Three identities
Karpathy’s résumé is often reduced to “OpenAI founding member + Tesla AI Director,” but those two titles conceal the identity that has remained constant throughout: teacher.
In 2015 he opened CS231n at Stanford—the school’s first deep-learning course. For a long time, the course notes were the default material for learning computer vision around the world. Before he became an executive at any company, he had already proved that he could make complex things clear.
What five years at Tesla gave him
In 2017 he joined Tesla to lead the vision system for Autopilot. Those five years occupy a special place in his intellectual structure: they turned him from a “researcher” into someone who knew how far research and product were from each other.
Autonomous driving is one of the few fields where the gap between a convincing demo and dependable mass production is enormous—between getting a demo to work and putting it on the road lies a long stretch of unglamorous reliability engineering. The reason his later judgments about the limits of AI were more restrained than most people’s lies here.
After leaving, he did something counterintuitive
After leaving Tesla in 2022, he did not start a company immediately, nor did he join one of the AI companies whose valuations were soaring. He began teaching for free on YouTube, moving from hand-written backpropagation to building a GPT from scratch, while open-sourcing minimal implementations such as nanoGPT.
nanoGPT’s design philosophy says a great deal: it does not optimize for peak performance, but for helping people truly understand what GPT is doing in a few hundred lines of code. That is a teaching mindset, not an engineering mindset.
Two years passed without commercial return, but they accumulated one of the densest stores of trust in the industry.
Eureka Labs: productizing his method
In July 2024 he founded Eureka Labs, an “AI-native school.” The central question is not how to build a stronger model, but how to deliver education at scale: excellent teachers will always be scarce, so can AI teaching assistants multiply the reach of one top teacher by several orders of magnitude?
This entry point is almost a direct product of his own experience. He spent a decade testing how many people one good teacher can change, then tried to turn that work from a craft into an industry.
The power of naming
“Software 2.0” (2017) and “vibe coding” (2025) are examples eight years apart, but the mechanism is identical: give an accurate, easy-to-spread definition to a phenomenon that is happening but still has no name.
Neither term is complicated, and neither is an entirely original insight—many people had already sensed the change. But only after a phenomenon is named can it be discussed, cited, and written into a business plan. This is the cheapest lever in a technical person’s hands, and the one most often wasted.