What Is AI, Really?

So What Is AI, Actually?

You have probably heard people throw around the term "AI" like it explains everything and nothing at the same time. Self-driving cars? AI. Chatbots? AI. That weird photo filter that ages your face? Also AI, apparently.

But the AI we are going to work with throughout this course is a specific kind: large language models, or LLMs. And they are both simpler and stranger than most people think.

Here is the most honest way to describe what an LLM does: it predicts the next word.

That is it. Seriously. When you type a message to Claude or ChatGPT, the model reads what you wrote and then generates a response one word (well, one token) at a time, each time asking itself: "Given everything so far, what word is most likely to come next?"

The Library Analogy

Imagine someone who has read every book in the largest library in the world. Every novel, every textbook, every manual, every blog post, every forum thread. Billions and billions of pages.

Now imagine you walk up to this person and say, "Hey, can you write me a marketing email for a cat cafe?"

They have never run a cat cafe. They have never sent a marketing email. But they have read thousands of marketing emails and thousands of things about cats. So they can produce something that looks and sounds exactly like what you asked for, because they have seen the patterns so many times.

That is essentially what an LLM does. It was trained on an enormous amount of text from the internet, books, code repositories, and more. It learned patterns: how sentences are structured, how arguments are built, how code follows syntax rules, how emails are formatted. When you ask it something, it draws on all those patterns to generate a plausible response.

Pattern Recognition, Not Understanding

This is the most important thing to internalize early: LLMs do not "understand" things the way you do. They recognize and reproduce patterns. When an LLM writes a correct Python function, it is not because it understands programming logic the way a software engineer does. It is because it has seen millions of Python functions and learned the patterns of what correct code looks like.

This distinction matters because it explains both why AI is shockingly good at some things and surprisingly bad at others.

What AI Can Do (Really Well)

Let us be concrete about the things LLMs genuinely excel at:

• Writing and editing text — Drafts, emails, summaries, rewrites, tone adjustments. This is where LLMs feel almost magical.
• Answering questions — Especially when the answer exists somewhere in their training data.
• Writing code — From simple scripts to complex applications, LLMs can generate functional code across dozens of programming languages.
• Brainstorming — Need 20 ideas for a product name? 10 ways to restructure your team? LLMs are tireless brainstorming partners.
• Explaining concepts — They can break down complicated topics into simple language (which is basically what this lesson is doing right now).
• Translating and reformatting — Between languages, between formats (turn this CSV into JSON), between styles (make this formal email casual).
• Analyzing and summarizing — Give it a long document and ask for the key points. This works remarkably well.

What AI Cannot Do (Yet)

Now the part most AI hype ignores:

• Access real-time information — Unless specifically connected to the internet or tools, LLMs only know what was in their training data. They do not know today's weather or yesterday's stock price.
• Do math reliably — This surprises people. LLMs can look like they are doing math, but they are actually pattern-matching, not calculating. They frequently get arithmetic wrong, especially with large numbers.
• Remember previous conversations — Each conversation starts fresh unless you are using a tool with memory features. The AI does not remember what you told it last Tuesday.
• Guarantee accuracy — This is the big one. LLMs can and do produce confident, well-written, completely wrong answers. This is called hallucination, and we will talk about it more in the safety lesson.
• Reason about novel situations — If a problem requires genuine reasoning about something truly new (not a variation of something in the training data), LLMs struggle.
• Replace human judgment — They can inform your decisions, but they should not make them for you. Especially for anything high-stakes.

The Key Vocabulary

Before we go further, let us nail down the terms you will hear constantly throughout this course. You do not need to memorize these right now; they will become second nature as you practice.

How Tokens Actually Work

Let us dig a little deeper into tokens because they affect how you use AI in practice.

When you send a message to Claude, your text gets broken into tokens before the model processes it. The word "hamburger" might become three tokens: "ham", "bur", "ger". Simple words like "the" or "cat" are usually one token each. A number like "2847" might be one or two tokens.

Why does this matter? Because every model has a context window — a maximum number of tokens it can handle at once. That window has to fit everything: your prompt, any documents you uploaded, the conversation history so far, and the response the model is generating.

Claude currently has one of the largest context windows available at about 200,000 tokens, which is roughly 500 pages of text. That is a very big desk. But it is still finite, and being mindful of it will make you a better AI user.

Temperature: The Creativity Dial

Temperature is a setting you will not interact with directly very often, but understanding it helps you make sense of AI behavior.

At low temperature (close to 0), the model almost always picks the most likely next token. This produces consistent, predictable, focused responses. Great for factual questions, code generation, and tasks where accuracy matters.

At high temperature (close to 1 or higher), the model is more willing to pick less likely tokens. This produces more creative, varied, and sometimes surprising responses. Useful for brainstorming, creative writing, and when you want diverse options.

Most AI tools set a reasonable default temperature for you. But if you ever notice that an AI is giving you the same answer every time and you want more variety, or it is being too "wild" and you want it to settle down, temperature is the concept behind that behavior.

The "Stochastic Parrot" Debate

You might hear critics call LLMs "stochastic parrots" — fancy words meaning "random repeaters." The argument is that LLMs are just remixing their training data without any real understanding.

There is truth in this critique, and it is healthy to keep it in mind. But here is the practical reality: whether or not an LLM truly "understands" anything is a philosophical question. What matters for us is whether it can help us get useful work done. And the answer to that is a resounding yes — as long as you know its limits.

Throughout this course, we will treat AI as a powerful tool: incredibly useful when used well, potentially misleading when used carelessly. That is the right framing.

Putting It All Together

Here is your mental model for the rest of this course:

AI is a pattern-matching engine with a massive training set. It is extraordinarily good at producing text that matches the patterns it learned. It does not understand, reason, or know things the way humans do. But its pattern-matching ability is so powerful that for many practical tasks — writing, coding, brainstorming, analyzing — it produces results that are genuinely useful.

Your job as an AI user is to learn how to give it the right patterns to match against (that is what prompting is), how to verify its outputs (that is what critical thinking is), and how to integrate it into your work in ways that make you more productive without making you sloppy.

That is what the rest of Rangitoto is all about.