Abstract:
This post explores the journey of using AI to brainstorm a novel theory: leveraging the I-Ching’s King Wen sequence as a potential framework for self-directed learning in Artificial General Intelligence (AGI). It details the iterative questioning process that led to the theory, key insights into how the sequence might embody advanced learning principles, and the potential implications for AGI development, linking to a resulting research paper.

Estimated reading time: 5 minutes

The I-Ching as a Framework for Self-Directed AGI Learning: A Journey Through Questions

The Questions That Led to Discovery

The theory came from asking a series of deep questions. Each question used what was learned from the one before it:

“If you were to focus on one area for AGI, what would it be?”
- This first question led to exploring self-directed learning as a very important skill for AGI.
“If you were to design a highly effective self-directed learning system, how would you start?”
- This question helped figure out the main parts needed for good learning.
“I will be building an autonomous agent for trading using the I-Ching. Is there a way to design it to help with this research?”
- Here, we started to see a link between old knowledge and new AI.
“Is there something to the King Wen sequence that can be used to predict outcomes?”
- This question started the search for hidden patterns in the sequence.
“If we were to apply Fibonacci to try to find a natural order to the hexagrams, would it be closer to the King Wen ordering or something else?”
- This led to studying the math behind different ways of ordering the hexagrams.
“If the idea is to improve learning by maximizing new information (Bayesian surprise), can we use this idea to cover the missing hexagrams?”
- This question linked ideas from information theory to how the sequence is built.
“Is the King Wen sequence already ordering for maximum new information?”
- This key question showed how advanced the sequence is.
“Is gradient descent the best learning algorithm available today? Does this connect to the King Wen sequence ordering somehow?”
- This final question helped bring together the link between old knowledge and new machine learning.

Each question built upon previous insights. This finally led to the idea that the King Wen sequence could be an advanced learning plan, older than our current information theory.

The Theory: I-Ching as a Framework for Self-Directed AGI Learning: Insights from the King Wen Sequence

New progress in artificial general intelligence (AGI) has made many people interested in systems that can learn on their own. While looking at how old knowledge and new AI meet, I found an unexpected link: the I-Ching’s King Wen sequence could be an advanced guide for AGI systems to learn by themselves.

The Journey of Discovery

We started by asking a basic question about how AGI systems learn on their own. While looking at ways to make learning better, we decided to study the I-Ching’s King Wen sequence using ideas from modern information theory and machine learning.

Key Insights

The King Wen Sequence: An Ancient Meta-Learning Algorithm?

What we found was surprising: the King Wen sequence seems to use advanced learning ideas that we are just now finding again in AGI research. The sequence shows:

A good balance between new things (surprise) and known things (familiarity).
It naturally avoids getting stuck in small problems by making smart jumps.
It changes how fast it learns as needed.
It can see patterns in many different ways at once.
It uses smart loops to go back and understand things better.

Beyond Gradient Descent

Gradient descent is seen as a top method in machine learning now. But our study suggests the King Wen sequence might use an even more advanced method. It naturally handles challenges that modern algorithms often struggle with, such as:

Getting unstuck from small problems without needing someone to fix it.
Changing its learning speed on its own.
Improving in many areas at the same time.
Making good plans for how to learn.

Implications for AGI Development

How advanced the King Wen sequence is suggests it could be a model for building better self-learning systems. Key ideas that could be applied include:

Learning in a path that isn’t a straight line.
Mixing opposite ideas in a balanced way.
Going back to ideas to understand them more deeply.
Making it look at things differently to avoid getting stuck.
Seeing patterns in many ways, both big and small.

Implementation Framework

interface KingWenLearning {
  state: {
    currentPattern: Pattern;
    oppositePattern: Pattern;
    nucleusPattern: Pattern;
    transformationState: number;
  };
  transitions: {
    smallSteps: Pattern[]; // Adjacent hexagram learning
    leaps: Pattern[]; // Non-adjacent insights
    returns: Pattern[]; // Revisiting patterns at deeper levels
  };
  dimensions: {
    direct: boolean; // Straightforward relationships
    complementary: boolean; // Opposite patterns
    nuclear: boolean; // Core patterns within patterns
    transformative: boolean; // Change patterns
  };
}

Future Research Directions

This study opens up some interesting ideas for future research:

A detailed study of the King Wen sequence using information theory.
Creating learning programs based on the sequence’s ideas.
Looking for other old systems that might have hidden ideas about how to learn better.
Adding these ideas into the AGI systems we are building now.

Conclusion

Finding out that the King Wen sequence could be an advanced learning plan, thousands of years older than our information theory, suggests we can find useful ideas for AGI in surprising places. As we keep building self-learning systems, the ideas in this old sequence could give us good advice for making better learning setups.

Here’s the zenodo link to the resultant research paper:

This post came from exploring self-learning systems with an AI assistant. It led to surprising ideas about how old knowledge could be used for new AGI development.

Brainstorming a New Theory of Self-Directed Learning for AI, with AI