Inside GPTs: Understanding Large Language Models
summary
Full pipeline from raw web to deployed assistant with concrete numbers: raw filtered Common Crawl becomes 44 terabytes of training text, tokenized into ~50k vocab historically and up to 200,000 tokens for GPT5/6, so English compresses far better than Macedonian (28 vs 60 tokens example); pretraining predicts next-token over trillions of tokens, then supervised fine-tuning with human conversations and RLHF to teach “I don’t know” and prefer answers; techniques include sparse activation (2 billion active of 500 billion), repeating layers to think more, large context windows and agents that execute APIs and manage context.
AI-written summary · about
chapters (14)
- 0:00Introduction and overview
- 1:06Pre-training the model
- 5:44Training the neural network
- 8:36Inference and token prediction
- 10:24Post-training and fine-tuning
- 13:18Reinforcement learning from feedback
- 16:50Knowledge in parameters vs context
- 21:05Training process and thinking steps
- 23:56Reinforcement learning details
- 25:42How models process language
- 27:08Extreme thinking and repetition
- 28:33What is an agent
- 31:03Agent training and risks
- 33:51Future of agents and safety
