The Sommelier

A person can learn to do something well without being able to explain the rule they follow.

A trainee sommelier begins with wild guesses about the price of a wine. She tastes a bottle, names a price, hears the correct answer, and discovers how wrong she was. After repeating that process many times, her guesses improve, even on wines she has never tasted before.

What changed?

In this chapter, you are going to build and train the smallest model that can make one possible answer visible.

The model receives four numbers describing a wine. Each input has an adjustable number called a weight, which decides how strongly that input affects the price. Multiply each input by its weight, add the results, and the model has a guess.

Then you will train it by hand. You will show it a bottle, compare its guess with the real price, and use the error to adjust its weights. After one bottle, you will do it again with another.

We are keeping the model this small for a reason. A large AI model can contain billions of adjustable numbers, changing through calculations too numerous to follow. This model has four. You can watch every number move, understand why it moved, and see how many small corrections gradually recover a pricing rule the model was never shown directly.

By the end, you will understand the central idea behind learning from examples: nobody has to write the finished rule into the model. Examples, mistakes, and corrections can shape that rule into its weights.

This is not a literal picture of a human brain, and it is not the whole story of AI at the frontier. It is one learning mechanism made small enough to hold completely.

The sommelier and her four tastes come from The Emergent Mind by Gaurav Suri and Jay McClelland. The example is used here with gratitude.