Deep Dive into LLMs like ChatGPT
by Andrej Karpathy
Karpathy's 3.5-hour walkthrough of the entire LLM training stack, from pretraining to RLHF.
Overview
Deep Dive into LLMs like ChatGPT is a 3-hour-31-minute lecture published on YouTube in February 2025 by Andrej Karpathy (former Tesla AI director and OpenAI founding member). It walks through the complete lifecycle of a large language model: pretraining data collection and tokenization, the neural-network (Transformer) architecture and how inference works, then the post-training stages — supervised fine-tuning (SFT) and reinforcement learning, including reinforcement learning from human feedback (RLHF) and reasoning-style RL. Along the way Karpathy builds intuition for why models hallucinate, how knowledge is stored in parameters versus retrieved via tools, the 'psychology' and failure modes of LLMs, and practical implications for using them effectively, with concrete demonstrations referencing models like GPT-2 and Llama 3.1. It is accessible to a general audience while remaining technically substantive, making it one of the best single resources for a grounded end-to-end understanding of modern LLMs.
At a Glance
- Topic
- Models
- Level
- All Levels
- Format
- Video
- Cost
- Free
- Duration
- ~3h 31m video
- Provider
- Andrej Karpathy
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How LLM pretraining works: data, tokenization, and the Transformer network
- ✓The post-training stack: supervised fine-tuning, RLHF, and reasoning RL
- ✓Why models hallucinate and how knowledge lives in weights vs tools
- ✓The 'psychology' and failure modes of LLMs and how to use them effectively
- ✓How pieces like GPT-2 and Llama 3.1 illustrate the full training pipeline
Highlights
- •From Andrej Karpathy — one of the most respected LLM educators
- •End-to-end coverage of the training stack in a single 3.5-hour session
- •Bridges intuitive explanation and real technical depth; free on YouTube
Who It’s For
Best For
- ✓Engineers who want a complete mental model of how LLMs are built
- ✓Practitioners moving from using LLMs to understanding their training
- ✓Anyone onboarding to LLM/ML work who wants an authoritative overview
Prerequisites
- •Curiosity about how LLMs work; no coding required
- •Basic ML familiarity helps but is not required
FAQ
What is Deep Dive into LLMs like ChatGPT?
Andrej Karpathy's ~3.5-hour YouTube deep dive (Feb 2025) into how LLMs like ChatGPT are actually built — a comprehensive, general-to-technical tour of the full training stack for anyone who wants a real mental model of how these systems work.
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: Curiosity about how LLMs work; no coding required, Basic ML familiarity helps but is not required.
How long does Deep Dive into LLMs like ChatGPT take?
Expect roughly ~3h 31m video. Most learners work through it at their own pace.
What will I learn from Deep Dive into LLMs like ChatGPT?
You'll learn: How LLM pretraining works: data, tokenization, and the Transformer network; The post-training stack: supervised fine-tuning, RLHF, and reasoning RL; Why models hallucinate and how knowledge lives in weights vs tools; The 'psychology' and failure modes of LLMs and how to use them effectively; How pieces like GPT-2 and Llama 3.1 illustrate the full training pipeline.