No Priors: Artificial Intelligence | Technology | Startups Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
3491 snips
Mar 20, 2026 
Andrej Karpathy, AI researcher and engineer known for OpenAI and Tesla, dives into coding by directing agent teams, his AutoResearch project for autonomous research loops, and why natural language may replace many app interfaces. He also explores model speciation, open vs. closed models, shifting job markets, robotics, and agent-driven education.
AI Snips
Chapters
Books
Transcript
Episode notes
AutoResearch Removes The Human From Model Tuning
- AutoResearch aims to remove the human researcher from the optimization loop and let agents run longer autonomously.
- Karpathy says an overnight run found better GPT-2 tuning than he had by hand, including weight decay on value embeddings and Adam beta adjustments.
Research Orgs Can Be Tuned Like Software
- Karpathy sees research organizations as code expressed in markdown roles, queues, and procedures that can themselves be optimized.
- Different ProgramMDs could encode risk tolerance, workflow, and coordination, then compete for improvement like any other system.
AI Progress Is Strong But Deeply Jagged
- Karpathy says autonomous loops work best where outcomes are objectively measurable and verifiable, such as kernel optimization or training loss.
- He argues model intelligence remains jagged: the same system can act like a brilliant programmer yet still tell stale jokes from years ago.
