You'll architect the reasoning core at the heart of Luma's world-modeling - the intelligence governing the world-simulations behind products like Dream Machine and Ray3. This is core research that powers everything Luma ships.
The role treats science and engineering as one discipline: you'll work across modeling, data, systems, and evaluation on problems where no playbook exists. It fits someone with first-principles intuition for scaling who understands why architectures succeed or fail, not just what the literature says. If you want to specialize narrowly in only the science or only the engineering, this asks for both.
What You'll Own- Drive the core research powering all of Luma's products - co-design multimodal representations, advance long-context training, and establish rigorous scaling laws.
- Close the gap between training loss and user experience with proxy tasks and automated metrics that guide research decisions.
- Build the research infrastructure for high-velocity work: production-research parity, reproducibility, and systems for rapid experimentation.
First 90 DaysOne way the first 90 could unfold.
- Days 1-30 - Immerse & Diagnose: Learn the current models and evals, and where the biggest scaling and quality gaps are.
- Days 30-60 - Ship & Validate: Land a modeling or evaluation improvement that moves a metric that matters to users.
- Days 60-90 - Scale & Systemize: Turn it into scaling laws and infrastructure the whole research team benefits from.
What You Bring- A Bachelor's, Master's, or PhD in Computer Science, Machine Learning, Physics, or Mathematics.
- First-principles intuition for scaling: you understand why architectures succeed or fail at scale.
- Fluency across frontier AI research and engineering as a single discipline.
- Proven ability to design and rigorously analyze experiments and articulate complex concepts.
- Practical experience with distributed or high-performance computing and optimizing training runs on large GPU clusters.
Nice to Have- A track record of publishing at top-tier venues (NeurIPS, ICML, ICLR).
- Proven ability to build and lead research infrastructure with production-research parity.
- Strong software-engineering practices: readable, reusable code, tests, and documentation.
- Experience with low-precision training and hardware-aware optimization for next-gen clusters.
