Which Door

Most companies aim AI at their cheapest work. Ask an executive where AI saves money and you get a list of jobs the company already did, only cheaper: the support queue, the first draft of the contract, the month-end close. That is automation, the smallest thing this technology does. Daron Acemoglu’s 2024 estimate puts the ten-year productivity gain from using AI this way at two-thirds of one percent. Those tasks were routine and cheap already. The expensive part of a business was knowing whether an initiative would succeed or fail before you paid for it. AI flipped the order: you find out first, then double down.

Microsoft counted how often an approved idea turns out to be worth building. The company built a platform to test ideas instead of shipping them. Ronny Kohavi ran that platform. He reported in 2009 that a third of ideas improved the metric they targeted, a third did nothing, and a third made that metric worse. At Bing and Google only ten to twenty percent helped. Capable people proposed every one of those ideas. Reviewers approved them without knowing which third they were in. In 2012 a Bing employee proposed a change to ad headlines. Managers rated it low priority and it sat for six months. Then an engineer saw it would take a few hours to build and tested it. The change raised revenue 12%, more than $100 million a year in the U.S. By its own standards the review worked perfectly. It also buried the best idea Bing had that year. Booking.com now runs more than 25,000 experiments a year. Its people are not bad at judging ideas. Judging ideas in advance is hard for everyone.

When managers decide which ideas to build they guess correctly only a third of the time. In a fast market the guess goes stale sooner. The cost of a test sets how fast a company learns. Expensive tests mean you run only a few. You write the plan and spend the money before you learn anything. Cheap tests mean you run hundreds. You learn before you have a plan to defend.

Samuel Langley had the plan. He was Secretary of the Smithsonian and the most credentialed scientist in the country. In 1898 the War Department gave him $50,000 to build a flying machine. The Smithsonian added $20,000. He spent it all on one full-scale Aerodrome and scaled it up from his models by calculation. A catapult threw the Aerodrome off a houseboat into the Potomac on October 7, 1903. It went into the river again on December 8. Two bicycle mechanics from Dayton spent about $1,000 over the same years. They flew gliders at Kitty Hawk for three seasons and crashed cheaply and often. The first two seasons went badly. Every wing designer then worked from the same published tables of how much lift a given shape would produce. Those numbers came from Otto Lilienthal, the German glider pioneer. The Wrights’ wings lifted far less than his tables promised. So in the winter of 1901 they built a wind tunnel from a wooden box and a fan and measured two hundred wing shapes themselves. Nine days after Langley’s second crash they flew. Langley bought two expensive answers. The Wrights bought several thousand cheap ones.

Jeff Bezos put the rule in his 2015 letter to shareholders. Some decisions are one-way doors. They are consequential and irreversible, and you make them slowly and with care. Most are two-way doors. Walk through one, look, and walk back if you were wrong. The next year’s letter added the number. Make most decisions with 70% of the information you wish you had. Waiting for 90% makes you slow.

Generative AI made one of the most expensive parts of a business cheaper: testing ideas and initiatives. Last year Procter & Gamble randomized 776 professionals to work on real product-development problems. Some worked alone. Some worked alone with AI. Some worked in pairs, and some in pairs with AI. Individuals with AI matched pairs without it. Pairs with AI were three times as likely to land in the top 10% of solutions. Both AI groups finished 12 to 16% faster. Without AI, R&D people proposed technical ideas and commercial people proposed commercial ones. With AI, both proposed balanced ones. One person could carry a team’s worth of first draft.

A prototype you could put in front of a customer used to sit at the end of a six-month process. Now it costs less than the meeting about whether to build it. Cheap prototypes are the leverage, not automation. Any company can now run the Wrights’ wind tunnel at the strategic layer: the price, the positioning, the product, the market you think you are in. Companies used to argue each of those from calculation because testing cost too much. Many had been two-way doors for years. They passed for one-way doors because of the price of walking through.

Cheap tests also change what thinking is. Feynman’s blackboard read at his death: what I cannot create, I do not understand. When the thing you are making talks back in an afternoon instead of a quarter, the plan stops being the way you think. It becomes the slowest way. From outside this looks like less thinking and more doing. From inside, the doing is the thinking. The version you can hold beats the version you can argue for, and it wins by more every year.

All of that is true. The cost of it lands on us.

When the price of trying falls toward zero, thinking before you try looks like waste. The reasoning is arithmetic, not cynicism. The company that cycles fastest wins. So the person who cycles fastest gets promoted, and everyone else learns to. Value used to be partly time. Now it is mostly time.

Hartmut Rosa spent a career on what speed does to a society. In his account acceleration runs as a loop. Faster technology speeds up social change. Social change speeds up the pace of life. The pace of life demands faster technology. Time-saving technology rarely produces time, because the gains get spent on more speed. The measure turns relative along the way: not good but better than the firm next door, not fast but faster than last quarter. A relative measure has no finish line. Hannah Arendt described the person that loop produces. Her homo faber judges everything by use, so each thing becomes a means to the next and the “in order to” swallows the “for the sake of”. She concluded that utility established as meaning generates meaninglessness. Fail fast is a good rule for a product. Practised for a lifetime it describes a life.

A loop that leaves no time to think about what just happened also works less well than it looks. At Wipro, call-center trainees spent the last fifteen minutes of each day writing down what they had learned. They scored 22.8% higher on the final assessment than trainees who spent the same fifteen minutes practising. Doing gives you experience and reflection turns that experience into learning, so a loop that cuts reflection to run faster ends up learning less. A whole field can lose the same way one person does. When researchers gave 293 writers story ideas from a language model, the stories got better and more alike, which the researchers called a social dilemma: each writer better off, the field worse off. Many companies now fail fast with the same model toward the same local optimum, so the market gets faster and narrower at once.

Speed is a discount rate. When you value time you are saying a result now beats the same result later, and the rate you put on that decides how far ahead you can see. The largest argument in the economics of climate change is an argument about that rate. In 2006 the Stern Review concluded the world should act hard and now, and William Nordhaus replied that the same models supported a slow ramp. They agreed on the models and differed on the rate: about 1.4% a year for Stern and about 5.5% for Nordhaus. At those rates a dollar of damage a century out is worth about 25 cents today to Stern and half a cent to Nordhaus. Most of the debate is not about what will happen, but about how much a person alive now gets to care about it.

Work at that speed long enough and you start measuring time in weeks. The things that decide a life move too slowly to show up at that scale. Climate runs on the scale of a century. The welfare of the next generation runs on the scale of a life. The person you are becoming runs on the scale of years. At those scales testing is hard and rolling back is harder. Aristotle says character builds the way a skill does: you do just acts until you are just and brave acts until you are brave, and in the end you are whatever you repeatedly did. If you repeatedly ship, measure and move on, that is what you become. A person is not a prototype, and there is no cheap version of yourself to try first.

Bezos wrote the 70% rule. He also put $42 million into a clock inside a mountain in West Texas, five hundred feet tall and built entirely from machinery. It ticks once a year. The century hand moves once every hundred years. The cuckoo comes out once a millennium. He wrote that over the lifetime of this clock the United States won’t exist, and that a species which can now build at civilization scale needs to think at that scale too. The man who taught business to move fast built a clock that moves once a year, and he was not contradicting himself. The letter already said it: two-way doors fast and often, with 70% of the information, and one-way doors slowly, with as much as you can get. Much of the skill is telling which door you are at. AI turned a great many one-way doors into two-way doors, and the right response is to walk through them faster than your competitors do. AI did nothing for the doors that were one-way to begin with. It only made them harder to recognize. You reach them carrying the speed you built up on the two-way doors, and you are through before you notice the difference.

Fail fast is the right rule for what you can try again. The climate, the next generation, the person you are becoming: none of them gets a second attempt, and they matter most. The last two essays argued that judgment is what remains for people once the procedure is automated. You spend it at the door, deciding which kind you are facing.