PinnedInBecoming Human: Artificial Intelligence MagazinebyHaitham Bou Ammar·Nov 14, 2022Safe Reinforcement Learning — Part IMany practitioners using reinforcement learning (RL) are often concerned about the safety of SOTA deep RL techniques. With that in mind…
Haitham Bou Ammar·May 11LLMs Don’t Need Infinite Context. They Need Better Recursion!Every few months, the AI world gets excited about a bigger context window.
Haitham Bou Ammar·Feb 15Beyond LLMs: Sample-Efficient AI That Actually Moves Hard MathSphere packing is a 2000-year-old problem. We reframed a state-of-the-art proof technique as a game and used Bayesian optimisation + tree…
Haitham Bou Ammar·Feb 11Multi-Task GRPO: How to Post-Train LLMs Without Leaving Any Task BehindA robustness-first way to make reasoning models reliably good across tasks and fix for the “zero-gradient” trap.
Haitham Bou Ammar·Jan 31We Didn’t Train the Model. It Started Reasoning Better AnywayThe uncomfortable question: What if RL isn’t teaching LLMs how to reason?A response icon1A response icon1
Haitham Bou Ammar·Apr 5, 2025From VR to Grippers: How We Built a Bimanual Robot That LearnsIn recent years, we’ve seen enormous progress in AI — from large language models to powerful vision systems. However, when it comes to…
Haitham Bou Ammar·Oct 11, 2024Accelerating Language Model Inference with Mixture of Attentions: Speculative Decoding 2.0Language Models (LLMs) are revolutionizing the world, but their immense computational demands have made them expensive and slow to deploy…
Haitham Bou Ammar·Sep 9, 2024From MCTS to Alpha-Zero with PyTorch — Part I (Building a Tic-Tac-Toe’r)AlphaZero is a deep reinforcement learning algorithm developed by DeepMind that has achieved superhuman performance in games like Chess…