Andrej Karpathy, the AI researcher and educator known for co-founding OpenAI and leading Tesla's Autopilot AI division, has released an open-source interactive treemap at karpathy.ai/jobs/ that visualizes 342 US occupations drawn from Bureau of Labor Statistics Occupational Outlook Handbook data, representing approximately 143 million jobs. Each tile in the map scales to total employment in that occupation, and users can toggle color layers between BLS projected growth, median pay, education requirements, and a novel "Digital AI Exposure" score generated by an LLM scoring pipeline. Karpathy is explicit that the tool is a development and research instrument, not a formal economic study, and cautions against treating its outputs as rigorous predictions.
The technical centerpiece of the project is a <a href="/news/2026-03-14-andrej-karpathy-scores-ai-exposure-of-342-us-occupations-using-gemini-flash-llm">general-purpose LLM-powered pipeline</a> included in the source code. Users can supply any prompt — scoring occupations by humanoid robotics exposure, offshoring risk, or climate impact, for instance — and rerun the pipeline to recolor the map accordingly. The included "Digital AI Exposure" prompt instructs an LLM to rate each occupation on a 0–10 scale, anchoring fully physical roles such as roofers and landscapers near 0 and fully digital routine roles like data entry clerks and telemarketers at 10. Software developers score 9/10, alongside graphic designers, translators, and data analysts. Karpathy frames a high score as predicting restructuring rather than elimination, explicitly noting that demand elasticity — software demand growing as each developer becomes more productive — can offset displacement effects.
The release hit a nerve on Hacker News, where commenters flagged a sharp irony: software developers, scoring 9/10 on AI exposure, are simultaneously experiencing one of the worst hiring markets in recent memory, with multiple people reporting job searches longer than twelve months. That sits awkwardly next to BLS projections showing above-average growth for the occupation. The gap suggests AI-driven restructuring may already be compressing software hiring faster than aggregate labor statistics can capture — a problem official government forecasts are structurally slow to reflect.
The tool connects to a thread Karpathy has been pulling for nearly a decade. His 2017 "Software 2.0" essay argued neural weights would progressively replace human-written code. In 2025 he coined "vibe coding" to describe AI-assisted development workflows. Now there is a quantitative scoring layer mapped onto BLS occupational data. The pipeline architecture itself is arguably the most reusable output: a lightweight pattern for using LLMs as a semantic scoring layer on top of structured government datasets. Karpathy described it as a "Saturday morning vibe coded project." A tool that rated software developers 9/10 on AI exposure was built by one of them, using exactly that workflow.