One bottle is not enough

The nudge has only ever been tried on one bottle. That plummy cabernet has been tasted, guessed at, corrected, and nudged from 10 dollars out to 8 dollars 75 out.

Three bottles went past earlier, but those were the greedy version, and the greedy version fell apart. The careful method has faced exactly one wine.

One wine cannot teach it anything worth having. Keep nudging the weights against that same cabernet and they will drift toward whatever prices that one bottle correctly, and nothing else. That is not tasting. That is memorising one answer. A model corrected only on the cabernet ends up knowing the cabernet costs 22 dollars, in the way a parrot knows it.

What Vincenza actually learned was how to price wines she had never tasted. The model needs the same thing, and one bottle cannot give it.

Here is what happens instead.

Gather a collection of bottles, each with its correct price attached. There are eight of them in this chapter, chosen to cover both grapes and a range of fruit: a plain cabernet, a plain viognier, a wine drenched in apricot, and several with no apricot at all. Then run the whole four-step loop on every single one:

1. Taste it.
2. Guess.
3. Be told the price.
4. Correct the weights a little.

Then start again from the first bottle and go round the entire collection a second time. Then a third. The weights keep shifting, a nudge at a time.

This whole process has a name: training. The collection of priced bottles is the training set.

That is the description. The next few slides are the thing itself.