There's a factor he doesn't mention, which is which set is removed from the table. Norvig's code removes the first set it finds. Another strategy would be to identity all sets and remove one at random. And who knows how humans play? It's not obvious whether these are equivalent regarding the rate at which the table gets "worse".
I've seen moderately serious folk play. They will recognize multiple sets simultaneously and either go for the one closest to them or the one with three cards tightly clustered. And they move fast! Watching them play is like watching a mathematically inclined cobra fight.
