If you play Entangled, Anchorhead, All Things Devours... the gameplay it's far above an AI can do. In ...devours... you have a time travelling puzzle where you must avoid a paradox with yourself. Good luck with that. And that's a game that will run under Frotz from DOS/Windows 95/old Linux/NetBSD in a 486. Or your basement Raspberry Pi, without even sweating a cycle.
Put a Pentium MMX or a basic RpiA, play Galatea https://www.ifwiki.org/Galatea or Counterfeit Monkey https://www.ifwiki.org/Counterfeit_Monkey and then enforcing an AI to create a coherent game with narrativeness, feelings and drama (without being formulaic on speech or narrative) without shitting itself over a few iterations of gameplays it's a dauntisk task.
And these are text games, I repeat, an Amiga, Atari or a 486 could run without too much hazzle. Yet they are created with trivial tools also runable in a Pentium machine from 1997.
Yet an AI can't reach these levels of gameplay even if they tried. Because maintaing coherence on described media with full of twists and easy to create "hard" puzzles like nothing (for an Inform6 progrmmer) it's a very hard task for a machine having to do plays on semantics.
On the other hand, I do find myself a bit skeptical that such a sophisticated game is likely to be written as quickly as you describe. If a game designed in a week is more basic, then perhaps that would be doable by a good LLM, either now or in the near future.
Later with Inform 7 they created an English like language, describing the game by just using English like sentences. Emily Short created Galatea with Inform6 which is more a conversation experiment than a game, think of Eliza but with stereoids, having even "feelings" and "topics" on the conversation. Depending on how you chat with a statue and what topics you uncover, after a long chat you will reach one ending or another. OFC the topic and the feelings, previous chat... it's dynamic and not pre-generated like a graphical adventure menu.
Then there is Counterfeit Monkey, where you can turn an object into anagrams of others in order to solve clever puzzles, and the object count it's really big. That could be really hard to implement with an LLM.
There are some other authors doing creative things with the parser. Junior Arithmancer by Mike Spivey is fun, but not as narratively rich as Emily Short's work.
Andrew Plotkin's Lists and Lists also turns Inform into a parser.
I think a future LLM could accomplish these feats. There's definitely a theme or class of Interactive Fiction: call it "simulation" or "meta" at play there. Some classic works like A Mind Forever Voyaging used it. And old Lucas Arts games like Day of the Tentacle even embedded games within the game. So given those kind of prompts something could be generated in a couple generations.
Not to take away from the current work. They build upon the shoulders of classic Interactive Fiction while being more forgiving and enjoyable.
I mean, on installing DOTT in any system you are actually installing both games of the series. DOTT just made a fancy way to launch the old MM while being ingame. You can launch it outside from DOTT by hand without opening DOTT first.
Is this "All Things Devours" or a different game? I'm very interested in trying it.