Guessing the next word from everything that came before it — not understanding, just very well-calibrated pattern completion.
Popular discourse treats large language models as if they reason or “think.” Underneath, the mechanism is simpler and older than the current hype suggests: a model estimates the statistical likelihood of the next token given everything before it, then picks or samples accordingly.
This isn’t a new idea invented by OpenAI or Google. Claude Shannon demonstrated the core principle in 1951, showing that a fluent English speaker’s implicit knowledge of language statistics lets them guess the next letter in a sentence with high accuracy — effectively describing prediction as compression of language’s redundancy, decades before “language model” was a term. Modern LLMs formalize and scale the same wager across enormous corpora.
The distinction matters because prediction is not the same operation as judgment, meaning, or truth — see Intelligence for why the vault treats it as a category error when a predictive system is granted epistemic authority it hasn’t earned.
Source: Shannon, C.E. (1951), “Prediction and Entropy of Printed English,” Bell System Technical Journal, 30(1), 50–64.