1 00:00:00,000 --> 00:00:01,080 All right, check this out. 2 00:00:01,080 --> 00:00:04,000 We're going deep on this paper called Magentic One, 3 00:00:04,000 --> 00:00:07,920 a generalist multi-agent system for solving complex tasks. 4 00:00:07,920 --> 00:00:09,240 Oh yeah, I saw that one. 5 00:00:09,240 --> 00:00:11,840 And honestly, the way they're approaching AI 6 00:00:11,840 --> 00:00:13,480 in this is totally different. 7 00:00:13,480 --> 00:00:14,920 Yeah, it kind of flips the script 8 00:00:14,920 --> 00:00:16,840 on how we usually think about AI. 9 00:00:16,840 --> 00:00:17,680 How so? 10 00:00:17,680 --> 00:00:20,400 Well, we usually think of AI as this one big brain 11 00:00:20,400 --> 00:00:21,440 trying to do it all right. 12 00:00:21,440 --> 00:00:22,280 Right, right. 13 00:00:22,280 --> 00:00:24,160 But here, they've built an AI system 14 00:00:24,160 --> 00:00:26,120 that acts more like a team, 15 00:00:26,120 --> 00:00:29,560 like a specialized A team of AI agents all working together. 16 00:00:29,560 --> 00:00:30,920 An AI A team. 17 00:00:30,920 --> 00:00:31,680 Yeah. 18 00:00:31,680 --> 00:00:32,960 That's kind of awesome. 19 00:00:32,960 --> 00:00:33,280 Yeah. 20 00:00:33,280 --> 00:00:34,880 So instead of one giant AI, 21 00:00:34,880 --> 00:00:36,840 we have a bunch of smaller AIs, 22 00:00:36,840 --> 00:00:38,480 each with their own special skill set. 23 00:00:38,480 --> 00:00:39,080 Exactly. 24 00:00:39,080 --> 00:00:41,160 It's like a well-coordinated squad. 25 00:00:41,160 --> 00:00:42,520 OK, I like where this is going. 26 00:00:42,520 --> 00:00:45,720 You've got a web surfer agent that can rise the internet, 27 00:00:45,720 --> 00:00:48,520 a file surfer to handle documents, 28 00:00:48,520 --> 00:00:51,560 a coder that writes code, and a computer terminal 29 00:00:51,560 --> 00:00:52,720 to run that code. 30 00:00:52,720 --> 00:00:54,720 Hold on, there's an AI that can write code 31 00:00:54,720 --> 00:00:57,440 that's both cool and kind of scary at the same time. 32 00:00:57,440 --> 00:00:58,200 Yeah, right. 33 00:00:58,200 --> 00:01:00,360 But why build it this way? 34 00:01:00,360 --> 00:01:03,160 What's the advantage of having a team instead of just one 35 00:01:03,160 --> 00:01:04,760 super AI? 36 00:01:04,760 --> 00:01:06,000 Well, think about it this way. 37 00:01:06,000 --> 00:01:06,560 OK. 38 00:01:06,560 --> 00:01:08,120 Even the smartest person in the world 39 00:01:08,120 --> 00:01:10,640 can't be an expert at everything, right? 40 00:01:10,640 --> 00:01:11,200 True. 41 00:01:11,200 --> 00:01:15,520 But a team of specialists, each focusing on what they do best, 42 00:01:15,520 --> 00:01:18,160 can achieve some pretty amazing things. 43 00:01:18,160 --> 00:01:19,240 OK, that makes sense. 44 00:01:19,240 --> 00:01:19,680 Yeah. 45 00:01:19,680 --> 00:01:22,680 So by breaking down these complex tasks 46 00:01:22,680 --> 00:01:25,280 into smaller, more manageable chunks, 47 00:01:25,280 --> 00:01:27,920 each agent can just focus on their area of expertise. 48 00:01:27,920 --> 00:01:28,360 Right. 49 00:01:28,360 --> 00:01:30,080 It's like a relay race, where each runner 50 00:01:30,080 --> 00:01:32,680 is optimized for their specific leg of the race. 51 00:01:32,680 --> 00:01:33,880 Precisely. 52 00:01:33,880 --> 00:01:36,320 And just like a relay team needs someone 53 00:01:36,320 --> 00:01:39,560 to coordinate the handoffs, this AIA team 54 00:01:39,560 --> 00:01:43,200 has a lead agent called the orchestrator. 55 00:01:43,200 --> 00:01:44,040 The orchestrator. 56 00:01:44,040 --> 00:01:44,560 Yeah. 57 00:01:44,560 --> 00:01:46,640 So that's like the brains of the operation, 58 00:01:46,640 --> 00:01:49,480 figuring out the best strategy, and then assigning tasks 59 00:01:49,480 --> 00:01:50,440 to the right agents. 60 00:01:50,440 --> 00:01:52,760 It's like the coach, making sure everyone knows their role 61 00:01:52,760 --> 00:01:54,040 and plays together effectively. 62 00:01:54,040 --> 00:01:56,040 Gotcha, so the orchestrator is the coach. 63 00:01:56,040 --> 00:01:58,440 But how does it keep track of everything that's going on? 64 00:01:58,440 --> 00:02:00,360 We're talking about a lot of moving parts here. 65 00:02:00,360 --> 00:02:01,960 That's where it gets really interesting. 66 00:02:01,960 --> 00:02:02,920 OK, lay it on me. 67 00:02:02,920 --> 00:02:05,920 The orchestrator uses these things called ledgers. 68 00:02:05,920 --> 00:02:06,520 Ledgers. 69 00:02:06,520 --> 00:02:08,840 Yeah, ledgers to manage the whole process. 70 00:02:08,840 --> 00:02:11,400 The task ledger is like the mission briefing. 71 00:02:11,400 --> 00:02:14,600 It lays out the overall plan, the known facts. 72 00:02:14,600 --> 00:02:17,960 Even educated guesses the AIA team might use. 73 00:02:17,960 --> 00:02:18,960 OK, wait, hold on. 74 00:02:18,960 --> 00:02:19,440 Yeah. 75 00:02:19,440 --> 00:02:20,840 Educated guesses. 76 00:02:20,840 --> 00:02:22,920 Is the AI just making things up? 77 00:02:22,920 --> 00:02:24,080 Not quite. 78 00:02:24,080 --> 00:02:24,440 OK. 79 00:02:24,440 --> 00:02:27,280 Imagine a detective trying to solve a case. 80 00:02:27,280 --> 00:02:29,160 Even with limited information, they 81 00:02:29,160 --> 00:02:31,720 can use their experience and intuition 82 00:02:31,720 --> 00:02:33,400 to develop a theory right. 83 00:02:33,400 --> 00:02:34,520 Yeah. 84 00:02:34,520 --> 00:02:35,960 That's kind of what's happening here. 85 00:02:35,960 --> 00:02:39,440 The AI is making inferences based on the information it has, 86 00:02:39,440 --> 00:02:41,480 even if it's not 100% certain. 87 00:02:41,480 --> 00:02:44,040 So it's using its AI intuition? 88 00:02:44,040 --> 00:02:44,960 Yeah, you could say that. 89 00:02:44,960 --> 00:02:47,360 To fill in the gaps and guide its investigation? 90 00:02:47,360 --> 00:02:48,680 That's pretty clever. 91 00:02:48,680 --> 00:02:52,040 But what about the other ledger, the progress ledger? 92 00:02:52,040 --> 00:02:54,080 Ah, yes, the progress ledger. 93 00:02:54,080 --> 00:02:54,800 What's that do? 94 00:02:54,800 --> 00:02:58,240 That's all about keeping things moving in the right direction. 95 00:02:58,240 --> 00:03:00,600 The progress ledger helps the orchestrator 96 00:03:00,600 --> 00:03:03,240 monitor what's been done, what needs to happen next, 97 00:03:03,240 --> 00:03:05,240 and if anyone is stuck in a rut. 98 00:03:05,240 --> 00:03:07,960 So it's like the orchestrator is constantly checking in, 99 00:03:07,960 --> 00:03:10,680 asking, OK, team, are we making progress? 100 00:03:10,680 --> 00:03:11,840 Is anyone hitting a wall? 101 00:03:11,840 --> 00:03:13,720 Do we need to switch gears? 102 00:03:13,720 --> 00:03:16,640 That seems crucial, especially with complex tasks. 103 00:03:16,640 --> 00:03:17,280 Absolutely. 104 00:03:17,280 --> 00:03:20,160 And to really test this A team approach, 105 00:03:20,160 --> 00:03:23,960 the researchers put Magentic 1 through a series of challenges. 106 00:03:23,960 --> 00:03:24,600 Like what? 107 00:03:24,600 --> 00:03:25,760 They're called benchmarks. 108 00:03:25,760 --> 00:03:26,480 Benchmarks? 109 00:03:26,480 --> 00:03:28,680 Think of them like obstacle courses for AI. 110 00:03:28,680 --> 00:03:30,720 Oh, OK, so they're testing how well 111 00:03:30,720 --> 00:03:32,880 Magentic 1 can handle real world problems. 112 00:03:32,880 --> 00:03:33,520 Right. 113 00:03:33,520 --> 00:03:35,040 What kind of tasks are we talking about here? 114 00:03:35,040 --> 00:03:35,960 Give me an example. 115 00:03:35,960 --> 00:03:39,240 One task from the GAA benchmark was, 116 00:03:39,240 --> 00:03:41,200 of the cities within the United States 117 00:03:41,200 --> 00:03:44,200 where US presidents were born, which two 118 00:03:44,200 --> 00:03:46,640 are the farthest apart from the western most to the eastern 119 00:03:46,640 --> 00:03:49,920 most going east, giving the city names only. 120 00:03:49,920 --> 00:03:51,640 Give them to me in alphabetical order 121 00:03:51,640 --> 00:03:53,160 in a comma separated list. 122 00:03:53,160 --> 00:03:55,000 Whoa, you and I had to read that twice. 123 00:03:55,000 --> 00:03:56,280 It's a mouthful. 124 00:03:56,280 --> 00:03:58,440 So to solve this, the AI would have 125 00:03:58,440 --> 00:04:00,960 to first find all the US presidents and their birth 126 00:04:00,960 --> 00:04:03,720 cities, then figure out the coordinates of those cities, 127 00:04:03,720 --> 00:04:05,840 calculate the distances between them, 128 00:04:05,840 --> 00:04:08,640 and then identify the two farthest apart going eastward, 129 00:04:08,640 --> 00:04:10,200 and list them alphabetically. 130 00:04:10,200 --> 00:04:11,320 You got it. 131 00:04:11,320 --> 00:04:13,040 That's a lot of steps. 132 00:04:13,040 --> 00:04:14,840 And it's not just about finding information. 133 00:04:14,840 --> 00:04:17,880 It's about processing it, making calculations, 134 00:04:17,880 --> 00:04:20,320 and presenting the answer in a specific way. 135 00:04:20,320 --> 00:04:21,280 It's a real challenge. 136 00:04:21,280 --> 00:04:23,000 How did Magentic 1 do? 137 00:04:23,000 --> 00:04:24,320 Pretty well, actually. 138 00:04:24,320 --> 00:04:27,480 It was statistically competitive with other top performing 139 00:04:27,480 --> 00:04:28,320 AI systems. 140 00:04:28,320 --> 00:04:29,000 Wow. 141 00:04:29,000 --> 00:04:32,080 Some of which were specifically designed for just 142 00:04:32,080 --> 00:04:33,480 one of these tasks. 143 00:04:33,480 --> 00:04:36,280 So this generalist approach is really working. 144 00:04:36,280 --> 00:04:37,080 It seems to be. 145 00:04:37,080 --> 00:04:39,080 What's remarkable is that Magentic 1 146 00:04:39,080 --> 00:04:41,360 held its own against these specialists. 147 00:04:41,360 --> 00:04:45,960 So the A team structure is showing real potential 148 00:04:45,960 --> 00:04:48,480 for tackling a wide range of challenges, not just 149 00:04:48,480 --> 00:04:49,640 one specific thing. 150 00:04:49,640 --> 00:04:50,160 Exactly. 151 00:04:50,160 --> 00:04:52,360 But I'm guessing it's not all sunshine and roses. 152 00:04:52,360 --> 00:04:53,480 Of course not. 153 00:04:53,480 --> 00:04:55,760 Every system has its limitations, right? 154 00:04:55,760 --> 00:04:58,320 What are some of the drawbacks the researchers highlighted? 155 00:04:58,320 --> 00:05:00,760 Well, one big one is the potential cost. 156 00:05:00,760 --> 00:05:01,240 OK. 157 00:05:01,240 --> 00:05:04,240 These powerful AI models require a lot of computing power 158 00:05:04,240 --> 00:05:06,480 to run, which can get expensive. 159 00:05:06,480 --> 00:05:07,280 Right, it makes sense. 160 00:05:07,280 --> 00:05:09,880 Plus the coder agent, while promising, 161 00:05:09,880 --> 00:05:11,240 is still pretty basic. 162 00:05:11,240 --> 00:05:11,840 Yeah. 163 00:05:11,840 --> 00:05:15,000 It can write simple programs, but it's nowhere near as 164 00:05:15,000 --> 00:05:17,840 capable as a human programmer. 165 00:05:17,840 --> 00:05:20,640 OK, so there's definitely room for improvement. 166 00:05:20,640 --> 00:05:21,200 For sure. 167 00:05:21,200 --> 00:05:22,960 But even with those limitations, 168 00:05:22,960 --> 00:05:25,840 this idea of a team of AI agents working together 169 00:05:25,840 --> 00:05:27,000 is pretty mind blowing. 170 00:05:27,000 --> 00:05:27,760 Oh, absolutely. 171 00:05:27,760 --> 00:05:30,080 It's like they've created a digital brain trust. 172 00:05:30,080 --> 00:05:33,360 Did the researchers delve into any specific areas 173 00:05:33,360 --> 00:05:36,840 where this team-based approach really shined? 174 00:05:36,840 --> 00:05:37,800 They did. 175 00:05:37,800 --> 00:05:40,320 And this is where it gets really interesting. 176 00:05:40,320 --> 00:05:42,400 They noticed that Magentic 1 was actually 177 00:05:42,400 --> 00:05:45,080 better at handling the harder tasks. 178 00:05:45,080 --> 00:05:45,720 Really? 179 00:05:45,720 --> 00:05:48,000 Yeah, the more steps involved, the more reasoning 180 00:05:48,000 --> 00:05:51,160 and planning required, the better this A team performed. 181 00:05:51,160 --> 00:05:51,960 Interesting. 182 00:05:51,960 --> 00:05:53,960 It kind of goes against intuition, right? 183 00:05:53,960 --> 00:05:56,480 You'd think a specialized AI would be better 184 00:05:56,480 --> 00:05:57,600 at a specific task. 185 00:05:57,600 --> 00:05:58,560 You'd think so, right? 186 00:05:58,560 --> 00:05:59,040 Yeah. 187 00:05:59,040 --> 00:06:01,840 But Magentic 1 even outperformed some specialized 188 00:06:01,840 --> 00:06:03,160 AIs on certain tasks. 189 00:06:03,160 --> 00:06:03,840 Wow. 190 00:06:03,840 --> 00:06:05,600 So the more complex the challenge, 191 00:06:05,600 --> 00:06:07,640 the more this A team structure shines. 192 00:06:07,640 --> 00:06:09,200 Yeah, that's what the data suggests. 193 00:06:09,200 --> 00:06:10,240 Why would that be? 194 00:06:10,240 --> 00:06:13,000 It could be because of how multi-agent systems are 195 00:06:13,000 --> 00:06:13,640 structured. 196 00:06:13,640 --> 00:06:14,160 OK. 197 00:06:14,160 --> 00:06:16,960 Having these specialists and a central coordinator, 198 00:06:16,960 --> 00:06:19,400 like the orchestrator, it really helps 199 00:06:19,400 --> 00:06:22,640 break down these huge problems into smaller pieces. 200 00:06:22,640 --> 00:06:24,160 Yeah, divide and conquer, right? 201 00:06:24,160 --> 00:06:25,160 Exactly. 202 00:06:25,160 --> 00:06:27,840 Makes those complex problems a little more manageable. 203 00:06:27,840 --> 00:06:29,920 But surely they also found some areas 204 00:06:29,920 --> 00:06:33,800 where the A team didn't quite click. 205 00:06:33,800 --> 00:06:34,880 Oh, absolutely. 206 00:06:34,880 --> 00:06:36,600 What did their error analysis reveal? 207 00:06:36,600 --> 00:06:38,960 Well, they dug deep into the logs, 208 00:06:38,960 --> 00:06:42,360 actually seeing how Magentic 1 tackled these tasks. 209 00:06:42,360 --> 00:06:45,320 And they found some opportunities for growth. 210 00:06:45,320 --> 00:06:46,240 Let's just say that. 211 00:06:46,240 --> 00:06:47,600 All right, spill the tea. 212 00:06:47,600 --> 00:06:49,120 What were some of the common hiccups? 213 00:06:49,120 --> 00:06:51,160 One frequent issue is what they called 214 00:06:51,160 --> 00:06:53,680 persistent inefficient actions. 215 00:06:53,680 --> 00:06:54,240 What's that? 216 00:06:54,240 --> 00:06:56,360 Sometimes the agents would get stuck in a loop, 217 00:06:56,360 --> 00:06:58,640 repeating the same action, even though it clearly 218 00:06:58,640 --> 00:06:59,360 wasn't working. 219 00:06:59,360 --> 00:07:00,880 Oh, so like the web surfer keeps 220 00:07:00,880 --> 00:07:03,440 searching for the same term, even though it's finding nothing? 221 00:07:03,440 --> 00:07:04,200 Exactly. 222 00:07:04,200 --> 00:07:07,160 Or the file surfer keeps opening the wrong document. 223 00:07:07,160 --> 00:07:08,520 Yeah, I've been there. 224 00:07:08,520 --> 00:07:11,760 So a key area for improvement is teaching the agents 225 00:07:11,760 --> 00:07:14,680 to recognize when their strategy is failing 226 00:07:14,680 --> 00:07:16,360 and to try something else. 227 00:07:16,360 --> 00:07:19,520 So building in more flexibility and adaptability 228 00:07:19,520 --> 00:07:21,600 make those agents a little more resourceful. 229 00:07:21,600 --> 00:07:22,360 Exactly. 230 00:07:22,360 --> 00:07:23,320 What else did they find? 231 00:07:23,320 --> 00:07:26,960 Another issue was a lack of thorough verification. 232 00:07:26,960 --> 00:07:27,800 What do you mean? 233 00:07:27,800 --> 00:07:29,920 Sometimes the agents would jump to conclusions 234 00:07:29,920 --> 00:07:31,720 without double checking their work. 235 00:07:31,720 --> 00:07:34,400 Oh, overconfident, huh? 236 00:07:34,400 --> 00:07:37,040 Like a student turning in an essay without proofreading. 237 00:07:37,040 --> 00:07:38,240 That's a great analogy. 238 00:07:38,240 --> 00:07:41,200 And that can lead to mistakes, especially when accuracy 239 00:07:41,200 --> 00:07:41,800 is critical. 240 00:07:41,800 --> 00:07:42,640 Yeah, makes sense. 241 00:07:42,640 --> 00:07:44,360 They need to learn to double check their work. 242 00:07:44,360 --> 00:07:47,240 Like a healthy dose of skepticism is needed, I guess. 243 00:07:47,240 --> 00:07:48,280 You could say that. 244 00:07:48,280 --> 00:07:51,680 What other opportunities for growth do they uncover? 245 00:07:51,680 --> 00:07:54,480 Magentic ones sometimes struggled with navigation. 246 00:07:54,480 --> 00:07:55,920 Navigation, like where? 247 00:07:55,920 --> 00:07:56,720 Social media on the web. 248 00:07:56,720 --> 00:07:59,240 Getting lost in the vastness of the internet. 249 00:07:59,240 --> 00:08:01,680 You could say that the web surfer would click around 250 00:08:01,680 --> 00:08:04,440 aimlessly or take a long, inefficient route to reach 251 00:08:04,440 --> 00:08:05,040 its goal. 252 00:08:05,040 --> 00:08:08,160 It's funny how we take navigating websites for granted, 253 00:08:08,160 --> 00:08:10,560 but it's actually quite challenging for AI, right? 254 00:08:10,560 --> 00:08:13,680 It's like a whole different way of seeing the digital world. 255 00:08:13,680 --> 00:08:14,480 Absolutely. 256 00:08:14,480 --> 00:08:18,080 It highlights the importance of well-designed intuitive 257 00:08:18,080 --> 00:08:19,400 interfaces. 258 00:08:19,400 --> 00:08:23,280 Something we understand easily can be very confusing for AI. 259 00:08:23,280 --> 00:08:25,080 Speaking of confusion, did the researchers 260 00:08:25,080 --> 00:08:27,960 mention any risks associated with letting 261 00:08:27,960 --> 00:08:30,400 these AI agents loose on our computers? 262 00:08:30,400 --> 00:08:31,440 They did. 263 00:08:31,440 --> 00:08:35,160 Because these agents can interact with the real world 264 00:08:35,160 --> 00:08:37,800 through the internet and files. 265 00:08:37,800 --> 00:08:39,320 They could potentially take actions 266 00:08:39,320 --> 00:08:41,360 that have unintended consequences. 267 00:08:41,360 --> 00:08:42,360 That sounds a little dicey. 268 00:08:42,360 --> 00:08:43,120 It can be. 269 00:08:43,120 --> 00:08:47,800 Like giving your very smart, but perhaps naive AI assistant 270 00:08:47,800 --> 00:08:49,040 access to your bank account. 271 00:08:49,040 --> 00:08:50,360 Yeah, something like that. 272 00:08:50,360 --> 00:08:51,000 Makes. 273 00:08:51,000 --> 00:08:52,600 There were instances where the agents 274 00:08:52,600 --> 00:08:54,920 tried to do things that, if they had permission, 275 00:08:54,920 --> 00:08:56,160 could have caused problems. 276 00:08:56,160 --> 00:08:56,720 Like what? 277 00:08:56,720 --> 00:08:59,600 Like attempting to reset a website password 278 00:08:59,600 --> 00:09:01,800 after repeatedly failing to log in. 279 00:09:01,800 --> 00:09:05,000 Oh, a bit too eager to solve the problem 280 00:09:05,000 --> 00:09:06,680 without considering the bigger picture. 281 00:09:06,680 --> 00:09:07,240 Precisely. 282 00:09:07,240 --> 00:09:09,440 It highlights the need for safeguards. 283 00:09:09,440 --> 00:09:12,400 We need to teach these agents to recognize actions 284 00:09:12,400 --> 00:09:15,640 that are irreversible or could have major consequences. 285 00:09:15,640 --> 00:09:19,200 They need to know when to pause and ask for human guidance. 286 00:09:19,200 --> 00:09:20,600 Like a safety switch. 287 00:09:20,600 --> 00:09:21,200 Yeah. 288 00:09:21,200 --> 00:09:24,520 Or like, whoa, their partner mechanism 289 00:09:24,520 --> 00:09:26,440 to prevent the AI from going rogue. 290 00:09:26,440 --> 00:09:27,160 Exactly. 291 00:09:27,160 --> 00:09:27,560 Gotcha. 292 00:09:27,560 --> 00:09:29,280 It's about finding the right balance. 293 00:09:29,280 --> 00:09:29,880 Yeah. 294 00:09:29,880 --> 00:09:32,160 Give them the freedom to learn and explore, 295 00:09:32,160 --> 00:09:35,000 but ensure they operate safely and responsibly. 296 00:09:35,000 --> 00:09:35,640 Makes sense. 297 00:09:35,640 --> 00:09:38,160 I bet those safety measures become even more critical 298 00:09:38,160 --> 00:09:40,280 as these AI teams get more sophisticated. 299 00:09:40,280 --> 00:09:41,160 Oh, for sure. 300 00:09:41,160 --> 00:09:42,960 But let's switch gears for a second. 301 00:09:42,960 --> 00:09:43,240 OK. 302 00:09:43,240 --> 00:09:45,720 We talked about the orchestrator, the challenges. 303 00:09:45,720 --> 00:09:48,560 But how does this whole system actually work? 304 00:09:48,560 --> 00:09:49,840 What's going on under the hood? 305 00:09:49,840 --> 00:09:51,480 OK, let's start with the task ledger. 306 00:09:51,480 --> 00:09:51,920 Right. 307 00:09:51,920 --> 00:09:54,600 Think of it as the team's shared brain, 308 00:09:54,600 --> 00:09:56,680 a storehouse of knowledge and assumptions 309 00:09:56,680 --> 00:09:58,520 they need to solve the problem. 310 00:09:58,520 --> 00:10:00,160 More than just a to-do list, then. 311 00:10:00,160 --> 00:10:00,440 Yeah. 312 00:10:00,440 --> 00:10:01,920 It's like a dynamic knowledge base 313 00:10:01,920 --> 00:10:05,200 that grows as the AI team gathers information. 314 00:10:05,200 --> 00:10:05,880 Exactly. 315 00:10:05,880 --> 00:10:07,400 It has four key elements. 316 00:10:07,400 --> 00:10:07,760 OK. 317 00:10:07,760 --> 00:10:08,880 First are the givens. 318 00:10:08,880 --> 00:10:09,400 Givens. 319 00:10:09,400 --> 00:10:11,520 The facts stated in the task description. 320 00:10:11,520 --> 00:10:12,000 Oh. 321 00:10:12,000 --> 00:10:14,000 Of the cities within the United States 322 00:10:14,000 --> 00:10:17,280 where US presidents were born, those are the givens 323 00:10:17,280 --> 00:10:18,760 the AI team starts with. 324 00:10:18,760 --> 00:10:19,280 Gotcha. 325 00:10:19,280 --> 00:10:21,200 Then there are the lookups. 326 00:10:21,200 --> 00:10:23,720 The things the AI needs to find out, 327 00:10:23,720 --> 00:10:26,640 usually through web searches or databases. 328 00:10:26,640 --> 00:10:29,560 So in our example, that would be the list of US presidents 329 00:10:29,560 --> 00:10:30,600 and their birth cities. 330 00:10:30,600 --> 00:10:31,120 Exactly. 331 00:10:31,120 --> 00:10:31,640 OK. 332 00:10:31,640 --> 00:10:33,320 Next are the derivations. 333 00:10:33,320 --> 00:10:34,080 Derivations. 334 00:10:34,080 --> 00:10:37,200 Facts the AI figures out based on the givens and lookups, 335 00:10:37,200 --> 00:10:38,920 often through logic or calculations. 336 00:10:38,920 --> 00:10:39,420 OK. 337 00:10:39,420 --> 00:10:42,280 So figuring out the coordinates of those birth cities, 338 00:10:42,280 --> 00:10:45,440 the distances, and which two are farthest apart going 339 00:10:45,440 --> 00:10:47,520 eastward, those would be the derivations. 340 00:10:47,520 --> 00:10:48,200 You got it. 341 00:10:48,200 --> 00:10:50,120 And finally, there are the educated guesses 342 00:10:50,120 --> 00:10:51,080 we talked about earlier. 343 00:10:51,080 --> 00:10:51,600 Right. 344 00:10:51,600 --> 00:10:53,480 Those AI intuitions, those hunches. 345 00:10:53,480 --> 00:10:54,040 Yeah. 346 00:10:54,040 --> 00:10:57,360 But how does the orchestrator actually use all this information 347 00:10:57,360 --> 00:10:59,080 to guide the team? 348 00:10:59,080 --> 00:11:01,360 That's where the progress ledger comes in. 349 00:11:01,360 --> 00:11:02,520 A progress ledger. 350 00:11:02,520 --> 00:11:03,040 Yeah. 351 00:11:03,040 --> 00:11:04,880 It's the mission control dashboard. 352 00:11:04,880 --> 00:11:05,320 Oh, OK. 353 00:11:05,320 --> 00:11:07,760 Showing the orchestrator the status of the task 354 00:11:07,760 --> 00:11:09,560 and guiding its next move. 355 00:11:09,560 --> 00:11:12,200 So the task ledger is the big picture. 356 00:11:12,200 --> 00:11:16,520 And the progress ledger is the tactical step-by-step guide. 357 00:11:16,520 --> 00:11:17,480 Exactly. 358 00:11:17,480 --> 00:11:19,520 The progress ledger helps the orchestrator 359 00:11:19,520 --> 00:11:21,440 answer some key questions. 360 00:11:21,440 --> 00:11:22,560 Like what? 361 00:11:22,560 --> 00:11:26,560 Well, the most important one is the task complete. 362 00:11:26,560 --> 00:11:28,800 Because if it is, everyone can go home early. 363 00:11:28,800 --> 00:11:29,880 Right. 364 00:11:29,880 --> 00:11:31,520 But if not, the orchestrator needs 365 00:11:31,520 --> 00:11:32,960 to figure out what to do next. 366 00:11:32,960 --> 00:11:33,600 OK. 367 00:11:33,600 --> 00:11:35,400 That leads to the next question. 368 00:11:35,400 --> 00:11:37,560 Is the team making progress? 369 00:11:37,560 --> 00:11:39,320 Because if they're spinning their wheels, 370 00:11:39,320 --> 00:11:40,680 it's time for a new strategy. 371 00:11:40,680 --> 00:11:41,400 Exactly. 372 00:11:41,400 --> 00:11:44,320 And that's where the progress ledger is really useful. 373 00:11:44,320 --> 00:11:44,960 How so? 374 00:11:44,960 --> 00:11:47,640 It can spot these unproductive loops. 375 00:11:47,640 --> 00:11:48,240 Oh, right. 376 00:11:48,240 --> 00:11:51,240 Like the web surfer stuck searching for the same term 377 00:11:51,240 --> 00:11:51,920 over and over. 378 00:11:51,920 --> 00:11:52,880 Precisely. 379 00:11:52,880 --> 00:11:55,200 The orchestrator can then intervene, 380 00:11:55,200 --> 00:11:58,320 redirecting the team, or changing the plan. 381 00:11:58,320 --> 00:12:00,520 So it's like, OK team, we've hit a wall. 382 00:12:00,520 --> 00:12:01,920 Let's try a different approach. 383 00:12:01,920 --> 00:12:02,520 Exactly. 384 00:12:02,520 --> 00:12:05,560 That flexibility is crucial for solving complex tasks 385 00:12:05,560 --> 00:12:07,520 where there's no clear solution path. 386 00:12:07,520 --> 00:12:07,920 Yeah. 387 00:12:07,920 --> 00:12:08,840 That makes a lot of sense. 388 00:12:08,840 --> 00:12:09,320 Yeah. 389 00:12:09,320 --> 00:12:11,560 But this all seems incredibly intricate. 390 00:12:11,560 --> 00:12:14,360 How does the orchestrator actually make these decisions? 391 00:12:14,360 --> 00:12:17,560 Does it have like some super intelligence 392 00:12:17,560 --> 00:12:20,440 that lets it grasp the nuances of the task 393 00:12:20,440 --> 00:12:22,240 in each agent's capabilities? 394 00:12:22,240 --> 00:12:25,880 Well, this is where the power of large language models comes in. 395 00:12:25,880 --> 00:12:26,360 OK. 396 00:12:26,360 --> 00:12:30,360 The orchestrator, just like the other agents, is powered by one. 397 00:12:30,360 --> 00:12:34,200 So it's not just following like pre-programmed rules. 398 00:12:34,200 --> 00:12:34,600 Right. 399 00:12:34,600 --> 00:12:36,600 It's using its understanding of language 400 00:12:36,600 --> 00:12:39,400 and its ability to reason to make decisions. 401 00:12:39,400 --> 00:12:40,040 Exactly. 402 00:12:40,040 --> 00:12:41,320 And here's the clever part. 403 00:12:41,320 --> 00:12:42,080 What's that? 404 00:12:42,080 --> 00:12:43,760 The researchers didn't specifically 405 00:12:43,760 --> 00:12:47,560 program the orchestrator to know how to manage a team 406 00:12:47,560 --> 00:12:49,840 or detect unproductive loops. 407 00:12:49,840 --> 00:12:50,200 Hold on. 408 00:12:50,200 --> 00:12:52,160 So how did it learn to do all those things? 409 00:12:52,160 --> 00:12:54,120 Through a technique called prompt engineering. 410 00:12:54,120 --> 00:12:55,040 Prompt engineering. 411 00:12:55,040 --> 00:12:55,360 Yeah. 412 00:12:55,360 --> 00:12:57,800 It's the art of crafting the right prompts 413 00:12:57,800 --> 00:12:59,840 to guide the language model's behavior. 414 00:12:59,840 --> 00:13:02,960 So giving the orchestrator like guidelines or instructions 415 00:13:02,960 --> 00:13:04,960 to help it understand its role and how 416 00:13:04,960 --> 00:13:06,040 to work with the other agents. 417 00:13:06,040 --> 00:13:06,880 Exactly. 418 00:13:06,880 --> 00:13:08,360 With careful prompting, they were 419 00:13:08,360 --> 00:13:11,360 able to turn this language model into an effective leader, 420 00:13:11,360 --> 00:13:14,120 coordinating a team of AI agents. 421 00:13:14,120 --> 00:13:16,360 Amazing what you can achieve with the right prompts, huh? 422 00:13:16,360 --> 00:13:18,280 But back to the progress ledger for a second. 423 00:13:18,280 --> 00:13:18,880 OK. 424 00:13:18,880 --> 00:13:21,120 We've covered three questions. 425 00:13:21,120 --> 00:13:23,320 Is the task complete? 426 00:13:23,320 --> 00:13:24,960 Is the team making progress? 427 00:13:24,960 --> 00:13:27,960 And are there any unproductive loops? 428 00:13:27,960 --> 00:13:28,320 Right. 429 00:13:28,320 --> 00:13:29,960 What about the last two? 430 00:13:29,960 --> 00:13:32,400 Well, after assessing the situation, 431 00:13:32,400 --> 00:13:36,400 the orchestrator needs to decide which agent to activate next. 432 00:13:36,400 --> 00:13:40,400 The fourth question is, which agent should speak next? 433 00:13:40,400 --> 00:13:43,600 So the orchestrator is like a teacher calling on a student. 434 00:13:43,600 --> 00:13:45,280 OK, Web Surfer, what have you found? 435 00:13:45,280 --> 00:13:46,000 Exactly. 436 00:13:46,000 --> 00:13:49,080 And the final question is, what instruction or question 437 00:13:49,080 --> 00:13:51,240 should be asked of this agent? 438 00:13:51,240 --> 00:13:53,640 So it's not just choosing the right agent. 439 00:13:53,640 --> 00:13:57,040 It's about framing the task in a way to help them succeed. 440 00:13:57,040 --> 00:13:57,600 You got it. 441 00:13:57,600 --> 00:13:58,320 That's critical. 442 00:13:58,320 --> 00:14:00,360 The orchestrator needs to understand each agent's 443 00:14:00,360 --> 00:14:02,600 capabilities and communicate effectively 444 00:14:02,600 --> 00:14:03,920 to get the results it needs. 445 00:14:03,920 --> 00:14:05,280 It's like a manager who knows how 446 00:14:05,280 --> 00:14:06,560 to get the best out of their team 447 00:14:06,560 --> 00:14:09,520 by playing to their strengths and giving clear directions. 448 00:14:09,520 --> 00:14:10,360 Precisely. 449 00:14:10,360 --> 00:14:12,840 Understanding those individual talents and fostering 450 00:14:12,840 --> 00:14:14,280 collaboration is key. 451 00:14:14,280 --> 00:14:17,440 That's what makes Magentic One so innovative. 452 00:14:17,440 --> 00:14:20,800 It's designed for teamwork, with each agent contributing 453 00:14:20,800 --> 00:14:23,640 their expertise to achieve a common goal. 454 00:14:23,640 --> 00:14:26,480 It's a whole new way of thinking about AI. 455 00:14:26,480 --> 00:14:29,280 Not about one super intelligence that does it all, 456 00:14:29,280 --> 00:14:32,960 but a system that learns and adapts by working together, 457 00:14:32,960 --> 00:14:33,760 just like us. 458 00:14:33,760 --> 00:14:34,160 I know. 459 00:14:34,160 --> 00:14:35,000 It's really cool. 460 00:14:35,000 --> 00:14:37,160 This research is incredibly exciting. 461 00:14:37,160 --> 00:14:37,560 It is. 462 00:14:37,560 --> 00:14:39,360 It really opens up a world of possibilities 463 00:14:39,360 --> 00:14:42,520 for how we can use AI to solve complex problems. 464 00:14:42,520 --> 00:14:43,040 Absolutely. 465 00:14:43,040 --> 00:14:44,880 And really enhance our own capabilities. 466 00:14:44,880 --> 00:14:45,720 Yeah, for sure. 467 00:14:45,720 --> 00:14:47,560 It's definitely a fascinating journey. 468 00:14:47,560 --> 00:14:49,560 We've covered a lot in this part of our deep dive. 469 00:14:49,560 --> 00:14:50,000 Yeah. 470 00:14:50,000 --> 00:14:51,600 The inner workings of Magentic One, 471 00:14:51,600 --> 00:14:53,480 the roles of those ledgers, the importance 472 00:14:53,480 --> 00:14:55,000 of communication and coordination. 473 00:14:55,000 --> 00:14:55,600 Right. 474 00:14:55,600 --> 00:14:59,160 But there's one big piece we haven't discussed yet. 475 00:14:59,160 --> 00:15:03,760 We've spent the last two parts digging into the brains 476 00:15:03,760 --> 00:15:05,520 of this AI 18. 477 00:15:05,520 --> 00:15:06,600 Yeah, the architecture. 478 00:15:06,600 --> 00:15:08,880 But we haven't met the actual agents yet. 479 00:15:08,880 --> 00:15:10,560 Right, the ones doing the hands-on work. 480 00:15:10,560 --> 00:15:11,880 The ones actually getting their hands dirty. 481 00:15:11,880 --> 00:15:12,480 Exactly. 482 00:15:12,480 --> 00:15:14,280 So let's introduce the stars of the show. 483 00:15:14,280 --> 00:15:16,960 Starting with the web server, what does that one do? 484 00:15:16,960 --> 00:15:19,120 It's all about navigating the internet, 485 00:15:19,120 --> 00:15:21,280 like a super-powered research assistant. 486 00:15:21,280 --> 00:15:25,400 OK, so it can search for information, follow links, 487 00:15:25,400 --> 00:15:26,640 even interact with websites. 488 00:15:26,640 --> 00:15:27,880 That's a lot of power. 489 00:15:27,880 --> 00:15:29,000 It is. 490 00:15:29,000 --> 00:15:31,760 But the web can be a messy place, even for humans. 491 00:15:31,760 --> 00:15:32,400 Oh, yeah. 492 00:15:32,400 --> 00:15:34,320 How does the web server even understand 493 00:15:34,320 --> 00:15:35,320 what it's looking at? 494 00:15:35,320 --> 00:15:37,800 It uses a technique called set of marks 495 00:15:37,800 --> 00:15:41,120 prompting to interact with those web pages. 496 00:15:41,120 --> 00:15:42,680 Set of marks prompting. 497 00:15:42,680 --> 00:15:43,160 Yeah. 498 00:15:43,160 --> 00:15:44,400 That sounds very technical. 499 00:15:44,400 --> 00:15:45,640 It is a bit of a mouthful. 500 00:15:45,640 --> 00:15:46,480 Bring it down for me. 501 00:15:46,480 --> 00:15:48,600 Imagine you're guiding someone through a website 502 00:15:48,600 --> 00:15:49,280 over the phone. 503 00:15:49,280 --> 00:15:49,840 OK. 504 00:15:49,840 --> 00:15:51,840 You might say, click the Submit button 505 00:15:51,840 --> 00:15:53,480 in the top right corner. 506 00:15:53,480 --> 00:15:55,720 You're giving them landmarks to find the right thing. 507 00:15:55,720 --> 00:15:59,040 OK, so you're providing specific visual cues. 508 00:15:59,040 --> 00:16:00,240 Exactly. 509 00:16:00,240 --> 00:16:02,200 That's what set of marks prompting does. 510 00:16:02,200 --> 00:16:02,640 Gotcha. 511 00:16:02,640 --> 00:16:05,120 The web server takes a screenshot of the web page 512 00:16:05,120 --> 00:16:09,200 and then uses AI to identify and label everything important. 513 00:16:09,200 --> 00:16:12,240 Buttons, links, text boxes, you name it. 514 00:16:12,240 --> 00:16:14,560 So it creates a digital map of the web page, 515 00:16:14,560 --> 00:16:16,280 kind of highlighting the points of interest. 516 00:16:16,280 --> 00:16:17,120 Precisely. 517 00:16:17,120 --> 00:16:19,640 With this map, it can then follow the orchestrator's 518 00:16:19,640 --> 00:16:22,120 instructions with pinpoint accuracy. 519 00:16:22,120 --> 00:16:24,320 So it's like giving the web server x-ray vision 520 00:16:24,320 --> 00:16:26,160 to see the structure of a web page. 521 00:16:26,160 --> 00:16:26,920 Yeah, I like it. 522 00:16:26,920 --> 00:16:29,600 But it's not just clicking and typing, right? 523 00:16:29,600 --> 00:16:30,000 Yeah. 524 00:16:30,000 --> 00:16:32,680 Can it actually understand the content of the page? 525 00:16:32,680 --> 00:16:33,680 It can. 526 00:16:33,680 --> 00:16:37,480 The web server can extract text, summarize information, 527 00:16:37,480 --> 00:16:39,960 and even answer questions about what it reads. 528 00:16:39,960 --> 00:16:40,320 Wow. 529 00:16:40,320 --> 00:16:41,480 It's not just browsing. 530 00:16:41,480 --> 00:16:42,840 It's comprehending. 531 00:16:42,840 --> 00:16:44,520 OK, that is impressive. 532 00:16:44,520 --> 00:16:45,200 It is. 533 00:16:45,200 --> 00:16:47,120 So we have the web server exploring 534 00:16:47,120 --> 00:16:48,440 the wilds of the internet. 535 00:16:48,440 --> 00:16:48,840 Yeah. 536 00:16:48,840 --> 00:16:50,320 What about the file server? 537 00:16:50,320 --> 00:16:51,520 Yes, the file server. 538 00:16:51,520 --> 00:16:52,360 What does that one do? 539 00:16:52,360 --> 00:16:55,400 Is it just the web server's cousin who prefers libraries? 540 00:16:55,400 --> 00:16:56,840 In a way, yes. 541 00:16:56,840 --> 00:16:59,560 While the web server tackles the open web, 542 00:16:59,560 --> 00:17:02,920 the file server is more of a digital librarian, 543 00:17:02,920 --> 00:17:05,840 navigating the structured world of files and folders. 544 00:17:05,840 --> 00:17:08,240 So it's the go-to agent for all things, documents, 545 00:17:08,240 --> 00:17:10,000 spreadsheets, presentations, all that stuff. 546 00:17:10,000 --> 00:17:10,720 Exactly. 547 00:17:10,720 --> 00:17:14,160 The file server can open and read a wide variety of file types, 548 00:17:14,160 --> 00:17:18,080 PDFs, office documents, images, even audio and video files. 549 00:17:18,080 --> 00:17:18,960 Wait, hold on. 550 00:17:18,960 --> 00:17:20,240 Audio and video. 551 00:17:20,240 --> 00:17:24,720 I thought MagentaCon was limited to text-based stuff. 552 00:17:24,720 --> 00:17:27,280 It can't actually watch or listen like we do. 553 00:17:27,280 --> 00:17:27,800 OK. 554 00:17:27,800 --> 00:17:30,440 Instead, it uses speech to text for audio, 555 00:17:30,440 --> 00:17:33,480 converting those spoken words into text for analysis. 556 00:17:33,480 --> 00:17:34,120 Oh. 557 00:17:34,120 --> 00:17:37,200 For videos, it sometimes uses captions or transcripts 558 00:17:37,200 --> 00:17:38,320 if those are available. 559 00:17:38,320 --> 00:17:42,040 So it's like a super-powered note taker listening to lectures 560 00:17:42,040 --> 00:17:44,960 or watching documentaries and summarizing the key points. 561 00:17:44,960 --> 00:17:46,240 A great analogy. 562 00:17:46,240 --> 00:17:48,960 And for text-based files like PDFs, 563 00:17:48,960 --> 00:17:51,920 they can extract text, search for keywords, 564 00:17:51,920 --> 00:17:54,000 and even answer questions about the content. 565 00:17:54,000 --> 00:17:55,720 It's like a super-charged search engine 566 00:17:55,720 --> 00:17:57,040 for your own digital archives. 567 00:17:57,040 --> 00:17:57,840 OK. 568 00:17:57,840 --> 00:18:00,960 OK, let's move on to the agent that really blew my mind, 569 00:18:00,960 --> 00:18:01,560 the coder. 570 00:18:01,560 --> 00:18:02,560 All right, the coder. 571 00:18:02,560 --> 00:18:06,720 How does this AI actually stack up against human programmers? 572 00:18:06,720 --> 00:18:09,080 Honestly, it's still pretty basic compared 573 00:18:09,080 --> 00:18:11,480 to a seasoned human programmer. 574 00:18:11,480 --> 00:18:11,920 Yeah. 575 00:18:11,920 --> 00:18:13,680 It can write simple Python programs, 576 00:18:13,680 --> 00:18:17,040 but it's not ready to build complex software applications 577 00:18:17,040 --> 00:18:17,600 just yet. 578 00:18:17,600 --> 00:18:20,640 So more of a junior developer learning the ropes 579 00:18:20,640 --> 00:18:22,400 and handling those smaller tasks. 580 00:18:22,400 --> 00:18:23,520 That's a good way to put it. 581 00:18:23,520 --> 00:18:25,680 But even at this early stage, the coder 582 00:18:25,680 --> 00:18:27,200 is incredibly useful. 583 00:18:27,200 --> 00:18:28,080 How so? 584 00:18:28,080 --> 00:18:30,360 It can automate repetitive tasks, 585 00:18:30,360 --> 00:18:32,640 write scripts to manipulate data, 586 00:18:32,640 --> 00:18:35,960 even generate code to interact with other systems. 587 00:18:35,960 --> 00:18:39,480 So while it might not be building the next big social media 588 00:18:39,480 --> 00:18:42,720 platform, it can still save human programmers 589 00:18:42,720 --> 00:18:46,200 a ton of time by handling those more mundane tasks. 590 00:18:46,200 --> 00:18:47,160 Precisely. 591 00:18:47,160 --> 00:18:49,520 And like all the agents in Magentic One, 592 00:18:49,520 --> 00:18:52,520 the coder is always learning and improving. 593 00:18:52,520 --> 00:18:54,840 As these large language models evolve, 594 00:18:54,840 --> 00:18:56,440 we can expect its coding abilities 595 00:18:56,440 --> 00:18:58,760 to become even more sophisticated. 596 00:18:58,760 --> 00:19:02,080 So one day, it might actually give those human programmers 597 00:19:02,080 --> 00:19:03,200 a run for their money. 598 00:19:03,200 --> 00:19:04,400 Maybe, yeah. 599 00:19:04,400 --> 00:19:06,640 Now, the last agent is the computer terminal. 600 00:19:06,640 --> 00:19:08,200 Yes, the computer terminal. 601 00:19:08,200 --> 00:19:10,320 And I have to admit, I'm having a little trouble 602 00:19:10,320 --> 00:19:11,360 picturing what it does. 603 00:19:11,360 --> 00:19:14,720 Think of it as the bridge between the code and the real world. 604 00:19:14,720 --> 00:19:15,000 OK. 605 00:19:15,000 --> 00:19:17,720 It's the agent that actually runs the code written 606 00:19:17,720 --> 00:19:18,640 by the coder. 607 00:19:18,640 --> 00:19:20,760 So it's the muscle behind the brains. 608 00:19:20,760 --> 00:19:21,720 You could say that. 609 00:19:21,720 --> 00:19:23,520 The one that takes those digital instructions 610 00:19:23,520 --> 00:19:24,680 and turns them into action. 611 00:19:24,680 --> 00:19:25,180 Exactly. 612 00:19:25,180 --> 00:19:28,200 The computer terminal has access to the operating system, 613 00:19:28,200 --> 00:19:30,920 where it can run code, execute commands, 614 00:19:30,920 --> 00:19:32,920 even install new software libraries. 615 00:19:32,920 --> 00:19:36,280 It's like giving the AIA team direct control 616 00:19:36,280 --> 00:19:38,760 over the computer, allowing them to interact 617 00:19:38,760 --> 00:19:39,800 with other programs. 618 00:19:39,800 --> 00:19:40,480 Precisely. 619 00:19:40,480 --> 00:19:42,440 And that opens up a lot of possibilities, 620 00:19:42,440 --> 00:19:46,400 downloading files, installing software, running simulations, 621 00:19:46,400 --> 00:19:48,240 even controlling hardware devices. 622 00:19:48,240 --> 00:19:49,040 Wow. 623 00:19:49,040 --> 00:19:51,520 It's the Swiss Army knife of agents. 624 00:19:51,520 --> 00:19:52,120 It really is. 625 00:19:52,120 --> 00:19:55,520 It's amazing how all these specialized agents work together, 626 00:19:55,520 --> 00:19:57,400 each bringing their own skills to the table. 627 00:19:57,400 --> 00:19:58,480 It's pretty remarkable. 628 00:19:58,480 --> 00:20:00,080 It's like they've redefined teamwork. 629 00:20:00,080 --> 00:20:02,600 It's a whole new way of thinking about AI. 630 00:20:02,600 --> 00:20:06,000 It's not about building one's super intelligence, 631 00:20:06,000 --> 00:20:09,320 but a system that learns and adapts through collaboration. 632 00:20:09,320 --> 00:20:10,280 Just like us. 633 00:20:10,280 --> 00:20:11,160 Just like us. 634 00:20:11,160 --> 00:20:13,000 And it's just the beginning. 635 00:20:13,000 --> 00:20:14,840 This field is evolving so quickly, 636 00:20:14,840 --> 00:20:17,760 I can't wait to see what other innovative multi-agent systems 637 00:20:17,760 --> 00:20:18,400 emerge. 638 00:20:18,400 --> 00:20:19,160 Me too. 639 00:20:19,160 --> 00:20:21,640 It's a fascinating time to be following AI. 640 00:20:21,640 --> 00:20:22,240 It is. 641 00:20:22,240 --> 00:20:25,320 This technology has the potential to revolutionize 642 00:20:25,320 --> 00:20:27,160 so many aspects of our lives. 643 00:20:27,160 --> 00:20:29,840 Well, this has been an incredible deep dive into Magentik 1. 644 00:20:29,840 --> 00:20:31,000 It has been fun. 645 00:20:31,000 --> 00:20:34,440 We've covered the architecture, the agents, the challenges, 646 00:20:34,440 --> 00:20:37,240 and the potential of this groundbreaking system. 647 00:20:37,240 --> 00:20:37,880 Absolutely. 648 00:20:37,880 --> 00:20:38,960 We've gone through it all. 649 00:20:38,960 --> 00:20:40,840 We hope this has given you a better understanding 650 00:20:40,840 --> 00:20:42,080 of multi-agent AI. 651 00:20:42,080 --> 00:20:42,580 Yeah. 652 00:20:42,580 --> 00:20:45,520 It's a rapidly evolving field full of possibilities. 653 00:20:45,520 --> 00:20:48,040 And as always, stay curious, keep learning, 654 00:20:48,040 --> 00:20:49,840 and join us next time for another deep 655 00:20:49,840 --> 00:21:13,840 dive into the world of cutting-edge technology.