Andrej Karpathy

Timeline by topic

Topics follow the event kinds supplied in this index. Each entry links to its full record in the chronological timeline. Dates retain the precision supplied by the publisher.

Timeline

1. 1986-10-23Born in Bratislava, Czechoslovakia (now Slovakia)birthBratislava

   Moved with his family to Toronto at age 15.

   [48][47]
2. 2009BSc, University of TorontoeducationUniversity of TorontoToronto

   Bachelor's degrees in computer science and physics.

   [48]
3. 2011MSc, University of British ColumbiaeducationUniversity of British Columbia

   Master's work on physically simulated figures with Michiel van de Panne.

   [48]
4. 2011 – 2016-08PhD in computer science, Stanford UniversityeducationStanford University

   Advised by Fei-Fei Li at the Stanford Vision Lab; dissertation 'Connecting Images and Natural Language' completed August 2016.

   [12][2]
5. 2015Created and taught CS231neducationStanford University

   With Fei-Fei Li, designed Stanford's first deep learning course and served as primary instructor; enrollment grew from 150 (2015) to 750 (2017).

   [30][2]
6. 2015-05-21'The Unreasonable Effectiveness of Recurrent Neural Networks'publication

   Blog post plus char-rnn release showing character-level LSTM models writing plausible Shakespeare, LaTeX and code.

   [14][11]
7. 2015-12-11 – 2017-06Founding member and research scientist, OpenAIfoundedOpenAI

   Named among the founding research engineers and scientists in OpenAI's launch post.

   [28]
8. 2016-12-05OpenAI Universe / World of Bits releasedprojectOpenAI

   Led the World of Bits browser-task benchmark within OpenAI's Universe agent platform.

   [29]
9. 2016-12-19'Yes you should understand backprop'publication

   Essay arguing backpropagation is a leaky abstraction practitioners must internalize.

   [15]
10. 2017-06-20 – 2022-07-13Director (later Sr. Director) of AI, TeslaroleTesla

    Led Autopilot computer vision — data labeling, training, on-chip deployment — reporting to Elon Musk; briefly also Tesla Optimus.

    [39][1][40]
11. 2017-11-11'Software 2.0' essaypublication

    Argued neural networks mark a shift from hand-written code to weights compiled from data.

    [16]
12. 2019-04-22Presented at Tesla Autonomy DaymediaTesla

    Explained the neural-network stack, fleet learning and the data engine to investors.

    [31]
13. 2020MIT Technology Review Innovator Under 35awardMIT Technology Review

    Honored for neural networks that let automated cars 'see.'

    [50]
14. 2020-08Released minGPTproject

    Minimal, educational PyTorch re-implementation of GPT training and inference.

    [5]
15. 2021-08-19Led the vision deep dive at Tesla AI DaymediaTesla

    Presented the eight-camera vision stack producing a 3D vector-space representation for Autopilot.

    [32]
16. 2022-07-13Left Teslamilestone

    Announced his departure after a months-long sabbatical; said he would revisit AI technical work, open source and education.

    [40]
17. 2022-08-16Started 'Neural Networks: Zero to Hero'project

    The micrograd lecture opened a from-scratch course building neural networks step by step.

    [20][4]
18. 2022-10-29Lex Fridman Podcast #333media

    Long-form interview on Tesla AI, self-driving, Optimus, Software 2.0 and AGI.

    [33]
19. 2022-12Created nanoGPTproject

    A minGPT rewrite 'with teeth': ~300-line training loop reproducing GPT-2 (124M).

    [6]
20. 2023-01-24'The hottest new programming language is English'milestone

    Viral post anticipating natural-language programming of LLMs.

    [18]
21. 2023-02-09 – 2024-02-12Rejoined OpenAIroleOpenAI

    Built a new team working on midtraining and synthetic data generation, per his own bio.

    [1][41]
22. 2023-05'State of GPT' at Microsoft BuildmediaMicrosoft

    Explained the GPT assistant training pipeline from pretraining through RLHF.

    [22]
23. 2023-11-23'Intro to Large Language Models' videopublication

    One-hour general-audience talk covering what LLMs are, where they are headed, and security issues.

    [21]
24. 2024Started llm.cproject

    LLM pretraining in raw C/CUDA reproducing the GPT-2/GPT-3 miniseries.

    [7]
25. 2024-06-09'Let's reproduce GPT-2 (124M)' videomedia

    Four-hour from-scratch build of GPT-2 ending in an overnight training run.

    [23]
26. 2024-07-16Announced Eureka LabsfoundedEureka Labs

    An 'AI native' school built around a teacher + AI-assistant symbiosis; first product the announced LLM101n course.

    [42][3]
27. 2024-09-05Named to TIME100 AI 2024awardTIME

    Recognized among the year's most influential people in AI, with emphasis on his teaching.

    [43]
28. 2025-02-02Coined 'vibe coding'milestone

    X post naming the fully-delegated, code-forgetting style of AI-assisted programming.

    [19][51]
29. 2025-06-19'Software Is Changing (Again)' keynotemediaY Combinator

    Y Combinator AI Startup School talk introducing the 'Software 3.0' framing.

    [26]
30. 2025-10-13Released nanochatproject

    'The best ChatGPT that $100 can buy' — a minimal end-to-end LLM harness.

    [8]
31. 2025-10-17Dwarkesh Podcast interviewmedia

    Two-and-a-half-hour conversation: 'AGI is still a decade away,' RL critique, ghosts vs. animals, education.

    [35][34]
32. 2025-11-06'Vibe coding' named Collins Word of the Year 2025milestone

    The dictionary credited Karpathy with coining the term in February 2025.

    [45][49]
33. 2026-05-19Joined Anthropic's pretraining teamroleAnthropic

    Announced on X; building a team that uses Claude to accelerate pretraining research, under lead Nick Joseph.

    [46]

Themes

• Education as a first-class engineering projectstatedphilosophy

  From CS231n's browser demos through the Zero to Hero series to Eureka Labs' 'teacher + AI symbiosis,' he treats teaching not as outreach but as the core work — and calls Eureka 'the culmination of my passion in both AI and education over ~2 decades.'

  [30][4][42][3][1]
• Software 1.0, 2.0, 3.0statedphilosophy

  A decade-long framing project: 2017's 'Software 2.0' cast weights-trained-from-data as the new code; the 2025 'Software 3.0' talk casts English-prompted LLMs as the next major version — a new kind of computer programmed in natural language.

  [16][26][18]
• Human-readable machine learningstatedmethod

  His libraries are sized for reading, not just running: char-rnn, micrograd (~100 lines), minGPT (~300-line model), nanoGPT, llm.c in raw C/CUDA, nanochat end-to-end. 'GPT is not a complicated model,' the minGPT README insists.

  [11][10][5][6][7][8]
• Autonomy skepticismstatedbelief

  He repeatedly cools agent and self-driving hype: keep AI 'on the leash,' ship partial autonomy with an autonomy slider, expect a decade of agents — not a year — because the demo-to-product gap in self-driving took a decade to close.

  [26][35][36][33]
• Know the leaky abstractionsstatedbelief

  Backprop is a 'leaky abstraction'; neural nets are not off-the-shelf. His recipe: inspect data before model code, build dumb baselines, add verified complexity in small steps. Understanding is the part that cannot be outsourced.

  [15][17][38]
• LLMs as a new kind of computerstatedphilosophy

  LLMs are utilities, fabs and early operating systems rolled into one — 'people spirits' simulated by an autoregressive Transformer — so software should be rebuilt for agents, GUIs should speed human verification, and psychology is as relevant as architecture.

  [26][21]
• Ghosts, not animalsstatedbelief

  LLMs are stochastic simulations of people distilled from internet text — jagged intelligence with superhuman recall and real deficits — not artificial animals or brains. This shapes how he expects the technology to mature: summoned entities needing taste and oversight, not trusted autonomous creatures.

  [35][34][38]
• Self-driving as the reference lessonstatedinfluence

  Five years on Autopilot supply his analogies for AGI: demos deceive, products take a decade, globalization takes longer, and safety-critical deployment punishes the over-confident.

  [36][35][33]
• Building and teaching in publicreportedpractice

  Repos, lectures and announcements ship together on GitHub, YouTube and X; Eureka Labs was announced early specifically so he could 'build publicly.' Coverage consistently frames him as the field's most-watched explainer.

  [42][43][1][8]

Works and projects

• Connecting Images and Natural Languagepaperpublished2016-08Stanford University

  PhD dissertation: multimodal embeddings, image captioning and dense captioning connecting visual data to natural language.

  [12]
• Deep Visual-Semantic Alignments for Generating Image Descriptionspaperpublished2015

  CVPR 2015 paper with Fei-Fei Li aligning image regions and sentence snippets, a foundation of his dissertation.

  [13]
• CS231n: Convolutional Neural Networks for Visual Recognitionothercompleted2015Stanford University

  Stanford's first deep learning course, co-designed with Fei-Fei Li; he was primary instructor as enrollment grew 150→750 across three offerings.

  [30][2]
• char-rnnprojectreleased2015-05

  Multi-layer RNN/LSTM/GRU character-level language models in Torch, released with the 'Unreasonable Effectiveness' post.

  [11][14]
• ConvNetJS and browser deep learning librariesprojectreleased

  ConvNetJS, RecurrentJS, REINFORCEjs and t-sneJS — neural networks trained entirely in the browser, powering CS231n's live demos.

  [2]
• World of Bits (OpenAI Universe)projectreleased2016-12-05OpenAI

  Browser-task benchmark inside OpenAI's Universe agent platform; he led the sub-project.

  [29]
• 'Software 2.0' essayotherpublished2017-11-11

  The essay that named the paradigm of writing software in neural network weights rather than explicit code.

  [16]
• microgradprojectreleased2020

  A ~100-line scalar-valued autograd engine with a ~50-line neural-net library; the opening artifact of Zero to Hero.

  [10][20]
• minGPTprojectreleased2020-08

  Minimal PyTorch re-implementation of GPT training and inference, built for readability; later semi-archived in favor of nanoGPT.

  [5]
• Neural Networks: Zero to Herootherongoing2022-08

  Video course coding neural networks from scratch — backprop through makemore to GPT and tokenizers — with notebooks and exercises.

  [4][20]
• nanoGPTprojectreleased2022-12

  'The simplest, fastest repository for training/finetuning medium-sized GPTs'; now deprecated in favor of nanochat.

  [6]
• llm.cprojectreleased2024

  LLM pretraining in raw C/CUDA — no heavy framework dependency — reproducing the GPT-2/GPT-3 miniseries.

  [7]
• nanochatprojectreleased2025-10-13

  Minimal end-to-end LLM harness — tokenization through RL and inference — pitched as 'the best ChatGPT that $100 can buy.'

  [8]
• LLM101n: Let's build a Storytellerprojectproposed2024-07Eureka Labs

  The announced first Eureka Labs course — an undergraduate-level class building a storyteller LLM end to end; repo archived while under development.

  [9][42]
• Eureka Labsotherongoing2024-07-16San Francisco

  His company: 'a new kind of school that is AI native,' built on teacher + AI teaching-assistant symbiosis.

  [3][42]
• arxiv-sanityprojectreleased

  Search and ranking tool over tens of thousands of machine learning arXiv papers, maintained since his PhD years.

  [2]

Appearances

• From Vibe Coding to Agentic EngineeringSequoia AI Ascent 20262026-04-30

  with Stephanie Zhan

  A year after coining 'vibe coding': agentic engineering as the serious discipline, LLMs as ghosts, and why understanding can't be outsourced.

  • video: https://www.youtube.com/watch?v=96jN2OCOfLs
  [38]
• Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AINo Priors2026-03-20

  with Sarah Guo

  On code-agent mastery, autonomous research loops, jobs, and agentic education.

  • video: https://www.youtube.com/watch?v=kwSVtQ7dziU
  [37]
• Andrej Karpathy — AGI is still a decade awayDwarkesh Podcast2025-10-17

  with Dwarkesh Patel

  'We're summoning ghosts, not building animals': why RL is terrible, why agents need a decade, and the future of education.

  • video: https://www.youtube.com/watch?v=lXUZvyajciY
  • transcript: https://www.dwarkesh.com/p/andrej-karpathy
  [35][34]
• Software Is Changing (Again)Y Combinator AI Startup School, San Francisco2025-06-19

  Keynote introducing 'Software 3.0': LLMs as a new kind of computer, partial autonomy, and building for agents.

  • video: https://www.youtube.com/watch?v=LCEmiRjPEtQ
  [26]
• Deep Dive into LLMs like ChatGPTAndrej Karpathy (YouTube)2025-02-05

  Three-and-a-half-hour walkthrough of the full LLM training stack: pretraining, tokenization, SFT, RL and RLHF.

  • video: https://www.youtube.com/watch?v=7xTGNNLPyMI
  [24]
• The Road to Autonomous Intelligence (No Priors Ep. 80)No Priors2024-09-05

  with Sarah Guo, Elad Gil

  Self-driving lessons for AGI, Tesla vs. Waymo, Optimus, and Eureka Labs' education plans.

  • video: https://www.youtube.com/watch?v=hM_h0UA7upI
  [36]
• Let's reproduce GPT-2 (124M)Andrej Karpathy (YouTube)2024-06-09

  Four-hour build of GPT-2 124M from scratch — implementation, speedups, hyperparameters, then an overnight training run.

  • video: https://www.youtube.com/watch?v=l8pRSuU81PU
  [23]
• [1hr Talk] Intro to Large Language ModelsAndrej Karpathy (YouTube)2023-11-23

  General-audience talk on what LLMs are, the LLM-OS analogy, and security issues like jailbreaks and prompt injection.

  • video: https://www.youtube.com/watch?v=zjkBMFhNj_g
  [21]
• State of GPTMicrosoft Build 20232023-05

  Breakout talk on the GPT assistant training pipeline and practical mental models for using LLMs.

  • video: https://www.youtube.com/watch?v=bZQun8Y4L2A
  [22]
• #333 – Andrej Karpathy: Tesla AI, Self-Driving, Optimus, Aliens, and AGILex Fridman Podcast2022-10-29

  with Lex Fridman

  Long-form interview spanning neural nets, Transformers, Software 2.0, Tesla's data engine, leaving Tesla, and AGI.

  • video: https://www.youtube.com/watch?v=cdiD-9MMpb0
  [33]
• The spelled-out intro to neural networks and backpropagation: building microgradAndrej Karpathy (YouTube)2022-08-16

  The first Zero to Hero lecture: build a scalar autograd engine, then a neural net, from a blank notebook.

  • video: https://www.youtube.com/watch?v=VMj-3S1tku0
  [20]
• Tesla AI Day 2021Tesla (Palo Alto)2021-08-19

  with Elon Musk, Ashok Elluswamy, Milan Kovac, Ganesh Venkataramanan

  Opened the technical deep dives with the Autopilot vision stack: eight cameras fused into a 3D vector space.

  • video: https://www.youtube.com/watch?v=j0z4FweCy4M
  [32]
• CVPR'20 Workshop on Scalability in Autonomous Driving keynoteCVPR 20202020-06

  Keynote on Tesla's multi-task neural networks and data engine for autonomous driving.

  • video: https://www.youtube.com/watch?v=g2R2T631x7k
  [27]
• Tesla Autonomy DayTesla investor event (livestream)2019-04-22

  with Elon Musk, Pete Bannon, Stuart Bowers

  Presented the neural-network stack and fleet-learning 'data engine' behind Autopilot.

  • video: https://www.youtube.com/watch?v=Ucp0TTmvqOE
  [31]

Relations

• Lex Fridmaninterviewed by

  Lex Fridman Podcast #333, October 2022.

  [33]
• Dwarkesh Patelinterviewed by

  October 2025 episode of the Dwarkesh Podcast — 'AGI is still a decade away.'

  [35][34]

Indexed in

• Dwarkesh Patelinterviewed

  October 2025 episode — 'AGI is still a decade away.'

  cited by the Dwarkesh Patel index

Claims

• factAndrej Karpathy was born on October 23, 1986, in Bratislava, then Czechoslovakia (now Slovakia), and moved with his family to Toronto when he was 15.[48][47]
• factHe completed bachelor's degrees in computer science and physics at the University of Toronto in 2009, a master's degree at the University of British Columbia in 2011 working on physically simulated figures with Michiel van de Panne, and a Stanford computer science PhD in 2016 advised by Fei-Fei Li.[48][2][12]
• factHis 2016 dissertation, 'Connecting Images and Natural Language,' developed models linking visual data and language: a multimodal embedding for image–sentence retrieval, an image captioning model, and a dense-captioning model that localizes and describes salient image regions.[12]
• factWith Fei-Fei Li he published 'Deep Visual-Semantic Alignments for Generating Image Descriptions' at CVPR 2015, aligning image regions with sentence snippets through a multimodal embedding and generating region-level descriptions with a multimodal RNN.[13]
• factDuring his PhD he interned twice at Google working on large-scale feature learning over YouTube videos, and in 2015 interned at DeepMind on the deep reinforcement learning team.[2]
• factTogether with Fei-Fei Li he designed and was primary instructor for CS231n (Convolutional Neural Networks for Visual Recognition), Stanford's first deep learning course; his Stanford page reports enrollment grew from 150 in 2015 to 330 in 2016 and 750 in 2017.[30][2]
• factOn May 21, 2015 he published the blog post 'The Unreasonable Effectiveness of Recurrent Neural Networks,' which trained character-level LSTM language models on Shakespeare, Wikipedia, LaTeX and source code, and released the accompanying char-rnn code on GitHub.[14][11]
• factAs a PhD student he wrote a family of browser-based deep learning libraries in JavaScript — ConvNetJS, RecurrentJS, REINFORCEjs and t-sneJS — and built arxiv-sanity, a tool for searching tens of thousands of machine learning arXiv papers.[2]
• factKarpathy was a founding member of OpenAI, announced on December 11, 2015; OpenAI's launch post lists him among the founding research engineers and scientists under research director Ilya Sutskever.[28][48]
• factAt OpenAI he led the World of Bits sub-project of Universe, released December 5, 2016 — a benchmark of browser-based tasks meant to train agents that operate real websites.[29]
• factIn June 2017 Tesla hired him as Director of AI and Autopilot Vision, reporting directly to Elon Musk, replacing part of the role vacated by departing Autopilot software head Chris Lattner.[39]
• factBy his own account he led the computer vision team of Tesla Autopilot and briefly Tesla Optimus; the team handled in-house data labeling, neural network training, and deployment onto Tesla's custom inference chip.[1]
• factAt Tesla's Autonomy Day on April 22, 2019 he presented the neural-network and fleet-learning approach, including imitation learning from fleet driving data and the 'data engine' loop.[31]
• factAt Tesla AI Day on August 19, 2021 he opened the technical deep dives with the Autopilot vision stack — eight-camera inputs fused into a 3D 'vector space' representation.[32][1]
• factHe announced his departure from Tesla on July 13, 2022, after about five years and a months-long sabbatical, saying he had no concrete plans but wanted to revisit 'technical work in AI, open source and education.'[40]
• factHe rejoined OpenAI in February 2023 and, by his own account, built a new team working on midtraining and synthetic data generation.[1][48]
• factHe left OpenAI again in February 2024 — announced February 13 — stating that nothing had 'happened' and that the departure was not the result of any particular event, issue or drama; he planned to work on personal projects.[41]
• factOn July 16, 2024 he announced Eureka Labs, an 'AI native' school pairing human-designed course materials with AI teaching assistants; its first announced product is the undergraduate-level course LLM101n.[42][3][9]
• factOn May 19, 2026 he announced he had joined Anthropic's pretraining team under lead Nick Joseph; Anthropic said he would build a team focused on using Claude to accelerate pretraining research.[46]
• factOn November 11, 2017 he published the essay 'Software 2.0' on Medium, arguing that neural networks represent 'the beginning of a fundamental shift in how we develop software' — code written as learned weights rather than explicit instructions.[16]
• factOn January 24, 2023 he posted 'The hottest new programming language is English,' followed by a thread of supporting examples of prompts functioning as programs.[18][49]
• factOn February 2, 2025 he coined the term 'vibe coding' in an X post describing a workflow of fully giving in to the vibes, accepting all generated diffs, and forgetting 'that the code even exists,' using Cursor's Composer with a voice dictation tool.[19][51][44]
• fact'Vibe coding' was named Collins Dictionary's Word of the Year for 2025, announced November 6, 2025; coverage credits Karpathy with coining the term in February 2025.[45][49]
• factHis June 2025 Y Combinator AI Startup School keynote 'Software Is Changing (Again)' introduced 'Software 3.0': natural language as the programming interface, with LLMs doing the rest.[26]
• factThe 'Neural Networks: Zero to Hero' video course began on August 16, 2022 with the micrograd lecture — a step-by-step build of a scalar-valued autograd engine — and grew into a series coding neural networks from scratch.[20][4]
• factHe released minGPT in August 2020, a minimal PyTorch re-implementation of OpenAI's GPT training and inference — about 300 lines for the Transformer model itself — aimed at being 'small, clean, interpretable and educational.'[5]
• factHe created nanoGPT in December 2022 as a rewrite of minGPT 'that prioritizes teeth over education'; its roughly 300-line training loop reproduces GPT-2 (124M) on OpenWebText, and the repo is now deprecated in favor of nanochat.[6]
• factIn 2024 he started llm.c, an LLM training implementation in raw C/CUDA with no heavy framework dependency, focused on reproducing the GPT-2 and GPT-3 miniseries alongside a parallel PyTorch reference.[7]
• factHe released nanochat on October 13, 2025 — 'the best ChatGPT that $100 can buy' — a minimal, hackable harness covering tokenization, pretraining, finetuning, evaluation and inference, advertised as able to train a GPT-2-capability model for about $48 on one 8-GPU node.[8]
• factHis general-audience video track includes '[1hr Talk] Intro to Large Language Models' (November 2023), 'Deep Dive into LLMs like ChatGPT' (February 2025, ~3.5 hours covering the full training stack), and 'How I use LLMs' (February 2025).[21][24][25][1]
• fact'Let's reproduce GPT-2 (124M),' published June 9, 2024, is a four-hour build of the GPT-2 124M model from scratch ending in a live training run; he describes it as building the nanoGPT repo about 90% of the way.[23]
• factAt Microsoft Build in May 2023 he gave the 'State of GPT' talk, walking through the GPT assistant training pipeline — pretraining, supervised finetuning, reward modeling and RLHF — and mental models for using the models.[22]
• factMIT Technology Review named him to its 2020 Innovators Under 35 list for 'employing neural networks to allow automated cars to see.'[50]
• factTIME named him to the 2024 TIME100 AI list, citing his trajectory from OpenAI founding member through Tesla and Eureka Labs and calling him one of the internet's most-beloved AI instructors.[43]
• stated beliefHe argues neural networks are not 'just another tool in your machine learning toolbox' but a new programming paradigm: in Software 2.0 the programmer specifies goals and datasets, and training compiles the dataset into the binary — the final network.[16]
• stated beliefHe holds that backpropagation is a 'leaky abstraction' and that neural nets are not off-the-shelf technology: students and practitioners should understand the internals — initialization, gradient flow, dead neurons — or they will likely fail when applying them.[15][17]
• stated beliefHis documented training process is deliberately skeptical: thoroughly inspect the data before touching model code, set up the full train/eval skeleton with dumb baselines, and add complexity only in small, verified steps.[17]
• stated beliefHe frames LLMs as a new kind of computer — with properties of utilities, fabs and operating systems — programmed in English, and analogizes the present moment to computing circa the 1960s.[26][21]
• stated beliefHe pushes back on agent hype: 'the decade of agents,' not the year — early agents like Claude and Codex are impressive and used daily, but lack continual learning and remain 'cognitively lacking,' with roughly a decade of work ahead; AGI, in his estimate, is still about a decade away.[35][34]
• stated beliefHe says reinforcement learning is 'terrible' — but everything else is much worse — and argues LLMs do not learn the way humans do, in part because naive self-training produces model collapse.[35]
• stated beliefHe describes LLMs as 'ghosts' rather than animals — 'we're summoning ghosts, not building animals': stochastic simulations of people with superhuman recall alongside jagged, unreliable deficits.[35][38]
• stated beliefOn products he advocates 'partial autonomy': keep the AI on a leash, make the human verification loop fast with GUIs, and give users an 'autonomy slider' rather than accepting 10,000-line diffs.[26]
• stated beliefHe describes Eureka Labs as 'the culmination of my passion in both AI and education over ~2 decades,' tracing a line from early YouTube tutorials through CS231n to Zero to Hero.[42][3]
• stated beliefHis model for AI education is 'teacher + AI symbiosis': human experts still design course material while AI teaching assistants guide students through it, aiming to let 'anyone learn anything.'[42]
• stated beliefAt AI Ascent 2026 he argued that agentic engineering — not vibe coding — is the serious discipline forming on top of natural-language coding, and that 'you can outsource your thinking but never your understanding.'[38]
• stated beliefOn leaving Tesla he told Lex Fridman he loves the company and Musk, found the decision hard, and would be interested in returning 'at some point' for a potential 'Act 2.'[33]
• stated beliefHe uses self-driving as the template for AI timelines: a Waymo demo impressed him around 2014 yet it took roughly a decade to become a payable product — a 'demo to product' gap he expects to recur with AGI, before any 'globalization' of the capability.[36]
• stated beliefHe is hesitant about brain analogies for neural nets: trained models arrive through a very different optimization process than biological evolution, so he calls them 'a very complicated alien artifact.'[33]
• patternAcross a decade his signature artifact is the minimal, readable reference implementation paired with a step-by-step walkthrough: char-rnn (2015), ConvNetJS and the CS231n assignments, micrograd (2020), minGPT (2020), nanoGPT (2022), llm.c (2024) and nanochat (2025).[11][2][10][5][6][7][8]
• patternHe repeatedly names paradigm shifts in ways that stick: 'Software 2.0' (2017), 'the hottest new programming language is English' (2023), 'vibe coding' (2025), and 'Software 3.0' (2025).[16][18][19][26]
• patternAcross Lex Fridman (2022), No Priors (2024), Dwarkesh (2025) and his YC keynote he consistently dampens full-autonomy hype while crediting real progress — the same 'demo vs. product' lesson he draws from five years in self-driving.[33][36][35][26]
• patternHe builds in public: repos, lecture videos and X announcements ship together, and he announced Eureka Labs early 'so that I can build publicly instead of keeping a secret that isn't.'[42][8][1]
• patternPress coverage of his career moves consistently foregrounds his educator role — TIME calls him 'one of the internet's most-beloved AI instructors' — and his tutorial videos are measured in millions of views.[43][41]
• patternHis pedagogy escalates one stack at a time: scalar autograd (micrograd) → character models (makemore) → GPT-2 scale (nanoGPT, llm.c) → full pipeline with post-training and evaluation (nanochat), mirroring the industry's own scale-up.[10][6][7][8][4]
• speculationNeither his Tesla nor his second OpenAI departure has a publicly stated cause; he called both amicable. Reporting notes the Tesla exit followed a sabbatical and coincided with the San Mateo Autopilot layoffs, and the OpenAI exit came months after the Altman ouster — but no source establishes a causal link.[40][41]
• speculationThe state of Eureka Labs is thin in the public record: the LLM101n repo is archived with a note that the course 'does not yet exist,' and after joining Anthropic in May 2026 he said he plans to resume education work 'in time' — the company's current roadmap is undetermined.[9][46]
• speculationThe precise line between his personal contributions and the Autopilot team's collective work is not publicly delineated; his own framing is that he 'led the computer vision team,' and technical details come largely from his own talks rather than independent engineering records.[1][32][27]
• speculationWhat his 2023–2024 OpenAI midtraining and synthetic-data team shipped is undocumented outside his own bio; OpenAI said only that his responsibilities transitioned to a senior researcher who worked closely with him.[1][41]

Sources

1. Andrej Karpathykarpathy.ai · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

   The subject's own bio/timeline page; self-reported. Also carries a tongue-in-cheek 'Order of the Unicorn' paragraph aimed at LLM scrapers.

2. Andrej Karpathy Academic WebsiteStanford University · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

   His deprecated Stanford-era academic page; records the CS231n enrollment growth, internships, ConvNetJS, and arxiv-sanity.

3. Introducing Eureka LabsEureka Labs · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

   The subject's company site: 'a new kind of school that is AI native.'

4. Neural Networks: Zero to Herokarpathy.ai · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

   Landing page for his from-scratch neural networks course.

5. karpathy/minGPTGitHub · 2020 · subject controlled · webpage · accessed 2026-09-17T00:00:00Z
6. karpathy/nanoGPTGitHub · 2022 · subject controlled · webpage · accessed 2026-09-17T00:00:00Z
7. karpathy/llm.cGitHub · 2024 · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

   LLM training in raw C/CUDA; reproduces the GPT-2 and GPT-3 miniseries without heavy framework dependencies.

8. karpathy/nanochatGitHub · 2025 · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

   'The best ChatGPT that $100 can buy' — a minimal end-to-end LLM harness.

9. karpathy/LLM101nGitHub · 2024 · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

   Eureka Labs' announced first course ('Let's build a Storyteller'); the repo is archived while the course is under development.

10. karpathy/microgradGitHub · 2020 · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

    A tiny scalar-valued autograd engine (~100 lines) with a small neural-net library on top (~50 lines).

11. karpathy/char-rnnGitHub · 2015 · subject controlled · webpage · accessed 2026-09-17T00:00:00Z

    Multi-layer RNN/LSTM/GRU character-level language models in Torch; released alongside the 'Unreasonable Effectiveness' post.

12. Connecting Images and Natural Language (PhD dissertation)Stanford University · 2016 · first person · pdf · accessed 2026-09-17T00:00:00Z
13. Deep Visual-Semantic Alignments for Generating Image DescriptionsCVPR 2015 (CVF Open Access) · 2015 · first person · pdf · accessed 2026-09-17T00:00:00Z
14. The Unreasonable Effectiveness of Recurrent Neural NetworksAndrej Karpathy blog · 2015 · first person · article · accessed 2026-09-17T00:00:00Z
15. Yes you should understand backpropMedium · 2016 · first person · article · accessed 2026-09-17T00:00:00Z
16. Software 2.0Medium · 2017 · first person · article · accessed 2026-09-17T00:00:00Z
17. A Recipe for Training Neural NetworksAndrej Karpathy blog · 2019 · first person · article · accessed 2026-09-17T00:00:00Z
18. X post: 'The hottest new programming language is English'X · 2023 · first person · webpage · accessed 2026-09-17T00:00:00Z
19. X post coining 'vibe coding'X · 2025 · first person · webpage · accessed 2026-09-17T00:00:00Z
20. The spelled-out intro to neural networks and backpropagation: building microgradAndrej Karpathy (YouTube) · 2022 · first person · video · accessed 2026-09-17T00:00:00Z

    First lecture of the Zero to Hero series.

21. [1hr Talk] Intro to Large Language ModelsAndrej Karpathy (YouTube) · 2023 · first person · video · accessed 2026-09-17T00:00:00Z
22. State of GPT | BRK216HFSMicrosoft Build · 2023 · first person · video · accessed 2026-09-17T00:00:00Z

    Microsoft Build 2023 breakout talk: the GPT assistant training pipeline and how to use it.

23. Let's reproduce GPT-2 (124M)Andrej Karpathy (YouTube) · 2024 · first person · video · accessed 2026-09-17T00:00:00Z
24. Deep Dive into LLMs like ChatGPTAndrej Karpathy (YouTube) · 2025 · first person · video · accessed 2026-09-17T00:00:00Z
25. How I use LLMsAndrej Karpathy (YouTube) · 2025 · first person · video · accessed 2026-09-17T00:00:00Z
26. Andrej Karpathy: Software Is Changing (Again)Y Combinator · 2025 · first person · video · accessed 2026-09-17T00:00:00Z

    AI Startup School keynote in San Francisco introducing the 'Software 3.0' framing.

27. [CVPR'20 Workshop on Scalability in Autonomous Driving] Keynote - Andrej KarpathyCVPR 2020 Workshop · 2020 · first person · video · accessed 2026-09-17T00:00:00Z

    On Tesla's multi-task neural networks and the data engine.

28. Introducing OpenAIOpenAI · 2015 · primary record · article · accessed 2026-09-17T00:00:00Z

    OpenAI's launch post; lists Karpathy among the founding members.

29. UniverseOpenAI · 2016 · primary record · article · accessed 2026-09-17T00:00:00Z

    Release of OpenAI's Universe platform; credits Karpathy on the World of Bits browser-task benchmark.

30. CS231n: Convolutional Neural Networks for Visual Recognition (Spring 2015)Stanford University · 2015 · primary record · webpage · accessed 2026-09-17T00:00:00Z

    The first offering of the course he designed and taught.

31. Tesla Autonomy DayTesla · 2019 · primary record · video · accessed 2026-09-17T00:00:00Z

    Investor event livestream; Karpathy presented the neural-network and fleet-learning stack.

32. Tesla AI Day 2021Tesla · 2021 · primary record · video · accessed 2026-09-17T00:00:00Z

    Recruiting-focused technical showcase; Karpathy opened the deep dives with the vision stack.

33. #333 – Andrej Karpathy: Tesla AI, Self-Driving, Optimus, Aliens, and AGILex Fridman Podcast · 2022 · interview · video · accessed 2026-09-17T00:00:00Z
34. Andrej Karpathy — 'We're summoning ghosts, not building animals'Dwarkesh Podcast · 2025 · interview · video · accessed 2026-09-17T00:00:00Z
35. Andrej Karpathy — AGI is still a decade awayDwarkesh Podcast · 2025 · interview · article · accessed 2026-09-17T00:00:00Z

    Episode page with the full transcript of the interview.

36. The Road to Autonomous Intelligence with Andrej Karpathy (No Priors Ep. 80)No Priors · 2024 · interview · video · accessed 2026-09-17T00:00:00Z
37. Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AINo Priors · 2026 · interview · video · accessed 2026-09-17T00:00:00Z
38. Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie ZhanSequoia Capital — Training Data / AI Ascent 2026 · 2026 · interview · video · accessed 2026-09-17T00:00:00Z
39. Tesla hires deep learning expert Andrej Karpathy to lead Autopilot visionTechCrunch · 2017 · reporting · article · accessed 2026-09-17T00:00:00Z
40. Tesla's high-profile Autopilot executive departsReuters · 2022 · reporting · article · accessed 2026-09-17T00:00:00Z
41. Andrej Karpathy is leaving OpenAI again — but he says there was no dramaTechCrunch · 2024 · reporting · article · accessed 2026-09-17T00:00:00Z
42. Former OpenAI, Tesla engineer Andrej Karpathy starts AI education platformReuters · 2024 · reporting · article · accessed 2026-09-17T00:00:00Z
43. TIME100 AI 2024: Andrej KarpathyTIME · 2024 · reporting · article · accessed 2026-09-17T00:00:00Z
44. What is vibe coding, exactly?MIT Technology Review · 2025 · reporting · article · accessed 2026-09-17T00:00:00Z
45. 'Vibe coding' named word of the year by Collins DictionaryBBC News · 2025 · reporting · article · accessed 2026-09-17T00:00:00Z
46. OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training teamTechCrunch · 2026 · reporting · article · accessed 2026-09-17T00:00:00Z
47. Andrej Karpathy (Q56037405)Wikidata · reference · dataset · accessed 2026-09-17T00:00:00Z
48. Andrej KarpathyWikipedia · reference · article · accessed 2026-09-17T00:00:00Z

    Used for discovery and date cross-checks; not the sole authority for any claim.

49. Vibe codingWikipedia · reference · article · accessed 2026-09-17T00:00:00Z

    Reference entry tracking the term's coinage and adoption.

50. Andrej Karpathy | Innovators Under 35MIT Technology Review · 2020 · reference · webpage · accessed 2026-09-17T00:00:00Z

    2020 Innovators Under 35 honoree profile.

51. Archived snapshot of the 'vibe coding' X postarchive.today · 2025 · archive · webpage · accessed 2026-09-17T00:00:00Z

    Capture of the original vibe-coding post taken the day after it was published.

Coverage and method

Scope as of

2026-09-17T03:00:00Z

Coverage supplied by the publisher

• biography
• work
• philosophy
• beliefs
• projects
• media

Method

public-person-index-v1

Tool

soulscrape

Contributors

• Soulscrape research workflow

Assembled

2026-09-17T03:00:00Z

Human review

No review status or review date is supplied. An assembly timestamp does not establish human review.

Open questions

See the supplied open questions

Open questions

• The precise boundary between his personal engineering contributions and the Autopilot team's collective work is not publicly delineated; most technical detail comes from his own presentations.
• No public source states why he left Tesla in July 2022 or OpenAI in February 2024; both exits were framed as amicable, and reporting notes only adjacent events (the San Mateo layoffs, the post-ouster reorganization).
• What his 2023–2024 OpenAI midtraining/synthetic-data team shipped is undocumented; the only account is his own bio.
• Eureka Labs' status is unresolved: the LLM101n repo is archived pending development, and after joining Anthropic in May 2026 he said education work would resume 'in time' without specifics.
• Sources vary between 'founding member' (OpenAI's own launch post) and 'co-founder' (later press) when describing his OpenAI role.
• His birth date (October 23, 1986) is carried by reference sources — Wikidata and Wikipedia — without a primary public record in this catalog.
• Video view counts, repository star counts, and enrollment figures cited in coverage are point-in-time self-reports or platform counters, not audited figures.
• The 2026 Anthropic role is described at team level (pretraining, under Nick Joseph); its charter beyond 'using Claude to accelerate pretraining research' is not yet public.