OpenAI have just added a new model, text-davinci-003, to their GPT-3 text generators. It is supposedly much higher quality than previous models, so I decided to generate some random encounters and some nanodeities using the same prompts as before. I also, for shits & giggles, decided to ask another AI to make pictures of the encounters, because: why wouldn't I?
Here are the results:
Random encounters
* At the edge of a nearby lake, a giant 3-headed turtle is sunning itself on a rock. It has a large shell with powerful scutes, and its 3 heads have different expressions: one is smiling, one is frowning, and one is screaming. The turtle is friendly, and will willingly carry up to four PCs across the lake on its back. It is also a repository of knowledge, and will answer any question they ask it, but will only do so if they answer three of its own questions first.
The turtle's questions are:
1. What is the most beautiful thing you have ever seen?
2. What is the most powerful emotion you have ever felt?
3. What is the most important lesson life has taught you?
* The PCs come across a clearing in a forest, where four wooden pillars have been erected around a raised platform. On the platform, a circle of five humanoid figures stand, their heads bowed and their hands clasped together. As the PCs approach, the five figures begin to chant a strange, eerie melody in a language none of them have ever heard before. The chanting becomes more and more frantic, and soon the five figures are spinning in a circle and holding hands. The ground suddenly trembles and a shimmering red portal appears in the centre of the circle. The five figures continue to chant, and the portal slowly starts to open, revealing a strange and alien landscape beyond. If the PCs step through the portal, they will find themselves in a place called the Shadow Realm, where they will be able to explore, interact with the inhabitants, and find powerful magical items. However, they must be careful, as the Shadow Realm is a dangerous place, and they must be careful to not become lost in its depths.
* A small rock-strewn valley, at the bottom of which lies a shallow lake. In the lake is a small island, upon which stands a single, ancient oak tree. As the PCs approach, they notice that the lake is home to a small colony of giant fire-breathing dragonflies. The dragonflies are friendly and will offer to give the PCs a ride across the lake, in exchange for a few pieces of copper. Once on the island, the PCs discover a small shrine to the god of fire, surrounded by a ring of standing stones. In the centre of the shrine is a small altar, upon which sits a single, flaming gem. If a PC takes the gem, they will gain a permanent increase to their fire resistance.
Nanodeities
* Gurf, a deity of long-forgotten front doors made from oak and yew, has domain over the forgotten and misplaced. Gurf appears as a moss-covered door with a lion’s head knocker and a keyhole with a glint of starlight. To appease Gurf, kneel in front of an ancient door and whisper a prayer for the forgotten and the lost. Gurf will reward you with an entirely new door, with a shiny brass knob and a lock that can never be picked.
* Marilath is the god of half-forgotten memories, and can be found in the far corners of one's mind. Marilath looks like a silver-scaled catfish, but with a human face and a beaked nose. To worship Marilath, one must sit in a comfortable chair and recite forgotten memories in a soft voice. If successful, Marilath will reward you with a vision of forgotten memories, providing a glimpse into the past.
* Gorack the Great is the god of forgotten roads, appearing as a giant centipede with a glowing yellow eye. To appease Gorack, one must draw a map on a piece of paper and trace a route along a forgotten road. Once the map is complete, Gorack will reward the worshipper with an altered perspective on the world, and will offer guidance on how to find forgotten roads.
Conclusion
These responses do feel slightly superior to the set I generated with an earlier text model. The differences are subtle, but the biggest difference is the one you can't see: with the previous model, I generated perhaps 10 or 15 responses in order to get three which I felt were adequately creative and believable. This time around, I only had to generate the three which you see above.
As for the image AI: that was predictably terrible, but a lot of fun. The images themselves are nice enough, but scarcely representative of the text (although the deities are a lot closer than the random encounters). I could have made the images a little more accurate by tweaking and re-tweaking the text, but life's too short and, as my attempts to get an AI to depict a viking longship made of thorns have taught me, AIs are terrible at generating anything other than simple portraits or landscapes.
However, I adore the fact that the picture of Gorack the Great, god of forgotten roads, incorporates what appears to be a prehistoric version of Google Maps' location pin.
I also learned a new word, "scutes".