요약
Mozilla’s Mark Surman explains why most AI models aren’t truly open source. The post Mozilla President Mark Surman on what “open-source AI” really means, and why it’s still evolving appeared first on MEDIANAMA.
본문
As debates around artificial intelligence intensify, one of the most contested ideas in the ecosystem is “open-source AI.” While companies increasingly label their models as open, questions remain around transparency, access to training data, and whether these systems truly meet the traditional definition of open source.
In this conversation, Nikhil Pahwa, Editor of MediaNama, speaks with Mark Surman, President of Mozilla, on the sidelines of the India AI Impact Summit. Surman breaks down what constitutes open-source AI, the limitations of current “open-weight” models, the economics of open ecosystems, and how Mozilla is thinking about building an open AI stack through developer tools and browser integrations.
The discussion also touches on tensions within the open-source community around licensing, corporate use of open models, and whether browsers like Firefox could become gateways to an open AI ecosystem.
You can watch the full interview here
This section of the interview has been lightly edited for clarity and brevity. Other parts of this interview will be published soon.
Read Part 1 of the interview: [link]
Read Part 2 of the interview: [link]
Read Part 3 of the interview: [link]
What is Open-Source AI?
Nikhil Pahwa: Let’s talk about open source. What is open-source AI, and what is not open-source AI?
Mark Surman: What is open-source AI is a hotly debated question, although, in many ways, not many people are actually in that hot debate. And so, to kind of bring it into the nerdiness of the people who do talk about this:
Open source at its core is about four things. It’s that I can use it freely. It’s that I can study it, that it’s transparent, that I can look inside how it works. I can modify it, I can change it for myself, I can build something new from it. And then I can redistribute that, I can sell it or I can give it away to somebody else.
That’s the essence of what open source is. And that’s how Firefox and Linux and so many things have grown, right? Because I can do those things.
So many people use it, so many people build on it and make it better, so many people spread it again, and it takes over the whole world.
So ultimately, open-source AI is the things that have those properties. And ultimately, I think they can take over the whole world. I think open source can be the winning paradigm if we get it right in AI.
And so, many of those properties are in the open source, or the things that get called open source but are really open-weight models, whether that is something like LLaMA, or something like Qwen, something like Mistral.
And then more importantly, or I think the thing that people miss, is many of the things that matter in the rest of the open-source AI stack: the things we use to train models, the things we use to pull data, the things we use to orchestrate, are open source in the traditional sense.
So what I would say is there’s a lot of open-source activity in AI, to the point where I could build, in fact, we have built, many people have built, their own ChatGPT-like things in mostly open-source ways.
And so, three years ago, you couldn’t have imagined that. Now, there are thousands of them.
The one limitation, or the one thing that really is a sticking point, is that most of the things that get called open-source models, like LLaMA or Qwen, the underlying training process, what the data is, how it was trained, the pre-training, is opaque. And that is not open source in any traditional sense, in that I can’t really look inside that part.
So those things, I think they’re still incredibly valuable. We use them, we build on them. They’re better than something that’s just locked behind an API that I can’t modify on top of.
But I think there is a real question: can we get to the point of highly performant, what people often call truly open models, where I see what the training data is, what the pre-training checkpoints were, and can I modify them and fork them at that level?
And I think, in the current architecture of LLMs as one type of AI, that is a harder thing because there’s so much money that goes into that process.
And what you’re seeing more is open source in relation to models being open source in terms of the ability to take them freely, modify them dramatically, and then be able to choose between and orchestrate multiple models.
So, just to take a case: if I pull together something on top of LLaMA or Qwen, and then I do fine-tuning or transform it in some way, or combine it with other models and create my own new small model and put it under a very good open-source license, some kind of Apache or MIT license, is that thing open source or not?
Not totally, truly, because the thing underneath is opaque. But I’ve changed it so much and made it something new, and taken these different pieces of raw material, I can say how I did that. Maybe it is.
And so I think, in that part of the creation and transformation of models, it’s almost like the materiality of that part of the stack, we have to wrestle with what we mean by open source.
And I think, frankly, my instinct is it’s going in the right direction, but we’re still babies in this, and we don’t quite know what we mean.
Licensing, Commons, and Corporate Use
Nikhil Pahwa: So one fear that I’ve seen from the open-source community in India that works on AI is around licensing, saying that if they truly build a completely open AI model, then all of that information, the weights, the data, will just be taken by proprietary large language models.
So they’re struggling with figuring out how to create a license for this.
Mark Surman: That doesn’t seem like a licensing question as much as a ‘do you believe in open source or not’ question.
Nikhil Pahwa: I think they want to be useful, but they don’t want to empower the proprietary system at the same time.
Mark Surman: But open source has always done that. I mean, the way that open source works is that anybody can use it, anybody can modify it, anybody can share it, right?
And so that becomes the question of how do you balance having a commons versus corporate benefit, or arguments about corporate capture.
So you can have a debate, but I don’t think it’s any different in AI, or whether it’s a good thing or a bad thing that the whole tech industry and most governments run on the Linux ecosystem, run on the Linux stack.
I think that’s good. We have a common infrastructure for humanity that individuals, companies, and governments all contribute to in different ways.
It’s highly leveraged when you think about the overall open-source ecosystem. The chief economist of the Linux Foundation has shown, Frank Nagle, that there’s only about $4 billion to create most of the open-source stack that we have now, and about $8 trillion in value that is unlocked from that.
So you have this highly leveraged commons, and yes, of course, companies benefit from it, but they are also the biggest contributors back to it.
And if that is how the AI infrastructure of the world works, that it’s a commons that, yes, companies use, but I can use too, and that they contribute to it as well, that’s kind of a better version of the economics of a thing than we have in most other parts of society.
I’m happy to make that trade.
Mozilla’s Approach to Open AI Ecosystems
Nikhil Pahwa: Just on openness, I’m curious, because we’ve seen a few things happen in the last many years. For example, the advent of F-Droid as an open-source app store, the advent of Hugging Face as a source for open-source models.
Has Mozilla considered either integrating, working? Is there an easy or quick way that I can implement a small language model within the browser itself?
Have you had those conversations? Have you thought of going down that path of becoming the open space where the open AI ecosystem sort of congregates, and the app ecosystem congregates? Because not everyone knows about F-Droid, but the browser is a great way for discovery of that integration.
Mark Surman: That’s a good question. It’s like, what’s the role of the browser, and what’s the role of Mozilla? And I would separate those in creating, basically, optionality and a way for there to be an open ecosystem of people pulling in different things, different agents, different models, different digital objects.
And we absolutely thought of that.
I will say, the first step in that is Mozilla AI, the separate company we spun out for developers. Its focus is on exactly that kind of choice.
So it’s putting out libraries like “any LLM” or “any agent,” which are basically developer tools oriented toward building dynamic choice into the architecture of what you do.
So it means that I can test against five different models. I can build an actual production-grade system that dynamically switches between models based on different parameters, including cost.
And if we start to build that as an architecture of choice, and start to build that out as standards in how developers work, then you’re going to start to have some of the substrate of what you’re talking about, which is choice. Choice in how AI works in the underlying architecture.
And we need that first, and we need that to emerge in the market amongst developers before it can really be practical.
And so that’s the first step we’re doing, and we’re trying to build a public-benefit business around that in the developer space.
How does that then show up for consumers, which the browser is the obvious vector for that? I think actually the email client might be another place for that.
I think we’re talking about it. The path is not yet quite clear. The values are clear.
So as we build out AI mode in Firefox, which we haven’t talked about yet, it being a place where I have agency to choose different services, combine things, I have choice, I can use different models, I can have my own controlled memory, all of those things are a part of what we’re talking about there.
Does it actually look like a Hugging Face or an F-Droid, or connect you to those things? I think we’re too early to know exactly how those kinds of things would work out.
Read More:
- Mozilla President Mark Surman on AI agents, the future of browsers, and a new browser war
- Mozilla President Mark Surman on AI, the future of the open web, and collapsing publisher models
- Mozilla President Mark Surman on AI Disrupting Search and the Future of Browser Revenue
For You
- Read Reasoned by Nikhil Pahwa: How AI is changing our world
- Sign up for MediaNama's Daily Newsletter to receive regular updates
- Sponsor a MediaNama Event