What ChatGPT is actually doing when it answers you

It doesn't "understand" you. It does one thing, over and over: guess the next token, append it, guess again. Once that clicks, the weird behavior stops being mysterious:

• Tokens, not letters — that's why it can't count the r's in "strawberry".
• Meaning is geometry — words become vectors; "king" sits near "queen".
• Attention decides which earlier words matter for the next one.
• Pretraining makes a wild autocomplete; RLHF turns it into an assistant.
• Temperature is why the same prompt gives different answers.
• The context window is why it "forgets" — the oldest text literally falls off.
• Hallucination is the model doing its job (plausible text) where plausible ≠ true.

Nine minutes, fully animated, no math required:

Which of these surprised you the most when you first learned it?