$VVV Community Call | September

Hey, Bankless Nation, we've got some bonus content for you here on the weekend every month, Bankless, and Venice partner together to do a VVV Community Call. That's a one hour long episode that I host live on X with John the head of strategy at Venice. In this call, this episode is really meant for VVV holders. We talk about VVV economics, Venice product updates, and cover the news and the AI industry as it relates to privacy and dignity in the world of AI. And, as you might guess, there's always a lot to talk about.

Well, typically we don't place content with our partners on the RSS feed. We've got enough requests from the community that we do so since that is where they consume their content. So we're adding the monthly VVV call to the RSS feed on the weekend after it's recorded. If you like the episode and you want to join the VVV telegram where we talk, learn, and educate about VVV, you can hit me up in the bankless discord or in my Twitter account and I'll show you the link. So here is the September VVV call with John from Venice.

Welcome to the September VV call every month. John the head of strategy at Venice and I go through the developments, the news, the events in the Venice and private AI space, keeping you up to speed with all things VVVV and everything related to user privacy and sovereignty in the AI age. John, happy September. How are things over at Venice? Yeah, happy September, David.

Things are good. It's been another crazy that it's been a month already, but things happen fast. I think this week or this month specifically has been when the social zeitgeist around AI has hit new highs with attention on anthropic and attention on open AI. Dario is having a meeting with Donald Trump, the big AI labs, the market conversation around the big AI labs is looking at their IPO coming down the pike at the end of this year or early next year. I think as a societal conversation, I think we're having the AI conversation more than ever, which feels crazy to say, but because we've been having this AI conversation for years now, but it seems to be nonstop in terms of just how ready AI is just a conversation on the daily podcast and the New York Times.

Everyone is talking about AI everywhere. Yeah, I think as these labs push forward, as the models continue to get better, I just I don't see how that doesn't continue to be the case. The thing just becomes more and more relevant as they become more and more capable and more people actually wake up to the capability. I'd actually argue that only a very small slice of the population has even woken up to what they can do today, much less what they're going to be able to do, you know, about three months from now. So I don't see that train slowing down.

In these VVV calls we're going to go through just updates in the Venice ecosystem, updates in the VVVV and the economics updates in the Venice products, but we're also going to talk about some AI current events where I want to pick the John's brain on some of the things happening around the Venice space in some adjacent companies just to kind of keep us up with the news. But I want to start with this tweet from Eric here. This came in pretty early this month, I think September 16th was when this was tweeted. So two weeks ago, Venice now at 250 billion daily tokens. This is Eric for his one of the founders of Venice.

He's retweeting himself from a tweet a while ago when he tweeted out that Venice is at 100 billion daily tokens gone through the Venice platform. Now, Venice was processing 100 billion daily tokens on June 25th, 250 billion daily tokens in September 16th. And if you look at this curve, I see a skateboard ramp. I see a curve going upward. So it's not just growth, but the growth itself is growing.

John, talk to me about what this chart means for you as strategy and for Venice as a company. Why is this chart so significant? Yeah. So I mean, at a high level, I mean, every time Eric posts this chart, it's always one of my favorites. And I always like to sort of, you know, half a ceaselessly, but not really like to say it's the best chart in crypto because this is just a chart of the fundamental business at the very base level.

There's obviously a lot more that goes into Venice's business than just how many tokens are we doing a day, but at a high level, directionally, like, this is the growth. This is where everything is. And like, you know, and I think also when you look at comps, when you look at the open routers, et cetera, they have, you know, similar charts. AI growth is AI usage is growing across the industry. We are certainly experiencing that event.

I think this chart is one of the best examples of just fundamentally, we're just simply doing more tokens every day. And this chart, I believe is specifically LLM tokens too. So like, you're not even talking about image models or video models as part of this, but at a base level, it just means more users on the app, more users on the API, more overall usage, the models themselves get better or cheaper. And so that can also push a bit more token usage. But yeah, I mean, this story has been pretty consistent, certainly since the beginning of this year.

When you say that this is your best chart in crypto, your favorite chart, is it accurate to say that like, if there was one chart to distill down the health of Venice, it would be just one. Would you say that's accurate? Yeah, I think at a high level, this is, yeah, that is accurate. There's obviously things that go into this of like, you know, how much revenue is Venice actually doing? What do the burns actually look like in those derivatives?

But at a high level, this drives everything, you know, this is growing. It's very likely everything else of the business is going to continue growing. I didn't realize that this is only LLM models and image and video models are actually not contained in these tokens because those have to be pretty consumptive in terms of inference of a total inference. Do you have any sort of ballpark metric about what this chart encapsulates? Yeah.

So I mean, what I can tell you is image and video have been growing in similar looking charts in terms of total amount of tokens, just because of the way those models work. They're actually, it's not going to be as many tokens. LLM's consume basically the most tokens because they're especially identical. They tend to be doing multiple terms and they might go do a bunch of things and pull out a bunch of contacts and they might even work themselves, you know, on an image model or a video like the LLM may actually interact with those things. So the actual amount of tokens on the video and the image are not going to be as much, but they're obviously, they're much more spendy too.

You know, you spend far more on, you know, a turn on a video model than you do on LLM generally. Yeah. I want to focus on the fact that this chart is bending upwards. If you kind of linearly extrapolate the growth on daily tokens, you get to about 1 trillion tokens a day by towards the end of January. But if you lean into the fact that this chart is curving upwards or the growth is growing, you get to 1 trillion tokens by December.

So I guess the question to you, John, that I have is when will Venice hit 1 trillion tokens a day before or after 2027? Yeah. And it's hard. You know, that would still be, you know, like a nearly 4X in that time from this point. So I don't want to be overly aggressive.

And I'll say probably this happens after December, but I've, I, you know, myself and many of us internally have been very wrong about the growth rates a number of times this year with it accelerating far past what we expected. So I don't think it's really by the end of December would be at all crazy. But if I had to guess, I think it's probably more likely like something like January or February of next year, just based on the current growth rates. But yeah, I mean, the growth rate in general has just been nuts and a lot of our models and forecasting we are beginning of this year. We assume something like, you know, if things were going amazing that we'd be doing something like 15 to 20% month over month growth and the actuality of this year, I think it's been closer to like 35 or 40%.

So we, we just tend to underestimate our own growth. Yeah. Yeah. One thing that is nice that we have in the crypto ecosystem. I'm speaking as an crypto investor here is that like AI is very exciting and frothic open AI is very exciting.

People really want exposure to those things. It's nice to be able to see the growth of AI demand in a chart. This is like what that chart would look like, you know, slightly up into the right with, with a slight crescendo of acceleration on growth. And it's also nice to be able to have an instrument that has exposure to the economics of those tokens. And like maybe there's other tokens out there in the crypto ecosystem that I'm not familiar with like how comes to mind, but like having something, having an instrument that is has drawn correlation to just AI inference demand.

It's nice that we have that option in the crypto sphere. Yeah. Yeah. I mean, I agree. That's definitely one of the reasons, you know, we have done the things we have done and we think it's important.

And, you know, even as you started this conversation talking about these IPOs coming up from anthropic and open AI and all these things, you know, we think it's important that people have access to something like this so that they can hopefully help participate in all of this enormous wealth that AI is creating that, you know, many people, if they try to buy into those frontier of the lab IPOs, they're not going to get any exposure to until it's, you know, well-known or at the $1 trillion most likely. Yeah. Yeah. Let's talk about VVV emissions. We've talked about this for the last two VVV calls, but as we all knew, VV emissions has gone down from 3 million a year down to 2.5 million a year.

This is the first of two emission reductions in VVV issuance will be going down to 2 million a year again later in October in a few days, actually. So if people on Thursday, yeah, so it's Monday on Thursday, it'll be down to 2 million a year. What do we know about the final equilibrium of VVV issues? Do we like the 2 million number? How do we determine that that was correct?

What would it change your mind about that needing to either like increase or decrease further? What are the inputs here? Yeah. I mean, we look at a number of things as these emissions come down. So obviously that affects basically staking yields the most.

What the yield is both the determinative, what is the emissions, but also how much VVV is actually staked. So those things kind of balance each other a little bit. So we'll be certainly watching closely what happens after this next emission reduction, how this all plays into things I would say in terms of where we want to end up. We're still, I guess, is even a 2 million a year, we're still going to probably be higher than we need to be. The goal is still a deflationary asset and that's likely going to require more emissions reduction.

So this is what we have announced in planned right now, but there will likely be more communication from that on us maybe as soon as mid-October or something like that. Okay. So not official comms from Venice, but John's hunch is that 2 million might be on the high side. So the near-term plan for you and Venice is to get down to 2 million yearly issuance number on Thursday and then see how the market reacts, see if people stake more, see if people do, see what people do, let the dust settle for a little bit, collect some data and then proceed from there. Yep.

I would say that's pretty much accurate. No decisions have been made past the current ones that have been announced, but yeah, we're always watching this closely and we really mean it when we say we want this asset to become a deflationary and that's going to mean being very careful with emissions in the supply in general. Speaking of the supply, the DM target has been going up. It's now up to 40,000 as the new DM supply target. So same kind of question.

What are the inputs? What's the data that you're looking at? Has anything because the DM supply target has been going up for I think six weeks in a row now? What data have you got? What have you observed to inform your decision-making about whether this equilibrium of 40,000 is correct or not?

Yeah. So same kind of thing. We observe the changes in real time. We see what impact this actually has on people staking, locking their stake to mint whether they burn some of their DM to get their stake back and just where how much we actually locked. And what we had been seeing is that there was definitely an acceleration on the locking as this had happened, which is kind of what we hope to and expected to see.

But that has definitely come back down a little bit in the last week, mostly driven I think by price action and some people wanting to unstake their VDD understandably, which required them to burn some DM. So this one actually, it's going to be really neat a little more time. So we only just had the last target increase take effect like a week or two ago. And really, this one is a more sensitive one in some ways. I want to see some months of data of how the market actually responds to this thing long term where the VVV supply actually ends up now that we have this new target and have made it easier for people to mint some DM.

But that's not in the vacuum. You have all these other market effects going on. So I think over time, this will be less noisy in the short term. It's a little bit noisy. Right.

Yeah. It was something that you said people will want to unstake their VVV when the VVV price goes up. September has been an incredible month for your VV price performance. And so does that imply that when VVV goes up in price, there is a DM demand because people who want to get liquidity at that higher VVV price have to go buy back DM off the market so they can unlock their VVVV. And so that can kind of like throw a little curveball I think in this ability of the VVVV DM relationship when there's a bunch of VVV volatility to the upside.

Yeah. This is very much what I would call a feature, not a bug of the tokenomics in that if anybody mints and sells VVV, what they're effectively doing is betting that the DM price at least will be lower at some point, that they'll be able to buy that DM back at a lower price and go from there when you have things like this, when the price is going up, it basically creates an incentive for those to want to buy back their DM, which naturally means it brings some more buy demand to DM because generally unless they're still holding that DM, they're going to have to buy it off the market somewhere if they ever want to unlock their VVV. So yes, it creates kind of this natural pendulum swing where when people want to unlock, they have to buy DM when VVV price and DM are accelerating.

It creates the opposite natural incentive for people to want to mint and potentially sell that into the market. And so the idea is that those kind of always that pendulum swings back and forth and you get a relative equilibrium. And I think it's interesting to point out, obviously these things are correlated, but I when we first announced that we were doing the DM supply, there was at least some feedback from a few community members who were concerned that we were quote unquote inflating DM. I still don't think that's really the right way to look at this. And I think the market kind of bear that out like DM price is actually up until the right even as supply increased.

Yeah. Yeah. In the last like six months to a year or so since the VVV price went from like a dollar or two up to where it is now at like $30. This has been the first real substantive like price appreciation that has also kind of has wagged the dog with impacting the VVV and DM economics. Now with any future price appreciation, at least we have like some hindsight to look back on.

We have some data to look back on. But nonetheless, there's always just like some inherent instability in the smoothness of every thing when there is like rapid price appreciation and also the DM economy so young as well. Like this is really the first data set that you guys are really able to work with to make sort of like informed decisions. Is that right? Yeah.

I mean, the DM targets supply in particular had stayed the same for you know, a year since launch and then this was our first actual change to the target and for us to see a chance of how that actually impacts the economics, how that actually impacts how many people are minting what people are willing to pay for a DM all these things that the market gives us good feedback on. But yeah, I do think it'll take a little time just like the initial DM tokenomics took, you know, I would say some months to kind of settle out into an equilibrium. I think we're still going to probably need a few months of time as people adjust all these various changing variables to settle out what I think the long term impact of this change actually is. So today, like I said earlier, I think it's a little bit noisy. There's so much going on that it's hard to draw too many conclusions from any of this right now.

But I think within a couple of months we'll have a better view. Now on previous VVV calls, we've talked plenty about the different mechanisms for how VVV gets burned. There's an automatic burn triggered by any time a new subscription comes. Somebody signs up for Venice, they buy a subscription, an amount of VVV is burned. And also a 5% of API credits that people spend.

So somebody buys $100 of API credits, 5% of that goes to buy and burn VVV. One thing we haven't covered, one mechanism we haven't covered is the discretionary burn. Discretionary can mean a lot of things. It's a pretty big term. What does it mean for the burn to be discretionary?

Can we discretionary burn more? Do we need discretionary burn less? What are the variables that go into this discretionary word? Yeah, so the variables, it's right there in the word and that this component of the burn today is ultimately up to the team and the team's discretion. Generally, this has tracked more or less loosely, not exactly, but loosely revenue growth.

And this was the first burn bucket that was introduced into VVV's tokenomics before we added any of the programmatic burns. So we have continued to use this lever as a way to effectively capture some of the revenue appreciation and bind burn that we have not directly put into anything programmatic yet. What's interesting is I think this might have been the first or second month where the programmatic burns, I think, eclipsed the discretionary burn, which is great to see even though the discretionary burn has continued to grow. That's also a great sign. Exactly what goes into it is it's ultimately the team looking at what we think makes sense for based on high level revenue numbers, high level growth numbers, what we think we can put into this and how the programmatic is scaling.

I think in the long run, over the long term, it might make sense. Right now we want to keep growing discretionary, I think. Over the long term, it might make sense for a discretionary to take a backseat to programmatic as the business really continues to grow and take off, hopefully, if that's the case, then I think we want more and more of that to be on the programmatic side and less up to us or or anything like that. But for now, it's a nice lever for us to basically add on to the burn every month and we're going to continue doing that at least for a bit. It's the idea here that on a month to month basis or on a quarterly basis, the team opens up the Venice P&L.

The Venice team has an amount of burn associated with it. You guys have to pay salaries. You guys have costs. You have data center costs. You have to pay for.

But then you also have the revenue and you also have already the programmatic burns. And then when you net everything together, there's some money left over. And out of that money left over, you discretionarily decide to use some of that money, some arbitrary amount to buy and burn VV as you guys so choose. Is that going to help works? Yeah, at a high level, I think that's a good way to put it.

It's some portion of our free cash flow after all those various things you mentioned, all the various expenses, all the inference we need to buy, the salaries we need to pay. In addition to our programmatic burns, what else can we put back here into the burn? And it's not like we decide to be clear on one number and then on one day this happens. These discretionary burns, just like programmatic burns are actually happening all the time. Effectively, there's just a constant T-WAP going in addition to the programmatic burns.

It's just buying all the time. And that adds up over the course of a month. And then you end up with a discretionary burn member based on that. Right. So there is the excess cash.

It's not like a market smash the buy button on VVV. It's a T-WAP. And the idea is there's excess cash that Venice has rightly created a buffer with and ultimately at the end of a month, I don't know how long a quarter you guys have some like slash, slash fun built up and you're like, okay, let's see what that into VVV and burn it. And that's a discretionary burn. Yeah.

I think at a high level, that's the right way to think about it. I like my things being like 90% accurate. If I'm like 90% direction, incorrect, I think we're doing good. One other change that I saw that happened this week is the increase of the programmatic VV burns from pro-structions now burns $2 if anyone signs up for a pro. So I've been now burned $2 in VVV, which is doubling from a $1 VVVVV.

Same kind of question. What is the input into what gives you guys the room, the tolerance, the flexibility to feel that it's safe to like a double the amount of VV burn in the economics of Venice? Yeah. So this isn't, this is back actually in April. It looks like.

Oh, this is in April. Yeah. Oh, this is old. Okay. Well, it's a similar kind of question though because like this does kind of show that there's going to be changing VVV economics with the programmatic burns.

Maybe you guys rate it up, rate it down. How do you guys make those decisions? Yeah. So I mean, we're still, you know, this is all kind of a work in progress, but we try to right-size these based on it. So when we first released these programmatic burns, it was just a basic $1 for whether it was pro-plus max.

So we finally released these programmatic burns. We made this change so that we wanted to stratify it so that there's a bit more burned per sub and that it's actually more in accordance with the level of sub that people are buying. Part of this too was just, this also allows the on-chain data to be more readable and knowable by something like a Venice stats. You know, when it was all a dollar, you couldn't tell what was pro, what was pro-plus, what was maxed with the different amounts. That's much easier for that on-chain data analysis to happen.

And then obviously, you know, for us, it was just like we wanted to have kind of a roughly similar proportion of the sub payment that goes to the burn. So the higher ones burn more. That was really it. This is something we will periodically reevaluate if something makes sense. And certainly if we did something like change the subscription pricing, this would probably change with it.

But right now it's kind of just based off the cost of those subs and what we think makes sense. When you guys meet decisions like this, it's one of the inputs into your guys' decision making. If we make it this number, that puts more data on-chain so that the VVV community mainly get go on Venice stats, but the VVV community gets more data to work with. And so it's like a dysplasia of the internals of Venice. And you guys actually do think about when and when not to put data on-chain.

Is that something a conversation that you guys have? Yeah, I would say that's definitely a conversation we have about what is the right data to put it on-chain, what's the right way to do it and how to differentiate it. So like, yeah, actually, you know, like there was actually a problem when we added the credit purchases as burns because they had the same $5 number as ProPlus. So you'll actually, if you look on-chain, you'll notice there's actually a slight difference between those. It's like a five or six cent different, but it's enough of a difference that those who are doing on-chain analytics can tell the difference between those things.

So it is definitely something that we think about. Nice. I like the idea of this fan is intentionally leaving very explicit breadcrumbs for those who look for it and have the data that they want to find. That's something to some of the market commentary that happened this week. On September 8th, I believe, VVV broke through all time highs at the same day as OpenAI had this Navier-Stokes Millennium Price problem controversy.

This was a big day. I think if anyone was holding VVV, they remember where they were reading the news this morning because maybe they did what I do and I looked at the charts and I'm like, oh, wow, wait a second. That's a very different price. Or they were looking at crypto Twitter and crypto Twitter was blowing up all about this anyways, but I'll kind of go through the story here. VVV started the day at $18 and peaked at $30 while a couple of mathematicians accused OpenAI of buying and stealing their work that they had prompted into OpenAI codecs.

So the TLDR of the controversy was that OpenAI claimed they found a solution to the $1 million Millennium Price math problem. There's a handful of these math problems and this is one of them. This particular math problem involves fluid dynamics that had stumped mathematicians for something like 200 years. Roughly 12 hours before OpenAI claimed credit, two mathematicians made a public statement said that they had been closing in and on solving the same problem using AI tools including OpenAI's codecs. Tristan, one of the mathematicians insinuated that OpenAI may have secretly accessed his notes that he had deposited into codecs because OpenAI's solution to the problem basically had the same logic processed that Tristan was using with.

So the claim is that OpenAI just used Tristan's work as a jumping off problem after spying on his work using his work that he had submitted into OpenAI. OpenAI didn't explicitly deny that this was a possibility. And overall, I think I woken a lot of people outside of crypto Twitter, outside of people around Venice to the idea that the big AI labs are kind of looking over your shoulder at your homework to see what value that they can get from whatever you submit into OpenAI and drop it or whatever. Giving a peek into the Venice Slack when this happened. What did you guys talk about?

What did you think? What was your reactions? What was being said? Yeah, I mean, I think this was another one of those events. I mean, I feel like almost every month we end up talking about like this one was particularly big obviously and there was a big reaction online about it.

You know, obviously outside of just crypto Twitter but across all AI people were very interested in this subject. And I think it's just another one of those reminders to people that like, yeah, actually keeping your thoughts private might actually matter, especially if you're working on novel interesting things like novel mathematic solutions. When you put into a frontier model generally is kind of just getting hovered up by default and they don't always even know themselves what that was kind of part of the of the conversation here is open AI was like, well, we didn't specifically do anything here, but we can't rule out that that information basically had been taken and put into the training data of our newer models. And that's kind of always the case. Like they don't they don't have like a whole list where they understand every single piece of information that goes into the training data.

And these models, they want, you know, at a core level, they want novel information. Like that's kind of the best thing to put into the training of these things is any sort of novel novel ideas, novel theories, like novel mathematics solutions. So it's really hard to divorce the fact that these guys had been working in Codex and like likely supplying a lot of these and then opening I just happens to announce that they've solved this long standing mathematics problem. I just I think it was another reminder to folks about the importance of privacy and AI. Obviously, we were very, you know, interested and excited about this from a standpoint of for us, it's kind of like another day, you know, another day at the office, another example of exactly why Venice exists, exactly to solve problems like this so that people don't have to worry that everything they're doing, the companies looking over their shoulder that every thought they have is just getting taken and put into AI training data and like even when someone, you know, checks the thing on the front of your labs, it says don't train on this.

It's just like what level of verification, if any, do they have that that's actually what's happening and that something doesn't get sprouted and anonymized and just brought into the training data because it's useful. Like we just don't know. And so you have to trust the labs that they're word and even that they're word, they don't really know. So I think it's just a good reminder to everyone about the fact that when you talk to these things that they are likely taking everything you get it. Yeah.

I think the conversation downstream of how actually the data got into hands of OpenAI was the most interesting because OpenAI distanced themselves from like, hey, we didn't quite know that this exactly was happening. But nonetheless, we can't rule it out that our models weren't trained on the data. And when they have to distance themselves from how they can't rule it out to me, that shows how structural it is as in this is built into the way that the company works as in they must steal your data in order to have an edge over their competition, who by the way, they are in an arms race with. And so it's kind of this like toxic arms race that forces the frontier AI labs to need to be able to access as much data as possible.

And I think this is why I think society's like heckles are being raised as some of the AI labs because they've seen this pattern before, right? Like Zuckerberg with Facebook, you know, if you're, if your product is free, maybe the product is you actually. And so people are kind of pattern recognizing with the frontier AI labs. And I think there's the frontier AI labs are losing a lot of trust in society because of stories like this. Yeah, I think exactly.

I think it's just a reminder. So many people use these products and don't think about these problems until they get stories like this and they either are materially affected or in a similar field or something where they actually have to worry about this and they think harder about it or something else comes along. You know, sometimes this story is a subpoena because people realize that the company has all their information and comes up in like some various lawsuit where that information is discoverable or sometimes it's a story like this where, you know, some novel solution to a mathematics problem is seemingly taken from the people who had actually been working on it and the big labs came, you know, credit for it. And I think the labs themselves are even reconsidering how they put out information about solutions like this because of this exact problem and worrying about how people respond to that.

But that's not a fairly good thing either. Like, you know, to the degree that we are getting things like novel solutions and mathematics. That's something we want widely shared. That's good for society. Those are who knows what things could be built on top of things like that.

And we don't want the labs to be in an incentive where they don't even want to share that such a thing has occurred or not share the proof. Like, we absolutely want that. But we also want them to, you know, respect those who are actually working on these things and have, you know, novel views in their field and not just getting hoovered up and people just, it's so easy to forget that exactly what you said. Like, these are built into how these companies work. This is not like a one off or a, they forgot to click the button or whatever it is.

This was an obipsis. No, these are products and companies that are built off of these massive, you know, or hugely expansive data collection because that's how you train the best model and they have such competition with each other with China across the world. And these things are really not slowing down. Obviously there was some various discussion that happened, you know, around potential slow downs and among like Dario and Sam and some of the big labs. But the reality is they just, they cannot slow down.

They're just everybody. Everyone in the world, in my opinion, is caught in this like prisoner's dilemma where if you slow down, you know, the result seemingly is worse. And so they're all trying to find a way to balance this. And it generally means they're going to use every edge of the day on including every piece of data that get their hands on. Yeah.

You use the illustration of when a lawsuit happens to one of these AI labs and you as a user, your data gets caught up in that lawsuit. Well, John, one of the news was that AOP and AI must turn over 20 million chat GPT logs according to a judge. So this is not just an illustration, not just an anecdote. This is real that opening I, equal have to turn over 20 million anonymized and anonymized GPT logs. So there's something there and a consolidated AI copyright case after it failed to convince a federal judge to throw out the judge's order about insufficient privacy concerns.

And so this is happening in reality. And this was earlier this year in January, but I think the test met evidence is here nonetheless. Yeah. I'm sure you could do a search or a AI about it and find 10 or 20 other such examples just like this. I feel like I hear one every other week.

There's always something going on. The legal cases are kind of canaries in the coal mine because they will push for whatever data a company actually has. This is one of our parameters at Venice is like how secure actually is the data, how effectively subpoena proof is it? We can only turn over what we have. We get a legal request to Venice.

We're going to know what we have. But the key is that we just simply don't have anything. There's nothing to turn over if you don't have it. But all these companies, they very much have it because their whole business model is built off having and using that data. I think two calls ago, John, we talked about the luxury that Venice has of not having to be in the arms race of model development.

You guys don't develop models. That's not what you guys do. It's not the business that you guys have. You guys allow the rest of the world to develop models and then you just plug it into the Venice platform. And I think that removes you from the AI arms race in a way that must feel kind of nice because, hey, you don't have to have the burden of training the models.

That's not a cost that you guys have to bear. And as a result of not being in the arms race, you don't have to feel compelled to retain users data or do anything you can to get your hands on more data. So I feel like it gives you guys a much more relaxed position in the market where you're don't have to worry about these sorts of things. That's my intuition. Is it feel that way?

Yeah, I mean, this has very much been a strategic decision on Venice's part to not wait into this field of model training and everything that comes with it. And I think so far it's very much been the right decision for Venice for the product we want to offer. This is not necessarily our expertise or a sales set. It doesn't mean we can never look at it or never lean into such a thing. But generally, there's just so many, so many labs spending so much money on all these models that it's better for us to just see them come out.

And I'm very happy personally that we don't have the stress of training these models because these companies, they spend so much time and everyone training these models. At this point, it's like, if they're lucky, they get to basically, their model gets to shine in the sunlight for maybe two weeks before it's surpassed by something else. So it's just so much stress to get that thing out there as fast as possible so that you can maximize that time that you get in the sunlight. Where for us, Eric has this example that, or this analogy that I like to use where he basically says it's like, man, that I'm raining from heaven of all these models that just get dropped and plugged into Venice and useful for our users and great for us. And we don't care which one is the hot one right now.

That's not, it's just not really our game. So the fact that we've kind of opted out of that I think is a big structural advantage for us and how our company works compared to a lot of other AI projects. Mana, raining from heaven is very poetic from Eric, as always. There's one more in the Trad AI space. I want more news about that.

I want to talk about this is actually from anthropics, own privacy.cloud.com website. This is updated, I think on September 8th, anthropic released a new data retention policy. And so this is a quote from the page that I have pulled up. To ensure we are responsibly deploying covered models, we are requiring limited data retention and review as part of our safety work prompts submitted to and outputs generated by covered models are retained for 30 days to support our safety work on every platform where these models are offered. So we are on a consumer plan from anthropic that's like cloud free pro or max.

You don't have to worry because your data was already being collected and retained permanently by anthropic. But according to anthropic, this is for enterprises and organizations that have already set up workspaces with anthropic with a zero data retention policy. So if you have that zero data retention agreement with anthropic, it is now actually a 30 day data retention agreement and anthropic is citing safety concerns as the motivation behind this. Now, to put a negative spin on this, when I see anthropic citing safety concerns, I kind of equate it to the government citing state security or this is because of terrorism or the kind of protected children or something, something, something, something that the government wants.

And I see that same sort of structure where anthropic is retained for 30 days because of safety concerns. But then if it gets retained data, it gets the value of that data. That's my like in sippin way of looking at this, but I'm sure it's probably somewhat accurate too. Yeah, unfortunately, I don't think that's an overly cynical view. I think it's relatively accurate.

And yeah, you always have to wonder, you know, certainly with anthropic and open AI of just like, what do they mean when they say these words safety? You know, it's everything they talk about lately, safety alignment, what does it actually mean? Safety for who alignment for who? What does that mean? So it's this very vague term that they use to justify something like this.

Like we can't even, you know, they can't even describe to their own ZDR policies anymore because now they have to do 30 days of retention for safety. And you know, and there may very well be like a few legitimate reasons why they have to do that, but they then turn it into this broader catch all that seemingly catches everybody in it and all of their data and just so happens that they also need that data to try. So you know, it's hard to not be slightly cynical in terms of their positioning and what they're actually doing and what they actually mean and how concerned are they actually about the safety components of these things, especially when they are already building so many guardrails into their models to begin with about what they will actually respond to, how they will respond to things, even on things that are often innocuous. Like again, this remains one of the leading kind of funnels for Venice is people using open AI using a drop in getting a refusal over something that they really don't think they should be refused on tends to lead them to search for alternatives and things like Venice.

So for us, yeah, it's hard to not be a little skeptical of what they mean. Ultimately, they're their own private company. They're going to do what they want and what they feel they need to do. And I think that they are, I think that they just don't want to give up that that training data feels like more of the root of this than actually, you know, operating as safely as possible in my opinion. So John, I think this is our fourth call together the VVV call do every single month.

I think this month is definitely the hottest in terms of the rest of the AI world's conversations around privacy and concerns around like user alignment, user alignment with like the AI platform. Has that turned into any sort of a change in signups or conversations with enterprises? How is that manifesting on the venous side of fence? Yeah, I would say it's it seemingly just keeps adding to the growth that we are seeing, which is just really not slowed down this year. I think even internally we keep waiting to see if we're going to hit some sort of growth plateau or you know, things are going to go sideways for a little bit.

And it just really hasn't happened to be honest. And I think I think part of that is these conversations and people waking up to the importance of privacy and AI. And it's still, you know, you know, we live in a bubble, especially those of us who live on Twitter, et cetera. So we're seeing, you know, these loud conversations that are probably, you know, only representative of a very small percentage of the population even being aware of them or involved in them. But they're at least becoming louder.

And I think that does wake up more people. And then you just, you get another story like this every week or two and people realize that. So for us, we just kind of are seeing continually steady, you know, adjustments to the growth rate that seemingly keeps going up and to the right. And so we're not complaining about that. But also AI itself just continues to grow and just becomes so much more capable.

And some of the things you can do with these models compared to what you could do even three months ago is just so wildly different. You know, some of my own tests that I run on some of these internal models, it's just wild how much better they are getting at things and the types of things they could do that you couldn't even imagine them doing three or four months ago. And then a lot of those things tend to drive a lot more revenue and token use itself. So it's like the better they get, the better they are also are at burning tokens. So you see just a lot more of that.

It goes back to the chart we showed earlier, like all these things coalesce and then you just have growth of these things. And I just, again, it's hard for me to imagine that train slowing down anytime soon because these things are just becoming so much more capable and so many more people are becoming aware of that capability. But at some point, we're going to hit like an inflection point where like a large, a real large, you know, percentage of at least the American population, but really the world's population becomes more understanding of what these things could do. Because you know, we all feel like we understand it, but we live in this bubble where, you know, we're using things like harnesses, you know, like the cloth coats and coatexes and hermise agents and all these things that really are rockboxed. They're now views and like all these things that make these things far more accessible and useful.

But it's such a small percentage of people that are actually actually using any of those tools. So like my best example is always that there's like, there's such a wide swath of the population that don't realize AI is anything more than souped up Google. Like that's how they use AI. Mostly they ask the questions like they would ask Google, they use chat GPTs for products, they get an answer back. They're like, wow, this is really cool.

I can have a conversation with it. But it's really just question answer, question answer. But the gap that's been forming is those who use AI in that way, which is basically really where AI was, you know, nearly 18 months ago. And then the people who especially in December of last year, January of this year realized that there was this leap where AI stopped answering questions and started doing things. And I just think so few people are actually turned on to that yet.

It's growing dramatically, but it's still such a small slice. I want to have a discussion with you actually about the way that the utility of all these models, where it gets expressed in like the tech stack. I think people will like the most normal thing for the unexposed to AI person, they'll go to OpenAI or Intropic, just like one of the two AI apps, it's likely their first stop. And if they want to learn how useful an LLM is, how useful an AI agent is that's what they go and they would tinker around there. But Venice has this product called Minds that allows for LLMs to get combined together.

And that kind of opens up the conversation of like, is the most useful place for AI at the like the layer of an AI agent with an AI lab? Or is there a layer, I kind of think a Venice is an abstraction layer around above all the models, is that you guys are placing a bet on that's actually where a lot of the value of AI gets expressed because that's where agents get to talk to each other, that's where LLMs get to combine in weird ways. Do you think that's a more creative layer than just seemingly something on top of the AI labs that's super powerful? Yeah, I think there's kind of a, there's kind of a couple schools of thought on this question. So there's kind of the view of the way the Frontier Labs have viewed it and they kind of have to view things this way because of the way their companies and their products work.

Their general view is that these kind of specialized agents or outputs or even things that use multiple models that they think that everything's going is it's just like one big fat, super smart model and that that's where everything will be. But that is not our view and I did not really the view that I think long term is going to be correct. So something like Minds, which is a product that we have in beta right now, we're going to be pushing that a lot more as we get some additional features shipped here over the coming weeks. But the idea of Minds is that a creator can assemble a model or various models together in a way with an agent that produces a far better output for a specific task than the kind of quote unquote naked model would by itself.

I personally think that's more where things are going and that being able to direct models in the way like effectively orchestrate or like give the agent instructions about how to orchestrate in specific ways for specific tasks will generally get the average kind of call it norm the user a much better output for the thing that they're looking for. And that's kind of what the bet of something like Minds is, is that they're really smart people out there who understand these models really well. And then there's a lot of people who just want the output and they have no idea how to get it. They don't know how to prompt well. They don't realize how much their prompt affects things.

The models in general are getting better even with very bad prompts. But the you can you can be a wide gaps though in turn and I think this will persist of the outputs that are possible for those who know how to assemble things in a specific way and those who just are asking a naked model and really don't know what they're asking for. So the idea of Minds is to make that far easier like creators kind of ship these AI experiences, these types of AI products that are assembled in specific ways to help you get a specific output. And I do think in many cases that's going to be better for the average user. They don't and part of it is it demonstrates what these things can do.

A lot of people they just they get to that chat GPT screen or whatever and they don't they just don't even know what to do. And that's why they end up using it like Google that they don't know what to do other than asking a question. They haven't they just aren't in that world where these models can do things. So something like Minds will more demonstrate. Hey, you want something that does X, you know you want trading a trading agent that's going to help you with your trading or you want an agent that's going to help you with your accounting or spreadsheets or you know a PDF grader or whatever it is.

And someone has assembled it in such a way to make that thing super you know, dead easy for you where all you have to do is put the input in and get exactly the output you want. That's kind of the whole thrust in idea of Minds. And I do think that that is there's something very powerful there that is a very different place than the labs will go themselves because all of their resources are dedicated into just making their one model so smart. And I think we saw an example of this even in the industry of the last few weeks because we had this whole thing with like this new class of models come out from Jev a former co-creator of chat GPT who effectively created rather than a LM text based model effectively a classifier or decision based model which is something that had existed before but not exactly in this way.

And now people are combining these things with the LM's and realizing that these constituent parts like an LM plus using Jev in the right places can be a lot smarter a lot more efficient a lot cheaper than just using the LM on a bunch of things where it might get confused non-deterministically. So I do think you're going to see more and more of this type of thing where the real power comes from combining various types of models and kind of system instructions in various ways to get a specific output. And then you know the word biodiverse when he comes to mind. I think you started talking you said there's like two schools of thought maybe the way to define this goes to that I got from you speaking just now is you have the frontier AI labs which they're trying to produce the one model to rule them all.

They're trying to produce God basically. They want to do that. They want to make digital God and it's just one model. And then there's the open router Venice side of things like I remember Alex Attala from Open Router he was pretty early to this he was like no no there's actually going to be thousands of models there's going to be more models than we can count. And what you're talking about with minds is like the combinatorial value of unique specialized models when you sum all the parts together it's going to be greater than the whole and that's going to be greater than God.

And these are the two races going on both can be true like they're not mutually exclusive but like maybe one is more valuable than the other. And I think what you're saying is like with unique specialized models the combination is the valuable thing and that's what we are that's why Venice is trying to plan to flag and make a bet on with minds. Yeah, I think that's a really good way to put it. It's the combination and the power of those who understand these and their ability to help deliver that output to someone who doesn't have any understanding of these things. You know most people are not going to be good at prompting most people are not going to be good at it.

Yeah, this is a long standing problem even at Venice and how we're evolving the product is like you know it's great that you can come to Venice and choose from all these models but the average user gets in and they don't know the difference between you know these hundreds of different models and what to choose. So we have to be smart about how we recommend them the right model, how we get them the right actual functionality they want. The average user doesn't care what model they're using. What they care about is are they getting the intelligence that they want and are they getting the outputs that they want. So the idea is like yes the digital God thesis that a lot of these labs are going for I think there's obviously something important and powerful about that but you know for specific use cases these combinations could be a lot better for a lot of specific use cases.

They're not going to be generalized better. They're not going to be better at everything but they can be better in very specific areas where the user really cares about you know the user really cares about do I get the output I want for the thing I'm looking for not is the model capable of doing everything under the sun that's you know most users are going to use only such a small fraction of what a model is capable of anyway what they really care about is can I get the thing I want and can I get it relatively fast and easy. That's where I think something like mine is to be very powerful. That's where we are explored. Is there any other company out there who kind of has a similar vision of combining models together to create combatorial outputs or is Venice really the pioneer on this frontier.

I definitely think we are in the phase of pioneering on this. I wouldn't say that there's nobody else. I mean there are other there are other companies that are doing things like you know that are effectively councils of models or ways to get models to route to each other you know in specific ways so that you use this model on this specific task and a smarter model on this specific task all the model routing is a combination of this but I don't think anyone's combining these things in the way that we are in like a consumer focused package even though there are intentional direction towards the combination idea. Yeah I think that's definitely something hopefully we are leading on and that we are very excited about and that we think is really interesting. I would say the harnesses kind of steps into this world you know the various harnesses do this a little bit but even those are more generally about you know very broad use cases very broad generalised usage.

We're interested in these very specific use cases that you otherwise cannot necessarily get as good of an output for unless you really know what you're doing. That's fascinating. I think that's for any VDV holder I think that's something that to actually keep an eye on to see the progress of this. If you accept the spectrum between the God models from the AI labs or the biodiversity like realm of the spectrum I think that is something that a VDV holder would be very interested in seeing is like the biodiverse models coming up and coming together like making some sort of weird outputs kind of like life. I always like the biodiversity metaphors.

We need to get into the investor Q&A John we got a few questions from the first two from GW Jack in the VDV telegram. How can Venice create additional demand to stake VVV? So as Venice improves the way the way holders can use state VDV how does the team think about designing these features so that they can create incremental demand to own and stake VDV? Yeah thinking about this in a number of different ways you know obviously over time VDV is kind of been up until the right there's like you know sometimes with various price actually this will go up and down but over any long time for any of these you amount enough this is kind of been one directional since the genesis of Venice. It's definitely something I spend a lot of time thinking about what are what is additional utility that might make sense for the staking sphere you know obviously one of the big things even right now is just the fact that you can get effectively lifetime pro access if you have a hundred stake DVV that's something a lot of people like that functionality and then of course the primary ability to mid DM and get that functionality out is something that we expect to drive over time.

I think that in some ways this will actually be this question of like what drives VDVV stake will be kind of downstream of DM utility like the more useful DM is the more useful things you can do with it the more places you can effectively resell that when you're not using it or get yield on it in various ways the more demand there is to mid Venice or mid DM rather from VVVV the more likely you're going to see staking continue to rise and people walking up to do those things there's certainly a number of things nothing I'm necessarily ready to talk about today but a number of things that we're considering on that front that you know could get out of it over time and it's something we pay a lot of attention to. This next question comes from Malcolm from the VVV telegram excuse me no sorry this one comes from Iguan Arquist and also Gecko Gecko.io the creator of Venice that's good at reoccurring subscriptions become the next expansion of VVV burn so I think people are asking about if reoccurring subscriptions is where that is on the docket in the conversation of VV burn.

Yeah it's definitely something that you know obviously we've talked about including even on these calls before it's definitely something we're looking at I don't know if it will be the next one necessarily there's actually something interesting going on one of our focuses that I think is going to I think people are going to love we're not necessarily ready to talk too much about these details yet but Eric has hinted at this kind of incentive and reward program that we've talked about and I think that that will likely be one of the next big drivers of effectively additional VV buys for that reason so something to stay tuned for future calls but yeah some really cool things coming on that front and then you know certainly things like renewables or something we're considering and could be on the docket sooner than later but it may not be the next one.

Another one from GW Jack who wants to know about any timeline updates on the data center bill about in any details that you can give about the data that Venice data center. Yeah so I mean this has been ongoing you know we've kind of since we announced the raise in back in July this has been one of our you know big capital expenditures is building out this initial data center and then looking even at multiple potential for additional data centers after that. I'd have to follow up with my team on the exact timeline but I believe that the first data center is supposed to come online hopefully before the end of October so maybe like late October maybe if there's delays as there can be with a lot of these things maybe that leaks into November but it's pretty soon it's coming up that we're going to be able to turn one of these online we won't necessarily know how that affects things like margins and then how that affects things like VV burns downstream right away it'll take some time of actually getting these things online optimizing them seeing which models we can run most efficiently and experimentation on that front but you'll definitely be hearing more from us as that gets built out and I think you know over the next couple of months.

For somebody who wants to pay attention to that topic of conversation with Fennest the increased margins from the data center because you guys own your own data center you're vertically integrating why are you doing that because you get better margins that way you like absorb some of your own costs the is that the part of the VV burn where that would be expressed is the discretionary VV burns is that right. It might be in the discretionary or it might there might be a world where we decide that you know our margins are good enough that we can increase the amount on something like the credit purchase or the subs like we talked earlier so it could it could come in that way and then yeah probably initially it might come via discretionary if we feel it's creating enough additional free cash flow that we can route that back into the. Last question from Iguan artist and also Cameron in the VV chat what role does near play in Venice's infrastructure and basically I think I think that's a question I would like to know too is like I know Venice uses near for accessing some very much of the top tier privacy AI products how do you guys use near how does it work into the Venice stack.

Yeah so near is used right now as kind of a at our top privacy kind of levels of T and end encrypted inference near is one of our providers for those types of models so if you go into the Venice app and you try to look for things like T and and to end encrypted models you'll see some of these models and some of those are offered by near so that's kind of built into our stack this is definitely a what I would say is a smaller part of Venice's overall inference these are not necessarily things that everybody cares about or wants but it's really important and critical to our offering that it is available and that those who really care about kind of that cryptographic proof of the privacy that they're looking for that they have those options so near has been a great partner in allowing us to offer such a thing.

That's like for the parts of Venice that touches the near stack that's for like the nuclear levels of privacy right like the most hardcore probably also the most expensive because now we have to start to talk to distributed systems. Yeah it's the highest level of privacy some ways the functionality is a little more limited because of that because there's certain things you can't get out of the context or that has to all be delivered at once and yes sometimes that means those are a little more expensive as a result but it depends on the model I think yours generally done a good job of being able to deliver those relatively competitive pricing so sometimes they're just good models to use if you care about privacy and you want that cryptographic proof. Well John this has been the September VVV call I have learned a lot there's just so much data to talk about with Venice and also just the wider world of AI there's one last topic to talk about before we go.

I think you and I are going to hang out here in about a month in New York because of the Lumara Film Festival. Talk to me about the Lumara Film Festival what is it why are you excited about it and what are you hoping to see there? Yeah so this is a film festival that Venice has collaborated with Moonpay to put on it was something internally that like our team really cared about doing something like this because you know we talk to a lot of creators every day creators especially since the advent of you know the better video models they've really loved creating creating the Venice because we don't block things that they get blocked on a lot of other platforms and so their ability to effectively express themselves artistically is a lot less limited through Venice and we want to give you know a demonstration of how you know AI video has come along and how artistic some of these things are actually getting that this is real you know we're getting to the point where we can get real short films and soon enough like real feature link films produced with AI that people are going to be really impressed of obviously a far fraction of the cost of traditional film development and film production so yeah I think we've gotten something crazy like over 700 submissions to this thing last I heard and there's basically you know a process going down now of the judges of whittling those down and figuring out what's

actually going to be aired at the film festival but it's really cool to see this happening it's really cool to see how excited people are about it and it's going to be really fun to see that and you know we've come a long way from you know Will Smith eating spaghetti in various terrible ways to things that actually look like they could have been produced by like you know a high tier film production. Were you saying that creators creating on Venice are at the absolute frontier of what it means to be film creators with AI? I think yes I mean obviously there's varying levels of skill and professionality depending on who you're talking to but some of these high tier creators absolutely you know some of the initial previews of things I've seen and I've only seen a small fraction of what's actually been submitted but some of them are like incredibly impressive in hit you know aesthetics and you know continuity in a way that you just would not necessarily expect or not necessarily call out that it was AI unless you already knew it was and I just think it's going to kind of democratize the ability for artists you know singular artists to create full films that they're really proud of and that really speaks to what they're trying to express where you know basically the production costs go from you know millions or tens or hundreds of millions of dollars to you know a few thousand dollars to produce something of similar quality and I

think you know like we always say in AI this is the worst it will ever be it's only going to get better than me. Well if you are in New York in 28 days I say yeah third week of October this is where in the Lamar film festival is so come check it out get to see what the frontier of film development with AI is and then one last CTA for the audience you can talk about the VVV telegram that is a telegram that we host to just talk about VVV AI inference DM tokenize economics if you want and invite into that telegram shoot me a DM on Twitter and I'll throw you in there. John spent another month another exciting month with Venice thank you for doing what you do it's very valuable to have Venice here on the earth and I can't wait to see everything coming down the pipe with minds and the data center and everything else and we will talk about all that stuff in a month.

Yep thank you David talk to you more soon.