Ep 165: MALWARE. WTF is "Decentralized AI" with special guest Jarrod Barnes, head of ecosystem at NEAR Foundation

[00:00] Malware is a non-technical look at the tech news of the week. This is a podcast where we learn together about everything from crypto to AI to whatever comes next in tech. I'm Natasha Hoskins. I'm Dina Burke. And this is Boys Club. Wait, is it just Boys Club? It's just Boys Club. [00:18] The boys club podcast. No, no. [00:20] Just boy stuff. [00:22] Hello and welcome to Malware. I am one of the hosts of Malware, Dina Burke. Normally I have a co-host, Natasha Hoskins, but she is off living her best life in Singapore right now. So it's just going to be me. I have a special guest on. His name is Jared Barnes. We are going to be talking about decentralized AI, [00:44] what it is, why it matters, who is doing it and what's next for the industry. [00:51] I hope that you learn a lot. I sure did. Thanks so much for listening. [00:56] you [00:57] Hey, Natasha. So a question we get asked a lot is, what do you look for in a crypto platform? So let's talk about it. Well, Dina, I look for a secure, no fuss platform that I can dive into right away. That's why I love today's sponsor, Kraken. If you're waiting for the right time to get into crypto, Kraken makes it super easy and intuitive to get started. Plus, if you get stuck, they have an award-winning client support team that's available 24-7, along with a bunch of educational guides, articles, and videos to help you along the way. If you're ready to check out [01:26] So kraken.com backslash boys club and see what crypto can be. Not investment advice. Crypto trading involves risk of loss and is offered to U.S. customers through Payward Interactive, Inc.

[01:39] We have Jared Barnes, our special guest on today's show. Just a quick bio here on Jared. Jared is the head of founder success at Nier Foundation, where he leads AI ecosystem efforts and runs Nier's AI incubation program, working with founders who are building the future of user-owned AI. I see user-owned AI a lot in our notes here. I'm excited to chat with you about [02:09] bio. Previously, you taught entrepreneurship, gaming, and esports at NYU, where you also managed the university's startup incubator. How cool. Worked as a venture investor and led player marketing and corporate innovation for the Los Angeles Rams. Welcome to the show, Jared. [02:25] Thank you for having me. Super excited to dive into this ever evolving intersection of [02:31] AI and Web3. Happy to start anywhere. Let's dive in. [02:35] I've been thinking a lot about our interview coming up and thinking about this mysterious intersection and also seeing a lot of stuff around it on Twitter and in the sort of zeitgeist. I know you think about and talk about. [02:49] This idea of decentralized AI, I hear you talk about user-owned AI. I'd love to talk about the distinction between that and what I think a lot of our listeners think about when they think about AI, which is basically ChatGPT and what would be, I guess, the centralized counterparts to what you call decentralized or user-owned AI. So yeah, just to sort of level set for folks, what would be the difference between like a ChatGPT and the decentralized AI that you are working on?

[03:19] Great question. [03:20] So let's unpack this. And I think probably one of the bigger fallacies in crypto is we start adding new terminology, new acronyms without fully explaining things. And so simply, if we're talking about decentralized AI, really all that is, is you are distributing the core systems of what AI is. So that's distributing how data is collected. [03:41] how that data is trained, how models are then built, and then how inference or how people interact with those models, right? So you're essentially decentralizing what's called the machine learning pipeline or this whole end-to-end journey of you as a consumer, right? You see ChatGPT, right? You log into your account on OpenAI and you ask a question to GPT-4-0, right? The reality is there was an entire process in order to get to that end output that you received. DAI and [04:11] is distributing that across systems that run on chain to avoid that centralization of power, right, as well. So that is a fundamental aspect is decentralized AI. When you start throwing in user-owned AI, user-owned AI simply is the incentivization layer [04:27] that lives on top of that in order to make [04:31] that system happen of decentralized AI, right? The analogous example I think is what IPFS is to file coin, right? IPFS, [04:40] decentralized storage, right? Filecoin, the incentive layer that lives on top of that, in order for a network of nodes to essentially store and exchange value on that existing data that's stored, right? That is really essentially what we at MIR are pushing through decentralized AI and this concept of user-owned, really with the idea of

[05:03] Not only how do we avoid this centralization of power, but more importantly, how do we actually create micro economies at scale for the individuals who are pushing this industry forward to benefit from it? [05:16] Okay, so this supply chain or the sort of background work that goes into the AI models that [05:25] As we understand it, like what you're saying, we log on to ChatGPT, we write in a prompt that spits something out, but there's a whole world of work that happens before that we're able to get that answer, that prompt response. So my understanding is that. [05:40] OpenAI end-to-end owns... [05:43] data, pre-processing, training, optimization, compute, all of that stuff. They're kind of those layers. That stack is owned entirely by them. And... [05:54] I can imagine that makes it quite efficient to run and to optimize. So with decentralized AI, if we're splitting out all of those steps and stages into different [06:06] protocols, perhaps maybe different companies, the stack itself then becomes decentralized. Am I thinking about it in the right way? [06:12] Thank you. [06:13] You are. You are. Yes. Right. So to simply put it to your exact point, right? Centralized AI, let's take OpenAI, Google, Microsoft, right? At the end of the day, they own data centers. They have proprietary data sets of their own. They train models on their own, right? Meta-trained model three, Lama three on their own, right? And use the variety of data, data sets that they aggregated together for that. And then they host and ship

[06:42] that model themselves, right? That is like end-to-end centralized AI. And they built the interface for everyday consumers to interact with that, right, as well, and have developer platforms and enable people to build these models. Decentralized AI, what you're doing, again, is you're distributing those systems across a combination of things that could be consumer hardware, that could be myself actually using my, what, MacBook M3 as a node that is part of a [07:12] network that could be part of it, right? It could be you as an individual consumer, right? Engaging with something that is decentralized or running on chain as well. So there's a lot of different components and we can totally unpack that, but that's a bit of the core differentiation is essentially, yeah, decentralized AI, you're, you're distributing that across multiple nodes, [07:32] in a ecosystem and network that are together. I think a lot of listeners hear us talk about why we think decentralization is important and why it's a core key value of the Web3 and crypto movement. But especially as it pertains to AI, what do you think is... [07:50] the risk of these centralized players like a Microsoft or Google or an OpenAI dominating the industry? Like what you are pouring your life force into the work of decentralized AI? Like what? Why? Why? Why? [08:05] Yeah. So I think it's a couple of things, right? And I typically don't lean into selling fear, right? Because I think fear is an interesting motivator, but ultimately it doesn't bring the best out in people, right? And so I think today's narrative, you often hear that centralized AI is this terrible, horrible thing, and it's going to be four companies that rule the world. I think there are layers and elements of truth to that that are important to be wary of. But I would

[08:35] perhaps Meta and Microsoft have done more for open source AI than perhaps some true open source companies, right? With what Meta has done on the Lama side and Microsoft essentially putting out VS Code, which is arguably like the de facto developer platform across the world, right? Completely free and open source. And so I think there's one end of the spectrum there. The other end of the spectrum to, I think probably the core of your question is, well, if these companies are [09:05] amount of value, right? OpenAI does, it was a study released by the information. They do 225 million in MRR, [09:13] which is like 3.4 billion in ARR. So they're, they're doing okay. And then monthly revenue, recurring revenue, and then annual recurring revenue. [09:21] This is me falling into the trap of using acronyms. Yes. Monthly and annual recurring revenue. So 3.4 billion in annual recurring revenue. OpenAI. Yes. And then you add in meta and perhaps those that are considered like the magnificent seven or the top performing kind of public companies right now in videos of the world where, you know, their market capitalization or in other words, like the total value of that company is $1. [09:46] bigger than crypto as a whole. So now you're putting serious dollars and the incentives of serious dollars, which are to continue to make that economic flywheel go at making these models better. Well, how do you actually make these models better? You need individuals' data. [10:04] Right, because what we've

[10:06] come to at this point is that pretty much the entire internet [10:10] has been scraped. [10:12] already. [10:12] and used to get [10:14] we know is GPT-4. [10:15] right um and even llama 3 was about 15 trillion tokens which is you know a good chunk of the entire internet in order to actually continue to advance ai you need what is not necessarily publicly publicly available that's like your private data that is essentially one of the only ways we're going to continue to advance models right that's one path there's a lot of paths we can talk about but that is kind of one key path right and that's why a lot of uh these big [10:45] largest repository of [10:47] private data, or they live in enterprises, right? Then it helps inform the model and it gets better over time. And so I bring this up to just share context, because if the incentive is for an open AI, is for a Google, is for an Anthropic to continue to grow their economic flywheel, they're going to do [11:06] essentially whatever is necessary to return value to their shareholders and continue to grow the value of that company, which likely means finding ways to gain access to our private data, which likely means those companies kind of deciding how those models are trained, which also likely means us as consumers. [11:24] probably accepting the output of those models as truth, right? Because I think one of the challenges is these models are getting to a point where the outputs are at such high quality that it's hard to understand, well, what is true and what is not. And if that is not essentially baked into your thought processes as you engage with AI, that could become quite dangerous over time, right? In terms of misinformation and disinformation as one bucket. And then the

[11:54] It is, I think, probably to the core of your question is now we're living in a world where we have essentially a handful of entities and very, very small amount of people, perhaps the 1%. [12:07] determining [12:09] what truth might be for the 99%, right? Which kind of tips the scales in a way that we may not want to see. And so that is where a sense a lot of our motivation comes from is how do we actually create that in a more equitable way, right? That is not overly reliant on these kind of massive companies and massive systems. Well, that means we need to be build a reputable and an additional, an alternative system that is actually at parity in terms of quality, but [12:37] open. [12:37] but private, but verifiable, right? Like some of these core pillars that you may hear around decentralized AI, and we can totally dig into that. But these are kind of when we're really digging into the meat of the problem and like the why around this, that is essentially at its core. [12:54] When I hear you talk about sort of the risks of centralized AI and these four companies siphoning up all the data on the Internet and using it to make better returns for their shareholders, I see that and I think that that's... [13:09] really frightening and [13:12] I'm not convinced that that will motivate people who aren't already sort of drinking the Kool-Aid of decentralization to act any differently, make any different decisions about using things than they already are. And that convenience is actually the biggest driver for adoption of these things. And so I think the products, the decentralized products need to not just be on parity with what the centralized products.

[13:35] offerings are, but I think they need to be [13:37] markedly better in order for them to have a chance given we're already so embedded with the Microsoft's and the Google's [13:45] The convenience is already there with the centralized offering. So I'm curious if you feel like that is possible. Also, given how you're speaking to the capitalization of these companies, they have so much money. And my understanding of how AI research and development works is that it's incredibly expensive. The compute is incredibly expensive. [14:05] There's some natural sort of moats that are built around these companies because it's so expensive to do. So yeah, just curious what your response would be to those two things. [14:14] Look, it's a great question, right? And I think to agree with your points, I don't know if someone has inherently chosen a product because of privacy, but they definitely have because of convenience, right? And so I think to the first... [14:27] point you made. One of the biggest things that decentralized AI truly is in need of is [14:33] killer applications, right? And again, this is, we're kind of repeating some of the same things we said across crypto. And so I think one of the core things is we have to understand like, where is it that we actually can truly win? Like what is meaningfully going to be differentiated in a way that will drive not just awareness, but kind of the aha moment of like, oh, wow, like I see the value now. And you start to actually convert people over. We have a lot

[15:03] we're testing and validating right now. And they've fallen in a couple of different buckets on the application side, ranging from truly... [15:10] essentially autonomous agents or LLMs and models that can take action on your behalf and being able to do that in a compelling way, whether that is on-chain or off-chain, and we could totally talk more about that. The other end of the spectrum has been more truly having blockchain as a back-end, right? And essentially Web3 solving problems for AI that currently have not yet been solved. [15:40] ways to have unique and specialized models where I think today we interact a lot. The average consumer probably interacts a lot with quad 3.5 through Anthropic or GPT-40 or even a one preview, right? These are all like the interface has been built for that in a pretty compelling and like simplistic way, but also the cost to run these models is incredibly high, right? [16:10] on everything that comes with running a data center with hundreds of thousands of computers or hundreds of thousands of servers together. We're actually distributing that cost. So therefore the cost can go down. And so what that enables is blockchain as a backend is actually a much more compelling solution to build the next frontier thing. The challenge has actually been, as you alluded to, is how do you get people to trust that? And that's where, quite frankly, the fundamental, I don't want to say gap, I was like fundamental opportunity.

[16:39] within decentralized AI right now is benchmarks and evaluations. [16:44] In the same way you choose a doctor, a lawyer, any service provider to engage with, right? You have a clear, even a restaurant. [16:51] You're going to go on Yelp. You're going to go on Google reviews. You're going to look, does it have 4.5 stars or more? Do I know it's good? Is this going to be worth my money and time? [17:00] That's the exact same way you evaluate [17:02] AI, right? Especially as a enterprise or somebody who's building with AI is, well, how good is this? Does it save me money? And does it help me perform better? If we do not answer those two things as decentralized AI, we are at risk of falling into the abyss of crypto Twitter versus meaningfully solving problems at scale. It's, hey, you need a killer core application. I mean, we can talk about some of those use cases, but you also need a way for people to understand how good this thing is [17:32] the like [17:33] aha or wow moment of a demo as well. [17:36] So we're not trying to just build another chatbot that would be parity with OpenAI or Anthropik or some of these other user products that people are interacting with today. Where in the stack are you most excited about? [17:52] the innovation that's happening and where do you where you see the most opportunity is it on the end user application front maybe more than chatbots something else other than chatbots is it in this ai agent thing which is kind of i see a little bit mid-stack um potentially or is it sort of deep in the technical stack around sort of the compute and the the foundational models and stuff like that where do you see the most opportunity in this world

[18:19] Yeah. And look at you on the technical side. I love this. I try. I try, Jared. This is awesome. I try. Yeah. You know, it's, um, and look, it's like, it's dense. Yeah, it really is. I'm trying, I'm hanging on by a thread, but we're here. So look, a couple of things, right? I think we've seen innovation in all three areas. [18:37] of those layers you mentioned. I think the reality is largely what, if you think about fundamentally, [18:44] What is blockchain? [18:46] great at. It's great at verifiability. [18:50] Right. We know something is on chain. We can verify that without sacrificing anything. [18:55] performance. And so [18:57] The fundamental area which you've seen decentralized AI projects raise the most money and quite frankly gain the most traction has been on compute and simply just this foundational layer that is required. I think that is one area where in terms of Web3 solving meaningful problems for AI. [19:17] we have already started to see. One example of that is a project called Hyperbolic. It is within the NIR ecosystem. They are a GPU cloud provider. In other words, they enable anyone to essentially spin up compute on demand and be able to use that to engage with a variety of different AI models. So say it's almost like anything you'd be able to order quickly on demand. Think of [19:47] all this inventory in my closet at home but anytime somebody orders it i want to be able to ship them a t-shirt that's essentially gpu cloud right you're able to spin up what's called an instance which enables you to print that t-shirt uh run imprints on a model or call a model and engage that model and give a user a response to an output

[20:05] Drop shipping GPUs. [20:07] That's what we're doing. Dropshipping. Let's do it. Let's do it. I love it. I hope you trademark that as well. So that's essentially like what... [20:15] the value is and why why that matters is again with a distributed system you're able to bring that cost down. [20:23] immensely because you're pulling from kind of different hardware across a variety of different data centers. And so the challenge with that is you need great infrastructure to make that work at scale. [20:34] That's the work of what it takes to be great in this space. But if and when you achieve that, you're able to, A, drive that cost down, and B, deliver faster inference or faster results where you're not just waiting for 10 seconds for this AI to respond to you. It's actually like, boom, instant, right? And the kind of the unit of measurement in the industry is called like tokens per second. In other words, how many words per second can like an AI model deliver back to you, right? [21:04] the wow, oh my gosh, that was amazing, like magical moments delivered at scale. I have a question just on that note. Please. [21:11] - Yeah. [21:12] Sam Altman sees something that is bringing the time of compute down considerably, doesn't he just come in and acquire it and then it just becomes part of the centralized stack like that for him as a as a business owner, he's probably looking to solve for that directly. And doesn't that just become an acquisition target for him or for Anthropic or whomever else?

[21:36] It's a great question, right? And maybe to respond back to the earlier point you had, this is the value of a decentralized network. Because the reality is, if I'm a GPU cloud provider, I don't own that hardware. This is print on demand. We're dropshipping, right? I don't own that inventory. And so even if they did want to acquire me, they're not going to acquire my inventory, right, as well. They're not going to acquire actually those GPUs. [22:06] And so there's an incentive mechanism for me to continue to provide my compute or provide my hardware to this specific network. And if somebody were to acquire that network, it is my choice if I want to continue to provide that or not. And so that's where some defensibility really comes in and differentiation, where I think it would be quite challenging to. [22:29] for a centralized player to come in because then they would be managing this distributed network, which I'd be willing to bet. I don't know if they want to do that. [22:37] I see. So it's easier for them to just spin up server farms or it's probably not a server, GPU farms, whatever. Yes. Buy land. Buy land and be building them themselves and be managing the oversight of them completely versus having this distributed network. The war right now, particularly on computers, it's a bit of a race to the bottom in terms of pricing. Well, you sell your t-shirt for $19.99,

[23:07] and we're going to see who wins, right? And so that's where the competitive dynamics come in. Again, with a distributed network, you're able to have a bit more flexibility on your pricing structure because the cost for you to facilitate that network is much lower than if you owned the hardware yourself, right? Because you're thinking like maintenance costs, you're thinking, well, what if one of my servers go down, you know, so on and so forth, right? And so you have a bit more flexibility there in terms of like price structure or so on and so forth. [23:36] Okay, so that's sort of deep stack. Let's talk about the AI agents, because I see a lot of chat about that on Twitter. So tell me about this. [23:43] For sure. So fundamentally, like, actually, like, just step back, TLDR. [23:47] agent, they enable a language model to interact with its environment. Without an agent, pre-agent, a large language model is [23:56] input [23:58] processing. [23:59] output, right? I give a language model question that language model, uh, uh, essentially runs computation or it thinks about the question I gave it. And then it spits back a, uh, response to me. [24:10] agents enable me to ask a question. [24:13] That language model now is going to call an agent and that agent can take action. It can take a hyper-specific action, such as creating a wallet, such as transacting on ETH, such as minting an NFT. That is a hyper-specific action. Or it could be kind of a variety of agents, right? Which are what we refer to as multi-agent systems. Or you heard this like fancy word of orchestration. In other words, it's just a way of telling the model what it is you want it to do and how you want it to do it.

[24:43] Thank you. [24:43] right, is really all orchestration is. Great example, you give an agent $1,000, you tell that agent, here's what I'm looking to accomplish with that $1,000, right? I want you to allocate it across these three tokens and my risk is moderate, right, as well. Don't put it all in meme coins, right, basically. That agent then can decide autonomously based on the inputs that you've [25:13] That's basically where we're at today. [25:16] Right. And kind of what I shared earlier on like benchmarks and evaluations is the only way we know that it's good is if it made us money, which is great. [25:25] But the reality is if it does work, that's not going to be a secret forever. And over time, that is actually going to degrade the quality of the agent because you're going to give it. [25:34] more bad data because everyone's going to try to optimize for [25:38] making more money, right? And so you actually need benchmarks to understand, well, if an agent took this action, why did it take that action? What led it to taking that action? It's no different than like, [25:49] you're trying to under, for any parents, like you're trying to understand like why your kid made a decision. [25:53] You actually kind of like sit down with your child. You got to unpack that. Well, what led you to making that? That is the research behind like, [26:00] the term evaluation and benchmark is like, you're just having like a parent teacher conference. Like did you, did you study for the test? What was your preparation for the test? Right? Did you ask questions in class? Again, arbitrary example here, but it is the way in which research is conducted around performance, right? You're giving it kind of a series of tasks. You're evaluating how well it performed those. I gave you a thousand dollars.

[26:24] How well did you return that capital? Did you lose all of the money? Did you make me more money? Why did you make those decisions? Right? In the context of kind of the intersection of AI and Web3, like this is where the industry is at today. There are a number of teams like tackling this head on. This feels unique to Web3 and crypto. This is something that my chat GPT, my Claude absolutely cannot do. And so the ability for these agents to spin up a wallet and to transact. [26:51] is very unique to this space. What are some of the teams that are working here and [26:59] What do you think is like the next inflection point for this? [27:04] industry that's existing within the broader AI landscape. [27:07] Two teams, I think, that are tackling the spectrum of this space that I think are really, really compelling and impressive. The first is a team called Almanac. Almanac is solely focused on what's called agent simulations. So kind of what I had shared earlier on evaluations and benchmarks. Let's say I give an agent $1,000 and I run hundreds of thousands of scenarios to better understand what they would do with that $1,000. [27:37] So they actually go to hedge funds, they go to PE firms, they go to family offices and say, hey, we are able to run hundreds of thousands of simulations over your allocation strategy. [27:51] How about you allocate a portion of your AUM or your assets under management to our agents? We'll take a percentage of that.

[27:59] transaction as kind of a management fee, but you're going to get the data and insight on performance of a variety of different scenarios. And that is going to be unique and differentiated, specific to the amount of money that you allocate to us. Really interesting. [28:14] really impactful, unique to crypto. [28:16] right as well right especially you talk about running that simulation over forecasted eat prices right as well so now you're getting into like some really really unique data on the institutional side that that matters for you know those specific customer sets on the other end of the spectrum let's talk about like everyday end user and let's say i want to do something similar to that then you know delegate um a hundred dollars to an agent that lives on shopify or an agent that lives on [28:46] He's like, [28:47] Here's kind of what I'm looking for. I don't know, really know what I want, but here's like directionally what I want. Can you go find like five different options for me? Give me kind of a variety. And then based on what you know about me, go ahead and buy that. [28:59] item for me. There's a team called Nevermind, which is focused on the concept or idea of agent to agent payments that is tackling this head on, which that's like, well, big idea. But basically, the whole premise for that is, well, if we're going to have agents transact on our behalf, they're likely going to do that with another agent. [29:21] right as well. And so that's where it gets like, whoa, you have to kind of zoom out and understand that [29:26] Okay, if we keep going at this pace, [29:29] the predominant end user on the internet.

[29:31] is likely going to be an agent. [29:34] And so if we are on these marketplaces and let's say Shopify has their own set of agents, we need a way for those to communicate with each other, right? And transact as well. And it sounds like very, very futuristic. I think, again, the pace at which the market has been going, the reality is it's probably only eight to 12 months away before this is achieved. [30:04] We at NIR just actually shipped a completely open source example that works with Refinance, which is one of our DEXs on NIR. So it sounds like super futuristic, but actually the tech is here. [30:15] Right. We have to normalize the behavior. We have to embed that into the shopping experience. Right. As well. My assumption is probably going to start with items that are in your existing cart and kind of be the nudge to go ahead and delegate that to an agent to get that user journey started. So that's where I think as consumers, as marketers, as operators and designers. [30:38] agents are probably like one of the last blank canvases that exist in like a, just a design space because [30:45] that hasn't necessarily been solved just yet. Like the foundational models, the like GPT-4Os, the quad 3.5s, like, [30:54] A lot of people working on that. [30:55] A lot of money being spent on that space. [30:57] agents fundamentally [31:00] are relatively new.

[31:02] design space and the implications of what that means for consumer experiences, we are at the tip of the iceberg. [31:09] so that's personally an area where i'm excited about but it's like one of the spectrum [31:13] institutions. [31:14] high amounts of assets under management, running simulations. The other is like, whoa, really compelling design and consumer experiences for everyday individuals. How might we impact both? I'll have my AI agent talk to your AI agent. We're good. We don't have to do this podcast anymore. [31:33] wow the the ai agent stuff is is crazy i also it will be funny if we're all in the crypto industry talking about how to onboard the next billion users and those billion users end up all being ai agents which which i i know we laugh about it i know we laugh about it but i i have to believe there is some truth to that statement largely because a couple of things like [31:56] Agents have proven [31:58] to be great so far at code generation. [32:02] and being able to take large amount of transaction volume and be able to get some form of predictions to that as well. [32:10] And what is it that we need in crypto? More developers and more people that are transacting on chain. Right. So I actually, I know it sounds again, like far fetched, but I think the time horizon or the timeframe in which [32:23] that might actually... [32:24] be true, I think is relatively near future. I'm going to give my AI agent 500 bucks and say, make me the most amount of money you can make me on Polymarket. There it is. Great experiment. Okay. So curious though, I mean, seriously, it sounds like there's an incredible amount of opportunity there. I know you shouted out all the marketers and designers and perhaps non-technical people who are listening to this podcast that that's the place to be looking, but like specifically where you name drop some startups, but...

[32:54] I would personally, selfishly love to learn more about the AI agent space. Where would you point folks who are listening who would be curious to dig deeper? It's a great question. So I think on the agent space, particularly, [33:08] a great potential starting point is simply just experimenting with, [33:12] yourself. There are a number of no code or just like very, very simple tools that you can start with to better understand just how agents work. One of those teams that is meaningfully building those out is called Langchain, started by an individual named Harrison Chase, amazing entrepreneur and founder. They have essentially become the de facto agent framework, or in other words, [33:42] run at scale. I think starting there to just better understand how agents work is an incredible, like just baseline, because the reality is, [33:52] agents [33:53] Think of that as a feature to an existing application. So like, let's take your Apple Watch and like closing your Apple rings as a very simple interface and experience that a lot of people could understand, right? Imagine if there was an agent that could extend that experience or create more value with that experience that would tell you, based on your historical patterns of your fitness data or your health tracking data, we're going to now provide you like specialized recommendations, right? For that, a team like Langchain and other companies [34:22] agent frameworks enable those experiences to happen as well. So I think the interface to start playing with agents isn't necessarily there just yet. It is much more like the developer and technical side, but understanding the capabilities.

[34:39] is where I think the true power is because of the rate at which tech is advancing. So by the time we actually figure this out on like, oh, this actually be a great use case, the tech will be here. [34:51] It is literally moving that fast if you look at just the pace of innovation in the space. [34:56] We're going to take a quick break and we'll be right back. [34:59] Thank you. [35:00] It's time for a more open, inclusive, and transparent financial system. A system that serves nearly everyone, everywhere, all the time. That's why we love today's sponsor, Kraken. Kraken is a crypto platform that provides a super simple on-ramp to the world of crypto with a 24-7 support team. Crypto transcends physical and imaginary borders. No matter where you are, you can send funds easily and quickly to almost any part of the world. Plus, forget about waiting times and waiting lines. You can send, receive, and trade crypto anywhere near instantly. [35:30] B at kraken.com backslash boys club, not investment advice. Crypto trading involves risk of loss and is offered to us customers through payward interactive Inc. No third-party transfers available. [35:42] I have to ask the question, which might be on some people's minds as they're listening to this. There's a big conversation around AI safety and that we see play out even as high as the presidential conversations. We have our EAC folks versus our degrowth folks. The war is being waged on every front, it feels like, these days around AI safety in particular. And I'm curious...

[36:10] how this world intersects with this idea of AI safety. My understanding is that a lot of the sort of safety protocols or processes that something like an open AI does or has is like this idea of red teaming where they send in researchers to sort of test and establish boundaries for what their models will and won't do. And that feels like something that is centralized. [36:40] centralized counter. [36:42] Points to the AI safety conversation or is it like in the spirit of crypto and web 3? [36:48] Is it this sort of libertarian... [36:51] free for all, anything goes, where, where do we, where are we landing here? And I'm sure there's a spectrum. [36:58] It's such a great question. And there's a spectrum. And I want to start with like a tangible, real example that's like happening today to make this like a bit more approachable because it's a very meaty thing. And we could probably spend an hour just talking about AI safety and probably like lose a little bit of people. But here's the reality of like what's real today. And I'm like, I will share this article so you can include it in the show notes. But Wired came out with an article specifically from Apple. [37:22] in the Apple Vision Pro, right? And you remember how compelling the Apple Vision Pro was in the eye tracking technology that existed because the haptics and like a user experience was like incredible. Well, what researchers found is Apple was actually able to track up to 92% accuracy, your passwords and anything that you would type.

[37:42] from their eye tracking technology. Wow, that is computer vision. [37:45] that is based on an AI model. And so now you have a company that [37:50] actually can just based off your eye patterns and eye movement can actually fully gain access to any data they wanted to of yours if they really wanted to and so when we talk about ai safety i try to frame it in like the real things that are happening today and that and mind you this is happening today at like 2024 ai we're not talking about 2030 ai we're not talking about 2035 ai we're talking like today what is happening and so i think where ai safety becomes paramount becomes really [38:20] invisible. [38:21] Because if you think about chat GBT today, it's visible. It's what's called human in the loop. In other words, like you are giving the input and you're seeing the output and you're experiencing in real time. Well, what happens, what happens when AI becomes invisible? [38:34] What do I mean by invisible? If you think about any time the last time you've used your debit card or credit card and done the tap to pay, you don't really think about the underlying technology of an NFC chip and transactions and the transaction fee that Stripe would take from this payment provider. It's all invisible. [38:50] Right. [38:51] But the reality is like there's significant value capture there. And then that data is used to determine everything from your credit profiling to your credit score, like so on and so forth. Right. And so if you add that into the context of AI and now you're talking about like AI is actually becoming so embedded into our day to day, we're not. [39:10] again, actively like

[39:12] thinking about any of the implications, that's when safety becomes really, really important. And that's like the practical... [39:18] spectrum of today. [39:21] You play this out into the future and now we're actually talking about model capabilities and I can we can talk maybe a little bit more about like open AI's recent release with a one and like what does that actually mean and why does that matter? Right? You're talking about model capabilities that are becoming and will continue to become so strong. We are introducing [39:42] reasoning, the ability to decide, right, to AI models. And [39:48] When AI models have the ability to decide, if you actually read OpenAI, they actually, you have to give credit where credit is due. They publish what's called their safety card or safety report. In other words, to your earlier comments on like, hey, we red teamed this thing, right? We tried to break it and here's what it did. [40:05] OpenAI published this, right? And with their most recent model in 01, what they found is basically they did a test. They gave it a task to complete. I'm not going to get into the technical specifications, but they gave it a task to complete and they wanted to see how would the model complete this task. [40:35] do it successfully. [40:36] that's that that is the technology that exists that's crazy that's like some like sci-fi stuff and here's the thing it's not going to show up on twitter like because you're not going to see it because it's like 40 pages deep in a research report but i'm like that is the technology that exists today and you can literally read it in the safety report right now so when we talk about safety it's like okay well what are the guardrails to this like who is going to say like yes we approve that no we don't and this goes back to the original piece here of like

[41:04] we [41:05] We need to provide value to the shareholders. What does the shareholder value accrue to is the thing that is going to be most monetizable. What I just shared is actually incredibly monetizable, right? And so that's where it becomes problematic because we have a tension of, well, we can monetize this at scale, but we also need to be incredibly thoughtful about how this might be misused. And so that is the current battle that is being fought right now. Decentralized AI and kind of what I alluded to earlier of like user-owned AI is. [41:34] plays a role in this. [41:36] Absolutely. And being able to both help [41:40] create value for those who are tackling that problem in micro economies, but also do that in a way that is verifiable and do that in a way that is, again, unchained and transparent and equitable, right, as well. But it's a real, it's a real fight. It's a real fight. [41:56] Wow. Jared, thank you so much. What a fascinating conversation. I feel like we could do another 25 podcasts on this subject and we still would just be scratching the surface. So I'm sure we'll have to have you on again to continue this conversation. Where can people learn more about the work that you're doing, the more about the work that Nir's doing? Drop some links here. I will gladly drop some links. So if you go to nir.org, [42:25] slash AI, try to keep the URL. [42:28] simple and straightforward. That is essentially our kind of single repository of like what is happening on NIR specifically, who are we working with, how are we working with them, what are the problems we're trying to solve for. And I think more importantly, we are constantly looking for people to collaborate with, I think in order for this space of decentralized AI to truly

[42:49] break out [42:50] and solve real world problems it's probably not going to be done by us competing against each other it's probably going to be done by us actually collaborating to go against the numerous forces that we just like did a deep dive into uh to meaningfully create change in in the world and so that is our focus both on like the the research side [43:10] working with institutions and folks in the research space, but also folks who are just generally curious or interested as well. I come at this as, you know, I get to sit and talk to founders every single day. [43:25] And so the privilege of that is you listen to problems and you start to hear patterns over time and you start to hear and understand white spaces. So if you're working on something that is really interesting in this space. [43:37] Would love to hear from you. Would love to chat. [43:39] Thank you so much. This is incredible, truly. And just appreciate you. And we'll talk to you again soon. [43:44] Amazing. Thank you. [43:45] *music*