NEAR AI Office Hours #32 w PIN AI

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

1. 0:05

   hello everyone uh welcome to another installment of near AI office hours we have today Bill and David from pin AI uh do you guys want to introduce yourself and we can dive in what you guys are working on and also how it kind of uh in interesting ways intersect with what neari is working and in General Vision for the future of AI assistants what get started sure hi uh my name is David krais I'm one of the

2. 0:35

   co-founder of uh P um before this uh my background is I was doing uh core research at theum Foundation mainly focusing on like me fee markets and uh before that uh I was running some ml engineering teams at Lyft so we were designing the Lyft Marketplace platform my team actually started that and then it grew uh so everything at the intersection of like uh economic mechanism and machine

3. 1:06

   learning um this was like before Transformers but we were using comets for like uh the key uh modeling the key function of the marketplace and utility function of users yeah yeah yeah har I'm bill um yeah uh I'm one of the co-founder at pii and chief scientist and uh before P I have been doing AI startup since uh uh like early 2023 and before that uh I've been in one of the large H fund run a like 20% team uh and uh have like A1 billion dollar

4. 1:44

   like basically running ml to trade us equity and before that I was a Stanford AI PhD actually I'm a mass PhD that um like spend my whole like Mass research career in deep learning when I start 2014 I was looking to different research topic and feel like okay there's only two really interesting topic one is deep learning one is Bitcoin as a maass PhD I feel like deep learning seems like there is more mathematical things that I can truly try to understand although even

5. 2:17

   today I don't consider I understand the much of actually how de learning works but uh I spend the entire PhD career doing those research and in 2016 I I was actually at Google brain doing question answering and working very closely with um people like asish and I think uh El you were also at uh Google brain we didn't interact much I was mostly with the machine translation folks and uh uh I'm the uh one that is doing question answering and

6. 2:51

   I was one of the latest uh like Engineers because I come from math background I'm not that good at COD tensor flow which is super hard to use at that time but I basically decide okay let me copy the state of art machine translation architecture uh that's before the self attention architecture so I copied the B directional STM model and immediately just like get it working on question answering basically I'm the one that discover okay by not changing a few

7. 3:23

   lines of code just like modifying the data set you can get a machine translation code work perfectly as a sequence sequence mode model um question answering I think when I look backwards it's a pretty interesting discovery of like how like versatile sequence SE model is so when Ashish showed me the self attention architecture I'm the first one actually got it work on question answering so I also like

8. 3:46

   witnessed the part of History together with Elia with lot of the other like Transformer authors yeah and um fast forward after um Google brain I was like okay um proving any ser about real industry deploying system seem hopeless right like the speed of involvement of like all this architecture is so fast and the systems that I saw in Google brain is so complex so I'm looking much more into applications I yeah like did a

9. 4:19

   lot of different research on like how to use AI in trading how to use AI in like uh cryptocurrency designs and all that and that fast forward lead me to uh where we are today in this triangle of like AI finance and like blockchain protocols awesome yeah I mean obviously yeah an interesting Journey so uh maybe explain what Pine is trying to achieve and we can kind of dive in from there maybe I can take this so p is the

10. 4:55

    first personal AI Network so we are essentially building personal AI which means that uh it's an AI that is like very close to the user has access to like a broad context from uh user uh data um and our focus is on first usability we want to be like uh build an application that is like serves uh um user first but also on privacy so both the data uh and the model uh resides on uh the user device

11. 5:35

    right now and um uh all the processing we do is on RT network uh that we built uh in house which serves our application and then like looking a little bit into the future our vision is that essentially with all this um um crypto AI field kind of exploding right now but we can see a trajectory where a lot of agent services are going to be registered on blockchains are going to be paying on blockchains Etc we want the

12. 6:07

    personal AI of pin to be also a Gateway into this whole universe of services yeah let me add one more thing which is our long-term Vision on the AI side where we're using all different uh Technologies but truly what we want to land is we want to have the most personalized the AI both the AI that help you save time like your AI executive assistant um like different type of AI agents like apple intelligence all the AI that help you

13. 6:39

    kill time like you're like most uh uh kind of close to Heart companion your like um yeah favorite AI that is like the movie her right those AI we want to have both AI that um like actually work as the way that we see in science fiction and we believe the biggest missing element is to have every users's versatile personal data all aggregated together uh from their shopping history

14. 7:09

    their chat history on different social platforms uh their trading history both in centralized exchanges and in all the wallets um like everything you can imagine right like your um video watching history and all that put all of them that together in a streaming fashion so that it's a source of um like your entire like Digital Life and uh so that all different AI developers and small companies can actually build different versions of AI agents AI

15. 7:42

    aventar uh applications that is um Gonna Know You much better than without using pi I have a couple of questions so you mentioned that the model and the data are on the user device but then you said that there's like a key Network so do you split the the work between the te Network and the local local machine yeah so the te Network right now like it basically does uh two things data fetching and processing um so what the user can do uh

16. 8:16

    is we have data connectors uh for all the main like Web Two uh applications uh we built like a dozen of them and then this is going to be open source protocol where like we invite more people to uh so the te right now is focused on the data processing part not on the agent part so essentially like fetches the data and then processes and it extracts like user preferences so kind of compresses the data and then the data is

17. 8:45

    then sent back to the user device and it's only saved there as context for the personal AI so so in this case the T machines they not like they don't have like H2 200s or whatnot it's just a uh CP not a requirement to have gpus no yeah eventually when there are enough h00 with confidential mode which is the second batch of h00 and then ever since I think all the Nvidia GPU have the confidential mode uh say then I think

18. 9:17

    slowly there will be more of those te that have strong computer ability and as a user so like the the I assume ke only makes sense if as a user I can you know get quote uh you know have the image of the virtual machine so how how does that work is there like a repo with the VM available or how how how do I as a user know that uh my you know the processing happens inside of the

19. 9:43

    key yeah the data connector protocol and uh the processing it's all open source uh yeah I mean it's not yet but it's going to be in a few weeks when we like launch the test that yeah yeah yeah yeah that's cool we we've been diving quite a bit into kees right and and like ensuring this protocol where I as a user can actually fully confirm that something is running inside of the KE is not as trivial right like I guess it's it's trivial to get the quote right but the quote is a testing only to that much

20. 10:16

    and uh having the whole protocol where I can confirm that it's actually the VM that there's no man in the middle it's not that trival I guess and educating users also is a is going to be an interesting it's going to be an interesting challenge right because like I think it's an interesting question right because you're running on device so I would be interested to know how good

21. 10:39

    over a quality you manage to get on Device and what models you're running right but I think like the way I think about it more and more models will be running in tees I users need to be educated on how to confirm that they actually interacting with the te and not someone who's uh you know like the virtual machine inside of the te doesn't have back doors doesn't have a station aable and what not right yeah yeah for sure yeah so I think two points uh the

22. 11:04

    first one is that our vision is the same actually like we also believe that like most agents and agent services are not going to be on the local device right there is too much computational constraints uh there to like fit all of them um that's why like um the personal AI for us I use this term before Gateway right uh the personal can do some things for you there will be some Integrations we want people to build like some

23. 11:33

    integration to maybe like control some apps from your Ro device but the main uh use um for that is that it has access to all your context and can act can act as like personal broker as well to like uh services that are in the cloud or in tees that are registered in like maybe the near blockchain or like these other registries that are popping up in different blockchains um yeah so this is the first point I think our view is aign there and then the second Point around

24. 12:07

    like um basically security of uh uh these agents I think it's very important right like I don't think that te is the only solution right like uh to me there is like kind of like three levels there is like first is reputation for like low stakes uh actions then it's like economics like uh staking and slushing or like uh putting some Fund in escrow uh like for agent you don't need to verify all the code like for things that are not high stakes and then like for the high stakes

25. 12:44

    thing you want like cryptographic security and I think that's where like te comes uh in I guess I guess at some point we need to move past them once like uh the cryptography is more production ready but uh yeah and so so today what what model do you run locally is it like llama or uh yeah we're actually trying a few different models some is uh like our in-house fine T the uh like tiny Lama

26. 13:18

    and also in-house fine T the 10 the Alibaba model and we uh basically trying to figure out which small model uh has a to run on the phone but also have the capability to do strong UND device like rag since lot of our task is about unstructured data summarization and structured information extraction and basically blind preference labeling without saying users data can we uh do uh like user preference labeling through

27. 13:50

    question answering style so the key is basically very strong Rag and summarization and uh uh yeah it requires some fine training and also it requires testing over different models so we are pretty um like kind of base model agnostic in terms of our pipeline designs and I guess completely orthogonal question right but uh I think something interesting to establish early is what like can you give like you know a couple of examples of something that someone would use p in the next year

28. 14:23

    right so so what of the most uh like in your view which uh scenarios in our lives today are are the most you know ripe for being solved with the you know personal assistant who has all the context and you know like constantly running on your device uh I think there are two types on High level in my mind one is save time the other is kill time in terms of save time it's more like uh your assistant your like uh executive assistant or your like even say like the boss of your

29. 14:59

    family office for someone successful like right you more familiar in fact there are people who runs your personal life like help you book trips helps you like uh right like arrange restaurant and help you like do day-to-day shopping so that it more like save your time and energy and uh protect you from the ads and the spamming and all that you just need to give it your high level instruction do everything uh even like with a family office it helps you like

30. 15:26

    say um say pray and manage your money in crypto and even in traditional stock market right allocate your wealth into different say like uh ETFs or like even in different meme coins if you are very up to it in terms of risk adverti right it's kind of like similar to shopping it's a expression of your value system yeah uh and the other type is The Killing Time killing time is more about companion it's more about Sol the loneliness and

31. 16:00

    really know you and really be there with you and provide the emotional value th those ones I think today open eye is not pushing there like my why Focus that open eye on more like the 0102 0304 lines of work it's more like pushing the frontier of human reasoning capability and uh pushing the frontier of like how good silicon intelligence can get to but there's another Direction which is how can we build AI that really ker towards every single human being and uh make

32. 16:31

    them feel like this is truly my like AI besty that knows me and the be there with me when I'm depressed be there with me when I'm happy and this is the one that lessen to me um regardless of like how much complaint I have and how much negative emotion I have I think those AI are pretty hard to build if you look at character Ai and stuff they were more about engaging towards a specific Target user group and if you look at some of the past trials like inflections piie

33. 17:02

    they were trying to build a good emotional AI but apparently the founder has a becoming Microsoft Chief scientist so I guess it's like half halfway down uh so it's actually very hard uh AI research problem and I believe the hardest part is to actually uh get a streaming uh refreshed version of users versatile data so that you truly uh understand the user and with the understanding then you also uh attract the top AI researchers to figure out how to make the AI R provide the emotional

34. 17:35

    value on the long actually yeah like initially when we started the conversation I was going to ask how it is how it compares to character but I see that it's a it's sort of in a very different uh very different space right it's it's less about having your actually that's that's a question I still have right so so given given what you just said right so so character the way I always understood character right is like hey now you have this body right

35. 18:07

    especially like if you're like an introverted person or what not you actually have someone to talk to right and like in itally when you were you were talking about like you know buying stocks my mental model of pin was slightly different right more of like helper right but you effectively saying that uh you know uh it's not yeah it is helping you to buy your stock but it's also helping you to to keep your mental

36. 18:29

    health in check right so so like how how do you do you see like P as the sort of also partially attacking the market of character or do you think it's uh completely different uh I think it's slightly different in the sense character is not trying to do personalization at least not as of today right like I'm pretty familiar with the space character uh talk chai they were all mostly trying to uh optimize for the engagement of their model so that you have even longer

37. 19:00

    rounds of conversation and I think a good fraction of their user is actually teenagers who are kind of L and are still in school and cannot really watch Tik Tok videos and just like chatting in classroom and I think there optimization is towards the distribution of that group uh to make the model R engaging to them and uh their uh characterization is not what I consider as personalization

38. 19:29

    right there is a uh thing about what who the AI talks like right and the other is who does the AI care about um in their cases more like their AI can talk very much like Elam mask or Taylor Swift but they do not specifically care about you and know your information um what we are trying to build is more like AI that really cares about you it can still be in the form of Taylor Swift because you are such a big

39. 19:58

    F Taylor Swift but it's like your personal Taylor Swift that R is like your best and knows you yeah I actually because I I I mean like I described user own AI as effectively AI that you know cares about your success not you know trying to make money of you which is you know most of these companies are right now but I'm curious how are you guys thinking about the you know cares about you how do you measure that how do you you know what a loss function in this case or reward

40. 20:32

    function yeah I think like make basically you can say that in like AI that does not make money of you it makes money for you maybe in the future right uh yeah so I think um the loss function kind of depends on the application itself right so I think a third category that bill didn't talk about is like this more like uh it's more futuristic slash like AI native use cases where like um there's going to be like real AI Services out there and they may like recruit

41. 21:13

    different AIS or they may want access to like your personal data asset like your personal IP Etc so the personal AI there can be like a key uh uh piece of uh infrastructure for you to connect to these new type of services this is like new applications that don't look like the ones we have today right they don't look like booking a trip uh or anything is like AI native non skew morphic apps uh so there um you can have like clear reward like in the

42. 21:48

    case of for example like participating in like uh uh investment Collective you can get the feedback from like uh the profit in your wallet I think that's that to me is like one clear loss function that like uh probably like if we're able to like uh financialized a lot of these interactions uh maybe you can use it not just for like one applications but for multiple ones um yeah I don't know more

43. 22:19

    in general maybe at some point we need to give people the ability to kind of program in uh uh their goals uh so we are not there yet um as with P but that could be also an option like in the future right that uh we give people the ability of specify how much they care about like different features yeah so currently as of today we're trying to understand okay uh how much your versatile personal data collection uh provides value about

44. 22:52

    letting your personal knows about you right so we actually designed this concept called God dog mode it's essentially like a like a exam exam that come up with uh like a paper test that is like yeah 100 question about you and uh it's like in question answer inform about yeah like uh yeah where is your favorite blah blah blah and yeah like um who did you talk to on Tinder last week and uh they response to them and stuff like this basically you can design uh

45. 23:30

    different styles of question targeting different subset of uh usage category and the key is to have the personal AI slowly learns about you and R it's more like if you are R important person you you are trying to uh like curate a long-term loyal servant that like really knows all your detail preference even better than yourself yeah and this part like we're doing research on it right now like we've had like some interesting result but like these type of evaluation Frameworks are like um like they have

46. 24:04

    some challenges to build but we feel that if we keep making progress to a point where like it's really robust this can be like uh integrated into our protocol right and people can um essentially have a better generic measurement of like what their personally capable that goes beyond just okay how many data assets like it has access to so Alex and I actually uh were building a predictive AI assistant back in Pre

47. 24:36

    not just blockchain but pretty near iay and so one of the things that we found which was extremely I mean for few people at least was extremely useful and it was an interesting signal is at the end of the day asking how people felt or how did they their day go that effed is like an NPS time score but it gives you a very uh like clear signal and then you can use that to kind of analyze you know all of the things that person had throughout the day and uh you know suggest maybe

48. 25:08

    things to like improve from there obviously back when we were building it we didn't have any any kind of level of sophistication of models to understand the context of personals day and data uh but uh yeah that was kind of one of the things we we looked at back back in the day because yeah we were actually trying to do some of the things back in 2017 um in yeah very different different

49. 25:36

    age yeah yeah at that time there it's completely impossible to just understand unstructured data right you have to create a very clean structure training data and then Dev set test set all that now it's mostly like a treated as a like a blind unstructured reader and then like you can ask any questions and hope it can do a decent job yeah back in the day you can still get some insights right like you can measure how how many hours the person slept or or or whether their geographical location coincided with any

50. 26:10

    bars right and then comp to their yeah so yeah I mean I think that there's a few different directions here we can dive in a little bit more technically right because indeed like we had near as well as Alec mentioned right looking a lot in tees including for AI inference because it does seem like a prerequisite for this kind of agentic uh VAB that you know we we all want to achieve uh and then the other

51. 26:46

    aspect of this is also like how will all of the agents work together coordinate and communicate as well as with your uh kind of personal assistant um so I know you guys also looking at kind of intent protocol uh to enable that so maybe uh you guys can start kind of how you look at it and we can you know contrast and compare yeah maybe I can start so we haven't done too much development work

52. 27:19

    on this right we have a research team that is like spending some Cycles on it um I feel that it's a key question as like the uh Agents come to life right so we have seen the first wave right we have seen the first wave where like if you remove all the noise uh what's left is essentially there is like these initial agents that have like some utility and it's kind of more in the entertainment category right it's like uh in the kind

53. 27:51

    of uh content generation on social media Twitter posting yeah maybe the posting is like uh it has utility for some projects right like if done right utao right yeah or k yeah marketing so that bucket and then there is something interesting around like these more kind of art like music generation or like um um yeah so only these two categories like I feel that like we will quickly move past these and as like base models improve like all the agents that I call

54. 28:35

    utility agents or maybe you can call Service agents are going to come to life so at that point I agree that we need like uh uh intermediation protocol between agents right it's not just the communication but it's also the ability to make commitments uh to like have access to like uh assets on different chains um yeah I'm curious also from your perspective I cannot really see how this will emerge right because standards are kind of hard to form and we cannot sit in circle and design a standard like

55. 29:09

    it's probably going to emerge from like the protocols that are more most successful uh in terms of volume of like these agents coming online but I'm also curious on your perspective on this since in near you are hting a lot of these projects and kind of creating like a bigger ecosystem yeah I mean the way we kind of think of this is effectively two punch approach right so as you said like for for this to take hold you need a lot of a like

56. 29:39

    Service agents I call them useful agents to contrast to useless I mean to um mecoin agents uh but uh so you need a way to like okay well where this go where's the service is going to run how they going to be able to uh you know register themselves indicate what they do uh you know get paid um kind of settle run inference you know pay for that cost Etc like all of that mechanisms and it's you know it's

57. 30:08

    totally possible for somebody to run all of that infrastructure and run it and you know pay for those things and kind of settle costs Etc but it's all complicated and like if we assume you know every restaurant and every you know Uber driver and every whatever should have their own agent who is effectively representing them uh on this in this agentic lab uh well this should be super easy for anybody to start you know to

58. 30:33

    configure to run right and so that's where the agent hosting comes in and that's where you know te infrastructure with GPU inference needs to come in because we want to ensure it's entend private it's verifiable you know if you want to run autonomously you can if you want to you know obviously this agent to represent you and you have full control you also can so that's kind of you know

59. 30:55

    the app. near the offering you effective an agent hosting we actually uh sponsor AI inference costs there as well and you know you can like Fork any agent you know configure it it can use blockchains it can use Web Two Services it has Secrets Etc now and kind of that's where the you know part one comes in and part two is like okay well now that we have agents and as well they can use Services Etc uh there needs to be a way to like kind of anyone to say hey I want some

60. 31:27

    somebody to run marketing for me right or I want you know pizza or whatever so like indicated intent and so there is you know a network of this agents who can be like yes I can do that this is something that I'm like capable of doing uh now the other agent or you user's assistant right selects the you know which one it wants to transact with they sign commitment you know it goes on chain and there is a mechanisms to you know reason if If the product wasn't delivered if the service wasn't

61. 32:00

    delivered if it's not on par not quality Etc so that's kind of the like two indeed two-sided thing and there we also like bootstrapping it with just defi because kind of that's the easiest place to start and uh there's a massive Demand on just like a multi-chain kind of Swap and Bridge which this thing very easily offers but then already we're going to have actually example in next day or so uh actually uh or or in the past

62. 32:30

    already um is the uh ability for an AI agent to effectively respond to the intense as well as onchain so like I call it amm where a stands for AI so AI Market maker and so it's AI Market maker that effectively uh runs in Te You Know can respond to like trading intents or onchain transactions you know decide how it wants to like what price it wants to quote you and then you know uh kind of

63. 33:03

    settle that transaction right so this is like just just like kind of building up from those two directions like more agents that can do more things and then defi into more generic intents kind of bringing them together yeah that's a very interesting Vision I initially also thought about this like how to grow this Bart marketplace right this like a network where one side is consumer users for our case is like every user have their personal AI represent them on the other

64. 33:34

    hand it's every single small business should have a service AI represent them and you need to uh like grow this bipart network and uh I was looking at the internet and feel like okay to convert all the service AI um it's probably the easiest to start from the Legacy web and the Legacy mobile service who has already apis right so we have this concept of agent link that basically you can wrap a older like website into API and then API into a agent and then

65. 34:08

    attach it wallet and everything um that's kind of the initial design but in our first phase when we look at how to grow a bipar network we feel like okay we should only grow one side which is the supply side of like the personal Ai and just have every user uh on board their data and build their own personal and get the personal very well since I realize among the consumer side even if

66. 34:33

    there is not much Service as of 2025 maybe it will happen in 2026 uh in 2025 it can still be like a consumer to Consumer Network like two% can have like social interaction can form all different relationships if you look at internet today there are like Facebook there are LinkedIn there are Tinder there are like all different kinds of like consumer to Consumer interactions so I was like okay maybe let's just focus on uh the consumer side and grow only one side of network that's kind of

67. 35:04

    my reasoning so I feel like n is taking a slightly different way which is attacking the hard part of like land uh kind of business internet right yeah I mean we kind as I said we kind of attracting it like attacking it from two directions because you know we have this kind of financial ecosystem right that's been built out anyway and intense is the right approach like intent is the right approach for anyone

68. 35:30

    with anybody else coming in into some kind of transaction right um not even Financial right it can be anything and so but like from Financial is pretty clear path and so like we're expanding that protocol making it more generic and then uh I think you're right that so far right it hasn't been very useful but I think like the the things are getting really close where you can you know buy from Amazon and you know like book travel and do all those things and so

69. 36:01

    those examples that I think uh we'll see personally I being able to do really easily on your behalf uh you know either through what you we call them shims yeah like effectively wrapping an API with an agent or uh or web agent so like I I I have this like thesis effectively that you know if your if your website or an app that's against AI right you will get web agent meaning like there will be web agent that will just go and click on your website and do stuff and so so if you're against it

70. 36:37

    you're G to get like automated or you you know adopted and you you actually capture some value from that and become part of this right so you you have two choices and so web agent to me is it's like the yeah the the back stop the shim is effectively like a way to you know really easily to upgrade someone into this world and then being AI native is effectively when you build build your

71. 37:00

    business AI first right so that that's to me that's like a progression for for anybody and and kind of for use cases and I think we're gonna like given what you know we see like in YC as well like I mean again outside of web3 like there's a whole world of people building effective AI startups right now ai first startups replacing ton of services right you know marketing automation uh like

72. 37:22

    phone like all the phone call services all the stuff is like effectively getting uh repl place right now was AI companies that are like AI first AI native and so one of our other advisors and um friend is uh Tim sh from from Cresta so they were building like this cold Center agents like uh very early they are pretty like they're not a startup anymore um I have one thing which I think could be interesting to

73. 37:55

    double click is like this notion of intent right because like especially in web 3 like it's kind of an overloaded term like for some time like everything was about like this intents which were just like sometimes just partial transactions or like very like defi focused right but you mentioned something that like we've also been thinking a lot about I started calling them like generalized intents like just to like separate them from like only defi transaction I don't know like maybe

74. 38:25

    we need a new name for this uh but it seems like you are alluding at like uh a new object which is like actually very different from a transaction right a blockchain transaction like can you say more around like what's kind of like the design principle and like uh how it looks like even for developers and what kind of this abstraction is trying to achieve yeah yeah I'm I'm I'm actually behind on a blog post about this

75. 38:57

    yeah so I mean and I agree that like intent in blockchain somehow became like this is just like a rap limit order and like that that became a meme but I I Define intent as like somebody else to get some done for you um or or in short get done uh the to me it's effectively a interactive offchain interactive protocol that involves uh three steps so first step is uh intent request so this is when you you share that you want something done and you either directly share it with

76. 39:37

    some specific party party or parties or you broadcast it to a you know some network of of uh kind of intent solvers right so so to give you an example I mean obviously you know I want to uh somebody to run a marketing campaign for me and generate 100,000 Impressions on my website right definitely not a limit order um or I want to buy an apartment and with a mortgage again very like very much an intent some constraints on it yeah like I'm in I'm in Lisbon and etc etc right

77. 40:12

    so okay so then you have a this is interactive so respones whoever sees an intent and says like you know this is something I can do or or interested to doing they respond with a quot like we call them quotes uh so this this is effectively an option to fulfill this intent uh which can be a composite right it can be that like you know I got an intent of like I want to buy an apartment as a mortgage I actually queried somebody else who is selling apartments and I quered somebody else who does mortgage and so yeah so

78. 40:43

    I'm returning actually a combination of these things which is you know this apartment um as a collateral to this mortgage provider you know this yeah you need to pay this much as as you know deposit and this much interest rate for this apartment that you know you want to buy and you have multiple options right so so this gets returned back to the caller requester and so that's when you

79. 41:07

    know again if it's a user they can select an option if it's their AI assistant it can like pre-core it and maybe show the user on these selected options whatever that thing is and then when user selects uh or AI agent selects uh that's when kind of the sign happen and so this becomes a commitment so this now has a commitment of the solver and the requester that goes on chain and so on chain actually the interesting thing is on chain can not just be like okay

80. 41:39

    let's exchange a v like money but it can be as complex as like running an arbitrary program on chain right because the the condition of the success can be arbitrary you can literally attach a JavaScript piece of code into this thing which will run on chain over time you know like in six days you know check if the apartment ownership was transferred or whatever including you can actually

81. 42:02

    run an AI agent yourself to check that something happened like oh go to registry of apartments and check that this happened and if yes then release money from escro and if not then start a dispute process which then you know goes to another agent who's like okay evaluate the the log of the you know what what they were trying to achieve and see who is wrong and if not go to court like if you cannot do it go to

82. 42:27

    court and actually this is like legally signed document because you have two signatures on both sides and there's like underlying kind of chat history that happened right so so the way kind of I think of this is effectively what you traditionally would do as a company to company right you would have a legal document you would sign it on both sides you would have you know force majour entry you would have you know like on

83. 42:51

    not delivery Etc like all of those things it it would be natural it would be a natural language program right it programs effectively a lawyer who will review it later to decide what happened and uh it would have cryptographic signatures and then you would have billing and invoicing in all the escrow Etc so all of that tooling can be done by blockchain all of the interpretation of this thing can be done by I agents and you also have this mechanisms of

84. 43:18

    Discovery and kind of participation so that's it it's a pretty you know it sounds complex and you know there's some complicated parts of it but it allows to actually move a lot of the things we usually expect smart contracts to do to kind of this offchain protocol which is way easier and it allows to do effectively everything in VB 2 right now in this protocol way more effectively I see yeah makes sense and

85. 43:45

    then you use the chain just like for the harm property that it's built for exactly it's execution settlement and like in like Val it's effectively a reg register of you know assets and value and it's settlement and kind of state transition execution yeah that's really rather interesting you thinking about such heavy like real world tide things that easily will need to go to like a real world Oracle like cord and stuff right

86. 44:15

    like I I was more thinking like can we uh replace today's like say internet eyeball matching things that happen on say like uh uh how like user gets to see this Twitter post rather than that Twitter post whether the meta push you this ads or that ads right today the like hidden contract is you use a web two internet a web one internet product and they do not give you any value they know your personalizations through a first party Clos Loop data flly will and in the meantime they send ads to you and

87. 44:50

    like that's your cost of using their free uh product and service but I was thinking okay uh is it possible to have a different way of doing it where user can actually sell their eyeballs their attention and their intent which in my mind is more like a uh attention but like associated with a transaction purpose right like Z would be some manit value that happen with this attention uh and I feel like uh in pis design were're more thinking can we let

88. 45:22

    every user basically run their own Facebook like uh as auction algorithm and the sell say like their like needs to curate the top five uh service provider that are all qualified to fulfill their intent and then they come and beat for like who gets this final um like order of like ability to serve you this opportunity to serve you this is actually not that different from the key

89. 45:53

    difference here is that like you are talking about like a purely dig example part about maybe that youting is that okay for like real world goods and asset like maybe more challenging because for purely digital Goods then you can they have a cryptographic yeah you have some some authentication yeah so that I mean the pure pure simplest example for Pure digital is obviously swaps right and that just that already works right now we have a product effectively running on top of that the actually the eyeballs and advertisement

90. 46:27

    is interesting because uh well outside of me hating hating advertisement but it is a useful it is a useful tool and so one of the things we actually thought and you guys are probably pretty close to that is you can actually use the secure Enclave on the mobile device as a way to ensure that the user actually watch the ad uh and and do all the scoring indeed on the local device as well Brave actually I don't know if

91. 46:53

    you're familiar Brave is doing that model as well which I think is is a a pretty good model and like how how do you send then the question becomes like how do you send and settle receipts around that user watch that ad right and like all the you know tracking around that and there's some interesting like cryptographic questions like how to accumulate like how to build an accumulator of you know receipts because you don't want to you know if you're having like a million impressions on

92. 47:17

    some ad like you don't want to settle a million transactions out of that although you can on near but um uh you C I mean there's still like there's value in like you know aggregating that into like a payment Channel style system so which is also something we've been look looking at around te as well like T become kind of being one of the parties uh in the payment Channel which allows you to do

93. 47:41

    monetization very effectively as well cool I mean I think we're over time so it was really awesome conversation thanks for uh kind of explaining a vision and obviously very align with it uh and uh yeah I'm excited to work more together and kind of do a lot of interesting things in in bringing all of the species together yeah yeah definitely yeah it's a very hard mission to actually push as we discuss you can say like there are more and more gravity pull towards real world uh and uh yeah

94. 48:12

    like as we move out of like pure Financial def I feel like it's it's probably going to be 99% of the future transaction 10 years from now right like if you look at real internet it start from lot of like a digital financial transaction slowly it becomes much more about uh like games and ads and like post and all that right I feel like with blockchain we might go through all that let's push together exactly we can only

95. 48:44

    win as an industry this is something convinced of right like yeah