On January 23, 1863, the commander of the Army of the Potomac, Ambrose Burnside, asked to see Abraham Lincoln alone, after midnight. He brought a draft order. Its first item dismissed one of his generals, Joseph Hooker, “as a man unfit to hold an important commission during a crisis like the present.” Burnside told the president that if the orders weren’t accepted, he would resign.
Lincoln relieved Burnside instead. Three days later he gave the army to Hooker.
He also gave Hooker a letter. Hooker was “a brave and a skilful soldier,” Lincoln wrote, with confidence in himself, “a valuable, if not an indispensable quality.” He had also “taken counsel of your ambition” and thwarted Burnside as much as he could. And Lincoln had heard, “in such way as to believe it,” that Hooker had been saying the army and the government needed a dictator.
Then came the turn. “Of course it was not for this, but in spite of it, that I have given you the command. Only those generals who gain successes, can set up dictators. What I now ask of you is military success, and I will risk the dictatorship.”
Peaks and valleys
Peter Drucker built a chapter of The Effective Executive, “Making Strength Productive,” on choices like that one. “Strong people always have strong weaknesses too,” he wrote. “Where there are peaks, there are valleys.” Anyone who tries to “place a man or staff an organization to avoid weakness will end up at best with mediocrity.” The question his effective executives ask about a candidate, he wrote, is always “What can he do uncommonly well?”
Hire for what someone does uncommonly well, then make their weak spots matter less.
Drucker’s own Lincoln story is about Grant. In his telling, Lincoln was told that Ulysses S. Grant, his new commander-in-chief, was fond of the bottle, and answered with the famous joke about finding out Grant’s brand and sending a barrel to his other generals. Drucker wrote that Lincoln chose Grant for his tested “ability to win battles and not for his sobriety.” Before Grant, Drucker said, Lincoln had appointed, in succession, three or four generals “whose main qualifications were their lack of major weaknesses,” and the North “had not made any headway for three long years from 1861 to 1864.”
The record is messier. Whether Lincoln ever made the barrel joke is unsettled. According to Quote Investigator, the earliest known printing, in the New York Herald in September 1863, was unsourced, and an 1882 history of the military telegraph quoted Major Thomas Eckert as saying Lincoln had denied it. And Hooker is hard to describe as a general chosen for his lack of major weaknesses. Lincoln listed them in writing and took him anyway.
The bet at Chancellorsville
It didn’t pay. In early May, at Chancellorsville, Hooker had about 90,000 men around the crossroads against Robert E. Lee’s divided 48,000, by the National Park Service’s count. He fought on the defensive and pulled the army back across the Rappahannock. The Park Service’s history of the battle says he “had lost his nerve.” Before Gettysburg that July, Lincoln had replaced him.
That’s the part Drucker’s version skips. Staffing for strength can raise the ceiling. It doesn’t guarantee the result. Hooker’s record, the Park Service says, was strong in the Peninsula Campaign and at Antietam. Grant’s evidence was of another kind: a campaign he had run himself. Lincoln’s July letter shows it. “I do not remember that you and I ever met personally,” it begins. Grant had just taken Vicksburg, in a campaign whose moves Lincoln had doubted more than once. “I now wish to make the personal acknowledgment that you were right, and I was wrong.” In March 1864, Lincoln assigned Grant to command the armies of the United States.
Drucker also drew a hard line. Character and integrity, he wrote, accomplish nothing by themselves, “But their absence faults everything else.” That is the one area, in his view, where a weakness is “a disqualification by itself.”
The engineer who can’t write the update
Imagine you’re running a product launch and choosing who takes the incident desk on launch day. Your strongest engineer can find almost any bug in minutes. She also writes customer updates that read like a stack trace, and sometimes like a scolding.
Staffing to avoid weakness would mean someone pleasant and slower. You give her the desk and pair her with a support lead who writes every customer update from her notes. On launch day, a payment bug is found and fixed in forty minutes. Customers get two calm emails. Her weakness is still there. It just has nothing to do.
A model that confabulates, and a scorer
DeepMind built that arrangement for a language model. Such models, its team wrote in Nature in December 2023, “tend to confabulate.” So in a system called FunSearch, a code model, Codey, built on PaLM 2, was given only the job of proposing short programs. An automated evaluator ran every program and scored the result, which the authors said “guards against confabulations and incorrect ideas.” The best programs went back into the prompt for the next round.
On the cap set problem in mathematics, FunSearch found a set in eight dimensions larger than any previously known. Only four of 140 experiments reached that size. The authors said the method works best when there is an efficient evaluator, and that tasks such as generating proofs fall outside it.
You can set up the same split. Before asking a model for a regex or a SQL query, write the test cases it must pass, and have it propose several versions. Run each one against your tests and keep only what passes. Then read the survivor, because tests only check what you thought to write down.
Lincoln had no test suite for generals. He handed Hooker the army, and in the last line of his letter all he could write was a request: “go forward, and give us victories.”