The Word Load, and What It Actually Does

There is a word this room has been using and it is doing more damage than the material it appears in. The word is load. It is used by every guide, every application, every setup instruction, every progress bar that has ever crossed a screen while a model is opening. And most of what a person thinks when they see the word is wrong.

The word sounds like moving. Load the truck. Load the dishwasher. Load the file. The English of it points at picking a thing up in one place and putting it down in another, and the place it came from is now empty because the thing is no longer there.

That is not what happens.

When a machine loads the model file, the file in storage does not go anywhere. It stays exactly where it was, at its full 6.63 gigabytes, and if the power were cut the next second the file would still be there when the machine came back on. What happens instead is a second version of the file appears in VRAM, on the card, and the machine begins reading from that second version. Same file, two addresses. The picture with the wrong drawing beside the right drawing is on the panel today because saying this in words is not enough. The wrong drawing has an arrow leaving storage and a dashed outline where the file used to be. That is the mental picture the word load produces on its own. The right drawing has no arrow at all. The file is drawn twice, once in each place, both intact.

Two things follow from this that are worth stating plainly.

The first is that VRAM has to be large enough to hold the file even though storage already is. This is why the ladder settled the question the way it did. The Chromebook has enough storage to hold this file if it had nothing else on it, and it still fails, because storage holding a copy is not what the arithmetic requires. VRAM has to hold its own copy while the machine works on it. There are two capacities in play, not one, and the numbers can differ by a factor of eight in ordinary hardware.

The second is that pulling the power out of the wall does not destroy the file. It destroys the copy. VRAM empties completely the instant the power stops, which is the reason the machine has to load the file again the next time it wants to run the model. Storage does not do that. Storage is what non-volatile means. The file will still be there at full strength the next time the machine turns on, and this is what makes loading a repeatable step rather than a permanent one.

There is one more piece the wrong picture obscures. When the model file gets larger than the card can hold, some machines will run it anyway by leaving part of the file in RAM and fetching pieces of it across a narrow connection each time they are needed. This works and it is not fast. The speed penalty is somewhere between five and thirty times slower than a file that fits on the card. That penalty exists precisely because the file is now in two active places at once, and the second place is much further from the GPU than VRAM is.

So the word load hides a shape that is nothing like the word. A copy appears. Nothing leaves. Everything stays. The file is in two places at once and one of them is not doing arithmetic on it. The picture drawn wrong on purpose is on the panel because until the room can see the wrong picture and the right picture at the same time, the word will keep pulling the mental model back to the wrong shape.

Every question on this half of the classwork asks about that difference in a different way. If a student can answer them from the shape rather than from the word, the day landed.

View

The first item is your name. Answer it — that is what puts you on the work now that email collection is off.

Exit ticket · on your own
Could This Class Run Its Own AI? · Day 3 of 7