Ask people how ChatGPT works and you'll hear guesses: it searches the internet (no), it has a database of answers (no), it understands language like we do (also no). The real mechanism is simultaneously simpler and stranger, and it fits in one sentence: a large language model does one thing — predict the next word.
Autocomplete, taken to its logical extreme
Your phone keyboard already does a tiny version: type "see you" and it suggests "later." A large language model (LLM) is that same idea trained on a colossal slice of human text — books, articles, code, forums. Show it "the capital of France is" and it predicts "Paris," not because it looked anything up, but because in millions of training examples, that's the word that followed. Generate one word, append it, predict the next, repeat — and out flows a paragraph.
The surprise that created the AI boom
Here's the genuinely weird part, the part that surprised the researchers themselves: predict-the-next-word, done at sufficient scale, starts producing behaviour that looks like reasoning. To predict text well, the model was forced to internalise the patterns underneath text — grammar, facts, logical structures, the shape of an argument, the conventions of working code. Nobody programmed it to translate French or write Python. Those abilities emerged because translation and code exist in the training data as patterns, and the model got extraordinarily good at patterns. (If you read the first post on this trail — pattern-matching without understanding — this is that idea at maximum volume.)
This mechanism explains the model's strangest behaviours. Why does it sometimes state nonsense with total confidence? Because it's producing plausible text, not checked facts — confidence is a writing style, not a truth signal. Why can't it reliably tell you about yesterday's news? Its knowledge froze when training ended. Why does it never say "I don't know" unprompted? Because the training data — human writing — rarely does. The full post on AI's made-up "facts" is the next waypoint on this trail.
The vocabulary you'll actually encounter
Four terms cover most conversations. Training — the months-long, enormously expensive process of learning patterns from data. Parameters — the billions of internal numbers where those patterns live (the "large" in large language model). Tokens — the word-chunks the model actually reads and writes; pricing and limits are measured in them. Context window — the model's working memory for the current conversation: exceed it and the earliest parts fall out, which is why very long chats "forget" their beginnings. None of this is magic — and knowing it's not magic is exactly what makes these tools usable in real work, which is where this trail heads next.