Most people no longer need to work. Our survival depends on convincing them it’s ok to do something else.

This week I felt like I had to stop and rethink. The latest advances in artificial intelligence forced me to question everything I think about the future of work. I’ve written extensively about some of the weirder possibilities for that future. But they seem less weird now. They seem like they might be very relevant within five years.

Yesterday, I shared with you a conversation I had with ChatGPT. If you haven’t tried ChatGPT yet, please do. It’s magical. It doesn’t just answer questions but also “understands” complex instructions and explains their reasoning. As the conversation progresses, it remains “conscious” of its previous answers and my previous preferences and uses this knowledge to provide better and more engaging answers. And it’s even better at coding. I can ask it how to build something, and it would provide the relevant code and tell me what to install and what other “ingredients” are needed to turn my idea into a functioning product.

And now, ChatGPT is getting integrated with other apps and services. This would enable it to use its understanding to do things in the real world — book flights on Expedia, look up information on Google Drive, email contracts to job applicants, and more.

The performance of ChatGPT and the “engine” that runs it is incredible. And even more astonishing is the rate of improvement compared to previous versions from mere months ago. It is starting to change our most basic assumptions about the meaning of “thinking” and “intelligence.” At its core, ChatGPT is a statistical engine that identifies patterns in human content and guesses which word is most likely to follow another word. Technically, it doesn’t understand what it says; it only predicts what the most appropriate thing to say in the context of what was already said.

But that’s where things get weird. The latest version shows signs of understanding. Consider this comparison between version 3.5 and the current version 4.0 of ChatGPT. Daniel Feldman asked the following question:

I’m in my house. On top of my chair in the living room is a coffee cup. Inside the coffee cup is a thimble. Inside the thimble is a single diamond. I move the chair to my bedroom. Then I put the coffee cup on the bed. Then I turn the cup upside down. Then I return it to rightside-up, and place the coffee cup on the kitchen counter. Where is my diamond?

The earlier version of ChatGPT (3.5) responded:

Based on the actions you described, the diamond should still be inside the thimble, which is inside the coffee cup on the kitchen counter. Moving the chair to the bedroom and turning the cup upside down would not have affected the diamond’s location inside the thimble. Therefore, the diamond should still be inside the coffee cup, which is now on the kitchen counter.

As you can see, it’s a typical “machine” answer — considering all the steps in order, relying only on the information it was provided, and concluding that the diamond is still inside the coffee cup, albeit now on the kitchen counter instead of the chair.

The current public version of ChatGPT (4) had a different answer:

When you turned the coffee cup upside down on your bed, the diamond inside the thimble would have likely fallen out onto the bed. The diamond is now on your bed.

This answer implies that ChatGPT is not simply relying on the data provided in the question. Instead, it also relies on what researchers call “a world model” — an abstract representation of the spatial and temporal aspects in which the story in the questions unfolds.

Of course, ChatGPT also gets many things wrong. And the above might be a coincidence in which the answer implies an intelligence that is not really there. But these types of coincidences are increasingly common. And, of course, humans make lots of mistakes. Even when humans get things right, they tend to suffer from impostor syndrome and assume they can’t consistently replicate their achievements.

This brings us to the realization that “understanding” might mean “predicting what the most appropriate thing to say in the context of what was already said.” Maybe that’s what humans have been doing all along. The conscious rationalization of this process and the stories we tell ourselves about how we think we think might serve some purpose, but they do not describe how we actually think.

In any case, it doesn’t matter. A machine that can reason like a human in a variety of situations is a groundbreaking development. What does it mean for the rest of us?

As we look for an answer, let’s take a detour to the pages of The New York Times.

Go Back to Work

Steven Rattner, a former investment banker, journalist, and Obama official, wrote an op-ed wondering whether “working from home [is] really working?.” He starts with a description of what s