The Noisy Desk and the Birth of Decentralized AI
It all began with a clamor emanating from a desk at Northumbria University, located in the northern part of England. This desk belonged to a young AI researcher who was embarking on his PhD journey back in 2015. The researcher was none other than Ben Fielding, who had ingeniously assembled a massive machine filled with early-generation GPUs to facilitate his AI development work. Unfortunately, this machine was so loud that it became a source of irritation for Fielding’s lab mates. He had to awkwardly position the bulky machine beneath his desk, leaving little room for his legs.
Fielding was known for his unconventional ideas. He was intrigued by the concept of “swarms” of AI—groups of various models that could communicate with one another and learn collectively, thereby enhancing their overall effectiveness. However, there was a significant obstacle: he was constrained by the limitations of that noisy machine hidden away under his desk. Fielding realized he was at a disadvantage. “Google was engaged in similar research,” he recalls. “They had thousands of GPUs in a data center. The methods they were employing were not far-fetched. I was aware of the techniques… I had numerous proposals, but I simply couldn’t run them.”
Ben Fielding, CEO of Gensyn, is scheduled to speak at Consensus 2025 in Toronto.
Jeff Wilser, host of The People’s AI: The Decentralized AI Podcast, will also lead The AI Summit at Consensus 2025.
Fast forward a decade, Fielding came to a crucial realization: Compute constraints would always pose a challenge. In 2015, he understood that if computing power was a significant constraint in academia, it would undoubtedly be an even larger hurdle as AI technology entered mainstream usage.
The Solution: Decentralized AI
In 2020, Fielding co-founded Gensyn alongside Harry Grieve, well ahead of the trend towards decentralized AI. Initially, their project focused on building decentralized computing solutions. However, as I have discussed with Fielding in previous conversations for CoinDesk and during various panels at conferences, their vision extends far beyond just that: “We aim to create the network for machine intelligence.” They are developing solutions that span the entire tech stack.
Now, a decade later, after Fielding’s noisy desk had disturbed his peers, Gensyn has launched its early tools. Recently, they introduced the “RL Swarms” protocol, a direct descendant of Fielding’s PhD research, and unveiled their Testnet that incorporates blockchain technology.
In this dialogue leading up to the AI Summit at Consensus in Toronto, Fielding provides an overview of AI Swarms, elucidates how blockchain fits into the broader picture, and emphasizes the belief that innovation should be accessible to all—not just the tech behemoths.
This interview has been edited for brevity and clarity.
Congratulations on the Testnet launch! What’s the main takeaway?
Ben Fielding: This marks the introduction of the first MVP features integrating blockchain with our existing offerings.
What were the original features before blockchain integration?
We recently launched the RL [Reinforcement Learning] Swarm, which allows for reinforcement learning post-training in a peer-to-peer network. Here’s a simplified breakdown: when a pre-trained model undergoes reasoning training—like DeepSeek-R1—it learns to evaluate its own thought processes and continuously refine its performance on a given task. It becomes capable of enhancing its own outputs.
We take this concept a step further by asking, “What if models could communicate with each other to evaluate one another’s reasoning?” When a group of models can interact and share insights, they collectively learn to disseminate information, ultimately aiming to enhance the entire swarm’s capabilities.
That makes sense, hence the name “Swarm.”
Exactly! This training approach enables numerous models to collaborate in parallel, improving the results of a final meta-model derived from those individual models. Simultaneously, each model continues to improve on its own. Imagine a scenario where a model on a MacBook joins a swarm for just an hour and then leaves: it would exit with an enhanced local model, enriched by the swarm’s collective knowledge, while also contributing to the other models’ advancements. This collaborative training process is open to any model that wishes to participate. That’s the essence of RL Swarm.
So, that was the recent release. How does blockchain fit into this framework?
With blockchain, we are advancing some of the foundational elements of the system.
For someone unfamiliar with the term “foundational elements,” could you clarify?
Certainly! I’m referring to the components that are closest to the underlying resources. In the software stack, you’ve got a GPU stack in a data center, drivers on top of the GPU, operating systems, virtual machines, and so on.
So, foundational elements represent the core of the tech stack. Is that correct?
Absolutely! The RL Swarm serves as a demonstration of the potential that exists. It’s essentially a prototype showcasing fascinating large-scale, scalable machine learning capabilities. However, for the past four-plus years, Gensyn has concentrated on building robust infrastructure. We’ve now reached a point where our infrastructure is in a beta phase, ready to be unveiled. Our challenge is to show the world the possibilities that arise from this significant shift in how we perceive machine learning.
It sounds like your efforts extend beyond decentralized computing and infrastructure?
Indeed, we have three core components underpinning our infrastructure. First is execution—we provide consistent execution libraries, our own compiler, and reproducible libraries for various hardware targets.
The second component is communication. If you can run a model on any compatible device globally, how do you facilitate communication between them? If everyone adheres to the same standards, it’s akin to the TCP/IP protocols of the internet. We develop those libraries, and RL Swarm exemplifies that communication.
Lastly, we have verification.
And this is where blockchain comes into play…
Picture a scenario where every device globally executes tasks consistently. They could interlink their models. But can they trust one another? If I connect my MacBook with yours, we can execute identical tasks and exchange tensors. However, how do we ensure that what we transmit is indeed being processed accurately on the other device?
In today’s world, we would likely draft a contract to agree on ensuring our devices perform correctly. In the realm of machines, this must happen programmatically. That’s the final aspect we are addressing: implementing cryptographic proofs, probabilistic proofs, and game-theoretic proofs to automate trust.
Blockchain offers all the advantages you might expect: persistent identity, seamless payments, consensus mechanisms, etc. Currently, with our Testnet, we are integrating RL Swarm with these foundational components, establishing a decentralized ledger where joining a swarm grants you a persistent identity.
In the future, we plan to introduce payment capabilities, but for now, this MVP provides a trust consensus mechanism that resolves disputes. This is just the initial phase of our future Gensyn infrastructure, where we will continue to add components as we progress.
Can you give us a glimpse of what’s on the horizon?
As we transition to mainnet, our software and infrastructure will be fully operational on blockchain, serving as the source of trust, payments, consensus, and identity. This is the first step, introducing identity so that when you join a swarm, you can register as the same individual, known to all without needing a centralized server or website for verification.
Let’s get visionary—what does the future hold one, two, or even five years from now? What is your ultimate goal?
Our grand vision is to make all resources essential for machine learning instantly and programmatically accessible to everyone. Currently, machine learning suffers from significant constraints tied to its fundamental resources, creating formidable barriers for centralized AI corporations. However, this need not be the case. With the right software, we can foster an open-source environment. At Gensyn, our mission is to construct all the foundational infrastructure needed to bring this vision closer to reality. Everyone should have the opportunity to develop machine learning technologies.
