I was at LessOnline. The thing that got me most excited was attending @Raemon’s very abridged Metastrategy workshop. To my lack of surprise, he has written up the general idea already, but that report focuses on people’s response to trying fractal strategy and I would not have found it very actionable had I seen it before attending the workshop. The core idea immediately spoke to me when I saw it in-person, and I think I can slam out a quick-and-dirty get-started guide that explicitly lays out what you can put on paper to try this out today if you want. I want to try to use the method, but I won’t really know whether it works for me for months, and I want to make sure I know what I’m attempting first. Caveat: it is likely that I misunderstand some aspect of the method after receiving all of two hours of instruction.

What is a plan?

A plan is a thing you want to do. However, in order to do that thing you have several options for how to do the thing. Each of those options is itself a plan which might have several options for how to do it. It’s probably a good idea to stop at three levels of meta, but you could do more. When trying to accomplish some goal, you could make a tree graph of all the ways it is possible to accomplish that goal, where there is one path which goes all the way down the tree which describes what you want to accomplish, the first subgoal you intend to target that gets you there, the default subsubplan you will start with to achieve that subgoal, and so on. If you’re trying to achieve anything difficult, it is very likely that you will have a brilliant idea for how to accomplish your goal that fails, so you should make multiple approaches to accomplishing your goal, and if each of those approaches has multiple options for what order to go in or who to work with or what toolset to use, then you can easily have a very twisted tree. I’m a graduate student, so a (slightly facetious) plan might look something like this:

Note that I have only listed below each plan potential immediate next steps toward the plan. Leaves which are both below another leaf are each a next step I could take. Usually, the levels of meta correspond to lengths of time. For example, the top level could be a thing you want to accomplish in a year, the next level plan you might spend the next month on, and the level below that is what you will do for the next few days in order to get to the one month goal. The idea is not to know every step of your plan until you’re done; the idea is to know what you’re working on right now and how it ties into your ultimate goal, even if you aren't sure what exact further steps are going to get you to a thing after you have done the thing directly below it. If you already know all of the steps you need to do in order to get to your end goal, you probably do not need metastrategy; you need willpower and maybe a good to-do list.

The smallest unit of metastrategy

This is a metastrategy, so the smallest unit of it should not just be one leaf. For your first attempt at the method, you should probably start with something about the size of the subtree I consider in this section, but in theory, I think you can build a large plan tree by treating each leaf as the top of such a subtree.

When you are considering strategy toward some (sub)goal, you should be considering one default plan and one good alternative plan. You should also write your next steps to take under each of plans, because otherwise you aren’t being explicit enough. So most of the time you’re dealing with a subtree of your entire plan which looks something like this:

You will notice some additional things associated with the default plans in the leftmost boxes. For the top of a subtree which you are considering, you should list out why you want to accomplish this goal. If the top of this subtree is not the absolute top of your plan, the "why's" might connect to the actual top of your tree. Alternatively, they might be direct appeals to your personal utility function. You should be deeply considering why this is a thing you want to do in the first place and why this plan appeals to you (why this plan instead of some other plan), because knowing why the plan is happening may help you distinguish between the subplans you have available or help you prioritize between the plans you have if you are juggling multiple long-term plans.

In your default plans, you need to have a crux. This takes the form of a prediction of a thing you expect to observe that will tell you whether this was a good plan or not. It can be positive or negative (I tried to put one example of each), but it should be a thing you can evaluate as soon as possible. This is not just a way to evaluate whether your plan worked, this is a thing you will be using to decide your next actions on a day-to-day and week-to-week basis. The idea is that you should pick a thing that if you see it (or don't see it) after a specified time, you will abandon the plan it is attached to for your secondary plan. It should not be horribly abstract; it should be a tangible thing you expect to have accomplished or an explicit observation you expect to make if something is going wrong. Because it will send you to a secondary plan, it may not actually be a thing that says your default plan isn't working, but it may rather be an observation that tells you that your backup plan is a better option. After all, your backup plan should be something you're excited about too, and it would be a waste of time to keep doing a good plan when you could be doing a great plan. Your crux should have a literal date attached to it, and you will be evaluating whether it has happened on that date and using the results to guide your next steps. You will be logging all of your predictions and the probability that you expect them to happen. A good resource here is fatebook.io, which will keep track of your predictions over time and prod you to evaluate them as they come due. As a bonus, you will be able to see whether you are well-calibrated. If you keep predicting things will happen with 50% certainty, and they only happen 30% of the time, then that is a personal fault you should try to fix.

I think each default plan should have at least a default first approach and one backup approach. The default first approach should have a crux on it which you can use to push yourself over to your secondary approach before the crux on the default plan it is attached to is triggered. More explicitly, your default plan and your default subplan and your default subsubplan and so on should all have cruxes on them. You should explicitly know what you're doing next (literally what you will do the next hour you work on this thing), what the goal of doing it is, and what will make you stop doing it to try something else on multiple levels of meta. If you do end up executing a plan switch based on a crux, you should immediately make sure that your new default plan has a crux and alternative plan lined up before you start it.

What do I do if something works?

I don't remember hearing this in the workshop, so I'm filling in the blank myself. Once you have accomplished some box in your tree, you need to replace it with the next thing you intend to do and as many recursive levels of sub-default options you need to know what you will be doing in the next hour you work on the plan. Sometimes this means moving your backup option over to your default option, if they were two different approaches to your ultimate goal which can build on each other. Sometimes it means you have to erase all of your backup options because they were mutually exclusive with your default. Regardless, once you finish something, it's time to replace it with your next step, or if the subplan you just completed has accomplished the plan which lies above it, replace the highest leaf on your plan tree which has been completed with your next immediate goal on the timescale appropriate to the level of meta you just erased. Then you write as many layers of subgoal under the new leaf as necessary to know what you will be doing the next hour you work on the problem. Don’t forget to fill in appropriate alternative plans at every level of meta. Eventually, you will either accomplish the goal at the top of your tree, or you will abandon the overarching goal entirely and start from scratch. I suppose you might integrate the top goal into an even longer-term goal once you get better at long-term planning or shift more of your focus to a more unified purpose.

What is the point?

The two meta-principles given to us to keep in mind during the workshop were that you should have at least 2 real (actionable) plans you believe in and you should experience a crux to change your plan. The crux is to a huge extent the actual point. You are making an actual commitment to abandon your clever plan if it isn't working, and you are devising an evaluation of whether it is working. Humans are great at rationalizing why the thing they are doing is the best thing they could be doing, and this is a way to hold yourself accountable to reality. I also think this framework is good for your ability to plan for reasons beyond the mechanics of it.

I think that metastrategy operates similarly to explicit Bayesian updating. (I have not advanced to the level of habitually doing these calculations on paper myself, so take this comparison with a grain of salt). If you have a lot of confusing evidence and you don't immediately know what you should believe based on it, you can sit down and write down all of your priors and update them based off of recently acquired evidence to get posterior beliefs with a certainty score, but it is completely acceptable to throw away the paper afterwards and go with your gut after you have seriously considered all of the evidence and the effect of your updates according to the laws of probability. It's not quite that the updates are just an excuse to consider the problem deeply, because numerically evaluating how surprising each bit of evidence is actually provides valuable insight. In the same way, forcing yourself to think concretely about what you expect to see if you try something and what success or failure would look like seems like half of the benefit of metastrategy, although the actual mechanics of evaluating probabilities after you've put your plan into action is helpful in its own right. Forcing yourself to come up with a backup plan helps you understand the space of actions you can take and choose better default plans, but it is also absolutely supposed to be a plan you believe in and will enthusiastically switch to if your default is (measurably!) not working out.

Note that the fatebook predictions are a mechanism for evaluating whether you should switch your current plan, but they are also a record of both how often your plans are working out over time and how good you are at predicting how well plans work. That explicit record will let you know if metastrategy is working for you, and if it is there are opportunities to level-up. Maybe you can get to the point where you can explicitly weigh your (well-calibrated) probabilities of success against how aligned a plan is with your utility. Maybe you can occasionally make the trade-off of taking a plan which is more likely to succeed, but is less aligned with your goals. Or the other way around, but either way, you have a framework for keeping track of that information.

What do I expect to get out of this?

I expect this to be an emotionally brutal experience for a while. I expect to be constantly second-guessing my deepest feelings about the things which are most important to me and running up against my own failures. In return, I hope to get better at making plans which I will follow through on. I hope to spend less time with things I "should" do sitting in my to-do list while I don't do them. I hope to spend less time chasing down plans I have become attached to when I should have cut losses and tried something new sooner. I hope to be better at coming up with alternate plans that I am nearly as excited about as my default plan. I am hopeful that this metastrategy approach will get me to my goals faster.