Privacy, Censorship and AI: A Conversation with Venice.ai

0:00Today on the AI Daily Brief, I'm joined by the leaders of the new OpenAI project, Venice. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show notes.

0:23When I heard about Venice AI a couple months ago, I was very excited. For those of you outside the crypto world, the team behind this company are some of the most fiercely philosophically respected in the space. These are folks who have long been banners of the principles that underlie that industry. Principles of openness, decentralization, permissionless systems, and individual sovereignty. Given the capacity for AI to be a force for centralization, I was incredibly interested to see what they would do. Well, what they're doing is called Venice.ai, and they introduced it with a piece called The Separation of Mind and State.

0:55The piece reads, It is clear that central powers shall always desire to capture human institutions. Money, religion, education, business, even language and mathematics itself. If monopoly control over God or language or money should be granted to no one, then at the dawn of powerful machine intelligence, we should ask ourselves, what of monopoly control over mind? To whom do we grant license over intelligence itself? The piece continues. All good people care about safety. The important question is safety best achieved through coercive centralization or through open decentralization. Some well-organized voices are certain it's the former.

1:27Observe the enthusiastic alliance between large tech firms, which naturally wish to curtail competition from upstarts, and the state, which has never seen a thing over which it didn't desire jurisdiction. There is another way. Transparency and decentralization are better safety mechanism than appeals to experts in state licensure. The former tends towards iterative improvement, while the latter tends towards complacency and stagnation. So what they came up with is called Venice, a private, permissionless LLM. Venice, they write, utilizes leading open-source AI models to deliver text code and image generation to your web browser or mobile app.

1:59The second point, they say, is about censorship. Every person who has used the leading AI apps has observed the weird, creepy, paternalistic censorship, and it's getting worse. Are you interacting with AI or a multi-billion dollar bias simulator? We don't believe they continue. The thoughts you develop in your mind are our business to regulate and censor. It follows that we don't believe the thoughts you develop with the help of a machine mind are our business to regulate and censor either. The solution is not to force tech companies to act in certain ways. the solution is to build alternatives and humbly offer them.

2:27Venice is joining as the newest sponsor of the AI Daily Brief. This bonus episode is a conversation with Eric Voorhees and Tiana Baker-Taylor of Venice about where the project came from and how they see the current shape of the AI landscape. Despite the fact that these folks come from the world of crypto, I don't believe you have to be the least bit interested in cryptocurrency to be interested in what they're building with Venice. So let's dive into the conversation and hear about what a different approach to LLMs could look like. All right, guys, welcome to the AI Daily Brief. Eric, Tiana, how are you doing?

2:57Doing great. Hey, Nathan. Yeah, great. Thanks so much for having us. Yeah, no, I'm super excited about this. So, you know, as all my listeners know, you guys are a new sponsor of the show. But I think that what, you know, I was mentioning before this in the introduction is that you guys are both people that I've known for a really long time. The project is one that's super exciting. It's the type of thing that when I was doing marketing consulting full-time, I would have been beating down your door to actually try to work with you guys because I think it's an important and a really interesting project.

3:29So what I wanted to do with this conversation is give a lot of folks who I think don't know you guys yet the chance to learn more about you and where this company comes from because its roots go back a lot farther than ChatGPT launching in November 2022, I think. But so let's get started maybe with just some personal introductions, you know, who you guys are, and then we'll talk about how Venice came to be. Yeah. T, you want to go first? Sure. So I am a reformed banker, spent the majority of my career in financial services. And about eight years ago, I noticed that a lot of the projects that I was working on that were very focused on financial inclusion weren't going anywhere.

4:17And it was quite frustrating. And so one of the alternatives that was really coming into its own was cryptocurrency and how crypto could be used for people that don't have access to bank accounts, that have to walk miles after getting paid in cash home and are targets for nefarious actors and how crypto could solve some of those problems. And so that's how I originally got interested in cryptocurrency and started my career full time in that. So I've spent a lot of time working with policymakers and a lot of time with companies building actual interfaces for people to be able to purchase crypto and trade it and invest it, etc.

5:03So still kind of in that financial services space. And recently, I had been speaking to Eric for a couple of years. He was definitely someone in the industry that I looked up to and admired and said to him, if you decided to do a new project after he had ended his last project, that I would love to work with him in the future. And so simultaneously, I had been really kind of doing deep dives into some of the generative AI options that people were using for different types of efficiencies and was a little concerned around the level of censorship and the way that my data seemed to be kind of triangulated with other types of services that I was using, like search, for example.

5:52And simultaneously, Eric called me and said, hey, would you like to join me? I'm going to build an open source generative AI app. And I just thought it was serendipity. Amazing. Yeah, so I come from the crypto world. It's been, you know, my hobby and passion in life for 12 years or so, 12, 13 years. and fell in love with it as a way to solve what I felt is the most pressing issue in the world, which is the debasement of currency. And really through that learned about permissionless systems and trustless systems and open source software. And it really like led me down a path of just a new way of thinking about good and bad software.

6:42And so that's been my whole world for a while. I built a couple of companies in the crypto world. Satoshi Dice was the first one. It was like a Bitcoin casino back in the day. It was half of all the Bitcoin transactions in the world up through 2013. And built a company called Shapeshift, which was an exchange for digital assets that didn't hold customer money. So I've always cared about protecting people through the way that software is designed. And so, you know, after after Shapeshift, these last couple of years, I've gotten into the AI world just as a hobby because, you know, who hasn't? It's so cool and started realizing that like it's actually heading into kind of a dystopian future if certain things aren't changed.

7:28and primarily I wanted to make an open source app that allowed people access to very powerful AI technologies without centralization and censorship. So that's where the idea for Venice came from and we can dive into that whenever, but that's my background. Amazing. So I think actually, Eric, it's useful perhaps for folks who don't know the history of shapeshift to, you know, part of what made it interesting as an observer from the outside is it was sort of constantly moving towards greater decentralization throughout the course of the project, you know, sort of finally ending up in this state where you guys kind of maxed out how decentralized the system had been.

8:13And I feel like, you know, to the extent that that was the, you know, the immediate predecessor to Venice, it might be useful to sort of get an understanding of how you're thinking about software design and system design had been evolving sort of at the end of that journey moving into when you started thinking about AI? Yeah, I think maybe a good analogy is like things like mathematics or language are decentralized technologies, right? No one controls math. No one controls language. These are tools that all people in the world can use openly. They can use them differently from each other. They can change them with time, improve on them.

8:57And because these technologies are open and accessible to everyone, you know, without borders, humanity is better for it. And it doesn't really matter like what any politician says about language or mathematics, like anyone can engage in these things. And I think that's very important for humanity. Something like money, I feel like belongs in the same category. And so that's why I cared about decentralization in the context of cryptocurrency. And when it comes to AI, we're talking about seriously powerful intelligent systems. And I think all sorts of good and bad consequences will come from this.

9:34But one thing that I feel very strongly about is that this technology is so powerful, it really needs to be democratized. Like we don't want to live in a world where one or several large central companies are providing, you know, the AI services for 90 % of the planet. So that's really like the genesis of Venice. Let's talk about Venice and, you know, let's try to give the listeners a sense of exactly how it's set up and what makes it different or how you guys try to design it differently sort of from the ground up. yeah well so the the elevator pitch is it's like um a chat gpt app and it just doesn't spy on you and it doesn't censor the information um that's pretty pretty simple pitch right and this is basically something i wanted for myself you know like i i'm a big fan of chat gpt one of the most incredible tech products ever built um and yet everything that you talk about with chat gpt or any of the other leading ones.

10:39It's all being recorded in one central location where anyone at that company can read everything that you say to the machines and everything that's said back to you. AI systems in the future will be able to read everything that you wrote and know you on a level that is perhaps more intimate than you even know yourself. And so I just wanted to engage with these AI systems, but I didn't want that degree of privacy taken away from me. But I didn't really know what to do until I learned about some of these open source models. You know, I kind of assumed that the leading models from Anthropic or OpenAI were so far beyond anything in the open source realm that it would be silly to try.

11:23And a few years ago, that was absolutely true. But the open source models have become highly competitive. and for most use cases they are roughly as good not necessarily if measured by like certain scientific studies but if you just ask someone to compare two systems that may interact with it casually you know they both feel in the same week so when i realized that i was like wow okay let's let's put an app together that lets people have the experience of interacting with AI, you know, text, image code, but not at all the Orwellian stuff. When it comes to how you think about Venice as a product, how, or maybe just the category of product is a better way to think about it.

12:07What is your bet on how consumers are going to make decisions about which of these applications to use? I mean, inherently, you're sort of making a bet that for a lot of use cases, open source is going to provide close enough to parity that it will be other factors that matter to people. But when you're designing Venice, obviously there's sort of this core of not spying on you, this core of non-censorship. Are there other product decisions that you guys have made to try to make it more like the experience that you would have wanted out of an LLM, out of a chatbot? Yeah. Well, from the privacy angle, it's a tricky thing because everyone says they care about it, but very few people will actually like take action to use something differently for the sake of privacy.

12:54Um, so, you know, I kind of knew that like just saying, Oh, we're, we're a private version of chat GPT, like that'll get some people, but it's not really going to be, it's not going to be emotionally compelling in the sense that to change people's behaviors. Uh, but what we're finding is that the censorship side is actually something that people connect with really quickly. Like anyone who's used chat GPT has felt the paternalism of, the answers that come back to you or don't come back to you or come back in a way where you can tell like that's not the machine replying to you that's some human or some committee that doesn't want that to be discussed uh everyone's had that experience and i think most people just tolerate it so when they try venice they have some of these interactions and that's not present and they immediately see it and we get people commenting all the time and like just so like just grateful that they can have these kind of interactions without that paternalism so that um the lack of censorship i think is what people will feel immediately um and then it's just about you know making sure that we are uh roughly as good as the the other models right in terms of images in terms of text um so we we think we're we're in that league and uh we also have built it in such a way that the speed of response is very quick um so everyone that uses it just like sees how fast and easy they can have an open source LLM and sees how powerful it is immediately.

14:16I think the other thing that I would add on the design aspect is it really serves multiple audiences in a very simple and sleek design. So it provides access to open source LLMs that anyone who is familiar with AI and familiar with open source software can go on to Hug and Face and really drill into how the models were curated and fine-tuned and trained in the first place, right? And so all that information is there completely transparently for somebody that knows what they're doing or someone who is curious and learning. Equally, the app then delivers an interaction with these open source models that you could potentially run on your own machine if you had the GPU power to do it, but it delivers it in a way for a mass market audience.

15:07So it's very simple to use. You don't have to toggle between an image generator and a chat generator for text. It's all sleekly embedded in one line. So you can very easily move between those two types of interactions that you might have. Equally within the image generation, there's a number of incredibly cool and every day it seems like something, you know cooler than what came out yesterday comes out today that is allowing people to really play with imagery and art and video um and some of the things that you can create are extraordinary but you kind of have to know what you're doing right you need to understand lauras and you need to understand how to prompt the the lm in a way that that's going to produce the image that you're looking for so um what's interesting within venice is there are kind of pre-programmed um different styles that you can generate images from.

16:02And I think for the new user, it's very accessible. So there's a balance between it appeals to a sophisticated user and you can really kind of dig into it and use it in a way that is more advanced and it equally caters to somebody who is brand new to AI. And I think the idea around, you know, how are we going to change behavior? I think we're in a very different situation today than we might have been with money, for example. We all have learned behavior that has been ingrained within us since childhood. Generational knowledge around how we save, how we spend is distilled to us from our parents, where AI is brand new.

16:45So I think that if people understand the differences, then they are more apt to make an informed choice. And I think that's where some of the privacy things come into play. Maybe that's not the first thing that they're attracted to, but once they learn the benefits, I think it becomes an even more attractive value proposition. It sounds almost like part of the product thesis is that the productization of LLMs in terms of making it a better user experience right now or up to Venice has only been available with this sort of at the cost of censorship and privacy, right? So the companies that have offered an experience that is kind of user-friendly and consumer-friendly have done so in the context of a particular set of restrictions around what people can and can't do and how they can and can't interact with it, as well as these sort of privacy concerns.

17:43It sounds, though, that you guys are bringing a product perspective to this, where you're saying, look, sure, you could use, if you are comfortable running an open source model on your computer, great, you know, but the vast majority of people are not going to be able to do that. And so there needs to be a sort of something that has the same level or even greater user, you know, usability, user friendliness, but without the costs of sort of, you know, data loss or, you know, just censorship more broadly. Yeah, people are always going to choose the easy product. And we can make judgments about that for better or worse, but that's clearly what people prefer just empirically through their own behaviors.

18:26So we knew with Venice, like we have to make something that's as easy or easier to use than a chat GPT. And if we can accomplish that and then give people privacy by default and a lack of censorship, like we'll be on to something. So, okay. So let's, let's zoom out a little bit because I'm interested in, and sort of broadening the conversation. So if you look right now at the sort of leading contenders when it comes to this category of LLMs and chatbots, there are a couple of different theses for what will ultimately win the market. On the one end of the spectrum are the folks who are clearly focused on being the state of the art, right?

19:08And so this is the chat GPTs of the world. I think OpenAI is almost their entire premise for how they're going to win this space is being ahead, right? And it's that sort of race for capabilities. On the other end of the spectrum, big tech is clearly making a bet that is saying it's actually about distribution and existing installs and how you plug models into that. And so they're aggressively leveraging their existing user base with effectively kind of good enough approaches or being willing to plug in other people. You know, Microsoft sort of sits somewhere in the middle of this being as close as they are to open AI, whereas Apple with Apple intelligence is just extremely focused clearly on the distribution and the installs.

19:53It feels like Venice is trying to carve a third path a little bit, which is, you know, has a similar philosophy of for a lot of different types of use cases, open source and sort of, you know, the state of the art, the pure state of the art is not the required condition. It's being, you know, sort of good enough, far enough along, but that there is a market, an opportunity for the context in which that software sits in terms of what people can and can't do. I mean, is that how you guys see it as this sort of, you know, a third path alternative that you're trying to carve out? Yeah. So I will give you an example of this in principle.

20:35So the image models on Venice are not quite as good as the image models that you'd use on Mid Journey. And the LLMs that you're using on Venice are not quite as good as like Claude Opus. However, if you go to any of those services and try to create an image as innocuous as Trump and Biden having tea, like you try to generate an image of that, you can't do it. these services simply will not they refuse to generate an image as innocuous as as two politicians having tea together um so we don't need to be like necessarily the cutting edge we just need to like allow people to actually use the system for things that they want to do um so that's that's our niche and uh just providing like open access to machine intelligence and these things are moving so quickly that like you know the state of the art is only maybe a couple months away from any open source model.

21:32So if you can deal with being a couple months behind the curve, but you can actually access everything you want to access and use it freely and openly without being spied on, I think there's a large market for that. And I think too, again, if you look at how people are using this technology today, it's not so much a third lane, I think, of users, but if you are a graphic designer or you're a documentary filmmaker, yes, this product isn't probably going to be fit for purpose for some of that more advanced programmatic stuff. But I would argue that the majority of kind of everyday people use chat LLMs almost like in place of search, right?

22:17So you're able to ask a question and get an answer without having to kind of sort through a whole bunch of possible answers, right? And so that is incredibly more efficient. If you're just trying to make a funny picture of a dog or capture a moment or recreate a moment, then this is very sufficient for doing that. So I think that they're like with any type of technology, it depends on what you're using it for. And over time, obviously, we will be enhancing the product with new features and capabilities as they become available and certainly as we have time to build them. But I don't think that the majority of people are writing, you know, PhD dissertations and, you know, creating documentary films with AI.

23:11It's kind of a niche group. And I think that there will always be companies that will cater to those niche groups of people and deliver the top of the, you know, state of the art technology for those particular purposes. Yeah. Yeah. I mean, listen, I think that to some extent from where I'm sitting, it's as simple as, you know, if AI were to stop advancing right here, I still think it would disrupt basically every industry in terms of how people did things, right? And that is a very strong argument that huge swaths of use cases are, you know, do not require the pure farthest sort of state of the art to be extraordinarily disruptive, which creates an opportunity for people to make different types of decisions within the market of options around which of these things that can deliver comparable results are they most interested in using, right?

23:58I don't think that most people are looking at avals when they're deciding which of these tools they like, you know, better to, you know, to the extent that they're even making that decision, they're just seeing which one feels more correct, right? You know, which one feels closer to the tone or, you know, whatever they're going for. So I do, I do think which one will even answer you in the first place. Right. And I think that that, you know, so let's actually talk about how you guys, what your perception of where the discourse around this is, you know, obviously, there's going to be some set of people who are extremely strong and loud about these types of issues.

24:32Obviously, Elon Musk is trying to raise a flag, you know, for some of this, but I think a lot of people are, you know, one, he's a Rorschach test that mostly anything that comes out of his mouth, you know, says more about whether you like Elon or not than it does about whatever he's saying. However, he is, you know, I think when he's talking about these things, part of the reason that they're, you know, that not everyone is listening is, is there's a perception that is sort of being self-serving with Grok and that he's just sort of got an ax to grind with open AI. But, you know, have you guys, I guess, witnessed an evolution in how people are discussing this?

25:07Are you seeing it get louder in terms of these questions of censorship. And, you know, when you think about who the market of people who are receptive to this message right now are, how do you think about it? Is it a group that's sort of defined in any knowable way? Or is it just sort of an ever expanding group that, you know, gets kind of bigger and bigger the more that these issues come up? Yeah, I think that people are noticing it across the board. There are obviously privacy evangelists that notice these things first, Right. And and they're very open about it. Edward Stodan is one of those people that, you know, very often will send out a tweet when something happens in the news and he's like, this is crazy.

25:49Are people paying attention to what's happening? And it doesn't necessarily matter what the specific instance was. But Edward Stodan is kind of one end of the spectrum. But I think, you know, we hear from our, you know, mainstream users that they find it refreshing that not only will then answer your question, but equally, it won't tell you that you shouldn't have asked the question in the first place. And I think that that is where people start to notice that kind of parochial patronizing tone that sometimes some of the big tech apps, their models can take. Right. So it's not just we won't answer your question, but it's, you know, you really shouldn't be thinking about this.

26:36And so when technology is telling us not only, you know, the supposed truth, but also curating what we should even be exploring, you know, intellectually, I think that's very problematic. And people do notice when they're being minded, you know, by a parent and this parent happens to be a piece of technology. So I think that's the one thing that has stood out kind of across the mix from people that really care about privacy and that really care about information being censored or free speech to just your average everyday person that is trying to potentially challenge their own views or to learn about a topic that some people consider taboo.

27:19And I think that becomes the discussion. Like, who is the arbiter of discretion? Who decides what is taboo? Does that vary by culture? Of course it does. Does that vary by socioeconomic class? Potentially. And so when we start limiting people's critical thinking and the ability to try to extend their knowledge base, I think that's very problematic and it's really obvious. Yeah, I'll add that if you look forward, you know, just a few years, it's not hard to imagine various governments in your country or in other countries, which puts pressure on these centralized AI companies, such that the censorship and the bias in the models isn't just coming from that company anymore, but is now coming from politicians behind the scenes.

28:16um i think everyone can understand it like if these systems had been around during uh covid uh what the ai could tell you about that virus and outbreak and the response to it would have been highly curated by various government authorities um and so if you trust all the government authorities like you know maybe you don't care about that uh if you're skeptical And if you think that many different voices need to emerge in order for truth to be discovered, then that kind of censorship is really, really dangerous. And when you combine it with like the world's leading artificial intelligence machine learning systems, it can get pretty dystopian.

28:59And so all we want to do is build an alternative so that if that starts happening, you know, people have a different way of interacting with these systems. For what it's worth, there's also the extremely banal reality, which is I have to create thumbnails every day. And a lot of times I talk about Trump or Biden as relates to crypto or as relates to, you know, AI or as relates to whatever the topic. And when I go to try to create an image that has one of those guys in it, I'm not trying to start some deepfake war. You know, I want Trump shaking hands with a frigging robot so I can put it on the YouTube thumbnail.