About a year ago, I started using the term "Proof Boom" to describe a new era in mathematics. AI would lead to an unprecedented increase in the production of mathematical proofs, fundamentally changing how mathematics is done.

Perhaps people working in frontier AI labs saw this coming before I did, but many mathematicians have only recently started to grapple with it.

I mention this for a simple reason: some era-defining developments can be seen before they fully arrive, if we look at them from the right perspective and reason from where they appear to be heading.

I think the same kind of change is now coming for intellectual work more generally.

A lot of people are naturally wondering about what AI means for their future. Until recently, I'd spend most of my answer on what the technology could and couldn't do. These days my answer is shorter: start making the preparations you'd make if AGI—artificial general intelligence, broadly meaning AI with human-level capabilities across a wide range of cognitive tasks—were already here. I am not claiming that a private model has become a reliably general intelligence. I certainly do not think this is the case for any single public model today.

I'm saying that the possibility is now concrete enough to shape decisions you make today. Some of those decisions, like what to study or which career to pursue, will matter long after today's AI models have been replaced by much more capable ones.

Working backwards

Game theorists often use a concept called backward induction to solve dynamic games. You start at the end of a game and reason your way back to the present. It only works if the game has an end, though. In an infinitely or indefinitely repeated game, there's no final round, so the argument doesn't work.

Our thinking about future events has a similar flavour. "Perhaps machines will reach AGI" is too vague to prepare for now. However, when a future event becomes sufficiently concrete for practical purposes, we can ask what we would need to have done before reaching it.

The future I have in mind is one in which AI systems can do almost any intellectual task humans currently do.

This is what I think has now become a concrete endpoint—or a new starting point, whatever one may call it. Such a concrete endpoint gives us a point from which to make the backward-induction calculations.

If a large part of your job depends on producing a particular intellectual output, what would happen if AI could produce it more efficiently? What would you need to learn and do differently?

I believe it is time to backward induct from an AGI event.

Specialisation

That said, AGI may not be a single machine. It's not clear at all that a lab capable of producing true AGI would choose to build it.

A seemingly safer strategy is to build domain-specific AIs that don't have general capabilities. However, it's possible that a single AI may have emergent capabilities that domain-specific AIs don't have. This could provide incentives for building a general AI and making sure it is built safely.

I want to make the point that specialisation can create great wealth, as the founder of modern economics Adam Smith noted: production can be maximised through the division of labour. The state provides institutions and safety, and prices help with coordination.

Thus, instead of a single AGI machine, we may see an AGI system in which coordination is achieved with humans in the loop.

There is something slightly counterintuitive about this. We tend to imagine progress towards AGI as progress towards a single system. But the economy works largely in the opposite way.

The same may happen with intelligence.

A collection of specialised AIs, humans, and institutions could collectively have capabilities that none of its components has alone.

There may never be a single AGI moment, just as there was never a single Industrial Revolution moment in England. Similarly, perhaps future historians won't identify a single AGI moment and instead will say this was all part of the electronic computer revolution.

Benefits of superhuman machines

When making your plans, please consider that in the human-vs-AI debate, while humans' current skills and knowledge are essentially fixed, humans are an incredibly adaptable species and can learn from AIs. So a system reaching human capabilities today doesn't mean that humans will immediately become obsolete tomorrow.

As AI systems substitute for some of our output, working with such systems will also change what we can achieve. This is one reason for optimism.

The next generation matters even more. A kid who grows up with advanced AI systems can develop new skills compared with many of us who encountered them halfway through our careers.

So there are several different AGI moments. AGI could be reached with respect to humans today. There is also AGI with respect to humans who work with AI. And there will be a different AGI threshold with respect to future humans who grow up learning from these first AGIs.

Even if machines eventually possess a persistent absolute advantage across all intellectual activities—meaning that even a new generation of humans cannot add anything to them despite growing up with them—it does not immediately follow that humans cannot produce economically meaningful outputs. As Ricardo famously illustrated, mutually beneficial trade can occur even when one party has an absolute advantage in producing everything, because what matters for specialisation is comparative rather than absolute advantage.

That said, comparative advantage does not guarantee that your current job or salary will exist. And it certainly doesn't guarantee that the wealth created by AI will be distributed equally.

In any case, humans will have to adapt to a new division of cognitive labour. Some tasks today will be entirely automated; others will be shared by humans and machines. Through creative destruction, entirely new economic opportunities may arise. And as humans, we will have to adapt to the new reality.

As they say, it is hard to make predictions, especially about the future. We cannot know exactly what that equilibrium looks like. That is another reason why I don't think the right response is to guess which professions are "AI-proof."

There may be no such thing.

A better question is: what becomes valuable when the thing I currently provide becomes abundant?

The Human Experience

This brings us to another question that has less to do with economics.

What happens if machines simply become better than us at every intellectual task?

Chess already gives us one answer.

Computers became much better at chess than humans. Deep Blue beat Garry Kasparov in 1997. However, after that there was a period when humans working with machines were better than machines alone. That period largely ended by the early 2010s. Until very recently, there were still certain chess positions that humans understood better. However, Magnus Carlsen recently said that there is no position he understands better than the best chess engines. It's game over for humans. Similarly, we are in a transition period in which, on many tasks, humans with AIs can produce better outcomes than AIs alone.

Despite all these AI developments in chess, humans didn't stop playing chess. That's because along the way we have come to realise that producing the best possible move was never the whole point.

We care about the human aspect of the competition and the process of learning and improving. In fact, computers made learning chess more democratic, and humans all over the world learned from them. Today's strongest players understand chess better than past players in part because they grew up with superhuman engines.

I think this distinction will become much more important.

There will be things we ask machines to do because we simply want the best possible outcome. And there will be things we continue doing ourselves because the process of doing them is the important thing.

Let's call the latter the Human Experience.

In a good scenario of beneficial AI, I think the Human Experience will become abundant. Games will be abundant. People will continue writing, learning, debating, and researching even when a machine could produce a better result.

We already accept this in physical activities. A car can travel much faster than Usain Bolt, but that doesn't make the Olympics less interesting.

The same may increasingly become true of intellectual activities.

It is the time

So what should you do?

I think that waiting for a concrete AGI moment to start thinking about it is the wrong strategy.

It is now time to backward induct from the AGI moment and take seriously the point at which most, if not all, intellectual outputs will become much cheaper to produce.

It is the perfect time to ask what you would want to have learned before that happens. Ask how to use these systems to produce more with them and to learn more from them.

And don't assume that today's human capabilities are final. The new generation will strike back.

Maybe a single AGI moment will eventually occur. Maybe it won't. Perhaps future historians will argue endlessly about when the threshold was actually crossed.

That doesn't matter very much for the decision I'm talking about.

When a future event becomes concrete enough, it can become rational to change our behaviour in anticipation of it.

I think we have reached that point with AGI.

Thus, rational people should start planning as if AGI has arrived.

Good luck, humanity.

Further reading