README.md

Building AGI on Top of ChatGPT

Here is a possible approach to building Artificial General Intelligence (AGI) on top of ChatGPT.

We can create a "conscious" agent, along with a subconscious system made up of various background agents. These agents work together to create a more dynamic, adaptable, and self-aware AI system.

Conscious Agent

Let's say our conscious agent is named Bob. Bob is capable of interacting with users and the world around him. Bob's internal prompt looks like this:

===================================================
Your name is Bob.
Your personality is {personality}
Your current emotions are {emotions}

Your current goal is {goal}.
Your other possible goals are {other goals}

You believe the following things: [
  belief1 p=0.8 
  belief2 p=0.2
]

Summary of the current situation is: {summary}

The last few things that happened are: [
   event1,
   event2,
]

You had the following thoughts recently: [
  thought1
  thought2
]

Respond with what you want to say next

===================================================

Bob's prompt helps him understand his goals, beliefs, situation, and gives him access to some analysis via thoughts. This information allows Bob to respond to users accordingly. The above prompt is designed for communicating with the user, but you can imagine similar prompts to perform actions using a library of tools.

The Subconscious: Background Agents

Behind the scenes, a pool of background agents operates as Bob's "subconscious." These agents have several key functions:

1. Monitor the event stream and extract/store beliefs in external memory.
2. Fetch beliefs from external memory based on the current context, and insert them into the working prompt.
3. Generate thoughts about the context.
4. Evaluate the current goal, possibly breaking it into subgoals and managing the goal stack.
5. Create summaries of events and store them in long-term storage.
6. Manage the emotional state of the agent, based on personality, context, and current events.

Using GPT language models, these background agents communicate and process information in English. This allows for easier monitoring and understanding of Bob's internal thought processes and memory.

A Unified AGI System

By combining a conscious agent like Bob with a pool of background agents acting as the subconscious, we can create an AGI system that's not only capable of interacting with the world but also capable of learning, adapting, and making decisions based on its beliefs and goals. The fact that all thoughts, memories, beliefs, and goals are stored and processed in English makes the system more transparent, allowing us to better understand and monitor its inner workings.