A theory of what “why” or “how” means

Learning is often viewed as being able to accurately predict the future. If you learn the current state of the world, and then learn a model of the dynamics of the world, you can “run” your model forward and predict the future. I think this is a basically good and correct meaning of learning or understanding something, though it bears some elaboration.

Sometimes you learn how to predict something but still don’t feel like you understand “why” or “how” it works that way. I actually think that insisting on asking “why” separate from “what” is a kind of epistemic red flag and means you’re confused about what “why” means. But here’s a framing that I think sometimes explains why something feels like a “why” question.

Imagine you’re watching the insides of a clock as it ticks. To predict the future of this system, you really only need a small number of parts. Mostly, everything is periodic or linear; the pendulum swings back and forth at a certain rate, each gear turns at a certain rate, and the weight drops, very very slowly, but also at a constant rate. If you want your predictions to be accurate for longer, then you have to model the fact that once the weight hits the end of its chain, the whole clock will slow down over the course of several seconds and eventually stop. And that if someone takes the key and winds the clock back up, raising the weight again, then it will start over with your periodic model.

This is a good example of a time when it’s easy (almost involuntary) to form a good predictive model but then be left with basically no idea how the clock works. Your predictive model would do just as well if each gear were being turned independently, behind the scenes.

Some “how” questions you might have are, which parts of the clock are causing which other parts of the clock to move? What happens if you change parts of the clock? Why would two clocks have different shapes, sizes or placements of their gears? And how could anyone build or come up with the idea for a clock in the first place?

If you want to study causality or the effect of interventions like these, then you can do that by treating your dynamical system as having parameters. In the clock system, the parameters are all the physical things that you might want to vary about the clock, like the size and locations of the gears. For a fixed set of parameters, you still need to know the dynamical equations and the state, but that will tell you everything about the future.

In practice, when you’re modeling part of the world like this, you’re actually studying a system that you only know the approximate parameters for, or a system which is only an approximately isolated sub-system, and in the future the rest of the system (maybe you) will intervene on it, thereby changing its parameters. These are the reasons humans are built to want to know “why” something is true.

An unfortunate side effect of having machinery for wondering “why” is that it can become detached from any conceivably sensible counterfactual parameters of the system, and people can be left wondering things like “why did it have to be Jake who died instead of me” or “why does God allow there to be evil in the world” or “why does anything exist at all” with a deep inner compulsion that there must be an answer.

Leave a Reply