Treat the model like a printer that constructs an image using pixels. Think of each pixel as a piece of a puzzle. You start by providing the AI with context and instructions, which it uses to print the first layer. This initial layer is like placing the corner pieces of a puzzle—setting the foundation. As you continue providing the same or updated context and instructions for each subsequent layer—layer 2, layer 3, and so on, up to layer n—the AI uses this information like puzzle pieces, fitting each piece together to understand the overall direction and gradually assemble a complete, detailed image. Each layer adds more clarity and resolution, enhancing the image until it achieves FULL HD quality.
At the start, set some pixels as the reference points, and as you add more layers, you will start to see a picture becoming FULL HD.
-
They hallucinate because they have a gap in values (let’s treat it like an
uncharted zone, like in the Minesweeper Game). To avoid this, you have to give it more information and a specific direction. -
They are lazy because we do not give them a direction →
we must be specific enough, or"it will chase its own tail"if we are not specific enough,it will use its values to generate something that is in its bubble. Simple → add more knowledge and specific instructions. -
They are lazy when there is a limit on context length under the hood (this is set to avoid so much waste of computation made by the user. I still think that on the API we should have control of this, and if it is above the limit, in a range set by the user, to pay extra).
I saw some companies do marketing based on the length of the “context” 
. Probably, after the marketing campaign ends and they get a number of subscribers, you will see other fees.
Explore AI Deductions:
Engage with AI through practical exercises to better understand its decision-making processes. Try this project from CS50 on AI, where you’ll employ strategies similar to playing Minesweeper, and see firsthand how AI can be guided: CS50 AI Minesweeper Project.