--- generator: contextjamming-llm-context title: "Andrej Karpathy" description: "The teacher who made rebuilding GPT by hand a public utility went back to a frontier lab to teach the machine to do the rebuilding." url: https://www.contextjamming.com/founder-files/karpathy markdown_url: https://www.contextjamming.com/founder-files/karpathy.md section: "Founder Files" content_type: "FounderFile" author: "Bret Kerr" publisher: "Context Jamming (ACRA Insight LLC)" language: en-US word_count: 2904 image: https://www.contextjamming.com/photos/ff-karpathy.jpg --- > **Andrej Karpathy** — Founder Files, Context Jamming. This is a plain-Markdown copy of https://www.contextjamming.com/founder-files/karpathy, prepared for LLM context windows, agents and retrieval. > > Cite as: Bret Kerr, "Andrej Karpathy," *Context Jamming*, https://www.contextjamming.com/founder-files/karpathy FounderFiles·N°069·Vision · Pedagogy · Pretraining 1986 — ![Andrej Karpathy at a lectern at Stanford, 2019](https://www.contextjamming.com/photos/ff-karpathy.jpg) *Fig. · The lecturer[Photo: Gladwin Analytics / CC BY 3.0](https://commons.wikimedia.org/wiki/File:Andrej_Karpathy,_OpenAI.png)* Subject·Andrej Karpathy·Founding member, OpenAI · Former Director of AI, Tesla · Pretraining research, Anthropic # Andrej Karpathy Karpathy taught a generation to understand neural networks by rebuilding them by hand — then went back to a frontier lab to teach the machine to do the rebuilding. *He built Stanford’s first deep-learning course, ran Tesla’s camera-only bet, and rebuilt GPT from scratch on camera for anyone with a laptop. In February 2025 he coined vibe coding in a post about throwaway weekend projects; by November it was Collins’s word of the year. In October 2025 he told a podcaster that agent code was slop and autocomplete was his sweet spot. Seven months later he joined Anthropic to use Claude to speed up pretraining research.* TRAINED Toronto · UBC · Stanford (PhD, 2016) AT Anthropic · Pretraining FILE N°069 § 01 · The Student ## Hinton’s class, then Fei-Fei Li’s lab He was born in Bratislava, then Czechoslovakia, in 1986 \[single-source\]. At the University of Toronto he took a double major in computer science and physics, with a minor in math. He got into deep learning the way a lot of people in this file did: through Geoffrey Hinton’s class and reading groups, years before the field had a name anyone outside it used. See the [Hinton FounderFile](https://www.contextjamming.com/founder-files/geoffrey-hinton). A master’s at the University of British Columbia followed, with Michiel van de Panne, on controllers for physically simulated figures. Then Stanford, 2011, a PhD advised by Fei-Fei Li — the woman who had just built ImageNet and was about to watch it break computer vision open. See the [Fei-Fei Li FounderFile](https://www.contextjamming.com/founder-files/fei-fei-li). His thesis was *Connecting Images and Natural Language*. The flagship result paired a convolutional network with a recurrent one and taught them to caption photographs. Internships along the way: Google Brain in 2011, Google Research in 2013, DeepMind in 2015. In May 2015 he published a blog post called *The Unreasonable Effectiveness of Recurrent Neural Networks*. It opened: *There’s something magical about Recurrent Neural Networks.* Then it showed the magic was a few hundred lines of code. He released char-rnn the same day. A character-level model wrote fake Shakespeare, fake Linux source, fake algebraic geometry in LaTeX. The post did something papers rarely do. It let the reader run the experiment that night. He did not explain the network. He handed it over and said: train it yourself. That would become the method for the rest of his career. § 02 · The Course ## CS231n and the open notebook In winter 2015, still a graduate student, he designed and became primary instructor for CS231n, *Convolutional Neural Networks for Visual Recognition* — by his own account the first deep-learning course at Stanford. He designed it and was its primary instructor, with Li as faculty of record and, by 2016, Justin Johnson co-teaching. Enrollment went from 150 in 2015 to 750 in 2017 \[single-source\]. That is the number Stanford saw. It is not the number that matters. The course notes went up on GitHub Pages. The assignments shipped as Python notebooks you could run on a laptop. The 2016 lectures went up on YouTube. A student in Lagos or Pune got the same material as a student in Gates Hall, minus the office hours. The notes are not elegant mathematics. They are mechanics. How backprop flows through a single multiply gate. Why you check your gradient numerically before trusting it. What the loss curve looks like when the learning rate is too high. The pedagogy is the engineer’s — show the moving part, then let the student break it. The people who took it online, not the people in the room, are the ones who went on to staff the labs. That is the pattern that never shows up on a CV. A teaching footprint is counted in seats; it is felt in commits. CS231n was the first time a frontier subject was taught in public at the speed it was being invented. It set the template. Every course he made afterward was a variation on the open notebook. § 03 · The Data Engine ## Software 2.0, at fleet scale December 2015, he was a founding member of OpenAI — research scientist, the period that put him in the same rooms as Ilya Sutskever. See the [Sutskever FounderFile](https://www.contextjamming.com/founder-files/sutskever). He stayed eighteen months. June 2017, Tesla hired him as Director of AI, reporting to Elon Musk, in the same week Chris Lattner left Autopilot. He was thirty. In November he published *Software 2.0* on Medium. The claim: neural networks were not one more classifier in the toolbox; they were a new way of writing software. Humans would stop writing the logic. They would curate the data, and an optimizer would write the weights. At Tesla he ran the argument at the scale of a car fleet. The *data engine*: deploy a network, find where it fails in the wild, pull those clips back, label them, retrain, redeploy. Code was what you threw away. Data was what you kept. The camera-only bet followed from it. In May 2021 Tesla stopped fitting forward radar to North American Model 3 and Model Y. At CVPR that June, Karpathy defended vision alone, with a supercomputer and a dataset behind it. Waymo kept its lidar. The argument is not settled. He presented again at AI Day in August 2021. In March 2022 he took a four-month sabbatical. On July 13 he announced he was leaving. The Software 2.0 essay said engineers would become data curators. Five years at Tesla tested it on the hardest deployment target there is. Then he left to go back to writing code by hand — on camera. > “If I can’t build it, I don’t understand it.” > Karpathy, Dwarkesh Podcast, Oct 2025 (citing Feynman) § 04 · Zero to Hero ## Rebuilding GPT in an afternoon August 16, 2022, a month after Tesla: a two-and-a-half-hour video called *The spelled-out intro to neural networks and backpropagation: building micrograd*. A scalar autograd engine in about a hundred lines of Python. 4.39 million views as of November 2026. It was the first lecture of Neural Networks: Zero to Hero. The series built bigram models, then MLPs, then a tokenizer. January 17, 2023: *Let’s build GPT: from scratch, in code, spelled out.* Two hours. A transformer typed out line by line from an empty file. 8.18 million views. The repositories carried the lectures into production. minGPT in 2020. nanoGPT in December 2022, now 64,397 GitHub stars — a few hundred lines that train a small GPT with no framework in the way. llm.c in 2024, GPT-2 in raw C and CUDA. nanochat in October 2025, the full pipeline from tokenizer to chat UI, which by its own README trains a GPT-2-grade model for about $48 against roughly $43,000 in 2019 \[single-source\]. None of it is novel research. All of it is reconstruction. The claim embedded in every repo is that you do not understand a thing until you can rebuild it with nothing borrowed. He said so, plainly, on the Dwarkesh podcast in October 2025, after borrowing a line from Feynman: *Build the code, arrange it, get it to work. It’s the only way to go.* The meme puts him in the tier of people AI owes a thank-you. The more exact debt is to the method. He made reconstruction a public utility, and a generation of engineers learned the transformer by typing it out behind him. § 05 · The Vibes ## A shower thought becomes a word of the year Sunday, February 2, 2025, 3:17 p.m. Pacific. A post on X: *There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.* Read the rest of it. He talks to Cursor through a voice app. He hits *Accept All* and no longer reads the diffs. He pastes errors back in without comment. When a bug will not die he asks for random changes until it goes away. The verdict is modest: *not too bad for throwaway weekend projects.* The post has 7.38 million views. The careful work — *Let’s build GPT*, two hours of typing a transformer from an empty file — has 8.18 million. The throwaway nearly caught the lecture. Then it lapped him. Collins named *vibe coding* its word of the year on November 6, 2025, beating *clanker* and *aura farming*. In his 2025 year-in-review he admitted he coined it *totally oblivious to how far it would go*. In June 2025 at YC’s AI Startup School he gave the theory a version number. *Software 3.0*: prompts are programs, and English is the language. 2.62 million views. The irony is structural. The educator whose whole method is *read every line* named the practice of reading none of them. He meant it as a joke about weekend apps. The market took it as permission. Karpathy’s most durable contribution is rigor-by-reconstruction. His most famous is two words he tossed off on a Sunday. The field remembers the second and runs on the first. > “There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” > Karpathy, post on X, Feb 2, 2025 § 06 · The Decade, Compressed ## From “slop” to the pretraining team October 17, 2025. On Dwarkesh Patel’s podcast, he cooled the room. This was *the decade of agents*, not the year. Agents *just don’t work*. They are *cognitively lacking*. The industry was *trying to pretend like this is amazing, and it’s not. It’s slop.* For his own code: *For now, autocomplete is my sweet spot.* Same interview: *we’re not building animals. We’re building ghosts* or spirits. He had written it two weeks earlier on his blog — *Pretraining is our crappy evolution.* Then the timeline moved. March 2026 he released autoresearch: an agent gets a small nanochat training setup and a five-minute budget per run, edits the training script, keeps what helps, discards what doesn’t. Two days unattended, about 700 changes, around twenty that stacked, and the time-to-GPT-2 leaderboard dropped 11 percent \[single-source\]. The human edits a Markdown file. The agent edits the Python. 99,104 stars — his most-starred repository. May 19, 2026: *Personal update: I’ve joined Anthropic.* He would *get back to R&D*. He works under pretraining lead Nick Joseph, and Anthropic said he would start a team using Claude to accelerate pretraining research. On education: *I remain deeply passionate about education and plan to resume my work on it in time.* See the [Kaplan FounderFile](https://www.contextjamming.com/founder-files/kaplan) for the lab he joined. The announcement drew 28.1 million views, nearly four times the vibe-coding post. He did not recant the decade. He compressed it. Seven months after calling agent code slop, he signed on to point agents at the most expensive code in the industry — the pretraining stack. § 07 · The Wall of Slop ## A course that does not exist, and a filter that finds him Friday night, October 30, 2026, 11 p.m. Karpathy opened YouTube and found *an enshittified wall of slop*. Next morning he gave Claude Opus 5.5 the job of building an alternative out of the YouTube videos recommended in Hacker News comments, 2007 to 2026. Two hours later he had HackerTube — 6,454 videos across 150 topics \[single-source\]. In the same post: *I don’t actually do anything resembling traditional engineering anymore.* No reading code. No writing it. No terminal. He made one API key, booted a server and bought the domain. Claude did the rest. Search HackerTube for his name. *Building micrograd* — the hundred lines of hand-written autograd — has seventeen Hacker News recommenders and a score of 21.0. *Let’s build GPT* has fourteen. The Y Combinator video titled *Vibe Coding Is The Future* has none. The machine he no longer needs to code for surfaces, as the best of what he made, the video of him coding by hand. One repository sits beside it. LLM101n, *Let’s build a Storyteller*, was announced with Eureka Labs in July 2024 and archived that August. Its README says, in a bolded note, that the course *does not yet exist*. It has 37,476 stars — more than llm.c. Thirty-seven thousand people bookmarked the promise of being taught. The tier list says AI owes him a thank you. The record says something stranger: the people who learned from him are still waiting for the next lesson, and the teacher is teaching the machine. The course does not exist. The stars are real. > “I don’t actually do anything resembling traditional engineering anymore.” > Karpathy, “HackerTube”, Oct 31, 2026 Timeline - 2009University of Toronto — BSc, computer science and physics; first deep learning via Hinton’s class - 2011UBC — MSc with Michiel van de Panne; begins Stanford PhD under Fei-Fei Li - 2015CS231n first taught; “The Unreasonable Effectiveness of RNNs” and char-rnn (May 21); founding member of OpenAI (Dec) - 2016Stanford PhD — *Connecting Images and Natural Language* - 2017Tesla — Director of AI, reporting to Musk (June); “Software 2.0” (Nov 11) - 2021Tesla Vision drops radar on North American Model 3/Y (May); CVPR keynote; Tesla AI Day (Aug) - 2022Leaves Tesla (July 13); micrograd lecture opens Zero to Hero (Aug 16); nanoGPT (Dec) - 2023“Let’s build GPT” (Jan 17); rejoins OpenAI (Feb 9); “Intro to Large Language Models” (Nov) - 2024Leaves OpenAI again (Feb); llm.c (Apr); Eureka Labs and LLM101n announced (July 16); LLM101n archived (Aug) - 2025Coins “vibe coding” on X (Feb 2); Software 3.0 keynote at YC AI Startup School (June) - Oct 2025nanochat (Oct 13); Dwarkesh podcast, “the decade of agents” (Oct 17); Collins word of the year (Nov 6) - Mar 2026autoresearch — agents tune nanochat unattended - May 2026Joins Anthropic’s pretraining team under Nick Joseph (May 19) - Oct 2026HackerTube — a YouTube alternative built by Claude in about two hours (Oct 31) The Index 150 → 750 CS231n enrollment, 2015 to 2017 \[ss\] 8.18M Views, “Let’s build GPT” (2023), as of Nov 2, 2026 64,397 GitHub stars, nanoGPT, as of Nov 2, 2026 7.38M Views of the Feb 2025 “vibe coding” post, as of Nov 2, 2026 $48 nanochat cost to train GPT-2 grade, vs ~$43K in 2019 \[ss\] 99,104 Stars, autoresearch (Mar 2026) — his most-starred repo, Nov 2, 2026 28.1M Views, “I’ve joined Anthropic,” May 2026, as of Nov 2, 2026 37,476 Stars on LLM101n, a course that “does not yet exist,” Nov 2, 2026 Reading list / Key works - [2015The Unreasonable Effectiveness of Recurrent Neural Networkskarpathy.github.io →](https://karpathy.github.io/2015/05/21/rnn-effectiveness/) - [2015CS231n: Convolutional Neural Networks for Visual Recognition (course notes)Stanford →](https://cs231n.github.io/) - [2017Software 2.0Medium →](https://karpathy.medium.com/software-2-0-a64152b37c35) - [2022Neural Networks: Zero to Hero (lectures + code)YouTube / GitHub →](https://github.com/karpathy/nn-zero-to-hero) - [2022nanoGPTGitHub →](https://github.com/karpathy/nanoGPT) - [2025The original “vibe coding” postX →](https://x.com/karpathy/status/1886192184808149383) - [2025AGI is still a decade away — interview with Dwarkesh PatelDwarkesh Podcast →](https://www.dwarkesh.com/p/andrej-karpathy) - [2026HackerTubekarpathy.bearblog.dev →](https://karpathy.bearblog.dev/hacker-tube/) - [2026Featured in: What AI Owes Them — the tier list, annotatedContext Jamming →](https://www.contextjamming.com/Binkgmeme) Dossier **Education.** University of Toronto (BSc, computer science and physics, minor in math, 2009). University of British Columbia (MSc, 2011; with Michiel van de Panne). Stanford University (PhD, 2016; thesis: *Connecting Images and Natural Language*; advisor: Fei-Fei Li — Stanford’s library record also lists Christopher Manning \[single-source\]). **Affiliations.** Internships at Google Brain (2011), Google Research (2013), DeepMind (2015). OpenAI (founding member, research scientist, 2015–2017; returned 2023–2024). Tesla (Director of AI, Autopilot vision, 2017–2022). Eureka Labs (founder, 2024; education work paused, per his own post, “to resume in time”). Anthropic (pretraining research, since May 2026). **Mentors.** [Geoffrey Hinton](https://www.contextjamming.com/founder-files/geoffrey-hinton), whose Toronto class and reading groups were his entry point. [Fei-Fei Li](https://www.contextjamming.com/founder-files/fei-fei-li), his Stanford advisor and CS231n co-instructor of record. Michiel van de Panne at UBC. **Collaborators / peers worth naming.** Justin Johnson (CS231n co-instructor). [Ilya Sutskever](https://www.contextjamming.com/founder-files/sutskever) (OpenAI founding cohort). Elon Musk (his manager at Tesla). Nick Joseph (his team lead on Anthropic pretraining). **Open source.** char-rnn (2015), convnetjs, micrograd (2020), minGPT (2020), nanoGPT (2022), llama2.c (2023), minbpe and llm.c (2024), nanochat (2025), autoresearch (2026). Combined stars across those ten repositories: more than 350,000 as of November 2, 2026. Further reading [““Animals vs Ghosts.””](https://karpathy.bearblog.dev/animals-vs-ghosts/) karpathy.bearblog.dev · Oct 1, 2025 · essay Career Shape tree-shaped — depth across many branches of one field ### Tree-Canopy Autodidact Self-credentialed; grows depth across branches of a single field and makes the opaque legible to everyone else. Role-Model Reference Class - Richard Feynman (by his own invocation) - Fei-Fei Li - Geoffrey Hinton Founder Context · JSON A small reasoning persona distilled from this file. Inject it into a chat or deep-research context to assess a business problem the way Karpathy would. Reason like Andrej Karpathy: before trusting any system, rebuild the smallest version of it from scratch and get it to work; prefer a few hundred legible lines over a framework. Treat neural networks as software written by optimization, so put effort into the data and the failure cases rather than hand-written logic. Be blunt about what current agents can and cannot do, measure everything against a clear metric, and hand to agents the loops that can be verified automatically. Publish the result so others can learn by rebuilding it too. ``` { "$schema": "https://www.contextjamming.com/schemas/founder-context-v1.json", "file": "N°069", "persona": "Andrej Karpathy", "archetype": "tree-canopy", "shape": "Y", "one_line": "An educator-engineer who understands a system only by rebuilding it from an empty file, publishes the rebuild for everyone, and now points agents at the rebuilding itself.", "cognitive_basis": { "credentialPath": "doctoral", "abstractionDirection": "bottom-up", "exitHorizon": "non-commercial", "moatInstinct": "optimization-and-systems", "capitalPosture": "not-established" }, "operating_questions": [ "Can I rebuild this from scratch in a few hundred lines, with nothing borrowed?", "What is the smallest working version that still contains the whole idea?", "Where does the system fail in the wild, and can that failure become training data?", "Is this agent out … ``` Share FounderFiles N°069 · Andrej Karpathy Filed by Bret Kerr · ACRA Insight LLC · Franklin, MA contextjamming.com · @bretkerr [← back to Context Jamming](https://www.contextjamming.com/)