The future of AI should be a swarm

I don’t want to live in a future where I need to rent intelligence every time I want to think, create, build, or solve a problem with technology. I don’t want every meaningful interaction with my computer to require permission from a distant server. I don’t want my work, source code, documents, or personal knowledge to become useless the moment a subscription expires, an API changes, or an internet connection disappears.

And I certainly don’t want a future where the only way to access increasingly capable artificial intelligence is to continuously send my information to enormous, privately controlled data centers and pay for every token that comes back. Yet this is the direction much of the AI industry is currently moving. We are building astonishingly capable systems, but we are also concentrating the infrastructure required to operate them. I believe we should be exploring a different future.

A future where artificial intelligence lives primarily on our own devices, but can connect to decentralized swarms of computing power and specialized intelligence whenever it needs additional capabilities.

Not a single supercomputer owned by one company. Not millions of disconnected personal assistants limited by their individual hardware. Something between the two. A distributed, collaborative intelligence network that people can participate in, contribute to, and benefit from. And perhaps more importantly, one that doesn’t require us to surrender ownership of our digital lives.

Your personal AI should belong to you

Your AI remains yours. The network simply extends what it can accomplish.

Imagine having a capable AI running directly on your computer. It knows how to work with your files, software, projects, and personal preferences. It helps you write, research, organize information, develop applications, and automate everyday tasks. It works even when your internet connection is unavailable.

Your conversations, documents, and long-term memory can remain on your device instead of being stored exclusively on somebody else’s infrastructure. For everyday tasks, its local model is perfectly capable. But what happens when you ask it to solve something considerably more complicated? Maybe you’re developing a complex application, analyzing a large scientific dataset, running simulations, or tackling a problem that benefits from several specialized models working together.

Your local AI recognizes that the task exceeds its current capabilities. Instead of refusing, struggling through the problem, or requiring you to manually select another subscription-based model, it connects to a larger network of available intelligence. It identifies the necessary computing resources, delegates appropriate parts of the task, receives the results, and continues working locally. When the work is finished, the external resources are released.

Your AI remains yours. The network simply extends what it can accomplish. That is the fundamental distinction I want to see in the future of computing.

What if AI worked more like BitTorrent?

We have already demonstrated that enormous networks of independently operated computers can collaborate without relying on one central infrastructure provider. BitTorrent is one of the most recognizable examples. Instead of downloading everything from a single server, participants connect directly to other participants, sharing resources and distributing the workload.

What if we took some of those principles and applied them to artificial intelligence? Imagine a network containing millions of devices. Some participants contribute powerful GPUs. Others contribute ordinary desktop computers, laptops, or smaller devices such as smartphones. Data centers could participate too, contributing the capabilities that consumer hardware cannot efficiently provide.

Not every device needs to perform the same work. A smartphone might run a small local model or contribute lightweight tasks when charging and explicitly permitted by its owner. A laptop could process smaller inference requests or execute specialized tools.

A gaming PC with a powerful GPU could host a considerably larger model. A dedicated server might run advanced reasoning models or process demanding workloads. And a data center could provide large-scale computation when those distributed resources are insufficient.

The network would coordinate these capabilities, matching tasks with suitable resources while accounting for performance, availability, energy consumption, and privacy. It would be a heterogeneous computing network rather than a collection of identical machines.

It would be a heterogeneous computing network rather than a collection of identical machines.

Of course, sharing computation is substantially more difficult than sharing files. Tasks must be scheduled, results verified, failures handled, and communication overhead managed. But the basic idea is not science fiction.

Research projects such as Petals (opens in a new tab) have demonstrated collaborative inference of large language models across geographically distributed devices. Exo (opens in a new tab) explores connecting everyday devices into shared AI clusters, including the ability to distribute model execution according to available resources. And Gensyn’s RL Swarm (opens in a new tab) explores collaborative model training over decentralized networks.

These systems solve different parts of the problem. None represents the complete vision I’m describing, but they show that meaningful parts of it are technically possible. What is missing is a broadly accessible architecture connecting personal AI, distributed computing, secure collaboration, and user ownership into one coherent experience.

A swarm of intelligence is more than a swarm of processors

There is another distinction worth making. Connecting thousands of computers does not automatically create a more intelligent AI. More processing power alone does not guarantee better reasoning, more reliable answers, or new capabilities. We need to consider how intelligence itself could be distributed.

One approach would be to divide the execution of a large model across multiple machines. Another would allow independent machines to host complete models of different sizes and capabilities. A third possibility interests me even more: collaboration between specialized AI systems.

Imagine your personal AI coordinating a collection of specialized models. One handles mathematical reasoning. Another analyzes code. Another researches technical documentation. Another independently verifies a proposed solution.

These systems could work on different parts of the problem, compare results, challenge assumptions, and return useful conclusions to your personal assistant. Some tasks could run in parallel. Others would require sequential collaboration.

This doesn’t mean ten small models necessarily outperform one powerful model. That would depend on the task, the models, and the quality of their coordination. But it opens the possibility of building intelligence networks around cooperation rather than simply making every model larger. Perhaps the future isn’t one enormous artificial brain. Perhaps it’s millions of capable systems that know when and how to cooperate.

Privacy must be part of the architecture, not an optional feature

There is an obvious problem with sending computation to computers operated by strangers. What happens to your information? If your personal AI delegates a task involving private documents, source code, medical information, or business data, you need confidence that the participating machines cannot simply collect that information.

Encrypting network traffic is necessary, but it is not enough. A remote computer normally has to decrypt the relevant input to process it, which means the operator of that computer may potentially access it. That makes privacy-preserving computation one of the most important challenges in this vision.

Fortunately, research is advancing in several relevant areas, including confidential computing, trusted execution environments, secure multiparty computation, and fully homomorphic encryption. These approaches offer different forms of protection, although they also involve technical limitations and performance costs.

I would envision a system with different privacy levels. Sensitive operations could remain entirely local. Other tasks could be delegated only to verified confidential-computing environments. Public information and nonsensitive workloads could use ordinary computing nodes across the network.

Privacy should not depend on trusting every stranger who contributes a GPU.

Users should control what can leave their devices, which resources are allowed to process it, and what information is retained. And until genuinely secure computation is available for a particular task, the system must be honest about that limitation. Privacy should not depend on trusting every stranger who contributes a GPU. It must be designed into the system from the beginning.

We still need data centers, but they shouldn’t own the future

I am not suggesting that data centers should disappear. Modern AI models require enormous computing resources, specialized hardware, high-bandwidth interconnections, and substantial energy infrastructure. Training the largest models across unreliable home internet connections remains a major technical challenge.

Data centers are extremely efficient at certain kinds of tightly coordinated computation, and decentralized systems will not automatically outperform them. We also shouldn’t assume that quantum computing is about to eliminate these requirements. Quantum computers have not demonstrated a general practical advantage for today’s production language-model workloads. For the foreseeable future, conventional CPUs, GPUs, and specialized accelerators will remain essential.

But that doesn’t mean everything must be centralized. We can use data centers where they make technical sense while allowing independently operated computers, local networks, universities, research institutions, and communities to participate in the same broader computing ecosystem. The goal isn’t to eliminate powerful infrastructure. The goal is to prevent powerful infrastructure from becoming the only gateway to intelligence.

We’ve seen what happens when technology becomes concentrated

This is where my concerns extend beyond artificial intelligence. Throughout history, control over valuable resources, knowledge, infrastructure, and technology has provided enormous advantages to those who possess them. Sometimes that concentration has enabled remarkable progress. It has also repeatedly created opportunities for exploitation, exclusion, and dependency.

I don’t believe artificial intelligence will somehow be immune to these dynamics. In fact, I worry it could amplify them. We are training AI systems on vast amounts of information produced by human civilization. That information contains our scientific discoveries, creativity, cooperation, and extraordinary achievements. It also contains propaganda, manipulation, discrimination, exploitation, conflict, and countless examples of people using superior knowledge or power to gain control over others.

AI models learn statistical patterns from this information. That doesn’t mean they are destined to repeat every harmful human behavior, or that they possess human motivations. But we should not assume that training on human history automatically produces systems free from human biases, power structures, and failures.

The objectives we give these systems, the incentives surrounding them, and the organizations controlling their deployment matter enormously. And those incentives become especially important when increasingly capable intelligence is concentrated inside a relatively small number of institutions.

Imagine a future where a few providers control most advanced AI capabilities. They could determine the price of access, the applications permitted, the data required, the models available, and the conditions under which people can use them. Even without malicious intent, ordinary business decisions, technical failures, policy changes, or service restrictions could affect millions of people who depend on those systems.

Today, losing access to a cloud service is inconvenient. In a world where intelligence is integrated into education, employment, research, software development, communication, and everyday decision-making, losing access could be considerably more consequential. We are approaching a point where access to computing intelligence may become as important as access to information itself. I don’t think we should build that future around complete dependence on a handful of providers.

We already surrendered too much control over our software

This concern also influences how I think about software development. Over the years, we have gradually moved more of our applications, workflows, and information into cloud platforms. There are excellent reasons for that. Cloud infrastructure makes collaboration, deployment, synchronization, and access across devices considerably easier. But somewhere along the way, we also became comfortable with the idea that software and information should exist primarily inside platforms controlled by someone else.

Our documents are in one service. Our project information is in another. Our development environment depends on another. Our data is accessible through APIs whose capabilities and conditions can change. And now we are introducing AI agents that must repeatedly connect to these systems, request information, interpret responses, and reconstruct the context necessary to perform useful work. Sometimes that’s unavoidable. But sometimes it feels like we’ve created enormous complexity simply to access information that could have remained directly available to us.

This is one of the ideas influencing my work on Skylit.DEV (opens in a new tab). Although Skylit.DEV is a WordPress development framework rather than a decentralized AI network, its architecture follows a related principle. The goal is to make a structured project folder the accessible foundation of a website, containing the critical templates, layouts, styling, code, configuration, and other project structures.

Those files remain synchronized with WordPress and its database, allowing developers and AI tools to work directly with readable project structures rather than relying exclusively on layers of remote interfaces. Working with AI during development has reinforced this approach for me. When a model can access a well-organized project folder, understand the relevant files, inspect their relationships, and make focused changes, the workflow can be considerably more effective.

It doesn’t always need to retrieve enormous JSON representations of information through tool calls simply to understand how a project is organized. Protocols such as MCP are valuable for integrating external systems, but they should complement direct access to understandable data, not automatically replace it. The point isn’t that every application must operate without a database or a cloud service. It’s that the essential structure of our work should remain accessible, portable, inspectable, and independent of any single provider.

If I change my IDE, AI model, hosting company, or development workflow, I should still possess my project. The same principle should apply to personal AI. Your information, memory, projects, and tools should not become inseparable from whichever AI company currently provides your subscription. The intelligence can change. The infrastructure can change. Your work should remain yours.

Decentralization is not automatically freedom

I also don’t believe that simply attaching a blockchain to a system makes it decentralized, secure, or beneficial. A network can claim to be decentralized while depending on a handful of validators, infrastructure providers, developers, or hardware owners. A distributed AI platform could even reproduce the same concentration of power it was supposed to challenge.

There are difficult questions around governance, malicious participants, computation verification, incentives, network security, and the unequal distribution of hardware. A volunteer network must also respect participants’ resources. Contributing computational power costs electricity, produces heat, and can affect battery life and hardware availability. People should be able to choose exactly what they contribute and when.

A well-designed system could recognize useful contributions and allow participants to exchange computing capacity without requiring every interaction to pass through a centralized commercial API. Blockchain might help with identity, accounting, verification records, or decentralized settlement. But it should be used where it provides a genuine technical benefit, not simply because decentralization has become associated with cryptocurrency. The real objective is not creating another token economy. It is creating infrastructure that distributes capability and control.

The internet connected people. Could the next step connect intelligence?

I keep returning to the early promise of the internet. A global network where independent computers could communicate, exchange information, and collaborate across geographical boundaries. That vision was not without problems, and central services have brought genuine advantages. But much of our digital activity now flows through platforms operated by relatively few organizations.

AI presents an opportunity to reconsider that direction. We could build increasingly powerful intelligence almost entirely behind centralized APIs. Or we could invest more effort in open models, local inference, peer-to-peer networking, distributed computing, privacy-preserving computation, and collaborative AI systems. These approaches are not mutually exclusive.

A decentralized intelligence network could include everything from smartphones to research laboratories and massive data centers. Some computation would happen locally. Some would be delegated to trusted infrastructure. Some could be distributed among independent peers. Different models could collaborate when doing so improves results. And people could retain control over their information and digital environments rather than being required to surrender that control as the price of participation.

Some computation would happen locally. Some would be delegated to trusted infrastructure. Some could be distributed among independent peers.

Will such a system be easy to build? Absolutely not. Will every device meaningfully contribute to every type of workload? No. Will decentralized AI immediately replace highly optimized data centers or frontier AI services? I doubt it.

But none of those limitations is a convincing reason to abandon the idea. Distributed systems were difficult long before cloud computing became commonplace. Privacy-preserving technologies are difficult. Open protocols are difficult. Making complex technologies accessible to ordinary people is difficult. That doesn’t make them unworthy of development. It makes them important engineering problems.

I want intelligence to become more accessible, not more controlled

I support AI development. I use AI extensively, and I see enormous potential in what these systems can help us achieve. My concern isn’t that artificial intelligence will become too capable. It’s that we may gradually surrender too much independence in exchange for those capabilities.

I don’t want to carry a screen connected to a wireless network and spend the rest of my life renting access to intelligence, paying for every interaction, and depending on a distant infrastructure provider to keep my digital life functioning. I want my devices to become more capable. I want software that I can understand, modify, move, and preserve.

I want personal AI that operates locally whenever possible. And when that AI needs capabilities beyond my own hardware, I want it to be able to connect to a wider network of intelligence without automatically surrendering my privacy or control.

We should be pursuing systems where communities, independent developers, universities, researchers, ordinary people, and data centers can contribute to a shared infrastructure. Not because decentralization magically solves every technical or social problem. But because distributing control gives us more ways to resist dependency, build alternatives, and preserve individual choice.

We have spent decades connecting the world’s computers. Perhaps the next challenge is teaching their intelligence to cooperate without requiring everyone to surrender control to the same few institutions.

The future of AI does not have to be one enormous fortress that everyone pays to enter. It could be a network of independent, interconnected systems, capable on their own and more capable together. A swarm. And I believe that is a future worth exploring.