Hey everyone 👋 Wrote up an analysis of DeepSeek’s latest paper on Manifold-Constrained Hyper-Connections (mHC) TL;DR: At 27B parameters, their Hyper-Connected models were crashing during training (3000x signal amplification, gradient explosions at step 12k). They fixed it using constraints from a 1967 matrix algorithm. Full breakdown with code simulations here: https://parthsharmaai.substack.com/p/deepseek-mhc-how-a-1967-algorithm