We need to reframe how we think about LLMs. They are not people. They are probabilistic computers.

When you type a prompt, you are writing a function. The input is text; the output is text.

Context Windows = RAM

The "Context Window" is the AI's short-term memory. If you exceed it, it forgets. Managing this window is like managing RAM. Don't feed it junk.

Few-Shot = Compiling

Giving examples (Few-Shot) is compiling the logic for the AI. "Here is pattern A, here is pattern B. Now do pattern C." It's pattern matching, not understanding.

System Prompts = Operating System

The "System Prompt" (e.g., "You are a helpful assistant") is the OS. It sets the rules. Change the OS, change the behavior. Hack the system prompt, and you hack the model.