Predictive and Causal Learning

To achieve AGI, a system must move beyond simple "curve fitting" or pattern recognition. It needs to understand how the world works.

Predictive Learning (System 1)

Most modern AI (like Large Language Models) is primarily predictive. It learns to predict the next token, pixel, or state based on historical patterns.

• Strengths: Incredible speed, handles high-dimensional data (images/text), associative.
• Weakness: Lacks a "mental model" of physical reality. It can generate physically impossible scenarios because it only knows "what follows what" in data, not "why."

Causal Learning (System 2)

Causal learning involves understanding the underlying mechanisms that govern events. It answers the question: "If I do X, what will happen to Y?"

• Intervention: The ability to test hypotheses (in reality or simulation).
• Counterfactuals: Reasoning about "what might have been" if a different action was taken.

The Structural Difference

graph LR
    subgraph "Predictive (Correlation)"
    A[Data A] -->|Pattern| B[Data B]
    end
    
    subgraph "Causal (World Model)"
    Cause[Cause] -->|Mechanism| Effect[Effect]
    Action[Intervention] -->|Changes| Cause
    end

Why Causal Learning is Essential for AGI

1. Safety: An AGI must understand the causal consequences of its actions to avoid unintended harm.
2. Transfer Learning: If you understand the causal rules of gravity, you can apply them on Earth or Mars. Predictive models often fail when the context changes slightly (distribution shift).
3. Planning: True planning requires a "World Model" where the agent can run mental simulations of cause and effect.

Example: The Broken Lamp

• Predictive AI: Sees a picture of a lamp and a cat nearby. Predicts "Broken Lamp" because in many training images, cats knock things over.
• Causal AGI: Understands the physics of mass, force, and gravity. It knows that if the cat exerts X force on the lamp, the lamp will fall because it exceeds the stability threshold.

Next: Architectures Overview