Building the Bitcoin Equivalent for AI Inference Compute with David and Daniil Liberman (Gonka)

Austin Federa (0:05)

David, Daniil, welcome to Validated. This is a type of episode we've been trying to do for a while and trying to find the right people to dive into it more.

As everyone knows, as you know, the audience knows, the rise of AI and compute networks has been fast and profound. It has changed most major manufacturing sectors very quickly, just in terms of the AI datacenter build out. Crypto is sort of for a long time tried to play in this market. There's distributed AI projects that actually are just centralized AI projects. There's been a lot of distributed training that was maybe distributed, but it wasn't actually doing any training. There's been a lot of forward compute products that have maybe or maybe not had any resemblance to the actual underlying compute. But there's also some folks doing real work in this space too. And so I wanted to talk about a huge number of these different types of things and dive into it a bit with you guys who've actually been working on building stuff in this space for a little while now with Gonka. There are many companies that have, I would say, started with some type of AI ethics or some type of moral framework as an underpinning, right, Vartan? Famously, this was where OpenAI began. It's debatable where they sit now on that spectrum. But when you were looking around at various technologies to underpin what you guys are doing, and I want to get into that as well.

What about crypto? What about blockchain? What about tokens? Not the AI type, but the crypto type actually felt like this was a stack that might be applicable here.

David Liberman (1:51)

Companies can easily change their statements.

Google promised all of us back in 2004 that there never will be an ad in the main feed of search query.

So where is this promise right now? And the same with, yeah, and especially in general, these structures, they are not built for immutability. Yes, they're not built for making the promise set in code, set in stone. Yes, and crypto is actually originally started from that. Bitcoin is this premise that we will put it in the code, and it will never change, and the rules will never change so that we can all entrust it, and we can then invest more and more efforts in this structure together. So the same was our original attraction for us. When you decide, like, who should trust us, that we will start with the ethical statement and then we'll not change it. It should be a protocol. It should be a certain code, and that's where crypto already proven itself.

Daniil Liberman (3:10)

In the very beginning, we ourselves considered, should we build a decentralized frontier lab? Should we build a technology which will allow training of open source, truly open source, completely open source, open, not just open weighted as Chinese models and some of the American models, but with open pre-context, with open alignment rules, with open everything there open. But then we quickly turned away from the idea of doing decentralized training. It is hard, but it's likely possible. But why we turned away from this? Because we realize it, in any sense, it will require a lot of resources. Tens of thousands in the beginning and then hundreds of thousands of thousands of GPUs, high-end GPUs later on, to basically even train the model, a lot of data centers, a lot of electricity. Tens of billions if not hundreds of billions of dollars investment to train those models. And if it's so, then there would need to be a revenue behind it.

And because there would need to be a revenue behind it, revenue is clearly not in building the models themselves, but in providing the service out of these models. So the revenue would come from the inference. That was our conclusion around three years ago, that eventually there cannot be other paths, but the path in which major sources of revenue is going to be inference. And therefore, the focus of any project which wants to be dominant or to take any share of the market space in the future would have to be inference, would have to be providing the API access or any type of a service access to AI rather than just basically training the AI.

And what was interesting is that we saw that Bitcoin for us turned into a financial framework rather than anything else, rather than a cryptocurrency. We moved all of this aside a little bit. Yes, this is a philosophical stone, bare bone of what Bitcoin is.

But we put it aside and said like, if we look at this through the perspective of financial framework, it would say that whoever adds more compute to the network, to the infrastructure behind Bitcoin, they will get the higher share of the mined coins. And as the network grow, as the cost of mining of every coin grow as well with the size of the network, with the sheer investment into the infrastructure, owners of these coins will win significantly. So whoever invest in the infrastructure, they will win over time. The most of value which currently already in trillions of dollars, and the same will happen with AI. The AI infrastructure will be in trillions of dollars. We all understand that now. Like three years later, and we are in probably over $700 billion of investment into AI infrastructure, three years. Like three years ago, no one would have even thought that it's going to be that fast. Hundreds of billions of dollars invested.