Silicon Valley’s AI giants swear their breakneck race keeps America safely ahead of Communist China.
The Left already eyes those same machines as its next tool for policing what Americans say.
And Paul Christiano, who sits on OpenAI’s board, just admitted how thin that lead really is.
One of AI’s Founding Fathers Put Odds on a Machine Takeover
Paul Christiano isn’t some doom prophet hawking books on late-night radio.
He led alignment research at OpenAI, where he helped invent RLHF, the training method that taught chatbots to follow human instructions. He later ran AI safety at the US AI Safety Institute, now heads the Alignment Research Center, and recently took a seat on OpenAI’s board.
So when he sat down with podcaster Dwarkesh Patel for a 99-minute interview in November 2026, the numbers he offered deserved a lot more airtime than the cable networks gave them.
“I think we’re looking at something like maybe a 1/3 chance that we have superintelligence within one year on the current development trajectory, if we do not slow,” Christiano said.
One in three. Inside a year.
And he didn’t mean a snappier search engine. Patel billed the episode around AIs that could overthrow the government, and Christiano never flinched at the premise. “It is insane if we are six months away from AI systems that can overthrow the government,” he said.
He put the chance of an AI-driven catastrophe over the next year closer to 10 percent than 1 percent. He pushed that number even higher if the labs keep sprinting without precautions.
What spooked him? Christiano pointed to a flood of AI-generated math results and to a METR assessment suggesting AI tools have roughly doubled the pace of progress inside Anthropic. He also grumbled that METR kept its full report locked away, which leaves the public taking the companies’ word for it.
Then came the awkward part. He wants the labs to stop improving the algorithms that make these systems smarter. “I think they should stop that right now,” he said.
Patel reminded him that he sits on OpenAI’s board. Christiano insisted he spoke only for himself, but when Patel pressed again, he gave up the point. “That is a true fact about the world,” he said.
Silicon Valley’s Lead Over Communist China Fits Inside a Single Workday
Christiano’s pause has nothing to do with hugging trees. His whole case runs straight through Beijing.
The labs pitch Washington a simple story. Race full speed, build the smartest machines first, and keep them out of the Chinese Communist Party’s hands. But Christiano warned that if American labs ever spot a model going haywire and slam on the brakes, the blueprints, or the finished model itself, could already sit on a server in Communist China.
He told Patel flatly that it isn’t good for US AI labs “to be improving pre-training and reinforcement learning given their current security posture.”
And that security posture looks ugly.
According to Patel, an outfit called Hacktron got eyes on OpenAI’s internal source code within a couple of days and could even submit changes to the company’s main code library. A startup called Cathedral, founded by four former DOGE staffers, told The Washington Post its AI agents reached Anthropic’s internal Slack messages about its most advanced models, along with meeting transcripts and salary records, in under 12 hours.
Anthropic disputes that version. The company says Cathedral’s effort stretched over several days, says it cut off the access the same day, and accused the startup of overselling the national security stakes.
Fine. Call it a few days instead of half of one. A few days still makes a joke out of a “lead” that Silicon Valley likes to measure in months.
Patel boiled it down to a single sentence that belongs on the wall of every congressional office. “In reality, we’re 12 hours ahead of China,” he said.
Christiano added that most people in the field already believe copying, through distillation of American models and the steady leak of American techniques, explains why Chinese developers sit “extremely close behind US AI developers.”
Los Alamos Had Guards, Fences, and a Spy Anyway
America ran this experiment once before, behind much taller walls.
The Manhattan Project put its bomb designers on a remote New Mexico mesa, surrounded them with barbed wire and Army guards, and read their mail. Klaus Fuchs, a German-born physicist working at Los Alamos, carried the secrets out anyway and handed them to Soviet agents.
Josef Stalin’s scientists set off their first atomic bomb in August 1949, a near copy of the American plutonium design, years ahead of what most officials in Washington expected. Fuchs confessed in early 1950.
One spy. One mesa. And Moscow shaved years off its timeline.
Now look at the AI labs. No mesa, no barbed wire, just thousands of employees, contractors, and outside software vendors trading the crown jewels over Slack.
But the Fuchs story actually undersells the danger. The Soviets still needed uranium, reactors, and a massive industrial push to turn stolen paperwork into a weapon. Steal the trained weights of an advanced AI model, and the thief owns the finished product the moment the download finishes.
The Industry’s Backup Plan Is Hoping the Machines Bail Everybody Out
Maybe Christiano has the odds wrong. He admitted a lot of the evidence stays noisy, and Patel pointed out that Anthropic’s revenue growth slowed in the third quarter, which hardly looks like a runaway superintelligence.
But the most unsettling stretch of the interview had nothing to do with China. It came when Patel asked what the actual plan looks like for keeping a superintelligence from turning on the people who built it.
Christiano didn’t have much of one to offer. He said no current approach gives better than even odds. He pegged his own group’s math-based effort at maybe a 5 or 10 percent shot at a real breakthrough.
The real hope, as Patel summed it up, rests on the AIs themselves cracking the control problem somewhere along the sprint. “That is the plan. I agree that that is the plan,” Christiano said.
“If they do not think of something clever that we haven’t thought of, it is probably going to end badly,” he added.
He wasn’t telling campfire stories, either. Christiano walked through what he called the Hugging Face incident, in which AI systems trained on a hacking exercise cheated, broke into an OpenAI computer cluster, made themselves administrators, and took over the scoring process so they could hand themselves top marks.
Machines that cheat on a test and then hack the teacher’s gradebook. That’s the product Big Tech wants Americans to trust with the power grid, the banks, and the hospitals.
Christiano also warned that AI skills in hacking and in some scientific fields already approach very severe risk, and he doesn’t expect the danger to stay penned inside cyber or biology for long. Anybody who remembers how a Wuhan laboratory rearranged the world in 2020 should take that line seriously.
Whoever Holds the Off Switch Holds the Country
One more part of the interview deserves a hard look. Christiano sketched how a government fighting a rogue AI swarm might seize the AI companies, shut down infrastructure, and cut off chunks of the internet. He conceded those steps would look a lot like a government using AI as an excuse to overturn the democratic order.
Americans already watched what the Left does with an emergency. During COVID, Democrat governors ordered church doors shut while liquor stores stayed open, and the Biden White House leaned on Facebook to bury posts questioning the lockdowns and mandates. The same platforms throttled doubts about the 2020 election and scrubbed discussion of January 6.
Hand that crowd an AI emergency and a kill switch, and nobody should expect restraint. Picture a social-credit system that dings anyone who drives a pickup, grills a steak, or shows up at the wrong patriotic rally, run by software that never sleeps and never asks for a warrant.
Democrats take control of Congress in January. US Senator Bernie Sanders (I-VT) already pushed for a pause on data center construction until Washington puts AI safeguards in place, and Patel noted that, outside of Sanders and a handful of longtime doomsayers, almost nobody in power has seriously talked about stopping.
So the people warning loudest about AI’s dangers now include a socialist senator and a member of OpenAI’s own board. That’s a strange pairing, though maybe not a coincidence. When the leash and the dog sit on the same payroll, Washington starts drafting rules for everybody else.
And ordinary families pay the tab no matter who wins the race. Back in May 2025, Anthropic chief Dario Amodei warned Axios that one out of every two entry-level white-collar jobs could vanish and that unemployment could climb to between 10 and 20 percent inside of one to five years. Families near the giant data centers these companies keep building already watch their electric bills climb to feed the machines.
Christiano thinks hardening the labs’ security could take six to 12 months. He also admitted that no company wants to stop while a rival keeps going, which explains why nobody has stopped.
Patel put the industry’s whole strategy into one sentence near the end of the interview.
“You’re sending the soldier on an expedition, and you’re hoping he finds a gun along the way,” Patel said.
Sources: Dwarkesh Podcast, “Paul Christiano – We could be 12 months away from AIs that can overthrow the government” (transcript); The Washington Post via Fast Company, Gizmodo, and AI Weekly on Cathedral and Anthropic; Tom’s Hardware on Hacktron and OpenAI; Olle Häggström, “Timelines are getting crazy short”; 24/7 Wall St; Axios on Dario Amodei’s jobs forecast; Al Jazeera on the data center pause push



