The delete test

There is a quick way to feel the difference, and it is to try to remove something.

Start with the ordinary case. A database holds a row saying Paris is the capital of France, and you want it gone. Find the row, delete it. Paris is removed, cleanly and completely. France is untouched. Every other row sits exactly where it was. The hole you made is precisely the shape of the thing you took out, and you can prove the removal by going to the address and finding nothing there.

Now the model. There is no row.

Go looking for the numbers responsible for Paris and there is nowhere to arrive. The weights carrying something about Paris are also carrying something about France, about European capitals, about the Seine, about the Eiffel Tower, about the word Paris in a novel, about croissants, about a small town in Texas with the same name.

None of those are stored separately from each other. They are leanings in the same enormous set of numbers, overlapping everywhere.

So the removal you want has no clean form. You would be nudging millions of numbers that are simultaneously holding up all of that, and the question is what happens to the rest when you do.

That question has an answer, and it is easier to feel than to describe. The next slide hands you the dial.