The one-sentence version
A language model reads everything you have given it so far and repeatedly guesses the next small chunk of text. That is the whole trick. Every behaviour you will meet — the fluency, the confident errors, the sensitivity to phrasing — falls out of that single mechanism operating at enormous scale.
Tokens and context
Models do not see letters or words. They see tokens: fragments roughly three-quarters of a word long. Your prompt, any attached document and the reply all consume tokens from a fixed budget called the context window. Two practical consequences follow. First, a long pasted transcript competes for space with the model's own answer, which is why very long inputs sometimes produce oddly short replies. Second, anything outside the window simply does not exist for the model — it is not being ignored, it was never there.
Why it feels intelligent
Predicting the next token well at scale requires implicitly modelling grammar, facts, tone, code syntax, argument structure and conversational norms. The model was never taught these as rules; they are the cheapest way to make good predictions. That is why the same system can draft an apology email and refactor a function: both are text-shaped problems.
Where the illusion breaks
The model optimises for plausible, not true. It has no internal fact-checker and no awareness of what it does not know. A fabricated citation is generated by exactly the same process as a correct one, with exactly the same confidence. It also has a knowledge cutoff and, in this app, no live web access — so recent events, prices and version numbers are unreliable by construction.
Arithmetic and multi-step logic are the second weak spot, because a single forward pass has no scratchpad. This is why asking for step-by-step reasoning genuinely improves accuracy: each written step becomes context for the next one.
What this means for your prompts
Four habits follow directly from the mechanism:
- Front-load context. The model has no memory of your project, your audience or your constraints unless you supply them.
- State the output shape. "Three bullets, under 40 words each" is a constraint the model can satisfy; "keep it short" is not.
- Give it a scratchpad for hard problems. Ask for reasoning before the answer on anything numerical or logical.
- Verify anything checkable. Names, numbers, quotes, citations, legal and medical claims.
Try it now: open the AI Assistant and send the same question twice — once bare, once with audience, purpose and format specified. The gap between the two answers is the entire skill of prompting, visible in thirty seconds.