What Betty Guessed Right

Claude Shannon pulled a Raymond Chandler novel off the shelf, put his finger on a random line, and asked his wife to guess the next letter, then the next, then the next. She got most of them. English, it turns out, is roughly seventy-five percent redundant — three-quarters of what you're reading right now could, in principle, be reconstructed from the quarter that survives. Gleick tells this story to explain entropy, but in 2026 it reads like something else entirely: a manual demonstration, by hand, with a paperback and a spouse, of exactly what a transformer model does at scale with a trillion parameters. Shannon's "approximations of English" — his zero-order gibberish giving way to digram frequencies giving way to word-level plausibility ("REPRESENTING AND SPEEDILY IS AN GOOD APT OR COME CAN") — are the direct, unbroken ancestor of anything a language model outputs. Nobody in 1948 was trying to build a machine that writes. Shannon was trying to measure a telegraph line. He built the blueprint for the thing anyway, as a side effect of asking a narrower question, which may be the most reliable way blueprints get built.

The chapter's real hinge, though, isn't the bit. It's the sentence Gleick renders in what he calls "germicidal quotation marks": "The 'meaning' of a message is generally irrelevant." Shannon needed to say this to make information measurable — you can't put a number on significance, but you can put one on surprise. It was a methodological amputation, performed for a specific engineering purpose, and it worked spectacularly. The problem, which the book states but doesn't quite sit with, is that the amputation outlived its occasion. We now run enormous portions of collective judgment — what gets recommended, translated, summarized, flagged, believed — through systems descended from a framework that was explicitly, proudly indifferent to whether any of it meant anything. That indifference was Shannon's discipline. It has become, ambiently, everyone else's default setting.

The chapter's other seam is quieter and, I think, sharper: the history of complaining about too much information, which Gleick traces from Leibniz fearing "a return to barbarism" from the proliferation of books, through Alexander Pope's "deluge of Authors," through Vincent of Beauvais justifying his encyclopedia by citing "the shortness of time and the slipperiness of memory," all the way to David Foster Wallace's "Total Noise" in 2007. Gleick notes, without much comment, that Amazon sells Data Smog — a book warning about information glut — with the marketing line "Start reading Data Smog on your Kindle in under a minute." That detail is the whole argument in miniature. The warning against the flood becomes an item carried by the flood. This isn't new to the algorithmic era; it's at least four hundred years old, which is either reassuring (we've survived worse ratios of noise to signal) or considerably worse than reassuring, since it means the absorption of critique into content isn't a bug of platforms — it's what information abundance has always done to its own critics, mechanically, regardless