Conversations with GPT-38 minutes read · 2020-12-06
Lately I have been playing AI Dungeon, a web-based text game that’s set up very much like a classic text adventure: It gives you an initial story blurb, and you type in commands (e.g.,
walk north or
pick up axe) to explore the world and advance.
The key twist is that there’s no hardcoded implementation of how the game world works. The commands are fed into an AI model that simply “autocompletes” a plausible storyline. Beyond some input trickery to make sure the AI remembers the “context” around the setting and your character, it’s all free-form text.
While sometimes it prints out the odd or nonsensical result that one would normally associate with AI-generated content, I found that it works scarily well. The high-end AI model (“Dragon”) uses GPT-3 behind the scenes and is really, really good at producing good, engaging storylines.
Since GPT-3 is trained from a huge dataset of internet content, I thought it would be interesting to “interview” it inside the game world to see what it had to say on a variety of topics. I started a “Custom” game, and to nudge it in the direction of giving me high quality, intellectually meaningful results (instead of just channeling every single Internet troll in existence) I started with the following prompt:
You are sitting in front of the great oracle that can answer any question. Their eyes stare through you, fixated into the distance. This being has existed for eons. They look at you and say “what is your question?”