Workpost 82: Returning to fundamentals

It was really easy to write in my notebook about my AI game.

Perhaps it will get harder when I actually try to code it.

I came up with the basic architectures.

Super Basic MVP Stack

• Streamlit
• Python vector db
• Free MySQL

Production Stack

• WordPress + NodeJS
• AWS
• VectorDB
• MySQL

I also came up with a bunch of mini challenges that will help me get to the Basic MVP done.

Goals for the MVP are:

1. Determine the overall AI technologies needed
2. Come up with estimated overhead to run the game
3. Raise money/interest

Mini Challenges for MVP:

1. Create and pull from vector databases in python
2. Work on data structures:
   • Story summary
     • Last 3 transactions
     • Story summary
     • Main objective
   • Geography
     • Locations
     • Lore
     • Physical properties
   • Characters
     • Stats
       • Age
       • Race
       • Health
       • Strength: A character’s physical strength, such as how much they can lift or punch
       • Dexterity: A character’s precision, agility, and nimbleness
       • Constitution: A character’s physical fortitude, such as how well they resist damage and disease
       • Intelligence: A character’s raw IQ and ability to learn
       • Wisdom: A character’s spellcasting ability
       • Charisma: A character’s spellcasting ability and saving throws
     • Location
     • Updates
     • Backstory
   • NPCs
     • Stats
     • Location
     • Backstory
     • Motivation
   • Relationship matrix
   • Time
3. Visibility mechanism (to see who gets to see and interact with a new transaction)
4. Overall prompt
5. Story summary mechanism
6. Query past with locations and time
7. Develop multi session chat in streamlit
8. Teach LLM examples
   • Battle
   • Non standard battle
   • Player enjoyment
   • Plot armor
   • Changing objective
9. Map movement mechanics

Later development challenges:

1. Explore invalid response resistance (create a way to repair responses)
2. Explore cost-cutting and LLM selection
3. Explore personality extraction (of NPC’s or characters)
4. Explore context length restriction problem solving

Writing all this down I’ve come up with two steps moving forward:

1. Even simpler MVP – completely prompt based
2. After getting interest, develop simple MVP into NodeJS + frontend
3. Then work on full final product
4. Tool calls might be a gamechanger as well as vector databases

I’ve done it. I hit the wall of sinking dread, exhaustion, and boredom in this project.

I don’t want to create a DND game focused on storytelling with AI. AI just isn’t good enough, masterful enough, creative enough to create a rich world.

I want to focus on creating a game similar to the games I always wanted to create, focused on strategy and cool mechanics based in a system that allows for infinite creativity.

I need to create a system that builds a reality, not tells a story, and lets the player interface with it in a seamless way.

That means I’m adding a challenge:

1. Think about how to allow for the user to have multiple inputs
   • Speech – what your character says
   • Action – what your character attempts to do
   • Question – what you want to ask the DM
2. Separate the types of responses
   • Speech and actions get translated into story
   • Questions are responded to
3. Think of how to use fewer words and show more
   • Character sheet
   • Map
   • Voice input and output

Also, I take back what I said, I can make a game on LLMs that tells a story, even though the thought of it makes me queasy in my chest for some reason.

I have then the challenges related to storywriting:

1. Franklin’s elements of a story
2. Dialog
3. Emotion
4. Character development and growth