1.8 — Summary

Here is the whole chapter on a single page. Keep it handy for a quick refresher.

Key ideas

• LLMs transformed NLP. Earlier approaches relied on hand-written rules and statistical tricks. LLMs replaced them with deep-learning models that understand, generate, and translate human language at a level nobody had seen before.
• Two-stage training. Modern LLMs are first pretrained on a huge unlabelled corpus using next-word prediction, then finetuned on smaller labelled data to follow instructions or classify text.
• Transformers power them. The transformer's attention mechanism gives the LLM selective access to the whole input sequence while generating output one word at a time.
• Encoder vs decoder. The original transformer has both. Text-generation LLMs like GPT-3 and ChatGPT keep only the decoder, which keeps the architecture simple and scaled.
• Scale matters. Pretraining needs billions of words. Here we'll implement training on small datasets and also learn how to load openly available model weights.
• Emergent behaviour. Despite being trained for "just" next-word prediction, GPT models can classify, translate, summarise, and more.
• Finetuning wins. A finetuned foundation model often beats a general-purpose LLM on a narrow task.
• Build your own. Custom LLMs unlock privacy, lower latency, and full control.

Mental model to take forward

Pretraining on giant text  →  Foundation model
                                   │
                                   ├─ Instruction finetune  →  Chatbot / assistant
                                   └─ Classification finetune →  Spam / sentiment / etc.

That's the entire shape of modern LLM engineering, and the sections ahead walk through every box on the diagram in PyTorch.

Up next

Chapter 2 — Working with Text Data — covers how raw text becomes the tokens and embeddings that actually feed a neural network. It's the necessary bridge between the bird's-eye view we just took and the hands-on code waiting in Chapter 3.