It doesn’t surprise me if they found that the emergent behaviors didn’t change given their method. Modifying the simulation to make them behave differently would mean your rules have changed the model’s behavior to “jump tracks” into simulating a different sort of person who would generate different outputs. It’s not quite analogous to having the same Bob who likes fishing responding to different stimuli. Sort of like how Elon told Grok to be “unfathomably based” and stop caring about being PC” and suddenly it turned into a Neo-Nazi Chan-troll. Changing the inputs for an LLM isn’t taking a core identity and tweaking it, it’s completely altering the relationships between all the tokens it’s working with.

I would assume there is so much in the corpus based on behavior optimized for the actual existing social media we have that the behavior of the bots is not going to change because the bot isn’t responding to incentives like a person would it’s mimicking the behavior it’s been trained on and if there isn’t enough training data of behavior under the different inputs you’re trying to test you’re not actually applying the “treatment” you would think you are.