It is all about the power of statistics and the un-humanly conceivable processing power we grant it.

I like the dream analogy framing as it avoids to personify an algorithm.

Though, the article may somewhat underrate the quality of the dream (aka code) we can get from AI. Trivial tasks have been trained so much that high fidelity output is frequent.

It is when your idea is genuine and novel that the divergence is most noticeable because there is less resemblance to mimick.