Skip to content

Andrej Karpathy

AI researcher and educator. Co-founder of OpenAI (offices described as the original OpenAI space in SF), former head of AI/Autopilot at Tesla, now independent — runs Eureka Labs and ships open pedagogical code (nanochat, micrograd, etc.). Coined "vibe-coding" (2024) and the follow-up framing "agentic-engineering" (2026).

Shape of his current thinking (2026 Sequoia talk)

  • Felt more behind, not less. December 2025 was a phase shift for him — chunks of agent-generated code started "just coming out fine" and he kept asking for more. Implication: the delta between AI-native and non-AI-native engineers is widening, not shrinking.
  • Software 1.0 → 2.0 → 3.0. See software-3-0. Programming is now context-engineering over an LLM interpreter.
  • LLMs are jagged. See jagged-intelligence. Peaky in verifiable domains (math/code), rough elsewhere — partly because frontier labs RL against verifiable rewards.
  • Animals vs ghosts. See animals-vs-ghosts. We are summoning ghosts — shaped by data + reward, not evolution — not building animals. Framing matters for how you deploy and trust them.
  • The bottleneck is you, the director. Even with agents "filling in the blanks," the human still owns taste, design, and what's worth building. Hence his enthusiasm for llm-knowledge-bases as cognitive scaffolding.
  • Education thesis: "You can outsource your thinking, but you can't outsource your understanding." Informs Eureka Labs' direction.

Running projects / artifacts referenced

  • nanochat / micrograd / menu-gen — minimalist training/demo stacks; he notes LLMs hate the simplification direction (they bloat code instead of shrinking it).
  • LLM knowledge bases — wiki-as-synthetic-data-generator over articles he reads. (llm-knowledge-bases)
  • MenuGen blog post — used throughout the talk as the anchoring example of agentic-engineering friction (Vercel deploy hell, Stripe↔Google email mismatch, auth plumbing).

Contrasts / positions

  • Against the "10x engineer" framing — thinks the ceiling of agentic engineering is much higher than 10x for those who invest in tooling and setup.
  • Skeptical that today's agent stack substitutes for real understanding; the UI/docs world is still written for humans and needs to become agent-native (sensors + actuators over the world).