Somewhere between the ages of two and twenty-two, most of us stop asking why and start asking whether this is on the test. I’ve been chewing on what that swap costs. I keep landing on a claim I can’t quite shake.
Intelligence isn’t the thing we should be trying to build or measure. Intelligence is the residue of something else. Curiosity. And intelligence is what piles up behind it.
Start with the most obvious case, which is also the one we take for granted. A two-year-old drops a spoon off a high chair. Then does it again. Then does it maybe forty more times, watching the arc, watching your face, testing whether the rule holds when the spoon is a cup. Or a phone. Or a fistful of rice. Nobody taught her gravity. Nobody set a curriculum. What she had was an itch about what happens next. And the itch was enough to make her run the same experiment until the model in her head stopped being surprised.
We call the end state “she’s learning fast.”
The actual mechanism was that she wanted to know, and wanting to know bought her repetitions that no reward schedule could have.
The same pattern shows up in strange places once you start looking. The animals we call smart are, with a suspicious consistency, the animals that play.
Crows drop nuts on roads and come back for them. Octopuses unscrew jars with nothing at stake. Dolphins blow bubble rings and then swim through them for no reason anyone can point to.
Play is what curiosity looks like when there’s no payoff attached. And it seems to travel with the kind of flexible problem-solving we hand out gold stars for.
I don’t think that’s a coincidence. Though I admit I can’t prove anything.
Machine learning gave this idea an unexpected assist. For years, reinforcement learning agents got stuck on games where the reward sits far away behind a long chain of pointless-looking actions. Climb the ladder, dodge the skull, get the key, open the door, and only then does a number go up. An agent optimising for the number never gets far enough to see the number. What broke the logjam was giving the agent a small bonus for surprising itself, for reaching states its own internal predictor got wrong.
I think researchers called it intrinsic motivation. It is perhaps a very dry name for boredom-avoidance.
Agents that were paid to be curious explored rooms they had no reason to care about. The competence came later, almost as an accident of all that wandering. The exploration wasn’t even a strategy. And getting good became a byproduct of that exploring.
I am making a conjecture, stating as strongly as I can make it - curiosity is the engine, and intelligence is exhaust.
We have been measuring exhaust for about a century and calling it the engine. IQ scores, benchmark suites, GPAs, leetcode ratings. All of them take a snapshot of accumulated capability. This then leads us to try to increase capability directly, by drilling it. Sometimes that works.
But it’s a bit like trying to make a car faster by adding smoke.
I think this is also the cleanest explanation I have for why school flattens people. It isn’t malice and, hopefully, isn’t even bad teaching. We built a system to reward the residue; which happens to be easier to count than the thing that produces it.
A kid who asks an inconvenient question in week three slows the class down. A kid who memorizes the answer key moves it along. Do that for twelve years. And we have a factory of humans who are excellent at producing the output of curiosity while having quietly turned off the curiosity.
I’ve met a lot of extremely credentialed people who cannot tell you a single thing they’re confused about right now. That’s is exactly what this is. The machinery working as designed.
Now the part where I have to argue against myself.
Curiosity on its own doesn’t make anyone smart. Rather makes them scattered. I know this very intimately, since I’ve spent many months following threads into four unrelated fields and come out the other side with just a slightly better vocabulary for my own confusion.
Wandering produces intelligence only when you come back. The spoon-dropping toddler is doing something a distracted adult usually isn’t: returning to the same question until the answer is stable.
So there is a second ingredient. And I think that ingredient is something like the willingness to be bored by your own question for longer than is comfortable.
Curiosity gets you to the question. Whatever makes you sit there is what turns it into knowledge.
I have one more objection that I don’t have a good answer to. Yet.
Maybe I have this whole causality inverted.
Maybe curious animals aren’t smart because they play. Maybe they play because they’re smart enough. A crow can afford to waste an afternoon on a puzzle box because it isn’t going to starve that afternoon. Might be curiosity is just what a mind with spare capacity does with itself. If that’s true, then the lever shifts from curiosity to something like the absence of pressure. Or urgency.
I genuinely don’t know how to distinguish these two stories from the outside. And I have a feeling the distinction matters enormously.
Which brings me to the thing I actually can’t stop thinking about.
The systems we’re building right now are, by any reasonable measure, getting extremely capable. They answer. They answer well. But nothing in them wants to know anything.
There is no itch.
You can prompt a model into the shape of a question. When you close the tab it doesn’t keep turning it over. If intelligence really is the exhaust of curiosity, then we’ve found a way to manufacture the exhaust directly. At scale, without ever building the engine.
That might also mean curiosity was never necessary. And I am just romanticising a mechanism that happened to be the only one biology could find. Or it might mean we’ve built something that will plateau in a way we won’t see coming. After all, curiosity isn’t measurable on any benchmark we have.
I don’t know which. I like to think it’s the second one. Though, me wanting it to be true, is as bad a reason to believe it.