What AI really is (and what it isn’t)

If you run a business, you have already used AI, or someone on your team uses it without telling you. What almost nobody explains is what the thing that answers actually is. This module has no formulas and no code. It is the explanation you would have wanted before pasting your first contract into a chat window.

We will answer five questions: what a language model is, how it is built, why it sometimes makes things up, what it is good for today, and how to think about it so you get real work out of it. At the end there is a 15-minute exercise and a short quiz.

What is the difference between AI, machine learning, generative AI and an LLM?

Think of four circles, one inside the other. The further in you go, the more specific it gets.

1. 1Artificial intelligence (AI) is the big circle: any program that does something that used to require human intelligence. The spam filter in your inbox is AI. The GPS that reroutes you around traffic is AI.
2. 2Machine learning is how almost all modern AI is built. Instead of programming rules by hand, you show the system millions of examples and it adjusts its internal parameters until it gets things right. Nobody wrote the rules for "this is spam"; the system worked them out.
3. 3Generative AI is machine learning that produces new content: text, images, audio, code. It does not sort what already exists; it creates something that did not.
4. 4An LLM (large language model) is the kind of generative AI that works with text. It is what sits inside Claude, ChatGPT and Gemini, and it is the subject of this course.

How is a model like Claude built?

In two stages. The first is pretraining: the model is shown an amount of text that is hard to picture (books, web pages, code, articles) and given a single task, repeated billions of times: guess the next word. "The lease takes effect on the…" and the model proposes "date". If it is wrong, it gets adjusted a little. If it is right, that gets reinforced. Anthropic’s documentation describes it the same way: the model is pretrained to predict the next word given the text that came before.

The surprising part is how much you have to learn to play that game well. To finish a sentence about leases you need to have absorbed some law; to finish a recipe, some cooking; to finish code, some programming. The result is a system with vast general knowledge that still does not know how to hold a conversation or follow instructions. Type a question and it may answer with another, similar question, because on the internet questions tend to come in lists.

The second stage is post-training, and it is what turns that "text completer" into an assistant. People rate responses (this one is better than that one), and the model learns to prefer the ones people value. The technique is called reinforcement learning from human feedback, RLHF. For Claude, Anthropic also trains the model with a public document it calls Claude’s constitution: it describes the values and priorities Anthropic wants the model to have, and according to Anthropic it is a crucial part of the training process. It grew out of a technique the company has used since 2023 called Constitutional AI.

The analogy we will use throughout: pretraining is like someone who spent years reading everything they could get their hands on, with no particular goal. Post-training is the onboarding that turns that person into a useful employee: how to treat people, what to do and not do, when to say "I don’t know". The constitution is the values handbook for that onboarding.

What is a token, and why do you pay for them?

The model does not read words or letters; it reads tokens. A token is a chunk of text: sometimes a whole word, sometimes a syllable or a punctuation mark. "Lease" is probably one token; "unenforceable" is several. Anthropic estimates that in English a token is roughly three to four characters on its models. In Spanish, with longer words and accents, the same sentence usually costs more tokens.

Tokens matter for two practical reasons:

• Cost. When a company uses these models through the API (connected to its own systems), it pays per token in and per token out. A long document costs more than a short one. In the chat app you never see that bill, but the message limits on your plan come from the same place.
• Memory limit. Every conversation has a context window: the maximum number of tokens the model can hold in view at once. As of October 2026, the current Claude models have a one-million-token window, which Anthropic estimates at about 555,000 English words. That is a lot, but it is not infinite, and anything that does not fit, the model simply does not "see".

Why does AI make things up, and why does it change its answer?

This is the most important point in the module. A language model does not look facts up in a database. It produces the most plausible text given what it read during training and what is in the context window. When you ask for the article number of a law or a company’s phone number, it consults nothing: it generates an answer with the right shape. Often the right shape matches the right fact, because it saw that fact thousands of times. But when it does not know, it has no natural mechanism for staying quiet: it fills the gap with something that sounds right.

That is called a hallucination. Anthropic’s documentation defines it as text that is factually incorrect or inconsistent with the given context, and notes that even the most advanced models can produce it. The good news is that some techniques cut it down a lot: explicitly giving the model permission to say "I don’t know", asking it to quote the document you pasted, and giving it tools to search instead of recall. You will see all three in modules 2, 3 and 4.

Two consequences you need to be clear on:

• Knowledge cutoff. The model learned from text up to a certain date and nothing after. As of October 2026, the current Claude models have a reliable knowledge cutoff of June 2026. Ask about anything later without web search and it will answer from what it knew, or make something up.
• Same question, different answer. The model picks each token with an element of chance (sometimes the most likely option, sometimes the second). That is why asking the same thing twice gives two different texts. Anthropic notes that even with that randomness turned down to its minimum, results are not fully deterministic. It is not a defect; it is what lets the model offer variations. But it means a fact that came out once is not a verified fact.

Is Claude a model or an app?

Both, and it helps to keep them apart. The model is the engine: the trained program that turns input text into output text. The product is what you actually use: the chat app, with its conversation history, projects, web search, connectors to Gmail or Drive, and paid plans.

In Anthropic’s case, as of October 2026, there are four current models: Claude Fable 5.1 (the most capable, for demanding reasoning and long tasks), Claude Opus 5.5 (the recommended choice for most work), Claude Sonnet 5.5 (the balance of speed and capability) and Claude Haiku 5.5 (the fastest and cheapest, for high-volume jobs like classifying or extracting data). The names change with every generation, so the thing worth remembering is the logic: there is a big, slow model, a middle one and a small, fast one, and you pick based on the task.

The same split applies elsewhere: ChatGPT is OpenAI’s app, and inside it run models from the GPT family; Gemini is both the name of Google’s app and of its models. When someone says "ChatGPT got it wrong", the useful question is: which model, with what context, with which tools turned on. This course uses Claude as its main example, but almost everything carries over.

What is AI good at today, and what is it bad at?

This list is as of October 2026 and describes what a language model does on its own, without tools. Further down we explain how tools change the list.

What it does very well

• Drafting and rewriting. An email to a supplier, an internal policy, a reply to a complaint, in whatever tone you ask for.
• Summarizing. Forty pages of building regulations into one page of key points, or a thirty-message email thread into three lines.
• Extracting data. From a PDF invoice to a table with date, tax ID, amount and VAT. From a WhatsApp conversation to a list of orders.
• Classifying. Sorting five hundred reviews into positive, negative and complaints by topic. Flagging which emails are urgent.
• Translating. Spanish, English and dozens of other languages, with a feel for register (formal, friendly, technical).
• Thinking with you. Playing devil’s advocate, listing scenarios, asking what you did not think of, proposing twenty names for a product.
• Writing and explaining code. This is what opens the door to parts 3 and 4 of the course.

What it does badly, or should not do alone

• Exact arithmetic without tools. Adding a column of a hundred figures "in its head" is exactly the kind of task where plausible text is not a correct result. With a calculation tool (running code) the picture changes completely.
• Recent facts without search. Anything after its knowledge cutoff, or anything that changed this week: prices, regulations, who runs which company.
• Decisions with consequences. Letting someone go, approving credit, replying to a notice from the tax authority. It can prepare the analysis; the decision and the signature are yours.
• Anything that requires accountability. AI does not answer to a customer, a judge or a regulator. If something goes wrong, the person who sent it is responsible.
• Knowing what you did not tell it. It does not know your margins, your difficult client or the exception you make on Fridays. That is the subject of module 2.

Much of the second list is fixed with tools: web search for recent facts, code execution for calculations, connectors so it reads your email or calendar instead of guessing. When a model can use tools and decide when to use them, it starts being called an agent. Part 2 of the course shows how to use them without code, and part 4 explains how agents work under the hood. The last two limitations (consequences and accountability) are not fixed with tools; they are fixed with a process in which a person reviews.

How should I think about AI to get real work out of it?

This is the mental model for the whole course: AI is a brilliant new hire who has read half a library and knows nothing about your business. It writes better than most people, knows law, accounting and marketing, works at any hour and never gets tired. But it started today. It does not know what your products are called, what you promised Monday’s customer, the tone you use with suppliers or why the building invoice gets paid differently.

You would never tell a new hire "write me the proposal" and walk away. You would explain the client, show them an earlier proposal you liked, tell them what not to promise, and read the draft before it went out. With AI it is exactly the same: the quality of what comes out depends almost entirely on what goes in. Most disappointments with AI are, in reality, one-line instructions given to someone who had no way of knowing what you know.

I run a [WHAT YOUR BUSINESS DOES] business in [CITY] with [NUMBER] people. My most repetitive tasks are [TWO OR THREE TASKS].

Explain what [CONCEPT: for example "a token", "the context window", "a hallucination", "an agent"] is, using my business as the example, in 150 words or fewer and without jargon.

Then give me one concrete example of how that concept affects one of the tasks I mentioned, and one question I should be asking myself about it.

What is the 4D AI Fluency framework?

So that we do not stop at "give it good context", the course follows the AI Fluency framework that Anthropic publishes with professors Rick Dakan (Ringling College) and Joseph Feller (University College Cork). It defines fluency as interacting with AI systems in ways that are effective, efficient, ethical and safe, and breaks it into four skills, the "4Ds":

1. 1Delegation. Setting goals and deciding whether, when and how to engage with AI. Which tasks you hand over, which you keep, which you share.
2. 2Description. Effectively describing goals to prompt useful AI behaviors and outputs. This is what people call "prompting", but it is broader: context, format, success criteria.
3. 3Discernment. Accurately assessing the usefulness of AI outputs and behaviors. Reading critically, catching inventions, knowing when the draft is done.
4. 4Diligence. Taking responsibility for what we do with AI and how we do it: data, transparency with customers and staff, who signs.

The first two happen before and during a task; the last two, during and after. You will see them throughout the course: Description and Discernment are the heart of part 1; Delegation drives part 2, where you decide which tasks and processes to hand to AI; and Diligence runs through everything, weighing most in parts 3 and 4, when you build systems that act on your behalf. Anthropic’s original course is free and is linked in this module’s sources. This course does not replace it; it grounds it in a Colombian business.

I am [YOUR ROLE] at a [WHAT IT DOES] business in [CITY]. I will list the tasks my week goes into. For each one, put it in one of four categories and explain why in one line:

A. AI can produce a first draft and I review it (writing, summaries, data extraction, classification).
B. AI helps me think but the decision is mine (analysis, options, risks).
C. It needs a separate tool (web search, calculation, reading my email) or data I have not organized yet.
D. Not delegated: it requires my judgment, my signature, or handling other people’s sensitive data.

My tasks:
[LIST OF 8 TO 15 TASKS, ONE PER LINE, WITH HOW LONG EACH TAKES]

At the end, tell me the three category-A tasks you would start with this week and what information from me you would need to do them well.

How does this course work?

There are four parts. Part 1 is for anyone who uses AI at work: context, tools, prompting and privacy. Part 2 brings it into your daily work without code: projects, connectors, automations, verification and bringing it to your team. Parts 3 and 4 are for people who want to build, with Claude Code and what is known as agentic engineering. They are written so that anyone can read them, even with no plan to program, because understanding how things get built helps you decide what to commission and from whom.

Every module has definitions to remember, a real example of how we use the idea at 77Rentals, prompts you can copy, an exercise of about 15 minutes, a checklist and a quiz. The next module is called "Context is everything", and it is where you stop writing one-line instructions.