Meta By The Bay

Meta, of course, needs no introduction. Known globally for platforms like Facebook and Instagram, Meta is now doubling down on AI infrastructure, open-source innovation, and building the metaverse.

RSVP

Our connections with Meta have grown over the years, particularly around open-source AI and production-ready systems. Engineers from Meta regularly participate in our meetups, sharing insights and practical strategies for deploying AI at scale.

On August 5th, at AI Agent SF Meetup #5 – Agents in Production, Kai Wu & Nikhil Maddirala presented “Llama Stack in Production” demonstrating how Llama Stack enables easy building and deployment of applications with Llama models using APIs and a plugin architecture. (pro tip start from 46:07).

As the creator of PyTorch, Meta plays a central role in advancing the framework, from core engineering improvements to empowering the global AI developer community.

This year Matt White, Executive Director of the PyTorch Foundation, will discuss the shift to fully permissive open models, the role of frameworks and licenses like OpenMDW in assessing AI model usability, and the importance of open standards for building responsible, agentic, and autonomous AI.

At Scale By The Bay 2023, Meta demonstrated this vision with two standout talks focused on their work with PyTorch, a leading deep learning framework originally developed at the company.

Nikita Shulga, an open-source maintainer of PyTorch, shared how developers can improve PyTorch from the inside—reporting issues, contributing code, and understanding the lifecycle of pull requests.

Supriya Rao, engineering manager on the PyTorch team, presented the latest performance upgrades in PyTorch 2.1—including quantization, pruning, and memory-efficient attention. She showed how these techniques can boost inference speeds by up to 8.5× on real-world generative AI models like Segment Anything and LLaMA 2.

What is Meta Building Now?

While Meta is widely recognized for its social platforms, its current strategic direction is centered on building the metaverse, accelerating AI infrastructure, and developing next-generation computing platforms. At its core is a belief in open innovation and hardware-software co-design.

PyTorch, now a Linux Foundation project, remains a cornerstone of Meta’s AI strategy. By investing in PyTorch’s performance and usability, Meta is enabling researchers and developers worldwide to train and deploy cutting-edge AI models—from generative AI to vision and multimodal systems.

And it doesn’t stop at software. Meta’s AI teams are also involved in designing custom ASICs, like the MTIA v1, to speed up AI workloads. Their vision is ambitious: enabling AI at scale in a way that’s open, efficient, and accessible to all. Meta’s AI Research (FAIR) lab is one of the most respected in the world—and its code is often open by default.

Despite pivoting toward the metaverse, Meta continues to invest deeply in foundational AI technologies, particularly multimodal models and efficient training/inference systems.

RSVP