Vlad Feinberg is a lead for Gemini pretraining at Google DeepMind. His public work explains language-model scaling, inference efficiency and practical routes into systems and agent research.[1]
Start with A Frontier Lab Insider's Path: Kernels, Agents, and Math Maturity →
X · feinbergvlad Career
After a computer-science degree at Princeton, Feinberg began an EECS doctorate in Berkeley’s RISELab. He left for Sisu, where he was Head of ML, then worked in Google Research’s Cerebra group before Google DeepMind. His work has moved between university research, startup engineering and model pretraining.[1]
Contributions
Demonstrate a skill at the edge of the model
His frontier-lab essay proposes two accessible directions: make model execution more efficient, or run controlled experiments with agent loops. Kernel work requires reasoning about memory, communication and physical hardware constraints. Agent work requires experiments that reveal behavior rather than simply using an assistant. He pairs those projects with mathematical foundations and close reading of existing research.[2]
Account for the model’s life after training
His pretraining lecture contrasts training-compute scaling with the constraints of serving a model. Latency, memory bandwidth, expected demand and data reuse can change the useful model size and training horizon. The slides connect these tradeoffs with smaller models and inference-efficient design, while warning that fitted scaling laws depend on the experimental setup.[3]
Tools, writing & talks
Writing, research & talks
Start here
A systems-oriented study plan built around kernels, controlled agent experiments and mathematical maturity.
Start with the kernel and agent-work sections. Programming and linear-algebra experience help; the later exercises require JAX and access to the specified accelerator environment.
Author[2]Original source- Gemini Pretrainingvladfeinberg.com
A lecture on scaling, serving constraints and inference-efficient model design.
Lecturer[3]