Skip to content

Andrey Karpathy

Karpathy is the prototypical high-reach clarifier. From Stanford PhD to Tesla AI Director to 'Zero-to-Hero,' his method is consistent: code it from scratch, explain it visually, and kill the magic.
🌱 seedling tended 2026-05-17 case biography AI education Tesla OpenAI 
flowchart LR
  stanford[Stanford · CS231n] --> openai1[OpenAI · founding]
  openai1 --> tesla[Tesla · Autopilot]
  tesla --> openai2[OpenAI · LLMs]
  openai2 --> eureka[Eureka Labs · Edu]
  eureka --> clarifier[high-reach clarifier]
Read next

Public records · CS231n history · Karpathy.ai · Twitter/YouTube trace. S550 opening.

"The hottest new programming language is English."

Karpathy is the first modern "high-reach clarifier" in this folder. His career is a trace of taking the most complex, murky frontier of AI and leaving it smaller, clearer, and more shared for the next million developers.

1 · the life in three sentences

Born in Czechoslovakia (1986), moved to Toronto at 15; PhD at Stanford under Fei-Fei Li where he designed and taught CS231n (Convolutional Neural Networks for Visual Recognition). Founding member of OpenAI (2016), Director of AI at Tesla (2017–2022) where he led the Autopilot team through the transition to pure vision, returned to OpenAI to work on LLMs, and left in 2024 to found Eureka Labs. He is best known for his "LLM Zero to Hero" education series and minimal codebases like micrograd and nanoGPT.

2 · timeline (the verifiable spine)

year event
1986 Born in Czechoslovakia.
2005–2009 B.Sc. Computer Science & Physics, University of Toronto. Mentored by Geoffrey Hinton.
2009–2011 M.Sc. Computer Science, University of British Columbia.
2011–2016 PhD, Stanford University. Thesis: Connecting Image and Natural Language with Deep Neural Networks.
2015 Lead Instructor for CS231n at Stanford. First massive-scale DL course.
2016 Founding member and Research Scientist at OpenAI.
2017–2022 Director of AI at Tesla. Led the Autopilot Vision team. Scaled "Data Engine" and "HydraNet".
2022 Sabbatical, then leaves Tesla. Starts "LLM Zero to Hero" on YouTube.
2023–2024 Second stint at OpenAI.
2024 Leaves OpenAI. Founds Eureka Labs (AI + Education).

3 · how he did what he did

Operating layer — habits, environment, communication.

writing as thinking

Karpathy's blog (karpathy.ai) and Twitter are not "marketing"; they are his public thinking substrate. The Unreasonable Effectiveness of Recurrent Neural Networks (2015) was a godding move for RNNs before the "Zero to Hero" era. Repo frame: godding/method — write to compress your own understanding.

code as the only truth

micrograd, nanoGPT, llm.c. He avoids heavy frameworks in his educational work. By stripping the library, he forces the reader to meet the math. Repo frame: COMPRESSIONS — the smallest possible trace that still runs.

visual bandwidth

His diagrams are legendary for their clarity. He doesn't use "art"; he uses "spatial logic" to show where the gradients flow. Repo frame: JUST-GODDING — visuals as a sprinkle on top of text.

the "1-hour" format

Taking a field (LLMs) and compressing it into a 1-hour "Intro" video. This is high-reach godding: reducing the time-to-legibility for a million people.

4 · thinking evolution (the field-jumps)

period field tool produced tool that unlocked the next jump
2011–2016 Computer Vision CS231n · Image Captioning ability to teach DL at scale
2016–2017 Generative Models RNN blog posts · OpenAI early research understanding the latent space
2017–2022 Industrial AI (Tesla) Data Engine · HydraNet scaling AI to real-world edge cases
2022–2024 LLM Clarification nanoGPT · Zero to Hero compressing the LLM revolution
2024– AI Education Eureka Labs the ultimate godding of the school

5 · what he was wrong about, and what he refused

  • Tesla FSD timelines. Like everyone at Tesla during that era, he was over-optimistic about the "solved" date for Level 5 autonomy.
  • Left the "frontier" labs twice. He has twice left the "frontier" (OpenAI) to pursue education or personal projects. This suggests a refusal of the "AGI-at-all-costs" race in favor of the "legibility-for-all" mission.
  • Did not build a massive library. Unlike Jeremy Howard (fast.ai), Karpathy doesn't maintain a heavy wrapper library. He prefers "copy-pasteable minimal files."

6 · the substrate

  • Stanford/OpenAI/Tesla. He was at the three most important nodes of the AI revolution at exactly the right times.
  • Geoffrey Hinton/Fei-Fei Li. Mentored by the giants of the field.
  • The "Great Clarifier" persona. He has a rare combination of top-tier research talent and top-tier teaching talent.

7 · copyable vs not-copyable

copyable not copyable
Build from scratch (micrograd style) to learn. Being a founding member of OpenAI.
Write public blog posts to compress your thoughts. Mentorship by Geoffrey Hinton.
Use high-bandwidth visuals instead of just text/code. The Tesla compute budget for Autopilot.
Explain as you learn — the "Zero to Hero" move. His innate pedagogical talent.
Kill dependencies to find the atomic truth of a project. Being at Stanford during the 2012 ImageNet moment.

8 · epistemic status

confidence claim
attested Career timeline, project list, publications, blog posts, YouTube series.
inferred "Legibility-for-all" as a personal mission.
guess That he left OpenAI to "god" the education sector.

9 · sources & see-also

— Case opened S550.