Concept Space vs Action Space

In the LW/rationality-sphere, it's common to say that the map is not the territory. This means our representations of objects are distinct from the objects themselves. This is an important, (and perhaps obvious) distinction. If you don't pay attention to this, you end up prey to a host of fallacious reasoning. For example, a conniving army general looking to cross a treacherous mountain pass might have the "clever" idea of erasing the marked mountain range from her map. This, of course, doesn't work because the mountains inform her map and not the other way around.

I think a similar map vs territory error arises when we assume that just because two things are conceptually similar, they can be solved in the same or similar way. Here's a concrete example of where this confusion might arise:

First, let's consider the concept of temporal discounting, a type of time inconsistent preference. Temporal discounting arises in situations where we make a decision and then, later on, wish we had acted differently. Specifically, it refers to instances where we see people value the present situation much more than the future. It's often what people have in mind when they refer to things like "instant gratification".

Some situations which exhibit temporal discounting include:

  1. A student puts off studying until the last moment. Instead, they choose to read a fun novel in the interim.
  2. A partygoer drinks more than they can handle, leaving their future self to deal with the resulting hangover.
  3. A dieter opts for yet another dessert during a meal, telling themselves they can make up for it with more exercise later.

In all three of these cases, there is indeed a commonality we can abstract—a human considers doing X and prioritizes over a future event Y. All three cases, then, fit our definition of "temporal discounting". However, does knowing that all three cases are instances of the same category imply a common solution?

I think the answer is mostly "no", and I'll try to explain why:

First off, what does knowing about the general case of temporal discounting tell us? Well, we know that this whole phenomenon relates to when our future selves prefer a different action than one we take in the present. There is a pattern we can abstract.

How does this pattern help us find a general strategy? Maybe we could try to create a heuristic like, "When you're about to do something that you're going to regret...don't."


That doesn't seem like enough.

I think there's an asymmetry here. It can make sense to abstract from many situations to come up with the concept of a General Problem (like what "temporal discounting" does). But trying to find a General Solution to the General Problem and then applying that General Solution to your specific situation is much more difficult.

In mathematics, this is readily apparent. We might look at finding the best-fit line to a set of points as well as finding the shortest path between two nodes in a graph. It is clear we can abstract the idea of "optimization" from the two problems (we care about some optimal value, relative to a metric), but the ways in which we optimize the two are going to be very different.

Indeed, in most cases, I don't think that it's clear there will be a clean General Solution, even when a common pattern can be abstracted.

Back to our examples:

For our student, it might be that they could reexamine their priorities. Perhaps this class isn't important to them, and doing poorly on this exam has few consequences. Or perhaps our student could rearrange their schedule around and study with a friend to shave off some of the aversion they had towards studying.

For our partygoer, they may want to consider the sort of circumstances which brought them to said party in the first place. Perhaps avoiding certain problematic friends would allow them to sidestep potential binge opportunities entirely. Or perhaps they could stick to non-alcoholic drinks and still have fun at the party.

For the dieter, perhaps planning out their meals in advance would allow them to avoid restaurants with unhealthy desserts. Or perhaps filling up before going out to eat with friends can help curb their desire for desserts. Or perhaps they tend to do a good job following their exercise plans, so eating another dessert was fine.

The point I want to get across is that, when looking for a solution, one might consider dramatically different factors, even though all three situations can be part of the same phenomenon. As a concept, temporal discounting is a descriptive category, not a prescriptive one. It is designed to help us see a common pattern between different situations, not necessarily to tell us how to act accordingly. I'm not saying we should avoid descriptive categories when thinking about self-help. Just that, in these situations where you’ve got a descriptive classification, it’s actually the specific details (and not the ability to recognize that you’re engaging in a more general phenomenon) which provide the most leverage towards solving your problem.

In other words, it's the considerations like "What led me, specifically in this case, to this situation?" rather than "If this is a temporal discounting situation, what do I do?" that lead to the most useful insights.

Caveat: Thinking about the general concept can be useful as a springboard for what to do next. For example, say you want to start flossing regularly and correctly recognize that you are trying to form a habit. Then, based on what you know about habit formation in general, you would know that you need a strong sensory trigger to start off the habit. From there, you can think about what items in your bathroom or daily routine could act as a good trigger.

(And indeed, this is how some rationality techniques, like TAPs, are created.)

But note that the majority of the work comes after noticing the relevant category of what you are trying to do is "habit creation". In a way, concepts can serve as a template. Once you've identified the appropriate concept, you still need to put in the effort to fill it out.

A final word of caution: It can be easy, I think, to do the relevant conceptual categorization, and then just stop there. Figuring out the general structure of whatever situation you're in usually results in a flash of insight. And it's not "fake insight", insofar as it really can be the correct conceptual bucket.

If your goal is to resolve the problem, though, then recognition isn't the whole story. It's important, then, once you've got a toolbox of mental models and helpful concepts, to make sure to follow through. There’s a sort of mental misstep that can happen where being able to simply identify the generalized principle at work can give the false impression that you also know how to solve the problem. Or, if you're not thinking hard, the flash of insight gives a sign to stop investigating. It's this misstep that I want to caution against.

By now, I hope the title makes sense: "Concept Space" is an imaginary space that groups items based on how similar they are, relative to some concept in question. Likewise, "Action Space" groups items based the similarity of their solutions (i.e. actionables you can take). Items close together in Concept Space can be very far apart in Action Space because they are two different ways of categorizing. But our defaults don't necessarily respect different notions of distance. Concept Space is much less constrained then Action Space. We can think of many ways to see patterns between two things, but only a small subset of them will correspond to things which are practically similar.

The default self-help paradigm, which jumps straight into giving you models and frameworks doesn't seem to directly address this, which is why I think this is worth being clear about.

Last Updated: 2020-06-20 12:14
First Published: 2020-01-26 15:37