What AI is and how an answer is born

AI from scratch: your digital assistant · Lesson 1 / 20

What AI is and how an answer is born

The "AI" you chat with is a large language model (LLM). It was trained on a vast amount of text to do one thing: predict which word comes next. That sounds modest — but everything else grows from it: conversation, emails, translation, explanations, ideas, even code.

It's worth understanding this mechanism in your first five minutes. People who don't spend years puzzled: "why does it lie?", "why doesn't it know about yesterday's news?", "why are the answers so generic?" All three answers live right here.

An analogy: an extremely well-read companion

Picture someone who has read a giant library — books, articles, forums, manuals — and remembers not the pages but the patterns: which words and ideas tend to follow which. You ask a question; they don't flip through a reference book, they assemble the most plausible answer from memory. Fast. Confidently. And sometimes confidently wrong.

The model doesn't "recall a page." It assembles the most plausible continuation of your text on the fly.

Three consequences to accept right away

A knowledge cutoff. The model knows the world roughly up to when it was trained. It may not know about last week at all — or it may start guessing, which is worse.
Confidence ≠ correctness. The tone is equally firm whether it's right or inventing. It has no inner sense of doubt the way a person does.
No telepathy. It sees exactly what you wrote in this chat. Not your situation, not your client, not your preferences — until you spell them out.

What every answer is assembled from

How an answer is born: prompt, context and the model's "experience" → result.

Each answer comes from three sources:

Your prompt — what you just wrote. The one factor you control completely.
Conversation context — earlier messages in this same chat. The model sees and uses them.
The model's training — the "experience" baked into its parameters. You can't change it, but you can steer it with your wording.

Notice: two of the three sources are in your hands. That's exactly why the skill of phrasing a request has such a strong effect — and why a good result isn't magic.

Example: why the same question gives different results

Tell me about Paris

Paris is the capital of France, one of Europe's major cultural centres… (half a page of encyclopedia-style facts)

I'm going to Paris for 2 days with a 7-year-old. Plan day one: 3–4 places close together, no all-day museums.

Morning: Eiffel Tower — views plus the carousel nearby. Midday: picnic on the Champ de Mars. Evening: a Seine boat ride, kids usually love it…

The model's knowledge was identical both times. All that changed was how much useful context you handed it. That's the skill we build across this whole course.

Insight. The model works with what you wrote, not what you meant. This is the master key to the entire course.

Common mistake. Assuming AI "searches the internet" or "stores a database of facts." By default it does neither — which is where invented details come from.

Pro tip. If an answer comes out watery, don't blame the model — ask yourself which specifics you left out. Nine times out of ten, that's the problem.

Cheat sheet

An LLM predicts text continuation; it doesn't "know the truth."
Its knowledge has a cutoff date.
A confident tone ≠ a correct answer.
Answer = your prompt + chat context + the model's training.

1. What is a language model actually doing when it answers?

2. Why might the model not know about last week's events?

3. Which of these do you control completely?

🔒 Answer the question correctly to move on to the next lesson.