Discuss: Meta-Thinking Skills
When do you go meta? When do you stop going meta?
In the video Q and A Eliezer offered some advice about this (the emphasis is mine):
I tend to focus on thinking, and it’s only when my thinking gets stuck or I run into a particular problem that I will resort to meta-thinking. Unless it’s a particular meta-skill that I already have, in which case I’ll just execute it. For example, the meta-skill of trying to focus on the original problem. In one sense, a whole chunk of Less Wrong is more or less my meta-thinking skills.
I guess on reflection I would say that there is a lot of routine meta-thinking that I already know how to do, and I do without really thinking of it as meta-thinking. On the other hand original meta-thinking, which is the time consuming part, is something that I tend to resort to only when my current meta-thinking skills have broken down. And that’s probably a reasonably exceptional circumstance, even though it’s something of a comparative advantage and so I expect that I do it a bit more of it than average.
Even so, when I’m trying to work on an object-level problem at any given point I’m probably not doing original meta-level questioning about how to execute these meta-level skills. If I’m bogged down in writing something I may execute my existing meta-level skill of try to step back and look at this from a more abstract level. If that fails then I may have to think about what sort of abstract levels can you view this problem on, and similar problems as opposed to tasks, and in that sense go into original meta-level thinking mode. But one of those meta-level skills I would say is the notion that your meta-level problem comes from an object-level problem, and you’re supposed to keep one eye on the object-level problem the whole time you’re working on the meta-level.
In his discussion post “Are you doing what you should be doing?”, Will_Newsome identified what seems to be an important guiding principle of meta-thinking:
Yay for going meta! I should repeat this process until going meta no longer produces time-saving results.
(where “time-saving results” can be replaced with “greater marginal utility” to obtain a form that is more generally applicable)
Some questions we could explore:
How can you identify when you are “stuck”, and when going meta has greater marginal utility than continuing on your current level of thought?
How many levels of meta can people regularly think on? Can this be extended (via drugs, mental exercises, external tools, etc.)?
Is it possible to develop a set of meta-thinking skills that help us execute or create Less Wrong meta-thinking skills?
Should one distinguish between meta-thinking and meta-meta-thinking?
(I plan to try to compile the insights and advice here into a top-level post discussing the principles of, and heuristics for, effective meta-level thinking)
Edit: Changed minor wording and altered the third question posed.
Incidentally this is a very valuable instrumental topic for doing well on job interviews, specifically in software areas. The problems are usually a mix of skills, “basic” knowledge and a thinly veiled IQ test questions; the latter type rely on “going meta” at some point of problem solving. If I had a good heuristic for this I think I could pass almost any interview :).
E.g. (classic) how do you delete a node from a linked list if you only have (pointer) to that node, not the previous one?
Sadly, a lot of interviews are more about “likeability” and people skills than pure critical thinking skills. (I wish it were that easy...)
At top software companies they seem to weigh the technical aspects more, though (somewhat understandably) the interviewers want to imagine working with the candidate as a future positive interaction.
Personally I think I pass the personality test, but recently blown an interview due to being stale in some areas and insufficient mental flexibility (“going meta”) on a couple of questions.
This is a specific application of “meta”, but a valuable one for many people here and is probably generalizable to other areas:
Pretty much the best programmer I know (an author of a number of Boost libraries) once reduced his “secret to success” to DRY Principle.
The way I understood it that he looked for higher level of abstraction even if it was locally suboptimal (in small programs cut n’ paste works fine); it might have taken longer in some instances, but made him a way better at creating abstractions.
Here’s a simple meta-thinking skill that aims at this: recursively ask yourself “How could I have thought of that?”.
I’ve been playing with this in the context of solving math problems. I’ll work on a problem for a set amount of time and then look at it’s solution. I’ll then think back to the thinking I did, and figure out what kind of meta-skill or meta-heuristic would have allowed me to make it to the next step of the solution. And so on, until the end of the solution if possible.
I think not. I feel that it would be reasonable to define meta-thinking as “thinking about patterns of thought”, with the content of these lower-level thoughts having no particular importance. They could be object-level thoughts or meta-level thoughts; in either case reflecting on their patterns would be considered meta-thinking.
This also fits my experience with meta-thinking skills on LW. I’ve found that techniques like Hug the Query can be applied regardless of the current level of abstraction.
Let’s speculate about how Eliezer developed his meta-thinking skills. I imagine that he was working toward FAI (he had something to protect) and frequently ran into problems (and could recognize them as problems). Unable to make progress he might ask “where have I seen a similar problem?”. Over time he could bunch together similar confusing problems. Eventually he might abstract their common essence (by asking “what makes them similar?”), and perhaps find a solution to them in the form of a meta-thinking technique.
This suggests a procedure for creating meta-thinking skills:
Work towards your goals and learn to identify confusing problems.
Don’t rush to “solve” them. Cluster them by identifying which ones are similar.
Discern the pattern of a cluster, and devise a technique to solve it and similar problems in the future.
I’m surprised this post hasn’t become more popular!
Eliezer’s quote seems important:
I think that this quote delivers a strong conceptual unification of various teachings and techniques on LW.
I also think that an investigation of the questions I raised might have high instrumental value.
Is there something I’m missing? Do people already have an understanding of these issues? Alternatively, do people have good reason to believe that this discussion will be fruitless?
One way I’ve done this is via the meta-level skill of recognizing a familiar pattern in the structure of my thoughts about a problem. This familiar pattern can be addressed with my existing meta-skills (which is one reason I remember it), and so I just apply them. This is what I think Eliezer means by “routine meta-thinking”.
It’s harder to identify when one should engage in original meta-level thinking. If you try to use the outside view, and recognize thinking patterns that can be treated on the meta-level, you most likely end up doing what I described above. If you can recognize a thinking pattern it means you’ve seen it (at least) a few times before, and probably know how to deal with it.
One could also approach this decision using a coarser outside view. You could ask yourself:
Generally, how much more productive am I when I go meta? How about when I go meta more than once?
Does going meta tend to be productive in this large class of problems? (i.e. in household problems, or in math problems)
An simple heuristic to use is to go meta periodically, with longer periods between higher meta-level thinking modes.
In this comment nickerst mentions using this heuristic in structured planning:
I don’t like meta—I instinctively agree with Robin’s post Doubting My Far Mind.