AX-100 · Lesson 1 of 12 · Module 1: How Assistants Work · 3:00

What an assistant really does

An AI assistant runs on a language model that writes by predicting the most likely next piece of text (a token), again and again. It does not look answers up unless a tool such as web search is switched on. Each pick has some chance in it, so the same question gets different wording and different invented details. Where a prompt leaves a gap, the model fills it with what usually goes there, so a fluent answer is not a checked one. Worked example: Sam’s spring-sale post gains a date, a discount and free coffee that nobody gave it.

Frame from lesson 1 of Sophyx Certified AI Fundamentals, "What an assistant really does"

What you will learn

  • Explain how an assistant builds an answer
  • Explain why answers vary between tries
  • Spot details an answer made up

The one thing to remember

It writes what’s likely, not what’s true.

Common mistake: Treating a fluent answer as a checked one

Fluent text is what prediction produces; it says nothing about whether the details are true.

The fix: Give it the facts it needs, and check every detail it adds.

Key terms

Prompt
What you type to tell the assistant what you want.
Language model
Software that writes by predicting likely next words.
Token
A small piece of text, often part of a word.

Transcript

The brands in the worked examples are fictional, and the numbers are illustrative.

Sam asked an assistant for one short post about the spring plant sale.

It came back with a date, a discount and free coffee that nobody had mentioned.

The sale is on the twenty-fifth, and there's no coffee. So where did those details come from?

Welcome to AI Fundamentals. This first lesson shows what an assistant really does.

You'll learn how an assistant builds an answer, why answers vary between tries, and how to spot details it made up.

First, a question to think about. When you ask an assistant something, does it look the answer up?

Usually it doesn't, and this lesson shows you what it does instead.

What you type is called a prompt. Behind the assistant, a language model reads it and starts to write.

It writes one small piece of text at a time, called a token, like a very well-read autocomplete.

At join us on Saturday, it weighs the likely next words, and the eleventh of April wins.

The real date, the twenty-fifth, was never among the options, because nobody gave it.

Then it moves on to the next gap, and twenty percent off sounds likely too.

Each pick has a little chance in it, so the same prompt can get a different answer every time.

In a new chat, it says this Sunday, the nineteenth, with twenty-five percent off everything.

Every invented detail changed, but the cheerful tone didn't. That's fine for ideas, but a warning for facts.

Here's a quick check. In Sam's first post, where did twenty percent off come from?

It was a likely-sounding guess. Sam never gave a discount, and nothing was searched.

Where your prompt leaves a gap, the model fills it with what usually goes there.

Sam gave no date and no offer, so each gap got a typical guess.

Fluent is not the same as true, so Sam marks every detail that never came from Greenhollow.

A common mistake is treating a fluent answer as a checked one.

Instead, give it the facts it needs, and check every detail it adds.

This week, ask one work question three times in new chats, and mark what changes and every detail you never gave it.

Before we wrap up, can you name the three big ideas from this lesson?

It writes likely next words, answers vary between tries, and gaps get filled with likely details.

So remember, it writes what's likely, not what's true.

Now try the practice quiz, and then join Sam in lesson two.

Part of AI Fundamentals (AX-100)

For anyone at work new to AI: how assistants answer, how to ask well, how to check the answer and keep data safe.

12 lessons, practice quizzes, a final exam and the Sophyx Certified AI Fundamentals certificate. Regular price USD 49.99. Your first certification is free. Free with any paid Sophyx plan.

Start the course free
Book a Demo