Joe Spisak

Joe Spisak is vice president of product and head of open source at Reflection AI. He previously led product efforts for PyTorch and the Llama open-source initiative at Meta, helping guide PyTorch’s transition to the Linux Foundation in collaboration with industry partners. His work connects AI technology with the arrangements that let developers build on it: hardware support, usable interfaces, distribution, and shared governance.

From hardware to shared AI infrastructure

Joe’s early career included a decade in semiconductor design, strategy, and product management. He led machine-learning strategy for Intel Cloud, then became an early member of Amazon AI, leading a product team and global AI partnerships. His subsequent work at Meta included product leadership for PyTorch and AI for Science within FAIR. He also held a leadership role at Google.

At Meta, Joe’s responsibilities extended beyond the company’s own use of AI. He helped guide PyTorch’s transition to the Linux Foundation with industry collaborators, giving the framework an organizational home beyond its originating company. He also led the Llama open-source initiative, extending his work from software for building models to models developers could adapt and deploy. His later responsibilities in Meta’s Superintelligence Labs included product efforts across PyTorch, Llama, and Meta’s agentic platform.

Joe and Luca Antiga’s account of PyTorch’s development explains how the framework’s priorities expanded. Its early emphasis was helping researchers express ideas and test them quickly. PyTorch 1.0 incorporated Caffe2 and expanded toward production deployment; PyTorch 2.0 put greater emphasis on compilation, performance, and support for different hardware. These were collaborative engineering achievements. Joe’s product perspective connects that evolution to the need for a diverse contributor community that can sustain the framework through technical change and commercial competition.

His coauthored PyTorch/XLA 2.0 article gives a concrete example of work across those company boundaries. The project began as a Google–Meta collaboration in 2018 to bring Cloud TPUs to PyTorch users, then expanded to include Amazon and other contributors. Supporting another accelerator required a usable software path between the framework and the hardware, alongside cooperation among the organizations developing it.

What developers need from an open stack

Joe’s writing and projects address several practical conditions for an open AI ecosystem:

• Hardware choice with a familiar development experience. His priorities for 2026 connect heterogeneous hardware support with communication, training, inference, and deployment software. Access to another accelerator becomes more useful when developers can adopt it without giving up the framework and workflows they already know.
• Reusable interfaces for reinforcement learning. As an OpenEnv maintainer and coauthor of its guide to training agents on real-world tasks, Joe works on a common interface between agents and task environments. This lets different tasks share integration machinery while retaining their own actions, observations, and scoring.
• Ownership with maintenance costs. In his enterprise argument for open models, Joe argues that businesses should consider control over their accumulated knowledge when choosing models and infrastructure. He also acknowledges the staffing required to maintain an open stack. Independence carries an operating responsibility.

OpenEnv makes the interface work tangible. An environment defines the actions an agent can take, the information it can observe, and how completion is judged. A client starts an episode, submits actions, and receives observations and rewards through a consistent interface. A word game and a calendar task can therefore use the same integration pattern while keeping different rules and measures of success. The interface reduces repeated integration work; environment authors still have to decide what behavior deserves a reward.

Joe’s April 2026 keynote, Community Led Open Source RL, continued this focus on collaboration in reinforcement learning.

Building around open models at Reflection

At Reflection, Joe leads product and open-source work at a company developing open-weight models and the software around them. Reflection’s first model, Beam, targets coding, reasoning, and agentic work, with availability through open weights, an API, distribution partners, and developer-tool integrations. The company also develops deployment software, customization recipes, and agent harnesses.

That scope extends the concerns running through Joe’s PyTorch and Llama work: making a model available is one step; developers also need ways to run it, adapt it, evaluate it, and connect it to their existing systems.