It began with a loud desk. The desk was a wood cubicle in a lab at Northumbria College, in northern England, the place a younger AI researcher started his PhD observe. This was in 2015. The researcher was Ben Fielding, who had constructed a big machine filled with early GPUs to develop AI. The machine was so loud it aggravated Fielding’s lab-mates. Fielding crammed the machine beneath the desk, nevertheless it was so large he needed to awkwardly stick his legs to the facet.
Fielding had some unorthodox concepts. He explored how “swarms” of AI — clusters of many alternative fashions — might speak to one another and be taught from one another, which could enhance the collective complete. There was only one downside: He was handcuffed by the realities of that noisy machine beneath his desk. And he knew he was outgunned. “Google was doing this analysis as effectively,” Fielding says now. “And so they had 1000’s [of GPUs] in a knowledge heart. The issues they had been doing weren’t loopy. I knew the strategies… I had a lot of proposals, however I couldn’t run them.”
Ben Fielding, CEO of Gensyn, is a speaker at Consensus 2025 in Toronto.
Jeff Wilser is the host of The Folks’s AI: The Decentralized AI Podcast and can host The AI Summit at Consensus 2025.
So a decade in the past, it dawned on Fielding: Compute constraints would all the time be a difficulty. In 2015, he knew that if compute was a tough constraint in academia, it could completely be a tough constraint when AI went mainstream.
The answer?
Decentralized AI.
Fielding co-founded Gensyn (together with Harry Grieve) in 2020, or years earlier than Decentralized AI grew to become modern. The venture was initially recognized for constructing decentralized compute – and I’ve spoken with Fielding about this for CoinDesk and on panel after panel at conferences – however the imaginative and prescient is definitely one thing wider: “The community for machine intelligence.” They’re constructing options up and down the tech stack.
And now, a decade after Fielding’s noisy desk aggravated his lab-mates, the early instruments of Gensyn are out within the wild. Gensyn not too long ago launched its “RL Swarms” protocol (a descendant of Fielding’s PhD work) and simply launched its Testnet — which brings blockchain into the fold.
On this dialog main as much as the AI Summit, at Consensus in Toronto, Fielding provides a primer on AI Swarms, explains how blockchain snaps into the puzzle, and shares why all innovators — not simply tech giants — “ought to have the best to construct machine studying applied sciences.”
This interview has been condensed and calmly edited for readability.
Congrats on the testnet launch. What’s the gist of what it’s?
Ben Fielding: It’s the addition of the primary MVP options of blockchain integration with what we have launched to this point.
What had been these authentic options, pre-blockchain?
So we launched RL [Reinforcement Learning] Swarm a couple of weeks in the past, which is reinforcement studying, post-training, as a peer-to-peer community.
Right here’s the best manner to consider it. When a pre-trained mannequin goes by means of reasoning coaching – like DeepSeek-R1 – it learns to critique its personal pondering and recursively enhance in opposition to the duty. It will probably then enhance its personal reply.
We take that course of one step additional and say, “It’s nice for fashions to critique their very own pondering and recursively enhance. What if they will speak to different fashions and critique one another’s pondering?” For those who get many fashions collectively in a gaggle that may all speak to one another, they will begin studying tips on how to ship data to the opposite fashions… with the general purpose of bettering the whole swarm itself.
Gotcha, which explains the title “Swarm.”
Proper. It’s this coaching technique which permits many fashions to type of mix, in parallel, to enhance the result of a ultimate meta-model that you may create from these fashions. However on the similar time, you’ve got each single particular person mannequin simply bettering by itself. So should you had been to return together with a mannequin on a MacBook, be part of a swarm for an hour after which drop again out once more, you’d have an improved native mannequin based mostly on the data within the swarm, and you’d have additionally improved the opposite fashions within the swarm. It’s this collaborative coaching course of that any mannequin can be part of and any mannequin can do. So that is what RL Swarm is.
Okay, in order that’s what you launched a couple of weeks in the past. Now the place does blockchain are available?
So the blockchain is us transferring ahead a number of the lower-level primitives into the system.
Let’s simply faux that somebody doesn’t perceive the phrase “lower-level primitives.” What do you imply by that?
Yeah, so I imply, very near the useful resource itself. So if you consider the software program stack, you’ve got acquired a GPU stack in a knowledge heart. You’ve got acquired drivers on high of the GPU. You’ve got acquired working methods, digital machines. You’ve got acquired all these items going up.
So a lower-level primitive is the closest to the underside basis within the tech stack. Am I getting that proper?
Sure, precisely. And the RL Swarm is an illustration of what is doable, mainly. It is only a considerably hacky demo of doing actually fascinating large-scale, scalable machine studying. However what Gensyn’s been doing for the previous four-plus years, realistically, is constructing infrastructure. And so we’re on this interval now the place the infrastructure is all at that v0.1 form of beta stage. It is all achieved. It is able to go. We’ve got to determine tips on how to present the world what’s doable when it is fairly a giant shift to the way in which folks consider machine studying.
It sounds such as you guys are doing much more than decentralized compute, and even infrastructure?
We’ve got three important elements that sit beneath our infrastructure. Execution – we’ve got constant execution libraries. We’ve got our personal compiler. We’ve got reproducible libraries for any {hardware} goal.
The second piece is communication. So assume you’ll be able to simply run a mannequin on any system on this planet that is appropriate, are you able to get them to speak to one another? If everyone opts into the identical commonplace, everyone can talk like TCP/IP from the web, mainly. So we construct these libraries and RL Swarm is an instance of that communication.
After which, lastly, verification.
Ah, and I’m guessing that is the place blockchain is available in…
Think about a situation the place each system on this planet is executing persistently. They might hyperlink fashions collectively. However can they belief one another? If I related my MacBook to yours, sure, they might execute the identical duties. Sure, they might ship tensors forwards and backwards, however do they know that what they ship to the opposite system is definitely taking place on the opposite system or not?
Within the present world, you and I might in all probability signal a contract to say, sure, we agree that we’ll make sure that our units do the best factor. Within the machine world, it must occur programmatically. In order that’s the ultimate piece we construct, cryptographic proofs, probabilistic proofs, sport theoretic proofs to make that course of completely programmatic.
In order that’s the place the blockchain is available in. It provides us all the advantages of blockchain you’ll be able to think about, like persistent identification, funds, consensus, and so forth. And so what we’re doing with the testnet now’s taking RL Swarm and the primitives of the opposite infrastructure and we’re including within the blockchain elements and saying, ‘Hey, while you be part of a swarm now, you’ve got a persistent identification, which exists on the market on a decentralized ledger.’
Sooner or later you’ll have the flexibility to make funds, however proper now, you’ve got that belief consensus mechanism the place we will terminate disputes. So, it is type of an MVP of the long run Gensyn infrastructure, the place we’re going so as to add in elements as we go.
Give us a tease of what’s coming down the pipeline?
Once we attain main-net, all the software program and infrastructure is stay in opposition to blockchain because the supply of belief, funds, consensus, and so forth., identification. This is step one of that. It is including identification in and saying while you be part of a swarm, you’ll be able to register as the identical particular person. Everybody is aware of who you might be with out having to test some centralized server or web site someplace.
Now let’s get wild and speak additional sooner or later. What does this seem like one 12 months from now, two years from now, 5 years from now? What’s your North Star?
Certain. The last word imaginative and prescient is to take all the assets that sit below machine studying and make them instantaneously programmatically accessible to everybody. Machine studying is closely constrained by its core assets. This creates this big moat for centralized AI firms, nevertheless it does not must exist. It may be open-sourced if we will construct the best software program. So our view is Gensyn builds all the low-level infrastructure to permit that to get as near open-source because it presumably can. Folks ought to have the best to construct machine studying applied sciences.
