Everyday AI Podcast – An AI and ChatGPT Podcast - EP 396: How Public AI Can Shape a Safer, Smarter Future for All

Episode Date: November 6, 2024

Ya know how in big cities there’s a mix of cars and buses? It’s a symbiotic symphony of private vehicles and public transportation sharing the road. Can AI be like that? Nik Marda, the Technical Lead, AI Governance at Mozilla joins us to help us build that path. Newsletter: Sign up for our free daily newsletterMore on this Episode: Episode PageJoin the discussion: Ask Jordan and Nik questions on AI governanceUpcoming Episodes: Check out the upcoming Everyday AI Livestream lineupWebsite: YourEverydayAI.comEmail The Show: [e-mail] with Jordan on LinkedInTopics Covered in This Episode:1. Public AI Explained2. Public AI Implementations and Limitations3. Trustworthiness and Concerns in AI Models.4. AI Integration in Daily Life5. Funding and Investment Needs for Public AITimestamps:00:00 Public AI promises a safer, smarter future.04:00 AI needs balance: commercial and public infrastructure.07:41 Public-private AI ecosystems can symbiotically coexist.12:34 Prioritize AI safety with smaller, quality datasets.15:38 Public AI models, data, and infrastructure needed.19:36 Public AI prioritizes trust, safety, accessibility, accountability.23:51 Public AI approaches offer control and cost benefits.27:56 Diversify AI sources for improved organizational resiliency.30:12 Public AI offers open-source, mission-focused alternatives.Keywords:Public AI, Private AI, AI application development, Llama, OpenAI, Google, investment in AI, public AI initiatives, Olmo, Falcon 40b, AlphaFold 2, Mozilla's Common Voice, National AI Research Resource, FAST, India AI, AI Factories, AI vendors, AI systems, AI Trust, AI Accountability, AI Governance, Mozilla, commercial AI infrastructure, sensitive data, public-private structures, closed proprietary AI models, open-source AI models, AI integration, incentive structure, proprietary AI systems.Send Everyday AI and Jordan a text message. (We can't reply back unless you leave contact info) Start Here ▶️Not sure where to start when it comes to AI? Start with our Start Here Series. You can listen to the first drop -- Episode 691 -- or get free access to our Inner Cricle community and all episodes: StartHereSeries.com Also, here's a link to the entire series on a Spotify playlist. 

This is the Everyday AI Show, the Everyday Podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life. Meet Firefly AI Assistant, now live and Adobe Firefly, the All In One Creative AI Studio. Just describe what you want to create and the assistant handles the rest, orchestrating multi-step workflows across Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome. The assistant accelerates execution. Do you know where your AI's been, right?

If you're just using a large language model at your company to grow your department, and it's proprietary, a closed model, which is what so many of us use, do you have trust there? Do you feel confident in kind of where this company got their data, what they're doing? with it? I don't know. Even if you're maybe using open source or open weights models, I think there's still some accountability and some trustworthiness that needs to be improved there. So that's why I'm excited today to talk about how public AI. Yeah, another kind of AI, just when you thought you couldn't keep up, how public AI can actually shape a safer and smarter future for us all. So I am excited for today's conversation. I hope you are too.

Welcome to Everyday AI. What's going on, y'all? I'm Jordan, and this is Everyday AI, and this is for you. This is your daily live stream podcast and free daily newsletter, helping everyday people not just learn about generative AI, but how we can all actually leverage it to grow our companies and to grow our careers. If that sounds like something you want to do, something you're trying to do, bet like this is the spot to do it.

If you haven't already, go to Your EverydayAI.com, sign it for the free daily newsletter because in addition to recapping all of the valuable insights from today's show, we put out a newsletter every single day, well, Monday through Friday, helping you keep up with everything else that matters in the world of artificial intelligence. So if you're joining on the podcast, we always leave that in the show notes. You can just go click that. If you're joining live, thank you for joining live. Today is technically pre-recorded.

So if you want your AI news, yeah, generally I rattle that off before the show. but we are debuting this live. But to go get the latest news, make sure to go check out our newsletter. All right, enough chit-chat. I am excited for today's show, something that I think I'm going to learn a lot about, and I hope you are too.

So please help me welcome to the show. There we go. We have him, Nick Marta, the technical lead for AI and AI governments at Mozilla. Nick, thank you so much for joining the Everyday AI show. Jordan, thanks for having me. I'm excited to be here. All right. So let's just talk real quick. What's your role? Technical lead AI governance at Mozilla. What do you do? I think everyone knows Mozilla from Firefox and some other initiatives, but what do you do in your role?

Yeah. So Mozilla broadly is trying to make the internet a more open and public place. And more and more of these days, that means thinking about how AI is shaping the internet and making sure AI is open and safe and a public resource for everyone. And so I'm focused on thinking about both how do we help it make it easier for developers to build AI more responsibly? So how do we do technical work and create technical tools that helps folks build AI safely? And then the flip side of that as well, how do we advance the policy and governance conversation to promote the broader societal structures we need to also promote trust-worthy AI? So thinking about both the technical and governance sides of AI governance. So I kind of alluded to it there in the intro, Nick, but why is this needed, right? Why do we need public AI, right? There's so many AI choices, right?

It's like there's new large language models or LLM updates every day. Why do we need something public? So most of the AI folks interact with, build with, use in their everyday lives are built on top of commercial AI infrastructures. And many times that works just great. It solves a lot of problems for folks. But there are many problems that folks want to or could use AI for that they can't, or there are many places where you want to be able to build with AI in a particular way. For example, you're trying to use AI on some proprietary data or sensitive data that you can't put in the cloud. Well, you're going to want to do that locally on your computer, and then it starts to get

more challenging to use a lot of these commercial solutions, where you actually probably want full access to the model and its weights and its code, and you want to be able to develop that locally. So there are places like that where there are gaps in the commercial ecosystem for developers and consumers, but then there's also this broader structural and societal conversation we need to have about who's building AI, where is the power in this very powerful ecosystem, And what is the infrastructure we're creating for the future of AI? Is it one where it's geared toward just solely commercial interests, some of which are good, some of which are bad? Or do we also want to have alternatives where you're building on top of public infrastructure as well?

And I think the easiest way to get at this is just through a couple quick analogies. So one is for transportation, right? You step outside, you look, if you're in a city anyway, and you look on the road and you'll see both cars and buses. right? Or you turn on the TV and you flip the channels and you'll see, at least in the U.S., you'll see both PBS and NBC, right, public broadcasting and a private network. And so you see a lot of these places where we have infrastructure that supports both public and private use cases. And this is the sort of thing we've seen time and time again with innovation, where often the private sector is pushing the frontiers of capabilities, but it's the public sector that's making those capabilities actually accessible and usable across a lot of the parts of the ecosystem where the commercial incentives just aren't there to actually promote the public interest use cases we care about. So help me understand that. I think that's a great analogy that you brought up there, right, like private cars and then public city buses can share the road. And there's

pros and cons for each and different use cases for each. How does that work with AI, right? Because doesn't ultimately, maybe a private company, have to be very involved in a public AI? Or is this something that, you know, governments or, you know, states, countries, is that how it should be? Like, so then they have that kind of not ownership, but in the same way that a bus is usually owned by a city, right? So who can own, you know, public AI to make it truly public? Yeah, so when we talk about public AI, we don't just mean governments.

Governments are, of course, a core part of public, but there are many other public institutions, right? So think about the early days of the Internet, and even what's come from that to today, right? You have the Mozilla Foundation backing Firefox. You have the Apache Foundation and the Linux Foundation backing those respective services, right? So you have these nonprofits that are creating a sustainable open source product that folks can build on top of. And you can imagine that sort of thing happening with AI as well. And we're seeing that happen in spots. I think the Allen Institute is a perfect example of this where it's a nonprofit backed development model for creating both AI models and AI training data that is part of this broader public AI.

ecosystem. But then on top of that, you can imagine commercial actors using bits and pieces of public AI artifacts to build their products and solutions. So I think if we do this right, the public and private AI ecosystems will be symbiotic with each other. They can build on top of each other. If you're a startup innovator trying to build your AI product, you could choose between more of a public AI stack or more of a private AI stack or the pieces that make the most sense for your particular application. These two things don't need to be contradictory, and they can actually work together,

and you can kind of have more of a modular approach and how you pick and choose how to use different parts of the ecosystem. So maybe could you help explain Nick a little bit more, right? Because I think many of the models that a lot of us have heard of are closed proprietary models, right? The GPT models from OpenAI, Gemini models from Google, Claude models from Anthropic, et cetera, right? But then there's also this other class of, you know,

either open source or open weights models, right? So as an example, Lama from meta. So what's the difference between those types of models and then true public models? Right. So the way we define public AI is we lay out three criteria for what public AI should be advancing. One of those is public goods,

that public AI should be creating open and accessible goods and shared resources at all levels of the AI stack. Another is public orientation that public AI should be centering the needs of communities and people, and public use that public AI should be prioritizing AI applications in the public interest. And I think if you look at Lama, for example, I think it's very clear that Lama has really enabled a lot of people to build AI applications and infrastructure that advances those three goals I just outlined. But Lama itself doesn't quite meet those three criteria. And we can do a longer breakdown of where and why, but I think a good example is just on public goods, right? Lama might be open weights, but when you look at all of the components that are needed to be able to use Lama as a true public good,

something that you can really build on top of, unencumbered, and not have it have to sort of fit certain constraints, right, the same sort of fights you saw in the early days of open source software, right? Lama doesn't quite meet those criteria. And so it's not public AI in and of itself, but I think we should be appropriately complementary to Mata and Lama as well for helping enable a lot of public AI innovation that has happened on top of Lama. Right? We can hold these thoughts in parallel and say that this is really helpful even if it's not public AI in and of itself.

So, you know, going back to your analogies that you gave, Nick, a lot of the public versus, you know, private comparisons made sense, right? So, you know, a bus in a car in a city makes perfect sense. You know, PBS, you know, in NBC, right? Like as an example, that makes great sense, you know, because they get somewhat similar. resources on a smaller scale. But when it comes to AI, how can that make sense or how can that work out in the long run? Because you know, you have these companies like Open AI and in Google and Anthropic with seemingly unlimited budgets, right? Billions or tens of billions of dollars and, you know, they're hiring, you know, dozens and hundreds of the world's top engineers.

And then with public AI, I don't think they have that that, that, that, that, that, that, many resources, right? So how can this still work? How can, you know, public AI still keep pace, right? How can the city bus still keep pace with these private cars that are getting all these investments? Yeah. So there are a couple parts of this. One is that public AI is partly about building differently, right? You're building on top of different infrastructure. You're building for different needs. So it's not necessarily trying to compete with the same types of public AI applications we're seeing in the world. Right. You're building on top of different infrastructure. You're You're not necessarily trying to create another chat GPT.

You might be trying to create a smaller language model that's trained on open access and openly licensed data that can then be used in high-risk, sensitive domains where you need to have very strong data prominence and be able to show clearly, like, here's what's in the data set. We promise this is not going to generate dangerous or illegal content. And so that's a very different development approach, right? you're not trying to scrape the entire internet to create the largest data set as possible to train your AI model. You might be looking for very high quality, specific data, smaller data sets, and you're going to create models that might not necessarily achieve the same scores on the same benchmarks, but comes with a set of guarantees that the private models don't, about trust, about safety, about what's in the training data set.

So there's an example of where you don't necessarily need to be hitting the same. model size and development costs as private companies. But at the same time, you're also right that there's a huge mismatch between the public AI ecosystem and the private AI ecosystem. And if we want this public AI ecosystem to actually be a meaningful counterpart to the private AI ecosystem, the same way you can choose if you're trying to go from the suburbs to downtown in a medium-sized city, you can choose between a car or a bus, we should be able to make those choices. We should be able to choose the alternative that makes the most sense for our

needs. And right now, the public AI ecosystem is just not at that level of being a true bus ecosystem that counterparts private cars. And so that's going to require major investment that can come from nonprofits. It can come from academia. But really, you're also going to need large investments from government because of the scale that we're talking about here, tens of billions, if not hundreds of billions of dollars across the world that are going into specific types of private AI innovation, you're going to need similar amounts of money to actually create a meaningful public alternative. And to reach that level of investment, governments are going to have to play a key role as well on top of philanthropy, nonprofits, and the broader public ecosystem.

So I think it's both alternatives that are built differently. on top of meaningful sort of systemic shifts in how we're investing in AI across society. So, Nick, you kind of already gave us, you know, one good example with the Allen Institute for AI in their open language model. Could you maybe give us whether they're, you know, models that are, you know, kind of quote unquote, out yet public models or either examples, right, or use cases of kind of this, you know, this shared road analogy, right? So maybe what are some other instances to help our audience better understand the importance of public models, right? Where, oh, you might not want to use a chat GPT or a Gemini for this type of work or this type of tasks.

So can you maybe, you know, give us kind of that breakdown of, okay, where are these public models and some example of use cases of where they would actually be, you know, very beneficial for that safer, you know, kind of use and public trust? Yeah, so I'll give you a couple examples, but I also want to broaden the scope here from just models to public alternatives at all levels of the AI stack, because we don't just need public models. We also need public data and public infrastructure broadly, right? So for models, for example, you've got everything from Olmo, from the Allen Institute, as we were talking about earlier, to Falcon 40B and Alpha Fold 2, right? And I give AlphaFault. photo is a specific example because it was developed in a commercial AI lab, but it was properly open sourced. And it's been used in all of these important biological applications to help advance the state of the science on so many important health and medical topics, right? So that's where you start to see some real world impacts with these models, both the large language models that are general purpose, all the way down to the more specific models trained on particular datasets and for particular applications. I would also give examples across the stack.

So common voice is a good example of something that Mozilla supports, which is this data set where we are essentially crowdsourcing voice language from around the world. So folks can donate their voice to this data set in their native language. And we're creating this data set that is open access, openly licensed, where folks can then come and build AI models on top of this data. So they can make their AI models actually work in all languages rather than just English, which is what you often see either as the only supported language or the best supported language in so many of the AI models that are being developed at commercial actors. So it's like that we see the importance of public alternatives across the stack.

I want to give one more example, which is we're starting to see a lot of governments put time and money and energy into building compute infrastructure. So in the U.S., you have initiatives like the National AI Research Resource and a program called Fast at the Department of Energy. internationally you have programs like India AI and the AI factories in the European Union that are starting to do a similar thing and these programs are essentially trying to expand access the compute needed to train AI models making that available to often academics who are trying to build AI models for their particular scientific tasks but also sometimes

for startups and nonprofits that want to use AI to solve some problem but just don't have the money to do so. So that's another type of public AI, except it's a very different structure, right? There it's government subsidizing access to compute rather than in the Allen Institute case where the Allen Institute is actually developing AI models and data directly for the AI ecosystem. All of those things can be called public AI. That's a good point, right? It's not just models, it's data sets, it's compute, right? because I think that, I mean, if you're a daily listener of this show, I think, you know, you probably realize now how much AI has kind of infiltrated our daily lives, right?

Our devices now, you know, the newest iPhones have small language models and large language models integrated, our operating systems, et cetera. You know, so as we look forward in the future, Nick, not going to ask you to bust out your crystal ball, right? But how can public AI really help, right, can help, especially with safety, accountability, and trustworthiness as AI starts to become more and more integrated into our daily lives? Yeah. So I think a core part of public AI is not just about expanding public access to AI, but it's

about making it easier for everyone to build trustworthy AI. So how do you make the tools needed, the approaches needed for safe AI, to be incorporated from the start of the development process, to be something that's layered in throughout the development process and the deployment process. So you can imagine, for example, as public institutions make these models more accessible and release them to the public, that they could be built on a different incentive structure where there's greater safety testing throughout the process of development and deployment.

There are tools that are put out alongside the model, to make it even easier for folks who are building with the model to build safely. Or I think an example of data sets, as we were talking about earlier, you can imagine this public AI ecosystem really prioritizing open access and openly licensed data that doesn't have some of the dangerous and illegal material that we're seeing in some of the data sets that are being used to train commercial AI models right now. And that's going to help promote safety and accountability across the ecosystem. So public AI done right will help create AI tools and alternatives and infrastructure that has prioritized safety from the start.

That is not going to be enough for safety and enough itself, but coupled with a better incentive structure too for developing and deploying AI applications in higher sensitivity, use cases in public domains, in places where people's rights and safety are particularly impacted. If in those spaces in particular, we're really leaning into this public AI approach, I think it'll really help create better incentives and better approaches to actually prioritize safety throughout the process, rather than sort of bringing commercial actors to spaces where there's so much sensitivity and care needed to actually use AI responsibly, that there's often a mismatch between the profit incentive and the public interest application that's trying to be. serve in that purpose. Adobe just introduced an entirely new way to create, bringing the power and

precision of its creative suite into one conversational experience. Meet Firefly AI Assistant, now live in the Adobe Firefly app, the all-in-one creative AI studio. Powered by Adobe's creative agent, Firefly AI Assistant lets you start with your vision, just describe what you want, and shape the outcome as it takes form with the assistant. The assistant orchestrates multi-step workflows, drawing on 60-plus pro-grade tools across Adobe Creative Cloud apps, including Photoshop, Illustrator, Premiere, Lightroom Express, and more to help bring your ideas to life. You can also get started with creative skills, a growing library of pre-built workflows for common creative tasks,

like batch editing photos, creating mood boards, portrait retouching, and creating social variations. Every step the assistant takes is visible, so you can refine, redirect or take over at any time. You stay in the driver's seat as the creative director. Adobe Firefly AI assistant now in public beta. See it today at firefly.adobie.com. Maybe what are some of the unknown dangers or risks that maybe business leaders aren't fully aware of when using proprietary closed large language models like I talked about?

I think now, and especially as we, you know, head into 2025, I think it's actually going to be hard for a lot of people not to use closed proprietary models. And kind of like you said, right, there's always there's always room to use each in conjunction or each for the best purposes. But, you know, moving forward, how can business leaders really make sure or find use cases where, hey, maybe public AI is something that we need to look into, you know, for this specific use case? How can they plan that out when it's, hey, these closed proprietary sources are kind of, you know, taking over our day-to-day operations? Yeah, so I think there's a real commercial interest to try to use more public approaches to AI in your own company. And what I'd point to there is think about that you're building infrastructure for the future of your company, right? I imagine many of the listeners today are not only using AI in some of their core. missions and core products, but also as an enabler of various business operations internally,

and it's sort of cross-cutting throughout the organization in a number of ways. And you can imagine, and we've seen this happen in many spots, where if Claude goes down for a day or Open AI changes its infrastructure and how its model works, the lack of transparency and openness and auditing that these open and public models can have, coupled with the fact that you're reliant on this private, sometimes finicky infrastructure, is often not a great approach for the long-term viability and uptime and reliability that you want your company and your services to have. So there are many, I think, commercial reasons why you would try to shift toward a more open, a more public infrastructure,

because it puts your company in more direct control over the AI model and the AI application itself. On top of that, sometimes cheaper. It's often cheaper, right? You're not paying for all the costs of cloud hosting and inference and API tokens that often end up really racking up a lot of costs. You'll still have many of those costs in an open model,

to be clear. But we've heard anecdotally that often an open approach to some of these problems can be 10 times cheaper in the long run. It just might require a little more work at the beginning to get it set up correctly. So a lot of commercial reasons. And then also more broadly, think about what we've seen throughout the history of the Internet over the last three, four decades, where, for example, now so many commercial applications are built on top of Linux, right? There's been so much value created through this open source infrastructure that is backed by nonprofits and or governments that can sustain those, that infrastructure over time. And then you're creating this, like, resilient infrastructure where they can't bait and switch

you into using their products and then trying to sort of gouge money. out of you over the long term. And that's one of my big fears out of the private AI ecosystem right now is that a couple of companies are going to capture this market. And then over time, because you've built everything on top of their infrastructure, that you can just keep charging you more and more because it's really expensive to change products and services at that point. And so before you get locked into a particular vendor or a particular approach, I think it's worth thinking from the start, okay, if we were to pay a little bit more, take a little more time from the beginning to build on top of open and public infrastructure, that might actually be really good for us 10, 15, 20 years

down the line. So it sounds like, you know, even if companies are already, you know, very ingrained in whatever, you know, their API of choices, whether it's, you know, Open AI or, you know, Anthropic or Google, it might be good for companies that are already seeing great productivity, efficiency gains. So almost have like a backup, right? kind of a public AI backup and to start, you know, hey, just in case, right, just in case this model completely changes and, you know, our business comes screeching to a halt, is it almost maybe a good idea for business leaders to start looking at public AI, at least for now, you know,

if they can't identify those quick use cases right away, you don't want to be locked in. If something changes, if prices go up, right? Is that maybe a use case that we're not paying enough attention to? So I certainly think it's really good for companies to start building that muscle of acquiring AI solutions, not just from public AI, but also broadly from other vendors on the primary vendor you're using, right? Start building that muscle of we're not going to just work with one company for everything on AI because it's just not good for the resiliency of your organization. So start trying to work with more folks. And within that, look at where you're you can use public AI approaches for different parts of your AI ecosystem. And I don't want to say that you have to use public AI. I think there are a lot of commercial AI applications where it makes a lot of sense to build on top of private AI infrastructure.

But there are places where you have a real choice to make and places where I think public AI will actually be a better alternative for you, places where you want that transparency about the AI services you're using, places where you want the greater safety tooling and auditing that you want in different AI models and data sets before you deploy it, where public AI alternatives can provide that and provide that with the transparency you need while private AI counterparts haven't. And in those spots, instead of just defaulting to the vendor that you use all the time, think about looking to alternatives.

Think about looking at places where you can build on top of infrastructure, that's going to be both more resilient and more trustworthy over the next many years, rather than what looks like it's the highest, best performing thing at this specific particular moment. All right. So, Nick, we've covered a ton in today's conversation. We kind of simplified and really looked at public AI and the differences between proprietary models, you know, open source, open weight models.

And then we also talked about some of the pros and cons and some of the instances where public models can coexist, right, with your existing solutions. But, you know, as we wrap up today's conversation, what's, you know, the one most important takeaway that you want people to remember when we're talking about how public AI can help us lead to a safer and maybe smarter future? Yeah. The one takeaway is that the commercial AI ecosystem we see thriving today is not the only way to build AI.

There is another way that is starting to emerge that we call public AI. that is built on top of open and accessible AI artifacts that are really creating a different future where you can build AI without always having to seek a profit, where you can put safety and trust at the center of the mission. And that public AI is not out to replace private AI, but it's there to provide you an alternative for when you're trying to use AI for your mission. public AI can provide a different way of building AI that can actually be a better fit for what you're trying to do and what you're trying to do to advance the public interest and the greater common good.

Love it. Love it. I think we all got a much needed lesson today, Nick, on the importance of public AI, especially when it comes to safety, accountability, and trust. So thank you so much for joining the Everyday AI show. We really appreciate your time. Thank you for having me, Jordan. And hey, as a reminder, y'all, we covered a ton.

If you missed a little detail there, don't worry. It's going to be in our newsletter. So if you haven't already, please make sure you go to your EverydayaI.com. Sign up for the free daily newsletter. Also, if you're listening on the podcast, thank you. Please make sure to subscribe and leave us a rating. If this was helpful, which I'm sure it was, don't keep that information to yourself. Hey, it's public, right?

Go put this out, share it with your company, share it with your friends. Thank you for tuning in. Please join us back tomorrow and every day for more everyday AI. Thanks y'all. Meet Firefly AI Assistant. Now live in Adobe Firefly, the Allman One Creative AI Studio. Just describe what you want to create in your own words and the assistant handles the rest, orchestrating multi-step workflows across Adobe Creative Cloud apps, including Photoshop, Premiere Express,

and more in one conversational interface. You direct the outcome while the assistant accelerates execution. Stand control with the ability to step in. in and refine at any time. See it today at firefly.adobie.com. And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit Your EverydayAI.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.