Andrej Karpathy
AI researcher, engineer, and educator. Known for making deep technical concepts in machine learning accessible to general audiences.
Background
- Co-founder of OpenAI (2015).
- Director of AI at Tesla; led the Autopilot computer vision programme.
- PhD from Stanford (2015) under Fei-Fei Li; dissertation on convolutional neural networks for visual recognition.
- Created the widely followed
neural networks: zero to heroYouTube series. - Coined the term vibe coding (early 2025).
Talks in this wiki
| Title | Date | Topic |
|---|---|---|
| Intro to Large Language Models | Nov 2023 | Busy person’s one-hour intro to LLMs; LLM OS framing |
| Deep Dive into LLMs like ChatGPT | Feb 2025 | Full pipeline — pretraining through security |
| How I Use LLMs | Feb 2025 | Practical ecosystem tour; tools, workflows, habits |
| From Vibe Coding to Agentic Engineering | 2025 | Software 3.0; vibe coding vs agentic engineering |
| Andrej Karpathy on AGI Timelines, RL's Limits, and the Future of Education | Oct 2025 | Dwarkesh interview; decade-of-agents, RL’s limits, education |
Recurring themes across talks
LLMs as compression. Karpathy consistently frames pretraining as lossy compression of the internet. The parameters store statistical patterns, not facts; recall is vague and probabilistic.
Jagged intelligence. A recurring preoccupation: LLMs are brilliant in most domains, with arbitrary, unexplained gaps — the reversal curse, the strawberry R-count, the car-wash question. The gaps reflect both structural limits (verifiability) and what labs chose to invest in.
Verifiability as the axis. From the RL discussions in ‘Deep Dive’ to the founder advice in ‘Vibe Coding’: progress clusters where results can be verified. This explains why maths and code improved fastest, and it points practitioners toward tractable opportunities.
Tool use as amplification. In every practical talk, the model alone is weaker than the model plus tools. Web search, Python interpreter, code execution: these convert LLMs from recall engines into problem-solving systems.
The LLM as a new computing substrate. From ‘LLM OS’ (2023) to ‘Software 3.0’ and the ‘neural computer’ (2025): Karpathy frames LLMs not as a tool within existing computing but as the foundation of a new computing paradigm.
The decade, not the year. In the 2025 Dwarkesh interview he resists the industry’s short timelines: agents are intern-grade, reinforcement learning is brittle (‘sucking supervision through a straw’), models collapse and lack machine culture, and the payoff diffuses into the existing 2% GDP exponential rather than arriving as a step function. The deficits echo his Jagged Intelligence preoccupation; his answer to them is education.
Key concepts originating from or emphasised by Karpathy
- Vibe Coding — coined by Karpathy, Jan 2025.
- Agentic Engineering — his framing for professional-quality agent-assisted development.
- Software 1.0, 2.0, 3.0 — his three-paradigm framework.
- Jagged Intelligence — his term for the Swiss-cheese capability profile of LLMs.
- LLM OS — the operating-system analogy introduced in 2023.