About the role
Physics pretraining is the part of Aether that gives the model its different-from-text scaling exponent. We curate solver-run corpora across CFD, FEA, MD, DFT and RTL; we design the next-state-prediction objective and the conservation-aware regularisation; we run the scaling-law experiments that decide where the next compute budget goes.
The team is six engineers and growing. The work is technical, empirical, and unusually high-leverage. A successful experiment shapes the next year of company direction.
What you'll do
- Design and run pretraining experiments at the scale of 7B to 280B parameters.
- Curate and balance the physics-pretraining corpus across our five core modalities.
- Maintain and extend the conservation-aware regularisation and grid-aware attention primitives.
- Co-author the per-release scaling-law and architecture-ablation papers.
- Pair with the applied team when our research output meets a real customer workload.
Who we're looking for
- Strong empirical machine-learning background. Trained models at meaningful scale; comfortable with the operations side of large training.
- Curious about the underlying physics; happy to spend a week with a numerical-methods textbook when the data demands it.
- Writes clearly. Our papers are published; the writing matters.
Bonus
- Background in computational fluid dynamics, molecular dynamics, or quantum chemistry.
- Prior contribution to a published scaling-law analysis.
- Experience with frontier-scale training infrastructure (parallelism, optimiser-state management, mixed precision).