Recent advancements in artificial intelligence demonstrate that true machine understanding is achieved through data compression, specifically by isolating fundamental physical structures from noise. In the first case, the Chinese Academy of Sciences' Institute of Automation introduced PhiZero, a world model that utilizes a "physical language" tokenizer to predict video evolution. By converting a 4-second, 33-frame video into just 256 discrete physical-language tokens—representing a 175x reduction compared to the 44,800 continuous tokens used by standard VAEs—the model separates dynamic state transitions from static appearance noise like texture and lighting. This enables the Reasoner (based on Qwen3-VL-4B) to predict symbolic physics that generalizes across different embodiments and visual appearances. Similarly, researcher Zhang Hongliang has applied AI to nuclear materials science to compress the complex evolution rules of irradiation-induced microstructural changes. Instead of relying on computationally expensive, decades-long physical simulations or empirical testing, his AI models the mechanisms of embrittlement and deformation at the micro/nano scale. Both examples illustrate that by discarding non-essential information—such as pixel appearance in video or brute-force simulation data in materials science—AI models capture the underlying governing principles, allowing for efficient prediction and cross-domain transferability.