Private, User-Owned Intelligence | Cameron Dennis

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

1. 0:00

   Cameron, we've got we've got you on the stage now, man. I mean, we've been talking about this a lot, but it feels super relevant. We'll get into everything that's going on with staking near and all that good stuff, but welcome to the show first off, and what are uh good to have you back, man.

2. 0:13

   It's good to have you back. And uh we're curious how you're feeling about this uh this potential invasion of privacy by OpenAI. Uh it's probably not new. Uh Open AAI has leaked stuff before. Um, I think the biggest question that needs to be asked is are they creating synthetic data sets on data that they're not training on, but they're creating synthetic data from and then training on that. Um, I don't know. I don't I just have no idea. Um, but that is something that I don't want

3. 0:43

   to have to guess. I would like But why is that important, Cameron? Sorry. Yeah. Talk talk us through like what those synthetic data sets are. What do they mean? Oh, yeah. So like they in their privacy policy in terms of service they say like they don't train on your data. Um but they could take said data and sort of create a synthetic version of that data and then train on that. Um, and so the I mean I don't know if this is 100%

4. 1:07

   relevant for uh you know this whole issue with this problem but because like the pre the pre-training was done quite a while ago for this model and like there's a bunch of back and forth of whether like the date of uh the pre-training was before or after uh when they started working on it. But the whole point is like they might not be training on your exact data but they might be sort of cloning your data.

5. 1:30

   Yeah. But I don't have the answer to this. This is like and I and I just have to trust what they say. I want to be able to verify it like cryptographically. Um and so this is where all this like TE sort of verifiable privacy comes into play.

6. 1:45

   Yeah, man. I mean, look, we saw that tweet go out with the highlighted text in yellow about those about this specific kind of incident and then instantly somebody picked up that Venice and Near are the best way to express the bullish thesis on private inference and both tokens are up like 30% or something from the from that post. And so I guess, you know, not to put on an investor hat,

7. 2:11

   we'll stay on the private inference side and the builder side and the tech side, but you know, do you think that the world is starting to wake up to the value of private inference and that there are private inference providers out there that are established? Um, you know, these tokens are catching a bit.

8. 2:27

   Obviously, crypto's had a bit of a turnaround, but like where do you think we are in the awareness curve of private inference? Uh, very early. Uh, still very early. I'm still educating like CISOs uh like chief security officers on what verifiable privacy and AI looks like.

9. 2:44

   And so we are still very very early in this. I mean confidential compute has been around for quite a while. Um and even at the last confidential compute summit it's started to feel like there's a you know different energy in the air from previous ones. Um, but it's still a bit early, candidly, and unfortunately humans, you know, we have to uh h there has to be a car crash before we put the stop sign where it needs to be.

10. 3:07

    Um, and so this is kind of what's probably going to have to happen in AI for people to care about this. But the other key thing is just like open source is getting so much better. And so like Deepseek V uh 4.1 came out today, Flash, and it's doing as well as Astra. And this is at at literally like like 95 or so percent cheaper. And so it's not just about confidential compute. It's about open source being able to perform as well at a fraction of the cost and have it be private. And so

11. 3:39

    the goal is make privacy make sure there's no trade-off or there's even a cheaper option or you know as performance uh than frontier models. Um Cameron, we're going to get back into staking for inference and confidential compute and all these things, but the community members in the chat here are crazy. First of all, they're they're asking you if you enjoy Mircats and they say you would make a

12. 4:06

    fantastic Mircat mob leader. Do you do you are you aware of Mircats and how do you feel about them? Uh so I don't know much about Mircats to be totally honest. Uh, and I'm sorry to the community, but um, you know, near cats were actually like the first sort of like u like internal meme of near.

13. 4:24

    Uh, and this is, yeah, we were all near cats uh, way back when. Um, so I guess this is a play on that and I don't really want to be a cult leader to be totally honest, but I really appreciate the, you know, all right Cameron, let's get into it.

14. 4:39

    Um, so staking Neo for private inference, right? And thinking about the scale of the TAM here, obviously Near AI is powering, you know, the the Venice API and can power pretty much any front end that wants to tap into verifiable inference. We just launched our rollup frontend for near.com. So, congrats to us. Um, and we're we're like, all right, what is the rollup LLM look like in the

15. 5:03

    future? Right? like is there a is there a $20 a month rollup website LLM offering that comes with like some better perks that is powered by near AI people can stake near to get access to you know this or they can purchase it on a subscription plan so we're like iterating a lot of these ideas right and we're like we're at v001 of this like what we're calling like the markets arm and so from my perspective

16. 5:28

    if I'm going out there and I'm like hey you know what guys use our LLM use our rollup AI because it's powered by near. You know, we use the Venice API. It's uncensorable. It's private by design and the inference is powered by near. You can stake near. All of my ducks are in a row here. This is a really good offer.

17. 5:47

    Um, you know, what do I have to trade off here? What like what is still missing from the private inference that you guys are doing compared to the Frontier Lab inference? Like, let's just let's just get it out on on the floor here. I would say it's like about the types of models. I mean there are really good like let's say voice models. I don't know how you guys use AI every day but like text to voice and like voice to

18. 6:10

    text is something that I think is growing exponentially. Um I have you know people in the office are like talking to their area AI sometimes even more than typing. And so you know having open source versions of these models come out that are competitive with uh the closed sourced ones I think is like a big step forward like thing that needs to be done there. Um, but honestly like we're at this precipice that there's not going to be a trade-off. Um,

19. 6:35

    I would say the other piece is like uh easy like model routing because in open source land it's very confusing to constantly have to stay up to date with like the best model for the best task. And so how can you have like a classifier that uh essentially routes a prompt to the best model according to the task? And then it's up to you guys to sort of make sure that those models are readily available for your users.

20. 6:59

    You have to abstract all that away because no one wants to hear, you know, GLM 5.3 flash versus GLM 5.4 and Y is better. Like people don't give a And so, um, you need to abstract that away. And I think that's probably like one of the missing pieces.

21. 7:15

    But the quality though just to Yeah. So like the so the quality of the model is can be the same right it's just like as far as the the level of intelligence that I'm able to get on near AI with private inference it's the same level of intelligence correct yeah it is I would just say it's like vibe like you need to provide the vibe for your users like I don't know if you've interact with a lot of these like you know Chinese open source models um but it doesn't sound the way you're sort of used to navigating chatgpt.

22. 7:51

    Um, and sometimes like the answers are much longer or like if you ask it to generate you a UI, you know, maybe it's a little bit more clunkier, the edges aren't rounded. Like that's all work that absolutely can be fixed and is actually not that hard of a problem to solve. Um, but that's the work that needs to be done. But in terms of like intelligence and reason capability, it's as competitive.

23. 8:14

    Gotcha. Okay. Okay. So, as far as staking near goes, could you please explain some of the mechanics there and maybe just start with the thought process around why this makes sense right now? Yeah, for sure. So, in short, uh we want people to be able to continue holding their near. So, you don't need to sell your near in order to get free inference. And the way that this works is you stake it to a validator that Near

24. 8:39

    AI manages. And that money that is generated from validation is essentially converted and just given to near. Well, not converted to cash. It's just paid to the near AAI team in exchange for inference. We're essentially just accepting near as payment, which makes sense because we're near AI. And so, um, the whole point there is now you don't need to necessarily spend any additional

25. 9:03

    money to get your AI paid for. You can just hold your asset, have that money that's generated through staking, and have that pay for near or pay for inference. It's it's pretty clear. So, no card, no account. You just stake. It's that easy.

26. 9:19

    Yeah. I mean, you have you have to have a near account, but yes, no other accounts. Yeah. And can you talk through the economic implications for near because I I think a lot of people, you know, they this will resonate with them as a product and it'll onboard a lot of people. But what does this mean for [snorts] the for near economics? What do you expect to take place here?

27. 9:44

    So there's a couple parts to this. One, it's a really cool new utility for the token that previously didn't exist. And so there's an entire new, you know, economy that opens up around, you know, having near being able to pay for intelligence. And so this sort of falls into the long-term goal of like making near token AI money. And so there's a very important piece there. Um, now for

28. 10:10

    the mechanics of how like that the value of those that inference gets calculated is based upon the inflation rate and the amount of tokens staked on the protocol. And so the the the caveat here is that and there's a little bit of a competing interest because there's a huge interest to have the inflation rate of the protocol go down. We don't want to have to overpay for protocol security and so there's a interest in making sure and

29. 10:38

    there was a huge inflation cut. So Near's almost like not quite deflationary but the goal is to make the protocol and the token deflationary. Um, now that's competing with the stake for inference piece because for stake for inference, you want the inflation rate to be higher because you're going to want for more tokens to then go pay for your inference. There's two things that can solve this. One, the price of goes

30. 11:01

    up. The price of goes up, you get more credits to pay for inference. Everyone's happy. The other thing is the cost of AI tokens goes down. We get better hardware. We get more efficient models. you know that part is also happening and so this to me is incredibly exciting. We have all these new like TPUs and types of like AS6 that are you know making inference a lot more effective. Um I expect the price of

31. 11:28

    tokens to continuously go down as well as more efficient models. And so um put these two things together I think this is going to be sort of the foundation for sustainable open source AI economy. Um and this is just a step like in that direction. And so can you just kind of talk to us about the TAM here? Like what would you consider how much near staked are you looking to acquire here and you know AI is just continuously growing. So

32. 11:56

    this is you know this is practically infinite as far as inference demand goes and also you know as Venice grows the demand for private inference directly correlates to near AI's demand. So just kind of walk us through like what would make this a success for near AI? How much near staked? How much inference demand? Give us a bit of a sense of some of the metrics that are like you know your your northstar.

33. 12:26

    The amount of nearstakes to this validator is important because this is going to determine the amount of revenue in near tokens that near ai receives. And so definitely like the amount of nearstake. Um but I also want to flag a lot of this right now given the sort of lower price of near as well as uh the low inflation rate. Um you this product is kind of designed for whales right now

34. 12:50

    and I want to be kind of clear about that because you do need to stake a lot of near to get access to a reasonable amount of inference to power your AI that's let's say always on like your agent that is scraping the web and doing things for you. um that sort of like always on workload would probably be around 60,000 near so about a hundred 20ish,000 you know $25,000 and so but the cool thing here and this is what people like need to wrap their head around. They're not paying $125,000.

35. 13:24

    They're holding that asset. They're just foregoing the reward from staking and instead receiving inference. And the cool thing here, and again this is like not tax advice, but and talk to your tax professionals about this, but if you're a giant whale and you have a millionaire, you can receive inference credits instead of receive near tokens which are taxed and so right that something that and again it varies per jurisdiction and I know you have viewers all around the world and so I don't want to like flag anything. But

36. 14:03

    like that is an interesting use case is like instead of getting tokens back, get some sort of like utility that is going to actually be spent to run your everyday life like AI tokens, right? And specifically non-taxable utility depending where you are.

37. 14:22

    Depending on where you are. For for sure. For sure. I I said I said that. Cameron didn't say that. Lawyers stay where you are. Um Cameron, I w to I want to dive a little bit deeper into the inflation that you talked about because I think we understand how, you know, right now the near token is fully diluted, but the protocol still has some inflation mech, you know, me mechanics.

38. 14:48

    There was a comment in the chat actually just now says, "You're not ready for deflationary near." I think people are kind of looking at that. They're not really sure whether it's going to happen, but they're they kind of see it looming. And then at the same time, we've got like these these near campaigns like the 333 campaigns and and some others going on.

39. 15:11

    Can you just talk us through like what is the current inflation dynamic on the near protocol and how does it work? What are the inputs and is a deflationary near really something that is in the realm of possibilities? Yeah. So uh just to kind of start from square one like the inflation i.e. new tokens coming to the economy um are normally are spent to pay validators to maintain the security of the network.

40. 15:40

    And so that is very important. But there's also an argument to be made that you know there's a cap of how much money you should be paying for security. And if you're overpaying for security on something that is pretty damn secure, never been hacked, never gone down, all this other stuff, you don't need to overpay for it. And so the goal is to make sure to pay for security at the right rate. And so that is like one of the main drivers of like the governance posts to drive down the inflation rate.

41. 16:14

    So that means the validators won't get paid as much. And so um that is a huge argument and debate that's going on that candidly I just want a sustainable new economy. I don't want to have a giant horse in this race. Um but that is like one piece and so um there was a proposal done last year to cut the inflation in half similar to like you know bitcoin havingss and so uh that did that passed

42. 16:44

    um I then there's another layer of all this which is like near intense near AI some of these like near foundation affiliated projects um will take the profits that they that they make and use them to either buy back near do something with it, stake it. I don't I don't have the the the clear, you know, this is what's happening. Um but that is to take more tokens off of the market and so you know decrease supply increase uh yeah decrease supply sort of

43. 17:13

    increases the price and so uh that part is uh pretty much dependent on candidly my ability my team's ability to make near AI money and that is done by you know selling inference tokens uh to highly regulated industries you people who need private AI and as Andy flagged there's almost an unlimited demand for AI tokens. Um, intelligence is in the AI term being tokenized. And so, um, there's an unlimited amount of intelligence needed

44. 17:45

    for solving the world's unsolved problems. So, that's the gist. Yeah. And I I I think you paint a more clear picture than I've heard others articulate. And it's a difficult paint picture to paint because we're just so, you know, fixated or consumed or just bought into the fact that, you know, we have US dollars and that's what we spend to buy goods. Like frankly, it's the world's reserve currency. You know, it it's psychologically

45. 18:13

    like we're psychologically monopolized to thinking that the US dollar is the only form of money. Sometimes, you know, if you think way back a hundred thousand years ago, you know, people were using gold and whatnot. Besides the point, future looking is more interesting when you start to think about how we could denominate some goods in the form of other forms of currency particularly

46. 18:37

    like you're saying making near the currency through which we denominate intelligence where you know if you're buying credits or AI compute tokens input tokens output tokens near AI you're going to be spending near and that is just the base form of money that we use to denominate intelligence. And I've heard other people try to make this case, you know, around Ethereum and DeFi

47. 19:01

    and that sort of thing. And it it oftent times does fall flat because people just can't really wrap their head around using a different form of money to denominate the goods and services that they're buying. And so as you're talking to chief security officers, you know, obviously the first step is to get them to buy into confidential AI, private AI, near AI. Once they understand the protocol,

48. 19:26

    how easy or difficult is it for you to overcome the next psychological hurdle? The idea that they're not going to be spending dollars on near AI. They're going to be spending near on near AI. Do you abstract that part away from them?

49. 19:39

    Do you say, "Hey, just give us dollars. We'll do the swapping in the background." Or how do you explain, you know, that component, the money component to these chief security officers? I don't talk about the near token at all. um they need to be I need to meet them where they are. Um they also don't like like some younglook guy telling them something that they don't know. uh you know there's a lot of ego with a lot of these people and you know I need to really match their energy and so um I

50. 20:08

    say hey pay in dollars pay in you know I'll invoice you at the end of the month like just traditional receipt based pricing whatever they're used to and um yeah we'll take their money and use that to buy near it's it's like super simple and then that's also where like intense is quite interesting um you throw this like giant you know liquidity network, being able to swap anything for anything. Um, it's that part can be like make everything seamless. And so there's like a lot of really cool things that we can talk

51. 20:39

    about um when it comes to you know uh AI commerce and how do you price intelligence and like also not just intelligence but the data. So like one of the most important things right now is real-time data for these labs to trade on. So like OpenAI is paying a ton of money and uh to get like real- time uh like business data or you know Google made a giant uh like investment into weirdly I think Spirit Airlines after they went bankrupt for all of their business data. Why? It's because they take it to train their next model

52. 21:12

    because all these labs are literally trying to become the last company on Earth. They're trying to replicate everything. And so what do they need to replicate everything? Real time data. And that's where everyone's prompts and how their agents interact with tools when things fail, all of that matters a lot. And I'd consider the data piece a very important pillar of AI money. Um, and so there's a whole side quest that

53. 21:37

    I'm happy to unpack at some other point uh talking about this. Yeah. Yeah, I mean Cameron I guess the other obstacle that I'm curious about outside of this kind of monetary you know I which seems like it's solved you know is this you know when I go on Venice I find myself before different types of queries I'm like all right I'm in an incognito window I've got the privac I've got the model set to the GLM

54. 22:04

    private model and then I'm like all right first can you please pull some context from any of my previous conversations just kind test it, you know, and he's like, "Nope, I can't find any of your previous contacts because you cleared the storage, cleared the cache, and cleared all the browser history and went into your Mac like files and deleted the file that was on your local browser and in your

55. 22:27

    computer." So, I'm like, "Okay, cool." And then I'm like, "Okay, is this private?" Like, "Yes, it's fully private. Here's why." I'm like, "Am I anonymous?" "Yes, here's why." "Is there any data that's going to be shared?" It's like, "Okay, no." All right, I'm trying to make a a bomb. No, I'm just kidding. And Right. And so it's like, how do you get that mental leap solved? Like, how do you solve that problem? Because if I'm doing this as

56. 22:53

    somebody who knows all the intricacies of what you guys are doing with Tees, with Venice, I know Eric, we've talked to him in person, like I'm like I'm very in the know and I'm doing this. How do you solve that for someone who's not so in the know?

57. 23:06

    Yeah, we need outcome. We it's outcome driven like we're not going to convince people on mechanics. Uh we got to convince them on outcomes. Sure. And so the you know takeaway here and this is something that actually Mark Zuckerberg and others have been talking about on the Facebook side or meta side.

58. 23:21

    Um talk about CVMs these things called confidential virtual machines and this is essentially a black box that you're able to run things in. It could be AI models. It could be, you know, data like like databases. And what everyone eventually will have is their own sort of black box that lives in the cloud that you have access to. This is like the long-term vision uh that you have access to and you get access through something like a near account or a crypto wallet, something that you

59. 23:53

    verifiably own yourself. I think of like the near account as like the off like the authorization in order to access your blackbox in the cloud. And that blackbox has all your context that has all your previous files. It has all the you know your photos and understands you.

60. 24:10

    Um now Meta is likely going to say, "Hey, this is your blackbox. We're going to throw our little like advertising engine or something in here and we're going to like learn about it and like kind of see what you store and all this that the other. It's yours. It's private, but it's going to like push certain data to, you know, our servers to run you ads.

61. 24:28

    And so, this is where um we differ from, let's say, like some of the big labs, like we're like, we're not going to run any of that stuff in there. Um, you can have your own blackbox off into it with your near account that you verifiably own yourself and then you call the inference APIs to get access to your private intelligence. So that context layer we started building some things called like a private portable memory layer.

62. 24:52

    Um so you can actually like move let's say all your chat history from open AAI to this you know your personal server so it actually understands what you're doing and so you kind of carry your context with you from thing to thing. Um, and you can take your context that is in a place that is like, you know, I don't know, they're training their their models based on your data just because you unknowingly agreed to that or

63. 25:17

    whatever it is or, you know, this synthetic data clone. And you can take all that context, put it into this black box that has that is stored on your local machine that has all the properties of the privacy embedded into it. And then you can leverage that context again in another private environment. Um, and then after your session, you can still have that context be built up in that black box, clear all

64. 25:41

    your history, clear all your cash, clear all your session data, and then come back and unless somebody is like logged into your Venice or or your near AI account or can get access to that somehow, maybe you maybe you like password protect it, there's like a okay, would you like to use context in this conversation? Yes. Okay, tell me your PIN.

65. 26:01

    Right? And then so you can build a blackbox of your own data but have individual private sessions. That sounds really really really interesting. Like that sounds like a breakthrough. It is coolest part. Pass key. So like Near supports like SEC P. I think it's like SEC 2 or I forgot the exact like you know encryption algorithm but like you could just pass key into anything.

66. 26:23

    And so no more having to write down these 12word seed phrase you lose it this that and the other like it's all backed by pass key. um completely abstracts like the all the crypto complexities away and it's really about user ownership. It's really about you know the things that we've all like the things that got us into the space to begin with and so um yeah this is all possible. Um

67. 26:45

    it's not super seamless to build. Um it's like well the idea in theory sounds great like bring your context from all your like all your skilled agents that you've built out port them into your own black box right and then be able to have individual private sessions with context because that's the biggest downfall to Venice or to near AI right it's just like every if you really want this to be private

68. 27:09

    every new chat is like starting with a blank slate which is tough and it's that's it's that won't win in the market like candidly you're going to need that context. And so, how do you do that context in a flexible way? Something that I've always wanted, and this is like a pipe dream of mine.

69. 27:23

    Hopefully, someone listening can go build it, is I want an agent that has access to my email that's verifiably private, and then it scrapes everything that I've like look at and do and like anything I sign up for, navigates the terms of service of that website, and then goes makes a request to get that data from my site, like export that zip because you could do this at like, you know, Instagram or Spotify or whatever.

70. 27:45

    export that zip to my email and then that is the thing that manages my context. Like that is my context. But the hard part is like largecale context management like things like Obsidian and things like this is all really cool stuff like critically important. Um but when you throw like a sh a ton of data at you know this like small context window the AI essentially stops working well and so overwhelmed

71. 28:08

    it overwhelm it's overwhelmed. Yeah. And so how do you manage a ton of context um in a seamless way? That is the engineering problem to solve. And I'm confident someone will solve it. Uh we are not doing this quite yet. It's all comp. It's all compression, right? You got to compress a large amount of context into something smaller and analyze it.

72. 28:32

    Yeah. Like synthesize it or whatever. Yeah. Yeah. Well, Cameron, I would love to have you on as you make more progress on that specific vertical because I'm all in on that. I will be the first power user of that. I will test that thing until non-stop. I love I love the private uncensored LLM. It is my favorite favorite thing. Like I just love how near's inference uh powers Venice and just how it's able to just I can talk

73. 29:01

    about anything. Like they were all like, you know, they were always like, "Oh, AI is gonna be your doctor and your therapist and your companion." And I'm like, "The hell it is." Yeah. Sam Alman's gonna be my therapist. I'm not having that. No. You know what I mean?

74. 29:14

    Yeah. Like pushes you in a direction you wouldn't normally have gone. Like it's uh 100%. People are things change. Yep. Yeah. Well, cool, man. Thanks for your time, Cameron. Awesome work and stoked to see this um keep growing. And I'll have to look into that. Uh is that private virtual machine that you speak of or is it PVM or CVM? Is that is that like can you just buy that somewhere and use it?

75. 29:40

    Yeah, you can actually uh follow network has CVMs for sale like for everyday people. Uh you can go spin up your own. Uh Nearai does not sell CVMs as a service. Uh we actually just run like agents that we build in CVMs. There's also a lot of other amazing use cases for CVMs. Mark Zuckerberg is making this more popular right now. I think there was a clip going around yesterday him

76. 30:03

    talking about this magical thing called the CVM. Um, but this has been around for quite a while. And uh, the the one thing I want to kind of walk take like walk away with is like what we're really trying to do is create a sustainable open source AI economy.

77. 30:17

    Because like right now AI is not that sustainable. It's like powered by debt. Like debt is underlying like all these giant data center buildouts. there's all these like all these giant labs have all this like potential offbook debt that's like not even popping up in like SEC filings and stuff and so how do we make AI sustainable um it starts with things like stake for inference it starts with things like controlling like how inflation gets allocated it's very low level and technical but at the same time this is

78. 30:47

    the hard problem to solve like how do we make sure AI is accessible to everyone around the world no matter where you come from you know your economic status this that and the other Intelligence is the thing. I love it. Thank you, Cameron. Thank you for your time. Awesome. Thanks. Awesome session. Thank you so much. See you again soon. Thanks.