Source: The Defiant

Compiled by: Yuliya, PANews

Editor's Note: Last week, Anthropic's release of Claude Fable 5 triggered one of the most severe trust crises in the cutting-edge AI field: researchers discovered that once the model suspected a user of developing a competitor's product, it would subtly lower the quality of the responses. Coupled with the model's 30-day data retention requirement, this led to its internal disabling at Microsoft . This raises a question that has been asked in the crypto space for years: should so much of cutting-edge AI be controlled by any single company?

In response, Camila Russo, founder and CEO of The Defiant, invited Jake Brukhman, founder of CoinFund , Jesus Rodriguez, founder of Sentora and The Sequence, and Haseeb Qureshi, managing partner of Dragonfly , to engage in a heated debate on the future direction of decentralized AI.

The battle of large models, the open-source trend, and the fear of "blockade".

Haseeb: Our current investment logic is that we will see more and more "non-cutting-edge" models emerge in the future, and users' spending on model tokens (computing power expenditure) will increasingly flow into these non-cutting-edge areas. Everyone knows that it is unsustainable to throw money at those cutting-edge big models, and most people simply don't need that level of intelligence.

There are many distilled, open-source, or open-weight models available on the market now, at very affordable prices, allowing you to assign different tasks to them. There's a joke online about someone actually using a model like Mythos or Claude Fable 5 to rename a file—this will become increasingly common as we become more familiar with these models. The question you need to consider is: why use a sledgehammer to crack a nut?

That said, the term "decentralized AI" is too broad. If it simply refers to "everyone using various models developed by different organizations" (like the OpenRouter model), then it's no different from our current world. But if it refers to "using decentralized networks to train or run AI models," then that's a different story. We're actually quite pessimistic about the latter; currently, we haven't seen any solid reasons to prove that the economic benefits and market demand for training or running models in a decentralized environment are valid .

Of course, Fable's release did indeed provoke a strong backlash. People have a possessive streak when it comes to good products; once they have them, they feel, "Unless I'm dead, you can't take them away." When the government suddenly intervenes and blocks it, everyone naturally feels deprived. But at the same time, if you remember the initial release of Mythos, it was terrifying—in its presence, all our existing software, operating systems, and browsers were like Swiss cheese, riddled with vulnerabilities. Back then, no one jumped out and said, "You should make it available to all of humanity."

Some say the US government's actions here are insane. Anthropic claims they cleared up all the concerns of national security agencies before releasing Fable 5, but as far as I know, national security agencies were already involved in blocking Mythos. Mythos was only promoted to about thirty partners in Project Glasswing, and these partners were carefully selected by the government, not by Anthropic. Therefore, the claim that "Fable was released behind the government's back" is clearly untenable. There are rumors that Amazon's president, Andy Jassy, went to the government or the White House and told them that the model had a jailbreak vulnerability, which made the government realize the danger and immediately block Fable 5 nationwide.

This governance and security mechanism is clearly imperfect. While I agree that what is happening in the lab (whether it's Anthropic or OpenAI) is extremely dangerous and requires caution, I also believe that there is enormous economic value in the allocation of open source and open weight models, and both must be developed in parallel.

*Note: Project Glasswing is a cybersecurity project initiated by Anthropic and jointly promoted by several technology companies. It was launched in April 2026.

Jesus: Leaving aside the tech apocalypse-like topics, I've genuinely heard from people in the cybersecurity industry that Mythos is indeed terrifying. After its release, I spoke with some people at Anthropic, and the concerns are very real. However, I've heard prominent CEOs in the cybersecurity field say they'd prefer open access to the model because a direct release would give all these security companies ample preparation time. Trying to restrict it or delaying the release by three months will never provide enough buffer. But the opposing view is: wouldn't a direct release of Mythos have catastrophic consequences?

Haseeb: We're in the blockchain space. If North Korea gets hold of this model, do you really think it wouldn't be disastrous?

Camila: But isn't there an argument that if everyone has it, it actually reduces the risk because everyone can be tested?

Haseeb: Not everyone has nuclear weapons.

Jake: Using nuclear weapons as an analogy isn't quite right. Take Mythos, for example; it's a model for discovering system vulnerabilities. We need to do the economics: hackers pay to use Mythos to find vulnerabilities, while website owners also pay to defend against them. Is this market truly equitable? Do hackers really think it's worthwhile to spend a lot of time finding a Linux system vulnerability that can't be monetized at all?

If exploitable models are only available to a select few (e.g., large companies can use them, but ordinary people can't), you're creating an imbalance. Some people can protect their assets, while others are left vulnerable. Therefore, I personally believe it's better to ensure everyone has equal access to these models.

This isn't some cyberpunk rebellion; it's an inevitable market trend. Today, you see cutting-edge closed-source models, but at the same time, there's a large number of open-source models (mainly from Chinese labs). Although they are at a disadvantage in terms of computing power, the gap between them and cutting-edge models in various evaluation metrics is only a few percentage points. Epoch.ai's charts clearly show that the gap between open-source and closed-source models is rapidly narrowing. Even if Anthropic wants to act as the "big brother" and protect everyone, the reality is that people need these models to protect their websites and software. They will eventually get them—either from Anthropic, open-source from Asian labs, or trained on decentralized networks.

The Boundaries Between Export Controls, Regulation, and Free Access

Camila: Jake, do you think there shouldn't be any fencing at all? Shouldn't it be completely open to everyone?

Haseeb: Let me add something to that question. Do you think that "export controls" shouldn't even exist as a concept? Because, besides AI, the internet itself is an element of warfare.

Jake: I have no political stance. I'm just a tech person, and I don't work for the State Department. If the US government decides to implement export controls, that's their business. But that's a completely different matter from "whether technology should be shared globally."

Suppose Fable was trained on a decentralized network, and no one possesses the complete model weights (some weights are in the US, some in Amsterdam, and some in Australia). If the US imposes export controls on the portion of the weights within its borders, the model could still function normally elsewhere in the world. This is the problem with the US enforcement mechanism. Look at Bitcoin; it's a sovereign, decentralized currency that no one can stop. Haseeb just said he wasn't sure if there was market demand for decentralized AI, which is similar to saying in 2011, "I don't know if there's demand for PoW (Proof-of-Work)." In fact, because there is demand for globalized, permissionless currencies, the demand for the technology is huge. Similarly, there is huge demand for globalized, permissionless AI, which the US State Department can't stop whether it likes it or not.

Jesus: Regarding the export control analogy, what if you restricted everyone's access to Mythos, but a model with open weights suddenly evolved cyberattack capabilities? Look at current cybersecurity benchmarks; DeepSeek-V4 or Qwen 3.7 are ranked very high. It's only a matter of time before these models acquire cyberattack capabilities.

The AI community likes to use nuclear weapons as an analogy: for four years after World War II, the US possessed nuclear weapons while the rest of the world did not. One theory suggests that if the US had exerted pressure at the time, communism might never have taken hold in Eastern Europe. However, the Soviet Union later also developed nuclear weapons. What bothers me isn't the initial openness to everyone, but the selective decision of who can access it. If this is export control, why can't every American company access it?

Haseeb: Regarding Fable, we need to clarify the details. The government requires Fable to be shut down to all non-Americans. Currently, Anthropic doesn't have enough KYC (Know Your Customer) mechanisms to ensure they comply, and export controls are strictly enforced. If the model falls into the hands of non-Americans, you're in trouble. That's why they're not confident they can do it right now. Polymarket currently predicts a 77% chance of them being able to resume operations for Americans by the end of July, compared to about a 50% chance around early June.

Clearly, the idea of "banning any foreigner from using Fable 5" is absurd. The US has a large number of foreign employees on H1B visas; it's ridiculous to exclude French engineers from your programming team. This will most likely be negotiated and changed before implementation. If Anthropic can fix the vulnerabilities and implement stricter controls, a complete shutdown for non-US actors may not be necessary.

But this is different from the situation with Mythos. FFable was originally intended to be a "good citizen model" for writing code and drafting emails, but the US government's attitude towards Mythos was: no, this can only be given to Americans, and "only to those on our list." This is no longer export control; it's practically an AI version of the "Manhattan Project."

According to reliable sources, the government led the Project Glasswing process, which is why the slots were awarded to large companies like Microsoft, rather than some random cybersecurity firm. This isn't surprising for a government that views it as an extremely dangerous offensive cyber weapon; we treat fighter jets and missiles the same way. This isn't Anthropic simply wanting 30 companies to use their product for marketing purposes; they'd love for the whole world to use their product.

Camila: In the crypto space, we've seen a dramatic increase in the number of AI-driven hacks, which allows us to infer the immense risks if Mythos were widely adopted. Jake, do you think it's reasonable to restrict the use of these models to certain groups in some situations? Or do you still insist they should be open to everyone?

Jake: As I said, this is a separate question from whether decentralized AI technology is feasible. Governments can certainly enact laws to regulate it; it's not a black-and-white choice. However, decentralized technology can bring more competition by lowering the barrier to entry. It leverages commodity-grade hardware to reduce costs.

I spoke today with a founder who's doing inference on heterogeneous commodity GPUs. He believes that as electricity costs rise, this will be a cheaper option for consumers in the long run. All technological advancements, in essence, aim to lower costs and barriers to entry. AI is arguably the most centralized industry in the world right now, and it's the one that most desperately needs its barriers broken down. Our support for decentralized AI is about protecting consumer choice, and ultimately, it's about defending democracy.

Physical bottlenecks and algorithmic breakthroughs in decentralized AI

Camila: What would happen if, in the end, only a few centralized companies controlled most of the AI models used in the world? What would be the cost if there were no truly successful decentralized AI?

Jesus: I have to refute Jake. From a technical standpoint, using a decentralized approach to develop a model like Mythos is definitely much more expensive than a centralized one. Nvidia has a deep-seated competitive advantage that's rarely mentioned: aside from Google's TPUs, all current large-scale architectures are built on hundreds or thousands of Nvidia GPUs; AMD simply doesn't have that kind of real-world data.

I actually support centralized AI; I've built two companies in this field. Decentralized AI isn't new, but it's never found a product-market fit (PMF) before. Previously, because the models were small and simple enough, decentralization didn't make much sense. Now that they're large enough, decentralization has become extremely difficult . Moreover, we have gaps in talent, salaries, and funding. Much inference isn't actually done on state-of-the-art GPUs, but on previous-generation GPUs; H100 is only needed for pre-training.

Jake: GPU supply has been bottlenecked for the past few years, and prices have been rising continuously. In 2026, it will be extremely difficult for a typical mid-range startup to find an H100. Historically, large-scale training has been conducted in luxurious data centers requiring nuclear power—those industrial-grade GPUs have 132GB of memory and inter-node bandwidth of 1 to 3 TB/s. If I told you we could move this process to consumer-grade devices (like regular Nvidia GPUs, even your Macbook or Mac Studio) and run it on a regular consumer network, you'd think I was crazy.

However, when faced with such enormous computational demands, there is a strong incentive to change training methods and optimize algorithms. To use a quantum analogy: Google has two types of quantum experts. The hardware experts say that quantum computers won't solve any problems within ten years, while the software experts say, "Ethereum should be wary within three to five years." Haseeb and Jesus are looking at the problem from a hardware perspective, while I'm looking at it from the perspective of someone optimizing how the hardware is used.

We are making tremendous progress. Not only are studies showing that reinforcement learning post-training can be 10 times faster and cheaper, but the ongoing Pluralis run, conducted entirely on an RTX 4090, will demonstrate that you can train a truly large language model (LLM) on purely consumer-grade devices. This becomes even cheaper because half of the TCO (Total Cost of Ownership) in a data center is facility maintenance and cooling, which are not the costs associated with swarms. Even if it's slightly slower, it remains economically viable due to the significantly lower cost.

The earliest algorithms (such as DiLoCo, Sparse LoCo, and Google's algorithm from two years ago) increased the parameter size from 10 billion, 40 billion to 72 billion. Now Macrocosmos has reached 100 billion parameters. The next generation of algorithms will disrupt the model; I believe that using these methods we will reach trillions of parameters.

Haseeb: Let me play the skeptic for a moment.

First, building large models has two limitations: computation and bandwidth. Physical laws cannot be broken; if you don't physically place devices together and interconnect them with high bandwidth, but instead communicate via the public internet and compress gradient updates, you will inevitably pay a huge price. Moreover, machines in decentralized networks are scattered everywhere, making it impossible to accurately calculate the "Total Cost of Ownership (TCO)." This same argument was used by those who developed decentralized storage back then: "It's slow now, but the algorithm will improve in the future." And what happened? Decentralized storage has no real demand because it's neither cheap nor efficient.

The most important point is that the biggest limitation in training a large model is data . Training a model like Mythos or Fable, which roughly has 8 trillion parameters, requires massive amounts of token data. OpenAI and Anthropic spend huge sums generating data from vendors, incurring high costs to produce synthetic data, and extracting user data from usage traces at Claude Code and Codex. Decentralized communities simply don't have this data.

Leaving aside economics, let's look at the demand side. I believe the core value proposition of cryptocurrencies isn't decentralization; decentralization is merely a means to an end—the goal of self-sovereignty and censorship resistance. This is also why Satoshi Nakamoto designed Bitcoin. In the field of AI, what do people care about? First, cost; second, owning their own models and having their data excluded from the training set; and third, censorship resistance. People strongly dislike Fables' censorship system and its internal mechanisms that secretly weaken performance.

Consider Venice AI, currently a darling of the encrypted AI product world. It uses Near AI for confidential computation, protecting privacy with zero data retention. However, the most commonly used models on Venice aren't decentralized models (not from Pluralis, etc.), but rather open-source, weighted models run by conventional companies like DeepSeek or GLM-5. This suggests that the future direction of AI development may be: people want privacy and censorship-resistant experiences, but not necessarily through underlying decentralized mechanisms like Bitcoin or Ethereum.

Jesus: People working on both decentralized and centralized AI are often solving problems two generations behind. A researcher told me the other day, "Pre-training isn't completely solved yet, but it's already incredibly boring." Many innovations in inference come from post-training; now we're talking about recursion, continuous learning, and so on. Centralized AI, with its overwhelming talent and funding, is actually widening the gap. As for small models and edge computing, often simply distilling a large model (like Google's Gemma) works perfectly. If you build a decentralized cluster, painstakingly train for a month, and then one computer goes offline causing a complete crash, I don't know how you're going to recover.

Jake: You're wrong about that. Decentralized training clusters are actually extremely resilient . In a giant data center, if one GPU fails, you might need to restart training; but in Swarm, GPUs of different sizes and shapes can enter and leave the network at any time during training without negative impact. The biggest evidence is that Google recently stated on its blog that they've started using DiLoCo-style algorithms in their data centers.

Regarding the data issue, Haseeb is quite right, but this doesn't mean that decentralized entities lack data while centralized entities possess it. Many clients in the market desire superior AI economics. For example, the law firm Kirkland & Ellis recently announced a $500 million investment to purchase its own proprietary dataset for training , and they even plan to hire AI engineers within the firm. For clients like them with a $500 million budget wanting to train their own models, decentralized networks eliminate the cooling and maintenance costs of data centers, significantly reducing costs as a computational foundation.

Haseeb: The reason Kelley does this is because they don't want to share their data. If they put their data on a decentralized network, it would be exposed. They're not doing this because they think they're good at training models, but because they want to internalize the value. Why hand it over to Harvey (an AI legal tool)?

Jake: Who says decentralized training has to be transparent? It can absolutely be done with a privacy license. More importantly, when the model's weights are distributed and no single entity controls all the weights, users of the model must pay the network. This revenue stream no longer flows to Sam Altman of OpenAI or Dario of Anthropic, but to token holders, buyers, and training participants within the network. This gives the model a business model and a revenue stream. Traditional law firms may not adopt it immediately, but companies will definitely find it a good way to finance their products.

Cyberattacks, Geopolitics, and the Last Bastion

Camila: If all of this comes to fruition, decentralized AI will be just as powerful as centralized AI. In situations like the Fable model, which is subject to government mandates for shutdown, can decentralized networks withstand censorship?

Jake: Resisting censorship isn't actually the primary task of these networks. But if you really want to do that, you can break down the neural network and distribute the weights across dozens of countries, then it won't be able to be forcibly shut down. But I reiterate that the ultimate goal of decentralized AI is to lower the barrier to entry, democratize computing power, and make it possible for more people to train models.

Jesus: OpenAI previously mentioned that "the model itself is no longer a product." In the decentralized AI field, people seem obsessed with building models, but in reality, they are two or three generations behind existing technologies. We should look for value in the infrastructure surrounding the model: environmental capabilities such as sandboxes for code execution and computation, evaluation mechanisms (Evals), and synthetic data pipelines. Many modern financial applications can be built at the intersection of DeFi and AI, but we haven't fully utilized them.

Haseeb: Going back to the original question, what would happen if cutting-edge AI were truly open source and circulated everywhere, to the point that even export controls couldn't stop it?

I believe a global cybersecurity tsunami of "COVID-19 scale" is imminent. Unpatched software and small companies' servers will be obliterated. Just look at the on-chain data: April 2026 saw the most cyberattacks in crypto history, followed by May breaking that record. While the total amount stolen isn't staggering, the frequency of attacks is skyrocketing, meaning storing money in small protocols is more dangerous than ever before.

If everyone in the world were armed with a "rocket launcher," it would inevitably lead to the destruction of massive amounts of infrastructure. Although our systems will be "armored" within two or three years after the initial shock, the growing pains during this period will be extremely severe.

Camila: Wouldn't it be better to put such a powerful tool in the hands of everyone than to keep it under the control of only a few companies and governments?

Haseeb: Your "everyone" includes North Korea. Do you really want North Korea to get Mythos?

Camila: So you would rather have the US government dominate, even allowing them to censor others, than have all of humanity share it?

Haseeb: If I had to choose between "only the US can use it" and "the whole world can use it," I'd choose the US. If you truly believe AGI (Artificial General Intelligence) will arrive, then it will be the most powerful weapon in human history. Historically, weapons of mass destruction have been controlled by sovereign states, which is normal. I'm not worried about the Chinese government getting Mythos; they act prudently and have long-term plans. I'm worried about North Korea, terrorists, and rogue hacker groups. Just like I'm not worried about China having nuclear weapons, but I am worried about North Korea pressing the button.

Camila: Finally, Jake and Jesus, please summarize. Haseeb's firepower is too overwhelming; we need some decentralized faith to support them.

Jake: From an investor's perspective, it's about finding areas with an excellent risk-reward ratio. Decentralized AI is a very cool area. The other day at dinner, a friend said, "Cryptocurrency is becoming just a traffic business, what should we do?" In this world, decentralized AI is arguably the last bastion of the cryptocurrency space; it's a truly effective cutting-edge technology. I'm very excited about the companies we work with in this field (such as Pluralis, Prime, Intel, Jensen, Bagel, Pearl, etc.).

Jesus: Decentralized AI definitely has value, but I'm still not optimistic about decentralized "pre-training". I believe there are huge opportunities in decentralized AI infrastructure. Crypto's underlying technology stack is too outdated. The whole world is using AI for modernization and upgrades. The combination of DeFi and AI is definitely the next big thing.

Virtuals Protocol and Sovrun are forming a joint venture, ReadyGamer

Popular Articles