Lesson 1: What Is AI, Really?

A note on this series. This is written for beginners, on purpose. I’m keeping explanations at a practical, conceptual level rather than a technical one. Several ideas here are described in generic terms rather than their full technical depth. The goal of these seven lessons is to demystify AI agents enough that the hype becomes understandable and ideally to spark enough curiosity that you go looking for the deeper technical material afterward.

Before talking about agents, I need to define what I mean by “AI” in the first place. The word gets used for everything from spam filters to self-driving cars, and that looseness is a big part of why the hype cycle feels so confusing. So I’m starting at the bottom.

The model is a prediction engine

Most of the AI you’ve heard about in the last few years, ChatGPT, Claude, Gemini, is built on something called a large language model, or LLM. Strip away the marketing and a modern language model is a very large statistical system trained primarily to predict the next token in a sequence. Those tokens can be pieces of words, whole words, punctuation, or other pieces of structured input. During generation, the model repeatedly predicts what should come next, producing the text that we see.

That next-token prediction is the core mechanism, but what emerges from it is considerably more complicated than ordinary autocomplete. Models can represent patterns, relationships, and structures in ways that allow them to produce surprisingly capable reasoning, coding, and problem-solving behavior. None of that means they have human-like beliefs or desires, but it does mean ‘autocomplete’ can be a misleadingly simple picture of what is happening inside the model. Anthropic’s own explainer on language models is a good next stop if you want more depth on this.

Training is where the “knowledge” comes from

An LLM learns by being shown huge volumes of text, books, articles, code, conversations, and adjusting itself millions of times until its predictions get good. The main training process is expensive and happens before a model is deployed. Once a particular model version is deployed, its parameters normally don’t update themselves from each conversation. But companies can later fine-tune, retrain, or replace models with newer versions, and some systems can also use information from conversations through separate memory or retrieval systems.

A model’s training data has a cutoff: information that wasn’t included in its training isn’t part of its pre-trained knowledge. But that doesn’t mean the system using the model is permanently stuck in the past. Search, retrieval, APIs, uploaded files, and other tools can give an agent access to information that appeared after the model was trained.

Prompting is how you talk to it

A prompt is just the text you type in. The model doesn’t have a separate “understanding module” that parses your intent, then a separate “answering module” that writes a response. It’s one continuous process. Your prompt becomes the beginning of a sequence, and the model keeps predicting what comes next, which are its replies.

That matters. Vague prompts get vague answers because there’s less pattern to latch onto. Specific prompts narrow the prediction down to something more useful, context, examples, a clear ask. This is the entire basis of what’s called “prompt engineering,” and it stops being mysterious once you see the model as a very sophisticated autocomplete rather than a mind reader.

Where the “intelligence” actually shows up

Language models turn out to be surprisingly general-purpose. The same next-word-prediction mechanism that writes a poem can also write working code, summarize a legal document, or explain a medical term, because all of those are just different patterns of text the model saw during training. That generality is why LLMs became the foundation for almost everything built after them, including agents.

But a plain LLM has real limits. A model operating on its own can’t check today’s weather, look up a live stock price, or send an email. Those capabilities require access to external data or tools. It can’t remember what you told it five conversations ago unless something outside the model stores that for it. On its own, a model is a closed box: text goes in, text comes out, and nothing in the world changes.

That gap, between a model that can only talk and a system that can do something, is where the idea of an “agent” starts. MIT Sloan’s explainer on agentic AI frames this as the shift from tools that answer questions to systems that act on their own. That’s the line I’m about to cross.

Where the model gets it wrong

One limit matters more than the rest: a language model will sometimes state something false with the same confidence it uses for something true. This is usually called hallucination, though fabrication is probably the more accurate word. The model has no concept of lying. It’s the same next-word-prediction mechanism producing a plausible sounding sequence that happens not to match reality. A made-up citation, a fabricated statistic, a court case that doesn’t exist, all come from the same process that produces correct answers. Nothing in the mechanism itself distinguishes a confident correct answer from a confident wrong one.

This matters more, not less, as models get better because fluent wrong answers are harder to catch than obviously wrong ones. Build the habit early: verify anything the model tells you that you plan to rely on, especially names, numbers, dates, and sources. Keep this in mind as the series moves into agents that act on what a model outputs. A wrong answer in a chat window is an inconvenience. A wrong answer an agent acts on is a different problem entirely.

What to carry into Lesson 2

An LLM predicts text. It doesn’t inherently “do” anything in the outside world.
Training happens once and is frozen; prompting is how you use what’s already there.
The model’s generality is what makes it useful as a foundation, but usefulness alone doesn’t make it an agent.
A model can state something false as confidently as something true. Verify before you rely on it.

In the next lesson, I’ll draw the line between a chatbot and an agent, and it’s a sharper line than the marketing suggests.