Modern models get better mainly by adding more numbers. We don't think that's the only way. MatterWave AI is a small research effort testing whether a sharper structure, not a bigger budget, can do the same work with far less.
Inside a normal model are large grids of numbers, and each one is free to become anything during training. That sounds powerful, but it's mostly wasted. The model relearns the same patterns over and over, and nothing stops a grid from quietly making a signal much bigger or smaller as it passes through.
Our design, the Geometric Rotation Transformer, takes that freedom away on purpose. Each grid is replaced by a set of rotations. It uses far fewer parameters to do the same work, and those properties hold before any training even starts.
We'd rather find the right shape for a model than spend more compute on the wrong one. Every part of it has to earn the numbers it uses.
When something is proven, we say so. When it's still being tested, we say that too, and leave the result blank until the number is real.
The idea is simple at heart. We describe it honestly, without dressing it up as more than it is.
The full explanation, the diagrams, the size comparison, and the honest limitations are all in our first research note.
THE GRT NOTE ↘Questions, collaborations, or press — reach us at david@matterwave.ai.