Maybe a strategy for the case when the BVHNode is internal would be to order the children with the probability of getting ray-primitive hits (e.g. the one with more primitives or occupy more space)?

I think it might be better if the primitives occupy more space, but having more primitives might actually be a reason not to check a child first.

I also wonder if there are ways to cache the results of the closest query to avoid having to rerun an "any intersection" query

a guess on the answer to the question on slide: enter the child with more primitives first? Although there seems to be a trade off as it could take up time to go through a larger child without intersecting any primitives and having to search in the other child eventually

It was mentioned in the lecture that this method is used to compute shadows. However, I was wondering if the appearance of the shadow is affected by how many ray intersections there are and how can that be considered when ray-tracing.