Training Regime Day 25: Recursive Self-Improvement

post by Mark Xu (mark-xu) · 2020-04-29T18:22:03.677Z · LW · GW · 0 comments

Contents

  Introduction
  Recursive Self-Improvement
  Exercise
None
No comments

Epistemic status: highly experimental

Introduction

I have this theory that humans have this implicit skill that's something like "doing stuff". If you're getting at "doing stuff", then you can reliably get what you want. If you're not good at "doing stuff", you often fail to get what you want, either by failing to get it or by failing to try.

Being able to "do stuff" gives you the ability to change your day-to-day behavior. For example, if you can do stuff, then you'll be able to begin an exercise routine without much difficulty. If you can do stuff, then you'll be able to consistently practice piano, if you want to learn piano. If you can do stuff, then you'll be able to consistently read books, should you desire. Being able to do stuff makes your life better.

Applied rationality is supposed to increase your ability to "do stuff". Murphyjitsu [? · GW] makes it easier to get what you want. TAPs [? · GW] make certain actions automatic. Systemization [? · GW] makes getting what you want closer to the default action. Goal factoring [? · GW] helps you access what you want more directly.

The lesson we learn from video games is that it's more efficient to grind skills directly. So how do you grind "doing stuff" directly?

Recursive Self-Improvement

Doing murphyjitsu makes you better at it, so you want to do murphyjitsu more. As the saying goes, "there's a TAP for that". You decide to make a TAP to do murphyjitsu every time you make a plan. But making a TAP is a plan - you can murphyjitsu it!

You keep forgetting rationality techniques, so you want to systemize. You list all the rationality techniques in a checklist. You think of a TAP to consult this list every time you make a decision. Making a TAP is a plan, so you murphyjitsu it.

You notice [? · GW] feel a bit strange, but you ignore it and move on. Then you realize you wanted to do some focusing [? · GW]. There's a TAP for that, triggering your meta-TAP that results in murphyjitsuing your plan to install a focusing TAP.

(All roads lead to murphyjitsu.)


On the object level, all I'm doing is pointing out that you can use rationality techniques on themselves - cute, but probably not that important.

On the meta level, I claim that this is what it takes. In order for applied rationality techniques to actually make your life better, you have to actually use them. The point of rationality techniques is to make it easier to actually do things. Anything short of using rationality tools on themselves means you aren't putting in your full effort (which is ok, if that is not something you desire).

E.T. Jaynes says that reasoning must use all available evidence. By analogy, to get what you want, you should use all available tools.

The type of behavior I'm getting at results in (and maybe is equivalent to) overlearning - the training of a skill beyond the point of mastery, with the goal of transferring knowledge directly into instinct and intuition.

Exercise

Pick a bug [? · GW] and apply the full force of your rationality ability to solving it at every possible level (or at least object and meta levels).

0 comments

Comments sorted by top scores.