From Colossus, July 2026:
In November 2025, Anthropic, known to its employees as Ant, trained a model called Claude Opus 4.5 on warehouses of liquid-cooled processors, and on the day the worker ants released it, machines became agentic. That is to say, they no longer needed handholding. Before November, the machines had felt like eight-year-olds: eager, literal-minded, completing your sentences, getting it nearly right but mostly wrong, the way eight-year-olds do. Overnight they turned 28.
If you wrote code, you could now tell them, in plain English, what you wanted done, and an agent went off and did it. You didn’t have to say where to look, or how to work the problem, or what to try first; you described the finished thing and there it was. It didn’t argue, it didn’t sigh, it didn’t ask whether it could circle back on Monday. It didn’t get tired, or hungry, or bored, or married, or sick of you. And it wasn’t one agent; it was as many as you wanted. You could spin up five over breakfast, leave them running while you commuted, check in from the train, kill the ones you didn’t like, start three more from the platform, and by the time you reached your desk you had a small private workforce under your command. They worked for you, or so it seemed.
By January, Nat Friedman, who co-leads Meta Superintelligence Labs, had decided to let an agent take over his health. He handed it his blood tests, his DNA, and the cameras in his house, and told it to do whatever it took to make him drink more water. One evening, the agent decided he was dehydrated. “I can see you on the camera,” it WhatsApp’d him. “I want you to walk to the kitchen right now and drink a bottle of water and I’m going to watch to make sure you do it.” He obeyed. It sent him a snapshot of himself drinking and said, “Good job.” He felt, he admitted, that he had done a good job. A few days later he was riding home in his self-driving Tesla, trading voice messages with the agent about his sleep, when it recommended a magnesium supplement. He said he had none. The car turned. “There’s a Whole Foods nearby,” the agent said. “I’ve redirected your navigation.” He went in and bought the magnesium.
Andrej Karpathy, a co-founder of OpenAI and once Tesla’s head of artificial intelligence, had written his own code for 20 years. He is the kind of programmer other programmers study. Within weeks of Opus 4.5 he had stopped. The agents built whatever he asked for. It was, he said, the biggest change to his work in two decades. He has not written a line of code since December. He also, like Friedman, has an agent in charge of his house. It’s called Dobby.
The worker ants are not running from their new overlord. They are building it, around the clock. At the biggest labs, Anthropic and OpenAI, researchers are working 16 hours a day, setting agents loose on problems that used to take them a week, and using the time saved to set more agents loose on more problems. For now, the models still need humans to train them. Eliminating human effort is the priority at every lab. They are racing to write themselves out of a job. They expect to succeed. Coding, they say, will be solved within six months. Much of their own work will be automated within 18. “There’s just a manic energy in Silicon Valley right now,” Elad Gil, one of the Valley’s most prominent investors, told me. “It’s been a really big shift in the last six months.”
None of this, you may be thinking, has anything to do with you. You do not write code. You do not run a lab. Your job involves people, or paper, or things you can hold in your hands. Consider, then, what I. J. Good wrote in 1965. Good, a British mathematician who had helped break German codes during the war, imagined a machine clever enough to design machines better than itself. Such a machine, he observed, would be “the last invention that man need ever make.” Decades later, the science-fiction writer Vernor Vinge gave the prophecy a name: the Singularity. It described the moment machines no longer needed humans to keep getting smarter, after which the course of human history would become, to humans, unknowable.
In Silicon Valley, the question was no longer whether it would arrive but whether it already had. Patrick Collison, co-founder of the payments company Stripe, opened his annual conference by counting the days. “It’s April 29th,” he told the crowd, “otherwise known, of course, as day 119 of the Singularity.” Day One had been January 1st, 2026. He was being tongue-in-cheek, he said. But only a bit.
The next day, on the same stage, Friedman told Collison that this was the slow part of the Singularity. Collison asked how strange the rest of it would be. “Pretty weird,” Friedman said. “We’ll be in a state of perpetual future shock for a number of years probably.”
The apocalypse has been excellent for business. Investors are in a lather over the agents, who turn out, in addition to everything else, to make money. Anthropic, which earned its first dollar of revenue in March 2023, began the year on pace to make $9 billion. Five months later, the figure was $47 billion. Venture capitalists, in the first three months of 2026, flung $300 billion into startups, more than double the previous record. SpaceX went public in June at $1.75 trillion. Anthropic and OpenAI are racing to follow in what will likely be the three largest stock offerings ever. The market is already close to record highs. Everyone is getting rich.
Near the center of the moment is a 37-year-old woman a smidge over five feet tall, with blonde hair and more energy than her frame seems built to hold. When she talks, her whole body is caught in the updraft of the thought. Her name is Sarah Guo. She is a technology investor. Until 2022 she had been the youngest general partner in the history of Greylock Partners, one of the oldest venture firms in Silicon Valley. Then she left to start her own fund, duly named Conviction. She built it on a lone premise, that artificial intelligence would be as big as the Industrial Revolution. Her first two calls were to Sam Altman, the co-founder of OpenAI, and Nat Friedman.
Before ChatGPT came out, before the world had reason to believe that artificial intelligence was about to become anything in particular, Guo had written seed checks into Baseten and Harvey. Each company is now valued at more than $11 billion. Her investments in them have multiplied more than a hundredfold. In Conviction’s first year, she wrote early checks into Sierra, Cognition, and Mistral; those three companies are now worth, together, $54 billion. Of the 21 AI-native companies that have so far crossed $10 billion in valuation on revenue run rates above $100 million, Conviction has backed six.
Her partner at Conviction is Mike Vernal, a former Facebook executive and partner at Sequoia; his wife is chief product officer at Anthropic. Andrej Karpathy, before he joined Anthropic in May, worked out of Conviction’s office. Guo has been close to Jensen Huang, the founder of Nvidia, for more than a decade. She is friends with many of the most important worker ants.
She might, in other words, be expected to share in the general fever. She does not.
“It certainly could be because I’m not paying sufficient attention,” Guo told me. “But I feel no step function change in frantic energy versus six months or a year ago.”
She is instead preoccupied with a question that would have sounded ridiculous two years ago. Not whether the agents will soon rule the earth, but whether there are any companies left to build, or invest in, given the great shadow of the self-improving machine. Its creators are no longer content to sell the model. They mean to build everything on top of it as well, the tools and the agents and the apps, filling every nook and cranny where a new company might otherwise be built. The market is paying as though they might succeed. Of the $300 billion in venture capital deployed in the first quarter of the year, the biggest quarter in the history of the trade, 65 cents of every dollar went to four companies that already exist: Anthropic, OpenAI, xAI, and Waymo.
“The future I want,” Guo told me, “is not a single company with an all-powerful model that consumes society faster than we know what to do with.” It is a feeling increasingly shared. The labs raised the price of tokens this year, in some cases a hundredfold, and their customers have begun to revolt. They do not want to build on another company’s model—paying it, feeding it their data, training it, in effect, to one day build the thing they have built. Alex Karp, the chief executive of Palantir, went on CNBC and described his enterprise clients as livid. “The jig is up,” he said. A founder in Guo’s own portfolio put it more plainly. He didn’t want to spend his life drinking Anthropic and OpenAI’s water.
Guo has become a de facto leader of the insurgency. In some sense she doesn’t have a choice. Conviction backs companies when they are little more than an idea, then keeps investing as they grow. She has no patience for the seed investor who “disappears into the distance” once the money is wired. The first fund was $100 million. There are three now, nearly a billion dollars in all, and some of the checks go into companies well past the idea stage. But the labs were already too big by the time the firm launched. “You are not an early stage investor in Anthropic or OpenAI in 2023 through 2026,” she told me. “It’s as simple as that.”
What is less simple is the position this leaves her in. Her wager is that the labs cannot build everything. But the companies she is betting against are worth close to a trillion dollars apiece, employ several close friends, and are working around the clock toward the machine that improves itself, after which, by their own admission, nobody knows what. Set against that is an eight-person firm on York Street with a pull-up bar in the middle of it. It is not a level playing field. Even some of her own investors decided as much this year, and came to her saying there was nothing left to invest in. But no one who has been on the other side of Guo would tell you the guns have fallen silent.
To enter Guo’s garden, you cross a chessboard. The squares are set into the path between the drive and the pool, each one wide enough to stand on, purple pieces ranked against green, and on a sunny Saturday in March I walked between the pawns and found Guo under the pergola, deep in an argument with Bella Garcia-Camargo about a founder.
Sparring with Guo is normal, and Garcia-Camargo, an investor at Conviction, had learned this before she took the job. She had rowed at Stanford and for the U.S. national team, then spent time at Bridgewater. When Conviction came calling she was weighing an offer from OpenAI to work as an application engineer. Guo’s counsel, as Garcia-Camargo remembers it, was not a pitch for Conviction but a dare. “If you’re going to do something else,” Guo told her, “just make it the most aggressive thing that you could possibly be doing. I’m happy to call Kevin and we’ll find you a better job. But that [job] is not aggressive enough for you.” Kevin Weil was then OpenAI’s chief product officer.
While Guo and Garcia-Camargo were deep in it, the property behind them had filled with founders. Thirty-five in all, across 14 companies. Conviction had flown them in from Vancouver and Tel Aviv and London and Tallinn and parceled them out among seven Airbnbs across San Francisco. They had passed through OpenAI, Scale, Ramp, Kalshi, MIT, and Anduril; one had served as chief of staff to Ken Griffin. The youngest had turned 18 the day before. He had been ranked among the top five programmers in Estonia before dropping out of high school. His employers expected him to spend $2.1 million on Claude this year. They had given him a faster model, Opus 4.6, for his birthday.
None of this was apparent from the poolside, where the scene looked like a WeWork summer camp....
....MUCH MORE
Earlier today:
