No Priors/Pod/26/03/20 Andrej Karpathy

• Person/Andrej Karpathy here
  • Projects like Micrograd, NanoGPT, and now MicroGPT are all part of a long-running obsession to compress LLMs down to their conceptual essence.
  • The core insight: the actual algorithm behind LLM training is surprisingly small (~200 lines of Python). Most real-world complexity is just engineering for speed and scale, not conceptual necessity.
  • Historically, he would explain these systems directly to humans (videos, walkthroughs), but that’s changing.
  • Now his shift is:
    • Don’t explain to humans → explain to agents
    • If an agent understands the code, it can translate, teach, and adapt explanations infinitely better than a static tutorial
  • This leads to a new model of education:
    • Instead of docs for humans → docs for agents (e.g., markdown)
    • Instead of lectures → “skills” that guide how an agent should teach something
    • The agent becomes the interface layer between knowledge and the learner
  • His role changes too:
    • Humans contribute the hard, irreducible insights (like designing MicroGPT itself)
    • Everything else (explanation, onboarding, teaching) gets outsourced to agents
  • The punchline:
    • “The things agents can do, they’ll do better than you. The things they can’t do — that’s your job.”