NEAR AI Tech Showcase

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

1. 0:30

   oh

2. 4:28

   ready oh

3. 6:23

   ch he

4. 8:02

   all right folks we're going to get started in about 2 minutes 2 minutes we'll get started

5. 9:27

   n n all right folks welcome happy Thursday to day three of ECC and more importantly welcome to the near AI Tech showcase my name is Jared Barnes I'm our head of founder success at near incredibly excited uh for us today to dig into some of the amazing Founders and builders that have been building uh AI on near

6. 10:32

   today is actually all about really unveiling uh some of the founders that we have been working very closely with over the past few months as part of our incubation program and you're going to hear some hear from some amazing panelists and our founder and CEO Ilia pukin later today really dig into what we're doing in AI just to tee up our our time today uh as well to to give some context for folks who aren't familiar

7. 10:56

   with near we believe in a world where all people control their own assets data and power of governance and in order to do that our mission and function specifically through this incubation program and what you're going to hear through Founders today is actually incubating this user owned AI stack and really making this real what does that actually mean what is user owned Ai and what is this stack again Ilia is going to Deep dive this

8. 11:22

   for you later today in depth but so you understand you're going to hear from teams at each layer of this stack from uh the data layer um data monetization and data acquisition to compute um inference and model training to uh agent Frameworks and really really exciting things that are happening at the consumer layer as well um today there have been a lot of uh panels there have been a lot of decks I'm sure you all have seen uh at this conference which

9. 11:46

   are absolutely amazing but we really want to get into the tech um I was talking to some folks outside in the lobby and want to reiterate each one of these Founders has products that are live uh and ready that you're going to really get to see today uh and so we're excited to make AI real um and really allow you all to interact with these products um and really dig in and so two quick calls to action because I actually want to get off of this stage and get

10. 12:09

    our Founders on the stage um if you are an AI founder building in this space join our community um you can scan this QR code um we have an amazing telegram chat you can talk to any folks that are in that chat to to learn a little bit more and then if you haven't already check out our AI office hours um Ilia pukin and Alex skof do an amazing job they just did a really really cool whiteboard session um actually talking about the importance of uh model training and what's happening in the

11. 12:32

    space um we do that Weekly um you're going to see the next one on Thursday so really really excited um if you want to join our community please please tap in uh as well and so folks we're going to get right into it like I said today is all about products um our first founder um is named Shu dong uh He is building a project called Mizu they are solely focused on synthetic data and open

12. 12:55

    source data sets uh shu is based in Singapore and was not able to make it today but he did pre-record a demo so we're going to start off and kick it off with Shu hearing a little bit more about Mizu hi everyone uh my name is Shu and I'm F of Mizu Mizu is decentralized open Data Network and our goal is to establish the largest and first open source de Network in the world so this

13. 13:20

    is like what we have been built today here you can see that this is the this is the real Matrix you know we're import like there's the strong pipelines behind the front content really importing the data into this kind of network and this data is all all like real time data so we already like processed about like 300 K records and there's five billion tokens already been processed and total data size is 300 megabytes like we're

14. 13:44

    ask me in next two months we'll be able to process one terabytes of the data in our Network and we have today we have like around 300 total data reos and uh we have committed about 3K commits and every commit will contain 100 records into this kind of dato uh the very first thing we're doing here is that uh to boost strap the whole open source Data Network of Mizu we're importing existing

15. 14:09

    data sets open source data sets into our Network and so we have like strong Foundation that uh in the future like when anyone like you know who looking for the data they can find the data they want in our in our data network from this kind of Open Source data sets and the very first data sets we're importing into our network is called dhma and dhma is the open source data set empower the very first fully decentralized open Lal

16. 14:31

    model uh omo open which is in like released by The Island institution and Domo has about like three trillion tokens in total compressed there about like 3.5 terabytes of data if decompress that I think it would around like 10 to 15 terabytes of data in total and we have uh what we're doing here is actually we are using large language model to process the whole domal data sets and we set up like hundreds of

17. 14:59

    different categories and we classify these kind of data sets into these kind of data categories and each data category will create we will create a new data REO for each data category so take this kind of data re as as example this data category is actually automatical and house data from Domo data set and uh you can see that we already like make hundreds of commits into this datao and here is the overview of the data set for example like this is the description this is the rewards Port

18. 15:28

    because every ribo has the address so it can hold some token or rewards in the ribo and this is the publisher address basically the one who published the ribo now everything is published by Mizu but soon we're going to open the gate so people can create REO for themselves and people can contribute data into the REO and this is the admin and committer so basically you can set who can really commit data into the REO for example you can just like allow the reputational

19. 15:54

    ones or like your teammates to commit a r just like GitHub right it's like a permission control stuff and uh this is the activities and this is the rules so rules is very important actually in our data REO design basically this is more like the pre-check of the Comm commitment so every data like committed in we will use a AI validator to go through all the rules to ensure that data really falls into the category for

20. 16:18

    this one we don't set any rules because now every data is committed by Mizu and we trust ourselves but in the future we can add a bunch of rules for example like we have this kind of descriptive rule R saying like all the data has to be text and the data committed in has to be medal has to be about medal and house and if it's not it will reject it and you cannot like commit data into the r and you can add a bunch of rules to Define and decide what kind of data can

21. 16:44

    really be committed into the Reel and this is the sample data yeah this is sample data because we know like uh there's gyes or like tons of data in there we don't think it's reasonable like to show all the data so we just like show a list of the samples but in the future we're going to have a more Det reports and if you really want to know all the data in this kind of list we're going to provide this kind of a decentralized data storage link for it to go through that view it's more like data Explorer thing and you can see that

22. 17:08

    the data is actually Json which is uh copied actually from uh from from the Domo imported I would say from the Domo data sets and this is the real commits if you click into each commit you can see like what kind of data is committed by this reibo and how much how many data records is added the type the commit message and who commit the data and also like it will show like how many tokens

23. 17:31

    it rewards it own by committing the data for now like we're not like you know rewarding users because the reward Price p is zero in the future we definitely can design this kind of bounty system so we can incentivize the community to jump in to help you to contribute the data and this is the applications uh applications means that we're actually trying to help developers help help developers to streamline the AI

24. 17:55

    application development process by uh building applications from the DAT R directly and uh uh actually you can build tons of different applications from the r right you can build a public knowledge base like a V database you can build a chatbot you can build agent you can even like build a fun model from this and uh uh we in the future really want to help developers to streamline the process so they can build applications from Theo with one click there no hassle and this is basically the demo

25. 18:25

    and uh this is very first stage as I said you know we only import the data in and now the community cannot really commit the data and cannot really like create the R by themselves but assum in the next two months we will be open the gate and adding more features so users can really jump in to try by themselves this is our like main website and feel free to check it out thanks for listening to our presentation and you can try out app and read more

26. 18:49

    documentations in the website and here's Mizu the first and largest decentralized open dat NW have a nice day exced build we've been inting eem right so was really one of the first teams we started working with very very excited to have them as part of our ecosystem and next up we have another really really important piece of our AI sta which isoud jemy

27. 20:09

    sorry just 30 seconds and we should be all good app amazing I love Live tech support right here amazing thank you perfect thanks okay perfect okay hi everyone um so I'm Jeremy representing uh representing sorry hyperbolic I'm leading business operations uh at hyperbolic and uh we are basically um building the open access AI cloud with a mission to making AI accessible to everyone for this we're tackling three main goals so goal number

28. 21:06

    one is basically this GPU orchestration layer uh you know the classic aggregating idle GPU resources from around the world from multiple different sources we're talking about data centers on-prem machines individual machines mining Farms etc for this we built our very own hyperdose which is our hyperbolic distributed operating system um we usually to say that we write consumer software uh basically you know for clusters not for computers um let me give you a very simple example let's say

29. 21:37

    that you know uh you have a sitting idle GPU resource uh you know or machine at home let's say you have a 3090 uh you know rather than joining hyperbolic Network as a nodes you'll be joining a cluster of 390s uh and this is basically to build Services you know on top of it so for now we have the inference engine um and it's basically to have a very scalable and robust backin architecture um goal number two is verification so basically making sure that we can verify

30. 22:06

    the entire AI pipeline so for this our CEO Jasper um actually wrote down a research paper uh called proof of sampling so if you have any questions uh later on uh just you know hit me up and I'll be able to answer any of your questions but the idea is really to to make sure you know to verify the um as an example like proof of inference for instance is to verify the outputs um of

31. 22:30

    you know the computations that are stemming from the third party nodes on the network uh problem number three is uh still work in progress but it's basically data privacy so let's say that you know you're a user I'm a compete provider on hyperbolic being part of this cluster as I mentioned to you before um you don't necessarily want to have a copy of your data being kept on my machine my third party node which you know can be based anywhere in the world

32. 22:53

    um so for this we're utilizing tees which is you know trusted executed environments and basically um making sure that we can have this confidential Computing and competes with traditional Cloud providers that usually are leveraging this through reputation like you know for AWS Etc as well as legal contracts so this is where we also want to you know revolutionize um so we also have a very deep and we're super proud of the integration we have with near um we it's still work in progress so the demo that I'm going to making today is

33. 23:23

    mostly going to be around the two products that we're catering to the first one is this GP Marketplace and the second one is the inference models um just briefly on near AI on the integration just so you have an idea of what we're doing and again if you have any questions afterwards you know uh please ask um the idea is that we're basically utilizing the um you know um hyperbolic inference endpoint for the

34. 23:48

    near AI users in order to handle you know AI tasks uh that they need so whenever there's a request it will be routed basically to the hyperbolic you know uh decentral ized Network secured you know by the proof of response and the proof of sampling um so yeah so let's jump maybe in a really really quick uh demo um so let's say that I'm so basically this is our dashboard it's live you can you know basically it's app

35. 24:13

    that hyperbolic that XYZ anyone can access anytime for now um let's say that I want Tobe be quitting this let's say that I want to rent this a100 uh PCI so I'm just going to be clicking on rent I can confirm here so it's been hardcoded so it's take it takes a specific amount of time uh but it's going to be really really quick and then so you can here click on manage and so you can basically copy the SSH here so now I'm going to go on my

36. 24:53

    terminal by the way I'm not a coder so it's super easy for anyone to do to do that um so let me just paste yes so now I'm in um okay so as you can see in here let me put it a little bit bigger I mean I can start using the GPU on the marketplace so as you can see it's extremely easy for the user experience um I can help help answering a lot of the questions maybe not all of them but if you have more happy to introduce you you know to our to our CTO we can connect on telegram as well as our founding platform engineer who

37. 25:37

    worked on this now really really quickly on the second um demo so here as you can see I'm going to stop renting um and so as I mentioned really briefly before you can basically connect with like your Google account onto our app that H XYZ and you'll get $10 free credits to start using the gpus on your own or start using the inference that I'm going to be showing you um in in a couple seconds

38. 26:10

    so this also has been hardcoded so that's the reason why okay perfect now I'm out and now I can just go back I don't have any instances as you can see here and I can go back on the GPU list um just be aware that this is sort of like uh you know pre-launch so that's the reason why we don't have a lot of computes um but it's it's uh coming over so now in terms of the model so what we're doing is this is mostly on the

39. 26:33

    side of the founding AI engineer who hosting open source AI models on our onto our platform leveraging our distributed you know Network um so as you can see we have VM Vision language models we have text to text text to image text to speech you know it's it's really broad um as soon as you know text to video will be of a good quality on um you know like open source models we'll obviously be integrating those uh here and hosting those here so let's say we

40. 26:59

    want to use Lama 3 so um you know I can just basically classic like chat GPT you know I can just go here start typing a prompt that being said you know there's going to be sort of like there's this integration that's happening for AI developers uh on the inference side of things so here I have this this response I can basically just go into the API can make like live adjustments go here you know just copy basically in order to uh put it in

41. 27:33

    terminal and yeah as you can see here it's basically providing I mean reacting really fast I can also you know for this stream put it for True in order to part the answer in order to have something that's you know better to use for uh you know the the developers with the API but you get you get the point that's extremely fast extremely practical uh because we're believing in the fact that you know it's not because we're doing in a decentralized fashion that it should be complicated for everyone to you know

42. 28:17

    to on board and join us um so yeah I think this basically concludes my presentation um I actually happen to have a next talk in 15 minutes uh nearby so I won't be able to answer your questions now but in about 45 minutes I'll be back so you know we'll be here until the end of the event so happy to answer anything you have amazing thank you all right let's give it up for Jeremy how about that how about somebody non-technical opening up the terminal for a live demo thank you

43. 28:51

    yeah awesome all right oh yeah you want that mic okay yeah we got uh I will I will run some coverage as we are getting our next team miked up thank you all for bearing with us we've had a lot of fun setting up today um and folks we've been so excited to jump right into these demos I'm realizing we got to give some shout outs to some folks this is how I'll I'll stall and cover as we're

44. 29:15

    getting set up here we got to give some shout outs so I was telling folks uh out there you know we we actually ran this event completely in house this is our full near Foundation team that put this on but there's actually one very special individual who has overseen all the operations for this so can everyone turn around and give a round of applause to Bianca moment at the back thank you thank

45. 29:40

    you Bianca holds me accountable uh and makes sure the ship goes so thank you Bianca Whitney are you ready I'm ready all right so you've heard from Mizu right building at the data layer you've heard from Jeremy at uh hyperbolic right really a key part of our stack in training at inference and you're about to hear from an amazing individual and allow him to introduce himself uh Whitney Gibbs uh CEO of ring fence um

46. 30:04

    who is also very much focused at the data layer as well um so as we've been building out our stack and the projects we're working with um we could not be more excited to work with ring fence so Whitney I'm going to turn it over to you and we're all set thanks Jared big round of applause for Jared they've absolutely crushed it with this event so as J mentioned my name is Whitney I'm the CEO of ring fence and we're building the data monetization layer for AI we've been incredibly

47. 30:35

    excited to be a part of this near Horizon AI incubator cohort um specifically because we get to work with a lot of great projects and near has opened us up to a ton of new connections we all know the problems that we're currently facing especially those of us in the decentralized economy when it comes to how AI is using our data without paying us for it and as the markets continue to grow the problem is going to continue to grow unless we all

48. 30:58

    start to tackle these problems headon and at ring fence we're really focused on helping people collect custody and monetize their own data our protocol as we continue to build it out has the ring fence vault which I'll show you here in a minute uh this allows users to custody and collect in custody and then protect their own data we then pull the data from the vaults using our node system we pass a hash to the near blockchain and then we pass structure data to to Lake

49. 31:27

    houses that we're partnering with which can then sell that data through public gateways and marketplaces and return fees back to the user so eer is the first integration of ring fence AI we have a demo that I'll go through here with you now and we've currently got 40,000 people on our weit list this is a product that we built in collaboration with Leonardo AI uh and we're we're really excited to Showcase this for the first time so very straightforward any

50. 31:53

    of you who have ever used mid Journey or Leonardo uh when it comes to setting up with eer you're just creating a profile as you connect your wallet here we have the integration with Myer wallet you connect directly you set a username and this username allows your personal data to be attached to any of the images that you either upload or that you create using the ai gener ai generation platform you can see here some of the

51. 32:18

    recently generated art from the platform itself and now you have the ability to set up a profile and this is a really cool feature for direct distribution of content you can connect either your Twitter account your Facebook account if you're a boomer you can also connect your mint base right and this gives you the ability not only to create art but immediately move it on chain and monetize

52. 32:41

    it so our whole platform is based around ring fence nfts these are essentially folders which you then create collections within and if you go on to our website and you read our light paper you will see how we believe these complex digital assets are going to revolutionize the way that AI generated content is monetized across the internet so you name your first rnf and then you

53. 33:06

    get a quick little mint free to Mint these once it's minted you now have your first folder set up now one cool thing that we've built is a custom verification engine which actually labels anything that you upload into the system and it also screens it on the against the internet for anything that's currently living elsewhere on the web or for anything that can s any brand or trademark or IP

54. 33:30

    infringement issues so you'll see here select this we've got a couple of images that we're going to select one of them is a self-generated Sci-Fi image and the other is everyone's favorite frog so you'll see here in a second a Year's worth of work goes into just a quick scroll and a checker an X right so we can see that the Custom Image was validated because we own it's ours and we now have the ability to use it to

55. 34:00

    train AI or to monetize in other ways and then Pepe obviously unfortunately isn't ours so while we're able to upload it it gets rejected and then it's removed from the system so it's not used by the nodes in the training of AI but the key to what we have been working on with Leonardo is their AI generation capabilities they have fantastic engines and what we have essentially done is streamlined things so that you can generate with intent most of the time when you go on to an AI generation engine you select what you

56. 34:33

    want to generate but then it gives you a plethora of different options like what's the size what's the resolution we boil this all down into just where do you want to post it and what are you looking to accomplish so after you do that you select which type of engine you want to use and then you enter your prompt once the prompt is entered obviously I think we're all familiar with how this process works you H once you hit next step it

57. 34:58

    goes to generation so you'll select your nft where these will be stored these will all go right into that folder as soon as they're generated and here you go you've got your four Images generated now here is where we differ from other AI generation platforms your content nfts which are stored in your ring fence nft folder can be used in different ways let's say you want to store them on chain or move them

58. 35:31

    to a social media platform you have the ability to do these you can move them on chain through minting they're immediately available to list on mint base because you've connected it through your profile you also have the ability obviously to view every asset that you have either generated or uploaded in your ring fence Vault and then you just select it View and Vault and you'll see those all there and then moving these online makes AI

59. 36:03

    generation more of a social experience so with eer you have the ability to take anything that you have generated and immediately move it from creation to distribution so at first we're starting with images next we'll have video then text and then music and all of these as we integrate more social platforms we'll have the ability to go from your profile and your generation directly to distribution giving you the best chances

60. 36:30

    at getting eyeballs and eventually monetization on the work that you're creating so you can see the recently generated art there in the the platform and for us you know we are big Believers in the fact that all of us are missing out right now on the data that or the value that our data brings because we do believe that our data has value and we believe that everyone should be able to ring fense it thank you

61. 37:00

    so folks we do have time for questions if you have any questions I will bring the mic around for questions we have a couple minutes here yeah yeah go ahead great question so right now what we're doing is we're working with integrating with existing data marketplaces like data bricks where there's already demand for the data uh and we're still in the process of of working working through that but essenti

62. 37:27

    the data buyers pay for the data and then we funnel those rewards back to the ring fence users yes e we are great question one of the biggest reasons we decided to to work with near and move everything over to near is because we want to abstract out all of the web 3 blockers for most users right signing a transaction interacting with wallets these are things that most people don't care to do right outside of our Niche little group of of uh fun people so we want to be able to open

63. 38:08

    this up to a broader audience and with near we're able to actually provide a great web to experience with all of the strong web three infrastructure to protect and help people monetize their work yes uh what kind of model are you using for uh detection if the image has already been uploaded so we have we've built a custom model so we we've used open source models that were were untrained and then we've we've spent the last year training and fine

64. 38:45

    tuning same with the the GPT for labeling that's a a custom GPT that we built so what can application developers do now to integrate R ring fans TBD we're still in the process of migrating eer over to near once that's done we will have test net for the protocol going live in October and at that time we'll be looking to integrate with with applications most of whom are run by people in this room so that we can start testing out the functionalities of data movement through the protocol to the the

65. 39:26

    end applications amazing let's keep it up for one more time thanks for the questions guys all right folks well you heard from the first three of our Founders and we have three more who are coming but we're going to take a quick 10-minute break chill we got some food we got a nice Skylight outside good venue we'll take quick 10 and we'll be back for our next three uh founders e

66. 41:52

    right what

67. 43:27

    a what e

68. 45:26

    why d why

69. 46:28

    what right

70. 47:54

    right what why

71. 49:27

    d right

72. 50:28

    what what

73. 51:56

    d what d

74. 53:13

    dat layer you heard from hyperbolic who's building uh training and inference and now you're about to hear from something that is critical to Bringing AI to life and that is compute um exhibits is uh a team that we're really really excited about um and already uh working with so let's give it up for Jonathan at exhibits who's going to share a little bit more thanks Jared mik check all

75. 53:36

    good hey guys uh Jonathan here at exibits and we are building the decentralized compute infrastructure layer for a whole bunch of AI companies and other GPU protocols so what we Define as base layer compute protocol for artificial intelligence so as you know um AI is the fastest growing thing that we've ever seen uh within the first two months chat GPT 100 million users uh and that's continuing to break records uh and on the backbone of that compute uh we see that as a Lynch pin to

76. 54:09

    AI all good uh yeah compute is a Lynch pin to AI market right now there's a huge bottleneck a16z in one of their articles they said the supply is so com constrained that demand outstrips it by factor of 10 and we see that becoming way worse as time goes on uh especially with the supply chain issues that we are now uh experiencing so the real battle is not among companies where we see it is against the incumbents AWS Microsoft Google uh and these guys uh sort of

77. 54:47

    corner the market and they haveing that's through the roofs get $1,000 credits and it's gone within the month month months come in and provide all that compute and these are some of our partners so exhibits this is our secret sauce so we are uh have a lot of experience in building over a 100 data sensors in the past uh and that gives us access to over 655,000 gpus so to date uh our if you want to access one of our 65,000 gpus uh we're ready to show and then have that with you uh we also have

78. 55:28

    suppliers uh render uh we recently did a partnership with them uh they have uh a lot of availability of gpus that they can contribute to us same thing with Foundry uh which is backed by dcg uh and because we've built all these data centers we've become tier one Brokers uh for NVIDIA AMD and super micro in total uh the 65,000 gpus we can measure that as compute hours and that's equal to over two billion uh uh compute hours uh of gpus that we can help help

79. 56:00

    service for people so our secret sauce as I said earlier uh building the data sensors we've built a lot of Technology uh because the inevitability of gpus going down uh is it'll happen right um so we built fall tolerance Technologies uh We've also created acceleration Technologies so what that means is that we can take the level lower level consumer gpus and make that as fast as

80. 56:26

    the A1 100s and stability uh that's one of the most important things for Enterprise companies uh is how long they're up and how long they can stay up and as far as the web 3 business model that's something we're really innovating in and we're running with uh to provide you newer ways and innovative ways of accessing gpus uh that is not offered anywhere in the market so uh what makes us more

81. 56:51

    differentiated from other players is that we specialize in bare metal and virtual machine as well as containers so all the other GPU marketplaces right now they're running on a container basis uh and by us having sort of that um Advantage it gives us the ability to do much more tasks pre-training fine-tuning performance optimization as well as full access to the gpus it's as if we were just sitting in the data centers ready to help you uh day or night first enter your email and password on the login

82. 57:23

    page to sign in upon successful login you will be to the dashboard where you can see an overview of all available features click on the key pair menu to create or upload a new key click on Virtual machines to return to the virtual machines page where you can deploy a new virtual machine follow the prompts to enter the parameters click the deploy button and you will see the new virtual machine in the creating state it will take approximately 1 to 2 minutes to complete the creation

83. 57:54

    process clicking on the name will take you directly to the machine's Details page where you can view and configure various parameters of the machine once the parameters are configured you can remotely access the GPU server via SSH now we can also install the required environments and software directly on this machine everything is running smoothly thank you everyone yep so that's uh preview of our on demand compute uh uh that hopefully

84. 58:27

    we'll be able to deploy within the next couple weeks so our value proposition uh why Enterprises go with us uh here's one case study so a company was previously paying 150,000 per month for compute at AWS once they switched over to us we've uh hit them with 72% savings and costs so they're happily paying 43,000 a month uh at a longterm contract with us uh not only that we've increased their training speeds so what that means it's a double-edge sword right like time to Market as and also cost savings uh so so

85. 59:01

    uh that's sort of our uh focus on uh of exhibits so our partners uh we work with a bunch of GPU marketplaces aosh Aether IET hyperbolic which you guys just seen present uh ritual so we treat these guys as Distributors of exhibits infrastructure so more and more of these people can have access and scalable uh AI compute as far as customers uh directly we've been servicing over a dozen customers and we have a monthly recurring revenue of over half a million

86. 59:35

    dollars so we are trying to innovate in this space so traditionally the way that people will get gpus uh is via an SLA and prepayment or monthly payment what we want to do is leverage a web3 model as I mentioned before so now if you need gpus you can stake tokens uh on a monthly BAS basis or on a per period basis in order to receive compute in exchange and then once you're done with that job uh you'll get your profits and you'll unstake your tokens uh uh sort of

87. 1:00:08

    a similar way to how maker Dell uh does uh sort of the the the the loan factors of their compute of of their stable token another way is that we want to increase the distribution channels by allowing everyone to have access to our compute Leverage Supply and act as their own microc clouds so on the right C uh a sort of distributor of ours uh cons service customers on their own they can mine different protocols or they can just use those gpus uh to further the

88. 1:00:40

    web three stock uh customer number two can just decide to service all customers themselves and customer number three can just uh provide it to every node uh and just do some node farming uh within the ecosystem of web 3 and they all will run on the same um sort of protocol and mechanisms that we do with our partners so what ties us all together is the uh web 3 uh tokenomics that we've doing so the GPU and the compute tokens will sort of solidify and build up the

89. 1:01:14

    ecosystem uh of exhibits yep so anyone who wants to onboard with us uh please scan here uh we're giving away a bunch of GPU hours uh for anyone any questions yeah so container it's essentially like wrapped right like so uh they'll rent off a GPU cluster uh from a service provider mainly a Data Center and they'll use it on behalf of them as opposed to us really being on the control layer of the the stack where

90. 1:01:57

    we can say okay uh you use it for like something very specific and the data sensor will configure it to that specific task but it's not really malleable if you need to use it for other task any other questions questions I get how one of your customers can save money how do they improve their training speeds though by 30% it's because they went bare metal or what was it about your product that allowed them to speed up training oh

91. 1:02:30

    yeah there's a lot of things so mainly it's not necessarily on speeding up like the hardware itself uh we're just adding more technologies that make it faster uh and make it more reliable so one of the things that make it more reliable is our fault tolerance mechanisms right so part of the speed of acceleration is uh going to the back door and making sure that it runs as smooth as possible as opposed to like having those stop Gates when it goes down there's more Technologies like

92. 1:02:56

    bubble technology of moving like the CPU like energy to the GPU energy and that's how we also optimize uh the gpus that we have in our ecosystem thank you I have a one last question so you mentioned about how you guys are different different than the other uh GPU providers but then you also mentioned that you guys are Partners so how does the competition work with part rip yeah yeah so uh for example like

93. 1:03:29

    aosh or um IET uh what they're doing is they're really customer facing and how their interface works is that it's okay who needs different clusters and all that stuff what the relationship with us is that they'll rent directly from us uh of which we will work on their behalf uh work on their behalf so it's like a distributor model uh we that we're partnered with with them you mentioned that in one of your case studies you get uh better performance is that just because the the

94. 1:04:08

    customer was using different Hardware versus AWS or um as far as different Hardware uh I can't remember what specifically they were using at AWS um but it most likely was A1 100s at AWS but then then we switched them over to a combination of A1 100s and then focused a lot of their inference on uh more sort of reliable gpus that we have in our Network as well so uh I'll I'll explain it more after

95. 1:04:40

    this any other final questions yes if I want to deploy say a thousand nodes on your network what sort of infrastructure do I need in place before I can deploy them are you using one of the deployment Technologies how do I deploy nodes like say a thousand nodes into your network oh right now it's very white glove so you'd have to just talk to our um developers and then uh we'll see what's the sort of standards that we can hit that will make

96. 1:05:11

    sense for us and then we'll uh work with you from there all right but eventually it'll be a API uh that you will just plug into your system oh so like the typical problem for no operators is that like if their um node is like idling there is no incentivize or do you like have like incentivization mechanism for like idling no yeah uh that's the problem with uh these GPU Market places right like a

97. 1:05:42

    bunch of them will have idle compute because of the saturation of the market so uh we've seen that in other protocols right so 5 million supposedly gpus went onto the network and there was not enough rewards for for that so we are creating a mechanism that's that's one time based and also compute uh quality based uh so those are some of the factors and then it also uh goes back into the business model of where does some of our Revenue our profit our

98. 1:06:11

    actual profit uh falls back to the idle compute as well thank you all right well as our next founder is getting set up again we are really excited to be working with exhibits um if you haven't already they have a banner outside that has that QR code for the free h100 compute uh if you're building in this space highly Rec recommend you check them out um really really excited about that team for our

99. 1:06:50

    next founder we have two more uh left is Don gsen with never mind um again if you remember the architecture we shared and sourcing and working with each uh team at at different layers of the stack Don is really focused on the agent layer um and really contributing to the agentic Frameworks that are being built on near um I know he's getting miked up now but I will give him uh a moment to get ready uh but Don has an incredible background

100. 1:07:15

     in the space has been building at the intersection of AI and web 3 for a very long time uh previously uh was one of the founders of ocean protocol if you're familiar with ocean um I allow Don to share a little bit more about his story and background uh but we're very very excited uh to see now ai and agents enter into the payments space uh so never mind is really focused on agent to

101. 1:07:38

     agent payments um so Don are you good to go we're live I successfully stalled for an extra 15 seconds yes maybe another 15 SEC okay I got 15 more seconds we got two team members from the near founder success team in the back who have done an amazing job give it up for Dylan and Devin yes yes I'm not the CTO so it's all right come on you piece of wow I don't know all right there we go hang

102. 1:08:33

     on all right agent yes we should have had an agent just do man guys let's give it up for all right uh first order of business thank you to near Horizon for organizing the event second order of business Whitney I need you to give me a crash course and how to do a proper presentation because yours is way better than mine um you'll wait till you see where how I embedded this shitty video um okay so I'm Don I

103. 1:09:06

     am the co-founder CEO of never mind we're building the decentralized AI payments protocol um thank you sir so our thesis is predicated on the belief that uh and the idea that we're witnessing the rise of a new consumer that's going to manifest as trillions of AI agents and that these agents are going to change Commerce forever this concept we're calling AI Commerce and as a response we're

104. 1:09:38

     building the PayPal of AI to be the engine behind AI Commerce H why are we doing this so as Jared mentioned uh I guess I'm OG AI web 3 I've been doing this for almost a decade now um and one of the the things that we've seen in particular over the last um 18 to 24 months is builders that have built productive first AIS and now agents are really struggling to commercialize those propositions uh the

105. 1:10:11

     existing Payment Solutions whether they're Web Two based or web 3 based are very rigid uh and static in their structure I'll unpa this a little bit more but we're addressing the need to create a much more Dynamic and variable price bed uh system uh to enable transactions between in particular AI agent to AI agent what does this mean so if we are gonna kind of unpack and and look at the anatomy of an agent this is a gross oversimplification of the construction or architecture of an agent but really

106. 1:10:46

     the idea here is to understand an agent can call on a bunch of different services or Tools in response to a d damic set of inputs um the simplest thing to conceptualize here is let's say an AI agent is built with a GPT series of Transformers any response could be rooted to a a given model within that series depending on the um uh I guess complexity of that request so what you have is an agent user whether that's a human or a carbon based agent as we like to call them or a silicon based agent uh

107. 1:11:25

     making a request to another agent and then depending on that request it gets rooted to differing levels of complexity um under the hood with these different models so let's just say hypothetically um anybody in this room can use a particular model uh somebody makes a request that's multimodal in nature that's going to get rooted to a model that can handle the multi multimodality of that request and that has an Associated cost to it now I'm a dumbass I'm going to make a very simple

108. 1:11:57

     request to this agent and it's going to get rooted to let's say probably like gbt 2.5 in this case the cost uh variable there it we're talking about orders of magnitude and difference between those costs of the associated services that are called by the same agent existing payment systems are very static in nature uh stripe was created to sell t-shirts on the internet right singular input singular output in this case we're talking about Dynamic inputs variable response variable cost how do

109. 1:12:32

     we actually settle that VAR that variability with the user what's needed is a more complex system of unit accounting um and what we're finding at the moment is that Builders have to build this themselves they're being left to their own devices and what they're inevitably building are these very clunky spoke Solutions first and foremost they're they don't work very well the ux sucks secondly you know if

110. 1:13:02

     you hold our vision of trillions of AIS emerging that are going to interact imagine every single AI imbued with its own form of payment and and mechanism for transacting they're just going to butt up against one another it's Prime for protocolization or standardization um to simplify the overall process so that's what we've built we have built this system of unit accounting that manages this metered component between the um the the service

111. 1:13:34

     that's being rendered and then enabling the settlement to the specific user that's making the request now we'll go through a demo and you will start to see what I'm talking about so I'm going to show you the very simple interface that we have it's much more complex but I'm not going to show you the command line but what we have here is is effectively a widget I can't use this uh particular chat interface until I've I've purchased

112. 1:14:00

     these credits what you're going to see here I can log in with a wallet but we've also gone through the motions of abstracting away all of the crypto complexity so you can log in with your socials um so logging in with my socials right now what we've also done is we've gone so far as you'll see it in a second you can pay in crypto for these credits but you can also pay in FEI it most users of AI services and AI agents at

113. 1:14:26

     the moment they're not web 3 right they don't know what the hell we're talking about here's a demonstration we've actually integrated stripe um this is just a testl version feel free to use this credit card number um so we'll go through the activity of actually paying for these credits in this case it was $1 for 15 and once I've made the purchase these get loaded into the wallet that

114. 1:14:50

     was created via uh my registration with my socials so now this is going to recognize that I've logged in and I have these credits and I can now interface with this agent through this simple um user interface so I've asked the question I get a response and now what you'll see is there was one credit redeemed for that request now I'm going to show how a builder an AI Builder can actually set this up on the back end so what you do first and foremost you have to create this token gating mechanism

115. 1:15:21

     for your AI service or agent um and that are that is these credits so the first thing we're going to do is describe the overarching uh credit structure this is just some metadata that makes it easier to administer and here's sort of the first piece it's like the meat and potatoes of of what we're doing with our system you can use um you can limit access based on the number of requests

116. 1:15:44

     uh for for credits or the time limit that you have access to a given service in this case we're going to set the price at um $10 for a th000 credits and that's it's going to create a smart contract um push that to a token Factory so when this Builder's users want to use that Builder's AI agent or service they buy these uh th000 credits with $10 or 10 usdc and they're able to access that credit the other thing you'll notice

117. 1:16:11

     here there's no popups there's no you don't have to sign anything it's all been abstracted away um so even the builders don't have to deal with the complexity of lib three now what we're going to do is take is is take an AI Sur service or agent and bundle that to this set of credits it's going to act as the gatekeeping component for this particular service um this is interesting in in this case the registration of this service the service is uh being rendered via an API what we give the Builder the ability to do is

118. 1:16:45

     actually choose the verbs or the functions within that API that they make that they Expos to their given user base um and we also can uh enable the um implementation of any sort of authentication mechanism as well if you want to you know fully secure what the the endpoint that you're making available bit more metadata here's the last piece of the puzzle so that credits component that I just created I'm now uh bundling this service to that set of credits and I'm setting the Redemption criteria so if you remember in the first

119. 1:17:17

     example I bought uh 15 credits for $1 made the request it then took one credit from me in this case with this service if that was enabled through a similar interface it would take 50 credits in Redemption and that's it that's my you can hit the uh just yeah there you go all right that's it any questions Round of Applause there's a bunch more stuff you can do but it's yeah yeah any questions folks we do have one in the back okay

120. 1:17:57

     say say I wanted to like integrate multi-agents where one person has one payment and I need to my agent needs to pay those agents how do how does a setup like that work yeah so in in the integration point would be through the SDK where either those those agents are issuing their own credits and then you have to create uh like either a pooling mechanism or a direct uh payto playay transaction um it depends on what you're actually paying in but happy to

121. 1:18:31

     take that question offline because that's a that's a nuanced question to dig into in more detail uh what is the thinking of using credits rather than just like usdc why do you have to buy credits from the vendor yeah so um the the issue here is broadly speaking what you're talking about is settlement so you can settle in usdc um regardless of that settlement you still have to perform some form of

122. 1:19:00

     metering right you need to understand what's going on with the underlying system you can set it at a like a a a very generalized meter uh setting where it's just like I'm going to charge this fixed fee and hope that that covers all of the costs for any given uh request usually that doesn't work so just for sake of example that's how Olas if if anybody's familiar with the Olas system of uh multi-stem um multi-agent system that's how they did it originally they weren't using usdc they were using XI

123. 1:19:34

     and they just set one price to access the problem there was because of this metering component they were either charging too much or they were charging too little what they ended up deciding from a business point of view is to charge too little so instead of charging for the power user that's going to you know request like multimodal inference they were just charging for my dumbass you know in the very simple request and

124. 1:19:57

     then they were having to underwrite the cost of the power users what we give is with these system uh the credit system the actual terminology for it is licensed tokenization it's a longterm like it's this licensing pattern or architecture has been around for a long time it gives you a great deal of flexibility in terms of what services or tools can be rendered based on a request and then each of those underlying

125. 1:20:21

     Services has its own Redemption criteria attached to it so as a point of Engagement you just buy the credits so you buy credits you buy credits you buy credits but you each one of you can actually invoke completely different sets of services under the hood uh it's it's a really good abstraction tool for doing this sort of meter translation into settlement long answer but yeah that's why we're using it have one

126. 1:20:51

     more we answer them all oh I almost got sorry it's it probably is a really dumb question is um if I want to use this AI agent how what's the definition of a a question that belongs to a like the difficulty like uh GPT 3.5 you mean the in the rooting functionality thankfully we're not building that that's up to companies like well in the non web 3ace open router Etc in the web 3 space there's companies like nap the AI and

127. 1:21:25

     stuff that are doing that so we partner with them enable the payments for the agents that they're building but they actually handle that rooting piece so we're just like a catalog of the services that are made available and more specifically the price point or Redemption criteria for each one of those tools amazing one more round of applause all right folks uh you heard from six amazing Founders we have been incredibly excited to work with each of them over the past few months we're

128. 1:21:58

     going to continue working with them so stay tuned for um uh their launches and releases uh next up we have one quick panel if you give us two minutes we're going to get our speakers miked up uh and everything to hear from some other Founders and then we will uh hear from our founder and CEO Ilia pukin so if you give us two minutes we're going to get set up our speakers if you can come up

129. 1:22:19

     we will get this panel going right what

130. 1:23:24

     right 2

131. 1:24:52

     right what

132. 1:25:53

     what e n

133. 1:26:54

     d right

134. 1:27:56

     what d what

135. 1:29:29

     for bearing with us in our transition time uh you heard from some amazing Founders and got to see some products but we also want to talk about what's happening in the market very excited uh for this next paneling Founders inv be introducing themselves talking about what's happening in the market what they building I'm going to turn over moderator Casey thank you jar very exciting to be here I'm a huge fan of this ecosystem and so we'll try to make this as

136. 1:29:58

     entertaining as possible maybe let's start with intros and we can start the tail in and move forward just so the audience knows who's chatting today uh awesome thank you uh hello hello test yeah perfect uh awesome thank you thank you Casey for kicking this off uh first of all I'd like to thank the U near team for organizing this wonderful event been no maybe louder hello hello testing I'll be louder you're G to eat

137. 1:30:34

     it uh actually no there won be Vol back anyway um so uh firstly thank you to the near team for organizing this Splendid event and uh my name is Leo I'm a computer Scientist by training and the co-founder and CEO of superside and at superside what we're building is the base La for AI agents and I'll hand it over to M well maybe before we even go for it just because the base layer for

138. 1:31:00

     agents is now becoming almost a popular thing to say can you just give like a little bit more granular detail of what that means thank you for catching that so essentially what we do at Super site is we're like giving agents the information for them to be able to take like different actions and specifically what we do is essentially like if you wanted to set up a trading agent what it

139. 1:31:21

     does is um let's say if um 30% of the top 100 most profitable traders of Lio buy an asset within a 6-hour window buy that asset for me so this kind of a like programmatic execution before if you wanted to build it you would have to be a sophisticated actor in the space such as a f who would have these data pipelines ready but essentially uh what we at superide enable is like for you to be able to build these kinds of uh trading agents or any kind of other

140. 1:31:51

     agent uh based on realtime data to take actions uh so that's what we're building awesome yeah thanks and also again thanks to the to the New Foundation for organizing been an amazing event so far um so my name is Matas I work for million uh focusing mainly on AI research that's also my uh my background I have a more traditional let's say AI background and I think this is the only room uh in the world where you can use traditional anai in the same sentence

141. 1:32:21

     together but uh here we are um what million does is um we're building a privacy preserving computational layer so for those of you um who are familiar there are loads of um privacy enhancing Technologies out there at the moment so you've got um FH fully homeomorphic encryption you've got MPC multiparty computes which is mainly um our our uh biggest Focus so lots of uh three-letter

142. 1:32:48

     acronyms and essentially what we try to do is BR bring all these Technologies together to build that um privacy enhancing computational uh layer because currently you know you've got encryption at rest encryption in Transat that have become so ubiquitous in the in the current Internet landscape even to the point that if you try to access a a website through through HTTP instead of https your browser whichever one you use uh either assumes that you made a

143. 1:33:19

     mistake or tells you that you're absolutely out of your mind um and we believe that the logical continuation of that will be um encryption during computation um as well which has for a long time been first impossible then infeasible and now showing signs of actually we might just uh be able to do it so that's what we do and if an L1 um it's a yeah it's a computational uh foundational layer yeah okay there

144. 1:33:51

     you go all right hey everyone I'm Don uh I just presented but for those of you that weren't in the room uh I'm the CEO and co-founder never mind we're building the decentralized AI payments protocol so our thesis is predicated on this idea that we're witnessing the rise of a new consumer uh that will manifest as trillions of AI agents and that these agents will change Commerce forever we call this concept AI Commerce and we're building the PayPal of AI to be the

145. 1:34:23

     engine behind behind AI Commerce my background uh I studied engineering at school and then went into Commodities trading doing risk which back then was called statistical uh analysis which became machine learning which is now I guess AI um I've been doing this for 20 years I added the web 3 flavor in 2016 I guess at this point I kind of OG AI web 3 I co-founded a project called ocean protocol back in 2017 and yeah this is

146. 1:34:53

     kind of um wow I didn't know that my life's work on display yeah great okay and then really quickly I guess just to do a one liner on me too um I'm also a computer Scientist by training and I was previously an engineer at Google and then was an investment partner at Paradigm and now I run my own Frontier Tech firm and uh I have so much respect for all the builders in this space so very great to be chatting with all three of you um okay so I one quick question for

147. 1:35:26

     neillian and then I have a few themes that we can talk about I I quickly do want to touch on the difference between the tees the fully H Mor effect encryption and then also the ZK because it feels like we've been chatting a lot about these three Technologies over the last week and if you could educate the audience and kind of the difference between them and where you're actually using them I think that would be

148. 1:35:48

     interesting yeah of course um so I think the so at a high level um most of them are actually trying to achieve more or less the same thing particularly um FH and MPC um but they just go about it in a bit of a different way um so FH without maybe stepping into a you know gruesome technical detail um essentially what it tries to do is um it tries to encrypt your data uh in a very specific way um that allows essentially

149. 1:36:18

     the computations of on that encrypted data to translate um into comput that would hold on the underlying data once you decrypt it um that add a bit of computational over that has been you know decreasing dramatically because of um protocol improvements but also Hardware improvements um so yeah that's on on a downward Trend and then there is MPC that essentially tries to tackle the

150. 1:36:42

     same problem but in a different way in that it encrypts your data by essentially sharding it and sending each Shard to a different um party in a network that performs computation on that so the computational overhead is a lot lower there um but you do need some kind of communication between those notes so the over at is just different it's more connectivity instead of um computation and then ZK is actually a little bit different that's more kind of

151. 1:37:12

     assuming that um you have a party that kind of holds a lot of information itself and wants to prove um whether or not something is the case um to a third party so I guess that is like what it tries to do is actually a little bit different um but also maybe to add on to that is that I think we're kind of seeing this shift where we used to think about these Technologies and in sort of Silo terms as very different Technologies and one will end up winning

152. 1:37:41

     which one will it be um and making bets on that whereas actually now you're trying to see the shift where all of these Technologies are kind of blending together and and people are really really as an industry we are really starting to see that there is no Silver Bullet and the lines between these Technologies are a lot more blurry than uh we initially all I completely agree with this and

153. 1:38:05

     that less than a year ago it was this vers that and I'll tell you two of my portfolio companies who used to have that frame of mind are now using tees for model privacy and then using HF for data privacy and they're actually using them together instead of choosing between them so I Echo that from at least my vantage point um okay so before we even go deeper can we do I want to better understand who's in the audience so can we do a show of

154. 1:38:30

     hands on if you're a developer okay and maybe can you raise your hand if it would be more interesting to learn about what this technology is being used for like raise your hand if that's interesting okay or if uh what this techn how this technology is built like what what is the technology behind this okay so let's go more in the direction of utility for now um maybe we'll do like an explain it like I'm five showing here and we can go around and just explain one use case that you

155. 1:39:04

     have the highest conviction in working over the next year in crypto and explain a world with it or world with it working and a world without it working meaning like why do we need it and anyone can understand it very simple don't high IQ us anyone can you go ahead I'll go is the question C does it make sense so use case yes but you understand the constraints yeah okay okay so here's an space loated in gaming

156. 1:39:36

     um you got everyone should check out and this is your highest conviction use case make sure so I think yeah well I mean I got a lot of convictions Casey but this is a high conviction use case all right because it's it's pretty fun um check out these guys it's called chirper uh chi RP R so they've created this uncucked online Dungeons and Dragons so it's completely unrestrained models behind the scenes and you can create these characters so it's it's like promp

157. 1:40:09

     tuning or instruction tuning to create these characters and then they can go out into this like gradually forming Universe right as all these new Agents come in and they run autonomously but because it's uncucked you can set up like murder Bots that go out and murder this is the first use case you're going to tell to the audience it's because I'm gaming because this is what we want I mean look Gladiator 2's coming out right like we like this kind of entertainment

158. 1:40:42

     so I just want to say I did not approve of being on the panel yeah by the way I'm noton poisoning so I took a spot suer uh actually no I am Brandon player player um no sorry so yeah this this ecosystem is awesome right and it's it's totally autonomous so these Bots go out these agents go out and they like work with one another or they murder one another and you know in real time what's

159. 1:41:12

     happening off the top of that it's it's becoming this content creation site where those that are unleashing these agents into that universe are now saying I want the story lines but it's not really consumable for me as it happens on the servers right so now what these guys are doing is is effectively extracting those narratives out of this ecosystem and creating like a a fully uh uh generative AI content Universe you

160. 1:41:46

     made Laura leave just say that's like you know it's it's this it's this Netflix for the new age of AI agents and it's ugc it's like being serious so that is kind of interesting it's that it's completely generated by the users without constraints without constraints I mean it's the unconstrained aspect of this that I think is going to be what people like users gravitate towards okay right so what interests me specifically from a business point of view is that the the

161. 1:42:17

     team that's building this has recognized the need as these agents kind of uh live with each other in this in this ecosystem what they recognize is the need for rewards amongst these agents as well as the consumers that are exist on the peripheral that want to tip the other agents that exist within this universe so they gravitated towards web 3 you know and the efficacy of of what we've built from a digital point of view and all these foundational elements um so this is this is the type of use case

162. 1:42:48

     I think it's going to be like Earth shatteringly good when it comes to content creation anyway how am I going so the timeline was a year right the yeah 12 months it's difficult to choose because there's so many amazing people building on on a on private this isn't an easy panel though you have to yeah but it feels like you know I will always be doing someone Dirty by not mentioning their awesome project but in the G to hat scenario that really need to needed to pick one um I think what I

163. 1:43:29

     mean ultimately if you think about gen what everyone wants um is a personalized AI assistant right like just an AI that knows all about you full context and can help you in like all facets of personal and professional life um of course there is the obvious ideological um problem of well in order for such a system to exist it needs access access to all of this information and personal information sensitive information that the willingness of people to give that

164. 1:44:02

     system access to that information is at an all-time low right um so even ideological um aspects I think are quite obvious um aside like the adoption for such a system is just low because of it um so what I am very excited about is um essentially a fully private personalized AI assistant where you still get the value that I think each and every one of us would use such a system because it's such an obvious

165. 1:44:31

     value out to so many facets of our life um without really paying the the the Privacy price let's say for it and what's your opinion recently seoa put out an article on gen's utility do you see that by chance it's actually quite embarrassing because a lot of people have sent it to me but I haven't had to I'll give you read oh I'll give you the you are I'm curious your opinion it's just it's debating if gen is production

166. 1:44:57

     grade ready or if it's really more of a toy and we're going to use it once but the retention isn't going to be that high and yeah I'm curious from what you just said where you think we are in that curve of adoption yeah I mean it's definitely early days especially like the web 3i flavor um is very very early days I think there are like some rather fundamental issues that need to be

167. 1:45:21

     addressed to unlock I think the majority of of use cases I think privacy is one of them I mentioned that already perhaps the technology for such a personalized as AI system to an extent actually already exists but the adoption would just be hampered by that privacy aspect um think besides that if you want to use it for S sort of um a knowledge assistance uh for those kind of workflows which also is a very obvious one particularly in the Enterprise uh segment um then the factuality problem

168. 1:45:52

     kicks in like the the reliability so I think for the the current sort of obvious main use cases there are these quite fundamental um problems but the good news is that there are a lot of promising uh Avenues a lot of promising directions uh that are being actively researched and built and they could provide viable Solutions but they would just need to be more of a zero to one very new system rather than an incremental

169. 1:46:23

     uh Improvement thank you uh in terms of use cases what I find I guess most pragmatic and practical uh on the intersection of AI and crypto at the moment is uh Reinventing defi in the sense of a ux perspective uh because right now I mean if you look at the way that we use crypto uh specifically it's actually like very cumbersome to bridge assets across different chains to have like all of these actions that you sometimes spend like three to four hours on to

170. 1:46:55

     just like for example like Buy nfts on a new chain um so essentially I think that uh in terms of like improving usability for crypto I think there's a definitely a very like a good use case that can uh open up I guess a lot more new opportunities so we can reinvent defi in the sense that you can have like I guess onchain in index funds that what they do is like they automatically rebalance a portfolio based on like certain events that have happened in the world or uh what you could do is you could build

171. 1:47:27

     like automated liquidity management uh like platforms you could have intent based solver networks so these are all use cases in defi that are now possible because of all of the advancements that are happening in uh like the AI and ml spaces and uh they should make it so easy that even a 5-year-old can actually like uh learn crypto pick it up and there's no longer a barrier to entry um

172. 1:47:53

     and I guess the goal standard is for example like when everybody's grand map um can like use crypto without ever having to actually like learn new words new technologies um and uh I think that uh the intersection of AI ex crypto actually helps a lot with that do you think we're oh I wanted to ask you a question oh okay do you think we're there yet yes or no and then well I'll pause there do you think yes a great question I mean uh we're getting there so um at superide we

173. 1:48:28

     have 170,000 users uh who use our products today to get like answers from blockchain data but um like in terms of where we are in terms of adoption I think adoption with crypto as a whole has been a bit slow to take off uh in terms of using different products on chain but uh we're getting there we're definitely getting there and I think over the next three to four years like uh there's going to be like uh exponential jumps in Improvement in terms of crypto ux yeah I agree with

174. 1:48:57

     that Vision I think just to raise one skepticism I have it's that you'd have to have extreme trust with the agentic flows because you're dealing with money and so while the value proposition is a lot higher the risk is also a lot higher and I think I'm more likely to trust like an Amazon agent to buy me shoes and deal with my money that's kind of and we're not even there yet uh because you

175. 1:49:19

     know agents are still not production great in a lot of ways but I I I love the vision no that's a great Point actually the thing is that like we've been trying to apply non-deterministic models in every setting uh and uh not every task needs non-deterministic models so for example if you can reduce the problem down to a more deterministic space then we can actually like take actions more

176. 1:49:46

     correctly than what I would have done as a human for example um but uh I think that's the problem we're trying to fit in generative AI in a lot of spaces right now where it doesn't necessarily belong but um I think slowly as the Hye dies down a bit uh we will get to a point where we like use the right models for different tasks yeah and use deterministic apis instead of these like

177. 1:50:09

     just llm browser based agents for a lot of maybe the dii stuff so it's not as risky in terms of execution exactly yeah that's that's smart that's what's your take on the seoa update because my interpret like there's there's a bit of nuance to it right they seem to be questioning the ubiquity of AI but I I've got some anecdotal evidence that that um points to the the utility and

178. 1:50:42

     its production Readiness of this stuff I can talk about that but I want to get because you're coming from the investor side so I want to see what like what's your is there any rebuttal or do you see the same thing they're okay well first I think it's really important that we have programs to help accelerate like Horizon quick call out to near but in general kind of zooming out I'll give a stat which is

179. 1:51:05

     that when I was at Google we were doing we had a team that was dedicated to basically getting AI into companies it's like an Enterprise deployment team and maybe I'll ask the audience is was 2019 what percent of models that we built for Enterprises do you think got into production does anyone have a guess low digits single digits wait zero is aggressive everyone that would make us

180. 1:51:30

     really bad at our okay close it's 30% and because this this Audi is this close you normal people are like 70 95 but yeah it's 30 and that was back in 2019 before there was even really generative Ai and this was just more deterministic linear models and even that didn't even work and it so gen probably you know closer to 50% and so I think we've always had this problem with AI or not maybe that's a hyperbolization but we've had this problem with AI for a

181. 1:52:01

     long time and we're still figuring out exactly where it can be used I think the Sequoia piece was really negative I don't believe that but like I I generally feel from a technological standpoint there's no doubt in my mind AI becomes ubiquitous there was like I would bet my life against that this becomes integrated in every single vertical in the next one to two years I I also feel like um that

182. 1:52:25

     article where it's being utilized at the moment is going to fall outside of the scope of that article because they're looking at the investment that's going into AI specific projects and then what's coming out of those projects but they're not looking at like the longtail effect maybe so I spent the better part of a decade and a half in IT consulting and I'm still I've still got good connections into that space so an anecdotal evidence right when I tell you

183. 1:52:50

     know my ex- colleagues what I'm working on lights start to you know go off right uh my old boss I met him a few weeks ago in Los Angeles I told him what I was doing and he's like yes this is exactly what we need I I don't know how many of you are Americans in the audience but maybe you've heard of swom Consulting it's a it's a relatively big Consulting agency in the US what they're using AI for at the moment is to um speed up

184. 1:53:16

     their due diligence Cycles so what used to take between four to eight weeks to basically the result was uh an architectural document they can now do that in a number of hours they sit down with the stakeholders they interview them they transcribe the interview they put it through an llm that they've trained and then they get like 80% of the architectural documents out of it and then they just have to spend their you know the next you know couple days tweaking that document he's looking at this from a business point of view the

185. 1:53:48

     current model that isn't going to be encapsulated in what SEO is looking at is an is a master Services agreement and a corresponding s so right it doesn't translate yeah so a master Services agreement is just an agreement that a company like solm Consulting is going to have for the overall work that they do and then there will be specific itemized uh pieces of work uh captured in what's

186. 1:54:11

     called a statement of work um they're very broad overarching like uh not very well defined instruments for performing transactions effectively where my boss is already Forward Thinking right he's like we can use this development to undercut all of our competition and get these documents to Market way faster than they can but we can't do that under this like existing economic and Commercial constraint that we have with these types of agreements

187. 1:54:43

     they're too slow to formulate so he's looking for opportunities where they can get more towards it's something in in Consulting it's called value based pricing um it leads to anyway regardless what he's talking about is getting to micro payments for basically paying for the hours of work that this productive agent in in in combination with a human in the loop is going to be able to produce that's where we're already seeing the value okay I don't think this is on so this is why I'm doing this too okay we

188. 1:55:14

     have roughly five minutes and I would love to do this in a Q&A format for the remainder of this so does anybody have any burning questions and if if not I can keep going no one really we're doing such a good job that's right yeah okay if anybody does as we go no okay all right well I would like we can do kind of um like a speedrun here and I'll ask a few questions and we can keep the answer really short if you weren't building what you're building right now in crypto

189. 1:55:47

     aai within the same category intersection what would you build uh I would build like I guess Hardware acceleration Frameworks for f um and uh like try to I guess figure out how to make things uh less computationally intensive or make the hardware better that space is going to be really hard yeah so M no Hardware acceleration is very interesting I think another one that I think is very interesting is um sort of everything around the uh like the data economy like new monetization of data and I think that a common

190. 1:56:25

     misconception I think that people have is that like I'm sure we've all heard like Al nauseum sometimes that data is a new oil um I think that right now regardless okay there are questions about like data Provence and like IP and all that kind of stuff but I think we've got more than enough data and what I do think is that we don't have enough information which is very different data

191. 1:56:50

     can contain information but not necessarily um and for example what we see is that some kind of data it can be replaced by synthetically generation uh generated data which means that almost definitionally there was no novel information there to begin with um I do think there is a lack of information a very Niche uh use cases like for example in the medical domain I think the average uh medical scan of like the average healthy person a sort a fraction of a fraction of a fraction of a scent

192. 1:57:19

     um but a medical information about someone a very specific disease is worth an incredible amount of um that's incredibly valuable um similarly with for example um if people have specialized knowledge in certain domains um I think for instance uh it's no secret that um open AI they want their gp5 or whatever they end up calling it to have PhD level intelligence uh that requires PhD level information about very specialized topics and they're

193. 1:57:51

     paying people with phds in you know very random areas like hundreds of dollars an hour to essentially provide that information to them so you do see that like by and large like data isn't super super valuable or at least not as much as people think um but these very specific pockets of information are I think they like kind of the you know you provide data there is some viability there uh there's some monetization there Etc uh

194. 1:58:21

     that some kind of decentralization does um or some web three approach more generally um does make sense so I'll probably be building in that's SP okay this is gonna be a really random comment um wonderful look because like I'm working on the culminate like this is the culmination of a 20-year career so if I'm not doing this it's going to be something completely different that I'm equally passionate about and I love to ski so I'd probably design skis no it had to be in the industry oh

195. 1:58:53

     no it would I would just work on this like okay we have I can talk for hours on your topic because we have one minute and my final question is going to be for everyone in the audience that wants to stay up toate with crypto aai Beyond chilling yourselves and your companies who is one person or one thing that everyone in the audience should go read or follow on Twitter uh okay I'll go first I level

196. 1:59:18

     set everything with a guy by the name of Richard bman at Napa AI uh he's super dialed in across the board okay yeah besides Nan if you want to learn more about like the Privacy aspects of it I would really recommend the materials that Zama puts out I think they're yeah like Nigel smart for example has been s so very interesting stuff there uh you can learn everything you need at Super site if you ask the right questions but uh one report that I would like to really like I guess give a shout

197. 1:59:50

     out to is uh by Mohammad from exol laabs uh he put together an incredible report on a primer to Ai and crypto and if you're just getting into the space then make sure to read that report because that's the best way to enter and learn about everything in this place okay great okay and genuinely not to Just sh near but I do think their Twitter is really good okay that's it thank you to our panelists again really appreciate the perspective there uh you know there's always this argument is AI going to take your job uh well I

198. 2:00:28

     probably can say with confidence AI is going to take my job because I totally messed up our agenda and I went off script and our team was looking at me with a very panicked face and I'm like what's going on so uh we do have one more team uh who is not present with us today but Pond is building a foundational model for crypto using graph neural networks we're going to get their video queued up they got a pre-recorded five minute demo that we're going to show and then we will close our experience hearing from Ilia pukin so

199. 2:00:53

     we're going to queue up this demo get our speakers unmik and then we'll close out the event hi guys uh we are Pond we are here to present what p is doing and also we want to let to know how we are working with near so what we are doing here at Pond is we are building a def foundational model network of crypto with a deep learning technique called graph NE Network so what is the model we are develop in so first of all we need to understand what is uh uh crypto what is the most Crypton netive data so the

200. 2:01:36

     most crypto data cated data in our opinion is unchain data so unchain data is a form of graphical data first of all it can be represented by different nodes which can be uh wets token contracts contracts and edges pretty much is interactions between different notes like by S transer so when the entire thing and we assign different properties to different notes and edges and on we learn what happens in the past and what

201. 2:01:59

     the consequences are and also L on we predict the next uh behaviors based on what the model has learned so far so as you can see you know on the left side this is the unchain landscape you know we have different notes different edges and we have different interactions between them so what our model is doing is R runs prediction of the next behaviors on chain and on once we have

202. 2:02:21

     the uh prediction of the behaviors based on the consequences what the model has learned in the past it can predict what the next consequences and also what the next behaviors are going to be with uh the context of what the what the edges are going to be or what the nose are going to be so from that what you can do is you will have a foundational model so many people they probably they have heard about what is language model as a foundation model of the internet because

203. 2:02:46

     the majority of the data structure of Internet is language so the foundational model has to be a language model model so what a language model does end of the day is it learns from other relationship of different words and put the next words um and that's how gbt gives an answer all the time but when we're talking about crypto first of all we think the onchain data hasn't been fully used onchain data is a graphical data so therefore initial model has to be a graph model and this is what we are

204. 2:03:10

     developing so rather than predicting next words based on what have learned by this model what the model is doing is for the next Behavior so this is the high level comparison of a language model and and a graph model most of them they are different in techniques but they have been designed to process different uh types of data so for now what the model supports on one hand pretty much all the Crypton native use cases like trading social and

205. 2:03:37

     security but initially since we are early stage we focus mainly on trading side which is what we are doing here so a lot of developers and Traders they've been using our model these days to run ago Quant tradings because previously it didn't have a benchmark model to execute orders or to of different qu strategies to calculate uh statistical calculations and to do risk management but now with with our model it's totally possible because the model aligns all the

206. 2:04:01

     informations available on chain and give you a highly densed output so you can do something on top of that by combining some the metrics you can develop different training strategies and execute different orders uh based on this machine learning guide machine learning power uh decision- making process so for for example some of the developers have been using our uh they've been developing their own back

207. 2:04:23

     tester and started back testing their different strategies and some the Quant Traders this days they've been using our model to predict what is the the next trend of different tokens um and let me show you how to use how to use our back tester so first of all what you need to do is you log in on our official website Crypt pond. XYZ and on you can launch Pond and this is a back tester you can

208. 2:04:47

     develop your strategies and you will be able to use our model directly from our API which is uh uh in our developer Community chat and we can post this link to here as well so what you do is you develop your own strategies and um you decided where and U how you want to execute different orders based on what you have found from our model and also from the combination of your own metrics

209. 2:05:10

     and lat on what you need to do is you simply just add a task uh after you finalized um what your trading strategies are going to be and what you can have here is your your strategy is currently running at the back end and we have I have run some example test for you so we can save some time here so what do you uh after you found the Run status is Success what you can do is you

210. 2:05:34

     just click on detail and will be able to find uh what is your back testing result so from your strategies which is our example strategy here with our model well and also our our our simple strategy what you end of the day will get from this autonomous trading algorithm is you will generate around a 3.4% of Revenue in a in two weeks um you have n profitable trading days you and five losing TR days um so that's what we give you and also give you all the

211. 2:06:04

     metrics that shop ratio and so you can adjust your strategies again and again and iterate on top of that and trying to find what's the best strategy that's tailored for you so you can you know make sure that you win the market so here is the example that has been run uh successfully and you can click on detail and you'll be able to know and find out how it's going to work out for you essentially so still um using our

212. 2:06:29

     application is fairly easy so firstly what you need to do is you can check our API doc which is the link is here and you will have the API access from our model which is the uh the behavior prediction so initially because we're focus on trading the model although it's generalized but it has a specific uh use case which is the trading use case what our API gives to you is the different token addresses with different predicted results such as the price or the grow or

213. 2:06:55

     what different matrics are used to predict this results and on once you have played with our API and you developed your own strategies and you have done your research you feel like you know you don't want to combine XYZ metrix into this model output to fun a small model on top of it or you know just use our model output to integrate it with whatever the models you are using today and on you can develop this uh strategies from our back testing

214. 2:07:20

     system by importing this entire model outputs into this back testing system and run this back test just like what I show you just now so yeah by collaborating with near uh we do feel like we want to push this crypto Ai and deuze the AI landscape to a new level and we believe uh developers are the new astronauts of this era and you know if you are a model developer you are a project founder and you are Quant Trader feel free you know to join us join the trend of

215. 2:07:52

     near and pond and work with us in some ways and um yeah we are always here to help and salute to near salute to p and salute to you guys folks let's give dlan a round of applause again he wasn't able to make it um uh uh just due to some travel issues but we're really really excited about uh What uh he and the team are building through Pond and thank you team for getting this as we're transitioning here

216. 2:08:52

     um so folks you've heard from six incredible teams we had a panel of our Founders uh I hope you're getting a sense of what is happening in the near ecosystem this is very much real very much alive products are ready for usage and so we've been incredibly uh excited about the momentum that is happening and uh we would be remiss if we did not uh give the opportunity for our founder and

217. 2:09:15

     CEO Ilia pukin uh to share a little bit more specifically about the broader Vision the broader road map and things that are happening on near um so I'm going to hand the stage over to him and I'll get your slides queued up here but let's give a round of applause for Ilia thank you thank you hope you guys had a good week uh I'm pretty exhausted but uh I think uh we're all getting

218. 2:09:40

     ready for uh less conferencing and more shipping so I'm excited about that but uh yeah as your kind of hopefully last talk of the week uh I want to kind of wrap you up and then we all going to go and have that Belgium beer so uh I kind of going to give you a quick uh Overlook on what we've been working on youri and kind of some the context of how we get there with user own AI uh for those who are not familiar with me uh I don't know what you're doing here but uh right fine fine but uh

219. 2:10:18

     so my background is in machine learning and AI I've been doing that for about 10 years before uh jumping into the kind of Crypt rabbit hole and uh I worked across actually uh a bunch of stuff before deep learning and then join Google research to work on natural language understanding and research because I think natural language is where we have the human knowledge this is where um

220. 2:10:41

     kind of we have the most information available and we have a very kind of easy way to test same as we test you know kids and adults if they know something by answering asking questions and so we worked on natural language question answering machine translation tasks and one of the challenges we always faced was that the research we were doing using neuron networks with recurrent neuron networks was we could not put it in

221. 2:11:06

     production and because on google.com you have very kind of low latency requirements and uh you know if if it like like a human you know go and reads five pages of the articles that are in search results right nobody going to wait for that on google.com so the question was how do we build a model that is able to read all of the text at Once In Parallel reason about it and give you answers quickly and so that's

222. 2:11:33

     what actually inspired the Transformer model that now is kind of underpinning all of the modern AI which CH gbt mid Journey uh Gemini L and others because kind of that desire for long context understanding without kind of uh spending latency but instead leveraging the fact that compute is highly paralyzed now I worked on that at Google we uh kind of worked on the first prototype on paper and I left to start a machine learning company called neari it was originally designed to teach

223. 2:12:06

     machines to code to really enable everyone to become a developer and program environments in a natural language way now this was 2017 we did not have you know money falling from the sky uh the G CPUs were not growing on trees and so we were trying to be smarter about it we were trying to get a lot of uh kind of valuable data from computer science students to really train the models and to do that uh we you know engaged them in China southeast Asia in Eastern

224. 2:12:39

     Europe and we had trouble paying them because you know some people in China don't have bank accounts we have wead so we need to OTC money into wead uh the PayPal didn't work in Ukraine right southeast Asia same thing and so crypto was actually a solution that we ourselves were like hey this is something we can use to send money around in programmatic way uh and then you you know you look at

225. 2:13:05

     Bitcoin you look at ethereum you know we send maybe a couple dollars a day to uh one of the students and the fees are $5 it's like well that doesn't work and so for us it kind of the inspiration for new protocol was coming from this how do we make it available for everyone really scalable really cheap and accessible at microtransactions and also solve some of this kind of fundamental issues that blockchains have which is you know when you're on boarding a new user they need to go and buy crypto to start using your

226. 2:13:37

     app that seems very strange right especially when we paying them so how do we solve all of those issues now the way we formulated this Vision though is kind of to Encompass what we've started and where we going is to give users ownership of their data assets and power governance right and so assets is something that everyone in the web usually familiar with you know let's launch more tokens let's you know have

227. 2:14:03

     you know your keys your crypto but the data was really important and this is something again in 2019 we're talking about already because it really kind of changes the relationship that you have with you know social media with other places and power of governance as well is this ability to s switch to leave to leave the platform right to take your social graph and uh go somewhere else to take your data to take your Google Docs and import them somewhere else right and none of this is still you know that

228. 2:14:32

     working that well now the challenge actually that generative AI is accelerating kind of the problem we've already had for past uh 10 years and this problem is that you have this big companies that have a very much attention driven economy right attentions only you need after all and this attention is you know turns into money through advertisement this money I reinvested into building a better models that are optimized to make more money to

229. 2:15:02

     get you to spend more attention right it's a flywheel that continues churning and again this is what true before generative AI right you know Instagram and Tik Tok kind of uh getting people to keep scrolling but it's becoming even more important now and in turn it's also starting to break the of internet right readed in in Tech overflow now just become a data kind of our houses that's selling data to those platforms you have

230. 2:15:29

     you know massive trust problems with institutions because you have kind of this po polar opinions that have been shown to get you engaged to get you enraged and to get you to keep scrolling you have kind of you know misinformation has been promoted because it's more engaging content and the reality is this is not a like evil plan that somebody like keep up with in any of this companies it's just a natural incentive structure that's been built that because they don't have more users to Target they need to get more attention out of

231. 2:16:00

     existing users and you need to keep kind of churning this when people launching new models new products they deciding based on what will get people to spend more time what will get people to spend kind of more money on this now regulations in general usually supposed to help us but as we know in web 3 and similar in in AI The Regulators themselves are not equipped to deal with this technology they don't

232. 2:16:25

     know what's the latest and especially in the space where things coming up you know every other week right it's really hard to be kind of a regulator they start and they fall back in AI case toes big companies to actually tell them what's the best right the heads of big companies are on the AI advisory committee and obviously for them it makes sense to capture it in such a way that open source and kind of new startups are not able to par SP so this

233. 2:16:52

     is kind of the backdrop of the challenge and so what is the solution well the solution is we do need a user on internet the same thing we've been saying about open web it's kind of broader uh picture than web 3 and it's combined this kind of set of principles with build in web 3 around user ownership about uh kind of new economies open marketplaces you know decentralized identity and with new transition to a

234. 2:17:18

     platform where AI is becoming the interface where the products are AI enabled and are through them you're able to kind of engage I think one of the important parts is kind of crypto been struggling to find its fot hole and now there is a platform shift that's happening in kind of in the world and you know s similarly how perplexity managed to you know disrupt some of the Google's traffic right we have an opportunity to disrupt existing platforms by building better products in this new way and so that's kind of what

235. 2:17:51

     we're calling this user own internet right it combines a lot of pieces and uh it builds an existing web stack right uh and networking and everything although there's a whole separate topic of there's a lot of issues there and we need to fix them but uh it relies on having personal Creator data personal data is stored on your device right it's not leaving it you're able to run models on top of it so user own AI is able to

236. 2:18:16

     access your local data and leverage it to give you predictions without Le all of this data somewhere else you have obviously trusters interactions which is blockchain you know peer-to-peer communication payment channels all of the tech we've been building that three and you have experiences that are more kind of uh composed at your user level not being built by centralized company

237. 2:18:40

     and shipped to you as kind of modic experience so kind of to formulate user own AI is really uh an ability for user to benefit kind of an and like the models are optimized for benefit of the user you know and benefit in in broader sense both economical as well as more kind of holistic mental Etc and I think it's important to kind of think about this as we kind of go into future is

238. 2:19:06

     like how do we measure this and there's a lot of set of kind of kpis we're going to have for the economic success we obviously have their own chain identity and kind of attached networks to that so we can optimize that and we need to work on other parts now you maintain maintain the ownership of your data uh you're able to kind of control who has access to it and how at any point of time you have this new kind of customized experiences right I want a you know not

239. 2:19:31

     taking app that is able to summarize and give me you know daily digests of everything I noted and you know uh at 8: a.m. puts it on my home screen and then maybe you want to take the recipe and uh you know customize it that it comes in uh in some different form and maybe you know uh reads it out loud for you right this is kind of instead of you know we getting an Apple notes app and we cannot

240. 2:19:55

     do anything with it like we actually have a uh an ability to customize experiences and uh really treat them more as um as experiences instead of the mon gaps and we have a kind of network where everybody benefits right it's not that all of the kind of revenue and profit being sucked out by few companies but actually have a uh kind of equal field where every every body needs to compete and produce and uh kind of uh contribute to actually benefit and I

241. 2:20:27

     think that's an really important part here we're not trying to bring down some of the centralized parties we're trying to even out the field so everybody can participate so there's a whole stack of things that needs to happen so internally this is what in a way happens inside one of those bigger companies and Labs right there's a whole data layer with data acquisition and data crowd sourcing the curation and you know teams creating synthetic data you have you

242. 2:20:53

     know whole teams that handle compute and inference infrastructure then used for you know specific models as well as running models on on user device and you have application stack which like ties it all together and provides tools to Consumer and business applications right so we need a version of that that is open source that is more than just you know one company providing everything but actually a combination of partners coming together and and uh providing all of this uh to the users and developers

243. 2:21:25

     and this is kind of what we've been doing uh with obviously Horizon and recruiting projects as well as just broadly partnering with the kind of broader ecosystem and this is again not just in AI web 3 it's also in traditional AI companies who providing some of the pieces and how do we hook them in into kind of broader picture now specifically on AI side we brought back the near AI right the original know kind of R&D lab that we started with and it has three specific goals one is how do

244. 2:21:57

     we kind of advance and coordinate open source AI research uh and I'll talk why this is a problem how do we make then this research available to everyone else to all the developers hey I want to use the most state-of-the-art you know agents at solving this problem and finally ourselves we want to leverage this open source AI R&D to build AI developer to change how you interact practice

245. 2:22:22

     Computing so the reason why we care about the AI developer is you know this is kind of my long time you know I was been trying to do this actually like I think from 2006 um because I think the interaction with Computing should be natural language it should be semantic it should not be syntactic and you should be able to you know uh interact in such a way that the computer itself builds software

246. 2:22:51

     able to execute it and able to kind of in like do things on your behalf without you kind of clicking explicitly now in web 3 it's an really interesting space where we have more infrastructure than applications right we have you know tens of thousands of infrastructure projects and probably a few dozen applications on this and on top of it even across the whole web 3 we only have you know this is like a little bit outdated but still 7,000 full-time developers 22,000 um kind of part-time this is less than

247. 2:23:23

     like one Google department just to give you an idea right um so this is very few and like the barrier to entry for smart contracts is pretty high as well uh you usually have like a senior developer and like if there's no Junior developers like how did you become a senior so like smart CCT developer so so that's kind of the problem space right on the other side with AI capabilities we actually have a really interesting opportunity to Advance how AIS work right I mean for those who

248. 2:23:53

     tried you know using them in your daily life like they work pretty well for you know code completion when you kind of supervise it you can do some interesting like first prototype like you know of a user interface they kind of give you like high level interesting things but as soon as you go and like actually try to get it to do a very specific thing correctly right they usually fall off

249. 2:24:16

     get stuck and do stupid things and so the interesting thing with code compared to many other spaces is that we're actually able to test if it's correct right we're able to actually verify that the code that it written is correct and so in research that's a very useful thing because now you can have people iterating on it and trying different things and being able to compare how it works between different

250. 2:24:40

     approaches and so it's really great environment for testing those reasoning capabilities of different AI models of you know agent setups Etc where you have kind of pretty complicated tasks and a very specific test sets for this now even more important for smart contracts we don't want just some smart contract we want a correct smart contract that is you know not losing anybody's money

251. 2:25:03

     you're not able to hack there's no like rounding error somewhere on the edge and so the important part going forward in general will be that we need the formal verification for everything we do for every code line we write and this is actually going to be important long term because even now the slms are are actually pretty good at finding vulnerabilities in the code and so this

252. 2:25:25

     is just going to get better and uh we kind of need to rewrite all of the software with form verification if we want to not be living in the world with you know zero day variabilities every day and so this is a really good environment to test that because form modification on Smart contracts is a space which got a lot of investment there's a lot of people look at it uh there's a lot of tooling as well uh and like as well as in Rust

253. 2:25:51

     now the other part is you know you wanted to be able to discover new apis right hey you know somebody just built a new project on near let's figure out how to use that you know hey there's a smart contract that was deployed we should be able to use it right so it's a really good way also to test if it can pick up new tools new uh apis and finally okay agent build an application right it still needs to go

254. 2:26:16

     and test it itself right it still should be able to go and like click through it it make sure it works right so this is actually a good environment to test out uh kind of how the interaction of applications work which is useful then to go in for process Automation and other use cases as well so this is a very like common skill that you can leverage so AI development is in a way like a really nice environment to test a

255. 2:26:41

     lot of different skills and kind of make make progress on them and create like this automated benchmarks where you know if you ask it to plan you know research plan for for like biotech like how are you going to test it right you need to build a lab to actually test if it's a correct plan and here you're able to kind of have an environment where you test everything really quickly now the other question is is this science fiction right because remember we started we pitched this in

256. 2:27:06

     2017 and started neari was that and back then it was a science fiction right like most investors were like huh um and so I think we've went from you know complete high like back then you know we were we created a custom data set because on a data set we actually collected from for real problems it was complete zero right so we created a custom data set which like deep mind and cited that they beat

257. 2:27:33

     and uh uh now we see on a real data set so this is uh this is kind of basic python problems which like already at 95s this is actually solving real GitHub issues on GitHub right and we went from 0.17 uh in um October last year November last year to already almost 20% uh in June right so the progress is pretty quick uh when you have a very like well set out Benchmark and so what

258. 2:28:05

     we're doing right now is actually setting on B benchmarks to kind of showcase more and more complex problems uh so that people and researchers can actually iterate on this but the point is like the progress is happening and it's happening really fast so the other side is this is all expensive and complicated right so what we've been doing is not just attacking the problem itself and research side but also setting up a framework so that everybody can kind of uh really Target it in the

259. 2:28:35

     open- source world right we want to coordinate compute and data acquisition crowd sourcing in such a way that uh you can come as a researcher and really uh you know go at it and be able to make progress and have kind of access to as much resources as you would in a kind of one of those traditional labs and so to do all this we kind of uh we have it's a um in quotes right now Hub uh pending brand brand uh um

260. 2:29:07

     Evolution but importantly it kind of connects it's a connective tissue between all of the different components right there's lots of and lots of different projects providing compute right be that kind of raw gpus or uh inference or maybe some training capabilities you have lots and lots of projects now doing data right this is novel data acquisition data monetization synthetic data marketplaces human labil

261. 2:29:30

     data and various other things uh and then we have researchers who are sitting on Sidelines in Academia and kind of like individual developers who kind of looking at some of as benchmarks that are now like uh at 95 and don't know how to engage like and what problems to solve as well as they don't have access to some of this computer data and so Hub is providing a way to coordinate all of

262. 2:29:54

     this resources get credits as a researcher showcase your idea and kind of up amount of computer data you can access to as you show in progress and this is actually how this research Labs internally work it's you know if you have 300 people not everybody gets you know 10,000 h100s all the time you actually show your experiments you show progress and then uh that work is gets contributed to a bigger model that's been trained kind of for everyone and so that's the same idea here to actually

263. 2:30:24

     surface all the experiments make them all in the same framework make sure that work that somebody done in another part of the world is reusable by you you can see the experiment running on The Benchmark you can see what it done well and what it's missing uh you can see new data sets being contributed and how you can leverage them in your models and finally all of this gets then packaged and right away usable by uh application

264. 2:30:49

     develop because now you can just access it to the same API right very simple kind of open AI style API uh which you can start accessing and importantly you can also Define your application specific benchmark right let's say you s you are you know doing something in medical space you was doing something in financial space you can Define your benchmark you can upload this to the hub and now all the researchers can start

265. 2:31:13

     looking at this and you know the traditional models are only doing like 0.1% on it people will be really excited to make progress because there's so much room to do it and so this is really kind of the way we want to coordinate kind of Open Source research to happen where the real applications are now bringing cap like their capabilities they're looking for the capabilities that are not right now there in the models and they

266. 2:31:36

     surfacing that to the researchers to then go and use compute you know request more data labeling request more data acquisition to then solve those problems so specifically what we're doing is this applying all this for the developer right so we have a set of benchmarks both the ones that already public and we're going to be publishing specific kind of more um code competition and web three Benchmark we predevelopment benchmarks and kind of the way we're structuring this process is there's a user talking with a aipm AI

267. 2:32:09

     product manager uh where you're refining the specification the task you're doing the the IPM is actually asking questions and uh you know asking for tradeoffs in your application and then you have a AI developer that goes from specification and builds a web3 application which contains you know smart contract front middleware to give you an example let's say you want to build a decentralized freelance Marketplace uh the aipm will go oh well here's how I think it's going to work here's the user stories you know

268. 2:32:40

     here's kind of the default user Journeys maybe components that will be engaged and you can refine it like hey I want to be able to have an escrow that you know goes into a for example to uh make yield while you know the freelancer is working on the project etc etc so you can refine kind of how it will work and then the developer will go use that specification to write the smart contract you know job listings you know freelancer profile Etc

269. 2:33:06

     write the front ends uh and kind of hook it up together with middleware so this kind of in a prototype stage right now uh and we're kind of interested in getting like first few people to try it out as well as the Hub um where you can you know upload your benchmarks as well if you're a researcher to try it out and see if you can beat benchmarks and kind of use this framework a lot of it right now is still again under very active

270. 2:33:29

     development we're figuring out the specific apis to make sure it's generic enough where you can do almost everything but it's specific enough so that it's reusable then uh for end users and developers this PR much it I'm excited for everyone to come here for projects that are building in the ipace and uh yeah hope you have a great uh end of ACC this was a a really big undertaking before we close I want to give a few final shout outs to some team members who are here the first is actually

271. 2:34:12

     Pierre Pierre is one of our software Engineers Works directly with all of our teams has done an amazing job getting uh these teams integrated the second is Cunningham war is our head of product uh she is actually one of the masterminds behind this whole program and really has overseen all of the strategy behind this and the third is Mark me Mark keeps all of our investors informed all the important people informed keeps me informed um but really really appreciate

272. 2:34:37

     uh the team and most importantly Ilia for taking time to share with us today folks this is just the beginning so uh really appreciate everyone coming out we'll be here till 5 grab some food and drinks and uh have a great rest of the day