Imagine two discussion participants.

Imagine two discussion participants. What's the most interesting thing for them to talk about?

If they talk about things they already fully agree on, they'll waste time and expensive conversation tokens.

If they talk about things they fundamentally disagree on--that are so outside each others' current mental model that they track as incomprehensible noise or actively clash--then they'll lose trust and come out of the conversation with nothing to show for it.

There's a goldilocks zone: things they don't currently agree on but are close enough that they could come to agree on them through sharing information.