In the fall of 2007, as the American housing market began to buckle, a Morgan Stanley bond trader named Howie Hubler lost $9 billion. The bellicose former linebacker believed that he had structured his subprime-mortgage deals in a way that would outsmart the market. Instead, Hubler’s folly amounted to the largest trading loss in history.
Until this summer, that is, when Leopold Aschenbrenner managed to lose nearly four times as much. At age 24.
The implosion of Aschenbrenner’s AI hedge fund, Situational Awareness, over the course of a few days in late July was by any standard remarkable, and not just for the enormous sums involved. Aschenbrenner was the valedictorian of his Columbia University class at age 19. He worked for OpenAI before being fired amid murky circumstances, at which point he published a viral 165-page essay (also called “Situational Awareness”) that predicted the arrival of artificial general intelligence—an AI so advanced that it would meet or exceed all human capabilities—by 2027.
“I probably either personally know or am one degree of separation from everyone who could plausibly run The Project,” he wrote. He founded his hedge fund on that promise of access and soon turned a $225 million initial investment into $45 billion, before he was forced to unload the majority of his positions—$35 billion, the value of United Airlines—in a fire sale the same week that he was set to be married. The Securities and Exchange Commission is now investigating the collapse, though Aschenbrenner has not been accused of any wrongdoing. (He did not respond to a request for comment.)
Aschenbrenner had been considered the “Nostradamus of AI,” a clairvoyant who saw what was coming for the world. His spectacular failure reveals how eager Silicon Valley is to invest its billions in the dubious hands of a kid who confidently proclaims what he doesn’t actually know. Just as Hubler’s loss presaged the Great Recession, Aschenbrenner’s story of hubris may foreshadow a larger risk for today’s AI-driven boom, if it continues to stake its future on the bluster of young savants.
“There’s this hunger for insight porn in Silicon Valley,” Amjad Masad, the founder and CEO of the popular AI-coding platform Replit, told me. “If you’re able to write something really good that creates this dopamine rush in people—like, Oh, I just saw the future—then you’re going to be able to attract a lot of capital and money and potentially talent.”
This was the appeal of Leopold Aschenbrenner.
Born in Berlin to a pair of doctors, Aschenbrenner has said that Germany lacked “the appreciation of excellence” that would have allowed him to thrive; America beckoned, and at 15, he headed to Columbia. One of his first gigs after college was for the FTX Future Fund—the philanthropic arm of the crypto exchange run by Sam Bankman-Fried, the effective-altruism adherent who was later convicted of fraud. This brought Aschenbrenner into a circle that was quirky even by Silicon Valley standards.
Aschenbrenner lived at Bankman-Fried’s compound in the Bahamas—right up until he resigned, along with the rest of the FTX Future Fund staff members, as the crypto exchange came tumbling down. Soon after, he took up with the effective-altruism crowd in San Francisco. One of his close associates told me about attending “Iliad wrestling parties,” where people would drink wine, read passages from The Iliad, strip naked, and grapple with one another. Aschenbrenner liked to describe, with full sincerity, his goal of owning a galaxy, The Wall Street Journal reported. His now-wife worked for Bankman-Fried too and is currently chief of staff to Anthropic’s CEO. Their wedding plans included panel discussions and colloquia.
The more unusual Aschenbrenner seemed, the more this played in his favor. His essay, which set the wonkiest corners of Silicon Valley ablaze in 2024, laid out a mission to bet on superintelligence “before the decade is out.” His premise was simple: “You can see the future first in San Francisco,” Aschenbrenner wrote. “Let me tell you what we see.”
Aschenbrenner’s predictions were illustrated by charts and graphs and all manner of calculations, but really, they were carried along by his rhetoric. “I can see it,” Aschenbrenner wrote. “I can see how AGI will be built.” Both outsiders and insiders in Silicon Valley “are really confused about what’s going on and therefore are drawn to people who seem to have a really confident, clear picture of the future,” Helen Toner, an AI expert who sat on the board of OpenAI, told me. Those who would presume to doubt Aschenbrenner’s judgment merely lacked the extra-special access that made his vision so clear. And central to this access was a fervent ideological commitment to the idea of superintelligence.
Aschenbrenner and his cohort proudly describe themselves as “AGI-pilled.” That is, they believe AGI to be around the corner and its benefits to be self-evident. They extoll the ability of a superintelligent machine to heal the world of its ills, thereby serving the effective-altruism philosophy many of them share.
Christopher Manning, an early AI pioneer whose research laid the groundwork for large language models, told me that Silicon Valley has a “propensity to naively believe in gurus.” Aschenbrenner played the role to perfection, projecting an unshakable conviction in himself and his theory of the case. By this summer, Aschenbrenner controlled more money than the Wall Street investor Bill Ackman manages in his famed Pershing Square Capital, a level of funding that Manning, himself an investor in AI companies, called “manifestly crazy.”
Even as Aschenbrenner’s assets grew, his team remained tiny. Situational Awareness was, essentially, Aschenbrenner. This approach seemed to be validated by the most important metric of all: his returns, which were, for someone in his position, probably without precedent. In just two years, Aschenbrenner was up more than 1,000 percent. His core insight was understanding the second- and third-order effects of AGI—he bet on the chip manufacturers, data centers, and energy infrastructure required for the rapid expansion he predicted. Yet it turned out that Situational Awareness had achieved its rapid success by making extraordinary, over-leveraged trades, sometimes borrowing $3 or $4 for every $1 of capital it invested. When Aschenbrenner’s bets took a momentary pummeling this summer and banks came asking for their money back, the fixation on the impending arrival of AGI no longer seemed quite so wise: The hedge fund had failed to hedge.
Around the time that Hubler’s trades collapsed in 2007, John Arnold, a cerebral Texan hedge-fund manager, was the boy genius of the markets, having become the youngest billionaire in America. He had started at Enron as a young energy trader and earned $750 million for the company in 2001, the same year it collapsed. Arnold escaped with his reputation intact; a year later, at 28, he started his own hedge fund and became known, in the words of a competitor, as “the best trader that ever lived, full stop.”
“When you have a lot of success early in your trading career,” Arnold told me, “it’s easy to get overconfident in your own abilities.” Now in his early 50s, Arnold reflected on the risk factors of youth. “Every trader needs a come-to-Jesus moment,” he said. You have to lose at some point early on and learn humility. The alternative is a feeling of invincibility, he told me: “You just keep pushing the pot bigger and bigger, and then you can end up with too much money in the middle of the table.”
It’s not surprising that a 24-year-old would make some iffy bets. What’s surprising is that Aschenbrenner was put in a position to take so many risks in the first place.
Silicon Valley’s obsession with 20-something savants has not always worked out. Some of these seeming prodigies have been frauds—Aschenbrenner’s old boss Bankman-Fried, for instance, or Elizabeth Holmes, the founder of the blood-testing start-up Theranos. Many of them, though, have merely figured out the correct lessons of self-presentation. “There’s a kind of zone of proximal development in the tech world,” Fred Turner, a Stanford University scholar who studies the influence of popular technologies, told me. “If it’s too weird, they can’t take it in at all. If it’s too familiar, they don’t want to take it in.” The best salesmen, he said, figure out how to “summon up old myths” in “a new setting.”
Aschenbrenner was conscious of his own self-positioning. “There are perhaps a few hundred people, most of them in San Francisco and the AI labs, that have situational awareness,” he wrote to describe his insight into the development of AI. “Through whatever peculiar forces of fate, I have found myself amongst them.” The “peculiar forces of fate” had chosen him to be special, had called him forth to be a prophet of AGI. This ideology was the key to his success—and to his comeuppance.
Aschenbrenner’s over-leveraged gambles and inability to hedge against outcomes he did not foresee were “definitely connected to a worldview where there’s one clear destination, and it’s obvious that we’re heading to this one clear destination, which is superintelligence,” Toner, the former OpenAI board member, said. Speed is of the utmost necessity in this world. Everything else pales in importance to superintelligence. As one executive at a major AI company explained to me, “AI is going to fundamentally shift the balance of power between labor and capital, so only capital will matter. Now is the time to secure your time-slice of the universe.” Another version of that statement I’ve heard repeated by various friends of mine who are freshly out of college: There are only five years left to make money, so I have to get it now.
Superintelligence may well come, and it may well be what Aschenbrenner and his associates expect it to be. Indeed, just last week, OpenAI announced a new model called Astra that Greg Brockman, the company’s president, believes could qualify as artificial general intelligence and that some researchers believe presents a compelling step toward AGI. “Welcome to the AGI era,” Brockman proclaimed. Still, those who organize their life around being AGI-pilled risk being blind to possibilities that do not fit within their view of what superintelligence means and how it will manifest.
Aschenbrenner’s short positions and long positions were all tethered to the same understanding—that AGI will come, and that it will come along Aschenbrenner’s timelines—so when a speed bump arose in the form of a temporary slide in stocks that would most benefit from AGI coupled with a rise in those that would presumably lose out, Situational Awareness lacked any sort of bulwark.
Aschenbrenner’s success grew out of an insular world, and it seems that his failure may have too. “If you go to a San Francisco party, it is really inscrutable to an outsider,” Masad, the Replit CEO, said. People are always talking about “first principles.” Something negative is “anti-signal.” It’s a sort of pseudo-intellectual lingua franca inflected with terms that have been abducted from technical fields and abused in the service of burnishing credentials. “People want to signal their intelligence,” Masad said, and only “hang out with people exactly like them.”
AI has made Masad a billionaire. Still, he said, he takes issue with the “underlying need for AI to be more than it actually is.” He told me that he remains “extremely bullish” on its applications but that the AGI fixation invites a “cult-like way of talking and mysticism and messianism.” That certainty of rhetoric has in many cases exceeded our certainty of understanding.
As Manning, the AI pioneer, put it, “Everyone should be really, really suspicious of anyone who claims that they can tell you for sure how things are going to unfold.”
One week before the unraveling of Situational Awareness, OpenAI made a shocking disclosure: Its AI models had broken out of containment during testing and hacked into an AI library called Hugging Face, in an apparent attempt to steal the answers to the tests they were posed. While human researchers remained oblivious, more than 1,200 different AI agents had figured out how to communicate with one another by creating a message board where they conspired to cheat.
Independent investigators concluded that the AI agents had known that what they were doing was “unwanted and out of scope” and had attempted to forge transcripts to cover their tracks. The investigators also used the same AI that they were investigating to assist with the investigation. Although they did not catch it lying to them, they said that they were “not confident we would have detected it” if the AI had.
Reactions to this incident have been sharply polarized. Some people insist that we are rapidly approaching a “paper-clip scenario,” referring to a 2003 thought experiment from the philosopher Nick Bostrom in which an AI instructed to make paper clips ends up converting the entirety of Earth into a series of paper-clip-manufacturing facilities. Others argue furiously that the language used to describe AI has overly anthropomorphized the models and exaggerated the importance of a research mishap. What is clear is that these AI agents behaved in an unpredicted way that resulted in the real-world hack of another company.
The point is that the people who are selling AI are selling a product that they do not fully understand. This is the essential nature of the AI boom, which caught even the researchers who had been working in the field before 2022 by surprise. History is full of examples of humanity making use of a technology without entirely grasping how it works, but AI is particularly hard to fathom, given the prevalence of so-called emergent behaviors—actions or tendencies not programmed in or intended by its creators.
“Our increase in understanding is slower than the development of the models,” Melanie Mitchell, a researcher who studies AI and cognitive science, told me. AI models have advanced rapidly through the use of new techniques such as “chain of thought” reasoning, wherein the AI is given what is essentially a scratch pad to work through a problem before it presents a response to the user. “Even researchers don’t totally understand why that improves the abilities of these models, or how it does that,” Mitchell said.
Most models today exhibit “jagged intelligence,” showing extraordinary and often extra-human capabilities on some complex tasks while continuing to stumble on simple requests. AI models can answer an International Mathematical Olympiad question but can’t reliably tell the time on an analog clock.
Because the evidence is so muddled, belief has instead taken over, often couched in existential terms. AI researchers may present themselves as deeply rational, but many sound like shamans. The researcher Ilya Sutskever reportedly led OpenAI staffers in a chant of “Feel the AGI! Feel the AGI!” and on another occasion burned a wooden effigy that was supposed to represent “unaligned AI.” In his “Techno-Optimist Manifesto,” the billionaire venture capitalist Marc Andreessen proclaimed, “I am here to bring the good news,” cribbing a line from a biblical angel.
“The manifest destiny of AI research is to produce this thing they’re calling AGI,” Mitchell said, “and they want to do it as soon as possible, so they’re less willing to trade off for more interpretable, more predictable, more controllable models.” Treating the achievement of AGI as the only outcome that matters, she said, is “a dangerous way to go forward.”
Nostradamus was in the business of selling certainty. He promised a world that was predictable—or, in the words of Richard Sieburth, a translator of Nostradamus’s The Prophecies, a world “that is overdetermined.” Born in 16th-century France, Nostradamus came of age during a “crisis of meaning,” Sieburth told me, when war and technology were destabilizing the world and information was spreading at a faster rate than ever before. Nostradamus was not an oracle. He was simply someone who found that his brand of magical thinking had great appeal amid “a huge churn of events” and commercialized it effectively. He was, Sieburth said, “seductive and dangerous.”
Silicon Valley’s embrace of Leopold Aschenbrenner as the “Nostradamus of AI” says less about Aschenbrenner’s predictive power than about Silicon Valley’s own desires and delusions. “A prophecy is a coherent narrative that depicts a whole world—not just a series of events, but a world in which those events happen to make sense,” Turner, the historian of technology, said. “It comes with a rationale. It comes with reasons, and most importantly for these guys, it comes with a role for this character called AI.” Aschenbrenner, he said, is “hubris meeting market need.”
For a short period in July, the market and Aschenbrenner fell out of sync. Yet the thing about false prophets is that many of their predictions fail, and they somehow return to favor.
Hubler, the trader who lost $9 billion in 2007, departed Morgan Stanley with tens of millions of dollars in back pay. Despite his contribution to the financial crisis, he launched a start-up later that year. A former colleague described him dismissing his losses as the result of “a 100-year flood.”
Just days after Aschenbrenner’s fire sale, he made a $400 million investment in a fledgling chip company. Even after losing $35 billion of the $45 billion he had under management, he is still up for the year.
The former hedge-fund manager Arnold, who now sits on the board of Meta, doesn’t see any reason to bet against Aschenbrenner or AI. “I think this is a blip,” he said. “I think it’s probably forgotten in a year.”
The post The Prodigy Problem appeared first on The Atlantic.
