The point Eliezer is addressing is the one that RichardKennaway brought up separately. Causal models can still function with feedback (in Causality, Pearl works through an economic model where price and quantity both cause each other, and have their own independent causes), but it’s a bit more bothersome.
A model where the three are one-time events- like, say, whether a person has a particular gene, whether or not they were breastfed, and their height as an adult- won’t have the problem of being cyclic, but will have the pedagogical problem that the causation is obvious from the timing of the events.
One could have, say, the weather witch’s prediction of whether or not there will be rain, whether or not you brought an umbrella with you, and whether or not it rained. Aside from learning, this will be an acyclic system that has a number of plausible underlying causal diagrams (with the presence of the witch making the example clearly fictional and muddying our causal intuitions, so we can only rely on the math).
In my opinion, the best fix would be to steelman the argument as much as possible.
The concept of inferential distance suggests to me that posts should try and make their pathways as short and straight as possible. Why write a double-length post that explains both causal models and metabolism, when you could write a single-length post that explains only causal models? (And, if metabolism takes longer to discuss than causal models, the post will mostly be about the illustrative detour, not the concept itself!)
The concept of inferential distance suggests to me that posts should try and make their pathways as short and straight as possible. Why write a double-length post that explains both causal models and metabolism, when you could write a single-length post that explains only causal models? (And, if metabolism takes longer to discuss than causal models, the post will mostly be about the illustrative detour, not the concept itself!)
You’ve convinced me. I now agree that EY should go back and edit the post to use a different more conventional example.
The point Eliezer is addressing is the one that RichardKennaway brought up separately. Causal models can still function with feedback (in Causality, Pearl works through an economic model where price and quantity both cause each other, and have their own independent causes), but it’s a bit more bothersome.
A model where the three are one-time events- like, say, whether a person has a particular gene, whether or not they were breastfed, and their height as an adult- won’t have the problem of being cyclic, but will have the pedagogical problem that the causation is obvious from the timing of the events.
One could have, say, the weather witch’s prediction of whether or not there will be rain, whether or not you brought an umbrella with you, and whether or not it rained. Aside from learning, this will be an acyclic system that has a number of plausible underlying causal diagrams (with the presence of the witch making the example clearly fictional and muddying our causal intuitions, so we can only rely on the math).
The concept of inferential distance suggests to me that posts should try and make their pathways as short and straight as possible. Why write a double-length post that explains both causal models and metabolism, when you could write a single-length post that explains only causal models? (And, if metabolism takes longer to discuss than causal models, the post will mostly be about the illustrative detour, not the concept itself!)
Should that be “cyclic”? I take it from Richard’s post that “acyclic” is what we want.
Yes, it should. Thanks for catching the typo!
You’ve convinced me. I now agree that EY should go back and edit the post to use a different more conventional example.