Muon-catalyzed fusion may never produce limitless energy, writes George Musser in this week’s Qualia, but the exotic atoms and molecules created in this exciting research program are stretching the boundaries of physics.
.@MuonSpace is turning space infrastructure from bespoke projects into a production system.
Eclipse is proud to lead the company’s $250M Series C as it builds the infrastructure layer for the industrialization of space.
More on why we invested from Eclipse Partner and Head of Growth Joe Fath:
They also use Muon throughout the network
- Router AdamW for stability
- GR projections AdamW (suspected due to their weird shape)
- ngram table Adam without weight decay
- per head orthogornalization
They trained the model with TP so for the parameter update, they come up with a parameter assigner to balance flops across ranks. The ranks then exchange via all to all
GDN-2 scales nicely with Muon 🚀
We trained hybrid MoE models at three scales (2B-7B) with Muon.
GDN-2 matches Mamba2's loss with ~10% less training compute!
Steepest Descent on Manifolds: 6. Muon + Double Rotation
Introduces MuonR — a Muon variant that constrains updates to left & right rotation matrices. This preserves the singular value distribution of weights, providing a clean, elegant way to maintain training stability.
From its origins to its next chapter of growth, @MuonSpace CEO Jonny Dyer sits down with Nasdaq to discuss the company’s journey and its vision to redefine how mission-critical intelligence is delivered from space.