The ‘AI Takeoff’ Is Coming for Office Jobs — and the Future of Work

The world may be just months away from the biggest economic disruption in history.

“It could be the case that in six months from now, we have the first system that can recursively self-improve and create AI systems that we humans can’t even conceive right now because they’re so much more complicated,” says Anton Korinek, a professor at the University of Virginia’s Darden School of Business and one of TIME magazine’s 100 most influential people in AI.

“Once we have that, all cognitive jobs in the economy will suddenly be at risk and prone to automation.”

Korinek has a name for this tipping point. In the latest episode of UVA’s “Hoos in STEM” podcast, hosted by mathematician and STEM Advisor to the Provost Ken Ono, he described the moment where AI systems’ capabilities rapidly advance to the point that they soon outperform humans in most cognitive tasks as the “AI takeoff.”

It’s a future that feels closer by the day. But Korinek cautions that the timing is anyone’s guess — and uncertainty abounds.

“We don’t know when it’s going to happen,” he said. “We don’t know for sure if it’s going to happen, although it looks increasingly likely, and we don’t know if it’s going to play out within a couple of days, a couple of weeks, months, years.”

Not everyone thinks the future is still ahead of us. Sam Altman, OpenAI’s CEO, believes the moment has already arrived. “We are past the event horizon; the takeoff has started,” he wrote on his blog in June. “Humanity is close to building digital superintelligence, and at least so far it’s much less weird than it seems like it should be.”

What is clear, though, is that we are entering a new chapter — one in which advanced AI systems are helping to build even smarter ones, accelerating progress with the power to reshape our economy, society and everyday lives.

A New Industrial Revolution for the Mind

While this may sound alarming, Korinek points out that we have been through many episodes in history where automation caused a significant disruption in the labor market — people lost their jobs and the economy adapted.

“We economists have been telling people these stories for 200 years that we shouldn’t be worried about automation, because we have always created new jobs and there have been difficult adjustment periods, but after that adjustment, society as a whole was better off and living standards rose,” he said.

But, Ono wondered, is that still true today?

“That’s the trillion-dollar question,” Korinek replied. “I’m not so sure.”

He explained that in the first century of the Industrial Revolution, we automated a lot of physical labor and built superhuman machines that could manipulate the physical world. And then in the middle of the 20th century, we started, at first very slowly but then progressively faster, the process of automating cognitive labor.

At first, that meant automating routine tasks that perhaps people didn’t appreciate as much. A good example is humble — but ubiquitous — spreadsheet.

“From an economic perspective, the spreadsheet revolutionized the work world because so much of what companies do when it comes to cognitive labor is essentially manipulating numbers,” says Korinek.

While we have had many advances in computing that performed routine cognitive tasks, he adds, progress was relatively slow over the past few decades, and labor markets were able to adjust and keep up with the job destruction by creating new jobs we couldn’t even dream of 100 years ago.

 “Now it feels like things may be different,” he cautions.

The Next Leap: When AI Learns Without Us

One of the reasons is that when we automated physical labor, it made our cognitive labor more valuable.

“If you look at our economy at large, there are goods and services being produced, and they require both physical and cognitive labor. And as humans, we rely on both. We need the brain, and we need the brawn in order to keep us alive and in order to keep us enjoy[ing] life,” says Korinek. “We have automated a lot of the brawn, and that means the brain became scarcer, and because of that, we have been preaching for the past few decades ‘Oh, we need to educate. We need more and more cognitive workers in the economy, because that's what the machines can't do yet.’”

But when the moment of “AI takeoff” arrives, all cognitive jobs in the economy will be at risk, he warns.

That, in turn, will put a premium on physical labor.

“We will see the opposite of what I described before in the 20th century, when we had automated so much of the physical labor, the cognitive labor, became more valuable,” Korinek says. “If we now automate all the cognitive labor, the physical labor is going to become more valuable, and it's going to create massive economic incentives to advance robotics even more quickly than it has already been advancing the past few years. And I think we are seeing robotic systems that have kind of general human-like capabilities already in the labs right now.”

The challenge, he argues, is to plan for uncertainty rather than cling to precedent. “My advice to all decision-makers is to engage in scenario planning for this,” he says.

The Limits of Today’s AI

For now, there are limits to current AI systems. Korinek sees two broads areas: the first has to do with lagging vision capabilities. “They just can't do what we can do, and that's a big handicap in things like you may have seen this computer use applications where, essentially, the AI tries to take over your desktop and perform work for you on your screen, with you just watching. They need better eyes for that — they need better vision processing,” he says.

The second has to do with the learning process.

 “Whenever you open a new chatbot window, it forgets everything it has done before, and so they do not dynamically learn, and that's a significant shortcoming compared to humans, and that limits their economic usefulness right now,” says Korinek. “But of course, there are major efforts underway to address that.”

He illustrates the point with a bird-and-plane analogy.

“When we built flying machines, there were very smart people like Leonardo da Vinci who thought, ‘Oh, we need to create good wings that flap’. But in the end, that’s not how the Wright Brothers solved the problem of flight,” he said. “In the space of artificial intelligence, it’s been the same, and it will continue to be the same. The way that the AI solves problems is going to be different from the way that we humans do.”

The big question now, he adds, is: “With this dynamic learning, for example, how much of that needs to occur in the precise same way that our brains do it, which is to constantly update their brains in real time, and how much of it can be done through some different process that leverages the specific strengths that the AIs have that our brains just would have never had access to due to our biological evolution?”

Preparing the Economy for Superintelligence

To better understand and prepare for this shift, Korinek recently launched the Economics of Transformative AI (EconTAI) Initiative at UVA — a research center focused on understanding and preparing our economy for AI systems that may soon match or exceed human-level intelligence.

“Right now, we are in this unique moment in time where the next couple of years are going to determine a lot about where AI will go in the future,” he said. “I think it's a somewhat frightening time, but it's also an exciting time, because there are so many urgent and important questions that we need to tackle. Those are technological questions, but there are also social questions. And for me, as an economist, [there are] economic questions.”

His course, “AI and the Future of Work,” invites students to study the advances that are at the very frontier of AI — “because one of the most fundamental and most important things is to be aware of what’s happening. That awareness is kind of a secret power — [it] allows you to see further into the future and to make better decisions about the future.”

You can listen to the full episode of “Hoos in STEM,” hosted by Ken Ono, and browse the archive for more content.