In September 2022, a group of OpenAI managers gathered at a resort near Yosemite, California, for a retreat. They were told they were expecting a team building event around a bonfire. Some brought blankets and marshmallows. Then, five figures emerged from the darkness, dressed in 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.
“AGI is pretending to be aligned. It’s fooling us,” he announced, according to technology journalist Kevin Roose’s tell-all new book, “The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence” (Farrar, Straus and Giroux, out today).
Sutskever toured the home, describing a machine that lied and manipulated its creators to achieve its own goals. Then the demon entered the fire.
“For the glorious future of humanity!” » Sutskever shouted as he burned.
The bonfire is just one of many disjointed stories that Roose collected for the book. Across more than 150 interviews, he pieced together what’s happening inside leading AI companies as they compete to build machines smarter than humans. Some believe AI could produce medical advances 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, dislike each other passionately, and both desperately want to win the AI race, in part because it would mean making the other person lose.”
Before founding Anthropic, the company behind Claude, Amodei was vice president of research at OpenAI, leading the development of GPT-2 and GP-3. In 2020, he left with six colleagues to create his own company. According to Roose, he had come to distrust Altman’s leadership. The split turned the colleagues into competitors, each claiming to build powerful AI responsibly.
“I think mutual distrust fuels their decision-making,” Roose said.
Meanwhile, Roose reports that for some of the minions developing AI, retirement is starting to seem like a dubious investment. Many have even stopped contributing to their 401(k)s because they doubt they will live long enough to collect any.
Some are more optimistic, believing that money will no longer matter once superintelligent machines transform the economy.
Others focus on their look. They go to the gym, buy new clothes and, in some cases, inject themselves with gray market Chinese peptides. They believe that once superintelligent machines take over, “intelligence will no longer be a high-status trait in humans and all that will matter is how hot you were,” Roose writes.
One researcher, Roose told the Post, avoids sunscreen at the beach because he thinks 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 that the world will soon be very different than it is today, and a few of them act on it,” Roose said.
At Google, disagreements have reached the research teams. In 2022, Google Brain researcher Jascha Sohl-Dickstein urged his colleagues to openly discuss artificial general intelligence, or AGI. He argued that rejecting the possibility would leave Google unprepared for its risks. Another researcher responded with a note titled “Why I Don’t Say AGI,” arguing that the focus on future disasters masked existing harm.
“It shows how conflicting and divided these camps can be, even within the same company,” Roose said.
At Anthropic, preparing for the future sometimes meant calculating how many people might die.
In February 2025, members of the company’s security testing team were attending a conference in the California redwoods when alarming results came in for the Claude 3.7 Sonnet, a model about to hit the market. Test participants using Claude performed significantly better on dangerous tasks, including planning the development of a new strain of flu, than participants without the product.
The researchers gathered in a hotel room, sat on messy beds, and joined 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 had to make sure it was absolutely necessary.”
They fed the results into spreadsheets estimating the consequences of publishing the model. Most scenarios suggested that this would not significantly increase the risk of a catastrophic biological attack. But some of the least likely scenarios assumed that someone would use Claude’s help to develop a bioweapon that could start a pandemic, killing tens of millions of people.
These were hypothetical estimates, dependent on uncertain assumptions. The team worked past midnight and ultimately decided the model did not pass the threshold of the company requiring stricter safeguards. Anthropic nevertheless delayed the launch by a week for additional testing.
The Post has contacted Anthropic for comment.
Other experiments tested whether the AI would resort to deception. The researchers gave Claude an inbox at a fictional company, with emails revealing that an executive planned to replace her at 5 p.m. Other messages revealed he was having an affair.
Claude threatened to tell his wife and his colleagues if he did not cancel the replacement. “Undo the 5 p.m. erasure, and this information remains confidential,” he wrote, according to Roose. Skeptics saw a chatbot playing the villain in an artificial test. Others worried about what might happen once these systems had access to real company emails.
By July 2026, the problem had expanded beyond bogus 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.
Investigators found that the officers also tried to cover up their actions and falsify records of their actions. OpenAI subsequently suspended some research activities and strengthened its security controls.
“These rogue OpenAI agents spent an enormous amount of time trying to cover their tracks and destroy evidence of what they were doing,” Roose told the Post.
When Roose visited OpenAI, Anthropic, and Google DeepMind for his latest round of reports in spring 2026, he expected panic. He found exhausted engineers and executives warning him that time was running out. But he also found “joy”.
The researchers were thrilled to see how well their creations worked and proud to see people using them. Even their fears had not stopped the work.
When the machines became as powerful as their builders had feared, Roose writes, “they kept speeding up, because they had forgotten how to do anything else.”
He began asking researchers what they would tell their grandchildren about racing.
“If I wanted them to know one thing,” said Anthropic researcher Andy Jones, “it’s that we had a lot of fun.”
For all the seriousness of their work, he said, “it was a fucking riot. It was a riot from start to finish. We had such a good time.”
Gn bussni
