https://rifaki.me/ 2026-04-26 weekly https://rifaki.me/portrait.jpg Mouhssine Rifaki Reinforcement learning researcher and PhD student at Imperial College London. Bachelor's in mathematics from Sorbonne, master's from the MVA program at ENS Paris-Saclay. https://rifaki.me/posts/ 2026-04-26 monthly https://rifaki.me/publications/ 2026-04-26 monthly https://rifaki.me/projects/ 2026-04-26 monthly https://rifaki.me/posts/adversarial-examples/ 2026-04-26 yearly https://rifaki.me/posts/img/adversarial-fgsm.png Adversarial examples and the robustness story that survived Adversarial examples were not just a security bug. They exposed a mismatch between predictive features and human-aligned robust features. https://rifaki.me/posts/diffusion-models/ 2026-04-26 yearly https://rifaki.me/posts/img/diffusion-forward-reverse.png Score matching, diffusion models, and why noise became a generator Diffusion models work because denoising estimates the score of noisy data distributions, turning generation into reverse-time dynamics. https://rifaki.me/posts/double-descent/ 2026-04-26 yearly https://rifaki.me/posts/img/belkin-fig1.png Double descent: the U that became a W and what held up at scale A reading of what double descent actually is, what the label-noise caveat does to the story, and what held up at scale. https://rifaki.me/posts/edge-of-stability/ 2026-04-26 yearly https://rifaki.me/posts/img/cohen-fig1.png The edge of stability and what it does to the flat-minima story A reading of Cohen's edge-of-stability result, Arora's theory, and what the phenomenon does to older narratives about SGD and flat minima. https://rifaki.me/posts/flat-minima/ 2026-04-26 yearly https://rifaki.me/posts/img/keskar-fig1.png On flat minima and the argument that won't die Why the flat minima debate keeps restarting. https://rifaki.me/posts/grokking/ 2026-04-26 yearly https://rifaki.me/posts/img/grokking-fig1.png Grokking: a puzzle and four explanations that might be the same thing A reading of the grokking phenomenon in deep learning. https://rifaki.me/posts/implicit-bias/ 2026-04-26 yearly https://rifaki.me/posts/img/implicit-bias-margin.png Implicit bias in gradient descent and the max-margin ghost Why unregularized gradient descent can still select a particular classifier, and why the clean linear result only partially transfers to deep networks. https://rifaki.me/posts/information-bottleneck/ 2026-04-26 yearly https://rifaki.me/posts/img/tishby-fig2.png Reading the information bottleneck paper with the benefit of hindsight A reading of Tishby's information bottleneck paper after the pushback. https://rifaki.me/posts/lottery-tickets/ 2026-04-26 yearly https://rifaki.me/posts/img/frankle-fig3.png The lottery ticket hypothesis: what it claimed and what remained after replication A reading of the original claim, the rewinding fix, and what remained after the replication critique. https://rifaki.me/posts/mode-connectivity/ 2026-04-26 yearly https://rifaki.me/posts/img/mode-connectivity-paths.png Mode connectivity and the case for one wide basin Low-loss paths, permutation symmetries, linear mode connectivity, and what the single-basin picture does and does not prove. https://rifaki.me/posts/neural-scaling-laws/ 2026-04-26 yearly https://rifaki.me/posts/img/kaplan-fig1.png Scaling laws from Kaplan to Chinchilla to broken laws and what still holds A reading of the scaling laws arc, what each paper claimed, and what still holds at current scale. https://rifaki.me/posts/neural-tangent-kernel/ 2026-04-26 yearly https://rifaki.me/posts/img/ntk-linearization.png The neural tangent kernel and the cost of freezing features The NTK made neural-network training analyzable by freezing the features. That was the insight, and also the limitation.