Why do science? | September 2026
Our capacity for invention depends on our understanding of the natural world. We learn about the natural world by conducting experiments. The more experiments we do, the more we learn about the shape of the world; its edges, the space that remains open to play, and the particular rules that govern that space. Our ability to create functional technologies fundamentally depends on our understanding of these three things, and scientific progress is a function of the volume and variety of experiments that get run as we learn about them.
Today’s frontier sits disproportionately in domains governed by emergence, contingency, and combinatorial search space–biology, materials, neuroscience, ecology–areas that don’t lend themselves well to theory. For example, we can compute how tightly a molecule binds to a protein but that won’t tell us whether the drug works in a person. You’d need to conduct an experiment to find that out.
Models like AlphaFold are certainly accelerating these fields. But they rely on humongous amounts of data–data derived from experiments. That makes each individual data point more valuable than it was before, and raises the bar for the reliability of that data.
Automated labs have the potential to dramatically increase the volume of experiments run. But they are expensive to set up, the set up limits experimental variability, and the cost of other materials, like reagents, are still something to account for. And crucially, robots don’t think, lest in divergent ways.
Variety is important because you need to be able to understand the limits of what you’re investigating–where it breaks, and where it behaves differently. This allows you to understand how generalisable something is and the rules of play. You find these out by stepping outside the conventional understanding of a phenomenon. When you do this for phenomena we understand well, it’s more likely you’ll get null or negative results by doing this. But these null and negative results are information themselves. And crucially, this data makes for more comprehensive models. If you only trained a model on positive results, it wouldn’t understand the specifics of a failure and therefore couldn’t reach for something new. You’d get a model that’s very good at telling you what you already know, and results that mimic what our scientific institutions prioritise; not results that reflect how the world actually works outside of these biases.
This post is part of my work with the Independent Science Society. Thanks Afiq Hatta for the many conversations about this topic.