Is AI Curious?

Random thoughts are something we all know. But does AI?

A train of thought can lead humans into a lot of rabbit holes. Some are useless. Some are funny. Some somehow become the thing that helps us understand a problem differently.

Today I was thinking about how to make a model learn someone over time. A few minutes later, I found myself reading about the history of why blue and pink became associated with boys and girls at gender reveal parties. Apparently, at one point, most babies wore white, and for a while people were not even consistent about whether pink was “for boys” or “for girls.”

I don’t think that particular curiosity got me closer to solving the model problem. It was definitely not the most efficient way to think. But it made me wonder whether these detours are part of what makes human thinking good in the first place.

Some of the smartest people I’ve met are also the most curious.

Not curious in the sense that they know a lot of random facts, although they usually do. Curious in the sense that you can put almost anything in front of them and they want to understand it. They ask questions. They keep pulling on the thread.

They are comfortable being confused and I think that matters more than we give it credit for.

Computer science is full of ideas that are uncomfortable to sit with. For example, huge parts of modern society are built on the assumption that some mathematical problems are difficult enough that an attacker cannot solve them in useful time. You start asking why RSA works, and you are already thinking about prime factorization, modular arithmetic, computational hardness, and what it even means for something to be secure.

Being unusually willing to sit with ideas like this, without immediately demanding that they become intuitive, is a superpower.

But in the process of sitting with ideas, you inevitably try to understand them through your own life context. That is where the rabbit holes start. We do not only search the knowledge directly associated with the problem. We pull from everything we have seen. The connection can seem stupid at first. But sometimes that unrelated thing gives you the useful abstraction.

If you solve a problem using only the standard knowledge of that field, there is a decent chance you arrive at roughly the same solution as everyone else in that field. Novel ideas often come from transferring a structure from somewhere else.

The important part is not simply knowing those other subjects exist. It is having spent enough time being curious about them that you have some mental model available when you need it.

Immanuel Kant defined genius as the ability to independently arrive at and understand concepts that would normally have to be taught by another person.

I like that definition because it makes genius sound less like knowing more and more like being able to reconstruct understanding for yourself. To do that, I think you need to relate ideas to your own life. You need to follow rabbit holes. You need to be curious enough to look at something unfamiliar, stay with the confusion, ask enough questions, connect it to things you already know, and eventually build the concept inside your own head.

Which brings me back to AI.

AI already has an exposure to an absurd fraction of recorded human knowledge. It knows about cryptography, dinosaurs, relativity, coffee history, supply chains, hike paths, compiler design, honestly everything.

In principle, it has all the ingredients for insane cross-domain connections; yet, when you ask AI to solve a new problem, it often produces something surprisingly conventional.

Knowledge is clearly not the missing ingredient.

Curiosity is.

Curious people wander. They have a search process that is inefficient. Most connections lead nowhere. But in that inefficiency is exactly where originality comes from.

That is Kant’s definition broken down. Intelligence is not only about absorbing concepts once someone else has organized them for you. It is about independently finding your way to them.

We spend a lot of time trying to make AI more knowledgeable, more accurate, and more goal-directed.

I wonder if we should also be trying to make it curious. Not just an AI that knows everything. An AI that can go down a rabbit hole and independently arrive somewhere new.