Note: This post is also shared on LinkedIn.

What if the path to better AI is not bigger models, but geometric constraints?

Another “Sputnik moment?” Maybe.

DeepSeek’s “mHC: Manifold-Constrained Hyper-Connections” boils down to a surprisingly simple intuition: do not let the signal spill outside the valid space.

In other words: stay on the donut. 🍩