Part 1: LLM Fundamentals
This section covers what’s actually happening inside a large language model, from raw text to generated output. By the end, terms like “attention head,” “softmax,” and “embedding dimension” will mean something concrete to you, not just buzzwords.
We’ll build up the concepts in four steps:
How LLMs convert text into numbers, and why that representation matters for fine-tuning.
The architecture behind every modern LLM. Attention, feed-forward networks, and how they fit together.
From probability distributions to actual text. Temperature, top-k, top-p, and why generation is slow.
Pre-training vs. fine-tuning vs. alignment. The 10,000-foot view before we get hands-on in Part 2.
Prerequisites for this section: You should be comfortable with basic linear algebra concepts (vectors, matrices, dot products). If “multiply two matrices together” makes sense to you, you’re good. We’ll explain everything else as we go.