How well do you actually know ChatGPT?

Generative (it writes), Pre-trained (it learned from data before you ever met it), Transformer (the neural-network architecture underneath). The name is a spec sheet, not a brand.

A hallucination isn't a malfunction — the model is doing exactly what it does, predicting plausible text. It has no internal signal for “I don't know”, which is why the fabrications sound so sure of themselves.

One concrete example of format, length, and tone beats any pile of adjectives. “Be accurate” gives the model nothing to act on; a sample gives it a target to match.

Custom instructions apply to every new chat; memory accumulates facts about you across chats. Together they end the ritual of retyping who you are and how you like answers formatted.

The free tier can use conversations to improve models unless you turn that off in settings. It's the single most consequential setting most users have never opened.

Low temperature picks the likeliest next word every time — consistent and a bit flat. High temperature loosens it up. Neither setting makes anything truer.

Built-in knowledge freezes at the cutoff; web search papers over the gap when it's available. If it isn't, last month's news simply doesn't exist to the model.

Formats drift because the model re-improvises them every time. Pin the structure down once — a template, a filled-in example — and reuse it. That's the whole trick behind consistent output.

Rewriting, explaining, and outlining run on what the model already knows. Today's news doesn't — without live search it will either decline or, worse, guess.

The system prompt is standing instructions that outrank anything you type — tone, role, rules. Custom instructions are essentially the version of it that you get to write.