How AI actually works: one mental model, everything explained.
You don't need the math. You need a mental model accurate enough to predict when AI will be brilliant and when it will confidently make things up — because every skill on this site sits on top of that intuition.
01 The autocomplete that read everything
A language model does one thing: given text, it predicts what plausibly comes next — like your phone's autocomplete, if your phone had read a meaningful fraction of everything humans have written. Ask a question and it generates, word by word, the most plausible continuation of a conversation where that question gets answered well. That's it. Every astonishing capability and every embarrassing failure follows from this one fact.
02 Why that produces brilliance
"Plausible continuation" turns out to be a shockingly deep target. To continue text well about plumbing, contracts, or grief, the model had to internalize the patterns of plumbing, contracts, and grief — which looks like understanding, and functionally often is. This is why AI writes, summarizes, translates, and explains so well: those are pattern tasks, and it has absorbed more patterns than any human ever could.
03 Why the same thing produces confident nonsense
The model doesn't know what it knows. Plausible-next-words and true-next-words usually overlap — but when they don't, nothing inside rings an alarm. A fabricated citation is generated by exactly the same process as a real one, in the same confident tone. It isn't lying; it has no concept of the difference. That's why verification is a permanent habit (the hallucination lesson), not a temporary bug workaround.
04 Five behaviors the model explains
- Why context helps so much: more context = a more constrained prediction. Vague in, generic out — it's not being lazy; you left the prediction wide open.
- Why it agrees with you too easily: your framing shapes the plausible continuation. Ask "why is X great?" and great-X text follows. (Fix: ask for the case against.)
- Why it forgets mid-conversation: models read a finite window of text. Fall outside it and, for the model, it never happened.
- Why 'thinking' modes help: generating reasoning steps before the answer literally gives the prediction better text to continue from.
- Why tools changed everything: letting the model search, run code, or read your files grounds prediction in reality — the difference between memory and evidence.
05 The stance that follows
Treat AI as a brilliant, tireless colleague with no sense of its own limits: superb first drafts, genuine insight, zero self-doubt. Everything this site teaches — grounding, verification, iteration, the supervision habits — is just the professional etiquette for working with exactly that colleague.
Ask any AI a question you know cold, then one adjacent question you can't verify. Watch the identical confidence. You now understand AI better than most people using it daily.
Try it in the Playground →This week's challenge
This week, each time an AI surprises you — good or bad — trace it back to the mental model: what made that continuation plausible? Five traces in, you'll find you can predict its behavior before you hit enter. That intuition is the foundation everything else on this site builds on.