A language model does one narrow thing extremely well: it estimates which chunk of text most plausibly comes next. Everything that feels like reasoning is built on top of that single ability, repeated billions of times.
Tokens, not words
Text is split into tokens — fragments roughly three to four characters long. The model never sees letters or meaning, only numeric IDs and the statistical company each one keeps.
Training is compression
During training the model adjusts billions of weights so that its predictions match real text more often. The result is a lossy compression of patterns in language, not a database of facts you can look up.
Why it invents things
When the pattern is strong but the fact is missing, the most plausible continuation is still produced confidently. That is why sources, dates and numbers always need checking against a primary reference.
