Interesting. I did the same but with Police Quest 1 (the AGI version). I vibe-coded an AGI interpreter which my harness could drive an had it run the game. I gave it a scan of the manual for initial information. It's silly how much it fumbles whereas 15 year old me could beat the game fairly quickly. https://youtu.be/JIs1QCcmQIY
I did manage to give it information via. screenshots, planning and several other things. Harness engineering is fun to do and a worthwhile exercise. It finally managed to score 93 of 245 points and then just choke out. I figured that a large language model would be able to burn through a 1987 primarily text based game with a primitive parser but it's not as easy as it sounds. My takeaway was that the LLM is definitely intelligent but not "generally" so.
I did manage to give it information via. screenshots, planning and several other things. Harness engineering is fun to do and a worthwhile exercise. It finally managed to score 93 of 245 points and then just choke out. I figured that a large language model would be able to burn through a 1987 primarily text based game with a primitive parser but it's not as easy as it sounds. My takeaway was that the LLM is definitely intelligent but not "generally" so.