But zero-shot classifiers with this level of intelligence, world knowledge, ergonomics, cost profile, and ease of use are new.

I feel like good engineering doesn't just ignore those things, or at least it didn't before recently. Now I guess social media has added a pressure to reduce everything to a hot take.

So it is tuned specifically to classify content for safeguarding? I'm not familiar with this particular model, but it most likely has a specific classifier head that is tuned for the safeguard task. This is completely different to zero-shot classification.