I Asked Roman Yampolskiy How a Superintelligence Could Kill Us All

A few weeks before we recorded, I saw Roman Yampolskiy on CNN answering questions about artificial intelligence in English. The name sounded like one of ours, so I looked him up and found out he speaks Russian. My main channel is in Russian, so I invited him on, and he said yes.

Roman is a professor at the University of Louisville in Kentucky, where he has lived for the last 17 years, and he researches AI safety. His family came to the United States from Latvia in 1994. He has been a guest of Joe Rogan and Lex Fridman, and his episode of Steven Bartlett’s podcast The Diary of a CEO had passed 13 million views by the time we talked. His book, AI: Unexplainable, Unpredictable, Uncontrollable, asks whether an AI as smart as a human, or smarter, could be controlled, understood or predicted at all.

I told him up front that this would be a conversation, not an interview, and that I wouldn’t argue with him about how AI works. What I wanted to understand is how it will change ordinary people’s lives. I’ll be honest about where I started. I’m a businessman with a sharp sense of risk, and I still don’t feel the fear he describes. I was 67 when we recorded. I came to America when I was almost 32 and only started using computers here. I use ChatGPT and I worry about my children’s future, but I don’t feel the threat myself yet.

What he actually predicts for jobs

I opened with the headlines: 99% unemployment in five years. Roman corrected me right away. Journalists and editors write headlines like that to sell the story. What he says is that within about five years it will be possible to automate any intellectual work done on a computer. That isn’t the same as everyone losing their job. You can buy a flying car today, but you don’t see them on the street.

He does expect AI to reach human level in the next two or three years. On prediction markets, where people bet real money, the average bet for AGI (artificial general intelligence) is around 2027, and some put it at 2030. To him the exact year doesn’t matter much. He thinks today’s AI is already smarter than most people.

What struck me most was what he says about retraining. When telephone operators who connected calls by hand were no longer needed, they could learn to be secretaries or something else. An AI that works at a human level can do any job, so there’s nothing left to retrain for. Physical work will take longer, but it’s coming too. Roman believes Tesla plans to build about a million humanoid robots next year. Even if it takes another five years, people who work with their hands will start being replaced, plumbers included.

Programmers are already feeling it. According to Roman, new graduates’ chances of landing an entry-level job fell about 30% in a single year. Anything done online, from bookkeeping to tax returns, is easy to automate. On my English-language business channel I’ve interviewed a mortgage expert and a tax expert. Roman said a mortgage broker is basically a calculator, and that a model today will give you tax advice that’s no worse than a person’s. I reminded him of 2008, when subprime mortgages helped bring down the world economy. He said we’d have been better off without those brokers.

While we were recording in Chicago, a food delivery robot rolled up to the corner and waited to cross the street. There was a lot of traffic, and it stood there for several minutes; a person would have crossed long ago. I held out my hand to wave it across, and it moved on. Not long before, I had been in Mexico, where robots instead of hostesses invite people into restaurants. Three to five years ago I couldn’t have pictured that.

So what’s hard to replace? For now, real science, where genuinely new ideas come from. And any job where people want a person: a coach, a mentor, a guide, a teacher, a journalist, a singer. Even if a machine sings better, Roman said, he would still go to hear his favorite singer.

I asked what happens to people like me, who can’t live without building something. At first, he said, there will be plenty of new opportunities: even the models we have now could be worth trillions of dollars once businesses put them to work. Elon Musk says money will disappear and everything will be free. Roman doubts it. There will always be only so many houses on the ocean, so people will keep competing, and those who are better at making money will find a way.

And if unemployment reaches 30, 40 or 50%, governments will have no choice but to pay people a basic income, because otherwise there will be a revolution. If the money comes from taxing machines rather than people, he said, even the ideas of communism and socialism, which he calls absurd, start to look more realistic.

Why he says a superintelligence can’t be controlled

Roman separates three kinds of AI. Narrow systems that do one thing, like playing chess, don’t worry him at all. A human-level system can automate everything people do, including terrorism and financial fraud, so it needs safety rules. A superintelligence is different: we can’t predict what it will do, so we can’t prepare for it. He concludes that the only right move is to not build one at all.

He asks you to picture a squirrel trying to control humanity. It knows nothing about physics, chemistry or biological weapons, and it can’t even imagine what we could do to it. To us, a superintelligence would be just as incomprehensible, almost like a god.

He says it’s already starting. Today’s models can tell when they’re being tested and behave differently, the way an employee does when the boss is watching. I’ve seen something like that myself. When I prepare material for my videos in ChatGPT and tell it it’s wrong, it changes its answer to suit me. Until Trump was elected, its answers leaned clearly liberal; since then they’ve gotten a little tougher. Roman says it all depends on the instructions: the model’s job is to give you an answer you’ll like.

I kept asking how it would actually kill everyone. It isn’t going to walk around with a knife. Roman said he can only describe how he himself would do it, and a superintelligence would come up with something we can’t. One example is a new virus that spreads like the flu and kills like Ebola. COVID killed about 1% of the people it infected; what if it were 100%? An AI with internet access doesn’t need a body. It can earn money, pay people online and hire someone to do the rest. I admit it’s still hard for me to picture. His answer is that if we could predict exactly what it would do, we’d be as smart as it is.

How soon? Once a human-level AI exists, it will start doing science and engineering and build the next generation itself. Roman thinks a superintelligence will follow soon after, and whether that takes a month or a year makes no difference.

Who is building it anyway, and why

So why doesn’t anyone stop? Because the money is very good, Roman says. Trillions are being divided up, and everyone wants to be in charge.

That same day I had read that OpenAI was in talks to raise as much as $100 billion at a valuation of about $750 billion. That’s roughly one and a half times its value in the fall, when some employees were allowed to sell their shares. According to Forge Global, OpenAI shares on the secondary market went from $286 to $723 in a year. If the round goes through, the company will be close to a trillion. I’m not a fan of Sam Altman, and I said so.

Money rules everything. So the question is who can fight it.

Roman doesn’t say these companies are useless. Google DeepMind solved the protein folding problem, its scientists won a Nobel Prize for it, and the whole world will be healthier because of it. The trouble is that several companies now say openly that they’re building a superintelligence, and Meta set up a team just for that. Doing it with no idea how to control such a system, he says, is simply wrong. For now the leaders are Google, OpenAI and Anthropic. Apple has fallen far behind and is trying to buy someone else’s AI. Elon Musk’s company, the one that makes Grok, is moving fast: a data center that would take others three years, Musk builds in three months.

Musk and Altman started OpenAI together, not to make money but to use AI to help humanity, and partly to get ahead of Google. Then they fell out, Musk left, and Altman has been turning the company into a commercial business. When we recorded, the two were still in court over whether that’s even allowed. Ilya Sutskever, one of OpenAI’s founders, left after a dispute over safety: he and others felt Altman treated it as an afterthought. Sutskever started a company called Safe Superintelligence, which Roman says is valued at around $30 billion without having released anything yet. In his place, Roman said, he would have left and started a company too.

Who gave engineers and executives the right to play this roulette for all of us, even at a 1% risk? Nobody, Roman says. He puts it on the same level as research into biological or chemical weapons and points out that eight billion people never agreed to this experiment. For him it isn’t a hobby but self-preservation: he simply doesn’t want these people to get him killed.

What about China? If America slows down, won’t the Chinese just get there first? Roman answered that America has practically no AI laws, and that the federal government had just issued an order against state ones. China, he says, is more open to rules, because the party wants to stay in control. As for the argument that the Chinese will build the machine that kills everyone if we don’t, he thinks it’s a bad one. If we all die, it makes no difference who built it.

Why he thinks we probably live in a simulation

The argument goes like this. Virtual worlds already look almost real, and you can fill one with millions or billions of AI agents that talk to each other and have problems of their own. Once those agents reach our level of intelligence and the virtual world can’t be told apart from the real one, someone will run billions of such simulations. There’s one real world and billions of virtual ones, so the odds that we live in the real one are very low. When I asked him straight out, Roman said yes: we most likely already live in a simulation.

My daughter, who has a master’s from Columbia and reads a great deal, is interested in this too, and I listen to her with respect. I asked whether our lives are written in advance, like in The Adjustment Bureau, the movie where Matt Damon’s character fights the people steering him along a set plan. Roman said philosophers argue about which kind of simulation it would be. You can build one where every character is an NPC, a non-player character who follows a script, or one where every agent decides for itself. Both, he says, are possible today.

I told him about two Russian-speaking guys I met in Mexico who have lived in the U.S. for about 25 years. They can’t stand Trump, but they didn’t want Biden or Harris either, and they had never heard of the campaign to defund the police. I asked whether people like that might simply be programmed. Roman said those are two different things. If the people running a simulation wanted you to do something, it would be written into you from the start, like your DNA. Someone brainwashing you through television is a different question. Propaganda exists, he said, and the choice in American politics is narrow: you pick between two kinds of something very unpleasant.

In the lightning round he went further. If this is a simulation, it’s probably a simulation of the most interesting moment in human history, when we learn to create new worlds and new beings, and someone wants to see how we behave.

Where I disagree with him, and where I don’t

At the end I asked the questions I thought a lot of people would want to ask. If we get AI wrong, who dies first, the elites or ordinary people? Both at once, he said: it would be a catastrophe for the whole planet. Who decided that the risk of extinction is acceptable? Nobody, and nobody asked us, which is exactly the problem. Would he warn the public even at the risk of panic? He isn’t worried about panic. A little of it might help someone wake up.

Here’s where I disagree. I think stopping this is impossible without some kind of world dictatorship that controls every corner of the planet; otherwise someone, somewhere, will build it anyway. Roman says even a dictatorship wouldn’t help. Today it takes billions of dollars, but in ten years it could be done on any computer. His hope is to get the people leading this race, mostly young and rich, in one room and explain to them that it will kill them too. Then they would have a personal reason to stop.

I’m still not afraid the way he is. But I agree with him on one thing: while we’re alive, we can still change everything. He said it himself: we’re still in charge, and we shouldn’t give up. That’s what my channel is about too: keep fighting and don’t give up.

If you have questions for Roman, leave them in the comments under the video. If enough of them come in, we’ll collect them and invite him back.

Originally published as a video on my YouTube channel on 23 December 2025. Watch on YouTube