An agent is, at its core, an LLM making decisions in a loop. Before any of that will make sense — why a model sometimes hallucinates, why prompts matter so much, why context windows have limits, why tool calls work the way they do — you need a real mental model of what an LLM is and how it actually works under the hood.
This module builds that model from the ground up. It starts with a light, intuitive overview of what large language models are and how they're trained, then gets concrete with tokens and embeddings — how raw text becomes the numbers a model can actually compute on. From there it covers the transformer architecture that every modern LLM is built on, and zooms into attention, the mechanism at its heart that lets a model weigh which words matter to which. It closes with a full, honest tour of how these models are trained end to end — pretraining, fine-tuning, and the security issues that come with the territory.
Watch the lessons in order, then continue on to AI Agents 101.
A light, visual intro to what an LLM actually is: predicting the next word, chatbots, pretraining, and a first look at transformers — the fastest way to build real intuition before going deeper.
How words become numbers a model can work with: one-hot encoding, why it falls short, and how word embeddings place similar words closer together in vector space — the same idea behind tokens and embeddings in every LLM.
A visual walkthrough of the transformer architecture behind GPT-style models: tokenization, embeddings, attention blocks, and how a model turns all of that into a prediction for the next token.
A deep dive into self-attention, the mechanism that lets a transformer decide which parts of the input matter most to each word — the single idea that made modern LLMs possible.
A comprehensive, general-audience tour of the full LLM pipeline: pretraining and fine-tuning, scaling laws, tool use and multimodality, and the security issues — jailbreaks, prompt injection, data poisoning — you'll need to know before building agents.