Causal-Dynamical-AI: Building Agents with Causal Reasoning Capabilities from First Principles

Current large language models (LLMs) are essentially statistical pattern matchers, lacking causal reasoning, forward-looking planning, and structural extrapolation capabilities. The Causal-Dynamical-AI project starts from first principles, integrating dynamical systems, probabilistic graphical models, and deep learning, aiming to build agents with world modeling capabilities. The project provides a structured learning path covering mathematical foundations, causal logic, state-space models, etc., and as an open learning journey, community contributions are welcome.

Current AI systems have three core limitations: 1. Lack of causality: Confusing correlation with causality, suffering from the "causal parrot" problem as termed by Judea Pearl, with weak counterfactual and intervention reasoning capabilities; 2. Weak planning ability: The autoregressive mechanism of the Transformer architecture lacks true multi-step forward planning, and Chain-of-Thought technology does not fundamentally solve the problem; 3. Limited extrapolation ability: Out-of-distribution (OOD) generalization performance drops sharply, making it difficult to handle structural changes in open environments.

The project proposes world models as a solution path, focusing on three technical routes: 1. Latent world models (DreamerV3/MuZero): Learn low-dimensional latent states to represent the world, supporting efficient planning and long-term prediction; 2. Joint Embedding Predictive Architecture (V-JEPA): Promoted by Yann LeCun of Meta, predicts missing representations in the concept space and captures the semantic structure of scenes; 3. State-space models (Mamba/S4): Based on control theory and signal processing, handle long sequences with linear complexity and have infinite memory capacity.

The project adopts a phased learning architecture: 0. Mathematical engine: Master ordinary differential equations, vector calculus, etc., with reference to Strogatz's Nonlinear Dynamics and Chaos; 1-3. Causal logic: Learn probabilistic graphical models (Bayesian networks, MCMC, etc.), with reference to Kevin Murphy's Probabilistic Machine Learning and Judea Pearl's works; 4. State-space models: Dive into architectures like Mamba/S4, mandatory reading of Gu & Dao's Mamba paper; 5. Emergent phenomena: Explore cellular automata, attractors, etc.; 6. Reasoning ability: Learn Monte Carlo Tree Search (MCTS) and PDDL planning; 7. Embodied intelligence: Apply to tasks like V-JEPA and world models; 8. Meta-learning: Explore technologies like MAML to quickly adapt to new tasks.

Code implementation and learning methodology: 1. LaTeX derivation: Re-derive the mathematical principles of each concept to build a solid foundation; 2. Manual implementation first: Write from scratch with NumPy (e.g., Euler solver) before using libraries like scipy/torchdiffeq; 3. Visualization-driven: Understand the dynamic behavior of dynamical systems (e.g., attractors, bifurcations) through interactive visualization. Core tech stack: PyTorch, pgmpy, mamba-ssm, matplotlib, sympy.

Research significance and community value: - Represents a research attitude that delves into underlying principles, not chasing short-term SOTA metrics; - Provides systematic learning resources for AI researchers/students, covering a wide range of topics from differential equations to meta-learning; - Open community, contributions for corrections and improvements are welcome. Limitations: - Resource constraints: Difficult to compete with large-scale industrial experiments; - Verification difficulties: Lack of standardized evaluation benchmarks to assess concepts like energy minimization and structural extrapolation; - Engineering complexity: Great challenges in designing a unified system integrating dynamical systems, probabilistic graphical models, and deep learning.

Summary: This project embodies a research orientation that inquires into the essence of intelligence, reminding us that AI progress needs to focus on core capabilities like causality and planning. It is a map leading to the forefront of autonomous machine intelligence, providing an entry point for basic research. Outlook: As a learning journey, researchers interested in AI basic research are welcome to join the community, contribute improvements, and jointly explore the core ideas of the next-generation intelligent systems.