@Richard_Ngo Did anyone (except me?) write about cultural hegemony on LW and did anyone invent the epistemic equivalent of preference cascades?
Suppose, as a simplified model, that individuals reason according to what they have learned as Bayesians/Garrabrant inductors would, but are bottlenecked on communication. Then I expect that phase transitions can be constructed even out of that.
Indeed, suppose that the Bayesians who haven't heard of two Suprising Facts converged on betting 3:1 on the Western description of the Molotov-Ribbentrop pact as an agreement between the USSR and Nazi Germany to divide up the spoils of Europe in wars of aggression,as opposed to the USSR-like description implying that the MRP was a sane idea, while there are two facts[1] that weren't spread and thus were unaccounted in the individuals' world models. Suppose also that, in the oversimplified model, any of the two facts would cause the Bayesian reasoning individuals to double the ratio of the Soviet perspective's weight to the Western perspective's weight.
If a Bayesian heard only one Surprising Fact, then the posterior probability ratio would stay 3:2 in favor of the West, and the Surprising Fact wouldn't propagate further. On the other hand, knowing both such facts would lead the Bayesian to have a probability ratio of 4:3 in favor of the USSR, try to propagate the facts and be bottlenecked on communicating both facts at once.
What is the main difference between the real-world cascades and the model, except for individuals having worse rationality/integrity?
Fact 1 relates to territories claimed by the USSR and the Axis. For example, the Second Polish Republic was split between the USSR and Nazi Germany roughly along the Curzon Line. Additionally, the only lands in Europe which were controlled by the USSR by 21 June 1941 but not by the Russian Empire in 1913 are Eastern Galicia, Northern Bukovina, the Hertsa region and a batch of islands in the Arctic Ocean.