1 00:00:00,000 --> 00:00:04,920 Okay, so let's dive into this world of open source AI. 2 00:00:04,920 --> 00:00:05,760 Sounds good. 3 00:00:05,760 --> 00:00:08,760 We're taking a deep dive into the two U3. 4 00:00:08,760 --> 00:00:10,920 Pushing Frontiers in Open Language Model 5 00:00:10,920 --> 00:00:12,640 post-training research paper. 6 00:00:12,640 --> 00:00:13,600 Okay. 7 00:00:13,600 --> 00:00:14,960 I think you've probably heard whispers 8 00:00:14,960 --> 00:00:17,120 about how this model has kind of shaken things up. 9 00:00:17,120 --> 00:00:17,960 Right. 10 00:00:17,960 --> 00:00:19,520 And you're in the right place to find out why. 11 00:00:19,520 --> 00:00:21,960 Yeah, this is definitely a really interesting paper. 12 00:00:21,960 --> 00:00:24,000 And what's really capturing everyone's attention 13 00:00:24,000 --> 00:00:28,080 is how this model, which was built by just like a small team, 14 00:00:28,080 --> 00:00:31,000 is challenging some of the biggest names in AI. 15 00:00:31,000 --> 00:00:31,840 Oh, wow. 16 00:00:31,840 --> 00:00:34,480 Even those like super secret closed source ones. 17 00:00:34,480 --> 00:00:37,560 It's like a David and Goliath situation in the AI world. 18 00:00:37,560 --> 00:00:38,400 Exactly. 19 00:00:38,400 --> 00:00:41,600 So for those of us who aren't, you know, AI experts, 20 00:00:41,600 --> 00:00:43,200 what exactly is two U3? 21 00:00:43,200 --> 00:00:45,320 So imagine a family of language models 22 00:00:45,320 --> 00:00:48,400 that anyone can access and tinker with. 23 00:00:48,400 --> 00:00:50,360 That's what two U3 is. 24 00:00:50,360 --> 00:00:52,280 Open source AI. 25 00:00:52,280 --> 00:00:55,760 They started with a solid foundation called Lama 3.1. 26 00:00:55,760 --> 00:00:56,600 Okay. 27 00:00:56,600 --> 00:00:58,320 They added their own secret sauce 28 00:00:58,320 --> 00:01:01,080 to create something truly extraordinary. 29 00:01:01,080 --> 00:01:02,920 Okay, so they're not building from scratch. 30 00:01:02,920 --> 00:01:04,720 They're taking something that already exists 31 00:01:04,720 --> 00:01:05,560 and improving it. 32 00:01:05,560 --> 00:01:06,400 Exactly. 33 00:01:06,400 --> 00:01:07,240 Okay. 34 00:01:07,240 --> 00:01:09,360 Think of it like taking a basic cake recipe. 35 00:01:09,360 --> 00:01:10,200 Okay. 36 00:01:10,200 --> 00:01:12,320 And adding unique ingredients and techniques 37 00:01:12,320 --> 00:01:14,960 to transform it into a gourmet masterpiece. 38 00:01:14,960 --> 00:01:15,920 I like that analogy. 39 00:01:15,920 --> 00:01:16,760 Yeah. 40 00:01:16,760 --> 00:01:20,760 So in this case, the team used a four stage process 41 00:01:20,760 --> 00:01:22,240 to train two U3. 42 00:01:22,240 --> 00:01:23,080 Right. 43 00:01:23,080 --> 00:01:25,240 It starts with the called supervised fine tuning. 44 00:01:25,240 --> 00:01:26,080 Right. 45 00:01:26,080 --> 00:01:28,800 And they teach the model to follow instructions 46 00:01:28,800 --> 00:01:30,920 using really carefully selected data. 47 00:01:30,920 --> 00:01:34,320 Yeah, it's like showing the model good examples 48 00:01:34,320 --> 00:01:35,520 of how to respond. 49 00:01:35,520 --> 00:01:36,360 Okay. 50 00:01:36,360 --> 00:01:38,440 Like a teacher giving a student model answers. 51 00:01:38,440 --> 00:01:39,720 Yeah, like a study guide almost. 52 00:01:39,720 --> 00:01:40,560 Exactly. 53 00:01:40,560 --> 00:01:42,680 But here's where it gets really interesting. 54 00:01:42,680 --> 00:01:44,680 They went above and beyond to make sure 55 00:01:44,680 --> 00:01:46,800 that the data they used was top notch. 56 00:01:46,800 --> 00:01:47,640 Okay. 57 00:01:47,640 --> 00:01:48,480 They even developed tools 58 00:01:48,480 --> 00:01:51,280 to decontaminate existing data sets, 59 00:01:51,280 --> 00:01:54,400 removing any overlap with benchmark tests 60 00:01:54,400 --> 00:01:56,320 to avoid it inflated scores. 61 00:01:56,320 --> 00:01:57,640 So it's like making sure the students 62 00:01:57,640 --> 00:02:00,280 aren't sneaking peaks at the answer key before the test. 63 00:02:00,280 --> 00:02:01,120 Exactly. 64 00:02:01,120 --> 00:02:01,960 Yeah. 65 00:02:01,960 --> 00:02:03,720 That level of rigor is impressive. 66 00:02:03,720 --> 00:02:04,560 It is. 67 00:02:04,560 --> 00:02:05,400 It is. 68 00:02:05,400 --> 00:02:06,240 Yeah. 69 00:02:06,240 --> 00:02:07,060 They're committed to building a model 70 00:02:07,060 --> 00:02:10,200 that can truly generalize and tackle new challenges. 71 00:02:10,200 --> 00:02:11,040 Okay. 72 00:02:11,040 --> 00:02:12,400 After the supervised fine tuning, 73 00:02:12,400 --> 00:02:15,200 they move on to what's called preference fine tuning. 74 00:02:15,200 --> 00:02:16,040 Right. 75 00:02:16,040 --> 00:02:18,120 Which is all about refining the model's responses 76 00:02:18,120 --> 00:02:20,520 based on what humans prefer. 77 00:02:20,520 --> 00:02:23,240 So they're taking into account what we, 78 00:02:23,240 --> 00:02:27,440 like what we humans find helpful, interesting, or enjoyable. 79 00:02:27,440 --> 00:02:28,520 You got it. 80 00:02:28,520 --> 00:02:29,800 And for this stage, 81 00:02:29,800 --> 00:02:32,680 they used a fascinating mix of human 82 00:02:32,680 --> 00:02:34,640 and AI-generated feedback. 83 00:02:34,640 --> 00:02:35,720 Oh, wow. 84 00:02:35,720 --> 00:02:38,280 It's like having a panel of judges 85 00:02:38,280 --> 00:02:39,620 with different perspectives, 86 00:02:39,620 --> 00:02:41,320 helping the model to improve. 87 00:02:41,320 --> 00:02:42,360 That's really clever. 88 00:02:42,360 --> 00:02:43,200 Yeah. 89 00:02:43,200 --> 00:02:45,120 It's like they're combining the best of both worlds. 90 00:02:45,120 --> 00:02:48,000 Human intuition and AI's ability to process 91 00:02:48,000 --> 00:02:49,760 massive amounts of information. 92 00:02:49,760 --> 00:02:50,600 Right. 93 00:02:50,600 --> 00:02:51,760 What's the third stage of training? 94 00:02:51,760 --> 00:02:54,200 The third stage is where they really break new ground. 95 00:02:54,200 --> 00:02:55,040 Okay. 96 00:02:55,040 --> 00:02:56,280 It's called reinforcement learning 97 00:02:56,280 --> 00:02:59,560 with verifiable rewards, or RLVR for short. 98 00:02:59,560 --> 00:03:00,400 RLVR. 99 00:03:00,400 --> 00:03:01,240 Yeah. 100 00:03:01,240 --> 00:03:02,080 Okay. 101 00:03:02,080 --> 00:03:02,920 And this technique is designed for tasks 102 00:03:02,920 --> 00:03:04,640 with clear right or wrong answers, 103 00:03:04,640 --> 00:03:07,040 like math problems or coding challenges. 104 00:03:07,040 --> 00:03:08,660 So it's like giving the model points 105 00:03:08,660 --> 00:03:10,280 for every correct solution. 106 00:03:10,280 --> 00:03:12,520 Solutioning it to become a master problem solver. 107 00:03:12,520 --> 00:03:13,360 Right. 108 00:03:13,360 --> 00:03:14,200 Okay. 109 00:03:14,200 --> 00:03:16,160 And the final stage is all about transparency. 110 00:03:16,160 --> 00:03:17,000 Okay. 111 00:03:17,000 --> 00:03:18,920 They've made all of their data, code, 112 00:03:18,920 --> 00:03:20,960 and even their training recipes public. 113 00:03:20,960 --> 00:03:21,720 Wow. 114 00:03:21,720 --> 00:03:23,760 This is a huge deal in the AI world 115 00:03:23,760 --> 00:03:25,560 where things are often kept secret. 116 00:03:25,560 --> 00:03:26,400 Yeah. 117 00:03:26,400 --> 00:03:27,240 Very secretive. 118 00:03:27,240 --> 00:03:28,080 Yeah. 119 00:03:28,080 --> 00:03:30,120 So they're basically inviting everyone to explore, 120 00:03:30,120 --> 00:03:32,800 learn from, and build upon their work. 121 00:03:32,800 --> 00:03:33,640 Exactly. 122 00:03:33,640 --> 00:03:35,120 That's the essence of open source, right? 123 00:03:35,120 --> 00:03:35,960 That's it. 124 00:03:35,960 --> 00:03:36,800 Okay. 125 00:03:36,800 --> 00:03:37,640 Let's back up for a second. 126 00:03:37,640 --> 00:03:38,480 Sure. 127 00:03:38,480 --> 00:03:40,280 You mentioned how they decontaminated data. 128 00:03:40,280 --> 00:03:41,120 Right. 129 00:03:41,120 --> 00:03:42,240 What exactly does that mean, 130 00:03:42,240 --> 00:03:44,080 and why is so important? 131 00:03:44,080 --> 00:03:46,320 So it turns out that a lot of data sets 132 00:03:46,320 --> 00:03:49,840 used to train AI models contain information 133 00:03:49,840 --> 00:03:54,120 that overlaps with the benchmark tests used to evaluate them. 134 00:03:54,120 --> 00:03:54,960 Oh, okay. 135 00:03:54,960 --> 00:03:57,680 It's like giving a student the test questions in advance. 136 00:03:57,680 --> 00:03:58,600 Oh, I see, yeah. 137 00:03:58,600 --> 00:04:01,960 This contamination can lead to artificially inflated scores. 138 00:04:01,960 --> 00:04:02,800 Yeah. 139 00:04:02,800 --> 00:04:05,080 Making a model seem smarter than it actually is. 140 00:04:05,080 --> 00:04:05,920 Right. 141 00:04:05,920 --> 00:04:08,000 Ah, so they're ensuring that the model 142 00:04:08,000 --> 00:04:10,200 isn't just memorizing answers, 143 00:04:10,200 --> 00:04:12,520 but actually learning how to solve problems. 144 00:04:12,520 --> 00:04:14,200 Like that's a crucial distinction. 145 00:04:14,200 --> 00:04:15,200 Yeah, it is, yeah. 146 00:04:15,200 --> 00:04:19,320 Did they talk about how they chose the data for training? 147 00:04:19,320 --> 00:04:20,160 They did. 148 00:04:20,160 --> 00:04:21,000 Okay. 149 00:04:21,000 --> 00:04:22,280 It seems like that's a key factor 150 00:04:22,280 --> 00:04:24,120 in building a well-rounded AI. 151 00:04:24,120 --> 00:04:24,960 It is. 152 00:04:24,960 --> 00:04:27,680 They emphasized using data from a wide range of sources. 153 00:04:27,680 --> 00:04:28,520 Okay. 154 00:04:28,520 --> 00:04:31,080 They tapped into real user interactions with models. 155 00:04:31,080 --> 00:04:32,200 Oh, wow. 156 00:04:32,200 --> 00:04:35,120 Imagine learning from how people actually talk to AI. 157 00:04:35,120 --> 00:04:36,000 That's cool. 158 00:04:36,000 --> 00:04:38,800 They also included data sets for specific skills 159 00:04:38,800 --> 00:04:42,080 like math, coding, and even multiple languages. 160 00:04:42,080 --> 00:04:42,920 Nice. 161 00:04:42,920 --> 00:04:45,000 It's like giving the model a diverse education 162 00:04:45,000 --> 00:04:46,880 to make it as versatile as possible. 163 00:04:46,880 --> 00:04:47,960 That makes a lot of sense. 164 00:04:47,960 --> 00:04:48,800 Yeah. 165 00:04:48,800 --> 00:04:50,480 So, is there anything about how they generated 166 00:04:50,480 --> 00:04:51,840 some of their own data? 167 00:04:51,840 --> 00:04:52,680 They did, yeah. 168 00:04:52,680 --> 00:04:53,560 Okay. 169 00:04:53,560 --> 00:04:55,080 That sounds intriguing. 170 00:04:55,080 --> 00:04:55,920 It is, yeah. 171 00:04:55,920 --> 00:04:58,720 They actually used AI to generate some of the training data. 172 00:04:58,720 --> 00:04:59,680 Oh, wow. 173 00:04:59,680 --> 00:05:02,680 But they were careful to avoid creating repetitive patterns 174 00:05:02,680 --> 00:05:05,920 that can make AI-generated data less effective. 175 00:05:05,920 --> 00:05:06,760 Right. 176 00:05:06,760 --> 00:05:08,640 So they used a clever approach 177 00:05:08,640 --> 00:05:11,040 where they created different personas 178 00:05:11,040 --> 00:05:12,600 for the AI to role play. 179 00:05:12,600 --> 00:05:13,440 Oh, interesting. 180 00:05:13,440 --> 00:05:16,600 Like a machine learning researcher or a beginner coder. 181 00:05:16,600 --> 00:05:19,160 This helped to diversify the generated data 182 00:05:19,160 --> 00:05:20,720 and make it more realistic. 183 00:05:20,720 --> 00:05:22,640 So they've carefully curated their data. 184 00:05:22,640 --> 00:05:23,480 Right. 185 00:05:23,480 --> 00:05:25,040 Now let's talk about the specifics 186 00:05:25,040 --> 00:05:26,400 of how they train the model. 187 00:05:26,400 --> 00:05:27,240 Yeah. 188 00:05:27,240 --> 00:05:29,320 It seems like this multi-stage process 189 00:05:29,320 --> 00:05:32,040 is where their innovation really shines through. 190 00:05:32,040 --> 00:05:32,880 I think so too. 191 00:05:32,880 --> 00:05:35,760 In the first stage, supervised fine tuning, 192 00:05:35,760 --> 00:05:38,000 they collect a massive amount of data 193 00:05:38,000 --> 00:05:40,360 containing prompts or instructions 194 00:05:40,360 --> 00:05:42,120 paired with the desired responses. 195 00:05:42,120 --> 00:05:42,960 Okay. 196 00:05:42,960 --> 00:05:44,520 It's like giving the model examples 197 00:05:44,520 --> 00:05:46,640 of good conversations to learn from. 198 00:05:46,640 --> 00:05:47,480 I see. 199 00:05:47,480 --> 00:05:49,720 And this is where all those diverse data sources 200 00:05:49,720 --> 00:05:50,880 come in handy, right? 201 00:05:50,880 --> 00:05:51,720 Yeah. 202 00:05:51,720 --> 00:05:54,480 Real user interactions, skill-specific data sets, 203 00:05:54,480 --> 00:05:56,360 and even the persona-generated data. 204 00:05:56,360 --> 00:05:57,200 All coming together. 205 00:05:57,200 --> 00:05:59,400 Do not just throwing everything in at random, are they? 206 00:05:59,400 --> 00:06:01,200 They're not, no, they're very strategic about it. 207 00:06:01,200 --> 00:06:02,040 Okay. 208 00:06:02,040 --> 00:06:03,280 They refine the data mix 209 00:06:03,280 --> 00:06:05,600 through multiple rounds of training, 210 00:06:05,600 --> 00:06:08,280 tweaking the recipe as they go to get the best results. 211 00:06:08,280 --> 00:06:10,240 So it's a very iterative process. 212 00:06:10,240 --> 00:06:11,080 It is. 213 00:06:11,080 --> 00:06:12,480 Constantly adjusting and improving. 214 00:06:12,480 --> 00:06:13,320 Yeah. 215 00:06:13,320 --> 00:06:14,560 What kind of results did they see 216 00:06:14,560 --> 00:06:16,440 after this first stage of training? 217 00:06:16,440 --> 00:06:18,360 Well, the paper includes a table 218 00:06:18,360 --> 00:06:20,720 that compares the performance of their model 219 00:06:20,720 --> 00:06:24,520 called 2U38B-SFT. 220 00:06:24,520 --> 00:06:25,360 Okay. 221 00:06:25,360 --> 00:06:27,920 Against other open source and even some closed models. 222 00:06:27,920 --> 00:06:28,760 Okay. 223 00:06:28,760 --> 00:06:30,280 And let me tell you, the results are impressive. 224 00:06:30,280 --> 00:06:32,400 I'm guessing 2U3 came out on top? 225 00:06:32,400 --> 00:06:33,240 You bet. 226 00:06:33,240 --> 00:06:34,060 Nice. 227 00:06:34,060 --> 00:06:37,120 It's significantly outperformed prior state-of-the-art models 228 00:06:37,120 --> 00:06:38,960 across a range of benchmarks, 229 00:06:38,960 --> 00:06:42,760 including MMEO, Truthful QA, and GSMA-K, 230 00:06:42,760 --> 00:06:45,200 which tests math problem solving skills. 231 00:06:45,200 --> 00:06:47,600 Wow, those are some tough benchmarks to beat. 232 00:06:47,600 --> 00:06:48,440 They are. 233 00:06:48,440 --> 00:06:50,360 That really speaks to the effectiveness 234 00:06:50,360 --> 00:06:54,320 of their data curation and multi-stage training approach. 235 00:06:54,320 --> 00:06:56,160 Yeah, they're really putting a lot of thought into this. 236 00:06:56,160 --> 00:06:57,000 Yeah. 237 00:06:57,000 --> 00:06:59,520 But I'm curious about this length normalized, 238 00:06:59,520 --> 00:07:01,480 direct preference optimization. 239 00:07:01,480 --> 00:07:02,320 Okay. 240 00:07:02,320 --> 00:07:03,160 What is that exactly? 241 00:07:03,160 --> 00:07:04,000 Sounds pretty technical, right? 242 00:07:04,000 --> 00:07:04,840 Yeah, a little bit. 243 00:07:04,840 --> 00:07:07,400 Well, direct preference optimization, or DPO, 244 00:07:07,400 --> 00:07:08,800 is the key algorithm they use 245 00:07:08,800 --> 00:07:10,560 for the second stage of training. 246 00:07:10,560 --> 00:07:11,400 Right. 247 00:07:11,400 --> 00:07:12,720 Preference fine tuning. 248 00:07:12,720 --> 00:07:15,640 Remember, this stage is all about making the AI's responses 249 00:07:15,640 --> 00:07:17,640 more aligned with human preferences. 250 00:07:17,640 --> 00:07:19,840 Right, so they're essentially teaching the model 251 00:07:19,840 --> 00:07:22,880 to be more human-like in its communication. 252 00:07:22,880 --> 00:07:23,720 Yeah. 253 00:07:23,720 --> 00:07:24,560 Okay. 254 00:07:24,560 --> 00:07:26,240 How does this DPO method work? 255 00:07:26,240 --> 00:07:30,920 It's a way of training the AI directly on human preferences 256 00:07:30,920 --> 00:07:33,800 without needing to create a separate reward model, 257 00:07:33,800 --> 00:07:36,160 which can be very complex and time-consuming. 258 00:07:36,160 --> 00:07:37,000 I see. 259 00:07:37,000 --> 00:07:39,640 The length normalized part is a clever tweak they made 260 00:07:39,640 --> 00:07:42,280 to the algorithm to make it even more effective. 261 00:07:42,280 --> 00:07:44,080 So they're simplifying the training process 262 00:07:44,080 --> 00:07:45,560 while making it more effective. 263 00:07:45,560 --> 00:07:46,440 Exactly. 264 00:07:46,440 --> 00:07:47,280 That's impressive. 265 00:07:47,280 --> 00:07:48,360 Yeah, they're pretty clever. 266 00:07:48,360 --> 00:07:49,640 What kind of data do they use 267 00:07:49,640 --> 00:07:52,080 for this preference fine tuning stage? 268 00:07:52,080 --> 00:07:53,520 They use a mix of sources. 269 00:07:53,520 --> 00:07:54,360 Okay. 270 00:07:54,360 --> 00:07:56,680 But they really focus on capturing human judgments 271 00:07:56,680 --> 00:07:59,280 about the quality of the AI's responses. 272 00:07:59,280 --> 00:08:00,120 Okay. 273 00:08:00,120 --> 00:08:01,240 They even developed a system 274 00:08:01,240 --> 00:08:04,760 where they use different AI models as judges 275 00:08:04,760 --> 00:08:07,840 to rate the responses generated by their model. 276 00:08:07,840 --> 00:08:10,560 Wait, they're having AI judge other AI. 277 00:08:10,560 --> 00:08:12,200 Yeah, it's a little bit meadow, isn't it? 278 00:08:12,200 --> 00:08:13,040 It is, yeah. 279 00:08:13,040 --> 00:08:15,040 But they're essentially leveraging the capabilities 280 00:08:15,040 --> 00:08:18,440 of different AI models to provide diverse perspectives 281 00:08:18,440 --> 00:08:20,920 on the quality of their model's responses. 282 00:08:20,920 --> 00:08:23,160 It's like having a panel of expert judges 283 00:08:23,160 --> 00:08:24,880 with different areas of expertise. 284 00:08:24,880 --> 00:08:27,200 That's an interesting way to get a more well-rounded 285 00:08:27,200 --> 00:08:29,160 assessment of the AI's performance. 286 00:08:30,120 --> 00:08:30,960 It is. 287 00:08:30,960 --> 00:08:32,400 Do they mention what kind of improvements they saw 288 00:08:32,400 --> 00:08:33,960 after using this DPO method? 289 00:08:33,960 --> 00:08:36,760 They found that DPO led to significant improvements 290 00:08:36,760 --> 00:08:39,320 in several areas, including Alpacaevigl, 291 00:08:39,320 --> 00:08:42,520 which tests the model's ability to follow instructions, 292 00:08:42,520 --> 00:08:45,440 and Safety, which evaluates its ability 293 00:08:45,440 --> 00:08:48,640 to avoid generating harmful or offensive content. 294 00:08:48,640 --> 00:08:51,600 So after two rounds of intense training, 295 00:08:51,600 --> 00:08:53,640 they've got a model that's not only smart, 296 00:08:53,640 --> 00:08:55,680 but also safe and well-behaved. 297 00:08:55,680 --> 00:08:56,520 That's right. 298 00:08:56,520 --> 00:08:57,920 Now let's move on to their final. 299 00:08:57,920 --> 00:08:58,760 Okay. 300 00:08:58,760 --> 00:09:01,040 And perhaps most innovative stage of training. 301 00:09:01,040 --> 00:09:04,880 Reinforcement learning with verifiable rewards, 302 00:09:04,880 --> 00:09:06,440 or RLVR. 303 00:09:06,440 --> 00:09:07,280 Right. 304 00:09:07,280 --> 00:09:08,760 This one sounds like it's pushing the boundaries 305 00:09:08,760 --> 00:09:10,720 of what's possible with AI training. 306 00:09:10,720 --> 00:09:11,560 It is. 307 00:09:11,560 --> 00:09:12,600 It's where it gets really interesting. 308 00:09:12,600 --> 00:09:15,360 RLVR is all about focusing on tasks 309 00:09:15,360 --> 00:09:16,840 where we can clearly determine 310 00:09:16,840 --> 00:09:18,400 if the AI got the answer right. 311 00:09:18,400 --> 00:09:19,240 Okay. 312 00:09:19,240 --> 00:09:20,800 Like math problems or coding challenges. 313 00:09:20,800 --> 00:09:22,760 Right, like a very specific grading rubric. 314 00:09:22,760 --> 00:09:23,600 Exactly. 315 00:09:23,600 --> 00:09:25,160 No room for subjectivity here. 316 00:09:25,160 --> 00:09:26,000 Right. 317 00:09:26,000 --> 00:09:26,840 How does it actually work? 318 00:09:26,840 --> 00:09:30,240 Well, in RLVR, the AI receives a reward. 319 00:09:30,240 --> 00:09:31,400 Think of it like a point. 320 00:09:31,400 --> 00:09:32,240 Okay. 321 00:09:32,240 --> 00:09:33,640 Only when it produces a correct solution. 322 00:09:33,640 --> 00:09:34,480 Right. 323 00:09:34,480 --> 00:09:36,640 This reward system helps to sharpen 324 00:09:36,640 --> 00:09:38,480 the AI's focus on accuracy. 325 00:09:38,480 --> 00:09:39,320 Okay. 326 00:09:39,320 --> 00:09:40,560 Making it a master problem solver. 327 00:09:40,560 --> 00:09:41,400 And they're training it 328 00:09:41,400 --> 00:09:42,960 on some challenging problems, right? 329 00:09:42,960 --> 00:09:43,800 They are. 330 00:09:43,800 --> 00:09:46,720 They specifically focus on three domains. 331 00:09:46,720 --> 00:09:49,680 Math problems from the GSM 8K dataset. 332 00:09:49,680 --> 00:09:50,520 Okay. 333 00:09:50,520 --> 00:09:53,000 More complex math problems from the math dataset. 334 00:09:53,000 --> 00:09:54,160 Wow. 335 00:09:54,160 --> 00:09:55,720 And precise instruction following 336 00:09:55,720 --> 00:09:57,240 from the IEVL dataset. 337 00:09:57,240 --> 00:09:59,720 So they're covering a broad spectrum of tasks 338 00:09:59,720 --> 00:10:03,160 that require accurate solutions and logical reasoning. 339 00:10:03,160 --> 00:10:04,000 Yeah. 340 00:10:04,000 --> 00:10:05,200 Did they encounter any hurdles 341 00:10:05,200 --> 00:10:07,360 while implementing this RLVR technique? 342 00:10:07,360 --> 00:10:08,200 They did. 343 00:10:08,200 --> 00:10:09,040 Yeah. 344 00:10:09,040 --> 00:10:10,160 Training a large language model 345 00:10:10,160 --> 00:10:11,320 with reinforcement learning 346 00:10:11,320 --> 00:10:13,560 can be very computationally expensive. 347 00:10:13,560 --> 00:10:14,640 Yeah, that makes sense. 348 00:10:14,640 --> 00:10:16,120 It takes a lot of processing power 349 00:10:16,120 --> 00:10:18,240 to teach a massive AI system 350 00:10:18,240 --> 00:10:20,800 by giving it points for every correct answer. 351 00:10:20,800 --> 00:10:21,640 Right. 352 00:10:21,640 --> 00:10:23,320 So what do they do to overcome this challenge? 353 00:10:23,320 --> 00:10:25,320 Well, they built a specialized system 354 00:10:25,320 --> 00:10:27,680 to run the RLVR training process. 355 00:10:27,680 --> 00:10:28,520 Okay. 356 00:10:28,520 --> 00:10:30,080 This system leverages a technique 357 00:10:30,080 --> 00:10:32,200 called asynchronous RL training. 358 00:10:32,200 --> 00:10:33,040 Okay. 359 00:10:33,040 --> 00:10:33,880 Which improves efficiency 360 00:10:33,880 --> 00:10:36,040 by allowing different parts of the training process 361 00:10:36,040 --> 00:10:37,240 to run concurrently. 362 00:10:37,240 --> 00:10:38,840 So it's like having multiple coaches 363 00:10:38,840 --> 00:10:40,480 working with an athlete at the same time. 364 00:10:40,480 --> 00:10:41,360 Exactly. 365 00:10:41,360 --> 00:10:42,920 A much faster way to train. 366 00:10:42,920 --> 00:10:44,040 What were the outcomes 367 00:10:44,040 --> 00:10:47,000 of this final stage of training? 368 00:10:47,000 --> 00:10:48,280 The results are impressive. 369 00:10:48,280 --> 00:10:49,120 Okay. 370 00:10:49,120 --> 00:10:51,680 They found that RLVR further enhanced 371 00:10:51,680 --> 00:10:54,840 the model's performance on those targeted tasks, 372 00:10:54,840 --> 00:10:59,280 particularly on GSMAK, math, and IEVL. 373 00:10:59,280 --> 00:11:00,120 Is that working? 374 00:11:00,120 --> 00:11:00,960 It's working. 375 00:11:00,960 --> 00:11:01,800 Okay. 376 00:11:01,800 --> 00:11:03,120 This shows that their approach is effective 377 00:11:03,120 --> 00:11:05,400 for both general language tasks 378 00:11:05,400 --> 00:11:07,320 and those requiring precise reasoning. 379 00:11:07,320 --> 00:11:08,160 That's incredible. 380 00:11:08,160 --> 00:11:09,000 It is. 381 00:11:09,000 --> 00:11:10,760 So they've taken this open source AI, 382 00:11:10,760 --> 00:11:13,600 put it through this really rigorous training program, 383 00:11:13,600 --> 00:11:16,760 and it's performing at a level that rivals 384 00:11:16,760 --> 00:11:20,000 or even surpasses some of the big names in AI. 385 00:11:20,000 --> 00:11:20,840 You're absolutely right. 386 00:11:20,840 --> 00:11:21,800 And it all comes back 387 00:11:21,800 --> 00:11:23,920 to their commitment to transparency. 388 00:11:23,920 --> 00:11:26,240 They've openly shared their data, 389 00:11:26,240 --> 00:11:28,720 code, and training methods, 390 00:11:28,720 --> 00:11:31,080 which is a game changer in the world of AI. 391 00:11:31,080 --> 00:11:31,920 It is. 392 00:11:31,920 --> 00:11:34,680 They're essentially inviting the entire AI community 393 00:11:34,680 --> 00:11:37,760 to learn from their work and build upon it. 394 00:11:37,760 --> 00:11:38,600 It's like they're saying, 395 00:11:38,600 --> 00:11:41,520 let's all work together to build a better AI future. 396 00:11:41,520 --> 00:11:42,360 Exactly. 397 00:11:42,360 --> 00:11:43,600 But hold on a second. 398 00:11:43,600 --> 00:11:46,320 We've talked a lot about training and data. 399 00:11:46,320 --> 00:11:47,160 Right. 400 00:11:47,160 --> 00:11:50,560 But what about evaluating 2U3's performance? 401 00:11:50,560 --> 00:11:51,400 Oh, that's a great point. 402 00:11:51,400 --> 00:11:53,240 How do we know it actually works? 403 00:11:53,240 --> 00:11:54,600 You raise a great point. 404 00:11:54,600 --> 00:11:58,240 Evaluation is a critical part of any AI research. 405 00:11:58,240 --> 00:12:02,360 And the 2U3 team didn't just rely on standard benchmarks. 406 00:12:02,360 --> 00:12:03,360 Okay. 407 00:12:03,360 --> 00:12:05,080 They went a step further 408 00:12:05,080 --> 00:12:07,440 and created a new evaluation framework 409 00:12:07,440 --> 00:12:10,160 called the Open Language Model Evaluation System, 410 00:12:10,160 --> 00:12:11,000 or Ulmes. 411 00:12:11,000 --> 00:12:11,840 Ulmes. 412 00:12:11,840 --> 00:12:12,680 Yeah. 413 00:12:12,680 --> 00:12:14,520 Okay, so it's like designing a standardized testing system 414 00:12:14,520 --> 00:12:17,760 for AI to ensure that everyone's playing by the same rules. 415 00:12:17,760 --> 00:12:18,600 Exactly. 416 00:12:18,600 --> 00:12:19,600 What's special about Ulmes? 417 00:12:19,600 --> 00:12:22,640 Ulmes includes a collection of carefully selected benchmarks. 418 00:12:22,640 --> 00:12:23,480 Okay. 419 00:12:23,480 --> 00:12:25,840 Covering a diverse range of tasks and skills. 420 00:12:25,840 --> 00:12:26,680 Right. 421 00:12:26,680 --> 00:12:28,000 But here's the key. 422 00:12:28,000 --> 00:12:31,200 They divided the evaluation suite into two sets. 423 00:12:31,200 --> 00:12:32,040 Oh, okay. 424 00:12:32,040 --> 00:12:34,520 One set, called the Development Suite, 425 00:12:34,520 --> 00:12:36,840 is used throughout the training process 426 00:12:36,840 --> 00:12:38,800 to monitor the model's progress 427 00:12:38,800 --> 00:12:40,560 and make adjustments as needed. 428 00:12:40,560 --> 00:12:43,680 So it's like giving the AI regular quizzes 429 00:12:43,680 --> 00:12:45,720 and practice tests throughout its training. 430 00:12:45,720 --> 00:12:46,560 Exactly. 431 00:12:46,560 --> 00:12:48,480 But what about the other set of evaluations? 432 00:12:48,480 --> 00:12:51,240 The second set, called the Unseen Suite, 433 00:12:51,240 --> 00:12:53,040 is kept hidden until the very end. 434 00:12:53,040 --> 00:12:53,880 Okay. 435 00:12:53,880 --> 00:12:54,720 It's like a final exam. 436 00:12:54,720 --> 00:12:55,560 Oh. 437 00:12:55,560 --> 00:12:57,760 Where the AI has never seen the questions before. 438 00:12:57,760 --> 00:12:58,640 Okay. 439 00:12:58,640 --> 00:13:01,480 This is the ultimate test of the model's ability 440 00:13:01,480 --> 00:13:03,960 to generalize and solve problems 441 00:13:03,960 --> 00:13:06,200 it hasn't specifically been trained on. 442 00:13:06,200 --> 00:13:08,920 That's a clever way to assess its true capabilities. 443 00:13:08,920 --> 00:13:09,760 It is. 444 00:13:09,760 --> 00:13:11,000 What kind of challenges do they include 445 00:13:11,000 --> 00:13:12,560 in this Unseen Suite? 446 00:13:12,560 --> 00:13:15,560 They included a range of new and challenging tasks, 447 00:13:15,560 --> 00:13:19,480 like AGEvol, which focuses on complex reasoning, 448 00:13:19,480 --> 00:13:22,560 and HREF, which evaluates the model's ability 449 00:13:22,560 --> 00:13:26,080 to follow instructions across different categories. 450 00:13:26,080 --> 00:13:26,920 Wow. 451 00:13:26,920 --> 00:13:28,560 They really wanted to push the model to its limits 452 00:13:28,560 --> 00:13:29,400 and see what it could do. 453 00:13:29,400 --> 00:13:31,840 So they didn't just create this powerful AI. 454 00:13:31,840 --> 00:13:32,680 Right. 455 00:13:32,680 --> 00:13:35,360 They also developed a really rigorous way 456 00:13:35,360 --> 00:13:37,040 to assess its capabilities. 457 00:13:37,040 --> 00:13:37,880 That's right. 458 00:13:37,880 --> 00:13:39,920 That's a commendable approach to research. 459 00:13:39,920 --> 00:13:40,760 It is. 460 00:13:40,760 --> 00:13:42,840 And in the spirit of transparency, 461 00:13:42,840 --> 00:13:45,920 they even shared some insights into what didn't work 462 00:13:45,920 --> 00:13:47,560 during the development process. 463 00:13:47,560 --> 00:13:48,440 Oh, interesting. 464 00:13:48,440 --> 00:13:50,880 For example, they experimented with different methods 465 00:13:50,880 --> 00:13:54,040 for generating data for preference fine tuning, 466 00:13:54,040 --> 00:13:56,200 but found that some approaches weren't as effective 467 00:13:56,200 --> 00:13:57,040 as others. 468 00:13:57,040 --> 00:14:00,360 So they're not just showcasing their successes. 469 00:14:00,360 --> 00:14:02,120 They're also highlighting their challenges 470 00:14:02,120 --> 00:14:03,920 and the lessons they learned along the way. 471 00:14:03,920 --> 00:14:04,520 Exactly. 472 00:14:04,520 --> 00:14:06,880 That's incredibly valuable for other researchers 473 00:14:06,880 --> 00:14:08,720 and developers building on their work. 474 00:14:08,720 --> 00:14:09,600 It is. 475 00:14:09,600 --> 00:14:12,640 And they also discussed areas for future research, 476 00:14:12,640 --> 00:14:15,320 like improving the model's ability to handle 477 00:14:15,320 --> 00:14:19,400 longer conversations, expanding its multilingual capability, 478 00:14:19,400 --> 00:14:22,560 and exploring ways to integrate it with tools and agents. 479 00:14:22,560 --> 00:14:24,200 It sounds like they've only just begun 480 00:14:24,200 --> 00:14:29,520 to explore the possibilities of open source AI with 2U3. 481 00:14:29,520 --> 00:14:30,680 It does, yeah. 482 00:14:30,680 --> 00:14:33,520 This research is truly groundbreaking, 483 00:14:33,520 --> 00:14:35,400 not just for the technical advancements, 484 00:14:35,400 --> 00:14:39,120 but also for their commitment to transparency and collaboration. 485 00:14:39,120 --> 00:14:39,680 Yeah. 486 00:14:39,680 --> 00:14:42,520 It's an exciting glimpse into the future of AI. 487 00:14:42,520 --> 00:14:43,160 It is. 488 00:14:43,160 --> 00:14:44,480 It is. 489 00:14:44,480 --> 00:14:45,640 Welcome back. 490 00:14:45,640 --> 00:14:48,720 I'm ready to hear how well 2U3 did on those hidden tests. 491 00:14:48,720 --> 00:14:49,240 Right. 492 00:14:49,240 --> 00:14:51,400 It's like waiting for the results of a big exam. 493 00:14:51,400 --> 00:14:52,800 Well, the results are in. 494 00:14:52,800 --> 00:14:54,160 And it's safe. 495 00:14:54,160 --> 00:14:57,040 I have to say that 2U3 passed with flying colors. 496 00:14:57,040 --> 00:14:58,360 OK, that's a relief. 497 00:14:58,360 --> 00:14:58,720 Yeah. 498 00:14:58,720 --> 00:15:00,400 But I'm guessing it wasn't a perfect score. 499 00:15:00,400 --> 00:15:01,080 Right. 500 00:15:01,080 --> 00:15:04,200 Were there any areas where 2U3 struggled? 501 00:15:04,200 --> 00:15:06,400 So there were a few interesting observations. 502 00:15:06,400 --> 00:15:06,800 OK. 503 00:15:06,800 --> 00:15:09,840 You see, while 2U3 excelled at math problems 504 00:15:09,840 --> 00:15:12,240 in the development suite, its performance 505 00:15:12,240 --> 00:15:15,760 dipped a bit on a similar but unseen set of math problems 506 00:15:15,760 --> 00:15:17,760 called deep-mind mathematics. 507 00:15:17,760 --> 00:15:20,640 So even though it aced the practice tests, 508 00:15:20,640 --> 00:15:23,280 it stumbled a bit when faced with brand new challenges. 509 00:15:23,280 --> 00:15:24,120 Exactly. 510 00:15:24,120 --> 00:15:25,280 Why do you think that happened? 511 00:15:25,280 --> 00:15:28,200 Well, the researchers noticed that 2U3 had a tendency 512 00:15:28,200 --> 00:15:30,400 to over-apply formatting rules. 513 00:15:30,400 --> 00:15:30,960 OK. 514 00:15:30,960 --> 00:15:33,040 It had learned during training. 515 00:15:33,040 --> 00:15:35,280 For instance, in the math data set, 516 00:15:35,280 --> 00:15:38,720 answers are often expected in latex format, 517 00:15:38,720 --> 00:15:41,800 which is a specific way of writing mathematical formulas. 518 00:15:41,800 --> 00:15:42,560 I see. 519 00:15:42,560 --> 00:15:45,720 2U3 seemed to apply this formatting even when 520 00:15:45,720 --> 00:15:48,720 it wasn't necessary, and that sometimes tripped it up. 521 00:15:48,720 --> 00:15:52,160 It's like a student who's so focused on using perfect grammar 522 00:15:52,160 --> 00:15:54,120 that they forget to answer the question itself. 523 00:15:54,120 --> 00:15:54,960 Exactly. 524 00:15:54,960 --> 00:15:55,600 That's funny. 525 00:15:55,600 --> 00:16:00,360 It highlights how even subtle aspects of the training data 526 00:16:00,360 --> 00:16:02,280 can influence the model's behavior. 527 00:16:02,280 --> 00:16:02,840 Right. 528 00:16:02,840 --> 00:16:05,720 Yeah, because it's picking up on these patterns, 529 00:16:05,720 --> 00:16:07,040 but not necessarily understanding 530 00:16:07,040 --> 00:16:08,240 like the underlying concept. 531 00:16:08,240 --> 00:16:08,880 Exactly. 532 00:16:08,880 --> 00:16:10,800 Were there any other unexpected findings 533 00:16:10,800 --> 00:16:12,360 in these unseen evaluations? 534 00:16:12,360 --> 00:16:13,160 Yes. 535 00:16:13,160 --> 00:16:15,720 They found a significant difference in performance 536 00:16:15,720 --> 00:16:20,080 between iFEVL, a benchmark they used during development, 537 00:16:20,080 --> 00:16:23,560 and a similar but unseen benchmark they created called 538 00:16:23,560 --> 00:16:24,800 iFEVLode. 539 00:16:24,800 --> 00:16:25,200 OK. 540 00:16:25,200 --> 00:16:27,800 I remember we talked about out of distribution tasks. 541 00:16:27,800 --> 00:16:28,240 Right. 542 00:16:28,240 --> 00:16:31,160 So iFEVLode is designed to test how well the model handles 543 00:16:31,160 --> 00:16:32,840 instructions it's never seen before. 544 00:16:32,840 --> 00:16:33,800 Exactly. 545 00:16:33,800 --> 00:16:37,800 Both benchmarks focus on precise instruction following, 546 00:16:37,800 --> 00:16:40,680 meaning the AI has to understand and execute 547 00:16:40,680 --> 00:16:42,840 very detailed instructions. 548 00:16:42,840 --> 00:16:43,080 Right. 549 00:16:43,080 --> 00:16:46,240 So it's like giving the AI a very specific recipe 550 00:16:46,240 --> 00:16:47,920 and seeing if it can follow it perfectly. 551 00:16:47,920 --> 00:16:48,480 Exactly. 552 00:16:48,480 --> 00:16:50,120 So how did two year three do? 553 00:16:50,120 --> 00:16:51,520 Well, here's the surprising part. 554 00:16:51,520 --> 00:16:52,040 OK. 555 00:16:52,040 --> 00:16:55,280 Even models that did exceptionally well on iFEVL, 556 00:16:55,280 --> 00:16:59,160 including two year three, saw a noticeable drop 557 00:16:59,160 --> 00:17:01,440 in performance on iFEVLode. 558 00:17:01,440 --> 00:17:04,080 So it's like they ace the practice test 559 00:17:04,080 --> 00:17:06,080 but struggle with the real exam. 560 00:17:06,080 --> 00:17:06,560 Right. 561 00:17:06,560 --> 00:17:07,480 What does that tell us? 562 00:17:07,480 --> 00:17:10,280 Well, it suggests that even though the model appears 563 00:17:10,280 --> 00:17:13,280 to be mastering precise instruction following, 564 00:17:13,280 --> 00:17:16,200 it might be overfitting to the specific examples 565 00:17:16,200 --> 00:17:17,680 it encountered during training. 566 00:17:17,680 --> 00:17:18,080 OK. 567 00:17:18,080 --> 00:17:19,880 In other words, it's not truly grasping 568 00:17:19,880 --> 00:17:21,960 the underlying principles, which 569 00:17:21,960 --> 00:17:24,880 makes it harder to generalize to new instructions. 570 00:17:24,880 --> 00:17:27,840 It seems like the AI is still learning by rote to some extent 571 00:17:27,840 --> 00:17:30,160 rather than developing a deeper understanding of how 572 00:17:30,160 --> 00:17:31,480 to follow instructions. 573 00:17:31,480 --> 00:17:32,680 That seems to be the case. 574 00:17:32,680 --> 00:17:33,000 OK. 575 00:17:33,000 --> 00:17:35,560 And it highlights an ongoing challenge in AI, 576 00:17:35,560 --> 00:17:38,440 developing models that can truly generalize and learn 577 00:17:38,440 --> 00:17:41,360 like humans do, adapting to new situations 578 00:17:41,360 --> 00:17:43,120 and applying knowledge flexibly. 579 00:17:43,120 --> 00:17:44,640 A good reality check that we're still 580 00:17:44,640 --> 00:17:46,040 in the early stages of AI. 581 00:17:46,040 --> 00:17:46,520 Yeah. 582 00:17:46,520 --> 00:17:47,640 And there's a lot more to learn. 583 00:17:47,640 --> 00:17:48,120 Right. 584 00:17:48,120 --> 00:17:52,400 But even with those limitations, the results of 2U3's 585 00:17:52,400 --> 00:17:54,720 evaluations are still pretty impressive. 586 00:17:54,720 --> 00:17:55,640 They are. 587 00:17:55,640 --> 00:18:00,000 They show that open source AI can be just as good as or even 588 00:18:00,000 --> 00:18:02,880 better than proprietary models in many areas. 589 00:18:02,880 --> 00:18:03,560 Absolutely. 590 00:18:03,560 --> 00:18:06,640 And the team's commitment to transparency 591 00:18:06,640 --> 00:18:09,560 is incredibly valuable for the entire AI community. 592 00:18:09,560 --> 00:18:10,240 It is. 593 00:18:10,240 --> 00:18:12,120 By sharing their methods and data, 594 00:18:12,120 --> 00:18:15,120 they're encouraging collaboration and driving innovation. 595 00:18:15,120 --> 00:18:15,600 Exactly. 596 00:18:15,600 --> 00:18:16,760 It's like they're saying, let's all 597 00:18:16,760 --> 00:18:19,400 work together to build a better AI future. 598 00:18:19,400 --> 00:18:20,760 So where do we go from here? 599 00:18:20,760 --> 00:18:21,920 That's a great question. 600 00:18:21,920 --> 00:18:23,960 What are the next steps for 2U3? 601 00:18:23,960 --> 00:18:25,560 And what are the broader implications 602 00:18:25,560 --> 00:18:28,360 of this research for the future of open source AI? 603 00:18:28,360 --> 00:18:30,360 Well, they've identified several exciting avenues 604 00:18:30,360 --> 00:18:31,520 for future research. 605 00:18:31,520 --> 00:18:32,080 OK. 606 00:18:32,080 --> 00:18:33,440 One area they're keen on exploring 607 00:18:33,440 --> 00:18:35,760 is enhancing the model's ability to handle longer 608 00:18:35,760 --> 00:18:38,440 conversations and more complex interactions. 609 00:18:38,440 --> 00:18:43,000 Right now, 2U3 is primarily trained on shorter conversations. 610 00:18:43,000 --> 00:18:43,720 Right. 611 00:18:43,720 --> 00:18:47,280 But imagine an AI that can understand and engage 612 00:18:47,280 --> 00:18:48,560 in more nuanced dialogues. 613 00:18:48,560 --> 00:18:49,000 Right. 614 00:18:49,000 --> 00:18:50,680 Like a real back and forth conversation. 615 00:18:50,680 --> 00:18:51,760 Exactly. 616 00:18:51,760 --> 00:18:55,560 That would be huge advancement for things like chatbots, 617 00:18:55,560 --> 00:18:59,640 virtual assistants, and AI-powered writing tools. 618 00:18:59,640 --> 00:19:01,000 All sorts of things. 619 00:19:01,000 --> 00:19:03,080 They're also looking into expanding 620 00:19:03,080 --> 00:19:05,200 its multilingual capabilities. 621 00:19:05,200 --> 00:19:05,760 Yes. 622 00:19:05,760 --> 00:19:07,160 So beyond English. 623 00:19:07,160 --> 00:19:07,960 Beyond English. 624 00:19:07,960 --> 00:19:10,640 Imagine an AI that can seamlessly communicate 625 00:19:10,640 --> 00:19:11,920 in multiple languages. 626 00:19:11,920 --> 00:19:12,680 Wow. 627 00:19:12,680 --> 00:19:14,640 It would open up so many possibilities 628 00:19:14,640 --> 00:19:17,520 for cross-cultural communication and collaboration. 629 00:19:17,520 --> 00:19:19,000 That's an inspiring vision. 630 00:19:19,000 --> 00:19:19,400 It is. 631 00:19:19,400 --> 00:19:22,320 Making AI accessible to people from all over the world. 632 00:19:22,320 --> 00:19:22,920 Precisely. 633 00:19:22,920 --> 00:19:23,720 That's amazing. 634 00:19:23,720 --> 00:19:26,120 It aligns with the idea of democratizing AI. 635 00:19:26,120 --> 00:19:26,600 Right. 636 00:19:26,600 --> 00:19:29,360 And ensuring that everyone can benefit from its potential. 637 00:19:29,360 --> 00:19:30,000 Yeah. 638 00:19:30,000 --> 00:19:33,680 Another area they're exploring is integrating 2U3 639 00:19:33,680 --> 00:19:36,560 with tools and agents. 640 00:19:36,560 --> 00:19:38,240 So not just communicating with words, 641 00:19:38,240 --> 00:19:40,440 but also interacting with the world around it. 642 00:19:40,440 --> 00:19:41,360 That's the idea. 643 00:19:41,360 --> 00:19:42,000 Wow. 644 00:19:42,000 --> 00:19:46,280 Imagine an AI that can access external tools and systems. 645 00:19:46,280 --> 00:19:46,760 Correct. 646 00:19:46,760 --> 00:19:50,280 Search engines, databases, even physical devices. 647 00:19:50,280 --> 00:19:50,880 Wow. 648 00:19:50,880 --> 00:19:53,160 So it could actually take actions based on what 649 00:19:53,160 --> 00:19:54,480 it learns and understands. 650 00:19:54,480 --> 00:19:55,280 Right. 651 00:19:55,280 --> 00:19:57,040 The possibilities seem endless. 652 00:19:57,040 --> 00:19:58,000 They really are. 653 00:19:58,000 --> 00:20:01,280 It shows how open source AI can be a driving force 654 00:20:01,280 --> 00:20:03,840 in developing really versatile AI systems that 655 00:20:03,840 --> 00:20:06,160 can address a wide range of real world problems. 656 00:20:06,160 --> 00:20:07,280 I agree. 657 00:20:07,280 --> 00:20:09,720 This research is truly pushing the boundaries of what 658 00:20:09,720 --> 00:20:10,880 we thought was possible. 659 00:20:10,880 --> 00:20:11,920 It is. 660 00:20:11,920 --> 00:20:16,040 It's exciting to think about how 2U3 will continue to evolve 661 00:20:16,040 --> 00:20:18,880 and inspire new innovations in the AI community. 662 00:20:18,880 --> 00:20:19,280 It is. 663 00:20:19,280 --> 00:20:20,800 It is a very exciting time. 664 00:20:20,800 --> 00:20:22,000 I'm glad we're here to talk about it. 665 00:20:22,000 --> 00:20:22,520 Me too. 666 00:20:22,520 --> 00:20:25,720 Welcome back for the final part of our deep dive into 2U3. 667 00:20:25,720 --> 00:20:27,040 It's been quite a journey. 668 00:20:27,040 --> 00:20:28,240 We've covered a lot of ground. 669 00:20:28,240 --> 00:20:28,760 We have. 670 00:20:28,760 --> 00:20:31,880 Exploring this technology, the training process, 671 00:20:31,880 --> 00:20:33,680 and even those unseen evaluations. 672 00:20:33,680 --> 00:20:34,000 Yeah. 673 00:20:34,000 --> 00:20:36,600 It's really fascinating to see how this open source model is 674 00:20:36,600 --> 00:20:38,520 kind of shaking up the AI world. 675 00:20:38,520 --> 00:20:38,920 Yeah. 676 00:20:38,920 --> 00:20:40,840 Challenging those preconceived notions. 677 00:20:40,840 --> 00:20:41,280 Right. 678 00:20:41,280 --> 00:20:43,560 And pushing those boundaries of what's possible. 679 00:20:43,560 --> 00:20:44,680 Exactly. 680 00:20:44,680 --> 00:20:48,240 But now it's time to step back and consider that bigger picture. 681 00:20:48,240 --> 00:20:48,720 Right. 682 00:20:48,720 --> 00:20:53,920 What does the rise of powerful open source AI mean for us? 683 00:20:53,920 --> 00:20:56,440 Like what are the implications for society, 684 00:20:56,440 --> 00:20:59,120 for the future of work, and even for our understanding 685 00:20:59,120 --> 00:21:00,440 of intelligence itself? 686 00:21:00,440 --> 00:21:00,640 Yeah. 687 00:21:00,640 --> 00:21:02,480 Those are the million dollar questions, aren't they? 688 00:21:02,480 --> 00:21:03,160 They are. 689 00:21:03,160 --> 00:21:04,880 And they're not easy to answer. 690 00:21:04,880 --> 00:21:05,360 No. 691 00:21:05,360 --> 00:21:08,680 But it is crucial that we start having these conversations 692 00:21:08,680 --> 00:21:10,960 now as this technology continues 693 00:21:10,960 --> 00:21:12,880 to evolve at such an incredible pace. 694 00:21:12,880 --> 00:21:13,760 Absolutely. 695 00:21:13,760 --> 00:21:16,000 So let's dive into some of these big picture questions. 696 00:21:16,000 --> 00:21:16,640 OK. 697 00:21:16,640 --> 00:21:19,840 First up, what does the rise of open source AI 698 00:21:19,840 --> 00:21:22,040 mean for the future of work? 699 00:21:22,040 --> 00:21:24,960 Well, one of the most immediate impacts we're likely to see 700 00:21:24,960 --> 00:21:26,880 is an acceleration of automation. 701 00:21:26,880 --> 00:21:27,200 OK. 702 00:21:27,200 --> 00:21:30,200 Tasks that were once thought to require human intelligence, 703 00:21:30,200 --> 00:21:34,000 like writing, coding, and data analysis, 704 00:21:34,000 --> 00:21:37,680 might become increasingly automated as AI, like tool 705 00:21:37,680 --> 00:21:40,240 U3, becomes more widely available. 706 00:21:40,240 --> 00:21:40,920 That makes sense. 707 00:21:40,920 --> 00:21:41,400 Yeah. 708 00:21:41,400 --> 00:21:43,960 If AI can do these things faster and more efficiently. 709 00:21:43,960 --> 00:21:47,040 It's only natural that businesses would adopt these tools. 710 00:21:47,040 --> 00:21:47,480 Right. 711 00:21:47,480 --> 00:21:49,600 But I can also see how that could lead to concerns 712 00:21:49,600 --> 00:21:51,400 about job displacement. 713 00:21:51,400 --> 00:21:51,920 You're right. 714 00:21:51,920 --> 00:21:53,000 It's a valid concern. 715 00:21:53,000 --> 00:21:53,440 Yeah. 716 00:21:53,440 --> 00:21:54,680 That's why it's essential that we 717 00:21:54,680 --> 00:21:57,760 start thinking about how we can adapt our workforce, 718 00:21:57,760 --> 00:21:59,920 invest in retraining and education, 719 00:21:59,920 --> 00:22:02,320 and create new opportunities that leverage 720 00:22:02,320 --> 00:22:05,120 the unique strengths of both humans and AI. 721 00:22:05,120 --> 00:22:08,400 It's about finding that synergy, that sweet spot, 722 00:22:08,400 --> 00:22:11,320 where humans and AI can work together 723 00:22:11,320 --> 00:22:12,880 to achieve incredible things. 724 00:22:12,880 --> 00:22:13,600 Exactly. 725 00:22:13,600 --> 00:22:15,960 And open source AI could play a key role 726 00:22:15,960 --> 00:22:19,280 in making these advanced tools more accessible 727 00:22:19,280 --> 00:22:21,960 to a wider range of businesses and individuals. 728 00:22:21,960 --> 00:22:22,200 Right? 729 00:22:22,200 --> 00:22:22,840 Absolutely. 730 00:22:22,840 --> 00:22:25,120 It's not just large corporations that 731 00:22:25,120 --> 00:22:26,680 can benefit from this technology. 732 00:22:26,680 --> 00:22:27,120 Right. 733 00:22:27,120 --> 00:22:30,680 Smaller businesses, startups, even individual entrepreneurs 734 00:22:30,680 --> 00:22:33,000 could have access to these powerful tools. 735 00:22:33,000 --> 00:22:33,520 That's great. 736 00:22:33,520 --> 00:22:37,280 Which could lead to a surge in innovation and creativity 737 00:22:37,280 --> 00:22:38,920 across all sectors. 738 00:22:38,920 --> 00:22:40,360 That's a really exciting prospect. 739 00:22:40,360 --> 00:22:42,120 It's like leveling the playing field. 740 00:22:42,120 --> 00:22:42,600 It is. 741 00:22:42,600 --> 00:22:46,560 And empowering everyone to participate in this AI revolution. 742 00:22:46,560 --> 00:22:47,400 Exactly. 743 00:22:47,400 --> 00:22:49,640 What about the implications for education? 744 00:22:49,640 --> 00:22:50,280 Oh, yeah. 745 00:22:50,280 --> 00:22:52,840 I can imagine AI having a profound impact 746 00:22:52,840 --> 00:22:54,320 on how we learn and teach. 747 00:22:54,320 --> 00:22:57,120 Imagine a world where students have access 748 00:22:57,120 --> 00:22:59,440 to personalized AI tutors. 749 00:22:59,440 --> 00:23:00,440 OK, yeah. 750 00:23:00,440 --> 00:23:03,520 That can adapt to their individual learning styles, 751 00:23:03,520 --> 00:23:05,280 provide instant feedback, and make 752 00:23:05,280 --> 00:23:07,480 learning more engaging and interactive. 753 00:23:07,480 --> 00:23:10,360 It's like having a personal coach for every subject, 754 00:23:10,360 --> 00:23:11,640 guiding you along the way. 755 00:23:11,640 --> 00:23:12,360 Exactly. 756 00:23:12,360 --> 00:23:13,440 That could be incredible. 757 00:23:13,440 --> 00:23:14,000 Yeah. 758 00:23:14,000 --> 00:23:16,440 But I can also see potential downsides. 759 00:23:16,440 --> 00:23:21,160 What about concerns that AI could stifle creativity 760 00:23:21,160 --> 00:23:23,000 or discourage critical thinking? 761 00:23:23,000 --> 00:23:24,680 Those are valid concerns. 762 00:23:24,680 --> 00:23:27,360 And it's essential that we address them thoughtfully. 763 00:23:27,360 --> 00:23:30,640 We need to ensure that AI tools are used responsibly 764 00:23:30,640 --> 00:23:34,200 in education, that they complement rather than replace 765 00:23:34,200 --> 00:23:36,520 human teachers, and that they encourage students 766 00:23:36,520 --> 00:23:38,760 to develop their critical thinking skills 767 00:23:38,760 --> 00:23:41,840 and explore their own unique creative potential. 768 00:23:41,840 --> 00:23:43,720 It's about striking that balance, 769 00:23:43,720 --> 00:23:47,080 using AI to enhance the learning experience, 770 00:23:47,080 --> 00:23:51,200 while preserving those really essential human qualities. 771 00:23:51,200 --> 00:23:51,920 Absolutely. 772 00:23:51,920 --> 00:23:53,920 Now, I can't let you off the hook without tackling 773 00:23:53,920 --> 00:23:55,160 one of the biggest questions of all. 774 00:23:55,160 --> 00:23:55,960 Oh, OK. 775 00:23:55,960 --> 00:23:59,280 What does the rise of AI, like 2U3, 776 00:23:59,280 --> 00:24:03,360 mean for our understanding of intelligence itself? 777 00:24:03,360 --> 00:24:05,600 The age-old question. 778 00:24:05,600 --> 00:24:09,200 As AI systems become increasingly sophisticated, 779 00:24:09,200 --> 00:24:11,520 they're challenging our traditional notions of what 780 00:24:11,520 --> 00:24:13,400 it means to be intelligent. 781 00:24:13,400 --> 00:24:15,760 They can perform tasks that were once thought 782 00:24:15,760 --> 00:24:17,320 to be uniquely human. 783 00:24:17,320 --> 00:24:19,640 And they're constantly learning and evolving. 784 00:24:19,640 --> 00:24:21,960 It's like we're witnessing a new form of intelligence 785 00:24:21,960 --> 00:24:25,160 emerging, and it's forcing us to rethink our own place 786 00:24:25,160 --> 00:24:25,760 in the world. 787 00:24:25,760 --> 00:24:27,040 Exactly. 788 00:24:27,040 --> 00:24:29,280 And the beauty of open source AI is 789 00:24:29,280 --> 00:24:32,600 that it allows us to explore these questions together 790 00:24:32,600 --> 00:24:33,800 as a global community. 791 00:24:33,800 --> 00:24:34,240 That's great. 792 00:24:34,240 --> 00:24:36,680 By sharing our knowledge and insights. 793 00:24:36,680 --> 00:24:39,920 We can collectively navigate the complexities of AI 794 00:24:39,920 --> 00:24:43,320 and ensure that its development benefits all of humanity. 795 00:24:43,320 --> 00:24:44,520 That's a powerful message. 796 00:24:44,520 --> 00:24:45,040 It is. 797 00:24:45,040 --> 00:24:46,800 It's not just about the technology itself. 798 00:24:46,800 --> 00:24:47,320 Right. 799 00:24:47,320 --> 00:24:50,160 It's about the collaborative spirit, the open exchange 800 00:24:50,160 --> 00:24:52,680 of ideas, and the shared responsibility 801 00:24:52,680 --> 00:24:55,600 we have in shaping the future of AI. 802 00:24:55,600 --> 00:24:56,920 Well said. 803 00:24:56,920 --> 00:24:59,880 Well, it's been a privilege to explore the world of 2U3 804 00:24:59,880 --> 00:25:00,800 with you today. 805 00:25:00,800 --> 00:25:02,000 Likewise. 806 00:25:02,000 --> 00:25:04,680 This deep dive has been a fascinating journey 807 00:25:04,680 --> 00:25:07,640 from the technical details to the philosophical implications. 808 00:25:07,640 --> 00:25:08,480 I agree. 809 00:25:08,480 --> 00:25:10,920 And I think it's safe to say that open source AI, 810 00:25:10,920 --> 00:25:13,960 like 2U3, will continue to shape our world 811 00:25:13,960 --> 00:25:15,440 in really profound ways. 812 00:25:15,440 --> 00:25:16,120 It will. 813 00:25:16,120 --> 00:25:18,800 It's an exciting time to be witnessing these developments 814 00:25:18,800 --> 00:25:20,520 and participating in this conversation. 815 00:25:20,520 --> 00:25:21,080 It is. 816 00:25:21,080 --> 00:25:22,000 Thanks for joining us. 817 00:25:22,000 --> 00:25:32,560 Thanks for having me.