As artificial intelligence advances beyond vision and language processing toward embodied systems—including autonomous driving and robotics—prevailing deep learning paradigms are confronting fundamental challenges. This talk presents a systematic overview of physics-native intelligence (Phi), covering its core design principles and methodologies for constructing its world model, data structure, and training algorithm. Four representative physics-native algorithms are elaborated: (1) RAD: A neural network optimizer engineered with built-in symplectic preservation to guarantee long-term training stability. (2) DACER: A reinforcement learning algorithm that equips neural network policies with the capacity to model multi-modal action distributions. (3) RACS: A ternary iteration framework for addressing safety constraints, which jointly learns the feasible region and optimal policy with provable monotonicity and convergence guarantees. (4) BOOM: An algorithm enabling bidirectional co-improvement with reliance on world model. By internalizing planning capabilities directly into the policy, it substantially enhances sample efficiency.