Two-minute read; links to the underlying papers, model cards and case studies inline.
Aether v1.0 is now generally available across every discipline surface. The 280B sparse-MoE checkpoint is in production. We published the architecture diagram in full at /simos — including the modality fan-in and the discipline-head decoder routing. Loss curves and ablation tables in the paper at /research.
A named aerospace prime moved a structural-margin workload from the incumbent. Eight-week pilot. Result: 6.4× cycle-time reduction at iso-fidelity vs the legacy CAE solution. Comparison memo co-authored with their CAE director. Story at /customers.
Refusal corpus extended with two CBRN-adjacent categories under the tool-policy layer. False-positive rate on benign chemistry queries held at <0.2% on the internal red-team set. Edge 1.3B graduated from preview to GA. Full diff at /simos#card.
What is the right eval cadence for a workload that takes weeks to produce ground truth? Reply to this email with your view. We will publish a synthesis in the August issue.
- (1) On compute-bound vs IO-bound forward-rolling — Tanaka et al.
- (2) The case for sovereign foundation models — Holst & Mercier.
- (3) Why public benchmarks plateau — Bauer.