Maybe we need a different approach or maybe more is different.

More training, more data, more parameters, more compute power... and voilà.

Hard to say... but we've been surprised more than once in machine learning history.