statement still holds value.

The whole network exploded with discussion, and Karpathy responded

Last night, this chart went viral, with many predicting white-collar workers are doomed.

Half a month ago, Anthropic released a report titled “The Impact of AI on the Labor Market: New Metrics and Early Evidence.”

Similar to Karpathy’s data, it states that AI covers up to 75% of tasks for current programmers.

Close behind are customer service reps, data entry clerks, and medical record clerks—those most impacted by AI.

In contrast, about 30% of jobs are relatively unaffected, such as chefs, lifeguards, and dishwashers, because these require significant physical human effort.

However, current AI adoption is only a small fraction of its theoretical potential.

This caused widespread panic on social media, prompting Karpathy to delete the data quickly.

He explained, “This was just a weekend hobby project I coded in 2 hours based on intuition, overinterpreted by everyone.”

Harvard’s Hard Data: AI is Not Just “Killing” Jobs

The panic is real, but it’s not the whole picture.

Harvard Business School professor Suraj Srinivasan, along with researchers from HKUST and Ohio State University, published a significant working paper titled “Replacement or Complement? The Impact of Generative AI on the Labor Market,” offering a more rigorous and complex answer.

Link:

The research team compiled a dataset covering nearly all online job postings in the U.S., tracking real-time supply and demand from 2019 to March 2025.

First, look at substitution.

After ChatGPT’s release, the top 25% of jobs with the highest automation potential saw an average quarterly decline of 95 hires per company, a 17% drop.

Finance and tech sectors lead the way, with roles like clerks, payroll processors, medical transcriptionists, and telemarketers—“screen-based” jobs—being systematically phased out by AI.

Next, enhancement.

In the same period, the top 25% of jobs with the highest augmentation potential saw an average quarterly increase of 80 hires per company, a 22% rise.

Roles like microbiologists, financial analysts, and clinical neuropsychologists share a common trait: some tasks can be accelerated by AI, while others rely on human experience, intuition, and social skills.

Behind these figures is a sophisticated quantification method.

The team used GPT-4o to evaluate over 19,000 tasks across 900+ occupations, classifying each task into four levels—“no exposure,” “direct exposure,” “application exposure,” and “image exposure”—based on whether AI can halve the task completion time, then weighted by task importance within each job to calculate “automation scores” and “augmentation scores.”

Skill-level differentiation is even more striking.

High-automation jobs see a 24% drop in AI-related skill demand, with overall skill requirements shrinking and new skills appearing less frequently.

These roles are being “stripped,” and as AI takes over most structured tasks, remaining work becomes simpler and more standardized, reducing the demand for human workers.

In high-augmentation roles, the trend reverses. AI-related skill demand increases by 15%, with overall skill requirements and new skill development rising.

These jobs are becoming more complex; workers need to use AI tools, supervise AI outputs, and coordinate human-AI collaboration. For example, in finance, investment managers and analysts use AI to process vast market data, but final judgments and decisions still rest with humans.

AI hasn’t uniformly cut into all white-collar jobs. It resembles a “career reshuffle”: routine information handlers are being eliminated, while those who can collaborate with AI are becoming more valuable.

How Long Is the Window?

Karpathy deleted the post, but the data can’t be erased. Harvard’s study is more cautious but equally conclusive.

Whether looking at Gemini Flash’s scores or empirical data from the U.S. job market, the message is clear: AI-driven restructuring of white-collar jobs is already underway.

It’s not a blanket slaughter but a process of differentiation.

Jobs with tasks that can be fully described and processes that can be broken down and standardized are being cut.

Jobs that require judgment in ambiguous situations, building trust between people, or making final decisions based on AI outputs are becoming more valuable.

This differentiation has a harsh consequence.

The traditional white-collar career ladder’s first rung—routine tasks like data entry, report writing, entry-level coding, basic analysis—is being pulled away.

Young people start here, doing repetitive work, accumulating experience and judgment, eventually becoming irreplaceable.

Now, AI is removing this first step.

The entry is narrowing, but the rewards at the top are greater.

For everyone still in the workforce, the only real question is:

What proportion of your work cannot be done by AI?

If the answer makes you uneasy, the time to act is not tomorrow but now.