It's kind of bonkers how much signal is encoded in cooccurrences of words!

But there's a clear alignment of sentences that are useful in the world.

A weak but consistent alignment based on the ground truth of useful utterances in the real world, which allows that inherent structure to pop out at scale.

People say things in the real world to make things happen, which means those utterances have to align with ground truth usefulness, which means they have an implicit structure to extract meaning out of.