WEBVTT 1 00:00:00.080 --> 00:00:03.040 In a world just overflowing with information, wouldn't it be 2 00:00:03.080 --> 00:00:05.960 great to have like a shortcut, a way to really 3 00:00:06.000 --> 00:00:08.320 get a handle on complex topics. 4 00:00:08.400 --> 00:00:11.880 Yeah, cut through the noise and find what actually matters exactly. 5 00:00:11.919 --> 00:00:13.759 And that's what we're aiming for today. We're taking a 6 00:00:13.759 --> 00:00:19.079 deep dive into something truly fascinating distributed artificial intelligence or DAI. 7 00:00:19.399 --> 00:00:21.760 And we've got a great set of resources to work from, 8 00:00:21.920 --> 00:00:26.199 academic chapters, research papers, really cutting edge stuff, all looking 9 00:00:26.239 --> 00:00:31.320 at how AI systems can work together cooperatively in a 10 00:00:31.359 --> 00:00:32.399 decentralized way. 11 00:00:32.520 --> 00:00:34.479 So our mission today is to pull out the key 12 00:00:34.560 --> 00:00:38.359 insights about DAI, what it is, it's building blocks, how 13 00:00:38.399 --> 00:00:42.479 these intelligent bits communicate, and how they're already changing things right. 14 00:00:42.560 --> 00:00:46.159 Healthcare, voting, even managing traffic in smart cities. It's pretty 15 00:00:46.159 --> 00:00:46.719 wide ranging. 16 00:00:46.759 --> 00:00:49.960 Yeah, get ready for some aha moments hopefully. Yeah, this 17 00:00:50.000 --> 00:00:51.119 should be pretty illuminating. 18 00:00:51.200 --> 00:00:51.719 Let's jump in. 19 00:00:51.840 --> 00:00:54.759 Okay, so let's unpack this. When we talk about distributed 20 00:00:54.799 --> 00:01:00.479 AI DAI, it feels different from the typical idea of AI. Yeah, 21 00:01:00.520 --> 00:01:05.000 that single super powerful brain. What's the core idea here? 22 00:01:05.239 --> 00:01:08.319 That's a really good starting point because traditional AI, Yeah, 23 00:01:08.359 --> 00:01:11.079 often it's about one system mimicking human thought for a 24 00:01:11.120 --> 00:01:16.799 specific task. But DAI, it sees AI more like a group, 25 00:01:16.959 --> 00:01:18.879 a collective, a group of what execs, a group of 26 00:01:18.920 --> 00:01:22.599 intelligent agents. Think of them as distinct software entities, each 27 00:01:22.680 --> 00:01:27.280 with some smarts, and they interact, they cooperate, maybe they negotiate. 28 00:01:27.359 --> 00:01:30.120 Sometimes they even compete. Okay, but they're all working within 29 00:01:30.159 --> 00:01:34.040 a larger system. So DAI is really a whole subfield 30 00:01:34.159 --> 00:01:37.239 focused on solving problems in this distributed way. 31 00:01:37.400 --> 00:01:40.079 And why go distributed? Why not just build one? You know? 32 00:01:40.239 --> 00:01:43.840 Really big AI? Is it just about handling more data. 33 00:01:43.920 --> 00:01:47.200 It's more fundamental than just scale, though that's part of it. 34 00:01:47.239 --> 00:01:51.159 Imagine a problem that's just huge or incredibly complex, maybe 35 00:01:51.200 --> 00:01:54.799 too much information for any single system to manage. Effectively 36 00:01:55.319 --> 00:01:58.799 breaking it down into collaborating parts gives you flexibility and modularity. 37 00:01:58.799 --> 00:02:01.040 It's often just much more efficient than trying to build 38 00:02:01.079 --> 00:02:03.400 one giant, monolithic AI. 39 00:02:03.840 --> 00:02:06.159 That makes sense for new systems, but what about all 40 00:02:06.159 --> 00:02:08.039 the tech that's already out there? Are we talking rip 41 00:02:08.120 --> 00:02:11.120 and replace or can DAI integrate? 42 00:02:11.400 --> 00:02:14.080 That's one of the really powerful things about it. DAI 43 00:02:14.319 --> 00:02:18.960 actively considers how to connect with and use existing systems 44 00:02:19.360 --> 00:02:21.680 these legacy systems. 45 00:02:21.240 --> 00:02:25.680 As they're sometimes called, so old databases, existing software stuff 46 00:02:25.719 --> 00:02:27.960 that works, but is an AI smart exactly. 47 00:02:28.039 --> 00:02:30.560 DAI offers ways to build what you could call an 48 00:02:30.599 --> 00:02:33.879 agent rapper or an agent sheath around them. It lets 49 00:02:33.919 --> 00:02:36.439 these older systems join the party, become part of this 50 00:02:36.560 --> 00:02:39.599 intelligent network without needing a complete overhaul. 51 00:02:40.240 --> 00:02:43.840 Okay, like giving your reliable old car a smart navigation 52 00:02:43.960 --> 00:02:45.280 system without swapping the. 53 00:02:45.240 --> 00:02:47.560 Engine precisely, it's a very practical approach. 54 00:02:47.759 --> 00:02:50.520 So that sounds like a massive upgrade for existing infrastructure. 55 00:02:50.560 --> 00:02:53.319 Then what kinds of performance benefits are we talking about? 56 00:02:53.400 --> 00:02:56.159 Well, the sources list quite a few. It improves overall 57 00:02:56.400 --> 00:03:00.000 system performance, computational effectiveness, definitely. 58 00:03:00.159 --> 00:03:03.560 Ability dependability, Yeah, meaning less likely to crash. 59 00:03:03.719 --> 00:03:06.439 Right if one agent fails, the whole system doesn't necessarily 60 00:03:06.439 --> 00:03:12.439 go down. It also enhances extensibility, easier to add new bits, responsiveness, adaptability, 61 00:03:12.479 --> 00:03:15.039 and even the reuse of components you already have. Think 62 00:03:15.039 --> 00:03:17.800 of it like weaving together lots of separate threads into 63 00:03:17.840 --> 00:03:19.759 a really strong, resilient fabric. 64 00:03:19.960 --> 00:03:22.439 And it also sounds perfect for problems that are naturally 65 00:03:22.479 --> 00:03:26.199 spread out, like geographically, or maybe data collected. 66 00:03:25.840 --> 00:03:29.879 Over time exactly right. DAI is particularly good for those 67 00:03:29.919 --> 00:03:34.439 situations where capabilities or data are spatially distributed, spread across 68 00:03:34.479 --> 00:03:38.159 different locations, or temporarily distributed happening at different times. It 69 00:03:38.159 --> 00:03:38.960 connects the dots. 70 00:03:39.120 --> 00:03:41.439 Okay, so we've got the what's and the why. Let's 71 00:03:41.439 --> 00:03:45.560 dig into the who these intelligent agents. What makes an 72 00:03:45.599 --> 00:03:50.159 agent intelligent in this context? Is it just complex code? 73 00:03:50.400 --> 00:03:54.039 Good question, It's not just complexity. Fundamentally, an agent is 74 00:03:54.319 --> 00:03:57.800 a piece of software doing tasks for something or someone else, 75 00:03:57.879 --> 00:04:00.719 could be other software, hardware, or a person. Okay, but 76 00:04:00.800 --> 00:04:04.639 even a basic agent shows some level of autonomy, some intelligence. 77 00:04:05.000 --> 00:04:07.719 But the key properties that really define them In DAI, 78 00:04:07.800 --> 00:04:09.960 there are usually four highlighted. 79 00:04:09.520 --> 00:04:10.400 Right lay them on us. 80 00:04:10.479 --> 00:04:14.159 First, autonomy, they can make decisions based on their environment 81 00:04:14.240 --> 00:04:19.040 without needing constant human handholding. A degree of independence. 82 00:04:18.839 --> 00:04:20.240 Makes sense, they're just puppets. 83 00:04:20.439 --> 00:04:25.240 Second, reactivity They notice and respond to changes happening around them. 84 00:04:25.240 --> 00:04:27.519 They don't just operate in a vacuum. 85 00:04:27.040 --> 00:04:29.360 So they react. But do they plan? 86 00:04:29.680 --> 00:04:33.639 They do. That's the third property. Proactiveness. They can plan ahead, 87 00:04:33.959 --> 00:04:37.839 anticipate future states, and work towards their goals strategically, not 88 00:04:37.920 --> 00:04:38.720 just reactively. 89 00:04:38.920 --> 00:04:42.439 And the fourth, this feels crucial for the distributed part. 90 00:04:42.560 --> 00:04:47.360 Absolutely, it's social ability, the capacity to communicate, coordinate, cooperate, 91 00:04:47.879 --> 00:04:51.399 maybe even negotiate with other agents to achieve shared goals 92 00:04:51.560 --> 00:04:53.319 or at least goals that require interaction. 93 00:04:53.480 --> 00:04:55.800 That's where the real power comes in, isn't it The interaction? 94 00:04:55.959 --> 00:04:58.920 Definitely? And some agents even take it a step further. 95 00:04:59.000 --> 00:05:01.879 They can learn from past experiences, adapt their behavior and 96 00:05:01.920 --> 00:05:05.480 improve over time. We call those evolutionary agents. 97 00:05:05.519 --> 00:05:07.680 And when you get a bunch of these agents working together, 98 00:05:07.879 --> 00:05:10.959 that's a multi agent system or MAS correct. I can 99 00:05:11.000 --> 00:05:14.600 imagine that getting messy fast. If they're all autonomous. How 100 00:05:14.600 --> 00:05:17.399 do you stop it from becoming chaos? 101 00:05:17.519 --> 00:05:21.519 Yeah, or mobocracy as one source puts it, You absolutely 102 00:05:21.600 --> 00:05:24.360 need rules of engagement. Interaction protocols are essential. 103 00:05:24.560 --> 00:05:26.959 Why just to keep order partly. 104 00:05:26.759 --> 00:05:31.040 But also to avoid overwhelming any single agent. If one 105 00:05:31.079 --> 00:05:33.720 agent had to manage all the planning and knowledge sharing, 106 00:05:34.120 --> 00:05:38.600 it defeats the purpose of being distributed. So protocols help 107 00:05:38.680 --> 00:05:42.279 spread that load. Sometimes you even have a special facilitator 108 00:05:42.279 --> 00:05:44.839 agent whose job is just to route messages and keep 109 00:05:44.879 --> 00:05:47.399 track of who can do what, like an air traffic 110 00:05:47.439 --> 00:05:48.600 controller for agents. 111 00:05:48.959 --> 00:05:52.560 So coordination isn't just about avoiding a meltdown. It's about 112 00:05:53.079 --> 00:05:56.680 achieving the overall goal effectively, like making sure everyone stays 113 00:05:56.680 --> 00:05:59.240 within budget or reacts quickly enough precisely. 114 00:05:59.319 --> 00:06:02.519 Groups of age often have to meet these overall constraints 115 00:06:02.800 --> 00:06:06.560 and think about how skills and information are often naturally distributed. 116 00:06:06.160 --> 00:06:07.800 Like your example of the operating. 117 00:06:07.399 --> 00:06:10.879 Room exactly, the cardiologists, the anesthetis the nurse. They all 118 00:06:10.879 --> 00:06:15.279 have different knowledge, different viewpoints, different data feeds. Effective coordination 119 00:06:15.360 --> 00:06:19.199 makes their combined effort successful. It's not just individuals doing task, 120 00:06:19.319 --> 00:06:19.639 it's a. 121 00:06:19.600 --> 00:06:22.000 Team effort, and it can just be faster. Right, even 122 00:06:22.040 --> 00:06:24.560 if one super agent could do at all, sharing the 123 00:06:24.600 --> 00:06:26.480 work and knowledge might get the answer quicker. 124 00:06:26.720 --> 00:06:30.680 Absolutely, efficiency is a major driver, which leads us nicely 125 00:06:30.680 --> 00:06:34.519 into how they actually communicate. What are the mechanisms right? 126 00:06:34.600 --> 00:06:35.439 What are the options? 127 00:06:35.639 --> 00:06:39.959 One classic approach is the blackboard architecture, sometimes called tupple 128 00:06:40.000 --> 00:06:44.199 space communication. Blackboard like in a classroom, kind of imagine 129 00:06:44.240 --> 00:06:48.319 a shared space. The blackboard agents don't talk directly to 130 00:06:48.399 --> 00:06:52.199 each other. Instead, they write information tuples onto the board 131 00:06:52.480 --> 00:06:54.480 and they read information left by others. 132 00:06:54.720 --> 00:06:57.399 Ah, So it's indirect communication exactly. 133 00:06:57.480 --> 00:07:00.639 It's very flexible, great for problems where the loution path 134 00:07:00.680 --> 00:07:05.639 isn't clear upfront, like pattern recognition or forecasting. The Hearsay 135 00:07:05.639 --> 00:07:09.079 two speech system was a famous early example using this okay, 136 00:07:09.319 --> 00:07:12.279 what else? Then you have remote procedure call or RPC. 137 00:07:12.920 --> 00:07:16.639 This is probably more familiar from traditional client server system. 138 00:07:16.519 --> 00:07:19.000 Right where one machine tells another machine to run some code. 139 00:07:19.120 --> 00:07:21.879 Essentially, yes, it lets a process on one node execute 140 00:07:21.879 --> 00:07:24.000 code on another node as if it were a local call. 141 00:07:24.560 --> 00:07:27.279 Makes distributed executions seem more seamless. 142 00:07:27.560 --> 00:07:29.199 Are there different types of RPC? 143 00:07:29.600 --> 00:07:32.360 Yeah, usually grouped into things like batch mode where you 144 00:07:32.439 --> 00:07:36.199 queue up requests, broadcast where you send a request out 145 00:07:36.199 --> 00:07:39.360 to multiple systems hoping for answers, and call back where 146 00:07:39.360 --> 00:07:42.319 the server actually calls back to the client for input 147 00:07:42.439 --> 00:07:43.560 or more processing. 148 00:07:43.800 --> 00:07:46.120 Seems like a more direct approach than the blackboard. 149 00:07:46.600 --> 00:07:49.639 It can be. It's very established now. With all these 150 00:07:49.680 --> 00:07:53.120 agents potentially talking in different ways, people naturally thought about 151 00:07:53.160 --> 00:07:54.920 standardization makes sense. 152 00:07:55.000 --> 00:07:57.279 Try to get everyone speak in the same language exactly. 153 00:07:57.639 --> 00:08:00.879 That led to things like the FEPA standards, the Foundation 154 00:08:01.040 --> 00:08:04.920 for Intelligent Physical Agents. They started in ninety six trying 155 00:08:04.920 --> 00:08:08.720 to create benchmarks so different agent systems could interact smoothly 156 00:08:09.160 --> 00:08:09.920 and did it work. 157 00:08:10.160 --> 00:08:11.439 Is FIPA the standard now? 158 00:08:12.000 --> 00:08:14.800 Well not really. It's a common story in tech standards. 159 00:08:14.920 --> 00:08:17.639 Some platforms adopted them, but FIPA never got the broad 160 00:08:17.680 --> 00:08:21.240 industry buy in they hoped for. The organization dissolved in 161 00:08:21.279 --> 00:08:23.920 two thousand and five, though an IE committee took over 162 00:08:23.959 --> 00:08:27.639 some of the work. It shows how hard standardization can be. 163 00:08:28.079 --> 00:08:32.000 Okay, And the last communication concept you mentioned sounded pretty complex. 164 00:08:32.519 --> 00:08:34.559 Dynamic possible world semantics. 165 00:08:34.919 --> 00:08:38.039 Yeah, that one's a bit more abstract but powerful. Think 166 00:08:38.039 --> 00:08:42.279 about traditional logic. The rules of the world are usually fixed. Yeah, 167 00:08:42.399 --> 00:08:47.039 Dynamic possible world semantics allows the possibilities and the relationships 168 00:08:47.320 --> 00:08:50.480 between agents and their environment to change over time based 169 00:08:50.519 --> 00:08:53.120 on interactions. Can you give you an example, like in chess, 170 00:08:53.480 --> 00:08:56.799 the possibility of castling exists initially, but once the king moves, 171 00:08:56.840 --> 00:09:01.679 that possibility disappears. This semantic framework helps model how agents 172 00:09:01.720 --> 00:09:04.759 reason and adapt in situations where the context, the rules, 173 00:09:04.799 --> 00:09:06.879 the possibilities themselves are evolving. 174 00:09:07.080 --> 00:09:10.799 Wow, okay, that is deep. So these are the concepts 175 00:09:10.799 --> 00:09:14.039 the mechanics. How is DAI actually being used? Where does 176 00:09:14.039 --> 00:09:15.320 it touch the real world? 177 00:09:15.440 --> 00:09:18.360 Oh? It's already in a surprising number of places healthcare 178 00:09:18.440 --> 00:09:19.240 is a massive one. 179 00:09:19.360 --> 00:09:21.960 Right. You mentioned that big data in AI in medicine. 180 00:09:22.039 --> 00:09:23.120 How does DEI fit in? 181 00:09:23.200 --> 00:09:25.679 Well, think about all the different data sources. You've got 182 00:09:25.720 --> 00:09:28.720 social media maybe hinting at disease outbreaks. 183 00:09:28.320 --> 00:09:30.559 Like tracking flu mentions on Twitter exactly. 184 00:09:31.000 --> 00:09:35.080 Then personal health apps, step counters, glucose monitors, mood trackers 185 00:09:35.120 --> 00:09:39.159 constantly generating data, plus genetic services like twenty three a me. 186 00:09:39.440 --> 00:09:41.639 It's an explosion of information. 187 00:09:41.279 --> 00:09:43.159 All coming from different places, different. 188 00:09:42.919 --> 00:09:46.799 Formats precisely, and it all needs storing and processing. Cloud 189 00:09:46.840 --> 00:09:50.519 computing is basically essential here for the scale and flexibility needed. 190 00:09:51.039 --> 00:09:55.080 DAI principles help manage and analyze this distributed data and. 191 00:09:55.080 --> 00:09:57.480 The analysis itself. What kind of insights are they getting? 192 00:09:57.639 --> 00:10:01.039 There's a whole spectrum. Descriptive analytics just telling you what happened, 193 00:10:01.320 --> 00:10:06.840 diagnostic figuring out why, predictive forecasting what might happen, like 194 00:10:06.960 --> 00:10:11.960 predicting patient risk okay, Then prescriptive suggesting actions or treatment options, 195 00:10:12.120 --> 00:10:15.840 and finally cognitive analytics aiming for more human like reasoning 196 00:10:15.919 --> 00:10:17.639 to find really subtle patterns. 197 00:10:17.759 --> 00:10:20.279 That's a lot of complex information. How do doctors or 198 00:10:20.279 --> 00:10:21.720 researchers actually make sense of it. 199 00:10:21.960 --> 00:10:25.879 Visualization is absolutely key. Tools like IBM, Watson and analysis 200 00:10:25.960 --> 00:10:28.759 or others like graphs are side escape. They turn this 201 00:10:28.840 --> 00:10:32.120 flood of data into charts, graphs, networks that humans can 202 00:10:32.159 --> 00:10:34.360 actually understand and use for decision making. 203 00:10:34.759 --> 00:10:38.919 But healthcare data is incredibly sensitive. What about the downsides? 204 00:10:39.399 --> 00:10:41.639 Privacy security? 205 00:10:41.960 --> 00:10:46.320 Huge challenges, absolutely, privacy concerns, data security, the risk of 206 00:10:46.399 --> 00:10:50.320 data loss, just efficiently handling massive files like medical images. 207 00:10:51.159 --> 00:10:54.519 These are all major hurdles that need constant attention and 208 00:10:54.639 --> 00:10:58.879 robust solutions. The potential is huge, but so are the responsibilities. 209 00:10:59.120 --> 00:11:03.159 Okay, shift gears a bit. What about finding information like 210 00:11:03.759 --> 00:11:04.559 search engines? 211 00:11:05.080 --> 00:11:08.600 But smarter Yes, DAI is playing a role in document 212 00:11:08.639 --> 00:11:11.480 and information retrieval too, And it's not just about matching 213 00:11:11.519 --> 00:11:13.480 keywords like a basic database search. 214 00:11:13.559 --> 00:11:14.240 How is it different? 215 00:11:14.360 --> 00:11:18.480 It's about finding semantically relevant information, understanding the meaning behind 216 00:11:18.480 --> 00:11:21.120 your query, and finding documents that relate conceptually, even if 217 00:11:21.159 --> 00:11:23.200 they don't use your exact words. Much more like how 218 00:11:23.320 --> 00:11:24.120 humans look for. 219 00:11:24.080 --> 00:11:27.279 Information, how do they achieve that make it more well relevant? 220 00:11:27.440 --> 00:11:30.919 Various techniques things like stemming reducing words to their root form, 221 00:11:31.039 --> 00:11:35.240 so fishing fish fisher all relate to fish relevance feedback 222 00:11:35.240 --> 00:11:37.240 where you tell the system yes, this result was good 223 00:11:37.360 --> 00:11:40.279 or no, this was bad, and it learns to refine 224 00:11:40.279 --> 00:11:43.799 the search and using the story sometimes built using AI 225 00:11:43.960 --> 00:11:47.879 like genetic algorithms to understand synonyms and related terms. 226 00:11:48.039 --> 00:11:51.759 Interesting, And there's also talk about decentralized search engines themselves. 227 00:11:52.000 --> 00:11:57.200 Yes, that's another angle. Using distributed maybe blockchain based approaches 228 00:11:57.200 --> 00:12:00.480 for search could offer enhanced privacy your searches aren't tracked 229 00:12:00.480 --> 00:12:03.320 by one company. It could allow for more community ownership 230 00:12:03.759 --> 00:12:07.440 and shared access to public data ledgers. Still early days, 231 00:12:07.480 --> 00:12:09.200 but a potentially different model. 232 00:12:09.440 --> 00:12:12.840 Now, what about AI making actual decisions? We see recommendation 233 00:12:12.879 --> 00:12:14.320 systems everywhere, right. 234 00:12:14.159 --> 00:12:16.679 From suggesting news articles or products. 235 00:12:16.240 --> 00:12:18.679 Online too much higher stake stuff. 236 00:12:18.399 --> 00:12:22.240 Exactly, all the way to potentially fully automated systems involved 237 00:12:22.240 --> 00:12:25.440 in things like judicial sentencing guidelines or even aspects of 238 00:12:25.480 --> 00:12:26.600 healthcare decision support. 239 00:12:26.759 --> 00:12:29.360 That sounds impactful. What are the pros and cons? 240 00:12:29.720 --> 00:12:33.399 Well, the pros can be significant reduced costs, increase speed, 241 00:12:33.519 --> 00:12:38.440 potential for greater consistency and objectivity by removing human emotional bias. 242 00:12:38.519 --> 00:12:41.440 Perhaps the the cons a major one is the risk 243 00:12:41.480 --> 00:12:45.320 of introducing new biases, often hidden within the data or algorithms. 244 00:12:45.840 --> 00:12:48.600 If the training data is bias, the AI's decisions will 245 00:12:48.639 --> 00:12:51.279 likely be biased too, and sometimes in ways that are 246 00:12:51.279 --> 00:12:54.639 hard to detect or correct. It's a huge ethical challenge. 247 00:12:54.679 --> 00:12:58.679 So it's not about AI replacing humans in decision making entirely. 248 00:12:59.039 --> 00:13:02.159 Not usually No, Well, the most effective models often see 249 00:13:02.200 --> 00:13:06.919 AI handling the heavy lifting, processing vast data, identifying patterns, 250 00:13:07.399 --> 00:13:13.759 but humans remain crucial for the strategic thinking, the value judgments, understanding, context, ethics, 251 00:13:14.360 --> 00:13:18.360 the uniquely human stuff. It's more about augmenting human capabilities. 252 00:13:18.360 --> 00:13:19.039 Okay, that makes sense. 253 00:13:19.039 --> 00:13:22.679 It's a partnership, a partnership exactly. Now, digging deeper into 254 00:13:22.720 --> 00:13:26.480 how agents actually work together, the sources mentioned the IC architecture. 255 00:13:26.519 --> 00:13:27.320 What's that about? 256 00:13:28.039 --> 00:13:31.879 ICE was specifically designed to help multiple specialist agents communicate 257 00:13:31.960 --> 00:13:35.840 and cooperate. It really focuses on how their internal states, 258 00:13:36.039 --> 00:13:40.360 their information, their intentions change dynamically as they communicate, So. 259 00:13:40.360 --> 00:13:42.919 It's not just passing data. It's about influencing each other's 260 00:13:42.960 --> 00:13:44.799 goals and plans precisely. 261 00:13:45.480 --> 00:13:48.679 It enables that deeper level of teamwork. And to make 262 00:13:48.720 --> 00:13:52.360 sure this all works reliably, they often use formal semantics, 263 00:13:52.919 --> 00:13:57.200 things like Tarski semantics or Cripke possible world semantics, which 264 00:13:57.240 --> 00:13:59.919 connects back to that dynamic possibilities. 265 00:13:59.320 --> 00:14:04.360 Idea we discussed right, providing a solid logical foundation, yeah, frameworks. 266 00:14:03.879 --> 00:14:07.639 For defining truth and possibility in these complex, changing multi 267 00:14:07.639 --> 00:14:08.440 agent worlds. 268 00:14:08.720 --> 00:14:12.600 And this allows for agents pursuing genuinely shared goals, right, 269 00:14:12.759 --> 00:14:15.440 not just individual ones that happen to align exactly. 270 00:14:15.679 --> 00:14:19.080 That requires real mutual responsiveness, commitment to the joint goal, 271 00:14:19.120 --> 00:14:22.240 supporting each other. It's sophisticated cooperation. 272 00:14:21.960 --> 00:14:26.200 And they don't just cooperate randomly. There are often organizational structures. 273 00:14:25.600 --> 00:14:29.240 Involved, yes, just like in human organizations. The structure helps 274 00:14:29.240 --> 00:14:32.759 divide tasks, define roles, coordinate activities. 275 00:14:32.799 --> 00:14:35.240 What are the components of such a structure for agents? 276 00:14:35.320 --> 00:14:39.759 Things like defined obligations, what tasks an agent must do, assets, 277 00:14:39.960 --> 00:14:44.399 what resources like software or hardware they control, information databases, 278 00:14:44.480 --> 00:14:46.679 expertise they possess and tools they. 279 00:14:46.559 --> 00:14:48.840 Can use, and rules about who talks to whom. 280 00:14:49.039 --> 00:14:53.279 Right. Relations like correspondence define the communication channels. An agent 281 00:14:53.360 --> 00:14:56.039 knows where it gets this input and who needs its output. 282 00:14:56.879 --> 00:14:59.440 The sources give an example of a hierarchy for signal 283 00:14:59.440 --> 00:15:03.639 interpretation and sensing nodes feed data up to synthesizing nodes, 284 00:15:03.879 --> 00:15:05.919 which feed up to integrating nodes. 285 00:15:06.279 --> 00:15:08.840 Very structured. Another architecture mentioned is Agora. 286 00:15:09.000 --> 00:15:11.559 What's its niche A gore is interesting. It's a layered 287 00:15:11.600 --> 00:15:15.279 architecture designed for building and running parallel applications across different 288 00:15:15.360 --> 00:15:17.120 kinds of hardware, even virtual machines. 289 00:15:17.320 --> 00:15:18.399 And where is it used? 290 00:15:18.679 --> 00:15:24.720 A key area is intelligent transport systems. Its smart traffic management. Okay, 291 00:15:24.759 --> 00:15:26.639 so how does Agora help with traffic? 292 00:15:26.799 --> 00:15:31.440 It contributes to safety, think collision avoidance, better traffic monitoring efficiency, 293 00:15:31.440 --> 00:15:35.879 optimizing traffic flow, navigation, maybe reducing fuel consumption, and even 294 00:15:35.879 --> 00:15:40.000 commercial stuff infotainment, location based services for cars and pedestrians 295 00:15:40.000 --> 00:15:41.000 with smart devices. 296 00:15:41.279 --> 00:15:43.559 So it's kind of the backbone for some smart city 297 00:15:43.600 --> 00:15:44.720 traffic applications. 298 00:15:45.000 --> 00:15:48.600 Could be Yeah, and interestingly, Agora also has an application 299 00:15:48.679 --> 00:15:52.600 in e commerce. It provides a protocol for bulk transactions 300 00:15:52.720 --> 00:15:57.320 that's designed to be minimal, distributed, secure, and non repudiated. 301 00:15:57.519 --> 00:15:58.919 Non repudiated meaning. 302 00:15:58.840 --> 00:16:01.679 Meaning you can prove who do what. It even includes 303 00:16:01.720 --> 00:16:04.799 things like online arbitration for disputes, and uses a neat 304 00:16:04.879 --> 00:16:08.720 credit based system where actual money doesn't necessarily change hands 305 00:16:08.759 --> 00:16:12.679 for every tiny transaction. Just an account, identify or transfer 306 00:16:13.039 --> 00:16:14.879 makes micro transactions more efficient. 307 00:16:15.399 --> 00:16:19.919 Fascinating now sticking with vehicles, but focusing on security van nets. 308 00:16:20.120 --> 00:16:23.519 These vehicular networks, they seem crucial for safety features. 309 00:16:23.759 --> 00:16:26.360 They are vehicle to vehicle V two V and vehicle 310 00:16:26.360 --> 00:16:31.519 to infrastructure VTI. Communication enables real time warnings about traffic, weather, accidents, 311 00:16:31.799 --> 00:16:33.320 potentially life saving stuff. 312 00:16:33.360 --> 00:16:35.679 Well, they must be a target for attacks absolutely. 313 00:16:35.720 --> 00:16:38.159 The sources highlight a couple the Sibyl. 314 00:16:37.879 --> 00:16:40.480 Attacks sibol like the multiple personalities. 315 00:16:40.240 --> 00:16:43.759 Exactly one attacker creates tons of freak identities on the 316 00:16:43.759 --> 00:16:47.919 network to gain disproportionate influence, maybe outvote honest nodes or 317 00:16:47.960 --> 00:16:52.639 spread misinformation like soft puppet accounts online, but potentially much 318 00:16:52.679 --> 00:16:54.399 more dangerous in a vehicle network. 319 00:16:54.519 --> 00:16:56.919 In the other one eclipse attack. 320 00:16:57.080 --> 00:16:59.600 That's described as a more targeted variant often seen in 321 00:16:59.600 --> 00:17:03.159 blockchain too. Instead of flooding the whole network, an attacker 322 00:17:03.240 --> 00:17:07.519 surrounds a single node, controlling all its connections, effectively isolating it. 323 00:17:07.799 --> 00:17:11.079 They can then manipulate the information that node sees or 324 00:17:11.160 --> 00:17:12.079 sins nasty. 325 00:17:12.279 --> 00:17:14.759 How do you defend vehicle networks against these? 326 00:17:15.000 --> 00:17:19.279 One promising approach mention is using intrusion detection systems IDs 327 00:17:19.480 --> 00:17:22.160 powered by deep neural networks DNNs. 328 00:17:21.880 --> 00:17:24.119 So AI security guards for the car's network. 329 00:17:24.200 --> 00:17:27.599 Pretty much, these systems learn what normal communication on the 330 00:17:27.680 --> 00:17:31.119 vehicle's internal network, like the cambalas looks like, and then 331 00:17:31.160 --> 00:17:34.359 they can flag malicious or abnormal packets that might indicate 332 00:17:34.400 --> 00:17:35.480 an attack is underway. 333 00:17:35.599 --> 00:17:39.920 Okay, this next application feels incredibly relevant today. Secure e 334 00:17:40.039 --> 00:17:42.319 voting using decentralized tech. 335 00:17:42.519 --> 00:17:45.480 Yeah, traditional electronic voting has faced a lot of challenges, 336 00:17:45.519 --> 00:17:46.119 hasn't it. 337 00:17:46.200 --> 00:17:50.960 Definitely Worries about hacking, physical tampering with machines, the cost 338 00:17:51.000 --> 00:17:54.119 of ensuring integrity, maintaining voter anonymity. 339 00:17:55.440 --> 00:17:57.960 It's tough, it really is. And this is where blockchain 340 00:17:57.960 --> 00:18:00.240 technology comes in as a potential solution. 341 00:18:00.079 --> 00:18:01.799 Because it's distributed and hard to tamper with. 342 00:18:02.000 --> 00:18:05.839 Exactly, it offers a public distributed ledger, there's no single 343 00:18:05.839 --> 00:18:09.359 point of failure to attack. Control is spread out once 344 00:18:09.400 --> 00:18:12.759 a vote is recorded. It's essentially immutable, part of an 345 00:18:12.839 --> 00:18:18.440 unchangeable chain, and it relies on consensus mechanisms to validate transactions. 346 00:18:18.720 --> 00:18:22.200 How do you ensure core voting principles like one person, 347 00:18:22.279 --> 00:18:26.759 one vote, secret ballot, making sure only eligible people vote 348 00:18:27.039 --> 00:18:30.480 blockchain seems transparent, which feels counter to secrecy. 349 00:18:30.599 --> 00:18:34.319 It's a clever balancing act. In the proposed designs, anonymity 350 00:18:34.400 --> 00:18:36.559 is preserved because while the vote is recorded on the 351 00:18:36.559 --> 00:18:40.759 public chain, it's cryptographically separated from the voter's actual identity. 352 00:18:41.359 --> 00:18:43.880 Your identity might be verified off chain or through a 353 00:18:43.920 --> 00:18:47.079 secure token, but your choice isn't linked to you publicly. 354 00:18:47.440 --> 00:18:50.759 One person, one vote can be enforced by requiring eligibility 355 00:18:50.839 --> 00:18:55.559 verification first and maybe associating a unique, non reusable token 356 00:18:56.079 --> 00:18:59.319 or even a tiny crypto transaction fee like one ether 357 00:18:59.400 --> 00:19:02.640 in one exam with casting a ballot just enough to 358 00:19:02.680 --> 00:19:05.079 prevent mass fraudulent voting without being. 359 00:19:04.920 --> 00:19:06.680 A real barrier and transparency. 360 00:19:06.720 --> 00:19:09.759 Transparency comes from the fact that the tally is publicly 361 00:19:09.880 --> 00:19:13.279 verifiable on the blockchain. Anyone can check that the votes 362 00:19:13.319 --> 00:19:17.480 were counted correctly according to the cryptographic rules, ensuring reliability 363 00:19:17.519 --> 00:19:19.319 and preventing tampering with the results. 364 00:19:19.480 --> 00:19:21.240 So how would worked for me as a voter? 365 00:19:21.519 --> 00:19:24.680 Typically, an election administrator would set up the election using 366 00:19:24.720 --> 00:19:29.200 a decentralized application DAP that interacts with a smart contract 367 00:19:29.279 --> 00:19:32.839 on the blockchain. You, the voter, would likely interact via 368 00:19:32.920 --> 00:19:36.400 this DAP, perhaps authenticate your eligibility and cast your vote. 369 00:19:36.759 --> 00:19:40.440 The smart contract records the vote anonymously. The administrator might 370 00:19:40.440 --> 00:19:43.079 see that a vote was cast from your district or precinct, 371 00:19:43.079 --> 00:19:44.200 but not how you voted. 372 00:19:44.400 --> 00:19:48.640 It's a really interesting potential application. Okay, so DAI is 373 00:19:48.680 --> 00:19:51.480 clearly being applied in many areas, but it's still evolving. 374 00:19:52.039 --> 00:19:55.160 What about the research side. How do scientists actually test 375 00:19:55.680 --> 00:19:57.279 these multi agent ideas. 376 00:19:57.519 --> 00:20:00.519 That's where test beds come in. The sources men several 377 00:20:00.559 --> 00:20:07.880 specialized platforms, mz adcl one, actalcmec. 378 00:20:06.799 --> 00:20:09.000 Arcas acronyms. To me, what do they do? 379 00:20:09.160 --> 00:20:13.160 Think of them as virtual laboratories or simulators specifically designed 380 00:20:13.160 --> 00:20:17.400 for multi agent systems. They let researchers create complex scenarios, 381 00:20:17.519 --> 00:20:20.960 deploy different types of agents, and experiment with various communication 382 00:20:21.200 --> 00:20:23.279 and coordination strategies. 383 00:20:22.680 --> 00:20:24.559 So they can see what works and what doesn't without 384 00:20:24.559 --> 00:20:26.599 building a whole real world system first. 385 00:20:26.759 --> 00:20:29.720 Exactly, they can test how agents handle errors, how they 386 00:20:29.720 --> 00:20:33.559 communicate under stress, how they adapt to changing environments. It's 387 00:20:33.599 --> 00:20:36.519 crucial for refining the theories and algorithms before trying to 388 00:20:36.519 --> 00:20:40.079 implement them, for say, controlling a power grid or managing 389 00:20:40.079 --> 00:20:41.000 disaster response. 390 00:20:41.160 --> 00:20:43.720 Makes sense So looking at the big picture, what are 391 00:20:43.720 --> 00:20:46.759 the main roadblocks or challenges still facing AI in general 392 00:20:46.920 --> 00:20:48.400 and maybe DII specifically. 393 00:20:48.640 --> 00:20:53.599 There are quite a few persistent ones. Handling unstructured data, text, images, 394 00:20:53.920 --> 00:20:57.559 video remains a huge challenge. It's just messy, right, The 395 00:20:57.640 --> 00:21:01.279 need for continuous training and often human oversight to interpret 396 00:21:01.319 --> 00:21:06.599 results correctly. AI isn't magic. It needs guidance and reality. 397 00:21:06.160 --> 00:21:08.599 Checks and technical limitations. 398 00:21:08.000 --> 00:21:11.680 Sure, risks of hardware failure, the sheer time and processing 399 00:21:11.720 --> 00:21:15.319 power needed for huge data sets and complex models, needing 400 00:21:15.359 --> 00:21:19.720 specialized chips or tons of RAM. Sometimes poor network infrastructure 401 00:21:19.759 --> 00:21:23.440