Karpathy’s free 2h11m LLM lecture is being recast as a crash course on agents, loops, harnesses, and self-improving systems

Andrej Karpathy’s free, 2-hour-11-minute YouTube lecture How I use LLMs is circulating again as a practical map of how language models become agents: not as a single prompt, but as loops, a surrounding harness, and systems that can keep improving without a person at the keyboard. The official video, posted on his channel as a general-audience follow-up to his training-focused deep dive, walks through the LLM product stack he actually uses—models and pricing tiers, thinking modes, search and deep research, files, code interpreters, Cursor, voice, images, video, memory, and Custom GPTs.