NEAR AI Ecosystem - What Did You Ship This Week 5 - AI Agents That Apply to Jobs for You

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

1. 0:16

   Near AI, what did you ship this week? Hello everybody. Welcome to the fifth episode of What Did You Ship This Week? My name is Cameron Dennis. I will be mainly just facilitating introductions and uh giving a quick update on what's going on in near AI land uh at the end of the call. But to kick it off, uh we do this about every week just to have people show off what they're building in the ecosystem. Kind of treat it like a

2. 0:44

   decentralized uh community standup. Um most the time people display demos of what they've actually built. Um sometimes it's just general updates from a lot of the teams working in the ecosystem. And today we have an amazing set of builders uh joining us. Some of them are have been working on new implementations for quite a while but not quite ready yet with something like X-Ray for encrypted memory and private rag. Uh Silverstream talking about

3. 1:09

   browser agents and uh we actually have a guest here that is uh won the hackathon or was a winner of the hackathon that we ran over the weekend from instant apply. Looking forward to hearing from uh Jivvon. Next, we have Matt Locker from Proximity Labs to unpack some progress being made on the shade agent side of things and hot wallet service, which I'm really excited to hear about. And then we'll move on to Asimov Protocol where

4. 1:35

   Matt Hamilton, who's a new member uh on the team, will be unpacking uh kind of what they've been focused on with their CLI tools, their SDKs, and uh some of I guess recent experience uh being in San Francisco fundraising. And close to last but definitely not least is follow network uh giving a quick update on the uh red pill and follow cloud integration their cloud CLI and a couple other

5. 1:59

   things. And then uh Jutsu Zahedan from Jutsu will be talking about the near typescript SDK and then maybe a little bit about the hackathon experience because he was also at the hackathon. he won the hackathon actually first place uh for building a AI agent that manages your LinkedIn uh campaigns and then at the end I will be giving a quick lowdown on what's happening in the near AI uh

6. 2:21

   company. So without further ado uh would love to bring on our first guest Leewin uh bringing you up to the stage and just FYI if you're going to share your screen make sure you're sharing the specific tab on your screen uh not your whole screen. So, yeah. Yeah. Um, but I I need to kind of switch back and forth from a a a text editor and my browser. That's totally fine. Just make sure you're not sharing anything you don't want to

7. 2:53

   accidentally share. So, that is that is the only request, but um Okay. Uh just real quick, it would be great if you can quickly introduce yourself, give a little bit of background on X-Trace, and then unpack what you've been shipping. Yeah. Uh so hello everyone. Uh my name is Leewen. Uh I'm the co-founder of X-rays. Uh my responsibility is mostly on building stuff and yeah so X-rays is basically um building a uh private memory uh which is

8. 3:24

   in for AI agents which is enabled by our uh encrypted vector database. So I guess without further ado, I'll showcase uh our product which is not open yet but is pretty functional I would say. Yeah, let me start sharing my screen.

9. 3:53

   Okay, can you guys see my screen right now? Yeah. Yes, we can. Yep. Okay. Okay. So basically we have this um so how you would use X-rays is very much similar to how you would use any vector database. So you will have a user uh port uh developer portal. Once you log in you can see all uh the you know knowledge bases and API keys uh on your portal. So so in order to use X-rays X SDK or vector database you first need to create

10. 4:24

    a API key and give it a name. Let's say demo whatever and then yeah you need to create a API key and copy it and save it somewhere so that you can use it later with the X-ray SDK and then in addition to API key you also need to create a knowledge base so let's say near AI demo knowledge base create right and you also would need the knowledge base ID in order to specify which knowledge base you want to um post your data to, right? Um so once you have

11. 5:06

    these two things uh yeah you probably also need to um manage so given API key you can manage the permissions of this API key to m multiple knowledge bases. So right now I'm going to give this API key the right access the read and write access uh to uh to the knowledge bases I just created. So right so these are all the setup you need before you can use our SDK. So now let's dive into our SDK. So basically our SDK

12. 5:39

    has like um two major components. One is data loader and one is uh retriever. So the data loader basically uh kind of takes in bunch of crypto primitives and a storage and compute unit uh which kind of decides where to post your data to and then you know it runs uh the compute script on your data locally and then post the encrypted data to whatever storage you specify as you can see here um from my screen uh in the Jupyter

13. 6:14

    notebook. So I'm going to run this cell, right? So let's wait. Oh, in the meantime, uh I'm using uh a news article I scraped, which is this one, as the contact as the document that is going to be posted to X-ray server. So yeah, the data loader is now running. Let's wait.

14. 6:39

    I think it's going to Yeah. Okay. So now uh it's done running. So now the data should be successfully processed and loaded onto X-ray server. So right now I want to show you how the data actually looks like on our X-ray server. Right?

15. 6:57

    Just to convince you that they are all ciphers, right? So these are the data that is loaded onto X-ray server. As you can see the chunk, the index, they are just you know a bunch of ciphers. So which means actress doesn't know anything about your data. Right? Okay.

16. 7:16

    Uh switching back I'm going to talk about how you act once your data is successfully loaded onto X-ray server how you are going to use it. So to use the data stored on X-rays is as simple as you know import a retriever from our SDK. So right now we only support a simple simple retriever which basically does near uh nearest neighbor search on your ciphers but in the future we're planning to roll out more advanced retrievers such as you know uh the ones that can do metadata search can do like

17. 7:49

    complex queries like range queries and all that. Yeah. So that's the the plan. So now let me run this cell. Once you have the retriever and the query, you can just retrieve the you know necessary data. So yeah, so now you can see it retrie ask why is elamas partnering uh partnering with x-rays sorry Microsoft uh the retriever successfully retrieves a bunch of you know context that is

18. 8:20

    semantically similar to the query. So basically it retrieves the context for the agent so the agent can use all this text to answer this question. Now switching back to the server side, I want to show you um yeah I want to show you the log so that you know um the queries are really encrypted. Right? So this is the log of that query and if if you check um the API key here matches the API key I specified here. So that

19. 8:54

    means the the the demo is actually a demo not a stage thing. Okay. Um so let's go to the body. As you can see the query is this. So it's just a cipher right. So that basically means X-rays when you query X-rays database X-rays doesn't know your query either. So X-rays doesn't know your query. X-rays doesn't know your data. But still X-rays can return the necessary context to your agents. Right? So once you have the uh context uh you can just you know fit it

20. 9:26

    to whatever inference engine you want to use. Right? So for example, X-rays provides you know a a simple implementation of deepseek. So if you run this it is going to take the context return here f it to deepsek and deepse will use this context to answer the question why is it part partnering uh with Microsoft and yeah so this is going to take a while because I don't have GPU on my machine but I think you get a you get the idea right and uh yeah I think

21. 9:58

    that's it this is amazing uh and just to be crystal clear the primary use case of this is for people to better train their AI agents with their personal data so they can be as helpful as possible for you. Is that correct? Yes. Yes. Um I wouldn't I wouldn't say it's training per se because I'm not X-ray is not modifying the weights of your model.

22. 10:20

    It's more Yeah, it happens uh it's it's rag it happens during inference. But yeah, but the idea is you can use your personal data uh to make your um AI agents better suited for whatever use case you want without sacrificing any uh privacy.

23. 10:38

    Cool. So this is a pretty important foundational prerequisite for AI agents to actually solve problems for people like you need a encrypted data store. Yeah. Yeah. Yeah. So yeah, pretty amazing. So it's incredible, man. Yeah. Yeah. Yeah. I'm working tirelessly on making it better, adding more features.

24. 10:54

    But yeah, happy to be here. This is amazing. Yeah, uh would love to have you back on. This is super exciting. Uh we do have to move on to the next uh builder, but thank you Leewin for coming. What's the best way for people to kind of stay updated on the project and you uh I guess uh I'll post uh my my Telegram handle here so people can just, you know, DM me. Cool. We'll make sure people follow your your Twitter account.

25. 11:22

    Oh, yeah. Right, right, right. Also, I need to Awesome. Cool. Awesome. Well, thank you for coming. Uh, we're gonna move on to Manuel from Silverstream. Manuel, how's it going? Hello. Hey, it's good to have you here. But so, just as a quick reminder, uh, quick overview, what is Silverstream doing? Uh, who are you? Why are you building Silverream?

26. 11:46

    And then what have you shipped recently? Yeah. So, uh, nice to meet you, Emanuel. I'm a scientist. I come from, uh, a strong AI background. And we're building Silverstein to make sure people can have their agents go on the internet and interact with the web. As X-rays shows us, people are going to share their personal information with agents and so we want these agents to do something with that information, which mostly is going to the internet by things or do

27. 12:15

    something in the name of the users. So at Silverstone, we focus on getting things reliable. You want an agent that does something every time, but also general. That's why we focus on giving tools to agent developers like you all um to use web web based interactions, right? What we've been we've been focusing Oh, hey, just real quick, sorry. Hi. Um so this just to be kind of give some like an analogy. Is this most

28. 12:42

    similar to like an anthropic computer use or an open AI operator? Yes, cool. Like the agent goes on a website, he clicks around for you. And let me let me skip ahead and share a demo um of one of our agents on the near platform. Let me share only the things I need to share. There we go. Can you see it? Camera, is it showing up? Yeah.

29. 13:09

    Yeah, it shows up now. Nice. So, this is our agent on the near platform. You can see on the bottom here um there is a request. This request can be directly from the user or it could be from another another agent. And this video is sped up for X because we don't want to be waiting for too long. But the agent is going to stream back a sequence of interactions with the web to achieve some task. In

30. 13:36

    this case booking an Airbnb. So if you call our agent as a tool, you will get a stream of in of images or video. Um we're still working on integration with near on that and the agent goes and does something for you. Then it will eventually get to to the checkout page.

31. 13:55

    We don't do payments yet. I hope we can present that maybe next month in one of these conversations. But this is the direction in which you're building. And the focus of right now the focus is to get things reliably working every time.

32. 14:11

    So in this case for example you see the agent at a translation window pop up which is what was unexpected and it will get out of it by itself. Now we want this to happen 100% of the times. Right? Let me skip forward. There you go. And then it goes to the booking. So what we are going to release this week is a data set for people to fine-tune and train their own agents mostly because coming from a science point of view the most important thing is data like you want to be able to fine-tune your own agent from a good

33. 14:45

    enough starting agent and for that you need data that's why we decided to open our data sets to people that need it in fact it's not public yet but if you go to silver streamai you can sign up for the waiting list which is going to we unlock very quickly uh to get access to our data sets. So you can fine-tune your own agents. And the whole point of user own agents is that you need to be able to see how the agent works but also to

34. 15:11

    create your own. Maybe start from an initial open weight model but then fine-tune it in an agent that you care about from some specific interactions. Right? With the data set, we come also some kind of uh collabs like training kits so people can learn how to train their own agents. But our focus is really to get the most reliable agents, deploy them on the near platform so everybody else can

35. 15:36

    use it. Do you That's basic. Yeah, super cool. So just to be crystal clear, uh people should be going to silverstream.ai to sign up for the demo. Yeah, we we'll throw the link somewhere here. There are two waiting lists. One is for the data which is the one that is going to be released soon. The other one is for to be a user of our agent which is going to come later on in time but you can sign up so you get the early

36. 16:05

    early spots. Cool. Awesome man. Well, this is super exciting. Um where can people learn more about Silver Stream generally besides the website? It will be on the website. We have a Twitter account. It's a kind of stale now. Like it doesn't say much. And also there is my own Twitter account which I guess is you can Google my name. Uh can I say can people read the chat? Uh no but we can follow up in a thread after. Yeah. Cool.

37. 16:33

    Awesome. Cool. Check out Manuel's Twitter account. Sign up for the beta. Uh this is all super exciting. Looking forward to having you on to talk about payments sometime. Like let's make these us these agents, you know, useful for the real world.

38. 16:46

    So, thank you for coming on. Well, um, next up we actually have one of the hackathon winners, uh, who is building with Silverstream. And so, uh, Jivvon, welcome to the show and would love to kind of get you to one demo what you're building or just talk about it, but first introduce yourself, uh, what you normally work on and then, uh, what you shipped last week. Uh, hello. Uh, I'm and actually real quick, I see Juliana

39. 17:12

    is also on the in in the chat, so I'm gonna bring her on stage as well. Thank you. Yeah, take it away. Yeah. Hello. Uh I'm Jven uh and me and Juliana we built instant apply uh last week in the hackathon. Uh both of us are CS majors at Minute by University. We are juniors right now. Uh normally I just build web apps addins or try to play with some ML algorithms and you introduce. Yeah, I

40. 17:41

    also normally work more in science. I currently work as a research assistant at Harvard Psychology Lab. So that was an unusual project to make, but we're both very excited about it. And the idea about the project is uh to help people apply to jobs. So the audience is basically everyone who needs to apply for jobs. Uh the way it works is basically it collects the information about you uh from your resume and from the separate fields that you fill out

41. 18:11

    like volunt uh veteran status that you don't normally put on the resume but that are often asked on the job applications and it fills out the job applications fully for you. So the from start to finish, it finds the job applications for you and Chan will talk and so the search in a bit. Um finds the links to job pages that fit your resume and description of what you want to find

42. 18:42

    like the job like the company size um city you want to work from etc. It goes to the link. It navigates any popup windows, multiple pages in the application, clicks through them in a way similar to what Manuel was just presenting was Silver Stream.

43. 19:03

    Um, so but specifically for job applications. Um, yeah. Do you want do you want to show the uh Yes. So just for the sake of time I already fil uh my profile has basic information that they would ask in a job application uh your summary uh regime uh like job titles portfolio links all of the details and demographic information which is hidden uh for now and the first thing that you do is what type of job you want. uh say I want uh software

44. 19:39

    engineering intern and I want it to be in San Francisco remote uh I would search for jobs uh and right now we're using API uh but next week our plan is to use uh silver stream uh to scrape and find all of the job applications that are relevant takes some time uh from the API And the way it works on the back end is it saves all the found uh jobs to a database from which it then pulls to

45. 20:11

    apply for them. But you can also check out the jobs manually. Um I might have a little trouble sharing my screen. Um let me see. And I can show you how we actually apply to the jobs. Um, have you guys used the tool to apply to jobs yet? Like yourselves?

46. 20:39

    Not just yet. It's a little bit rough yet. And there is a lot of for many job application websites. There are many different structures for how it works. So, it's difficult to optimize for every website. But it's working quite well for uh single page sites right now. Uh and also like sometimes they ask for capta and you have to login and we just don't want to like randomly like put their password for user and save it in the database. Uh but if it does not have a signup page it

47. 21:12

    works pretty well so far. Got it. Soon with with X-trace you will be able to put credentials in and ensure their privacy. So yeah, you can see all these things are coming together or uh I am Am I sharing my screen? Can you see it? Yeah, I ignore the rough front end. That is not what we're all about. Uh so I like Oh, no. Wait, can you actually see it? Cannot. Can you? No, I only see the apply to jobs page.

48. 21:44

    Um sorry about that. It's okay. I'll just go ahead and do what you just told us not to do. I was Yeah. Okay, that sounds great. So, I go back to the apply page. I like the apply to jobs and it starts the whole process all by itself and it fills it. It's not from my resume. I'm not an but um yeah it can select stuff from the data folders. It's not currently filled cuz it was a fake resume but it does work and it will do

49. 22:29

    the same thing with a resume. We were a little bit short on time because we just built it on the hagodon literally, but we're very excited about the idea and I think for now we have a solid proof of concept kind of situation that is actually working and can interact with different websites. So that's where we're at now and we plan to build on it.

50. 22:54

    That's super exciting. What next? Like what are you doing to improve it? So one of the things uh yesterday majority of what I worked was uh working on the near aentto if seeing if we can actually have the source functionality stuff there. We tried a bit uh but I guess the next goal is uh to use silver stream and like some other like existing aents we have uh to first like safely store the user password so you can login uh like go through capture and also work

51. 23:22

    on multi-page uh like website like workday I think is really hard to work with. Uh so over the next week we have a break so we are planning to utilize that week to like make it a proper MVP. I love it. Using your spring break to ship. Uh love to hear it. Cool. Uh Lyanna Jivvon, is there anything that you guys want to add? Any place that people should follow you? Uh should they reach out to you on LinkedIn, Twitter?

52. 23:47

    Uh anything else you want to say before we hop off to Matt? Yeah, I think people can follow us on LinkedIn definitely. Uh, we'll put links in the chat. That works. Perfect. Yeah. Awesome. Anything on your end? Same. I'll share my Telegram and LinkedIn in the chat as well. Thank you.

53. 24:09

    Beautiful. Awesome, guys. Cool. Well, thank you for joining. I hope to see you on in the future to, you know, highlight exactly all the progress we've made. And, uh, without further ado, want to bring Matt Locker to the stage. Matt has been an OG nearbuilder for a very long time and now I think he is shipping one of the most impressive and I guess inspiring uh pieces of technology in

54. 24:32

    crypto right now uh which are shade agents and so Matt would love for you to kind of further explain what these shade agents are and uh what you've been working on. Well, thanks Cameron for the the glowing intro there. Um, yeah. So, I'm I'm working on shade agents, and I'll give a quick refresher on that in just a bit. Um, and then I'm going to talk about some ideas we have for shade

55. 24:55

    agents, uh, that are coming up, uh, really shortly, and we'd love to get the Near community and other builders involved. Um, so I'm going to share screen window. Um, and basically just a quick overview if you haven't seen shade agents yet.

56. 25:14

    Um, we basically have created a a framework on near that allows you to deploy arbitrary worker agent code in a trusted execution environment. So this code can talk to LLMs which can also be in the T. Um, so they're they're running kind of in a confidential compute situation or I mean the worker agent can talk to external APIs. it could be talking to external language models and

57. 25:41

    stuff as well. The real like core part of the sheet agent stack is that the worker agent will derive an ephemeral near account um so that it can basically talk to a a corresponding near smart contract down here and it will get the remote attestation from the trusted execution environment including the code hash of the actual code that it's running. Um now the reason that's important is that once we register this

58. 26:11

    agent in the smart contract this uh arbitrary worker code can just go ahead and call uh methods in the smart contract and uh those methods are protected by virtue of this worker agent running that particular codebase. So uh I can't come up with a better way to describe it other than there is a very tight coupling between the code that is run in the trusted execution environment

59. 26:37

    for the worker agent and the smart contract. So we have this real uh sort of tight connection there that we think is pretty verifiable and and pretty like sort of secure and um that would look like that that sort of like manifests itself in some cool design patterns where you can uh do method access control in a protocol based on the worker agent having a particular code hash. Uh you can have multiple worker agents working on different parts of the smart contract.

60. 27:08

    Okay, so that's just a little overview on shade agents. Um, and I just want to jump back and basically say there's a couple of different things we've been doing with this. So, one for ETH Denver, we launched a shade agent that uh basically would listen for tweets on uh on X um zets or whatever whatever they're calling them these days. I'm I'm old school. I'm just going to call it

61. 27:31

    Twitter. But basically, um, it would listen for tweets and it would, uh, it would look for people mentioning shade agents and whether they were sort of throwing shade on AI slop that's out there or whether they were mentioning shade agents in a favorable tone, they would get rewarded with uh, tokens that were minted on hyperlquid. So, they would get either base tokens if it's positive sentiment or they would get shade tokens if it's negative sentiment.

62. 28:00

    Um now a lot of people in the near community were like why why are you minting these tokens on hyperlquid? We're showing off chain signatures as well. So you know we can do this agent in a T that has this tight coupling with the near smart contract but then from that near contract we can make calls to the MPC uh the chain signatures protocol and we can basically get signatures for

63. 28:23

    any chain. So, our next iteration of these agents is we're going to be launching this next week is an agent that will register a name for you on base. So, uh, on base they have, you know, base names or or base names and, uh, and basically all you have to do is tweet at this agent and it'll create a one-time deposit address that you can, uh, that you can send the requisite, you know, funds to based on, you know, the

64. 28:52

    price of the name. And once you fund that account, those funds will be put towards registering a name on base. So the agent will just sit there and uh you can tweet, hey shade agent, you know, register my name on base. And um what we're going to do is we're going to have an interaction here with Bankerbot as well. So you can send the funds from your Bankerbot account and we'll

65. 29:17

    register the base name to your Bankerbot address. So it it'll create a little bit of like agent to agent stuff as well. And again, like we're we're trying to show off that this is like a multi-chain solution. um as well. And so the last thing I wanted to talk about is we've got this um uh this is kind of a bit more early early kind of green field idea here, but uh basically we've we've got these agents out there like Bankerbot and and

66. 29:45

    there's a whole bunch of other ones coming out where they all take sort of your ex username or your um you know your Telegram handle or whatever and they create a wallet for you based on your ability to authenticate. through that service. Now, a lot of people are going to create agents that do that. And so, for one Twitter handle like my or X handle, whatever my my name is like Matt Dlock here, I might have if

67. 30:14

    there's 10 different agents, I might have 10 different accounts. Um what we're trying to do because we have this shade agent stack and we can sort of create a secure connection between something running in a T and the uh underlying smart contract that can create these accounts for you. Um what we're going to try to do is we're going to try to get the like some of the near community and and maybe beyond multi-chain community to rally behind kind of a hot wallet service. So, um,

68. 30:47

    one like like one developer could launch multiple agents that do different tasks, but they all have the ability to access the underlying same wallet. So, for my handle, Matt Dlock here, I'll have I'll have one account, but it can be accessed by hundreds of different agents. And what we're trying to avoid here is like everyone launches an agent that talks to a wallet that's derived from your from your X username. There you go. You know, you've got you've got hundreds of agents. Now you have hundreds of

69. 31:21

    wallets. That's going to create a big problem. So, we're trying to kind of propose this hot wallet service uh by using the shade agent stack as sort of the core primitive to create this underlying um like liquidity layer that can be accessed by hundreds or thousands of agents. And I think there's going to be hundreds or thousands of agents that are specified to do different tasks. And then, you know, people might say, hey, what about security this and that?

70. 31:47

    Obviously, there's there's some issues there, but the whole uh hot wallet service would run on the premise that you would have to mention the um the agent's you uh username to basically allow that agent to sort of get access to your underlying wallet and create a transaction call for chain signature and get that signature. So, we can do this.

71. 32:13

    Um, like I've thought through it kind of technically and I think we can do it, but we obviously want some support and we want to know that there's like a market kind of reason to do it. Um, so we're seeing what the community says about it. Um, we have a shade agents telegram group as well. Um, yeah, if you want an invite to that, uh, send me like a DM on twi on Twitter x whatever. Um, anyway, yeah, happy to chat about it. Cameron.

72. 32:40

    Yeah. So, I guess uh let's say someone wants to create a shade agent today. Where should they actually go besides this? Oh, you can start at near.aishade. Um we've got a little notion page up there with all the info about shade agents. Um yeah, and the stack that I showed earlier, that's a great question. The stack I showed earlier, it um it's it's a framework that's ready to go, ready to be built on today. You basically just fork the shade

73. 33:06

    agent template and away you go. And and um something like the hyperlquid uh token minting agent that took me a couple of days after forking the template. And the base names agent, it's only taken me a couple of days since forking the template. So this is very uh fast and easy to use sort of hackathon fodder.

74. 33:31

    Besides some of the agents that you had highlighted, are there anything is there anything that you guys want to see built? Like you guys don't have the internal capacity to build. Um, you know, there's a lot of builders probably watching this. I think there's like over 2,000 right now. Uh, what are you looking for? Absolutely. So, I I haven't I haven't been able to get really really deep into into LLMs and and using them with the shade agent framework. So, I'd

75. 33:56

    love to see somebody connecting to an LLM doing some meaningful stuff. So, uh there's a lot you could do. You could have shade agents doing, you know, image generation or, you know, writing poetry or whatever and and just doing some fun kind of interactions and behaviors around uh around, you know, the LLMs themselves and the generative aspect of them. I've mostly basically been using the uh the sort of the Twitter kind of command prompt and then basically do

76. 34:26

    something. So I I've been sort of doing sort of seek and destroy kind of example use cases, but I'd like to see somebody do something more expressive. Um and also uh um mind share trading agents whatever you know um um yeah any sort of like web markets or whatever creating uh even like peer-to-peer like uh betting kind of things like you could have people kind of talking on X basically

77. 34:53

    saying hey I think you know the price of whatever coin is going to go to this do you want to do the overunder and somebody else joins up deposit into some chain signatures derived accounts and the the agent basically finishes it off. And because you can kind of trust the agent code, the worker agent code, um you know, it can basically sit there, listen, and resolve the market and then

78. 35:18

    basically execute on the chain signatures accounts. Super cool. It sounds like you're using Twitter's distribution channel to get essentially the apps built and distribution is everything in AI and crypto. So, uh yeah, this is super exciting. Um, last but not least, is there anything else that you want to add before we pass it off to Matt from Asimov?

79. 35:40

    No, I think we're good. And I'm just really excited about the whole Near AI community and what everyone's building and and I love that you're doing this Cameron, so thanks for having me. Absolutely. Actually, other thing, does Proximity invest in teams building with shade agents? Uh, I believe we do. Um, yeah. So, we're uh you know, we're we're involved in in sort of a lot of the investments in the near AI space as as

80. 36:04

    well. So, yeah, just uh you know, reach out. I think we have something on our on our page, proximity.dev. Cool. Amazing. Agent builders out there, go build shade agents and try to get investments. Uh but first, you obviously have to build something useful that will ideally make money. So, um yeah, awesome, Matt. Thank you for coming. Uh, we'll love to have you on in the future.

81. 36:27

    Got to pass it off to other Matt. Um, thanks for joining. Great. Other Matt Matt, we got a double header here. So, Oh, yeah. Okay, let me just uh share my screen here. Um, beautiful. Um, and just as a quick quick quick intro, uh, introduce yourself, introduce what ASMO does, and then let's unpack what you've been working on this week. Great. Okay. Uh, you can see slides.

82. 37:01

    Yeah. Hopefully. Yes. Yes. Okay. Cool. Uh, hi folks. Uh, my name's Matt. Uh, it's great to be here. Um, I'm going to give you a quick update on Azimoff Protocol. A little bit about myself. Uh, I am new to the Near ecosystem. uh something I'm genuinely excited about.

83. 37:19

    Um I've recently oh well I've worked in a vast number of uh web3 projects and and crypto projects uh over the years um from uh the XRP ledger and uh and Ripple through to protocol labs at Filecoin um Arbitum Foundation Arbitum um Bit Tensor uh doing decentralized AI and and blockchain stuff and uh now most uh recently Azimov and uh I'm currently here in San Francisco uh we're doing a US fundraising trip. So, this is going

84. 37:50

    to be slightly different uh what did we ship because in this case it's really uh uh more about shipping kind of ideas and and sort of progress towards what we're building. Uh but I thought it'd be good to give us you know give everybody an update as to what we're up to at the moment. So, uh yeah, I joined Azimoff uh shortly after meeting Artto and Joshua who you may know from the near ecosystem already uh at ETH Denver. their vision

85. 38:15

    and energy uh kind of immediately resonated with me and drew me into their mission of decentralizing knowledge and tackling some of AI's most pressing problems. So, Azimoff was specifically focused in in addressing the challenge of highquality trustworthy structured data uh something that we feel is currently underserved uh yet is critical for reliable AI. So our platform leverages decentralized verified uh knowledge in the shape of knowledge graphs uh which not only help uh store

86. 38:48

    and retrieve structured data efficiently but enable powerful reasoning uh capabilities. So this approach allows us to integrate diverse data sets dynamically uh and ensure that information is being used both accurate and verifiable. So at the moment we're uh wrapping up our fundraising round. Uh like I said we're we're in San Francisco. We've been running around San Francisco and and and the Bay Area um uh

87. 39:11

    chatting to many people, going to a lot of events. Um so we're wrapping up our fundraising round at the moment. I'm thrilled to announce that things are going really positively. Uh we have an anchor investor uh Brian Pellegrino from layer zero has come on board which has given us significant momentum to kind of continue on and uh when we get back home to roll up our sleeves and continue building. So, uh, the last three weeks, um, a lot of major events, uh, ETH Denver, Ethereum SF, uh, Stanford

88. 39:37

    Blockchain AI Summit, uh, GDC, uh, we've had numerous discussions that have been incredibly valuable in helping us clearly frame our MVP. Uh, we've identified our beach head use case, which is addressing the critical issue of hallucinations in LLMs, uh, specifically when answering factual questions. Right? Right. So, at the moment, you've probably experienced uh you ask a factual question of an LLM.

89. 40:02

    Um, and the answer it comes back with may uh sound reasonable, may seem feasible, but may be factually incorrect. Uh, the example I often give is, you know, you ask it for a flight route uh from A to B, and it'll come back and say, you know, fly this particular airline to this airport, this aine to to that airport. Uh, sounds all reasonable. And then you check it out and find out that uh actually those airlines don't fly those routes.

90. 40:26

    uh so by ensuring the integrity of data fed into the models through uh like I said decentralized verified knowledge graphs we can significantly reduce uh these hallucinations inaccuracies when dealing with factual information and boost reliability. This is crucial because you know misinformation inaccuracies in AI outputs can have significant real world implications uh you know affecting everything from daily tasks to critical decision-making uh processes. So whilst we've been here,

91. 40:56

    we've opened a whole lot of promising conversations with uh several institutional investors. Uh the enthusiasm we're seeing from these groups is a fantastic validation of our vision. Uh one event I asked people who had, you know, experienced hallucinations. 90% of the audience put their hands up. I imagine the other 10% are probably lying. And it's been fantastic validation of our vision. Um and gives

92. 41:19

    us a lot of confidence in the strategic direction. We've talked to a lot of people from a lot of different industries about how uh verified uh knowledge can can can help them. So on the community and side event side, the last three weeks been really hectic uh but extremely rewarding. We've been traversing all over the the Bay Area um actively participating in a whole load of events uh from specialized hackathons

93. 41:43

    to unique community gatherings. So we intended uh things like Near AI's useful agents hackathon. you've already seen uh on this uh stream some outputs from that. Um we'll chat to people there. Uh we've also joined really diverse events such as a local art exhibition at the former Boston Dynamics building um and uh which blended like robotics and creativity and uh art at the moment is at the opening of a robotics co-working space. So these kind of things are offering us useful insights into

94. 42:13

    different sectors that Azimoff can can help with. Uh, one particularly fascinating uh, visit was to AGI House, which is a four-story multi-million dollar mansion overlooking San Francisco. Imagine the geekiest frat house you've ever seen uh, filled with the greatest minds working on cutting edge artificial intelligence, talking about things like consciousness and and you know really uh, sort of high level concepts and the the energy and collaboration and depth of conversation

95. 42:38

    there was truly inspiring for for what we're doing. So we made significant uh connections uh spread our you know our vision at prominent industry gatherings Ethereum SF prime intellects decentralized AI day shelling point ETH rising uh we've met a bunch of influential figures and innovative thinkers Robert Scoble uh Alex Skidnoff as well from the community Alex Chima um and these have sparked a lot of interesting ideas around collaborations

96. 43:04

    and and what we're doing here um so yeah we so went to Stanford blockchain AI summit And we participated in Domaga's developer camp uh where exchange insights alongside um you know investors and industry leaders like Bill Thai and Amanda Terry and uh most recently we've been at the intersection of AI and web 3 with the GDC conference that's happening here. So we've been going along to GDC uh events uh community meetups like Orbus 86 as well. So uh we love to

97. 43:34

    continue the momentum. Um you can follow us uh both the Azimoff protocol uh main account uh here and this QR code will take you there. It's also azimoff protocol uh on Twitter or X and uh my personal account Hammerto as well um that you can uh find me there and follow me there. Amazing Matt. Well, thank you so much for giving the lowdown for everyone else listening. Uh this is how you do fundraising. If you are an agent builder and want to actually get, you know, money, uh, you have to show up.

98. 44:07

    You have to go to all these events. You have to talk to customers. You really have to actually get your ass out there and we we were trying to make a list of all the different events we've been to. It's been like three events a day we've been going to. Uh, uh, so yeah, it's it's but yeah, it's it's paying off, but it's a balance, right? So like you do this now so you can focus on shipping in the next couple months. to go ahead and also you

99. 44:32

    know hopefully you know connections for future funding rounds as well. Um that's that's that's one of the main things uh you know that we're doing here as well. So uh yeah it's been very good but also just validating the idea right talking to investors talking to so many different industries uh talk health care talking education you know what does uh you know verified knowledge and and large language models mean to them you know what are the problems they're

100. 44:55

     seeing um around using LLMs in uh in you know in in the case of something like healthcare or even legal or like I said education as well so that's been really good and if you can solve hallucinations holy crap that is a really big deal um I is make I'll make this story very short, but I was in Thailand ask an LLM, hey, are there any like dangerous animals around? I'm going to the jungle and it said no. And then I go talk to the

101. 45:18

     person that's working in the front desk of the hostel that I was staying at and they said, oh yeah, right across the street there's a whole family that was trampled by elephants yesterday. And so I'm like, wow. Uh I wish that this LM let me know that I have to watch out for elephants. And so it literally can mean life or death in a lot of situations.

102. 45:37

     That's that's a that's a fantastic use case. I mean, we're not going to be able to solve hallucinations in in general for like everything, but definitely for factual information for questions like that. This is this is perfect. That's that's the kind of thing we're going to be uh we're going to be looking to solve. Amazing, man. Well, thank you for coming on. Uh a little bit different

103. 45:53

     than the average what did you ship this week, but builders, take note. This is how you fundra. Yeah, great. Great. Thanks for having me on. Awesome, Matt. Thanks. and want to bring on Hong from Falla. And just give me one sec while I bring people to the stage. How's it going, man? Hey, great. Great. I feel like I see you, you know, a couple times a week now. It's I'm very lucky. Yeah, but uh just to give everyone a quick

104. 46:18

     lowdown, remind them, uh what is Follow Network? Who are you? What's your role with FA? And what did you ship this week? Okay, sure. Uh yeah, I think this is actually the second time I've been here uh to share what we have uh built last week. Uh so I'm H uh co-founder and the CTO of FA Network. Uh FA Network is uh TE based solution uh uh uh uh blockchain as well as the cloud. So our major product so far is called the Fala cloud. So it's kind of like AWS but as

105. 46:52

     long as you put your program deployed on our cloud we make it fully uh trustless. So how how does trustless mean? Although it's not like a fully like a distributed fully decentralized uh cloud is still hosted on some data centers but at the same time you get two properties. One is uh your program deployed on this cloud is verifiable. So integrity is protected and second you also get the privacy. So all the inputs outputs and the

106. 47:23

     intermediate states is all encrypted and nobody else except the program and the user can see the raw data. So by using the TE there's actually a lot of uh interesting use cases. One of that is has a huge overlap with near AI. So like uh when we develop AI, we always want to make sure the AI is not controlled by any single entity. So uh if you uh just run AI on AWS, then no matter how secure your AI is, it's still limited by the uh

107. 47:56

     security of the infrastructure as a company, you can always as a admin, you can always access the AWS and change whatever that's already deployed uh on the server. So essentially, although it's a AI, but you have no way to prove it's the real AI and it's not going to uh do anything evil, but if you put it on a TE, TE is like a super super secure box. So you can prove that this is uh

108. 48:21

     all controlled by the AI and you can even trace it back to your program, your source code. You can trace it back to the RM you're using and even how you train this RM. So everything can be uh transparent and uh fully auditable and fully secure. So this is the power of te and in uh today's uh event I'm going to show you uh a bunch of uh interesting progress we have made last week. So

109. 48:55

     here all right so um under the hood uh our uh our team do a few uh things. First of all, we do the development on open source projects. So no matter you're using our product or not, you can just feel free to use our open source project including DStack which is the uh backbone of all the TE infrastructure.

110. 49:16

     It's a open source uh uh project. You can deploy it on any capable hardware to turn that into something that can be used easily. Uh on top of that we also have the collaboration with near to ship the private MR SDK. It's another layer built on top of uh DStack. So uh last week a big updates from Dstack side is that uh now we enable the integration with the private MR SDK. So that means uh you can use the GPU uh with uh our uh

111. 49:45

     uh the the features already provided by the other uh underlying open source SDK. So give you some uh uh uh feeling. So now if you use the private MR SDK you're going to do something like this. So uh there's a combined new and give all the parameters and then you run combine the dstack run your CVM then you have the uh uh your your program loaded into the uh GPU but the same time if you check our cloud then we actually give user very

112. 50:17

     user friendly uh web UI. So you don't need to write any code or write any uh command line you just click deploy and then everything is already there. So uh the updates of this integration actually enable uh to uh connect the GPU SDK with the CPU SDK. So that in the CPU SDK we give you all the uh like uh convenience feature including the web UI in this CPU SDK. So without even without using our

113. 50:48

     commercial product, you can also use the exact same uh user interface um on your uh uh customized build or onremise deployment. So this is the first thing. We do also have a few other uh updates but that's a little bit technical deterministic RTMR3. This make verifying uh your uh TE application onchain easily. and another one add TCB status check on the onchain KMS and this can

114. 51:17

     make the verification of your TE program onchain more secure. So yeah, this is updates on the open source SDK side. And second one I want to mention is that uh we last week we also did a really interesting thing. If you're familiar with uh AI agents, one of the very popular AI agent SDK is called Eliza and they just ship V2. And here we actually have a very brief demo how you can run Eliza V2 which supports multiple agent

115. 51:48

     running in the same program. It's like a swarm just not not just a single uh uh single uh um agent and you can already deploy that with our CRI tool uh integrated to your GitHub action and you can make sure it's like a user experience is pretty much like a versell you make the change and push it to your GitHub then everything can be automatically uh pick up by the GitHub action compile uh build and push it to the cloud Everything can be done within

116. 52:20

     just uh around 10 minutes. So uh if you uh are interested in that uh you can uh check our blog post uh on our website so that you can learn how to use that. So this is the second one Eliza V2 on the FA cloud and the third one is actually um when you're using a TE very important uh thing is you want to check how uh you can prove this program is indeed running in the TE and it's secure. In the past

117. 52:52

     we do have a uh proof uh uh web UI where you can upload the proof and see the technical parameters of the proof. But as you as you can see all those fields are just uh the hashes and for nontechnical people they have no idea what it is. All they can do is to check okay it shows verify. It seems like very secure but we want to make this even better for the uh uh average users. So here on the cloud we do have a new

118. 53:25

     version of the testation explore. So uh as long as you turn on the uh uh your your application showing on this dashboard uh it can be a part of it and if you check for example this one time lock if you click view testation we're going to bring you to a page with a detailed explanation and for example it shows uh this is running in the verifiable hardware the system is good system means the operating system as well as the uh infrastructure and the

119. 53:59

     KMS which is uh another components that secure your program and in the future we're also going to give you the application level verification that means you can link from your the hash of your docker container into your uh original source code and this can make the auditor or even non-technical people uh way easier to be convinced this is the real TE not a fake TE and there are

120. 54:24

     bunch of uh explanation here so I'm not going to talk too about that. And finally uh last week we launched our uh CRI uh cloud uh the the the cloud CRI. So uh for some of the people uh you use uh you you you feel really good to just use the web UI but if you go into a professional then uh you want to automate everything. So that's why we bring the cloud CRI. So in this CRI you can uh almost do everything

121. 54:56

     that's that's available on the web UI but always the command line. So all you need to do is to first install the CRI and then uh you can use the command to do everything you need. For example, link to your cloud account uh create virtual machines, manage virtual machines or uh retrieve the attestation of that. So uh yeah because time is really limited so I'm not going to talk too much about that. Oh yeah finally one more thing I want to mention is that uh

122. 55:29

     we just uh uh let me check if I can find it. Uh yeah here it is. So one more thing is uh when you are using the uh uh you when you are using uh uh uh like building the AI agent usually you not only build your agent itself you also need to use the uh AI uh models right so uh another thing we just shipped last week but actually not shipped it yet this is the preview uh uh version is that now we have integrated the uh TE GPU inference into the cloud as well through another

123. 56:09

     project called redpill. So uh that is uh it's kind of like a open router where you can use uh just a one uh account to access many many uh uh different AI models available uh uh on the on on on the platform. Uh but here uh what we want to uh uh highlight is that now we have the models running in the GPU TE.

124. 56:33

     So not only you can get your AI agent fully secure, you can also get the model fully secured with the TE. So nobody else including some Ottoman or you know whatever whoever you you can tell they have no way to say what's running in this model and what is the input or the output. So yeah, if you want to do something uh sensitive deal with the user privacy data, use this one. This will be uh uh bring your uh AI agent to the next level. All right. Okay. So

125. 57:08

     that's a very quick uh overview of uh what we have done last week. Yeah, you guys are are doing too many things to uh to fit in a single session. So it's it's we'll make sure to follow up with in greater detail uh with a thread because uh there's so much going on. So congratulations on all the progress. Um and thank you for joining as always. Uh, as I've asked everyone else, what is the best way for people to stay in touch with VA?

126. 57:36

     What is the best place for people to like, you know, learn more? Oh, yeah, of course. Uh, just check our Twitter account and we're going to push a lot of updates every day and also a lot of memes. All right. If you like memes and updates, go follow follow Twitter. Thank you, Hong. Thank you. Yeah. And last but very much so not least, we have Zahed.

127. 57:58

     uh he is a frequent attendee of the what did you ship this week because he's always shipping and it's always exciting and also huge congratulations on winning the useful agents hackathon uh your agent was by far well indeed the most useful uh and I if you want to kind of give a quick overview about what that is and then also then talk about what you uh what else you've been working on okay so hey every everyone I'm Zah uh I am uh

128. 58:27

     founder of jutsu and kind of a we are working on two main things at this moment. So the the first thing is the main priority is kind of a uh building a really good uh typescript SDK for uh for builders because uh uh Cameron uh Nia did a AI did a great job of the hackathon and it was great for two different reason.

129. 58:54

     one um I get some time to kind of hack on something random stuff that typically I will not do in my day job as a founder and the other thing was like I had the opportunity to kind of see how people are building and using the near AI Python SDK and what are the problem they have. So based on all this learning uh we are kind of a working and improving the TypeScript SDK. So it's not at this moment ready but I'm going to give you a

130. 59:25

     preview of how it will look like. Um let me start sharing my screen uh and then share screen. I'm going to share entire screen and I promise I don't have anything that I don't want to share. U so this is um I'll run a video. Uh so it's about the uh one of the main thing I learned during the hackathon was getting started to build the agent was not easy because people had to fight through the Python uh Python version mismanagement and uh kind of a from zero

131. 1:00:07

     to a agent took lot of time for a lot of people uh even for experienced PE Python developer. So our goal is to build a kind of a a mechanism where all you do is just a npm install uh something and then uh then you just create and um start building and start testing locally. So uh if you see the command line so for near TS it will be near EITS and then you just it's a similar uh same kind of a way to

132. 1:00:42

     create a uh agent but all of a sudden you have now a project with two file where you can run any uh TypeScript and you you can run any kind of TypeScript but you get all all the functionality for uh for near AI SDK through the core and then uh you can add your code to run. I'll speed it up. So you probably missed that one. So you will just say npm run. So and then it it will run the project and in the pro and then you can

133. 1:01:19

     kind of give your input. We also saw like how difficult it is to do the multi-line thing. So because if you if your agent have a multi-line kind of a input local testing is a nightmare. So uh with this one uh it's as simple as you don't have to as a user don't have to do anything. Uh the framework will figure it out. Um so two big uh important kind of a thing. one is uh you just have to do npm install and then it

134. 1:01:51

     will feel like any other uh standard uh typescript project like if you are familiar with uh create react app or or nextjs app so so it should feel like that for javascript developer so that's kind of the uh kind of the SDK update and we are working on it and u we'll give more update the second thing I'll show is the the h uh the hackathon project. So the hackathon project the main idea was uh I don't like to do social media especially LinkedIn but I understand how important

135. 1:02:30

     it is to do LinkedIn post. So imagine like I have a event that will be next week. So in fairness like I am going to run SF node on next Thursday. So now I need to kind of start posting for that event so that I can attract uh more attendees. So it's a it's a work like every day I have to do. So instead what if I could have a agent something that kind of a do it for me. So that's kind of the main goal that was the main goal. Uh so that's why like

136. 1:03:02

     we added this campaign feature where you will say like a in the description section maybe you will uh just copy paste your uh event description and then you will say what is the goal you have what kind of writing style and what is the call to action and which date to which state you want to have this event and then you will see like a uh so I asked from 24 to 28 and then it created

137. 1:03:26

     all this post for me and then uh if you are interested you can go and so the this is about the post product launch so you will see like a um this is I did not add anything um kind of a uh but you are like a you are always welcome to kind of a check the the generated output and then uh kind of a let's say I want here a probably uh a new line or you want to add something else but in few second you get uh post

138. 1:04:01

     for the whole week. So that's kind of a one of the important that was one of the main goal. So that's what we did. And now during the hackathon and outside hackathon when I'm showing this feature and it looks like a lot of other people are like me who don't have time or do who don't necessarily like to do social media and uh they want to use this tool uh for that. So uh it is available. So if you want to use just go app.post with

139. 1:04:32

     2t.ai. So the other feature that I will uh think like a is super interesting is um if you want to create kind of a a single post. So you not only just create a text post but also can add images. Um so and then let's say this is a generated post but it looks like too long. And if you want to make it shorter, so it's just as simple as like choosing it to make it shorter. So now AI will kind of make it short. Uh you can generate images, you can generate

140. 1:05:04

     more realistic images or if so and then this is probably because we all are engineers so we put a diff. So you'll be able to see like a what are the things that uh this AI agent kind of a change. So you can see this new version is uh pretty much half of the old version. And then uh if you are okay with that accept and then you can uh kind of improve it if you want. Image is same. Uh so if you

141. 1:05:32

     want to generate a new image so in this case I am regenerating you can I can regenerate but of course you can always generate with all the different other style and other features. Uh and then the last thing is we are working with videos. So if you want to add a video so that's then you can just use the video.

142. 1:05:53

     Uh so yeah so and the last thing I would say like some people also I found like a through research like a carousel is uh very important for LinkedIn and I have no time or interest to build a carousel. So with this one you can just say um okay near AI agents and then it will kind of do a Google search and if you it will try to find information about near AI and then I don't know what it will find welcome to the future of AI introduction to near

143. 1:06:25

     a hen capabilities um so whatever it found it's it's kind of a but you can see like you can uh it gives you think of it this tool as V 0 so where you get the version zero and then you can quickly iterate because I don't know about you but for me uh writer blocks is a thing and English is not my first language so I always find this kind of tool useful and I don't have to copy paste from uh uh chat GPD or kind of a h create a really nice

144. 1:07:01

     prompt in chat GPD so that's kind of the main reason for building this tool And if you want to use it, go to uh kind of a uh app.post.ai. Um and if you are using if you want to take LinkedIn seriously, so start using this tool and follow the post uh page in LinkedIn. Thank you.

145. 1:07:28

     Awesome, man. Super exciting stuff. Uh looking forward to the improvements to the TypeScript SDK. Um, people come to me all the time saying, "Hey, I'd like to deploy an agent I built on the near AI hub, but Typescript." So, um, I think this would be a big unlock for the ecosystem. So, thank you for coming. Is there any last words that you want to add? Any place that people should follow you? You know, obviously try post. I'll be talking a bit more about that later.

146. 1:07:52

     Um, is there anything else? Uh, LinkedIn. Uh, like I'm probably more active in LinkedIn. Uh, yeah. And send send me a message in Telegram. You can send me message in telegram anything about typescript update or if you have any suggestions. So we have open source repo so you can follow uh the project our progress. Um and any feedback constructive feedback are welcome. So awesome. Well thank you.

147. 1:08:23

     Thank you. Uh, I guess I'm the last one to go and the TLDDR is that there's been a lot going on in the near AI world. Uh, right now, let me just share my screen. Um, going to be giving you guys a quick update. Everything can generally be found on Twitter um on exactly what we shipped last week. Like we do our best to build the public. And so, uh, kicking off from I guess this week is the hackathon. We ran this on Saturday.

148. 1:08:53

     Uh we had over about 280 people sign up about 140 people actually attended. I had about 18 project submissions. Uh the whole focus of this hackathon was to build useful agents. Useful agents. The youth utility piece is defined by some sort of benchmark that you create to well prove its usefulness. And so uh that was really wellreceived. That was in partnership with Electric Capital and

149. 1:09:17

     SF Compute. Uh, Electric Capital is probably the best, if not one of the best uh, VC funds in cryptoland and they've been doing a lot more deals in AI. One of their portfolio companies is SF Compute where you can actually buy or sorry, rent or lease GPUs by the hour without any long-term contract. This is a really big deal for anyone that is, you know, training models or, you know,

150. 1:09:41

     needs very flexible uh, compute contracts for inference depending if you like get a lot of inference at a given time. And so both of those partners are great. Uh we had an amazing set of bounty partners. There's a whole thread kind of unpacking the photos. Uh we had uh uh Matt Wolf, no uh Adam Wolf from Anthropic give an amazing presentation about uh he actually helps run cloud code. Yeah. Anthropic. Um talking about what makes a useful agent. Evan from SF

151. 1:10:10

     Computes, Ilia from Near of course. And uh that was really, you know, impactful and educational talk. These are some of the winners. Uh you just heard from Zahed with post. Uh second was diligence AI. They built a uh VC due diligence platform and a set of agents that you know decreases the time that they actually need to take to do quality technical and nontechnical due diligence. And this was a winner uh near

152. 1:10:35

     food is an AI agent that can help you find free food uh based on Luma. uh if people are organizing events, it essentially scrapes Luma and uh uses and recommends you events to go to for free quality food. And I will say there was amazing food at the near at the useful agents hackathon uh which some people may have just gone for the food. So hopefully uh this doesn't make a bunch of food grifters uh come to our feature events, but it's a cool agent. I'm not

153. 1:11:08

     going to walk through all of the agents, but I will say it was a pretty cool and eclectic group of builders. People everything building from like research tools for biology to uh test time compute routing agents to, you know, improve reasoning uh on for agents and other LLMs to, you know, vacation mode, which is an AI agent that can try to save corporate knowledge bases. Uh so if someone goes out of office for a while, like they're not totally lost when they come back. uh you know the classic

154. 1:11:37

     wallet portfolio managers uh other apps to agents to identify the cheapest uh version of an an item you're trying to buy. Uh factbased uh like agents to to fact check news AI agents to demonstrate how uh you can conf confidentially navigate uh robotics. So it's like agents doing everything. It was pretty cool. Uh really enjoyed the event. And uh next up, another big thing that Alex Kafford from the Near AI team did talk

155. 1:12:10

     about last week is Near's implementation of MCP. This is anthropics model context protocol. And the whole point is now if you are using any type of agent, any type of building framework that is supportive of Anthropics MCP, you can now use those tools to build agents that can interact with Near. uh they can do things like manage accounts, check balances, send tokens. You can read this

156. 1:12:35

     thing here. Uh but it's pretty cool and I think a really great way to sort of plug into the greater MCP uh ecosystem which is growing very quickly and obviously there's a lot of future work to be done. Uh this is I think really exciting stuff. There's a bunch of like really big sensitive projects building in the near AI world right now or the company Neurai that I can't talk about

157. 1:12:59

     but hopefully uh in the coming months or weeks uh we'll be have more to share there. Um, oh, I forgot to mention Ilia uh did speak. He had his own keynote at Nvidia GTC uh highlighting a new paper that was announced called uh decentralized confidential machine learning uh that essentially highlights how uh model builders and fine-tuners and really agent builders as well can get paid for their contributions where uh they can even like have a closed sourced model where you're able to

158. 1:13:31

     verify the inputs and reward people based on their contributions to fine-tuning a model. pretty exciting stuff. Uh, a lot of it is TE based. Uh, there's a whole paper here that will unpack it. Uh, if you are looking to kind of get the TLDDR, I suggest checking out Ilia's Twitter thread and you can check him out and follow him at ill black dragon. Uh, this was what he pitched at the Nvidia GTC conference and you know more more to come on this front. This is sort of just the paper

159. 1:14:00

     explaining how everything would work uh when we are trying to incentivize uh model builders and finetuners with money because everyone knows uh the whole world of like AI right now is is the business model is not truly figured out yet and I do think that crypto and uh payments based on attributions uh contributions are pretty much a prerequisite to making AI business models sustainable.

160. 1:14:27

     So really cool paper unpacking that. Uh what else do I have here? Uh oh, we are, you know, I'm still working on the uh development of an AI auditor agent. This was all uh reviewed and announced in a RFP. There will be more RFPs coming out.

161. 1:14:42

     So if you are an agent builder looking for very clear scoped work, uh there will be more RFPs coming out for agents that we want built and hosted in the near AI hub. Besides that, um I mean I manage near's Twitter account even though I'm not really a marketing manager or you know really have a ton of experience. So I I do manage all of that and that's sort of what I uh some of the things I've been doing this week. Uh right now really just following up with a lot of the

162. 1:15:09

     partners that you hear here today. Um there's a couple other really big integrations coming out. Um, I'm also the whole new AI team will be start hiring a lot of people soon and so at least some of the stuff I'm working on is uh building out entire hiring plans and you know scaling the team because we are going to need a lot more capacity if we're trying to build truly user owned

163. 1:15:31

     AI. Um, these are some of the things I'm working on and obviously uh we'll share updates next week but that's about it for now. Um hopefully we can get more near AAI engineers on this call in the future. Um trying to do it, but sometimes they would rather focus on shipping than talking about what they're shipping. So guess my role. That's about it, guys. Um shout out to Near Week for facilitating this every week. Um and

164. 1:15:58

     everyone who came to participate. Uh we do this again every Thursday around 11:00 a.m. Pacific. That is 2 p.m. Eastern. And uh come back next week. Hope to see you then. Thanks for coming. Thanks for watching. Subscribe to Nearweek.