L03. How AI puts an answer together

You will distinguish supplied facts from guesses in an AI answer. Allow about 30 minutes for the explanation and cup example. Prepare this book and paper; you do not need to open a chat yet.

In class

The teacher offers several endings to the sentence about Olena’s cup. Together, separate plausible wording from known facts, then discuss how training helps a model produce an answer.

Training happens before our conversation

Imagine recognising a cup in a photograph. Cups vary in colour, angle and lighting, so listing every appearance is impractical. During machine learning, examples help adjust internal numerical settings called parameters. The system gradually learns useful patterns.

One main training task for a language model is predicting text from its surroundings. Many examples teach language relationships, styles and ways of completing tasks. Further training helps it follow instructions. [F210]

When you send a message, the model uses its existing settings and your text. A clarification changes the material available for the next reply; it does not instantly retrain the model. Whether conversations contribute to future training depends on the service’s policies.

Where the next word comes from

Complete this sentence: “Olena put the cup on the…” You might suggest “desk” or “shelf”. If the previous sentence says she is organising her workspace, one choice may seem more appropriate. The surrounding information helps you choose.

A model calculates likely continuations using tokens: words, word parts or symbols. Each selected piece joins the text already produced, and the process continues. This simplifies the mechanism; modern systems can add reasoning steps and tool use. [F211]

A fluent continuation need not be true. Without information about the cup’s position, the model cannot establish it. Invented detail presented as fact is a model hallucination.

[[VIS:f5_model_operation]]

Model, application and agent

A model produces a response. An application provides the message field, buttons and conversation history; it may offer several models. ChatGPT is therefore a product name, not the name of one fixed model.

An agent combines a model with a cycle of actions: receive a task, choose a tool, inspect its result and continue. It might use a calculator to work out a total. Opening a file requires access to that file.

On paper, try separating known information from an assumption. Here are the starting facts:

Olena has a cup, a notebook and a phone.
She wants to make space for work.
The cup's current position is not given.

A suitable instruction for a future chat is: “Suggest a tidying sequence. Do not describe the objects’ starting positions as known facts.” Moving the cup somewhere safe could be a useful suggestion. “The cup is beside the laptop” is an unsupported statement: no laptop was mentioned.

> [FIX] An answer adds detail absent from your material. Identify it directly: “The cup’s location is unknown. Remove that claim and keep the general suggestion.”

Choose by the result you need

For an explanation or draft, start with ChatGPT, Claude or Gemini; recent facts need search and opened sources. Codex, Claude Code or Antigravity can work on website files with folder access. An image tool creates or edits pictures; speech recognition turns audio into text for subtitles. These remain different tasks even within one application. The content lessons introduce design in Canva and subtitles in Clipchamp.

Home practice

• Write three true details about another object, such as a book: its title, cover colour and language. Do not state where it is lying.
• Write two sentences: one using only those facts, and another adding a guess about its location. On a sheet headed “Fact and guess”, underline the unsupported detail and explain what information is missing. Your check should catch it even if it sounds plausible.

Self-test

• Are training and answering a message the same?
• The assistant places the cup beside a laptop, although neither its location nor a laptop was supplied. Is this a fact or a guess?
• ChatGPT lets you choose different models. Why does the application’s name not identify one model?

Words you will need

• Parameters: the numerical settings of a trained model.
• Generation: producing an output from the current request.
• Token: a unit used by the model to process data.
• Hallucination: false or unsupported content presented as fact.