Physics Of AI
Physics of AI is [[LiuZiming|Liu Ziming]]’s name in 149. 亲历中美 New Labs 资本狂潮，和清华刘子鸣聊：AI for AI、机制可解释性和 Max Tegmark for using physics-style methods to understand, predict, and design AI systems. It reverses AI For Science: instead of applying neural models to physics, Liu wants scientific structure, abstraction, controlled experiments, and theory-like compression to improve AI.
The source breaks Physics of AI into three perspectives. The spatial perspective looks inside models at neurons, features, representations, and internal structure. The temporal perspective watches model properties evolve through training. The “parallel universe” perspective uses controlled experiments, phase diagrams, and condition changes to ask where a trick or architecture actually works.
Key Claims
- AI research is currently too empirical and too concentrated around a few architectures to count as a mature science.
- Physics of AI should identify stable structures and laws behind model behavior rather than only benchmark winners.
- It overlaps with Mechanistic Interpretability but does not require every explanation to be neuron-level.
- Training dynamics are part of the object of study, not just a path to a final checkpoint.
- Controlled variation across data, architectures, optimizers, and random seeds matters because single-run explanations can be brittle.
- Physics of AI is the intermediate layer Liu wants before AI For AI becomes reliable.
Connections
- [[LiuZiming|Liu Ziming]], Max Tegmark, and MIT — source people and research context.
- AI For Science — reversed direction: AI for science versus science for AI.
- Mechanistic Interpretability and AI Interpretability By AI — adjacent model-understanding routes.
- [[KolmogorovArnoldNetworks|KAN]], Transformer Architecture, and Hardware Lottery — architecture and model-design cases.
- OPHIS Research Workflow, Meta-Model Training Curve Prediction, and AI For AI — automation path built on physics-like structure.