Deep Dive into LLMs like ChatGPT
by Andrej Karpathy
The whole LLM training stack in one sitting, demonstrated live in real tools.
Overview
Deep Dive into LLMs like ChatGPT is Andrej Karpathy's 3-hour-31-minute walkthrough of the entire LLM training stack, published to YouTube on 5 February 2025 and described by him as a general-audience deep dive. Karpathy was a founding member of OpenAI and Director of AI at Tesla, and this video is the successor to his earlier Intro to LLMs talk. It proceeds in the order the models are actually built. The first hour covers pretraining: where the data comes from, with FineWeb opened and scrolled on screen; byte-pair tokenization demonstrated live in Tiktokenizer; what the network takes in and emits; the internals shown through bbycroft.net's 3D transformer visualizer; then inference and a GPT-2 training run. He then loads a Llama 3.1 base model on Hyperbolic to show what a raw pretrained model does before any alignment. The second hour covers post-training — conversation data and chat templates — followed by long sections on hallucination, tool use, knowledge versus working memory, a model's knowledge of itself, why models need output tokens to think, and why tokenization makes spelling and character counting hard. The final hour moves from supervised finetuning to reinforcement learning, then DeepSeek-R1, the AlphaGo analogy, and RLHF including where reward models get gamed. Nearly every claim is demonstrated on screen rather than asserted, and the timestamped chapters make it usable as a reference.
At a Glance
- Topic
- Models
- Level
- All Levels
- Format
- Video
- Cost
- Free
- Duration
- 3 hours 31 minutes, single video with timestamped chapters
- Provider
- Andrej Karpathy
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Trace the full pipeline from raw crawled web data to a deployed assistant model
- ✓Explain byte-pair tokenization and why it breaks spelling and character-counting tasks
- ✓Describe what a base model actually does before any post-training or alignment
- ✓Understand hallucination as an artifact of training data rather than a mystery
- ✓See why reasoning models need output tokens as working memory to compute anything
- ✓Distinguish supervised finetuning, reinforcement learning and RLHF and when each applies
- ✓Recognize why RLHF reward models can be gamed and are not true reinforcement learning
Highlights
- •Taught by Andrej Karpathy — founding OpenAI member and former Director of AI at Tesla
- •Everything demonstrated live in real tools: FineWeb, Tiktokenizer, bbycroft's visualizer, Hyperbolic, LM Arena
- •Reached 582 points on Hacker News and spawned widely-shared community TL;DR write-ups
- •Timestamped chapters make it a reference you return to, not a one-time watch
- •Free with no signup, covering ground most paid courses split across several modules
Who It’s For
Best For
- ✓Engineers who use LLM APIs daily and want to understand what sits underneath
- ✓Anyone building intuition before starting a finetuning or reinforcement learning project
- ✓Technical people who need to explain LLM behaviour and limits to non-specialists
Prerequisites
- •None — Karpathy builds up from first principles for a general technical audience
- •Some prior exposure to neural networks makes the middle hour easier to follow
FAQ
What is Deep Dive into LLMs like ChatGPT?
Andrej Karpathy's three-and-a-half-hour walkthrough of how models like ChatGPT are actually built, from web-scale pretraining data through tokenization, inference, post-training, reinforcement learning and RLHF. Afterwards you can explain why an LLM hallucinates, why it cannot spell reliably, why reasoning models emit long outputs, and what actually separates a base model from an assistant.
Is Deep Dive into LLMs like ChatGPT free?
Deep Dive into LLMs like ChatGPT is free to access.
What level is Deep Dive into LLMs like ChatGPT for?
Deep Dive into LLMs like ChatGPT is aimed at a all levels audience. Recommended background: None — Karpathy builds up from first principles for a general technical audience, Some prior exposure to neural networks makes the middle hour easier to follow.
How long does Deep Dive into LLMs like ChatGPT take?
Expect roughly 3 hours 31 minutes, single video with timestamped chapters. Most learners work through it at their own pace.
What will I learn from Deep Dive into LLMs like ChatGPT?
You'll learn: Trace the full pipeline from raw crawled web data to a deployed assistant model; Explain byte-pair tokenization and why it breaks spelling and character-counting tasks; Describe what a base model actually does before any post-training or alignment; Understand hallucination as an artifact of training data rather than a mystery; See why reasoning models need output tokens as working memory to compute anything; Distinguish supervised finetuning, reinforcement learning and RLHF and when each applies; Recognize why RLHF reward models can be gamed and are not true reinforcement learning.
Topics
Sources
This page was written from 3 sources, 2 on domains other than youtube.com.