Cohere Labs - Franz Srambical, Co-founder of pdoom.org

World models offer a compelling framework for training agents by learning a simulation of the environment, a paradigm quickly gaining traction for its potential to solve data scarcity in fields like robotics. However, a fundamental open question remains: Is it truly feasible to build world models that can accurately simulate complex environments for agent training?
In this talk, I will introduce Jasmine, a performant open-source JAX-based world modeling codebase, representing our first milestone towards answering this question. Jasmine is a production-ready framework built for scalability, enabling researchers to train diverse world model architectures—from Transformers to diffusion models—across hundreds of accelerators. I will discuss the engineering decisions behind Jasmine that make it a robust tool for large-scale research and how we are using it to establish "Empirical Environment Complexity Scaling Laws” by systematically measuring the compute and data requirements to model environments of increasing complexity, providing a concrete path towards understanding the future of world models and their role in building more capable agents.

Franz is co-founder of pdoom.org, an open research community focused on addressing core blockers towards general intelligence that will not be solved by scaling up compute. His work spans the entire stack from kernel-level optimization to large-scale distributed systems. His previous work includes open ML infrastructure, crowd-sourced software engineering datasets as well as pre-training and reinforcement learning of models at scale. He previously worked at Celonis, is founding member of neuroTUM, and has contributed to the Linux kernel.

Add event to calendar

Apple Google Office 365 Outlook Outlook.com Yahoo