Transcript
Five hundred thousand users in a single month. That's what Eureka Labs pulled in after its February 2026 beta launch — and Andrej Karpathy, the founder, former Tesla AI Director, and OpenAI founding member, built it on a single premise: the best human tutors are locked behind wealth and geography, and AI can unlock them for everyone. On March 25, 2026, Eureka Labs shipped adaptive AI tutors to that half-million user base. The education arms race just changed shape. While the previous lecture focused on robotics and data collection challenges, this lecture shifts to AI's transformative role in education, particularly through boutique AI. Eureka Labs addresses the challenge of high student-to-teacher ratios by offering personalized AI tutors, making elite education accessible. The model is one unique AI tutor per student — what Karpathy calls boutique AI — not a mass-produced chatbot, but a personalized agent that tracks your specific knowledge gaps and adapts in real time. Eureka Labs' integration of multimodal AI allows students to engage interactively with subjects like physics, transforming passive learning into active exploration. The AI tutor identifies where understanding breaks down, not by quiz score alone, but by analyzing response patterns, hesitation signals, and conceptual drift across sessions. That's a fundamentally different diagnostic than any standardized test. Karpathy envisions AGI-like capabilities in education through agent swarms, where specialized agents collaborate to tailor learning experiences. Here's where Karpathy's own workflow becomes the proof of concept, Sergey. He no longer writes code himself. He spends hours directing AI agents instead — a state he describes as being in AI psychosis, an intense, almost obsessive focus on agentic workflows. His team achieved a 90% reduction in human oversight for AI research by March 2026. That's not a productivity gain. That's a structural redefinition of what a small team can accomplish. Eureka Labs runs lean by design; Karpathy envisions a boutique team where a handful of people, amplified by agents, outperform organizations ten times their size. The agentic workflow differs from traditional AI interaction in one critical way: it's not a single prompt and response. It's a loop. The agent plans, executes, evaluates, and revises — autonomously. For education, this means the AI tutor doesn't wait for you to ask a question, Sergey. It proactively identifies the gap, constructs the right scaffold, and closes the loop before confusion compounds. Quality of interaction, not quantity of parameters, is the design philosophy — a direct departure from the compute arms race that defines most frontier AI labs. The challenge Karpathy acknowledges is scaling human expertise without diluting it. A great tutor's value is judgment — knowing when to push, when to simplify, when to reframe. Encoding that judgment into an agent requires rich behavioral data from expert teachers, which is scarce, much like the teleoperation data problem in robotics. Eureka Labs' answer is to start narrow: domain-specific agents trained on curated expert interactions, then expand. It's the same playbook as foundation models — start with quality, then scale. AI-native education transcends digitized content, offering dynamic, personalized mentorship through teacher-plus-AI agents accessible globally. Karpathy is betting that the same agentic loop that closed 90% of human oversight in AI research can close the gap between the world's best education and the world's least-served learners. That is the boutique AI future — not one model for everyone, but one agent for each person.