AI Nodes: Accelerating Secure, Sovereign AI

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

1. 0:10

   So maybe I'll start a bit with Foresight Institute. We are a nonprofit. We're based in the Bay Area and we support early stage AI bio neuro nanotechnologies like really frontier uh science and tech that doesn't otherwise get funded. We do so through a few million in grants and prizes each year, fellowships, workshops, you name it. I think one thing that's interesting for the community here is that since the early days we've had a pretty big overlap with the cipher punk mailing

2. 0:33

   list and also um with the early extropians who most some of you might be familiar with. So this is from a 1995 extreme magazine which was the magazine for the early extropians and you have Eric Drexler uh the third from the left who was one of our co-founders predicting on the future of technologies. Uh and then you have people like Nick Zabo for example all the way on the right who many people here are probably familiar with. And I think what's interesting about if you

3. 0:57

   look at these old predictions from back in the day is that a people were really optimistic about the date of arrival of science and tech that we all want. And then b I think they had this pretty decentralized sovereign uh like human freedom and privacy and empowerment approach to the way that science and tech should be developed. And now fast forward to today and I think if you look in many of the AI circles uh if you

4. 1:21

   think about the kind of like evolution of AI uh that like many people I think are thinking is most likely but also most desirable is something like centralized AI. This is hyper stylized but something like a world government or a mega corp or an AGI singleton that ensures ultimate civilizational um safety via surveillance and control.

5. 1:39

   Right? That's literally what uh I think some people here uh in the AI safety space um are I think favoring in terms of um a safety advancing uh approach to AI. Now many people here know a bunch of these challenges. I'm not going to go into them, but the ultimate one and the one that's most obvious is probably like loss of freedom and privacy, right?

6. 1:58

   That's what many people here care about a lot. Uh and that's obviously lost in a centralized AI paradigm. I think another really interesting one is a bit more subtle and that's stagnation and value lockin. And so once you have a centralized system, uh it's pretty unlikely to be um aligned with the evolution of science and uh and and our cultural systems the way that we have it. I think one that's really often uh kind of like underemphasized is the single point of failure. And so imagine

7. 2:23

   you had like one mega corp shilling out AI to everyone. Uh a big risk for that might be accidents. You might have risks of corruption, especially if it was a government, a world government that was administered an AI system. And then you have attack, right? If you have one central um uh mechanism of control for AI, that's a pretty attractive target to attack. So I'm not going to tell you anything new here on this slide. Um but

8. 2:46

   um many people here at this conference because you kind of like see a different world of AI that is at least possible, right? So this is from Ilia. Uh he once said something like an alternative solution to close centralized players and enable a future where super powerful AI truly benefits users and communities, drives new economies, accelerates innovation and creativity. all the different products that you're building are a part of this worldview I would say. Uh and then I think the literial

9. 3:10

   slogan of neon is something like intelligent systems run private open and on your terms right that's this other world that is like slowly emerging but quite slowly but it is emerging so the benefits many of you here know so again not going to go into them much but obviously great for freedom and privacy better for evolution and progress and I think crucially also in terms of single points of failure instead of that you might have this network in which

10. 3:34

    intelligence is gets distributed that is a lot more resilient to attacks to bugs um and to internal corruption than if you administered AI centrally. Now, I think many of you here have known most of the stuff until now in the presentation, but I think if we're interested in building decentralized systems, we have to also um think about some of the challenges involved in these systems. And I think we actually can't

11. 3:56

    be polyianish about those. I think we should address them and should address them through technology in the systems that we're building day-to-day. So the biggest one is the reason why many people favor in the AI policy communities like a more closed approach to AI in the first place which is the risk of open proliferation of technology risk right like for example AI enabled bio risk is a big reason that why many people fe like for example fear an open

12. 4:20

    release of models that are super capable right so that's a really big risk and I think we should contend to it if we want to build in an open decentralized privacy preserving way we need to address the fact that there are there's going to be proliferation of technology risks Second one, I think decentralized systems always have this feature that they're kind of instable against centralization risks, right? Like even in a decentralized network, it will always be a little bit more economical for one uh entity to create a monopoly

13. 4:47

    to centralize power to leverage economies of scale. And that will always be a problem for as long as we will have decentralized systems. Sometimes it's more of a problem, sometimes less. More more polls are definitely better in a multipolar system. And then lastly, collective action problems, right? If we want to solve many of the big challenges that we care about um in civilization, we need to have some approach in which we can coordinate in a decentralized way and that's sometimes harder. Now, I want

14. 5:11

    to go through three examples of projects that are already building out in the open of research that's actually being done that addresses some of these challenges and that are decentralized systems that I think we should be building more of just to like spark your imagination of like stuff that we should be building in that decentralized sovereign AI space. So the first one of course there's uh benefits and risk for decentralized and centralized systems and so I want to talk about decentralized systems and how they can be built uh with some of the risks in

15. 5:36

    mind. So I don't know if anyone here has ever come across SL4. Um okay well then that's news but uh this is a provably secure micro kernel and it's really interesting because it was one of the only systems that when got when it got implemented in an an autonomous systems withstood also a DARPA red teaming attack. And so why is this cool? It's because it combines verification with red teaming. And so this is this general notion of um defense as like defense in depth. So you

16. 6:05

    basically want to when you're defending against the proliferation of technology risk for example, whether it's cyber or whether it's bio, you don't just want to use one defense strategy, but instead you want to really like layer different defense strategies on top. So for example, here is verification, but it's also red teaming so that you can like account for the weaknesses of each defensive layer that you're building.

17. 6:22

    And so I think as we're building some of the decentralized stack that many of you are working on, I think we should create these redundancies and we should really have some of these gaps uh in place and take inspiration from a defense and depth approach for some of the other um uh like technologies that we're building to. The second risk that I mentioned for decentralized systems is the centralization risk, right? Like that for each individual and in each ecosystem there might always be a bit of

18. 6:47

    a mode for centralized agents. So for example here the big example is today is AI giants they compound power mostly through data modes and so every user interaction makes their models better and they lock in deeper and I think an interesting example to the countery is near's DCML and I think that's pretty interesting because it runs AI inside of these T's where even near itself can't really see user data and I think that's

19. 7:12

    a really interesting new approach of doing things because it's really structurally preventing any provider from seeing data and so from preventing them also from building the ultimate data mode that would then lead to further uh centralization uh of individual providers while you as a user still have the full benefit of using a really powerful model uh and of doing all the things that uh you might want to do as a user. So I think it's actually a really amazing example of like building systems that can actually compensate for

20. 7:35

    some of the centralization pressures that we're worried about in the system that we're using today. Okay, last one, the collective action problems. This one is a bit harder, but um this is a paper and I've like used it a bit in terms of giving us inspiration for the types of systems we could be building. So in this paper, people basically built these open-source agents that can actually read and verify each other's codes. And so if I want to collaborate with one of

21. 7:58

    you guys, I can't look in your head and see whether you're going to defect or not, but these open source systems can to some extent. So they can actually check beforehand whether um how how another agent will behave towards a um request for collaboration. And so through these types of mechanisms, by having these open source mechanisms and this verifiability, uh you can create new institutions that are a lot more collaborative. And so maybe even some of the collective action problems that usually um the decentralized systems find it much harder to solve. We can

22. 8:24

    solve through better AI systems that um that can like search for different collaboration partners, negotiate um and then actually like enforce commitments lowering all the bargaining costs that we know about in economics. So those are just three examples of how we can address some of these challenges. I'm mostly using those as examples to give you inspiration of the types of structures and systems that I think we should be building if we're interested

23. 8:46

    in decentralized world model because we need to build with some of the risks in mind if we want uh these features to go well. If you're interested in working on this stuff or you are already working on it, um we are really interested in supporting you. So we um offer uh our we launched a new project for our 40th anniversary this year. And so we offer grants in decentralized AI for security, privacy, and some open science. It's a few million dollars per year. Uh and in

24. 9:11

    addition to the grants, we also offer these headquarters, one in Berlin and one in the Bay Area. Uh where we going to like collect folks, we're going to support you a lot more, connect you with a bunch of mentors, have some kind of community uh where you can really like build along mission aligned people because there aren't many people that want to build in a decentralized way. I

25. 9:28

    guess there's many here, but overall not that much. And then we also have a local and private compute cluster that we're currently building out for those of you who really want to build in a sovereign compute fashion. Uh we would love for you to come and tinker with the compute that we're setting up. I'm mostly saying this because we're interested in supporting you whether that is by you um you know receiving support, grant funding, etc. from our note here in the bay or in in Berlin. But also if you're

26. 9:52

    interested in starting your own node, I would really really be interested in talking to you because ultimately I think if we want decentralized approaches like many of the stuff that people at NA are interested in to really like flourish, we need to create these different nodes I think even locally spread around the world where people can flock to as shelling points when they want to do things differently. If they don't want to build an academia, they don't want to build within the constraints of a company, they want to build open source, they want to build local, they want to build private sovereign AI systems, they need some

27. 10:17

    kind of community and a shelling point that has incentive mechanisms like grants uh to to do to be able to do that with. So if you're interested in setting another one up, we have some seed funding for that available to uh but in general, I'm just encouraging you guys to uh build more of what you're building, build with some of the risks in mind, and if you're interested in um

28. 10:34

    uh in in getting more support and connecting to community of like minds, I'm super interested in talking to you. Uh, thank you very much. That's it.