It started with a noisy desk. The desk was a wooden cabinet in a laboratory at Northumbria University, in northern England, where a young AI researcher began his PhD. -track. This was in 2015. The researcher was Ben Fielding, who had built a large machine filled with early GPUs to develop AI. The machine was so high that the irritated Fielding’s laboratory buddies. Fielding stuffed the machine under the desk, but it was so big that he had to awkward to stick his legs to the side.

Fielding had some unorthodox ideas. He explored how “swarms” of AI – clusters of many different models – could talk to each other and learn from each other, which could possibly improve the collective whole. There was only one problem: he was handcuffed by the realities of the noisy machine under his desk. And he knew he was outgunned. “Google also did this research,” Fielding says now. “And they had thousands [of GPUs] in a data center. The things they did were not crazy. I knew the methods … I had lots of suggestions but I couldn’t run them. “

Ben Fielding, CEO of reunion, is a speaker in Consensus 2025 in Toronto.

Jeff Wilser hosts People’s AI: The Decentralized Ai Podcast and will host the AI ​​summit in Consensus 2025.

So a decade ago it went up on field: Calculate restrictions would always be a problem. In 2015, he knew that if Compute was a tough restriction on the academy, it would definitely be a tough restriction when AI went mainstream.

The solution?

Decentralized ai.

Fielding co -founded reunion (along with Harry Grieve) in 2020 or years before decentralized AI became fashionable. The project was originally known for building decentralized calculation – and I have talked to Fielding about this for Coindesk and on panel after panel at conferences – but the vision is actually somewhat broader: “The network for machine information.” They build solutions up and down the tech stack.

And now, a decade after Fielding’s noisy desk irritated his lab friends, the early tools are in the nature. Reunit recently released its “RL Swarms” protocol (a descendant of Fielding’s PhD work) and just launched its test network -which brings blockchain into the fold.

In this conversation, leading up to the AI ​​summit, by consensus in Toronto, Fielding gives a primer on AI swarms, explains how blockchain clicks into the puzzle, and shares why all innovators -not just tech giants -“should have the right to build machine learning technologies.”

This interview has been condensed and easily edited for clarity.

Congratulations on the test network -launch. What is the essence of what it is?

Ben Fielding: That’s the addition of the first MVP features in blockchain integration with what we’ve launched so far.

What were the original features, pre-blockchain?

So we launched rl [Reinforcement Learning] Swarm a few weeks ago, which is reinforcement learning, post-training, like a peer-to-peer network.

Here’s the easiest way to think about it. When a pre-formed model reviews reasoning training-as Deepseek-R1 teaching it to criticize his own thinking and recursively improved against the task. It can then improve its own answer.

We take this process a step further and say, “It’s great for models to criticize their own thinking and improves recursively. What if they can talk to other models and criticize each other’s thinking?” If you get a lot of models together in a group that can all talk to each other, they can start learning how to send information to the other models … with the overall goal of improving the whole sword itself.

Gotcha explaining the name “Swarm.”

Right. It is this training method that allows many models to combine, in parallel, to improve the result of a last metam model that you could create from these models. But at the same time, you have every individual model that just improves on your own. So if you had to get along with a model on a MacBook, take part in a sword for an hour and then fall back out again, you would have an enhanced local model based on knowledge in swarm and you would also have improved the other models in the swarm. It is this collaborative education process that any model can participate in and any model can do. So that’s what rl swarm is.

Okay, so that’s what you released a few weeks ago. Where does blockchain come in now?

So blockchain are us who move forward some of the lower level primitives into the system.

Just let’s Pretend That someone does not understand the term “primitives at the lower level.” What do you mean by that?

Yes, so I mean, very close to the resource itself. So if you think of the software stack you have a GPU stack in a data center. You have drivers on top of GPU. You have operating systems, virtual machines. You have all these things to go up.

So a primitive lower level is the closest to the lower foundation of the tech stack. Do i get it right?

Yes, exactly. And RL Swarm is a demonstration of what is possible, basically. It’s just a somewhat hacky demo about making really interesting large -scale, scalable machine learning. But what reunion has done in the last four-Plus years, realistically, builds infrastructure. And then during this period we are now that the infrastructure is all at the V0.1 kind of beta level. It’s all done. It’s ready to go. We need to find out how to show the world what is possible when it is quite a big shift to the way people think of machine learning.

Sounds like you are doing much more than decentralized or even infrastructure?

We have three main components that sit under our infrastructure. Execution – We have consistent execution libraries. We have our own compiler. We have reproducible libraries for any hardware target.

The second piece is communication. So suppose you can just run a model on any device in the world that is compatible can you make them talk to each other? If everyone chooses the same standard, everyone can communicate as TCP/IP from the Internet, basically. So we build these libraries and RL Swarm is an example of this communication.

And then eventually verification.

Ah, and I guess this is where blockchain comes in …

Imagine a scenario where any unity in the world performs consistently. They could link models together. But can they trust each other? If I connected my MacBook to yours, well, they could perform the same tasks. Yes, they could send tensors back and forth, but do they know that what they send to the other device is actually happening on the other device or not?

In the current world, you and I would probably sign a contract to say, yes, we agree that we make sure our devices are doing the right thing. In the world of machine it must be programmatically. So this is the last piece we build, cryptographic evidence, probable evidence, game theory evidence to make this process completely programmatic.

So this is where blockchain comes in. It gives us all the benefits of blockchain, you can imagine, such as sustained identity, payments, consensus, etc., and then what we do with the test network now takes RL -SVERM, and the primitives of the other infrastructure and we add to the blockchain components and say, ‘Hi, when you end up for a swarm now, you have a persistent identity, as we are there. Ledger. ‘

In the future you have the ability to make payments, but right now you have the confidence consensus mechanism where we can terminate disputes. So it’s a kind of MVP for the future reunion -infrastructure where we will add components while we go.

Give us a drill of what comes down the pipeline?

When we reach the main network, all software and infrastructure are live against blockchain as a source of trust, payments, consensus, etc., identity. This is the first step in it. It adds identity in and says that when you participate in a sword, you can register as the same person. Everyone knows who you are without having to check any centralized server or site somewhere.

Now let’s get lost and talk longer in the future. What does this look like a year from now, two years from now, five years from now? What is your North Star?

Sure. The ultimate vision is to take all the resources that are under machine learning and make them immediately programmatically accessible to everyone. Machine learning is severely limited by its core resources. This creates this huge Vollgrav for centralized AI companies, but it doesn’t have to exist. It can be open sourced if we can build the right software. So our point of view is that reunion builds all the low -level infrastructure, so it can get as close to open source as it might. People must have the right to build machine learning technologies.