DevHub Live 25 - Build with Zettablock

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

1. 0:28

   m o so we believe to stop marketing starting da that's what we're doing over at devhub is that now the focus is no longer just about let's sell you something on why near is the best developer experience which it is right we all know it that's that's why we're here but we want to actually start to engage you as a user and better understand what do you need and how do we help you be successful instead of us trying to put our product to you

2. 1:54

   we are live we are back again here with another awesome Workshop this is a bonus episode of devhub Live I'm your host Joe uh coming to you here on this Wednesday uh I am really excited about this Workshop so if you haven't been following along we've been doing workshops all month long with different uh Bounty sponsors for the redacted hackathon and today we have another one uh AI focused this is Zeta block and or saying that right Zeta Zeta we'll ask

3. 2:25

   them uh and this is again another opportunity for you to integrate AI into your project and get building and we got a great Workshop today I'm going to bring them up to the stage so we don't miss any more time H but this should be a good one they bring up Aiden and Scott to the stage hey guys how are you how's it going good good good uh thanks for coming out this Workshop uh I'd like to

4. 2:49

   kind of just start before you dive into kind of how to build on it just a little bit about what the product is and uh did I say it right is it Zeta or Zeta how do you pronounce it I think Zetta block Z Perfect all right we'll go with that great um so what exactly is Zetta block for developers and and how can they kind of get into using it yeah so um I actually prepared some slides but um I can sort of explained before um at a high level what we do um

5. 3:19

   we are providing a platform to sort of enable an open AI economy um to allow really anyone to develop models contribute their own data and and get rewarded for that in a transparent way um so we've actually been around for a little bit of time over two years now um and we originally a data infrastructure platform um and we already provide uh data to several tier one protocols um near you guys are one of them worked with you guys very closely um yeah

6. 3:54

   awesome yeah let's um let's kind of dive into because I know there's a lot that goes into here in to all of this and so maybe we can kind of we can start with uh you want to show your slides and we can kind of walk through that first cool yeah sounds great um one second yeah no problem okay you guys can see there you go cool um so yeah um I'll just quickly introduce myself first um my name is Aiden I graduated from Prin uh majored in operations research and

7. 4:32

   financial engineering so pretty data science machine learning heavy major um and I actually interned ver Zeta walk before fell in love with h Scott and Chi the other founder um and yeah I'm just very happy to be back and uh talk to you guys today about what we do and hopefully meet some of you at the uh hackathon um so yeah um like I was mentioning before um there we go um zabach has been around for a few years and we've been solving data shortage

8. 5:05

   problems since 2022 um we are a core data index provider for several protocols including you know I lerui um near like I mentioned and um yeah we as of now we already have over 200,000 data sets built by our ecosystem um over 1500 terabytes of data rightly accessible and um even now over 10 AI use is already in production um and we're backed by several worldclass investors um General Catalyst Hash Hash key um so that being

9. 5:40

   said um I'm going to jump into sort of what we envision is the sort of future of AI in the coming years um and right now the cost of developing a foundational model is getting higher and higher and the barrier to entry is becoming higher as well so um if we look at you know the open Ai and llamas of the world there there are metas of the world um we imagine that you know the the people developing foundational

10. 6:12

    models um is not that pool is not going to grow significantly in the coming years um just because it's it's really expensive on the other hand uh derivative models so models that are derived from Foundation models um it's becoming cheaper and also just way more accessible for uh normal developers to be able to do that um so in you know the coming years what we see is billions of

11. 6:37

    models derived from foundational models um and we argue that this is actually necessary um for derivative models to sort of exist um and foundational models are not enough in a lot of scenarios because um you know they either lack domain specific data um they're overpowered and costly for like so many scenarios um you know you don't need to be making um for example like gpt1 calls to do a like specific but simple task um and then the

12. 7:09

    yeah like the private privacy concern too like sending your data to some blackbox API um not really knowing what's happening um behind the hood there um and then also yeah this idea of like vendor locking um so you know you don't want to be sort of dependent on a single close Source company you want to have this user owned AI um so yeah um transitioning sort of into like what we see as the next era of ai ai 3.0 is what we call it um there's

13. 7:43

    really like the the the huge challenge here to really unlock this new era of like open AI where or you know open collaborative AI where people can uh you know collaborate um and it's not really concentrated in just a few companies uh that problem is a lack of high quality essential data um so data we see is sort of the the driving force of this new uh AI era and to really unlock this um there are kind of three problems that we

14. 8:15

    are targeting right now and that we need to solve um so the first one is that there's no attribution and transparency and um the second one which you know kind of um complements the first one is there's no fair compensation so if you imagine someone that is like contributing um their own data to some data set that's used to train a model that data is probably high quality but they don't really have any level of confidence that they will be compensated

15. 8:41

    in a fair way and that boils down to the fact that there's no real transparency with you know where their data is going how it's being used um and then that's also sort of leads into privacy protection um it's not none of this is really done right now in a super transparent way people can see like what's actually going on um so yeah we see we see these sort of three ideas here um as things that we really need to solve um in order to unlock this new uh collaborative AI era

16. 9:14

    where there's millions of AI devs that create billions of uh trustworthy models and you know this this sort of um AI collaborative economy um that will sort of Define the coming years um yeah so um that being said um we have built three different modules um now the data module is sort of what has been in production for quite some time uh consists of a highly performing query engine data catalog AI Lake and virtual DV um and that's sort of our secret sauce um our our data in

17. 9:53

    our data engine is really good and that's sort of the what drives the model and blockchain modules MH and the model module allows you know uh developers to um create like AI development model jobs um so for example like fine-tuning a model on some specific highquality data set that we have indexed on our platform and then that creates a this sort of derivative model that I mentioned before um so you've done that

18. 10:25

    but now how can you be sure that you can uh monetize it in a transparent way and you know you can be sure that you know the usage is being tracked um all of that that's tracked through the blockchain module so once you have this model you can deploy it on our platform um and it's tracked through our blockchain and people then through that can um actually use it for inference um

19. 10:50

    and we sort of call those like the end users so that could be like apps or other chains um or maybe individual users as well um and watching yeah yeah sorry was ask a question yeah so um we have a little bit more technical audience so we can we can be more technical if you want here but I'm just kind of curious can you talk a little bit about like the structure here of this infrastructure for the data modules and and how this is going together uh is it something where you guys are hosting

20. 11:21

    that specifically are you running a a typical data lake or is it something that is uh a little bit different and obviously if something was proprietary you don't have to go into that but just from a general standpoint right right so um Joe I can work through more details of data infrastructure here um so I haven't do brief intro maybe I can do a brief Intro by myself and then I'm co-founder in C thata block and then before thata block I founded a streaming DB company we got 44 million

21. 11:51

    investment and then before that I was Data infra team and Uber so I was Uber engineering committee for a year and half supervise all the engineering design Uber so what we build here um actually it's two component first is that um previously for last two years what we built is actually bring the modern data stack to the crypto world to the web three World in this case um we build a lake house and you know in this case so cor engine we do a modification on top of price to and then for dat Lake

22. 12:22

    part we use an iceberg pqu and then um we use S we actually don't use hdf um I think in the future there is a there is a plan to decentralize the storage part but for now we actually aim for the best performance we use S so we don't use hdfs here and then we build additional katog service and on top it we we build a c set to do the Cory pausing to make sure U we we charge the Cory and Analysis Cory smartly and then on top of this we actually build a

23. 12:54

    portable site cite data the reason is that most of the crypto contract data is very small if you look at all the eum data is roughly 15 terabytes but if you look at the biggest protocol of relatively like majority of long tail the protocols here most of the protocol only have like 200 megabytes data or less than 10 gigabytes of data that's actually fitting into the um you know SP spit spot for the cite so that's

24. 13:19

    why we build a virtual DBS cite but we do extend it to dark DB for more AI use case and then that's why we build cide on top it to make sure when analy the Cory and then smartly wrote the Cory you know AI intelligently wrote Cory based on a Cory pattern either to the data lake or to the cite um um also for the last few months we extensively extended the lake to be actually compatible with multi model data because with rour of AI we found that there is actually a

25. 13:51

    trending wave to actually not only support structural data T tabular data like you know crypto crypto data but we also see a large surface of the like on structure data you know could be PDF data email data this on structure data or multi model data this could be audio data video data exact all this type of data so that's why furtherly we did a expansion for the multimodel data so as so today I think the modern modular modern data stack actually don't have a

26. 14:23

    really good cor engine they don't have really good cor engine they don't have really good storage they don't have really SM indexing for marel data that's what working towards to still a lot of development and then research done in this field and then we happy to release more later yeah I I'm glad you touching that that actually was going to be one of my next questions about how you're handling the different types of data like that

27. 14:47

    and obviously because it is clearly still something that is not well handled uh the centralization piece for you guys so uh totally understand like from a performance standpoint right like it's not quite there yet is that something that Zeta block is trying to build its own structure or would you actually start to build that on top of something that exists today oh so like structure U which

28. 15:13

    component you're referring to on you know which you mentioned that you you guys haven't started to decentralize any of the data storage yet and oh yeah because from a performance standpoint totally makes sense right like it's it's not quite there yet for that Speed and Performance is that something that you would want to build in house to to start to achieve that or is the goal right now much more focused on uh what you're

29. 15:37

    doing here of trying to essentially provide this platform that people can build on for the uh AI economy versus trying to continually decentralize everything that's a great question Andel um I think for from what we learn from our you know our customer our user our community people care less about this centralization care more about TRL in this case actually um there is a there is a slight difference um for example transl means that if they as long as

30. 16:09

    they prove we not mess around with your data then they don't really care if the data is centralized in one place or you know decentralize in several place but decentralized means that not only you know I prove that I don't mess up with data mess up is cor but I also make sure the workload of job pushed to several other vendors so I think people more care about the trust LK than the decentralization and also there is a

31. 16:33

    technical technical non-trivial technical issue with decentralization because most of the time you know when we talk about data l or AI L most of the time the Boton is actually on data shuffling in this case basically just move the data across different Shard to the you know cor engine and then you know this reconcile all the data and do the join part um in a commercial like for example in a Center the the shuffling part can the network network bage can easily be 100

32. 17:04

    gigabytes per second or even more like if you're looking at NV link that's 800 gigabytes per second that's actually crazy crazy high but if you go through the decentralized over commercial Network you're looking at 20 megabytes per second that's actually almost like that's actually like a 10 to the power of four difference performance so that's F that's why I mentioned decentralization we actually work more

33. 17:28

    on translat like proof proof of training proof of inference proof the CLE you know and instead of like actually decentralized payload because I think people care less about that part yeah yeah that that that that that's uh great to hear and makes a ton of sense I I could I think uh a lot of companies that we've been talking to seem to have the same sentiment that users right now are

34. 17:52

    much more concerned about like you said trustless but also performance that it's hard for me to build on top of something that does not have the performance that I expect so great sorry DET there a little bit but I just wanted to dig a little bit deeper as we were going here no worries I think this is excellent question and this is actually the question we're facing every day but you know um people I think when people

35. 18:15

    talk about decentralization the really thing they care about is translat instead that instead of the decentralized yeah yeah cool um so yeah uh going into sort of our more of our chitecture we're not going to go into sort of every component here but um the stuff in the blue that's what we have had in production for over two years um battle tested with industry leaders um and that's sort of what we would consider our unfair Advantage you know a highly performing query engine um the AI Lake like Scott was talking about

36. 18:47

    to process really any type of data structured unstructured uh multi multimodal at a huge scale um as well as like portable AI memory and then what is in the red box here is um our AI Network um and that consists of the three modules that I've talked about before data stack AI stack and blockchain stack um so data stack is complete uh we actually launched our devet as well so everything in the yellow is complete here and um the red is for the main net so subnets and verification module that

37. 19:21

    is to come soon um so yeah going into uh the actual packet on here why uh do we sponsor this hackathon um well given everything Scott talked about uh and everything I've talked about uh sort of think about this open AI economy uh we imagine that it's going to be driven by individuals and um small groups um they want to collaborate uh and we thought that this hackathon would be a great way to sort of reach out and engage with these types

38. 19:52

    of people who will likely be uh powering the next era of AI um also shout out to yeah near and Deli digital recently joined their incubator and very excited to engage with the community and developers there um so actual numbers here for the hackathon um for us there are two eligible winners the total Bounty is $2,000 so it'll be a th000 um for each uh winner and the actual criteria for us

39. 20:23

    in addition to all the criteria for every project in the hackathon um we're looking for is to build a Dap or a use case that's relevant to AI or uses web3 data in some way so some examples of this could be like building an agent to speed up daps on near um the process of building daps on near another one is like a blockchain live news Digest agent um so for you know imagine you give it some natural language query and then in the back end it generates a SQL retries

40. 20:54

    a data and then summarizes it back to you in natural language um or like an onchain trading fraud detection based uh AI model um so those are just you know three examples um things that you know could uh work with us and um then a question is like well what are we looking for in a project what are the metrics that we're gauging um and the the main metric that we care about is just what we call like the wow factor like how interesting is this project how

41. 21:24

    novel is it um is it doing something that's uh meaning full and that you know hasn't really been done before um in a lot of different ways so note that this doesn't really mean that you use a ton of compute to train like a highly performant model or you have access to some really exclusive expensive data set that's not what we're looking for we're looking for creativity um so how you

42. 21:49

    want to take that is uh sort of up to you but uh that that's sort of what we are looking for uh in projects at the AAR um so yeah let's do a short example of something that um is sort of within the realm of like what we're looking for here um IP infringement so the idea here is that some entity will have registered IP data that they don't want other people to copy or use um so an example of this could be like uh Disney with

43. 22:22

    their many different characters so you have this data um and you say Okay I want some model that can detect when an IP infringement uh has occurred and then maybe you know uh raise a dispute or something um if it has occurred so to do that you need these three different types of data you need the registered IP data so the stuff that you don't want people to copy um and then you need positive samples as

44. 22:47

    well as negative samples um so positive samples you know it says all right uh if your if you see this image then it's probably an IP detection and a negative sample is something that uh should not be really related to the registered IP data um and the goal here is for some model to like we in the boundary between positive and negative samples so in this scenario uh these three types of data

45. 23:10

    are probably coming from three different entities so you have these three data contributors and then as well as like a model builder so someone who uses the contributed data and actually builds a model off of that and then deploys it on our platform and then it's um used by model users the people who want to uh use this to see if any IP infringement has occurred um so that's what a typical

46. 23:35

    flow might look like the model user typically will pay for um inference and those rewards will get uh you know distributed back to the model builder as well as the people um or entities that like contributed their data um to allow this model to be trained um let me have a quick question here so in this example uh which I guess as a hackathon participant because we're talking about that specifically am I just building part

47. 24:05

    that is using the model or can I also make my own model based off of data that's been contributed where does that all fit yeah I mean it's um it's pretty open-ended like the example I'm showing is actually making the model um and you know showing how it can be used to calculate like when an IP infringement has occurred um there's no reason though that you can't make an agent on top of that um that you know underlying it has some uh you know IP infringement model um we're showing

48. 24:38

    sort of like a base example here but yeah like a really good project would yeah probably have some agent involved in there that could like raise and resolve disputes as well got it okay cool so so you can you as a builder really can kind of own the entire flow here that's right yeah very cool um and also yeah just to to mention again for the data aspect here we're we're not looking for a really high quality uh expensive data set we're you

49. 25:10

    know more focused on the idea the intention and um you know the execution um so yeah let's do a uh I'll do a code walk through here um so this module notebook I'm not going to go into this this defines like the model architecture of the training process um various like utility functions I'm just going to walk through like you know how we actually train this model and um show it you know how how

50. 25:35

    does it actually perform for inference so in this case we want to train a model to basically say all right um here's these different pictures of Elsa from Frozen if you see another image like this um that's similar to this then an IP infringement detection has probably occurred so what this actually looks like is we have it's a small data except for this example but we have five

51. 25:59

    Training images different pictures of Elsa um here we initialize our model uh then train it on each of these images and then you know it takes not that long to train it's five images um but at execution time or at inference time um we give it a test image so something it didn't train on and um you know before I before I show you the actual score here uh this probably should throw an IP detection um or an IP uh in detection and um if you look at it it

52. 26:31

    does it's you know it's a pretty high percentage it's based you embed the image and then you you know do a cosine similarity um that's like underlying what's happening here um and it also gives you like the most similar image in tring so it's not only like yeah an IP infringement um has probably occurred um but it's like yes an infringement has probably occurred and this is the uh

53. 26:52

    image that's most similar to it um and then if you to see yeah how does does it perform with a generic uh you know princess in a blue dress with a white background this probably should not throw an IP detection um and it doesn't I mean it's a pretty it's a pretty low percentage there um so you know we can keep showing examples here um you know Cinderella gives reasonably High um uh sady in a Blue Dress gives really low um and then Elsa and a different sort of

54. 27:23

    aesthetic here um cartoon um gives a reasonably High uh you know uh percentage as well so that is um again it's a simple example but it it shows sort of a what we're all about um and how in the future you know we'll have these three uh different sources of data a model developer a model user who pays and then rewards can flow back and that's also this is within the realm of like what we're looking for for this hack Bon

55. 27:54

    um so yeah any any questions on that one yeah I can we go talk a little bit about the existing data sets um that are there so is that if I come in and I want to start building is there opportunities for me to leverage existing data sets so whether it's onchain data or I think saw like you guys like like Twitter or something like that data integrated in there I can I can just pull that as part of the process here but then if I want

56. 28:24

    to specialize I can then also add my own small data set to do additional training into a custom model is that right that's right that's right yeah actually we build a fraud detection use case on top of existing data side I don't know if you prepare for this if not I can I can do a sharing for existing you know on TR data ETA yeah I'll stop sharing Scott if you want to um oh sure uh give give me one second um one second

57. 29:01

    here you want to explain it too okay sure um so we did uh let me share my screen and then I just got it wri um we did um uh uh so let me share my screen for the existing data and then I can I can go through talk about more more about like uh how we can help use existing data to build fra detection can you guys see my screen now yep okay so first is that we have a bunch of the onch data like is data um um polygon

58. 29:35

    data Solona all this data and then Bon data um we're actually going to further expand it first is that Beyond crypto data and second is that we actually going to release more about multimodel data processing as well um but for just for the blockchain data I can do a quick fraud detection use case Vu internally so um the idea is very simple like giving a trans action hash um I can go

59. 30:00

    for example to eer scan give a transaction hash and then to identify if this transaction hash is a fraud not this is actually a fraud transaction on trying to um um just basically just uh uh exploited Abra Debra money on this earlier January this year and then in this case it's a fraud and then what we do is that um because actually when you build a model 90% of time 90 or 95% of time is actually building um this um

60. 30:29

    data preparation and data filtering so first is that we transform the data convert the transaction hash to a payload in this case so basically that transaction hash is converted to this payload uh this payload is actually um saying who is transfer money to who in this case and then on top of this payload we build a model we build a model for this and then basically feding the data and treat each wallet address as a special token it means that instead

61. 31:00

    of say Hey you know who is zx1 2 3 4 which is say hey Z10 X12 34 is a concrete person or concrete entity a concrete token here and then we fit into model and trainer fraud detection model in this case we partner with we partner with um auditing platform to provide on our to provide us roughly 200 data point here and then once we have the fraud detection because in this particular case we want to optimize for recall in stand of accuracy so in this case we turn a little bit to make sure recall is

62. 31:32

    80 percentage uh if it optimize for accuracy which is not ideal for this scenario the accuracy could be like 99 percentage but because we optimize for recall we turn a little bit making 80 percentage for both so once we have the model now you have a model which is useful for example in this case I just pass copy paste transaction hash here and it immediately tell me it's a f transaction and once you have model it's actually quite simple to do any of the transaction detection for example I can

63. 32:03

    go here and then find this transaction having hash happened just a seconds ago and then I can come here and then say here could you tell me if this transaction is fraud or not um you know I can just uh copy paste the transaction hash and evaluate immediately so it will give me very high confidence level is not a fraud because the pattern for example if you copy paste the pattern here here um the pattern for this transaction is very different than the

64. 32:33

    pattern is saw in the past realize this is UN likely to be a fraud transaction so all this is actually enabl on our platform as well today that's really neat and um so for one I guess again from a hackathon standpoint developers who are not maybe comfortable with fine tuning or under you know they're not really in that part they can can just jump in and start leveraging some of the stuff that's already here sounds like which is great two it looks like just from what

65. 33:05

    you're just showing me that the act of actually updating these models a little bit doesn't seem overly difficult here it really just looks like small blocks of python code to really get across like what you're trying to change is that right exactly so if a if they break down all the step for training uh the most important part is they pration that's actually 90 to 98 percentage time 90 to 95% time so that's why uh this actually this work we actually referring to a

66. 33:37

    Berkeley us Berkeley's research called block gbt we build on top of the research result and then we follow the similar we follow the similar methodology here and then basically majority time it's massager data then once we have the data and then the training here is more like a boiler plate code you know boiler plate code and this happens to almost all the AI case you know 90% time you probably spend massaging the data and find the

67. 34:02

    right data and then clean the data and then there is some hyper parameter fan tuning but there is like like a formula for it uh you can just tune a little bit and then basically majority of time you just tune for the trade-off between accuracy and Recall now you guys mentioned that um you because you're actually leveraging based off of these uh base models do do I as a developer have a choice of which

68. 34:29

    model I'm going to use that's a really good question um so it's so really depends on use case like for fraud detection um previous you know if you look at the fraud detection most of the Lama model you know you know gen you know gen am model you cannot use for this particular fraud detection we actually use Bird model because Bird model has a classifier in this case you can see distilled Bird model it has a classifier um but model choosing and also how do you actually um you know how

69. 35:02

    do you actually design your structure U you know this you know neural network structure you know right that part code is uh it requires some experience but uh once that part is settled the majority I'm still focus on the data you know data data preparation part um yeah but U model you know Foundation model tring is actually a big part in this part yeah very cool very cool um so the agent part of this so a lot of people have asked us about different agent we have a few different companies

70. 35:35

    Come on talk about building agents so can I just jump in and just start building an agent as well well that's a really good question um I think the agent here is model plus perceptor plus actor here and then in this case u i I do saw a very popular project recently called um on t there is a very popular project just to keep AI generated tweets um in this case you can build large language model percept PRP all the new Twitter tweets and then do the action as a you know integrate with the Twitter

71. 36:11

    API to the Twitter tweets um from that perspective I think everyone can uh as any app developer can um integrate with a third party API to make a model to like to a agent in this case nice now all the stuff that's here um what is the the uh setup for a developer do I need to paint is like kind of like a SAS service can I just dive in and start playing with it what does that look like today and maybe in the future oh sorry Joe could you repeat the question yeah is so to start using Zeta

72. 36:51

    block and building on top of it is it like a SAS model where I have to buy credits and start building or can I just kind of jump in and see what you know play around with it for a while sure so I think irine is actually um give $50 per month credit 30 or $50 month per credit it probably cover majority of part um we we're actually going to change our business model to a token based model as well so currently it's us

73. 37:16

    model but we're going to change it we're going to launch a new protocol and change it to like a token based model ver so nice Okay and one of the things and I know this is not related NE to the hackathon but just kind of my own curiosity of just how you're stru so part of it is with the blockchain model was being able to uh make sure people are compensated for things like the data sets and models how is that all fitting

74. 37:43

    together and as a developer if I if I'm bringing different things to the table is it is this an opportunity for me to earn as well or am I real is my focus really just on building an end product that's a good question Aiden do you want to answer this part um yeah so model Builders as well as like people that contribute data um the idea here is that you know you may do it because you you want to do something useful um you want to find tuna model

75. 38:09

    that solves a specific use case um but the angle we're taking is just very collaborative um in the sense that like I maybe I have something I want to do um but I don't myself have the data to do it and maybe I don't even have the technical expertise to develop a model um but in this sort of more open AI economy I can find someone that has the right data I can find another person who can develop the model um and then I can

76. 38:35

    uh pay for inference and the rewards can flow back to them uh for their contributions very neat that's really neat so long term it sounds like that really the goal here is maybe it's not the best word for but it's more like democratizing the ability to build AI data sets so right right so again like you say it's not about decentralization the fact that like there's all these nodes running everything but the data is

77. 39:06

    trustless uh everyone can contribute I am compensated for my stuff you know it's eliminating the things that are people I feel like are uh getting frustrated with with the more centralized things where they feel like those are all just kind of ingesting and taking we're now it's a two-way street here yeah EX exactly yeah I mean yeah you're as an individual your your data is so valuable um and there should be an easy transparent way to actually uh earn

78. 39:37

    based off of that so yeah we we fig out as Aiden show earlier in order to reach as3.0 the the next biggest Gap is all this data I'm not talking about Wikipedia data this is actually like all the image data video data like for cancer detection data or it could be like AI girlfriend type data and then could be some robotic data in this case robotic data uh this type of data is actually you know great you know basically just in in great demand but on the other hand the data contributor side you know they have no way to contribute

79. 40:11

    data confidently if they upload the data in huging phase we basically give the data for free and then um if you know if they if they list the data outwhere they don't know how to charge a data as well should I charge 50k should I charge 10K or should I charge a million 2 milli million dollar it's very hard to charge a data so what we provide is a TRL platform so model builder can use the

80. 40:34

    data to explore the data but they cannot download the whole data site they can explore the data they can build the model they cannot download the whole data site but once they build the model they don't need to pay a great big up front fee they can just build a model and use the data set so every single inference fee will go give percentage percentage back to the data contributor in this case data contributor don't don't need to worry about mispriced the data you know mispriced data and motor

81. 41:01

    Builder don't need to pay a huge like a upfront fee for the data side uh yeah yeah that I mean makes a ton of sense to to me to start building that way and does do you feel and this is just purely from a philosophical standpoint this allows for better more recent data because I know I can both have control and be compensated that's right that's right um I I think it's going to be um it's much easier

82. 41:33

    actually working with 10 plus business partner to provide the data in this case their data is more you know massage and then they actually need to remove some sensitive information as well it's a little bit harder to get data from Community individual user because that's actually you know massage data make a data use for categorize data filter data is a lot to work in this case so that's why we start with business partner but I think the long long run Community

83. 41:58

    contributor data is possible um this is particularly true for like some multimod data mhm makes sense makes ton of sense uh I'd love to just kind of uh before we wrap up here uh so I have the link to your docs down below can you just do like a quick maybe overview of like what's available on your docs uh kind of getting started what do I need to have ready to uh start building something

84. 42:29

    sure um yeah so let me let me briefly work through the dog so the currently um just let me share my screen again the existing dog is actually uh focus on the data infa product so in this case uh we have a different recipes to actually uh use the data send s quy or build a backbone for the backbone for the uh like indexing the data build a backbone for the your application here and then we have some prebuild API for you ready

85. 43:00

    to use and then one one of the one of the casing we support is that you can easily convert a CLE to a graph K API um there is also a detailed uh detailed in uh uh description here so how to convert SQL to a graph API here um so all all the AI component we haven't release it yet so the the all the AI component including the um train you know training the data like a training data and

86. 43:30

    inference inference of the model EXA we haven't released it yet we target release at the end of the November and then we have we're going to have a early early release for first 100 developer or first 100 business partner in this case like the developer or or Community member uh we're going to release in phes and then I think the first first pH release probably going to be uh uh around sometime November so if you're interested please sign up in the in the in the website and then we're going to

87. 43:59

    release in batch in this case but for all the existing data product um um there is a you know comprehensive developer Doc and then you can just come here and then sign into use and then currently we don't require you to uh we don't require you to um you know it's it's a it's a better open better access so you can use the existing data INF product um Self Serve that's awesome yeah I wanted to make sure we show

88. 44:23

    because I think for developers again uh a again from just from like anecdotal from conversation like some people are still kind of skittish when they hear AI they think oh I'm not that level of programmer I don't know python whatever it might be but like they can get started you have apis if you know how to you know ingest a rest API if you understand graphql like you can be off

89. 44:46

    and running and doing something really pretty cool uh right off the bat right now so I think um from a hackathon standpoint feels like that that's kind of an easy way to go uh at the bare minimum um yeah that's right that's right uh awesome this this was this was great I I'm I'm certainly very interested I hope people uh really got something out of this uh there is so much cool stuff

90. 45:11

    happening on the AI front here I'm glad you guys are part of the near ecosystem uh this is a a really great Advantage I hope people uh take time to check this out and and start building on this uh any last kind of thing we want to leave people with um again we'll share all this out make sure people kind of catch up on this but anything else no yeah I think I'm good on mind um yeah I hope to uh see a lot of you in

91. 45:36

    Bangkok should be fun um so yeah thank you Joe so much appreciate it yeah no my my pleasure Scott any last words that's all on my side thanks Joe yeah no thanks guys I really appreciate it uh and again folks uh if if you haven't checked it out go on and check it out uh appreciate thank you see you guys yeah uh folks I I I hope everyone took some time to check that out uh this is again now we've had a few different AI platforms come on here and I want to take a second here and actually

92. 46:15

    just uh just kind of go through again some different things that if you're building and you want to start hacking uh we can certainly jump in you know right off the bat and let let me let me show you something here so let's let's jump in so as Scott was showing here so in in the AI uh documentation here I'm sorry the API documentation here again when we talk about building a product for a hackathon one of the tips that we've continued to

93. 46:49

    harp on here at devhub is like don't try to do too much like don't try to over engineer and here's a great example where I can take maybe an existing product I can take something that maybe doesn't quite exist yet and I can jump in with their pre-built apis and right off the bat if I need to pull in different data sets right from chain data like that's an easy one like the example they just walked us through here with uh fraud detection all of that

94. 47:21

    stuff is readily available I can start pulling in some different information if I'm doing something with chain signatures I mean let's start putting those pieces together right if I'm building with chain signatures I'm doing stuff across chain and now I want to pull in some other data you know in real time you know through this and I want to have a model that maybe is again uh you

95. 47:45

    know whether it's fraud detection whatever else it is I can start pulling these pieces together and from a hackathon standpoint now like you're integrating AI you're integrating a chain signature you're you know integrating all the different components that you can all apply to to you know try to win prizes and so to me I feel like this is something where we're not just saying like hey check out these

96. 48:09

    different platforms for the sake of doing it it's really because it gives you an opportunity to really expand your product a little bit maybe think a little bit differently outside of what you're trying to do uh and see kind of where things kind of go from there one of the other things uh here too is that again uh you can actually you know as Scott mentioned here you can actually jump in and do some some Custom

97. 48:34

    Calls as well and that's uh again if you're comfortable writing graphql stuff that's really powerful as well and overall I mean I haven't really had time to check out the recipes just yet but I mean really simple stuff to kind of like query to a real-time database JavaScript like here it is like it's like if you're working in JavaScript you know here's the recipe they're making it super

98. 49:03

    simple for you and you're of and running and so I think these are just again like really great opportunities for for everyone to to jump in uh overall let's let's check out while we're here uh so if you have not seen this uh again the r haathon all of the different prizes are down here and you know we just talk to Zeta block and so here they are here is just highlighted everything

99. 49:37

    but you can see all the information here and now this is again where you can start pulling different pieces together and seeing what other bounties are there that kind of make sense for you uh tons of great resources all up here we have over 1300 participants uh there's plenty of time you haven't seen uh the deadline has been updated to October 31st uh that's a lot of time still to go out and

100. 50:03

     build uh a lot of the stuff that we've shown on these Workshops can be built you know get at least get started in a weekend like absolutely like you can dive right into this so really take that time jump into there if you are unsure if you have an idea that you want to just run by someone uh really come to any of our different chats so if you go to the resources and you come in and you want to jump into uh any of our different groups we have telegram groups here jump

101. 50:35

     in there uh ask some questions share your stuff you know uh we're all here to really help each other we you know from at the devhub standpoint we want to make sure that uh you feel really comfortable with uh everything that you're trying to do if you need technical help we're here and so come ask those questions come walk through through some of this different stuff uh we understand that

102. 51:01

     sometimes it's hard to have a a good idea for hackathon so you know Jump Right In and and please please ask questions uh we're we're available obviously and through our our Twitter as well if you don't use Telegram and you can message us uh there as well we I will say if you if you haven't already as well um make sure that you are looking at the general I know frequently asked questions here of what are we looking for in the different projects a lot of opportunities to win prizes again over

103. 51:33

     $700,000 worth the prizes so there's something for everybody um you know I I I really I think it's a great opportunity for everyone to to jump in even if you're not a builder you're not a coder uh you know find a team uh brainstorm uh that stuff is just as important presentation is important as well all those things so uh great way to get started we are back uh so we will have our regular scheduled uh devhub live next week I not

104. 52:11

     next week tomorrow the today is only Wednesday tomorrow on Thursday we will have uh again a special bonus uh on Friday evening we're going to do a devhub live uh coding with frol uh he is one of the uh early developers here on near uh we have meta and he's going to walk through and he's going to do a to-do app uh on near just to show how the contracts work all that stuff it's going to be a great time so your Friday evening is all set book it down and then

105. 52:47

     Monday we have meta pool so they they have a liquid staking platform but it's so much more than that and so uh really make sure you're following all socials here on your protocol near devhub so that you can stay up to dat uh we have a newsletter the link is down here below uh every single week we have great updates for developers uh everything that's going on in near ecosystem so make sure you take some time to uh sign

106. 53:12

     up for that get a quick summary of you know what hackathons are happening how you can help uh how you can jump in you know what do you need to know to be uh top shelf Builder here on here with that said uh this was another awesome Workshop I'm super excited everyone got to see what they're building uh again we will have uh even more great stuff all the way through the end of the month so keep building keep hacking uh it's going to be a great October here and we're excited to see

107. 53:45

     what everyone builds but until then we will see you again tomorrow and uh I am Joe your host for near devhub see you later