Explain in plain terms what a model is: something that turns inputs into outputs using settings it learned.
People who build AI use the word "model" all the time, and it sounds more mysterious than it is. Once you can picture what a model does, a lot of what you read about AI gets easier to follow. That includes why an assistant sometimes gives an answer nobody can explain, and why it doesn't pick up new facts by itself.
A model is something that takes an input and produces an output by following settings it learned from examples.
Auntie Lim, an invented example, runs a chicken rice stall in a hawker centre in Toa Payoh. Every evening she decides how many chickens to buy for the next day. She has never written a formula, but she carries one in her head.
She looks at three things. The first is the day of the week: weekdays are busier than Sundays because of the office crowd. The second is the weather, since rain keeps people at home and lowers sales. The third is how close it is to payday, because the week before payday is quieter. She weighs these together and arrives at a number: buy twenty chickens.
Machine learning models work like her mental formula. Only the inputs and outputs are different:
A spam filter takes an email as its input and gives a label, spam or not spam, as its output. Face recognition on your phone takes a camera image and outputs yes (this is the owner) or no. A language model takes the text so far and outputs a prediction of the next word or part of a word. ChatGPT, Claude and Gemini are built on language models, and Module 3 covers them in detail.
Mathematicians call anything that turns inputs into outputs a "function". You don't need the maths, only the shape: the input goes in, the model calculates, and the output comes out.
Picture Auntie Lim's reasoning as a machine with three dials on the front. One sets how much a weekday adds. One sets how much rain takes away. One sets the payday effect. If you turn the rain dial up, the machine buys fewer chickens on wet days. With all three dials set well, it buys about the right number on most days.
In a real model these dials are called parameters. Each parameter is just a number, and the calculation combines the input with all of these numbers to produce the output. A small model, for example one predicting chicken sales, might have a handful. Models that recognise faces or write text have millions. The large language models behind today's assistants have billions. One parameter does very little by itself, and the behaviour comes from all of them working together.
A model's skill lives entirely in its parameter values. Two models with the same design but different parameter settings behave differently. Two stallholders with the same stall but different instincts will buy different amounts of chicken in the same way.
With three dials, a person could set them by trial and error. With billions, nobody can. Parameters are set by training. The model is shown many examples, and an automatic process adjusts the numbers bit by bit. This continues until the outputs are right more often. When training stops, the parameters are fixed and saved as a file together with the model's design. Lesson 2.2, "Learning means guessing, measuring the miss and adjusting", walks through how this works.
The training process found the settings and no person chose them, so the people who built a large model cannot point to one parameter and say what it does. Researchers can study a model from the outside, and some work on looking inside it. Even so, nobody can fully explain why a model produced one particular output. When a company says it does not know exactly why its assistant gave a strange answer, that is usually the plain truth.
A trained model is a file that holds all the numbers, plus the design that says how to combine them. When you use an assistant, that file sits on the company's computers. Your input is run through the calculation using the fixed numbers, and the output comes back to you. The numbers do not change.
The model does not learn while it is being used. If you ask the same question tomorrow, it starts from the same settings. If the world changes, the model knows nothing about it until it is trained again, or until the new information is fed into the conversation. Module 5 covers feeding new information into the conversation so a model can use facts it was not trained on.
Many people believe an assistant is quietly learning from every chat. The underlying model is not. Some products save notes about you and show them to the model next time, but the parameters stay unchanged.
Auntie Lim's instincts did not come from a manual. They came from years of evenings. When she bought too many chickens, she threw some away and adjusted. When she bought too few, she sold out by one o'clock and adjusted. A model is built the same way, only faster and with far more examples.
A model takes inputs, runs a calculation using learned numbers, and gives an output. Its skill is in those numbers. Training sets them automatically from examples, and once training ends they stay fixed. Being able to say that clearly to someone who has never thought about it is a good test of whether you understand it, and that is what the activity asks you to do.
Write two sentences explaining what a model is, in words a parent or friend with no tech background would understand. Then read them out to that person.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).