Real-world Problems

Power vs Convenience

How far can you get with knowing just the causal structure, not the causal mechanisms or the categories making up the variables?

I’ve been assuming that the scientific method will be able to solve the problems I want it to, but I haven’t tested that assumption.

There are some real limitations getting in the way. If my aim is to write down causal models explicitly, so that I don’t have to memorize them, then I cannot build causal models faster than I can type.

So, there seems to be a tradeoff between convenience and predictive power. I can write a simple causal model cheaply but make less precise predictions, or I can painstakingly build an elaborate causal model and enjoy making accurate predictions.

All this assumes that I don’t want to use my own brain to do some of the work. For example, I might choose to write down the structure of the model on paper, but store the workings of the mechanisms in my own brain, which can effortlessly handle such complexity (see: bicycle riding). Brain time is costly, though, and not very scalable. If I can figure out how to squeeze the maximum predictive juice out of a given explicit model, then I can walk into another field, look at a causal structure written by somebody else and start making useful predictions immediately. I wouldn’t have to wait to understand all the details from the bottom up.

Reasoning with explicit models scales. It’s not limited by my brain’s knowledge-acquisition bottleneck.

Maybe all this is just a pipe dream. Maybe you can’t do any useful thinking with just the causal structure. Maybe you need to “understand” each mechanism works before you can begin, like knowing about cell structure, cell division, transport, and respiration, before you can start asking how to speed up the growth of your potted plant.

I’m betting that you don’t have to. That’s what abstraction is for. When you use a calculator, you don’t need to understand how the logic circuits inside combine to give you the answer to “2 + 2”. You just trust that they satisfy the abstraction of a “calculator” - something capable of answering (most of) your arithmetic questions.

In Evidence We Trust

What is the simplest causal model needed to solve a problem? I don’t even have to know the details of the model myself, just the amount of detail needed. Also, to push my theory really far, assume that you’re completely new to the field, so you don’t know any background details (like cell organization in biology). Just given a bare bones causal model and basic human thinking, how far can you go?

What sort of problems do I want to solve? Ones that pay off handsomely. Things that can save me money, or help me make more.

Personal problems include: waking up early everyday; getting a ripped body; having the most fun for the least money; working with deep focus everyday; cleaning up the house regularly; being happy, etc.

Professional problems (for humanity) include: solving technical problems in computer science, economics, physics, biology, engineering, math, psychology; designing great essays, movies, songs, products, buildings, tools; designing better public policies and so on.

I just remembered that you only need to know as much information as is valuable. If further details are not going to change your decisions in any way, don’t learn them. So, the answer to the question “how much do I need to know?” depends very strongly on how much value you can extract from further information, which further depends on the tools you have for manipulation. For example, knowing intimate details of digital logic circuits is probably useless unless can obtain access to a chip-manufacturing unit.

So, asking about the information needed to solve a problem in isolation is silly; you could just as well go into the depths of quarks and analyze the minute differences they would make. Instead, you need to know the decisions at hand and the value of information for the problem.

Therefore, the question should be: given a set of possible decisions, and a bunch of variables, what is the minimum information needed to get the best decision?