Intro to Large Language Models (1-hour talk)
by Andrej Karpathy
The best hour anyone can spend building a correct mental model of what an LLM actually is.
Overview
Andrej Karpathy recorded this 59-minute-23-second talk on 23 November 2023 from a presentation he gave at the AI Security Summit, and it remains the single best hour for a non-specialist to build a correct mental model of large language models. It runs in three parts. The first, to about the 25-minute mark, establishes what an LLM is at the file level — a parameters blob plus the code to run it — then covers inference, the cost and scale of pretraining, fine-tuning into an assistant, and the honest framing that the model is 'dreaming' plausible documents rather than looking anything up. The second part, to roughly 45 minutes, is his forecast: scaling laws, tool use, multimodality, System 1 versus System 2 reasoning, self-improvement, and the 'LLM OS' framing in which the model is the kernel process of an emergent operating system with the context window as RAM and tools as peripherals. The final quarter is security — jailbreaks, prompt injection and data poisoning — which is why the talk was given at a security summit at all. Slides are published as PDF and Keynote. Two things to know going in: it assumes no mathematics and no code, which is exactly what makes it shareable with non-engineers, and it predates reasoning models and production agent loops, so treat the middle section as a 2023 forecast rather than a description of the present.
At a Glance
- Topic
- Models
- Level
- Beginner
- Format
- Video
- Cost
- Free
- Duration
- 59 minutes, single sitting
- Provider
- Andrej Karpathy
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓What an LLM is concretely: a weights file plus the code to run it
- ✓How pretraining and assistant fine-tuning differ in cost, data and purpose
- ✓Why hallucination follows from next-token prediction rather than being a bug
- ✓The 'LLM OS' framing with context as RAM and tools as peripherals
- ✓Where scaling laws, multimodality and System 2 reasoning were heading in 2023
- ✓The main LLM attack surfaces: jailbreaks, prompt injection and data poisoning
Highlights
- •Karpathy built training pipelines at OpenAI and led Tesla's Autopilot vision team
- •No maths and no code — the primer you can send to a non-engineer colleague
- •Slides published as PDF and Keynote, so it doubles as a briefing template
- •Still spawning derivative work: a 2026 Show HN interactive guide built on it hit 245 points
- •Treats security as a first-class topic, not an afterthought appendix
Who It’s For
Best For
- ✓Engineers who need one correct end-to-end mental model before going deeper
- ✓Tech leads briefing non-technical stakeholders on what LLMs actually are
- ✓Anyone starting AI work who wants the big picture before the tooling
Prerequisites
- •None — the talk explicitly assumes no mathematics and no programming
- •Awareness that it was recorded in November 2023 and predates reasoning models
FAQ
What is Intro to Large Language Models (1-hour talk)?
Andrej Karpathy's 59-minute general-audience talk, recorded November 2023 at the AI Security Summit, giving a complete mental model of large language models in three parts: what they are and how they are trained, where capability is heading under his 'LLM OS' framing, and the security failure modes. No maths or code required. You finish able to reason about LLM behaviour instead of guessing at it.
Is Intro to Large Language Models (1-hour talk) free?
Intro to Large Language Models (1-hour talk) is free to access.
What level is Intro to Large Language Models (1-hour talk) for?
Intro to Large Language Models (1-hour talk) is aimed at a beginner audience. Recommended background: None — the talk explicitly assumes no mathematics and no programming, Awareness that it was recorded in November 2023 and predates reasoning models.
How long does Intro to Large Language Models (1-hour talk) take?
Expect roughly 59 minutes, single sitting. Most learners work through it at their own pace.
What will I learn from Intro to Large Language Models (1-hour talk)?
You'll learn: What an LLM is concretely: a weights file plus the code to run it; How pretraining and assistant fine-tuning differ in cost, data and purpose; Why hallucination follows from next-token prediction rather than being a bug; The 'LLM OS' framing with context as RAM and tools as peripherals; Where scaling laws, multimodality and System 2 reasoning were heading in 2023; The main LLM attack surfaces: jailbreaks, prompt injection and data poisoning.
Topics
Sources
This page was written from 3 sources, 2 on domains other than youtube.com.