AI Alliance Office Hours: PyTorch 2025 with Joseph Spisak
Join Joe Spisak for a deep dive into the latest advancements from PyTorch, focused on making distributed AI and Reinforcement Learning (RL) more standardized and scalable.
Key Announcements & Project Spotlights 🌟
OpenM Launch (RL Environment Standard): A new collaboration with Hugging Face aimed at defining a de facto standard for RL environments. This ensures consistency and simplifies sharing across the community.
TorchComms Project (Fault-Tolerant Communication): Introducing a single, standardized interface for fault-tolerant communication collectives. Designed to support diverse hardware, addressing common scaling constraints.
Monarch Project (Efficient Distributed Training): A powerful single controller that simplifies distributed training by treating GPUs as "meshes". It's key for efficient multi-data center pre-training and fully asynchronous RL.
Open-Sourcing Internal Stacks: PyTorch has open-sourced a vast amount of its internal technology, including kernels and production-grade RL post-training and agent-building methods.
PyTorch Foundation Growth: The Foundation welcomes Ray as a hosted project, and Dell Technologies joins the board, accelerating the drive toward an open and enterprise-ready ecosystem.
About our speaker
Joe Spisak is a Product Director in Meta’s Superintelligence Labs (MSL), where he leads product efforts across PyTorch and Meta’s Agentic platform. He has over a decade of experience building AI platforms, with leadership roles at Meta, Google, and Amazon. Joe helped make PyTorch the world’s leading open-source AI framework and guided its transition to the Linux Foundation in partnership with industry leaders. He also spearheaded the open-source strategy for Llama, making cutting-edge models broadly accessible and enabling the community to scale AI in an open and collaborative way. Joe is an active angel and advisor to next-generation AI startups such as Anthropic, General Reasoning, ReflectionAI and many others.