At 36 seconds after 4:00 a.m. on March 28, 1979, pumps stopped in Unit 2 of the Three Mile Island nuclear plant in Pennsylvania. A relief valve on the reactor’s cooling system opened to let off pressure. Thirteen seconds into the accident, it should have closed again. It didn’t.
On the control panel, a light showed that the electric power that opened the valve had gone off. The operators took that to mean the valve had shut.
Then came the alarms. Two operators were in the control room when the first one sounded. By the account of the presidential commission that investigated, a cascade followed that numbered 100 within minutes. One of them later told the commission: “I would have liked to have thrown away the alarm panel. It wasn’t giving us any useful information.”
Somewhere in that flood were the signs that the valve was still open. It would stay open for 2 hours and 22 minutes.
Rabbits and lettuce
In a lecture published in 1971, Herbert Simon talked about his neighbors’ rabbits. They had bought their daughter a pair at Easter, one male and one female, “and we now live in a rabbit-rich world.” A rabbit-rich world, he said, is a lettuce-poor world.
Then he swapped rabbits for information. “What information consumes is rather obvious: it consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention.”
Information uses up attention. The more of it arrives, the less each piece gets.
Simon was precise about who pays. In an information-rich world, he said, “most of the cost of information is the cost incurred by the recipient,” and human beings “can attend to only one thing at a time.”
In the same lecture, Simon cited a paper on time-sharing networks by John G. Kemeny, who had helped develop the Dartmouth Time-Sharing System. In 1979 Kemeny, by then Dartmouth’s president, chaired the commission on Three Mile Island. Its report put the main weight on the operators’ actions, then on what drove them: training it called greatly deficient, confusing procedures, lessons never turned into clear instructions. Then it came to the control room. “During the first few minutes of the accident, more than 100 alarms went off, and there was no system for suppressing the unimportant signals so that operators could concentrate on the significant alarms.”
A system that speaks less
Simon proposed a test for any new system, a computer or a new department: it reduces the demand on everyone else’s attention only if it absorbs more information than it produces, “if it listens and thinks more than it speaks.”“The crucial question is how much information it will allow to be withheld from the attention of other parts of the system.”
A filter has its own price, and one of the discussants on Simon’s panel named it. The political scientist Karl Deutsch pointed to Lyndon Johnson, “one of the best-filtered chief executives this country has ever had.”“The uncongenial information may be the vital information.” At Three Mile Island, the drain pipe below the valve usually ran a little hot from a slight leak, so the crew had learned to dismiss its high readings. That morning, those readings were the sign of the open valve.
The renewal buried in your inbox
Imagine you manage the accounts of a software company’s twelve biggest clients, and you’re copied on every thread about them. That’s a couple of hundred emails a day, and you skim the subject lines in the evening.
One of them reads “Re: Re: Q3 renewal.” In its fourth paragraph, a client’s head of operations writes that they’re talking to a competitor and need an answer on pricing by Friday. You find it the following Wednesday.
Nobody hid it. Twenty colleagues were keeping you informed, and every copy took a slice of the same evening. Suppose you ask for one line at the top of anything that needs you, “decision needed by Friday,” and let the rest go into a weekly digest someone else reads first. You see far less, and the renewal is one of the few things you see.
Hundreds of lines, usually accurate
curl, an open-source tool for moving data across the internet, has a seven-person security team. In July 2025 its lead developer, Daniel Stenberg, wrote that each security report engaged three or four of them, “Perhaps for 30 minutes, sometimes up to an hour or three. Each.”
The team was seeing what Stenberg later called “an explosion in AI slop reports.” In earlier years, north of 15 percent of submissions had been confirmed vulnerabilities; starting in 2025, under 5 percent were. He wondered whether the easy bugs had simply been found already. In January 2026 the project ended its bug bounty, to remove “the incentives for submitting made up lies.” Code contributions were less of a problem, he noted, because machines check them first: no human time goes to a pull request “until the quality is good enough to get green check-marks from 200 CI jobs.”
On October 22, 2026, he reported a twist: since spring, the made-up reports were “practically gone,” and the AI-assisted ones were typically accurate. The new challenge was “high volume, high quality.” Every report now ran to hundreds of lines. “AI talks too much,” he wrote. In a good report, the human writes a brief first paragraph, then submits the full report.
You can apply his rule before you hit send. When an AI helped you write something for a colleague, put a few sentences on top in your own words saying what’s wrong and how much it matters. If they’re hard to write, read the rest again first. If you’re the one receiving, note how long a typical report takes you to triage this week, ask for the summary first, and time it again next week.
The full report can sit underneath, for whoever turns out to need line 214.