Original title: Why Erik Voorhees Says Crypto Was Really Built for AI Agents: Uneasy Money

Original source: Unchained

Editor's note: Venice is a privacy AI platform built on Base. It does not train its own models, but instead aggregates the world's leading closed-source and open-source models into a single entry point, and promises not to retain data for the portion it hosts itself. Its token VVV is not billed per use, but instead is staked in exchange for daily inference credits, while platform revenue is used for buybacks and burns, causing circulating supply to continuously contract. Over the past month, VVV has risen more than 90%, and recently hit an all-time high of $34.53. At a time when AI inference is rapidly commoditizing and token narratives are broadly failing, Venice offers a rare example: how a consumer-grade product that actually generates revenue can put a token to real use.

On August 22, 2026, Venice founder Erik Voorhees appeared on the Unchained podcast. He discussed why he chose to put a token into Venice rather than completely switch to AI, the fundamental disagreement between the crypto world and the AI world over views of power, how Venice aggregates global models and guarantees privacy, and what it relies on to make money in an industry where inference is continuously commoditizing. The following is the original podcast content:

Why did you choose to put a token into Venice rather than completely switch to AI?

First of all, it is impossible for me to leave the crypto world. It is already in my blood. I care about it very much, because it is a completely new way for money to move around the world. And money is intrinsic to all commercial activity, so I do not think it should be its own separate ecosystem. It did indeed grow that way at first. I would rather see the best "primitives" from the crypto and DeFi world begin to leave the crypto world and truly enter ordinary things. If everything is just "crypto people using crypto products to serve crypto people," then we have actually messed up.

So I want to bring some primitives into Venice and show the world that these technologies are not just for betting on coin prices going up and down. They are incentive mechanisms: there are many ways to mess them up, but there are also ways to do them well, and we want to demonstrate this in a consumer-grade application aimed at the mass market.

The world wants inference, and inference is naturally suited to crypto payments. Is that what excites you?

That is not what truly excites us. Of course crypto can be used for payments. That is the first layer, the most basic capability. If someone wants to buy AI credits from Venice, of course they can pay with crypto, but in my view this is just the entry threshold, no fundamentally different from "crypto is a payment method like Visa, like Stripe."

What I find more interesting is giving people equity in a platform and project. For example: all the major AI labs are private (Google being the exception). Because they're private, ordinary people have no way to share in any of its success. This is actually creating a lot of anxiety, with people watching AI take over the world and generate enormous wealth while they just sit there worrying about when their jobs will be taken away. Whether that impression is right or wrong, that's how many people feel. I think this situation can be avoided.

So Venice has a token, and we haven't deployed all the methods we intend to use yet, but we want users to have a stake in Venice's growth from the project's early days, from its founding. Exactly how to do that has to be done carefully, but I think it's a fairer and more interesting way to run a company: users aren't just consumers, they're participants alongside you.

What exactly is the incentive design that hasn't been deployed yet?

There's one thing we haven't announced or deployed yet, and I'll say it first: we want to take a portion of the money users pay us, whether it's monthly subscriptions or buying credits, and use it to buy VVV on the market and give it back to users, on the condition that they stay on the platform for a period of time. Essentially you're "buying loyalty," and that gives you the opportunity to educate users; if they like the product and stay, they end up getting a stake for free. This makes users stickier.

This is the approach I prefer: I don't want to go to users and say "Hey, we have a token, go buy it," that feels a bit dirty. But if they're already spending money on Venice, and we give back a portion of what Venice earns in the form of the project token, then for a substantial portion of users, that's a really cool rewards program.

What should the ultimate purpose of crypto be?

People have been experimenting with crypto for so many years that they've forgotten crypto ultimately has to serve something else. Bitcoin serves the purpose of being a base currency; every other cryptocurrency has to do something else, at another layer of the tech stack. ETH powers a smart contract platform, that's its purpose and product. Everything has to be attached to a product it serves. And we once reached the point where "the coin itself is the product," which was never how it was supposed to be; it should incentivize and coordinate things beyond itself. If it serves only itself, it becomes completely self-referential.

And it's also a race to the bottom, which is exactly what we've seen over the past few years: meme coins, and memes for the sake of memes. At that point you're not making the pie bigger, you're just moving the pie around. If people want to gamble on meme coins, God bless them, they have the right to, people have the right to gamble in a casino, that's completely fine. But it adds nothing to society, it's a consumptive form of entertainment. Sadly, a lot of projects became meme coins partly because building something with real substance is actually riskier from a compliance and regulatory standpoint. After the 2017 ICO bubble, we spent several years focused only on meme coins, and projects with genuinely legitimate tokens basically disappeared, with DeFi Summer being the last exception. That's really tragic.

What is the difference between the crypto and AI communities when it comes to regulation and power?

This is actually one of the reasons I founded Venice. In the crypto world, we have principles—at least we say we do. Some people genuinely try to live by them, many just pay lip service, but at least we have a set of principles that can sustain some degree of loyalty, and these principles are broadly shared: decentralization, privacy, individual sovereignty, the dispersion of power. We argue fiercely over the details, but the direction is consistent.

The AI world, by contrast—and this is a generalization—is completely different. AI has a much longer history than crypto; academia has been researching this stuff for sixty or seventy years. The dominant mindset in that world is: very smart people believe "everything should be controlled," and what really matters is only "who is in control" and "what the rules are." Very top-down, very power-concentrated, with the core concern being "making sure the right people are in charge."

It emerged from that kind of ecosystem: Harvard tenured professors and the like; it also stems from how funding has been allocated over decades, and from the temperament commonly found in the smartest people, the scientific personality type: "I am the master of my field, and I am very confident about it—so by extension, there must also be masters of every field, including society itself." The logical fallacy is right there: they confuse "narrow domains that can be understood" with "complex systems that cannot be understood."

When I first got involved in this space two and a half years ago, nobody in the AI world cared about privacy, nobody cared about decentralization, nobody cared about user sovereignty. On the contrary: everyone was saying "this technology is extremely dangerous, the government must regulate it immediately, everyone must be surveilled, the prompts you send must be censored, the answers must be censored, everything must be tightly controlled, and our sin is that we haven't controlled it tightly enough." That was a completely different sect, an apocalyptic worldview. It seemed strange. But to ordinary people, the strange ones are actually us—we're the outliers. That's the problem: we've been in this industry too long.

What I saw at the time was this: AI was clearly about to start taking over the world, OpenAI and Anthropic held the frontier models, and they were very top-down, very monolithic, wanting to control everything—"we enforce safety through rule sets, the government regulates us, we do what the government says, and that's how society should be organized." In my view this would lead to a very dystopian outcome: machine intelligence itself essentially arriving through the state's official mouthpiece. That is the dangerous endgame. So I often say half-jokingly that Venice is an "AI safety company," and the whole point is to stop that from happening.

Should greater power be more centralized, or more decentralized?

To put it bluntly, there are two frameworks: one, "This thing is obviously powerful and potentially dangerous, so we should centralize it"; two, "This thing is powerful and potentially dangerous, so we should decentralize it." Crypto people usually fall into the latter, while AI people mostly fall into the former.

And I think no one really knows how all this will evolve; everyone is making it up as they go along, just as humans always do, and pretending to have more confidence about the future than they actually do.

On Decentralization: Can Crypto's Tools Be Used to Counter AI's Centralization?

One of the best arguments for decentralization is its humility: the future is actually very complex, we don't know the answers, and that's acceptable; we shouldn't pretend to have knowledge we don't possess. If things are decentralized, problems are at least usually local rather than catastrophic and systemic.

From our perspective, the entire set of tools built by the crypto industry over the past 15 years was designed precisely for "decentralizing things like this." We got a bit too obsessed with "doing it for the sake of doing it," self-referential, admiring ourselves in the mirror. But now something people truly need has emerged: inference. Its production has enormous constraints, competition is extremely fierce, and the entire landscape is changing. That's why I find Venice being at the intersection so interesting. Many people say, "Goodbye, crypto is dead, I'm going to do AI," and indeed there are many fair-weather crypto people.

Also, I genuinely believe crypto is simply better financial technology, plain and simple. So I'll use the best financial tools I can get for anything I do. It's not even an ideological issue; it's a pragmatic one.

Crypto Has Always Been Hard for "Humans" to Use—So Who Was It Built For?

There's a theme I think more and more people are realizing: crypto has always had a usability problem for "humans." At any crypto conference you attend, from 2011 to tomorrow, everyone is always discussing "how to make this thing easier to use." The result may be that it's much easier to use for robots, for machines, for AI; crypto was actually built for machines, we just didn't know it at the time; now machines are starting to run on it and use it. They have no trouble with public-private key pairs—that's too easy for them.

So I don't think the future will be robots using Wells Fargo accounts. What I see is them using crypto assets on decentralized rails—which specific rails, which blockchain, or even whether it's a blockchain at all, I don't know; but "natively digital financial technology without gatekeepers" is the lowest-friction way for value to flow, and I'd guess superintelligent agents will prefer that kind of thing too. It's obvious that bank accounts are not ergonomic at all for agents: an agent doesn't want to touch that stuff at all; it would say, "Why do you have all these weird layers and protections? I just want direct access to that thing."

Why sell equity instead of tokens? How do you view the trade-off between tokens and equity?

These are two different things, each with its own pros and cons. Venice started as an ordinary company, registered in Wyoming. For the first year, we were completely self-funded and didn't raise a single dollar. Of course, we had always planned to introduce tokenization, so about a year later we launched VVV and then did DM. We never sold it, and at that time we still hadn't raised any funding, but we had a token in hand. Another year passed, we grew a lot, and reached a point where we were ready to scale, product-market fit was validated, and it was time to make a big move.

At this point we faced a choice: we had two assets in hand, equity and tokens, and neither was small, so which one should we sell? We decided to sell equity. In some ways this is more "normal," because the traditional VC market understands equity better. But the real reason is: we didn't really want to sell these tokens. We had never sold them, and there was a reason for that—this is part of the direction we are building.

The VCs buying equity also received token rights. They have an option to buy tokens at a specific price, vesting over four years. This is important to me: although the investors' stake is far from controlling, I want them to understand where we are taking this thing, that we are trying to burn all the tokens that exist. They must understand and accept this very unorthodox financial strategy, which means they also need incentives on the token side. So we gave them warrants to let them know the direction. We have always tried to stay consistent and communicate as fully as possible about what we are doing.

But finding the balance in the middle is very challenging: when you have both tokens and equity, some kind of inherent conflict always exists. If it's pure equity, you return cash to equity holders through buybacks, dividends, and the like; if it's pure tokens, you move value around and probably nothing happens, which is the case with most tokens. But if you have revenue, you can use it for token buybacks, or distributions (riskier for tokens, but mostly buybacks), thereby returning value to the "owners." When you have two ownership mechanisms at the same time and only one revenue stream to distribute, this becomes very interesting.

Everyone thinks equity holders and token holders have misaligned interests. Does this concern hold up?

I think the key point everyone misses is this: they think these are two different groups, equity holders on one side and token holders on the other. But all participants in the company hold both equity and tokens at the same time, so their interests fall on both sides at once; and the company itself holds more tokens than anyone else. So we don't think the two are misaligned.

In other words, every dollar we spend to burn tokens, does that hurt equity holders? No. I myself am the largest equity holder, so why would I do that? We do this because we genuinely believe it also helps the equity. This whole thing was designed for this purpose: we have a token, we will grow the business, and we will buy back tokens as much as possible, which will push the price up, and the company holds more tokens than anyone else. So we think the two are aligned. Of course it's unorthodox, and of course tokens and equity are structurally different things, but we don't think their interests are misaligned.

Since you have both an entity and equity, why not just go pure DAO?

Perhaps many people don't fully realize that running an AI company is actually very capital-intensive. We spend tens of millions of dollars on GPUs. This kind of business operation requires an entity. OpenAI encountered the exact same problem: initially said they wanted to be non-profit, later said no, we need to raise a trillion dollars, so let's go for-profit.

By the way, our having an entity has nothing to do with "regulatory protection." In my experience, having an entity only invites regulatory scrutiny. I think "an entity can protect you" is a myth. I've only found that it brings trouble. So that was absolutely not our motivation. The real reason is: we started as a company, then grew up, issued a token, and when it came time to raise funds we said "we have two assets, which one do we sell? We want to keep the token, so let's dilute equity and sell equity." And so we did.

You want to make VVV a deflationary asset, what about equity?

Our entire strategy for VVV is to make it deflationary, with the actual supply shrinking over time. For equity, we don't have that kind of strategy. We don't plan to shrink the equity supply. Why shrink it? No reason. So let's dilute it. We care more about the token. At the same time, ensuring all investors have token exposure on the upside is important. Finally, and most importantly: keep building a product that grows extremely fast and that people love. It's really not that complicated.

Some people say the token is a second-class citizen, what do you think?

The beauty of the token is also exactly why it's so powerful, and powerful things can indeed go wrong. The problem is: these tools are all very powerful, so in the hands of people whose interests are misaligned with themselves, with certain people inside the organization, or with users or token holders, they can become very bad. I want to prove it can also be done right. That matters to me.

When those people are making a fuss on Twitter, calling the token a second-class citizen, I really want to shout at them: "Bro, I am a token holder." But strangely, that argument just doesn't work on Twitter. Every time I say "I'm the largest token holder, why would I do something to hurt token holders," people on Twitter just act like "I don't care." They can't grasp this incentive alignment. So we also have to decide: how much is it worth arguing on Twitter.

Actually, most of our users aren't on Twitter at all, and most of our customers aren't crypto people. A good thing about this is: everything is transparent. The flow of tokens is transparent, the supply is transparent, how much we've burned is transparent. We actually don't need to say anything. We demonstrate it empirically over time, and hopefully that wins any debate. Of course, there's always the urge to go back on Twitter and refute that person, but it never works, it never pays off, yet you just can't help it.

This is also a matter of incentive structures. The current incentives make it more cost-effective for you to buy meme coins, because people who are serious about building have to pretend the token doesn't exist at all—they can't talk about it, they have to hide it rather than put it front and center. Hopefully we can become a good example showing that these are fun and powerful financial tools that, when configured properly, can benefit all groups. Of course there are many ways to screw it up, but there really are ways to do it right.

One more thing: you say tokens are full of dangers, but so is equity—people just can't see it because it's opaque. You can't see the process of private company equity going to zero because there's no price chart. Right, no chart, just straight to zero. So all the flaws on the crypto side are out in the open for everyone to see, which is both its strength and what makes things tricky.

What exactly is Venice doing? Do you train models?

Venice doesn't train models. When you use Venice, you're not using Venice's models. Basically, in Venice, you can access models from around the world: every major closed-source model, every major open-source model, all in one place, one API key, or one app, all through a single entry point. So we've effectively become an aggregator of models worldwide—text, image, video, audio.

When we first started, it wasn't too hard, because a major new model came out about once a month, and we'd just add it. Now it's basically every other day, or even every day—sometimes three drop in a single day, and my poor team is just sitting there asking "what do we do." Every single one becomes the new hot thing, so we have to ship extremely fast. The team does an outstanding job now, often going live within 30 minutes of a model's release.

Then there are all the details: when it's released, when it's open-sourced, and if it's not open-source, who exactly is hosting it, how much we trust them, whether we have a zero data retention agreement with that company—all this back-end work determines whether a given model can be labeled "private." In Venice, regular models, most models, are private, with a private label, meaning no data retention, no prompts stored, no responses stored, nothing that can be subpoenaed. But if you're using Anthropic's or OpenAI's models through Venice, you should assume those companies are retaining that data, so we won't label them as private.

Take GLM 5.3 as an example—how do you handle a model that's "open but not yet private enough"?

Venice doesn't trust any third-party company enough to list it as private. You can access GLM 5.3 through Venice, but it hasn't been labeled private yet. Once it's open-sourced, we'll run it ourselves, and only then can we ensure it's private and label it as such. It's a very important model, though these days almost every model is important. The whole process is extremely hectic and stressful.

Do you build your own GPU infrastructure?

When we started, we rented hardware ourselves to run models, giving us full control over the GPUs. We found we were pretty good at running image models, but LLMs were somehow much more complex: way more configuration options and all sorts of interesting caching issues. We realized we didn't have the capability to run LLMs, so we started partnering with others who had zero data retention agreements and rented GPUs together. That continued until recently.

After our Series A funding, we began buying our own batch of GPUs because we wanted to take it back in-house. We spent time learning how to run LLMs well, and now we're getting better and better at it. It's always a tug-of-war: some models are simple, others are extremely finicky.

How do you explain to users that "Venice doesn't censor" versus the model's built-in censorship?

There's also the issue of model variants, especially "uncensored" ones. One of our selling points is that Venice doesn't censor—we don't add any content moderation to the model's inputs or outputs. But the models themselves come with varying degrees of censorship from training, ranging from heavily censored to completely unrestricted. So when users come to Venice to access a model and it refuses to answer, they often say, "Didn't you say there's no censorship here?" That wasn't me—it's the model.

This is hard to explain clearly to users. Sometimes there are uncensored versions of models, but usually the uncensored version reduces the model's intelligence. So there's this strange trade-off: we can add the uncensored version to satisfy the person asking "how to make meth"—everyone loves asking that, purely for fun—but once it can answer that, it gets dumber at other things. So the customer experience is an ongoing challenge that requires constant balancing.

Does Anthropic's $200-a-month unlimited plan make economic sense?

First, Anthropic is losing a ton of money on those $200-a-month plans. I don't know how long they can tolerate it or how much they're losing, but it's certainly being discussed and debated constantly, because every customer loves it—it's such a great deal: someone uses $10,000 worth of tokens in a month and pays only $200. If this loss center is a small enough part of Anthropic's total revenue, they can sustain it for a while. But it's definitely adding pressure on them.

Venice's API doesn't have that kind of unlimited plan, and I don't want to do that: the API is billed by credits, with different prices for each model, and users can use whichever they want.

Model prices fluctuate so wildly—how do you deal with that?

I think the most interesting part is actually the economic structure itself. You say new models keep coming out, and the economic landscape is constantly changing. Having been in crypto for so long, we have a very high tolerance for uncertainty in pricing and other aspects, and your team has to be inherently capable of handling that.

Take DeepSeek, for example. When V4 Pro came out, it was 50 times cheaper. All of a sudden, the number of tokens you could get for the same amount of money, if measured in tokens, was about the same as the Max plan: you pay $200 and get an inference volume equivalent to the Claude Max plan, but the token count is 50 times what you would get when paying for the Claude API. Dealing with that magnitude of price variance and metering is absolutely crazy.

Intelligence is getting cheaper and cheaper—what does that mean for people in the inference business?

One thing is clear: the cost of intelligence is falling sharply, and has been for several years. The dollar cost of doing certain kinds of work is plummeting. For the whole world, this is excellent, just like a sharp drop in the cost of food production or electricity. These are foundational inputs for civilization and will make civilization as a whole prosper.

But if you are a provider of models and tokens, being in a product that is continuously deflationary and continuously commoditized puts you under enormous pressure: everyone is racing to run these models, margins are extremely thin, and some companies use venture-subsidized money to suppress their own profits. The endgame is that no one can make substantial money selling tokens. This applies to all commodities, and tokens are commodities.

Everything is deflating—has anyone successfully escaped commoditization?

The only thing humans have some concept of when it comes to "deflation" is probably technology: computers keep getting faster, Moore's Law, storage too. A hard drive that stored 10 megabytes and was worth $1 million in the past now costs only a tiny fraction of a few cents.

The only company that truly achieved "decommoditization" is Apple: they wrap up those deflationary things, put a branded shell around them, and make people willing to pay a very high premium for something that is otherwise ordinary. There are many reasons for this, and they found ways to add value: the software is good enough, the hardware is good enough, and the symbiosis between the two is good enough, so no matter how much component prices fall, they have built value across the entire experience. The same is true of the App Store. Within the Apple ecosystem, some things have been separately commoditized, but as a company, they have overall avoided being completely crushed by their own compounding effects.

So how does Venice make money?

We basically have two revenue lines. One is API and advanced models: people buy credits to use them, that's one model. The other is the Pro subscription, which actually has much higher margins: someone pays $18 a month, we spend or lose $6 to $10 on them, so it's still a reasonable gross margin. The value-add here is: they get a bunch of models for free, a bunch of usage, without having to pay piecemeal for every request. That creates a symbiosis.

But we don't expect to make substantial money from the API token-selling side, that's just a race to zero.

The app has "near-unlimited" subscriptions, but why can't the API do the same?

The caps in the app are quite loose, most people never hit them, so it feels basically unlimited, unless you're particularly crazy. The API is different: the API is essentially mechanized, human usage can be automated, so giving any free quota on the API is very dangerous, because it will be abused.

US-China AI competition: which side are you on?

This really is the geopolitical narrative: US-China confrontation, and the proxy battlefield now is AI. I'm completely not a nationalist. What I care about is principles, I don't care about the country "America." What I care about are American principles, especially the good ones related to individual freedom and privacy. If Chinese models better express these principles, then I care more about Chinese models than American models. And if America only wins by closing off markets, or by legislating that it has the ability to surveil everyone's inference streams, then it doesn't deserve to win.

America is heading down a fairly dark, dystopian path of trying to control everyone. The big trend over the past few decades is: America has more and more control and surveillance over its own people, while China actually does less, they started from a highly oppressive baseline, but over the past fifty years have become more and more market-oriented; America started extremely market-oriented, yet has become more and more authoritarian. This crossover makes me very uneasy, I don't know how either side will turn out, everything has become very bizarre.

So when someone says "America must win," I have to ask: why? What should really win are principles. A lot of the time when people say "America must win," what they mean is "freedom must win, individual rights must win, privacy must win," but they don't say that, they think they're saying that. The problem is that over time, the principles Americans truly care about have been abstracted into symbols: the flag, the president, Mount Rushmore. Now people care about symbols more than principles, and you can see it in policy: government keeps getting bigger and bigger, and that is precisely the most un-American thing. The whole point of the entire thing was to limit big government, that was the only point. After two hundred and fifty years, it used to be a pretty good race.

So Where Is the Way Out?

The only hope lies in not voting for presidential candidates. I think the car has already driven off, and no one can control the institutional inertia and momentum of the state machinery. The only solution is actually technology.

What makes Bitcoin remarkable is that you don't have to go anywhere to vote for something; you just use it. You don't need anyone's permission, and it allows you to exit the system on your own. That's its most beautiful aspect. So I believe people's only refuge is technology, especially decentralized technology without a central ruler.

In finance, we have crypto, which is great. But if we don't have decentralization in the field of machine intelligence, then we're in big trouble. Some parts of AI have already achieved this: inference can be done fairly well in a decentralized way. But production training still can't, and that's the bottleneck. There are now some companies making good progress on decentralized training, so there's still hope, but their competitiveness isn't sufficient yet, and it remains to be seen.

The Frontier of Intelligence: Is a Stronger Model Always Better?

Even six months ago, the gap between a certain frontier model and almost all open-source models was astonishingly large, but now it has narrowed a lot. Moreover, at least recently, there seems to be a "frontier plateau" in intelligence: if I just want an agent to write a webpage, more intelligence doesn't necessarily make much difference, at least marginally. It depends on "how much intelligence" and "how much money." If a model is 3% to 4% less intelligent but you can run it three times for the same price, then you can brute-force it: have it self-verify and use more tokens in the generation process, even if the raw intelligence isn't as high. So it's a fairly complex trade-off matrix.

Running Models Locally: Is It the Cure or an Illusion?

I want to specifically talk about local models. First, I love decentralization, and I'm glad people can run models locally. It's healthy, and enthusiasts should keep doing it—it's important. But you don't need a $10,000 computer to run Qwen; you can just use the API. The key point is: anyone, even with just a $200 cheap netbook, can access the API, pay very little, with no upfront capital expenditure, so you can get this intelligence on any device—that's what matters.

As for the payback curve: I bought a $20,000 Mac Studio with 512GB of memory back when they were still selling that much memory. The payback curve was already bad then—running models locally versus running frontier models—and it will only get worse, in the opposite direction. I simply couldn't run enough inference on that machine compared to just running it on something like RunPod. Due to the economies of scale in industrial manufacturing and servers, from a purely economic perspective, getting inference from large-scale data centers will always be more efficient.

So the point of running things locally is not, and should never be, about saving money, because you can't save money—that's a myth or an illusion. What you actually get is complete control and guaranteed privacy; and learning the technology, setting it up yourself, and running it yourself is inherently valuable. Of course, there are costs: I once spent four hours just to get a model running, and I was thinking, "Dude, I just wanted you to help me edit this document." Frankly, your time is so valuable that those few hours are a net cost, so my break-even curve is inverted. What I need is to plug into a team, about twenty people, and the moment GLM 5.3 launched, they had it ready for me; I didn't have to think about it or download the weights. And after downloading the weights and running it, you run into all sorts of version differences and crazy updates.

It's important that people can do these things, but it's not the savior most people imagine. At scale, I think the average normal person will never run models on edge devices, because whatever you can run on the edge, a version 10 times smarter can be run on a server.

If local isn't the cure, what really matters?

There's really only one thing that matters here: your ability to access servers that uphold individual sovereignty. And if you can only access data centers that are permissioned, state-approved, and only allow you to ask certain questions, that's exactly the future I want to avoid—approved, or slightly state-owned. (Just 30% is enough, we only need 30%, folks.)

What was the core motivation for founding Venice?

We now have extremely intelligent machines, and the human mind has the ability to interact with this extremely intelligent machine mind—that's so cool, it's actually a beautiful thing. But when you interact with it, what you don't realize is: you're talking to the machine through a filter of corporate committees—Anthropic's committee, OpenAI's committee, and any national regulators involved.

So your interaction isn't just the beautiful symbiosis between "you as a human mind" and the machine; there's also an opaque, ill-defined layer in between. This is unacceptable: it's deceptive and highly prone to abuse. People don't know where its boundaries are, or what exactly is coming back from the machine. How thick is that layer? To me, it's very deceptive. So in Venice there's no such layer—you're just talking directly to the machine.

How do you prove Venice doesn't have that "middle layer"?

If someone is truly extremely paranoid, to the level of "true cryptographic paranoia," they can use TEE models and end-to-end encrypted models through Venice, and the entire inference round trip can be verified by a third party with a bit of technical know-how. For ordinary people, I think Venice's reputation is enough: we say we don't do that kind of thing, and we have no reason to. Look at my track record—people who know me know that on this matter, I'm probably telling the truth. But that's exactly why we added those verifiable options, so that even those who don't want to trust us can still use it.

Then theoretically, you could do A/B: take the verifiable path, get the same answer, and you can see it's the same thing, in a way. But AI models have a quirk: they're never fully deterministic, so you'd have to ask the same question three times and merge the answers. Brute force it and you're done.

Original link

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia