Wednesday, August 20, 2014

Small, Consistent Effort: Uncharted Waters In the Art of Rationality

Summary: I predict that there are powerful secrets yet to be uncovered in the area of rationality skills that are fairly easy but take a long time to learn.

I've been thinking about what sorts of things rationality skills are, and how they are gained. By "rationality skills", I mean patterns of thought and feeling that contribute to systematic improvement of the accuracy of beliefs, and of the satisfaction of values. 

The sort of categorization that most interests me is based on how the skills are acquired. I imagine a grid of rationality skill acquisition. It looks like this.


Things farther to the left take less time to learn, while things farther to the right require some combination of processing time, many iterations, and long strings of dependencies on other skills that must be acquired serially. While "difficult" and "takes a long time to learn" may be highly correlated, I don't think they're the same thing.

It can take a child quite a while to learn long division. You generally need to lean addition in order to learn subtraction and multiplication, multiplication in order to learn division, and the final procedure that leads to the right answer, which depends on multiplication and subtraction (and division, if you want to be efficient). All together that can take a long time.

But once you've got all the pieces of basic arithmetic, the final procedure is pretty easy. If you've got detailed instructions in front of you, it can even be carried out correctly on the very first try. And the pieces themselves are pretty straightforward, especially if you recognize that the execution of algorithms will suffice, and deep understanding isn't strictly necessary. It may be a long and complex process if you've never seen arithmetic before, but the greatest inferential gap is either between addition and multiplication or between multiplication and division. Those are leaps average gradeschoolers can make. No individual part is all that difficult to get your head around.

But consider the simplest problems in elementary algebra. In addition to the basic arithmetic operations, you need two more pieces: "doing the same thing to both sides of the equals sign", and "variable". "Doing the same thing to both sides of the equals sign" is a even easier than "the procedure for long division".

But "variable" is fundamentally different. It requires a new kind of idea. It requires abstraction, which is not only new but inferentially distant. It may even be the greatest inferential gap a child must cross in traditional math education up to pre-calculus. It isn't a complex idea, though, and there's not really such a thing as "half-way understanding variable". You get it or you don't, and when you get it, elementary algebra suddenly makes sense. "Variable" is probably an epiphany. And it's a difficult enough epiphany that, according to Jo Boaler, a great many adults never do have it.

I think the Lesswrong Sequences are mostly good for a few epiphanies. They're largely boot-strapping sorts of epiphanies, which re-order your mind in ways that allow for further epiphanies. But they're still epiphanies. They're skills that are difficult to gain but happen all at once, in this case over the course of reading a blog post. They're mostly things of the form "understanding X" or "realizing that Y". And most of the potential lessons of the sequences are fairly difficult unless you happen to have a mind with exactly the right arrangement, which is part of why most people don't have their whole mind rearranged once per post. So the Sequences mostly exist in the upper left corner of the skill acquisition grid.

CFAR workshops occupy the whole left half of the grid. Most of what's taught in the actual classes falls in the bottom left--quick and easy--because the lessons are only fifty minutes, and they're mostly practical instead of conceptual. Rather than lecturing you for an hour, as though reading several Sequence posts aloud, they're more like, "Here is a procedure that is surprisingly domain-generally useful. Let's practice."

For example, CFAR teaches Trigger-Action Planning, known in the Cog Sci literature as "implementation intentions". It's got even more bang for the effortful buck than memory palaces, because the effect size is similarly enormous, but it helps with anything that can be broken down into concrete triggers and concrete actions. And all it takes is learning to compose specific enough if-then statements, like so: "If I hear my alarm in the morning, then I will hop out of bed immediately." Other bug patches CFAR installs include Murphey Jitsu, Goal Factoring, Focused Grit, and Againstness. (Don't worry, I'll discuss exceptions to this in a minute.)

The rest of the CFAR experience, the socialization outside of classes, usually causes at least one epiphany. Participants have conversations with instructors and other participants, and since everybody there is carefully selected to be bright, curious, and interesting in diverse ways, there's always somebody saying, "Wow, I've never thought of that!"

CFAR teaches one lesson from the bottom right quadrant: Comfort Zone Expansion, or CoZE. CoZE is basically CFAR's take on exposure therapy. Exposure therapy can take a long time. Though you might see progress right away, you're usually not going to wipe out a deep fear or anxiety in a single go. It takes repeated exposure with a slow and steady increase in intensity.

But exposure therapy is fairly easy! Scary, though by design not very scary, but not difficult. The principle is not hard to understand, the procedure is straightforward, and there's just not much more to it than that. It takes time, is all. So CFAR devotes a lot more time to CoZE than to the other units. There's a standard 50minute CoZE prep class, and there's an entire evening devoted to the "CoZE outing", where everybody goes off for hours in search of repeated exposure to a feared stimulus. CoZE is a tortoise skill. "Slow and steady wins the race." It relies almost entirely on small, consistent efforts.

Some of CFAR's other lessons may be close to the middle of the X axis, but I don't think there are any others that must necessarily take many iterations to properly install.

There is one skillset CFAR attempts to impart in a class format that I think falls in the top right quadrent: Bayesian reasoning. It is not merely an epiphany, and if you want a version that works in real life, it is not a bug patch. When last I saw it (June 2014), the Bayes unit was not up to the same standard as the Bug Patch units or CoZE, and I think I may now understand a big chunk of why.

Bayesian reasoning depends on some pretty mind-twisty habits of thought. Not only are the skills difficult to attain, but they require a combination of long processing time, many iterations, and long strings of dependencies. It takes a couple epiphanies, a few bug patches, lots of habit installation, and the long and difficult process of weaving all of that together into fully Bayesian patterns of thought and feeling. A two-hour class is simply not the right format to get all of that done.

[CFAR does offer six weeks of 1-on-1s for all participants, so there's more room for imparting Tortoise skills than the workshop itself allows. But those are extremely personalized, more like counseling than the usual sort of teaching, and it's hard for them to scale in the same way as the Sequences or the standard batch of CFAR units, so I'm not discussing those so much.]

Wizard skillsets like Bayesian reasoning are definitely possible to attain. I think almost all of it, if not all of it, happens by acquiring components from the other three quadrants and weaving them together over time. If there are rationality skills that primitively require slow and difficult aquisition, I don't know what they are. Most of the really badass epistemic skills, I suspect, are Wizard skills. And so far, CFAR plus the Sequences seldom seem to be enough to get people there.

I've learned some hard things. I've learned to prove theorems of nonstandard mathematics that defy my most basic logical intuitions, for example. I've learned to interpret ancient, bizarre, abstract Indian philosophy. I've learned to follow Blues dance like nobody's business. And I can't think of a single skill I've gained that simply could not be broken down into quick and easy bug patches, getting-my-mind-around-it-ness, and boatloads of small, consistent efforts.

So maybe I'm wrong, and most of the Wizard skills worth having are primitively slow and difficult to attain. After all, that's one theory that explains why I lack Beisutzukai-level mastery. There's got to be something Anna Salamon and Eliezer Yudkowsky share that I lack, and maybe this is it.

But you know what Anna and Eliezer definitely have that I don't? Practice. Years and years of practice. I heard the word "rationalist" outside of Cartesian philosophy for the first time just two years ago. So maybe while I've had most of the epiphanies I'm going to from Lesswrong's material, and while I've installed most of CFAR's bug patches, there's a third class of easily attainable skills I must gain before I can weave all of it together and become far stronger as a rationalist.

If this is true, it's very good news! It means that if I can looks at the Wizard skills I desire and break them down into the epiphanies and bug patches I already have, I may be able to ask myself, "What part of this puzzle is going to take small, consistent effort?" And I might well come up with a useful answer!

With a single exception, all of the skills I've gained directly from Eliezer since I've lived with him over the past year confirm this hypothesis. (He gave me one all-or-nothing epiphany in person, which was "fail more".) All of the others followed more or less the same pattern:

  1. He emphasized the importance of something I already basically had my head around, both abstractly in principle and concretely in practice.
  2. I decided to practice CONSTANT VIGILANCE for a single failure mode associated with lack of the skill.
  3. I noticed the failure several times over the course of days or weeks until I could predict when I was about to experience the failure mode.
  4. I practiced CONSTANT VIGILANCE for times when I could feel that the failure mode was about to happen.
  5. I tested out a few ways of responding to the feeling that the failure mode was about to happen, to find out what overcoming the problem might feel like.
  6. I let the results of those tests process for a little while.
  7. Often, I ran my observations by Eliezer to get his feedback.
  8. I composed a trigger-action plan (though usually not in writing) with the trigger "I notice I'm about to experience the failure mode if I don't do anything to stop it", and an action I expect to avert the failure.
  9. I practice the trigger-action until it feels like a background habit.
  10. I weave my understanding of the problem and its import into my practice.
Imagine a master rationalist does the exercise I described above, picking a Wizard skill and sorting its components into the other quadrants. And imagine she wants to teach that skill to me. She can say some things about what must be understood, hoping to cause the relevant quick but difficult epiphanies. She can give me some simple bug patches to install if quick and easy solutions are part of it. Then, for every slow but easy tortoise component, she could drastically speed up my skill acquisition by providing me with, or otherwise helping me uncover, the following information.

  1. What it feels like to notice the failure mode itself, or how to find out what it feels like.
  2. What it feels like to notice that the failure mode is about to happen, or some things it might feel like.
  3. What to do when I notice that feeling, or a few options for what to try.
A compilation of such advice on tortoises, especially if it were presented in a way that encouraged consistent check-ins and small efforts toward improvement, would be a new kind of rationality resource.

It would not, however, be unprecedented in other domains. Without even doing research, I am aware of books approximating this concept focusing on yogamindfulness, writing, and physics. I think we need one of these for the art of rationality.