Large language models in plain language

Large language models (LLMs) generate text by repeatedly asking: what token is likely next? That simple idea explains a lot of their strengths and failures.

Tokens, not “thoughts”

Text is split into tokens — chunks that may be whole words or pieces. The model assigns probabilities to possible next tokens and samples from them (with settings that make answers more or less “creative”).

It is not looking up a single truth database. It is continuing a pattern that often looks like knowledge.

Why they sound confident

Training rewards fluent, helpful-sounding continuations. Confidence is a style of language, not a calibrated certainty meter. A wrong answer can still sound polished.

Hallucinations are expected

When the model invents a citation, a date, or a feature that doesn’t exist, that is often called a hallucination. It’s not a rare bug you can ignore — it’s a consequence of predicting plausible text under uncertainty.

Practical response: verify claims that matter. Prefer primary sources for facts that affect money, health, law, or someone’s reputation.

See also: Glossary — Hallucination, Token. Next lessons cover embeddings (how meaning becomes comparable numbers), transformers (the common architecture behind many LLMs), and a follow-on on how transformers work — still without heavy math.

Takeaway

LLMs are powerful pattern completers. Use them for drafting, exploring, and restructuring — then check anything you’d be embarrassed to get wrong.