Kartik Talwar & Erik Voohrees I Private AI With Venice AI I Agentic Ethereum 2025 Summit

[Music] we're GNA do a fir side chat with Eric and uh we're gonna talk about private AI Eric is the founder of Venice a lot of you who've been in this space before know him from the world of decentralized exchanges with chap shift we're going to bring him on and talk about how things have been evolving in the world of private AI so without further Ado let's welcome Eric welcome Eric hi how's it going things are going great um I'm excited to be uh to be having this chat here and um we don't have a lot of time so we'll kind of jump right into it because a lot of questions that came in from uh the community and just discussions and things from our own list and I'll just GNA get started um my my first one to you is uh what kind of made you uh go into spending more more time in AI come out from coming from your background in finance and exchanges and just the early world of decentralization and uh what kind of made you choose this as the next thing good question uh so yeah I was running shape shift for about eight years and we decentralized it in 2021 um and then I basically took a couple years off and just kind of um explored the world and uh tried to you know regain my sanity after after building a company in the crypto space uh and of course like everyone got very interested in the AI tools the chat Bots and the chat gbts um I thought it was very very impressive obviously uh but I have no background in AI so it was just kind of like a interesting Hobby and fun fun thing to play with but as it as it moved on I started realizing that like there's something a little dystopian about it uh and that was that you would have these conversations with with the AI um and I don't think people realize that those entire conversations are just going into the hands of the company producing producing the chatbot so like in in chat gbt's case all of your conversation is going directly to Y to chat gbt to open AI um and also the other weird thing was that the answers would be like censored in weird ways they would be um curated you know manicured some topics were off limits uh and and other things you couldn't talk about it all and then there were like more subtle forms of bias where you're like okay it's definitely not a machine conveying this answer it's some combination of like a machine and a human or a committee or a policy or like something like that um and that started feeling really weird to me and you know projecting that Trend forward you can see like how as the world starts using AI everywhere if everyone is using it through these centralized sources that um you know curate the content based on the policies of the companies and the governments that oversee them I can just see that going really badly so um I realized maybe I could help um why wasn't there a product that was basically like a chat gbt but doesn't buy on you and doesn't answer the answers and that's the road that led to build Ben um so in a way this is obviously very overlapping with kind of what we saw with the world of finance and and you're effectively fixing a problem that uh you kind of or if I were to kind of rephrase a lot of this from FR of my interpretation which is effectively uh limiting the user for what they could do and uh and or adding these constraints and uh this kind of we've seen this entire reason for why our crypto ecosystem exists and you kind of pointed out a similar parallel here um the the thing I guess I'm curious about is does that translate to enforcing things being private from the point of view of nobody should know what's going on or is it more about giving people a choice um and and do you kind of distinguish between the two here and uh and is that kind of what the goal of V is is just say I offer them a choice where I don't have to worry about saying oh I can't answer that for you try again or is it about the fact that like I should be able to limitlessly kind of see see what's going on and and also preserve that I have the full right to play with this Tech however I please and should people should not interfere yeah um on one hand it's about like protecting users and this was a concept that we formulated with shapeshift back in 2014 when we launched it um the the ethos is essentially like to protect people's privacy there's only really one good way to do it which is to not have their information in the first place um and every company in the world has all these privacy policies and they all swear that they take your privacy seriously and um they would never leak it out it would never be abused it would never be sold it would never be you know sent to someone else but of course that's wrong um and it does not matter how how big or small the company is it does not matter how um seriously it takes privacy the data generally leaks with time um sometimes in benign ways some sometimes in just commercial ways sometimes in really bad ways so so really we like the the way you protect people in this case is you just don't take the information in the first place and that's an that's a design Paradigm that that is common in crypto but is alien to most of the world outside of the crypto world where people take things like privacy seriously um people don't understand that concept they don't realize why it's valuable but uh I I felt that if if the intelligence of the world was was moving into the AI Paradigm that uh I at least wanted to try to prevent that from being the sort of dystopian surveillance world that it would by default become um so yeah that's that's basically has this kind of allowed you to take a different stance on what actually maybe to back up like what does it mean to have private AI as as what the the Venice offering is right now and and what's kind of the the version of how do say how does somebody interpret this and what what this means and uh how does this compare to the next alternative which is like the non-private version in terms of things you don't get or do get yeah so like when so everyone's use chat gbt when you have a conversation with chat gbt all of your questions are going to open AI they're stored in their databases and all the answers from the AI that come back to you are also stored by them so they have your entire chat history connected with your pii connected with your credit card your identity all your metadata and and everything and all the data across the different conversations that you have um not just now but into the future you know they will always have that information uh so in venice's case the experience is very similar it's a chat bot where you can talk to the AI um but the conversation is never stored anywhere other than in the user's browser so um Venice never has it never stores it and um that means that it doesn't matter like if we get hacked there's nothing to hack doesn't matter if we get subpoena there's nothing to subpoena um it doesn't matter if we have some employee snooping around on stuff because there's nothing on which to so um that's that's the fundamental difference and does this re does this require kind of a a way you kind of think about architecting this uh service or is it purely the fact that hey others are choosing not to but they can and and it was as simple as that but you realize that not only do we have to choose not to take in all this extra information but it also requires reinking how LMS work or how we expose these things to the the end user yeah it's obviously much harder to do this stuff without storing the data um one advantage that made this possible is that we we as in Venice are not trying to train models so um you know open AI has this ongoing incentive to build the better and better and better models one of the ways in which they do that is that they use the users conversations to approve their model um which is effective it's a good way to make the models better uh but to do that you need all the conversations of all the users so because we aren't trying to build models ourselves we don't have that incentive um that lets us really like treat privacy uh much more seriously we also have a very normal Revenue model where we just charge people for a pro account like a normal subscription and so we don't have any um desire or need to like sell people's data because we just make money when people sign up for the service so we remove those incentives first and foremost um as a start makes a lot of sense um and and uh obviously we're kind of getting uh a little bit ahead of ourselves here because uh you are kind of fairly early into the this just like that the history and the life cycle of uh the company and the and the product but U have you kind of seen um like I guess what's been the reaction from people in terms of uh is this Choice enough for them to kind of switch over and and they're they're really valuing this thing or does this mean that you only end up attracting uh the group that is not using chat gbt or other products because they are basically getting their data uh leaked or or spite on or kind of abused good question uh so we we basically have like two pillars on which benice is built one is that it's private which we've been talking about the other is that it's not censored we don't add any kind of like guardrails to the conversations or prohibit any kind of topics um and and sort of both of those we we bring that in our messaging equally to the market what's been interesting to see is that while most people say that they care about privacy they don't actually change much behavior in order to get privacy and there have been many privacy products in the in the world that just they never become big because of that yep reason um privacy is one of those kind of vague things where like it doesn't matter until it matters and you often don't really know when you've been violated or or how it's really hard to feel the danger of of that um but on the censorship side people feel that very quickly and so what we've what we've found is that like the the thing that people are telling us is so great about Venice isn't the Privacy stuff because they they all like that but what what I can tell emotionally clicks with them is when they send in a question into the AI and they get back an answer that they can tell wasn't censored right like and it can be as simple as like you know tell tell me an offensive joke yeah and our AI will do it um it's not necessarily a funny joke but it can be offensive and when people see that I think it's very refreshing for them they they're like removed out of the sort of paternal normal world where corporations and governments treat people like children always needing to be managed always needing to be overseen all sorts of rules and guardrails needing to be put everywhere and when people use Venice and they have these conversations and they realize that the speech is actually free um and open that's what really connects and uh that's that's the thing that has made people you know cancel their chat gbt subscription over to Venice it's amazing um what is the process of uncensoring this information is it the fact that these things by default I I assume you're using all the open source available models is by default that they are um coming in as is and and the platforms are applying the the security layer on top or or did you have to actually do some other fine tuning to uh it's it's a little of both uh so every model that's trained in the world whether open or closed Source um always has some degree of like bias sens and and some of that is just like what data you know the creators have it used in the set like you have to make trade-off decisions and um classification decisions so every model is a creation of humans and has has biases no model is perfect in that regard um and so the open source models that are on Venice um some of them are more censored than others because they have these kind of things built into them but the other layers of censorship that the centralized companies tend to add which live above the model the rules of what you can ask the rules of what can be answered the the various guardrails um we don't have any of that so what it what it means is the aggregate effect is much more free and open um even though it's it's never perfect like there's always people that ask a question they're like hey you said this was uncensored well I don't agree with you know like that's an asymptotic goal we can't perfect that but um we have a a vastly higher degree of open and anyone who's used these systems feels that pretty quickly I mean we' seen Chach refuse to um reply to some things that you like how is this even uh in the ballpark of uh being a thing that is not allowed and full spectrum too so it's really some of it's really silly like we we um one our little marketing tricks months ago during the election was we made images of um Trump and Biden and then Trump and Cala just in like very benign situations like having tea you know Biden and Trump sitting in in Venice having tea at a cafe um you can't produce that image with open AI it's like against their terms so um I just think that's so Preposterous like when your when your content restrictions are such that you can't create benign humorous images of public figures um you've gone way way too far and so we just use that to show people like uh yeah you can create images of of public people um through benis and and that alone you know gets a lot of attention um one thing we've kind of also seen along the same lines is uh uh this same point of uh kind of the bigger um chat apps effectively getting um criticism for the censorship and you kind of have the the gro saying look we are going to offer you a similar if not better model and we're also going to wind down on some of these card rails um have you kind of been in a scenario where you've either cross compare advantage to to how Croc offering is or the users have kind of looked at this comparison and sort of what what's your take on that side of the world where it still ized offered service they are still going to likely improve based off of the the feedback loops uh but way less to to no guard rails yeah um Gro is pretty good so Gro is definitely less censored than Umi or anthropic or Gemini um so they're they're decent in that regard they're not quite as uncensored as Venice but they're pretty good um but they're not private so they still suffer from the like collecting all of everyone's conversations and all the risks and problems of that side gotcha so so okay so so the value then is to Additionally you also want privacy of what you're what you're uh your interactions are understood um this kind of gets into a bit more philosophical side of things but um how do you kind of think about censorship in general uh and and if there should be a balance and and safety around this um is is it your stand s that it's just a binary kind of thing here which is we're not going to add anything and whatever the model says you going to proxy to the user as is or do you feel like there are definitely some scenarios where this may be relevant and uh there there should be the ability to do something and whether it's true now versus could be true later like how does that thinking evolve over time or or do you think there's a framework that you have on how to handle this thing moving forward yeah I mean philosophically one of my deeply held principles is that um if there isn't a victim then there isn't a crime and this is like a really good framework through which to see the world and so from the perspective of speech speech itself uh never has a victim um if two people are talking voluntarily with each other it does not matter what they say there is no victim in that conversation period they can talk about any anything they want as horrific or or Insidious or dangerous as that speech might be um there is no victim from the words there can be victims from actions taken from words and then sometimes it is very fair and reasonable to prohibit those actions which create victims but the um the speech itself should be uh free purely on philosophical grounds because there is no victim to it so that's um that's obviously woven into our our ethos ofen um just to poke that a little bit I do have other questions what I get to but this is now just a really interesting point um how does the definition evolve when in this case speech translates to a request to an action right I can say hey tell me how do I do X which could be a set of instructions that could be used to harm somebody and in that case is that a free speech case or is that a a there's no there's still no victim there you know um if I ask the thing like how do I murder someone um no victim has been created and uh I might want to know that for all sorts of reasons some of them very benign maybe I'm writing a book and I want to learn like how that's done so I can make my novel a little more effective maybe I'm just like intellectually curious in the topic and I want to explore things you know like the the one of the signs of a of an intelligent and well-formed mind is to be able to explore ideas without automatically accepting or believing them so what that means is to be able to like explore ideas that you think are wrong or that you don't have an opinion about um and not not necessarily thinking that that's true or that you should do it um but again the most important Point here is that any conversation that you have does not create a victim and so um no crime like that be that should be tolerated yeah okay no that's that's a good way to to frame that I I I was looking at it from the perspective of uh how do you kind of back apply this thing so in this case of okay asked instructions for a murder but then somebody commits a murder does that get traced back and uh looked into a postmortem or are we stopping at the interaction as it was happening live which is free speech and and conversation yeah it's just it's an extremely dangerous slope once you leave the actual action right so someone who someone who murders someone um that action is obviously wrong and should be should be criminal uh but all sorts of tools might be used in that murder maybe the AI was used to instruct how it could be done uh but what about the car that was used to drive to where the victim lived right or the knife that was used to kill the person like um someone manufactured the steel that went into that knife someone manufactured the knife someone manufactured a computer through which this person learned about how to do the horrible like all the way down yeah yeah it's like um all of these are contributing factors none of them are crimes themselves the crime is very clearly the act of the murder um and frankly I think if authorities would focus more on the acts which are crimes they might do a better job of actually reducing crime in society instead of policing all these sort of um a priority actions that that may or may not lead to them y that makes a lot of sense um I think we'll we'll shift back to the crypto side of things and uh the question I have is uh would love to kind of understand how do you um make this a lot more accessible and then you kind to have a really interesting hybrid token model we love to learn about that model um and uh sort of how are you kind of bringing the decentralized world into into the equation here yeah uh so my background of course is like crypto crypto crypto and um I've lived and breathed it since Bitcoin in 2011 and um when I started building Venice uh that was a little strange for me because like at its core it's not a crypto product um but I wanted to bring in the principles of decentralization where they made sense and not bring them in where they didn't so um you know Venice is a normal centralized company um we operate very much like any kind of SAS business where people pay for a subscription um but where the decentralization was important for us is in the provisioning of of gpus so um essentially you know like most most companies would just sign up with a bunch of servers at AWS or or similar um and what we wanted to do was build this to be much more resilient in case policy um changes in the world like we're trying to build Venice to be [Music] um which me it has to be able to Route Around uh all sorts of different types of friction and restrictions and so that meant that we needed to not build it with the Assumption of any particular GPU provider we needed to build it such that we could use a decentralized set of gpus so um that's one area in which we've done that um and then also crypto of course is great for payments so like um it works really well for that so you can pay for Venice from anywhere in the world uh using crypto and we don't have to deal with like oh does your credit card work in Nigeria kind of thing like anyone with crypto can pay and that side of the decentralization works great too um and then we just recently uh released our token a little over a week ago and uh that permits anyone to access the API without paying per request for the API um and in this case that was a good use of a of a decentralized asset as well because um that API is geared toward AI agents um who are you know incorporeal actors they don't have a credit card or an ID they need to be able to access an API and and for like agents crypto makes a hell of a lot of sense um so we we've tried to bring these things like in in where they make sense but we're not decentralizing elements of our stack that don't need to be no and and that's obviously the the sensible way to do it and uh I think you have the benefit of hindsight and just zooming out version of what we've seen so far happen and just improving from uh what we've seen um it's not been a lot of time but uh has there been kind of interesting uses for for the API that you've seen so far on the platform and uh or kind there are some interesting projects that you can kind of talk about that uh are leveraging V yeah so um so yeah the API is fairly new and we're just starting to see in like after the token launch we're just starting to see people start to play around with this thing um but one of the cool things is that we we have a concept in benis called characters and any anyone who's used character AI is uh familiar with that but um we made the characters function available through the API so what that means is that people can build character-based applications with our infrastructure uh very easily over the API and um it's been fun to watch all the different requests that come in about like different types of interactions that people want with these characters over the API that we hadn't we hadn't thought of like um creating special API keys that like have a countdown timer on them or um characters that are available on a certain key but not on another key so lots of like kind of nuanced stuff of people building on the on the API but it's been that's just the natural part of like exploring uh how you actually also it's also humbling to because you always think you know how something will be used it's never yeah it's often different and you got to really watch what people actually do instead of what you think they'll do U good lesson for anyone in business does it make it harder if you don't have a a good uh feedback system built in because you're not learning from uh usage data um yes and actually this was this was kind of a struggle in venice's early days so we Venice launched in May of last year um but the the token just launched a week ago um and with that we we created a Discord to like start building that community and other than the occasional like customer service emails which were you which were useful we didn't have this kind of real-time interaction with our with our hundreds of thousands of users before um and so now it's just like drinking from a fire hose and they're all in there talking about things and you know like many of the ideas aren't great or we wouldn't do but then are some real real nuggets in there um and so we're we're all as a company just trying to adapt to like now being able to talk in real time with all of our users that's awesome um I W to close off by a final question which is obviously a year of um of kind of giving people this amazing Choice a week of kind of the API and the token that we're launching um anything that you can kind of talk about on what we can look forward to for the rest of the year in terms of improvements things that you want to tackle or just the choice that people will have and um we can kind of end off in in that not um yeah we're just trying to roll out features as as quickly as we can there's like so many cool things to build with AI now it's really like an open open territory and there's so much stuff so um we're definitely moving our characters to be more like agents so that's like a very natural path our characters right now are just sort of personalities shaped by the system prompts that you can interact with and they're they're fun and they're compelling but they're not yet agents and so and is a definition of an agent here that they have a notion of memory or just the ability to chain multiple actions uh either and and I think the the thing that makes an agent most agentic to use that fancy when it can do something somewhat autonomously you know like like um a chatot you're always it's always sitting there and when you ask it something it responds but an agent should have some degree of uh you know deciding to to take an action and doing it without an input having come in to immediately spur it um and yes it needs some degree of memory it needs to be able to interact with the with the web and with different tools but it's that sort of like impetus toward behavior that that I think makes something agentic that's a that's a good definition I like that um awesome well Eric want to thank you so much for uh giving us uh your time and sharing about everything that's been going on Venice and uh can't wait for a lot of people to build it build on top of the API and just use uh a way more uncensored lolm yeah well thanks so much for having me on good good chat [Music]