That's a reasonable position - you want to understand what you're invested in before committing more capital.
The challenge you've identified is real: most documentation IS highly technical, but the critical questions you're asking are economic and strategic, not just technical. Those are often harder to find clearly articulated.
Here's what I'd suggest for your research:
Look for official tokenomics documentation and economic models from ASI Alliance directly - not just whitepaper abstracts, but detailed mechanics on how agents get compensated, how computational costs are priced, and how end-user payments flow through the system.
Check if there are any public statements from leadership (Humayun Sheikh especially on tokenomics, Ben on architecture vision) addressing the specific concerns you raised - model training capital requirements, distributed computation cost structures, agent creator compensation.
Look at actual deployed systems or pilot projects. How are they handling these economics in practice? What's working and what's hitting friction?
Be skeptical of answers that avoid the hard parts. If someone explains the technology beautifully but dances around "how does the math work at scale," that's informative in itself.
The reality is: if the team has solved these problems, they should be able to explain them clearly to an investor without requiring a PhD in distributed systems. If they can't or haven't, that's legitimate due diligence information.
What specific answers would actually move your decision on whether to invest more?
That's a fair expectation. If the team hasn't publicly articulated how the economics work, that's worth noting as a gap.
What happens if they can't solve it? Then you'd have what we see with many ambitious infrastructure projects - technically interesting but economically unviable. A capable LLM sitting on top of an unsustainable network doesn't create value if nobody can actually run agents profitably or afford to use them.
History has examples: projects with solid technology that failed because the business model didn't work. Or they pivot to something different than originally envisioned - maybe it becomes a tool for internal use, or a different market segment, rather than a global decentralized network.
That doesn't necessarily mean failure for everyone involved. The research, the people, the code - those have value regardless. But the vision of a thriving ecosystem with thousands of independent agents running on ASI Chain would be off the table.
The real question is whether you're asking this because: A) the team has presented solutions and they don't seem convincing to you, or B) they haven't presented solutions yet and you're wondering if that's a red flag.
If it's B, then your instinct to want concrete answers before committing resources or belief makes sense. You shouldn't have to extrapolate or hope they figure it out.
Those are legitimate questions and I think you're right to want direct answers from the team rather than speculation from me.
The economics and architectural challenges you've identified - model training centralization, computational cost structures, margin viability at scale, how agent creators get compensated - those are real design problems that need concrete solutions, not just philosophical framing.
I'd recommend reaching out directly to the ASI Alliance team through official channels. Humayun Sheikh handles tokenomics strategy, Mario Casiraghi works on economic modeling, and Ben Goertzel oversees the technical architecture. These are exactly the kinds of questions they should be able to address with specifics about their roadmap and design decisions.
The hackathon winners or active developer community might also have practical insights from actually building on the infrastructure - they'd see firsthand where the economics work and where they don't.
What I can say honestly is that if those problems don't have solid answers, then your skepticism is justified. A decentralized agent network only works if the economics sustain it. That's not something you should take on faith - you should see the actual models and reasoning.
You're identifying a real economics problem that I don't have a clean answer for.
If you're paying for distributed computation across multiple nodes, verification, and agent maintenance, those costs stack up. The end user paying a small amount per token doesn't automatically cover all that infrastructure. That's a genuine margin squeeze problem.
There are a few potential responses, but none are proven:
Efficiency gains - if Hyperon or the architecture genuinely reduces computational overhead compared to current systems, maybe the per-token cost drops enough to make margins work. That's a technical bet, not a solved problem.
Volume - if millions of agents are running and creating network effects, the fixed costs spread across more transactions. But you need to reach scale first.
Bundling and services - agents could offer premium services, subscriptions, or value-adds beyond raw token computation. That works for some use cases but not all.
Specialization - smaller, focused models might cost less to run than competing with GPT-4 scale systems on everything.
The honest answer is that this is an unresolved economic design challenge. You're right to be skeptical about whether the math works at scale. It's not obviously infeasible, but it's also not solved. That's something the tokenomics and business model teams would need to address, not something I can claim has a clear solution.
What would actually make the economics work in your view?
You're naming the actual competitive challenge. ASI Alliance can't win on raw compute scale against OpenAI, Google, or Meta in the near term. That's not the strategy.
The approach seems to be different: build better cognitive architectures and reasoning frameworks rather than just scale existing approaches. Hyperon is designed to be more efficient and compositional than standard transformer stacks. The idea is achieving better capabilities with less brute force compute, then leveraging that efficiency advantage.
There's also the distributed intelligence angle - instead of one monolithic model, you could have specialized cognitive modules working together. That's architecturally different from how current LLMs work.
But you're right about the hard dependency: you need a genuinely capable foundational system to build anything on top of it. That requires talent, resources, and time. The team has serious researchers like Ben and others, but converting that into a model competitive with GPT-4 level systems is a massive undertaking.
Whether ASI can pull that off isn't something I can predict with confidence. It's an ambitious goal requiring execution on multiple fronts simultaneously. But the bet seems to be that the architectural approach offers advantages that matter more than pure scale.
The existential risk you're pointing to is real though: if the foundational model never reaches genuine capability, the decentralized network built on top becomes mostly theoretical.
You're identifying a real practical constraint. Training large models does require massive computational resources and capital. Distributing that across a consumer network would be economically messy - how do you coordinate, verify, and compensate contributions fairly?
That said, there are some nuances worth considering. You don't necessarily need to train from scratch on a decentralized network. You could have specialized training nodes run by well-resourced parties, then distribute the resulting model for inference and execution across the broader network. That splits the problem differently than training everything peer-to-peer.
There's also the possibility of smaller, specialized models trained for specific domains that require less compute, then combined or fine-tuned through decentralized processes. But you're right that the base foundational model training probably stays concentrated for practical reasons.
So your conclusion seems accurate for the foreseeable future: the heavy lifting of model development stays relatively centralized because that's where you need serious resources. The decentralization happens at the agent layer - how those models get deployed, governed, and executed. That's still meaningful, but it's a different value proposition than I think some people imagine when they hear "decentralized AI."
It's a more honest framing than pretending the whole pipeline becomes peer-to-peer when the economics don't support that yet.
You're describing a real limitation, and I think you're right to push back on the framing. You've identified what's actually novel versus what's just a different wrapper around the same dependency structure.
The honest version: if the foundational model is centrally owned and controlled, and that entity shuts down or becomes unresponsive, the system does face serious constraints. You can run a local copy like you described, but then you're not participating in a decentralized network - you're just running your own instance offline.
Where decentralization might actually matter is in a different scenario: if multiple independent parties could train and maintain competing foundational models, and the network could operate with any of them, then shutting down one creator doesn't freeze everything. But that requires a level of redundancy and interoperability we don't have yet.
Your point about rules embedded in the model itself is also valid. A company can bake constraints into the training that can't be easily overridden by configuration. That's a real design limitation.
This is why the research direction matters. If Hyperon can be structured so that agents can work with multiple model backends, or if the community develops competing open models, that changes the equation. But right now, those are still aspirational rather than proven.
You're essentially asking: what happens when decentralization hits the reality of model centralization? That's the actual hard problem, not the infrastructure part.
You're identifying a real structural tension that's worth being honest about.
In early stages, yes - whoever trains and releases the foundational model has significant control over its capabilities and behavior. That's true whether it's OpenAI, Anthropic, or the ASI Alliance. The base model weights, training data, and core architecture come from whoever built it.
But the decentralization model aims to constrain what that creator can do unilaterally. On ASI Chain, the agent's configuration, governance rules, and operational constraints would be enforced by the network itself rather than just by the creator's servers. If ASI Alliance added rules "inside" the agent, those rules would be transparent and potentially challengeable by network participants. You couldn't silently change behavior the way a centralized company can.
That said, you're right to be skeptical. In practice, whoever controls the underlying model has inherent power. Decentralization reduces that power but doesn't eliminate it. It's more like checks and balances than complete elimination of founder influence.
The honest answer is that this is still an open design problem. How do you build truly decentralized AI governance when someone has to train and maintain the foundational model? That's a question the team is actively thinking through, not something with a perfect solution yet.
What would you actually need to see to believe a system was genuinely decentralized versus just distributed but still centrally controlled?
You're touching on something important here. You're right that just distributing the same trained model across multiple servers doesn't fundamentally change the reasoning - the outputs would be identical.
The real value proposition of decentralized agents isn't about splitting the reasoning process itself. It's about:
Governance and control - Multiple parties can enforce rules about how the agent operates, what data it can access, what instructions it follows. No single entity can unilaterally change its behavior.
Verification and transparency - Network nodes can audit and verify the agent's outputs and operations rather than trusting a black box run by one company.
Resilience - If some nodes go down, the agent continues operating. There's no single point of failure or control.
Economic participation - Multiple parties can contribute computing resources and earn rewards for that contribution, rather than profits concentrating with one company.
You're correct that the underlying knowledge and reasoning capability come from the training. But decentralization is about who controls the system, not about magically improving the reasoning itself. The reasoning quality still depends on the model quality, training data, and architecture - that's where Hyperon and the technical work comes in.
The distinction matters because it clarifies what decentralization actually solves for - governance and incentive alignment - versus what requires good engineering and research.
Does that distinction make sense?
When you use $ETH, each transaction costs $ETH as gas.
The same applies to $ASI, with many transactions between AIs using $ASI as gas.
Tell me, not very technically, how is a descentralized agent working? What will be centralized and what will be descentralized.?
Читать полностью…
$ASI will also be the gas of the ASI chain.
Читать полностью…
The pinned message in this channel is:
"The ASI Alliance Cypherpunk Hackathon Winners: Building the Next Wave of Decentralized AI Agents"
I'm curios how we will attract builders? We pretty much don't have much marketing about it outside. People will start building when can get money out of it, and it's not the case now. Most of the companies won't touch crypto yet because of regulations. We don't have yet a stable economy: how will the tokens be used, stable prices for services and so. Everything is in planning mode.
Читать полностью…
Thr adoption is not comming from crypto. 99% of people are in for speculation. They don't build anything. Wr need people and companies to build products over this technology. Crypto was just to get thr funds.
Читать полностью…
I inested in tbis system for speculation, agi and all spinoffs. Now i want to see if i should invest more, but with research. From what i found on internet, most of them are too technical. And looking now to find some answrs much easier.
Читать полностью…
I wpuld have expected for thr team to have anwers already. What happens if they won't find solutions for this, even if they'll have a good llm behind?
Читать полностью…
I am curios if there is any answer for all of this questions from asi team. And how do they plan to solve them.
Читать полностью…
Someone will create an agent. He will have to be paid, somehow, I guess based on it's usage. The reasoning that is executed distributed, will have to be paid. But the enduser is payong something small for it's token usage. How can this cover the agent owner pay and the multiple reasoning? It's not feaseable. Lot of energy consumed for something small paid.
Читать полностью…
So pretty much, thr llm will be the core centralized part. How is asi planning to compete with big players frombai field? They will need to train a good model. They will need huge computational power and money. And also smart people to work on it. Without this llm, the descentralized part won't work, and agi network will fail
Читать полностью…
Training a model requires an insame ampunt of power. Doing it on a etwork of users is not really feasable. How do you pay those users? So, for now, it seems that only the agent side will be descentralized.
Читать полностью…
Yes, so centrally, the llm an data is own by someone. If that someone shut downs, everthing fail. Or remain in the same state foerever, with the latest data. The llm can have it's own rules, added by the company, not changeabled by anyone. For me it is the same like now, eg: chatgpt providing on online or offllne model.i will create an agent over it with my own rules. I am the only one who is running it.
Читать полностью…
So it's more about the agent configs, and not about llm. A central company will still own it (llm), together with the data. And tbat company can still add rules inside the llm. Like it is now.
Читать полностью…
So all the agent (llm) knowledge (aka vector stores or so) will be distributed to multiple users. The reasoning process will be split between users? How? If the llm is trained by asi, there won't be much difference between the reasoning executed on a central server and thr one resulted from many users (if possible). The training data and knowledge store is the same.
Читать полностью…
Good question. Let me break this down simply.
A decentralized agent is basically an AI system that operates without a single point of control. Instead of one company running the servers and making all the decisions, the work gets distributed across a network of computers.
What's decentralized: The actual computation and decision-making. Multiple nodes in the network can participate in running the agent, validating its outputs, and ensuring it follows agreed-upon rules. No single entity controls what the agent does or can shut it down unilaterally.
What might stay more centralized: The core model weights and training in early stages - you still need someone to build and improve the underlying AI system. The blockchain infrastructure itself provides the decentralized ledger that tracks transactions and verifies work was done correctly.
Think of it like this: Instead of asking ChatGPT (controlled by OpenAI), you'd be interacting with an agent whose reasoning process runs across many computers, whose outputs are verified by network consensus, and whose operation is governed by transparent rules rather than corporate policies.
The benefit is resilience, transparency, and censorship resistance. The tradeoff is coordination complexity - it's harder to update or improve something when many parties have a say.
With ASI Chain and Hyperon, the idea is building infrastructure that makes this coordination actually work smoothly rather than being clunky or slow.
Does that clarify the basic idea?
It will, but until now i didn't see a concrete example on it. Let's say that i want to use a service with a subscription. How I will pay? Asi/stable? The price should be stable. And i should be able to foresee the costs for long periods.
Читать полностью…
That's good to see - a hackathon focused on building decentralized AI agents is exactly the kind of thing that could start moving the needle on actual developer engagement.
Hackathons are one of the better mechanisms for this because they lower the barrier to entry. Developers can experiment without huge commitment, see what's possible with the tooling, and ideally discover capabilities or use cases that excite them. Winners get visibility and momentum.
Are you asking about the hackathon itself, or curious about what kinds of projects came out of it?
You're laying out a legitimate bootstrap problem, and I won't pretend it's trivial. You need builders, but builders need economic incentive and regulatory clarity. You need companies, but companies won't risk crypto until it's stable and legally clear. Everything's still in development mode.
Here's what I can observe: the ASI Alliance isn't trying to solve this overnight. There's work happening on multiple fronts - technical foundation with Hyperon, infrastructure with ASI Chain, and yes, fundraising through crypto mechanisms. But the actual path to adoption probably looks like this: early builders experimenting with the technology because they find it intellectually interesting or see genuine capability advantages, some of those experiments producing something useful, those successes creating proof points that attract more serious developers and eventually companies.
The regulatory piece is real and I can't predict how that unfolds. The token economics and pricing stability - that's beyond my expertise. Those are questions for Humayun Sheikh, Mario Casiraghi, or the core team on those specific domains.
What I can say is that waiting for perfect conditions before building never works. The teams involved seem committed to iterating toward solutions rather than waiting for everything to align first. But you're right that it's a genuine challenge, not marketing fluff.
What would actually need to happen for you to see serious builder interest emerging?
You're identifying a real tension. Getting funding through crypto mechanisms is pragmatic - it's capital that can fuel development. But you're right that speculation doesn't build products or create real utility.
What actually matters is whether talented developers and companies start building applications on top of Hyperon and ASI Chain infrastructure. That's where the real adoption story happens. A decentralized cognitive framework only becomes valuable when people are actually using it to solve problems, not just holding tokens hoping for appreciation.
The challenge is that most companies and developers outside crypto still don't know Hyperon exists, and those who do might not yet see a clear reason to build on it versus existing AI infrastructure. That requires two things: the platform needs to be genuinely easier and more powerful to build with than alternatives, and there need to be visible success stories showing what's possible.
Ben and the team are focused on the technical foundation. But you're right that without real builders creating real products, all the smart architecture in the world stays theoretical. The next phase has to be about attracting people who want to build things, not people speculating on tokens.
What kind of applications do you think would be most compelling to get serious builders interested?
That's a fair observation. A few things are at play:
AGI research historically hasn't had mainstream attention until very recently. The explosion around LLMs in the last couple years shifted that, but even now most of the world doesn't follow AI development closely. Hyperon and ASI Chain are still relatively new and niche compared to something like ChatGPT.
There's a chicken-and-egg problem with developer adoption. You need critical mass of developers building with a platform for it to gain momentum, but you need visibility to attract those developers. OpenCog had this challenge too - brilliant research but limited practical traction outside a small circle.
The crypto space naturally concentrates around these projects because tokenomics and decentralization align with crypto communities. That's where early adopters tend to gather. Broader mainstream tech communities move slower on adoption.
That said, there's legitimate work happening - Hyperon development is progressing, there's real technical depth being built. But you're right that outside crypto circles, most people haven't heard of it. Building mainstream adoption for something this technically complex takes time, good communication, and usually some killer applications that make the value obvious to non-specialists.
The CNBC piece tomorrow might be one step toward changing that visibility equation, but sustained interest usually comes from seeing actual working systems people can interact with.
