No matter how much effort you put in to prompting and coaxing it, GPT-3 really couldn't manage anything but the simplest encounter-based game design.
There were things that GPT-3 did really well -- it could write you a 600-word essay on the characteristics and atmosphere of each chamber or passage you find yourself in -- but it couldn't stick to a narrative or engage in anything resembling long-term planning. So, ultimately, you're left with something quite strange and dreamlike: A campaign that's vividly -- even hyper-realistically -- described, but where nothing makes sense and there's no logical continuity over time.
GPT-4 knows what it's doing, though. It's very different. As far as I can tell right now, it combines the outstanding descriptive capabilities of GPT-3 with the ability to (a) remember things it has previously said and establish logical relations over time, and (b) make sure that the campaign appropriately mirrors a "normal" D&D campaign. GPT-3 would frequently mess up and have you fight dragons at Level 1; GPT-4 more appropriately sticks to goblins and bandits. And, when it comes to tactics, pacing, etc., it's really quite good. I suppose it's a little bit conventionally-minded, but that's to be expected.
Almost scary to think that it'll only get better from here on out, because GPT-4 is damn near "good human DM" tier, and already much better than a bad DM.
No matter how much effort you put in to prompting and coaxing it, GPT-3 really couldn't manage anything but the simplest encounter-based game design.
There were things that GPT-3 did really well -- it could write you a 600-word essay on the characteristics and atmosphere of each chamber or passage you find yourself in -- but it couldn't stick to a narrative or engage in anything resembling long-term planning. So, ultimately, you're left with something quite strange and dreamlike: A campaign that's vividly -- even hyper-realistically -- described, but where nothing makes sense and there's no logical continuity over time.
GPT-4 knows what it's doing, though. It's very different. As far as I can tell right now, it combines the outstanding descriptive capabilities of GPT-3 with the ability to (a) remember things it has previously said and establish logical relations over time, and (b) make sure that the campaign appropriately mirrors a "normal" D&D campaign. GPT-3 would frequently mess up and have you fight dragons at Level 1; GPT-4 more appropriately sticks to goblins and bandits. And, when it comes to tactics, pacing, etc., it's really quite good. I suppose it's a little bit conventionally-minded, but that's to be expected.
Almost scary to think that it'll only get better from here on out, because GPT-4 is damn near "good human DM" tier, and already much better than a bad DM.