The Case for User-Owned AI | Illia Polosukhin

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:05

   We've transitioned from AI being a really cool capability and kind of experimented a prototype to now becoming more and more a clear kind of important enabler or uh mission critical part of businesses and people's personal life.

2. 0:25

   And it's really important for people and companies to really own their AI and really control how how can they use it. Right now they're not. Right now if you're using a centralized provider right you're not really in control. You are kind of by definition giving up all all of it to this provider. To give you an example, I was trying to think uh research about mRNA and I was afraid to

3. 0:51

   use uh you know some of the AI centralized AI products because it was not clear if I'm going to get banned uh for asking these questions, right? Because there's no uh kind of sovereignty on this like you do need to use something that's private, verifiable and end to end yours. Similarly, uh if you're a company and using this for business critical, you never know when you're going to get deplatformed or if something changes underneath uh in the system prompt in kind of definitions and

4. 1:23

   you start getting different results that you didn't expect and if this is how you make financial decisions, how you make medical decisions, it's a very critical uh critical question. So it's important that we have a AI that is belongs to the user belongs to the business and what are the properties of that well it needs to be private it needs to be verifiable and ideally you want to know exactly what are the what data went in into it

5. 1:49

   to really kind of turn this into the useful product that you're using and so that's what near AI is doing really enabling these properties for everyone and uh you can use it yourself or you know your company and uh other businesses can uh rely on it as a service.

6. 2:14

   We see major enterprises starting to sell uh sovereign AI and I think they're starting to target part of the AI stack, right? they are effectively pointing out that indeed when people are using closour models they are not in control of what's happening and so they're providing um kind of a this as a service uh to the companies with openw rate models right so this is a good step in the right in the right direction but it's not far enough ideally we want um companies to

7. 2:46

   be able to tap in into a something they can definitely verify that is correct. Right? So what is required for that? Well, it needs to be open source first, right? If you cannot verify end to end what is going on, you're not able to actually know uh that there's nothing uh getting injected in the middle. Second one is it cannot be centralized even to the company to some extent, right? Imagine even if it's all set up in your own basement. The challenge is you may still uh lack

8. 3:20

   capacity, you need to upgrade, there's a lot of uh complexity that comes in with operating on uh infrastructure. And so you want infrastructure you can trust but that is not centralized, right? That doesn't go down that can handle dynamic uh load uh while verifiably in private.

9. 3:39

   And that's again what nearai is doing by offering combination of secure enclaves and cryptography having signed upgrades and governance that's used on blockchain to really enable this whole process to be done where everything's open source.

10. 3:55

    Everybody can verify for every individual inference where it came from, what hardware was used, where it's located, while uh knowing that nobody else seen the inference, the prompt, and knowing that there was no injection or changes to the model that happened between this blockchain is really serving as root of trust for this kind of AI solutions. And this is a kind of a general concept, right? Right now when we think of internet, right, we when you type in

11. 4:30

    google.com, you're actually trusting uh a few different services to resolve you to the right IP address, right? First, you're trusting the DNS to give you the IP address for the domain, right? And uh DNS is an is a dist distributed network but it is rooted in you know some authority that has actually said hey this is uh this is the server servers that are going to give you the correct answer but more importantly for encryption right uh when you communicate to google.com you're

12. 5:04

    trusting the certificate authority right this is effectively issuers of certificates and there has been examples where the certificates were wrongly issued or were hacked when somebody could have actually had a valid certificate for google.com domain uh to really inject their own malicious interaction. So what root of trust here is is really the dissipating removing this authority the centralized authority in most of the cases and turning this into a blockchain governed smart contract protocol and uh like to give

13. 5:38

    you an example right for AI uh let's say there's a new model came out somebody needs to review approve and upgrade uh so that everybody else can trust on it and so you want this process to be decentralized you want this process to be auditable you want this process to clear to everyone and exactly verifiable. That's what near kind of blockchain as a root of trust can offer.

14. 6:00

    And it's not possible to do if you're just doing it as a centralized enterprise.