scripod.com
Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

Overview

Shownote

Highlights

Transcript

Chapters

Pins

Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

No Priors: Artificial Intelligence | Technology | Startups

Mar 20
No Priors: Artificial Intelligence | Technology | Startups

No Priors: Artificial Intelligence | Technology | Startups

Mar 20
OverviewShownoteHighlightsTranscriptChaptersPins

Shownote

What happens when AI agents can design experiments, collect data, and improve — without a human in the loop? Andrej Karpathy joins Sarah Guo on the state of models, the future of engineering and education, thinking about impact on jobs, and his project Aut...

Highlights

Andrej Karpathy discusses his evolving relationship with AI, from delegating most of his coding to agents to exploring autonomous research. He shares insights on the current limits of AI, the future of engineering, and the potential impact on jobs and education.
00:00
Delegating most coding to AI agents
04:54
Shift from compute-bound to token-bound
10:19
AI agent controls lights, HVAC, shades, pool, spa, and security
13:29
AI agents can replace many custom apps
17:58
AI agents autonomously tuned hyperparameters better than he could manually
25:23
AI struggles with nuanced requests like humor.
28:33
Performance varies sharply across tasks.
35:00
Verification is cheap but generation is expensive.
44:15
Outside labs, one can be a free agent, aligned with humanity.
50:59
Centralization of intelligence carries systemic risks.
53:54
Physical robotics will lag behind digital AI.
1:03:12
Documentation will be written for agents rather than humans
1:05:40
Focus on what agents cannot do

Chapters

Andrej Karpathy Introduction
00:00
What Capability Limits Remain?
02:55
What Mastery of Coding Agents Looks Like
06:15
Second Order Effects of Natural Language Coding
11:16
Why AutoResearch
15:51
Relevant Skills in the AI Era
22:45
Model Speciation
28:25
Building More Collaboration Surfaces for Humans and AI
32:30
Analysis of Jobs Market Data
37:28
Open vs. Closed Source Models
48:25
Autonomous Robotics
53:51
MicroGPT and Agentic Education
1:00:59
Conclusion
1:05:40

Transcript

Sarah Guo: Code's not even the right verb anymore, right? But I have to express my will to my agents for 16 hours a day. Manifest. Andrej Karpathy: How can I have not just a single session of plot code or codex, or some of these agent harnesses? How can I...