XY and the Monkey's Paw

1. You have a goal, call it G.
2. To attain G, you think of a proxy P.
3. You ask an agent to give you the instructions for P.
4. The agent gives you long complicated instructions to solve P. You implement them, with a lot of time and effort.
5. You totally fail to see an alternative solution, because you were just focused on P.

For example:

• You want to access your home, but you forgot the key. You call a friend who is very good at picking locks and he spends 30 minutes on the phone giving you complicated instructions on how to unlock that precise lock that your door has. After that, you spend 2h finding the right tools and picking the lock until you succeed. Then, you realize that you could have just called another friend who lives 5 minutes from you and has a copy of the key.

This pattern is called the XY problem. It is well known in technical support, where a user who wants to do X asks how to solve Y instead.

This is similar to the Monkey’s Paw problem. The Monkey’s Paw would be:

1. You have a goal G
2. To attain G, you think of a proxy P
3. You ask an agent to do P for you
4. The agent is wildly successful at implementing P. However, P does not achieve G, or it does so in a way that has unintended consequences.
5. You failed because you did not achieve the real goal G, but a proxy

For example:

• You want to go out with the prettiest girl in the class. You ask the Monkey’s Paw to give you big muscles so that you can impress her. The Monkey’s Paw gives you an abnormally massive deltoid that looks deformed. You impress the girl, but not in the way that you expected. You fail to go out with the prettiest girl in the class

The pattern is similar, but not the same. In the XY problem, the problem is about efficiency. In the Monkey’s Paw problem, the issue is about unexpected consequences. A minor difference in the way I presented both problems is that in the XY problem the user implements the solution and in the Monkey’s Paw, it is the agent. I think this is not really the fundamental distinction, and whoever implements the solution is actually irrelevant.

I have encountered the XY problem multiple times when interacting with AIs. Luckily, no Monkey’s Paw problems. For now.

A solution to the XY problem is making sure that your goals are clearly explicit at all times! Goals sometimes have a hierarchical structure, and some of them might still be there without needing to be clarified. Some of our goals are sometimes not even obvious to ourselves. For example, I’m working on some research now. I could present my goals to the AIs by saying:

1. My goal is to solve this very specific technical detail

However, this can easily lead to an XY situation. A better one would be:

1. My goal is to better understand this general problem that exists in the field of microbiology
2. A subgoal is solving this very specific technical detail

A third option, which is still better, could be:

1. My goal is publishing a paper in (at least) a mid-tier journal, in the next 6-12 months
2. The paper will deal with this specific problem
3. To solve this specific problem, I think we need to solve this technical detail. However, if there is a different alternative path to this, please let’s explore that

This last structure is much better because it gives enough context. Most of my previous failures have been about not giving enough context. Of course, one could give “too much context” and then end up reconsidering some fundamental decisions in life, but that’s out of the scope of this post.