Oh yes my point was that you’re exactly right. Current neural net architectures need absolutely ridiculous amounts of labeled data to “learn” anything, and they’ll just repeat the biases in their training data.