This post is once again inspired by a Ted Chiang short story. This time, it’s “The Evolution of Human Science”, which imagines a world in which super-intelligent “metahumans” have become incomprehensible to the ordinary humans they’ve left behind. Human scientists in that world practice “hermeneutics“: instead of original research, they try to interpret what the metahumans are doing, reverse-engineering their devices and observing their experiments.
It’s a thought-provoking view of what science in the distant future could become. But it’s also oddly familiar.
You might think I’m talking about machine learning here. It’s true that in recent years people have started using machine learning in science, with occasionally mysterious results. There are even a few cases of physicists using machine-learning to suggest some property, say of Calabi-Yau manifolds, and then figuring out how to prove it. It’s not hard to imagine a day when scientists are reduced to just interpreting whatever the AIs throw at them…but I don’t think we’re quite there yet.
Instead, I’m thinking about my own work. I’m a particular type of theoretical physicist. I calculate scattering amplitudes, formulas that tell us the probabilities that subatomic particles collide in different ways. We have a way to calculate these, Feynman’s famous diagrams, but they’re inefficient, so researchers like me look for shortcuts.
How do we find those shortcuts? Often, it’s by doing calculations the old, inefficient way. We use older methods, look at the formulas we get, and try to find patterns. Each pattern is a hint at some new principle that can make our calculations easier. Sometimes we can understand the pattern fully, and prove it should hold. Other times, we observe it again and again and tentatively assume it will keep going, and see what happens if it does.
Either way, this isn’t so different from the hermeneutics scientists practice in the story. Feynman diagrams already “know” every pattern we find, like the metahumans in the story who already know every result the human scientists can discover. But that “knowledge” isn’t in a form we can understand or use. We have to learn to interpret it, to read between the lines and find underlying patterns, to end up with something we can hold in our own heads and put into action with our own hands. The truth may be “out there”, but scientists can’t be content with that. We need to get the truth “in here”. We need to interpret it for ourselves.