Just looking through the benchmarks, the model seems to be very high on visual/spatial capabilities. So I assumed they just tried some hard tasks on it and was surprised by wine identification and thought it was surprising that it could identify wines.