NEAR AI Office Hours #16 with ZettaBlock - Decentralized AI Models Reshaping AI Development

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

1. 0:01

   when near we'll find the flame empowered Minds rise and S use your own AI we explore hold the future hand in hand in this Tech driven Wonderland okay hello everyone welcome to this week's AI office hours we have H Zeda block co-founders today um very exciting I'll let you guys introduce yourself and then we'll dive in in what you're working on as well as what's happening overall awesome hey guys nice meeting you uh Hey IIA and Alex great seeing you

2. 0:52

   guys again and uh my name is chi I'm the co-founder and CEO of Zab block so a little bit about myself um before Z block I used to lead product management uh at a web2 AI company called Data bricks and then before that uh I did my PhD from UC Berkeley more at the time focus on machine learning and causal influence so right now it's probably called expandable Ai and uh uh right after my graduation actually um the first startup that was a funing member at was a Automated machine learning

3. 1:25

   platform company so have been spending uh the last decades in the AI Big Data space so nice meeting you guys hey IIA and great to see you guys again um yeah hi folks um so my name is Scott I'm co-founder in CTO Z blog um before this start of I found a streaming DB company and before that I work at Uber before so I built a skill AI inside Uber like human in the luopa data labeling platform at Uber and then it

4. 1:56

   grow actually from uh it grow like more than two years 10 use case actually with 200 full-time employee work kind to labeling all the all the all the things for the for the data and then for the AI use case and before that I work at Salesforce on Salesforce I inst antic Cloud I was one of the founding engineer um I graduate from UI graduate school yeah that's my background awesome well great to see you here and kind of introduce you to the

5. 2:26

   viewership of uh a office hours Let's uh uh well give us quick overview of what Zeta blocks is and what problems I trying to solve um awesome so uh Zeta block is we call ourself a decentralized the data Centric model development platform for trustless and open a development so what does that mean is basically we enable model developers to collaboratively curate manage and monetize data and AI

6. 2:57

   models with a robust governance and ownership principles uh and the ultimate goal here is to foster a uh open and accessible ecosystems e ecosystem of uh data sets and models so more specifically basically we allow data curators to share and monetize the data sets for AI access um efficiently and also model developers can build and a share models with uh um those usage based rewards and also like the model users such as like the app developers or

7. 3:27

   even let's say the data providers can access and deploy high quality models in their preferred environment awesome yeah I mean oh go ahead Alex no no I just wanted to mention that it's something that uh uh I I think El and I were discussing something along this lines I think like as far back as a couple years so it's exciting to see that you know uh materializing that's awesome what uh I mean to make it spicy

8. 4:15

   what do you think would be the uh like your competitive advantage to win here yeah that's a very good question so um apparently there's a lot of uh uh companies and uh Platforms in web to uh such as dats where NS scl where together AI are building uh robust the platforms for U model developments and data processing for model development um but the goal here or like the differentiation we think uh we can leverage blockchain and crypto to do is

9. 4:50

   actually bring the better Provence tracking guarantee and uh verification into the process and the reason why we think this is very important is is we believe in the future there's going to be billions of um customized models actually built on top of let's say Foundation models to empower the business use cases or business process or even like your you know daily operations work and in that way in order for uh you to trust uh models and agents

10. 5:20

    to run the job for you as a company as a person you want to make sure uh one the uh the the contribution you made into the process actually can be uh track and also like GA the you know loyalties and contribution back to what the percentage of contribution made in the process and second of all very importantly how you can have a more systematic way to guarantee that this process is in have like a high integrity is not like being

11. 5:50

    you know poisoned or being tempered with um like bad behavior or like just a let's say bad data uh by the way we can dive into that more because we're data Cent um platform but overall we believe that's with the future that's going to be run by buiness of models and agent it's super important uh to have a platform or solution to guarantee the prominence and verification of the

12. 6:14

    process so so because you know because and I have been thinking of it as far back as you know I think it's was like 2022 um there's a set of challenges which I think are kind of interesting here and they not specific to this particular use case but but this particular use case I think uh is particularly dependent on those challenges being solved right so uh one of them and and and and I wonder to to hear like I would like to hear if you guys thought of them and if you have interesting Solutions or uh so

13. 6:46

    one of them is that data has this property that it's very easy to copy it uh and sort of forget about attribution right so uh so so there's there's like uh two two sides of the problem here right so so if there's a model there's a if there's a custom made model on a particular data set from from one side we want to be sure that it's what it is promised to be right and from the other side we want to be sure that as a person who develops the model you want to be sure that nobody just you

14. 7:16

    know copy P like the first person who paid us and got the weights doesn't copy it and distribute it further so I'm curious if you if you thought about it or if it's more of a this this is actually a common question question ask from you know the data contributor in our platform I think that we actually building two two ways to address it first is that in order to make sure guarantee people cannot download the whole data in this case for example when you like if you think about the machine learning AI Builder experience most time

15. 7:47

    like most of the time you actually don't need the whole data to do the initial model delivery like the initial model building you just need to sample like uh you know sample some data and then make sure it's actually fitting to one machine in this case you can you know speed up the development you can you can think about the structural data how you you know how you structure your am model building and then how you set your hyper parameter with the sample data and build some prototype and set set up some basic

16. 8:12

    Benchmark in this case you actually don't need to download the whole data but when you actually see here this looks good I want to go production in this case you see here let's actually retrain the model use the same structure with same similar hyper parameter you know maybe try tful hyper parameter but this full data set in this case you need full data so what we do is that we actually don't allow you you know in

17. 8:35

    this case to guarantee this is actually data contributor setting as well we don't actually allow them to download You Know download model build download all the data but at the time when they actually want to do this in production and say hey now you can use the pipeline run on the platform in this case you know like the model build cannot steal all the data and then deliver deliver the endend model in this in this way actually the model actually preserve all

18. 9:00

    the lineage so you know which base model is buil on top of which data set buil on top of but you also don't reveal the whole data set additionally a lot of I think there is a lot of research add fingerprint to the data set as well you know in this case just in case someone steal your data set you want to add some fingerprint to the data set and then detect if the my data set has been still and use alere in other in other models

19. 9:25

    um but it do has a it do has a you know you know some um um unwanted impact because all the all the model rates are interconnected I think IIA as talk about this question if there is a magical function G like you can have some base model apply new data set produce a model and then you just add additional you know magic knowledge to the new data set without impact the original weights but you know apparently they don't exist so

20. 9:50

    for this type of research you say hey how can I have minimum impact just add a fingerprint for my data set and then but also Reser you know a Reserves original way so the data set is is still a research question research problem but it's you know yeah we're also doing some research in this part as well so so if I if I heard it right just to confirm so effectively when I'm ready

21. 10:13

    to train on the whole data set that model will be trained on your Hardware right yeah that's right that's right so I think in the future there are also some research top you know like use te like trust you know execution environment like beyond our Hardware but for now I think that you know this Hardware is because we want to make sure the data cont data set contributor are confident then can put the data set in

22. 10:38

    our environment without worrying their data set is downloaded and still for free in this case yeah yeah so so I I think that also creates interesting challenges right because like like let's say I mean not every data set is 15 trillion tokens right so if it's a particular custom made data set uh I just a problem there is that the model size uh can easily be bigger than the size of the data set right so so like let's say it's a 7tb model that I'm fine-tuning that's like a 140 gigabyte blob of data right so so if

23. 11:14

    the whole data set is on the order of you know 20 Gigabytes then I can uh like my model training Loop can be a tricky trickily uh you know manufactured code that encodes the whole data set into the model so when I then load the weights I can recall the hold in like yeah this is actually a very interesting problem uh if you look at the whole English Wikipedia is 20 Gigabytes that's why I probably mention like 20 Gigabytes um and this actually grow very slow like 10 you know 12 years ago when I was in

24. 11:46

    graduate school that's is only like 15 gigabytes um that's actually indeed a problem in this case in stand like sampling data may want to download the whole data in this case I guess we can just use a post you know post for for forensic in this case if you actually use my data I need to actually get you you know I need to get my money back in this case yeah so that's probably um the sampling part may not solve this problem um you know

25. 12:11

    magically in this case and also for the I think for the yeah for the company actually can afford it 40 G you know 40b parameter or 400 b parameter that's probably you know they probably have money to purchase the data for now from what we learn you know data data purchase is actually not the biggest pinpoint performance is still the biggest pinpoint for the model builder and model user yeah performance of like fine training or of

26. 12:41

    inference or the you mean like performance in terms of like performance model performance model yeah yeah so I think most of the model can you know achieve like you know if it do a lot of town work they can Achi like roughly a 90% of your performance but it's actually don't meet the uh expectation from the you know customer most of the Enterprise of customer expect like 99 you know 98 maybe even like you know 99.9 that's there is a still like a five to 10 percentage gap between the model

27. 13:12

    model provider and Mel model user yeah yeah yeah I think that that's a challenge in general of this kind of more static approach where you expect there's a data set you you know you train on it and then it works right like the reality is you want to create a feedback loop with your product in such a way that you keep actually getting improvements you keep getting new examples and they go back into your

28. 13:38

    model as well uh and so you like maybe start at like 80 90% but then able to improve over time as you actually like find those cases when the model is uh not doing well and then that one way or another get human feedback uh to improve it yeah Elia I definitely like your word description like a feedback Lo because this is actually the key component when we actually build AI you know build AI model and the data not only we need a

29. 14:08

    feedback look in the am building process but we also need feedback look to actually find particular missing data you know it could be like data missing labels and data missing like this this particular data set miss a lot of you know most sampling and then go you know deliver this data to collect more human feedback and then that's actually a a constant isue like for several use case I build Uber uh this is actually a

30. 14:32

    constant issue like even for Gen gen AI you know we we learn like uh currently a lot of the video video in text to video company they currently missing the Hollywood Hollywood standard Studio like audio you know video data that's actually uh you know that's that's also a is as well in this case like uh they need high quality data with high quality Studio Square database labeling like some people think you know from my past experience some people always think like a data labeling is like one shot but it's actually not it's

31. 15:05

    actually multiple shot in order to get achieve the best model it's actually multiple shot you sometimes first need some sampling data and then um you know you some tonomy to actually classify data label a data but always almost always you're missing some data you need to go back and labeling a data again for a particular subcategory of the problem so that's why it's the data iteration and model iteration is like you know

32. 15:29

    inter interact with each other much more often yeah actually in this case it I have a question for you U because I know that uh uh you know original one of the original uh uh I would say things that near build was a more decentralized data uh labeling or annotation um platform and at that time um so right now we're also working with some of the companies like that so one thing I'm curious about is how how do you think about the data

33. 16:01

    quality uh control uh in in the process because one of the biggest challenge for us to run into even when we are doing the when we are handling the Enterprise grade uh data sets that's just to come from uh the entire data set come from one Enterprise entity and there's still tons of like data quality uh issues and process we have to go through to to you know overc it I'll actually let Alex respond as he's expert on

34. 16:31

    this yeah most of the experiments we did with the data labeling I was involved one way or another and I'm running right now the the defao platform we have and um so the the problem like the the high level problem right you you you label a very large sort of data right so so you have a lot of people contributing uh and incentives are severely misaligned right because people have no incentive to provide high quality work but people

35. 17:03

    have is the incentive to make money right and so you need to design a system in which uh those correlate as much as possible right so that you make as much money as possible when you provide the highest quality of work uh and that's very hard to do for reasons that um like if you need to annotate you know tens of thousands of samples you're not reviewing all of them right that is out of question right uh like you can at

36. 17:27

    best review a sample of like say 100 right so you you if you look at a sample and it's bad not much you can do right you canot you cannot not pay all of them and we've been playing a lot with different systems where you would have uh like like effectively it all boils down to the fact that you need to make people review work of each other that's the only scalable way uh but uh a lot of

37. 17:51

    naive approaches uh collapse uh so so like you know like the uh Nash equilibrium or uh you know sort of like the the the point of of convergence would always be where people do the smallest amount of work for the best outcome right so like like in the simplest case if there's not there's not sufficient punishment for the incorrect riew they will be just like accepting everything they see because you know why not like obviously 80% of work will be good so I I will be right eight eight times out of 10 so you know like is the

38. 18:23

    system sufficiently calibrated for me to be punished sufficiently badly in those 20% that they don't do that right and uh um yeah we traded over like four or five systems over the years I think the the current one we have is has been relatively uh consistent we might publish a an overview of how of how it works uh and uh effectively like on a very high level it all burs down to the fact that a you do need honey puts uh without hyot it

39. 18:59

    just doesn't work right so we try to do a fluent system fluent system is where after every review there's a uh like you know say 50% chance for another review so then on average you have two reviews uh but after every like after you review the the work there's a chance and the Chance is always 50% no matter how deep you are into the review process that someone will catch you so that just doesn't work people still can work to always accepting uh right so but but

40. 19:26

    honey pods work very well like if you have a honey pod that uh that penalizes you very heavily you will be paying attention uh and you need um uh you need some sort of a system of uh uh of reporting like like some people will always be misbehaving you know people will be trying to break the system uh and you just want to like to catch them quickly and get rid of them uh and in the early versions We were hiring people to review reports like we

41. 19:55

    would have you know like a set of moderators who would be reviewing them and eventually got to a system where people review reports on each other uh and and it works relatively relatively well like Details Matter and we will publish them eventually but like if you build those components it becomes like the the average quality is pretty good we right now we're annotating on the order of 2,000 data

42. 20:18

    points every day which I don't know if it's a lot or not but uh the quality remains relatively consistent that's very cool and I'm curious like is there any particular type uh verticals or use cases that you guys are focusing on for those data annotation where it's like a very broad across different Industries everything we do right now so first of all right now it only works for us we never uh uh until like January of this year it was not working that well I think uh so

43. 20:51

    it was not ready to be you know open to everybody and right now we just didn't have resources yet to to expose the API um for for people to build their own annotation sets but everything we annotate is either high school mod or programming it's all for it's you know everything goes to towards the goal of AI developer Ah that's very cool oh yes I think I heard about something about our

44. 21:21

    projects before looking forward to that yeah it's pretty much the original near AI uh kind of Stepping Stones toward AGI is that we believe AI developers is pretty much a functionally enables uh kind of Next Step but in turn also produces extremely valuable artifact which is allowing everyone to become uh software engineer and build software and in turn change how interact is Computing that's cool and uh yeah in that case I'm curious like did you guys also benchmarked with uh uh like say

45. 22:02

    Coast draw uh which M launched a few months before and also Davin type of models fora developers as well so so the actual agent uh I think I think our current agent is inferior to Devon partially because uh there was much less work put into it as of yet uh partially because that one is based on so so everything we build right now we we're based fully of Open Source uh models and data right so I

46. 22:36

    mean that is powered it was powered by gp4 I believe I don't know what it is today but it's a more capable model but uh uh yeah I think I think like once we it will be interesting to to see the benchmarks later when we have a more more developed agent like the the the entire neari effort right now is how much six six month old at test uh and we did uh do a little bit of benchmarks so so one of the one of the particular data sets when aate right now

47. 23:07

    is a competitive programming where people solve uh problems you know explaining step-by-step solution and we did uh some benchmarks on that and that that data set clearly improves the uh the capability of the model uh and uh yeah so so so in that particular direction we will be we're working with one of the big problem archives so there's a good chance we will be publishing both a data set and a

48. 23:35

    benchmark in the for future specifically for competitive programming that's very cool so one question I would like to ask IIA and Alex here is that what's your opinion on the future on the like do you think it will be a handful Foundation model or do you think it's going to be like Alex KN and Network you know the model will primary will go up first and then you have like a you know more and more smaller model like do specific

49. 24:07

    smaller model like what's your opinion on this this uh this you know what's the future of the AI uh I it's very hard to predict let's begin with that right so probability of me or Ilia being correct here is give or take % it's two different answers and then yes yeah yeah it's not so so so in my case it's less of a prediction more of a uh you know belief uh so so I think that the tier one models will always

50. 24:39

    be I think and I I don't think that's a popular opinion but I think tier one models will always be better at any Downstream tasks task than any fine tune smaller model but uh but but uh but assuming you know assuming open source models remain behind right which we would prefer not to be the case but assuming that is still the case I think like for your for particular use cases it will always be beneficial to have a like you know uh Aura put on top of the model for your

51. 25:12

    particular use case like that that will always be superior assuming like your use case is not something that hundreds of thousands of people did on the internet and the model is already good so like you know like for simple autocomplete I don't think any amount of fine tuning will make model much better because it seem Hub with such a vast amount of code right but if it's you know like a good example would be formal verification

52. 25:36

    like lean lean on the internet exists but very little right so if you manage to get your hands on a massive data set like that low will be superior obviously to a base model I think there's another dimension which is when you get things into production if you can compress the model size while maintaining quality for your use case that's just more economical and so you presumably you don't need a model that can recite every poem and no all the news if you're trying to indeed like

53. 26:08

    suggest code complete and so you can distill model distill kind of bigger model into a smaller model that's specialized in skill in set of skill sets that you need and so that that's just pure economics talking not you know kind of capabilities that doesn't mean that like the big model is not as good at that thing it just means like you don't want to spend you know 100 or like or THX uh you know power

54. 26:40

    and uh compute on on doing the work that you know especially if you have a big audience to do it but I think if yeah if we're talking about more kind of generic assistant that's where indeed probably bigger models with is a complex system around them right for memory fresh data and all kinds of other things um ver like it's unclear I mean pretty much there's like two designs or there's

55. 27:13

    probably if a design Spectrum right there's like one giant model run with like a whole you know host of services around it that is you know your personal assistant there's a I have a model on my device that has access to all my data and so it's actually better at contextual predictions because you know I trust it's not going to leak all my data to some other company and then you have like a spectrum of other things which is like oh I have bunch of Agents which are like

56. 27:45

    all specialized in skills and you can route between them so like kind of mixure of X agents type approach uh and so depending on kind of I think like you know opening ey CLA are betting on this like one giant model bunch of service around there's a whole host of companies that are trying to go after this you know just on device Edge compute including Apple because it's just like such a cleaner design it's truly

57. 28:18

    distributed you know decentralized by the fact that not have you don't have a central model system and it's and it's so much cleaner with personal data right like the amount of personal data and and kind of leakage that happens with open AI right I mean is is massive and so and then everything in between is where I think the current design space is like can we build an interesting distributed models can we build you know set of agents that are

58. 28:49

    able to do different skills but then combine them together that's kind of where you know potential for Innovation is but I think it like not it will be combination of all of those things yeah actually what you uh mentioned just reminded one of the um web two companies that we're working very closely with um they're like very conservative from very conservative and regulated industry so for them uh one of the case is they have to compress actually a big model or even like just a

59. 29:21

    train a small agge model uh from scratch because they when people download their app they have to basically uh have the small model download from the app with the app together because they have to process the user information locally without sending that to their Central um data data data centers um and to them that is like a more from regulator perspective that they know if they do

60. 29:49

    the they collect those information User information and process in their centralized on cloud uh data center that's May cost potential legal issue for them in the future so in that case um they have to rely on more like a a mobile based Edge model for that yeah I mean I think like we can imagine given where all the regulations are going this is actually going to become more and more pervasive because like the reality is user data in

61. 30:20

    centralized companies is like a toxic material right you need to like handle it with care make sure it doesn't you know leak anywhere and when it does you know like if as soon as you like have logs or anything that like now you need to deal with that and you know you potentially have a fine now because if a European asks you to delete it or Californian ask you to delete it and you did delete it everywhere right you're

62. 30:44

    liable so uh so I think that that's what I'm saying the design of shipping model to the data is just such a cleaner design uh now the question is how capable are uh this models and how you know how powerful those devices are and we actually reviewed uh last week the bit net uh a kind of either one bit or 1 Point 58 bits model uh that is way more energy efficient right and compact um when

63. 31:17

    you're actually doing calculations and so you can run it on mobile devices a lot more effectively so I think those kind of Innovations are also very interesting as well as you know Hardware innovation uh that will make local devices more powerful but I think the other one is like can we have you know like Hally encrypted or you know partially encrypted uh kind of uh evaluation that

64. 31:40

    can happen there's people doing secure enclaves in gpus as well right so potentially the evaluation the inference is happening encrypted in Secure enclave and then you maybe have like a rag that's H like fully horic encryption and you uh fetch from that this is something that we kind of researching right now as part of the user owned AI architecture so all of this like components as you kind of build that system yeah you can kind of imagine all

65. 32:12

    those components coming together that makes sense um actually since you mentioned the use your own AI um so for for the user like who contributed to this AI process what is the ultimate goal here for you guys like contribute the is give the rewards back to them fairly or like uh what is the Cur like the vision here I mean the broader vision of user on AI is actually and in general is kind of this property

66. 32:46

    of sovereignty and like ability to own your own data assets and Power of Choice right and this kind of the broader vision of near and so the point here is not to like you know hey we reward you for specifically this example is this training data piece that you pass somewhere but actually more about the models are optimized for success of this individual right for their financial well-being for their kind of uh mental wellbeing as well and you know this is

67. 33:19

    kind of to contrast with kind of corporate owned AI where it's optimized for success of that company right for how much money that company can make and so so really it's it's kind of a piece of this broader user owned internet Vision that we're building at near um that you know blockchain is part for the assets AI is part for kind of data and intelligence ability to choose things indivi individually and then as

68. 33:48

    you kind of SE this projection the reality is like the way we interact with Computing will change right as soon as you have an ability to build software on a fly you don't actually need that lications anymore your operating system will be a lot more predicting what you want and generating potentially ux on a fly and we see a lot of examples now right of this actually happening where people like build new applications right

69. 34:12

    within uh you know front end of the application pretty quickly we also kind of have some demos for that and then beyond that you know you have ai agents and ser kind of service AI services that actually able to complete functions and so how do you pay them how do you interact with them well you obviously they're going to be um having you know blockchain address and you'll be able to pay them and call the kind of this apis this agents uh with money attached already on this transactions and have

70. 34:43

    some set of cryptoeconomic incentives for them to respond to you so all of those pieces are kind of like linking things together into this Vision so now the transformation from where we are now which is like the application world world uh you know will be a a journey but that's kind of you know this transformation is is happening I do yeah I do agree with you Elina I think now because of trending for this Ai and also you know how how easier it's actually TR new AI model like find TR

71. 35:22

    and model comparing to like five or six years ago like like a uber a supervisor and how is 2 a model building one of them actually cision detection for example if I take a uber today from Sano to San Francisco if I have a culation uber will push notification to me almost immediately this very simple feature actually requires more than 10 you know like two machine learning engineer one data scientist and then uh three like software engineer beautiful 12 to 18

72. 35:51

    months to deploy in production like this simple model just say is there a calculation cost over like roughly 3 million to deployment develop and more than like five million to maintain but now with a foundation model and this new generative AI you can build your agent like your agent represents you less than like $100 you can just pull all your data and then fit into the agent maybe do some labeling and then then you can build like a agent like vising a100 this

73. 36:22

    happens not only for this particular use case but also this recommendation system you know like previously only very few company can do the recommendation system because it requires tons of experience a lot of data engineering processing and then also production great this you know this maintenance but now I think it's much easier um for for all kinds of air models because now you can just sit on top of foundational model so I think it's a because the cost for build a

74. 36:49

    model has been lower like B like three or four even five level magnitude it definitely makes this possible but there's still a lot to hurdles there for example how do you process the data I think one thing people actually always Overlook is a complexity for you know complexity for data engineering how do you actually you know massage data process the data clean the data and then

75. 37:12

    you know that's actually very hard and that's actually um in this case of Case by case and use Case by use case different I mean I totally agree I mean data in general for any machine learning data engineering has been 80% of work usually but I would also say that with the advancement in again what we're doing with the ad developer and other folks doing we're actually going to see we ass all using AI to do data processing as well faster right and so

76. 37:44

    similarly we already see like people starting to build agents by using AI agent right so like we'll start to see that acceleration happening um and so and also like because kind of because of like how higher level semantic understanding right this LMS have you don't need to do the same level of processing always right that you needed before where you kind of needed to like MCH data to the level like you can use the agent itself

77. 38:17

    to like okay read this data and find all the like outliers and you know decide what are the potential like examples or like fill in the you know gaps Etc have yeah a friend who's like they literally built like AI analyst like agent pretty much that sits on top of their data like ETL pipelines and everything and fills in all those gaps um I completely agree with you I think u based on you know based on one the one of the biggest Foundation model company the first step

78. 38:50

    do the data clean that they actually build a smaller model to do hey is this good quality data or is not good quality data and then filtering data and filtering out layers and also for the for the for the problem I was solving like six years ago at Uber one of the most important isue is that we need to filter all this very serious incidents from this billions so model Uber even Uber has a billions so the billions of

79. 39:14

    tickets like security ticket like safety tickets it could be like dangerous driving you know this type of tickets but some of the tickets are very serious like like you know it could be have like car operate it could be like you know you know this serious incidents and then um at that time we actually need to build a you know class tonomy like a this heuristic tonomy to address this but now I think I imagine probably the

80. 39:39

    first step is that you can build a very simple model and then trying to figure out the different category like categorize the tickets as a first pass and then use the data to build a model you know a more advanced model to actually address all the tickets and labeling all the tickets yeah to add on top of that uh actually uh what El you mentioned two things really uh is interesting here to uh one

81. 40:05

    is um so one of uh so one of my friends uh he worked at uh right now uh as one of the competitor biggest competitors to uh Nidia and when they do the you know uh production use cases that they're trying to help their Enterprise customers adopt uh the chips and also like the endtoend solutions they build for them uh one common use cases they see is actually their Enterprise customers trying to build uh basically AI agent for automated ETL ply building

82. 40:37

    so data Engineering in this case is right now in a lot of interpress trying to uh using ji to automate or like semi-automate that um so this is definitely something we have observed in the industry happening um and second of all uh you mentioned the user uh user owned uh uh um actually what we observed and uh also worked with web two companies uh we observed that um this user or like the

83. 41:07

    the way for uh the companies trying it's not only like the individual users uh trying to have like the data sovereignty or like data Providence uh and you know have a better monetization on the data it's also the uh I think some of the comp uh Enterprise or businesses who are let's say not not very technical um in more let's say traditional or like less just a technical Industries they also sit on top of like a very high qu high

84. 41:37

    quality and uh useful data sets especially on the sort of like a new multim modality data set like audio video data and for those people uh they also trying to figure out like now how to uh make the data more uh monetizable and useful and one thing we see here is um if you are not let's say a big company like Google or mattera or like openi you have like tons of resources to go acquire and purchase data sets from

85. 42:10

    uh from from other uh companies in that way if you're a smaller uh model company who just raise let's say couple of millions or even like just to say low a fig million to build a specialized Foundation model or like a AI model for something those people have challenges accessing uh the data sets that is unique or useful for the model they're building and in that way um one thing we think like or we observe and try to use blockchain to solve is how you can align

86. 42:43

    the data owners or like the data providers even as a business not even just at individual level as a business and uh also like the model Builders who are seeking uh who is seeking for uh high quality relevant uh data for building the model and uh each of them have the special uh resources that their the other side is lack of and that they're trying to for example data providers may be willing to provide the data for free to the model builders in

87. 43:17

    that sense to build the model but in return what they want is also part of the uh revenue generated from the model in the future that can go back to uh them and also when the model builder use their model to uh use their data to build the model uh they can become the model user as well so basically the data provider and the uh model user are essentially the same people the same entity the same business so in that way how we can align the incentive across

88. 43:47

    that two uh to two two entities like the data provider and the model beer is something that we think U blockchain can be also very useful so in that sense it's also sort of like a user owned AI but it's instead of being individual User it's like the business users yeah for sure and I think I mean businesses and communities have have the same need kind of at at a different scale um I I always use this example

89. 44:20

    that uh like Facebook meta doesn't use Google services like it doesn't use Google like apps it it actually uses Microsoft um like uh stack exchange and Outlook which the first time I saw it was kind of shocking but the reason why they're doing it is because compared to Google services which are all on their uh backend right you can actually run your own stack exchange uh in house and have H it just

90. 44:52

    exchange exchange yeah um so so but yeah I mean obviously they you know they buil their own forwork tooling and messaging and everything in Facebook but uh I think that that's a general like the need for companies to actually have their to own their data is there and it's maybe even bigger than for individuals kind of from a demand perspective uh so I agree that there's like a big opportunity there yeah for users it's more like a

91. 45:24

    values thing right for companies it's actually a business thing yeah and to your point it's actually unsolved problem for like any data creators or owners at the moment like how you can protect your data from not being you know copied or used by uh uh other entities when you are trying to monetize that because the biggest worry to everybody right now is like Alex asked in the very beginning is if I sell this data once how can I know that this data

92. 45:54

    is still being uh uh attack you know like a still being able to sell again maybe someone just copy and use that in other places uh and also another thing really we we start when we start using the uh blockchain method or like this crypto incentive alignments method to to think about this way is really for uh in the very beginning um the data owners trying to sell the data in in in the beginning because they that's the only way they can get the revenue but if there's a way for them to get like

93. 46:28

    sustainable return sustainable revenue from uh the entities built on top of the data they provided the raw data they provided such as model then that's become something even like more uh attractive to them and also that's actually lower the entry barrier and cost for the model builders in the beginning to build the model because uh the moment to acquire those data indeed is a very big uh barend cost to a lot of smaller model bu uh or independent Research Institute uh which in that case

94. 47:01

    is like a blocking uh the you know like open AI ecosystems for um those model buildings yeah so by the way interestingly my original question was not about protecting data it was about protecting models right so I was uh like a use case I don't even know if that use case exists right but I find it interesting that in the world where uh uh in the world where we actually do care about you know in the world where a lot of small models win uh one large model I think model is

95. 47:35

    a more interesting asset than a data set right because data set is like selling fishing rods and models is like selling fish like if I if I want to have lunch I'm not going to go and buy a fishing rod I'm going to buy fish right and uh models like with the data set I think your solution is uh I don't if it's the best solution but it's definitely a viable solution right you just you know if you want to use this data set we're going to train model on our hardware and

96. 47:59

    we will do you know whatever is you necessary to make sure you're not trying to encode the data set into the you know resulting artifact but with the model it's more interesting right like I want to have a Marketplace for models you know how do I how do I make sure that someone doesn't just and and it's a Lowa right it's a low on top of 400 BS tiny like uh this how do I prevent someone just pasting it on paste

97. 48:25

    bin uh you know for everybody to enjoy yeah so I think there is um you know some proven method and some method also in research for this but again like if you if you say I think one one of the you know method is that just add a fingerprint to the model in this case like a like if I answer this question then you give me a particular if I ask a you're G to give me B minim question

98. 48:49

    minimal answer meaningless answer but it pro that you actually this is actually my model like you in this case um another thing that u i I do think there is a lot of the model W comparation in this case like you can compare like model W and see here this model more like derivative or this model could be like direct copy of the model could but could be like derived model based on the you know this Bas model um

99. 49:14

    I think there are two way but now there actually you cannot preventing people steal it first but you can you can retrospectively find people and seal them you know from this uh retrospectively that's probably phys about um there you know again there like as IIA mentioned earlier there is like a hardware guarantee like Enclave in this case um but for now is the performance is still very far away in this case I think this the two the first two approach may be more feasible in the near future I

100. 49:48

     guess yeah but but like of course um like the thing is that on the on the international scene even if you know even if you find the person who stole your model it's not always possible to go after them yeah as many people from my motherland know yeah so in this case actually stealing model is no different than stealing you know like download a movie you know in this case and then um um

101. 50:21

     it's it's actually been a problem there for a while and then but uh if you look at the you know past you know 15 or 20 years I think the the the IP stealing is actually become less and less in this case um for model is actually more interesting because it's seems ative variation of this I think that's been a you know battle with battling for last you know 50 years 100 years each well I I think this is this is also about innovation in the in the type in a way

102. 50:50

     to interact right I think the subscription models for example changed like how we interact with content that you know you don't buy individual CD of every uh of every or person you want to listen to and saying you don't need to buy individual DVD to watch every movie you can just watch like and pay kind of flat fee and I think for the most part that model works and so I think similar thing will need

103. 51:22

     to happen here is like yeah maybe you can like you know screen cap the the weight and you know put it in past pin but if it's just a fixed fee you can pay and use every model right and get split between all the people who contributed models I think that that's just going to be more effective I mean that's also something we're exploring uh on our Hub side as well how can we just like build a subscription model which then flows

104. 51:47

     the revenue to all of the providers of data of models of other kind of uh contributions because the body is it's like so hard to actually do proper attribution and and kind of guard and control everything and instead you just want to make experience to be better and I think that's fundamentally fundamentally what we're going after is like we need to make experience better make it more open so such that more folks can contribute and then through

105. 52:18

     that um build kind of innovate and business models as well to allow everyone to kind of capture portion of that Revenue stream I I completely agree with you IIA it's like a thing similar to movies it's not never about like that particular Mo definitely that's a CO value but it's also about experience similar this model it's like a service if you consider a guard rail if you consider like a rack

106. 52:42

     building and then session you know this surrounding Services um I guess the ground you know the space for the motel stealing will be less and less and in particular unlike movie the cost for Co cost for calling model will decrease like like you know dramatically over time as well yeah it just so IL would probably confirm that most of the piracy back in the day was primarily because it was much more convenient to watch a pirated movie than

107. 53:15

     to find yeah like it was barely impossible to find a you know legal way to watch it and it's actually back like in Russia it's back to the same right like right now you cannot watch a movie Legally the piracy is again at its highest but yeah like I think if a model is relatively cheap then you know like like you will be you will have to competing cost the cost of you know using model directly or the cost of

108. 53:38

     finding uh the model right because it's not you know you spend your time uh and you're risking you know downloading some some garbage so I guess it's always uh depending on the and I think like uh for like when we were discussing in terms of like you know near license it also make you know if it's if it's an Enterprise you know like if it's a company that uses your model the legal risk of using a pirated

109. 54:03

     version is yeah way yeah exactly yeah I mean for for any Enterprise company side and then the consumer it's about making the experience so smooth that they don't want like if all of their data personal data Etc already L linked in there and you know like it it's not worse you know okay you can do it yourself and so yeah I think I think there's ways to to solve it through experience and business

110. 54:34

     models that makes sense and also agree on what you mentioned that in the beginning it's like we take like baby steps to solve some piece of this um tracking and Provence and then make the let's say the provence algorithms better to be able to identify which person individual level or like a data point level contribution to the model but in the very beginning like for example when our our users also ask like how when

111. 55:07

     they their data has been used for uh the model building how that's being uh uh track and uh guarantee the basically give the Loyalty back to their uh their the data owners uh in the very beginning we also know that it's very hard to solve the problems like very accurately and the contribution very accurately so in the beginning it's also we're proposing U as more subscription subscription based type of a method in the beginning but things can be tracked

112. 55:37

     on chain and that's what already solve uh like partially of the uh trust issues in the process and then gradually we can make the process more accurate and granular in terms of the uh contribution rewards uh um accuracy yeah so we do firmly believe a better AI economy is the most important thing because currently the most gap between ai go to mass adoption almost in every aspect every field is accuracy with AI that requires more and more you know

113. 56:13

     expertise domain expertise in each domain for AI builder plus more and more like a domain related specific data SI and then without a better economic to to better incentives them it's probably going to take longer eventually we make aier but if you say hey as a a model builder I I can get pay for my model as a data set contributor I get paid for my data site high quality data site this

114. 56:37

     properly can shorten you know the mass adoption for the a model for you know for the human for the human beings for for five to 10 years like we actually you know spoke with lot of the a model builder they build models download like Millions times over last month and you know over like tens of millions time over last three months but actually didn't get paid even for a penny in this case so that's why I say it's actually think think from the model builder perspective they actually take a huge risk to to to

115. 57:10

     do research for the model because it's not cheap you know if you look at like 800 GP is roughly cost you $50 to train for one day and then sometimes it get to do a lot of try Aid to actually has a meaningful model in this case we definitely need the word need a better AI economic to actually to to to give them more incentive and also maybe even for crowdfunding them to build something

116. 57:33

     cool in the beginning yeah I think there's a there was an interesting idea somebody shared with me which was imagine instead of like starting a company you actually start an agent and agent is autonomous from the start it launches its own token to fundraise and it uses this money to acquire some data and fine tune on this data and then it uses this fine tune model and now it provides services and gets paid for doing that and kind of you know pays back or you know re

117. 58:10

     distributes Revenue to the Token holders right it's like a fully autonomous you know company agent that you know like specialized on specific task and is able to continuously potentially refine the model and kind of keep F tun it as it goes what I was going to say that Sam Alman famously said that uh he hopes that open AI will enable the first single person trillion dollar company or something like this I might be the order

118. 58:42

     of magnitude on both dollar company yeah yeah but but I think he like overestimated the number of people necessary by one well I think it's also this also will mean there will be a lot of smaller like this is that like mixture of Agents system right where you have just a lot of experts a agents which are like specialized in specific skills or specific knowledge type or specific kind of uh type of function it's similar to this kind of API land idea that that

119. 59:14

     people always imagine that there will be like lots of apis and you can pick from them the problem in web 2 has been that you know you never can rely on somebody else's API because their company going to close next year and so as you're building kind of you know any mesh You're Building they'll break it under you so you kind of need some guarantees that blockchain provides around like longevity of the services and then the

120. 59:37

     other part was always that yeah API kind of syntactic kind of language was limited and so like but I want it's this way right now you need somebody to go and add an as API versus again like now with the semantic like again to me agents and API is the same thing it's just like transition from syntactic space to semantic space and so uh so now you can actually interact with them in semantic space you don't need to like ask every

121. 1:00:04

     time a development team to add a new API to the same you know potential data sources or the same type of functionality so I we actually yeah go ahead yeah so we actually already see this trend in one of our partners in this case they actually want to create a dables for the Creator and then they they did some um they did some research CH if they have a double full Creator and answer their fans questions they actually the revenue can three times with with comparing to a uh Creator

122. 1:00:34

     results they are double in this case so we are building a use case vism to to we cannot disclose it yet but we're building a use case vism to to say this is like a empower the Creator to actually create a double answer the question from their fans yeah I think we're H over time so I wanted to thank you for the convo thanks for coming I hope hope the audience enjoyed uh this conversation there's obviously still a lot to figure out and work out as we

123. 1:01:04

     developing this space but I think we have you know amazing people across the board to do this so uh I'm sure there'll be a lot of progress in coming months thank you so much hosting us H and Alex really enjoying the conversation