It all kicked off with a loud desk. This desk was a wooden cubicle situated in a lab at Northumbria University in northern England, where a young AI researcher commenced his PhD journey back in 2015. The researcher, Ben Fielding, had assembled a robust machine filled with early GPU technology to advance AI developments. However, this machine’s noise was so disruptive that it irritated his lab colleagues. To accommodate it, Fielding squeezed the machine under the desk, which forced him to awkwardly position his legs to the side.
Fielding had some unconventional ideas. He investigated how groups of AI, or "swarms," comprising diverse models could communicate and learn from each other to enhance their collective capabilities. The issue was that he was limited by the constraints of that noisy machine lurking beneath his desk. He recognized that he was at a disadvantage. “Google was involved in similar research,” Fielding recalls. “And they had thousands of GPUs in dedicated data centers. What they were accomplishing wasn’t outlandish. I was familiar with the methods… I had numerous proposals, but I couldn’t implement them.”
Fast forward a decade from that noisy desk, and we’ve arrived at the emergence of Gensyn’s early tools. The company, which Fielding co-founded with Harry Grieve in 2020 — well before Decentralized AI became trendy — was initially recognized for its work on decentralized computing. However, the broader vision revolves around creating “the network for machine intelligence” with solutions spanning the entire technology stack.
Recently, Gensyn unveiled its “RL Swarms” protocol, a product of Fielding’s PhD research, and launched its Testnet, marking an integration of blockchain technology. In this discussion leading up to the AI Summit in Toronto, Fielding provides insights into AI Swarms, details on how blockchain fits into the equation, and advocates for the right of all innovators—not just tech giants—to construct machine learning technologies.
Kudos on the Testnet rollout. Can you summarize what it entails?
Ben Fielding: It’s the implementation of the initial MVP features of blockchain integration with what we’ve previously launched.
What were those original functions before blockchain?
Not long ago, we introduced the RL (Reinforcement Learning) Swarm, which applies post-training reinforcement learning as a peer-to-peer network.
What’s an easy way to conceptualize that?
Think of it like this: when a pre-trained model undergoes reasoning training—like DeepSeek-R1—it learns to evaluate its own thought processes and improve its performance on specific tasks. Subsequently, it can refine its responses.
We extend this idea further by asking: “What if these models could critique each other’s reasoning?” When multiple models are grouped together, communicating with one another, they can begin sharing information in a manner that ultimately enhances the entire swarm.
Ah, that makes sense and explains the term “Swarm.”
Exactly. This training approach allows models to work collectively, improving the results of a final meta-model derived from them. Meanwhile, individual models keep evolving on their own. So, if a model were to join a swarm using a MacBook for an hour and then leave, it would emerge with an improved local model based on the swarm’s collective knowledge, also contributing improvements to others in the group. This collaborative training process is open to any model that wishes to participate.
Got it. So that was what you released recently. Now how does blockchain fit in?
Blockchain represents our progress in integrating some fundamental elements into the system.
For someone unfamiliar with the term “lower-level primitives,” what does that mean?
It refers to elements that are closest to the resource itself. Picture the software stack: at the base, you have GPUs located in a data center. Then layered above that are drivers, operating systems, virtual machines, and so forth.
In other words, lower-level primitives are the foundational components in the tech stack, correct?
Precisely. The RL Swarm exemplifies what’s achievable—it’s a somewhat experimental demonstration of robust, scalable machine learning. Over the past four years, Gensyn has primarily focused on constructing the necessary infrastructure. We’re currently at a point where this infrastructure has reached a beta level, and now we aim to demonstrate what’s possible, prompting a significant shift in how people perceive machine learning.
It sounds like your ambitions extend beyond just decentralized computing or infrastructure.
Indeed, we have three key elements underpinning our infrastructure: Execution—which includes consistent execution libraries, our own compiler, and reproducible libraries for any hardware target; Communication—where we aim for models to run on compatible devices and seamlessly communicate, similar to how everyone agrees on the TCP/IP standard; and lastly, Verification.
And this is where blockchain comes in…
Imagine a scenario in which every device executes tasks consistently and can connect models together. The question arises: can they trust one another? If I connect my MacBook to yours, while they can execute identical tasks and exchange data, there’s no guarantee that the processes happening on your device align with what was agreed upon.
In our current world, we might sign a contract to ensure compliance. In the realm of machines, this verification must occur programmatically. Therefore, we create cryptographic proofs, probabilistic proofs, and game-theoretic proofs to automate this process.
This is where blockchain is effective, offering myriad benefits like persistent identity, payment solutions, and consensus mechanisms. With our Testnet rollout, we are integrating these blockchain elements within the RL Swarm and other foundational components, establishing a persistent identity stored on a decentralized ledger when joining a swarm.
In the future, we’ll incorporate payment functionalities, but for now, we provide a trust consensus mechanism to resolve disputes—a preliminary version of what Gensyn aims to achieve with its infrastructure.
What can we anticipate moving forward?
As we reach the main-net stage, our software and infrastructure will operate with blockchain as the foundation for trust, financial transactions, consensus, and identity. This initial step involves integrating identity when joining a swarm, ensuring that users are recognized without any need for verification from centralized servers or websites.
Looking ahead—what does the future hold in one year, two years, or even five years? What’s your ultimate goal?
The overarching vision is to make all necessary resources for machine learning instantly and programmatically accessible to everyone. Currently, machine learning faces significant constraints based on core resources, creating barriers for centralized AI firms. However, this need not be the case. It can be democratized through the right software solutions. Our mission at Gensyn is to construct low-level infrastructure to facilitate this transition to what’s as close to open-source as possible, ensuring everyone has the opportunity to innovate with machine learning technologies.