The Linux Foundation in the Age of AI

AMSTERDAM — This week, the New Stack Agents is all about open source AI. Jim Zemlin, the executive director of the Linux Foundation, joined us on the showfloor of the Open Source Summit in Amsterdam to talk about the state of open source AI, the role of the Linux Foundation in keeping AI models and tooling open, emerging standards and more.

One AI Foundation To Rule Them All?

With the Cloud Native Computing Foundation (CNCF), the Linux Foundation built a vendor-neutral home for the cloud-native ecosystem. The CNCF started with the donation of Kubernetes, but now hosts over 200 projects. But in the world of open source AI, there’s no central foundation yet. The Linux Foundation hosts the PyTorch Foundation, LF AI & Data, and numerous stand-alone projects and standards, but there is no CNCF-like umbrella foundation.

“The AI stack has been emerging around a few different vectors,” Zemlin told me when I asked him about this. Early on, he noted, there was a lot of interest in building foundation models and projects like PyTorch and TensorFlow. Then, more recently, the focus shifted to open-weight models from the likes of DeepSeek, Mistral, and others, with today’s focus moving to the inference stack, agents, and open protocols.

“I do think that we’ll see the emergence of a more collective effort around agentic AI, but I think it’ll take a little bit of time for everybody to congeal around what’s the approach to that stack? What are the different components? Whether it’s in agent-to-agent protocols, or things like MCP — and that’s clearly underway,” Zemlin said.

And while he said that it would be convenient to have a single open source AI foundation, he also noted that “it’s really hard to predetermine this stuff as the technology is emerging” and that his number one rule for when an open source community is emerging is “don’t screw it up.”

He also noted that a foundation has to ensure that it doesn’t get in the way of innovation by becoming too big. To make Kubernetes ubiquitous and move the world from virtual machines to containers was a massive undertaking that involved rebuilding the entire toolchain to support container-based applications.

“I think balancing big versus small agility is an important thing. And so what we try to do is be a little bit hands-off and let things grow. But eventually, if there is a utility in having a larger collective working on moving the entire ecosystem towards a perspective on agent technology that doesn’t get in the way of organic innovation, we want to support that. And those are things that we are kind of balancing right now,” Zemlin said.

Open Source Models

As for long-standing questions about the definition of open source models (and whether open-weight models are open source) and whether those models themselves have a place inside foundations, Zemlin noted the incredible amount of resources (and money) it takes to build frontier models. His approach here is mostly driven by pragmatism.

“I think that the open community would love to see all the way back to open data — everything being completely open. And who wouldn’t want that? But I think pragmatically speaking, given capital expenses, the access to data, you don’t want to look at the open weights models and say ‘well, you’re not pure enough’ and look a gifted horse in the mouth, to some extent,” Zemlin said.  “So I do think that the open-weight model trend is important and good. I do think we will eventually see end-to-end open models, open data all the way through.”

He noted that the Linux Foundation created the Model Openness Framework, which allows for a nuanced view of how open a model is, but also acknowledged that for most developers, that’s an academic discussion.

For the entire discussion, which also includes segments about China’s investments into open source software, European regulations like the Cyber Resilience Act and its effect on open source projects, as well as Zemlin’s own AI usage, subscribe to our podcast or head to The New Stack Agents Live.