MODULE 05 — REALITY CHECK · 18 AI Concepts

What AI still can’t do

Not a list of things it is bad at. Seven places where the map simply ends — each one the direct shadow of a mechanism you have already watched work.

because

The pattern

None of these is a bug

THE MECHANISM THE LIMIT IT PRODUCES trained once, then frozen a fixed window, rebuilt each turn predicts plausible text output feeds back in as input no consequences reach it learns from data exactly as given every token takes real compute cannot know today cannot remember you cannot tell true from plausible early errors harden into facts cannot feel that it matters can carry its data's biases length and size are never free

A bigger model shifts none of these lines. Each one is answered by adding a mechanism from the outside — a lookup, a store, a check, a person.

So what

You can now predict where it will fail before you watch it fail — which is the whole difference between being impressed by it and being able to use it.