March 27, 2024
I’m thrilled to share that I was recently featured as a guest on the De University of Ethereum channel, hosted by Tina Dai.
In this episode, we explored the groundbreaking intersection of blockchain and artificial intelligence, with a deep dive into my work with FreedomGPT.
It was an incredible opportunity to discuss how we’re pioneering a decentralized AI network to empower individuals and democratize technology.
Video
Watch the full conversation on YouTube by clicking here.
Or watch it below:
Full Transcript
Tina Dai: Hello everyone and welcome back to our Ueth weekly workshop. This session will be an exploration of how blockchain empowers decentralized AI, which is one of our most popular topics that we’ve hosted to date. This workshop will be titled Building Decentralized AI featuring John Arrow of FreedomGPT. We will dive into how FreedomGPT and its FNT token are pioneering a decentralized AI network. I’m Tina Dai, your Ueth host for the session. Over the past five years, I’ve mostly focused on the crypto space through early-stage investing and built a foundation in product and strategy prior to investing. I’m joined today by John Arrow, who is the co-founder of FreedomGPT. We’ll hear the initial spark that led to the creation of FreedomGPT and also better understand how the company is focused on decentralizing AI to create a more equitable and accessible future in this exciting innovation. We’ll start with John’s background and explore the vision and mechanics behind FreedomGPT. John will share more of the inception of FreedomGPT, its mission, and how it got to creation, creating this FNT token to incentivize participation in its compute, to democratize AI and reward contributors. So before we jump in, a few housekeeping items. If the conversation inspires any questions, please drop them in our general Discord channel, and we can address them asynchronously. And to make sure your chapter stays active, please comment on the workshop in our Discord within 24 hours at the University of Ethereum. Our goal is to really inspire our community with real innovations, and we do this by pushing you to learn through doing, reflect on how you can contribute to decentralizing power and knowledge and AI. With the story that you hear about FreedomGPT, please join me in giving a warm welcome to our guest John. I’m going to add John now to the stage. So John, it’s great to have you with us today. We would love to learn more about your background. What were you doing before FreedomGPT?
John Arrow: I’m really excited to be here. It’s just so fantastic that you’ve opened up this environment for teaching. So thanks again for having me on. My background in crypto is kind of an interesting one. I started a company called Mutual Mobile when I was at the University of Texas, and we bootstrapped that to 400 people. A lot of what we were focusing on was emerging technology, so whatever was new, we helped Fortune 1000 build it. And then when crypto came onto the scene, we kind of had a look at that, too. And so back in 2011 is when I first got involved in Bitcoin. And like a lot of people then, I didn’t know exactly how to purchase Bitcoin. So I heard about Mt. Gox, which we all know the crazy stories there. Fortunately, I got my Bitcoin out of Mt. Gox before it kind of did its implosion. But that was my first kind of entrance into it, not really knowing what it was about. I think I have some Bitcoins with a cost basis of close to $18, and that was out of sheer luck and curiosity rather than anything else. But along the way, I kind of kept in the back of my head what had happened to Mt. Gox, right? You had this really wonderful technology which was supposed to be decentralized, but then what did people do? They ran and they put it into a centralized exchange. And we keep seeing that story really repeat itself over and over and over again. It seems like you think we’d learn lessons from what happened with FTX, but you still have a ton of people who put crypto in a centralized exchange. And apparently, you know, they say history never repeats itself, but it sometimes rhymes. Well, I think we’re going to see a lot of that happen with AI, unfortunately. And so when we founded FreedomGPT, we recognized that AI and generative AI was going to be just as transformative as crypto. And unfortunately, in order to ensure that, it meant we had to take a book out of the crypto page and do a decentralized route there. As soon as AI launched, we saw governments and large organizations trying to figure out how to specify what it could and couldn’t say and when it could and couldn’t answer something. And so that was the impetus for us wanting to create the decentralized AI app store that has become FreedomGPT.
Tina Dai: Wow, so you’re a true OG in crypto. I hope you held onto that Bitcoin. But you kind of touched on the origin story of FreedomGPT, at least the philosophy behind it. Would love to hear a little bit more about what was that initial moment? What was that spark initially that led you to say, I’m going to commit my resources to building this, this product in app store?
John Arrow: Great question. And I think, like a lot of things in life, sometimes there’s serendipity in timing. So I had sold my company at the end of 2023, and a month before we sold it, ChatGPT was launched by OpenAI. And we had done a lot of work with AI while at Mutual Mobile, my prior company. But I think it was just perfect timing that when we sold the company, that was right when ChatGPT was becoming a household name. And it launched in November of 2022, for those who weren’t familiar with the launch. And one of the crazy things about that is I think everybody had this experience where they started using it and they’d ask a question, and ChatGPT, despite it being remarkable, would refuse to answer. And it reminded me of a similar moment that I had back when I was about 12 years old on AOL Instant Messenger. Now, I know a lot of the listeners are younger than me, but there used to be this thing that AOL put out there, America Online, called AOL Instant Messenger. We would all use it. And one day this company launched something called SmarterChild. And SmarterChild wasn’t AI, but it was one of the earliest chatbots that you could ask questions. And it was a remarkable piece of technology for its time. And you would ask it a question, it would most of the time answer, but every so often you’d ask something and it would refuse to answer. And I’m sure they had their reasons. Weirdly enough, Microsoft ended up acquiring that company, SmarterChild. But it had a huge impact on me when I had a computer for the first time in my life saying, no, I’m not going to answer that question. And again, that was when I was about 12 years old. We’ll fast forward to last year or a little over a year ago when I asked ChatGPT a question that was seemingly innocuous, and it told me, no, I’m not going to answer that. And it became this extremely kind of transformative moment for myself and my co-founders when we started just seeing what it would and wouldn’t answer, where we said, why should a computer be judging you? Why should a computer allow what you can and can’t ask? And AI will likely be the most significant technology for humanity since electricity. Now, if that’s the case, we need to make sure that people have access to a true version of it, not one that a large company or world government decides that you’re allowed to have. And so with FreedomGPT, one of our first models allowed you to ask anything without bias and without restriction. So nothing was off limits. And just by the sheer nature of this being salient, we had a ton of people start using it. It was featured in the New York Times. Over the last year, we’ve had over 2 million people interact with our product. And that really paved the way where we said, well, okay, if we have a text-based LLM that’s censor-free, what are some other models that we can add? And some of them we strive towards having open-source and censor-free models. But we also recognize that the important thing for our mission of giving everyone access to every AI is to allow people to access AI models that might be difficult to get their hands on otherwise. A lot of these models require you to have a substantial amount of compute available and some technical knowledge to set it up. But the great thing with freedomgpt.com is you can just go to the website, you can open up the browser, you have access to 80-plus models, the most cutting-edge ones in AI right now, and you could start using them with no knowledge. If you are more inclined from a technical standpoint and you do have substantial compute, you can actually use our models for free. You can download the FreedomGPT client to your desktop where it’s 100% private, it’s 100% censor-free for the Liberty model that we have out there, and you can run inference for others. And we’ll actually pay you with our token, which we’ll talk about later, so that you can help other people access models in a decentralized fashion.
Tina Dai: I love the story you shared about your youth. And it’s also fascinating to me how, like a lot of what I love about crypto, is this pursuit of freedom and pursuit of certain ideals. And it sounds like basically what you’re doing with FreedomGPT is bringing some of those ideals, chiefly like freedom, to this technology. And so I think this would be a good transition now to your presentation on FreedomGPT and what led you to this intersection of crypto and AI, as well as a lot of the challenges that you’re currently seeing in our landscape today. So I’m going to go on mute and have you start your presentation.
John Arrow: Excellent, Tina. Thank you. Before I dive into it, I want to completely agree with something you just said, that philosophically, ideologically, the mindset between crypto and decentralized AI is so, so similar. I would say in many cases, AI is the first real use case of crypto outside of just kind of financial products where we really need crypto to function. There’s no way that decentralized AI can work unless there is crypto. So I’m so glad that you brought this point up. And the crazy thing about what’s going on right now, if you want to kind of encapsulate the problem down into kind of one simple slide, that’s that because you have a transformative technology, everybody who is an incumbent in the world is scrutinizing it. Some people are scrutinizing it for the right reasons, some are scrutinizing it for the wrong reasons. And if you look at why AI is kind of provoking, for lack of a better word, the same autoimmune response that crypto did, it’s because it threatens kind of the status quo. It threatens the incumbents that are there. And again, some of them are operating with, I think, good reason, right? I mean, the FTC is a very important organization, as well as a lot of other government bodies. But whenever there’s a new technology and they don’t understand it, they can sometimes rush ahead and hamper innovation and kind of harm consumers rather than help them. So OpenAI, we’ve already seen them banned in several European countries. We’ve seen a ton of scrutiny. And this is still in its infancy. And so kind of our thesis here is that as the capability of AI becomes more pronounced, what’s inevitably going to happen is that autoimmune response is going to go from, you know, a scale of 0 to 100. I would say it’s at about a 5 now. I think it’s going to go to about a 75 before the end of 2024. And what that means is that we’ll be in a situation where unless you have a decentralized version of AI, you aren’t going to be using the latest and greatest version of it. And we anticipate kind of that level of censorship and restriction on AI will be probably directly proportional to the capability. It could be logarithmic. Again, in the news, we heard about Google Gemini, and obviously I think people are being a little bit overreactive when they say Google’s trying to rewrite history. I would say it was most likely more of an honest error, or maybe somebody got a bit overzealous. Nevertheless, I think it invoked the type of response where people are starting to see the danger of what happens when the traditional incumbents are able to be the kind of sole ones accessing this really, really powerful technology. And what scares me isn’t more incidents like Google Gemini. I think what’s scary to, you know, not just to the AI industry, but humanity in general, is when it becomes more difficult to discern. Obviously, if Google Gemini gives you a picture of George Washington that’s a different race and a different gender, you’re aware of it. But the next phase will be way more subtle. It’ll be way more insidious and be almost impossible to discern. Is that true or is that not true? And that’s really why we need a diversity of AI models, is to protect ourselves from one large organization saying our model is the right model. That’s really the only way around it. You need a diversity of voice and choice in the market, hence why you need a decentralized AI app store like FreedomGPT. And it’s amazing how arbitrary and capricious some of the censorship is that we’ve seen. Midjourney, an extremely successful company, already has kind of decided what it will and won’t allow. And some world leaders are allowed, some are not allowed. And we’re not going to kind of try to dive into why they chose different things other than to say that this problem is going to get way, way worse before it gets better. Any significant company is going to kind of have to bend to the sway of large incumbents. So really the only way around that is to ensure that there’s a decentralized approach. Any centralized company, big or small, will not be able to hold up to the immense amount of world pressure they have from large incumbents to basically ensure that their AI operates in a certain way. You really kind of have to decentralize the thing and throw away the keys so that can’t be changed. That’s our eventual goal with FreedomGPT here. What’s fascinating too is, like we talked about, that exponential curve as it relates to that censorship and that prohibition on AI’s capability increases. So naturally, it should be no surprise that there’s an inverse. Look, there’s an inverse correlation to being able to use capable AI in an unfettered way. We’re already seeing a huge, huge push to monitor people’s use of AI, and that’s only going to intensify. Unfortunately, a lot of people today rely on ChatGPT and other models that you can use at FreedomGPT, you know, for legal advice, especially if they don’t have access to legal representation. And unfortunately, that does not have legal privilege, right? And so you’re already kind of seeing a situation where you have a two-tiered justice system where people who ask a lawyer questions that can remain private, but if they ask an AI lawyer question, that’s part of the evidentiary record, and that’s really unfortunate. We’re also seeing governments around the world enact pretty far-reaching, I think, ill-informed regulations around what rights you have and don’t have to using AI privately. There’s an executive order that’s going to go into effect this summer in the United States that, for all intents and purposes, takes the whole KYC, Know Your Customer, AML, Anti-Money Laundering laws and will start applying it to AI, where you’ll have to fill out certain forms and identify yourself before you can use AI in a way that you want to use it. Now, the problem with that is it starts to have a really chilling effect. If you know the questions, the inferences that you ask, are going to be monitored and you know that they are subject to scrutiny, it’s going to start to interfere not just with your, the questions you ask, but really the way your mind works. It starts to be something that I think is very, very much a slippery slope as it relates to civil liberties. And again, one of the reasons why privacy is so important and the reason why we allow people to download FreedomGPT and have 100% private inference never leaves their computer. So again, you know, you may never need this, but if you download it today, at least if the internet goes off, or if, you know, some world power decides to ban AI, you’ll at least have that on your computer. We realize that, you know, the freedom, the way it works in the United States, is not quite understood. We think that freedom is something that’s guaranteed in a document, maybe the Bill of Rights or Declaration of Independence, whatever document you want to look at, and that might set the parameters and the stage for that is not how it is assured though. The only way you can assure freedom is making sure you have more powerful technology than the people who want to take it away. History is full of transgressions of that right. There have been societies that are way more just, way more ethical, moral by a lot of different standards, that were overrun by other civilizations that had better warfare technology, better economic engines. And so when we think about what preserves freedom in the United States and what will preserve freedom in most of the sovereign countries on this planet, it’s being able to access the tools, it’s being able to access the information, it’s being able to have freedom of speech to ensure that you’re not overrun. That’s why it’s so important to put this technology in the hands of ordinary citizens. It should not be segregated in the hands of a few powerful incumbents because if you do that, it will be misused, and it will be used in a way to subjugate the people who don’t have it. Just, it’s just human nature. When we launched FreedomGPT, we kind of weren’t aware and didn’t even expect the community that would rise to the occasion for us. It was just so heartwarming to see people understand immediately what we were trying to achieve. Yes, we got a lot of hate mail, for sure, by people who think that AI should be censored and should only have the right version of the answer. But overwhelmingly, we realized that most of the people kind of clamored to us, got the mission, and there were a lot of them who were people who had a crypto background. What I think is interesting too, is an insight that many of our users share, and they have a healthy respect for the concept that AI could give out misinformation. You hear time and time again that the danger of uncensored AI is misinformation. And that might be true, right? There are sometimes when misinformation might exist, for sure. But, you know, one person’s misinformation might be another person’s information, vice versa. That being said, the real reason when people want to censor AI, if you kind of dig deep, you really interrogate what’s going on, it’s not that they’re concerned about misinformation. They’re concerned that the AI might give a real truth that people aren’t ready for. Because imagine a hypothetical AI that, whenever you asked it a question, it always gave you the right answer, right? Every day for 100 days, you’re asking it different things. You’re asking it to predict the weather for you, you’re asking it for something that happened. And it always agrees with your worldview. You start to rely on it, you start to build trust with that AI. Then all of a sudden, if you ask a question that is a controversial question, it maybe has a lot of kind of disputed answers to it, but it gives you a certain unpopular answer unequivocally, that can cause social upheaval, right? That can cause a lot of civil unrest. And I suspect that’s what most people who, if you really dive into it, that’s why they’re concerned with kind of uncensored AI. And our belief is it’s the exact opposite. If you try to censor history, if you try to restrict what people can see, it will only end badly. There’s that famous quote that those who start by burning books will end by burning men. And I think AI very much will play out in the same way unless we keep it open and available and free for everyone on this planet. So we have a lot of users who come to us for different reasons. And I’d say early on, the reason people came to us most was because they were interested in getting AI where they can ask any question and always get an answer. And that very much evolved into, you know, this idea around privacy that whatever they ask will remain private and stay on their computer and not be subject to somebody else finding it or being leaked or something like that. I would say, though, today, most people who access FreedomGPT care more about the proliferation of all of the different models we have. Whenever there’s something cutting-edge new in AI, whether it’s text-to-music or text-to-video synthesis, or, you know, an AI that can make a phone call for you, which our Phone Call GPT that can do—phonecallgpt.com, it’s live on freedomgpt.com—and it can literally make phone calls and do things for you. They come to us for those types of applications, a new, new thing that they can’t get anywhere else. I would say that’s the lion’s share of reasons why people access us today. That being said, we were built on the foundation of being censorship-resistant and being private. So those are the underpinnings. I would say the main value prop today, though, is just this prolific access. You don’t need a subscription to a dozen different AI companies that can really add up with FreedomGPT. You can use all of those models and not need to pay a dozen companies. I think that’s super refreshing for users that want to be on the cutting edge of AI. Now, we are very much a company that did not set out to be a blockchain company. We stumbled across it out of necessity very early on in the FreedomGPT journey, back when we were getting millions of users finding out about it every day. We ended up in a scenario where somebody would ask a question and, in their view, the LLM gave them the wrong answer. Well, they would go complain to our centralized host provider and say, hey, you shouldn’t allow these guys to be hosted with you. And in the middle of the night, they would shut off our hosting, even though we were paying them many thousands of dollars a month. And this happened several times. So we realized, look, this is such a controversial, such a useful technology. There’s no way this can continue to exist and propagate in a centralized hosting fashion. It was at that moment we realized we needed to become a Web3 company, and we needed to figure out how to make it so that one person couldn’t change their mind and shut us down. And so if you look at our journey towards becoming a Web3 company, we kind of recognized several things. One, it was within the open-source community. We have a huge commitment to being open-source. We were number three on GitHub. We had nearly a million people download and install our free desktop applications so they could run private censor-free inference. We have a cloud version where we weren’t monetizing for a while, but then we decided to start monetizing to offset some of our hosting costs, where we’ve consistently been doing a thousand dollars a day and growing at a pretty good growth rate. And so we kind of took all these variables together and the fact that we had been banned from hosting providers, the fact that you can’t even mention freedomgpt.com on Twitter or X in a direct message, and all of this other kind of scrutiny. We said, look, we don’t really need the traditional computing rails. We can go completely decentralized. We have a ton of users with amazing compute. Why don’t we let them start hosting the FreedomGPT cloud models for us? Because this will achieve several things at once. Talk about killing two birds with one stone—all of a sudden, it means that we cannot need to rely on the centralized hosting providers. Number two, it means we can give back to our early supporters; we can figure out how to pay them the money that we were paying our hosting providers. And number three, we can ensure that this is done in a way that remains completely private, it can’t be compromised, and there isn’t one person who can change their mind and kind of say shut down FreedomGPT. And so that was the realization where we decided, look, we’re not like a typical crypto project. We already have a lot of revenue coming in. How can we kind of migrate and take what we’ve built and make it so that it has the benefits of what Web3 can offer? And that’s exactly where we are today. We’re right at, I’d say, the end of the beginning on that stage of decentralizing FreedomGPT. We did an airdrop of our native Freedom Network Token, and that is the economic unit with how we pay our nodes, our hosts, and anybody can be a node or host. All you need to do is download the desktop version of FreedomGPT. We just did an airdrop where we airdropped almost 50 million Freedom Network Tokens to eligible nodes during that period. We called that our pre-genesis period. And we’re just about to begin our next airdrop, where we are going to be rewarding nodes on our network who are able to process true inference for people who want to use the FreedomGPT cloud model. And those FreedomGPT cloud model people are—maybe they’re people who are not technical, they just want to go to a browser and use our 80-plus models. Maybe they want to use them on their phone. Maybe they don’t have the compute available. But if you are interested in kind of earning these tokens, all you need to do is download the application, keep your computer plugged in when you’re not using it, and then in a completely anonymous, safe way, when your computer has a spare cycle, it will process behind the scenes that inference for a user, and you’ll get rewarded with those tokens for the work instead of AWS. And so it’s a really kind of interesting way to give back to the ecosystem. There are going to be other ways to earn tokens right now, but the most important part is building out kind of that infrastructure layer where we can process inference in a completely decentralized manner. I encourage everybody to check it out. The easiest way to kind of start to learn about the Freedom Network Token is to go to freedomgpt.com/fnt. If you do that, you’ll see a world map of all of our nodes. You’ll see all of the nodes online, as well as kind of the amount of compute that we have. And we have—just again, we did this less than a month ago—and as a result of the airdrop, we have several million dollars’ worth of compute now online that we can use to process inference for our cloud users. And that will translate directly into earnings for our desktop users who are contributing. Here’s that map that I talked about right here. What’s so cool about this is that you can see there’s true distributed computing here, true decentralization. We’re all over the world. I’m not sure why the graphic shows Somalia as being highlighted, but yes, we’re, you know, we’re not there, but we’re in a lot of different places. And every time we add a new country and we add a new region, it increases the robustness of the network. What it also means too, as regulations and kind of new laws become clearer as it relates to AI, we can ensure that different regions are able to run different models. And so it provides a great comparative advantage. It also means that when people in the West are up and using their computers, we can use the nodes in the East when they’re sleeping and not using them, and vice versa. So it gives us a lot of just ability to do some pretty cool things that we couldn’t do if we were just kind of centralized in the Americas. There’s been a huge amount of movement, I think, in DePIN in the last couple of years. You started to see some fantastic examples of how you could take the early days of crypto where you were just brute-forcing, basically SHA-256 encryption algorithms, and start to do useful work. And I think this is the type of thing that’s going to really change the image of crypto when you can, for the first time, say, look, crypto isn’t just an energy hog, it’s not just an energy consumer, it’s building things that wouldn’t have existed otherwise. And you know, for my personal pursuit, that’s something that I care a lot about, something that my co-founder Tarun cares a lot about as well. And so I’m so happy that we get to stand kind of on the shoulders of giants who’ve been working so hard in the crypto space for the last, I guess, well over a decade now, and really figure out a way we can contribute to the world in a positive place and kind of change what it means to be a crypto company. So when people talk about AI tokens and they talk about what’s going on in AI, there’s a lot of elements to it. There are a lot of great companies in this space. Morpheus is one of them that we’ve gotten to know well, Kosh is one of them, Bittensor, there are a lot of different ones, and I think that’s a great thing. What many, many people, and what the casual observer might want to know more about these AI tokens, is that there’s a ton to bite off and chew here. When you’re talking about creating utility with AI, you have to look at the chain and kind of understand where, you know, where are the opportunities, where is the room for improvement? And so we decided to focus predominantly on the inference side of it. We wanted to figure out how we can create a great UI layer. Again, we have, you know, just, we’ve had over 2 million people use FreedomGPT. When there’s a brand-new model that’s out there, how can we ensure that that model owner gets distribution to a ton of captive users, and they don’t have to worry about the hosting side of it? And so that’s why we’ve decided to focus predominantly on kind of the inference side of it and the output side of it too; over time, we might look at the other value layers as well. That being said, we don’t want to reinvent the wheel. We want to rely on partners and be some additive to the ecosystem. I don’t think there’s going to be kind of a real competitive nature in the ecosystem for a while out because there’s just so much that needs to be done. And you have a ton of companies out there that are all kind of approaching it in their own manner. And we’ll get to some of the differences in the companies in a second. But here’s a really great kind of high-level thing, and I noticed one thing we need to clean up on that slide real quick is I think that logo is Bittensor and then right next to it is FreedomGPT. Again, very different companies working on different sides of the value chain, but they are, I think, very much complementary. And the other perspective to look at this, of course, is just by sectors. We have something that looks like one of the big think tanks could have put together here, but we see a ton of cross-pollination through all of these different kind of optics. Decentralized compute is important. That being said, we think decentralized compute is very much a race to the bottom. This is an age of abundance, and there is going to be way more decentralized compute out there than is needed for the current problem stack. So instead, we kind of have to turn this into an opportunity question and say, look, we have all of this decentralized unused compute now that we’ve kind of linked it together and things like FreedomGPT have launched. What are the novel problems that we can solve that we couldn’t solve before? One of our initiatives is we’re rewriting Wikipedia. We have something that we’re calling WikiFreedom, and we’ll talk more about that at a later date. But we realized that we could basically create a much larger, less biased, less censored version of Wikipedia in a fraction of the time using some of the compute that we have on our network. And so I think you’ll start to see these categories and this kind of gestalt way of viewing the world really start to blend together in some pretty remarkable ways in the future. And again, I encourage anybody who’s interested in crypto to explore Freedom Network Token. This is one of those rare tokens where you do, you know, we’re not trying to do like a launch or an ICO or anything like that. We are an established company, and we are launching the token because it enhances the privacy, the censor-resistance for our users. If we didn’t have this, we would have to deal in fiat. And so it’s a pretty unique token that you don’t need to pay any money, you don’t need to stake any Ethereum or anything like that. We literally will reward you with our token for simply downloading the desktop app and enabling your computer to run inference for our users. And by doing that, you don’t need to have your computer do brute-forcing of encryption algorithms like you’d have to do with Bitcoin and a lot of the other tokens out there. It’s something that most computers can handle, especially if you’ve purchased your computer in the last few years, and you don’t need to kind of risk anything. It’s only going to be given to you, you know, when you’re, you’re only going to need to keep your computer on when you’re not using it; otherwise, you won’t even know it’s there. And so I’d encourage anybody who’s interested in earning Freedom Network Token to check it out, and we’ll pause there and see if there’s any questions.
Tina Dai: Yeah, well, thank you so much for that, John. I have a list of questions prepared for you, but before we move on to that, I’d love to just hit on some of the points you touched on. I think the first being crypto is one of the first applications of crypto is AI. I think that’s a really interesting statement that you started off with because aside from things like payments, crypto has historically had some challenges in finding consumer applications, right? And I think that statement of blockchain being an equalizer for AI is so astute in that, like, if you think about AI, it’s like so centralized. It’s a black box. It’s very, like, centralized and monopolistic, right? There’s, like, a handful of companies that own all these, like, large assets. Whereas with blockchain, you get the opposite of that, where it’s very decentralized, very transparent, and then it allows for user monetization and accessibility. So I think tying together these two, like, giant technologies really has a ton of synergies that can be achieved independently on its own. And so I think your story itself is actually quite a good encapsulation of all those principles, right? You were very laser-focused initially on tackling censorship and access, but you realized that creating this decentralized network was not really just, like, something to do but a real necessity for you to operate in an ethos-aligned way that you want to be doing without getting deplatformed. And I also think this whole concept of ensuring everyone in the community gets rewarded fairly is my whole favorite thing about crypto. Decentralized resource networks in general, like you had touched on, like DePIN itself as a category, really fosters various communities where things don’t necessarily exist or resource networks don’t necessarily exist outside of this crypto paradigm. And I just love this foundation of giving ownership to users based on their actions, which truly highlights to me what’s exciting about the future of decentralized AI and decentralized networks in general. I think for this next section, we’re going to spend 10 or 15 minutes on Q&A and just dive a bit deeper into some of the thinking behind FreedomGPT, as well as the community-building aspects that might be crucial to its success. If this conversation—anyone who’s listening to it—has any questions, you can definitely drop them in our Discord, which is discord.gg/ueth, and then we can answer those asynchronously for you. But some of the questions I have prepared—so I think both AI and crypto are pretty new concepts. And for the average user who may be using your platform, how are you working to familiarize them with education either around crypto or AI, such that they understand the value proposition of your technology?
John Arrow: Well, you summarized what I just said so eloquently and so succinctly. So thank you for doing that. I should have just deferred to you. You’re really good at getting things quickly and summarizing them. And then as it relates to your question, I would challenge that question a bit around AI being a new technology. AI’s got quite a history. It’s approaching 100 years old if we think about the first vacuum-tube computers. And when we were looking at using computers for firing solutions, it was imagined that you could create computers that could think. And there’s been just an immense amount of research and failures to kind of get to the point that we are at. And what’s interesting is, even though we’ve seen some recent success with generative AI and it’s become a household name that everybody can use it, people already kind of envisioned the possibilities of robots, of all of the amazing optimistic scenarios for what AI could do, as well as all of the nightmare scenarios what people could do. So I think it’s a bit different from crypto in the way that, even though encryption has been around for a while, the concept of blockchain is extremely new, and it’s something that has not been around as long. So it’s kind of fascinating to see those two technologies merge in such a way because both are becoming household names. And I think it shows that we may be in a simulation that these two are growing together now at the exact right time, because I don’t think that they can exist well without each other. And you’re seeing an immense amount of cross-pollination there. And one of the interesting things is they kind of serve as a check for each other, right? If you think about AI in preventing the nightmare scenarios that people are concerned with, a lot of the nightmare scenarios are when one centralized organization gets a better AI than everybody else, right? If one of our adversaries gets better AI or some terrorist group gets more capable AI that can produce the killbot drones and take over. And crypto very much can stand as the inhibitor and the check on that. Because if we figure out a way to give it to everybody, that means that we have this kind of decentralized arms race that keeps everything in check. I wish we lived in a world where we didn’t need weapons. Unfortunately, that is not our world. Instead, we live in a world where there are more good guys than bad guys out there. And so if we know the bad guys are going to get the weapons, we want to make sure that the good guys get the weapons too. And AI in the hands of good people will serve to inhibit and block and protect the bad actors that are using AI. And crypto is the facilitator to ensure that.
Tina Dai: When you mentioned robots, it reminded me of the NVIDIA demo yesterday with robots. And I had this crazy thought. I was like, oh my gosh, we’re going to have robots in my lifetime. Which is an idea I maybe, like, loosely thought of on, like, a sci-fi TV show or movie, but never, you know, grappled with as reality.
John Arrow: So, well, for the audience, in case people are watching this in the future—neither of us are robots, we can confirm.
Tina Dai: Oh, well, one thing you had touched on in your presentation, which is you think, like, decentralizing compute is a race to the bottom. Can you expand on that? And you also mentioned a bit more about the different ways you’re using excess compute with FreedomGPT. So we’ll have to understand the vision behind the products that you’re perhaps pursuing or exploring with the excess compute. And also why you think it’s, like, a race to the bottom slash the implications for all this—you know, there’s so many classes of companies pursuing decentralized compute, so would be interested in understanding, like, the implications of what that would imply.
John Arrow: For sure. I mean, and I think right there, there are so many companies pursuing decentralized compute because they see the importance of it, and it will basically guarantee that there was going to be a lot of decentralized computing compute out there. If we look at the history of just centralized compute, we see this seesaw back and forth between hardware and software where, ultimately, whenever there’s something new from a software standpoint, it requires an immense amount of hardware to run it, right? If you think about, like, the software it took to do the moon landing, it was, you know, essentially a supercomputer that by today’s standards couldn’t run a wristwatch. And so we always see this inevitable—software does something new, and it requires the most gargantuan, most expensive hardware in the world to run it. AI has been no different. FreedomGPT has been no different. When we first launched our censor-free model, we calculated what it would cost to run in the cloud for the users for one day, and it was several hundred thousand dollars per day to run that. Today, it’s measured in kind of sub-$1,000, which is pretty insane that in the course of a year—really less than a year, I think FreedomGPT actually, it’s our birthday today, weirdly enough, we launched it one year ago today. Just good serendipity. Thank you.
Tina Dai: Wow.
John Arrow: And so if you think about, like, a one-year-old, the progress that a one-year-old human makes, it’s immense. And the same thing with software. In that time period, we saw several orders of magnitude in reduction of cost. And that isn’t because of magic, it’s because the software got better. It’s because people started competing and saying, how could we do things in a more efficient manner? And then the next, the next kind of permutation of that was, how can we pull more people, more resources into the inference side of it? So not only did the software get better, but we got more, more kind of smart and more creative in how to bring in more hardware and compute rather than going to NVIDIA and buying an H100 for $40,000. A bunch of people have extremely capable computers sitting at home unused on their desktop or sitting at home on the weekends in their office not being used. Or maybe their mobile phones in an old desk drawer somewhere not being used. And just by simply giving people—going back to the DePIN comment that you made—giving people the slightest incentive to turn those on and to spend a negligible amount of power, of energy on power to keep them plugged in, connected to the Wi-Fi—that solves all of our compute problems, and that’s that race to the bottom. There is no hardware shortage if everybody plugs in their hardware. And yes, it is true, there’s going to be software that pushes the boundaries again. If you look at what OpenAI is doing with Sora and a lot of the other text-to-movie synthesizers out there, you will see that other side of that seesaw go back where you need the most capable compute to run that inference. But economics are that the hardware will catch up, unlock that. So I do think it’s a race to the bottom as it relates to decentralized inference economics. But it’s a type of race to the bottom that benefits everyone in the space. It means that the people who are enrolling their computer are going to earn some type of compensation for it. It means we’re going to see more capable models that can be used by others for an amount approaching free. And it also means we’re going to see new models quicker for the open-source community that we wouldn’t see otherwise. And that’s one of the best things about capitalism, is that it ensures that the consumer starts to find a world where things they used to pay for become free, or in fact, they start getting paid for using them, even in the case of FreedomGPT.
Tina Dai: Yeah, that’s fascinating because effectively you’re basically saying that we’re going through this hardware-software cycle where software puts demand on hardware, and then hardware eventually catches up, and so it pushes out the software piece further up the stack to where now we’re getting more innovation on a consumer application front. And so your prediction is effectively that over time, as we sort of go through the cycles of catching up on hardware, the things that are available to consumers that compete for their attention will become, like, more and more interesting and compelling for the end consumer. Which I would love, I would love to see that—especially, like, in crypto, right? For instance, like, the idea of, like, a consumer application in crypto has yet to catch up. And with AI, you’re sort of seeing this, like, happen a lot faster where there’s already a ton of consumer applications that people are trying with, like, AI agents and things like this.
John Arrow: And so really, crypto and AI, they’re the perfect partner.
Tina Dai: Yep, exactly. I’m really excited to see how the ecosystem develops overall. But with this very rapidly growing decentralized AI ecosystem, what prospects do you see for potential collaboration with other Web3 AI projects?
John Arrow: Well, that’s what gets so exciting all of a sudden. If we think about everything in the world that contributes to morbidity and mortality and quality of life, you can kind of boil it down to energy, right? People’s access to light, sweet crude. People’s access to kilowatts—that really, more than anything else, controls humanity’s progress. If you have unlimited power, you can have unlimited desalination, so you can make fresh water wherever. If you had true unlimited power, you could control the weather and ensure that everywhere on the planet is perfect for growing crops and living and even do space exploration in an extremely efficient, fast way. Now, compute is not energy, but it’s close to it. Because once you start giving access to compute that approaches free and has really no bounds on how much you can get, it unlocks a lot of the same potential for humanity. So when Web3 and the crypto ecosystem start looking at what AI has done with decentralized compute and what crypto has done with decentralized compute, it allows you to start asking some pretty interesting questions. What would you do with close to unlimited free compute? What problems would you solve that couldn’t be solved before? Maybe it’s things relating to the genome and doing, you know, protein folding and solving that. Maybe it’s something relating to SETI and looking at the universe—there are already decentralized applications that do both of those things. But the level of decentralized compute that’s about to come online will mean that anybody can come up with an idea and access many multiple millions of dollars of compute for close to free in a way that society has never been able to do before. And there’s just such cool examples of what that can do. I’ll give just one, because otherwise we’d spend the whole day talking about all the fun things. I would encourage the audience to think of their own. One really cool one is, what about your own AI that listened to everything that you did throughout the day, and it only had your interest at heart? That data didn’t get sent anywhere else. It just said, how could John or how could Tina have the best versions of themselves? And that AI is looking out for your interest? Well, you previously couldn’t do that when you had to pay for compute. It would be prohibitively expensive for most people. But now you can do that in a way that’s close to free. And FreedomGPT is actually going to be launching a specific model that does effectively that soon. But that’s just one example. And I guarantee you there’s literally millions of other ones that could be just as useful. And so by having that decentralized AI app store, anybody—maybe in the audience who comes up with one of those ideas—they’ll be able to host it on FreedomGPT and immediately overnight access all of our captive, excited new AI users to test it out.
Tina Dai: We didn’t touch on this, but what other models does FreedomGPT support, like this AI agent?
John Arrow: We’re a big believer in multimodal models. Ultimately, the world that we live in isn’t just text-to-text. It isn’t just seeing things or hearing things or feeling things. It’s a bunch of the different senses. We’re trying to add models to support every one of the natural senses we have. In addition to many different LLMs for code generation, for writing, we also have text-image synthesizers, we have an early text-to-video synthesizer, we have text-to-music. We have an AI that can make phone calls for you. We have a lot of models that will kind of teach you how to do things. And the nice thing about it is the models that are being added, people can experiment with them, and they can ask the same inference, the same question to multiple models and then choose the one that gives them the best answer, the one that’s the most helpful to them. Over time, we think the biggest problem isn’t that there isn’t going to be enough models. We think it’s going to be the opposite—there’s going to be way too many AI models. So right now at FreedomGPT, we’re working on a router that will automatically try to understand what you want to do. It will learn in a secure and private fashion and then route your prompt to whatever model can give the most valuable inference. And by doing this, it takes the guesswork out of it, right? A lot of people have too much choice in their world today. You don’t want to have to sift through a thousand different models to figure out which is the best. You just want an AI that can sit at the top of the stack and say, look, I know what you’re trying to do here, let me go figure that out for you. You don’t need to worry about this. And we want that to happen in an increasingly passive way where eventually you won’t even need to prompt the AI. It’ll just go and infer itself that it needs to go do one of these models, and then it will perform the tasks for you and let you know, and that’s how it should work. It should be in the background. You shouldn’t even need to think about it.
Tina Dai: Wow. So much to look forward to. But we’re now coming up on time and would love to move towards a closing next step. So a couple more questions for you, John. So as we wrap up, we’d love to hear you share any parting thoughts or advice for our community, especially for those who are interested in this intersection of AI and crypto. Maybe you can share any resources or pathways you might recommend to get involved in the community, or how do they even engage with FreedomGPT?
John Arrow: What a great question. I mean, I think if you were just starting out, what a magical time to be looking at the world, because I can’t think of an easier ramp to get involved in something. With crypto and a lot of Web3, you had to have somewhat of an engineering background. That was pretty important. With AI, though, it’ll kind of teach you itself, right? It’s got an instructor built in, which means anybody who’s curious enough, anybody who’s willing to spend the time to start using and experimenting with the different products out there, can become self-taught, and then they can figure out how they can contribute and add to the community. Whereas with crypto, no amount of using crypto is going to teach you really about cryptology. It might teach you a few things, but you’re not going to be writing Web3 applications unless you really sit down and focus on it, versus AI will teach you how to do it. I think the biggest thing is just to start experimenting and seeing what are things that you can do today. Substitute out workflows that you have currently with AI. Initially, it might not be easier, but over time, the AI is going to get better. And I think people who make that leap now are going to be way, way ahead of people who are kind of dragging their feet on it, is the one point I’ll say. And then the second point I’ll say is that if you want to do an analogy to where we are in the timeframe of things for the new generative AI kind of wave, I think this feels a lot like 2011 again, like AI has been out there, but only the extreme early adopters have really played with, I think, generative AI in a meaningful manner. And we’ve already seen kind of some plateauing. If you look at the user growth around OpenAI, it’s largely plateaued. So this is the perfect time to get into it because people are having to build, and there’s some fundamentals there rather than just kind of market hysteria, which is great for bringing in new people, but it’s not great for kind of building and getting things done.
Tina Dai: How many users does OpenAI have?
John Arrow: Okay, that is a very disputed question. I heard that they have 100 million paying users—so subscriptions, I believe, is what was what the status, which is substantial, and this is substantial—2024, but let me just confirm that.
Tina Dai: Yeah, I’m just curious because you said it plateaued, and so verifying 100 million paying is actually lower than I would have expected.
John Arrow: Actually, 100 million people using it monthly, so it’s probably paying is actually a lot less.
Tina Dai: Oh, wow. So that’s actually a lot less than I would have thought. But very interesting points. I also think, like, for AI, it’s, like, yeah, we’re so early still in the entirety of the impact of AI and also crypto too, right? It’s, like, always good to get started, and both sort of have this open-source mindset, so anyone can get involved with quite low friction relative to any other industry or field, which will likely adopt some version of AI technology and crypto over time. So amazing session. Thank you again for all your insights. I always learn a lot whenever you answer questions, especially with your stories, and it makes my mind think in a different way than I would myself think or choose to answer a question. Thank you again for your time, but as always, we will share a recap of this workshop via our newsletter, which is The Digest, and we also will write a blog about this session via our Ueth blog, and that will come out in a few weeks. So keep an eye out on our Twitter for these content pieces and the announcements for our next workshop, as well as community spaces. So thank you for everyone who attended and joined us today, and we’ll see you for the next workshop. Thank you, John, for your time and for all your insight.