Level 1: for a 12-year-old
Imagine a friend who has read a huge library and is amazing at guessing the next word. You say "peanut butter and..." and they say "jelly." A language model does that over and over: it guesses one small piece of text, adds it, then guesses the next, until the answer is finished. It isn't looking anything up. It's guessing from patterns it learned.
Level 2: for a student in another field
A large language model works in tokens, chunks of text that are often a word or part of one. Given everything in the conversation so far (the context), it calculates a probability for every possible next token, picks one, adds it, and repeats. Those probabilities come from parameters, billions of numbers tuned during training so the model gets better at predicting the next token across enormous amounts of text. Later training, including human feedback, shapes it into a helpful assistant. From your economics background: it's closer to a fitted model than to a database, like a regression that predicts from patterns rather than storing every observation.
Level 3: for a practitioner
Modern LLMs are transformers. Tokens become vectors, and stacked attention layers let each position draw on earlier ones, producing a probability distribution over the next token. Generation is autoregressive: sample, append, recompute. Temperature and top-p reshape that distribution; lower temperature makes output more predictable, not more correct. The context window caps what the model can see at once. Post-training (instruction tuning and reinforcement learning from feedback) changes which continuations are likely, which is why models with similar architectures behave so differently as products. Tools like search are wrappers around the same core step. The misconception experts still correct: that the model retrieves stored facts. Knowledge is spread through the weights, so a fluent, confident answer can still be false.
Where the Level 1 analogy breaks: your friend could re-read a book to check; the model can't consult its training text at all. It also doesn't always pick its single best guess.
Test questions
1. If the model isn't looking things up, where does its answer come from?
1. Why can the same prompt give two different answers?
1. A colleague sets temperature to 0 "to make it accurate." What's wrong with that reasoning?
Answers
1. From patterns learned in training, used to guess one piece of text at a time.
1. The next token is sampled from probabilities, so a different pick early on leads down a different path.
1. Temperature 0 makes the model take its likeliest token each time, which reduces variety, not errors. The likeliest token can still be wrong.
