Deep Dive into LLMs like ChatGPT

Andrej Karpathy explains how large language models move from web-scale pretraining through supervised fine-tuning and reinforcement learning to become conversational assistants, while clarifying tokenization, hallucination, tools, and model limitations.

5 sections ยท 20 lessons

Course outline

Training Pipeline

  1. Three Training Stages
  2. Internet Data Curation
  3. Tokenization Fundamentals
  4. Neural Network Training
  5. Base Models Explained
  6. Text Generation Process

Post-Training

  1. Supervised Fine-Tuning
  2. Conversation Tokenization

Limitations

  1. Hallucinations
  2. Tool Use Integration
  3. Computational Limits
  4. Tokenization Blind Spots

Reinforcement Learning

  1. Reinforcement Learning Basics
  2. RL in Math & Code
  3. Chain of Thought
  4. RLHF for Subjective Tasks
  5. RLHF Limitations

Using and Evaluating LLMs

  1. Mental Model of AI
  2. LLM Ecosystem
  3. Future Capabilities

Start learning with Wondering