Perhaps I'm not in the right circles, but I see nobody discussing this: LLMs are incredible at linguistic games.

i.e., I yanked my entire game codebase into the context window, then asked Claude to write perk descriptions in haiku form. Not only it did so with results that are fairly poetic, in many cases it captures the essence of the functionality of said perks from the actual code logic tied to it.

It does that in a literal second per perk, whereas even the most impressive human would likely need several minutes to get the logic then conjure a haiku for it.

LLMs are incredible at reproducing the _structure_ of human language. Like remarkably good at it. 

It also _looks_ like intelligence but is fundamentally different. This is where most of the confusion lies. 

They’re basically giant Excel spreadsheets, where you put words in at the top and get convincing reply words out at the bottom