1 00:00:00,000 --> 00:00:02,920 All right, welcome back to the deep dive. 2 00:00:02,920 --> 00:00:03,920 We all know you're here 3 00:00:03,920 --> 00:00:05,560 because you wanna be ahead of the curve, right? 4 00:00:05,560 --> 00:00:07,760 You wanna understand the big things 5 00:00:07,760 --> 00:00:09,960 and you wanna understand them fast. 6 00:00:09,960 --> 00:00:12,320 And well, today, we're gonna dive into something 7 00:00:12,320 --> 00:00:14,760 that is seriously shaping the future of AI. 8 00:00:16,280 --> 00:00:18,720 How we make sense of how these AIs act 9 00:00:18,720 --> 00:00:19,800 when there's a whole lot of them. 10 00:00:19,800 --> 00:00:20,760 Yeah, that's it. 11 00:00:20,760 --> 00:00:22,720 It's like, you know, we're moving beyond 12 00:00:22,720 --> 00:00:24,720 just looking at a single AI doing its thing. 13 00:00:24,720 --> 00:00:26,920 We're talking about like frameworks. 14 00:00:26,920 --> 00:00:28,560 Think of them like these advanced labs 15 00:00:28,560 --> 00:00:31,120 where researchers can observe how tons of AIs behave 16 00:00:31,120 --> 00:00:32,160 when they're all together. 17 00:00:32,160 --> 00:00:33,920 So it's like instead of studying a single ant, 18 00:00:33,920 --> 00:00:36,200 we're looking at the entire ant colony. 19 00:00:36,200 --> 00:00:37,080 Exactly. 20 00:00:37,080 --> 00:00:38,000 And you know what we're using 21 00:00:38,000 --> 00:00:39,880 to look at these AI colonies today? 22 00:00:39,880 --> 00:00:40,720 Wait on me. 23 00:00:40,720 --> 00:00:41,540 Camel. 24 00:00:41,540 --> 00:00:42,380 Camel. 25 00:00:42,380 --> 00:00:43,220 Yeah. 26 00:00:43,220 --> 00:00:44,720 And it's not just another fancy AI term, you know, 27 00:00:44,720 --> 00:00:45,800 right from their GitHub page, 28 00:00:45,800 --> 00:00:48,200 which is where we're getting most of our info from today. 29 00:00:48,200 --> 00:00:50,800 Camel, which stands for, get this, 30 00:00:50,800 --> 00:00:53,880 communicative agents for mind exploration 31 00:00:53,880 --> 00:00:56,600 of large language model society. 32 00:00:56,600 --> 00:00:59,640 It says, and I quote, the first and the best 33 00:00:59,640 --> 00:01:01,920 multi-agent framework. 34 00:01:01,920 --> 00:01:03,080 Bold statement. 35 00:01:03,080 --> 00:01:04,080 I know, right? 36 00:01:04,080 --> 00:01:06,080 But what's really cool is it's got this huge 37 00:01:06,080 --> 00:01:07,800 open source community around it. 38 00:01:07,800 --> 00:01:10,040 Like there are over a hundred researchers 39 00:01:10,040 --> 00:01:13,040 all working together, trying to figure out the rules, 40 00:01:13,040 --> 00:01:15,960 what they call scaling laws of how these AIs behave 41 00:01:15,960 --> 00:01:17,600 when you have so many of them. 42 00:01:17,600 --> 00:01:21,840 And you can check out all their work at camel-ai.org. 43 00:01:21,840 --> 00:01:24,240 Okay, so to sum it up for our listeners, 44 00:01:24,240 --> 00:01:26,840 we're going deep on camel today. 45 00:01:26,840 --> 00:01:28,400 We're gonna figure out what it is, 46 00:01:28,400 --> 00:01:30,520 why all these researchers are so hyped about it, 47 00:01:30,520 --> 00:01:33,320 and what kind of doors it opens to study these AI agents. 48 00:01:33,320 --> 00:01:34,160 You got it. 49 00:01:34,160 --> 00:01:35,000 All from their GitHub. 50 00:01:35,000 --> 00:01:37,200 All right, so first things first, 51 00:01:37,200 --> 00:01:40,800 what exactly is camel, like in plain English? 52 00:01:40,800 --> 00:01:43,000 So basically think of camel as a toolkit, 53 00:01:43,000 --> 00:01:45,880 a super well-organized set of tools and instructions 54 00:01:45,880 --> 00:01:48,200 that researchers can use to really dig deep 55 00:01:48,200 --> 00:01:50,920 into how AIs behave, what they can do, 56 00:01:50,920 --> 00:01:53,080 and you know, any potential problems they might cause, 57 00:01:53,080 --> 00:01:54,520 especially when there's a lot of them. 58 00:01:54,520 --> 00:01:56,080 Gotcha, so the whole point is to study them 59 00:01:56,080 --> 00:01:58,560 on a large scale, not just individual AIs. 60 00:01:58,560 --> 00:01:59,400 That's exactly it. 61 00:01:59,400 --> 00:02:01,240 It's not just about making individual AIs 62 00:02:01,240 --> 00:02:03,240 a little bit better, it's about looking at what happens 63 00:02:03,240 --> 00:02:05,080 when you have a whole bunch of them working together, 64 00:02:05,080 --> 00:02:06,480 what new things emerge. 65 00:02:06,480 --> 00:02:09,720 And this isn't just some like academic thing, right? 66 00:02:09,720 --> 00:02:12,080 Their GitHub makes it clear that it's open source 67 00:02:12,080 --> 00:02:13,040 and it's all about communities. 68 00:02:13,040 --> 00:02:14,360 Oh yeah, totally. 69 00:02:14,360 --> 00:02:17,320 The fact that there are over a hundred researchers 70 00:02:17,320 --> 00:02:19,240 working on it, that says a lot. 71 00:02:19,240 --> 00:02:21,200 They really believe in this approach. 72 00:02:21,200 --> 00:02:23,840 The core idea is that you can't really understand 73 00:02:23,840 --> 00:02:27,040 how AIs act by just looking at one at a time. 74 00:02:27,040 --> 00:02:29,040 You've got to see how they act in a group. 75 00:02:29,040 --> 00:02:32,480 So it's like, less about one AI solving a problem 76 00:02:32,480 --> 00:02:34,440 more about how they work together, 77 00:02:34,440 --> 00:02:36,760 or maybe even like compete with each other. 78 00:02:36,760 --> 00:02:39,280 Exactly, and they're not just talking about it. 79 00:02:39,280 --> 00:02:42,600 The GitHub page shows they've made actual tools, 80 00:02:42,600 --> 00:02:45,720 different types of agents, like AIs with special skills, 81 00:02:45,720 --> 00:02:46,800 different tasks they can do, 82 00:02:46,800 --> 00:02:48,640 different ways to give them instructions, 83 00:02:48,640 --> 00:02:49,880 different models they can use, 84 00:02:49,880 --> 00:02:51,840 and even simulated worlds for them to live in. 85 00:02:51,840 --> 00:02:52,680 It's amazing. 86 00:02:52,680 --> 00:02:54,960 Wow, so they basically built a whole research lab 87 00:02:54,960 --> 00:02:56,680 for these multi-agent AIs. 88 00:02:56,680 --> 00:02:57,800 Yeah, pretty much. 89 00:02:57,800 --> 00:02:58,800 And you know how on GitHub 90 00:02:58,800 --> 00:03:00,200 you can see those stars and forks? 91 00:03:00,200 --> 00:03:04,600 Well, it has 8.8 thousand stars and 921 forks. 92 00:03:04,600 --> 00:03:06,760 And the open source world, that's a lot. 93 00:03:06,760 --> 00:03:08,680 So people are definitely paying attention. 94 00:03:08,680 --> 00:03:09,800 Big time. 95 00:03:09,800 --> 00:03:12,160 Plus, it's under the Apache 2.0 license, 96 00:03:12,160 --> 00:03:15,080 which basically means anyone can use it and build on it. 97 00:03:15,080 --> 00:03:18,840 Okay, so Camel's this open, community-driven framework 98 00:03:18,840 --> 00:03:21,080 to study lots of AIs together. 99 00:03:21,080 --> 00:03:23,200 But what are the like ground rules? 100 00:03:23,200 --> 00:03:24,240 How do they design it? 101 00:03:24,240 --> 00:03:25,920 What were the main ideas? 102 00:03:25,920 --> 00:03:28,320 So there are four big design principles 103 00:03:28,320 --> 00:03:29,680 they talk about on GitHub. 104 00:03:29,680 --> 00:03:31,440 The first one is evolvability. 105 00:03:31,440 --> 00:03:33,920 It means these multi-agent systems aren't fixed. 106 00:03:33,920 --> 00:03:35,880 They can actually get better over time. 107 00:03:35,880 --> 00:03:37,840 Either from the data they generate themselves 108 00:03:37,840 --> 00:03:40,640 or by interacting with these simulated environments. 109 00:03:40,640 --> 00:03:43,720 It's like they can learn using either reinforcement learning 110 00:03:43,720 --> 00:03:45,120 where they learn by getting rewards, 111 00:03:45,120 --> 00:03:47,720 or supervised learning where they learn from labeled examples. 112 00:03:47,720 --> 00:03:50,280 And it's not just like getting a little better. 113 00:03:50,280 --> 00:03:52,960 The exciting part is that totally new strategies 114 00:03:52,960 --> 00:03:55,880 or even like new types of intelligence could just pop up. 115 00:03:55,880 --> 00:03:58,080 Whoa, so the system can like evolve. 116 00:03:58,080 --> 00:03:59,800 Like the AIs can learn and change. 117 00:03:59,800 --> 00:04:01,480 Yeah, totally. 118 00:04:01,480 --> 00:04:04,000 The next principle is scalability. 119 00:04:04,000 --> 00:04:07,560 Since the whole point is to study huge groups of AIs, 120 00:04:07,560 --> 00:04:09,600 the framework is built to handle systems 121 00:04:09,600 --> 00:04:11,440 with millions of agents. 122 00:04:11,440 --> 00:04:12,880 And it's not just about the numbers, 123 00:04:12,880 --> 00:04:14,680 it's about making sure they can coordinate, 124 00:04:14,680 --> 00:04:17,160 communicate, and share resources efficiently. 125 00:04:17,160 --> 00:04:20,680 Imagine like millions of AIs all working at the same time. 126 00:04:20,680 --> 00:04:21,560 Millions. 127 00:04:21,560 --> 00:04:22,520 That's just mind-boggling. 128 00:04:22,520 --> 00:04:24,320 It's like managing a whole digital world. 129 00:04:24,320 --> 00:04:25,400 I know, right. 130 00:04:25,400 --> 00:04:28,120 Okay, next up is statefulness. 131 00:04:28,120 --> 00:04:29,960 This is all about memory. 132 00:04:29,960 --> 00:04:31,160 The agents in CAML, 133 00:04:31,160 --> 00:04:33,440 they don't just react to what's happening right now. 134 00:04:33,440 --> 00:04:36,360 They can remember past interactions and experiences. 135 00:04:36,360 --> 00:04:39,280 This helps them have longer, more complex conversations 136 00:04:39,280 --> 00:04:40,720 and tackle harder tasks 137 00:04:40,720 --> 00:04:42,680 that need them to remember what happened before. 138 00:04:42,680 --> 00:04:44,600 So they can learn from their mistakes 139 00:04:44,600 --> 00:04:46,600 and build on what worked before. 140 00:04:46,600 --> 00:04:47,440 Exactly. 141 00:04:47,440 --> 00:04:48,760 And that's super important 142 00:04:48,760 --> 00:04:50,320 for any kind of collaboration 143 00:04:50,320 --> 00:04:52,520 or complex problem solving. 144 00:04:52,520 --> 00:04:54,600 And the last principle is, this is a cool one, 145 00:04:54,600 --> 00:04:56,160 code as prompt. 146 00:04:56,160 --> 00:04:59,600 So in CAML, the actual code that's written for the agents, 147 00:04:59,600 --> 00:05:01,520 along with any comments in the code, 148 00:05:01,520 --> 00:05:03,240 is used as a way to guide them. 149 00:05:03,240 --> 00:05:05,040 The idea is that good, clear code 150 00:05:05,040 --> 00:05:07,040 isn't just for humans to understand. 151 00:05:07,040 --> 00:05:09,280 It can also help the AI agents themselves 152 00:05:09,280 --> 00:05:11,240 figure out what they're supposed to do. 153 00:05:11,240 --> 00:05:14,320 So like the code itself is a set of instructions for the AI. 154 00:05:14,320 --> 00:05:15,160 Yeah. 155 00:05:15,160 --> 00:05:16,880 So there's an agent that's supposed to write summaries, 156 00:05:16,880 --> 00:05:18,800 and in the code, there's a comment that says, 157 00:05:18,800 --> 00:05:21,200 like, focus on the main financial numbers. 158 00:05:21,200 --> 00:05:24,360 That comment becomes part of the instructions for the AI. 159 00:05:24,360 --> 00:05:26,440 That's a really neat connection between coding and AI. 160 00:05:26,440 --> 00:05:28,760 Okay, so we've got the design principles down, 161 00:05:28,760 --> 00:05:31,480 but why are all these researchers choosing CAML? 162 00:05:31,480 --> 00:05:34,640 What are the big reasons they list on their GitHub page? 163 00:05:34,640 --> 00:05:36,720 Okay, so first, the ability to create 164 00:05:36,720 --> 00:05:39,080 a large scale agent system is huge, 165 00:05:39,080 --> 00:05:41,240 simulating up to a million agents. 166 00:05:41,240 --> 00:05:43,880 That lets researchers look at emergent behavior, 167 00:05:43,880 --> 00:05:47,040 those unexpected patterns that only pop up in big groups, 168 00:05:47,040 --> 00:05:49,440 and see how these AI societies scale. 169 00:05:49,440 --> 00:05:50,760 Emergent behavior. 170 00:05:50,760 --> 00:05:54,240 So like, how individual neurons firing 171 00:05:54,240 --> 00:05:56,480 can lead to consciousness, but for AIs? 172 00:05:56,480 --> 00:05:57,320 Yeah, exactly. 173 00:05:57,320 --> 00:05:59,760 It's about understanding how complexity arises 174 00:05:59,760 --> 00:06:01,200 from these interactions. 175 00:06:01,200 --> 00:06:04,200 Next, CAML allows for dynamic communication. 176 00:06:04,200 --> 00:06:06,640 The agents can interact and work together in real time, 177 00:06:06,640 --> 00:06:08,760 so they can communicate, coordinate their actions, 178 00:06:08,760 --> 00:06:11,560 and really collaborate to achieve a shared goal. 179 00:06:11,560 --> 00:06:13,160 So they're not just working independently. 180 00:06:13,160 --> 00:06:15,640 They can chat and negotiate with each other. 181 00:06:15,640 --> 00:06:16,480 Exactly. 182 00:06:16,480 --> 00:06:18,440 It's all about teamwork for AIs. 183 00:06:18,440 --> 00:06:19,960 Then of course, there's stateful memory, 184 00:06:19,960 --> 00:06:22,040 which we talked about, lets them remember stuff 185 00:06:22,040 --> 00:06:23,720 and act more intelligently over time. 186 00:06:23,720 --> 00:06:24,800 Right, right. 187 00:06:24,800 --> 00:06:25,640 What else is there? 188 00:06:25,640 --> 00:06:26,480 Oh, this is important. 189 00:06:26,480 --> 00:06:28,520 Support for multiple benchmarks. 190 00:06:28,520 --> 00:06:30,400 This means researchers can test their agents 191 00:06:30,400 --> 00:06:32,360 on standard tasks and compare the results 192 00:06:32,360 --> 00:06:33,920 to others in a fair way. 193 00:06:33,920 --> 00:06:36,760 It's like standardized tests, but for AIs. 194 00:06:36,760 --> 00:06:39,040 Helps make sure everyone is measuring progress 195 00:06:39,040 --> 00:06:40,000 in the same way. 196 00:06:40,000 --> 00:06:42,720 So it's like standardized testing for AI societies. 197 00:06:42,720 --> 00:06:43,640 Makes sense. 198 00:06:43,640 --> 00:06:46,760 What other benefits does CAML have for researchers? 199 00:06:46,760 --> 00:06:48,880 So there's support for different agent types. 200 00:06:48,880 --> 00:06:50,840 CAML is super flexible. 201 00:06:50,840 --> 00:06:54,120 Researchers can use all sorts of agents, tasks, models, 202 00:06:54,120 --> 00:06:55,400 and environments. 203 00:06:55,400 --> 00:06:57,480 It can be used for all kinds of research questions 204 00:06:57,480 --> 00:06:59,000 in all kinds of fields. 205 00:06:59,000 --> 00:07:02,800 And lastly, there's data generation and tool integration. 206 00:07:02,800 --> 00:07:05,760 CAML can automatically create huge data sets, 207 00:07:05,760 --> 00:07:09,520 which is super helpful for training and testing AI models. 208 00:07:09,520 --> 00:07:11,920 It can also easily connect with other specialized tools, 209 00:07:11,920 --> 00:07:13,600 which makes the whole research process 210 00:07:13,600 --> 00:07:15,600 from start to finish way smoother. 211 00:07:15,600 --> 00:07:18,040 Sounds like CAML could really speed up AI research. 212 00:07:18,040 --> 00:07:19,760 Okay, so with all these features, 213 00:07:19,760 --> 00:07:21,200 what can you actually build with it? 214 00:07:21,200 --> 00:07:22,720 What are some applications they mentioned? 215 00:07:22,720 --> 00:07:25,000 They highlight three main areas. 216 00:07:25,000 --> 00:07:27,640 First, data generation, which we already talked about. 217 00:07:27,640 --> 00:07:29,720 You can use CAML to create tons of data 218 00:07:29,720 --> 00:07:32,480 to train and test AI models in all sorts of areas. 219 00:07:32,480 --> 00:07:33,840 Second is task automation. 220 00:07:33,840 --> 00:07:35,880 You can build systems with multiple AIs 221 00:07:35,880 --> 00:07:39,200 that can work together to automate really complex tasks. 222 00:07:39,200 --> 00:07:43,200 So imagine a team of AI assistants all working on a project, 223 00:07:43,200 --> 00:07:44,600 each with their own specialty. 224 00:07:44,600 --> 00:07:47,480 Exactly, they can each handle a different part of the task 225 00:07:47,480 --> 00:07:49,280 and work together to get it done. 226 00:07:49,280 --> 00:07:51,760 And third, world simulation. 227 00:07:51,760 --> 00:07:53,960 They can create simulated environments 228 00:07:53,960 --> 00:07:55,440 with AIs interacting, 229 00:07:55,440 --> 00:07:57,680 which allows them to study all sorts of things, 230 00:07:57,680 --> 00:08:01,120 like how markets work, how information spreads, 231 00:08:01,120 --> 00:08:02,880 or even how social rules develop. 232 00:08:02,880 --> 00:08:03,720 That's so cool. 233 00:08:03,720 --> 00:08:06,440 You can basically create virtual societies of AIs 234 00:08:06,440 --> 00:08:07,600 and see how they evolve. 235 00:08:07,600 --> 00:08:09,680 Yep, now let's get a bit more technical. 236 00:08:09,680 --> 00:08:11,160 They have a text stack section 237 00:08:11,160 --> 00:08:13,280 that describes all the different parts of CAML 238 00:08:13,280 --> 00:08:14,640 that make all this possible. 239 00:08:14,640 --> 00:08:15,720 Okay, let's hear it. 240 00:08:15,720 --> 00:08:18,120 First, there are the agents themselves, 241 00:08:18,120 --> 00:08:21,880 which is basically the code that defines how each AI acts. 242 00:08:21,880 --> 00:08:23,320 Then there's agent societies, 243 00:08:23,320 --> 00:08:26,320 which is the framework that helps manage all these AIs 244 00:08:26,320 --> 00:08:27,200 and their interactions. 245 00:08:27,200 --> 00:08:29,680 We already talked about the data generation module. 246 00:08:29,680 --> 00:08:31,240 Then there's the models module, 247 00:08:31,240 --> 00:08:33,880 which lets researchers customize the AI models 248 00:08:33,880 --> 00:08:35,800 that power the agents. 249 00:08:35,800 --> 00:08:37,520 Tool integration is really important, 250 00:08:37,520 --> 00:08:38,720 and they have a module for that, 251 00:08:38,720 --> 00:08:41,960 so the agents can use external tools for specific jobs. 252 00:08:41,960 --> 00:08:45,560 There's a memory module to handle the AI's memory. 253 00:08:45,560 --> 00:08:49,400 Oh, and they also have modules for storage, benchmarks, 254 00:08:49,400 --> 00:08:51,920 interpreters, data loaders, retrievers, 255 00:08:51,920 --> 00:08:53,040 which are used for this thing called 256 00:08:53,040 --> 00:08:55,880 retrieval augmented generation, a runtime environment, 257 00:08:55,880 --> 00:08:57,880 and even human in the loop components 258 00:08:57,880 --> 00:09:00,600 for when you need a human to step in and oversee things. 259 00:09:00,600 --> 00:09:01,960 Well, that's a lot. 260 00:09:01,960 --> 00:09:02,960 They've thought of everything. 261 00:09:02,960 --> 00:09:05,400 Right, it's a really well-designed system. 262 00:09:05,400 --> 00:09:07,640 Like we talked about, it's all community-driven. 263 00:09:07,640 --> 00:09:09,400 Yeah, that was really impressive. 264 00:09:09,400 --> 00:09:12,080 So what are they focused on researching with Camel, 265 00:09:12,080 --> 00:09:13,920 and how can someone join the party? 266 00:09:13,920 --> 00:09:16,160 So their main goal, as they say on their GitHub, 267 00:09:16,160 --> 00:09:18,640 is to learn more about how AI agents behave, 268 00:09:18,640 --> 00:09:21,520 what they can do, and any risks they might pose, 269 00:09:21,520 --> 00:09:23,680 all by studying them at a large scale. 270 00:09:23,680 --> 00:09:26,120 And they really want other people to use Camel 271 00:09:26,120 --> 00:09:28,440 for their own research and join their projects. 272 00:09:28,440 --> 00:09:29,840 They even give an email address 273 00:09:29,840 --> 00:09:31,320 for people who want to collaborate. 274 00:09:31,320 --> 00:09:33,320 It's all about working together. 275 00:09:33,320 --> 00:09:34,640 It's great that they're so open. 276 00:09:34,640 --> 00:09:37,200 Oh, and I saw they have a bunch of data sets on Huggingface. 277 00:09:37,200 --> 00:09:38,040 That's awesome. 278 00:09:38,040 --> 00:09:39,640 Oh, yeah, that's a great resource. 279 00:09:39,640 --> 00:09:42,680 They have data on AI interactions, code generation, 280 00:09:42,680 --> 00:09:46,080 math, physics, chemistry, even biology. 281 00:09:46,080 --> 00:09:48,400 And the data sets are in different formats, 282 00:09:48,400 --> 00:09:52,920 like chat logs, instructions, even translated conversations. 283 00:09:52,920 --> 00:09:55,560 They even have visualizations to help people understand 284 00:09:55,560 --> 00:09:56,400 the data. 285 00:09:56,400 --> 00:09:58,680 So researchers don't just get the tools, 286 00:09:58,680 --> 00:10:00,920 they get a bunch of data to play with, too. 287 00:10:00,920 --> 00:10:02,920 That's a great way to get people started. 288 00:10:02,920 --> 00:10:05,400 OK, now for the people who want to actually get their hands 289 00:10:05,400 --> 00:10:07,200 dirty, they have these cookbooks, right? 290 00:10:07,200 --> 00:10:09,920 Yeah, the cookbooks are like tutorials and guides 291 00:10:09,920 --> 00:10:11,080 for using Camel. 292 00:10:11,080 --> 00:10:12,640 They've got a bunch of different categories. 293 00:10:12,640 --> 00:10:15,080 There's basic concepts, which covers the basics, 294 00:10:15,080 --> 00:10:17,520 like how to create an agent, a group of agents, 295 00:10:17,520 --> 00:10:18,800 how they talk to each other. 296 00:10:18,800 --> 00:10:20,400 Then there's advanced features, which 297 00:10:20,400 --> 00:10:24,280 goes into more complex stuff, like integrating external tools, 298 00:10:24,280 --> 00:10:26,320 using memory, and implementing this thing 299 00:10:26,320 --> 00:10:28,760 called retrieval augmented generation. 300 00:10:28,760 --> 00:10:30,920 That involves using knowledge graphs 301 00:10:30,920 --> 00:10:32,880 to make the agents even smarter. 302 00:10:32,880 --> 00:10:35,920 So these advanced features are like the next level 303 00:10:35,920 --> 00:10:37,680 of multi-agent systems. 304 00:10:37,680 --> 00:10:38,680 Exactly. 305 00:10:38,680 --> 00:10:41,160 And they also have cookbooks on model training and data 306 00:10:41,160 --> 00:10:44,400 generation, where they show you how to create data with Camel 307 00:10:44,400 --> 00:10:48,000 and then fine-tune models using tools like Unsloth. 308 00:10:48,000 --> 00:10:50,800 They also show you how to work with specific data formats, 309 00:10:50,800 --> 00:10:53,560 generate data for this chain of thought, reasoning, 310 00:10:53,560 --> 00:10:56,000 and even how to upload your data sets to HuggingFace 311 00:10:56,000 --> 00:10:57,520 so everyone can use them. 312 00:10:57,520 --> 00:10:59,960 And then there's a section on multi-agent systems 313 00:10:59,960 --> 00:11:04,000 and applications with examples of how to build real-world things. 314 00:11:04,000 --> 00:11:05,800 They show you how to make web scrapers that 315 00:11:05,800 --> 00:11:07,800 can role play to get you information, 316 00:11:07,800 --> 00:11:10,440 build AI committees to judge hackathons, 317 00:11:10,440 --> 00:11:12,920 and even how to create customer service bots on Discord 318 00:11:12,920 --> 00:11:15,560 using both agentic retrieval augmented generation 319 00:11:15,560 --> 00:11:17,320 and local AI models. 320 00:11:17,320 --> 00:11:19,360 It's really amazing what you can do with it. 321 00:11:19,360 --> 00:11:20,680 Those are some cool examples. 322 00:11:20,680 --> 00:11:22,920 I like that they're showing people how to build practical stuff. 323 00:11:22,920 --> 00:11:24,440 What about that data processing section? 324 00:11:24,440 --> 00:11:25,360 What's that about? 325 00:11:25,360 --> 00:11:25,880 Oh, yeah. 326 00:11:25,880 --> 00:11:28,680 So that section is all about using agents 327 00:11:28,680 --> 00:11:30,680 to analyze video data, different ways 328 00:11:30,680 --> 00:11:34,920 to get and process data from websites using tools like FireCrawl, 329 00:11:34,920 --> 00:11:38,120 and even how to work with PDFs using tools like Chunker 330 00:11:38,120 --> 00:11:40,880 and these really powerful models for Mistral AI. 331 00:11:40,880 --> 00:11:42,560 They want to give people the skills to use 332 00:11:42,560 --> 00:11:44,360 Camel for real data problems. 333 00:11:44,360 --> 00:11:46,680 So they're really making Camel easy to use 334 00:11:46,680 --> 00:11:48,240 and showing its potential. 335 00:11:48,240 --> 00:11:50,680 What if someone wants to contribute to Camel itself? 336 00:11:50,680 --> 00:11:51,800 How do they do that? 337 00:11:51,800 --> 00:11:53,200 Well, on their GitHub page, they 338 00:11:53,200 --> 00:11:54,960 have a contributing to Camel section. 339 00:11:54,960 --> 00:11:56,880 They really encourage people to contribute code, 340 00:11:56,880 --> 00:11:59,160 and they have guidelines for how to do it. 341 00:11:59,160 --> 00:12:00,960 But you don't have to be a coder. 342 00:12:00,960 --> 00:12:04,000 They also say that just spreading the word about Camel, 343 00:12:04,000 --> 00:12:06,200 like talking about it on social media or ad events, 344 00:12:06,200 --> 00:12:07,680 is super helpful. 345 00:12:07,680 --> 00:12:09,360 So anyone can help coder or not. 346 00:12:09,360 --> 00:12:11,240 What about staying in touch with the Camel team? 347 00:12:11,240 --> 00:12:12,520 How can people reach out? 348 00:12:12,520 --> 00:12:13,520 They have a bunch of ways. 349 00:12:13,520 --> 00:12:16,280 You can use GitHub issues to report bugs, suggest features, 350 00:12:16,280 --> 00:12:18,160 and just see what's going on with the project. 351 00:12:18,160 --> 00:12:21,520 They have a Discord server for real-time help, discussions, 352 00:12:21,520 --> 00:12:24,240 and connecting with other people who are using Camel. 353 00:12:24,240 --> 00:12:27,440 They're also active on X, what used to be Twitter, 354 00:12:27,440 --> 00:12:29,560 where they share updates and announcements. 355 00:12:29,560 --> 00:12:31,600 And for the super passionate Camel fans, 356 00:12:31,600 --> 00:12:33,480 they even have an ambassador project. 357 00:12:33,480 --> 00:12:34,720 Wow, so many ways to connect. 358 00:12:34,720 --> 00:12:37,960 OK, so like any good research project, 359 00:12:37,960 --> 00:12:41,200 giving credit where credit is due is really important. 360 00:12:41,200 --> 00:12:43,440 Do they have info on how to cite Camel? 361 00:12:43,440 --> 00:12:45,080 Yes, they give a specific citation 362 00:12:45,080 --> 00:12:46,600 for their research paper, which was 363 00:12:46,600 --> 00:12:51,960 presented at this big AI conference called NURY PS in 2023. 364 00:12:51,960 --> 00:12:53,800 And they have a whole section on acknowledgments 365 00:12:53,800 --> 00:12:56,000 where they thank NOMIC AI for giving them access 366 00:12:56,000 --> 00:12:57,880 to their Atlas tool and Haya Hemud 367 00:12:57,880 --> 00:12:59,840 for creating the Camel logo. 368 00:12:59,840 --> 00:13:02,120 They also mentioned that they use ideas from other research, 369 00:13:02,120 --> 00:13:05,280 like Baby, AGI, Persona Hub, and Self-Instruct. 370 00:13:05,280 --> 00:13:07,480 And they ask people to cite those original papers 371 00:13:07,480 --> 00:13:10,680 if they use those specific parts of Camel in their own work. 372 00:13:10,680 --> 00:13:12,680 It's all about good research ethics. 373 00:13:12,680 --> 00:13:14,000 That's great to see. 374 00:13:14,000 --> 00:13:16,120 OK, so we've learned a lot about Camel today. 375 00:13:16,120 --> 00:13:18,400 What do you think is the most important thing about it? 376 00:13:18,400 --> 00:13:19,880 I think the biggest thing about Camel 377 00:13:19,880 --> 00:13:21,480 is that it gives us a powerful way 378 00:13:21,480 --> 00:13:26,320 to explore this whole new world of large-scale, multi-agent AI 379 00:13:26,320 --> 00:13:27,440 systems. 380 00:13:27,440 --> 00:13:29,080 It gives researchers the tools they 381 00:13:29,080 --> 00:13:33,160 need to understand how AIs behave when there are lots of them, 382 00:13:33,160 --> 00:13:35,440 how they work together, and how they interact 383 00:13:35,440 --> 00:13:37,040 in these shared spaces. 384 00:13:37,040 --> 00:13:39,440 It's not just about making AI more powerful. 385 00:13:39,440 --> 00:13:42,080 It's about understanding the complex dynamics that 386 00:13:42,080 --> 00:13:44,720 emerge when you have AI societies. 387 00:13:44,720 --> 00:13:46,960 And I think the fact that it's open source and community 388 00:13:46,960 --> 00:13:48,160 driven is super important. 389 00:13:48,160 --> 00:13:50,600 It makes this kind of cutting edge research accessible 390 00:13:50,600 --> 00:13:51,280 to everyone. 391 00:13:51,280 --> 00:13:52,040 Absolutely. 392 00:13:52,040 --> 00:13:55,840 It makes AI research more democratic and collaborative. 393 00:13:55,840 --> 00:13:58,640 The data sets, the documentation, the cookbooks, 394 00:13:58,640 --> 00:14:00,640 it all shows that they want to make Camel useful 395 00:14:00,640 --> 00:14:02,520 and have a real impact on AI research. 396 00:14:02,520 --> 00:14:05,120 OK, so for our listeners who are totally intrigued, 397 00:14:05,120 --> 00:14:07,320 where should they start if they want to learn more? 398 00:14:07,320 --> 00:14:09,800 Definitely check out the Camel GitHub page. 399 00:14:09,800 --> 00:14:12,160 Just search for Camel iCaml. 400 00:14:12,160 --> 00:14:14,480 You'll find the code, all the documentation, 401 00:14:14,480 --> 00:14:16,760 and links to all the community stuff we talked about. 402 00:14:16,760 --> 00:14:18,680 And if you want to see Camel in action, 403 00:14:18,680 --> 00:14:21,920 look for their demo of a conversation between two chat 404 00:14:21,920 --> 00:14:22,880 GPT agents. 405 00:14:22,880 --> 00:14:25,600 It's a really cool example of how these multi-agent 406 00:14:25,600 --> 00:14:27,080 interactions work. 407 00:14:27,080 --> 00:14:27,960 That sounds awesome. 408 00:14:27,960 --> 00:14:29,120 Thanks for all this info. 409 00:14:29,120 --> 00:14:31,080 And for our listeners, here's a final thought. 410 00:14:31,080 --> 00:14:33,280 As AI's get smarter and more connected 411 00:14:33,280 --> 00:14:36,200 and they start forming these complex digital societies, 412 00:14:36,200 --> 00:14:38,960 what kind of unexpected things will we discover about them 413 00:14:38,960 --> 00:14:40,760 using frameworks like Camel? 414 00:14:40,760 --> 00:14:42,560 And how will understanding all this 415 00:14:42,560 --> 00:14:45,680 change the way we develop AI and how it impacts our lives? 416 00:14:45,680 --> 00:14:47,360 That's something to think about. 417 00:14:47,360 --> 00:14:49,880 Thanks for joining us on the deep dive.