Library · paper

Prediction and Entropy of Printed English

Claude Shannon
1951·Bell System Technical Journal, 30(1), 50–64

Source: https://archive.org/details/bstj30-1-50

Next-token prediction, by hand, in 1951.

Shannon measures the entropy of English by turning it into a guessing game — show a reader some text, have them predict the next letter, and count how well they do — and from that derives how predictable, and how redundant, the language is.

That is precisely the objective a modern LLM optimizes (predict the next token; minimize the surprise) and the metric it is judged by (perplexity is Shannon's entropy in disguise).

Where his 1948 paper founded information theory, this one points it straight at language — which is why it reads as the missing link between Shannon and the transformer.

A rare primary source you can read in one sitting and never see next-token prediction the same way again.

Central argument

Shannon measures the entropy — the average information per letter — of printed English, and does it experimentally: by having human subjects predict successive letters of a text and recording how often they succeed. From this he derives bounds on the entropy and the redundancy of English, and models the language with n-gram statistics. It is next-symbol prediction as a way to quantify a language's structure.

Critique

It is a short, information-theoretic study of English orthography from 1951 — it says nothing about learning, neural networks, or generation, and its estimates are of their time. Its value is conceptual and historical: it establishes the prediction-equals-understanding-of-structure idea that language modeling would later mechanize, not any method you would use directly.

Why it matters for product

For anyone learning how LLMs work, this is the origin story of the training objective itself. It makes concrete why 'just predict the next token' is not a trick but a measure of how much a model has captured the structure of language — and it grounds perplexity, the metric you will meet in every training run, in Shannon's entropy. The clearest one-sitting read for seeing that modern language modeling rests on a 1951 idea.