Andrej Karpathy and the People Who Actually Build the AI

The headlines go to chatbots and CEOs. But somebody writes the training loop, labels the data, and ships the model. This is a record-first profile of Andrej Karpathy — one of the most visible of those builders — and the wider craft he represents. NU commentary and analysis of the public record.

1. Start with the feeling

Most people meet "AI" as a text box that talks back. It feels like magic, or like a threat, depending on the day. What it almost never feels like is work done by specific people — engineers staring at loss curves at 2 a.m., arguing about data quality, deciding what the thing should refuse to do. That gap between the polished output and the human labor behind it is where a lot of confusion lives. So NU does the boring, honest thing: name a builder, check the record, and separate what he did from what gets said about him.

2. The fair test

Claims about tech figures inflate fast — "father of," "genius behind," "predicted everything." The fair test is simple: what is documented? Where Karpathy's own writing, his employers' statements, or reliable biographies confirm a fact, NU states it. Where something is a coinage, an opinion, or folklore, NU labels it. No "the man who built ChatGPT" shortcuts — that is a team product, not one person's trophy.

3. The documented career

Andrej Karpathy was born in Bratislava, Slovakia, moved to Canada as a teenager, and studied at the University of Toronto before earning a PhD at Stanford, where he worked on deep learning and computer vision under Fei-Fei Li【1】【2】. At Stanford he designed and taught CS231n, "Convolutional Neural Networks for Visual Recognition," one of the first widely followed deep-learning courses【1】.

He was a founding member / research scientist at OpenAI, which launched in late 2015【1】【3】. In 2017 he left to become Director of AI at Tesla, where he led the team behind the Autopilot computer-vision stack until he departed in 2022【1】【4】. He returned to OpenAI in early 2023, then left again in February 2024, stating he was departing on good terms to pursue personal projects【1】【5】.

In July 2024 he announced Eureka Labs, an "AI-native" education startup aimed at building AI teaching assistants and courses; his first course material, LLM101n, has been developed in the open【6】. He also runs a heavily watched YouTube channel where he builds neural networks and language models from scratch on camera — including the "Let's build GPT" and "nanoGPT" walkthroughs【1】【7】. None of the above is in dispute.

4. The coinages — credited precisely

Two phrases that now float around the industry trace back to Karpathy specifically, so NU credits them carefully.

• "Software 2.0." In a 2017 essay, Karpathy argued that for many problems we no longer hand-write the logic; instead we specify a goal and let an optimizer find the program inside a neural network's weights. The data and the training become the "source code"【8】. The idea of learned programs is older; the memorable framing is his.
• "Vibe coding." In a February 2025 post on X, Karpathy described "a new kind of coding... where you fully give in to the vibes" — leaning on AI to generate code while the human steers loosely. The phrase spread fast and was even named a notable new term by dictionary watchers in 2025【9】【10】. NU notes the obvious caveat Karpathy himself implied: it is great for throwaway projects, riskier for production code — a style, not an engineering standard.

These are his coinages, not laws of nature. Useful labels; treat them as labels.

5. Why these builders matter

Karpathy is one name in a much larger crew. Modern AI rests on documented contributions from many people and labs: the "Attention Is All You Need" team at Google that introduced the Transformer in 2017【11】; researchers like Geoffrey Hinton, Yoshua Bengio, and Yann LeCun, who shared the 2018 Turing Award for deep learning【12】; Fei-Fei Li, whose ImageNet dataset helped ignite the field【2】; and the large, mostly uncredited workforce of data labelers and annotators whose work makes "training data" exist at all【13】.

Why surface them? Because agency lives with people, not with the product. When a model is biased, refuses, hallucinates, or impresses — that traces to choices someone made: what data, what guardrails, what objective. Naming the builders is how you keep AI a thing humans are accountable for, instead of a weather system that just "happens." Records over spin.

6. What NU is not claiming

• NU does not call Karpathy "the creator of ChatGPT" or any single product. He contributed to teams; the products are collective.
• "Software 2.0" and "vibe coding" are framings and opinions, not proven engineering doctrine.
• The roster in §5 is illustrative, not a ranking — AI has thousands of material contributors, and any "great man" telling of it is spin.

7. Read the record yourself

• Karpathy's own site, blog, and YouTube — primary source for his coinages and teaching【1】【7】【8】.
• Eureka Labs and the LLM101n repo — what he is building now, in the open【6】.
• The Turing Award and "Attention Is All You Need" — the broader builder record【11】【12】.

NU points to the people, dates, and documents, marks the coinages as coinages, and lets you weigh it — kooky till proven.

Note on imagery: the portrait is a real photograph, "Andrej Karpathy, OpenAI" from Wikimedia Commons, credited to Gladwin Analytics under CC BY 3.0 — file page linked in Sources. Exact direct URL flagged for verification.

Sources

1. Wikipedia — "Andrej Karpathy" — en.wikipedia.org/wiki/Andrej_Karpathy
2. Wikipedia — "Fei-Fei Li" / "ImageNet" — en.wikipedia.org/wiki/Fei-Fei_Li
3. OpenAI — "Introducing OpenAI" (Dec 11, 2015) — openai.com/index/introducing-openai
4. Tesla AI / press coverage of Karpathy's 2017–2022 Autopilot role — reuters.com (Karpathy Tesla departure, July 2022)
5. Karpathy departure post, Feb 2024 — x.com/karpathy ; coverage at theverge.com
6. Eureka Labs (official) — eurekalabs.ai ; LLM101n repo — github.com/karpathy/LLM101n
7. Andrej Karpathy — YouTube channel ("Let's build GPT", nanoGPT) — youtube.com/@AndrejKarpathy
8. Andrej Karpathy — "Software 2.0" (2017) — karpathy.medium.com/software-2-0-a64152b37c35
9. Andrej Karpathy — "vibe coding" post (Feb 2025) — x.com/karpathy
10. Merriam-Webster / dictionary "words to watch" coverage of "vibe coding," 2025 — merriam-webster.com
11. Vaswani et al. — "Attention Is All You Need" (2017) — arxiv.org/abs/1706.03762
12. ACM A.M. Turing Award 2018 (Hinton, Bengio, LeCun) — amturing.acm.org
13. Reporting on data-labeling labor (e.g., Time, "OpenAI Used Kenyan Workers," Jan 2023) — time.com

NU original — commentary and analysis of the public record, "records over spin, kooky till proven." Verify every cited source yourself.