Dwarkesh Podcast Andrej Karpathy — AGI is still a decade away
10186 snips
Oct 17, 2025 
Andrej Karpathy, a leading AI researcher and former Tesla Autopilot head, shares insights on the future of artificial general intelligence (AGI) and education. He discusses why AGI will likely take another decade to mature, highlighting the inefficiencies of reinforcement learning compared to other methods. He critiques the slow progress of self-driving technology, attributing it to safety requirements. Karpathy also emphasizes the importance of integrating AI into education, proposing a model that combines expert faculty and AI assistance for personalized learning.
AI Snips
Chapters
Books
Transcript
Episode notes
LLMs Excel At Boilerplate, Not Novel Architecture
- Current coding models help most with boilerplate and common patterns, not unique, tightly-structured repos.
- For intellectually intense code, human architects still outperform LLMs.
RL 'Sucks Supervision Through A Straw'
- Reinforcement learning is extremely noisy because credit is assigned to entire trajectories from sparse rewards.
- Humans instead review, distill, and assign partial credit — something RL lacks.
Favor Process Supervision Over Sparse Rewards
- Use process-level supervision instead of only final rewards to teach models stepwise behavior.
- Beware using LLM judges naively because models will find adversarial shortcuts.
