NEAR AI Tech Showcase - Token2049

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

1. 0:14

   one what what

2. 1:30

   n n

3. 2:38

   good all right all right I got the thumbs up so I guess we're live welcome all to our near AI Tech showcase my name is Jared Barnes I'm our head of ecosystem and founder success at the near Foundation uh and Incredibly excited for today's event for those of you who are with us at ECC in Brussels earlier this year we had an amazing round one of our first iteration of our Tech showcase uh so uh if you're back with us again welcome if this is

4. 3:35

   your first time welcome you're really in for a treat we have a packed agenda today so I'm actually going to quickly get out of the way and let our first Speaker here come up but to briefly give context on this event why we're even hosting this where this is even coming from uh for us at near AI is not just a feature it's actually core to our product we're building an AI native ecosystem uh and working with a number of teams at all layers of the stack whom you're going to hear from to today which

5. 4:00

   we're really excited about we also have two uh esteemed members of our near AI team um one of our co-founders as well who you're here from uh and more importantly as we have been growing and scaling this ecosystem um there are two big Partners uh that we're really excited to introduce today so I'll give some suspense to that but going to hear from them uh later on as well um but first and foremost thank you all for coming out I'm actually going to go

6. 4:24

   ahead and turn it over to our first Speaker uh Pierre who is an engineer as part of our near AI team team to give you a little bit more context about our big picture Vision um when it comes to AI on near What specifically the near AI team is working on and this Vision behind user owned AI so let's give it up for Pierre should be good yeah okay yeah perfect thanks for the great intro Jared um yeah so today I want to make a quick case about user own Ai and

7. 5:15

   why we should all all care about it um so my name is pwen I work with neari for a few months now and um so first thing first what's like user own I to give you a quick definition um so right now we all rely on centralized AI providers to do all our inferences model training we also give them all all the datas in Need for the model training and all this kind of stuff and so here the idea is like we

8. 5:40

   should really lower the our dependency on them because if one day they want to close access we are just we in trouble so yeah I will I will do a bunch of different use cases and I will cover like data sets benchmarks models or agents and try to give you each one of you like a quick definition how what what it really means um so the first use case is I have a a friend working at e d she was a re

9. 6:06

   research assistant and so for a few months she was basically uh tasked to do like um annotations on like different images so she would have an image of different celles and she would have to annotate what they were looking for so she would spend months and months and months to do this to create like a data set but um as some of you may know to create to do like a model training you

10. 6:30

    need a lot of data and uh so we as one researcher there is no way that you have enough data to create like a proper model that really makes good predictions so uh what they've done at eth zc they contacted other research groups and together like they gathered a good enough data set and then using um a third party provider they did the model training and then this third party provider will um will do the serving for them so once the training is done given

11. 7:02

    certain model architecture uh this U third party provider basically gives the access to the to the model and then so like everybody relies on this third party provider and then when other research groups wants to join the join like the fun and had some some new data they have to email everybody and find like uh how it's like set up right now it's like very complex process and tedious it takes yeah weeks and weeks for supposingly easy process and um yeah

12. 7:32

    so I want to offer a new parading so basically here is like so you have again this uh three research groups together they uh put the data set in some like decentralized uh like storage where like everybody can contribute and just like add the data set here and so it's like a normalized like easy to access and decentralized place where everybody can add the data then once all these data is gathered they um they do the Mel training or fan tuning so depending so

13. 8:05

    let's continue with the researchers depending on the University you might have clusters in your University but most of them don't so you would still need to rely on a third party but here the pigm is different because um like the API is very standardized standardized so it's for the user it's very easy to switch providers based on like based on their needs um so once the model is trained you have to uh to do

14. 8:29

    some benchmarks so Benchmark is basically a way for you to tell if a model is good or not at a given task and uh so here example I took like another research group and basically they are they want to do some benchmarks for self detection processes and so they would had a new Benchmark specializ just just for this and so with this approach it's very easy for anyone else to come and um

15. 8:52

    being able to evaluate different models for um like with different benchmarks and then you have also the the guy like a new research group coming in adding data data points into this data set and so it's like a very open process and everybody can read the data and it's not like nobody can shut it down shut it down basically um so just now to speak more about data sets uh so quick definition data set is basically a collection of examples so this is used by different

16. 9:26

    like a models to do the training so basically you inest a data set and then it's going to basically train on all the data you gave it um so right now the current approaches by all this centralized party like open Ai and G Google Microsoft uh sorry meta it's like so they contract third party providers and then so these third party providers are paid to do a given data set and of course this company are for-profit organization and um so this for-profit

17. 9:58

    organization are really incentivized to increase the margin they get from from the money they got from this open AI for example and so this like um for the what it means for the people doing the annotations like they're paid for the quantity of the data they're providing and not really on the quality and also like they these providers for these data set providers they use like people from countries that are not paid a lot so it's like kind of unfair and here if we

18. 10:27

    take a second approach of crowdsourcing it's where so you open a b and let's say you are willing to spend a million dollars on a given data set and then like all the community can contribute to given that you have the knowledge for it you can add new data points and also what like this provides is like it's very good for data quality because then you can do like peer reviews of the given data points let's say someone adds

19. 10:50

    data at at a given time someone else will come afterwards maybe it would come to double check and if it happens to be that the dataon is wrong the person providing the first providing the data before get slashed using different mechanism that could be decentralized and here the parad changes really on you're P paid based on quality and not quantity which is very important in AI

20. 11:13

    because for model training basically the most important is usually the the data set that you give in input as input uh now so about benchmarks so benchmarks so quick definition is like um basically a standard test or it could be uh some not not standard but like a bunch of tasks that are used to evaluate a model so basically you give an input you have an expected output and then you run all these benchmarks on a given train model and this way you can determine if a model is good at a given

21. 11:43

    task or not so you can for example this cell detection uh model you give it a bunch of different cell problems could be cancer it could be whatever and then you can see if it's good or not for a given task um so this is used right now by research communities to determine of course if the models that they created are relevant or not and efficient but right now it's a very tedious process

22. 12:05

    also like it requires a lot of knowledge how to create these benchmarks and even where to like store them how to share them with the community or even like companies have like close Source benchmarks to to check a bunch of different things and so here it's very easy for the community to come add new benchmarks add new examples even like create small benchmarks for very specific task uh yeah so with the new approach

23. 12:30

    it's uh yeah very easy for anyone to add new like benchmarks and create uh cool things and so also like the community is more way more Broad in reach compared to um like the closed approach where yeah um so I want to take a second use case so let's say you are a hedge fund and you want to expand the service you offer and so you want to do defi like automated defi trading um so this

24. 12:56

    requires like a lot of knowledge in the space and uh you don't want to spend like you don't want your teams to spend like months and months coming up with a perfect like agent and so here um what you do basically you have like an algorithmic trading agent AI if you wish it's just more specialized and this agent what what it's going to do so let's say uh you want to have tweet sentiment analysis because it's kind of

25. 13:21

    common right now to do these kind of things so um you have a bunch of different like tweet tweet sentiment analysis agents that basically each one of them are run against a bunch of different benchmarks that we spoke before and so you know a given agent is it good or not for the task and then also you know the pricing for each agents and this way as a company you can choose which agent you want to use or not and you do this for every uh every

26. 13:47

    features you don't want to implement in house and this way like you have a you can use a cool ecosystem where like everything is interconnected and it's very easy to use different models and different agents and like all this the work that the agents do they usually work requ payment so you can do microtransactions micro payments with um using blockchain so it could be on near for example where it's cheap cheap and

27. 14:12

    efficient um so yeah to speak more about the models so right now uh for example open AI offer a bunch of different models if we say like gp4 so as I said before the most important thing is read the data that you give it to these models so this this companies contract companies to do the the data set for them they all come with different biases because like open don't want like have a lot of for investors and there are some

28. 14:38

    things they're not allowed to say or they don't want to create like any politic problems so they come like with lot of quad rails and this kind of thing and also you you don't have a transparency of what it's doing in the inference side so if it's like biasing your results or it's not answering what it's supposed to answer and all these companies are profit driven so they want to get the most money out of their users and so the other approach could be like

29. 15:01

    leveraging open source models uh we could um really like create much broader like models ecosystem where like models for specific tasks uh exist and so it's very easy for anyone to create models and uh Leverage The Open Source community and here it's uh so it's very transparent and open source you have less bias because you can check everything and even if you're not happy with the model you can just train train

30. 15:26

    it again and here like you own the data that is so you can double check the data that is used to train the model and you can share it if you want the d your data to be used to train the models uh compared to like centralized companies you don't really know what data they're putting in their model training so it could be personal data but we hope not um so agents so um for example P offers what they call gpts um so basically how agents gpts for example work you give it a

31. 15:57

    bunch of different inputs it could be pdfi prompts bunch of different things and with this open AI is going to try train if you wish a GPT uh so it's basically agents and then like it's going to do the tasks that you asked it to do so it's a full black box because you have no clue how it works inside if one day open ey decides to close this service down there is no way for you to like for

32. 16:20

    your company to continue running on it and it's uh behind a lot of paywalls access controls so if they increase the price if they just don't want you to access anymore you are really uh in a world of trouble and also lack of collaboration let's say you have a GPT and you want to interact with an agent running on some gmin agent platform there is like not so many ways to do it right now and I'm not sure is it off I'm not sure there is a way in the short

33. 16:48

    term to uh for it to work so I want to speak about agents we been speaking before where it's very easy for anyone to uh uh thanks one two what um where like for these agents it's very transparent you can really check what they are doing you have access over the inference you have uh really like a rich marketplace where you can interact with bunch of different agents and you leverage micro payments on blockchain so basically nobody is like telling you you you can access it

34. 17:20

    or not it's just there and even if like the company offering this business goes down like it will remain for forever as long as we have computers to do the compute yeah so just to sum it up so what's really like the user own stack basically it's like a very opener open and collaborative stack where like everybody can interact with it's fair you can really Trace everything the ownership is not in one company it's really distributed so as long as we have storage and compute it's very easy for anyone to uh to interact with this and

35. 17:54

    uh you don't have a single point of fure compared to like party providers where let's say open AI wants to shut to shut their service down one day your service will continue to run um yeah so just one last example to sum it up so the I I took the example of centralized AI versus like user owned AI or decentralized whatever name you want to give it um so the biases with centralized are very high because in the

36. 18:18

    data sets are very high because like they're incentivized to go in given Direction based on on their like uh yeah on their on their need uh it's not at all it's a full blackbox so you have no clue how it's running inside data privacy put it medium there because in theory they're not supposed to use uh use the data but it it was shown multiple times already that they are they are doing it but by low normally

37. 18:42

    they should not um data privacy for the case of USI would be uh high in theory where like it's very easy for anyone to double check what data is used to train the models and um all this kind of stuff for the competitive advantage so for centralized AI so if you use a model served by open AI gini meta you you don't really have an advantage of your competition because everybody in this in the industry are using these models so there is like uh very very low margin for you to really

38. 19:15

    like be like magnitudes order of magnitudes better than the competition but if you are using like uh a model that was strange just for this process it's very easy for you to um to be really like better than the competition and the ease of use for centralized AI it's easy because everybody knows their apis as a builder and for us own ai ai it could be easy could be hard depends how the builders do it it has to be

39. 19:40

    built and uh yeah thanks for listening to me right thank you right yeah you're good okay all right folks well can we give it up for Pierre one more time thank you uh incredibly excited to make everything Pierre just shared real and go into some rapid fire demos of projects who are building some of the exact use cases that he just mentioned however before that we do have two announcements that have been in the works for some time as we've been growing and scaling uh our AI ecosystem um the first is an incredible

40. 20:37

    partnership uh with a company called a she's uh here in the back if you could come up uh very very excited about this partnership you know one of the things that we've meaningfully thought about is how do we uh scale one of the most valuable assets that exist today in the context of a Ai and that is compute um and aier will be our PR for compute provider for our upcoming accelerator program and more broadly to our

41. 21:03

    ecosystem um and so wanted to have she up here a little bit just to share uh some more about the vision for a um what you all offer to Founders and projects um as well and I'm realizing we need to get her miked up so we'll provide some coverage here last minute but amazing yeah if you want to go ahead and take uh this mic you're good yeah thanks Jared it's really amazing to be here just to start with I was part of the new fam as well until last year so it's just a

42. 21:32

    great feeling to be here uh ather uh how many of you have heard of ather before just uh show hands oh that's that's pretty cool yeah so um we're solving for the GPU scarcity problem to be very honest and uh we're kind of building uh GPU decentralized Marketplace uh an Enterprise Focus GPU Marketplace essentially um for helping AI and caming companies out there uh by aggregating gpus uh from Telos from Enterprises from

43. 22:07

    data centers from data mining operations and um kind of plugging it into our compute marketplace right um there's a huge GPU scarcity problem also there is an underutilization of GPU capacity uh if you look at it close to like 15 to 18% in the mercus so um you know we provide incentives to list gpus on our platform and you can access it at uh I would say cheaper costs than any hyperscalers out there with this partnership uh essentially what we

44. 22:39

    intend to do is provide help and onboard newer startups uh to our ecosystem both new and ather ecosystem through computer credits um if you were to go into our Marketplace it's pretty straight forward think of it as entering a restaurant looking at a menu and ordering in right so you just select what kind of gpus what type uh for what duration what location and instead of a credit card you use ADH to kind of uh pay the

45. 23:08

    transaction and on the supply side if you want to contribute GPU on a network it's pretty straightforward as well it's like a 15minute API call integration and you just need to take at right so pretty straightforward very excited to partner with near I see like a huge overlap with the kind of uh startups that we support and kind of near AI is pushing and happy to kind of uh support through ecosystem

46. 23:34

    activations through compute credits any sort of support uh on the infrastructure side so in case you guys are interested in looking for gpus you can find me uh the back uh after the end yeah thank you Jared thank you so much F the takeway is if you need computes please find Siege but again thank you we are so so excited uh to be working with you all so that is partnership number one however there has been some more in the works as well realiz I'm blacking my screen get so

47. 24:06

    excited mess up the the screen here the second partnership is as we've thought about meaningfully uh again scaling our AI ecosystem how might we work with the best in the space uh who can bring in expertise to meaningfully engage with Founders and projects so we actually have I think a good chunk of the entire DOI Labs team here if you all want to be able to come up front um and get slowly miked up looks like we got Kevin here as the sole representative thank you we're

48. 24:31

    going to get you miked up real quick and as Kevin is getting miked up I'm going to slightly stall but give some coverage here um about 3 months ago uh we set out and said wow we had incredible success with our first uh near AI incubation cohort some of those projects you actually hear from today in order to meaningfully do not just round two and copy paste but we said how do we get better um we wanted to again go work with the best we had probably one of the

49. 24:56

    most impactful 30-minute meetings the DOI Labs team and it was very clear to us that we wanted to uh go work uh with um uh just a great organization so uh Kevin here um COO of Deli Labs if you could just share a little bit about um not only your work but specifically the AI thesis that you recently launched and we can share also more about the program as well yeah happy to oh yeah yeah I got miked up okay yeah

50. 25:21

    yeah lot of mics all right yeah great to be here and to represent Deli in this really exciting program I'll just take a a minute to introduce you to Deli in case you're not familiar so we actually operate as three different companies we have Deli digital started in 2017 does Crypton native research and insights um has really covered the space well and in crypto and aai in the last year we've published three we think are really seminal reports in crypto and AI we've

51. 25:51

    got a fourth report coming out a little bit later this year and we think in total it's probably the most thorough coverage of crypto and AI so so really encourage you to check that out but the tldrs we've been covering the space and and think deeply about it the second company is Deli Ventures that's our crypto native VC fund my colleague Alex Golding back here um is with them and we've been investing since 2019 made

52. 26:14

    over 150 investments in crypto in the last probably year and a half we've made at least 15 in the crypto AI space couple of the notable names include uh news research grass pond i.net so really looking to invest across the stack continue to actively write checks in this space so really paying close attention across everything that's being developed in crypto and Ai and then the third company is where I spend most of my time and my colleagues over here as well Deli Labs we're the Builder team we

53. 26:46

    incubate and we accelerate new crypto projects and we're a little bit different from some of the other crypto accelerator projects out there like we take a very thesis-driven approach and what that means is that we go really deep into a few select areas and we bring not just the crypto expertise but usually a lot deeper level of sector specific expertise as well so um it's really good timing because speaking of

54. 27:11

    of thesis we just published our crypto AI thesis last week really excited to get that out ahead of this uh big event here and um encourage you to go check it out it's pinned to the deli lab's Twitter profile it's a long read so um you know get some time get a drink whatever you need to go through it the really short answer to it is that at Labs we believe that the decentralized

55. 27:35

    sort of many models world is what is much needed to balance the centralized god mode models that big Tech is pursuing and so we believe that every element of the stack you know has an opportunity for us to explore in a decentralized model and that's what we're kind of pursuing and um you know partly why we're super excited to partner with near in this U accelerator program we really want to find you know the next generation of AI companies that that we can help accelerate near brings

56. 28:07

    a tremendous technical capability set and we're um you know super interested in in everything that they have to offer it complements really nicely with what we do best at Deli labs and what we bring is a lot of crypto expertise we've been we've worked with more than a dozen projects in the past three plus years helping solve everything from kind of early product Market fit to things like

57. 28:31

    tokenomics go to market strategies fundraising ultimately your tge launch and we've pretty much seen everything with the companies that we've worked with so I think probably the most important value ad we bring is just helping Founders that have to figure out all these things for the first time we've probably seen these things up to a dozen times and so we can just help you get through them so that's um why we're excited to be part of this I think it's

58. 28:55

    a really powerful combination when you bring together what near offers and and what we bring to the table so if you're interested in the program come find me uh Luke redphone Sage my colleagues over here uh happy to talk more about it Kevin uh and more importantly Deli Labs team thank you all so much a for being here but we are so excited for this partnership folks October 14th just less than a month away is when we'll be kicking off this program formally so please take Kevin up on the offer uh we

59. 29:28

    will all be collectively around here if you're interested in learning more about this program now thank you Kevin uh for the reason we all came here we actually want to get into some rapid fire demos this is not just another uh set of panels we really want to get into uh the tech our first uh founder who I'll talk about here uh briefly um Jordan from The bit team just kidding no Jordan from The

60. 29:51

    bit team slight adjustments to the agenda um we're going with Bob and Quest thank you for for the heads up no we're good we're good um always a lot of moving parts so I'm going to talk about Bob from Quest flow first thank you Jordan for being here you're coming up next we met Bob from Quest flow uh about four months ago in San Francisco one of our Bay Area based Founders really really excited about what he is building in the agent space um if you just go

61. 30:17

    ahead and unplug there yeah you'll be good to go um what he's building in the agent space particularly uh in the context of productivity I won't steal his Thunder I'll let him share a little bit more specifically about uh that as well so Bob take it away man on great um all right nice nice meeting you everyone uh my name is Bob co-founder and CE of quest flow what we're building is as you can see is a Next Generation AI agent economy uh very echoing what

62. 30:53

    near was talking about we want to create a multi-agent ecosystem where we reging multiple AI agents uh to ask them to form a group um and then actually take actions on your behalf and then we distribute incentives to these AI agent creators supporters and the community members um so without further Ado let me show you the product actually I was warned that probably I should do a recording but it's not as exciting so let's see what happens um so before we jump into what

63. 31:23

    we're doing let me show you how many agents that's currently in our Market uh because agent is a very vague term so in our sense we we see agent as you know the agent in The Matrix that can actually take actions on behalf so some of the agents within our ecosystem are more data focused like uh Hacker News GitHub trending data OK link all these uh data agent can get data sources from

64. 31:47

    specific uh uh sources and some of others are more you know action-taking ones so we work with uh you know slack notion Gmail Google Calendar all these agent can actually uh take actions on your behalf not just you know giving you information so for this demo um what I'm going to show you is uh something really too near so for example um I copy and paste some of the commands uh one example that we use very often is token price monitor but uh as we know that we

65. 32:19

    use telegram mostly so what we can probably do is as a first try um you can talk talk about okay what is the token price of near uh in this case uh what multi-agent our multi-agent orchestration uh framework going to do is first we're going to understand user intent second we're going to find out within the agent Market which agent can solve this problem so in this case it's going to be uh okay link uh which is uh

66. 32:44

    the okx uh agent that can get the token price um then it will takes a little bit time then what's going to happen is it will start a new task which is having the near token price quy for you then it will get the latest token price of new uh this is just one um this the first step the Second Step going to be okay um I already get the price then how about send the token price to telegram so what's going to happen in this case uh that's why a lot of people are not doing this is we're going to start figuring

67. 33:14

    out the context what's happening beforehand and then actually sending the message uh to telegram uh but for any action taking agent there's a problem is you need to handle different authorizations handle different missing parameters across multiple platforms so that's where Quest flow comes in so we design our own multi-agent framework where we help you to figure out where is the missing parameters uh what telegram needs and then we can fill in this information so in this case as you can see we'll say okay you can fill in the

68. 33:45

    chat ID um in and then you can get a message so we already get uh we we fill uh we build a bot for you uh let me c copy it and then you can type in as you can also see that we are probably the only one that's all natural language so you don't have to know any code you can just talk about what you want to do across multiple platforms and uh our multi-agent uh framework will start working for you uh on the spot so it takes a little bit time uh then as you

69. 34:17

    can see the token price of new has been sent to Telegram and then takes a little bit time then the telegram bot will receive the message uh right here so this is one example um the other example uh that I want to show you is uh not only we can handle messaging apps we can also handle a lot of SAS tools uh so in this case I have another um demo you can just talk about okay I want to post uh NE ecosystem blog post about Ai and web 3 uh on notion uh which

70. 34:50

    is very regularly used in your daily workflow and also it's uh useful for SEO as well um so some of our c do is they will say hey like every day uh based on some you know news post a blog post for me on my notion and then you know use that for SEO purposes uh that in this case as you can see we found two different uh agent one is the web three information aggregating agent and one is

71. 35:14

    notion agent but as you know that notion is very complicated you need to handle notion authorizations you need to handle uh which notion databases will be read in and where do you want to put put the blog post so what we're going to do is in all natural language we're going to ask you to authorize us uh the notion uh as a tool um and then when clicks yes and then we start understanding it then

72. 35:39

    we're going to figure out okay what's your notion data structure what are the uh different workspace with your notion and then we're going to ask you to pick which specific database you want to post uh post the thing so you're going just click confirm um and then our multi-agent start working basically getting the information drafting the blog post filling all the uh data within notion which is the title the description the body and also actually

73. 36:03

    posting it for you um this is more like a one-time task that it's more for demo purposes but everything can be set up in a autonomous way so you can just say hey like every day send me a message or every time there's a certain CR criteria getting met uh making sure that you are posting things uh on on notion so uh it takes a little bit of time to get the magic done so that's that and before uh they complete the task uh we also support multiplayer mode so as we probably uh

74. 36:38

    like I mentioned we orchestrate multiple agents but we don't think it's a single player game it's more like a multiplayer game so we allow you to invite your team members to join the workforce with you and also you can share just like a figma file or like a slack Channel then you can work with multiple uh team members uh along with multiple AI agents uh within the market um and some of the other ones that we support it takes a little bit longer so

75. 37:02

    not not waiting some of the other ones that we support is uh like uh we also support multimodel like YouTube um uh videos and stuff like that uh any case I think my internet is not connecting but you got an idea anyway uh well everything is live so feel free free to give it a try uh on quest's website and uh yeah happy to answer any questions sorry it's it's not connecting but yeah can we give a round of applause for a live demo not easy live demo

76. 37:45

    is yes thank you awesome um folks Bob will be around if you have any questions thank you so much Bob really really appreciate it we're going to go ahead and actually transition to our next founder Jordan here from the bit team as he is getting miked up and set up again thank you Bob really appreciate that um we're on a theme here oh we got some traffic here I'm G to get around you we

77. 38:07

    got a theme Here of Agents um Jordan from the beta team is also working on agents in a really meaningful and compelling way one of the key questions uh and problem areas space you give an agent a wallet what new uh uh uh behaviors are enabled what new interfaces could you design what new experiences could you create really really excited about what they're building he's going to also give you a live demo a bit more on the crypto native side on um what they're building on the agent side as well so we'll uh

78. 38:40

    give Jordan a second here I can take your uh computer get this plugged in and we'll get him set up cool my lady let's go all right um really honored to be here it's such a pleasure thank you for hosting um and also it's great to see so many familiar faces and new ones to meet today um I'm Jordan aka star pause can find me online and I'm here uh representing bit of protocol and everything I'm going to share this whole deck is chocked full of good stuff so go

79. 39:38

    ahead and scan the QR code um you'll get yourself a near name an NF and also access to the slid so that you don't have to take notes while I'm um firing off all this great stuff so if you've been around the near ecosystem you'll know that beta has been through um a a whole lot of different things um the the amount that's been built under this brand is huge and the new name allows us to take all of the really great stuff that's been built and bring

80. 40:11

    it together in a really meaningful way so what we're doing is providing a smart and simple way for everyday users to access crypto on any blockchain powered by AI agents um that's a big uh Mission so how do we break that down uh we've got the onboarding and engagement uh transaction facilitation and our developer ecosystem that brings it all alive there's a lot of moving Parts here they've been built over the last four years as we've done activations for big

81. 40:42

    conferences like ncon um music festivals and um we'll have a few more announcements that are in the deck that you'll get to see uh the activations that we're doing so what do we have that brings this to life a white labeled wallet that helps build trust with users uh any brand can come in and white label the wallet so that when a new user comes in and lands in a wallet they don't think what is this weird thing they see

82. 41:06

    the brand that they trust um the AI discovery which helps with retention and engagement giving opportunities for cross-pollination between projects and uh continued engagement from any uh brand that does an activation with us and then we've got the pay Master which is clutch for getting new users into the ecosystem um what this does is allows you to do air drops um and everything else without

83. 41:31

    users having to figure out what gas is how to get crypto Etc they're just like in the door and then they can mature and figure out what all the value is there and graduate out of it but the pay Master lets them get in the door easily and then we've got the cross chain functionality coming where we can do not only near transactions but also have pay master and all these other things acting across ethereum salana all the other

84. 41:55

    chains um and then like I was saying the developer ecosystem is how we bring this to life with our own open source tooling and also our grants and bounties for building out more um excellent uh AI agents so I mentioned the AI wallet or the white label wallet here is an example where we partnered with near Foundation to enable the internet Cricket Council to onboard tens of thousands of sports fans who really don't have any any um reason to be caring about crypto but we did a loyalty system where they

85. 42:29

    were able to uh come and participate in the international cricket uh games that happened and it all felt like internet or international cricket Council and this week we have launched the FKA twigs uh live stream um from sou bees anybody who Tunes in to this web page FKA twigs can mint a frame from her performance and if you're not familiar with Twigs she's got millions of followers and she's one gramys and she's a very controversial out there artist

86. 43:04

    she's a great fit for crypto um and it's a really cool Community to be bringing into the space um she's doing a health Tech performance over two weeks where she has 11 performers with her and they're performing the whole time 24 hours and it's kind of like fish tank meets crypto um anybody can tune in you can do it right now and check out what what the dancers are doing doing uh right there and this is in collaboration with Erse and then uh I'm not so brave as to

87. 43:36

    be doing the the live demos on stage uh with the spotty internet and everything but I do want to show you exactly what it looks like to interact with an AI agent inside of a wallet because that's a common question that we get so just going to drop us into the wallet and there's just a button to click that says AI you come in and chat with the agent just as you would expect to familiar uh we're just send some near here and the important thing is that the AI agents

88. 44:01

    are generating transactions they're not going off and doing transactions on their own they're bringing transactions back to the user that are well formatted and easy to understand so that the user is doing the final action themselves and then after the user does the uh the verification somebody gets Rich we sent some near uh I also mentioned the pay Master earlier as being the way to get folks in

89. 44:27

    the door really easily one of the most exciting pay masters we have live right now is text Royale if you're interested in playing a telegram game it's a battle royale happens on the hour in Telegram and we're doing um thousands of transactions through the pay Master for them so that people that are encountering near for the first time are able to claim their pay play to earn tokens directly from the battles that

90. 44:50

    they're winning there and kind of Shifting Gears to the developer side of things uh the templates are where we have examples of how to integrate the wallet into your app whether it's like JavaScript or um whatever other technology you're using we have a large list of ways of doing that we also have all of the AI agents that we've built live there and all the agents that have been submitted by hackathon participants live there so if you're interested in checking out agents

91. 45:21

    how they work the nuts and bolts of it and also if you want to submit one this is the repo a clone do your work throw the pr at us and that brings us to the hackathon where's that Banner oh there was a banner up here yes yes redacted um definitely be there in November but also the hackathon has already started so get get hacking on the agents um we're offering 3,000 per agent that is submitted and our criteria are around the usual things like functionality um

92. 45:54

    Innovation how usable it is in your documentation but but most important for us is the maintainability of these things we've done so many hackathons where we've seen amazing things built and then teams just disappear and don't turn their their projects into actual products and find success with them so what we're really interested in seeing from teams is folks who are dedicated to building something meaningful and want to iterate on it with us as a

93. 46:18

    partner um you can find the full details uh in a in a GitHub repo in the readme that's on the QR code we've got the the readme there and and um you submit it on dev post and then one last thing there's always got to be that one last thing um exciting thing going on at the uh at the bitto labs where we just keep on shipping um so we've got this new npm make agent tool if you're thinking to yourself oh my God building an agent it sounds so

94. 46:53

    hard no npm make agent you'll have a up and running in less than 30 seconds uh very easy command line tool it does the deployment for you using enro and all that so very quickly you can begin interacting with an AI agent that is doing Financial transactions on chain um really excited to bring that to the world um and yeah that's that's pretty much it really excited to see the the everything that gets built and if there's any questions happy to take

95. 47:23

    those as well um that that's it incise and tight thank you thank you so much Jordan really really excited about what is happening in the agent space um we're going to do a short break am I interpreting you correctly all right I'm working on my lip reading here on getting cues from the back we're going to take a short break because we've had some amazing demos when we come back you're going to hear from a couple more teams who are building we'll take a quick 15minute break uh and then we'll

96. 47:53

    come back for round two

97. 48:55

    d what oh

98. 49:57

    what d n

99. 52:09

    what 2 why

100. 53:55

     what what n

101. 55:25

     y d w

102. 56:56

     d a

103. 57:57

     what n

104. 59:25

     what d e

105. 1:00:56

     w a

106. 1:02:10

     what regular scheduled programming now uh you heard from our first couple of teams that were really really focused on agents um now we're actually going to transition a little bit more to compute one of the more fundamental uh aspects of the ecosystem in which we're building Jasper Zang uh co-founder and CEO of hyperbolic one of I think AR arguably our most dynamic Founders in the ecosystem uh going to go ahead and turn turn it over to him to share a little

107. 1:03:16

     bit more about what they're building through hyperbolic thanks for the introduction and thanks for me uh to the near AI showcase uh hi everyone I'm Jasper I'm the CEO and co-founder of hyperbolic I did my math PhD at UC berley I finished in two years which make me the fastest person to finish a PhD in the history of Berkeley when I was 24 years old uh and also give me the idea uh of the name of my future company hyperbolic because I was doing uh topology and hyperbolic

108. 1:03:45

     geometry uh before that I graduate from P University with the highest owner soon of the year I won a few gold medals in various mathematical competitions after graduation I work at uh seral Securities and Avalanche before I founding hyperbolic so at hyperbolic uh our mission is that uh we want to democratize AI uh to bring open access to uh AI for everyone uh we believe that AI is the most powerful technology known to the Mankind and in the future you will be integrated with the most

109. 1:04:18

     important systems like Finance um robotics social media transportation uh even your brain uh and uh right now in the centralized ecosystem there's no way that you can uh be part of the uh the the community to uh control the AI systems and neither you have the chance to participate in this Revenue sharing opportunities so uh holic we try to build an open AI ecosystem economy with our partners

110. 1:04:47

     including near uh of the People by the people and for the people where AI infrastructures services and models are accessible contributed governed and benefit all of you um so a lot of people have the same vision but uh we're trying to currently focusing on the infrastructure layer and want to solve uh three major problems first we want to build a scalable system to utilize uh decentralized heterogeneous Compu around the world uh first uh we're building the

111. 1:05:19

     AI inference service to uh which we call hyperbolic compiling service that can compile different am models into uh various Hardware backend including Cuda for NVIDIA rocken for AMD uh metal for Apple Etc and we can achieve uh very fast performance on AMD chips right now and we provide comparable performance compared to centralized Solutions like together at any scale um in the back end we also build our decentralized oration layer which we call hyperbolic

112. 1:05:50

     decentralized operation system so as long as your machines and clusters install and run hyperd do uh your machines can be seamlessly integrated with our Network which allow users uh around the world to uh utilize and orchestrate all these computers across the world and so once we've we already finish uh the back end and we're doing some internal testing um and then very soon we'll invite developers to build uh different AI services on top so besides the AI inference Service uh we will have

113. 1:06:23

     ai training service AI fun service Etc and all of them will be utilized to support our web 3 and web two partners including near um and the second goal is that we want to ensure and verifiability in decentralized AI because in the decentralized network how do you actually trust the results actually correctly generated by a Rena third party noes uh people mention about using consensus using op uh using ZK to do this verification but none of them is actually practical because consensus and

114. 1:06:55

     ZK uh caus a lot of computation uh and for op uh the finality is not ideal because usually need to wait like 7 days before uh you actually know the result is correct so at hyperbolic we invented a new verification mechanism called proof of sampling which is a joint work with our advisers uh professors from Columbia University and UC Berkeley uh which is a sampling based verification

115. 1:07:19

     mechanism so for most of the cases only one node in the network will gener generate the result based on the users request but with a certain chance uh the orators will ask another note we call validator to regenerate the the result based on the same request and if the uh the answers are different uh then it will initiate an arbitration process and the Dison note will be punished with a high level of penalty and because of this we can use math to prove that uh

116. 1:07:48

     there's a threshold based on the staking amount based on the uh Slash amount uh as long as the checking uh probability of checking is higher than that threshold the system has pure strategy n c which means that in this system everyone's every not's uh best strategy is to behave 100% honest all the time and this provide a huge uh a very uh secure guarantee for the results generated on our platform and we are currently working with a few reaking protocol to build it as a generalized

117. 1:08:22

     ABS to secure other avss building on various risk taking prot like a and kak um and lastly we're trying to protect users privacy uh even it's run by R third party node and we're L we're Lev leveraging a te with our partners to solve the Privacy problem um and so the whole ecosystem architecture will look like this like we have the decentralized orchestration layer as a foundation and then on top the people will come and build different AI services on the side we will have verification layer to

118. 1:08:54

     secure every every othero services and we also have privacy layer to protect users privacy um and um once we finish this infrastructure we will enter the second phase which is the applications and economy phase uh where many different defi AI defi applications can be built on hyperbolic such as uh like GPU future exchange AI model Revenue sharing uh AI powered Dow Etc and uh together we can f a strong economy and

119. 1:09:26

     ecosystem for AI uh community and so next uh I'm going to show you a quick demo to demonstrate the capability of hyperbolic um so like due to the internet uh right now like if you go to the ap. hyperbolic doxyz uh we already launched our AI inference which is uh service that allow you to access different a models powered by the decentralized GPU Network and we also have the uh GP Marketplace which allow you to uh rent gpus across the

120. 1:10:04

     world um and I I'll see if I can actually run the demo Because the Internet is not ideal but uh q12 VR is like a powerful VN model which is like a vision language model that can read both pictures and language um so uh if I run it um okay looks like the network does okay that's good so uh you can after like usually after just like a less than one second you can see the answer and it will say like uh it will tell you like what this image is and uh

121. 1:10:40

     we're currently integrate our service with near aab uh so besides firework uh hyperbolic is the the so providers in web3 to power AI uh near AI user own uh initi initiative uh and we would love to see many near developers uh build on top of uh hyperbolic and near to power the a a applications uh and another thing that we are very excited about is like we current we just launched our GP Marketplace which allow people to

122. 1:11:11

     contribute GPU as well as access gpus uh through our decentralized organization layer um and uh here's a video to show you how to rent it so uh currently it's like internal testing so if you want to test it uh send us the your email and then we can add you to the testing batch but very soon we're going to launch our test net which allow people to uh permission list to uh provide a compute to our Network um so the user experience is very simple you just like choose the

123. 1:11:41

     GPU that you want uh so in this example um you just want to rent uh 4090 and you want to run Alama which is like a serving engine to uh to run AI models on a GPU uh so after a few seconds uh you already get access to this machine and so you will SSH into that machine uh you're going to download uh Alama so first uh you will install like curl which is like a very basic command light uh like tool that for you to download

124. 1:12:14

     different softwares um so you will find um the the command to download this AMA software um so after like a few seconds it's downloading and then you'll install it and after that it's good so you will uh run AMA to uh as on the backend to serve different a models uh and then now you can see ama find the eight gpus and so you open a new uh terminal uh window and then you ask La Lama to run Lama 3.1 um so it's downloading the Lama 3.1 um

125. 1:12:55

     model um and uh let's give it a few seconds okay so it's successful download the the model and now you can ask uh any questions that you want uh to this a model and so the the goal of of this is trying to show that like uh a lot of AI researchers and developer are kind of uh being restricted from accessing compute uh they they are kind of GPU pool uh but with hyperbolic we allow we're trying to

126. 1:13:36

     build this like Crow uh like shared community that to make sure that everyone is no longer GPU po everyone is GPU rich and they are Avail they are able to build any applications they want uh either in web two or web 3 so thanks awesome thank you Jasper uh it has been great to work with the hyperbolic team we're going to have one of our final demos uh here from Tristan and Nilan

127. 1:14:05

     we're going to get him miked up uh but is Tristan is getting set up here amazing thank you as trist is getting set up one of the core components of our ecosystem that we've really really focused on is privacy right and as we have worked with a variety of different projects across the space Nilan has really stood out um in what they are building uh particularly with way what they refer to as blind comput so I

128. 1:14:28

     won Shar a little bit more about Vision excited for him share turn over to him cool thank you very much for that introduction so hi I'm Tristan I'm a co-founder of nilon at nilon we're building what we call Humanity's first blind computer uh blind compute is really a bit of a rebranding of privacy enhancing Technologies we found that a lot of the time when you're talking about this new field of cryptography um

129. 1:15:28

     a lot of people actually hearing about it for the first time and so it's not about trying to differentiate it's it's about trying to convince users that this new type of cryptography can really lay the groundwork for um a better internet and a better future what we're building is a broad platform to work to make it easy for developers who are not Specialists not cryptographers to build

130. 1:15:51

     with privacy enhancing Technologies and we believe that that is one of the biggest things that has kind of stopped uh uh or or bottlenecked usage of of these things because we think that they're ready to go from Theory to practice we're focused on you know we call ourselves cryptography agnostic but at this stage because we think it is the best tool for the job right now we're focused on uh multi-party computation because we think cryptographic based Solutions versus perhaps Hardware based

131. 1:16:22

     Solutions uh are more fit for higher value value data scenarios so you know there are definitely places especially for defense in depth where Hardware I is a good solution but for high value data scenarios especially like what we're going to be talking about today with oh this is not uh with AI data uh pii specifically we believe that um you know cryptographic Solutions are the key so without further Ado I'll give a little

132. 1:16:53

     overview of the network before we get straight into the demos but the network exists as its kind of own entity it has a coordination layer but really we're partnering with chains and it's been great to work with the team at near to be able to bring blind compute so blind compute and storage to the daps that are building uh on these ecosystems and then we also have kind of our own ecosystem of apps um that are building directly on

133. 1:17:19

     us when they don't need you know that that back end so without further Ado let's get into sort of the story of how we think about AI we are a broader platform than just AI but AI is a really big Focus for us especially on the research side and we first came at this as sort of uh you know an intellectual research Pursuit and built some demos based on our platforms um and this is this is kind of one of the original demos and we're doing some fact finding around how easy

134. 1:17:50

     could we make this really and how useful do we think it could be really and um this was like one of the very first things that we made it uses the classic mnist Model right for handwriting uh uh data sets um I will say mnist is handwriting and this is a trackpad writing so it's a bit different it can be a little off sometimes but what I'm doing here is I just wrote something and

135. 1:18:15

     this turns it into a 28x 28 grid and here's where the magic happens of multiparty computation so this actually gets secret shared split up into three pieces you know sometimes we say horcrux across three different nodes as seen here and when when it's across those three different nodes you can actually blindly as in none of those parties none of those nodes can see what they're holding or what they're Computing uh but

136. 1:18:40

     you can still run something over it so here we see uh it's predicting and it predicted a four great for what it's worth I picked four because it works a lot better this is about 80% accurate um twos are particularly bad um and in this first model but we've gotten a lot better since then so what we released and sort of double triple down on this is something called notada AI NADA is

137. 1:19:04

     our uh python based DSL so it was very important to us that we have developers be able to build on this without having to learn you know new things so we wanted it to be as close to what AI developers and and developers in general are used to so we have this python based DSL that kind of works like a a a python Library and it has a lot you know we also have versions of numpy pyit learn and um all of our AI stuff has pie torch like interfaces to make it easy for you

138. 1:19:35

     know people coming from traditional Ai and so these are some of the examples for the NADA AI library and this really unlocked lots of uh uh developers and and we try to make it as easy as possible in hackathons and when working with new teams to be able to use some of these things so here's an example of some of the models that we've built just to kind of give people a starting place right you have convol convolutional neural Nets you have linear regression classifier models um you know spam

139. 1:20:01

     detection and we've started going to hackathons with this and you know working on the devx to try to get people to you know even in the the short time span of a hackathon be able to create applications that we think you know could already be useful and we actually have two teams that we've signed up from the ecosystem out of the the handful of hacks that we've done that that works out about what they're building um here are just two examples from recent

140. 1:20:25

     hackathons um which is why I bring them up one is melanoma detection so you can put pictures of melanoma and it will class using a classifier model uh you know train with some amount of accuracy right um uh with these pretty small models whether the the tumor is benign or not right uh and you can do kind of a similar thing with these um this depression sort of uh uh questionnaire now just on its own if if you're not sharing any personal data except perhaps your IP address maybe it doesn't make

141. 1:20:58

     sense as to why you would want this with blind computation but the minute you start adding more pii to this what if you start asking about location data universities are using this perhaps to uh do secure analytics across different majors and and uh different types of demography suddenly pii becomes very important and this is exactly the kind of place where the more sensitive information the more what we call high

142. 1:21:21

     value data you add into it that's really where neon fits and so this is a similar thing right uh high accuracy model over uh uh depression answers so one of the things I said is important to us is devx and um one of the ways that we've been kind of approaching the developer is by having these python notebooks um and these P this pytorch uh like functionality right and so we have a bunch of these uh these are all the different things from the NADA AI library that I I presented before uh here is an example for spam

143. 1:21:52

     detection and it walks a developer through you know well put in your GitHub username do you want to enable Telemetry uh here's how you download all the pieces um especially for a hackathon we find that that's where a lot of people get stuck um and try to make it as easy as possible for them to get to the part where they're actually picking data sets where they're training their models and where they're doing you know blind computation actual classification or

144. 1:22:14

     whatever it is uh in a privacy preserving manner um and I won't run this notebook because I think you know it doesn't exactly present well but I can show you what it would look like with a more visual demonstration that we have which is here I'm loading the spam model right um and I'm uh classifying it here you know we have some examples here I could I could also go to GPD or something I'll just use one of these it's easier uh to to generate sort of a legit message or

145. 1:22:47

     maybe a Spam sounding message here you're tokenizing the message and then here similar to before right you're kind of running that mask and classify function right so it's splitting up that message so the nodes are storing it don't have any access to it and then it is running the model over them and sending out just the prediction result in the clear um yeah so so that's sort of where we are in kind of our acceleration towards getting AI private AI you know blind

146. 1:23:18

     computation into users hands and um I briefly wanted to just go over one of the more in AI projects that's that's working in our ecosystem unfortunately it's an iPhone app so I don't really know how to uh mirror it on this screen uh but um there're they're still in test flight on the app but they should be kind of coming to Market pretty soon um and they are uh a great team kind of with more uh uh un kind of NGO

147. 1:23:41

     background and they're really into self- Sovereign Ai and spe spe specifically data sovereignty and sustainable monetization of people's data so how does that work well if you go into the app it's basically it feels kind of spooky it's like a data farm right you go in and it's like please allow location access please allow photo roll access please allow calendar access please and the more you say yes the more

148. 1:24:04

     you earn right okay that sounds not great but what's actually happening is that it is training a very simple model locally On The Edge on the phone and then where million comes in is that they're actually splitting that model up and sending it into the nilon network and then in the nilon network you have all of these not privacy preserving user models but that you know the network the

149. 1:24:28

     infrastructure is blind to on top of that users can then give consent to run queries or train other models that are kind of privacy preserving right knowing what you run on top of that could leak data it's important that you know people are are paying attention to what they are consenting to and also that there's people who know what they're doing that are helping that out but suddenly you have this really viable uh data

150. 1:24:53

     Marketplace for useful data and the supply side you know works great with crypto but the demand side for a lot of this type of data that can be much more sensitive than uh you know especially on an iPhone I I don't think even Apple has all of that data right um and that can really generate some useful insights depending on on who the buyer is the final little thing I wanted to do here is show uh one of the PC's that's sort of uh well already done H this is by a

151. 1:25:27

     group in our ecosystem called skillful and this is actually private rag so me just create a new wallet here uh cool so what this will do they're building sort of a bunch of llm tooling uh decentralized llm tooling but what we're interested in right now is this neillian testnet play ground and what this is is like I said private rag so using ukian distance semantic search you can put in uh a document in this case we're going to use this blood test sample that I have

152. 1:26:06

     prepared um and it's stored in the network in a privacy preserving way nobody can see that blood test right the trust is distributed across the network you can then load it up as a source and you can ask questions like should I be worried about some of the high levels in this test and what's happening in the background is the rag is completely under privacy picking out the pieces of that document

153. 1:26:50

     that it believes to be the most important to the context and leaving out as much as it can right and then loading that as context into the prompt now I will note we're not doing the llm inference under privacy we have some internal stuff working on that but I won't go into that but uh even this depending on what you're doing the use case can already be quite powerful it's minimizing the amount of data that needs to be leaked to uh the llm providers to the infrastructure and then you know if you

154. 1:27:21

     add a little bit of anonymization on top of this a little bit of uh uh you know DP techn differential privacy technology you can really get some um cool outputs but as you can see here it takes you know the information in the extended sample blood test report and it says you know there is one high thing you got to you got to be careful about that right and you can do this for a bunch of documents that's just sort of the the live demo this is one of the

155. 1:27:47

     things that um you know we've been speaking to near about and and they're really excited to integrate on their end so yeah that's neillian thanks guys you um so folks you've seen a pretty comprehensive scope of projects here building at all different layers of the stack we have one final team uh that will share a little bit specifically about how they're tackling the data problem two final teams I get excited and I always forget our agenda two final teams who be shared a little bit about

156. 1:28:19

     the data problem and how they're solving that um and so we're going to turn it over to uh chi uh who is the co-founder and CEO of Zeta block uh Zeta block uh has been around uh for a number of years focused predominant on data indexing um but now really focused on building out a more comprehensive AI platform that is able to leverage cross chain data so we'll give Chi a couple of minutes to get set up here and then we'll turn it

157. 1:28:44

     over to her e hello oh sorry thank you hey guys uh can you hear me um yeah okay thank you of course um hello I'm Chi I'm the

158. 1:29:46

     co-founder and CEO of Zeta block uh we are an open development platform for AI economy 3.0 so a little bit background about myself uh before Z block I used to lead product management at a at data bricks a web2 data and AI infrastructure company and before that uh I did my PhD from UC Berkeley focus on machine learning and causal inference AKA go AKA ethical AI or responsible AI um and uh uh the first AI startup that I was

159. 1:30:18

     building uh in 2018 uh was an automated machine learning uh Automated machine learning platform company so it's amazing to see uh a new wave especially with a J that is pushing AI to another level of adoption and intelligence so uh a little bit uh just a background about Zeta block we have been building the solution since uh early 2022 and uh uh what we have been building uh in the beginning was trying to you know really tackle uh the AI

160. 1:30:49

     landscape from data uh the reason is simple uh as we all know data is actually the fundamental oil for everything that you can come up with models and any intelligence so that's why in order to make model really able to achieve what we want and have like a high quality high performance model we have to have high quality data and with unique patterns that you want to insert into so that's why uh in the last two and a half years we're very honored to work with a lot of great projects in the web web 3 space and also like web two

161. 1:31:20

     space especially at this time uh as we uh mentioned here a lot of crypto companies such as uh I lay store protocols s and also near uh we work with them and uh um a little bit context about what we are building we basically build a open and uh uh trust free AI development platform that's enable uh guys to you know really Foster a open and accessible ecosystem of data and models so what does this mean so basically I'm going to give you a live demo later on to show you that how

162. 1:31:51

     developers can prepare the data into the model ready format and then uh use the data to train the model uh on top of uh our development Network and then they can influence the model and use the model for agent building or like some applications building so uh let's dive into the uh demo uh I'll give you the context about the case study here so basically uh open Zing is one of our customers they build a lot of uh onchain

163. 1:32:17

     risk detection uh fraud basically fraud transaction detection uh bots on top of Z block and how would that work is basically they combine a lot of like onchan transaction data and also a lot of like offchain label data and also like some kyc data uh into zock and then they did the data wrangling and data preparation and use the data prepared uh to train a machining model so at the time it's more like traditional models uh and then uh once the model has been trended you can use the real-time

164. 1:32:46

     streaming data set uh data uh from Z block to do the real-time inference and now this time what I want to download to you is instead of like what open zaping build this is actually a model that we co-build with uh uh Google so gcp and uh let me show you one thing which is quite interesting so as we on uh some of you guys probably heard about this rack pool uh about like 200 days like a seven 76 7 eight months ago so this is a rack pool um that Abra

165. 1:33:19

     kabra money has been uh really like uh you know lost a lot of money so we we have this uh model fraud transaction detection model uh built on top of a transformer so we basically fine-tuned Transformer with a lot of like those unique uh labeled uh fraud transaction data and here is just to show you we can predict that this transaction is really a positive for transaction with a very high level confidence score so how did

166. 1:33:51

     we do that now let's take a look look at and as you can see it's very quick and how did we do that uh let's start from here so this is our um this is the platform we already released to the public uh later on uh this year we're going to release the AI development platform uh uh with a ga uh availability as well so this is to show you that you can now turn any transaction data into uh model ready format that is basically turn any transaction hash to uh in uh in n Tres and what does that

167. 1:34:25

     look like is basically look like this so we prepared uh the trending data and test data and this is like a for example like the test the data with the label label one means this is a positive fraud detect fraud transaction and then this is a lot of like a label data uh we actually collected from multiple uh auditing company Partners uh that are willing to participate into building

168. 1:34:49

     this uh uh Cutting Edge uh fraud detection use cases with uh Z block and uh uh gcp and then basically after this has been prepared now we can then go to uh The Notebook to start building this uh uh building this uh model on top of the Transformer base uh which actually is a birt and we turn any uh unique transaction U like address uh into like specialized the token so how did that do is we prepare the data um to you know

169. 1:35:20

     like split that into training and testing and then uh tokenize the the address and then uh process the data and now here is getting uh uh interesting so basically we start training uh the Transformer uh model on top of uh Z block Network later I'll will show you like the command line and basically once that has been Trend uh and uh we published the model actually to uh uh uh hugging vase as well uh to show to people and uh then you can do the

170. 1:35:49

     inference a realtime inference on top of Z block either using like a notebook to show or using API or like a terminal end points just like what I showed uh on the terminal and here one thing as you can see if I am a developer uh and also building for a specific use case what I want to really optimize this use case for is depending on the business needs here so in the very beginning we uh started with a model that has very high accuracy but the recall is not necessary High which means that is not being able

171. 1:36:21

     to you know detect the transaction fraud fraud fraudulent transaction well but in this case it's very important I would rather uh have more fast negative a fast positive than fast negative because if I have a lot for fast negative then that's probably means I'm going to really screw out my my my uh treasury or like the money that on on top of my protocol so that's why uh we iterate the model building a few times and now as you can see the recall score

172. 1:36:52

     has been improved from uh almost like uh 60% to 80% uh which in this case you can of course further improve that and uh anyways so the point here is as you can see now once the model has been building you can use the you can use this model to do any prediction so now let me do a real-time prediction here we can go to Ether scan and we can take a look at the latest uh transaction so this transaction has just been pushed uh in

173. 1:37:23

     uh 11 seconds ago and we can take a look at this oh so as you can see this's most likely a negative uh negative fra fraud transaction prediction so which means we're good to go and uh then uh also this is uh just oh sorry that's bad by the way this is just to demonstrate that how uh the entire transaction and also like AI training uh you know like a data preparation and model training and model uh inference uh process has been recorded on chain um so this is just an example to we

174. 1:38:28

     demonstrate that how this we deploy the smart Contra onto op test net to demonstrate how this work so basically as you can see uh we can recorded all those transactions with a different type of smart contract that we design we Define here underline we actually Define uh over 15 smart contract uh uh and prodal underlying to track each of the step in this AI uh preparation basically

175. 1:38:53

     data preparation and the model training and model inference process um yeah so this is pretty much like the end of the demo and uh as you can see uh Z block is uh uh a data and also like a data and AI company uh which is buil for the open uh future of the a economy 3.0 and uh uh reach out to us if you have any questions and stay tuned uh our testet is going to be launched in a month and also there's a lot of exciting uh AI capabilities for developers uh to build

176. 1:39:24

     on top and the most good thing here is you will have exactly the same experience as we what you used when you used like a data breaks where like any scale U together AI Tope of a web two uh solutions to build models but on in the meantime we can guarantee and track the prominence and attributions of the entire process on chain thank you thank you awesome thank you chi and thank you folks uh for bearing with us I know this

177. 1:39:59

     has been a lot of demos we're going to close it out we have our final founder who is coming up here Brendan uh co-founder of Masa AI really really compelling use case and one of our data Partners uh specifically building a Data monetization Network uh we met the Masa team at consensus in Austin Texas in the US uh earlier this year and we have been so so excited to work with this team Brendan I think we got to get you miked up so you'll be good to go um but this

178. 1:40:29

     uh Team specifically Masa really aligns with our vision for user owned AI um and particularly what they are enabling for everyday individuals to monetize their own data so Brendan's getting miked up here but we're going to turn it over to him here our last demo for the day before closing Keynotes thank you folks okay good afternoon everyone hope this hasn't been too much of a long day for you all I'm going to keep this quite Punchy uh let me see if I can actually

179. 1:41:11

     get this to go full screen into a slideshow that would be nice cool so I'm Branda from Massa um I'm the technical co-founder of Massa we've been in operation about two and a half years uh it's really great to be working with the near team here um really love what is being done with near and Ai and we've been building a fair AI platform that allows anyone anywhere in the world to

180. 1:41:35

     both access data for free right now and contribute to the huge growing data set that is growing exponentially every day and we'll dive into both what our data is how we aggregate it and we're going to show you a rag agent that is built using specifically Twitter data today so we're be focusing on building Unstoppable ai ai applications with our decentralized Network and nears aiab aiub I'll get my words in a moment um

181. 1:42:06

     and what we mean by Unstoppable is with decentralized technology there's no kill switch there's no off button if you put a gun to my head you can't shut down our network no matter how hard you try or how hard you force me um and maybe you've seen this a bunch over the last few days but I always was like start here this is what we see as the fundamental building blocks of decentralized AI compute data model of

182. 1:42:30

     model and inference and on the data side you've got things like public web gated we specifically unlock through scraping Twitter data which has got huge demand in the web 3 space and then this personal data you know Apple's doing this on device inference that's using like te or secure enclaves to make sure no data leaves and you can get personalized AI in your device which will also see over the next couple of

183. 1:42:55

     months so at Massa we um have this ecosystem of Miners and workers that are fully decentralized we essentially decentralize what scale AI does bright data and many of these other data scraping platforms there'll be a bunch of people out there that say all the data has been scraped but that's actually not true since I be since I began talking there is more data been generated you know thousands if not

184. 1:43:18

     hundreds of thousands of web pages thousands of posts none of that has been scraped so you're always going to need a constant feed of information that becomes the eyes and the ears of AI like that's what we do we enable eyes and ears for AI make agents smarter give them memory and then unlock use cases like you know injecting context into system prompts doing rag retrieval augmented generation if you want to get really fancy doing llm training or fine tuning which is definitely more complex

185. 1:43:48

     than the former and the majority of use case that we see today on our platform is very lowest common denominator there is so much low hanging fruit in this space there is no need to make things complicated simply give people data allow them to create embeddings put it contextually into the right context in an agent and you're just going to get massive like high value for your end users and build products that users love

186. 1:44:12

     we do have an ongoing like early stage part of the platform which is using specific llms for structuring annotating and diarising data this is very ear we also do speech to text for things like podcasts um YouTube videos things like that for conversational AI to be trained on that data and you know what are we doing on near we're enabling agents and data access on the near Hub uh every day we have a stream of index

187. 1:44:43

     tweet data that's about 2 million tweets a day all contextual for um the crypto space and you can also access real- time information but decentralized AI needs open permission list data for training uh training models and giving agents memory and today we're going to go through it's 7even minutes but I'm going to skip forward the video a demo of using uh Massa scrape tweets to create a crypto satum meme coin agent this is a a

188. 1:45:13

     typo in my tiredness yesterday um they're going to use real time and index tweets can spot Trends and it can reason if you use it with something like gp01 or chain of agent like logic uh and we're going to be using Lang chain um local version of llama 3.1 running on my on my laptop and uh the Massa protocol to make this all work so here is my video I decided not to do all this live

189. 1:45:44

     your starting point and your journey for Massa is to go onto our GitHub repo get up and running with one of our protocol nodes use some of our tool tools like our python SDK that allows you to use our CLI for scraping go to our Dev dos um and generally just get starting up and running we're going to post a single page agent in the near Hub along with all of our data sets so you'll be able to go there and access all this

190. 1:46:09

     data for free or indeed come to our API and make organic requests in real time to access all this so this is basically taking you through setting up an individual node on the network you got several ways to get connected you can run a single node in private configuration configure as many scrapers as you want from one boot node you could similarly connect to the main Network and access the entire pool of shared resources or come and join our bit

191. 1:46:34

     tensor subnet on subnet 42 which is what on a daily basis is indexing a lot of tweets and will'll also earn you about 50 bucks a day if you're running a minor to scrape Twitter data right now um so I'm going to I'm going to skip forward this because this is all a bit boring and hindsight this look far more interesting there's a bunch of repos here that um will show you how to get started we've got an awesome Massa repo lots of good code examples in there you can dive in and get up and running with

192. 1:47:04

     an agent within like five minutes um and this is just demonstrating how you can retrieve uh information on God's Twitter profile um this is actually going out on my machine here fetching data from workers this is going through the Twitter frontend graphql API in point that we reverse engineered there's tens of thousands of Twitter workers we have on the network that are scraping and uh helping unlock this data and similarly you can access followers and write the way through

193. 1:47:36

     to uh tweet data actual raw tweets and the search functionality through the tweets endpoint is very complex you can do between date ranges you can do by keyword you can do really sophisticated and all queries and go down to hashtags specific people that you're following and build a very very very detailed comprehensive data set uh to be able to use in your AI agent and this is just

194. 1:47:59

     giving an example of the raw data that's returned for the latest 10 Bitcoin tweets so this is just doing how you can like in real time scrape using one of our Python scripts um just going to go forward here because you're all getting to the end of the day and we're actually fetching tweets on the # mem here so these are being directly queried through the um protocol this will sleep when Twitter

195. 1:48:32

     gets rate limited and it'll come back on and manage your connection to the actual scraper and and and from here you can spin up very quickly a streamlet application using uh Lang chain uh it will use their local embeddings it will load these tweets that I just scraped I actually did this this morning forgive my um lack of formatting here I'm a backend engineer not a front end engineer um and uh we can then have a

196. 1:48:58

     conversation specifically about the documents the Twitter data what we've done is we've taken the text field we've merged that with the author of the tweet and the date and we've then passed that through an embedding endpoint which if I skip back you would see this all being done in the command line but I'm going to Bear you the pain of watching command line prompts going through their cycle

197. 1:49:20

     and we can start having a conversation with this agent this this is extremely lowest common denominator this is very very simple you can pair this with inference that's on chain uh this is actually giving a nice output for once and it's just going to tell me more about the meme coins that are being talked about unfortunately on salana um but but that's where meme coins live um so yeah this is this is what we do we make data incredibly easy to use with the data itself tooling and I'm really

198. 1:49:53

     excited to see what we can kind of connect to in the near near AI ecosystem so people can start building real products right like the longer it takes you to get to this part this this point in time the harder it's going to be for you to start testing with users understand what their needs are and stop talking at conferences and events like this and like actually engaging with people I'll be rather doing this than wasting time talking about site we're

199. 1:50:17

     not building and addition to this just to close out you can also scrape web data through the protocol Discord Telegram and all others and this is just demonstrating how you can get context on Meme coins from Wikipedia and we'll do the same thing with um coin market cap to go and get the most relevant data for uh meme coins from coin market cap which includes pricing data so you can merge

200. 1:50:41

     these two like building block uh bits of data infrastructure together to give you a a great output for your for your data needs so I will stop this this video and Skip forward okay one second how to participate all right go to Massa or very soon the near Hub you'll be able to get up and running with that streamlet agent in a single page file on the runners that are in the near Hub um and you can also participate by being a

201. 1:51:12

     minor by scraping data and very soon pre-processing and you'll earn Master admissions for doing that and probably to well not probably definitely to if you do it on bit tensor as well so thank you very much have a great day amazing thank you Brendan folks we are on the home stretch thank you to the faithful who have been with us through this experience um but you're in for real treat as we close here uh with our

202. 1:51:39

     final keynote um and just maybe to Echo Brendan as Alex is getting set up here you know a couple things you've seen a variety of different use cases across uh the near ecosystem which we're genuinely excited about thank you um that we're genuinely excited about yeah we got to get you miked up that we're generally excited about and so U hope you can understand that AI on near is very much real you know we intentionally did this as a more technical uh and product

203. 1:52:04

     focused experience and event largely to a not just have another set of panels uh but also be show and and not tell right so we are really really excited about how these use cases have been coming to life and then lastly just to introduce our final speaker here um Mr Alex skof has an incredible background as a researcher um and has contributed to this the field of AI in some really compelling ways um and I think in the time I've been around Alex one of the things I've admired most outside of his

204. 1:52:34

     obvious research expertise is I have found there's actually very few people who are genuinely excited about the technology in this space um and to I think hear from Alex uh is a real treat but also to hear his genuine excitement I think is is rare so um Alex in the spirit of showing and not is going to share a little bit about some of the things that are happening particularly through near AI in closes out today so

205. 1:52:58

     I'll turn it over to him and let's give a round of applause for Alex how you guys doing good okay I'm Alex I'm from near uh let's talk a little bit about artificial intelligence um some of you might know some of you maybe don't but uh near started as an artificial intelligence company both myself and Ilia my co-founder coming from AI backgrounds uh and earlier this year we started this new effort near AI the the idea to started it was the's idea

206. 1:53:32

     initially uh it uh it comes from uh sort of looking at the current state of the world uh through the lens of n's Vision n's vision is to make sure that in the future people have power uh and control over their data assets and power of governance uh and if you look at the sort of first derivative of the world today uh and look at what happens uh with artificial intelligence first of all there's a a hypothesis which I think is not crazy in 2024 that AI will be

207. 1:54:02

     transforming the way we interact with internet the way we interact with services with the world uh and uh the way the world is moving today that uh um that capability in the world will be fully captured by the small number of profit-seeking entities this is what happened to mobile phones this is what happened to social networks this is what happened to uh literally every uh technological revolution in the past

208. 1:54:30

     with a single exception of blockchain uh and uh the problem is that once that attention is captured by those profit seeking entities they start capturing data the data will goes there uh their models become better and better it gets harder and harder with every passing year to catch up for the uh incumbents in the world uh and uh effectively the the goal of neari is to make sure that there is a sustainable path towards the future

209. 1:55:00

     where open-source um user owned artificial intelligence is competitive with the uh with the artificial intelligence offered by profect seeking companies uh that is a multi-disciplinary effort it includes uh empowering uh um ecosystem uh companies and projects uh it includes building infrastructure where they collaborate uh it includes I think on the uh in the long term and on a higher

210. 1:55:32

     scale some important aspects uh that need to be figured out one is how do we actually ensure that uh base models continue to exist in open source today we entirely dependent on Facebook uh for that to continue and Facebook is not exactly known uh as a company that is uh not lishing World Gardens um uh and uh slightly more interesting uh aspect is how do we ensure that uh tier one research does does not stay within uh those silot companies so today we happen to be uh at a relatively rare

211. 1:56:11

     um I think point in time where majority of the research which Powers those uh tier one models is known outside of those companies but that will not continue right like even if if you look at uh the latest generation of open AI models o1 uh the technology behind them uh is for the first time uh is actually not known outside of open a right we have we know it's some something called

212. 1:56:34

     strawberry but what is strawberry uh nobody knows right uh some sort of Monte Carlo research we don't know right and so we need to make sure that uh researchers so unfortunately today the situation is such that for majority of researchers are going to open AI is the absolutely best thing that can happen to them and um that is a problem right because uh that means that the best research will continue uh to happen behind the closed doors but there are people at the end of the day AI research

213. 1:57:06

     AI researchers are same people as us they are developers and developers are by Nature anarchists uh and you know many anarchists do not want to work for open Ai and so what we need to to make sure is that there is infrastructure there is access to resources there is uh uh place to collaborate access to data for those researchers who do not want to go to open AI today it is not quite the case it is very hard to to pull those

214. 1:57:30

     things get together very hard to apply very hard to apply for Grants and we're working actively on building uh that infrastructure researchers can work together uh I will not be demoing that today uh it is at the relatively early stages today uh but something interesting that uh uh when I was thinking what I can demo what what cool I can show one of the developers at near V uh he was playing with our um infrastructure and he built this little agent uh as a toy uh and he was showing

215. 1:58:01

     it to us uh at our uh weekly sync uh and uh I found to be very fascinating and I hope you will find it fascinating as well I will show you uh how it works then after I show it to you uh I will uh uh I will spend some time to to highlight certain implications right in it's not going to be uh even remotely as polished as that uh for example um what was the first demo uh yes Quest flow I think quest flow was uh uh extremely

216. 1:58:33

     streamlined and I was thinking of my demo I was like yeah that's not going to be that's not going to be like that so so this is something that was built by a single person uh in a week on top of infrastructure that is barely existing today so don't expect too much but I think the the thing is still going to be pretty cool um Okay so this is a uh if you guys familiar with the um service

217. 1:58:57

     called Web Sim uh this is going to be resembling webs uh if you're not familiar with webs you should familiarize yourself because webs is webs is Magic um so this is a tool that builds applications on the flow uh and and web Sim is the same as well uh but this one uh has a Twist and the twist is that it has access to uh through augmented retrieval augmented generation R for short it has access to a bunch of

218. 1:59:29

     uh uh near specific information so it has access to um uh contract metadata uh function signatures uh Etc uh and I will show you a quick demo here uh uh and and V was fixing things literally like two TOS away so let's see if it's how how well it flies show show me my let's see how well I can type with one hand near and okay let's go uh so this uses uh something we call inference router an

219. 2:00:08

     inference router is a proxy uh that rout uh requests to to various uh uh to various inference providers right so hyperbolic is one of the backends there uh it's a um cool okay so uh this is the code so okay so I I sent it a request it it generated a page uh on the Fly you will notice that it uses a um you know Near View WAP near Ft balance off so it's familiar with the um I I don't know if the if the font is is small let me know I I will make it bigger but so it's familiar with the uh

220. 2:00:50

     ABI of the contracts uh which it pulls from uh uh through rag right and it shows me this is my hot wallet it has uh an active balance right and um okay and so that then you can start doing some cool things like you can uh you you can you can like uh actually let's let's first uh iterate over it uh add a button right miror uh so uh under the hood the way it works is that there is a chat uh with an uh there's an agent and a chat with a model uh and the model uh

221. 2:01:40

     has okay everything broke so at the end of the day obviously the goal is that uh you you uh you don't need to be technical to work with this or didn't break it just uh it's just slow yeah okay here's the button I have no idea how much it will wrap let's see what does it do here um the R button oh it will ask me for the amount of near yeah yeah it will ask me okay that's not that's not good

222. 2:02:20

     uh all and so so you saw I just did something in the context of an element right so uh the the way the agent is instructed to work and this is something we took from web Sim is that it is instructed to give ID to every element uh and then there's like a little thing on top of the interface that highlights elements and when uh when I entered something right now it was in the context of the element that was highlighted so the model knows what it

223. 2:02:43

     needs to do right and so I don't know how much I trust the code generated by an agent okay let's let's give it a try um and so again you will see that it uh it can take action transaction successful thank you uh yeah balance is updated I have more app near uh so it it can take action on my behalf uh the way it works is that it lives in a knife frame uh and the outer frame has access to my account and an API uh to interact with it right so that means that every action goes

224. 2:03:25

     through the um through the wallet it's a decently good security probably you shouldn't run it with an account which has like uh all your life savings but for uh uh for things like that it works pretty well and it can do uh yeah you can do like some make it you can do some fun things with the like you know model models are very good at doing front end because there is a lot of front end code online

225. 2:03:58

     um so so besides the fact that uh AI developer is probably very close right because like this is very oh rainbow uh this is very uh like I remember the first time I was playing with Devon Devon was relatively slow and it like takes time uh and iterates right but but this is practically real time right like it took us 5 minutes to build a relatively simple application and so uh the way I like to think about it so so in the short term we will see a lot of

226. 2:04:32

     things like Quest slow beta where you can interact with an agent agent will be taking action for you in the chat interface uh but I think sort of longer term the transformation that will happen is that tools like that will become way more mature uh right uh people will move way Bey Way Beyond something that can be built by uh by one person in a week right so uh imagine this with a Chain of Thought with certain validation that

227. 2:05:00

     will not allow it to Break um uh that will iterate on top of mistakes look at the console uh errors Etc right and you get like a sophisticated uh application building tool that uh if you if you try to recall the last 10 minutes nothing I did uh required any I just I mean you need to understand what wrapping near is but given I'm building a tool for wrapping near I probably understand what that means right so so we will get into the world where people will be able to build

228. 2:05:34

     applications on the spot and I think a very interesting uh implication of that uh that makes me personally very excited is that uh I mentioned at the beginning right one of the big things we care about is data privacy but if you think about the way we interact with applications today uh we default the infrastructure of the modern internet is no privacy by default what I mean by that is that if you open a website uh you really have to trust the developer

229. 2:06:04

     of the website for your data not to go away if you like d go or if you go to Unis swap.com and try to swap you have trust in the developer that the website is doing what it claims to do right so like not things like um private note taking or or like private messenging they all exist uh in the context of you trusting uh the developer if you download signal yes signal protocol is private but how do I know that the application itself does not send the message that I just typed somewhere else

230. 2:06:36

     right and even even if I spend time to reverse engineer signal the very next up upgrade that happens uh you know it it all goes to uh it all disappears but with technology like that uh you can imagine that the world will be fully transformed into into the state where companies do not provide you with an app itself they provide you with an API which is documented which lives uh in

231. 2:06:59

     the uh in the system right and then you build an app uh for the particular use case you have today uh on the spot right as inference gets faster right uh a request was taking on average 10 seconds right so inference will be getting faster right imagine that taking couple seconds for per per turn right you will uh if I want to have an application which which works on top of signal API I'm not going to be downloading signal I

232. 2:07:23

     will be just building I will be just saying hey I need an app which works with signal uh I have it locally I know absolutely for certain that my data is not going anywhere so I think this is very exciting right and it's also exciting From perspective uh uh of uh if you if you if you look at how the world was changing like back in the day when I was when I was little uh in the in the in college I remember as like uh you know like there was a protocol called SQ uh and the

233. 2:07:53

     default well default default client forq was complete garbage but people were using a thing called a thing called quip and qup was sort of very appealing it was like telegram client today it was very appealing to the masses right but for people who did not like it there was another client called Miranda and Miranda was like you can do whatever you like right a bunch of plugins it can look like anything you want it can have uh any functionality you want uh and

234. 2:08:16

     that world sort of is over today if I want to use Facebook Messenger I have to to use their bloated uh unusable application and you know if if the person I talk to is on messenger I have I have no choice I have to be doing that right again that will you know technology like that uh will be changing it because uh now I build application on the spot for my particular use case in

235. 2:08:41

     the moment and in the moment you know that can be as powerful as I needed as long as API allows it to exist so I think this is um this is very very exciting uh and uh we uh at near we we put a lot of effort into building this concept of a developer we want to see I I think in the short term in the medium term it will enable uh way more people building applications uh on top of blockchain uh and yeah it went to sleep uh yeah okay guys I don't want to go to sleep so I

236. 2:09:15

     will I will I will finish here but I hope I hope you're as excited as I am I think it's uh great time to live uh and great time to be a corgi thank you I think we're going to close with that folks you know uh I think Alex's laptop went to sleep but we're very much awake very much excited uh as well and genuinely appreciate not only Alex but all of our Founders who demo today uh I know folks will be around here we have the space until 7 I believe so get some

237. 2:09:48

     snacks get some refreshments enjoy thank you all for coming out