The 2024 course by Stéphane Mallat, Professor at the Collège de France, revisits the theme of understanding 'Why it works' in mathematical terms. It focuses on comprehending the very recent results of generative models through Score Diffusion. Thus, learning and generation through random sampling are addressed. The mathematical framework in high dimension is probabilistic, and the key concept is independence. Monte Carlo methods lie at the heart of calculations in high dimension (Metropolis), especially with the Markov chains introduced in 2023. Notions introduced by Fisher (model, inference, etc.) and Shannon (entropy) are reviewed before tackling Score Matching and Score Diffusion algorithms with the Ornstein-Uhlenbeck equation. Neural networks (U-Net) used for denoising provide an excellent tool in this perspective.