Inside the Live Brain: Word Dominoes & The Creativity Thermometer!
When you chat with ChatGPT, why do the words pop out one by one instead of appearing all at once? And how can a simple dial turn a serious robot into a wild poet? Let's take a peek!
🎲 1. The Epic Game of Word Dominoes (Autoregressive Generation)
An AI cannot spit out an entire paragraph in a single second. Why? Because to know what the 5th word should be, it first needs to know what the 4th word was! It's just like toppling dominoes: each word topples forward to trigger the next word!
Click the button to watch the robot topple the domino chain word by word:
⚡ The Secret Memory Drawer (KV Cache): Notice how the robot doesn't get tired? It stores all the past words in a special "Memory Drawer" (called KV Cache), so it never has to re-read what it already typed!
🌡️ 2. The Creativity Thermometer: Serious vs Wacky!
Every time the AI picks the next word, it has several candidate words with different percentage probabilities. We can use a magic dial called Temperature to control its personality: