Abstract: Generative artificial intelligence (GenAI) is moving from “Knowledge Engineering” (2020–2023) to “Cognition Engineering” (since 2024). We examine what drives this shift and what it makes possible. First-act models relied on pre-training and fine-tuning to store and retrieve knowledge; they handled routine tasks but struggled with sustained reasoning. The second act marks a step change: models can reason for long periods—from minutes to hundreds of hours—and function as cognitive tools rather than databases. Three factors underlie this change: better, reasoning-dense data; test-time scaling methods; and advances in self-training. We identify five consequences: greater data efficiency; tighter coupling of pre-training and reinforcement learning; a focus on mining cognitive data; stronger performance on complex, real-world tasks; and interaction as a default mode of intelligence. Data engineering also shifts: “journey learning” replaces “shortcut learning,” capturing full decision paths with reflection, error correction, and backtracking. These changes reshape human-AI collaboration and open new ways to tackle difficult problems.