Minimum qualifications:- PhD in Computer Science, a related field, or equivalent practical experience.
- 4 years of experience with research agendas across multiple teams or projects.
- One or more scientific publication submission(s) for conferences, journals, or public repositories.
- Experience working with deep learning frameworks (e.g., JAX/Flax, PyTorch) for distributed systems and accelerator clusters.
Preferred qualifications:- 2 years of experience in coding and leading multiple research efforts and influencing research direction.
About the jobJust as grading olympiad proofs takes someone who can solve olympiad problems, judging an AI-proposed change to a frontier training algorithm takes a researcher who could have invented it. In this role, you are that researcher.
You will invent new machine learning (ML) architectures, optimizers, and training methods, and build the systems that let Gemini propose, implement, and test such changes on its own. As the team's final technical verifier, you will decide which AI-generated ideas are real advances and carry the winners into next-generation Gemini models.
This is a deep individual-contributor role. You'll combine original research in ML algorithm and architecture design with engineering across the full stack, from model code and accelerator kernels to distributed training and evaluation.
Individual pay is determined by factors including job-related skills, experience, and relevant education or training.
US: $207000 - $300000 (USD) 20% bonus target equity benefits
Learn more about benefits at Google .
Responsibilities- Invent, implement, and scale new Machine Learning (ML) algorithms, such as architectures, optimizers, training objectives, and recipes.
- Design and build automated research systems in which Gemini proposes, codes, and evaluates changes to its own training stack, moving toward a model that designs its successor.
- Serve as the final verifier for AI-proposed algorithm changes: reproduce results, design decisive ablations across model scales, rule out confounds such as evaluation leakage or unequal tuning budgets, and decide what moves forward.
- Build and own lean, high-performance training and evaluation code (JAX on Tensor Processing Units, with custom kernels where needed) allowing researchers and AI agents can test ideas at meaningful scale.
- Partner with Gemini pre-training, post-training, and infrastructure teams to land validated improvements in production models, and share results through internal talks.
