Andrej Karpathy

AI researcher and educator. Founding member of OpenAI and later Director of AI at Tesla, who moved to AI education and started Eureka Labs in 2024. He appears in this vault twice. The gist at karpathy-llm-wiki, which this vault’s schema instantiates, is his. And he is cited on air in 2026-10-05-youtube-poteto-software-factory.

What she borrows from him

lauren-tan uses two of his framings to place her method (18:17 to 18:42). The captions render his name as “Carpathy”.

• Hill climbing. She calls it “a term the labs talk about a lot”: you have a rubric or a way to score something, and with a loop an agent can keep improving against it. That is what verification delivers.
• autoresearch. She points at Karpathy’s autoresearch as “famously released” and carrying these ideas. The repo was created 2026-03-06 and its README signs it ”-@karpathy, March 2026.” It gives an agent a small but real single-GPU nanochat training setup, lets it edit one file (train.py, which holds the GPT model, the Muon-and-AdamW optimizer, and the loop), run a fixed five-minute experiment, keep the improvement or discard the regression, and repeat; the README says roughly twelve experiments an hour and about a hundred while you sleep. That loop is hill climbing in the literal sense. The human writes program.md, the agent’s instructions, which the README calls “essentially a super lightweight skill”, a line that ties straight into skills-as-process. See determinism-extraction.

The distinction the host draws

The host separates Karpathy’s original vibe coding, “the code almost doesn’t exist, you forget that code might be a thing,” from Tan’s method, where the code and the environment are essential and bad input gives bad output (54:49). Karpathy’s autoresearch sits between them: the human stops running experiments but still owns the environment, the reward, and the boundaries.

Sources

• 2026-10-05-youtube-poteto-software-factory, captions at 18:17 and 18:35
• karpathy/autoresearch README https://github.com/karpathy/autoresearch/blob/master/README.md for the March 2026 date, the train.py/program.md layout, the five-minute budget, the 12-per-hour / 100-overnight counts, and the “super lightweight skill” line, checked 2026-10-05
• karpathy.ai https://karpathy.ai and the Wikipedia record for the OpenAI, Tesla, and Eureka Labs roles, checked 2026-10-05