Engineering for Bounded Cognition

Browse<br>ForewordManifestoGlossaryGuidesPosts<br>Browse by tag<br>All posts (27)cross-cutting (1)implementing (16)investigate (2)operating (13)planning (7)retiring (3)retrospective (3)reviewing (10)triage (2)verify (4)<br>Engineering for Bounded Cognition<br>The mind that builds software is far smaller than the software it builds, and almost everything good in engineering is a way of living with that gap.<br>3 February 2026 cross-cutting<br>You have heard that the mind can hold seven things at once. It is the most repeated figure in popular psychology, and the man who found it did not believe it. George Miller titled his 1956 paper "The Magical Number Seven, Plus or Minus Two" with his tongue in his cheek; he wrote that the way the number kept turning up across unrelated experiments was probably nothing more than a coincidence, and said he felt persecuted by an integer. He was right to be suspicious. When later researchers stripped away the tricks we use to cheat, the rehearsing under your breath, the grouping of digits into chunks, the number fell. The honest estimate of how many separate things a person can hold in mind at one time, with no help, is about four.<br>Four. Set that against the device you are reading this on, running tens of millions of lines of code, and the whole problem of software fits into a single ratio. The systems we build are among the most intricate objects in human history, and they are built, and changed, and rescued under pressure, by minds that can keep about four things straight at once.<br>It is worse than small, it is narrow. The attention you point at those four slots is not a floodlight thrown across a room; it is a torch beam in a dark warehouse, and nearly everything is outside it. People were asked to watch a video and count how many times a team passed a basketball. Halfway through, a person in a gorilla suit walks into the middle of the shot, turns to the camera, beats their chest, and strolls off, nine seconds in plain view. About half the people watching never see it. They are looking straight at it. They are not blind; the beam was simply pointed at the ball. It gets stranger. Stop someone in the street to ask directions, and while they are answering have two workers carry a door between you, and behind the door swap yourself for a different person, different height, different shirt, different voice. About half of people carry on giving directions to a stranger, never registering that the human in front of them has been replaced. And the little you do hold, you hold for seconds: show someone three random letters, stop them rehearsing, and within twenty seconds most of it is gone. Four slots, a narrow beam, and a leak. That is the instrument we build software with.<br>Once you have felt how small the mind is, the surprising thing about software is not that it breaks. It is that it ever works. We have raised cathedrals of logic far too large for any person to hold, and we keep them standing with the four-slot, gorilla-missing minds we were issued at birth. The marvel is the uptime.<br>This puts a word the industry loves in a harsher light: human error. Read almost any incident report and you will find a person at the end of it who typed the wrong command, clicked the wrong button, missed the warning, and we file the whole thing under their carelessness and resolve to try harder next time. But a warning that half of attentive people will look straight at and not see is not a warning, it is a decoration. A button one tired click away from disaster was put there by a designer, not by the person who clicked it. The people who study this for a living, in aviation and medicine and the control rooms of power stations, where "just be more careful" gets people killed, worked it out long ago: when a capable person makes the same mistake over and over, the fault is almost never in the person. It is in a system built for an operator who does not exist, one who never tires, never glances away, and holds the entire machine in their head. That operator has never been born. Every system built for them is already broken; it simply has not met its bad day yet.<br>For a while you could comfort yourself that the machines would rescue us from this, that a computer at least did not have a four-slot mind. Then we built machines that read and write language, handed them our code, and watched them fail in a way that was uncomfortably familiar. A large language model has what is called a context window, a hard limit on how much it can take into account at once. It behaves less like long-term memory than like the thing you are holding in mind right now, and the moment something falls outside the window it is not faintly recalled, it is simply gone. Even inside the window, the model does not attend evenly. In a study with the flat title "Lost in the Middle", researchers found these systems answer well when the fact they need sits near the start or the end of a long input and noticeably worse...