#213 Mark Surman: How Mozilla Is Shaping the Future of Open-Source AI

0:00The open source world, including Mozilla, came to say we need an alternative and we need an alternative that lets people configure things, change things, that was the spirit of open. And we still stand for those things today. We were founded as a nonprofit foundation because we believe that those open source assets need to be there as public assets. But we've grown. We continue to make Firefox. We've built a commercial business or a social enterprise around Firefox. But we've also moved into whole areas of advocacy, really advocating for privacy and trustworthy AI, and increasingly are moving into how we take the values that were behind Firefox and bring them into the AI era.

0:40AI might be the most important new computer technology ever. It's storming every industry and literally billions of dollars are being invested. So buckle up. The problem is that AI needs a lot of speed and processing power. So how do you compete without costs spiraling out of control? It's time to upgrade to the next generation of the cloud, Oracle Cloud Infrastructure, or OCI. OCI is a single platform for your infrastructure, database, application development, and AI needs. OCI has four to eight times the bandwidth of other clouds, offers one consistent price instead of variable regional pricing, And of course, nobody does data better than Oracle.

1:27So now you can train your AI models at twice the speed and less than half the cost of other clouds. If you want to do more and spend less, like Uber, 8x8, and Databricks Mosaic, take a free test drive of OCI at oracle.com slash IonAI. That's E-Y-E-O-N-A-I, all run together. oracle.com slash IonAI. That's oracle.com slash IonAI. Well, hi, Greg. I'm Mark Sermon. I am the president of Mozilla. And my focus at Mozilla really is how do we show up for the era of AI in the way that we did in the web era, which is really to make the internet, the digital world, something that is open and accessible to everybody.

2:20And I have spent my whole career, even though I only came to Mozilla part way through, and I'm not one of the founders, really kind of working in this weird intersection between how do you balance the interests of society and the interests of private players that are making tech or earlier on in my career, you know, any kind of communication. I actually started out, you know, as a punk rock kid working in public access TV. And my whole idea was like, in the era of commercial television, you need an alternative. And the web came along and like, oh, there's a place where everybody could have a voice.

2:56And so that kind of desire for everybody to be able to use technology or use communication for their own purposes really drove me into the early internet. And, you know, I've done a lot of different things. So I ran an activist internet service provider, which sounds like a funny thing, like providing web hosting and email accounts to people like Greenpeace and Amnesty International back in the early days of the internet. And then did a lot of open source work, you know, through the late 90s and eventually wound my way to Mozilla. Yeah. And for people who are not familiar with Mozilla, give the origin story and how the organization has developed.

3:42Well, Mozilla, people mostly know for Firefox. Excuse me. Mozilla mostly people know Firefox. I'll keep saying it in a way you aren't sneezing. No, that's okay. We're recording on different streams. Oh, are we? Okay. Mozilla people mostly know for Firefox, which is the open source web browser that launched in 2003, really at a time where Microsoft dominated everything. It didn't just dominate the browser market, but it did have 98 % market share. but the server room with Windows NT and it owned the database world. I mean, really, the whole stack was Microsoft. And the open source world, including Mozilla, came to say, we need an alternative and we need an alternative that lets people configure things, change things.

4:29That was the spirit of open. And we still stand for those things today. We were founded as a nonprofit foundation because we believe that those open source assets need to be there as public assets, but we've grown. We continue to make Firefox. We've built a commercial business or a social enterprise around Firefox, but we've also moved into whole areas of advocacy, really advocating for privacy and trustworthy AI, and increasingly are moving into how we take the values that were behind Firefox and bring them into the AI era. So we've set up an AI company called Mozilla AI. We'll talk more about what it's thinking about doing right now, really hoping that open source can win in the AI era, and an investment arm focused on investing in other responsible tech companies called Mozilla Ventures.

5:25And when you talk about moving, yeah, talk about Mozilla AI. I mean, how does that fit with the original mission? Well, if you think about the original mission and what's in the Mozilla manifesto which is worth going back and and reading and you know it is about this idea of the internet open and accessible to all it's about things like privacy security open source inclusion you know how that played out in firefox was we created an open source browser that then web developers could use to you know create stuff with javascript and all the whole web stack which really you couldn't do on internet explorer and that was really about opening things up for or anybody to try anything on the web.

6:07Like if you think back to late 90s, early 2000s, it's pre-Facebook, pre-Gmail, pre-Twitter, pre like lots of things we assume today, pre-Google Docs, all the stuff that we actually do everyday business in a web browser wasn't possible before. And when Mozilla was about, it was like opening the app to everybody, letting everybody build a business on top of, you know, on top of the internet on their own terms. And actually, there's a Harvard study that talks about open source has created almost$9 trillion in value in terms of it being raw material for people to build businesses on top of. And so we believe in the AI era, you need the same thing.

6:50That if you just rely on open AI, just rely on the kind of the standard closed stack, which is becoming dominant in their APIs early on, you're gonna really narrow the kind of terms of who gets to do what. And you're not gonna get that$9 trillion effect of it being something that we create a Lego block, a Lego box for everybody to kind of just riff with. You're not going to get a lot of the power of open source in the AI era. So, you know, that's why we're really focused on AI that is open source. There's a lot of the people working on it too. We want to work with them. And also, you know, the other piece is the AI needs to be trustworthy.

7:30It needs to look at what is privacy going to mean to us in this era. It needs to look at how do we deal with some of these issues like bias and discrimination. So we see tackling AI in many ways with a Mozilla value set as being the same task of having tackled the web. Whatever the fabric of our digital society is, whatever we're building things from today, needs to be built with public interest values as well as commercial values. And that's why we've expanded our focus. Yeah. But you're not building models. You're more of an advocacy group. Is that right? Yeah, so Mozilla AI and Mozilla Ventures and actually the company that makes Firefox are building the technology, not models per se.

8:14So I'll get to that specifically. But just to kind of give you some of the history and to break it out and so people understand concretely what we're doing. About five years ago, me and a few other people really came to the conclusion I just described. If we wanted to advance our values and put the kind of future digital life in people's hands, we needed to figure out how to play in AI and not just the web. And we started out with advocacy. So for five years, we've written a paper. People can go check out on Trustworthy AI. We've been pushing for both governments to step up and companies to be responsible in how they're developing AI.

8:54And I would say the first, you know, up until about 2022, we were mostly moving into the AI space as an advocacy voice. Over the last couple of years, we've set up this venture arm, as I talked about. So that is actually investing in companies that are building trustworthy AI. So you think about one company, there's a company called Flower AI, which is the leading player in what's called federated learning or federated training, meaning you can have AI run on your device in a way that nobody can see the data or the inference. It's a lot of the technology actually that's inside of the Apple intelligence play, which they're positioning as being about privacy.

9:32So through those kind of investments, we're playing in the tech stack part of it. And then, you know, what we've started to do in Firefox and other core products is very carefully play with how you do trustworthy AI. So, for example, in a beta version of Firefox today, you could go and say, choose to turn on chat GPT or hugging chat from an open source provider. So what we're doing in there is like letting people pick models from different providers, much like you would use those models inside of Edge or Chrome. But because we're very focused on independence, on choice, on open source, you can pick from four or five different models to do something like summarize a web page for you, as opposed to being locked into the model of a particular, you know, that the browser vendor is tied to.

10:27And then more at the core, and this is where we're earliest stage, closest to the models themselves, is Mozilla AI. Mozilla AI is our commercial open source R &D lab, which is focused on making it easier for developers to adopt open source AI instead of, say, going to an anthropic or an open AI. And in that, at this stage, we're not developing models. We're looking at partnering because other people are going at this. We'll get to talk about Llama in a minute probably, but certainly Meta is going at the question of open source models. So why duplicate that? Because we don't have the resources.

11:11Although there is an answer to why duplicate it. It isn't fully open. So then there are other people like Falcon, which comes out of a university in the UAE, also something called ULMA, which comes out of something called the Allen Institute in Seattle. There's other, there's a French lab like that, that are building end-to-end fully open models. Our belief is partnering with them and then making those models more robust and attractive to developers is gonna be the fastest path for us to help get open source alternatives, truly open source alternatives into the market. So we're playing across a whole breadth of investing in people who are building the tech, playing with it or rolling it out cautiously in Firefox, and then working very closely with people developing the models to bring them to developers and make them more performant.

12:03Yeah. On the federated learning, are you building a toolbox for people that want to do federated learning or do you have? Yeah. So Flower, who is that company, I mean, we're not doing it directly. We're a key investor and a partner of theirs, is building the standard by which, you know, hopefully everyone will do federated learning. So they're at the front of the standard setting process. And they've also built the tech so that any company, I mean, they have clients and, you know, who are big companies who need this stuff, who are relying on their toolkit to just deploy federated learning in the context of their AI projects.

12:47So, yeah, absolutely. And the thing is, there's so many pieces to this puzzle. We can build them all. So, you know, there's an example of a company we've invested who's building that toolkit. And now there might be Credo AI, which is a company in California focused on AI governance. I mean, so many people are, their employees are just picking up AI tools and using them, exposing their data to stuff, all of that. Credo is building a whole product line around doing AI governance. So as people adopt AI in your company, it's following whatever rules that you have around how your company's data is used or how you're complying with regulations that you need to comply with.

13:26So, you know, there's so many pieces of the puzzle. We're trying to pick the ones that we can work on directly. So some of that's going to be developer adoption of core models. And then also where we can help others tackling other pieces of the puzzle through things like investment. Yeah.

13:46On Mozilla beta or Firefox beta, you were saying that, you know, people can choose between models. Do you give access to the, I mean, are you guys giving access to GPT-40 Pro, or is it the free version? I mean, how do you balance that with proprietary models? So how it works right now, and as I say, it's just a beta. So that by definition, we're trying to learn is similar to how search works in Firefox today. So if you go into Firefox, if you haven't used it, go try it. And you type, you know, search for Craig, and, you know, something about Craig, and what you've written in the past, or what the podcast is about, it's going to drop down a menu and, and, you know, you could just hit return, but also gives you a button to say search in Wikipedia, search in Bing, search in Google, search in a couple other options.

14:48And so unlike any other browser, automatically by default in Firefox, we're giving you choice. We're giving you choice in where you search from. So fast forward to now, what are we trying to do with these AI models? When I use Edge, I'm tied to whatever version of OpenAI they've got built into there. If I'm using Chrome, I'm tied to Gemini. we're letting you say oh i want to summarize this article i could choose gpt4 but it is the free version because you know we're not but i think it's also whatever version you have so it's it's more that it integrates it into the workflow of the browser so we're just offering you the free one if you have the pro one you can use it there you can pick gemini you can also pick mistral which is an open source one and isn't available in any browser or hugging chat which is an open source one and isn't available in any browser.

15:40So what we're trying to do in that experiment is give people the choice, give people the choice between chatbots on a kind of per task level, but also expose people to open source options in addition to commercial options. You know, open source versus closed, because some of those open source options are commercial. Yeah, that's fascinating because increasingly systems or, or even in individuals, you know, I, I use all these models, but I don't rely on one. I go back and forth and I have, you know, the output from one I give to another, have a critique it and go, you know, do that. Would, would you be able to do that in Mozilla?

16:28Uh, so that that's exactly the direction we're headed. So trying to not only, obviously you can just do that yourself and cut and paste, try to build that fluid ability to move between models in an integrated way into the browser and again it's early days but the thing that you're talking about your use case i think is a perfect use case for the kind of thing that you know something like firefox which isn't tied to a particular model and a particular business strategy um you know you have uh you'd be able to do that in firefox yeah that's fascinating and and with the advent of agents presumably you could you could build an agent that does that for the user that that will will you know trigger a debate between models and come up with a consensus answer is is that the kind of thing that that you're looking at Yeah, I would say, you know, not at all ready for prime time, but maybe a sneak peek on the podcast.

17:34There is a set of R &D projects. So when we talk about kind of having deeper R &D stuff, exactly looking at that. So we're working on an agent framework called Pulsar, which is looking at what would it look like in the context of the web to have a web native agent framework and let you do exactly that. So it's still exploratory. There's a handful of engineers working on it, but that's exactly where we see the future going. Yeah, that's fascinating. Well, I have to confess, I haven't used Firefox for a while, but I'm going to now. is uh is this beta available uh to yeah if you go if you go into i can't remember if you have to install the beta version i think you might still right now so you go and you install something called firefox nightly you can anybody can get it it's just out there uh and then you go into a tab that says firefox labs and firefox labs is just a way for you to turn it on and off different experimental features i think firefox labs is also available in the main release branch and so So I'm not sure if this experiment has been pushed into the mainstream experiments bucket or is in the early, early experiments bucket.

18:45So, yeah, maybe the first step, if you were looking, is see if it's in Firefox Labs in your mainstream version of Firefox. And if it's not, go and install the nightly version. yeah and uh you were talking about uh working with uh with companies on on uh on open source models uh and you mentioned llama are are you guys primarily focused on on llama working with llama or are you no not not at all so i'll say you know off the top we are huge fans of llama and llama despite what Mark Zuckerberg says is not open source in our opinion well in that in that the training data is is closed there's a couple of reasons that it's not so um you know if we kind of sit on this topic for a second then I'll get to your original question um there's something called the open source definition which has always defined or says since the late 90s has defined what counts as an open source license and there's you know four basic criteria that you can use it for free, that you can study it or that it's transparent, you can look under the hood, that you can modify it to be something else for your own purposes, and then that you can share that modification.

20:02And that over the last 20 years, that definition has stood the test of time. And frankly, that definition has given people the confidence to build businesses on top of open source, because you know, it's always going to work in these predictable ways. Nobody's going to pull the rug out from under you. You're always going to be able to modify it for how you want. So like if you want to build an industry and an economy on open source, you need those four rules. The Open Source Institute, which, you know, is the group that kind of sets this rule set up, has gone and done a new definition of what is open source AI mean?

20:38What's the open source definition for AI? It's not quite the same as software. It includes not only the software to make an AI model, but the models themselves and the data. And so, you know, basically they've said, here's how you interpret those four principles in this era. And LAMA does not at all fit into the mix. One, because, yes, the data is totally opaque. You know, you don't have to, you can't make the data fully open source in most cases because there are copyright questions or privacy questions or whatever. but you can't even see what data was used to pre-train LAMA and that in our view but also in the definition by the open source initiative is essential but the other thing they have in the in the LAMA license which is very odd um and and really just is by any definition of open source doesn't you know doesn't make it count as open source is it says in the LAMA license if you get to 700 million users on your product, it's not open anymore.

21:38And so imagine that. Imagine I built Azure or Gmail or like some famous product or frankly, Facebook on top of Linux. And as soon as it got to a certain user threshold or it's no longer open and Facebook is built on Linux. So imagine if that happened to Mark Zuckerberg, some unpredictable unknown cost, some rug pulling would happen once your business is successful. you're not going to build on open source if that's the rule set. And that is the rule set for Llama. So, I mean, we are huge fans. We actually produced a piece of open source software called Llama File that makes it easy to run Llama on a local machine, even a high-powered laptop.

22:19So, like, we think it's a real contribution and it's a big counterpoint to open AI. So we're really happy it happens. We wish they would make it fully open source, but they haven't. And so it is where we work with a lot of these other labs, like the Allen Institute, which are doing stuff that is fully, fully open from the data set all the way up to the model parameters. And often those are nonprofit labs, just like the Linux Foundation makes Linux or Mozilla Foundation makes Firefox. talks sometimes you want these core assets even if they're very commercially valuable and used by industry to be held by some steward who is not you know trying to back the the interest of a particular company yeah and the Allen Institute is um being transparent about the training data yeah they've got a fully open training data set called dolma that feeds into Olmo the model um so can totally see what that is you know you could use it to go and train the model differently because the software is also open um you know they're still on the heels of of somebody like llama but but actually you know their most recent smaller models actually have a lot of performance gains and and are really useful in cases where you've got a particular use case in mind that you don't need the most powerful model but you need full transparency and we think that's going to be a pretty meaningful market.

23:53And ultimately, that'll grow into something that can be a real competitor to both the closed models and the semi-open models, let's call them. Yeah. And what about Mistral? Is that a truly open source? You know, Mistral is, you know, they've changed it since they came out, so I'm not fully tracking it. They have some stuff that is more open source and then some stuff that they're really kind of keeping for themselves we haven't kind of talked with them and we're not the open source initiative so you know the open source initiative as they develop this um open source ai definition will kind of test all of the different licenses against things so i don't know if they've looked at mistrawl but you know definitely mr all is uh is a mix of um of you know open source and and closed approaches we really like them i think they're also an important player they're one of the players who is plugged into this Firefox experiment, you can get the Mistral chatbot there.

24:55Yeah. The larger question, I mean, regardless of whether Lama is truly open source or not, at what point do you think meta is going to run out of uh runway on on its spending to keep up with uh with open ai in particular i don't think i don't think they will i mean they're not talking this way and maybe they don't even think this way i think the the bigger threat is that you know some public ai models would would come up and compete with llama and replace it as the industry standard i mean their bat which i i totally believe it's their bet i just think it only works if they were fully open and they're not is to be the industry standard in you know what is the the kind of open ai that open source ai that people build on uh in the way that linux is the industry standard for what people build a you know operating system that they run their servers on so you know why does meta why does Amazon, why does Google, why does IBM contribute to Linux, which they all do?

26:13Why is Linux able to run the whole internet based on contributions from companies? It's because companies put resources in because it lowers their own operating costs and they don't compete in the operating system market. So Zuckerberg is playing the same game with this. I'm not going to compete in the chatbot market primarily. I'm going to add it as a feature to WhatsApp, to Facebook, to whatever. And so what I want to do is define the industry standard and lower my costs by having other people. I mean, I saw him in an interview with Jensen saying, oh yeah, NVIDIA is putting 200 engineers into Llama.

26:50That's the math of it, right? Is if everybody starts to rely on Llama, he's actually increasing the robustness and lowering the operating costs of Facebook. It's brilliant. And again, I think it would be a great, it is already a great contribution for it really to be the industry standard I think they need to be fully open otherwise if I'm a government or another big business do I really want to build on top of Lama that at some point I might pay some undetermined amount for in the future or that's owned by Facebook and not some independent third party that might change their mind in the future, ultimately I think people are going to start not choosing Lama for that reason, especially if more performant truly open models emerge yeah um although you know the the innovations continue to come out of open ai and for example with uh o1 or strawberry yeah do you expect that meta will keep up with that in its llama series or something i think the world will keep up i mean there was an article i don't know if it was in the financial times or the information today about Chinese models approaching the level of where open AI is at with their innovation.

28:09I think you're going to see a race on this stuff, but I don't think it's going to be a definitive race where somebody gets the end and then everybody falls over. You know, everybody else falls over. I think the nature of intellectual progress in science is people see what's going on. People are trying to catch up and innovate around each other and do different stuff. You know, the big two innovations of open AI were they took all this open science and like put it in a black box and hit it, right? They took the transformer paper and they started working on it for a bunch of years and they stopped contributing, you know, participating in the open research community.

28:47Other people continue to innovate outside of them. And Sam Altman is a brilliant product guy, right? I mean, he's been a part of the whole sort of product thinking of Silicon Valley through Y Combinator and so many other things. So you have them, they've locked up their innovation engine and not shared, and they've done great product work on top of it. But the rest of the world, or not the rest of the world, of course, but many people are still in the open innovation, open science space and AI. Collectively, I ultimately believe, maybe I'm a Pollyanna, but history bears this out. Collectively, those people will innovate as fast and actually eventually out-innovate open AI.

29:30yeah that's a that's the debate though yeah yeah so i know what side of the debate i'm on i also know what where i'm placing my bets uh i mean you know lots of other people are betting in the other direction and history will tell but i think we're early in this right if you think about you know the cycles of of history we're talking about five ten twenty years before we really know the answer to this question yeah and certainly linux uh or unix uh thrived and uh and uh an entire uh ecosystem developed around that um exactly and i think look look at who are the companies who are left in the unix business right their red hat and canonical and those how many people are still running sun you know, Sun OS or Sun, you know, Sun hardware doesn't exist anymore, right?

30:27And, and how many people are running Windows NT, which, you know, was Microsoft still trying to compete against Unix? It was actually the Unix, which was Linux that everybody used because everybody contributed to it that became the dominant standard. And so I think we'll see, is there a Firefox of the AI era? is there a linux of the ai era um you know what what is that thing and and does it win out yeah uh and um on uh well the difference again is that the cost of training of these models is so enormous that uh and that you need someone of the size of meta to to to develop the models for it's a really critical point and a critical question and you know it's one where i think we could really just uh you know miss a pitch uh in you know in the sense that the ball is being thrown as we've got to swing and we got to hit it in that uh yes it's expensive to train these things i mean the economics of that may change over time and there's lots of debate about that but let's say it's still very expensive you've got people like meta who are doing you know effectively a kind of uh a public good in many ways by putting llama out there and absorbing the cost and also trying to control the industry and in the process but that's good and then you have all these governments around the world who are saying oh compute is a problem let's you know put national research dollars towards that and that could go one of two ways that's where i talked about we could miss a pitch you could say all of those national research dollars for compute go towards open source and that you know ideally you get a small set of open source projects that are all leveraging that subsidized compute for stuff that's produced in the public interest we would call it public ai and that that actually adds up to something that's a real meaningful player for everybody but i think there are more than enough government dollars around the world already going into compute, that this kind of public AI lane could play a role.

32:46And if you look at this in really long timescales, like other big social, technological, economic transformations, Henry Ford doesn't create the national road network or the regulatory environment for automobile safety. right it is private players play a role but you know in terms of the bigger infrastructure that connects everybody that you know the public sector also plays a role um nonprofits play a role and so on so i'm increasingly thinking we i'm increasingly thinking we need this conversation going about what's actually the public side of this story that enables um you know all of the the commercial innovation to be both successful, but also, you know, democratic and spread broadly.

33:36Yeah. The, the, you mentioned the, the, you know, the, your preference would be to restrict the, in the U S is the national AI resource. There's our two R's resource. Yeah. Yeah. Yeah. Repository or whatever. Yeah. Nair. Are there restrictions on that that you have to? It's leaning in that direction. I mean, I think it's still early days where it's not so much restrictions, but like requirements, like I think is that you would have to make what you're producing open source. I'm not sure how robust that is and what extent it is. So, but certainly our encouragement of where it should go is that anything produced on AI resources, on NER resources or with NER funds is open source.

34:26And, you know, you have a precedent for that. The National Institutes for Health, which is a big, you know, government funded health research pool, requires that anything funded with NIH money be released as what's called open research, which means that you have to publish the findings of your scientific study on health topics in a way that anybody can access and not just in some private journal that if you don't have a big, expensive subscription, you can access. And so you do have precedents where you tie public dollars to open publishing. And I think that's in AI a principle that we need to apply.