> Model efficiency could improve faster than expected.

There are compounding incentives for this. I see this as the most likely outcome, though it will likely be stepwise rather than gradual.