Why counting hits the same wall

The oldest way to guess the next word needs no learning at all. Just count.

Take a huge pile of text. Every time dog appears, tally what came after it. Do the same for every word. To guess what follows dog, pick whatever won most often.

That works in a thin way. After dog, barked shows up thousands of times, so barked wins.

Now ask what follows puppy. You start from nothing. The counter has never met this word, and everything it worked out about dog sits on dog's card.

A puppy is a dog. Nearly everything true of one is true of the other. The marks on dog's card are exactly what puppy's card needs, and there is no way to carry them across. Two entries, nothing between them.

Even by hand you would be stuck choosing what travels. barked belongs on puppy's card. It does not belong on cat's. To make that call you would already need to know which words are related, which is the thing you were trying to work out.

The tally hits the same wall as the private slots. What it learns about one word can never help with another. Every word is an island again.

This is also the machine people picture when they call a language model just fancy autocomplete. We will come back to that claim.