[tagraves]

To be honest, I don’t find this particularly impressive. I didn’t read to the end, but the beginning parts of the game are littered with errors. Basically every command is followed by ChatGPT miscalculating the remaining resources and the player having to correct it. Could the player have just asserted they had different resources? I’d guess they could have. Another poster mentioned “good enough to be used” as a criterion, and this is far afield from that. It’s not good enough to be used for a game, much less any serious business practice.

[mfabbri77]

As already said, yes of course, because it's just a conversational model, not an AGI... it doesn't really understand the logic of the game. But it perform quite good overall, and it can be improved with more training data. Maybe it'll be playable with GPT4? The question is: can it become so good at pretending to be intelligent, to convince a human?