The Curiosity Engine

We’ve become very good at optimization. We optimize models, workflows, organizations, calendars, and inboxes, along with almost anything else we can measure. It’s become our default strategy for improvement, which makes sense. Optimization gives us a target, a way to track progress, and some confidence that we’re moving in the right direction.

A recent neuroscience preprint made me wonder whether intelligence may begin somewhere else.

The researchers trained artificial neural networks using two ingredients drawn from human cognition: a drive to explore and reduce uncertainty, and the biophysical constraints a real brain operates under. Neither one did much on its own. Together, they were enough for the networks to reproduce patterns of human synaptic development and features of adult brain architecture. The networks also did well on a harder test. Given tasks that recombined familiar pieces in unfamiliar ways, they solved them quickly, which is close to what we mean when we call someone a fast learner.

It’s an early result, and one preprint shouldn’t carry more weight than it can support. The question behind the work is still difficult to discount. We usually think of curiosity as something intelligence produces. People who think critically ask better questions because they understand enough to see where the gaps are. This research points the other way. The authors describe curiosity as a determinant of the network’s final architecture rather than a downstream consequence of it. On that view, curiosity is part of what builds intelligence, not just something that intelligence does once it exists.

That’s a different way of thinking about learning. Optimization begins with a definition of success. You choose an objective, measure performance against it, and improve the system over time. Curiosity begins without knowing where it will lead. It follows loose threads, spends time on questions with no obvious value, and accepts that much of the effort may go nowhere.

Organizations tend to reward the work they can see and measure. Execution produces milestones, metrics, and something to report at the end of the quarter. Curiosity rarely offers that. It consumes time without promising a result, and when a thread goes nowhere, the exploration is easy to dismiss as wasted effort. Even when it does lead somewhere important, the path is usually messier and less deliberate than the version we tell afterward.

Generative AI changes that relationship, because it can satisfy curiosity almost immediately. Questions that once sent us through books, papers, conversations, and experiments can now produce a polished answer in seconds. That’s an extraordinary capability, and in many cases it’s exactly what we need.

Speed also changes the experience of learning. The answer can teach us something, but so can everything we encounter while trying to find it. The search creates connections that weren’t part of the original question, and the wrong turns can become more valuable than the destination.

If the paper is pointing in the right direction, curiosity may be part of what creates intelligence rather than something it produces. As we build systems that can answer more of our questions, we may need to think just as hard about how they explore and what holds their attention. The question isn’t whether they can find answers. It’s whether they have any reason to keep looking once the answer they have is good enough.

📄 Curiosity Shapes Brain-like Architectures and Functions https://www.biorxiv.org/content/10.64898/2026.07.02.735826v1

Algorithm & Blues publishes on Sundays. If this was forwarded to you, you can subscribe at joefuqua.blog.