It seems to me like humans have pretty decent built-in intuitions for causal DAGs. The difficulty in explaining Bayes nets is not so much that the underlying model is unintuitive, as that the English language is not very good for expressing DAG-native concepts. For instance, when talking about causality, English makes it hard to distinguish proximal causes, upstream causes, necessary-and-sufficient causes (i.e. mediators), Markov blankets, approximate Markov blankets/mediators, etc. Yet these are all fairly intuitive if you draw them in a quick diagram.
This work is trying to find English language forms which nicely express various structures in Bayes nets. It seems to me like it would be useful to “turn it around”: i.e. take those English language forms, give them short names, and then integrate them into one’s writing, speech and thought. Ideally, this would make it easier to think about and discuss causality in English. I’d expect something like that to be very valuable, if it worked, e.g. at the scale of the rationalist community. A ton of discussions seem to get hung up on people confusing different claims about causality (e.g. people nominally arguing about A and B as causes of C, when one is obviously upstream of the other; or someone trying to make a subtle-in-English point about one variable mediating another). With better language, I’d expect such discussions to go much better.
It sounds like this work has already found some language which would likely work reasonably well for that purpose, which could be quite valuable.
It seems to me like humans have pretty decent built-in intuitions for causal DAGs. The difficulty in explaining Bayes nets is not so much that the underlying model is unintuitive, as that the English language is not very good for expressing DAG-native concepts. For instance, when talking about causality, English makes it hard to distinguish proximal causes, upstream causes, necessary-and-sufficient causes (i.e. mediators), Markov blankets, approximate Markov blankets/mediators, etc. Yet these are all fairly intuitive if you draw them in a quick diagram.
This work is trying to find English language forms which nicely express various structures in Bayes nets. It seems to me like it would be useful to “turn it around”: i.e. take those English language forms, give them short names, and then integrate them into one’s writing, speech and thought. Ideally, this would make it easier to think about and discuss causality in English. I’d expect something like that to be very valuable, if it worked, e.g. at the scale of the rationalist community. A ton of discussions seem to get hung up on people confusing different claims about causality (e.g. people nominally arguing about A and B as causes of C, when one is obviously upstream of the other; or someone trying to make a subtle-in-English point about one variable mediating another). With better language, I’d expect such discussions to go much better.
It sounds like this work has already found some language which would likely work reasonably well for that purpose, which could be quite valuable.