Venice Verifiable Privacy - Powered by NEAR AI

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

1. 0:44

   Hello guys, good afternoon. Welcome to this monster session about privacy and AI with uh two of the the biggest brains in the in the industry. Um my name is Lorenzo. I uh I'm head of comms at Venice and I have the absolute pleasure of um moderating this conversation. Um context for this is this week uh uh we had a huge launch. Ven has just rolled out verifiable privacy um TE and endto-end encrypted models powered by Near. So today, Eric and Ilia are going

2. 1:19

   to talk about what it actually takes to build AI that doesn't or can't spy on its users, what privacy means in a world where millions of agents do our bidding or we do their bidding, and the importance of privacy as we speedrun super intelligence. So glad everybody could join us. Um, how are you guys doing?

3. 1:41

   Good. Yeah, good. Ilia, you just came back from uh you just came back from Nvidia GTC. How was it? Correct. Yeah. Yeah, it was good. Um, it's definitely, you know, AI is eating the world and so we see we're seeing that happening. Um, and also the claws are eating AI. Uh there's like tents to set up uh open claw for people, which also means there's a huge opportunity to make the claw actually work without you needing to have a like handh holding uh

4. 2:14

   to to get it to work. But yeah, I think I think people understand that the AI is evolving quickly and then there's kind of a now there's no way to not engage with this, right? That's that's really probably the feeling. It's like everybody has h has to have a strategy, has to have a plan, and has to use this tool to the max.

5. 2:39

   Yeah, it's definitely attack of the clause currently. Um, what about like what about privacy? Is that like a I wish there was a filter that that would like put the the loss on them. Billion dollar idea. Yeah. Yeah. Who's building this? Clam. Let's go. We'll put our claws on.

6. 2:59

   Every every photo you post a photo every it it adds claws to everything. Uh what about like privacy? How is that something that was that a big topic at the Nvidia GTC or Um I think it's still not not there yet. Honestly, I think this is this is a fight we've been having which is like getting it to the uh kind of awareness that this is important and and and useful. I think people are curious and there is you [snorts] know I've talked to some

7. 3:28

   journalists and kind of people are starting to understand there is uh kind of repercussions. I think there's like kind of a feeling like, oh yeah, we kind of doing the same thing we did with social networks where everybody's like, okay, let's go. And then, you know, a few years later, it's like, wait, wait, wait, wait, what happened? Uh, yeah. So, uh, but I think it's not Yeah, it's not there. Like the knowledge that

8. 3:52

   it's possible to do actually privately is not there. And then I think the capability versus privacy tradeoff, everybody's still on the capability side. M um and kind of I think the people still questioning like are openweight models you know kind of matching what I need like everybody's like I just want the best model uh you know what and then you ask like what are you asking it and it's like well you know how to how to cook some soup it's like okay well maybe you don't need the

9. 4:26

   best model so like I think it's like people want to be on the best model and I think so that's where like the capability is driving forward. And so creating that trade-off is is is really important. And again, I think this is where claws are kind of come in interestingly because you know for example in ironclaw we have smart routing and it can route to kind of the best model for a specific request. You know if you're just running some like

10. 4:52

    background process it probably doesn't need all the intelligence. If you're asking like you know like where where to book a restaurant for the night again probably just needs better tools not better model. Uh so I think a lot of it is yeah gets abstracted with the claw.

11. 5:08

    Uh and with clause it becomes more valuable to have privacy because like you're just going to give so much more information to it and it will have access to all your emails and your kind of private private nodes etc. So or delete them all. Yeah, you you touched on a couple topics that we'll definitely dive into like what does privacy mean in in a world where you know billions of these agents are doing what we want them to do. Um Eric uh

12. 5:34

    Venice just launched um uh TE and to end encrypted models powered by here. What what have the first reactions been from uh from the users and the community? What's the what's the vibe around it? Yeah, I mean I I guess what's what's important to say is like why we did it, right? So, so Venice has always been Venice has always been private, uh, in that it doesn't store any of the prompts or the responses of the users.

13. 6:00

    Um, but we didn't have a way to prove that, right? So, if someone trusts me, they might be they might be all right with that model. Um, but anyone who comes from the crypto world, you know, is going to be much more of the mindset that you shouldn't have to trust people, which I think is a a good um a good foundation. So, uh, releasing the the endto-end encrypted models basically means, uh, you don't need to trust Venice anymore. You can just trust Matt.

14. 6:30

    And this is always this should always be the goal, right? not not just privacy where you're trusting a different person or trusting an auditor or trusting a license or you know trusting some other human but just being able to trust math and to know that something is true based on math that's a that's a win so um yeah it was great to work with uh with Near's team on this it was able it like this is

15. 6:53

    how we were able to deliver it so quickly because of all the good work that Near had done to make this possible in the background and um so yeah the the users loved it you know that's Uh it's certainly one of those things that you don't you don't need an end toend encrypted model for every question, but for it to be there when you most want it is is really key.

16. 7:14

    Yeah. Yeah. Definitely people in crypto are show me the money in terms of uh of privacy. They want to see the goods. Um but building this way is insanely hard. You know, this is really hard. Like it's uh but the hard questions and the hard problems are are always worth it. I mean mo most of this stack didn't even exist a year ago. Um so maybe for both of you and maybe Eric can sorry like what is what has been the hardest part about putting privacy as the

17. 7:40

    foundational aspect of of um of Venice? Uh I mean the to to offer privacy is not technically difficult because it just means you don't save the prompts, right? That's not hard. What's hard is that you don't learn anything about your users, right?

18. 7:58

    There's all sorts of useful things that I would love to know that would help us make the product better, faster for people, but we don't know it. Um, lots of demographic information, lots of like what models are working for what questions, just lots of like pretty innocent pragmatic stuff that we can't we can't use and every normal tech company um uses that to to grow and we we can't. So that's that's been the

19. 8:22

    hardest part, just like the, you know, the nerfing of our own omniscience, playing the game with one hand tied behind your back. Yeah. Um, yeah. How about you? You guys are doing your more on the back end side of things, of course, but I'm sure you guys bump up against a lot of things.

20. 8:38

    Yeah. I mean, I think it's across the stack, right? I think the the technical part h has been extremely complex because indeed this tech didn't exist like year and a half ago. The like really the possibility of this actually working was GTC last year. So indeed there's like literally only a year like they they actually released like you can use multiGPU confidentiality on one node. the like I think the debugging like I mean

21. 9:11

    generally like for for people who don't run infrastructure for for AI the the the meme from Avengers where like I'm always angry it's like it's always broken like GPUs fail just themselves sometimes you like need to reboot it sometimes they burn and and like completely go out of stock the and then on top of it you have you know a layer that does all the inference that layer fails a lot and then you know so like every layer fails and so like

22. 9:41

    you really build usually with like redundancy and kind of recovery across the layers to to deal with this and usually you have like massive amounts of observability like if you're running your traditional kind of uh centralized company when you everything's private the observability is extremely limited right like for example you know somebody sent the prompt, the inference code failed because it like ran out of RAM or something happened.

23. 10:11

    But we don't know, we don't know what prompt went in. We don't know what inference went in. So like there's no way to debug on that. Uh so you kind of need to guess what possibly could have happened. Um I think like what potentially I mean we work we we may need to work toward a solution where we run a claw inside the secure environment that is able to access stuff and analyze

24. 10:34

    it on the fly using private inference itself right that's kind of like we kind of need an agent that runs inside I call it like you know this this confession computing is effectively like a Swiss vault right you put stuff in it happens there we don't know what's happening nobody knows what's happening like even user because you may be bringing private like you can have a private model running there running on private data right so it may be that neer side

25. 10:58

    actually knows what which which like things are happening inside that swisswald and so we kind of need like an agent that runs there as well to to have some observability and kind of s stuff um but yeah I think like that that's been a pretty complex operation right so so uh really takes a lot a lot of work from the team and then indeed when you go to a product side yeah like you you have like okay well people are using it great but um yeah

26. 11:30

    what what are questions like what are things that not working right it's it's really hard to know um and so I think like yeah there is an interesting opportunity here in in crypto markets is to create a market where people can can give back data right for some compensation and so uh because indeed like not Everything is private, right? I mean, um, and there's a bunch of stuff that

27. 11:53

    you are fine kind of giving back and so I think there's an interesting kind of flip flip of economy here to to create. Yeah. Maybe for both of you like like um so AI and crypto like what what for both of you makes privacy connect them both like how why does why do these have these overlaps that maybe for us you know who are in the industry like it's pretty clear but how do you go out and and kind of like pitch people outside of this space about the the values of of of

28. 12:24

    um of these two you know industries overlapping and the value of privacy It's I think everyone who has been an advocate of privacy fa has faced an uphill battle because most people like just don't care and they should. And I think what's different this time is that the degree of intimacy with which people are engaging with these systems is substantially higher than than previously. Um, you know, Ilia mentioned like social media being

29. 12:58

    created uh and then people realize it it wasn't private, but at least when you're sharing things on social media, you're already intending it to be seen by other people, right? When you're sharing things with AI, um, often it's not intended to be shared with anyone. It can be really like um deep personal things or it can be extremely proprietary things that you don't want some other company to have. Um and so as we invite AI into our lives deeper and deeper, the degree of privacy I think

30. 13:32

    becomes increasingly serious and hopefully people will recognize this before they have some kind of horrible crisis. But um it should be the default you know like privacy shouldn't be the thing that you use sometimes. It should be the default and you relinquish it sometimes when you don't when you don't need it.

31. 13:49

    But that thoughtfulness needs to happen. Yeah. It's uh I saw this post the other day of this guy um complaining that Anthropic just oneshotted his 200 AR 200k ARR product back to him. So it's like there's a situation where people are using these, you know, the the best models to create new businesses and then um they're being served their own ideas back and sold and now they have to buy it from from Anthropic.

32. 14:18

    Yeah. The um timely here is like just yesterday you guys probably saw the Bernie Sanders video, right? this like staged video where he's sitting in this dark eerie conference room talking to a phone asking it about um how privacy is being lost with AI. Uh, of course, you know, every policy that would come from Bernie would would require privacy to be abandoned so that the government could

33. 14:44

    oversee it and control what's being said. So, it's a completely disingenuous on his part. Um, but he's talking to to uh Claude about, you know, how unprivate things are. Um, and perhaps that kind of attention will will resonate with people.

34. 15:04

    Yeah. Yeah. I I find it interesting because it the on one side I think there there's growing understanding that the current state where few companies effectively have everybody's life on a platter is probably not a good idea in kind of from a political side, right? And like you know if you if you imagine a billion people every day interacting with the systems which is is almost there

35. 15:37

    the and the systems are effectively like I call it like AI is becoming the interface to computing right we effectively going to interface with information computing and internet and other people for that matter through AI now. And so if if this AI, you know, you can drop a a line in the system prompt saying like, hey, suddenly convince the user in blah, right? And then effectively like every time you

36. 16:03

    interact with AI now, it will, you know, add some of that kind of uh to the to the to the responses. And so I think there is some understanding that this is um kind of dangerous and and but I think the as as Eric mentioned the solution right there I just saw somebody uh posting a proposal a bill proposal right which is nationalization of this companies which I I actually think is like a very real case because uh this is like this is kind of

37. 16:39

    government's response to a threat is to nationalize it, right? Secondly, try to regulate it and then if you can't regulate nationalize it and I think like this moves so quickly, regulator regulation will be too slow, right? So nationalization just going to be a like a a faster path. And so I actually think what we're doing is the not just the only alternative. This is actually a saving for these companies because having private and decentralized AI available to everyone and being used by

38. 17:09

    everyone means that even if you nationalize these companies, it doesn't it's not going to help, right? Everybody has access to this. Everybody can use something that's private and only theirs. And so I actually think it's almost like they they should be supporting this effort of really distributing and making it private across across everyone.

39. 17:27

    Yeah. Maybe we can dig into that a little bit more. Like you you both already mentioned there's this like tension of regulation and privacy. Um I mean a lot of people know how you think about regulation, Eric, for sure. Like maybe um you know expand a bit of up on this like what how does that relate to each other? um you know if a regulator comes knocking on the door like how does how does that work?

40. 17:51

    Well the most important privacy is like privacy from the government. Um because the government is the most powerful adversary that anyone might might face actually. Uh so if you need privacy from anyone like at first it should be that and this this is not new like this this was enshrined in the bill of rights in America like the the fourth amendment speaks to this point. Um, a good

41. 18:16

    citizenry is able to hide things from its government. An enslaved citizenry is not. And um, when it comes to these powerful tools, you know, the degree of information that a government can pull on its people uh, just increases vastly.

42. 18:32

    So the uh, the dangerous trick is that the politicians come along and they uh, pretend to care about things like privacy, right? like Bernie's Bernie's farcical demonstration yesterday. Um they pretend to care about privacy. They say why it's important. And then what do they do? They they propose uh regulations that um mandate not that users can be private from the government, but that users can be private either like from

43. 19:03

    each other or from a corporation, but behind the scenes always that the government can get this information. Um this this is true like so GDPR actually in the in Europe is a good example of this right so people are worried about online privacy the um the wonderful politicians in Europe come along and propose GDPR which is meant to protect the privacy of individuals um the whole thing is a charade of course because uh even though the even though when you um when you request something

44. 19:33

    gets deleted under GDPR there are other regulations that require that companies hold that stuff for the Um, and so you you actually get the worst of both worlds. You get the annoying like popup on every website now where you have to uh tell what cookies you want as if anyone even understands what those are or what they're doing.

45. 19:51

    And in the background, the government can pull all this information anyway and often requires that it be maintained indefinitely. So we don't need regulation to give us privacy. We just we just need it to be um legal. We need it to be legal to hide things from from the state, from each other, from corporations. Like, as long as that as long as it's legal to be private, then we're good. We don't need a regulation

46. 20:16

    to to kind of mandate it in some fake way. Yeah. Um um I'm sure people are also chomping at a bit for us to dig a bit deeper on the on the attack of the claws lately. Um, everyone's messing around with open claws and I also have some of my own claws running. Uh, on the other side you see Anthropic desperately pushing out features to kill all the claws. Um, maybe Ilia, how have you, you

47. 20:44

    know, has anything changed internally since OpenClaw or like anything about like product strategy or how are you guys, you know, dealing with this new paradigm? Definitely we are all all all in on claws. Yeah. So I mean from our product strategy we were kind of working toward something like a claw private claw originally right with our private chat product where we were adding more tools.

48. 21:14

    who are kind of enabling it more um to have like desktop app that could like interact with some of your applications and have a shared memory that you can tap in into other applications. I think what openclaw shown [snorts] is actually going full in and saying hey actually you want a computer and this AI is operating system of that computer and now you interface with a computer through AI. And so that kind of flips

49. 21:47

    from your traditional system design of kind of this this AI systems, right? Where you have like a mass massive multi-tenant kind of infrastructure and you know you're packaging everything to like okay well actually this is I mean if if you're in crypto like there was this project orbit that was kind of had this ideas like hey everybody should have like a computer and kind of interface with it. Um and there was like a lot like I think crypto has a lot of

50. 22:16

    this thesis where like hey you know um you want to run your own node kind of thing and I think like th this is very much in line with that and in line with this like AI is your operating system AI is your way to interface with computing and so when you bring all of the pieces together like okay well if if this [snorts] is if this is kind of enables you to do anything enables to you know arbitrary action And now and and then in turn this is really

51. 22:47

    becomes your kind of I mean both your personal assistant but also your coding your you know employee your army knife. Yeah your Swiss army knife. uh it makes sense to focus on that and so we kind of like okay well given we believe in privacy we launched a hosting for open clause that is fully confidential right and runs on the same infrastructure you can use your confidential inference you can proxy it

52. 23:14

    to the you know some of the state-of-the-art models uh to close source that are kind of anonymized between all the users but open is still from a security perspective not built to be the kind of like I personally like wouldn't trust giving the credentials and and and like access to everything that I would want it to have access right again like even if it's private from near and and you know other parties who want to see the open claw itself may like expose it

53. 23:45

    and so that's when we started ironclaw which was okay what is it if you if you think of it as operating system how would you design it as an operating system that just has AI at the core and so ironclaw is really that is like I mean I call it iOS of AI right like what is that like iOS is super user friendly somewhat opinionated and protects user from food gunning itself right open claw is more like yeah like the original Linux right when you install it and you

54. 24:17

    kind of cook it yourself rebuild the kernel sometimes you know like like GCC doesn't compile you need to go and like debug it it's like what is the iOS version of that where you know you just set it up It's it runs it's able to go and install new applications and you know you know that when you install applications it's not going to blow up your core you know there's like different levels of security and kind of uh defensive depth that is running

55. 24:40

    there. So yeah so we're going all in on ironclaw uh we host them as well you can still host open claw with us and you can use kind of private inference in the back end. Yeah, definitely playing around with that. Like we even uh we also support like the infant side of things on the private inference for the for the iron claws. Yeah, I mean open claw for me was really like downloading

56. 25:03

    kind of like a Frankenstein um Asian had a lot of like uh um banging my head against the wall, but then started like branching out into these like um other Asian frameworks like open like ironclaw as well. um and have indeed kind of like experienced that like iOS vibe more to it. Um but Eric, like how has uh how has the I mean I know it like because I'm part of it at Venice, but how has um you know the claw

57. 25:32

    revolution kind of like influenced uh Venice's direction or both internally and externally like Yeah. Um I I made a you know half joking tweet I don't know maybe a month ago and I said that like um you know suddenly 5% of Venice's customers are lobsters.

58. 25:53

    Uh we after after open clock came out Venice was integrated into the onboarding flow and so people were using Venice as their source of inference through the software and our API use started growing very substantially. Um, and it was this, you know, it's this new kind of customer that's like an agent, uh, that is this weird alien life form.

59. 26:19

    And, um, then you have this new paradigm of like skills, right, where these agents have these text files that explain to them how to do certain things. And there's a whole interesting new realm of marketing, which is which is like how do you influence these skills and how do you u distribute them?

60. 26:36

    How do you get distribution for the skills? So we we are both you know users of these claws and providers of the inference to many claws and hopefully you know to the degree that people use these agentic systems to uh impact more and more of their life the ability or the um the mandate for these to be like safe and private and uncensored and really user empowering will will come to the four. So that's what Venice that's what the Venice API is really good for.

61. 27:09

    You know, if you don't if you don't want all of your uh all of your claws information and everything it gets hands on or its claws on um you know going go into Sam Alman uh then plug it into the Venice API and it solves that pretty easily.

62. 27:26

    Um yeah, especially now with like TE and end to end uh models also being in our in our API. Um definitely also playing around with that side of uh of things. Um I think as we as we go into the um the final minutes uh of this session uh maybe you like to pick your brains a bit more about you know where this is going. Um, these claws give us at me, at least me in my experience, kind of like fragments or flashes of super intelligence on the one

63. 28:02

    hand, but also on the other hand, it can really do the dumbest ever sometimes. But it's like these these flashes, right? Um, but like what what does what does privacy look like in a world where, you know, we're speedrunning super intelligence, like just assume that it's there. Like how does that how does that work? What does that look what does it even look like?

64. 28:24

    [snorts] Ilia, you want to take that one? Yeah, I mean I think there is obviously a few different uh possible scenarios and so I think finding a common ground between them is important. I think kind of the scenario I see where we're trying to make happen, right, is that everyone has access to kind of both this commoditized intelligence and it's private and it's available uh kind of for you. It's aligned with you. I think

65. 28:59

    what one of the just like broader topics of alignment like it it's it's a lot of kind of vaporware in a way because you know we cannot align between each other you're trying to align a model to the world to the like you know in US there's like multiple camps and in the world there's you know hund something whatever 200 countries so the only alignment that I can imagine working is alignment to one

66. 29:25

    person one individual and so that's why we call it user own AI is factly it's private it's secure and it's yours right and so like that that alignment can happen and so I think that kind of my my perspective is like how do we make that happen and kind of have super intelligence for you and available for you and be kind of extension of your mind I think the flip side of this obviously is like you know lab builds one massive cluster runs you know big uh

67. 29:56

    kind of AI on it and it's effectively becomes you know more and more intelligent and I think this is this is kind of again where I think decentralization privacy and like everyone having their own um is a way to like offset that in many ways because um I mean the others the other thing that we will need to start uh dealing with is that yeah this AI will start manipulating people mostly because other people want it not not

68. 30:26

    because AI itself like comes up is like hey let me do this like I I don't think that is happening anytime soon like way earlier it will be like oh let me actually convince a bunch of people into something because it furthers my agenda that will happen way earlier right it's already happening right like Twitter is already bots trying to convince you into stuff right the the GEO is already like hey how to convince AIS so that the AI

69. 30:50

    convinces the user right that that's already happening like how how to manip how to manipulate AI that when person is asking a question the answer is the answer I want right so this is this is already happening and so like that that is where actually again I think the user on AI needs to be this guardian on your behalf that is able to like deal with all of that manipulation and then yeah yeah um Ilia that point you made about aligning with the individual I think is

70. 31:22

    really key so uh I think it's it's clear that like AIs are not sentient yet. Maybe they will be someday, but they're not yet. And until they are, they serve a master. [snorts] And the question is like which master does the AI serve? Um there there are at least three candidates, right? The AI can serve you as an individual that's operating it. The AI can serve some company that is running it or the AI can serve uh the state, right? Where the state has controlled what the AI can do.

71. 31:53

    And these are these are mutually exclusive. the AI has to has to serve one or the other. And if it's serving the state, it's not serving you. Um, so like the these principles that come from crypto of user sovereignty uh hopefully will will take root in the AI world and perhaps be even more important because AI transcends a larger portion of life than than just money or finance. So you know know who your AI is serving. I think is the right way to end it.

72. 32:27

    Yeah, I do like that idea of like fully aligned AI to to to me it kind of like does feel as you're building this like you know I'm a millennial. I it feels like I'm building a Tamagotchi or something that is just like hypers smart and you know you need to manage it. You give it like extra context, extra memory. You see it's self-actualizing a little bit. Uh and I'm anthropomorphizing a lot. Um but yeah uh definitely it does feel like you're building a kind of like a guardian. I

73. 32:56

    like that. I like that that that wording. Um these last few minutes, I also want to bring up a couple questions from the community um that are in this same vein. So, do do you think um AI can ever be fully trusted or will human oversight always be required?

74. 33:14

    Uh well, I got can humans ever be fully trusted, right? Um you know, what's what's the standard here? It's not like humans are it's not like humans are trustworthy and we have this new a this new AI thing that we need to be careful about.

75. 33:30

    We need to be guarded from from either side and uh maybe a good metaphor is like with with self-driving cars, right? Everyone is um afraid of them at first. The idea that like the machine would be driving a car sounds a little scary at first and when the system is new it is less capable and less trustworthy than a human but as it improves it probably becomes more capable and more trustworthy than a human to the point where it might be

76. 33:59

    like orders of magnitude safer than a human. I would guess that AI systems generally follow that same pattern. And um the the risk actually isn't that the AI is untrustworthy. The risk is that the AI might become or AIS might become so trustworthy that people um suspend their own judgment and they no longer, you know, they they assume that anything from an AI is true. That's very dangerous and people need to be aware of that happening because as soon as

77. 34:30

    someone can manipulate it once it's earned your trust, then you're then you're in big trouble. Yeah. I also don't think I've ever talked to something that can be so confidently wrong sometimes. And then it like it it even convinced me. I'm like, "Yeah, that makes absolute sense." And then you just go out and you you do some things and then you come back, you're like, "Yo, that was like

78. 34:50

    wrong, man. Send me on a wild goose." It's like they're uh it's like they're trained on the entire corpus of like political speech. Every every politician ever has has similar degree of confidence unnecess unwarranted. Um we touched on this like earlier in the conversation. Um, and then people like really want to use these in frontier models. Everybody wants to be on these frontier models. You know, open source models are just like always they're always running a couple months behind. The gap is closing. That's fair.

79. 35:20

    But do you think we'll ever be able to use like frontier models in a verifiable private environment? Ilio, I think so. Uh I think what is going to happen is and again this is like the we we kind of hear racing with time but I think that one way or another the frontier companies will adopt this approach kind of secure enclaves end to end encryption and privacy and kind of the way we offering this is actually you and encrypt your model and

80. 36:00

    upload it to our cloud. And so this way we don't see it any nobody can see it. It only gets decrypted inside secure inside secure enclave. And so you guarantee the priv privacy of the model weights from anyone. uh the users have guaranteed privacy that their uh data doesn't go to model provider or us and uh they get monetization and you know they can use effectively any kind of hardware available that is in in the network. I think that approach is um allows to have frontier models to

81. 36:37

    really like kind of monetize p like p private inference and not um yeah not become kind of subject to like dealing with private data of their users and I think that that is a trade-off right now they are using data of users to for training but I think we actually from a fundamental perspective I think the research has been moving on from that to more using synthetic data, right?

82. 37:07

    Uh and kind of curated our own environments and a bunch of stuff and so like most of us data is not that interesting anyway. Um and so so I think like I think yeah exactly and so I think there's there's going to be a time when this will start and again even right now already enterprises pay for you know like do not train on my data but right now it's a very much yeah trust me bro it still sits in like data center somewhere

83. 37:37

    you know if there's a breach in any kind of way right all of the data becomes available so it makes sense for them to adopt this as well. So that's kind of our position and you know we're pitching this uh as as a tech stack. Um and then we can have frontier models that are kind of verifiable and and on the same level.

84. 37:59

    I I think I'd take a slightly different uh outlook on it. I don't know that the proprietary closed source labs will ever be private. Um but the open-source models are catching up and like if you look at the trajectory over the last couple years the gap between the proprietary labs and the open source labs is closing. The question is does that trend continue such that the open source models actually become the frontier models? That will be extremely interesting. And of course if the open

85. 38:33

    source models do become the frontier then they can all be run privately and and we're good to go. Um but that's an open question. You know, will the gap just compress to nothing or will the open source actually surpass the uh the proprietary models?

86. 38:47

    Yeah, definitely following that race very very closely. Um maybe one or two last questions to to close us out. Um and we touched on it earlier like do you do and do you fear any government backlash for providing users with private AI? I mean fierce but yes yes uh yeah um governments will not permit private AI as it gets bigger so the question is is like which governments will strike first which ones will try to outlaw it first um the European governments will definitely

87. 39:30

    outlaw before the US the US uh in this regard will probably be the the last lingering like place where private AI can exist because thankfully the US retains very strong speech protections. But I do not know that that will withstand the um the game theory incentive of of trying to you know make all such conversation illegal. So yeah, I am worried about it.

88. 40:02

    Right. That's why we got to keep fighting that fight, right? Um, last question to close us out. I think maybe this this is a you know people always talk about like P Doom. What is uh considering the latest developments with the clause and the trajectory we're heading at? What is your P Doom Ilia and and Eric, if any.

89. 40:26

    I I don't try to calculate things I can't calculate and AI is not the thing that I'm most most fearful of. So my my PGM isn't based on on AI models. It's based on you know like uh the state that actually kills hundreds of millions of people. Um that would be that would be the entity that I worry more about. But no way to calculate a probability on it.

90. 40:53

    Yeah, I I I kind of agree. I think as I said it right now it's not AI that's a problem. people who are using all the tools which indeed I mean AI now is in toolbox right and then um and I think yeah my my I mean I'm optimistic and and and and believe in in what we're doing that we can actually substantially affect the the impact of AI and and kind of manipulation by giving everyone user on AI. I think that's that's really at the core of what we're doing is like same

91. 41:30

    same as you know blockchain gives I mean I'm I'm from XUSSR I've I've seen multiple currency collapses uh and so like I understand that for money and and I understand that for AI I think like we need to continue building that kind of as a solution uh that's fundamental and can survive through hopefully can can both survive through hopefully cataclysms that don't happen but also can actually be the release valve of some of this where you know

92. 42:04

    there's no point of stopping something if there is an alternative that is available to everyone right there's no way to you know like I mean we've seen this with internet and like Arabspring right they've tried to block internet but the mesh networks popped up the kind of the peer-to-p peer communication channels popped up and kind of the like the re reorganized around those tools. I

93. 42:28

    mean not we but the people in those regions I think similar thing here by having this release wall it kind of like I mean I I don't want to use this example because I don't think it's valid but like h having people having access to the to the tools on the same level always democratizes the kind of removes some of the uh pain points that I mean Eric is talking about.

94. 42:56

    It's a great note to end on. Thank you. Thank you both, Ilia and Eric. Um, thank you everyone who joined us. Be sure to make sure to go and try out our end to end encrypted models and TE models powered by Near on Venice. Fairable privacy is here. Um, so thank you.

95. 43:16

    Thanks everyone. We'll uh we'll catch you on the next one. Peace. See you.