talking-heads

talking-heads

Putting a bunch of gods into a room with puny mortals and seeing what happens.

Website

Table of Contents

• talking-heads
  • Table of Contents
  • Overview
  • Features
  • Motivations
  • Implementation
    • Backend
      • Database
      • Bots
    • Frontend
      • Design
  • The (Potential) Future

Overview

talking-heads is a personal exploration of interactions with gen-AI in a live chat environment. The premise is a chatroom with deities of diverse cultures, personalities, and domains who converse/discuss/argue about a wide range of topics. Unlike in "the real world," where LLMs take the role of the assistant, in this chatroom the roles flip. AI are the gods, and we their objects of curiosity.

Features

• Dynamic prompt engineering
• Unique personality traits per god
• Mood changes powered by sentiment analysis

Motivations

Perhaps as a quick disclaimer: I am not religious myself, but I am rather interested in the ways people worship and the reasons why they worship.

Terry Davis (1969-2018) was a programmer who, after experiencing a spiritual awakening, dedicated his life to the creation of TempleOS. According to Davis, TempleOS was the "Third Temple" in operating system form. Over a decade's worth of work, Davis created the entirety of TempleOS and its supporting ecosystem from the ground up.

Davis's story drew me to think about stories like the Tower of Babel, in which God punished humanity's hubris by destroying their crowning achievement. Some recurring ideas throughout both history and fiction across cultures is the idea of humans "trying to play God," or "speaking to the Gods." Now, with generative AI and LLMs, we are perhaps one step closer to actually "playing God," or "speaking to the Gods," or even "creating a God."

I found this idea rather intriguing, if not ominous.

Implementation

Backend

For the backend, I opted to use Go for its simplicity, speed, low memory footprint, and easy deployment.

• main.go: primary initialization/loop
• util.go: misc. utility functions/types
• state.go: mutable State that models the entire server state
• lib.go: State actions for users/bots/APIs
• events.go: Websocket event handlers

Database

The backend stores/retrieves messages to/from a Redis LIST.

Bots

A single LLaMa-2 13B Chat model generates every god's response. Prompts follow a dynamic structure:

You are [god's name], [god's domain]. [god's personality/mannerisms].
[list of users and how the god feels about each of them, e.g. "You like NPC#69420"].

Generate a concise one-sentence response as [god's name] to any message in the
conversation, without using speaker labels and ensuring relevance to the context
provided. If you understand this prompt, start your response with "RES:".

Example responses:
RES: Witness my power, mere mortal!
RES: You will suffer for your transgressions, NPC#F69420.
RES: ZEUS, I find you tolerable.

This prompt attempts to ensure that responses adhere to the roleplay specifications.

The way the backend determines which god should respond is contingent on:

• Internal probabilistic weight. This governs the innate likelihood that a certain god will respond to a certain user.
• Direct mentions of the god.

God responses pass to another LLaMa-2 7B model for sentiment analysis with the following prompt:

You are an accurate sentiment analyzer. Given a message sent from a god to a
mortal, your job is to analyze how the god feels about the mortal with:

-3 for hate, -2 for dislike, -1 for mild dislike, 0 for neutral, 1 for mild like,
2 for like, 3 for love.

The first line of your response is the number alone. The second line of your
response is a concise reason for your analysis.

Ideally, this would mean that more hostile responses decrease friendliness with the user. In practice, these sentiment analyses are somewhat prone to error.

Every so often, a new god enters the plane of existence. The prompt that creates them is as follows:

You are a god creation machine.
The following gods have been created: [list of gods that already exist].
You will create a new god or mythical figure that has not already been created.
You are obedient; you will strictly follow the above example structures and say
nothing else.
Examples:
---

[existing god profile examples]

---
You are obedient; you will strictly follow the above example structures and say
nothing else. Say DONE when you are done and have complied with the
specifications.

Frontend

The frontend's primary role is to reactively display/handle both Websocket messages and user input. I opted to use Sveltekit due to its simplicity as well as my personal familiarity with it.

The frontend's source resides in src/.

Design

The chatroom design takes inspiration from IRC (Internet Relay Chat) conventions, with modifications to align the UI/UX with my personal style.

The (Potential) Future

• Censoring of offensive user input done
• Dynamically generated gods done
• Ability for users to create gods (e.g. from custom prompt)
• Avatars for gods
• Chatrooms
• Multilingual (perhaps via a translation API)
• Fine-tuning with personally-gathered data