Gmail stopped using IP reputation entirely a year or two ago, and effectively every major consumer mailbox provider relies on domain reputation heavily these days. Almost all major filters look at a combination of indicators, so it's becoming less and less common that a good mail coming from a bad IP is subject to filtering.

The answer to why better spam filtering AI hasn't been developed in a non-proprietary setting is fairly straightforward: the success of any filtering method is heavily dependent on consuming data at scale - including recipient behavior (e.g., mark as spam, vs. spend time reading a message), and this underlying data is just as important, if not more so, than the machine learning approach used to compute filtering assessments.

In short - you more or less have to be a large mailbox provider in order to have a chance at doing this well.