Training vs. using

Two completely different events get called “AI”. One took months and happened once. The other takes a second and happens every time you type.

TRAINING · once Months of compute. Enormously expensive. Everything it will ever know is fixed here.

USING · every chat Instant. Cheap. Changes nothing about the model.

Drag the red line — that is the knowledge cutoff. Different models are trained at different times, so it sits in a different place for each one.

Asked about something before the line Answers from memory, instantly, and is usually right.

Asked about something after the line The dangerous case: it does not know that it does not know, so it often answers anyway.

The problem it explains

Talking to it does not teach it

People assume a correction sticks. It does not: nothing you type during Using ever reaches the numbers set during Training.

The missing arrow is the point. “I already told you that” fails because telling it something writes to the desk, not to the model.

How it works

TrainingIt ends on a date, and that date becomes the edge of what it knows.

UsingSame frozen model as everybody else’s, every single time.

The catch

The model stays frozen. Everything clever happens in what gets handed to it — which is the next few pages.

So what

A model with no lookup is a very well-read colleague who has been in a locked room since the cutoff date — and does not know they were locked in.