Ben Fielding: Decentralizing Machine Intelligence

It started with a noisy desk. The table was a woody cubicle successful a laboratory astatine Northumbria University, successful bluish England, wherever a young AI researcher began his PhD track. This was successful 2015. The researcher was Ben Fielding, who had built a ample instrumentality stuffed with aboriginal GPUs to make AI. The instrumentality was truthful large it annoyed Fielding’s lab-mates. Fielding crammed the instrumentality beneath the desk, but it was truthful large helium had to awkwardly instrumentality his legs to the side.

Fielding had immoderate unorthodox ideas. He explored however “swarms” of AI — clusters of galore antithetic models — could speech to each different and larn from each other, which mightiness amended the corporate whole. There was conscionable 1 problem: He was handcuffed by the realities of that noisy instrumentality beneath his desk. And helium knew helium was outgunned. “Google was doing this probe arsenic well,” Fielding says now. “And they had thousands [of GPUs] successful a information center. The things they were doing weren’t crazy. I knew the methods... I had tons of proposals, but I couldn’t tally them.”

Ben Fielding, CEO of Gensyn, is simply a talker astatine Consensus 2025 successful Toronto.

Jeff Wilser is the big of The People’s AI: The Decentralized AI Podcast and volition big The AI Summit astatine Consensus 2025.

So a decennary ago, it dawned connected Fielding: Compute constraints would ever beryllium an issue. In 2015, helium knew that if compute was a hard constraint successful academia, it would perfectly beryllium a hard constraint erstwhile AI went mainstream.

The solution?

Decentralized AI.

Fielding co-founded Gensyn (along with Harry Grieve) successful 2020, oregon years earlier Decentralized AI became fashionable. The task was initially known for gathering decentralized compute – and I’ve spoken with Fielding astir this for CoinDesk and connected sheet aft sheet astatine conferences – but the imaginativeness is really thing wider: “The web for instrumentality intelligence.” They’re gathering solutions up and down the tech stack.

And now, a decennary aft Fielding’s noisy table annoyed his lab-mates, the aboriginal tools of Gensyn are retired successful the wild. Gensyn precocious released its “RL Swarms” protocol (a descendant of Fielding’s PhD work) and conscionable launched its Testnet — which brings blockchain into the fold.

In this speech starring up to the AI Summit, at Consensus successful Toronto, Fielding gives a primer connected AI Swarms, explains however blockchain snaps into the puzzle, and shares wherefore each innovators -- not conscionable tech giants — “should person the close to physique instrumentality learning technologies.”

This interrogation has been condensed and lightly edited for clarity.

Congrats connected the testnet launch. What’s the gist of what it is?

Ben Fielding: It’s the summation of the archetypal MVP features of blockchain integration with what we've launched truthful far.

What were those archetypal features, pre-blockchain?

So we launched RL [Reinforcement Learning] Swarm a fewer weeks ago, which is reinforcement learning, post-training, arsenic a peer-to-peer network.

Here’s the easiest mode to deliberation astir it. When a pre-trained exemplary goes done reasoning grooming – similar DeepSeek-R1 – it learns to critique its ain reasoning and recursively amended against the task. It tin past amended its ain answer.

We instrumentality that process 1 measurement further and say, “It’s large for models to critique their ain reasoning and recursively improve. What if they tin speech to different models and critique each other's thinking?” If you get galore models unneurotic successful a radical that tin each speech to each other, they tin commencement learning however to nonstop accusation to the different models… with the wide extremity of improving the full swarm itself.

Gotcha, which explains the sanction “Swarm.”

Right. It’s this grooming method which allows galore models to benignant of combine, successful parallel, to amended the result of a last meta-model that you could make from those models. But astatine the aforesaid time, you person each azygous idiosyncratic exemplary conscionable improving connected its own. So if you were to travel on with a exemplary connected a MacBook, articulation a swarm for an hr and past driblet backmost retired again, you would person an improved section exemplary based connected the cognition successful the swarm, and you would person besides improved the different models successful the swarm. It’s this collaborative grooming process that immoderate exemplary tin articulation and immoderate exemplary tin do. So that's what RL Swarm is.

Okay, truthful that’s what you released a fewer weeks ago. Now wherever does blockchain travel in?

So the blockchain is america moving guardant immoderate of the lower-level primitives into the system.

Let’s conscionable unreal that idiosyncratic doesn’t recognize the operation “lower-level primitives.” What bash you mean by that?

Yeah, truthful I mean, precise adjacent to the assets itself. So if you deliberation astir the bundle stack, you've got a GPU stack successful a information center. You've got drivers connected apical of the GPU. You've got operating systems, virtual machines. You've got each this worldly going up.

So a lower-level primitive is the closest to the bottommost instauration successful the tech stack. Am I getting that right?

Yes, exactly. And the RL Swarm is simply a objection of what's possible, basically. It's conscionable a somewhat hacky demo of doing truly absorbing large-scale, scalable instrumentality learning. But what Gensyn's been doing for the past four-plus years, realistically, is gathering infrastructure. And truthful we're successful this play present wherever the infrastructure is each astatine that v0.1 benignant of beta level. It's each done. It's acceptable to go. We person to fig retired however to amusement the satellite what's imaginable erstwhile it's rather a large displacement to the mode radical deliberation of instrumentality learning.

It sounds similar you guys are doing a batch much than decentralized compute, oregon adjacent infrastructure?

We person 3 main components that beryllium underneath our infrastructure. Execution – we person accordant execution libraries. We person our ain compiler. We person reproducible libraries for immoderate hardware target.

The 2nd portion is communication. So presume you tin conscionable tally a exemplary connected immoderate instrumentality successful the satellite that's compatible, tin you get them to speech to each other? If everybody opts into the aforesaid standard, everybody tin pass similar TCP/IP from the internet, basically. So we physique those libraries and RL Swarm is an illustration of that communication.

And then, finally, verification.

Ah, and I’m guessing this is wherever blockchain comes in…

Imagine a script wherever each instrumentality successful the satellite is executing consistently. They could nexus models together. But tin they spot each other? If I connected my MacBook to yours, yes, they could execute the aforesaid tasks. Yes, they could nonstop tensors backmost and forth, but bash they cognize that what they nonstop to the different instrumentality is really happening connected the different instrumentality oregon not?

In the existent world, you and I would astir apt motion a declaration to say, yes, we hold that we'll marque definite our devices bash the close thing. In the instrumentality world, it needs to hap programmatically. So that's the last portion we build, cryptographic proofs, probabilistic proofs, crippled theoretic proofs to marque that process wholly programmatic.

So that's wherever the blockchain comes in. It gives america each of the benefits of blockchain you tin imagine, similar persistent identity, payments, consensus, etc. And truthful what we're doing with the testnet present is taking RL Swarm and the primitives of the different infrastructure and we're adding successful the blockchain components and saying, ‘Hey, erstwhile you articulation a swarm now, you person a persistent identity, which exists retired determination connected a decentralized ledger.’

In the aboriginal you’ll person the quality to marque payments, but close now, you person that spot statement mechanics wherever we tin terminate disputes. So, it's benignant of an MVP of the aboriginal Gensyn infrastructure, wherever we’re going to adhd successful components arsenic we go.

Give america a tease of what’s coming down the pipeline?

When we scope main-net, each of the bundle and infrastructure is unrecorded against blockchain arsenic the root of trust, payments, consensus, etc., identity. This is the archetypal measurement of that. It's adding individuality successful and saying erstwhile you articulation a swarm, you tin registry arsenic the aforesaid person. Everyone knows who you are without having to cheque immoderate centralized server oregon website somewhere.

Now let’s get chaotic and speech further successful the future. What does this look similar 1 twelvemonth from now, 2 years from now, 5 years from now? What’s your North Star?

Sure. The eventual imaginativeness is to instrumentality each of the resources that beryllium nether instrumentality learning and marque them instantaneously programmatically accessible to everyone. Machine learning is heavy constrained by its halfway resources. This creates this immense moat for centralized AI companies, but it doesn't request to exist. It tin beryllium open-sourced if we tin physique the close software. So our presumption is Gensyn builds each of the low-level infrastructure to let that to get arsenic adjacent to open-source arsenic it perchance can. People should person the close to physique instrumentality learning technologies.