Andrej Karpathy

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI is fundamentally a story about a change in the substrate of computation, and Karpathy is the person who named that change most precisely. Where Edo Segal describes the consumer-facing surface—a hundred dollars a month buying what once required a team of engineers—Karpathy describes the engine. The democratization Segal celebrates is downstream of the fact that the act of programming has changed so completely that anyone who can describe what they want can now attempt to build it. Claude Code and its siblings are Software 3.0 in operation; the orange pill moment is what Software 3.0 feels like from the user’s side.

His framework also supplies the cycle’s most rigorous account of why the transformation is partial. Karpathy insists the systems are jagged: expert in some domains, elementary in others, with no smooth surface connecting the peaks. The same model that drafts a competent legal brief will confidently miscount the letters in a word. This jaggedness is not a transitional bug but a structural feature of optimizing for narrow objectives—the systems were shaped by imitating human text, not by living in the world, which is why Karpathy calls them ghosts rather than animals. The cycle’s portrait of the human difference—consciousness, the wondering self that asks why—finds its engineering complement in Karpathy’s ghost framing: the difference is not a matter of degree along a single scale but a difference of origin, written into how each kind of thing came to exist.

His concept of the autonomy slider offers the cycle’s most practical answer to how humans and machines should work together during the decade of transition. Rather than a binary between full human control and full machine autonomy, he proposes a continuum calibrated to the cost of verification in each domain. Where checking the machine’s output is cheap, autonomy can be high; where a confident error causes real harm, the human must stay close. The Trivandrum training session Segal describes—twenty engineers redirected from execution to judgment—is the autonomy slider moved thoughtfully upward, with the human migrating from writing every line to supervising, verifying, and steering.

What places Karpathy in the cycle’s gallery of indispensable thinkers is the combination of deep technical authority and intellectual honesty about the limits of his own field. He could have let vibe coding stand as an uncomplicated celebration of AI-assisted development; instead he spent considerable effort explaining its scope conditions, insisting on the distinction between play and production. He could have extrapolated the steep early stretch of the capability curve into an imminent arrival of machines that match us across the board; instead he says the timeline is longer than the hype suggests, that reinforcement learning as currently practiced amounts to sucking supervision through a straw, and that the most important work ahead is not making the model bigger but building the surrounding system—the memory, the tools, the orchestration—that turns raw capability into reliable application.

Origin

Born in Bratislava and raised in Toronto, Karpathy studied computer science at the University of Toronto before earning his doctorate at Stanford under Fei-Fei Li, where he created and taught CS231n, the university’s first deep learning course. In that classroom he developed the pedagogical instinct that would define his public career: the conviction that you do not truly understand a thing until you have built it yourself from first principles, that the demystification of difficult machinery is itself an act of empowerment. He went on to become a founding member of OpenAI and then director of artificial intelligence at Tesla, where his years leading the Autopilot vision system gave him an education that purely linguistic AI does not provide—the education of trying to make a machine act reliably in the physical world, where mistakes have weight and the environment does not cooperate.

At Tesla, Karpathy confronted the long tail: the endless parade of rare situations, each individually improbable but collectively constant, that the system has never seen and must nonetheless handle correctly. A driving model that is brilliant in ninety-nine percent of situations and catastrophically wrong in the remaining one percent is not ninety-nine percent of the way to a solution; it is dangerous. His response was the data engine—a disciplined loop of identifying failures in the field, collecting and labeling the relevant data, and retraining to close the gap—and from this experience he drew the lesson that the decisive variable in most real-world AI applications is not the architecture but the diet, the quality and composition of what the system is trained on. Software 2.0 was not a theory he developed in a seminar. It was a description of what was already happening to the codebase under his supervision.

After Tesla he returned to OpenAI briefly before founding Eureka Labs in 2024 with the conviction that the same technology he had spent his life building could transform how people learn—that an AI tutor, properly designed, could offer every student the kind of patient, personalized instruction that has historically been available only to the privileged few. In 2026 he joined Anthropic’s pretraining team. Across every role, his most durable contribution may be neither a model nor a company but a sequence of lectures: Neural Networks: Zero to Hero, in which he builds the core machinery of modern AI from nothing, in plain code, assuming only basic Python and a hazy memory of high-school calculus, and carrying the learner all the way to a working miniature of the transformer.

Key Ideas

Software 2.0. Karpathy’s central claim is that software is undergoing a change of kind rather than degree. Software 2.0 replaces explicit human-written instructions with learned weights: humans specify a goal and a rough architecture, and an optimization process discovers the actual program. The source code is the dataset; the compiler is training; the binary is the trained network. This inverts the entire activity of programming, shifting prized skill from algorithmic cleverness to data judgment—knowing what examples the model needs, recognizing gaps in its experience, and curating them into existence. The programmer’s relationship to the resulting artifact is unlike anything in the history of the craft: you built it, but you cannot read it.

The LLM as a new kind of computer. Karpathy resists metaphors that frame large language models as minds. The metaphor he reaches for instead is computational: the LLM is a new kind of computer, and we are in the early, primitive period of its operating system. The context window functions like working memory; the model’s weights function like a vast read-only store of knowledge. He has compared frontier model capability to an electrical utility—produced by a small number of capital-intensive providers and distributed to vast numbers of consumers who pay by the token—and to a semiconductor fabrication plant: a small set of immensely sophisticated, expensive facilities at the frontier of what is physically and economically possible.

Jagged intelligence. The capability profile of large language models is not uniform. It is jagged—brilliant in places, startlingly weak in others, with no smooth surface connecting the peaks and valleys. The jaggedness is not a bug being slowly fixed but a structural feature of optimizing for narrow objectives, and it means that competence in one region licenses no confidence in any adjacent region. The fluency of a language model is seductive; it speaks in the register of expertise even when it is wrong, and its mistakes do not announce themselves. Jaggedness is a discipline of humility about a system that invites overconfidence.

Ghosts, not animals. Large language models are not digital animals shaped by evolution under survival pressure. They are ghosts—statistical echoes of human text, fluent because they were trained to reproduce the patterns of human expression, but lacking the embodied, evolved substrate that grounds a real mind. They have no body, no survival pressure, no continuous existence, no stake in the world. The ghost framing dissolves a great deal of confused thinking: systems that lack animal drives will not turn on us like predators, but they will also lack the embodied common sense that any creature learns through a body acting in the world.

The autonomy slider and the decade of agents. Rather than a binary between full human control and full machine autonomy, Karpathy proposes an autonomy slider—a continuum the human adjusts depending on the task, the stakes, and the trustworthiness of the system. The constraint on agent capability is not primarily the intelligence of the underlying model but the difficulty of verifying that the agent did the right thing. Where verification is cheap, autonomy can be high; where a confident error causes real harm, the human must stay close. Karpathy calls the coming period the decade of agents, choosing a decade deliberately as a rebuke to those announcing the age has already arrived.