← Home
← Back to Studies
STUDY 002

Andrej Karpathy

He works in his spare time so he can teach.

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.

The honest story is that he is a frontier-AI builder who keeps quitting frontier-AI labs to teach. Three exits — Stanford, Tesla, OpenAI (twice) — and at every transition the most-cited public artifact of his tenure was produced near the exit, 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 the institution where his compounding work could be the day-job: Eureka Labs.

The career has a single shape. Each transition reduces the share of his output that is institutional-private and increases the share that is public. The terminal value — Eureka Labs — is the position where the answer is 100%.

If you read only one sentence from this document, read this one: Karpathy's most cited work is his exit work, and the company he eventually built is the one where leaving is no longer the way to ship the public artifact.

The rest of this thesis is the unpacking of that sentence.

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. The signature artifacts of those months are The Unreasonable Effectiveness of Recurrent Neural Networks (May 2015) and Pong from Pixels (May 2016) blog posts, both pedagogical. 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 — at Tesla, hand-written C in the perception stack was being progressively replaced by neural nets, and the essay extracts the principle from the practice. 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 choose to do? He teaches. The implication is load-bearing: his teaching is not the day-job's leftover; it is the thing he does when the day-job is removed.

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. The repository is archived in August 2024 with a README stating the course is "currently being developed." As of May 2026, no public release date.

The 2025 — 2026 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 recent statement of his AGI timelines (~decade away, continuous not discontinuous, RL is terrible). The early-2026 Sequoia Ascent fireside introduces verifiability as the new constraint and retires "vibe coding" (his February 2025 coinage) as obsolete in favor of "agentic engineering."

Recognition: 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 — Pedagogy as Research

This is the central insight of this thesis.

Karpathy's most prolific creative period was the 2022 sabbatical. No employer. No grant. No research-deliverable obligations. The output: a from-scratch ML curriculum that has educated more LLM engineers than any other single corpus.

This is not a coincidence. It is a tell.

The conventional reading of his career is that he teaches in his spare time. The honest reading is that he works in his spare time so he can teach. The teaching is the research. The teaching is the thing the institutions interrupt.

The evidence converges from four directions.

First, the artifact-timing pattern. 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 is not as production environments but as ladders to climb so the next public artifact can be produced from a higher vantage.

Second, the structural choice in 2024. When Karpathy left OpenAI for the second time, he did not start a foundation-model lab (which his network and reputation could have funded). He did not start a venture firm (which his network would have welcomed). He did not return to academia (which Stanford would have offered). He started a school. The decision he made when he had the freest possible hand was to build the institution that lets him teach.

Third, the from-scratch instinct. micrograd, makemore, nanoGPT, llama2.c, llm.c. The same pattern across a decade: take a frontier system, rebuild a minimal working version, publish the code. This is not a teaching method. It is a research method. Karpathy understands a system by rebuilding it, and his rebuild is the public artifact. The pedagogy and the research are the same activity.

Fourth, his own framing in the corpus. Yes You Should Understand Backprop (Dec 2016) is the philosophical key: "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.

The implication of pedagogy-as-research is that Eureka Labs is not a career change. It is the structural position he has been working toward for fifteen years. Each prior institution was a partial solution — high resources, low public output (Tesla, OpenAI). Each public-facing rest period was the inverse — high public output, low resources (sabbatical). Eureka Labs is the synthesis: a private company whose product is the public artifact, where being open is the moat, where there is no longer a tension between "do the institutional work" and "ship the teaching."

This is what makes him difficult to imitate. The Karpathy posture is not "be smart enough to teach AI." It is "structure your career so that public-facing work is the day-job and not the side-project." Most working professionals — including most working AI researchers — cannot do this. He spent fifteen years bending an existing path before starting a new one.

The pedagogy-as-research claim, if accepted, also resolves a number of secondary puzzles:

  • Why he keeps leaving institutions cleanly (no public conflict in any departure): the institutions never had what he wanted.
  • Why he produces fewer papers than expected for his stature: papers are gated public artifacts; blog posts and YouTube lectures are not. He chose the un-gated channel.
  • Why his most-cited work is rarely his most-novel research: novelty is not what he is optimizing for. Comprehension transmitted to the next reader is what he is optimizing for. Software 2.0 is not novel ML — it is the highest-fidelity transmission of an idea that was already in the air.
  • Why Eureka Labs is a school and not a tool company: the bottleneck he is attacking is taste / curriculum design scarcity, which is the bottleneck pedagogy-as-research generates as it scales.

Read the rest of this thesis through this lens.

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; the whole stack fits on a page.

The central arc: 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 (the binary). 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 (his rhetorical question — "Is there space for a Software 2.0 Github?" — was answered by HuggingFace). The 2025 update predicts the next category: 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.

Neural nets as a new programming paradigm. "If training vanilla neural nets is optimization over functions, training recurrent nets is optimization over programs." The instruction set is minimalist by design — "only two operations: matrix multiplication and thresholding at zero (ReLU)." This collapses the supposed gap between ML research and software engineering. Training is programming. Data is source.

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." The mapping: LLM = CPU/kernel; context window = RAM; tools = peripherals; prompts = programs. The 2025 widening: LLMs as utilities (electricity-grid analogy), as fabs (only a few players can build the kernel), as timeshare mainframes (predicts the inevitable next phase of local LLMs). The framework's productive prediction: most consumer "AI products" today are shells/applications 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.

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.

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 (Tab → Cmd+K → Cmd+L → Cmd+I), Perplexity (search → research → deep research), Tesla Autopilot (levels 1 — 4) — the slider is the product.

The recipe (six-step training methodology). 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 DQN clip-by-value bug. The philosophical underpinning of his pedagogy: if the abstraction leaks, you must understand the layer below. Hence micrograd. Hence everything.

RL is brute force. "Policy gradients are a brute force solution… the approach we use is also really quite profoundly dumb." Updated 2025: "RL sucks supervision through a straw." Karpathy is RL-skeptical in a specific way — not "it doesn't work," but "it's the dumbest way that works, and the path forward requires model-based abstraction." RLHF works because the LLM brings the prior; pure RL fails because it has none.

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.

The bitter lesson, adjusted. Sutton was right; pure parameter-count scaling is no longer enough. The new bottlenecks are continual learning, multimodality, and computer use. Eureka Labs sits inside this — continual personalized learning is the product.

The whole stack fits on a page. That is the lesson. The frameworks that compound are the ones compact enough to carry. Karpathy's intellectual legacy will likely be that he made the field's vocabulary smaller, not larger.

The stack at a glance

  • Software 1.0 → 2.0 → 3.0. Code → weights → prompts. Three additive layers. Build at the layer that doesn't yet have its tooling.
  • Neural nets are programming. Training is compilation. Datasets are source.
  • LLM as operating system. Kernel + RAM + peripherals. Three structural positions in any AI product: kernel, shell, peripheral. No fourth scales.
  • Agents as "people spirits." Stochastic simulations with emergent psychology. Coworker with anterograde amnesia.
  • The autonomy slider. Demo is works.any(); product is works.all(). Ship partial autonomy with a verification loop.
  • The recipe. Become one with the data → skeleton + dumb baselines → overfit → regularize → tune → squeeze juice. ML is closer to systems debugging than to mathematics.
  • Backprop is a leaky abstraction. If the abstraction leaks, understand the layer below. The pedagogical core.
  • RL is brute force. "Sucks supervision through a straw." RLHF works because the LLM brings the prior; pure RL fails because it has none.
  • Verifiability is the new constraint. Software automates what you specify; AI automates what you can verify. If the user can't cheaply verify, you don't have a product.
  • Bitter lesson, adjusted. Scale was necessary; scale alone is no longer sufficient. New bottlenecks: continual learning, multimodality, computer use.

V. The Network and Its Center of Gravity

The professional graph has shifted decisively. In 2015 the dense cluster was Tier 1 (academic mentors: Fei-Fei Li, Hinton-orbit Toronto, the Stanford committee) and Tier 2 (OpenAI cohort: Altman, Brockman, Sutskever, Schulman, Zaremba, Musk). In 2026 the dense cluster is Tier 4 (public intellectual orbit: Lex Fridman, Dwarkesh Patel, Sarah Guo + Elad Gil, Stephanie Zhan, Garry Tan) and Tier 5 (students at scale: Justin Johnson and the CS231n cohort, the diffuse YouTube cohort, the Eureka Labs students).

Three observations sit on this shift.

He is the only OpenAI founding-cohort member who has become a public-facing teacher at scale. Sutskever has done occasional Dwarkesh appearances; Schulman is rarely seen outside research venues; Brockman and Altman are operators, not pedagogues. Karpathy occupies the bridge between the technical inner circle and the public-facing outside in a way no one else in his cohort does. This is a rare and load-bearing position. It is also why his framings shape the public discourse out of proportion to his publication count.

There are striking absences. No close edge to Yann LeCun, despite obvious shared interests in self-supervised learning. No edge to Demis Hassabis or DeepMind beyond his 2015 internship. No edge to François Chollet, despite shared pedagogical instincts. The absences suggest 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. A meaningful fraction of senior ML engineers at every major lab sat in CS231n. The diffuse YouTube cohort is operationally significant and literally untraceable. The pattern is that his "students" are now numerically dominated by people he has never met. This is the structural condition that makes Eureka Labs the natural next move — at some scale of teaching, you are no longer Karpathy-teaching-students, you are Karpathy-curriculum-being-delivered. The company exists to operationalize the position the field has already put him in.

VI. The Worldview's Texture

The worldview is held together by three commitments.

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 the Software 2.0 essay possible three years later — he is not arguing for neural nets as a clever new technique, he is arguing for them as the inevitable substrate. 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. You should let your model be huge. You should understand every gradient.

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

The rhetorical signature that holds it together: state the maximalist claim, then immediately temper it. "RNNs are Turing-complete" → "don't read too much into this." "AGI will certainly be written in Software 2.0" → in an essay otherwise grounded in tooling. "Decade of agents" → followed by careful enumeration of why current agents fail. "RL is terrible" → "(but everything else is much worse)." The pattern handles the asymmetric cost of being wrong in public — bold framing, hedged claim, and a willingness to be visibly less-than-certain about his own strongest framings. The hedge is built in. This is why his predictions age well.

VII. Eureka Labs — the Structural Answer

Eureka Labs is the resolution of a fifteen-year tension. The tension: frontier resources require institutions; open output requires independence. He spent fifteen years toggling between the two.

The company's stated thesis — "a new kind of school that is AI native," with human teachers designing curriculum and AI Teaching Assistants scaling delivery — is correct as far as it goes. But it does not fully describe what the company is for. The company exists for two reasons.

First, structurally: it is the position where 100% of his output is public. There is no proprietary internal codebase to leave behind, because everything is the curriculum, and the curriculum is open. The exit pattern that defined his prior career terminates here, because there is nothing to exit from.

Second, defensibly: the bet is that great teachers are the bottleneck, not great content delivery. Most EdTech is in the business of distributing average-quality instruction at scale and has consistently failed to do so profitably. Eureka inverts the premise. Karpathy's curriculum delivered at Khan Academy scale, with a Karpathy-trained TA in every student's pocket — that is the moat. The curriculum is the asset; the AI is the leverage. The defensibility analysis works only if the courses themselves are exceptional in a way that competitors cannot easily replicate. CS231n's longevity (still cited a decade later) is the proof point.

The strategic risks are real. The AI-TA layer may commoditize faster than the curriculum advantage compounds. If GPT-5 or Claude 5 can deliver excellent tutoring against any decent curriculum, Eureka's moat collapses entirely to curriculum quality — a defensible position only if Karpathy's taste is so good that even commodity AI delivery against it dominates competitor curriculum + frontier AI delivery. Bet-the-company on his taste continuing to be exceptional and inimitable.

The course design may be too narrow. LLM101n is for a very specific learner — undergrad-level, technically motivated, willing to train their own model. The market for "I want to seriously build AI from scratch" is real but small. The company's expansion path beyond LLM101n is the strategic question.

Karpathy is the bottleneck Eureka is supposedly removing. The whole company's defensibility rests on his curricular taste. He is the pre-LLM bottleneck the AI-TA scales — but he is also a single person, and one-person companies have one-person strategic risk.

Frontier labs are entering education directly. OpenAI, Anthropic, and Google all have nascent education plays. They have the LLM advantage, the distribution, and the ability to hire great teachers. Eureka's response is "we got there first with Karpathy" — which is true, and which may not be enough.

That LLM101n has been archived for ~21 months and has no public release date as of May 2026 admits two readings. Read A: slow build, deliberate. CS231n took two years to mature; Zero to Hero rolled out over 18 months; the 17-chapter LLM101n syllabus is more ambitious than either. Read B: productizing teaching is harder than productizing teaching content. The Zero to Hero videos are content (free, on YouTube); LLM101n is a product (paid, with AI-TA infrastructure, with cohort delivery, with measurement of student outcomes). The gap between these is large; many great teachers fail to make the transition. Both reads are probably partially true.

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. ~10 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 (§IV). Karpathy's signature is making the field's vocabulary smaller, not larger. The corpus is small enough that any motivated reader can consume the entire thing. By design.

The reading-order logic for first encounters (3 hours): Software 2.0 → Yes You Should Understand Backprop → Pong from Pixels → Intro to LLMs talk. 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. The order is not arbitrary; each stage is the prerequisite for understanding what makes the next stage possible.

Four observations on the corpus shape:

The blog is the load-bearing channel, not papers. Karpathy's intellectual influence is overwhelmingly via blog posts and YouTube lectures, not peer-reviewed publications. He chose the un-gated channel. This is consistent with the pedagogy-as-research thesis: the goal is comprehension transmitted, not novelty certified.

The repos are the philosophical core. Each one 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; llm.c is the training loop. The redundancy is the point. He is not trying to compete with PyTorch or HuggingFace at frontier; he is providing a comprehension path to the layer below.

The talks are increasingly the sharper signal than the essays. From 2017 — 2021, the blog was where the major framings appeared. From 2023 onward, the talks have done more of the work — Intro to LLMs (Nov 2023), State of GPT (May 2023), Software 3.0 keynote (June 2025), Sequoia Ascent fireside (early 2026). He may be moving from a written-essay-first cadence to a spoken-talk-first cadence. This is consistent with Eureka Labs's medium (spoken instruction at scale).

The Twitter/X presence is more load-bearing than expected. Several of his strongest 2024-2026 framings — "vibe coding," the System Prompt learning idea, the agentic-engineering pivot — appear first on X, not in essays. The X corpus deserves a closer read than this thesis gives it.

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 is not "make videos." The lesson is build the public artifact early, choose institutions that let you, and leave when they don't. Twelve operating rules, categorized.

Adopt — universal across high-craft work

  1. Pedagogy is research's most honest test. Whatever you cannot rebuild from scratch and explain, you do not yet understand. Backprop is a leaky abstraction generalizes far beyond ML — frameworks lie, abstractions leak, and the only defense is comprehension of the layer below. For non-ML work: whatever you cannot explain in a one-page memo, you do not yet understand.

  2. The public artifact is the asset. Three institutional exits, three signature artifacts each timed to the exit. The work that compounds is the work that survives the moment. When choosing between "polish this email" and "publish this framework," default to the framework.

  3. Choose the un-gated channel. Blog posts and YouTube lectures travel further than peer-reviewed papers because nothing gates them. Optimize for comprehension transmitted, not novelty certified.

  4. State the maximalist claim, then temper it. Strong claim, integrated hedge, fact-anchored. "X is true. Important caveat: Y. The implication is Z." The hedge is built in.

  5. Make the vocabulary smaller, not larger. Compact frameworks compound; sprawling ones don't. Whole-stack-on-a-page is the test for any model you'd want others to carry.

  6. Verifiability beats specifiability. If the recipient cannot cheaply verify the claim, the claim is not yet shippable. Code with tests; design with rendering; math with proofs.

  7. Build the kernel, the shell, or the peripheral. Three structural positions in any AI-built product. No fourth scales. Diagnostic for any "AI-powered X" pitch: which structural position?

Adapt — same principle, calibrated by context

  1. Configure the role so public-artifact output is the day-job, not the side-project. Most working professionals produce less compounding public work than they could because the institutional configuration squeezes it to the margins. The structural move is rarely to leave the institution; it is to reshape the role.

  2. From-scratch is the comprehension test. Karpathy's repos are pedagogy and research at once. The analogous move outside ML: write the one-page memo that rebuilds the system from first principles. If you can't, you don't yet understand.

  3. The autonomy slider as product strategy. Don't ship full autonomy. Ship partial autonomy with an escalating human-verification loop. Cursor, Perplexity, Tesla Autopilot — 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 that was already in the air. That is the higher contribution, not the lesser one.

  2. Don't accept resource-rich institutions if they squeeze your public output to the margins. The terminal value is the configuration where there is nothing to exit from.

The Karpathy posture is structural, not productivity-tactical. He spent fifteen years bending an existing path before starting a new one. The portable lesson is not the format; it is the configuration.

X. What Could Be Wrong

The thesis above could be wrong in three specific ways. Each is worth holding.

The pedagogy-as-research framing could be too clean. It is possible Karpathy is simply someone with unusually good intuition who happens to like teaching, and the artifact-timing pattern is coincidence rather than thesis. The data is consistent with both readings; the reason to prefer the structural reading is that the 2024 founding of Eureka Labs is otherwise hard to explain — it is a costly, irreversible, structural commitment to teaching, made when his alternative options were maximally lucrative. Costly commitments are the high-quality signal.

Eureka Labs could fail. LLM101n could not ship; the AI-TA infrastructure could prove unbuildable at quality; competitors could pre-empt. If Eureka fails, the thesis still holds in its core claim (pedagogy was his actual operating posture), but the company would no longer be the structural answer to the fifteen-year tension — it would be a failed attempt at the structural answer, and the next move would be ambiguous.

The "exit pattern" reading depends on what comes next. If Karpathy stays at Eureka for fifteen years, the pattern reading is confirmed — he found the position he was looking for. 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. The next test of the thesis is the durability of his commitment to Eureka.

XI. Open questions worth pulling on

Listed in rough priority order. These are the threads this thesis has not fully unwound and which would tighten the analysis if pulled.

The Dwarkesh interview deep-read. The October 2025 AGI is still a decade away interview is the most authoritative recent statement of his thinking, and this thesis sampled it strategically rather than transcript-deep-read it. A full pass would tighten §VI (worldview) and §IV (mental models) significantly. Likely to surface 2-3 new framings not yet captured here.

LLM101n watch. When (or if) it ships, the curriculum becomes a primary source for the Eureka Labs read in §VII. The thesis on the company is partial until then.

The Twitter/X corpus. Several recent framings ("vibe coding," System Prompt learning, agentic engineering) appear first on X. A systematic pass through 2024-26 X long-form 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 Toronto + Hinton lineage on both sides. Public material is sparse. Worth watching for any joint appearances; the conversational dynamic between them is likely the most informative single artifact about how Karpathy thinks under technical pressure.

The Tesla period's intellectual residue. Five years. Considerable archive material from AI Day 2021, Autonomy Day 2019, ICML 2019. This thesis treats them as supplementary to the blog; a Tesla-deep pass might surface mental models specific to embodied AI, fleet learning, and edge-case engineering that aren't on the blog.

The pedagogy-as-research thesis itself. The strongest test: does it predict Karpathy's next move? If the thesis is right, his behavior over the next 2 — 3 years should look like deepening the public-artifact channel (more Eureka courses, possibly a book, possibly more spoken talks) rather than returning to a private-research environment. Watch for confirming/disconfirming signals.

XII. The take-home, in one sentence

A frontier-AI builder who keeps quitting frontier-AI labs to teach — three exits, three signature artifacts each timed to the exit, then a company configured so there is nothing left to exit from. The career is not "AI researcher who happens to teach"; it is "teacher who happens to keep getting hired by frontier labs," with Eureka Labs as the structural answer to the fifteen-year tension between frontier resources and open output. What's portable is not the format (blog posts, lectures, repos) but the operating posture.

What's portable, in one phrase: pedagogy is research, the public artifact is the asset, and the vocabulary that compounds is small enough to carry.

As of May 2026.

© Rishi Dogra · CC-BY 4.0