Norm Growth

Reproducing mHC: The Art of Constraining Signal

We usually think constraints limit power. In deep learning, they channel it. Visualizing why Manifold-Constrained Hyper-Connections (mHC) stop signal explosion.

January 4, 2026 · 3 min · Shengzhe
The RL loop: agent interacts with environment

The RL Book That Finally Made Reinforcement Learning Click

Most RL intros drown you in notation. EasyRL does the opposite—it builds mental models that survive contact with real training runs. Here’s why it finally made reinforcement learning click for me.

January 1, 2026 · 5 min · Shengzhe