Is Open Source AI Winning The Enterprise? | BasedAI CEO

Transcript

We are live for our weekly premiere AI show, AI Supercycle episode number 33 powered by our friends at Near as you can see here here at the event powering their Iron Claw product as well as near.ai and their inference product with the work they're doing here with the Grenada government. That will be talked about soon. We are here with the Venice co-founder and now recently announced based AI founder Tina Baker Taylor. Tina, welcome to the show. >> Thanks, guys. Thanks for having me. It's a awesome conference. Fun to see you here.

>> Yeah, absolutely. It's a good crowd. How's the island treating you? >> Yeah, the island's great. It's been been a little rainy the last couple of days, but it's a good for us to be able to focus on the conference and >> Yeah. >> Well, let's not bury the lead. The folks want to hear about Venice's recent success. They want to hear about based AI. Um obviously this kind of this niche of decentralized private AI, specifically private inference is really really gaining traction as the kind of world wakes up to the potentially dystopian future that we're snowballing towards.

I would just love to hear your high-level kind of foundational viewpoint on why this is such an important uh challenge to tackle. >> Yeah, well you had Eric on a couple months ago. So um I'll probably echo a lot of what he said. I think when we founded Venice the promise was that we'd both been using AI really um finding a lot of advantages um and just the magic of what it was able to do. But quite concerned about some of the things the model would say, some of the things the models wouldn't do, um and equally like where is all this stuff going, right? Um and both of us coming from a crypto background just didn't really feel the trust-me-bro vibe, right? And thought I would rather not have to trust you. I

would rather just be able to access this magical tool and not be worried about um where my data was going and how it might be used against me in the future, quite honestly. Um, so yes, I think 2 years and 2 and 1/2 years into Venice, um, the platform has gone from strength to strength. There's over 2 million users, um, launched the token a year ago, um, which I think people are really starting to understand the value proposition of being able to buy an asset, stake it, and then receive that inference, which is private or end-to-end encrypted.

There's a variety of options, um, on Venice. Um, or equally access a frontier model in an anonymized way, so it might not be completely private, but at least you are abstracted away from your prompt. Um, to be able to access that tool and that inference in a perpetual way. So, you're making an investment once, um, and then you're able to stake that and use that inference, um, every single day without having to keep paying for it. Um, so I think that has been a really unique product market fit, bringing crypto and an AI together in the space, um, where it's meaningful.

It's tangible. You can use it. And I think that's been a challenge for crypto for a long time. Like, what is this good for anyway? Um, and AI really is a obvious use case. Machines know how to work with other machines. They can transact with each other. We're seeing that advent and the rise of agents. Um, so Venice kind of solved that challenge for the consumer. Being able to access AI in a private way that, uh, has a lot of those dystopian concerns removed. Um, and so for me, looking at this from a business perspective and an enterprise perspective, we've seen kind of the the rise of agents almost overnight. I'll be honest, it happened a little faster than I expected that it would. Um, and the

ability for those, uh, agents to undertake tasks on their own and act autonomously, and couple of that with, you know, primitives like XPRT 2 and stablecoins, now they can not just do it you um, do the research for you, but they can execute the task cuz they can pay for things. And so, as you said, the snowballing effect, I started become really worried about businesses just kind of sleepwalking into Claude code. And then for me, I think when Pete from Open Claw was aqua-hired into Open AI, I was just like, "Wow, okay, we have these cool open-source tools that people can use to deploy engines on their own, and now they're just being co-opted by these black boxes." And I was like, "All

right. So, Venice has solved this problem for consumers. How do we solve this problem for people and businesses and enterprises to be able to deploy agents to increase their workforce, to automate tasks without being locked in to more big box or black box providers?" And so, that's what Base AI is trying to solve. >> Amazing. And so, just kind of double downing on the success of Venice, you guys are doing about 60 billion tokens per day in production right now. It's been a very steady growth curve. I think Eric put up a chart today about it.

Where are you seeing the most demand come from in terms of that token production? What types of models, what types of use cases are you really seeing for this kind of private inference, private AI? >> So, I'm not day-to-day at Venice at currently, but the demand is coming from the API predominantly. And a lot I think there's a balance between agents, human-driven agents, some autonomous agents, but equally anybody that is trying to use a significant amount of inference and scale something. So, there are small businesses, there are individuals that have different maybe hobby projects,

but I think that people are now starting to deploy those agents personally to undertake all kinds of things. Manage my calendar, book my travel, um and some of them are maybe what we consider mundane tasks. So, it's it's simpler to or maybe not simpler, but you're probably more likely to start with these kind of small things that don't have like a material impact, right? On if something goes wrong. Um but we're also seeing people use agents quite a lot for um digital asset or financial service transactions, right? So, these agentic wallets are able to actually transact for you. Um go and find me the best yield on USDC and use this wallet to stake 30 grand,

right? Um and a company that Base AI uh recently acquired um Warden. So, Warden Labs um has built out an entire uh agent wallet infrastructure that will conduct all those tasks for you. So, bringing Warden into Base AI is going to enable businesses to also have access to an agentic wallet to be able to transact and undertake not just like give me advice, but actually so build me a website. Fine, build your website. But if you want it to build you a website and then actually deploy the website, like go get the domain, get it up and running. There's transactions that have to happen in there. Um and so agents are able to do that um obviously with some

human supervision. >> Right. >> Right? Um and so, I think that across the open-source space, um I was just on panel with Near, and they're seeing basically the same kind of usage, right? People are starting to automate some of the mundane tasks in their life. And I think that those tasks are going to become more complex and more important as we start to gain more confidence. >> Yeah, definitely. >> So, there's two [clears throat] um there's two tracks here. There's the privacy track, and Venice is you know, proficient at that. Uh they've been able to essentially turn every model private. Uh and then there's >> Well, they give access to um some of the frontier models, like you can access for Grok. Um and it's anonymized. Some of those models are anonymized. So, be

careful when you're scrolling through, um, some of them are completely private, um, you know, TEE enabled, um, which is, you know, a trusted uh, execution environment, um, and some of them are just anonymized because they're they're it's a pass-through into that. >> and actually I I saw some people pushing back on that saying that, "Look, Venice isn't fully private. Like you can't call it a private AI." But >> It depends on the models. That's correct. >> So could you just further ex- So there's an anonymized version of models available and then fully private versions?

>> Yes. >> Okay. >> Yeah. So you can, um, what Venice is very cool, um, in that there are all of the open-source models, um, and Venice can run those through very specific machines, right? That are able to um absolutely kind of guarantee and secure that privacy. But if you're if you're using Claude, for example, um, and you're going through the Venice architecture, then you're going through basically a reverse proxy and all of your prompts are stored on your machine locally and they're not, you know, stored on on the the GP at the end, right? Um, but in that case, you uh, your prompts aren't attached to you because within

Venice, like you are abstracted away from like you can use Venice anonymously. >> Right. >> Um, and you get a limited amount of of usage. If you don't sign up, you don't have an account, it is still available for you to use. Um, so you can be anonymous and those those prompts would be anonymized, um, but the model cannot be guaranteed private because it's going through uh, Anthropic in Claude's case, for example. So just but but it's very clearly noted. So when you're selecting your model, it will tell you right there um what what environment it'll be executed in, what your prompt will be executed in.

>> Got it. And so, this [clears throat] is the the privacy track and it it is model dependent to some extent. And then there's the the open source side of this, which is more where Based AI is focused. And there's also several open source models that are available on on Venice and you do have some more customizability and privacy features with those. >> When you're looking at the state of open source in AI today, it feels contradictory because typically the, you know, US has been steadfast and forward-thinking when it comes to open source models or at least the open source movement in general when it comes to software, but now we have China leading the way when it comes to open source models. I think this is fundamentally disconnected from the way that people typically think about software development globally. How do

you see the progression of open source models? And do you think that there's a tangible path for an open source model to compete with these large models of, you know, the the largest private companies in the world? >> I do. I do. Um and I think that that is going to be based on, you know, the wrappers that are built to use them. So, for example, uh the first product that Based AI is is launching is called HireBase. Um HireBase will allow you to hire um any AI worker across a variety of different um disciplines. So, they might be a sales person, a marketing person, content, um finance, uh HR. So, if you don't have that person in your team

already, let's let's say HR for example. You need some policies and you need some payroll stuff every once in a while, but do you have the um resources to be able to hire someone for a hundred grand a year? Maybe you don't. So, you can augment your, you know, CEO or CFO with a agent that is designed um and solely curated to perform those HR tasks. So, if you're using Claude or you're using GLM-5, um depending on the task, when you look at the performance of those open-source models, there are a number that are equally as performant, right? Now, that is not going to be true across every single task, right? There will be some models that are a bit

better for coding, some models that are better at maths, some better for finance, some models that are better at creative writing, so they'll be better for marketing activities. Um, so some of that is understanding what those models are good for, and pairing them with the um, agent or the virtual worker, so that that worker can be the most performant in that particular use case. Um, and so with Fire Higher Base, um, that's what Base AI is trying to solve for, either augment your team, so maybe you need uh, maybe your marketing manager needs some help with um, performance marketing, and they need to better understand where intentioned buyers are coming from, right? So, they're placing the ads, they're getting all the dashboards, it's

a lot of information. You can use an agent to be able to put that into context for you and say, "Well, actually, you should be spending more money on AdWords because you're getting more intentioned buyers from there." Right? So, that's not replacing a human, but it's making a human much more effective so that they can spend that time that they were spending on reporting on something way more value added in for that business. >> Mhm. >> Um, so in those cases, we're able to um, identify the best model for that type of activity. Um, and the key thing here is these open-source models are just significantly more or significantly less expensive.

>> Yeah. >> And so what we've seen is people, you know, they get their open claw out and they hook it up to their Mac Mini and away they go and their agents are doing all this stuff and it's really cool and then, you know, Anthropic decides they're going to cut you off, right? Or you're they're going to completely throttle how much um, inference they're going to give you. And then they're just going to decide the agents can't use that anymore. Um, but you can sign up for this other plan, right? And now you're paying, you know, per token. And then people get a 5,000, 10,000, 20,000 cloud bill and they're like, "Whoa." Right?

Um so, the the margin to be able to deliver nearly identical, not always exactly as good, but nearly identical outcomes for a significantly cheaper price, I think it's going to make this technology way more success or accessible to businesses, big and small, right? >> And and one of the things that you mentioned on on Based AI is that almost 70% of open model of enterprises are adopting open source models versus closed source.

You know, again, counterintuitive because you think these are institutions with bigger budgets, they can spend more on tokens, they can afford, you know, the the high Claude bill, but that's not what you're seeing on the ground. >> Now, and and I think a lot of that has to do with just data, right? So, the whole trust me, bro, analogy that I gave earlier. And Claude is continually learning from everything that you put in, right? Which makes it a little bit more smarter, a little bit more performance, the context is is sometimes richer. And so, that's the key difference.

The open source models are published on a date, and that's it, right? So, if you are a business and you're really concerned about your data being used to train models, then you might be putting some client information in there, you might be putting you know, your commercial strategy in there. You want it to come up with, you know, a a marketing strategy, go-to-market strategy. Like, that's really proprietary sensitive information. And if that's going to be used to train a model or potentially get caught in something like we've had a number of court cases in New York, for example, where the AG has said, "Okay, I I want all all the inference from this period of time." And you might not even be a party to that to that action, but now all of your inference is is caught

up in in discovery, and you know, things get leaked and things get out, and you weren't even like a party to this lawsuit. It's kind of nuts, right? Um so, if you're using open source models, there's it's just not possible to do that, right? The the model is done being trained. So, it can't use your data um to to further train. Now, what we can do is take those um open source models, especially in an agent context. If I want an agent that is designed specifically for I want it I want it to be a finance manager, right?

So, I I get the model that's the best for that, and then I can fine-tune it with specific information. And so, for example, it might be that I'm going to build an agent suite for a particular business. Let's say it's a fund, for example, you know, we're here in Bermuda, and that fund will then give us their data that we can fine-tune that model to be very responsive to their finance team, right? And so that we can do that in a way that again, there's no data leakage because it is defined to that client. It's defined to that agent.

It's defined to that instance. Um and you just can't do that with with a cloud because it's always just going to be hoovered up into their training. >> Maybe this is a dumb question, but with open source models, because they're they're released, they're done, they're not taking that data, training the model any further, why is privacy so important when you're using an open source model? >> Um well, so there's there's always the potential um and this is why inference tools like TEE are important. Um so, the the prompt goes to the machine, right?

And the machine does what the machine and then it sends it back to you. So, there's a couple points along the way where that could be intercepted, right? Um and somebody at a data center could, you know, potentially break into the machine and get access to that data. And so, something like TEE prevents some of that from from happening. Um, and that would happen in in any case, right? I think the open source is especially helpful in that you're not going to be worried about your data being used for training, um, which happens a lot, or your data being used to um, intercept your thinking based on advertising that it might push to you.

Um, the other thing that I think worries me a little bit with the non-open-source models that are continually evolving and you don't know at what state that is evolving. Um, the responses that you get, like we've all had the response where like it's just flat out, I'm not going to answer that question for you, which is a bit weird, right? You're like, why? Um, but I think what is, um, most concerning is the nuance of the response that you get that might be, um, a version of the truth, but whose version of the truth? So, you know, what's in the system prompt? What has Anthropic's board said you can uh, say about, you know, you you see

questions, I mean, I don't I I I want to be demonstrative, but like, you know, is a transwoman a woman, right? And and and a different model will give you a different answer, right? Um, and some of that is because of, you know, a safety feature or a board decision, um, where open-source models tend to give you they just answer the question, right? Um, and again, it it it depends on, you know, Venice is very committed to not filtering that intelligence. You're getting the intelligence back, um, as it was intended without any type of curation on top.

>> And I think kind of, you know, extrapolating on on that point, this is why the blockchain rails become important for >> Absolutely. this kind of primitive, right? Like we've been searching around for a really good use case and idea for like blockchain and AI, but it feels like this kind of censorship or this uncensorability becomes the number one primitive and then also like Near's working on like private enclaves in the cloud, so that becomes like another key point. >> Verifiability, I think is going to become important. >> And so maybe you could just kind of further define the the the core vision that you have for the intersection of blockchain and AI as it exists today, you know, you know, backpacking off this open source versus closed source

conversation. How do you really see whether it be tokens, like actual assets, um the actual blockchains themselves, the privacy, the verifiability? What is kind of like the match made in heaven here for this convergence and intersection and kind of like how how how has this vision changed? >> Yeah. So I spent the better part of my crypto career trying to get governments and regulators and policy makers to understand like what is crypto good for anyway, right? What's the use case? Is it This is just a speculative asset and no, it isn't and these are all the ways that this technology can improve the transfer of information, the tokenization of assets, the ability to create capital, like all of the things that we talk about all the

time, right? Um and for the most part, until very recently, they just weren't buying it, right? They just they just were not buying it. Um and so I think for me, the advent of AI makes it very clear to me that maybe crypto wasn't actually for people to use. Maybe it is a programmable, um you know, digital ready money. Could be tokenizations of funds, it could be I mean any any anything that you can tokenize um is now machine readable and is now able to be used by the machines and transferred by the machines. And so if you could take the humans out of that and I think that's where agents become really incredible. And they will provide

that kind of service layer where it's been hard for us to to to make these things work in our operably and I'm gave the analogy earlier on my panel that, you know, a few months ago I sent a whole bunch of tokens on the wrong chain. Don't know where they are. Never going to get them back and I know what I'm doing. So imagine the people don't know what they're doing, you know, how scared they are to use these things and and move big chunks of money. Equally if you could extrapolate that up to like a business or, you know, a sovereign wealth fund, um, you know, they have security concerns, um, they have regulatory concerns and so for me I just think that intelligence layer integrated into a verifiable blockchain infrastructure allows you to

tokenize anything and move it around and I think that AI will provide that that service layer where we're just checking what the AI is doing, but we don't have to worry. I mean, Near has Near intense now where you can just send it and you don't have to be so specific of what chain you're using because the robot's going to pick it up and is going to make sure it doesn't make a mistake. >> Yeah. Yeah. You know, I'm curious specifically about the way that the Venice token model works. Do Do you think that this model of tokens of staking the VV asset to to have inference available and ability to prompt? Do you think that's scalable to to the extent that where you guys are trying to take Venice and kind of like where Eric sees this going? Like is that

model sufficient for, you know, hundreds and hundreds of billions of tokens being produced each day by users or do you foresee a different incentive mechanism needing to be implemented there? >> Yeah, so I think what's been interesting with Venice over the last 6 months is people have really started to understand the value proposition of the VVV token and and by um by extension the DM token, which is actually the token that gives you the inference and you stake VVV to be able to mint the VV to mint the DM, um which is what you're using for inference. So, it sounds a little complex. It It isn't

quite as complex as it sounds, but >> Um and so uh there is a significant amount of the VVV in circulation staked. And I think that is a testament to the um that people believe that there is value. Um and then we're looking at how much DM is being used. And again, um that's also kind of a signal that that people think that the utility is there and they are using it. Um DM sits as a liability on Venice's balance sheet. So, the amount of DM issued Venice must provide that inference. So, you you have it on

demand, right, if you hold the DM. Um and so that scaling question I think is is going to um play out over time. Um Eric is a visionary um and he I would I would I would never discount um >> Sure. >> [laughter] >> his ability to ensure that um Venice is ready to scale. It is a consumer-based app. Um and so the majority of the users that it's that it's facing off to primarily today are individuals, right?

So, there's there's still plenty of room to scale there. >> Yeah. >> Um I think from a from an enterprise side when I when I look at what we're doing at Base AI now, um it would be difficult to uh have a a token >> attached, yeah. >> attached to it, yeah. And I I think it would be Yeah, it would be a a different sound. >> Well, it's also just like a different Yeah, exactly. It's a different kind of like bet, if you will. Like the the Venice bet is like almost piggybacking off the compute demand scaling, you know, equation. And sure, the based AI model is essentially the same, but it's in such it's much more of a boutique kind of specialized fashion. I feel like that, you know, it is is could

eventually surpass the demand from just the consumer side cuz businesses are obviously using a lot more power, a lot more in intelligence, a lot more people in terms of actual like for for a specific use case, but >> Well, and I I use I use Venice every day many times a day. Um, but if I were a fiduciary of a public company or of a company that had a lot of customer data, so let's say it was a wealth manager or financial advisor, or a doctor, or even like a personal trainer. It doesn't even have to be like a regulated business, but I have a lot of your information and your information. So, I have a responsibility to to protect that and secure it

and keep it segregated from my other clients' data. Um, Venice isn't necessarily designed I mean, it is designed to do that, but the business isn't focused on meeting those regulatory requirements that might exist um, in a jurisdiction. So, for example, to be HIPAA compliant or to be SOC 2 compliant, right? And so, for Based AI, for us to serve those customers, we will be, right?

And I would never dream of suggesting that that Venice should change its ethos and start to get on board with all of these government licenses, right? That's completely the antithesis to what Venice is trying to do. So, yeah, that for for the enterprise, I think we have to evolve a little bit and and identify that they have rules that they have to follow and so we have to meet them where they are. >> It it sounds like Base AI is like an extension of Venice in the sense that you can use it to help achieve some of these compliance forms, right? If you need HIPAA you and and meet SOC 2, there's certain privacy requirements, you can use Venice as a tool to help fulfill those requirements for enterprise clients. Is that the right way to think about it?

>> No, so Venice Venice doesn't have those accreditations. >> Right, right. >> Um so I think where they're similar is I obviously um was part of the design of what the power of Venice would be. Like what is the value proposition? And what's the problem we're trying to solve for people and that really was an alternative to chat GPT that wasn't going to spy on you, that wasn't going to try and change the way you think through inception of giving you back some filtered version of the truth. Um that wasn't going to set you up to be you know the party to some lawsuit because you sent a prompt to chat GPT 3 years ago and now your data got hoovered up in some action.

Um and you know and you look forward into kind of the thought police minority report, you know, future, you know, just asking a question might become some kind of weird thought crime, right? Um so Venice solves for that. So stepping back from Venice, I think one of the things that I thought still bothered me and especially with the like I said the the advent of of the open call moment was all of these businesses were trying to plug in all of these agents and they didn't really know what they were doing.

And they weren't focused on security and they're putting their credit card details into you know, just the the chat app and I thought well this is this is bananas. Somebody's really going to get sued. And so it's it's the the ethos that uh we built Venice with but stepping away from that into kind of a more enterprise mindset and building something specifically for businesses that do have those obligations. That doesn't mean we wouldn't use Venice inference, right? We could absolutely use the inference that the rails that Venice has built for product layers that Based AI is is putting out to the market.

>> Exactly. Okay. Okay. Exactly. >> Maybe maybe that was >> what I was trying to suggest maybe I was in in the wrong words. >> separate companies but of course we would use Venice inference if >> to help fulfill >> Absolutely. >> Yeah. Very cool. What do you what what do you think about the the way that the whole kind of ecosystem has played out just kind of closing out? You guys kind of sat down couple years ago. Look, we want to build an uncensored chat GPT. Now you're here today trying to bring a much larger idea with Based AI Based AI to enterprises. Obviously, you know, the Venice token has done well. The mind share has grown.

People are excited about this private decentralized inference. Just kind of looking back is this is this kind of what you guys thought was going to happen? >> Uh well, I remember when we went live with Venice. So Venice was I I think we were probably 5 months in. Took us 5 months to build it out. And we went live in May at Consensus and that was, you know, our big kind of out of stealth moment. Um and Eric and I were, you know, doing press and podcasts and talking to people and networking and trying to explain what we were doing.

Um it was kind of like Bitcoin, right? Where some people like totally got it straight away but most people didn't, right? Um and the product was really new but ChatGPT was pretty new, too, right? Um and so there were still some things that ChatGPT could do that Venice couldn't do and so, you know, there was it was it was really kind of bringing people along on the journey. And I knew what we were doing was really important, but it wasn't it wasn't like a viral aha moment, right? And one thing that I was really surprised by, honestly, was I thought that people would get the privacy imperative straight away. And they just didn't, right? So, in the beginning, I think the uncensoredness

of Venice was more attractive to people it was easier to see. It was more visible. You could ask the same question on two platforms, and the difference in the answer was really tangible. But privacy is a little bit more ephemeral, right? And if you ask most people, do you value privacy? Would you invest in privacy? Do you act in a Is privacy important to you? They'll all say yes, and then they'll do stupid stuff, right? You know, put their location on Instagram, right? Like Obviously, that's you're not really that concerned about privacy. And so, for me, I think there's a little bit of retraining people, and I think that has a lot to do with like social media, where I we think, "Oh, I don't have

anything to hide." But this is a very different, right? You put on social media what you want people to think and see about your life and and and your personality. The way that you engage with AI tools is deeply personal, right? It is who you actually are, not what you want people to see about you. And so, when I ask people, would you put your diary or, you know, your all of your, you know, financial statements on your Facebook? And they go, "Oh my god, of course not. I wouldn't do that." I'm like, "Well, essentially, that's what you're doing when you're putting all this information into ChatGPT. It can be accessed and viewed by people that you don't that you don't and used against you, right? So, for me now, in this

moment, I kind of I I have a B2B background, right? I I I joke that I'm a reformed banker. Um I spent a lot of time with, you know, governments and policy makers working on crypto policy. Um I built out some of the biggest exchanges. I was the head of strategy and regulatory affairs for Circle um during, you know, the the MECA licensing um period of time. And so, I understand what institutional challenges are. >> Sure.

>> Um maybe more than I do individual humans because they continue to surprise me with the silly stuff they do right? >> Tina, you've got a big, big job on your hands. >> Yeah. >> You've got [laughter] You've got a lot of private diaries to protect from enterprise secrets to plans to team management on the Venice side, everyone's dirtiest, darkest secrets. >> Yeah, they're important and they're yours and they should be yours and no one else's. >> Thank you, Tina. >> No, thank you, Andy. >> Hope you have a great event. Thanks for sharing some time with us. Ladies and gentlemen, that has been AI Supercycle episode number 33, our weekly premier AI show powered by our friends at Near. We are live at the Bermuda Digital Finance Forum day two.