Professor Yi Ma argues that intelligence is not about scale or statistical correlation, but rooted in two universal principles: parsimony (compression into low-dimensional, structured representations) and self-consistency (closed-loop memory that simulates and verifies reality). He contends that large language models excel at memorizing compressed human knowledge—not understanding—and lack the abstraction needed for scientific reasoning, as shown by failures on tasks like the ARC Challenge. Similarly, 3D generative models like Sora reconstruct appearances without grasping spatial semantics or frame-of-reference switching. Ma traces these ideas back to cybernetics and evolutionary biology, showing how noise, diffusion, and benign non-convex optimization landscapes all serve compression-driven learning. His CRATE framework derives architectures—including transformers—from first principles, enabling more interpretable, efficient, and scalable models. Ultimately, he calls for a shift from empirical engineering to a rigorous, unified science of intelligence grounded in geometry, information, and feedback.