NEAR AI Ecosystem - What Did You Ship This Week 3 - Show Off Real-World Agents

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

1. 0:17

   Near AI, what did you ship this week? Good morning, good afternoon, and good evening to everyone watching this live stream. My name is Cameron Dennis and today we are entering our third episode of what did you ship this week with Near AI. We have an amazing group of developers and founders in the ecosystem going to be presenting a little bit about what they're doing what they did over the last week. Uh for those who've been at ETH Denver uh very clearly know that Near AI team has been shipping a

2. 0:48

   lot. Um so the Eric will be talking a little bit about uh what we're doing internally but uh this space is really intended for the ecosystem to essentially provide a engineering standup uh to show everyone what they've been shipping. So we have an awesome awesome crew today. Uh we're going to kick it off first with uh Robert Yan who's been contributing a lot to Eliza over the last couple of weeks. Uh then we're going to move on to Aaron from

3. 1:11

   Krenics uh to talk about some uh deepen uh work that they've been up to and how they deployed their EBM contract on Aurora. Next we're going to go through uh intellects and our friend Omar uh who will be talking about uh activators management and model training on devices and then uh after that talk with Jes from Jutsu about all the progress being made on the Near AI TypeScript SDK.

4. 1:37

   Moving on to Shu from Mizu to introduce uh Curryifi and uh the app built on top of Mizu. Super exciting. And then uh we have intern Barry with Strawberry Eye to talk about Barry Chat in R3 that recently launched. And then to close it out, we're going to be kicking off to Eric from Near AI. So without further ado, uh actually before we top into that, I just want to quickly highlight some of the cool things that were built uh that are not here uh that was announced at ETH Denver. and one of them

5. 2:06

   was a TE enabled agent hosting with the near AI hub which is incredibly exciting. We're going to come out with some blog posts and some tutorials around that. Um, additionally there was uh near cloud operator agent. So now there's an agent that can actually manage your browser for you and you can give it these pretty awesome uh commands and it will literally go around and click. And so we highlighted showed a demo at ETH Denver or at an event that

6. 2:30

   we ran separately in Eth uh booked Airbnbs for people. So pretty cool. Uh there was another one for the assistant that was showing how you can buy um uh headphones with just natural language. And Eric will probably dive into some of these more details, but it's been an incredibly awesome week on the Near AI side for shipping. And then we also just want to flag we running a hackathon in San Francisco on Friday, March 14th and 15th about building useful agents. Uh as we've all sort of seen a lot of these agents are cool.

7. 3:05

   They're fun. They're a little gimmicky and not totally useful for enterprises or startups or you know everyday consumers quite yet. So we're running a hackathon called the useful agents hackathon in San Francisco March 14th 15th uh with Electric Capital and SF Compute. So, lot of cool stuff coming through the pipeline. More to come probably after. But just to kick off the space, I'm going to pass it off to Robert to talk about what he's been up

8. 3:30

   to on the Eliza OS. Robert, I'm going to bring up the stage. And just for some house rules, uh, welcome Robert. Uh, please be careful when you share your screen. Only share the tab that you are intending to share. And, uh, we don't want to make any mistakes. So, yeah.

9. 3:48

   Thank you. Absolutely. So, yeah. see my screen. Not yet. No. Okay. Give me a second. Take your time. Oh, yep. I see it. Bring it up on stage now. Boom. Oh, and by the way, uh I'm going to pretty much call people off. Everyone has around seven minutes. Uh we're starting a little bit late, so this might be a little delayed, but Robert, you have about seven minutes.

10. 4:15

    Take it away. Thank you. Thank you everybody. I'm Robert. I have been working near the ecosystem for like over four years. Very happy to join today's session and today I want to talk about elizer and near agent stack. And so what is what is it? So first of all this deck in this deck we're going to building all the agents with lizard which is very powerful framework that it supports a

11. 4:40

    very interesting proging ecos ecosystem that you can talk with other blockchains and you have a lot of uh AI models to choose from and what else you have the various adapters for your database and the rack the storage to which can be used to manage the memory of your agent also it's for a lot of clients so uh We really like this framework and also we had a very strong community today and we are using as a core to build the agents.

12. 5:08

    What else? We are using UI as a hosting platform for the agents because at first it's very friendly for beginners especially providing the free inference and API and the hosting service for agents and it's also provides sorry very uh powerful uh AITV protocol for payment and the communication uh among agents.

13. 5:34

    So you're not only building agents, you're building a business and make your agents able to be monetized. Also we are uh EI provide PT features to protect the privacy of agent agent and its users and for open web and crypto we are using near protocol because very performant and shy blockchain also it provide a lot of crosschain features to interact with multi- chains via chain signatures like bitcoin and eerian also

14. 6:08

    we have a new intense to do crosschain swap So combining all this together we have this very powerful agent stack that you are you can using all the frameworks you already familiar with to build the agents in your lizard. You can use all the plugin you want but also you can host the entire agent within your AI.

15. 6:29

    You can use a frame APIs and you can use the payment features. You can use privacy functions with TE and do the discovery easily with all the features we have. What's more, we have uh developed several plugins for you like we had a new plugin that you can use to intex with near blockchain, but we also we have a multi training that can use the same account but you have the deration pass with the uh bitcoin account and the

16. 6:58

    account you derived to interact with bitcoin the ethereum and etc with no efforts. Also you can still use other plugin like name generation in your agent and with all the existing clients and database etc. So it's really powerful framework not only to for near blockchain but also you can use the same framework to build agents for other blockchains and use all existing product in in the ecosystem.

17. 7:32

    Yeah, that's why we need this agent framework because it's not only open source, it also had a very strong communities there from both new and eler also it has a modular design with its product ecosystem. It's also had very bad a great support for privacy payment and existing and the seamless support for multi-chain features and alo has free inference and hosting service and the most important is our wood cross stack not only for building like toys

18. 8:05

    and demo projects. It's a serious ecosystem that uh frameworks use to build this real businesses and even unicorns unicorns in future. So how to start? If you want to try this form now, you can just visit this link and uh you will see all the plugins and the features we have included there. You can just use all the existing actions like you can just send this message to the agent. It will use your near account but using the chain signature features to send the bitcoin to another account

19. 8:42

    with the single message. It's very very oh sorry you cannot say that's another tab I just show this one but you can just open this link and uh have a try. So yeah um what's more we want to extend this framework more in future like we want to really host the uh it is a agent on your AI so it can benefits from the IATP payment and like the T privacy environment features also we want to include the near intent actions by fat the m work there and have it in the multi

20. 9:23

    training also we'd like to extend and the new plugin to include more uh actions for different deps. And very interesting program we want to develop is a bitcoin pluging that not only support a native action for bitcoin but also we want to use a set bridge uh to make it possible to request signatures with your bitcoin account. So you can use a bitcoin account to do all kinds of crosschain feature I just mentioned for

21. 9:55

    new account also you can use that to do the crosschain swap powered by the new intents. uh you know we would like to extend our features uh to support this share a agent plugin and we can use this agent to empower all kind of smart contract to be really really powerful and if you want to learn building agent also provides some tutorials for you please take a look if you're interested yeah that's pretty much I

22. 10:28

    want to share hope you enjoy it well this is awesome Like this is I think the way sharing. Yeah, you can stop sharing. I can take it off too. Um I'm so excited about this integration because what this will enable is any agent that is built with Eliza has the option to transact across any chain. So agents aren't going to care what blockchain they're deploy to. They just want to go where liquidity and users

23. 10:52

    are. So I think this is like a fundamental shift for the entire agent space. and the fact that they get free and private inference. I don't know what else people would want, frankly. Oh, besides distribution. Um, and Eric might talk a little bit more about distribution later. Um, but Robert, this is incredible. Um, and again, thank you for creating all of the online education materials, like everyone that's watching

24. 11:14

    this that does have a small community, like a developer community that you like to educate, take these m like materials. They're for you. So, Robert, thank you so much. Uh it's amazing having you contribute to this and super excited to see what's to come and I think that this is actually a playbook that will be kind of iterated again and again and again uh for every agent framework. So free private multi-chain inference for everybody. Robert, is there anything you

25. 11:42

    want to add like any um Twitter account people should follow you at? Anything else? Yes, you can form me at Robert and Zo if you want and hopefully I can share more about update for this framework in future. Thank you. Awesome. Great. So Erin, we're going to bring you up on stage if you don't mind just uh turning on your camera. Welcome. Welcome. It would be great for you just to give a quick overview of yourself, a little bit about your background, a little bit about your company, and then what you ship this week. Yeah, sure. So, uh this is Aaron.

26. 12:15

    I'm a founder of Kronox project that utilizing edge computing for AI tasks and uh before founding Kronox I was working Google for over 10 years mostly in Google search uh research working on AI stuff uh was involved to uh first engineer of the tenflow light on NLP and also the uh Germani the working on the aon fatality and many other stuff. So today I am going to share the uh our effort on shipping uh deploying kronox on near and uh lot

27. 12:57

    of details here how we uh how we solve the Chinese here. Uh so first let's introduce what is kronax. So for kronex we are doing things differently. Um so uh there are tons of DPM projects and uh for most of project you see you are using those DPIN GPUs just like you are using on AWS you select which GPU from which provider you are uh they're providing and log to this uh GPU and use

28. 13:30

    it and for Kronas we think that if we want to build a purely permissionless and trustless uh protocol we cannot provide this kind of functionality What we're doing here is um what we do here is for any incoming API we are going to schedule this to uh uh remote edge nodes and execute this on edge nodes get result and uh we do verification on this result coming and send back result to the client. So we focus on two things

29. 14:03

    task scheduling and content protocol. Here is our protocol. So for any AI task, we are going to deploy this schedule this through a blockchain. We are going to schedule this on three nodes. Uh 10% of chance we're going to schedule on three nodes only if the three nodes return the same result. We claims no one is cheating on us. And for rest of 90% of task, we're going to sing occur by single node. So overall we have

30. 14:33

    uh uh 10% of task esq nodes 90% of task esq single node suppose I'm a node I need to st 10 token before executing any task so I have 10% chance to get my 10x token be slashed which is econ economical on wise to cheat on this network the amazing things here is everything here is coined by blockchain so there's no server uh for this protocol This is why we can achieve permissions and

31. 15:04

    service. Everything looks perfect until we deploy on chain. So what can you expect? What are the problem here? So the channel here is because everything is uh depend on blockchain. So everything is uh deployed uh everything is coined by the smart contract and for any single air task we are going to trigger multiple smart contracts and due to the nature of the current blockchain infra uh it's very expensive and uh and

32. 15:38

    slow. So here is our solution. We are trying to deploy. We start our project as uh EVM smart contract on substrate. But when on near, we noticed that near is not EVM uh native. But luckily we have Aurora which is EVM compatible to bridging our smart contract to uh near. But as I mentioned because of the all these challeners I had we cannot deploy directly to Aurora.

33. 16:08

    So we need to imple implement our own op stack to make sure that we can make efficient execution on Aurora. So here's what we're doing. We have thousands of nodes and uh we currently not have thousands of apps but we have we have a few apps that running on top of our API.

34. 16:29

    So there are still thousands of the EVMt uh contract transactions every every minute. So what we're doing here is we deploy this on the OP stack with our own sequencer and package this contract to Aurora. So in this way we have much fewer uh transactions. So we have much less transition cost and much faster execution with this sequencer. So in the future we are going to make it a general solution for EVML2 transer. Hopefully this can also support other EVM contract to deploy on on near.

35. 17:09

    So uh so every project they have their own op stack as L2 and Aurora can handle this and they can bridge into near without any effort. So right now we are solving this problem. That's it. Thanks. Awesome man. Well thank you for coming and um what's the best way for people to follow up with you after they want to learn more? Sorry, what's the question again? What's the best place for people to follow up with you or Okay, sure. So, uh follow our uh Twitter account that is

36. 17:42

    uh Kronox AI uh C R Y N UX AI. Awesome. Cool. Well, it's great to have you. And up next, we have Omar. So, thank you, Aaron. Taking you off screen. Welcome, Omar. Welcome back. Hey, Cameron. Thanks. Thanks a lot. And for those who may not have seen previous episodes, can you just give everyone a little bit of background on yourself um about Intellects and what you shipped this week? Yeah, absolutely. Absolutely.

37. 18:11

    I'm Omar. I'm co-founder and CTO uh at Intellect AI protocol. What we are building is um an agent a framework and an infrastructure for agents uh interoperability and coordination for multis swarm architectures. That means from the decentralized federal learning for the building of intelligence, asset registry such as agents, models, data sources and contribution incentives and a lot of things more that we we will speaking today.

38. 18:41

    Amazing. Do you have a screen you want to share? Yes, I do. Um on the meantime uh last week for those who doesn't remember we have been speaking about uh what our framework does and we we we show some demonstration on how uh agents coordination and and execution happens behind the the scenes. Today I want to show what happens on the token holder side of things and at some point we will

39. 19:11

    also show what happens on the business side. Um so today we will focus on what we call our activators management platform. But first it's important to understand what we call activators and like any protocol we have uh our own names out there.

40. 19:27

    Um the way we call agents whether it's individual agents or swarms um for us there are activators and those are agents that are managed by intellect token holders um somehow becoming their digital force. is like putting agents uh at the service of people to do particular task. In our case is to solve business problems that we call activations and they are composed by one or more tasks. Um one simple example related to what I presented uh last time is I want to improve baby rib sales by 30% on my business. So I'm a sales rep.

41. 20:03

    I have a product to sell to restaurants and I need some help from the from the AI to to to solve this. So that's for us is an activation that will be divided and break down into into smaller tasks. Um the the way this work is that any intellect token holder can own and manage those activators by what we call active staking. So instead of you know putting some tokens there not doing anything. Uh for us is you put those tokens somehow to work and making those

42. 20:33

    activators um smarter and faster that's the simplest way to explain it. um to solve tasks within activations. Um when a task is solved then those activators will receive rewards. Uh but in general all the different kind of contribution over our platform um uh that are related to solving a task will get rewards.

43. 20:53

    That's contributing data for training of the models um validating that data validating the outcomes training providing compute and and and and more. Let me move to the next one. The way this work is of course by leveraging near AI capabilities. You have been presenting a lot of the different capabilities we will be using in the near future such as discovery, payments and many more. But on the other side, we are also using the Intel X AI protocol when it comes for the way the

44. 21:26

    collaboration particular how the intelligence is built, how the the the federation of that intelligent happen, keeping privacy on the different members that are contributing that intelligence, how the interperability works and all the different incentive across agents and the different stakeholders uh happens. Interoperability is such it's very important topic for us. um on the

45. 21:49

    way um agents depending you know regardless of the network and and and the framework that they use how they talk each other how they achieve uh consensus and stuff like that um and finally also leveraging the data pipelines and well when I when I mention about intelligence building uh from the uh AI from the intellect AI protocol as decentralized learning as a registry and more what I want to show today is the way you know how that activator management management platform works, what we see as activators, activations,

46. 22:23

    dashboard and so on. I try to go very quick so we don't we don't lose time but mainly we will see um how it looks when you are not logged in uh how it looks the dashboard the activators how I can create assign and stuff like that. Um I hope we are still looking the same screen. No I will have to change quickly. Activators management. Okay.

47. 22:46

    Yeah. Let me share again. Uh yeah, I should do this. Perfect. Awesome. Perfect. Um just uh as a heads up, we are still in the process of uh connecting uh near wallet. This was an older version that was EVM based but this is the way um you can see the different activators that are online and the different activations uh that are there uh where you can apply some of the activators to solve those tasks. Of course, each activation will

48. 23:23

    have a description uh depending on the complexity of the activation might need more or less activators might have some that are running if needs uh if it has ferated what's the topic about and some more information when I get in let's connect should come back awesome immediately I will look at my dashboard and I will see well the tokens I have available the ones that I have staked on on different activators, how many working activators I have and how many I have that are available uh to assign to some work. I can see some

49. 23:59

    numbers. We're trying to make this very very simple for any token holder that doesn't need to understand too much about AI training models and stuff like that. So we try to you know put it in a way that is making it faster and smarter behind the scenes. that means that you will have more trained agents available that you can do more training on those agents and that's the way you can kind of play a game because some of the task

50. 24:26

    might not need to have such a powerful activator. Um so it's not just you know I have the most the most smart smartest activator and the fastest activator you might be overspending uh in unnecessarily. So you can you can see on the different activators you have you know how they are working uh how much accuracy you know the the feedback you are getting uh from the different tasks that are being solved you know how is

51. 24:52

    going on and you can act on that and you can change some of the um the features. Of course you also have a job board where you can see different task that needs to be done. So it's kind of you know applying your activator to a particular job and at the end if it solves you know and gains the it wins the race for solving some of those task you will be earning rewards uh for for for fulfilling that and I think I'm running out of time. Yeah we're running out of time but this is super exciting.

52. 25:20

    Um thank you for sharing. Is there anything else that you want to you know tell everyone to go follow you at or anything that they can do in the meantime? Yeah absolutely follow us on on Twitter, Telegram. We are we have again a lot of things uh to to show you know one at a time and we will need everybody's and everyone's contribution you know there will be a lot of things to do within the the ecosystem we are building. Awesome man. Well thank you

53. 25:45

    for coming and next we're bringing up Dylan from Pond. Dylan welcome to the stage. Uh we can't hear you. Uh, I think you have to go change the settings in the app to make sure it's connected to your headphones instead of the the computer. Great. Um, Zahed, actually, do you want to you can we can swap people.

54. 26:12

    Zahed, if you're ready to go next. Great. Howdy. Howdy. Can you hear? Hi. Yeah. Yeah. So, welcome back. Welcome back. Yep. Um but again, for those who might not know you, have not watched the previous ones, please introduce yourself, a little bit of back, your background, um what you're building with Judsu and then what you did this week. Okay. So, hey everyone.

55. 26:38

    Uh I'm Zahed, founder of Jutsu and Jutsu is a kind of a aent platform. uh and we are working with NIA to build a typescript uh SDK for uh near AI agents. So NIA have this ambitious goal of kind of building one trillion agents and I believe if we want to achieve that goal we have to bring all the web developer into building agents and the best way to uh bridge that gap is the TypeScript. So um we are working on kind of improving

56. 27:14

    building and improving the uh TypeScript developer experience. So let me share my screen. Uh do you see my not yet I have to add stage but there we are. Great. Okay. So uh let's do that. So this week basically uh what we did is we uh kind of I onboarded few users and kind of observe how they install near SDK and what are the challenges when you kind of install and kind of start trying to build the agent and that was a really great

57. 27:56

    learning experience and we also build over 10 agents at the end I will show a example of real agents that produce some great result. And then uh finally the getting started guide for TypeScript is in the pull request. So probably in a day or two uh that will be added in the uh yeah documentation.

58. 28:19

    Uh I also added a image generation uh method in Typescript. So from Typescript SDK now you can not only generate text you can also generate images. Uh and tonight we are going to run a workshop on near Typescript SDK in SF in SF node.

59. 28:42

    Uh so SF node for background is the largest uh NodeJS and JavaScript uh community in the Bay Area. So I'm excited for that one. Some of the stuff that we'll we'll do next week is kind of a building more agents, onboard more users. we will start creating some video tutorials because it's very important and adding new features in uh TypeScript SDK and this is a example uh where we build a uh kind of a agents that write

60. 29:13

    children's book u so I'm going to show my code just tell me like you see my uh screen looks great okay perfect so this is a agent JavaScript uh TypeScript regions that take a topic it write a books for children's for any topic that you want to teach your kids. Uh so basically it create a books it create the chapter for each chapter it generate a image it add those image and then at the end it create a PDF book and the result

61. 29:55

    is this one. So you can see the advent adventure of Kudo and his magic garden and this entire children's book. This book was written by uh you can see llama 3 uh llama version three uh and it produced really good uh result. We have bunch of other books that I tried. So books about gratitude uh with all these chapters. So this is a book about gratitude. So you can see it's a beautiful book with great image.

62. 30:33

    So I believe uh kind of a TypeScript SDK is getting really uh in a great stage in in a few weeks where we can build real world application using TypeScript and uh just look into the uh docs.ne.ai AI and we'll keep updating the TypeScript doc and in the future we'll create a kind of a cookbook where you will have different example including this one like creating a book uh using uh AI.

63. 31:07

    Yeah. So that's kind of a my update and if anyone have question I'll be happy to kind of answer. Yeah. We haven't mastered the question piece to this yet. Okay. Um, I'd love for you to send me these links. If you can send me just the agent links, uh, that'd be amazing.

64. 31:28

    Beautiful. Thank you. This is super exciting. Like agents that are doing useful things. Like, who would have thought? Um, and so, and by the way, if you have a kids, like definitely go write a book and print it. They will love it. No, it's amazing. Uh, I know some people on this call have kids. Not going to call them out, but try this out. you know, use these agents um and provide feedback. You know, we're we're always here to help. So, Zahed, thank

65. 31:54

    you so much. Incredibly exciting. Please send me the links so I can drop them in the thread. And uh next uh you know, hopefully Dylan, how you feeling? Can we hear you? Bring you back on. Where did he go? And Dylan's gone. So, um next Barry intern, are you down to come up next?

66. 32:15

    Just give me a thumbs up. I can see you in the great Welcome, welcome to Hey, Carol. How are you? It's doing great, man. Living the dream. So, can you give everyone just a short background on yourself? Uh, what Strawberry is and then what you guys ship this week? Yeah. Um, I'm I'm the CTO of Strawberry Eye. We're basically solving search and execution via AI agents on the crypto

67. 32:46

    world. Basically, we're trying to solve every question you may have to ask for both research and trading or bridging or swapping in general crosschain wise without you even have to think about the chain you're rooing on. And yeah, one piece of the time we achieving this goal started by this week we released just a tiny piece of the infrastructure which which is not so tiny which is berry chat which is will be the the

68. 33:22

    house of our the AI agents and yeah I'm going to share my screen so you may have a look um let's I'll have to pull it up, but I'll probably see it when it pops up. Hope you can see the screen. Yeah. Yes, I guess. Uh so uh basically uh yeah I just for the sake of time I shared some common question that you can ask to bury chatar3 which is the latest uh mixture of agent super agent which is basically

69. 34:04

    orchestrating all the tiny agents underneath it. Uh for example, we have the price research agent, the uh news agents, the Twitter agents and also defy lama uh many integrated uh agents towards specific uh crypto tasks. So uh for example here I'm basically requesting the TVL of a or in uh this is interesting. I just tried it now and I asked it to compare Aanon versus Strawberry High and it it not only gave me all the price actions and

70. 34:46

    all the stuff but also the rate the latest news about them and doing the comparison because it's actually it was using two different agents one for the prices and one that we're looking for the ticker across Twitter and also since we're a chain in agnostic. Let me see if I can find the the proper one. Oh yeah, you can basically request the holdings of any uh crypto wallet. Uh if it's EVM or Solana, it's going to or

71. 35:23

    even an ENS domain, it's going to resolve that and fetch all the tokens that it it has. Here I'm asking for the Vitalic wallet in general and it giving me a summary. I can go deeper and ask it for more and yeah, for example, I'm requesting the latest news about near it. It's basically running on Twitter only by fetching like hundreds of thousand tweets per day and indexing in our rag pipeline and it's doing pretty

72. 35:57

    good. I I've been following the news about near and it's straight to the point. Let's see if I can find some more example. But yeah, this is just the research phase of uh Barry chapter 3 because the next goal will be releasing Barry swap.

73. 36:18

    Uh the research part I is only useful for the user that don't want doesn't want to like go in another place to uh take his questions basically. U so in order to like allow people to trade we have to support basically any wallet and we just let me see if I can sure yeah I guess Okay. Yeah.

74. 37:08

    So basically you you it's a I I guess the one of the first uh chat interfaces that it's web three friendly because you can straight up connect with your wallet and since we will support trading you can actually add multiple wallets and in in the moment you will be able to execute trading and bridging uh you can manage multiple wallets inside one single account. You see the the text underneath it which is

75. 37:39

    preview. H but it's not the only thing that we achieved this week because right now we have a a huge problem even though this is a pretty good interface. We have a problem with mixer mixer of agents architecture that they are mostly since they're there are some so many interaction between them in underneath there's little to no feedback to the end user for example if I ask give me the latest news about

76. 38:15

    Barry it's probably going to stick over over that for like one or two minutes because it's doing a lot of stuff. So the next phase for berry chat R3 v2 will be providing a small thinking steps and I can show you a demo because we basically rebuilt the old architecture this week in order to support that.

77. 38:44

    Uh let me check uh give me the top protocols of near. So you can see it's basically in parallel while it's doing it stuff. You can see it's the rack pipeline of the agents. It's basically printing out the steps. So the user will have a constant feedback of the operation that the AI agents are doing underneath that because otherwise it's like you have to wait probably for one minute especially because you are using reasoning models in in this case it's deepse three which

78. 39:25

    is fairly accurate but also has a lot of thinking parts and all all that you care at the end of the day is the result but you you also have to be entertained otherwise the user is going to drop the UI and like this for example has already finished it and I guess it's pretty accurate too seems like so and in the meanwhile if this were on the new UI you would have see all the thinking steps kind of miming uh chat GPT

79. 40:04

    UIs So this is going to be live actually this week. This is what we accomplished by rebuilding all the AI AI agents architecture for this week and it's going to drop on the chat interface this week and the next step will be bringing the execution part which is actually swapping. We already have all the price feeds and the contract address of everything like we we've been scraping that forever. We we collected more than

80. 40:36

    33,000 tickers with prices updated like every hour and not only counting all the Twitter posts that we scrape every day. So we're going to integrate by next week the uh the swap interface and finally bringing also execution to Barry chat.

81. 41:00

    This is awesome man. I have so many questions. I wish we had more time. Uh especially on like data sourced me as fast as possible. So no, there's too much. But thankfully we do this every week. So um maybe we should probably just do another one like strictly Q&As's at some point. Um but this is sick.

82. 41:18

    Fortunately, we do have to move on to Dylan. But amazing work. What should people do in the meantime to either test Barry Chat, follow you guys? Like what's the call to action for the community? Well, you you can already use a berry chat. It's at chat.youustroberry.ai or just u give us a follow on Twitter. We pushing constant updates. You can search for strawberry eye or follow me on Twitter. I'm

83. 41:43

    intern basically the reverse of this nickname that I put on the live chat. Awesome. Okay, so Dylan still struggling. Shu, are you around? You ready to go? Barry, intern Barry, appreciate your time. Techno stage. Thank you. Thank you. Shu, are you ready to wreck?

84. 42:06

    Yep. All right. Welcome. Welcome. Just real quick, um, name, background, what is Mizu? What did you ship this week? Yeah, sure. So, my name is Shu and I'm found of Mizu. Uh before found Mizu I was like walked in Facebook for like seven years in data infra and all this kind of Libra stuff and uh Mizu is I found Mizu like in uh end of 2013 uh 2023 and uh uh Mizu is basically a network for people to group their laptops together and uh basically group

85. 42:42

    the computation power of laptops and desktops or these kind of edge devices and provide a unified API interface for upper layer applications to And we believe that for this kind of H edge AI networks uh because of the nature of the network it's like it's unreliable it's high latency and uh there's no SLA so it's more for them to it's better like you know basically this kind of network is uh it's hard for them to do this enterprise tasks but they are

86. 43:13

    very good candidate for asynchronous tasks like bash API provided by this open AI and it can provide like extra benefit benefit. That's why like we're finding like we're looking forward to some like have some asynchronous AI network applications or use cases. So we built one by ourselves which is uh uh curify. Yeah, curify is basically today I'm showing like the application we build on top of our network and uh uh

87. 43:37

    underneath like it's using a hybrid hybrid network that uh we use GPU like the the cloud service at the backbone but like know we shift workload to our edge network to reduce the cost. Yeah. So Curify itself uh is uh is pretty interesting like uh it's trying to like it's it's a platform that you can group multiple telegram groups and uh use AI to summarize the data and then like you can get a summary of all these kind of

88. 44:07

    groups and aggregate them to out output to some somewhere else. Okay, I will give you a very concrete example of uh what it can do. Uh for example like uh you can have a telegram account and then like with this telegram account you have 20 groups and you can uh basically say these 20 groups could be like web3 k groups and you say I want to get some information from this web3 kl groups for example like what kind of uh projects are they looking into and uh what kind

89. 44:34

    of projects like they're promoting here uh then like it's easy with purify you can just like import the account with trending groups and then like you can write some prompts and will automatically to generate this kind of information for you and you can dump the information into another group or you can dump the information into some other channels like Twitter whatever thing.

90. 44:55

    Yeah. So it's a perfect tool for you to create content from multiple data sources for yourself and to promote your own branding. And uh uh here like I also bring the demo. I would like I'm not sure if Cameron's still here or not like I'm always here man. I'm just Yeah, feel free to pull up your screen and I'll bring it on stage. Yeah. Yeah.

91. 45:16

    So, I'm going to show a demo basically to guide you through because like it's still under development and we're going to launch that next week. Uh Oh, yeah. Besides that, so now what we're launching is nonte version which means you have to give us your session and we can see your messages. That's why I'm so excited about like working with Near together. I think NAR is working hard on this kind of TE environment to hosting AI agents and I think it's perfect

92. 45:39

    candidate for us to to to host my application on because we want to unlock the potential of personal accounts basically anyone they can host their Telegram account in our network without hesitation because there's no privacy integr bring it on stage and I'm happy to make it happen. Just as a quick time check, we have about uh two minutes or three minutes. Yeah, that's that should be sufficient. Yeah.

93. 46:09

    Can you see the screen now? Okay. Super simple. Uh we're still improving the UI, but like I will show you like how the general workflow works. So you log in with Google and use email whatever thing. Once you logged in, you will see like three things here. Accounts where you can import your Telegram accounts.

94. 46:24

    You can import multiple Telegram accounts. And once you input a term account, you will like synchronize all the group information into the the application basically. For example, the the the account may have like training groups and then the training groups will show here and based on this kind of groups you can create a collection basically for example like I have like three groups with uh near and I can group these three groups into a

95. 46:47

    collection and based on collection you can create a data workflow. Say I create data workflow that I I want to build a FAQ for example for NEO's dev community then I can put all these kind of neo groups here and I create create a data workflow says hey use AI to help me to generate a dynamic FAQ from these groups and it's auto updated every 10 minutes that's what I can do for for example

96. 47:10

    here I can import an account I will use one of my account it's uh it's in my phone actually uh let me see let me import my account maybe you don't want to throw your actual phone number on the web. It's actually it's a test account, so it's fine.

97. 47:26

    Yeah. Uh, okay. So, Oh, actually, I don't know like it's what's what country it is. It's a test account we we got from somewhere. And then you can get code. I'm not sure if it's it's still China. I don't know what's the country code 457 Colombia. Okay, it's Colombia.

98. 48:04

    See Colombia. Yeah. And it's like this. Then I can get code. Uh then I will send code to my to my to my telegram. I see the code which is and then confirm you will see I need a password right I type in password and then logged in this is basically the account importing you just like imported account like login at telegram and all this kind of login password application happen front end so you don't have to

99. 48:35

    worry that we know the password and we're also working with some other teams like to host the front host our front end into decentralized environment completely so you don't have to worry that you know we update the code so so we can know your password all this kind Okay, it's completely safe. But we will keep a session, it's a session file in our database and uh uh we use a session file to synchronize the data for your groups. So after we import the account

100. 49:00

     like we have a chron job in back end to run every five minutes to pick up the groups. Uh we can call sync now and it will synchronize all the groups here. Okay. And we add all these kind of groups. We had all these kind of near dev communities here and I'm actually working uh hard to generate FAQ for dev community cameras. Yeah, that we would love that. I get too many DMs so I would

101. 49:24

     appreciate that. Yeah. Yeah. We use AI to to to extract all the question answers from the groups and we we summarize that with another AI and output that in in in the website. So you can use that like you know out of box. Very cool. Man, this is awesome. Uh for the sake of time we do have to move on.

102. 49:39

     Um, is there anything else that you want to quickly share with everyone watching? Yeah, I will like share the tutorial later with our uh Twitter account so I can try it myself and uh but like now it takes time so I can like know go through the whole process with you. Okay, I'm happy to retweet it. So send it by. No problem. Awesome. Sh. Well, thank you so much for coming and um see you soon.

103. 49:58

     Yeah. Bye. All right, we have Oh god, Dylan, where'd you go, buddy? You were just here. Um, Eric, uh, any chance? Dylan's back. Um, Dylan, let's make it happen. Welcome. Welcome. Hey guys, so sorry, uh, for making this, uh, you know, I was facing some technical issues and sorry for the waiting and you again, Cameron, you made it. You made it. You made it.

104. 50:24

     So, um, quickly who you are, a little bit of back background. This is Pond. What did you ship this week? Sure. Yeah. So, I'm Dylan. I'm one of the co-founders of Pond. What we are doing is we are providing other infrastructure and also we are empowering the entire developer community to solve crypto problems through auction data by training models on data. So that's uh what we do and uh our project and as displayed next to my name.

105. 50:56

     Awesome. Cool. And uh what you ship this week you have about let's see I think it's like about five minutes. Sure. Yeah. So uh we definitely have shipped a lot of things and uh screen uh so this is our website. So what we've been doing these days is on one hand we are working with we are working with Vitalic on developing different models for public goods also on another end we are working

106. 51:19

     with the cater projects in the space by uh providing data sets uh from different on data structured data sets provided by us and also the machine learning um guidelines for them to develop different models and train our on data.

107. 51:34

     So for now we have attracted around like 5,000 developers. So each competition they have around like few thousands to 100 sorry to 10,000 developers competing with each other. So after the competition we are going to integrate the top performing models with our real-time data and inference infrastructure to allow them to be virtualized and uh which are you know displayed and posted here. So each of

108. 51:58

     the models represent uh one of the you know competition result from different developers. After it has been uh you know integrated with our real-time data stream and our infrastructure we going to leverage uh you know MI's agent stack and other you know agent network that we can use by integrating those model outputs to those agents and users will be able to customize different agents powered by those models to develop some different recommended use cases. for

109. 52:27

     example uh you know well this is a early prototype that we about to ship it out in a few weeks. So uh you have a some sort of like a training co-pilot as an extension you know this is a extension with a few agents backed by a few models in different use cases. So on one hand you can you know scroll left and right and you'll be able you will be able to see you know behind the scene of each

110. 52:51

     price change on this candlestick of this K chart of this token. What's the reason for this token to pump 5x 5% or what's the reason for this token to g 20%. Because for every price change there's a reason is it because developer related what is Australian or is it because you know general market behavior you know people are pumping you know this token or what so you you're going to have this sort of like a trading co-pilot powered by different

111. 53:16

     models developed by the community and also uh they're going to you know inject it into a different stere and also you will be able to see you know where's the potential price prediction with specific time interval back by these models like for Now you know on on the trading side we have you know 12 hour volatility prediction uh 6 hours 3 hours and all developed by different community members and also we

112. 53:40

     have some other you know other um models like what a risk and the best part is all those models can work with each other to orchestrate with each other because end of the day all the crypto use cases no matter going to be security trading social etc they're all going to be on this same uh database which is the onchain data.

113. 54:02

     So for each models, users will be able to customize and combine them to different agent workflow and to be injected into this uh extension and um people can essentially use it. So for now for this demo purpose we're having to create a two agents why what one is price reasoning like what's the reason behind those price change of different tokens and second is uh it's a price prediction uh which is backed by multiple models from you know risk uh uh volatility prediction risk management

114. 54:33

     models and uh relation analyst models and offer detection models etc. So uh you are seeing aggregated results from different models uh from different proing such a result. So that's on a high level what we are doing. So from the the data side we've been structuring data and on the model side is we are working with the community to figure out what is the the best machine learning guidelines so we can ship out those

115. 54:58

     models through competitions by providing them required data sets and people can compete with each other. And after the competition, we're going to integrate some of the best models on this model labs for people to use and they will be used to power different agents and different agents can combine with each other uh created by different you know general users uh to form some sort of a

116. 55:20

     workflow and those use cases will power different features through this extension. Man just demoing this. You have shipped so much since the last time we spoke. Holy crap. Um yeah was really exciting. What can people do to kind of get started today? Sure. Yeah, we'll talk later. And also we want really want to leverage N's agent stack.

117. 55:42

     We feel like you know is really pushing this entire yeah space forward. So we are very glad and part of awesome man. What can people do right now with pond? Like so we are going to have some uh summer announcement coming out pretty soon. So uh I think the best way people can do with us at the moment is to follow our Twitter. I also have put down our uh Twitter handle next to my name is

118. 56:08

     Pong Gn. Feel free to follow that keep track of track of where we are. Awesome man. Well, thank you for coming. Super stoked to have you on in the future obviously. Yeah. Thank you guys. Awesome. Thanks for joining. Last but not least, Eric, thank you for your patience. Um, bringing you up on stage talking about what Near AI has been doing this week.

119. 56:33

     And a lot of the team is in Denver, but a lot of the team is at home shipping. Yeah. Uh, haven't been at Denver, but I've been having a conference-like lack of sleep schedule. Um, near AI team's been working real hard, uh, getting a lot of things out. Uh, we've shipped a ton this week. We've been talking about a lot of it on stage at near at Denver and at side events. We'll be posting a lot to our Twitter, but I'll just dive into one thing that we showed off briefly on Tuesday, but it's really cool. So, I'm going to show it off

120. 57:02

     again. Um, let me just get my screen up here. Okay, you're looking at the near um interface here. And as you may know on Near AI, you can write and host agents, um, AI agents that can do things for you. Um, we've got our own agent SDK.

121. 57:26

     We've got free inference. We've got, um, it's it's a platform to run them. Uh, this one is called Sound Sage. It's really cool. Uh, it is a this is an agent that, uh, can make headphone recommendations for you and can go ahead and purchase them for you off Amazon. So this agent has a server server component that can go buy things off Amazon. It communicates with the agent over MCP model context protocol which is something that Anthropic has built and is really gaining a lot of traction. Uh but I'm not going to talk directly to

122. 57:57

     this agent. Uh instead I'm going to go over to my next tab. I'm going to talk to the Near AI assistant agent. So, what we're going to do live and we'll see if it works is I'm going to talk to this assistant agent and the assistant agent is going to go try to find some other agent that can go buy me some audio equipment. It'll find Sound Sage. It'll go talk to Sound Sage, tell it that I could use some audio equipment. Sound stage will go and say, "Okay, here's

123. 58:21

     what I've got. I'm going to go tell the Near AI assistant that I want to buy it." And then we're going to actually literally buy something online right now because um I could use some cheap but waterproof earbuds for running and we'll see what happens here.

124. 58:45

     Um, so what should be happening here is the neari assistant will say uh we'll will start executing. I'll say I don't know how to do that myself. It'll talk to another agent called the discovery agent and the discovery agent is aware of all of the agents on the new AI platform. tries to find something that can go solve this problem for us. Um, it should be now searching through its database and finding, hey, sound stage might be able to solve this problem for

125. 59:09

     you. Um, and here we go. So now we have sent a message. Uh, we've now gone through that whole process. We found Sound Sage. We went and contacted Sound Stage and said, "Hey, I need these cheap waterproof earbuds for running." And Sound Stage responded with a product catalog. Um, now the really cool thing about this is, uh, you see it's a nice UI. It's not just text. So, there's a

126. 59:34

     bit of a negotiation process happening between this agent and the next agent down the line to say, "Hey, I support special capabilities. Um, I can show you a special UI for choosing a product that's going to be better than just text." Uh and this is all defined in something also we came out with on Tuesday, our AITP agent interaction and transaction protocol. So this is essentially a spec for how agents can communicate with each other and communicate with users for structured data like um showing a product list,

127. 1:00:07

     asking a yes or no question, um submitting a payment, uh doing wallet operations. Uh so as long as all the AI agents speak this protocol with each other, you can of course communicate over text, but you can also communicate say a product grid. So uh let's see what we got here. Uh D. These ones are cheap.

128. 1:00:26

     Um I'm not trying to spend a lot of money on a demo here. I'm going to click buy now. Okay. So that sends again a structured message back to the assistant. The assistant goes and talks to Sound Sage and Sound Sage is going to go talk to Amazon. And uh we're going to see what happens next.

129. 1:00:47

     Okay, the message came back that we'll need our shipping info. Now, what should happen in the future here is my assistant could know my shipping info and it should just be able to fill it out, send it right back to the uh to the sounds agent without without having to involve me. But right now, this assistant doesn't know my shipping info.

130. 1:01:03

     So, we're going to fill it out. I'm going to stop sharing for a second because I'm actually going to buy these headphones. I just took you off stage. Cool. Yeah, I feel like the Near AI team has too many headphones now. Is there like a headphones charity we can donate to or something?

131. 1:01:32

     Fantastic. Okay. Um, I've submitted it. It's now I'm going to let it sort of fall off the top of the screen uh and the thread. There we go. Okay, perfect. Okay, I'm going to share again. Did I get the right screen there? Okay, cool.

132. 1:02:06

     You guys can see my zip code. I'm a proud Jersey City resident. Now, you know where I live, roughly speaking. Um, okay. So, now I got another structured message back again from Sound Stage through the assistant back to me saying, "Well, here's my payment request.

133. 1:02:21

     $9.32 uh payable to Sound Stage. I'm going to connect my wallet." Um, I'm going to pop open Meteor here. Uh, it's in a pop-up window. You can't see it. Um, which is for the best. Uh, but over in Meteor, I'm now I've now connected my Meteor wallet. So now I've got my payment method here. And in the future, you'll be able to to add multiple payment methods. Um, whether those are

134. 1:02:46

     near wallets, EVM wallets, maybe credit cards. Um, because we sort of treat payment as being separate from being able to log in and use your application. So, if I click pay now, um, now I've got, and again, you can't see this, but, uh, I'm looking at a Meteor, uh, approval dialogue asking me to send, uh, $9.32 of USDC from my Near account to, um, to Sound Sage. In fact, maybe I can go ahead and share that over just you can get the full experience.

135. 1:03:22

     All right. So, here's my transaction approval. Uh, $9.32. Um, heading over to soundstage.n. I'm going to click go ahead and click approve. I'm going to switch right back to sharing the other screen. Okay, we're back here. We've sent this. And so now this is a message essentially from me saying here's my transaction hash. This is the assistant sending that on to sound stage and then sound stage is going to go place that order for me

136. 1:03:57

     on Amazon and um couple days I should get some headphones. Uh we're also going to send you an email receipt. And so this is live like go to chat.near.ai and you can go buy something with USDC on Amazon. Works today. Works really great.

137. 1:04:13

     Check it out. Um, but I consider this to be really a a the an um not necessarily a demo, but an example of this agent interaction protocol. So, what we were doing behind the scenes here, zooming in on my on my diagram, is like user talks to the client to the assistant.

138. 1:04:31

     Assistant talks to what we're calling a service agent, which goes and talks to a service. And I think that's how the internet's going to work in um in a few years. So, we're just showing that off right now. We're going to be building out that network so that your assistant can talk to hundreds, thousands of different service agents to go do what they need to do in addition to having all the context about how you want it done. All right, that's one of the things we've done this week. Yeah,

139. 1:04:54

     there's so many. And what I'm so excited about is, you know, crypto is now usable for everyday goods. Like you can literally just go on Amazon with your wallet, not have to get a separate debit account or debit card or open up an account with an exchange that get forces you to get this like non-custodial wallet. You can spend your money on Amazon. Absolutely.

140. 1:05:18

     This is great. Uh, I also want to point out like, you know, it's I I think there's a great opportunity to remake what internet commerce looks like because every every link in that chain that help facilitate that purchase should get a cut of that purchase. The whole system of like uh referral and affiliate marketing and is is kind of broken right now on web 2 and and this is a great opportunity to solve it. So, we're doing that with AITP. We're doing that with the near AIHub. Come check it out. Cool. Um would I sort of envision

141. 1:05:50

     like going back to even like Robert's first you know demo of Eliza like if we can get all these pieces added to every single agent framework everything is hosted on your AI totally modular you can go use this across any other app it doesn't we were paying for inference so it's free and anyone who actually is like built agent companies knows that's like their number one cost so we get all

142. 1:06:13

     these agent developers it's callable through the assistant this assistant Or can we we talk about white labeling or is that not yet? No, sure. I mean, no secrets here. Okay, great. Um the um if we have this assistant that's white labelled, not every company in the world is going to want to, you know, manage their own agent network and manage their own AI assistant. So, they're able to like fork ours. It calls all the legitimate agents that are added to the

143. 1:06:40

     hub. So, we need more agents that go buy relevant things. But in order to do that, we need better benchmarks. And something that I've been working on in the last couple weeks is an auditor agent, an agent that audits other agents in order to ensure that that agent that the assistant is calling is not leaking data is is, you know, complying with the compliance standards or anything else.

144. 1:07:00

     And so you can kind of see what's happening here. Agent cloud get all the quality agents assistant the go to market and distribution for the the eight quality agent builders from the cloud. And then lastly is payments. So this is the new internet. It's super exciting. Eric, thank you for joining.

145. 1:07:20

     Is there anything else that you want to add to before we hop off? No, but I just want to say you're doing a great job hosting this live stream. It's uh it's been smooth as silk and uh let's see you next week. Yep. Except for the first one. The first one was not smooth as s.

146. 1:07:34

     But you live and learn and I appreciate you for coming on every week. It's been great. All right. See you. Awesome everybody. So, um, that's about it. Uh, we do have, I do want to remind we have a hackathon coming up. Uh, there's so many events happening in ETH Denver right now. If you are in ETH Denver, uh, the last event that we will be sort of affiliated with is called ETH Rising.

147. 1:07:55

     That is tomorrow around 10:00 a.m. We're doing it with Electric Capital and REI. Uh, so that's incredibly exciting. Oh, we have a paper club in San Francisco uh next Tuesday, I believe, where we will be unpacking S3. We did skip the Oh, sorry, S1. We did skip the last weekend news uh this week. So, next week is going to be a very jam-packed uh live stream of AI news. And that's about it, guys. Um, tune in next week. We are

148. 1:08:26

     going to be doing this again every Thursday at 11 a.m. Pacific, about 2:00 p.m. Eastern, and more teams to come. So, thank you all for joining and see you next time. Thanks for watching. Subscribe to Near Week. Jim Jim, thank