Foundations · Chapter 4 of 11
How does AI actually work?
A plain-language explanation of what large language models actually do — no jargon, just the mental model you need to use them well.
You don't need to be a technologist. But understanding the mechanics changes how you use it.
Trained on language
These models process massive volumes of text to learn patterns and structures. They're conversational tools, not just standard databases.
Generates by predicting
AI doesn't search databases. It predicts the most likely next word at extraordinary speed, constructing responses from learned patterns.
Multi-modal & reasoning
Modern models take more than text — images, documents, audio, even video. The latest can also pause to "reason" through a problem step by step before answering.
How a search engine works
- You type a question
- It scans an index of existing web pages
- It ranks and returns links to pages others wrote
- You read the source yourself and decide what's relevant
How an LLM works
- You type a prompt
- It draws on patterns absorbed from billions of texts during training
- It constructs an original response word by word
- You get a direct answer — but there's no single source page behind it
The response is constructed, not copied: an AI draws on patterns absorbed from billions of pages. It gives a considered answer, but it is speaking from patterns it has learned — not reading from a specific page. Its training also has a cut-off date, so it has no built-in knowledge of recent events unless you give it that information or connect it to live search. It might blend sources or misremember a detail. That flexibility is its strength, and also why it sometimes gets things wrong.