Who Owns Intelligence and Value in the Agentic Economy?

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

1. 0:00

   media company called the rollup. We'll be moderating the panels today uh and all of the talks. We have our first panel today covering AI and the agentic economy. We have our first three speakers ready. So, we're a little bit behind. Without further ado, come on up, ladies and gentlemen. Take a seat, enjoy, grab a a coffee, a water, and uh we're good to go. So, panelists number one, we are ready for you.

2. 0:30

   and we'll get the show on the road. Who owns intelligence in the agentic economy? We've got Alex Ruiz from Microsoft, Cameron Dennis of Near AI, and Raa Share Ascari from Grayscale. Guys, please come up. Oh, sorry. We've got Vidor Jen Selad.

3. 0:54

   Yes. Uh guys, welcome to the panel. who owns intelligence in the agentic economy. Um, quick introductions and your favorite part about this this trip so far. I'm Vider. I'm the co-founder of Sulfur. We built SLA's first wallet. Uh, we saw a massive evolution in wallets from being like tech maxis to user maxis, which I'm today now. The best part of this conference is near Miami.

4. 1:22

   All right. Love that. Um, so my name is Cameron Dennis. uh I've been a part of near for about five years or so currently running a lot of the initiatives at near a primarily focused on growth and compute. So if anyone is interested in confidential compute uh we'll talk more about this today and uh the thing that excites me most about uh this conference so far is just the interest in agentic commerce like this seems to be the thing that a lot of

5. 1:45

   protocols outside of here as well are interested in and the fact that we were so early is is helpful. Uh first of all, thank you for the invitation to be here. I'm Alex Ruiz and I'm one of the um senior cloud architects at Microsoft that works with our strategic healthcare customers. So I've been working the last seven years in helping regulated industries uh develop agentic platforms. So really

6. 2:08

   focused on the confidential computing and you know data processing and keeping everything compliant. So looking forward to being here. Awesome guys. Uh kind of opening question here. There are kind of two agentic economies kind of being created.

7. 2:21

   one where the core platforms are controlling a lot of the intelligence, a lot of the models, the inference and a lot of the value and kind of a new cohort of one where the users will eventually be the ownerships and the distributors of this intelligence of this value. How do you guys see this current evolution uh of the control and ownership of the intelligence as it's been for the last 5 to 10 years and kind

8. 2:45

   of as we go into the next one to two years? How do you see the the control of intelligence in uh evolving? Cameron, you can start and we'll go. Yeah, for sure. So, um right now anthropic and open AI own a lot of the intelligence. Uh it is possible for you to own your intelligence. Uh I mean hopefully everyone here does own their own intelligence, but uh yeah, this is a huge concern because they're the ones that are pushing the largest models and

9. 3:10

   they can easily cut you out. I was actually blocked from OpenAI right when it came out uh because I was looking up uh nuclear explosions and like the impacts of them. And I was uh very quickly then pushed to use anthropic and that actually like I was one of the first anthropic users and uh it became very clear to me how easy it is for these companies just to like revoke your access and if those companies are the ones that are powering frankly the future of knowledge work uh that's a huge concern. So, we need to make sure it's userowned.

10. 3:43

    I think it's going to go platform. I think you've already seen that. Um, the the prevalence that that I've seen customers that I work with just gravitate towards, you know, we just just give us an endpoint or give us an interface and we're not really concerned about the rest. Um, I think platform consolidation is something that we've seen over the last 3 four years and I don't really see that um going away.

11. 4:06

    There are cases where I think it's important um for users to kind of speak up about the context of these agents that they're building, but at the end of the day, platform teams are just simply better positioned to own the intelligence and kind of secure um these these use cases.

12. 4:24

    Can you define platform teams real quick? Platform teams are the people that manage clouds, manage the cloud at like an organization. They're the people that have traditionally, you know, there's different there's data platform.

13. 4:36

    Looking from a wallet perspective, just they have that background. So, so looking from a wallet perspective, uh I think this is the first time we're seeing something different uh in terms of AI evolution. This is the first time we're seeing something different uh in terms of AI evolution. When is the first time we're seeing something different uh in terms of AI evolution? When wallets and crypto started, probably everyone remembers you need to run your own RPC nodes. You need

14. 5:05

    to run your local servers. You had your MetaMask wallet. You could plug in your RPC wallets there. And that was like very sort of user centric. The user had the ability to to change wherever. Now fast forward 10 years, what's an RPC?

15. 5:21

    Does it really matter? It doesn't matter at all. And I think in this AI context, what we're seeing is that everyone's everything started from platform because it's the ultimate user experience. So the underlying is it an LLM? Is it this harness that harness at the end of the day doesn't really matter if it solves your problem at the same time and cryptos wallet that sort of allows you to send and receive USDC. The

16. 5:47

    users probably don't really care if it's near Salana or whatever blockchain if they can do it. Alex, I'm curious from the standpoint of Microsoft kind of one of the core components of uh of the blockchain landscape is kind of this aspect of composability. This idea of being able to mix all different types of assets, products, etc. on chain. I'm curious kind of in in a similar lens from the

17. 6:10

    Microsoft perspective, what does composability mean in the agentic economy? like how does a company like Microsoft create a nonfragmented or a composable agentic agentic world? So to if I understand your question like how are we giving users the pieces to build what they want to build as opposed to just like writing it yourself or consuming it. I think you're seeing that middle ground across all of Microsoft's stack, but from what I've seen in practice, it's not where people are at.

18. 6:46

    People really are like, "Hey, just give me the platform and I'll write code and deploy it onto it." Or, "I don't even know what you mean when you say platform, so just give me a UI and consume it." But we we do offer those kind of middle grounds where it's like hey if you want memory that you can configure but you don't really have to know how it works like we've got that or like here's an endpoint for a model like the pieces are there and you you can't

19. 7:11

    take that approach and it's available it's just not what I'm seeing especi on the user side especially I mean I only work with healthcare customers and traditionally have researchers are the biggest consumers of wanting to build agents but they don't have the backgrounds to to build them their theirelves and they're not thinking about security. They're not thinking about guard rails. They're just like, I want to solve this problem and I think I can use Claude or Chad GPT to do it.

20. 7:40

    Give me an endpoint. Right? They're not they're they're not ready or they're not they haven't made that leap to um to build it themselves in any capacity even with the pieces or even from scratch. Um in in a similar vein Cameron from the kind of uh idea of building a compu a composable agentic economy whether offchain or onchain I think a lot of folks in in the web 3 space are very privacy focused I think near has has

21. 8:06

    built a very privacy focused brand how do you think about the rise of private inference as it pertains to the growing need for uh privacy with usage of AI um there's a couple cool teams in the space building this also outside how does how does near AI get uh you know establish a strong position here.

22. 8:24

    Yeah. So right now the people who care about private inference are primarily governments uh financial services u sometimes healthcare organizations but uh the way that near is going about it is we care about distribution and so we're going to first start with like the people who need privacy most uh and then second is build these partnerships like uh like brave if you guys are who here uses Brave anyone it's a one of the largest browsers but 110 million users

23. 8:49

    uh we are working with them to actually enable private inst private inference for all their LEO there's actually going to there is a chat on the right hand bar that can summarize your page for you. And we see the future of AGI as a one where there's this agent harness that really cares about security. So I want to kind of fragment these things between security and privacy. Security is around like hey it's not going to like leak your stuff and then if it does leak your stuff you have this extra guarantee of

24. 9:13

    privacy. And so we built this thing called ironclaw which is the security harness which can manage your credentials. And we actually have uh one of our close partners here today from Abound. His name is Nishkcom. And uh this ironclaw implementation can actually manage their banking credentials. Something that you probably wouldn't give your, you know, open AI or anthropic right away. And the cool thing here is it is actually owned by the

25. 9:37

    user. And so the user can, you know, completely control this themselves. It's not being leaked. And then it uses the private inference to action. So think of it like the harness is the body, the inference is the brain. And so the inference piece is uh one of the main focuses but we're doing this through large distribution channels like Brave and then the harness is going to be through wallets ideally.

26. 9:58

    Do you think that the Brave summary use case is a good one for private inference or do you think that inference will be done on a local device in the future? So I mean I would love for inference to be done locally if it is powerful as the best models in the cloud. So right now you are sort of stuck if you want to use the best models you are stuck using cloud implementations. So that's not going to change for a long time. Um and people obviously they always want the

27. 10:25

    most intelligent models. And so what we enable is that that data privacy that you get from local deployments but with the flexibility of the cloud. So you're able to actually like expand your you know token capacity or even usage compared to being just stuck whatever hardware you have.

28. 10:41

    And do you think that that that there's going to be a proliferation of local models so that this kind of userowned intelligence becomes like almost like an asset of itself, right? There's kind of uh conversation I think Larry Frink yesterday or the day before said that he anticipates there's going to be global uh compute futures markets basically really kind of growing. Do do you see a similar kind of asset class of it itself

29. 11:04

    being local intelligence that you can have on a local machine? I mean, I would love this like just just to make it clear like I would love everything to be local and verifiably private. I just don't think that the hardware the models are there now or anytime soon. So then what's the incentive to do that?

30. 11:20

    Like why would you like what's going to cause that to actually happen? It's going to take a long I don't I don't know especially when the alternative is you can get the same guarantees with performance from the cloud without having to manage your own stuff. Like everyone here doesn't want to manage your own, you know, server at home. You don't want to manage your own AI at home. You're going to want to have it managed in the cloud and maybe

31. 11:41

    through, you know, even a provider like Microsoft or someone else, but with the privacy guarantees of the trusted hardware that Near enables. And so that's sort of the the thing, but ideally make this all invisible for users. Like users shouldn't have to see this.

32. 11:55

    You know, one thing that I think all of us can relate to is the speed of advancement of the intelligence of each model implementation. Just Opus 4.5 to to 4.6 6 felt like a big jump. People have kind of mixed feelings on 4.7. Now ChachiT's models 5.4 5.5. They've gotten just progressively better. And it feels like what was what was once the most in intelligent model feels like a silly kind of outdated archaic even stupid

33. 12:21

    model now. And so you can kind of think that the cost of intelligence is dramatically decreasing. What does that mean for the agentic economy? How do you think about the accelerating decreasing cost of intelligence as it pertains to actually getting agents out in the wild uh and conducting actions on our behalf?

34. 12:42

    Great question because this is what I'm living on a day-to-day basis right now. I mean, look, there's a lot of people still out there using GPT40 because they put a lot of effort into um into the prompt engineering, into the evaluations, building their applications around that. And in theory, you know, that it should just be a matter of swapping out an endpoint for a newer skew essentially, right? Let's jump from 40 to 51 or whatever. Um, but then you've got to do all that testing again.

35. 13:11

    So, I think that the and I think this is a good guardrail in place from just jumping up the intelligence ladder is you've got to redo your prompt engineering, you've got to put in your evaluations again, you've got to red team the application again. Like that that is a good thing. I I think that yes, the level of intelligence is increasing faster and faster, but adoption is kind of like there's a speed limit that a lot of organizations are

36. 13:36

    way under or even just operating at that speed limit. And I think that's okay. I think that's a tempered approach in this area is good. That's for enterprise. For enterprise enterprise, I'm talking about enterprise. That's that's the perspective that users are just hopping from new model to new model. And yes, got it.

37. 13:53

    And the problem there is that there's like a I wouldn't call it a branding problem, but like who knows like what's 40, who is 5.6, whatever. Like users don't know and they don't really care. And I think uh consequences of this is that they just there's this like meta like just use the the strongest model.

38. 14:14

    Burn the most tokens assuming that you'll do the job better than some others. Okay, maybe. that you'll end up massively overpaying. They care about money though. And so the the thing is with like these open source models, they're so much cheaper. Like Opus uh Opus 4.6 I believe it's like $25 per million output tokens. A similar model that's open source and 4.6 there's not a it doesn't hit quite hit the benchmark, but I think Deep Seek V4 is I think it's like $5. Like it's literally like five times cheaper. And so if

39. 14:47

    you're trying to run like huge workloads and especially in emerging economies and so right now like I just came back from I mean it's definitely not an emerging economy but like Korea the government's very price sensitive like they don't want to you know release this thing to every single government employee and have them rack up these crazy bills and so they would rather host these open source models either themselves or work with a provider like us to save money.

40. 15:10

    like these things are going to print money for these labs and open source models are much more financially you know intelligent option and and also kind of you know you get all the properties of what open source provides right I mean it this kind of the like the whole idea of AI is a kind of like this distributed collective intelligence of you know of a machine and open source provides that for you

41. 15:34

    know uh actually you know creating and building uh tools right and so that you when it comes to where you guys think the broader uh biggest breakthroughs for uh some of the implementations of AI in the or of the agentic economy. Is there a specific sector? Is there a specific type uh adapted by the agentic you'll be some of the first to be uh adapted by the agentic economy? How do you think about this?

42. 16:04

    Yeah. So, um I'm talking to CEOs almost every day like how are you using AI? What are like how are you actually going to, you know, not have to hire 10x more people to get 10x more returns and like grow 10x? And I actually think when an agentic executive coach and assistant allinone essentially being an agent well executive and assistant allin one and money and so but what do they care about out across their org because they're saving so much time and money and so but what do they care about? They want to

43. 16:34

    make their org because they're saving so much time and money. And so, but what do they care about? They want to make sure all their, you know, sensitive information is not being leaked. They want to make sure that like, you know, HR can't see everything that's happening in legal and everything in marketing can't be seen what's happening in sales or whatever, whatever. And so, um, I

44. 16:50

    think it's just going to start with leaders adopting this and scaling it across the or we're seeing it already. I think one of the challenges right now is the personnel and talent um challenge. I think that there is something to be said for the amount of effort necessary to retrain or upskill technical people across different industries to be able to build with these systems from a build perspective.

45. 17:14

    There needs to be a certain level of organizational maturity achieved before you're in a position to take advantage of these emerging technologies. However, one thing that AI has been good at from the very beginning and even before we started throwing around the term AI for everything is um recording transcription and so a lot of organizations are immediately looking at call center at any any sort of technical support or call center support that they offer to consumers.

46. 17:43

    That's a prime opportunity as like the first opportunity to take to take advantage of this technology. You get your feet wet with that. You start upskilling your people. Now we can start to tackle some other priorities um that are not as like um cutting edge but they're where the impact is.

47. 18:02

    I mean it's tough to say uh like when you think about like overall commerce I think the world has reached a specific equilibrium of how much money is being spent on what and this we cannot expect this to sort of like explode. there's like money circulation cannot accelerate so quickly. So I'm a bit sort of skeptical and afraid that we're gonna all be inclined to invent new things that we can charge for which then stacks up the bill which then makes stuff more

48. 18:33

    more expensive. Like imagine needing to sort of pay every time you visit a web page like X4 to consume an API pay pay pay. So it's going to stack up and I think sustainability then is like becoming a real problem. One of the beautiful things about uh the innovation that OpenClaw brought ironclaw as well is this idea of this prescriptive models, right? Rather than being um you know you you're requiring a lot of prompting to get to an end state that they actually almost prompt you and

49. 19:03

    force you to make choices rather than you having to always tell it what to do. It's like it it was a big breakthrough, right? So also, you know, with that in mind, you you have these like promptless models really, you know, starting to get a lot of traction. I'm curious how you guys think about that that shift in uh LLM behavior as it relates to actually having these things work for doing a day-to-day payment or doing an actual hey go buy me a flight or hey, how

50. 19:30

    important is that like prescriptive ability to kind of have that second order thinking in these models? So we this is part of the reason why near intense is called near intense is because AI is going to move away from prompts you having to write something to it knowing what you want and then it actioning on your behalf. So you have an intent. It's my mom's birthday. It's Mother's Day. Got to go buy her flowers.

51. 19:52

    Like that AI is going to automatically go route that intent to you know find the flower shop, find the delivery, all this other stuff. And that is the next step. And so like AGI is something that everyone's been sort of working towards, but people actually like people who know what's going on say it already exists. It's just not adopted to everyone yet.

52. 20:13

    And that user experience isn't quite there yet because it's so complicated. So when you say these like, you know, predefined, you know, structures, we call these agent harnesses. And this is what AGI is, I think, going to to this is the the vehicle that's going to be delivered by.

53. 20:28

    I think the you would use the word harness. I I like to lean on traditional software terminology like an operating system. I think that's the big piece missing. Like I think a lot of pe I think there's pieces in place for just talking to these agents or to these LLMs in a way that's very natural. I mean look look at how everybody is just so accustomed to chat interfaces that just won out. Um but what's missing is the is

54. 20:54

    the operating system for it. There's also a lot of privacy concerns. I think this goes back to what uh Cameron alluded to with the, you know, personal computing like this living in on your device. I'd feel more as a user, I would feel more comfortable about making payments or having an agent make payments on behalf of me if I know that like this is kind of all on my device.

55. 21:15

    But in the state that it is today without that OS offering some level of security, I I don't know how realistic that is today. Any closing thoughts? Yeah, I think related to this, we also before the panel mentioned about okay, it's it's a mess when you need to sort of maintain MD files to provide context and the what you refer to as this like onboarding experience which became much better and

56. 21:40

    if you're really going to commit to like using a specific agent or agentize part of uh some workflow or for example your your your mom's birthday example, there's not just like the intent that like something needs to happen. There's also a question of style and like the style of execution is sometimes also important like how can the agent know what flowers your mom likes like you need to spend some time with the agent

57. 22:02

    maybe it listens to you over a month or two or a year or your whole lifetime so it prompts itself correctly with and it actually like reads your mind. Ladies and gentlemen, give it up for your panelists Alex Cameron Budor. Thanks guys.