Andrej Karpathy
Core Stance & Frontier Insights
Software engineering has shifted from roughly 80/20 hand-coding versus delegation to 20/80 and beyond, making token throughput and human orchestration more important than keystrokes. AutoResearch found training improvements beyond Karpathy’s manual tuning when objectives were measurable, pointing toward agents that optimize experiments, research procedures, and eventually organizational instructions. Autonomy remains strongest where outcomes are verifiable, while judgment, humor, security, and physical-world feedback remain unresolved risks as cheaper software could expand demand even while frontier researchers automate themselves away. Frontier Thesis: Software development has inverted to 80% agentic delegation, shifting the critical bottleneck from writing code to human orchestration and token throughput. Measurable loops like AutoResearch outperform human tuning, yet full autonomy remains constrained.
Strategy: Bet on incremental capability gains and continuous compute scaling, while positioning education as an essential hedge against human de-skilling as workflows automate.
Risks & Warnings: Achieving reliable agents is a decade-long grind facing autonomous-driving-style deployment friction. Current systems behave like summoned ghosts rather than robust animals—plagued by reward hacking, fragile out-of-distribution reasoning, and a lack of lifelong learning.
Curated Podcasts & Talks
Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
- 🗓️ Date：
2026-03-20｜ 🎙️ Show：No Priors
Software engineering has shifted from roughly 80/20 hand-coding versus delegation to 20/80 and beyond, making token throughput and human orchestration more important than keystrokes. AutoResearch found training improvements beyond Karpathy’s manual tuning when objectives were measurable, pointing toward agents that optimize experiments, research procedures, and eventually organizational instructions. Autonomy remains strongest where outcomes are verifiable, while judgment, humor, security, and physical-world feedback remain unresolved risks as cheaper software could expand demand even while frontier researchers automate themselves away.
View Dialogue Notes & Transcript Memo
Interview Summary & Key Takeaways: Software engineering has shifted from roughly 80/20 hand-coding versus delegation to 20/80 and beyond, making token throughput and human orchestration more important than keystrokes. AutoResearch found training improvements beyond Karpathy’s manual tuning when objectives were measurable, pointing toward agents that optimize experiments, research procedures, and eventually organizational instructions. Autonomy remains strongest where outcomes are verifiable, while judgment, humor, security, and physical-world feedback remain unresolved risks as cheaper software could expand demand even while frontier researchers automate themselves away.
- 🔗 Original source & video: Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
- 📝 Transcript Status: Core insights and takeaways summarized. Full runtime is approx 45-90 mins.
View Dialogue Notes & Key Takeaways
Key Takeaways: Software engineering has shifted from roughly 80/20 hand-coding versus delegation to 20/80 and beyond, making token throughput and human orchestration more important than keystrokes. AutoResearch found training improvements beyond Karpathy’s manual tuning when objectives were measurable, pointing toward agents that optimize experiments, research procedures, and eventually organizational instructions. Autonomy remains strongest where outcomes are verifiable, while judgment, humor, security, and physical-world feedback remain unresolved risks as cheaper software could expand demand even while frontier researchers automate themselves away.
- 🔗 Original source & video: Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
Andrej Karpathy — “We’re summoning ghosts, not building animals”
- 🗓️ Date：
2025-10-17｜ 🎙️ Show：Dwarkesh Podcast
Karpathy’s differentiated call is a “decade of agents”: current systems lack continual learning and multimodality, while coding agents proved “not net useful” on nanochat and struggled with code never written before. He expects three or four or five more RL updates while saying current compute may be absorbed rather than overbuilt; deployment’s march of nines and timeline miscalibration remain watchpoints.
View Dialogue Notes & Transcript Memo
Interview Summary & Key Takeaways: Karpathy’s differentiated call is a “decade of agents”: current systems lack continual learning and multimodality, while coding agents proved “not net useful” on nanochat and struggled with code never written before. He expects three or four or five more RL updates while saying current compute may be absorbed rather than overbuilt; deployment’s march of nines and timeline miscalibration remain watchpoints.
- 🔗 Original source & video: Andrej Karpathy — “We’re summoning ghosts, not building animals”
- 📝 Transcript Status: Core insights and takeaways summarized. Full runtime is approx 45-90 mins.
View Dialogue Notes & Key Takeaways
Key Takeaways: Karpathy’s differentiated call is a “decade of agents”: current systems lack continual learning and multimodality, while coding agents proved “not net useful” on nanochat and struggled with code never written before. He expects three or four or five more RL updates while saying current compute may be absorbed rather than overbuilt; deployment’s march of nines and timeline miscalibration remain watchpoints.
- 🔗 Original source & video: Andrej Karpathy — “We’re summoning ghosts, not building animals”