Open Source Summit NA 2025 | Matt White, The Linux Foundation

Clips
News
More from Open Source Summit NA 2025

Matt White

Executive Director of the PyTorch Foundation and GM of AI

The Linux Foundation

Code, community and the future: 13 takeaways from Open Source Summit NA

Beyond uniting developers under shared ideologies, goals and pain points, Open Source Summit North America ignited conversations at the intersection of code, community and change.TheCUBE goes live with insights at Open Source Summit NA

play_circle_outline Overview of PyTorch as an open source machine learning framework for deep neural networks.

play_circle_outline Envisioning the Future: PyTorch Framework Evolution and Its Dynamic Ecosystem in the Rapidly Advancing AI Landscape

Info
Transcript

Matt White, The Linux Foundation

Matt White

Executive Director of the PyTorch Foundation and GM of AI The Linux Foundation

In this interview from the Open Source Summit North America 2025, Matt White, executive director of the PyTorch Foundation and GM of AI at The Linux Foundation, joins theCUBE Research’s Paul Nashawaty to explore the evolving landscape of open source AI governance. White offers a detailed breakdown of how PyTorch has grown from a Meta-originated framework into a cornerstone of the Linux Foundation's AI ecosystem, now operating under an umbrella structure that includes projects like vLLM and DeepSpeed.

The conversation dives into the Open Model Definit... Read more

explore Keep Exploring

What is the role and evolution of the PyTorch framework and its managing foundation? add

What are the expectations for the evolution of the PyTorch framework in relation to industry innovations and community growth? add

bolt Powered by CUBE AI

Matt White, The Linux Foundation

search

Paul Nashawaty

>> Welcome back to Open Source Summit North America 2025. My name is Paul Nashawaty and I'm covering the app dev space in all things open source. I'm joined today by Matt. Matt, welcome.

Matt White

>> Hey, thank you very much. Thanks for having me. Appreciate it.

Paul Nashawaty

>> Yeah, great. Thanks for being on. Why don't you tell the audience a little bit about who you are and what you do?

Matt White

>> Yeah, so I am the executive director at the PyTorch Foundation and the GM of AI at the Linux Foundation. So in my role at the PyTorch Foundation, I manage the day-to-day and our strategy around our projects, programs that we're developing, and fundamentally community building around to the fast-moving space of AI.

Paul Nashawaty

>> It's certainly fast-moving for sure, yeah. No, PyTorch is really, really cool technology. It's really cool tech. For those that may not know it, I am not going to do it any justice, but other than saying it's an open source machine learning framework that helps with deep neural networks. Right. But maybe you can explain a little bit better than I did, because I've tried to do any justice.

Matt White

>> Sure. In brief that's what it does. It's sort of the de facto framework for architecting building training models and distributing those models more broadly. And so the PyTorch framework itself obviously emerged from Meta around 2017, came to the Linux Foundation and under the PyTorch Foundation. And we've been managing and running the foundation for the last two and a half years. We recently became an umbrella foundation, and so now our scope is broader than just the PyTorch framework. So we've welcomed some new projects, vLLM, DeepSpeed and more exciting projects to come.

Paul Nashawaty

>> Yeah, that's awesome. I can't wait to see what's happening. This is a very fast-moving landscape here, so it's pretty cool stuff. So let's talk about, just going to jump right in and talk about the fundamental differences between the AI models versus traditional software licensed models. Let's talk about the new approach, OMDW and what that means and why that's important from a licensing perspective.

Matt White

>> Sure. So around two years ago, we started work on this open model openness framework. And we were looking at from the open source vantage point, the differences between conventional software and open models effectively. And so models have a lot of components. We've identified 16 different components that fall in the categories of code documentation or data. And so we look at model weights as like high-dimensional data, model architectures are actually sort a blend of software and configuration. And so there's all these sorts of adjacent components as well that you don't find in conventional software. And so the work with the model openness framework evolved. We saw in practice that the community was facing some challenges with it in terms of adopting and classifying their models. And so what we did is we looked at the industry, what do we have out there for licenses that would be appropriate for models? And so we initially started off looking at per content models, so OSI approved for software. We looked at like CDLA for data and then CC BY 4.0 as maybe a recommended license for documentation. And so we evolved from there. And we ended up starting a journey around nine months ago on several iterations of a license that would be specifically tailored for open models. And so take the permissive approach and look at covering all artifacts that you'll find in a distribution of a model.

Paul Nashawaty

>> Yeah. And I think this is important to note. If I'm going to simplify it to kind of bring it down to the model openness framework really helps standardize and simplify the AI artifacts that come into it. So if you take those AI artifacts and be prior to having this model, you had to manage and catalog these things individually. But when you look at it, in addition to simplifying and standardizing, which is very important in itself, what are some of the benefits compared to the current fragmented landscape? If people are using their own bespoke solutions or their own way of doing things, why should they go in this direction versus doing their own thing?

Matt White

>> Yeah. So I think one of the most important things when you're innovating or building on models is to understand what you can use them for. And so when we look at permissive licensing, we know that we can use, study, modify, redistribute, and build upon models. And so we have that experience in open source software, but we haven't really built that experience with open models. And so the benefit of having something like OpenMDW as a single license is that you drop a single license in your repo. And even more important is the familiarity and the comfortability that that license will bring, because folks today are very comfortable with MIT, very comfortable with Apache-2.0, so comfortable that they reach for those when they're releasing models. And those licenses don't conventionally cover all the artifacts that you'll find within a model and its components. And so having a single license in your repo provides that benefit of, hey, when I go and download this model, I know exactly that I can use it for whatever purpose I need and I can build an application on it. I can safely and responsibly use it within my enterprise. And so there is that very nice label that you get with it that says, "Hey, this is open source and you can do what you'd love to."

Paul Nashawaty

>> Well, and I also see it as the devils are in the details. If you miss something, that can be a big miss, and that's a big challenge. So I think having a unified approach or a standardized approach makes a lot of sense. I want to talk about, and this is kind of a big topic that comes up with regards to AI artifacts and such, is copyright, patents, trademarks, infringement and all this. How do you protect this with that model?

Matt White

>> Right. So the copyright expectation it's two-fold, is if you're going to include data with your distribution of your model, you need to make sure that you've cleared that you have rights to that data. If you don't, then the expectation is you should not redistribute that data. Or if you do redistribute the data, make sure that the license stays intact. When we look at big pre-training data sets, folks don't release those generally. We know where to grab them using the same sources. But when you redistribute a model, you're probably going to more likely find reinforcement learning type data and preference data. Data that's smaller data sets. And a lot of that is human curated. Some of it's machine generated, but a lot of it is human curated. And so you want to make sure that if you've curated that data that you're releasing it and you are the source of that data. With other respects, the OpenMDW license does cover off all of the patent and other considerations, database and so forth, that will protect downstream users from getting into trouble.

Paul Nashawaty

>> Well, let's talk about those downstream users. When we think about the collaboration and innovation that plays into this, why is no copy left or attribution only approach or design of this way of doing is significant for those downstream developers or even enterprises? And how is it going to affect the collaboration or innovation?

Matt White

>> Sure. If anything, we've seen, even with permissive licenses that we continue to have collaboration and contributions. And so we look at like PyTorch, VLM, these other software projects, they are released in a permissive licenses, Apache-2.0 and MIT are sort of prepared today for a lot of models. And we see that people are taking these models. They are fine-tuning them, they're quantizing them, they're applying different optimizations. And in many cases, in fact, they release their work back to the public. They repost the model to Hugging Face. And then in some cases maybe they just take it in-house, and we want the folks to innovate. We want them to be at the liberty to be able to do what they want to do, and not everyone is going to contribute back. We see a lot of companies using PyTorch, and not everybody makes contributions to it, but by and large, most of the industry does.

Paul Nashawaty

>> Yeah. Well, let me ask you a different question. I think using it, I can just from our brief conversation here, I can see the benefits of it. Let's talk about pitfalls or potential consequences that could emerge from adopting a license like this. What would be some of the pitfalls?

Matt White

>> So I think the alternative today is Apache-2.0 and MIT, which we've already established don't fully outlay all of the aspects of models and the harms and the potential misuses. And so OpenMDW covers off disclaimers around you take on certain responsibility when you're using the model and you understand that it could be possibly misused. And we like to fall back to the court system to make a lot of the decisions around legalities. We're not in the law enforcement business. But the great thing is that we outline specifically model outputs. We say we don't impose any restrictions on model outputs. That is really something that we think is very permissive, look at model usage. And when you look at some community-based licenses, they impose certain restrictions or they don't explicitly call this out. And people are left to understand, hey, if the model produces code, is that code subject to copyright? And right now under a conventional open source software license, it would be.

Paul Nashawaty

>> Yeah, there's accountability. Organizations need to own it. And it's like just turning away going, "Well, I pulled it from somewhere or I used it." You're still accountable. You push it out the door, I think it's important. So let me ask this, we look at this, it's an evolution. Things are changing very rapidly. Here at Open Source Summit we see a lot of innovation occurring, a lot of new paths to do things, but also convergings of technologies to get to where you need to go for these organizations. A year from now when we're sitting here and we're talking about PyTorch and we're talking about the adoption, where do you see it and where do you think it's going to go from here?

Matt White

>> Sure. So if we're talking about PyTorch the framework, I think the framework's going to evolve with the industry. We are always incorporating the best innovations out of industry and out of research. And then look at the foundation, the foundation we expect it to continue to grow. Grow our membership, grow more projects, create a really rich and robust ecosystem of AI projects that will have fairly low friction, because we have the benefit of having this very unified community working together to make sure that everything is interoperable. And as there's innovations that are coming very quickly that we probably can't predict today that maybe a year or two down the road we didn't have the foresight to predict. And we expect we're very fast adopters, AI's moving quick. We have adopt fast, very quickly as well. And so we see those innovations coming in the form of new projects, or we see them as features in existing projects.

Paul Nashawaty

>> Yeah. Well, I definitely think you're spot on. Your successes so far have really grown and shown the adoption has taken off. Well, I know we're at the end of our session here. I want to thank you for your insights. This has been really, really great. I think the audience got a lot out of what we talked about here, but there's still a lot to learn.

Matt White

>> Sure.

Paul Nashawaty

>> Where would they go to learn more information?

Matt White

>> The PyTorch website, we have tutorials and other pieces of information, FAQs and so forth that people can kind of get started with. We have a getting started page there as well. Also, joining the Discord as well is super important to get engaged with the community. But moreover, we actually we'll be launching our PyTorch training and certification programs this October at the PyTorch conference. And so welcome folks to come out and spend the day learning about PyTorch, and we're reaching out to folks that are entry level and folks that are more experienced. And we'll hopefully have content that they can use and get certified on.

Paul Nashawaty

>> Very cool. Well, I appreciate your time today. This has been really great. I appreciate you watching our session today. My name is Paul Nashawaty. I'm coming to you live from the show floor at Open Source Summit North America 2025. Thank you for watching theCube, the leading source in tech news.