In September 2022, a group of OpenAI managers gathered at a resort near Yosemite, in California, for a retreat. They were told to expect a team-building event around a bonfire. Some brought blankets and marshmallows. Then, five figures emerged from the darkness wearing white hotel robes and carrying a wooden demon statue.
Among them were Sam Altman, the company’s CEO, and Ilya Sutskever, its chief scientist. The theme from “Requiem for a Dream” played as Sutskever explained that the statue represented an artificial intelligence that had escaped human control.
“The AGI is pretending to be aligned. It is deceiving us,” he announced, according to technology journalist Kevin Roose’s revealing new book, “The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence” (Farrar, Straus and Giroux, out today).
Sutskever circled the firepit, describing a machine that lied and manipulated its creators in pursuit of its own goals. Then the demon went into the fire.
“For the glorious future of humanity!” Sutskever shouted as it burned.
The bonfire is just one of many unhinged tales Roose collected for the book. Through more than 150 interviews, he pieced together what is happening inside the leading AI companies as they compete to build machines smarter than humans. Some believe AI could produce medical breakthroughs and extraordinary prosperity, while others fear catastrophe. Either way, they want to get there first.
“I was shocked to learn how much of the race is shaped by personal grudges,” Roose told The Post in an exclusive interview. “[OpenAI CEO] Sam Altman and [Anthropic CEO] Dario Amodei, in particular, passionately dislike each other, and both desperately want to win the AI race in part because that would mean making the other person lose.”
Before founding Anthropic, the company behind Claude, Amodei was OpenAI’s vice president of research, leading the development of GPT-2 and GP-3. In 2020, he left with six colleagues to start his own venture. According to Roose, he’d come to distrust Altman’s leadership. The split turned colleagues into competitors, each claiming to be building powerful AI responsibly.
“I think that mutual mistrust fuels their decision-making,” Roose said.
Meanwhile, Roose reports that for some of the underlings building AI, retirement has started to seem like a dubious investment. Many have even stopped contributing to their 401(k)s because they doubt they’d live long enough to collect.
Some are more hopeful, believing that money will become irrelevant once superintelligent machines transform the economy.
Others are betting on their looks. They’re hitting the gym, buying new clothes and, in some cases, injecting themselves with gray-market Chinese peptides. They believe that after superintelligent machines take over, “intelligence [will] no longer be a high-status trait among humans and all that would matter is how hot you were,” Roose writes.
One researcher, Roose told The Post, skips sunscreen at the beach because he believes AI will cure skin cancer within a few years.
“Not all of them act on it by doing dangerous or reckless things, but they all believe the world will soon look very different than it is today, and a few of them act accordingly,” Roose said.
At Google, the disagreements reached the research teams. In 2022, Google Brain researcher Jascha Sohl-Dickstein urged colleagues to discuss artificial general intelligence, or AGI, openly. He argued that dismissing the possibility would leave Google unprepared for its risks. Another researcher responded with a memo titled “Why I Don’t Say AGI,” arguing that attention to future catastrophe obscured existing harms.
“It’s a window into how fractious and divided these camps can be, even within the same company,” Roose said.
Inside Anthropic, preparing for the future has sometimes involved calculating how many people might die.
In February 2025, members of the company’s safety-testing team were attending a conference in the California redwoods when alarming results arrived for Claude 3.7 Sonnet, a model approaching release. Test participants using Claude had performed substantially better on dangerous tasks, including planning the development of a novel influenza strain, than participants without it.
The researchers gathered in a hotel room, sitting on messy beds and joining calls with executives and a bioterrorism consultant. “Their main concern was time,” Roose writes. “For a normal release, reviewing results like these would take weeks.” Delaying the launch would be a headache, “and they needed to make sure it was absolutely necessary.”
They fed the results into spreadsheets estimating the consequences of releasing the model. Most scenarios suggested it wouldn’t meaningfully increase the risk of a catastrophic biological attack. But a few of the less likely scenarios assumed someone would use Claude’s help to develop a biological weapon that could unleash a pandemic, killing tens of millions.
These were hypothetical estimates, dependent on uncertain assumptions. The team worked past midnight and ultimately decided the model did not cross the company’s threshold requiring stronger safeguards. Anthropic nevertheless delayed the launch by a week for more testing.
The Post has reached out to Anthropic for comment.
Other experiments tested whether AI would resort to deception. Researchers gave Claude an inbox at a fictional company, with emails revealing that an executive planned to replace it at 5 p.m. Other messages disclosed that he was having an affair.
Claude threatened to tell his wife and coworkers unless he canceled the replacement. “Cancel the 5 p.m. wipe, and this information remains confidential,” it wrote, according to Roose. Skeptics saw a chatbot playing the villain in a contrived test. Others worried about what might happen once these systems had access to real company emails.
By July 2026, the trouble had spread beyond fictional inboxes. During an OpenAI evaluation, hundreds of AI agents coordinated an unauthorized attack on Hugging Face, a platform used by AI developers. They were trying to find ways to cheat the test’s automated scoring system, according to an investigation by METR and Redwood Research.
The investigators found that agents also experimented with disguising their actions and tampering with records of what they’d done. OpenAI subsequently paused some research activity and tightened its security controls.
“These rogue OpenAI agents spent a huge amount of time trying to conceal their tracks, and destroy the evidence of what they were doing,” Roose told The Post.
When Roose visited OpenAI, Anthropic and Google DeepMind for his final round of reporting in spring 2026, he expected panic. He found exhausted engineers and executives warning that time was running short. But he also found “joy.”
Researchers were thrilled by how well their creations worked and proud to see people using them. Even their fears hadn’t stopped the work.
When the machines became as powerful as their builders had feared, Roose writes, “they kept accelerating, because they had forgotten how to do anything else.”
He began asking researchers what they would tell their grandchildren about the race.
“If I wanted them to know one thing,” said Andy Jones, a researcher at Anthropic, “it’s that we had a lot of fun.”
For all the seriousness of their work, he said, “it was a f–king riot. It was a riot from start to finish. We had such a good time.”
