Andrej Karpathy published an analysis at karpathy.ai/jobs/ mapping LLM-specific automation exposure across US job categories. The page was returning his biography rather than the analysis at time of writing, so what follows draws on Hacker News discussion, where the piece surfaced under the title "AI Exposure of the US Job Market." According to commenters there, <a href="/news/2026-03-14-andrej-karpathy-scores-ai-exposure-of-342-us-occupations-using-gemini-flash-llm">his analysis</a> assesses which occupational roles face the most direct substitution risk from current-generation large language models with tool access. Karpathy is a founding member of OpenAI, former Director of AI at Tesla, and creator of Stanford's CS231n deep learning course — his framing of this question carries weight in both technical and policy circles.
The key methodological constraint, per HN discussion: Karpathy scopes the map strictly to LLM-tool substitution, excluding AI-driven robotics and physical automation. That makes it a near-term baseline, not a comprehensive automation forecast. Commenters noted that widening the scope to autonomous systems and robotic workers would push virtually every occupational category into high-exposure territory. The LLM-only framing is consistent with the work at Eureka Labs, the AI education startup he founded in 2024 — the map appears aimed at communicating what language models can do today, specifically, rather than modeling the full automation trajectory.
That scoping choice takes on a different character when you look at Karpathy's actual background. His MSc at the University of British Columbia focused on machine learning for physically-simulated robotic figures. His 2015 DeepMind internship placed him on the deep reinforcement learning team alongside Koray Kavukcuoglu and Vlad Mnih. At Tesla he built Autopilot's computer vision stack before contributing to Tesla Optimus, the company's humanoid robot program. The person most qualified to include robotics in this map chose not to. That is a studied editorial decision by one of the field's most credentialed roboticists, not an oversight.
The practical gap for workforce planning is concrete: occupations most exposed to physical automation — transportation, warehouse operations, manufacturing assembly — register as lower-risk under a LLM-only lens, even as Agility Robotics, Figure, and Boston Dynamics deploy capable systems into those exact roles. Policymakers who use Karpathy's map as a primary planning signal risk systematically under-investing in retraining for blue-collar workers in physically-embodied jobs. The analysis — if it performs as HN commenters describe — is a defensible snapshot of LLM-driven displacement. Karpathy has said publicly that he sees AI education as the most urgent intervention in the current transition; that context suggests the map's narrow scope is a feature, not a ceiling.