In 2015, Siri (and a number of other assistants) could tell you the weather tomorrow easily, but general question-answering was a pie-in-the-sky research dream.

In 2015, Siri (and a number of other assistants) could tell you the weather tomorrow easily, but general question-answering was a pie-in-the-sky research dream. Tons of labs were working on this problem using all kinds of approaches. These mostly fell over when you asked a different kind of question, like one with a structure that just wasn't in the training set. The most impressive ones seemingly cherry-picked their results pretty aggressively.