Karpathy
He set the human benchmark on ImageNet, built Tesla’s vision-only driving system, coined “vibe coding,” and kept returning to the same thing: finding the irreducible minimum and giving it away. A biopic on the teacher who builds.
Somewhere around 2014, Andrej Karpathy sat alone at a computer, manually classifying images of dogs. One thousand categories. Hundreds of photographs. He had built a custom labeling interface, tried to outsource the work to crowdsourced labor, tried recruiting undergraduates, even organized a lab “labeling party” that produced error rates too high to use. So he did it himself. By image 200, he later admitted, he was only continuing “for science.”
The exercise had a specific purpose. He wanted to establish the human benchmark on ImageNet, the standard visual recognition challenge, so that the research community would have a number to compare against. He classified roughly 1,500 images and landed at a 5.1% error rate. GoogLeNet, the leading neural network at the time, came in at 6.8%. Humans were still ahead.
Within a year, they weren’t. Microsoft’s ResNet model hit 3.57% in December 2015, decisively below Karpathy’s mark. His error rate ended up cited in a federal government report on the future of artificial intelligence. The man who set the bar became the bar that got cleared, and then he moved on to building the courses that would teach the next generation how to clear it.
That is, in compressed form, the whole story. But the compression loses the specific texture of who this person is and how he got here. That texture is the part worth understanding.
Andrej Karpathy was born in 1986 in Bratislava, in what was then Czechoslovakia. He was fifteen when his family moved to Toronto. In his PhD dissertation, he thanked his parents for what he called their “leap of faith,” the sacrifice of comfort and familiarity for the possibility of something better. He was, he wrote, partly motivated by a determination to vindicate that bet and make them proud. That sentence explains a certain relentlessness that shows up throughout his career: the sense that squandering the opportunity would be a specific kind of betrayal, not just of himself but of people who gave something up so he could have it.
He discovered programming around that same age, fifteen, in Toronto. A few years later, as an undergraduate at the University of Toronto, he had what he described as a library epiphany. Standing among the stacks, he felt the vertigo of realizing how much there was to know and how little of it he could hold. “I’m a blob of soft tissue with finite, leaky memory, a slow CPU, and damnit I felt hungry.” That sentence tells you everything about his voice: precise, irreverent, honest about the absurdity of the situation. The experience pointed him toward what he called “the ultimate meta problem.” If human cognition is the bottleneck, maybe the work worth doing is on intelligence itself.
He had considered quantum computing first. He gave it up because he couldn’t get his hands dirty. That detail is worth sitting with.
At the University of Toronto, he walked into Geoffrey Hinton’s neural networks class around 2007. Neural networks were, at that moment, considered by most of the field to be a dead end. Hinton’s approach, training layered networks to learn representations directly from data, was out of fashion. Karpathy’s initial reaction was that the digit classification problems Hinton kept demonstrating were “cute but useless toys.” What changed his mind was Hinton’s way of talking about the networks. Hinton would use phrases like “in the mind of the network,” and Karpathy found himself intrigued by the implication: that something like thought could be happening inside a machine. He absorbed from Hinton a preference for iteration over elegance, for working systems over theoretical neatness. That preference would define everything that came after.
He moved to Stanford for a PhD under Fei-Fei Li. His dissertation was called “Connecting Images and Natural Language.” During his first year, he rotated through labs run by Daphne Koller, Andrew Ng, Sebastian Thrun. He absorbed across all of them and settled with Li, who was building ImageNet into the benchmark that would crack the field open. Karpathy became both a contributor to and a product of that work.
The ImageNet human benchmark experiment happened during this period. And then came CS231n: Convolutional Neural Networks for Visual Recognition. Karpathy, Li, and Justin Johnson built Stanford’s first dedicated deep learning course. It started with around 150 students in 2015 and grew to 750 by 2017. The lecture videos were watched millions of times. The course notes became a standard reference. A generation of practitioners learned to build neural networks from those materials. That is not a metaphor. That is a literal description of how the field grew.
Before the dissertation was finished, Karpathy published a blog post called “The Unreasonable Effectiveness of Recurrent Neural Networks.” He had trained a character-level model on Shakespeare, Wikipedia, and the Linux kernel source code. The Linux-trained version generated C code with proper indentation and correct bracket matching. Individual neurons inside the network, without being told to, had developed specialized behaviors: one tracked whether the model was inside a quoted string, another detected line position. Nobody programmed that. It emerged from the training. The post was a gateway for a lot of people into deep learning, written in plain language, generous with code, honest about what was surprising.
In December 2015, Karpathy was one of eleven people who co-founded OpenAI. He took an early internal poll on AGI timelines. His vote was twenty years, “very long compared to others,” he later noted. The early culture was small and idealistic, focused on open research. He was there at the beginning.
He left for Tesla in June 2017.
What Karpathy did at Tesla deserves more attention than it typically gets. He became Director of AI, later Senior Director, leading a team of around 300 engineers. He was building a real system, not a research demo. The system had to work on public roads, in every weather condition, at scale across millions of vehicles.
The central technical bet he led was the decision to go vision-only. In May 2021, Tesla began shipping vehicles without radar. Competitors were building sensor suites that combined lidar, radar, and cameras. Karpathy’s argument was direct: when radar and vision disagree, you have to choose which one to believe, and vision has fundamentally higher precision. “Vision is necessary and sufficient,” he said at a computer vision conference. Strip the redundancy and build vision that actually works.
The infrastructure behind that bet was called the Data Engine. Roughly one million vehicles on the road, each equipped with eight cameras. Over 200 trigger classifiers flagging edge cases. Clips flowing back, getting labeled, feeding into the next training run. Six billion labeled objects. 1.5 petabytes of data. 5,760 GPUs. When an engineer mentioned needing more GPUs even twice, Musk didn’t ask for a budget proposal. He called the cluster manager directly: double the cluster right now. When bureaucracy said six months for delivery, Musk called Jensen Huang at NVIDIA and the constraint disappeared.
He has been candid about what working for Musk was like. Most CEOs are remote from their engineers, several layers of management removed. Musk would spend half his time talking directly to engineers. The team was kept intentionally small. Karpathy described having to “plead” to hire additional people. Whether or not you agree with that philosophy, Karpathy executed within it and built something that demonstrably works.
He left Tesla in July 2022. He briefly returned to OpenAI in early 2023, then left again in early 2024. His exit came just three months after the chaotic Sam Altman ouster and reinstatement, so speculation was inevitable. His response: “nothing happened and it’s not a result of any particular event, issue or drama, but please keep the conspiracy theories coming as they are highly entertaining.”
During and after Tesla, he was writing. Three ideas in particular have spread through the builder community in ways that are still working themselves out.
The first is Software 2.0, published in 2017. Traditional software consists of explicit instructions a programmer writes by hand. Neural networks are different: the programmer defines an objective and provides data, and the training process finds the weights that produce the desired behavior. “Datasets are the new source code. Model weights are the new binaries. Gradient descent is the new compiler.” He later extended the framework to Software 3.0, where the programs are written in natural language. “The hottest new programming language is English,” he wrote in early 2023. I build with these tools every day, and that framing actually changes how you think about the work. You’re not configuring software. You’re communicating with a system that can interpret intent, not just instruction.
The second idea is the LLM OS. A large language model is not best understood as a chatbot. It is more like the kernel of an operating system: the context window is the RAM, stored information is the filesystem, tool calls are system calls, natural language is the shell. That reframing shifts the question from “what can this chatbot do” to “what kind of platform is this.”
The third is the Autonomy Slider, from a 2025 keynote. How much decision-making should the system have? Not a fixed setting. A continuous spectrum, adjusted based on context, stakes, and earned trust. He paired it with another framing: build Iron Man suits, not Iron Man robots. The suit gives the person wearing it capabilities they didn’t have. The robot replaces the person. “You can outsource your thinking,” he said, “but you can’t outsource your understanding.”
I keep coming back to that line. In February 2025, he fired off what he called a “shower of thoughts throwaway tweet” coining the term “vibe coding,” describing a way of building software where you talk to the model and forget the code even exists. Collins Dictionary named it their Word of the Year. By December 2025, he posted that he had gone from writing 80% of his own code to essentially none, in a single month. Fourteen million views. And then in February 2026, he released MicroGPT: a complete language model in 243 lines of pure Python, zero dependencies. “The culmination of a decade-long obsession to simplify LLMs to their bare essentials.” The person who stopped writing code turned around and wrote the most pedagogically stripped-down version of a language model anyone had published. He was still at the controls. A month later, he released autoresearch: one markdown prompt, 630 lines of code, one GPU. It ran 700 experiments in two days and discovered optimizations he had missed over years of hand-tuning. He described his state as “AI psychosis,” the anxiety of not having agents actively working.
There is a thread running through all of this that is easy to miss if you follow the headline version of his career. The thread is teaching.
Before any of the research, there was badmephisto. That was his YouTube channel handle as an undergraduate. The name came from Diablo II. He wanted “Mephisto” as a Hotmail username but every variation was taken. His English vocabulary at the time consisted of “about 20 words, one of which was bad.” Since the demon Mephisto was certainly pretty bad, he typed in “badmephisto” and it worked. He made Rubik’s Cube tutorial videos because the tutorials he wanted didn’t exist. His videos were detailed, patient, and unusually clear. The channel drew millions of views. Competitive speedcubers, including two-time world champion Feliks Zemdegs, have cited his tutorials. He was a college student who had just learned English.
Then CS231n. Then a YouTube series called “Neural Networks: Zero to Hero,” now at 1.34 million subscribers and over 27 million views. Then a set of open-source implementations, each stripping away one more layer of abstraction: micrograd in 300 lines. nanoGPT, now at 58,000 GitHub stars. llm.c, a language model in pure C. MicroGPT, 243 lines, zero dependencies. About that last one he wrote: “Everything else is just efficiency. I cannot simplify this any further.”
That sentence is the clearest window into how he thinks. Every one of those projects is asking the same question: what is the irreducible minimum? What do you have to understand to actually understand this?
In July 2024, he founded Eureka Labs. The pitch is an AI-native school where a teacher designs course materials and an AI teaching assistant helps guide students through them. The reference he reached for was Feynman: imagine working through difficult material alongside someone with deep knowledge who is there at every step. The first course teaches students to build a storytelling language model from scratch. The thing teaching you is a miniature version of the thing you’re building.
He had every opportunity to optimize for money, status, and proximity to power. He was at OpenAI at the founding. He was at Tesla at the peak of its ambition. What he kept returning to was not accumulation. It was simplification. It was teaching. It was finding the version of the thing that a smart person with no prior knowledge could actually understand and build on.
The builder who teaches is common enough. The teacher who builds is less common. But the person who treats building and teaching as the same act, as two expressions of the same compulsion to find the bottom of things, is genuinely rare. Right now, when the tools are changing faster than most people can track and the best practices are being invented in real time, that combination matters more than it usually does.
Most of the people building right now are doing so without a clear mental model of what they’re actually working with. Karpathy’s entire career has been the sustained attempt to produce those mental models and give them away. Not sell them. Not gate-keep them. Give them away.
The thing that will compound over the longest period is the teaching infrastructure he has been building since he was a college student who couldn’t find the Rubik’s Cube tutorial he needed, so he made it himself.