
Wes Roth Andrej Karpathy GPT - Advice for building AI agents
23 snips
Nov 11, 2026 Andrej Karpathy, a computer scientist and AI researcher known for work at OpenAI and Tesla, explores why early AI agents struggled, how language models revived the field, and whether AGI will emerge as one agent or many. He discusses the decades-long challenge of building reliable products, neuroscience-inspired designs for memory and planning, and why agents remain an exciting frontier.
AI Snips
Chapters
Books
Transcript
Episode notes
OpenAI’s Early Agents Couldn’t Use Websites
- In 2016, Andrej Karpathy built World of Bits to teach reinforcement-learning agents computer tasks instead of games.
- Agents clicked through simple flight-booking and food-ordering pages, but stumbling toward rewards failed because the technology was immature.
Language Models Replaced Reinforcement Learning
- Agent development changed completely once language models became capable enough to serve as the foundation.
- Modern agent builders generally use language models rather than the reinforcement learning that dominated Andrej Karpathy’s early experiments.
AGI May Become Digital Civilizations
- AGI may look less like one artificial mind and more like agents cooperating in organizations or civilizations.
- This possibility makes agent research compelling because it connects individual capabilities with large-scale digital societies.



