I don't really see how the four categories would help with training spam filters. "Social" is half spam, half maybe not. "Promotions" is mostly spam, though Google would rather see it as ham. "Updates" and "Forums" are mostly not spam.
I believe the parent poster meant that the classifiers that train into the baked-in categories are trained like a spam filter and use cross-account global knowledge. In other words, it's a model that works if there's a small number of buckets and a huge dataset to train the bucket-targeting against and breaks down if every individual user has their own buckets with their own conflicting understanding of what those buckets should contain.
Ironically, the Inbox rollups worked absolutely beautifully and the gmail filters are really, really rough and don't work well. "Updates" are so incredibly nebulous that it's really a 50/50 crapshoot on whether or not google knows what should go in there. "Updates" and "Social" are commonly mixed up, and finance/purchases have zero place to go. Are they 'Updates' like when I get a 'Our ToS has been updated' email? Etc. The categories are too general to be useful.