In the autumn of 1999, someone in the FreeBSD project proposed a small fix. The sleep command, which tells a computer to wait, couldn’t handle fractions of a second. Let it take 0.5, the proposal said. FreeBSD would match its sister systems, OpenBSD and NetBSD, and no existing program would break.
The project’s mailing list treated it like a constitutional amendment. Objections, counter-proposals and changes piled up, Poul-Henning Kamp wrote, as if the change “would have plugged all the holes in swiss cheese or changed the taste of Coca Cola.” He guessed the thread had run a couple of weeks.
The change went in on October 1 anyway. Kamp, by his own description a kind of de facto spiritual leader of the project, said later that the thread made him angrier than almost anything in his life. The day after the commit, he emailed the project’s developers.
He didn’t defend the half-second. He reached for a book of satire written 42 years earlier, about a committee and an atomic reactor.
Item Nine
C. Northcote Parkinson was a history professor at the University of Malaya when he published Parkinson’s Law in 1957. One chapter stages a finance committee meeting, with invented members and invented numbers, and it is meant to be funny.
Item Nine is an atomic reactor, estimated at $10,000,000. Of the eleven members, Parkinson explains, four don’t know what a reactor is. Three more don’t know what it’s for. Of those who do, only two have any idea what it should cost. One of them, Mr. Brickworth, has doubts about that suspiciously round figure. But to raise them he would have to start by explaining what a reactor is, “and no one there would admit that he did not already know. Better to say nothing.”
Approved. Two and a half minutes.
Item Ten is a bicycle shed for the clerical staff: $2,350. Aluminum roof or asbestos? Galvanized iron? Does the staff even need a shed? “Everyone can visualize a bicycle shed,” Parkinson writes. The debate runs forty-five minutes and might save $300. Then comes the coffee for another committee’s meetings, $57 a year, and an hour and a quarter of argument that ends with a request for more information.
Parkinson named it the Law of Triviality: “the time spent on any item of the agenda will be in inverse proportion to the sum involved.”
Meetings tend to spend longest on what everyone understands, not on what costs the most.
Two pressures, one room
Parkinson never measured any of this. His committee is invented, and the “inverse proportion” is part of the joke. What makes the scene useful is that it pulls apart two separate pressures.
One works on the big item. Questioning it is expensive. You need knowledge most people in the room don’t have, and the person who has it would have to begin by exposing everyone else’s ignorance. Kamp’s version: people can’t grasp the reactor, so they assume “somebody else checked all the details before it got this far.”
The other works on the small item. After waving the reactor through, Parkinson’s members wonder “whether they have really been pulling their weight.” The shed is their chance to prove it. In Denmark, Kamp wrote, they call this “setting your fingerprint.”
Kamp also counted a cost Parkinson’s committee never shows. The developer who proposes the shed has to sit through all of it. His email ended with an apology to the people lurking on the list, “scared away from participating.”
The checkout button
Say a product lead brings two decisions to her team’s weekly review. One is whether to move all card payments to a new provider, which touches every customer’s money and which only one engineer has studied closely. The other is the wording on the checkout button.
If the button comes up second, the payment switch may get two and a half minutes. The engineer has worries about the new provider’s refund handling, but airing them means walking six people through how card refunds work. The button may get forty.
She can change the order of events. The engineer sends a one-page plain-language note on the payment switch two days before, so nobody has to admit in the room that they don’t know what a chargeback is. The switch goes first, with its own block of time, and each person names one risk before anyone argues. The button goes to the designer, who can choose any wording that fits the existing style guide without bringing it back to the meeting. That last move is Kamp’s wish: let the person building the shed pick the colour.
A judge with nothing to prove
A language model grading answers isn’t trying to show a committee it pulled its weight. Even so, a 2023 study found a judge’s attention drifting toward what was easy to see.
Minghao Wu and Alham Fikri Aji had GPT-4 and Claude 1.3, along with crowd workers and experts, compare pairs of answers to 40 everyday questions. Some answers were seeded with factual errors, others with spelling or grammar mistakes. With all criteria blended into one verdict, GPT-4 ranked correct answers written with grammatical mistakes below answers carrying several major factual errors. When a model missed the inaccuracies, the authors wrote, it “often favors flawed outputs over shorter or grammatically imprecise responses.” The expert humans in the study ranked them the other way round.
Their fix looks like a better agenda: score accuracy, helpfulness and language as separate items, so the visible flaws are less likely to crowd out the hidden ones. The authors report that this significantly improved GPT-4’s judgments, particularly on factual accuracy. For the crowd workers, it brought no statistically significant improvement.
In Parkinson’s meeting, the coffee was still undecided after an hour and a quarter. The reactor, all $10 million of it, had been signed off in two and a half minutes.