What the four numbers say

Eventually the corrections stop achieving much. Each new pass barely moves the weights because they are approaching the shared rule behind this training set. For the demonstration, that is the moment to stop.

So stop. Freeze the weights and correct nothing further. What is left is four numbers that do a job, and that is a trained model.

Before going further, look at those four numbers in the machine. They started identical at 8, and every difference between them now was put there by being wrong about wine.

Read them as a claim about what makes a bottle expensive.

What does this model think matters most?

Watch what that does to a price. Start with the cheapest thing the model can imagine, a plain viognier with no fruit at all, at about 6 dollars 50. Make it a cabernet instead and it goes to 14. Drench that cabernet in plum and it reaches 30. Drench it in apricot instead and it reaches 53.

Same wine, one change at a time. The grape is worth about 8 dollars, plum about 16, apricot about 39. Only one of those four bottles was ever in the training set.

Nobody told the model that. It has no opinion about wine, no taste, and no idea what an apricot is. It arrived there by being wrong about eight bottles, over and over, and following the errors wherever they pointed.

What the four numbers say

So stop. Freeze the weights and correct nothing further. What is left is four numbers that do a job, and that is a trained model.

Before going further, look at those four numbers in the machine. They started identical at 8, and every difference between them now was put there by being wrong about wine.

Read them as a claim about what makes a bottle expensive.

What does this model think matters most?

Same wine, one change at a time. The grape is worth about 8 dollars, plum about 16, apricot about 39. Only one of those four bottles was ever in the training set.