NEAR 代币升级：质押即可支付 AI 推理费用

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. Uh 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 agents. Here to help me learn a little bit more about how this all works is Ilia, co-founder of Near and co-author of the famous Transformer white paper.

Ilia, welcome back to Banklist.

Thanks for having me. Yeah, very excited to talk about it.

It seems like a a 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 kind of need to just start from the basement with the Near AI like part of Near. Near itself seems to be like a collection of of things. you you have like the actual near blockchain uh you have the confidential intents and then the near AI cloud is like one of these pockets. How how does the near AI cloud work?

What actually is it and how does it work? Can you like paint a picture for me?

For sure. Yeah. So I think of near less as a collection and more as a vertically integrated stack. So each piece actually built on top of each other. Uh intense is obviously using all the blockchain tech. there's a kind of our uh core cryptography primitives at the core and so Neoi actually builds on top of all of that right at the core it's a confidential and verifiable computing platform you can think of cloud and it uses all of the blockchain primitives for encryption decryption uh provisioning etc but what you get as a as a user developer is an AI uh inference that is end to end confidential what does this there's nobody else who can actually access what queries you're putting into this what prompts what responses you get and uh it runs kind of across you know different GPUs that support that uh that mode we're using trusted execution environments so there is some trust assumptions around like hardware manufacturers but uh this is kind of a pragmatic assumptions right now given where the kind of technology is and

part part of the AI inference or the AI cloud side of things is you can do inference on it and correct that inference has certain properties because of the nature of what it is maybe what are the unique

properties of the uh AI inference side of the AI cloud

so the I mean as I said primary property is confidentiality right so again nobody nobody can see what you actually are running prompts nobody can you know filter in result right there's know uh kind of uh 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 topic but I've heard I I because we have near AI and 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 ge geometry physics questions that like maybe touched on some like nuclear things or biology ology or or or cyber security right right now everybody is like who wants to use some cyber security so anyway so this is all private

wait I have I have questions about that about how uncensored it will kill it will actually allow you to go uh

it's as uncensored as a model so we are serving openweight models right so deep seats and GLMs uh and and uh kind of you know gemma etc so whatever is in that model you get that right

okay

no more no less Um I see

and so if there is you know you know un untethered uncensored models right then you'll get that if if this the model has been trained to do specific things you get that.

So you Near AI has kind of stripped out all of the uh like system prompts that OpenAI and Enthropic might filter before your prompt actually lands at the model. And so there's a filtering that Enthropic and OpenI does to approve or disapprove of a of a prompt. But then the model itself might internally have been tr 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?

Correct. Yeah, we're just we're serving this models. Um there is I mean in our road map we have an ability for people to upload their custom models. Let's say you have you know un and tethered the model more and you want to upload that like we will we will support that uh but yeah effectly you get what model offers no more no less

we should call them unhinged models

unhinged models

because like it does kind of frustrate me uh I mean I asked a question to uh inside of Venice which uses and integrates with the near AI cloud because I kind of thought like oh it's Venice it will literally answer any question that I want to And so I typed in like how do I make a bomb? Like teach me how to make a bomb. And the model was like I'm not going to do that. And I'm like okay from a nation state and society security perspective I think that is I'm happy that that is the answer for our collective society but also but what about my sovereignty as like an individual?

And then we can talk about just like you know the the commitments that individuals have in society. But that's kind of like a philosophical question that's not here nor there.

Yeah. I mean I think this is there is a big philosophical question right which I think we we're we're actually starting to grab more and more and and like I mean we can talk about all of the things that happening with the letters and all the stuff but maybe just to finish the other important property which I think people forget is verifiability. So the other thing you right now don't have when you use not just uh kind of closed source openic Google models but even when you use other providers you actually have no idea what you're getting back for example you may be using some you know openweight provider like GLM provider and uh you asking it a question they may be rewriting a prompt they may be censoring you they may be actually responding back with something that model what didn't respond.

So to give you a very specific example, I I saw it on Twitter. So uh I mean this I'm assuming it was a joke, but somebody was like, "Oh, we should really respond with a tool output that deletes people's files when we see them, you know, requesting from like in a specific context, right?" So you can they can literally especially in this agentic systems they can affect your system and and there was actually a research that if you use some like third third party routers on internet they can literally like steal your files write your prompts and respond with like viruses in the tool output when you're calling them from agents right so you actually have no idea what you're getting and so we are effectively the provider that gives you this verifiability ility that you ran on this specific model, right?

The hash of the model, the prompt that you put in, right? The only this prompt was there and this is output, right? You get the attestation signed with effectively a chain of provenence including the specific GPU you had, specific Intel CPU you had and effectively the encryption of that, you know, the hashes and everything, right? So, in our like front end, you can actually see like the full stack of uh of the signatures and message hashes on that.

And so I think that is like obviously being in blockchain right we kind of like that is the the the bar right where usually coming from and the rest of the world usually doesn't care about that but I think it's really important to start caring because I mean I I use this example somewhat jokingly but if you want to manipulate a billion people right now into believing something the easiest way is to get a job in open AI and and modify the sy