1 00:00:00,560 --> 00:00:05,720 This is Epicenter Episode 532 with guest David Minaj. 2 00:00:19,800 --> 00:00:22,200 Welcome to Epicenter, the show which talks about the 3 00:00:22,200 --> 00:00:24,960 technologies, projects, and people driving decentralization 4 00:00:24,960 --> 00:00:27,960 and the blockchain revolution. I'm Federica Ants and I'm here 5 00:00:27,960 --> 00:00:30,680 with Maher Roy. Today we're speaking with David 6 00:00:30,680 --> 00:00:35,240 Minash who is the Co founder and CEO of Valerie and founding 7 00:00:35,240 --> 00:00:38,840 member of Autonalus. And Autonalus is a funny 8 00:00:38,840 --> 00:00:42,080 project, it kind of. It's an AI slash blockchain 9 00:00:42,080 --> 00:00:47,160 crossover and why that is interesting we'll dive into in 10 00:00:47,160 --> 00:00:50,800 just a second. Let us tell you before about our 11 00:00:50,800 --> 00:00:55,400 sponsors this week though. This episode is brought to you 12 00:00:55,400 --> 00:00:58,080 by Gnosis. Nosis builds decentralized 13 00:00:58,080 --> 00:01:02,080 infrastructure for the Ethereum ecosystem with a rich history 14 00:01:02,080 --> 00:01:06,960 dating back to 2015 and products like Safe Cow Swap or Gnosis 15 00:01:06,960 --> 00:01:11,480 Chain, Nosis combines needs driven development with deep 16 00:01:11,640 --> 00:01:15,440 technical expertise. This year marks the launch of 17 00:01:15,440 --> 00:01:18,800 Gnosis Pay, the world's first decentralized payment network. 18 00:01:19,520 --> 00:01:24,040 With a Gnosis card you can spend self custody crypto at any Visa 19 00:01:24,040 --> 00:01:25,960 accepting merchant around the world. 20 00:01:26,600 --> 00:01:30,040 If you're an individual looking to live more on chain or 21 00:01:30,040 --> 00:01:34,160 business looking to white label the stack, visit gnosispay.com. 22 00:01:35,040 --> 00:01:38,000 There are lots of ways you can join the Gnosis journey. 23 00:01:38,480 --> 00:01:42,120 Drop in the Gnosis Dow Governance form, become a Gnosis 24 00:01:42,120 --> 00:01:46,320 validator with a single GNO token and low cost hardware, or 25 00:01:46,320 --> 00:01:49,800 deploy your product on the EVM compatible and highly 26 00:01:49,800 --> 00:01:54,520 decentralized Gnosis chain. Get started today at Gnosis dot 27 00:01:54,520 --> 00:01:58,320 IO. Chorus One is one of the biggest 28 00:01:58,320 --> 00:02:02,360 node operators globally and help you stake your tokens on 45 plus 29 00:02:02,360 --> 00:02:05,960 networks like Ethereum, Cosmos, Celestia and Dydx. 30 00:02:07,240 --> 00:02:10,720 More than 100,000 delegators stake with Chorus One, including 31 00:02:10,720 --> 00:02:13,200 institutions like Bit, Go and Ledger. 32 00:02:13,640 --> 00:02:17,680 Sticking with Chorus 1 not only gets you the highest years, but 33 00:02:17,680 --> 00:02:21,600 also the most robust security practices and infrastructure 34 00:02:21,840 --> 00:02:24,080 that are usually exclusive for institutions. 35 00:02:25,000 --> 00:02:27,880 You can stake directly to Chorus One's public note from your 36 00:02:27,880 --> 00:02:31,040 wallet, set up a white table note, or use the recently 37 00:02:31,040 --> 00:02:35,800 launched product Opus to stake up to 8000 ETH in a single 38 00:02:35,800 --> 00:02:38,880 transaction. You can even offer high yield 39 00:02:38,880 --> 00:02:42,240 staking to your own customers using their API. 40 00:02:42,520 --> 00:02:45,360 Your assets always remain in your custody so you can have 41 00:02:45,360 --> 00:02:48,960 complete Peace of Mind. Start staking today at Chorus 42 00:02:48,960 --> 00:02:53,160 .1. Hi David, it's a pleasure to 43 00:02:53,160 --> 00:02:55,640 have you on. Yeah, pleasure to be here. 44 00:02:55,640 --> 00:02:58,120 Thanks for having me. Absolutely. 45 00:02:58,440 --> 00:03:03,200 Tell us a little bit about yourself and your background. 46 00:03:04,160 --> 00:03:10,480 Sure. So I came to crypto from a sort 47 00:03:10,480 --> 00:03:13,520 of background in maths and economics. 48 00:03:14,280 --> 00:03:21,640 I did maths undergrads and then really got into applied game 49 00:03:21,640 --> 00:03:24,360 theory. There were some fantastic 50 00:03:24,440 --> 00:03:28,640 courses at UCL where I did that and one thing led to another and 51 00:03:28,640 --> 00:03:32,160 I ended up doing a PhD there and then that. 52 00:03:34,200 --> 00:03:39,040 You know, if you Fast forward quite some time that for me to 53 00:03:39,040 --> 00:03:43,440 discovering that I really like this intersection of game theory 54 00:03:43,440 --> 00:03:50,960 and machine learning which I've done a lot together with and 55 00:03:50,960 --> 00:03:54,280 interest which had sort of grown steadily in crypto and 56 00:03:54,280 --> 00:03:57,880 blockchain. And so I'm working in that space 57 00:03:57,880 --> 00:04:00,360 now for over 5 years, and particularly at this 58 00:04:00,360 --> 00:04:04,720 intersection of crypto and AI. Yeah, super interesting. 59 00:04:04,720 --> 00:04:07,280 Sounds like applied economics and maths. 60 00:04:08,640 --> 00:04:12,440 And yeah, sounds like the ideal background for getting into 61 00:04:12,440 --> 00:04:13,920 crypto. You. 62 00:04:14,000 --> 00:04:17,440 As we said in the intro, you are Co founder Valerie which is a Co 63 00:04:17,440 --> 00:04:24,040 contributor to Autonallers. You kind of Co founded with 64 00:04:24,720 --> 00:04:30,480 someone else also called David and what kind of what motivated 65 00:04:30,480 --> 00:04:34,840 you to kind of Co found this project kind of What's the 66 00:04:34,840 --> 00:04:36,520 problem you were setting out to solve? 67 00:04:37,280 --> 00:04:38,880 Yeah, it's a good, it's a great question. 68 00:04:38,880 --> 00:04:44,680 So Valerie's mission is to basically create open source 69 00:04:44,680 --> 00:04:51,240 software for people to Co own primarily agentic AI and we'll 70 00:04:51,240 --> 00:04:53,560 kind of uncover but what we mean by this. 71 00:04:54,400 --> 00:04:59,800 I actually have two Co founders 1 David Galindo has a background 72 00:04:59,800 --> 00:05:03,360 in cryptography. What's a cryptographer and so is 73 00:05:04,480 --> 00:05:08,320 and the other Co founder has a pseudonym called Oak Sprout 74 00:05:08,320 --> 00:05:12,840 Bataan. He has a product background and 75 00:05:12,840 --> 00:05:17,760 the three of us really kind of in different ways we're excited 76 00:05:17,760 --> 00:05:21,520 about autonomous agents and this general pressure which you see 77 00:05:21,520 --> 00:05:27,400 in AI towards agentic kind of AI systems. 78 00:05:28,000 --> 00:05:33,600 And we had different experiences with this topic, different 79 00:05:33,840 --> 00:05:38,560 insights and we came together to basically build a substrate on 80 00:05:38,560 --> 00:05:44,880 which you can as groups kind of Co own these agentic AI systems. 81 00:05:45,560 --> 00:05:49,400 So I think that's sort of the driving force is to provide this 82 00:05:49,400 --> 00:05:52,120 kind of software set which allows people to do that and 83 00:05:52,120 --> 00:05:55,840 also kind of create applications which people then can own in 84 00:05:55,840 --> 00:06:00,800 that way. You've used the term agentic AI 85 00:06:00,800 --> 00:06:07,240 system quite a few times there. What is an Agentic AI system? 86 00:06:08,800 --> 00:06:11,680 The way I think about it is that if you look at the sort of 87 00:06:11,680 --> 00:06:17,640 dominant forms of AI then you can sort of see maybe three 88 00:06:17,640 --> 00:06:19,720 ways. So you have like in the other 89 00:06:19,720 --> 00:06:23,200 parts of the last century like this dominance of rules based 90 00:06:23,200 --> 00:06:26,520 systems. And then and then they basically 91 00:06:26,520 --> 00:06:29,640 said you know that you have hard coded rules, often extremely 92 00:06:29,640 --> 00:06:33,640 sophisticated, which allow you to build sort of certain types 93 00:06:33,640 --> 00:06:36,280 of AI systems. And by no means has this kind of 94 00:06:37,560 --> 00:06:40,920 part of AI research and applications gone away. 95 00:06:41,560 --> 00:06:44,000 But at some point you then had more of these kind of learning 96 00:06:44,000 --> 00:06:46,760 systems emerge where you have like neural networks and deep 97 00:06:47,720 --> 00:06:51,120 learning and other forms of learning, reinforcement 98 00:06:51,120 --> 00:06:56,800 learning, where you effectively use data to construct part of 99 00:06:56,800 --> 00:07:00,400 the algorithm effectively, right. 100 00:07:00,400 --> 00:07:03,720 So the the system learns from data rather than all the rules 101 00:07:03,720 --> 00:07:07,720 being prescribed. And if you look at what's now 102 00:07:07,720 --> 00:07:12,080 happening as we have these very powerful large language models 103 00:07:12,080 --> 00:07:18,200 and other types of powerful AI models, but they by themselves 104 00:07:19,680 --> 00:07:24,120 are certainly now not, not authentic in the sense that 105 00:07:24,120 --> 00:07:28,440 they, you know, there's some some data which sits somewhere 106 00:07:28,440 --> 00:07:32,800 and then you effectively query these models, you instantiate 107 00:07:32,800 --> 00:07:35,560 them, you query them and then you get a response. 108 00:07:35,560 --> 00:07:37,840 And that can be a very sophisticated response, but 109 00:07:37,840 --> 00:07:41,800 that's it. What what's interesting is once 110 00:07:41,800 --> 00:07:48,000 you think about effectively systems that are having agency 111 00:07:48,240 --> 00:07:54,280 and can sort of autonomously act and there's an enormous pressure 112 00:07:54,280 --> 00:07:58,480 towards the system from a pure optimization point of view and 113 00:07:58,640 --> 00:08:00,880 evolution point of view. Like if you think about it, 114 00:08:01,480 --> 00:08:04,240 where can you make the most money? 115 00:08:04,240 --> 00:08:07,640 Where are the most exciting applications as well It's 116 00:08:07,640 --> 00:08:12,000 autonomous systems which can take actions by themselves and 117 00:08:12,000 --> 00:08:16,480 then not, you know maybe instructed by some third party, 118 00:08:16,480 --> 00:08:18,440 some human or or some other system. 119 00:08:18,640 --> 00:08:22,520 And so yeah, these kind of agentic systems we we can 120 00:08:22,520 --> 00:08:25,800 uncover but what what they look like conceptually. 121 00:08:26,440 --> 00:08:29,680 But that's I think where the train is headed, where where a 122 00:08:29,680 --> 00:08:34,159 lot of focus is going towards across like the AI fields. 123 00:08:35,559 --> 00:08:40,280 So David, maybe to kind of let me repackage this somewhat, 124 00:08:40,440 --> 00:08:45,080 would it be fair to say that in the genetic system is something 125 00:08:45,080 --> 00:08:51,080 where you kind of give an AI agent a goal but don't perfectly 126 00:08:51,080 --> 00:08:55,960 specify how it should, how it should go about achieving that 127 00:08:56,240 --> 00:09:01,000 or is that too simplistic? Yeah, I think that works. 128 00:09:01,000 --> 00:09:04,000 And in particular, I mean what we're interested in are sort of 129 00:09:04,000 --> 00:09:08,200 these autonomous agents. And so we can briefly define 130 00:09:08,200 --> 00:09:11,360 that sort of conceptually. So usually it's a sort of 131 00:09:11,360 --> 00:09:15,720 software system which is placed in some environment and from 132 00:09:15,720 --> 00:09:18,440 which it perceives certain information. 133 00:09:20,240 --> 00:09:24,240 They could be blockchain events, literally, or they could be 134 00:09:24,600 --> 00:09:27,440 things from an API. They could be something from a 135 00:09:27,440 --> 00:09:32,920 sensor, which it has locally. It then uses that information 136 00:09:32,920 --> 00:09:36,920 plus whatever its internal architecture looks like, to then 137 00:09:37,000 --> 00:09:40,680 take action again in its environment. 138 00:09:40,680 --> 00:09:45,760 And that environment again can be like a blockchain, another 139 00:09:45,760 --> 00:09:51,200 API, another agent, some actuator of any, any form. 140 00:09:51,200 --> 00:09:53,880 So this is what we would call like an autonomous agent. 141 00:09:54,320 --> 00:09:59,920 And effectively what I'm saying is that what we're seeing 142 00:09:59,920 --> 00:10:03,760 increasingly is that there's more focus to basically create 143 00:10:04,640 --> 00:10:08,640 models which can act as subsystems of such autonomous 144 00:10:08,640 --> 00:10:11,360 agents. Or even like, almost like 145 00:10:11,360 --> 00:10:14,240 Subsume and autonomous agents as a whole, right? 146 00:10:14,240 --> 00:10:17,840 And so there's this kind of pressure towards these kind of 147 00:10:17,840 --> 00:10:21,600 systems. So as an example of an 148 00:10:21,600 --> 00:10:23,640 autonomous agent, maybe we could think of. 149 00:10:25,360 --> 00:10:29,840 So imagine there's a there's diagnosis network and then 150 00:10:29,840 --> 00:10:32,240 there's the code base of diagnosis network. 151 00:10:32,240 --> 00:10:36,800 And one could imagine like a coding agent of some kind where 152 00:10:37,960 --> 00:10:41,280 somebody opens an issue on against the code base of 153 00:10:41,280 --> 00:10:45,080 diagnosis agent or against the code base of the network. 154 00:10:45,080 --> 00:10:48,640 I want to add this feature to the core protocol of diagnosis 155 00:10:48,640 --> 00:10:51,360 network. Then an agent could be something 156 00:10:51,360 --> 00:10:56,040 that kind of as a first step isolates the pieces of code that 157 00:10:56,040 --> 00:11:02,160 need to be changed, as a second step creates code making those 158 00:11:02,160 --> 00:11:05,760 changes. As a third step does some form 159 00:11:05,760 --> 00:11:08,120 of testing. So that could involve kind of 160 00:11:08,120 --> 00:11:11,280 like static analysis, but that could also involve like runtime 161 00:11:11,280 --> 00:11:15,320 analysis and gets feedback from the environment and then makes 162 00:11:15,320 --> 00:11:20,480 another set of changes and comes up with a draft like a draft 163 00:11:20,480 --> 00:11:24,360 change a draft pull request. Yeah, I like that. 164 00:11:25,160 --> 00:11:27,000 I think that's that's a good example. 165 00:11:27,440 --> 00:11:33,640 Also Speaking of Kinosis, Jane, So one autonomous agent which is 166 00:11:33,640 --> 00:11:36,600 running there, one type of autonomous agent which is built 167 00:11:36,600 --> 00:11:39,200 on the autonomous stack and it's running there every day is a 168 00:11:40,120 --> 00:11:42,720 agent which trades in prediction markets. 169 00:11:43,240 --> 00:11:46,240 And so if you kind of map that into this model which I was just 170 00:11:46,240 --> 00:11:47,920 describing, it might be quite helpful. 171 00:11:47,920 --> 00:11:53,200 So again here. What it's observing are sort of 172 00:11:54,000 --> 00:11:56,720 basically new markets. Opening. 173 00:11:56,720 --> 00:12:00,320 So it adds us to the list of markets which it kind of has a 174 00:12:00,320 --> 00:12:06,040 look at. It might then fetch information 175 00:12:06,800 --> 00:12:09,520 pertaining to the events which are referenced in these markets 176 00:12:09,520 --> 00:12:12,960 from really anywhere in the web, so a. 177 00:12:13,320 --> 00:12:17,360 Search API or just like crawling itself almost. 178 00:12:18,240 --> 00:12:23,240 It then uses that information, that context basically, on the 179 00:12:23,240 --> 00:12:27,800 event. As well as various AI models. 180 00:12:27,800 --> 00:12:32,880 At the moment most of the agents use some form of large language 181 00:12:32,880 --> 00:12:38,480 models to basically prompt these models with that kind of 182 00:12:38,480 --> 00:12:44,920 information and then once it arrives at a prediction for that 183 00:12:44,920 --> 00:12:48,200 event together with an accuracy and other kind of information 184 00:12:48,200 --> 00:12:52,920 which it estimates, then sorry confidence wizard estimates, 185 00:12:52,920 --> 00:12:57,960 then it will. Construct A transaction and then 186 00:12:57,960 --> 00:13:01,000 sort of act in that market IE kind of take a position in that 187 00:13:01,000 --> 00:13:02,240 market. So for instance if it's a 188 00:13:02,240 --> 00:13:07,600 binding market, yes or no by the relevant tokens which represent 189 00:13:07,600 --> 00:13:09,720 these events. And so here what you then have 190 00:13:10,200 --> 00:13:13,560 is this environment being sort of these smart contracts and the 191 00:13:14,360 --> 00:13:18,840 information endpoints where which are pertaining to these 192 00:13:18,840 --> 00:13:20,800 markets. And then the actions are these 193 00:13:21,160 --> 00:13:24,400 taking these positions in the markets and then sometime passes 194 00:13:24,400 --> 00:13:28,800 and then? The agent might actually make 195 00:13:28,800 --> 00:13:29,880 some money. Oh, OK. 196 00:13:29,880 --> 00:13:33,240 But that sounds primarily like kind of like automation 197 00:13:33,240 --> 00:13:35,920 technology, right? So basically people wouldn't 198 00:13:35,920 --> 00:13:41,600 necessarily know that I run this sort of software to do things 199 00:13:41,600 --> 00:13:45,640 just like kind of I run for instance, say trading scripts, 200 00:13:45,760 --> 00:13:51,160 right. So how how do we know that this 201 00:13:51,160 --> 00:13:54,200 is, I mean, I assume to some extent this is already 202 00:13:54,200 --> 00:13:57,400 happening, but kind of like where it gets really 203 00:13:57,400 --> 00:13:59,240 interesting. It's kind of when you kind of 204 00:13:59,240 --> 00:14:03,800 design systems where several of these agents kind of come 205 00:14:03,800 --> 00:14:12,120 together to kind of in a game theoretic way to kind of figure 206 00:14:12,120 --> 00:14:19,040 out you know something to do some conclusion or something, 207 00:14:19,040 --> 00:14:21,680 right. So yeah, a couple of things. 208 00:14:21,680 --> 00:14:24,720 So firstly I think. You're right that like there's 209 00:14:24,720 --> 00:14:27,120 sort of automation and then there's different levels of 210 00:14:27,120 --> 00:14:28,680 autonomy. And like if you think about the 211 00:14:28,680 --> 00:14:32,240 self driving car, they have these sort of levels and it's a 212 00:14:32,240 --> 00:14:34,840 bit similar to think here like you have different kind of 213 00:14:34,840 --> 00:14:38,840 levels and you can be closer to what people might describe as 214 00:14:38,840 --> 00:14:41,040 automation. And then there's also this thing 215 00:14:41,040 --> 00:14:46,880 where as time moves on, we tend to prescribe things which maybe 216 00:14:46,880 --> 00:14:49,640 we saw as more autonomous towards automation because they 217 00:14:49,640 --> 00:14:53,520 kind of get wrapped behind like an agent, for instance. 218 00:14:53,520 --> 00:14:57,640 And then I can just sort of see the act of interacting with this 219 00:14:57,640 --> 00:14:59,760 agent where the agent is actually autonomous. 220 00:14:59,760 --> 00:15:02,480 As for me, from from the perspective of the user, it's 221 00:15:02,480 --> 00:15:04,080 almost like just like automation. 222 00:15:04,080 --> 00:15:07,520 I'm just calling this API which then goes away and creates an 223 00:15:07,520 --> 00:15:10,080 outcome for me. And so I think there there's 224 00:15:10,080 --> 00:15:12,360 always. That but. 225 00:15:12,440 --> 00:15:16,520 But you're right and then. In the system which I was 226 00:15:16,520 --> 00:15:19,360 describing, actually the way it's practically implemented 227 00:15:19,360 --> 00:15:22,560 already is that there's already three types of agents today. 228 00:15:22,560 --> 00:15:28,000 So the trading agent itself doesn't actually come up with 229 00:15:28,000 --> 00:15:30,520 the prediction. It's other agents who specialize 230 00:15:30,520 --> 00:15:33,760 on that. And now we're even like picking 231 00:15:33,760 --> 00:15:41,120 apart that role because what we basically see and is that from a 232 00:15:41,120 --> 00:15:44,920 practical perspective if you can sort of specialized your agent 233 00:15:46,120 --> 00:15:49,600 that has its benefits like the same way we specialize as humans 234 00:15:49,600 --> 00:15:56,440 and but also from a sort of practical users perspective of 235 00:15:56,440 --> 00:15:58,800 running the the agent that can have its benefits. 236 00:15:58,800 --> 00:16:02,560 So for instance if I had an agent which has to have all 237 00:16:02,560 --> 00:16:05,960 sorts of let's say open source model which needs to run 238 00:16:05,960 --> 00:16:11,320 alongside it which it uses, then this can become quite obese To 239 00:16:11,320 --> 00:16:15,040 actually run this thing like quite impractical. 240 00:16:15,040 --> 00:16:19,200 Whereas if an agent can use other agents to get something 241 00:16:19,200 --> 00:16:23,800 done, then it might be as simple as making a small crypto payment 242 00:16:23,960 --> 00:16:26,080 to for instance get a prediction. 243 00:16:26,200 --> 00:16:28,840 And so that's the case here. And then you have to obviously 244 00:16:28,840 --> 00:16:32,760 trade that off with other design considerations of the system. 245 00:16:34,880 --> 00:16:37,840 So I'd like to state Frederica's question in like in different 246 00:16:37,840 --> 00:16:43,440 ways. So any standards taking company 247 00:16:43,920 --> 00:16:47,800 would would be running for example price oracles right? 248 00:16:48,280 --> 00:16:51,200 So in in in a price Oracle it's fetching the price from 249 00:16:51,200 --> 00:16:54,560 somewhere and submitting the price to the blockchain and it's 250 00:16:55,200 --> 00:17:00,840 getting paid in crypto to to do that and the one could imagine 251 00:17:00,840 --> 00:17:05,599 that entire. So it's like the the code of 252 00:17:05,599 --> 00:17:09,000 like a price Oracle is is highly mechanical it can it is 253 00:17:09,000 --> 00:17:14,520 specified entirely in in in a in a programming language the input 254 00:17:14,680 --> 00:17:19,319 to it is very structured. It is probably coming as like 255 00:17:19,480 --> 00:17:22,599 Jason files that that are structured in a particular and 256 00:17:22,599 --> 00:17:24,280 it's output is also very structured. 257 00:17:24,280 --> 00:17:27,000 It is producing transactions that have these fields and 258 00:17:27,000 --> 00:17:30,320 etcetera. Perhaps that is like actually 259 00:17:30,320 --> 00:17:34,600 like an agent itself except it's like a very dum agent. 260 00:17:34,840 --> 00:17:39,600 And the kinds of agents you are thinking of are like AI agents 261 00:17:39,600 --> 00:17:44,800 where we are trying to climb the hierarchy of well the inputs no 262 00:17:44,800 --> 00:17:46,600 longer need to be that structured. 263 00:17:46,600 --> 00:17:51,080 It may not come as Jason or or Protobuf or any of these 264 00:17:51,080 --> 00:17:53,280 protocols. It might come as a English 265 00:17:53,280 --> 00:17:55,680 language and it could be anything that comes in. 266 00:17:55,680 --> 00:17:58,040 So the inputs becomes unstructured. 267 00:17:59,040 --> 00:18:03,480 Then the processing logic instead of being structured in 268 00:18:03,480 --> 00:18:10,920 the form of code, you could have processing logic where the agent 269 00:18:11,400 --> 00:18:16,920 like DNO comes up with how to execute on a certain input and 270 00:18:17,400 --> 00:18:20,800 its execution path is kind of like invented for that 271 00:18:20,800 --> 00:18:24,480 particular input and it might be different from what it was 272 00:18:25,320 --> 00:18:28,120 previously. And then finally on the output 273 00:18:28,120 --> 00:18:32,920 side, it's outputs could also be unstructured, meaning it's it's 274 00:18:32,920 --> 00:18:37,560 producing output in the terms of English language which has like 275 00:18:37,720 --> 00:18:40,160 of course with English also has structure which has like but 276 00:18:40,160 --> 00:18:44,640 less structure than a programming language output or a 277 00:18:44,640 --> 00:18:49,000 Jason output would have. And so maybe the financing of 278 00:18:49,040 --> 00:18:53,440 the AI agent is we are trying to generalize the input, the 279 00:18:53,440 --> 00:18:58,920 processing and the output of what what is already kind of 280 00:18:58,920 --> 00:19:03,760 like a traditional crypto agent. So validators, price oracles, we 281 00:19:03,760 --> 00:19:05,720 might think of them as traditional crypto agents, but 282 00:19:05,720 --> 00:19:09,440 we are trying to kind of push their boundaries in like what 283 00:19:09,440 --> 00:19:12,880 they can do. Yeah. 284 00:19:12,960 --> 00:19:15,240 And there's these different dimensions which you're kind of 285 00:19:15,240 --> 00:19:17,840 pointing at, right? So you have like the levels of 286 00:19:17,840 --> 00:19:23,400 autonomy and then the levels also of the kind of how dynamic 287 00:19:23,400 --> 00:19:30,560 is the decision making and how open-ended is it, how structured 288 00:19:30,560 --> 00:19:32,640 and unstructured can be the input and output. 289 00:19:33,200 --> 00:19:37,160 And basically like if you look at it from our perspective, the 290 00:19:37,160 --> 00:19:42,320 way we look at it is our stack kind of allows you to build 291 00:19:43,160 --> 00:19:44,720 across a whole range of these things. 292 00:19:44,720 --> 00:19:49,400 So we have some products which are very, very structured. 293 00:19:49,400 --> 00:19:52,640 So they're basically rules based of the kind which you know in 294 00:19:52,640 --> 00:19:55,640 Oracle is actually one example we we you can build an Oracle on 295 00:19:55,640 --> 00:19:58,840 our stack. It's not like anyone is like 296 00:19:58,840 --> 00:20:02,080 majorly focused on it, but we have some demos of the sort and 297 00:20:02,080 --> 00:20:06,160 then all, you know, you go a bit to the right on that dimension 298 00:20:06,160 --> 00:20:09,760 and then you add this prediction agent where it becomes a bit 299 00:20:09,760 --> 00:20:13,400 less structured because. Yes, some of the flow of it is 300 00:20:13,400 --> 00:20:16,400 entirely structured in the sense that it will always sort of do 301 00:20:16,400 --> 00:20:20,000 certain actions in a certain sequence. 302 00:20:21,000 --> 00:20:22,640 I'll get to. That in a second as to how 303 00:20:22,640 --> 00:20:27,040 that's actually done on a code level and then inside of the 304 00:20:27,040 --> 00:20:29,920 states, actually let me explain it to them right now it's like 305 00:20:29,920 --> 00:20:32,240 we structured as a sort of finite state machine. 306 00:20:32,240 --> 00:20:36,640 So basically we say OK, the the overall agent is described as 307 00:20:36,800 --> 00:20:40,880 this graph like structure where transverses through these States 308 00:20:41,280 --> 00:20:45,160 and then in some of these states it might sort of dynamically 309 00:20:45,160 --> 00:20:48,280 choose which path to take going forward, but sort of the rails 310 00:20:48,280 --> 00:20:50,560 are given, right. So it can't just sort of totally 311 00:20:50,560 --> 00:20:54,280 go off the rails and suddenly say I was a prediction agent now 312 00:20:54,280 --> 00:20:57,600 and now I'm kind of doing this other thing shopping clothes or 313 00:20:57,600 --> 00:21:03,480 whatever. And this we see this as A 314 00:21:04,040 --> 00:21:07,600 basically a pragmatic approach and B also a big advantage 315 00:21:07,600 --> 00:21:14,640 because you know obviously these kind of AI enabled agents, 316 00:21:14,640 --> 00:21:18,680 autonomous agents is something relatively new in that form. 317 00:21:18,840 --> 00:21:21,840 They were stuck in the sort of beltrum for a long time where 318 00:21:22,040 --> 00:21:24,880 basically nothing much happened for for decades and multi agent 319 00:21:25,040 --> 00:21:29,120 systems research, I mean no, no sort of big move forward. 320 00:21:29,120 --> 00:21:34,480 And on the other hand, you now have these sort of AI agent 321 00:21:34,480 --> 00:21:37,160 models based mostly in large language models where it seems a 322 00:21:37,160 --> 00:21:39,120 lot is happening. But then when you dig a bit in, 323 00:21:39,720 --> 00:21:44,320 often if you leave them too unstructured, nothing, it's an 324 00:21:44,320 --> 00:21:47,160 interesting research exercise, but practically not too much 325 00:21:47,160 --> 00:21:50,040 happened. So the sweet spot is still in 326 00:21:50,040 --> 00:21:53,040 between, that's what we say where you provide a certain 327 00:21:53,040 --> 00:21:56,640 degree of structure and then within certain states the agent 328 00:21:56,640 --> 00:22:02,240 can be dealing with unstructured input or output and and and and 329 00:22:02,240 --> 00:22:04,160 can do what you were just describing. 330 00:22:05,760 --> 00:22:10,320 And that I think it's you know how long we will be in this 331 00:22:10,320 --> 00:22:14,200 phase where it's so an in between, I don't know, you know 332 00:22:14,200 --> 00:22:18,880 there's certainly attempts to build sort of like almost like a 333 00:22:19,920 --> 00:22:24,080 large language model but for actions where people sort of 334 00:22:24,080 --> 00:22:27,600 trained us sort of in in into the model itself. 335 00:22:29,480 --> 00:22:33,080 We'll have to see when when they're actually, I think 336 00:22:33,080 --> 00:22:36,040 usable, but if you want to use off the shelf technologies today 337 00:22:36,600 --> 00:22:39,400 and then you're sort of limited to still providing some degree 338 00:22:39,400 --> 00:22:41,920 of structure. The other way to look at this is 339 00:22:41,920 --> 00:22:43,520 also from an efficiency point of view. 340 00:22:43,800 --> 00:22:48,080 So once you actually know that your agent is meant to be an 341 00:22:48,080 --> 00:22:50,800 autonomous agent in prediction markets, that it's meant to make 342 00:22:50,800 --> 00:22:54,680 its money there and that's that you want to use it for that, 343 00:22:55,720 --> 00:22:59,120 then it's kind of pointless is if every time it's running it 344 00:22:59,120 --> 00:23:01,160 has to figure this out from first principles. 345 00:23:01,160 --> 00:23:02,720 That's a very dumb approach, right? 346 00:23:02,720 --> 00:23:05,520 The same way in in in in in in programming. 347 00:23:05,520 --> 00:23:09,480 If I write an efficient program, I might not generate everything 348 00:23:09,480 --> 00:23:11,320 dynamically. I might have like sort of hard 349 00:23:11,320 --> 00:23:15,800 coded, you know, look at tables or whatever where I just pull 350 00:23:15,800 --> 00:23:18,120 values out because they're way more efficient than if I were to 351 00:23:18,120 --> 00:23:20,000 generate them on the fly even if I can. 352 00:23:20,680 --> 00:23:23,160 And so the same thing is here, sometimes you might want to 353 00:23:23,160 --> 00:23:25,320 apply an agent actually at the building stage. 354 00:23:25,320 --> 00:23:30,160 So going back to what you were saying earlier, Meher applying 355 00:23:30,160 --> 00:23:34,400 agents to build agents is also something we are focused on. 356 00:23:34,400 --> 00:23:38,920 So we have like some internal tooling now where we are able to 357 00:23:38,920 --> 00:23:46,760 basically prompt our tooling and then it generates sort of half 358 00:23:46,760 --> 00:23:48,920 of the agent, like not all the code is finished. 359 00:23:48,920 --> 00:23:51,880 There's still like some software developer engagement needed, but 360 00:23:52,280 --> 00:23:56,560 it generates a lot of it. So there's this angle as well. 361 00:23:56,560 --> 00:23:58,320 From an efficiency point of view, where you don't 362 00:23:58,320 --> 00:24:00,600 necessarily always want to figure out everything at 363 00:24:00,600 --> 00:24:04,800 runtime, You might want to sort of ahead of time build a better 364 00:24:04,800 --> 00:24:08,840 agent which is then forced to act within these bounds given by 365 00:24:10,200 --> 00:24:13,720 that design. I'm still a little bit confused 366 00:24:13,920 --> 00:24:17,040 kind of As for the Asian terminology. 367 00:24:17,320 --> 00:24:21,040 So I think kind of there's these cases where kind of I can 368 00:24:21,040 --> 00:24:24,440 imagine you kind of you have large games that you kind of you 369 00:24:24,440 --> 00:24:27,440 optimize for and kind of that means kind of you don't, you 370 00:24:27,440 --> 00:24:30,160 don't have to do so much on chain because kind of you can 371 00:24:30,160 --> 00:24:33,920 optimize it off chain and kind of agents can keep each other in 372 00:24:33,920 --> 00:24:36,800 check, right. And that to me is kind of like 373 00:24:36,800 --> 00:24:43,360 the multi agent system, at least in kind of my lay understanding, 374 00:24:43,760 --> 00:24:47,000 but kind of in, in your description now it sounded like 375 00:24:48,960 --> 00:24:54,280 what I would have conceptualized as one agent, you guys often 376 00:24:54,280 --> 00:24:59,360 think of as different agents that are somehow amalgamated 377 00:24:59,360 --> 00:25:03,560 together into kind of like a super agent as you said, kind of 378 00:25:03,560 --> 00:25:06,160 like there's the prediction, there's kind of the research 379 00:25:06,160 --> 00:25:09,480 agent and the prediction making agent and so on. 380 00:25:09,640 --> 00:25:14,560 Maybe you can kind of delineate the the terms here a little bit 381 00:25:14,560 --> 00:25:19,040 for us. Let me zoom out even like a bit 382 00:25:19,040 --> 00:25:24,480 further. So one of the you know core 383 00:25:24,480 --> 00:25:27,320 things which I guess like the idea of multi agent systems is, 384 00:25:27,320 --> 00:25:30,640 is that you have. Multiple potentially different 385 00:25:30,640 --> 00:25:33,840 types of agents which generate some sort of emergent outcome. 386 00:25:34,000 --> 00:25:37,960 And so if you look at any individual of those agents then 387 00:25:37,960 --> 00:25:41,000 they themselves wouldn't bring about this outcome and then 388 00:25:41,000 --> 00:25:44,320 that's only the collection. So this is the example I was 389 00:25:44,320 --> 00:25:47,200 giving earlier where you have these three types of agents and 390 00:25:47,200 --> 00:25:50,040 then they kind of coming together and the outcome are 391 00:25:50,040 --> 00:25:53,800 sort of AI driven prediction markets where no human ever 392 00:25:53,800 --> 00:25:58,400 participates in. Now if we look at our stack 393 00:25:58,400 --> 00:26:02,840 specifically it gets a bit more interesting still which is that 394 00:26:03,280 --> 00:26:10,440 we basically say OK, going back to this idea of Co owned AI and 395 00:26:10,440 --> 00:26:15,240 Co autonomous agents like what motivates us there? 396 00:26:15,240 --> 00:26:19,080 Well, what motivates us is that we are a bit concerned that as 397 00:26:19,080 --> 00:26:23,560 there's this tremendous pressure to build better AI models and as 398 00:26:23,560 --> 00:26:28,240 there's this tremendous pressure to build better agentic AI 399 00:26:28,240 --> 00:26:32,080 systems, that ultimately they will be owned in a very 400 00:26:32,080 --> 00:26:34,720 centralised way and also operated in a very centralised 401 00:26:34,720 --> 00:26:37,480 way. And so the question is, can you 402 00:26:37,480 --> 00:26:40,680 create basically a substrate where people can own them in a 403 00:26:40,680 --> 00:26:44,240 decentralised way? So now one obvious answer is if 404 00:26:44,320 --> 00:26:47,480 you somehow can make a smart contract smarter. 405 00:26:47,560 --> 00:26:50,120 And there's a lot of exciting projects which are kind of 406 00:26:50,200 --> 00:26:53,920 trying to do that with CKML and other kind of technological 407 00:26:53,920 --> 00:26:58,920 approaches where and effectively you just use a blockchain, a 408 00:26:59,000 --> 00:27:03,840 public one, and you run some code on it which might have been 409 00:27:03,840 --> 00:27:09,160 sort of verified off chain as a verified on servershet might 410 00:27:09,160 --> 00:27:14,320 have been proved off chain. Now in our case what we offer is 411 00:27:14,320 --> 00:27:18,960 basically, OK, if you want to build an autonomous agent and 412 00:27:19,080 --> 00:27:23,560 you then want to run that code as a decentralized system, then 413 00:27:23,560 --> 00:27:26,560 you can do that. So in the Ola stack you can 414 00:27:26,560 --> 00:27:31,080 develop this trader agent which I mentioned earlier and you then 415 00:27:31,080 --> 00:27:33,080 can run it as a multi node system. 416 00:27:33,080 --> 00:27:36,800 So what basically happens is that the trader agent is like 417 00:27:36,800 --> 00:27:41,240 the the whole of all these agent nodes and here it's a bit 418 00:27:41,240 --> 00:27:43,400 different because these agent nodes are effectively like 419 00:27:43,400 --> 00:27:47,480 blockchain nodes. They're sort of replicating the 420 00:27:47,760 --> 00:27:52,960 the work and also the code. They're you know often quite 421 00:27:52,960 --> 00:27:56,640 identical or can even be fully identical instances of each 422 00:27:56,640 --> 00:27:58,920 other. And then they work together to 423 00:27:58,920 --> 00:28:02,640 effectively become the straighter agent and so on chain 424 00:28:02,640 --> 00:28:06,440 they're represented as a multi SEC and off chain there's this 425 00:28:06,440 --> 00:28:10,280 couple of nodes which have like a state synchronisation between 426 00:28:10,280 --> 00:28:12,520 them. So very practically what they 427 00:28:12,520 --> 00:28:16,960 use as tender meant and at the moment as a consensus gadget so 428 00:28:17,000 --> 00:28:23,960 that all the nodes in the system agree on the actions this agent 429 00:28:23,960 --> 00:28:29,840 should take next. So in the field of LLMS itself, 430 00:28:29,840 --> 00:28:35,240 right like and now I'm referring to let's say like the non crypto 431 00:28:35,240 --> 00:28:39,640 part of building on top of LLS which is probably 1000 times 432 00:28:39,640 --> 00:28:44,080 bigger than the the crypto part. There's like lots of different 433 00:28:44,080 --> 00:28:49,600 frameworks that are kind of like building agents using LLM. 434 00:28:49,600 --> 00:28:54,360 So blank chain is probably the most commercially successful. 435 00:28:54,920 --> 00:28:59,240 But then you'll go and find like Microsoft Autogen which which is 436 00:28:59,240 --> 00:29:03,200 a multi agent system in in how it's constructed. 437 00:29:03,880 --> 00:29:06,400 But there are, but there are like loads of others in fact. 438 00:29:06,920 --> 00:29:10,600 In fact the problem is it's a problem of plenty rather than a 439 00:29:10,600 --> 00:29:17,120 problem of problem of scarcity. So maybe to start with, in terms 440 00:29:17,120 --> 00:29:21,320 of like the agent framework you're building, what is like 441 00:29:21,320 --> 00:29:25,880 really different about your agent framework from the things 442 00:29:25,880 --> 00:29:29,440 that might be happening outside crypto as a whole? 443 00:29:31,680 --> 00:29:37,320 Yeah, I think one key thing is that we always want system which 444 00:29:37,320 --> 00:29:41,800 are sort of able to take action on chain like any other users. 445 00:29:42,480 --> 00:29:46,960 So we see like alternative agents as these sort of the 446 00:29:46,960 --> 00:29:49,600 active users of various protocols and we can talk about 447 00:29:49,600 --> 00:29:53,280 this later in a bit what benefits that has for the 448 00:29:53,280 --> 00:29:59,040 protocol, but that means that in our case. 449 00:29:59,160 --> 00:30:04,280 Sort of the crypto wallet and also the on chain representation 450 00:30:04,280 --> 00:30:06,520 of the off chain agent are like first class citizens. 451 00:30:06,520 --> 00:30:09,800 So we think of this from the design beginning and then that 452 00:30:09,800 --> 00:30:12,880 has implications. For instance when we come back 453 00:30:12,880 --> 00:30:16,720 to this trader if I want to Co own like let's say a long chain 454 00:30:16,720 --> 00:30:19,760 agent, well you'd have to basically build what we've built 455 00:30:21,000 --> 00:30:25,840 because you need some way of basically sharing ownership of 456 00:30:25,840 --> 00:30:29,920 let's say an on chain wallet like a a safe let's say a multi 457 00:30:29,920 --> 00:30:33,040 sic with these off chain instances of agent. 458 00:30:33,040 --> 00:30:36,440 So our our framework let's you do this that's that's one way to 459 00:30:36,440 --> 00:30:38,960 look at it. So it's just a sort of native 460 00:30:39,720 --> 00:30:42,680 crypto support. I guess the second thing is 461 00:30:42,680 --> 00:30:47,880 this, if you go further to Co ownership, there's sort of two 462 00:30:49,040 --> 00:30:53,480 extremes there again. So if Co ownership can be 463 00:30:53,480 --> 00:30:56,760 achieved entirely on chain. So for instance you have like a 464 00:30:56,760 --> 00:31:01,680 safe which has some assets and now you have a lot of let's say 465 00:31:01,680 --> 00:31:04,160 land chain agent or auto GPT agent or whatever. 466 00:31:04,320 --> 00:31:08,480 One of those framework agents all kind of holding a wallet and 467 00:31:09,520 --> 00:31:12,760 then being a signer on the safe. Then this could work right 468 00:31:12,760 --> 00:31:15,640 because they don't necessarily need off chain consensus 469 00:31:15,720 --> 00:31:17,160 depending on what the application is. 470 00:31:17,520 --> 00:31:20,280 But actually once you look into the interesting application, 471 00:31:20,280 --> 00:31:23,960 turns out that almost always once you go beyond like simple 472 00:31:23,960 --> 00:31:28,120 things which are done on chain, you need off chain consensus. 473 00:31:28,120 --> 00:31:30,920 Because often it's like things like oh like even on Oracle 474 00:31:30,920 --> 00:31:34,120 needs to agree off chain potentially on the data it wants 475 00:31:34,120 --> 00:31:37,600 to put on chain. Certainly efficient oracles want 476 00:31:37,600 --> 00:31:43,040 to do that off chain and then if you imagine this off chain 477 00:31:43,040 --> 00:31:45,760 system wanting to act upon something else off chain then 478 00:31:45,760 --> 00:31:47,760 for sure you also need off chain consensus. 479 00:31:47,760 --> 00:31:51,240 So there it then also again helps to have a stack which 480 00:31:51,280 --> 00:31:53,200 gives you this out-of-the-box which always does. 481 00:31:53,800 --> 00:31:58,960 Now a third way to look at it and this is sort of purely on 482 00:31:58,960 --> 00:32:02,480 the independent of crypto and more sort of on the structuring 483 00:32:02,480 --> 00:32:06,680 of the agent is that coming back to our discussion earlier, 484 00:32:06,680 --> 00:32:11,280 automation versus autonomy and like sort of fully AI based and 485 00:32:11,280 --> 00:32:18,920 dynamic agent systems rather than those which are maybe like 486 00:32:18,920 --> 00:32:20,760 sort of based on hard coded rules. 487 00:32:21,280 --> 00:32:24,360 The reality is if you want to build like really use cases 488 00:32:24,360 --> 00:32:27,520 which can use, which people can use today and which are actually 489 00:32:27,520 --> 00:32:32,040 meaningfully and securely achieving something, then you 490 00:32:32,040 --> 00:32:36,600 can't go yet in these sort of fully unstructured models where 491 00:32:36,600 --> 00:32:40,960 you just basically repeatedly prompt an LLM like you can do it 492 00:32:40,960 --> 00:32:44,160 but it's it doesn't work. I mean you need to provide 493 00:32:44,160 --> 00:32:46,840 structure and then if you look at the frameworks you know long 494 00:32:46,840 --> 00:32:52,520 chain is an interesting example and and you know I have no, no 495 00:32:52,880 --> 00:32:55,280 nothing bad to say about it but it's an interesting example. 496 00:32:55,280 --> 00:32:57,720 They're moving towards graph structures as well because it's 497 00:32:57,720 --> 00:32:59,720 pretty obvious that a chain won't cut it. 498 00:33:00,160 --> 00:33:02,880 Like your decision making is almost never a chain. 499 00:33:03,200 --> 00:33:07,240 That is the most basic kind of application where it's like ABCD 500 00:33:07,240 --> 00:33:10,360 and then going back to a right. The reality is you're going to 501 00:33:10,360 --> 00:33:13,920 have even in the most basic application you're going to have 502 00:33:13,920 --> 00:33:16,680 the happy path which might look like this, but then off the 503 00:33:16,680 --> 00:33:18,960 happy path you have all these error paths which need to sort 504 00:33:18,960 --> 00:33:20,760 of loop back to different States. 505 00:33:21,200 --> 00:33:23,400 And so you're basically in a graph structure. 506 00:33:24,720 --> 00:33:27,160 And so that's where we started our journey. 507 00:33:27,160 --> 00:33:30,280 We we basically like five years ago said, well, if you're 508 00:33:30,280 --> 00:33:35,480 building autonomous agent systems, then it's unlikely that 509 00:33:35,480 --> 00:33:38,160 we're going to have in the short term these sort of fully 510 00:33:38,160 --> 00:33:41,680 open-ended sort of just models which we need to trend or 511 00:33:41,680 --> 00:33:44,960 somehow the agentic system pops out. 512 00:33:44,960 --> 00:33:50,200 But instead we still need to provide some rails and then use 513 00:33:50,400 --> 00:33:54,280 models alongside those rails. And these rails in our case are 514 00:33:54,280 --> 00:33:58,040 these basic graphs along which the agent has to travel. 515 00:33:58,600 --> 00:34:02,560 And now if you put it back all together, I think one of the 516 00:34:02,560 --> 00:34:05,640 benefits you have our stack is that you can go and say, OK, I 517 00:34:06,240 --> 00:34:10,880 have a A use case where there's some states in which the agent 518 00:34:10,880 --> 00:34:12,880 is very free. There's other states where I 519 00:34:12,880 --> 00:34:16,960 want the agent to just travel along this track, then I can do 520 00:34:16,960 --> 00:34:20,600 this and now I also want this agent to take action on chain 521 00:34:22,239 --> 00:34:24,880 ever so often then it already comes out of the bots. 522 00:34:25,800 --> 00:34:28,280 So we obviously from the beginning when we built the 523 00:34:28,280 --> 00:34:31,320 framework, we're really heavy users of the safe. 524 00:34:31,520 --> 00:34:34,760 So we had like if you're in support with Safe since 525 00:34:34,760 --> 00:34:38,960 basically day one, since the framework is usable and now as 526 00:34:38,960 --> 00:34:42,880 we're sort of expanding it to other sort of types of 527 00:34:44,600 --> 00:34:48,320 blockchain ecosystems, we're always kind of having the same 528 00:34:48,320 --> 00:34:52,320 design podium again where we pick like a multi SEC which is 529 00:34:52,440 --> 00:34:56,520 dominant in this ecosystem and then build the compatibility of 530 00:34:56,520 --> 00:35:00,720 the stack around it. OK, I think I'm I'm now less 531 00:35:00,720 --> 00:35:04,960 confused about the agentic part, but I'm still confused about 532 00:35:04,960 --> 00:35:06,840 kind of like the protocol as a whole. 533 00:35:07,040 --> 00:35:10,080 So kind of like if you look at the stack now, now we kind of 534 00:35:10,080 --> 00:35:13,440 have some understanding of what these agents I can and can't do. 535 00:35:13,440 --> 00:35:16,840 I can't just give an agent, I don't know I can't just say here 536 00:35:16,840 --> 00:35:20,360 you have 100 die, you make me some money and basically the the 537 00:35:20,360 --> 00:35:25,160 agent will go and kind of like either kind of like but but you 538 00:35:25,160 --> 00:35:30,200 know build the arbitration bought or kind of make saucy 539 00:35:30,200 --> 00:35:36,120 pictures on mid on mid journey and kind of put them on only 540 00:35:36,120 --> 00:35:37,560 fans. And I mean, so basically it's 541 00:35:37,560 --> 00:35:40,040 like, it's like you have to give it some structure, I understand 542 00:35:40,040 --> 00:35:44,480 that now, but but how do you put this all into a protocol and 543 00:35:44,480 --> 00:35:46,840 kind of where does the Co ownership come in? 544 00:35:47,040 --> 00:35:50,640 Because this is something that in principle with an LLM model 545 00:35:50,640 --> 00:35:55,800 and like some dev background I could just do on my own, right? 546 00:35:55,800 --> 00:35:57,240 I don't need autonomous for that. 547 00:35:58,960 --> 00:36:02,200 Yeah. And that's a great question. 548 00:36:02,200 --> 00:36:09,600 So basically I think for like one of our core insights is that 549 00:36:09,600 --> 00:36:13,120 it's not about building individual agents, it's about 550 00:36:13,120 --> 00:36:18,520 building effectively many agentic systems which can 551 00:36:18,520 --> 00:36:20,480 interact. Because ultimately we're like 552 00:36:20,680 --> 00:36:23,480 big believers in the specialization and even like 553 00:36:24,440 --> 00:36:26,800 from a very practical point of view, we want to build better 554 00:36:26,800 --> 00:36:28,520 agents. People will build very different 555 00:36:28,520 --> 00:36:31,160 agents, so the framework will have to get get a very 556 00:36:31,160 --> 00:36:32,800 different. Sort of use cases so. 557 00:36:33,360 --> 00:36:37,240 So the protocol was always designed to enable basically 558 00:36:37,240 --> 00:36:40,840 entire Asian economies and enable their bootstrapping. 559 00:36:41,120 --> 00:36:44,520 So there is a couple of mechanisms which facilitate 560 00:36:44,520 --> 00:36:46,600 that. The stack itself is open source. 561 00:36:46,600 --> 00:36:50,680 So when you have an open source stack, there's never a forcing 562 00:36:50,680 --> 00:36:53,960 function to tell them or you have to use this protocol. 563 00:36:54,800 --> 00:37:00,160 So you have to basically create like a a reason on top where why 564 00:37:00,160 --> 00:37:02,760 it would make sense for people to engage with this protocol. 565 00:37:03,360 --> 00:37:08,880 So one thing which we noticed is if you want to have these busy 566 00:37:09,240 --> 00:37:14,040 autonomous agent use cases really grow, then we need 567 00:37:14,040 --> 00:37:18,880 obviously a lot of development, you know, developers who build 568 00:37:19,280 --> 00:37:22,200 on the stack. Why do developers have the 569 00:37:22,200 --> 00:37:23,600 benefit of building on the stack? 570 00:37:23,600 --> 00:37:26,480 Well, there's some of the technical reasons we mentioned 571 00:37:26,480 --> 00:37:28,760 before, but there's also one of composability. 572 00:37:29,160 --> 00:37:35,800 So we basically have created a very composable framework where 573 00:37:35,960 --> 00:37:40,240 it's not so much about composing arbitrary Python libraries, 574 00:37:40,240 --> 00:37:43,760 which is a focus of a lot of the other frameworks, but where it 575 00:37:43,760 --> 00:37:48,680 is the focus of the stack to compose business logic itself. 576 00:37:50,200 --> 00:37:53,240 And that's particularly with autonomous agents of the current 577 00:37:53,320 --> 00:37:55,640 generation. If you think about it, it's very 578 00:37:55,640 --> 00:37:58,880 important. So if I have like for instance 579 00:37:58,880 --> 00:38:01,240 this trader and at some point it's going to settle a 580 00:38:01,240 --> 00:38:05,120 transaction you might say well that is just a matter of sending 581 00:38:05,120 --> 00:38:07,320 the transaction. Well this is actually not true. 582 00:38:07,680 --> 00:38:11,760 There's around like 20 or 30 states in the finite state 583 00:38:11,760 --> 00:38:14,560 machine which takes care of settling the transaction because 584 00:38:14,560 --> 00:38:16,880 there is like on the happy path whereas things are you need to 585 00:38:16,880 --> 00:38:19,720 come, you need to sign it from all the agents, you need to then 586 00:38:19,720 --> 00:38:21,240 submit it. You don't need to wait for it to 587 00:38:21,240 --> 00:38:24,520 be settled. And if anything goes wrong in 588 00:38:24,520 --> 00:38:26,720 any of those states, the resolution looks different. 589 00:38:28,360 --> 00:38:31,800 Now you don't want a developer to re implement that. 590 00:38:32,080 --> 00:38:37,720 Similarly if you think about things like interacting with 591 00:38:37,760 --> 00:38:43,480 these prediction markets, that might actually be like something 592 00:38:43,480 --> 00:38:45,160 which you might want in another agent. 593 00:38:45,160 --> 00:38:48,840 So being able to kind of compose these things is is is very, very 594 00:38:48,840 --> 00:38:52,400 interesting. And so one big part of the 595 00:38:52,400 --> 00:38:55,960 protocol as a result is this focus on creating a developer 596 00:38:55,960 --> 00:39:00,120 developer incentive mechanism whereby developers get rewards 597 00:39:00,120 --> 00:39:04,960 for contributing these pieces of agents and entire agents into 598 00:39:04,960 --> 00:39:07,400 the stack. So that's this code side of 599 00:39:07,400 --> 00:39:10,240 things. They can do that 600 00:39:10,240 --> 00:39:13,880 permissionlessly. So very practically, you know 601 00:39:13,880 --> 00:39:17,920 you develop the stuff you registered on chain as these 602 00:39:17,920 --> 00:39:21,560 NFTS and then there's a sort of reward system which works sort 603 00:39:21,560 --> 00:39:25,800 of on a on an epoch basis. And on the other side of this 604 00:39:26,040 --> 00:39:33,240 is, is the question of capital. So obviously the developer 605 00:39:33,240 --> 00:39:37,200 rewards come partially from you know emissions, but over the 606 00:39:37,200 --> 00:39:40,880 longer term they will have to come from productive agent 607 00:39:40,880 --> 00:39:43,680 systems which the Dow kind of operates. 608 00:39:43,680 --> 00:39:47,840 I'll get to that in a moment, but even to get you there, 609 00:39:48,280 --> 00:39:51,480 basically you need a bootstrapping mechanism whereby 610 00:39:51,480 --> 00:39:54,680 the people can actually use this OLAS token. 611 00:39:54,720 --> 00:39:56,680 And so that's where bonding comes in. 612 00:39:56,680 --> 00:40:00,840 So whenever the protocol is deployed on a new chain, then 613 00:40:00,840 --> 00:40:04,920 effectively there's a bonding mechanism in place whereby 614 00:40:05,640 --> 00:40:09,600 people who use or believe in the protocol can provide liquidity 615 00:40:11,360 --> 00:40:15,120 in this token and the chains token and then. 616 00:40:15,480 --> 00:40:19,640 Return that LP token to the protocol and receive effectively 617 00:40:20,200 --> 00:40:22,440 all us. And what this does is that 618 00:40:22,440 --> 00:40:26,280 basically you have a very decentralized way of bringing 619 00:40:26,280 --> 00:40:28,120 that utility token. 2 more chains. 620 00:40:28,200 --> 00:40:30,000 Why do you want it on more chains? 621 00:40:30,000 --> 00:40:32,240 Because that's the third bit which is staking. 622 00:40:32,240 --> 00:40:37,480 So once I obviously have like code which does something useful 623 00:40:38,080 --> 00:40:43,360 and now I want to be able to actually operate these agents. 624 00:40:43,440 --> 00:40:46,360 And as we noted before, you can operate them like a 625 00:40:46,360 --> 00:40:48,680 decentralized system. So it ends up looking a bit like 626 00:40:48,680 --> 00:40:52,480 operating a blockchain. So you effectively then have the 627 00:40:52,480 --> 00:40:57,720 staking system where the operators of the nodes in any 628 00:40:57,720 --> 00:41:04,200 given agent can basically earn these staking rewards. 629 00:41:04,760 --> 00:41:09,000 And in order for that to obvious move, it helps when the token is 630 00:41:09,000 --> 00:41:11,200 basically accessible on that chain directly. 631 00:41:11,920 --> 00:41:16,280 So that's sort of the three mechanisms like staking being 632 00:41:16,280 --> 00:41:22,360 the last and then the code capital sort of pair and then we 633 00:41:22,360 --> 00:41:25,600 can dive in there if you want. Yeah. 634 00:41:25,600 --> 00:41:27,360 So actually that's a lot of things, right? 635 00:41:27,360 --> 00:41:32,080 So let's try to recap. So right at the layer of 636 00:41:32,080 --> 00:41:37,360 building the agent, what you're saying is like, OK, your 637 00:41:37,360 --> 00:41:42,440 framework in a sense like there are many frameworks which we can 638 00:41:43,040 --> 00:41:48,280 can be used to build agents. But the differentiation of the 639 00:41:48,280 --> 00:41:54,400 orthogonal AS framework is that it contains the differentiation 640 00:41:54,400 --> 00:41:57,000 is in two dimensions. The first dimension is it 641 00:41:57,000 --> 00:42:03,360 contains components that would make blockchain integration and 642 00:42:03,360 --> 00:42:07,280 blockchain transaction creation easy. 643 00:42:07,680 --> 00:42:11,560 So this could this could involve things like, OK, a blockchain, 644 00:42:12,360 --> 00:42:13,720 an agent. If it needs to interact with a 645 00:42:13,720 --> 00:42:15,600 blockchain, it needs to store a private key. 646 00:42:15,600 --> 00:42:19,320 So maybe it it needs some components for the securing of 647 00:42:19,320 --> 00:42:22,720 private keys, it needs components by which it can read 648 00:42:22,720 --> 00:42:26,080 blockchain data, it it needs components where transactions 649 00:42:26,080 --> 00:42:29,240 can be sent and it can it can figure out that they were 650 00:42:29,240 --> 00:42:32,080 confirmed or not. So there are like some standard 651 00:42:32,080 --> 00:42:36,960 pieces of logic that are use in a lot of different places, maybe 652 00:42:36,960 --> 00:42:40,280 in exchanges, you use it in the hot wallet or things like that, 653 00:42:41,280 --> 00:42:45,160 and you're going to build the standard versions of those 654 00:42:45,160 --> 00:42:47,840 components and integrate them into a framework so a developer 655 00:42:47,840 --> 00:42:50,280 doesn't have to worry about those aspects. 656 00:42:51,200 --> 00:42:56,120 Then the second thing your framework's providing is it is 657 00:42:56,120 --> 00:42:59,760 providing some kind of cognitive architecture. 658 00:43:00,480 --> 00:43:06,680 By that what we might mean is that you want the agent to 659 00:43:06,680 --> 00:43:09,880 basically apply its intelligence, but you want it to 660 00:43:10,400 --> 00:43:14,160 you want to constrain its intelligence in a certain manner 661 00:43:14,160 --> 00:43:18,720 which is like you know, for for a particular problem, always 662 00:43:18,720 --> 00:43:23,120 think, always create a tree and reason through the nodes of a 663 00:43:23,120 --> 00:43:27,960 tree or in this particular problem create a line and like 664 00:43:27,960 --> 00:43:30,200 there are nodes and reason through all of these nodes. 665 00:43:30,200 --> 00:43:35,880 So particular problems might have particular ways of thinking 666 00:43:35,880 --> 00:43:39,560 that if we constrain the agent to think in that particular way, 667 00:43:39,720 --> 00:43:43,120 it will produce better outputs. And so you are you're providing 668 00:43:43,120 --> 00:43:49,920 a way to develop against some of these of these like constraints 669 00:43:49,920 --> 00:43:52,920 right Like so a developer can put these constraints into their 670 00:43:52,920 --> 00:43:56,400 system and then they could they could use it. 671 00:43:57,080 --> 00:44:00,480 Those are on the framework itself and then what you're 672 00:44:00,480 --> 00:44:04,200 saying is actually like the the network itself. 673 00:44:04,200 --> 00:44:09,000 So, so now we jump from like the framework deals with the problem 674 00:44:09,000 --> 00:44:12,600 of how do you build a single agent or how do you build with 675 00:44:12,600 --> 00:44:16,600 two or three agents and they coordinate with each other. 676 00:44:17,160 --> 00:44:20,680 But then you jump to the network level, where it's the problem of 677 00:44:20,720 --> 00:44:26,080 ultimately you want thousands of agents to be to be built. 678 00:44:27,720 --> 00:44:31,640 And there the kinds of problems you're trying to solve are how 679 00:44:31,640 --> 00:44:35,480 to provide developer incentives for the improvement of your 680 00:44:35,520 --> 00:44:39,600 agent framework itself. Yeah, so this is. 681 00:44:39,760 --> 00:44:44,720 So if we zoom out a bit what we ultimately want is a machine to 682 00:44:44,720 --> 00:44:51,160 well even if you zoom out a bit more so the Co ownership of 683 00:44:51,360 --> 00:44:59,720 autonomous agents and agentic AI ends up being I think what if 684 00:44:59,720 --> 00:45:02,480 you think about like Tao like decentralized autonomous 685 00:45:02,480 --> 00:45:07,160 organization is sort of almost like the end state of that. 686 00:45:07,160 --> 00:45:11,120 So if you, if you think about this concept of like some 687 00:45:11,520 --> 00:45:15,080 organization which we own, which is in itself autonomous and 688 00:45:15,080 --> 00:45:18,480 which has the highest degree of decentralization we can achieve, 689 00:45:18,480 --> 00:45:24,720 then ultimately this will be using forms of AI and and be 690 00:45:24,720 --> 00:45:27,200 agentic right by by its definition. 691 00:45:27,200 --> 00:45:32,120 And so the different angle at which to come at this is to say, 692 00:45:32,120 --> 00:45:36,120 OK, how, how can you basically coordinate all the actors which 693 00:45:36,120 --> 00:45:39,640 need to make that happen, right? Because if it's just on chain, 694 00:45:39,640 --> 00:45:42,840 then you're always constrained by what you can do on chain, 695 00:45:42,840 --> 00:45:45,280 right. So if you just have a smart 696 00:45:45,280 --> 00:45:48,480 contract then there's always someone who has to call that 697 00:45:48,480 --> 00:45:50,320 smart contract for something to happen. 698 00:45:50,320 --> 00:45:54,040 And by necessity you will always be limited to what's possible on 699 00:45:54,040 --> 00:45:56,600 chain which I think will always be less than what's possible off 700 00:45:56,600 --> 00:46:00,240 chain. And so in a way that the other 701 00:46:00,240 --> 00:46:03,680 way to look at it is to say how can you create basically a 702 00:46:03,680 --> 00:46:08,320 protocol autonomous which allows the creation of these kind of Co 703 00:46:08,320 --> 00:46:11,880 ownable autonomous agents. And then this means you need to 704 00:46:11,880 --> 00:46:14,440 coordinate a bunch of actors, you need to coordinate those who 705 00:46:14,440 --> 00:46:16,280 are developing them. That's why you have the def 706 00:46:16,280 --> 00:46:20,160 incentive mechanism. You need those who operate them, 707 00:46:20,160 --> 00:46:26,560 which is around staking and you need those who basically provide 708 00:46:26,560 --> 00:46:31,840 this liquidity for the whole system to exist at any given 709 00:46:31,840 --> 00:46:35,480 point, which is the bunders. And so that's kind of what the 710 00:46:35,480 --> 00:46:38,520 roles of the protocol is to coordinate all these actors. 711 00:46:38,520 --> 00:46:41,920 Now obviously it's highly complex. 712 00:46:41,920 --> 00:46:43,840 So we should make it a bit concrete. 713 00:46:43,840 --> 00:46:48,440 If you think about what we had earlier discussed quite a lot 714 00:46:48,440 --> 00:46:52,400 the the trading agent use case with the prediction markets. 715 00:46:52,480 --> 00:46:56,440 And then there's the system called the Max inside of it, 716 00:46:56,680 --> 00:47:00,360 which is the third type of agent which basically just specialises 717 00:47:00,360 --> 00:47:05,840 on making the predictions. And these kind of agents are 718 00:47:05,840 --> 00:47:09,680 basically something which you can imagine running as this 719 00:47:09,680 --> 00:47:13,280 decentralized system which the autonomous style itself then can 720 00:47:13,280 --> 00:47:15,800 own. So you effectively then have a 721 00:47:15,800 --> 00:47:20,480 situation where the autonomous style can provide on an ongoing 722 00:47:20,480 --> 00:47:24,840 basis this kind of off chain system with configurable degrees 723 00:47:24,840 --> 00:47:29,200 of decentralization which offers these services to other agents 724 00:47:29,200 --> 00:47:33,080 in the autonomous ecosystem and then that allows you to sort of 725 00:47:34,120 --> 00:47:37,720 bootstrap this over time, if that makes sense. 726 00:47:39,120 --> 00:47:43,320 Can I think of it like this? That so today we have a few 727 00:47:43,320 --> 00:47:48,520 different chains that are trying to build what I call like puppet 728 00:47:48,520 --> 00:47:51,280 accounts or delegated account control. 729 00:47:51,680 --> 00:47:55,880 Those are those are two like 2 interchangeable words. 730 00:47:55,880 --> 00:48:01,040 But the essential idea behind it is so the Near network is trying 731 00:48:01,040 --> 00:48:04,920 to build this. So Near's idea is that, OK, 732 00:48:04,920 --> 00:48:08,040 there's a blockchain with a set of validators. 733 00:48:08,840 --> 00:48:13,120 And what if this blockchain itself could own a Bitcoin 734 00:48:13,120 --> 00:48:15,280 address? Not only a Bitcoin address, but 735 00:48:15,280 --> 00:48:19,840 another Ethereum address. And so from the perspective of 736 00:48:19,840 --> 00:48:23,680 Bitcoin, it's like a normal address with a private key. 737 00:48:24,440 --> 00:48:30,360 But the private key is actually split into the validator set of 738 00:48:30,360 --> 00:48:34,320 near by some really smart cryptographic protocol. 739 00:48:34,600 --> 00:48:39,200 So Bitcoin thinks this is like a single, it's a normal address, a 740 00:48:39,200 --> 00:48:41,840 single individual, but in reality underneath it is 741 00:48:42,760 --> 00:48:46,960 actually validator set of near that controls that account. 742 00:48:47,320 --> 00:48:50,960 And in a sense that you can say that OK, that in a near network 743 00:48:51,040 --> 00:48:54,760 by itself is kind of like owning this, owning this address on 744 00:48:54,760 --> 00:48:57,040 Bitcoin and this other address on Ethereum. 745 00:48:57,760 --> 00:49:02,240 If you start with that point and then kind of you layer on the 746 00:49:02,240 --> 00:49:08,000 idea that is it possible that OK, that there be a way by which 747 00:49:08,920 --> 00:49:15,880 a network could own not only an address but an address plus a 748 00:49:15,880 --> 00:49:22,160 piece of like running code. And that running code is 1 an 749 00:49:22,160 --> 00:49:25,320 economic agent. That running code is an 750 00:49:25,320 --> 00:49:31,360 autonolous, this framework agent, so it has an address and 751 00:49:31,360 --> 00:49:35,920 it has some kind of like structured and unstructured 752 00:49:35,920 --> 00:49:41,000 logic so you can actually message it, give it tasks and 753 00:49:41,000 --> 00:49:45,120 expect responses. And so autonolous is trying to 754 00:49:45,120 --> 00:49:48,160 do that ultimately. Like how do you have a DAO that 755 00:49:48,160 --> 00:49:55,760 can own an address plus some kind of code and so and it it 756 00:49:55,760 --> 00:49:57,920 owns both of those components together. 757 00:49:58,400 --> 00:50:01,560 And then it can also circle a set of like make money, make 758 00:50:01,560 --> 00:50:06,560 money through it, that that is what you're what you're seeking 759 00:50:06,560 --> 00:50:10,240 to achieve. Well, this this is, yeah. 760 00:50:10,280 --> 00:50:12,400 So we would call this like a protocol on tab. 761 00:50:12,440 --> 00:50:17,560 So it's basically if we go back to this concept which we said 762 00:50:17,560 --> 00:50:21,560 earlier, if you have an existing validator set, so basically you 763 00:50:21,560 --> 00:50:24,360 could say OK well let's just do this all on training, you know, 764 00:50:24,360 --> 00:50:28,320 like let's just somehow modify the chain so it can sort of run 765 00:50:28,320 --> 00:50:30,800 long running tasks. And then you will find very 766 00:50:30,800 --> 00:50:33,440 quickly that there's all these arguments as to why that cannot 767 00:50:33,440 --> 00:50:35,640 work. Like you need an application 768 00:50:35,640 --> 00:50:39,160 specific chain in order to have long running tasks, Because if 769 00:50:39,160 --> 00:50:41,280 you have a public chain it becomes a immediate. 770 00:50:42,040 --> 00:50:44,200 Basically at hack vector for for. 771 00:50:44,200 --> 00:50:47,080 Denial of service. Distribute denial of service 772 00:50:47,080 --> 00:50:51,560 because you can just sort of pre empt future blocks indefinitely 773 00:50:51,600 --> 00:50:54,640 by scheduling tasks for future blocks now. 774 00:50:55,080 --> 00:50:57,920 So effectively whatever nearest during there. 775 00:50:57,920 --> 00:51:02,080 I don't know in too much detail, but there's limits to kind of 776 00:51:02,080 --> 00:51:05,160 putting too much on a block public blockchain which is meant 777 00:51:05,160 --> 00:51:09,720 to run repeatedly or scheduled basically. 778 00:51:10,480 --> 00:51:13,800 So you need to do it on some sort of application specific 779 00:51:13,800 --> 00:51:15,400 layer. And now you could say, OK, well 780 00:51:15,400 --> 00:51:18,720 we can just run some sort of layer two or layer three or 781 00:51:18,720 --> 00:51:22,080 layer N or whatever. And ultimately there it's mostly 782 00:51:22,080 --> 00:51:27,560 about having sort of again an architecture where you can 783 00:51:28,120 --> 00:51:32,000 basically then inherit some degree of security right and 784 00:51:32,480 --> 00:51:34,520 execute some of those instructions. 785 00:51:34,520 --> 00:51:39,360 And in a way I think ultimately you know in in in the future one 786 00:51:39,360 --> 00:51:43,920 day an autonomous service will look quite similar, similar to 787 00:51:43,920 --> 00:51:47,000 an app specific roll up potentially because it will 788 00:51:47,000 --> 00:51:52,320 basically have a lot of degrees of verifiability and it will 789 00:51:52,760 --> 00:51:56,800 potentially even inherit some of the security as a result from 790 00:51:56,800 --> 00:52:01,240 the chains on which it acts. But it will have these more 791 00:52:01,240 --> 00:52:05,800 autonomous a long running tasks here which is executing, which 792 00:52:05,800 --> 00:52:09,040 is different from like a public blockchain where I always need 793 00:52:09,040 --> 00:52:13,560 to basically at any given time offer these blocks which accept 794 00:52:13,560 --> 00:52:17,600 a certain amount of basically bidding into them and then once 795 00:52:17,600 --> 00:52:20,480 they're full, they're full right, I can't guarantee you 796 00:52:20,480 --> 00:52:23,240 that I'll execute you, whereas an autonomous service can do 797 00:52:23,240 --> 00:52:24,400 that. It can say, well, I'm 798 00:52:24,400 --> 00:52:28,760 application specific. Ever so often I'm doing exactly 799 00:52:28,760 --> 00:52:34,440 this thing. OK, so I feel like this has 800 00:52:34,440 --> 00:52:40,600 become super abstract. Maybe that's kind of make some 801 00:52:40,600 --> 00:52:45,520 examples, right. So one of the main topics that 802 00:52:45,520 --> 00:52:51,840 kind of you posit this will be used for in the short term is 803 00:52:52,280 --> 00:52:55,840 optimization of Daos. Can you give us some examples 804 00:52:55,920 --> 00:53:00,240 how kind of things work in Daos today and how you see them 805 00:53:00,320 --> 00:53:05,480 improving by kind of putting these autonomous agent systems 806 00:53:06,080 --> 00:53:08,120 on top of them or kind of enmeshing them? 807 00:53:09,920 --> 00:53:12,440 Yeah. So actually there's no answer to 808 00:53:12,440 --> 00:53:15,760 this question in, in the sense that originally before when we 809 00:53:15,760 --> 00:53:18,880 started out with the stack that like Dow's are this primary 810 00:53:18,880 --> 00:53:21,200 custom for that, right. They have various of chain 811 00:53:21,440 --> 00:53:23,520 processes which are often quite centralized. 812 00:53:23,760 --> 00:53:27,960 That's helped them make them more decentralized and and more 813 00:53:27,960 --> 00:53:31,960 autonomous and both things would join their name turns out from a 814 00:53:31,960 --> 00:53:35,240 good to market point of view. And it's not particularly great 815 00:53:35,360 --> 00:53:39,280 because a lot of dials have actually a lot of things to do 816 00:53:39,520 --> 00:53:42,600 and they're maybe not the best organized entities always. 817 00:53:42,600 --> 00:53:46,720 And so it takes a lot of time and you you're not getting to 818 00:53:46,720 --> 00:53:49,160 the goal very fast. You also need to coordinate a 819 00:53:49,160 --> 00:53:51,040 lot of actors by the definition of it. 820 00:53:52,040 --> 00:53:57,000 So actually what we noticed is that what's was we still believe 821 00:53:57,000 --> 00:54:01,160 in this and I'll talk about an example is that it's better to 822 00:54:01,160 --> 00:54:05,160 focus on problems we see in our own Dow and make them as 823 00:54:05,160 --> 00:54:11,160 autonomous and decentralized and or just build basically users 824 00:54:11,200 --> 00:54:13,760 for other decentralized protocols. 825 00:54:13,840 --> 00:54:17,960 So what I mentioned earlier the use case, these autonomous 826 00:54:17,960 --> 00:54:21,520 agents are basically users of Omen, users of Knosis, users of 827 00:54:21,520 --> 00:54:24,080 Safe. You know they have done around 828 00:54:24,080 --> 00:54:28,760 70% of all SAFE transaction on Knosis since summer like on a on 829 00:54:28,960 --> 00:54:32,440 a on a weekly basis basically we've done hundreds of thousands 830 00:54:32,440 --> 00:54:36,040 of transactions which basically benefit these protocols on which 831 00:54:36,040 --> 00:54:38,680 they're deployed and obviously themselves as well because 832 00:54:38,680 --> 00:54:43,640 they're they're profitable. Now an example which I like 833 00:54:44,280 --> 00:54:47,960 because it's very easy to understand, which can apply to 834 00:54:48,080 --> 00:54:51,480 many dials and which they can adopt quite easily as 835 00:54:51,480 --> 00:54:54,280 governator. It was a bit of a joke project 836 00:54:54,360 --> 00:54:56,960 which is sort of slowly maturing. 837 00:54:57,560 --> 00:55:01,960 Basically it's built on the autonomous service stack which 838 00:55:02,160 --> 00:55:04,040 which OLAS offers. Autonomous. 839 00:55:05,800 --> 00:55:09,040 What it does is it's it basically replaces a human 840 00:55:09,640 --> 00:55:12,880 delegatee in in a DA. So if I obviously have tokens 841 00:55:12,880 --> 00:55:15,680 and I don't always want to vote, I could delegate them to someone 842 00:55:15,680 --> 00:55:19,760 I think will vote more or vote in my favourite like with my 843 00:55:19,760 --> 00:55:23,880 intent and so on. And we implemented that in code. 844 00:55:23,880 --> 00:55:26,440 So basically there's an autonomous service which 845 00:55:27,000 --> 00:55:29,920 continuously watches those styles for which it holds 846 00:55:29,960 --> 00:55:34,120 delegated tokens, and then when it sees those proposals either 847 00:55:34,120 --> 00:55:39,360 on snapshot or on chain, it can then vote in those proposals. 848 00:55:39,360 --> 00:55:43,280 And obviously in order to do that, it needs to use a large 849 00:55:43,280 --> 00:55:46,200 language model to actually read the proposal and reason about 850 00:55:46,200 --> 00:55:48,600 it. It also needs that in order to 851 00:55:48,600 --> 00:55:52,120 make sense of the preferences it is given and sort of bring those 852 00:55:52,120 --> 00:55:54,440 two things together to arrive at a voting decision. 853 00:55:54,840 --> 00:55:57,360 But the actual voting, coming back to the structured versus 854 00:55:57,360 --> 00:55:59,480 unstructured is a very structured process. 855 00:55:59,480 --> 00:56:02,640 There's zero point and having the agent figure this out every 856 00:56:02,640 --> 00:56:05,000 time because it will probably fail most of the time. 857 00:56:05,240 --> 00:56:07,600 Instead you just have that part hard coded right. 858 00:56:07,880 --> 00:56:11,200 So basically this is a nice example of what we were 859 00:56:11,200 --> 00:56:13,840 discussing about earlier in very abstract ways. 860 00:56:13,840 --> 00:56:16,800 You have these sort of structured bits which are 861 00:56:16,800 --> 00:56:21,160 defined very well and then you have these unstructured parts of 862 00:56:21,200 --> 00:56:24,720 of the logic where you're looking at this proposals making 863 00:56:24,720 --> 00:56:28,280 sense of it and so on. Are there new attack factors 864 00:56:28,280 --> 00:56:31,480 that are introduced here? So basically, if if I kind of, 865 00:56:32,400 --> 00:56:38,920 if I kind of trust an autonomous agent to kind of make voting 866 00:56:38,920 --> 00:56:44,200 decisions for me, I kind of, I rely heavily on the fact that 867 00:56:44,800 --> 00:56:50,720 this autonomous agent actually will act in the way that I would 868 00:56:50,720 --> 00:56:53,280 act if I were to look into it, right? 869 00:56:53,520 --> 00:57:00,320 So how do you make sure that the agents actually do what they are 870 00:57:00,320 --> 00:57:04,960 meant to do, on the face of it? It's a great question. 871 00:57:04,960 --> 00:57:08,640 So the there's two parts to this. 872 00:57:08,680 --> 00:57:12,840 Well, many, but I would split it into one is like the 873 00:57:12,840 --> 00:57:15,760 preferences. So that's where the Governator 874 00:57:15,760 --> 00:57:17,920 should fall short. It doesn't actually allow you to 875 00:57:17,920 --> 00:57:21,880 express very rich preferences at all and that's just a a matter 876 00:57:21,880 --> 00:57:25,520 of our time and effort which has gone into this part of the 877 00:57:25,520 --> 00:57:28,680 application. But one side where it exceeds on 878 00:57:28,920 --> 00:57:35,720 is the basically certainty that it implements the decisions, the 879 00:57:35,720 --> 00:57:37,920 decision logic which is meant to implement. 880 00:57:37,920 --> 00:57:40,240 So if you think about a human, if you delegate to them, you 881 00:57:40,240 --> 00:57:42,640 basically have no clue, right? It's all reputation based. 882 00:57:43,480 --> 00:57:46,600 If you were to think to delegate to a single long chain agent or 883 00:57:46,600 --> 00:57:50,280 auto GPT agent, well, it really depends on the developer. 884 00:57:50,280 --> 00:57:52,160 Who's running that, Are they even running it? 885 00:57:52,200 --> 00:57:54,400 If they're running it, are they running the code they told you 886 00:57:54,400 --> 00:57:56,080 they're running? Right, all this kind of stuff. 887 00:57:56,560 --> 00:57:59,240 Whereas with an autonomous service which has multiple nodes 888 00:57:59,520 --> 00:58:03,760 operated by different operators, you then start getting into a 889 00:58:03,760 --> 00:58:07,440 similar basically threat model which you have with like a, you 890 00:58:07,440 --> 00:58:09,440 know, like your Cosmos chain basically. 891 00:58:09,760 --> 00:58:15,120 Or any other sort of Byzantine fault tolerant system whereby 892 00:58:15,240 --> 00:58:19,200 you have to reason about OK, how many like operators are there, 893 00:58:19,200 --> 00:58:23,560 how decentralizes it and then is the majority of them honest. 894 00:58:23,560 --> 00:58:28,680 If the majority of them is honest, then you have very high 895 00:58:28,680 --> 00:58:32,200 security guarantees because you effectively what happens is that 896 00:58:32,640 --> 00:58:36,280 each one of them has to agree or the majority of them has to 897 00:58:36,280 --> 00:58:39,080 agree and each one of them uses these models. 898 00:58:39,560 --> 00:58:42,360 So you're not even relying on a single model instance, which is 899 00:58:42,360 --> 00:58:44,360 another issue with large language models, they're not 900 00:58:44,360 --> 00:58:46,080 necessarily deterministic at all. 901 00:58:46,520 --> 00:58:49,800 They sometimes can be configured to be, but like some of them 902 00:58:49,800 --> 00:58:52,720 can't even be configured to be deterministic. 903 00:58:52,760 --> 00:58:57,400 And so then having multiple agents each come to independent 904 00:58:57,400 --> 00:59:00,480 valuations and then sort of pool that decision making and then 905 00:59:00,480 --> 00:59:03,400 agree is actually like a massive improvement. 906 00:59:03,400 --> 00:59:06,680 So on that dimension I would say Governet is already better than 907 00:59:06,680 --> 00:59:09,800 a human, because a human could, you know, do whatever. 908 00:59:09,800 --> 00:59:12,240 And here you have like a node system implementing that 909 00:59:12,360 --> 00:59:14,800 decision logic. Yeah. 910 00:59:14,800 --> 00:59:18,080 So I think kind of what we often try to do in these episodes is 911 00:59:18,120 --> 00:59:21,440 kind of we try to understand how exactly things work. 912 00:59:21,440 --> 00:59:26,280 And I think this was more an episode about kind of talking 913 00:59:26,280 --> 00:59:29,600 about why it would make sense to have something like this. 914 00:59:29,760 --> 00:59:32,880 So kind of I kind of I know I want to change gears a little 915 00:59:32,880 --> 00:59:39,560 bit here and kind of ask about concerns you may have about 916 00:59:39,560 --> 00:59:42,720 this. So kind of like if you look at 917 00:59:42,720 --> 00:59:47,080 AIS the way that they have improved in the last couple of 918 00:59:47,080 --> 00:59:51,360 years at least kind of like in, in the popular mind. 919 00:59:51,360 --> 00:59:54,560 I know that kind of it's been a long time time coming and so on, 920 00:59:54,800 --> 00:59:56,880 but it's really impressive, right. 921 00:59:57,280 --> 01:00:01,960 It seems absolutely certain that they were kind of surpass human 922 01:00:01,960 --> 01:00:06,920 ingenuity and you know, capacity on all kinds of axes in. 923 01:00:07,360 --> 01:00:11,640 You know, the very short term, and if you talk to AI safety 924 01:00:11,640 --> 01:00:19,480 people, it kind of often they will tell you they're not so 925 01:00:19,480 --> 01:00:22,520 concerned because you can always switch it off and now kind of 926 01:00:22,520 --> 01:00:27,400 pairing it with a technology that by definition no one can 927 01:00:27,400 --> 01:00:29,600 turn off. Does that worry you? 928 01:00:31,680 --> 01:00:35,720 Yeah, I think it's a. A good topic to discuss and one 929 01:00:35,720 --> 01:00:39,360 we won't, will not be obviously sad about. 930 01:00:41,600 --> 01:00:46,840 I think the, the first thing which I strongly believe in is 931 01:00:46,840 --> 01:00:49,760 that it's very, very, very unlikely that there will be just 932 01:00:49,760 --> 01:00:53,800 sort of one model which kind of runs away and like takes over. 933 01:00:53,800 --> 01:00:58,360 And that's just even in like very favourable cases to the 934 01:00:58,360 --> 01:01:02,680 sort of super intelligence arising and being able to 935 01:01:02,680 --> 01:01:08,280 consume a lot of resources. There's like geographical 936 01:01:08,280 --> 01:01:12,200 physical sort of constraints which make it unlikely. 937 01:01:13,880 --> 01:01:17,160 I think what's much more likely is that it that we'll have a 938 01:01:17,160 --> 01:01:22,880 situation where certainly a lot of centralized players will own 939 01:01:22,920 --> 01:01:26,800 very, very powerful models. And so I think actually what we 940 01:01:26,800 --> 01:01:31,680 should be what I'm most concerned about is the economic 941 01:01:31,840 --> 01:01:37,920 impact of this kind of change in technology on people rather than 942 01:01:37,920 --> 01:01:42,080 these hypotheticals where some software's lasers all. 943 01:01:42,080 --> 01:01:45,000 I think it you know it's not it's important to kind of keep 944 01:01:45,000 --> 01:01:49,080 it in the back of our minds but and and like with every 945 01:01:49,080 --> 01:01:54,600 technology be mindful as to when these dangers become more 946 01:01:54,600 --> 01:01:56,760 apparent that we kind of think about them. 947 01:01:56,760 --> 01:01:59,960 But like the the much, much bigger concern I think is you 948 01:01:59,960 --> 01:02:04,760 know economic under economics. If if if you hit hit listen to 949 01:02:04,760 --> 01:02:08,800 someone like Sam Altman. It's this naivety of the 950 01:02:08,800 --> 01:02:10,720 economics which really riles me up. 951 01:02:10,720 --> 01:02:15,200 Like they all go around and say, you know, I mean that you know 952 01:02:15,240 --> 01:02:18,600 by all means like they're great like you know entrepreneurs and 953 01:02:18,600 --> 01:02:21,080 create great products and so on. But everyone has the weak ones. 954 01:02:21,080 --> 01:02:24,960 I think here it's like this kind of naivety around just because I 955 01:02:24,960 --> 01:02:27,000 create better technology, everyone will be better. 956 01:02:27,000 --> 01:02:28,560 Well, that never worked out that way. 957 01:02:28,880 --> 01:02:32,040 The reality is that it's always a distribution question. 958 01:02:32,040 --> 01:02:35,960 And if the distribution sacks of access to these kind of models 959 01:02:35,960 --> 01:02:42,000 and people's ability to use them for their lives and improving 960 01:02:42,000 --> 01:02:45,800 their own situation, then it doesn't matter how good the best 961 01:02:45,800 --> 01:02:48,360 model is. Then there will still be even 962 01:02:48,360 --> 01:02:53,040 bigger disparities in sort of income, health, wealth around 963 01:02:53,040 --> 01:02:55,080 the globe. And I think that's what we 964 01:02:55,080 --> 01:02:56,680 should all be really worried about. 965 01:02:56,680 --> 01:02:59,880 And that's kind of the mission of our entire business. 966 01:02:59,880 --> 01:03:02,840 And the mission of autonomous is about creating these kind of 967 01:03:03,160 --> 01:03:06,920 systems which can be Co owned so that there can be groups who can 968 01:03:06,920 --> 01:03:10,160 share these systems. That doesn't mean that all 969 01:03:10,160 --> 01:03:12,680 problems are solved because now you know, these groups could 970 01:03:12,720 --> 01:03:16,120 again be better off than others and you still have these kind of 971 01:03:16,120 --> 01:03:18,600 distributional issues. But at least it's it's a start. 972 01:03:18,760 --> 01:03:21,440 So I'm worried about the economic impact of this much, 973 01:03:21,440 --> 01:03:25,640 much, much, much more than these kind of hypotheticals which I 974 01:03:25,640 --> 01:03:28,240 think are interesting for dinner conversations but really don't 975 01:03:28,880 --> 01:03:32,440 kind of miss the point mostly. Having said that, I think, you 976 01:03:32,440 --> 01:03:35,480 know, let's say we Fast forward, there's like multiple 977 01:03:35,480 --> 01:03:40,920 generations of advances and like even like models which are 978 01:03:40,920 --> 01:03:45,040 basically agent in the model, sort of you know, like some call 979 01:03:45,040 --> 01:03:48,640 them large action models now I saw and others call them 980 01:03:48,640 --> 01:03:54,800 different D then you know Open AI has their reinforcement 981 01:03:54,800 --> 01:03:58,120 learning merged with large language models, there's 982 01:03:58,120 --> 01:04:01,000 different attempts, whatever it will be in the ultimate state. 983 01:04:01,520 --> 01:04:04,960 And if you imagine that to run in a sort of blockchain like way 984 01:04:06,240 --> 01:04:09,600 where it sort of has a bad intent and we can't turn it off, 985 01:04:09,600 --> 01:04:12,600 yes, I think it's something we should keep in the back of our 986 01:04:12,600 --> 01:04:15,840 mind and and think about solutions. 987 01:04:15,840 --> 01:04:22,360 But I think the flip side of this is again that if this model 988 01:04:22,360 --> 01:04:29,760 is used for good, then having transparency and kind of 989 01:04:29,760 --> 01:04:33,320 censorship resistance can bring many goods as well. 990 01:04:33,320 --> 01:04:37,120 So I think let's take it one step at a time basically and 991 01:04:37,120 --> 01:04:39,960 focus on the problems which we for sure know will happen, which 992 01:04:39,960 --> 01:04:43,720 I think are distributional. I feel like we've touched on 993 01:04:43,720 --> 01:04:47,640 many, many things. If people want to learn more or 994 01:04:47,640 --> 01:04:54,000 kind of build their own age genetic systems for autonlers, 995 01:04:54,000 --> 01:04:58,880 or kind of just use systems that are already there, where where 996 01:04:58,880 --> 01:05:01,960 should we send them? Yeah. 997 01:05:01,960 --> 01:05:05,400 So we have this thing called the Academy and that's a great 998 01:05:05,400 --> 01:05:08,880 start. So that's for people who want to 999 01:05:09,120 --> 01:05:11,400 basically have more like support as they're building. 1000 01:05:11,480 --> 01:05:14,720 We have the docs. All of that can be found on the 1001 01:05:14,720 --> 01:05:17,160 website. So Olas dot network and then if 1002 01:05:17,160 --> 01:05:21,720 you follow autonomous on Twitter as well, there's like weekly 1003 01:05:21,720 --> 01:05:26,280 updates where I think those two places are are the best. 1004 01:05:26,480 --> 01:05:29,920 Perfect. I am so curious to see how this 1005 01:05:29,920 --> 01:05:33,800 is going to evolve. I think we should pencil in kind 1006 01:05:33,800 --> 01:05:39,400 of a follow up soonish to just to see kind of like what people 1007 01:05:39,400 --> 01:05:43,680 build and kind of how it actually changes things because 1008 01:05:44,200 --> 01:05:48,640 the opportunity space here is absolutely enormous. 1009 01:05:49,880 --> 01:05:53,760 Yeah, let's do that. Yeah, it's been a pleasure to 1010 01:05:53,760 --> 01:05:55,840 have you on. Thank you very much. 1011 01:05:56,760 --> 01:06:01,360 Was a pleasure being on. Thank you for joining us on this 1012 01:06:01,360 --> 01:06:03,720 week's episode. We release new episodes every 1013 01:06:03,720 --> 01:06:05,760 week. You can find and subscribe to 1014 01:06:05,760 --> 01:06:09,520 the show on iTunes, Spotify, YouTube, SoundCloud, or wherever 1015 01:06:09,520 --> 01:06:11,920 you listen to podcasts. And if you have a Google Home or 1016 01:06:11,920 --> 01:06:14,640 Alexa device, you can tell it. To listen to the latest episode 1017 01:06:14,640 --> 01:06:18,720 of the Epicenter podcast, go to epicenter.tv/subscribe for a 1018 01:06:18,720 --> 01:06:20,400 full list of places where you can watch and listen. 1019 01:06:21,000 --> 01:06:23,320 While you're there, be sure to sign up for the newsletter so 1020 01:06:23,320 --> 01:06:25,640 you get new episodes in your inbox as they're released. 1021 01:06:25,640 --> 01:06:28,920 If you want to interact with us guests or other podcast 1022 01:06:28,920 --> 01:06:31,760 listeners, you can follow us on Twitter and please leave us a 1023 01:06:31,760 --> 01:06:33,760 review on iTunes. It helps people find the show 1024 01:06:34,000 --> 01:06:35,080 and we're always happy to read them. 1025 01:06:35,960 --> 01:06:38,480 So thanks so much and we look forward to being back next week.