Many years ago, I decided to try to build something like "Skyrim" (a complex, simulation-heavy RPG game) that would describe what's going on instead of rendering it in 3D. Inspired by old gamebooks and Choose Your Own Adventure stories, but also games like Dwarf Fortress or Rimworld, I explored many blind alleys and generally learned things the hard way. This year, I finally got to release a full, commercial game for iOS and Android, based on these concepts.
This video is about those blind alleys and about the strategies I learned to deal with the problems at hand.
I'm not mentioning it in the video, but this is no AI Dungeon. That game and the other GPT-2 and GPT-3 based projects are fascinating, but they also have no ground truth. So they end up being closer to these funny, dadaist improv sessions than to actual games. In my game, there is ground truth, and it's simulated to the level of individual body parts. This game does not use ML (like GPT-2 or GPT-3) to "render" its prose.
Skyrim rendered in text (2017 article): https://filiph.medium.com/skyrim-rendered-in-text-1899548ab2c4
The game (Knights of San Francisco, 2021): https://egamebook.com/knights
You can follow my gamedev exploits on Twitter at https://twitter.com/RaindeadGames
Huge thanks to Lycoris Radiata, who gave me the idea for making this video, and who was kind enough to review the first draft of it and gave me tips on how to improve it. You can reach Lycoris through https://github.com/TheLycorisRadiata.
0:00 Visualization intro
0:52 World state & rendering
1:57 Natural language generation
3:12 NLG & games
5:36 Repetition vs text
6:30 More realism
8:25 Game trailer
9:14 The problem with “radical effects”
10:45 The problem with choices
12:24 Spatial relations in text
13:07 Player reading in interactive works
15:40 But, why?
16:42 Indie & authorial
18:48 The result
20:09 Awkward outro
31 Comments