← Home
← Back to Studies
STUDY 002

Andrej Karpathy

He kept quitting frontier labs to teach. In 2026 he went back.

Andrej Karpathy

I. The Frame

The standard story about Andrej Karpathy is that he is the world's deep-learning teacher who happens to keep getting hired by frontier labs. The standard story is backwards. This study said so in an earlier version — and then the news overtook that version too, in a way that turns out to be the most revealing thing about him. So what follows is a correction of a correction, and the second mistake is the more instructive one.

The honest story starts here: he is a builder who keeps quitting frontier-AI labs, and the output keeps looking like teaching. Four departures across three institutions — Stanford, Tesla, and OpenAI twice — and at each one the most-cited public artifact of his tenure was produced near the transition, not in the middle. Stanford gave the world CS231n. Tesla gave the world Software 2.0. The 2022 sabbatical gave the world Neural Networks: Zero to Hero. OpenAI v2 gave the world Intro to LLMs and the LLM-OS framing. And in 2024 he stopped quitting institutions and built one of his own: Eureka Labs, a school where the compounding public work could be the day-job.

The first version of this thesis read that as a monotonic march. Each transition, it argued, reduces the share of his output that is institutional-private and increases the share that is public; Eureka Labs was the terminal value, the position where the answer is 100% and leaving is no longer the way to ship. That reading was clean, and it was wrong. On May 19, 2026, Karpathy left the terminal position. He deprioritized Eureka and joined Anthropic to build a team on its pre-training group — back inside a frontier lab, back into work that will mostly not be public.

A thesis is only worth writing if the world can falsify it. This one got falsified in the specific way its own final section predicted, and the falsification forces a better reading. The invariant was never teaching. Teaching was the channel that comprehension flowed through whenever Karpathy stood outside an institution. The invariant underneath it is older and simpler: he understands things by rebuilding them from scratch, and he goes wherever the most interesting thing to rebuild happens to be. For fifteen years that object sat outside the lab, so the output looked like teaching. In 2026 the most interesting thing to rebuild — a training run that improves itself — turned out to need frontier compute, so he went back in.

If you read only one sentence from this document, read this one: Karpathy's career is not a march toward teaching; it is an oscillation between resource-rich and output-rich environments, and he moves each time the object he most wants to comprehend crosses to the other side.

The rest of this thesis is the unpacking of that sentence, and the honest accounting of what the first version got wrong.

II. The Career as a Single Pattern

He was born in Bratislava, October 1986, raised in Toronto from age 15. The technical formation was Toronto undergrad (CS + Physics) → UBC MSc with Michiel van de Panne (physically simulated figures) → Stanford PhD with Fei-Fei Li (2011 — 2015/16), dissertation on Connecting Images and Natural Language. Three internships during the PhD — Google Brain, Google Research, DeepMind — gave him intimate contact with each of the three institutions that would define the next decade of AI.

In 2015, while still finishing the dissertation, he co-launched CS231n at Stanford with Fei-Fei Li and Justin Johnson. Enrollment grew from ~150 to ~750 over three years. The course notes at cs231n.github.io and the Karpathy-led 2016 video lectures became the canonical introduction to deep learning for vision; both are still cited a decade later. CS231n was open from day one. Stanford got the lectures; the internet got everything.

He joined OpenAI as a founding research scientist in December 2015 — the cohort of Sam Altman, Greg Brockman, Ilya Sutskever, John Schulman, Wojciech Zaremba, with Elon Musk as early board chair. The OpenAI v1 tenure (Dec 2015 — Jun 2017) is the least-archetypal stretch of his career. Its signature public artifacts are pedagogical blog posts — The Unreasonable Effectiveness of Recurrent Neural Networks (May 2015, written just before he joined) and Pong from Pixels (May 2016). Frontier research was happening; he was producing frontier teaching.

In June 2017 he joined Tesla as Director of AI / Autopilot Vision. Five years. He built the Autopilot vision team, the data engine, the deployment pipeline onto Tesla's custom inference silicon, and the HydraNet multi-task architecture. Six months in, he wrote the Software 2.0 essay — the single most-cited piece of his career. The essay reframes neural networks from a clever new ML technique into a new programming paradigm: code is now weights, datasets are now source files, training is now compilation. The essay reads like a Tesla engineering manifesto generalized to the field, because that is what it is. Note the timing: most-cited artifact, six months into a five-year tenure. He was already producing his public output in advance of his exit.

March 2022: sabbatical. July 13, 2022: announcement of departure. The public statement telegraphs Eureka Labs two years before it exists: "long-term passions around technical work in AI, open source and education." For seven months he had no employer. The output: micrograd. makemore. The "Let's build GPT" lecture. Zero to Hero, the from-scratch ML curriculum that re-established him as the field's primary teacher. This is the period to study if you want to understand his operating posture. Without an employer or institutional pressure, what does he do? He rebuilds the field's systems from scratch, and he publishes the rebuild.

February 2023 — February 2024: OpenAI v2. ~12 months on midtraining and synthetic data generation. The major public artifacts during this tenure are external, not internal: State of GPT at Microsoft Build, the Intro to LLMs 1-hour talk that becomes the field's standard explainer, the GPT Tokenizer lecture as the last Zero to Hero installment. The pattern repeats — public artifacts ship near the exit. Departure announced "nothing happened, no drama, working on personal projects."

July 16, 2024: Eureka Labs. "A new kind of school that is AI native." Human teachers design curriculum; AI Teaching Assistants scale delivery. First product: LLM101n — Let's build a storyteller, a 17-chapter syllabus from Python basics to a working model. The repository was archived in August 2024 with a README stating the course is "currently being developed," and it never shipped a public cohort.

The 2025 period is the Software 3.0 era. The June 2025 YC AI Startup School keynote — "Software Is Changing (Again)" — extends the 2017 essay into a third stack-layer: Software 1.0 = code, 2.0 = weights, 3.0 = prompts. English is the new programming language; LLMs are operating systems; the "decade of agents" (2025-2035) is the period when this OS finds its applications. The October 2025 Dwarkesh interview is the most authoritative statement of his AGI timelines (~decade away, continuous not discontinuous, RL is terrible).

Then, quietly, the pattern turned. In October 2025 he released nanochat — the from-scratch instinct scaled to a complete system, a full-stack ChatGPT clone (pretraining through inference) trainable for about $100. Through early 2026 he tuned it in public: FP8 support in February, and by early March 2026 a training run that reproduced a GPT-2-class model in two hours on a single 8×H100 node, down from three a month earlier, mostly on the strength of a dataset swap to NVIDIA's ClimbMix. Days later he shipped autoresearch: an agent that runs ML experiments on the nanochat codebase by itself. Left running two days, it ran ~700 experiments and found ~20 stackable improvements that transferred to a larger model and cut the leaderboard's "time to GPT-2" from 2.02 to 1.80 hours — about 11%. By mid-March the leaderboard read 1.65 hours, against the 168 hours OpenAI needed for the original GPT-2 in 2019.

That project is the hinge of the whole career. The from-scratch rebuild had always been a teaching artifact. Autoresearch turned it into a research engine — the minimal, readable codebase became the substrate on which an AI does the science. In late April 2026 the Sequoia Ascent fireside ("From Vibe Coding to Agentic Engineering") introduced verifiability as the new constraint and retired "vibe coding" (his February 2025 coinage) in favor of "agentic engineering." And on May 19, 2026 he announced he had joined Anthropic, building a team on the pre-training group under Nick Joseph, with a mandate an Anthropic spokesperson described as "using Claude to accelerate pre-training research" — autoresearch at frontier scale. His own words: "I remain deeply passionate about education and plan to resume my work on it in time."

Recognition along the way: Innovators Under 35 (MIT Tech Review, 2020); TIME 100 Most Influential People in AI (2024). The career is not over; he is 39.

III. The Load-Bearing Claim — Comprehension by Rebuilding

The first version of this thesis named the central insight "pedagogy as research." That was close, and the May 2026 reversal shows exactly how it was close. The teaching was real, but it was downstream of something more fundamental. Karpathy's actual operating primitive is this: he understands a system only by rebuilding a minimal version of it from scratch, and the rebuild is the artifact. When he is outside an institution, the rebuild gets published and looks like teaching. When the rebuild needs frontier compute, it pulls him back inside.

Start with the evidence that survives the reversal, because most of it does. His most-cited work is consistently produced near transitions, not deep into tenures. Software 2.0 six months into Tesla. Zero to Hero during the sabbatical. Intro to LLMs mid-OpenAI-v2. The institutions' role in his output was rarely as production environment; it was as vantage. And the from-scratch instinct is a decade-stable constant: micrograd, makemore, nanoGPT, llama2.c, llm.c, and now nanochat. Each is a from-scratch implementation of something the field already has at frontier. The redundancy is the point. He is not competing with PyTorch or HuggingFace; he is building a comprehension path to the layer below.

The philosophical key is Yes You Should Understand Backprop (Dec 2016): "frameworks that hide backprop create a false sense of reliability." If frameworks lie, the only defense is comprehension of the layer below. The injunction is not "for students" — it is for everyone, including himself. The from-scratch artifact is how he keeps his own understanding honest. The students benefit because the artifact is published.

Now the part the first version missed. If the primitive were "teaching," the 2024 founding of Eureka Labs would have been the end of the story — the structural position where teaching is the whole job. It wasn't the end. Between late 2025 and early 2026, the rebuild instinct produced nanochat and then autoresearch, and autoresearch is the tell. It is not a course. It is not aimed at students. It is a from-scratch training loop handed to an agent so the agent can improve it — the rebuild becoming a research engine that does the science itself. That is the most direct expression of comprehension-by-rebuilding he has ever produced, and it points away from the classroom, not toward it. Anthropic hired him to run precisely that idea at scale.

So the reframe is this. Pedagogy-as-research was a special case, true only when the interesting rebuild happened to be cheap enough to do outside a lab. The general case is comprehension-by-rebuilding, and it is indifferent to whether the output ends up public. That indifference is what the first version could not see, because until 2026 the output had always been public. The moment the most interesting rebuild required a frontier training budget, the invariant did exactly what the invariant does — it followed the object — and the public-output share, which the first version treated as a monotonic ratchet, dropped.

This is what makes him difficult to imitate, and the difficulty is now clearer than before. The Karpathy posture is not "teach AI" and it is not "stay independent." It is "organize your working life around the thing you most need to understand, and be willing to change environments completely when that thing moves." Most professionals cannot do this. They optimize the environment (title, resources, independence) and let the object of attention drift. He does the reverse. He spent fifteen years bending existing paths, then built his own, then left it — each time following the same underlying gradient.

Where the earlier version said the teaching is the research, read: the rebuild is the research, and teaching is only what the rebuild becomes when it is cheap and public.

IV. The Mental-Model Stack

Karpathy's intellectual contribution is not the discovery of new ML primitives — others contributed more there. It is the stable framework that lets others orient. A handful of recurring models compose his entire public output. They are unusually compact; a handful of them cover almost everything he has said.

Software 1.0 → 2.0 → 3.0. Three layers, additive not replacing. 1.0 = handwritten code → compiled binary. 2.0 = dataset + neural-net architecture → trained weights. 3.0 = English prompt → LLM execution. Today's products are mosaics of all three. The 2017 essay correctly anticipated the entire MLOps category before it existed; the 2025 update predicts the next one — prompt management, eval frameworks, agent infrastructure, llms.txt. If you want to know what dev-tools to build, find the layer that doesn't yet have its tooling.

LLM as the new operating system. "LLMs should not be thought of as a chatbot, but as the kernel process of an emergent operating system." LLM = CPU/kernel; context window = RAM; tools = peripherals; prompts = programs. The productive prediction: most consumer "AI products" are shells running on a kernel they don't own. Startup strategy collapses to three options — build the kernel, the shell, or the peripherals. No fourth option scales.

The cognitive core. (2025.) The provocative inversion of scale: the valuable object may be a small model that has deliberately given up memorized encyclopedic knowledge in exchange for reasoning, keeping the kernel of intelligence and looking facts up rather than storing them. If right, the future of "LLM personal computing" is not one giant model but a small, always-on core with tools and memory around it — the operating-system metaphor taken to its conclusion.

Agents as "people spirits." "Stochastic simulations of people, with a kind of emergent psychology — simultaneously superhuman in some ways, but also fallible in many others." The corollary: "LLMs are a bit like a coworker with anterograde amnesia." The framing refuses both dominant narratives — "stochastic parrot" denies the emergent psychology; "it's a person" denies the fallibility. The middle is more predictive than either end.

Ghosts, not animals. (2025.) The sharpest version of the same idea, and one of his most important recent framings. Animals arrive with billions of years of evolution compiled into their hardware; we are not building animals. We are summoning ghosts — entities distilled from imitating a snapshot of the internet, with no innate drives and a hazy recollection of their training rather than lived experience. Pretraining, in his phrase, is "crappy evolution." The frame explains both the eerie competence and the missing common sense, and it is why he expects progress to look like an engineering slog rather than a spark of life.

The autonomy slider. "Demo is works.any(), product is works.all()." Because LLMs are jaggedly fallible, full autonomy is not shippable today, but partial autonomy with an escalating human-verification loop is. Cursor, Perplexity, Tesla Autopilot — the slider is the product.

The march of nines. (From his Tesla years, sharpened in 2025.) Each additional nine of reliability — 90% to 99% to 99.9% — costs roughly the same large amount of work as the one before it. Reliability is not a finish line you approach but a ladder where every rung is a fresh climb. This is the engine under "the decade of agents": a demo that works nine times in ten is one nine in, and the useful product lives at four or five nines, which is years of nines away. It is also the most direct import from his self-driving decade, where he watched exactly this curve refuse to bend.

Verifiability is the new constraint. (2026 Sequoia.) "Traditional software automates what you can specify; AI automates what you can verify." The products that work are the ones where the human can cheaply verify the output: code with tests, design with rendering, math with proofs. If you cannot describe how a user will know the output is correct, you do not have a product yet. Its twin is jagged intelligence — capability that is superhuman and incompetent in adjacent cells of the same task, which is why verification, not specification, is the binding constraint.

Vibe coding vs. agentic engineering. (2026.) Vibe coding — his own February 2025 coinage for building by natural-language feel — raises the floor: anyone can now produce something that runs. Agentic engineering raises the ceiling: it is the professional discipline of coordinating fallible agents while holding correctness, security, and maintainability. The first is a democratization; the second is a craft. He retired the first term in favor of the second within a year, which is itself the rhetorical signature (below) in action.

Autoresearch — the rebuild becomes the researcher. (2026.) Give an agent a minimal, readable training codebase and let it run the scientific loop by itself: propose a change, train for five minutes, evaluate, keep or revert, repeat overnight. On nanochat this found ~20 stackable improvements in ~700 experiments and transferred them to a larger model. The frame generalizes past ML: the highest-leverage artifact is not the answer but the smallest legible system on which a machine can search for the answer. This is the model he carried into Anthropic.

The recipe (six-step training). Become one with the data → end-to-end skeleton + dumb baselines → overfit → regularize → tune → squeeze juice. The aphorism: "The qualities that correlate most strongly to success are patience and attention to detail." The deeper claim is that ML research is closer to systems debugging than to mathematics. This is why he can teach it — he treats it as a craft with reproducible heuristics, not a mystery.

Backprop is a leaky abstraction. Frameworks that hide it "create a false sense of reliability." Sigmoid saturation, dead ReLUs, RNN gradient explosion. The philosophical underpinning of his pedagogy: if the abstraction leaks, understand the layer below. Hence micrograd, and in the end all of it.

RL is brute force. "Policy gradients are a brute force solution… profoundly dumb." Updated 2025: "RL sucks supervision through a straw." Not "it doesn't work," but "it's the dumbest way that works." RLHF works because the LLM brings the prior; pure RL fails because it has none.

Bitter lesson, adjusted. Sutton was right; pure parameter-count scaling is no longer enough. The new bottlenecks are continual learning, multimodality, and computer use.

The stack at a glance

  • Software 1.0 → 2.0 → 3.0. Code → weights → prompts. Build at the layer that doesn't yet have its tooling.
  • LLM as operating system. Kernel + RAM + peripherals. Three structural positions in any AI product: kernel, shell, peripheral. No fourth scales.
  • The cognitive core. The valuable model may be a small one that trades memorized knowledge for reasoning and looks the rest up.
  • Agents as "people spirits" — ghosts, not animals. Stochastic simulations with emergent psychology; a coworker with anterograde amnesia. Summoned from internet text, not evolved.
  • The autonomy slider. Demo is works.any(); product is works.all(). Ship partial autonomy with a verification loop.
  • The march of nines. Each nine of reliability costs as much as the last. Useful autonomy is years of nines away — hence the decade of agents.
  • Verifiability & jagged intelligence. Software automates what you specify; AI automates what you can verify. If the user can't cheaply verify, you don't have a product.
  • Vibe coding raises the floor; agentic engineering raises the ceiling.
  • Autoresearch. The smallest legible system a machine can search on beats the hand-tuned answer.
  • The recipe. Become one with the data → skeleton + dumb baselines → overfit → regularize → tune → squeeze juice.
  • Backprop is a leaky abstraction. If the abstraction leaks, understand the layer below.
  • RL is brute force. RLHF works because the LLM brings the prior; pure RL fails because it has none.
  • Bitter lesson, adjusted. Scale was necessary; scale alone is no longer sufficient.

The whole stack fits on a page. That is the lesson. Karpathy's intellectual legacy will likely be that he made the field's vocabulary smaller, not larger.

V. The Network and Its Center of Gravity

The professional graph has now shifted twice. In 2015 the dense cluster was academic (Fei-Fei Li, the Toronto/Hinton orbit, the Stanford committee) and the OpenAI founding cohort (Altman, Brockman, Sutskever, Schulman, Zaremba, Musk). Through 2025 the dense cluster became the public-intellectual orbit (Lex Fridman, Dwarkesh Patel, Sarah Guo, Stephanie Zhan, Garry Tan) plus students at scale. As of May 2026 it has swung back inside a lab: Anthropic, reporting into the pre-training group under Nick Joseph, adjacent to the Claude research organization.

Three observations survive the move and one is new.

He remains the only OpenAI founding-cohort member who became a public-facing teacher at scale, and he now carries that public identity into a frontier lab. That is a genuinely rare position: an inside researcher whose framings shape the outside discourse out of proportion to his publication count. It also creates a new tension the field will watch — how much of a public teacher can a person remain while working on a lab's most competitively sensitive layer, pre-training.

The striking absences persist. No close edge to Yann LeCun, Demis Hassabis, or François Chollet despite obvious shared interests. The pattern is a deliberate keep-it-small posture — many followers, few co-conspirators. He talks to a small set of people deeply; the rest of the field reads him.

The students are now numerically dominant and largely unmet — a meaningful fraction of senior ML engineers at every major lab sat in CS231n, and the diffuse YouTube cohort is untraceable. The new fact is that some of those students are now his colleagues. The pre-training team he is building at Anthropic will almost certainly include engineers who first learned the material from him. The teacher went to work inside the institution his teaching helped populate.

VI. The Worldview's Texture

The worldview is held together by three commitments, and the third now sits in productive tension with his employer.

First, the bitter lesson, earned the hard way. The 2014 Feature Learning Escapades essay is the receipt: he started believing in elegant unsupervised feature learning, watched supervised CNNs on ImageNet win, and updated. "We may have to pay in labels." The intellectual honesty of that update is what makes Software 2.0 possible three years later. The 2025 refinement: scale was necessary; scale alone is no longer sufficient.

Second, the conviction that humans must understand what they build. Yes You Should Understand Backprop is the philosophical core; the from-scratch instinct (micrograd, llama2.c, nanochat) is the practice. The bitter lesson and the comprehension imperative coexist in tension, not contradiction: scale wins for the system; comprehension wins for the engineer. Let your model be huge; understand every gradient.

Third, AGI as continuous engineering, not discontinuous event. From the October 2025 Dwarkesh interview: AGI ~a decade away; it will "blend into" the existing ~2% GDP growth curve; the bottlenecks are continual learning, multimodality, computer use. He is doubly anti-doomerist and anti-utopian — no P(doom) numbers, no singularity timelines. AGI as a long-tail engineering project, with a working scientist's calibration.

Two 2025 arguments give that third commitment its current texture. In "Power to the People" he notes that LLMs invert the usual order of technology diffusion: where the internet, GPS, and cryptography reached militaries and large corporations first, ChatGPT reached hundreds of millions of ordinary people before it reshaped any institution — which he reads as unusual and hopeful about who the technology serves. And he is uneasy about the field feeding on itself: models are, in his phrase, "silently collapsed" onto a thin slice of their own distribution, low in entropy, and training them on their own output narrows it further. It is one reason he thinks human data, and human noise — even dreaming — still matter, and a rare note of caution in an otherwise builder's optimism.

Here is the tension worth naming plainly, because a careless reader will call it a contradiction. In late 2025 Karpathy was the field's most credible skeptic of the near-term hype — "RL is terrible," "AGI is a decade away," current agents are "brittle." In May 2026 he joined a frontier lab's pre-training team. The two facts only look opposed if you think people join labs to chase a singularity. On his own worldview they fit perfectly: if AGI is a decade-long engineering slog rather than an imminent event, then the rational place for someone who wants to do the engineering is inside a lab with the compute — and the specific problem he took, using models to accelerate their own pre-training, is the one concrete lever that could bend a decade-long curve. The skeptic did not convert. He went to work on the unglamorous middle of the curve he had always described.

The rhetorical signature that holds it all together: state the maximalist claim, then immediately temper it. "RNNs are Turing-complete" → "don't read too much into this." "Decade of agents" → followed by why current agents fail. "RL is terrible" → "(but everything else is much worse)." The hedge is built in. This thesis has now watched that habit operate on his own career: he stated a career pattern by living it, then updated it in public by reversing it. The willingness to be visibly wrong about his own strongest framing is the trait.

VII. Eureka Labs — the Answer That Wasn't Terminal

The first version of this thesis called Eureka Labs the resolution of a fifteen-year tension: frontier resources require institutions, open output requires independence, and Eureka was the synthesis where 100% of the output is public and there is nothing left to exit from. That reading was the load-bearing error, and the way it failed is instructive.

The company's stated thesis was sound as far as it went — "a new kind of school that is AI native," human teachers designing curriculum, AI Teaching Assistants scaling delivery, the bet that great teachers are the bottleneck, not content delivery. The defensibility argument was that Karpathy's curriculum delivered at Khan Academy scale, with a Karpathy-trained TA in every student's pocket, is a moat, because CS231n's decade-long longevity proves his curricular taste is inimitable.

What the first version treated as strategic risk turned out to be the actual outcome, and faster than expected. It flagged four risks: the AI-TA layer commoditizing before the curriculum advantage compounds; the course design being too narrow; Karpathy himself being the one-person bottleneck; and frontier labs entering education directly with better distribution. It did not consider the fifth and decisive possibility — that the founder would simply lose interest relative to a more compelling problem. That is what happened. LLM101n sat archived and "currently being developed" for nearly two years and never shipped a public cohort. Meanwhile nanochat and autoresearch consumed his public attention through early 2026, and by May he had taken a role at Anthropic.

Two things must be said carefully here, because the temptation is to overclaim. First, Eureka Labs has not been shut down. No public source confirms a wind-down, sale, or closure; the honest description is paused and deprioritized. Second, Karpathy's own framing keeps the door open: "I remain deeply passionate about education and plan to resume my work on it in time." Read charitably, this is a pause, not an abandonment — consistent with a career of leaving and returning. Read skeptically, "in time" is the phrase founders use about the thing they have stopped doing. The truth is not yet knowable, and this thesis will not pretend otherwise.

The correct conclusion is narrower and truer than the original. Eureka Labs was not the terminal position; it was the swing to maximal-public in an oscillation, and it was undone by the same instinct that created it. The moment a more interesting thing to rebuild appeared — one requiring compute a school cannot buy — the instinct that built the school walked out of it. Whether the school survives the departure is the open question §XI now leads with.

VIII. The Content Corpus as Structured Argument

The corpus is unusually compact: ~15 substantive blog posts, nine Zero to Hero videos, sixteen CS231n lectures, ~12 industry talks, six long-form interviews, ~12 GitHub repos. The whole canonical Karpathy fits on a single index page. This compactness is not accidental; it is the same intellectual move that produces the mental-model stack — make the vocabulary smaller, not larger. The corpus is small enough that any motivated reader can consume the entire thing. By design.

For a first encounter (~3 hours): Software 2.0 → Yes You Should Understand Backprop → Pong from Pixels → Intro to LLMs. Five artifacts give you the operating posture; everything else is detail. For the actual learning sequence (180 — 240 hours), the path is layered to mirror the abstraction stack: CS231n teaches the unit operations of deep learning; Zero to Hero rebuilds backprop atomically and layers PyTorch on top; the GPT lectures wire transformer/attention/LayerNorm onto the language-modeling foundation; the systems stack (llama2.c, llm.c, nanochat) extends into hardware, deployment, and post-training. Each stage is the prerequisite for the next.

Four observations on the corpus shape:

The blog is the load-bearing channel, not papers. His influence is overwhelmingly via blog posts and YouTube lectures, not peer-reviewed publications. He chose the un-gated channel — comprehension transmitted, not novelty certified.

The repos are the philosophical core. Each is a from-scratch implementation of something the field already has at frontier — micrograd is autograd, nanoGPT is GPT, llama2.c is Llama-2 inference, nanochat is a full ChatGPT pipeline. The redundancy has always been deliberate. And autoresearch is the repo that changed what the repos are for: the first one whose audience is a machine rather than a student.

The talks are increasingly the sharper signal than the essays. From 2023 onward — Intro to LLMs, State of GPT, the Software 3.0 keynote, the 2026 Sequoia fireside — the spoken artifact has done more of the framing work than the written one. And the teaching videos got longer and more ambitious: the four-hour Let's reproduce GPT-2 (2024), then the general-audience Deep Dive into LLMs like ChatGPT (~3.5 hours) and How I Use LLMs (2025), which document not just how the models work but how he personally works with them. He has moved toward a talk-first, and now video-first, cadence.

There is a second, quieter corpus — his own operating methods. Alongside the ML writing sits a run of personal-systems essays: an append-and-review note-taking method, a "digital hygiene" regime, "Biohacking Lite" on his own metabolism, a hunt for the best sleep tracker, a log of his own productivity. They read like hobby pieces, but they are the Recipe (§IV) turned on himself — instrument the system, become one with the data, iterate. For a reader trying to copy not just what he thinks but how he works, this is where the method is most literal: he runs his life the way he runs a training loop, and he publishes the run.

The Twitter/X presence is more load-bearing than expected. Several of his strongest 2024-2026 framings — "vibe coding," the nanochat and autoresearch releases, the agentic-engineering pivot, and finally the Anthropic announcement itself — appeared first on X, not in essays. For a figure this influential, X is now the primary publication channel.

IX. The Replication Question

What does Karpathy's career teach a working professional? The trap is to copy the format (YouTube videos, ML curriculum) rather than the posture. The lesson was never "make videos." After the 2026 reversal it is sharper still: organize your work around the thing you most need to understand, build the smallest honest version of it yourself, publish when you can, and change environments completely when the thing you're chasing moves.

Operating rules, categorized.

Adopt — universal across high-craft work.

  1. Rebuilding is the comprehension test. Whatever you cannot rebuild from scratch and explain, you do not yet understand. Outside ML: whatever you cannot reconstruct in a one-page memo, you do not yet understand.
  2. The public artifact is the asset. The work that compounds is the work that survives the moment. Between "polish this email" and "publish this framework," default to the framework.
  3. Choose the un-gated channel. Optimize for comprehension transmitted, not novelty certified.
  4. State the maximalist claim, then temper it. Strong claim, integrated hedge, fact-anchored. The caveat travels with the claim.
  5. Make the vocabulary smaller, not larger. Compact frameworks compound; sprawling ones don't. Whole-stack-on-a-page is the test.
  6. Verifiability beats specifiability. If the recipient cannot cheaply verify the claim, it is not yet shippable.
  7. Build the kernel, the shell, or the peripheral. Three structural positions in any AI-built product. No fourth scales.

Adapt — same principle, calibrated by context.

  1. Follow the object, not the environment. Most people optimize title, resources, and independence, and let the object of attention drift. Reverse it. Pick the thing you most need to understand and let it decide where you work — even if that means leaving the position you built.
  2. Configure the role so the artifact is the day-job — until it can't be. Reshape the role so compounding public work sits at the center rather than the margin. But hold this loosely: the point is the work, not the configuration, and sometimes the best work is temporarily private.
  3. The autonomy slider as product strategy. Don't ship full autonomy; ship partial autonomy with an escalating human-verification loop. The slider is the product.

Reject — don't import.

  1. Don't optimize for novelty when transmission fidelity is the goal. Software 2.0 isn't novel ML; it's the highest-fidelity transmission of an idea already in the air. That is the higher contribution.
  2. Don't mistake the position for the point. The first version of this thesis did exactly that — it treated Eureka Labs as a destination. The portable lesson is a gradient to follow, not a chair to reach.

X. What Could Be Wrong

This section is where the first version of the thesis earned its keep, so it is worth quoting what it said before the fact. It wrote: "If Karpathy stays at Eureka for fifteen years, the pattern reading is confirmed. If he leaves Eureka after three or four years, the pattern reading is partial — Eureka was not the structural answer either, and we do not yet know what is." He left after roughly two. The disconfirming signal fired early and hard. Honesty requires cashing that out rather than absorbing it smoothly.

What broke. The specific claim that his career is a monotonic march toward 100% public output, with Eureka as the terminus, is false. The public-output share is not a ratchet; it went down in May 2026. Any reading built on "he is done leaving institutions" is dead.

What survives, and why to trust it less than before. The oscillation reading — comprehension-by-rebuilding as the invariant, environment as the free variable — fits both the fifteen years and the reversal. But note the epistemic hazard: a thesis that fits both "he stays independent forever" and "he goes back inside a lab" is in danger of being unfalsifiable. The honest guard is to state in advance what would break the new reading. It would break if Karpathy's Anthropic tenure produced no comprehension-oriented public artifact over several years — if the man who has published a from-scratch rebuild at every prior station went fully dark. The invariant predicts he won't. That is the next test, and it is a real one.

Two further ways the current reading could be wrong. First, the mundane explanation: perhaps there is no deep invariant and Karpathy simply took an extraordinary offer at the most important lab of the moment, and the "oscillation" is a story imposed on ordinary careerism. The defense is weak-to-moderate — the offer is real, but the autoresearch-to-Anthropic throughline is specific enough that pure opportunism under-explains it. Second, the education question genuinely hangs. If he resumes Eureka within a couple of years, the oscillation reading strengthens; if "in time" quietly becomes never, then the teaching was more contingent than even this revised thesis claims, and the invariant is narrower still — closer to "he goes where the frontier compute is," which would be a smaller and less flattering claim than the one written here.

XI. Open Questions Worth Pulling On

Listed in rough priority order.

The Anthropic tenure's public output. The single most important watch item, because it is the falsification test for §III and §X. Does the from-scratch, comprehension-oriented public artifact survive inside a frontier pre-training team, or does he go dark? The invariant predicts at least one nanochat-shaped artifact emerges from this period. Watch for it.

The fate of Eureka Labs. Paused, not closed. Does "in time" become a return, a quiet shutdown, or a handoff? The answer decides whether the oscillation reading or the narrower "follows the compute" reading is correct.

What autoresearch becomes at frontier scale. The Anthropic mandate — using Claude to accelerate pre-training research — is autoresearch with a frontier budget. If it works, it is one of the more consequential ideas in AI, and Karpathy is running it. This is worth its own study within a year.

The Nick Joseph / pre-training-team dynamic. Reporting into a team lead rather than founding or directing is a new posture for him. How he operates as a senior individual inside someone else's structure is informative about a side of him the public artifacts don't show.

The X corpus, still under-read. Vibe coding, the nanochat/autoresearch narration, agentic engineering, the Anthropic announcement — all appeared first on X. A systematic pass through 2024-26 X threads would update the corpus index and likely add to the mental-model stack.

The Sutskever / Karpathy edge. The deepest single intellectual peer relationship in the OpenAI cohort. With both now running frontier-adjacent research at competing labs, any public signal about how they think under technical pressure remains the most informative unavailable artifact.

XII. The Take-Home, in One Sentence

A builder who understands systems only by rebuilding them from scratch, who for fifteen years found the most interesting rebuilds outside the lab and so looked like the field's great teacher — until, in 2026, the most interesting rebuild became a training run that improves itself, which required frontier compute, and pulled him back inside.

The career is not "AI researcher who happens to teach," and it is not the tidier "teacher who happens to keep getting hired by frontier labs" this thesis first proposed. It is an oscillation with a fixed center: comprehension by rebuilding. Teaching, labs, the school, the return — these are the swings; the from-scratch instinct is the pivot. What is portable is not the format (blog posts, lectures, repos) and not the position (independent teacher). It is the discipline of following the thing you most need to understand, wherever it goes, and building the smallest honest version of it with your own hands.

The first version of this study ended "As of May 2026" and was overtaken within weeks. This one states its own expiry plainly: the reading here holds until Karpathy's Anthropic period either produces the public rebuild the invariant predicts, or doesn't. As of August 2026.

© Rishi Dogra · CC-BY 4.0