NEAR’s New Token Utility and AI Economy | Illia Polosukhin

**David Hoffman** (0:02)
Bankless Nation, there was a big announcement and evolution in the NEAR part of crypto. The NEAR token has got a little bit of an upgrade.
There's now a more formal integration between the NEAR AI Cloud and the NEAR token. So you can now pay for inference on the NEAR AI Cloud by staking NEAR. So, stake NEAR, receive free inference for yourself or your agent. Here to help me learn a little bit more about how this all works is Illia, co-founder of NEAR and co-author of the famous transformer white paper. Illia, welcome back to Bankless.

**Illia Polosukhin** (0:34)
Thanks for having me. Yeah, very excited to talk about it.

**David Hoffman** (0:37)
It seems like one of the larger upgrades to the NEAR token that I've seen in a while. In order to really understand it, I think we need to just start from the basement with the NEAR AI part of NEAR. NEAR itself seems to be like a collection of things. You have the actual NEAR blockchain, you have the confidential and intense, and then the NEAR AI Cloud is like one of these pockets.
How does the NEAR AI Cloud work? What actually is it? How does it work? Can you like paint a picture for me?

**Illia Polosukhin** (1:07)
For sure, yeah. So I think of NEAR less as a collection and more as a vertically integrated stack. So each piece actually builds on top of each other. Intense is obviously using all the blockchain tech. There's kind of our core cryptography primitives at the core. And so NEAR AI actually builds on top of all of that.
At the core, it's a confidential and verifiable computing platform. Think of cloud and it uses all of the blockchain primitives for encryption, decryption, provisioning, etc.
But what you get as a user or developer is an AI inference that is end-to-end confidential. What does this mean? There's nobody else who can actually access what queries you're putting into this, what prompts, what responses you get. And it runs kind of across different GPUs that support that mode. We're using trusted execution environments, so there is some trust assumptions around like hardware manufacturers, but this is kind of a pragmatic assumptions right now given where the kind of technology is.

**David Hoffman** (2:12)
Part of the AI inference or the AI cloud side of things is you can do inference on it. And that inference has certain properties because of the nature of what it is. Maybe you can... What are the unique properties of the AI inference side of the AI cloud?

**Illia Polosukhin** (2:30)
So, I mean, as I said, primary property is confidentiality, right? So, again, nobody can see what you actually are running prompts. Nobody can filter in result, right? There's no kind of censorship, additional censorship or blocking or whatever that's happening on top of this, right? I don't know, you know, if you've tried asking some sensitive questions to, you know, open AI on Tropic, but I've heard...
Because we have NEAR AI, I mostly use that for any sensitive topics. But I've heard of multiple people who got banned for even pretty like reasonable, like, you know, Geometry, Physics questions that like maybe touched on some like nuclear things or biology or cybersecurity right now. Everybody is like who wants to use some cybersecurity. So anyway, so this is all private.

**David Hoffman** (3:19)
Wait, I have questions about that, about how uncensored it will kill, it will actually allow you to go?

**Illia Polosukhin** (3:26)
It's as uncensored as a model. So we are serving open weight models, right? So deep six and GLMs and kind of, you know, JAMA, et cetera. So whatever is in that model, you get that, right?

**David Hoffman** (3:38)
Okay.

**Illia Polosukhin** (3:38)
No more, no less.
I see. And so if there is, you know, you know, untethered, uncensored models, right, then you'll get that. If this is a model, it has been trained to do specific things, you get that.

**David Hoffman** (3:50)
So you, NEAR AI has kind of stripped out all of the like system prompts that OpenAI and Anthropic might filter before your prompt actually lands at the model. And so there's a filtering that Anthropic and OpenAI does to approve or disapprove of a prompt. But then the model itself might internally have been trained to like not answer specific questions or to answer specific questions in a certain way. And you don't really have any control over that because NEAR is really about the pipeline of traffic and data of prompts to models. Is that, that's accurate?

**Illia Polosukhin** (4:29)
Correct, yeah, we're just, we're serving these models. There is, I mean, in our roadmap, we have an ability for people to upload their custom models. Let's say you have, you know, untethered the model more and you want to upload that, like we will support that.