Changes:
- Overhauled the main gameplay loop. Before, each action was based on its own prompt which had no relation to the story. Now the game is driven by recent story events, world background, and current characters in the scene.
- Switched to the "adventure" versions of gpt-neo, which leads to much more interesting stories and combat similar to early versions of AI Dungeon. Thanks to mrseeker for helping create these fine-tuned gpt models.
- Improved event detection (somewhat fixes glitchy combat and other issues): Instead of using laundry-list-of-examples prompt engineering which only works for large OpenAI models, we now list the highest probability words after any event. When the words we’re looking for cross a certain threshold, e.g. “the enemy has been [injured]”, it is considered to have happened.
- Added the possible events to each story iteration: Player death, player injury, enemy death, enemy injury, npc death, being attacked by a new enemy, appearance of new NPC, change of location, and (in some cases) player being healed, player receiving a new item. Multiple things can happen in a turn; for example, if a big explosion happens during combat, there's a pretty good chance both you and the enemy are injured or killed.
- You can now one-shot enemies in combat due to the detection of sudden death in the story. This means the name of your item matters more than before. This goes both ways, so your enemy can instantly kill you as well. Fortunately, the player now has plot armor to protect you from sudden death from the story, which can be regenerated by healing.
- Overhauled NPC interactions which can actually affect the world now. For example, you can ask for or threaten someone for an item. If the story says they gave you an item, the AI could detect that you got a new item.
- Fixed an issue where bundling the default models didn’t work correctly. Now you shouldn’t have to download extra things if using the 125M, 1.3B, or 2.7B versions.
Known issues:
- The story prompt can sometimes end up in a bad state because of repeating words, causing the AI to produce low-quality text. We are currently looking into a fix.
- Accuracy of event detection got a lot better but still needs improvement
- 6B model isn't "calibrated" for detection so it's running on 2.7B's calibration which could be unpredictable