Karpathy used AI to score 342 US occupations, to see which professions are most affected by AI.
Then the project got too popular, so he deleted it.
He scraped data for 342 occupations from the US Bureau of Labor Statistics, then used the Gemini Flash model to give each a 0-10 AI exposure score.
Simply put, exposure means: the proportion of a job’s tasks that can be moved to a computer to complete.
Jobs that can be done entirely on a computer score high; those requiring physical work score low.
After finishing, he created a heatmap:
The rectangle’s area represents the number of people in that occupation, and the color represents the AI impact score.
You can switch between different dimensions: AI exposure, median salary, employment growth rate, education requirements.
The project exploded immediately after launch, Elon Musk retweeted it, and major media outlets reported on it.
Then Karpathy deleted the project.
He later posted a message on Twitter explaining:
“This was a Saturday morning two hour vibe coded project… It’s been wildly misinterpreted, so I took it down.”
Translation: A two-hour vibe coding project on a Saturday morning, was severely misinterpreted across the internet, so I took it down.
He also specifically emphasized one point:
“The ‘exposure’ was scored by an LLM based on how digital the job is. This has no bearing on what actually happens to these occupations.”
Exposure only measures how digital a job is, which is a separate issue from whether the job will disappear.
Later, someone cloned the project and redeployed it, and Karpathy restored the repository.
But this process of launch → viral → delete → restore itself already speaks volumes:<span leaf="" style="color: rgba(0, 0, 0, 0.9);font-family: mp-quote, 'PingFang SC', system-ui, -apple-system, BlinkMacSystemFont, 'Helvetica Neue', 'Hiragino Sans GB', 'Microsoft YaHei UI', 'Microsoft