The idea came to Chris Edelstein at three in the morning in a Stanford computer lab that smelled of stale pizza and ozone from overheating CRT monitors. He had been training a neural network on movie ratings — fifty thousand users, two million ratings, the MovieLens dataset that everyone in the lab was using — and the network had started doing something unexpected. It was no longer predicting what users would rate highly. It was predicting what users would rate highly six months later, after they had watched the films. The gap between those two numbers was the interesting part. The gap was growth.

He called it the Bridge. The name came to him fully formed, the way ideas did back then when everything seemed possible and the hundred-billion-dollar companies were still being founded in garages by kids who hadn't finished their dissertations. The Bridge would connect people not to what they wanted but to what they needed — the film that would challenge them, the book that would expand them, the music they didn't yet have the ears to hear. Taste was a trajectory, not a point. You just needed the right recommendation engine to plot the curve.

This was Vector A. This was the pure state. Chris was twenty-eight years old, six months out of his Stanford CS PhD program, living on a futon in a house on California Avenue with three other founders. His co-founder, a business school dropout named Marcus Okonkwo, handled the money and the pitch decks and the meetings on Sand Hill Road where the venture capitalists sat behind mahogany desks in glass-walled offices and asked the same three questions: how big is the market, how fast can you grow, and what is your exit strategy. Marcus was good at those questions. Chris was good at the code. The division of labor felt natural.

At ninety percent Vector A, Chris coded through the night while Napster churned through his headphones — Radiohead, The Smashing Pumpkins, early Wilco — and the Iomega Zip drive clicked its hundred-megabyte songs of progress. He was building something genuinely new. The Bridge's first alpha test showed that users who followed the algorithm's "growth path" recommendations reported measurably broader taste profiles after three months. They were listening to jazz and classical alongside their grunge. They were reading nonfiction alongside their fantasy novels. The algorithm was making people smarter. The algorithm was working.

At eighty percent Vector A, the seed round closed at two million dollars. Marcus brought champagne to the office — real champagne, Veuve Clicquot, not the André they'd been drinking at house parties — and they toasted on the rooftop of their new office on Sand Hill Road, a converted dental suite with a view of the 280 freeway. Chris felt the shift like a change in barometric pressure. The term sheet was generous by 1999 standards: two million at an eight-million pre-money valuation, standard liquidation preferences, a board seat for the lead investor, a company called Endeavor Capital whose managing partner, a man named Alan Fisk, wore Hermès ties and spoke in aphorisms about hockey-stick growth curves.

"Taste trajectories," Fisk said at the closing dinner, turning the phrase over in his mouth like a wine he wasn't sure he liked. "It's an elegant concept, Christopher. It's also not what's going to make us rich. What's going to make us rich is engagement. Time on site. Daily active users. The numbers."

Chris nodded. He understood the language. He just didn't speak it yet.

At seventy percent Vector A, the product launched. The Bridge was a browser-based recommendation engine — you rated fifty items, and it gave you a feed of recommendations organized not by predicted enjoyment but by predicted growth. The tagline Chris wrote for the landing page read: "What you want is who you are. What you need is who you'll become." It was pretentious and he knew it, but it was true, or it had been true, or it was still true enough that he could look at himself in the monitor reflection without flinching.

The early users loved it. The early adopter problem, Marcus called it. The people who signed up for a taste-expansion engine in 1999 were already the kind of people who wanted their taste expanded. They were self-selecting for growth. The real test would be the mainstream users, the people who opened AOL and clicked whatever Yahoo put in front of them, the people whose tastes were stable and satisfied and unchallenged. Chris knew this. He coded features to onboard those users gently, to introduce unfamiliar recommendations like a friend offering a book you'd never heard of: "Trust me, you might like this."

At sixty percent Vector A, the first board meeting arrived. Alan Fisk sat at the head of the conference table, his Hermès tie loosened at the collar, a BlackBerry — one of the early ones, the 850 with the tiny thumb keyboard — resting beside his untouched bottle of Fiji Water. The quarterly numbers were projected on the wall from a Dell laptop connected to a projector the size of a small suitcase. Monthly active users: twelve thousand. Average session duration: fourteen minutes. The growth path was shallow but positive. The curve looked like the beginning of something.

"Twelve thousand users," Fisk said. "Pets.com has four hundred thousand. And they sell dog food."

"Pets.com loses money on every bag," Marcus said smoothly, because Marcus was good at this, Marcus was the shield. "Our unit economics are software margins. We just need time to —"

"I'm not asking for Pets.com numbers tomorrow. I'm asking about the engagement metric." Fisk tapped the screen. "Fourteen minutes per session is abysmal. YouTube isn't even a thing yet and I can already tell you that fourteen minutes is abysmal. Users should be on your platform for an hour. Two hours. You're supposed to be addicting them to self-improvement."

"The recommendations are designed to be consumed slowly," Chris said. "The idea is that you read the book, you watch the film, you digest it, then you come back. The loop isn't supposed to be addictive. It's supposed to be —"

"Profitable," Fisk said. "The loop is supposed to be profitable. We're going into Series A in six months. If your engagement numbers look like this, no one will touch you at a valuation that doesn't wipe out my position. Figure out how to make them stay."

At fifty percent Vector A, Chris added the Feedback Optimizer. It was a simple change, he told himself. The algorithm would still recommend growth-path items, but it would also surface one "comfort item" for every three growth items — something the user was guaranteed to enjoy, something that would make them feel good, something that would keep them clicking. The engagement numbers went up. Session duration climbed to twenty-two minutes. At the second board meeting, Fisk smiled for the first time.

At forty percent Vector A, Marcus came back from a meeting with a potential acquirer — a media conglomerate that was buying up internet properties the way a whale consumes plankton, indiscriminately and at scale. Their offer was insulting, forty million for a company that hadn't yet proven its model, but the number lodged itself in Chris's brain like a splinter. Forty million. If he owned thirty percent after dilution, that was twelve million dollars. He could endow a chair at Stanford. He could fund ten startups. He could stop coding at three in the morning and start sleeping like a human being.

He didn't take the offer. He just started thinking about it. That was the shift.

At thirty percent Vector A, the algorithm changed. Chris rewrote the recommendation core to optimize for session duration rather than growth trajectory. It was a technical decision, he told himself, just a different loss function in the neural network, just a different weighting of the collaborative filtering matrix. The Bridge stopped asking "what would make this user grow?" and started asking "what would make this user stay?" The two questions were related. They had to be related. Weren't they?

The recommendations shifted. Where once the algorithm had suggested a challenging documentary about climate change, it now suggested a slightly more entertaining documentary about climate change. Where once it had suggested a dense literary novel in translation, it now suggested a literary novel originally written in English. The growth was still there, just gentler, just slower, just more palatable. The users stayed longer. The engagement numbers climbed past thirty minutes. Fisk sent a congratulatory email that Chris read three times, searching for the part where it became a warning.

At twenty percent Vector A, Chris interviewed a user. Her name was Lydia, a thirty-four-year-old marketing director in San Jose whom the algorithm had identified as a power user. She had been on the platform for four months. Her taste profile had expanded significantly — she was reading philosophy, listening to world music, watching foreign films. The algorithm was working. Chris asked her how she felt about the recommendations.

"I love them," she said. "They make me feel like I'm becoming a better version of myself. Like there's this smarter Lydia out there, and the app is helping me catch up to her."

"That's exactly what we designed it to do," Chris said, and for a moment he was back at sixty percent Vector A, back in the pure state, back when the words meant what they said.

"But sometimes," Lydia continued, "I wonder if the smarter Lydia is real, or if she's just a reflection of what the algorithm thinks I should be. Like, am I actually growing, or am I just being told I'm growing? And does it matter?"

Chris opened his mouth to answer. He had no answer. The question was the answer, and the answer was that it mattered, it mattered more than anything, and he had spent six months making it matter less.

At ten percent Vector A, the Series A term sheet arrived. Endeavor Capital was leading again, joined by two other Sand Hill Road firms whose names were interchangeable collections of nouns — Summit Partners, Benchmark Capital, Accel. The pre-money valuation was sixty million. Chris would own twenty-two percent after dilution. On paper, he was worth thirteen million dollars. He was thirty years old, which in internet years was practically geriatric, and he had built something that millions of people used every day, and the recommendations they received were mostly things they already liked, with just enough novelty to make them feel adventurous without actually requiring them to be.

The closing was in three weeks. Chris sat in his office — a real office now, with a door and a window and a Aeron chair that cost more than his first car — and stared at the dashboard on his monitor. The numbers were beautiful. Daily active users: 340,000. Average session duration: fifty-one minutes. Monthly growth rate: eighteen percent. Everything was trending up and right, the way Fisk had promised, the way the term sheets demanded, the way the market rewarded.

He clicked on the recommendation feed. He scrolled through what the algorithm was showing his users. A new Tom Hanks movie. A Stephen King novel. The Beatles' greatest hits. A TED Talk about productivity. A documentary about sharks. A self-help book about habits. Comfort items. Comfort items. Comfort items. The growth-path recommendations were still there, buried three or four deep in the feed, but users rarely scrolled past the third item. The algorithm knew this. The algorithm had optimized for this. The algorithm was doing exactly what he had programmed it to do.

At zero percent Vector A, Chris tried to find the moment. He sat in his office at midnight, the 280 freeway a ribbon of headlights beyond the window, and scrolled backward through six months of git commits, Slack messages, board meeting notes, term sheets, product specs. Every decision had been small. Every compromise had been reasonable. Every step from A to B had been a single step, nothing you'd notice in isolation, nothing you'd call a failure of principle. You add a comfort item. You tweak the loss function. You optimize for session duration. You take a meeting with an acquirer. You don't sell. You just think about selling. You just price the sale. You just imagine the number. You just let the number become a reference point, and then you let the reference point become a goal, and then you let the goal become the point, and then one day you look at your dashboard and you can't remember what the point was.

There was no moment. There was never a moment. The vector between want and need was continuous. You were always at every position along it simultaneously. The algorithm was you, and you were the algorithm, and neither of you knew which direction the arrow was pointing.

Chris opened a new file in Emacs. He stared at the blank buffer for a long time. Then he started typing, not code but something else, something he hadn't written since the Stanford days when he believed that ideas could be pure and products could be good and money was just a way to keep the lights on while you did the work that mattered.

He wrote: "What you want is who you are. What you need is who you'll become. But what if the bridge between them is a mirror, and all you see in it is the person someone else built you to be?"

He saved the file. He closed the laptop. In the morning, he would sign the term sheet, and the Bridge would become worth sixty million dollars, and the algorithm would keep feeding people what they wanted while telling them it was what they needed, and Chris would own enough of it to never have to code again.

The fog was rolling in from the Pacific, the way it did in Palo Alto when the marine layer crept over the Santa Cruz Mountains. It looked different at midnight than it did at dawn. Or maybe it looked the same. Maybe fog was just fog, and you were just you, and the only thing that changed was which vector you were willing to look at.