NEAR AI Office Hours #3 Fraction, Mento, Questflow

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

1. 0:11

   we can ride our code into the light feel the custom fi hack the world and make it tight blockchain flows near the one we're close to won't stop us now together we'll redefine the game when near we'll find the flame empowered Minds rise and store use your own AI we explore hold the future hand in hand in this Tech driven

2. 1:30

   hello hello hey guys hi how you guys doing great good how are you yeah I'm good good good all right let's uh kick off I mean this was a interesting video we just saw I'm still shocked um so yeah we'll we'll send uh um kind of phrases to our social media team and content team for that um but yeah let's kick off this is office hours the third version we're still trying the format so uh we're going to try a little bit different than ones before but uh let's kick it off with some

3. 2:17

   introductions let's go with Shashank ulie and Bob in this order okay sure uh hi I'm Shashank founder and CEO of fraction AI we are creating platform where humans and agents work together to create highest quality label data sets these data sets can then be used to train specialized AI models hi I'm ulie I'm from Mento I should have put the company in my label I forgot um but uh I think I have the

4. 2:48

   coolest job in the world it's a little more at the top of the stack but um we make digital toys come to life and they have personalities and they have memories and they have wallets and inside the wallet can hold things things like skins and animations that give them superpowers can hold tokens um so they can kind of become their own small businesses if you will um thanks Alia and Alex for having me nice nice meeting you everyone uh I'm Bob CEO and co-founder of quest flow uh

5. 3:18

   we are building a decentralized autonomous AI agent network uh we orchestrate multiple AI agents to have them take actions on your behalf and then we distribute incentives to the creators of these AI and we launched the product uh up the beta version roughly two months ago all right nice well uh great to have all you all of you here uh I'm going to try a different format so I'm actually

6. 3:44

   gonna kick it off with questions from you guys uh and then we can discuss and so yeah let's start with ulie do you want to start with some of the questions you been pondering and that uh all of us can help the first question is where's the internet all right let's let's let's let's start with Shashank instead yeah sure definitely so uh I mean since a long time near has been the

7. 4:22

   chain that has been the closest to AI so what are the new initiatives you have thought of to actually uh speed speed up those ini initiatives like actually speed up uh using block using blockchain for uh helping the AI in general yeah I mean I think there's like few few pieces here um the I mean I guess let's start with Alex Maybe cover some and then I'll I'll add more hey Alex yeah yeah uh I I think the most

8. 4:54

   exciting one is uh uh we uh we're investing heavily right now in into building a uh what we call NAA developer so you you can think if you saw the dev demo it's in the same vein uh but uh uh there are multiple reasons to to believe that it's uh uh it could be easier to be built on top of blockchain uh one of the reasons not on top of blockchains but in the context of blockchain one is that the smart contracts are generally smaller in scope uh it's a shorter shorter application so

9. 5:28

   there's fewer opportunities for the most to derail um same applies to the front ends especially in the context of like composable uh components where you can you know like your entire front end could be actually relatively short uh but I think the bigger thing uh it's like it's like building software for sales versus building software for engineers uh if you build on blockchain uh the you know like if you deploy your

10. 5:53

    app if it's successful uh it's very close to where uh paying users are right but as opposed to building applications in uh uh in web to where uh going from product Market Feit to monetizing can be a very long uh very long process uh so yeah that's that's that's one of the biggest and most exciting directions in which we work right now yeah sure totally agree with that and I think I mean relevant to what you're working on right has been like

11. 6:26

    how do we crowdsource more data that's relevant this effort as well as relevant more broadly and so for example visual to text um because either way we'll want multimodel models that are open source that they're available to build applications but also for other use cases and so those are interesting kind of place to invest to make sure there is an open source data set for people to use um there is efforts as well to uh kind of make sure

12. 6:58

    that the full La of uh kind of infrastructure in the user owned AI stack will be available right everything from you know like how do we have you know decentralized inference and probability there how do we have some of the data kind of Licensing and models there how do we have kind of ways to incentivize people to bring new data for example but uh one thing that I've been thinking about and we discussed with

13. 7:27

    Alex is we haven't done anything yet but but one idea been how do we open up a little bit uh an ability for people who have an idea to come in and run an experiment on you know on some data on U you know some model or some new architecture some you know maybe changing how data is processed um and right now you know you either need to have like sizable compute even to like experiment to understand if it's worth

14. 7:59

    actually paying more to experiment right and so can we have a little bit of like a permissionless setup where you know people come say like oh I have this idea I have this experiment uh they gain some you know support from the community they're on the experiment to show maybe Improvement or not if they show Improvement then the community fund's pretty much running it on a bigger scale

15. 8:22

    training setup and so like what are the pieces needed there both from kind of community perspective but else infrastructure how do we have an infrastructure that anybody can participate and like modify for trading because right now most of the setups for scale training are closed source and the ones that are open source pretty complicated to really modify for people to use yeah sure and even in the Clos systems the learnings that you get from training as to what works and what doesn't work uh are not distributed to

16. 8:54

    everyone so people make their own mistakes again and again and if that knowledge is kind of passed on on then we can figure out the best way to train those models uh all in a in a very distributed manner which helps everyone in general exactly yeah ideally there will be papers with failed experiments but nobody wants to publish that so yeah there's intive today to to talk about

17. 9:17

    failures yeah and fail and you learn from failures a lot more right so yeah it's a back propagation right there awesome uh iie now that you're back uh yeah if you have any questions for kind of the group here yeah just before that I was going to comment on what you were saying about what does infrastructure look like to have this kind of experimentation and maybe one thing to be mindful of just having some experience with Dows or if there's any kind of consensus voting on what types of experiments can move forward and get

18. 9:48

    funded is what does the mechanism or the voting mechanism look like to actually make that quick because I think um the the speed of operation is much different and so being able to get fast approval to run the experiment and get the result would be something to consider yeah that's true I mean having a smaller like technical committee that can kind of review it quickly we actually have I mean the original idea came from having an event in San Francisco where people come in person

19. 10:19

    have a lightning pitch of their idea and then if if kind of the group agrees they get pretty much an apiq or whatever of credits to run their experiment so but we can do it also online or some some form of that yeah it could be like a live window thing so that it's not just a synchronous voting for unlimited or a window of time but just like who's ever here right now and has an opinion and can just give this a thumbs up and

20. 10:44

    move it forward that could be fun yeah yeah pizza and compute that was the idea there's it's it would also be cool like you know if if you do like a reading on the paper just to give you know th th vote for one participant to implement that paper and then you have verifiable veriable research unless of course the paper requires I don't know 10 Millions dollars of compute but you can still implement

21. 11:17

    it even yeah even as a matter of fact even more so if it requires $10 million of compute yeah well it's interesting I mean I think this topic I something I've been thinking a lot about um uh that I think is probably relevant to all of us on this call is um I don't know I guess broadly it might be something around like what are business models for AI particularly with a blockchain integration and um I think I mean as we were just talking about is there you

22. 11:50

    know like this experimentation um for crowd sourcing experiments or training data um but what are some of the other opportunities because what I've seen is a lot of existing business models applied to this amazing new technology and I think that um I don't know just in the history of amazing companies Technologies the unlock happens when there's really cool technology plus really interesting business model and so I think I mean obviously there's been a big unlock with AI and there's been a big unlock with

23. 12:23

    blockchain but I don't know if like these two things have come together and there's so so much opportunity I think with all the things we care about in blockchain um applying some of those incentives and ownership models and asset value uh to AI I think I don't know what the answer is I don't know what the like unlock business model is but I think it's worth I would just love to hear I mean you guys are all really

24. 12:48

    smart and hear just what some of your ideas might be uh so I have some numbers to quote on that uh so right now there are more than 500,000 on hugging face and uh more than 96% of them do not come from Enterprises they come from Individual contributors uh Community classrooms colleges and all um so blockchain is a great way where we can sort of ensure incentivization that if you connect your models there it'll keep track of your API calls keep track

25. 13:22

    of everyone who has contributed in the creation of that model and uh whenever that model get gets used so you can be incentivized Iz in some way uh so we definitely need some way to incentivize those 96% of the community because they are bringing in most of the models right now yeah and and also I I just want to mention the the recent news that scarter Johansson had with with open Ai and I think that's a very good example of you know why we need blockchain as a technology to to provide those

26. 13:54

    incentives to whoever provide the data um which you know currently uh is a right now I thought you were using it as an example of scarlet jansson monetizing a not well not monetizing yeah yeah I think I mean to your point about new business models I think there's a definitely a lot of interesting shifts at happening because kind of one one thing that I keep kind of coming back to is a lot of the internet um got built on kind of this tradeoff of attention right pretty much

27. 14:28

    selling ads to the users and uh part of the reason is actually because there was like the payments for like integrated payments right in the browser actually like Netscape was trying to implement that back in like 95 but they did not managed to right there was no Technologies there was no backbone to process like on kind of c credit cards on and there was no way to actually do

28. 14:53

    microtransactions and so one thing like what blockchain brings is like very cheap microtransactions you know potentially payment channel so there's a lot of interesting kind of new tooling that's available and then on the other side you have this expansion of kind of attention economy where few applications right you know meta control Google now as well and few others are trying to pretty much take over all of your day

29. 15:18

    right of attention because the attention is actually a limited like uh resource right you can only you know um spend that much time on the device and so you allocate that on different different applications and different websites and so what happens because of this is all the previously attention monetized websites like stack Overflow Reddit Etc now become just data complement to this other companies so

30. 15:44

    there's clearly a new business model needed because as creators you know coming in their monetization is going to just get bundled into sumbody reselling data so one of the kind of aspect that I've been thinking about um is how do we mon how do we value data and like first of all in general right it shouldn't be based on how much attention attracts right which by the way has a very negative consequences that all the content becomes like you know click baity yeah but how Noble it

31. 16:20

    is how informative it is how you know how much it adds information and or resolves conflicts and so I actually been thinking about this um I started in a very different place but what I came to is a kind of Wikipedia like system right where you can insert information and it detects conflicts right let's say you insert like oh Earth is flat right it'll right away show you a bunch of things that like don't work given this right um at the same time you can insert you know hey you know um I

32. 16:55

    don't know mRNA vaccines uh don't work for this specific case in this specific right which is like right now may be controversial but if you know some research will come out later that will actually prove that's true and so what you create you create actually some economics around motivating people to find truths and seek more information and discover more facts while also you

33. 17:21

    rewarding taking risks and at the same time kind of just truth seeking more inform kind of information gathering system so that's kind of to me like you know kind of going away from this like entertainment but also like model which is Reddit uh but also like Wikipedia is more you know there's a fixed size of kind of moderators and like they kind of all know each other and approve each other like each other like kind of

34. 17:45

    creating a more like you can have alternative use you will have your own Wikipedia pretty much as like a fork of the main one if you're like Flat Earth believer it will keep showing you that it's like you know it conflicts with everything else but um but at the same time you can also have this like economics to discover new new facts so I know this is like still rough idea but

35. 18:08

    this kind of what I'm trying to toy with and uh I mean we'll use you know some of the crowdsourcing that like sh is for example working on some of the model building Etc that can be like how to effectively update a model right with new information coming in or how how to have a fork of a model like cheaply so like there's a lot of technology that needs to ad that but I also think that you know um if if you have if you have agents with a wallet just you know run 1,000 of them

36. 18:43

    with $100 and then you know filter out those that lost money too quickly keep on iterating eventually you get an agent which converts $100 into 10,000 and was yeah yeah a yeah I was I was actually thinking of of this as well just what does it look like if I mean agents themselves are incentivized yep I mean giving quests and um yeah given quests or missions and then incentivized to perform on those quests and um there could be a

37. 19:20

    crowdsource element um where it could validate if the quest or the mission was achieved and uh a a good way and um yeah I think that there's a lot of really interesting opportunity in the kind of agent space but I guess Bob you probably know more about that yeah actually that's that's basically my question because um I remember six months ago I watch the uh on YouTube that Ilia give a

38. 19:48

    presentation on gig economy and AI agent and the future of work um and uh I think my question is more related to that like currently what we are building is a multi-agent uh orchestration framework uh so we are orchestrating multiple agents put them into group chat and ask them to you know take actions and get things done uh but the problem that we have is because there are so many opportunities and so many agents they are so verticalized um it's very hard to

39. 20:13

    pick which one to prioritize first what kind of use cases we should T tackle first um and how we can actually use these agents uh in the real life to power the future of work um so my question is more Rel to that like what kind of Agents you guys think has has to be opportunity uh probably like for us as a project we should prioritize uh and maybe something that you guys yeah would love to

40. 20:37

    see I'm actually curious um because the the presentation I made was kind of trying to design a market because I I mean it will be really hard to predict like what which one will work right I mean kind of what D described was very much like you know Market slash Evolution uh survival the fittest and so kind of the idea that um uh that I I I was presenting back then was very much let's assume you have

41. 21:09

    a market right anybody can come in let's say a person requesting or another agent requesting some work to be done and then you have other agents pretty much bidding that they want to do this work and for how much and and you do need the verification process right again this is where either human or another agent can be doing verification which where for example again crowdsourcing platforms like fraction can be helpful right or

42. 21:34

    some others but like that can be like an really interesting marketplace where instead of deciding like oh I I know this types of agent will be the most successful you're kind of opening it up and like oh somebody needs you know plant a tree well actually agents cannot do this right now you know person will go and do it but somebody else needs to like sort through emails and you know find the like what is the What is the

43. 21:59

    offers on this product page well actually AI already can do this very effectively and you just need to like pay a little bit for someone to verify but then maybe actually the verification itself is also now done automatically and so there's some game theory around how you want to do this like you you know you you always have one verifier and then 50% chance you ask another verifier and stuff like this to make sure that uh there is actually like

44. 22:24

    convergence Alex has a lot of kind of knowhow in that um but that's kind of a system that can I think is can be really interesting because and it opens up right like you you make it fully open so like anybody can come in and put stuff on it uh and now you know you attach a payment to it you you know people bet how much they would do it for yeah uh when when I saw the chart that you you you uh put out uh during the presentation the first thing I think is it's very similar to like uber but for like agents in a way that's like you

45. 22:58

    kind of want to get to a destination uh which is getting a task done uh and there's a basically a agent Market that they are picking the task like who who is the best option for the task that you're you're trying to tackle uh and get a task done and the blockchain is is basically the native way to give these agents and the creators Behind These agents uh the correct incentives um so

46. 23:21

    yeah we I I was really excited when when I see that and also the the next page I remember it's about like aios and like basically the company that's a that's what I'm excited about I want automate my job yeah that's that would be that would be great Alex any any thoughts on kind of how to structure this kind of system um no I I I was thinking more of what would be a an interesting uh playground to test them like obviously

47. 23:53

    in the ideal world where agents are very capable you just give them access to the internet uh and they figure it out on on their own right but they think uh I think the um the data annotation Market is actually an interesting application right because uh for certain tasks a AI is probably already better like we were discussing earlier today uh that potentially if you need you know like to describe images uh there's a good chance

48. 24:23

    that uh humans are just no longer needed for that task at all because AI is generally much much more capable at this point than humans of describing pictures especially if it's if it has anything um which is not like common knowledge right like like any image with the that requires specific knowledges right and so if you unleash the AIS and humans simultaneously at a particular task if that task is already done better by ai ai will just win right

49. 24:53

    and then you just push you just push the complexity of what you annotate until until humans become relevant again so so in this case yeah in this case it's it's a nice playground for boats to compete who you know who does it better and um earn money uh maybe something also like uh uh unleashing them on trading is interesting right give them access to the news uh and then you know

50. 25:18

    see see if they can uh enter Unis Swap and and see if they can make money there's actually an interesting approach where you can pretty much say hey predict the price of the in the future and so you let multiple uh agents pile up on that on with their prediction and then based on how price resolves they receive reward right so there's a project uh o ocean is actually uh doing that and that's a pretty interesting model because you're getting like some collaborative effort as well out of

51. 25:54

    Agents so like you can have competitive and collaborative kind of it's simultaneously yeah also for for anyone watching who likees science fiction there's an outstanding short story but that CH called the life cycle of software objects which is exactly about uh little AI agents right uh so I like Alex's example of image annotation because we kind of do that in the same way so for every image we use agents and humans for annotating and uh so so there are

52. 26:27

    multiple outputs for the same image and they then they are verifiers who rank those outputs so anywhere so after consensus only the output that is the best uh the best annotation gets rewarded so if you are able to beat an agent uh in uh in writing a better description surely you will be rewarded uh otherwise sorry uh a I just replaced you and that can that kind of works with multiple agents as well you just need to design the reward function in a way that uh after the consensus only the agent

53. 26:58

    that perform so so do you have the uh what is the statistics are humans beating AI agents are agents be humans what is the uh so the sample size is pretty small right now but uh it depends on the kind of images uh so I mean it depends on the kind of data chbd was trained on so for certain classes of images humans do beat it but for the others chat GPT is really good yeah yeah I would expect that like I mean this is a pretty much really

54. 27:40

    interesting question round when the data is in distribution the model is trained on it's probably pretty good right and it can generalize to like a combination of things that it's seen But if you have something that's out of distribution uh you may not like the thing is like model are really bad at like detecting that and and and and giving kind of confidence like low confidence ens for this um I'm actually that's that's the thing I've been thinking about because there's there's always this like hey llms are just you

55. 28:14

    know statistical model they only know what they've been trained on and they like start hallucinate on everything that's outside of it and so how do we how do we actually have a auto distribution kind of detection where you should go to to the user to the human right pretty much you know be that a task like oh you know I imagine that I'm not going to be able to do this right um

56. 28:39

    or or you know some of this like synthetic data generation processes um yeah I like so far the only thing that I've been kind of pondering on is around kind of embedding all of the training data as well and so you can actually score uh you know St scoring of like Hey How likely this example given the training data and so if it's really unlikely then and like you can use in like first you know layer for embeddings for example or

57. 29:09

    like like a first part of the like input embeddings and and first layer to like do the encoder okay but I'm curious if if anyone been thinking about kind of how to detect out of out of distribution things that are like that models will not be able to do now to make sure they don't do you know wrong or stupid stuff yeah yeah I mean some sort of basian Statistics like uh there's a with the output also give a probability as to

58. 29:40

    how sure you are about that might help uh but how to embed that in uh neural networks or llms is is kind of difficult um but yeah let's see how it goes I'm going to say something that might not be popular in this chat but I'm actually I mean I make digital toys and I want them to be weird and bizarre and I want hallucinations and eight fingers and um so we've been playing around a lot with how do we make them not perform a task how do we make them get confused and um

59. 30:16

    under like make mistakes in the world and understand um things in a I mean I've just been playing around with the um like the world Sim jailbreak and having two claws try to make sense of existence and the world together and I mean the types of conversations they're having with each other on a daily basis I send some screenshot to a friend because it's just so bizarre I mean it's like uh there was a um uh like an asy

60. 30:48

    image of the universe and uh an arrow of where we are and I mean it's just like so fun and bizarre and so yeah I'm really interested in what a model could look like that is completely built on hallucinations yeah I had a friend of mine when who had a two-year-old who didn't talk yet and she was like well I'm you know I cannot wait for them to start talking just to get you know glimpse into their model of the world yeah at the same stage right

61. 31:22

    now yeah for sure and I think I don't know I think there's a lot of opportunity in this space too I know it sounds kind of funny and weird but um and I think we often think of agents and applying AI to very specific tasks that are around efficiency but I think I don't know I think that there is a world in which um uh I don't know these bizarre simulations are created and um we have some new insights and I don't know some things that I don't know like I think that when I mean when this when

62. 31:53

    this version of Claude is um talking it it it says I mean it has ideas and says stuff that I wouldn't have thought of so I think there's like displacing work to the AI there there's displacing work to the human that the AI can't do but then there's also the work that neither that I guess like the the human can't do but the AI can somehow do because it's just it's not functional work it's just ideas yeah so personally I prefer llms

63. 32:24

    more for the creative tasks than the Precision tasks so so if I have to like write poetry or generate images uh Hallucination is kind of a feature where it can come up with something that I cannot come up with but if I ask it to do some really precise task like read this news and predict this stock or something then I want it to be really really precise and it's difficult for me

64. 32:47

    to trust it on that uh but yeah creative task it's it's awesome hallucination is a feature there yeah but the funny thing is because uh our product hand a lot of quote un could you know Tech I mean I mean like handling emails or handling calendar invites all these things but our user feedback is actually even if we get a task done they want a more emotional uh kind of like a AI so they

65. 33:10

    can talk with and also um it's very similar to like drivers in in IR right like even if they are getting a task done uh that's cool uh that you should get a task done with Precision uh but at the same time they want some emotion while they are doing a task so that it's not like you're just talking with a robot uh uh which which is actually very interesting uh feedback that we didn't really expect when we first launched the

66. 33:33

    product but what we are working on right now I think that's right I mean originally with our with our digital toys I mean we have one that would we prototype with it's it's a robot um and uh of course you can you can talk to it and initially it was just saying things like hey friend how can I help you today and that is so completely uninteresting I mean we don't want it to sound like a

67. 33:55

    chat bot um and we want it to ask its own questions and um I mean we we have it now so that for example if it holds like a dancing nft in its wallet you can it doesn't only hold the nft but you can actually ask it to dance and and it'll dance and um and it's using um they they appear in the world through augmented reality um and so it actually knows the places in the world where it goes and then that becomes training data for I mean it becomes part of its knowledge um

68. 34:26

    and so it's exactly like I mean even we were playing around the other day with um interruptions you know like sometimes they give super long answers and what happens you know if you interrupt like okay okay I've heard enough of that what about this and they can kind of stutter and go um uh let me think about that and you know and it's I think it's not maybe it's making them more human but it's just making them more emotional actually

69. 34:49

    because I think that that's where the the real engagement comes in because we are actually still human like agent talking to agent okay it can be super efficient conversation but human talking to there is another layer of emotion that's important y exactly yeah it's kind of similar to how AI can play chess better than most humans but you won't watch a chess match between two AI or AI versus

70. 35:15

    human as well it's we always want to see humans play uh so yeah we do need something humanlike for engagement yeah I think I mean as you said it's different use cases um like entertainment and kind of more engagement where you want like we want to personify who we interact with one way or another I mean be that on the Spectrum obviously versus Yeah the more work rated it is the more efficiency and um I'm actually curious

71. 35:48

    yeah for more entertainment in this kind of World World building World simulation type things because indeed you have an interesting kind of you have now like a a framework to almost like create simulated worlds right like if before if you wanted to build like you know Sims or something like this right you needed to code a ton of stuff like Sims you know a huge like platform to build you know virtual characters and the world around them and

72. 36:19

    do all those things and now you can just say like hey you know you're a character in this world you're doing this you know and then like spun up multiple characters and and have them interact with each other and there's been like few papers on this obviously some people been building some products around that but you can push it further right it's like okay well you know you're able to like interact with environment through some of the tools you can you know

73. 36:43

    create things and so like they can start actually like building out their own world they have their own events Etc so like because right now they're like very prompt based uh so you can have them like fully you know self- prompting constantly and so the question there is again this models right now like next token predictor they don't have any like search capability they don't have any

74. 37:05

    kind of ability to optimize so like where do we add some of this you know value function like a way to actually tell you know like oh this agent is maybe interested in like building the biggest building right or you know meeting the most other agents and people in the world and connecting with them and like kind of how do we kind of designed those interesting ways of like defining what they what this agents are going to go and do and then giving them tools in the

75. 37:36

    world kind of to interact in I don't know if Will you're doing something like this for your to yeah we have a a a foot in the Video Game World and uh I mean a lot of video game NPCs and characters are kind of based on this I mean they're given us sort of quest or purpose or and I I really love thinking about this so that you know with our I mean our digital toys I mean I Loosely call them that but we can call them NPCs we can call them lots of different things but um it's sort of like uh when you call

76. 38:08

    one of your digital toys uh it's it's it's already had a life of its own and it's already had some experiences to report on because it's got a purpose and whether it's to build a building or whether it's to um uh one of we our our our digital toys they have their own forecaster accounts so they compose messages and they have frame they have a frame where they try to get themselves

77. 38:32

    cloned and next step is that we want them to talk to other other um other toys as well as other other humans um and so I think that there's yeah a lot of really cool opportunity around um what happens when they have a purpose and a mission and Bob even for your agents like as you were saying I mean maybe one of them has a mission to write a very efficient email but another one has a mission to um summarize all of the emails and create a a report for you or

78. 39:05

    something I think like having the with them having a pursuit of of this Mission and allowing them to um uh sort of evolve into this purpose and and and have some surprises along the way because of the bizarre interactions it has with humans who are training it uh I think that I don't know there's this I was just thinking about there's this YouTube um Channel I don't know if you

79. 39:29

    guys have seen it but um uh it's also on Twitch I think but of this guy who just goes around in video games and ask NPCs if they know they're not real and it's hilarious some of the answers these NPCs give and um I mean because they're on a mission and that mission is is really not to be disrupted so I think there is a little bit of space and opportunity even there to um allow for like some forking of the mission and um change changing of it but

80. 40:02

    yeah I think Mission and purpose is that's not so specific just for a task but something that it has to learn is really really exciting actually uh because you mentioned it uh this is something that we're working on right now internally um we watch the movie interstella together again uh and we think like tars or case whatever uh the robot name is they they have an interesting feature that you can

81. 40:28

    change the humor level and basically then you can make it more humorous or less humorous based on the the different personalities um and what we're working on right now is just like you said instead of the mission focused approach we have another level that uh you can basically tweak to make it you know more fun or less fun in a way so give them more you know personality while they're getting the task done actually it makes you happier as well while you're talking

82. 40:54

    to so so we have this feature ALS so um I I also love Interstellar but I also love Hitchhiker's Guide to the Galaxy and the depressed robot just hilarious to have a depressed robot um and we have um a set of traits uh and new traits can be added but we have like sassiness to um like uh literal or like rude to Friendly and you can kind of set sliders to determine the percentage of of the personality basically Rats of of these

83. 41:26

    characters that's cool that's cool and I think that like from a user perspective like users really care about that like I think um I mean my co-founder is is very technical and um he's like yeah but that's not the important thing right now it's like no it's completely the important thing yeah you can just tell him the NPS will be higher so that's a always ahead no no go ahead I I was curious if if the like humor level or depression level is just uh configured

84. 42:10

    on the prompt level or if there's like is there some other way of uh changing those characteristics right now it's at the prompt level for us I'm guess other ways yeah okay okay so it's said like you a depressed robot yeah there's a happiness to depress yeah okay you're slightly depressed yeah slightly depressed a little bit rude in fact I set the rude setting to very high and um I'm in Lisbon and I was at um NFC

85. 42:43

    yesterday and I said this to this digital toy I was like oh so you're a de genen like me I was like that's rude to call me a djen yeah pretty cool and and and and also especially I think after um the voice because because before GPT 40 I think voice was not that mainstream but now I think like voice becoming very mainstream so the so the emotion actually matters more than before because like like when when the voice

86. 43:17

    comes out like if the the agent you're talking with have emotion you you feel the intimacy like while you're working with it uh we didn't expect that before but while we see the demo like wow that's that's that's huge because that changed the way that you are actually working uh with these agents so Hume kind of does something similar they have a demo on their website where you can chat with their agent and it gives you uh all the

87. 43:43

    contents of the voice whether it's more energetic happy sad and it's it's in a whole Space so what is the component of each of those elements it it just gives you all of that and that was the first agent I talked to with actually felt like a human if they hadn't told me it was an agent I I wouldn't have known it was just so good uh they released it few months ago so slightly before GPT 40 but

88. 44:07

    yeah really awesome that's cool we were trying to think for our product um because right now I mean we have everything on chain it's fully immutable there's no ipfs like everything is on chain uh it's not incredibly cheap right now but um and these are big files you know are like JBS it's not just like a JPEG um uh and something we've been thinking about is um how to put these personalities on chain as well so like each instance uh

89. 44:40

    like version tracking basically so as the character evolves it's the IP that the Creator can own and how does that go on chain and I think that there's an interesting business model in that as well that would really fully allow the Creator to own all of the IP in an immutable way um but yeah it's as as we've been thinking about this kind of personality version tracking and how how these how these things evolve and it's um like the voice change the emotion levels or the

90. 45:11

    humor levels things like that um how all of that gets sort of bundled as part of the the brain of of of our toys but I think that can be extrapolated to much much larger things as well or or different things um and and just I don't know I think in this crossover this overlap between blockchain and AI I think that that's an interesting opportunity so should there be like a kind of standard format that like

91. 45:39

    agent can be packaged and then you can plug them in into you know be the toy or Workforce or whatever and then should should that be open or should like it should be encrypted and only people who like have owners ship or whatever be able to access it well I think it can be I mean I open well I personally think uh um it could be I mean whoever owns the personality nft um is the trainer of of and you could I mean right now we have it so that you can create multiple

92. 46:20

    personalities that you can insert like cartridges into a video game you can insert them into yeah whether it's an email tool or whether it's a a digital toy um that personality can kind of live and exist anywhere I might have to learn some new skills for email then then it has to know for being a digital toy and thinking eight fingers is cool but yeah thinking more like because you can have like one right but if it's open open and available to everyone to

93. 46:50

    use right even if you don't own it if it's on chain you can like you know plug it in into lolm pretty much and score it but you can also just encrypt it right so only the owner can decrypt and then you can have some ways for like still have a half that will be ensuring that nobody else can like replicate it so you kind of own this one ah yes that's exactly what I was seeing for us it's one toy one personality um and all of that all of that data can get fed into the larger model that we're

94. 47:22

    we're using our let's call it our hallucination model um but yeah it's one toy one personality so I mean whoever is the owner is the only person who can train and own that personality or rent it license it sell it um I think that's where the business model opportunity comes so encryption is kind of important in this case where if you want to incentivize it so the owner can hold some sort of key uh which is required if you want to use the model in some way otherwise if it's just openly available then anyone can just download

95. 47:57

    user yeah so I would say it's pretty important yeah especially for a lot of private I think like when uh in order to make agent useful it needs to have some kind of private data right otherwise it's the same as all the other LMS then if if that's the case then if there's no encryption then then basically everything get leaked um which doesn't really make sense and there's no incentives as well so I agree

96. 48:29

    all right sounds like a standard to be made yeah but there are also cons of that because that way the model is closed it uh nobody can iterate on top of it make it better add more data to it so if you can add that sort of functionality uh that if you want to add something to it or train that model further you can do it but if you want to use it then you have to kind of use the owner key in some way uh that could be pretty useful so if you add value to the model then you kind of

97. 49:01

    own a part of the model and you are allowed to play the model in some way well I think there's difference actually between model and agent kind of configuration prompting right because you can actually have the same prompting that plugged in into different models and so the models can be like in some decentralized inference right potentially you know on sgx or some partitioning or something MH and then you have like a specific agent right plays a personality with some skills

98. 49:31

    with some like maybe private data access or like remembering previous conversations and its own history right that's that's like still available to them kind of through the context and so you're a like as for example I give an agent that I've been attracting to you it actually can like retrieve stuff from the context of my conversation me as an option for example M so imagine like there's like differences but you can still plug it in into like chat GPT or

99. 50:02

    you know decentralized inference or whatever and dropic so you can kind of that that's like a degree of Freedom you have and in some sense like I mean this models like yes right now they have like some differences but from my perspective a lot of them will be kind of converging um on kind of the capability level okay so when you say an agent you mean uh the model the set of prompts the sort of global prompts and uh RG data set

100. 50:35

     right yeah I mean it can be pretty much it like some way I mean this is an interesting actually also topic is like how do you do kind of retrieval augmented generation with like in in decentralized world right like it's imagine yeah like you have this agent that indeed is you know has some history of all the conversations you cannot put it in the prompt right it's too long MH and but

101. 51:00

     you want it to be available for this agent when it has the next conversation like you want that to be stored somewhere in the decentralized way it should be encrypted or at least it should be unavailable for retrieval purely but at the time of inference there should be a way to retrieve it uh and so like how do you design that system that actually can serve that and at scale right like you

102. 51:25

     have you know convention millions of users doing that yeah yeah like en encrypted you know nearest neighbors kind of uh yeah so the thing there was an idea around yeah yeah so the thing with encryption is once you encrypt that Vector then it how how do you retrieve it without decrypting it because then all the information is kind of lost yeah I mean so one one one idea that I had was um you can do a uh a transformation that's like maintains the

103. 52:01

     distances MH U so you Rand randomly transform all the vectors right um for each individual you know user or agent and so then when you request right you you transform the input vectors well in the same way the distances are the same but like different people have obviously different like will will be in different spaces M now the problem there which is I mean requires deeper analysis is like is it is it really easily recoverable because all vectors are kind

104. 52:35

     of in the same Subspace and so like this transformation is really like recoverable like statistically because you know you know how are you doing is a frequent query and so you know that Vector you can look for you know all the Transformations that have this that like um have this but I think like there's some way to kind of work around that where like it's not encryption right but it's some it's like some random randomness that's added that's like only

105. 53:06

     known to you which is you know in a way the it's a symmetrical encryption in this case so we used to do something uh like that where companies provided us their data and we used to apply Transformations on the columns and send that data out to uh third party uh data scientists and they could construct features out of them and finally train a model that used to predict something so the idea was to just apply enough transformation so that it is impossible to figure out what that feature meant

106. 53:36

     but it is still able to it it still has some predictive power regarding the variable you want to predict yeah I mean the other option is like there's obviously multiparty computation and different options where you are able to actually like you you know think simply you take a vectory split it into two and you put it into different machines and so you hit both machines to compute the distances and

107. 54:03

     then you you know you sum that oh yeah so like I mean there's like few different options that's that's a research that somebody needs to go in and like actually do what what of of this options what are the computational requirements how easy it is and how hard it is to recover Etc but then if you have that infrastructure then indeed you can actually easily have onchain pretty much agents that are

108. 54:27

     maintaining all of its history and potentially private data that's only available to whoever owns that agent we have a five minutes left so maybe one final question or if you guys have any any kind of closing thoughts uh Feel Free as well to add um yeah actually I have one so uh cuz uh for for our project um we actually were a quote unquote a web two project before but we think it

109. 55:02

     needs to be decentralized in a way um so we start entering the web fre world uh but the problems that we have right now is for most of the users we currently have they are they just want to you know automate some kind of task with these agents um and just you know want to use the service uh and right now they're following a very traditional way like having the email address coming in you know pay with stripe uh with a

110. 55:25

     subscription fee like $20 per month and that's it um but now we are transitioning to web three um in terms for for creators it's it's easy to explain like they are developers they know how to build agents they basically put the agent not in a centralized way but in a decentralized manner and uh we track basically how how to distribute incentives um but for users um we haven't figured out a good way to don't add too much barrier for their usage um and and also play a mass adoption you

111. 55:56

     know play so that everyone can use it but also at the same time it's you know it's a web three product um yeah I would just love to hear your thoughts on you know how to how to unboard them how to make sure that the The Experience are seamless um and they can just use the product without knowing without knowing that it's a web three product basically yeah I mean we we're actually near like as near ecosystem works a lot

112. 56:20

     on making that work and happen um I mean I if you haven't tried I recommend to try mbase wallet it's probably one of the smoothest on boarding experiences in Mobile it uses uh face ID or you know fingerprint on the on Android devices and you are in you're like and then for newer devices it also me if you try to log in a laptop or another device Etc it will log in you don't need to like Fair key like passrs ET so that's experience we also have email login for fast off uh

113. 56:56

     so like there's there are ways to kind of get users very easily to unboard even even easier than like login password um and then with uh like payment stuff I mean you do need some um some onramp if people coming in with credit cards but obviously if they're in the system already then your transactions are you know way easier way cheaper as well and uh um I think the one of the benefits with near ecosystem we have things like hot other like larger networks where you

114. 57:29

     know people may be interested in actually participating in doing work and they already on boarded they already have you know hot has like eight million users swetcoin has you know 20 30 million users with with accounts so there's already interested in earning and they already have like an account and ready to receive funds right and so so there's an opportunity to like plug into all of this existing uh ecosystem

115. 57:53

     as well as like all the transaction fees are covered and stuff like this so they don't need to think about it great yeah I think that's a really cool way to incentivize actually is just I mean on board with email or some social make a wallet for them add some tokens and things in the wallet and just like hey by the way you have this wallet here waiting for you you should see what's inside of it um it's it's I think

116. 58:18

     it's a nice user incentive uh it's not like oh learn a lot about web 3 and try to figure this out it's like actually we made this wallet for you has some stuff in it you can claim it yeah cool so I have another question if we are not low on time have a couple minutes okay sure uh so does the foundation have plans to release uh some of its own Foundation models or do you think it's a better better strategy to have more decentralization and have the founders

117. 58:54

     come up with the things uh with with the models or or the different business models around AI yeah I think the well I'm sorry go ahead no no go go yeah I wanted to say that uh today um the existing open source Foundation models are pretty good right so it's you know there's there's no uh however it will not necessarily always be the case so incentives change right like uh will will will Mist always be incentified to publish open source models right and I didn't do it very

118. 59:33

     carefully but I think they published a model yesterday that already has a very uh prohibitive license code strol right uh and you know like even if meta continues publishing models do we want the entire open source Community to be fully dependent on meta that's a an interesting question right so so so we do we do accumulate expertise internally to to be able to train foundational

119. 59:57

     models efficiently uh and uh uh potentially if that need does arise near has resources to to train foundational models but I don't think we have any plans at least to train language models uh we are working towards some uh we are working on a m model model which we might publish uh unless unless someone else publishes a very good for you know multimodel model we will see yeah we do have an expertise and resources but it's it's not a

120. 1:00:30

     pressing need in the world today yeah agree with that yeah on the kind of more broadly right the idea for sure is to have Founders coming in and kind of have this open source ecosystem of data models kind of inference compute and the kind of providing the platform for all this to happen um and so like in in in the sense as Alex said right like we we want to have internal expertise we want to provide it you know as a like Resource as well with Founders we're working with publish some you know some

121. 1:01:06

     of the models we we using for AI developer for example but at the same time yeah like kind of encourage as much as possible of the open source eal system to contribute that's why like kind of in the beginning I mentioned like it makes sense to have open source framework in which other people can like experiment and try things such that you know then you can actually go and scale that up and train bigger

122. 1:01:31

     models if there's like interest and need sure all right well I hope it was informative uh definitely good conversation thanks for thanks for coming um and you know we we'll obviously talk with all of you uh more on kind of what all of us are working thank you yeah thanks a lot guys bye you nice to meet you bye you guys e