1 00:00:00,000 --> 00:00:03,200 Have you ever wished that you could just hit, 2 00:00:03,200 --> 00:00:06,200 like fast forward, you know, on learning a new skill? 3 00:00:06,200 --> 00:00:07,040 Oh yeah. 4 00:00:07,040 --> 00:00:08,440 Like skip all those hours of practice 5 00:00:08,440 --> 00:00:10,400 and just like jump right to being an expert. 6 00:00:10,400 --> 00:00:11,240 Absolutely. 7 00:00:11,240 --> 00:00:12,700 Well in the world of AI, 8 00:00:14,320 --> 00:00:16,120 that's what researchers are trying to do. 9 00:00:16,120 --> 00:00:16,960 Yeah. 10 00:00:16,960 --> 00:00:20,160 And there's this new breakthrough called B-Star. 11 00:00:20,160 --> 00:00:23,480 That is really shaking things up. 12 00:00:23,480 --> 00:00:24,320 That's right. 13 00:00:24,320 --> 00:00:25,480 So maybe you could tell us a little bit about 14 00:00:25,480 --> 00:00:28,000 what is so groundbreaking about B-Star? 15 00:00:28,000 --> 00:00:29,960 So B-Star is interesting because it tackles 16 00:00:29,960 --> 00:00:32,840 a really, really fundamental problem in AI, 17 00:00:32,840 --> 00:00:36,920 which is that as these models get more and more complex, 18 00:00:36,920 --> 00:00:40,120 they need more and more data to be able to learn. 19 00:00:40,120 --> 00:00:40,960 Yeah. 20 00:00:40,960 --> 00:00:44,160 And that data typically comes from us humans 21 00:00:44,160 --> 00:00:45,920 and it's a very, very expensive 22 00:00:45,920 --> 00:00:48,240 and time consuming process to create. 23 00:00:48,240 --> 00:00:49,080 Right. 24 00:00:49,080 --> 00:00:50,760 It's like we're constantly having to feed the AI 25 00:00:50,760 --> 00:00:51,600 That's right. 26 00:00:51,600 --> 00:00:52,560 with more and more and more information 27 00:00:52,560 --> 00:00:54,680 just to get them to learn a little bit more. 28 00:00:54,680 --> 00:00:55,520 That's right. 29 00:00:55,520 --> 00:00:58,520 But B-Star is all about making the AI more self-sufficient. 30 00:00:58,520 --> 00:00:59,360 That's right. 31 00:00:59,360 --> 00:01:00,200 Right. 32 00:01:00,200 --> 00:01:01,020 So how does that work? 33 00:01:01,020 --> 00:01:02,480 So we call it self-improvement 34 00:01:02,480 --> 00:01:05,680 and we've seen other attempts to try to improve AI's ability 35 00:01:05,680 --> 00:01:08,880 to self-improve, but typically they would plateau. 36 00:01:08,880 --> 00:01:10,440 B-Star takes a very different route. 37 00:01:10,440 --> 00:01:12,120 And so we're seeing a lot of really interesting 38 00:01:12,120 --> 00:01:13,120 results with it. 39 00:01:13,120 --> 00:01:14,040 That's really exciting. 40 00:01:14,040 --> 00:01:15,240 So how does it work? 41 00:01:15,240 --> 00:01:17,760 How is this AI self-improving? 42 00:01:17,760 --> 00:01:20,200 Okay, so imagine you're learning to play the guitar. 43 00:01:20,200 --> 00:01:21,040 Right. 44 00:01:21,040 --> 00:01:21,880 You could spend all of your time 45 00:01:21,880 --> 00:01:23,880 just practicing like the same three chords. 46 00:01:23,880 --> 00:01:24,720 Yeah. 47 00:01:24,720 --> 00:01:26,400 And that's something we call exploitation in AI. 48 00:01:26,400 --> 00:01:28,480 Or you could experiment with like new melodies 49 00:01:28,480 --> 00:01:31,240 or new techniques and then we call that exploration. 50 00:01:31,240 --> 00:01:34,520 And the key to actually getting better is this balance 51 00:01:34,520 --> 00:01:36,560 between sticking with what you know 52 00:01:36,560 --> 00:01:37,800 and trying something new. 53 00:01:37,800 --> 00:01:40,320 Right, so it's finding that sweet spot between the two. 54 00:01:40,320 --> 00:01:42,060 That's exactly what B-Star is doing. 55 00:01:42,060 --> 00:01:43,640 But unlike previous methods, 56 00:01:43,640 --> 00:01:46,520 which would kind of set that balance beforehand, 57 00:01:46,520 --> 00:01:50,520 B-Star will continuously monitor and adjust that balance 58 00:01:50,520 --> 00:01:52,120 throughout the training process. 59 00:01:52,120 --> 00:01:52,960 Oh, wow. 60 00:01:52,960 --> 00:01:54,860 So think of it like a self-driving car 61 00:01:54,860 --> 00:01:57,560 that is adjusting its course in real time 62 00:01:57,560 --> 00:02:00,200 to account for like traffic and other conditions. 63 00:02:00,200 --> 00:02:02,520 So B-Star is actually learning how to learn. 64 00:02:02,520 --> 00:02:03,360 That's right. 65 00:02:03,360 --> 00:02:04,840 Which is, that's wild. 66 00:02:04,840 --> 00:02:05,680 It is. 67 00:02:05,680 --> 00:02:09,320 Okay, but how does it know if it's on the right track? 68 00:02:09,320 --> 00:02:11,280 Like if it's actually improving. 69 00:02:11,280 --> 00:02:13,760 So we use a metric called the balance score. 70 00:02:13,760 --> 00:02:14,600 Okay. 71 00:02:14,600 --> 00:02:17,320 And it's essentially measuring how well it's balancing 72 00:02:17,320 --> 00:02:18,880 this exploration and exploitation. 73 00:02:18,880 --> 00:02:19,720 Right. 74 00:02:19,720 --> 00:02:21,960 And so we're looking at both like the quantity of outputs 75 00:02:21,960 --> 00:02:25,040 that it produces and then also the quality of those outputs. 76 00:02:25,040 --> 00:02:27,040 And if it's producing a lot of good outputs, 77 00:02:27,040 --> 00:02:28,760 we know that it's learning effectively. 78 00:02:28,760 --> 00:02:30,080 That's kind of a record card. 79 00:02:30,080 --> 00:02:30,920 Yeah. 80 00:02:30,920 --> 00:02:31,760 For the AI. 81 00:02:31,760 --> 00:02:32,600 That's a good way to think about it. 82 00:02:32,600 --> 00:02:33,440 Yeah. 83 00:02:33,440 --> 00:02:35,920 Okay, so does this approach actually produce better results? 84 00:02:35,920 --> 00:02:36,760 Yeah. 85 00:02:36,760 --> 00:02:38,760 So we see in the research that they tried this out 86 00:02:38,760 --> 00:02:42,160 on a variety of tasks, including math problems 87 00:02:42,160 --> 00:02:44,280 and writing code and answering common sense 88 00:02:44,280 --> 00:02:45,320 reasoning questions. 89 00:02:45,320 --> 00:02:47,120 Wow, that's a lot of really different challenges. 90 00:02:47,120 --> 00:02:47,960 It is. 91 00:02:47,960 --> 00:02:49,040 So how did it do? 92 00:02:49,040 --> 00:02:50,600 So first let's look at math problems. 93 00:02:50,600 --> 00:02:54,380 So for this, they use data sets called GSM8KNMath, 94 00:02:54,380 --> 00:02:57,560 which tests the model's problem solving abilities. 95 00:02:57,560 --> 00:03:01,320 And on GSM8K, BSTAR was able to achieve a pass 96 00:03:01,320 --> 00:03:04,680 at one accuracy of 53.8%, 97 00:03:04,680 --> 00:03:07,080 which is a really big jump from the previous best, 98 00:03:07,080 --> 00:03:08,960 which was 46.8%. 99 00:03:08,960 --> 00:03:12,120 And on math, it was able to get 27.8%. 100 00:03:12,120 --> 00:03:13,200 Wow. 101 00:03:13,200 --> 00:03:14,800 So figuring out math. 102 00:03:14,800 --> 00:03:15,640 It is. 103 00:03:15,640 --> 00:03:16,480 That's amazing. 104 00:03:16,480 --> 00:03:17,280 That's right. 105 00:03:17,280 --> 00:03:18,600 Okay, what about something like coding 106 00:03:18,600 --> 00:03:20,880 that seems like way more complicated? 107 00:03:20,880 --> 00:03:21,720 Yeah. 108 00:03:21,720 --> 00:03:24,640 Even with coding, which is a very, very difficult problem 109 00:03:24,640 --> 00:03:29,080 for AI, BSTAR was able to make some pretty significant 110 00:03:29,080 --> 00:03:30,240 improvements. 111 00:03:30,240 --> 00:03:33,140 So they used a data set called APPS. 112 00:03:33,140 --> 00:03:37,520 And with this BSTAR achieved an accuracy of 19.6%, 113 00:03:37,520 --> 00:03:39,900 which is again, significantly higher than the previous 114 00:03:39,900 --> 00:03:40,920 methods that they tried. 115 00:03:40,920 --> 00:03:42,280 So it's not just a one trick, Pony. 116 00:03:42,280 --> 00:03:43,120 No, it's not. 117 00:03:43,120 --> 00:03:44,480 It can do lots of different things. 118 00:03:44,480 --> 00:03:46,560 Okay, so you mentioned earlier that other attempts 119 00:03:46,560 --> 00:03:48,840 at self-improvement kind of hit a wall. 120 00:03:48,840 --> 00:03:50,480 Does BSTAR have the same problem? 121 00:03:50,480 --> 00:03:52,920 So that's what's really, really exciting about this, 122 00:03:52,920 --> 00:03:54,600 is that it keeps getting better. 123 00:03:54,600 --> 00:03:55,440 Oh, wow. 124 00:03:55,440 --> 00:03:58,760 It doesn't just improve for a few rounds and then plateau. 125 00:03:58,760 --> 00:04:02,080 It keeps getting better and better over multiple iterations. 126 00:04:02,080 --> 00:04:04,040 So for example, in the math training, 127 00:04:04,040 --> 00:04:06,080 they gradually increased something called 128 00:04:06,080 --> 00:04:07,360 the sampling temperature. 129 00:04:07,360 --> 00:04:08,560 Okay, sampling temperature. 130 00:04:08,560 --> 00:04:09,400 Yeah. 131 00:04:09,400 --> 00:04:10,240 That sounds very technical. 132 00:04:10,240 --> 00:04:13,800 It is, but basically imagine your like brainstorming ideas. 133 00:04:13,800 --> 00:04:16,800 A low temperature would keep your ideas kind of close 134 00:04:16,800 --> 00:04:17,880 to what you already know. 135 00:04:17,880 --> 00:04:18,720 Right. 136 00:04:18,720 --> 00:04:20,520 A low temperature would encourage you to think more 137 00:04:20,520 --> 00:04:21,680 outside the box. 138 00:04:21,680 --> 00:04:22,520 Okay. 139 00:04:22,520 --> 00:04:24,080 So that's what sampling temperature does. 140 00:04:24,080 --> 00:04:27,120 And as they gradually increase this temperature, 141 00:04:27,120 --> 00:04:29,560 BSTAR was able to avoid getting stuck in a rut 142 00:04:29,560 --> 00:04:32,200 and find new and better solutions. 143 00:04:32,200 --> 00:04:34,160 So it's like, it's constantly trying to push 144 00:04:34,160 --> 00:04:35,440 its own creative limits. 145 00:04:35,440 --> 00:04:37,600 Yeah, that's a really good way to think about it. 146 00:04:37,600 --> 00:04:38,680 Even as it gets better. 147 00:04:38,680 --> 00:04:39,720 That's right. 148 00:04:39,720 --> 00:04:42,440 Okay, so are there any other techniques that it uses 149 00:04:42,440 --> 00:04:44,120 to kind of boost its learning? 150 00:04:44,120 --> 00:04:46,320 So something else that's really interesting in the research 151 00:04:46,320 --> 00:04:49,240 is that they use additional reward models. 152 00:04:49,240 --> 00:04:50,080 Okay. 153 00:04:50,080 --> 00:04:52,440 So think of this as specialized feedback systems 154 00:04:52,440 --> 00:04:55,200 that give more nuance than just saying right or wrong. 155 00:04:55,200 --> 00:04:56,040 Okay. 156 00:04:56,040 --> 00:04:58,760 So for example, when they were testing its math skills, 157 00:04:58,760 --> 00:05:03,120 they used what's called a process reward model or a PRM. 158 00:05:03,120 --> 00:05:03,960 PRM. 159 00:05:03,960 --> 00:05:04,800 Okay, what is that? 160 00:05:04,800 --> 00:05:06,960 So instead of just telling the AI that like, 161 00:05:06,960 --> 00:05:08,840 its final answer is right or wrong. 162 00:05:08,840 --> 00:05:11,800 It actually looks at each individual step 163 00:05:11,800 --> 00:05:15,200 that AI takes to get to that answer and provides feedback. 164 00:05:15,200 --> 00:05:17,960 Oh, so it's like having a teacher that grades your work 165 00:05:17,960 --> 00:05:20,840 and then gives you notes on how you could improve. 166 00:05:20,840 --> 00:05:21,880 That's exactly right. 167 00:05:21,880 --> 00:05:22,720 Yeah. 168 00:05:22,720 --> 00:05:25,800 So it gives the model this really, really comprehensive 169 00:05:25,800 --> 00:05:27,640 understanding of what it's doing well 170 00:05:27,640 --> 00:05:29,000 and what it's not doing well 171 00:05:29,000 --> 00:05:32,160 and allows it to really learn from that feedback. 172 00:05:32,160 --> 00:05:33,440 So we've talked about the balance score 173 00:05:33,440 --> 00:05:36,080 and we've talked about how it adjusts this exploration 174 00:05:36,080 --> 00:05:36,920 and exploitation. 175 00:05:36,920 --> 00:05:38,160 Yeah. 176 00:05:38,160 --> 00:05:41,520 But why should we care about this whole self-improvement 177 00:05:41,520 --> 00:05:42,640 thing in AI? 178 00:05:42,640 --> 00:05:44,200 That's a really great question. 179 00:05:44,200 --> 00:05:47,560 So if we zoom out and think about the bigger picture, 180 00:05:47,560 --> 00:05:49,560 I think it has really, really big implications 181 00:05:49,560 --> 00:05:52,440 for how we develop and use AI in the future. 182 00:05:52,440 --> 00:05:53,280 Okay. 183 00:05:53,280 --> 00:05:56,120 So imagine AI that can learn and adapt on its own 184 00:05:56,120 --> 00:05:59,520 without this constant hand-holding from us humans. 185 00:05:59,520 --> 00:06:00,360 Right. 186 00:06:00,360 --> 00:06:01,560 That would open up possibilities 187 00:06:01,560 --> 00:06:03,160 in so many different fields. 188 00:06:03,160 --> 00:06:05,440 So it's less time spent on data prep. 189 00:06:05,440 --> 00:06:06,280 That's right. 190 00:06:06,280 --> 00:06:08,840 And more time spent on actually tackling real-world problems. 191 00:06:08,840 --> 00:06:10,160 Exactly, yeah. 192 00:06:10,160 --> 00:06:11,000 That's awesome. 193 00:06:11,000 --> 00:06:11,840 Yeah. 194 00:06:11,840 --> 00:06:15,720 So does this actually work with like larger AI models? 195 00:06:15,720 --> 00:06:17,680 So that's what's really encouraging about this 196 00:06:17,680 --> 00:06:19,520 is that it seems to scale. 197 00:06:19,520 --> 00:06:20,360 Okay. 198 00:06:20,360 --> 00:06:23,160 So they tested it on a model called Lama 3.18b, 199 00:06:23,160 --> 00:06:24,880 which is a fairly big model 200 00:06:24,880 --> 00:06:28,160 and they still saw these really significant improvements. 201 00:06:28,160 --> 00:06:29,120 Oh, it's actually scales. 202 00:06:29,120 --> 00:06:29,960 That's great. 203 00:06:29,960 --> 00:06:30,800 That's right. 204 00:06:30,800 --> 00:06:32,760 What else is really neat about this approach? 205 00:06:32,760 --> 00:06:34,920 So I think one of the really interesting things 206 00:06:34,920 --> 00:06:37,520 about MeSTAR is that it makes this whole self-improvement 207 00:06:37,520 --> 00:06:39,240 process more transparent. 208 00:06:39,240 --> 00:06:40,080 Okay. 209 00:06:40,080 --> 00:06:41,960 So in other methods, it's kind of like a black box. 210 00:06:41,960 --> 00:06:43,760 You just run the model and hope for the best. 211 00:06:43,760 --> 00:06:44,600 Right. 212 00:06:44,600 --> 00:06:46,840 MeSTAR allows researchers to really see 213 00:06:46,840 --> 00:06:48,040 what's happening at each step 214 00:06:48,040 --> 00:06:50,680 and why certain approaches are working better than others 215 00:06:50,680 --> 00:06:53,800 and how the model is adapting its strategy over time. 216 00:06:53,800 --> 00:06:55,480 So it's not just making the AI better, 217 00:06:55,480 --> 00:06:57,080 but it's making it so we can understand 218 00:06:57,080 --> 00:06:58,280 how that improvement is happening. 219 00:06:58,280 --> 00:06:59,120 Exactly. 220 00:06:59,120 --> 00:07:00,880 Okay, so we've covered a lot. 221 00:07:00,880 --> 00:07:03,440 MeSTAR is about self-improvement in AI. 222 00:07:03,440 --> 00:07:05,880 It balances exploration and exploitation. 223 00:07:05,880 --> 00:07:06,720 Uh-huh. 224 00:07:06,720 --> 00:07:08,720 It uses this balance score to stay on track. 225 00:07:08,720 --> 00:07:09,560 That's right. 226 00:07:09,560 --> 00:07:11,440 It's doing some really amazing results 227 00:07:11,440 --> 00:07:12,920 across all these different tasks. 228 00:07:12,920 --> 00:07:13,760 Yeah. 229 00:07:13,760 --> 00:07:14,720 So where do we go from here? 230 00:07:14,720 --> 00:07:17,040 Well, so this version of MeSTAR focuses 231 00:07:17,040 --> 00:07:19,520 on fine-tuning these key parameters 232 00:07:19,520 --> 00:07:22,200 like the sampling temperature and the reward thresholds. 233 00:07:22,200 --> 00:07:23,040 Okay. 234 00:07:23,040 --> 00:07:25,600 But there's tons of potential for future development. 235 00:07:25,600 --> 00:07:26,440 Like what? 236 00:07:26,440 --> 00:07:28,640 So for example, they could explore 237 00:07:28,640 --> 00:07:32,680 incorporating even more sophisticated decoding techniques. 238 00:07:32,680 --> 00:07:33,520 Okay. 239 00:07:33,520 --> 00:07:36,040 That could allow the model to have even finer control 240 00:07:36,040 --> 00:07:37,880 over its exploration and come up 241 00:07:37,880 --> 00:07:39,520 with even more diverse solutions. 242 00:07:39,520 --> 00:07:41,640 So like giving the AI a bigger toolbox. 243 00:07:41,640 --> 00:07:43,320 Yeah, that's a good way to think about it. 244 00:07:43,320 --> 00:07:44,160 It's a play with. 245 00:07:44,160 --> 00:07:45,000 That's right. 246 00:07:45,000 --> 00:07:46,560 And then on the exploitation side, 247 00:07:46,560 --> 00:07:48,920 we could look at using dynamic reward models 248 00:07:48,920 --> 00:07:51,520 that actually adapt and change over time. 249 00:07:51,520 --> 00:07:52,360 Okay. 250 00:07:52,360 --> 00:07:53,760 Which would allow the feedback to be 251 00:07:53,760 --> 00:07:57,120 even more nuanced and relevant as the model learns. 252 00:07:57,120 --> 00:07:59,160 So it's all about making it more adaptive. 253 00:07:59,160 --> 00:08:00,920 Yeah, that's a great way to put it. 254 00:08:00,920 --> 00:08:01,760 And responsive. 255 00:08:01,760 --> 00:08:02,680 Exactly. 256 00:08:02,680 --> 00:08:04,000 Okay, so we've talked a lot 257 00:08:04,000 --> 00:08:05,280 about the technical side of this. 258 00:08:05,280 --> 00:08:06,120 Right. 259 00:08:06,120 --> 00:08:09,040 But I'm really interested in the real world applications. 260 00:08:09,040 --> 00:08:11,640 Like what fields could be transformed by an AI 261 00:08:11,640 --> 00:08:12,480 that can do all this? 262 00:08:12,480 --> 00:08:14,800 Oh, the possibilities are really, really exciting. 263 00:08:14,800 --> 00:08:15,640 I bet. 264 00:08:15,640 --> 00:08:17,080 So think about something like robotics. 265 00:08:17,080 --> 00:08:20,760 Where machines have to adapt to these very complex 266 00:08:20,760 --> 00:08:22,440 and unpredictable environments. 267 00:08:22,440 --> 00:08:23,280 Yeah. 268 00:08:23,280 --> 00:08:27,960 Right now, most robots rely on pre-programmed instructions. 269 00:08:27,960 --> 00:08:30,920 But with B-Star, we could potentially see a future 270 00:08:30,920 --> 00:08:33,400 where they're constantly learning from their experiences 271 00:08:33,400 --> 00:08:35,400 and getting better at what they do. 272 00:08:35,400 --> 00:08:37,800 So as I consider them just being tools. 273 00:08:37,800 --> 00:08:38,640 Yeah. 274 00:08:38,640 --> 00:08:41,080 They become more like partners. 275 00:08:41,080 --> 00:08:42,200 Absolutely. 276 00:08:42,200 --> 00:08:43,040 That's amazing. 277 00:08:43,040 --> 00:08:45,360 Right, so imagine like a robot surgeon 278 00:08:45,360 --> 00:08:48,280 that could learn from each surgery that it performs. 279 00:08:48,280 --> 00:08:49,120 Wow. 280 00:08:49,120 --> 00:08:51,400 And continually refine its techniques. 281 00:08:51,400 --> 00:08:52,240 Yeah. 282 00:08:52,240 --> 00:08:53,640 And reduce the risk of error. 283 00:08:53,640 --> 00:08:54,480 That would be amazing. 284 00:08:54,480 --> 00:08:56,400 Or think about like a search and rescue robot. 285 00:08:56,400 --> 00:08:57,240 Okay. 286 00:08:57,240 --> 00:08:59,080 That could adapt to different environments 287 00:08:59,080 --> 00:09:02,000 and navigate this challenging terrain more effectively. 288 00:09:02,000 --> 00:09:04,680 Just in healthcare and disaster response alone, 289 00:09:04,680 --> 00:09:06,120 applications are huge. 290 00:09:06,120 --> 00:09:07,040 They are, yeah. 291 00:09:07,040 --> 00:09:07,880 Wow. 292 00:09:07,880 --> 00:09:09,280 So a lot of times we think about AI 293 00:09:09,280 --> 00:09:12,600 as being good at logic and reasoning. 294 00:09:12,600 --> 00:09:13,440 Right. 295 00:09:13,440 --> 00:09:14,400 What about creativity? 296 00:09:14,400 --> 00:09:15,880 That's a really great question. 297 00:09:15,880 --> 00:09:20,040 Like could B-Star help AI become more artistic? 298 00:09:20,040 --> 00:09:22,400 So we've already seen AI create some really impressive 299 00:09:22,400 --> 00:09:23,800 artwork and music. 300 00:09:23,800 --> 00:09:26,720 But a lot of times it feels a little bit mechanical. 301 00:09:26,720 --> 00:09:27,560 Right. 302 00:09:27,560 --> 00:09:29,680 Because it relies on these existing data sets and patterns. 303 00:09:29,680 --> 00:09:32,120 And it's kind of just remixing what's already out there. 304 00:09:32,120 --> 00:09:32,960 Yeah. 305 00:09:32,960 --> 00:09:35,200 But with B-Star, eventually AI could develop 306 00:09:35,200 --> 00:09:38,000 its own unique style, its own creative voice. 307 00:09:38,000 --> 00:09:39,480 So instead of just imitating us. 308 00:09:39,480 --> 00:09:40,800 Yeah. 309 00:09:40,800 --> 00:09:42,760 It could actually become an artist in its own right. 310 00:09:42,760 --> 00:09:43,600 That's right. 311 00:09:43,600 --> 00:09:44,440 Now that's a game changer. 312 00:09:44,440 --> 00:09:45,280 That's right. 313 00:09:45,280 --> 00:09:49,880 So imagine AI composing original music 314 00:09:49,880 --> 00:09:52,760 or writing stories or even designing new products. 315 00:09:52,760 --> 00:09:53,600 Absolutely. 316 00:09:53,600 --> 00:09:55,640 And because B-Star is so good at balancing 317 00:09:55,640 --> 00:09:57,280 exploration and exploitation. 318 00:09:57,280 --> 00:09:58,120 Wow. 319 00:09:58,120 --> 00:09:59,920 It could potentially find that sweet spot 320 00:09:59,920 --> 00:10:02,240 between pushing creative boundaries 321 00:10:02,240 --> 00:10:05,280 and still producing work that resonates with people. 322 00:10:05,280 --> 00:10:05,680 Right. 323 00:10:05,680 --> 00:10:07,800 It could help us to discover completely new forms 324 00:10:07,800 --> 00:10:09,640 of art and expression. 325 00:10:09,640 --> 00:10:11,080 That's incredible. 326 00:10:11,080 --> 00:10:12,840 This is making me think about education too. 327 00:10:12,840 --> 00:10:13,320 Oh yeah. 328 00:10:13,320 --> 00:10:17,040 Like what if we could use B-Star to create personalized learning 329 00:10:17,040 --> 00:10:17,800 experiences. 330 00:10:17,800 --> 00:10:18,400 All right. 331 00:10:18,400 --> 00:10:21,320 That adapt to each student's individual needs 332 00:10:21,320 --> 00:10:22,320 and learning style. 333 00:10:22,320 --> 00:10:23,440 That's a really cool idea. 334 00:10:23,440 --> 00:10:25,480 So imagine a tutoring system that 335 00:10:25,480 --> 00:10:28,800 can pinpoint a student's strengths and weaknesses. 336 00:10:28,800 --> 00:10:29,120 Yeah. 337 00:10:29,120 --> 00:10:31,360 And then tailor the curriculum accordingly. 338 00:10:31,360 --> 00:10:31,720 Right. 339 00:10:31,720 --> 00:10:33,920 So you could give customized feedback, 340 00:10:33,920 --> 00:10:36,360 adjust the pace of learning, and even recommend 341 00:10:36,360 --> 00:10:38,000 new topics to explore. 342 00:10:38,000 --> 00:10:40,600 To be like having a personal tutor for every student. 343 00:10:40,600 --> 00:10:41,360 Yeah. 344 00:10:41,360 --> 00:10:42,160 That's amazing. 345 00:10:42,160 --> 00:10:42,520 OK. 346 00:10:42,520 --> 00:10:42,720 OK. 347 00:10:42,720 --> 00:10:44,480 So this conversation is getting me really excited 348 00:10:44,480 --> 00:10:45,880 about the future of AI. 349 00:10:45,880 --> 00:10:47,200 It's a very exciting time. 350 00:10:47,200 --> 00:10:47,960 It is. 351 00:10:47,960 --> 00:10:50,400 But with all this talk about self-improvement 352 00:10:50,400 --> 00:10:52,920 and the systems constantly evolving, 353 00:10:52,920 --> 00:10:56,120 it does make me wonder, where does human control fit 354 00:10:56,120 --> 00:10:56,920 into all of this? 355 00:10:56,920 --> 00:10:58,400 That's a really important question. 356 00:10:58,400 --> 00:11:00,720 And as AI becomes more autonomous, 357 00:11:00,720 --> 00:11:03,040 it's really, really critical to ensure 358 00:11:03,040 --> 00:11:05,760 that we maintain a certain level of oversight. 359 00:11:05,760 --> 00:11:07,800 We don't want to create something that we can't 360 00:11:07,800 --> 00:11:09,200 understand or control. 361 00:11:09,200 --> 00:11:10,160 Hit the balancing act. 362 00:11:10,160 --> 00:11:10,640 Yeah. 363 00:11:10,640 --> 00:11:13,560 Between letting AI learn and evolve. 364 00:11:13,560 --> 00:11:14,040 Yeah. 365 00:11:14,040 --> 00:11:16,040 And also making sure that we're still 366 00:11:16,040 --> 00:11:17,280 steering it in the right direction. 367 00:11:17,280 --> 00:11:18,040 That's right. 368 00:11:18,040 --> 00:11:20,080 So it sounds like we need to be thinking not just 369 00:11:20,080 --> 00:11:24,840 about the technical aspects of AI, but also 370 00:11:24,840 --> 00:11:28,920 the social and ethical and even philosophical implications. 371 00:11:28,920 --> 00:11:30,120 Absolutely. 372 00:11:30,120 --> 00:11:31,440 AI is not just a tool. 373 00:11:31,440 --> 00:11:33,640 It's a reflection of ourselves and our values. 374 00:11:33,640 --> 00:11:37,880 And as we develop more and more advanced systems like B-Star, 375 00:11:37,880 --> 00:11:41,320 it's so important that we do it responsibly and thoughtfully. 376 00:11:41,320 --> 00:11:41,840 Yeah. 377 00:11:41,840 --> 00:11:43,560 We need to ask those tough questions. 378 00:11:43,560 --> 00:11:44,040 Yeah. 379 00:11:44,040 --> 00:11:45,600 And have those difficult conversations 380 00:11:45,600 --> 00:11:47,720 about what kind of future we want to build. 381 00:11:47,720 --> 00:11:48,680 Absolutely. 382 00:11:48,680 --> 00:11:50,560 So it's not just building a better AI. 383 00:11:50,560 --> 00:11:52,440 It's building a better future with AI. 384 00:11:52,440 --> 00:11:52,880 Yeah. 385 00:11:52,880 --> 00:11:54,440 I think that's a really nice way to put it. 386 00:11:54,440 --> 00:11:54,720 OK. 387 00:11:54,720 --> 00:11:57,400 So this has been a really fascinating deep dive. 388 00:11:57,400 --> 00:11:58,280 It has. 389 00:11:58,280 --> 00:12:00,560 We started by talking about B-Star's ability 390 00:12:00,560 --> 00:12:03,360 to help AI models learn more efficiently. 391 00:12:03,360 --> 00:12:03,840 Yeah. 392 00:12:03,840 --> 00:12:06,200 But this conversation has really made me rethink 393 00:12:06,200 --> 00:12:07,520 this whole field of AI. 394 00:12:07,520 --> 00:12:09,240 I think that's what's so exciting about these kinds 395 00:12:09,240 --> 00:12:10,480 of breakthroughs is that they really 396 00:12:10,480 --> 00:12:12,800 challenge our assumptions about what's possible. 397 00:12:12,800 --> 00:12:13,360 Absolutely. 398 00:12:13,360 --> 00:12:15,160 And I think that's something our listeners should 399 00:12:15,160 --> 00:12:16,000 take away from this episode. 400 00:12:16,000 --> 00:12:16,560 Yeah. 401 00:12:16,560 --> 00:12:21,840 B-Star is just one example of the incredible innovation that's 402 00:12:21,840 --> 00:12:23,560 happening right now in AI. 403 00:12:23,560 --> 00:12:24,520 That's right. 404 00:12:24,520 --> 00:12:26,240 It's a field that's constantly evolving. 405 00:12:26,240 --> 00:12:26,880 Yeah. 406 00:12:26,880 --> 00:12:29,640 And who knows what breakthroughs are just around the corner. 407 00:12:29,640 --> 00:12:31,440 That brings up a really important point. 408 00:12:31,440 --> 00:12:34,880 How do we keep up with all of this as individuals? 409 00:12:34,880 --> 00:12:35,320 Right. 410 00:12:35,320 --> 00:12:37,040 It can feel like a tidal wave. 411 00:12:37,040 --> 00:12:37,240 Yeah. 412 00:12:37,240 --> 00:12:38,360 It can feel really overwhelming. 413 00:12:38,360 --> 00:12:38,840 Yeah. 414 00:12:38,840 --> 00:12:41,520 So how do we even begin to keep up with it all? 415 00:12:41,520 --> 00:12:43,080 It starts with curiosity. 416 00:12:43,080 --> 00:12:43,600 OK. 417 00:12:43,600 --> 00:12:45,840 So read articles. 418 00:12:45,840 --> 00:12:47,920 Listen to podcasts like this one, of course. 419 00:12:47,920 --> 00:12:48,400 Of course. 420 00:12:48,400 --> 00:12:51,120 Seek out events related to AI. 421 00:12:51,120 --> 00:12:53,400 There are so many resources available now, 422 00:12:53,400 --> 00:12:55,600 both online and in your local community, 423 00:12:55,600 --> 00:12:58,720 that it's easier than ever to really tap into this world. 424 00:12:58,720 --> 00:13:01,760 And you don't need to be a computer scientist 425 00:13:01,760 --> 00:13:03,080 to understand this stuff. 426 00:13:03,080 --> 00:13:05,800 Or to even form an opinion about how 427 00:13:05,800 --> 00:13:07,760 AI is going to impact our lives. 428 00:13:07,760 --> 00:13:08,640 Not at all. 429 00:13:08,640 --> 00:13:10,800 And I think the more we understand about it, 430 00:13:10,800 --> 00:13:12,480 the more we can make informed decisions 431 00:13:12,480 --> 00:13:13,960 about its development and use. 432 00:13:13,960 --> 00:13:14,320 Yeah. 433 00:13:14,320 --> 00:13:16,240 It's like anything else, the more you learn. 434 00:13:16,240 --> 00:13:16,720 Yeah. 435 00:13:16,720 --> 00:13:17,840 The more empowered you are. 436 00:13:17,840 --> 00:13:18,360 Absolutely. 437 00:13:18,360 --> 00:13:19,120 To participate. 438 00:13:19,120 --> 00:13:19,640 That's right. 439 00:13:19,640 --> 00:13:21,480 And to have a voice in this conversation. 440 00:13:21,480 --> 00:13:21,880 Yeah. 441 00:13:21,880 --> 00:13:24,720 And don't underestimate the power of conversation itself. 442 00:13:24,720 --> 00:13:25,120 Oh, yeah. 443 00:13:25,120 --> 00:13:26,120 Talk to your friends. 444 00:13:26,120 --> 00:13:27,160 Talk to your family. 445 00:13:27,160 --> 00:13:27,520 Right? 446 00:13:27,520 --> 00:13:28,920 Talk to your colleagues. 447 00:13:28,920 --> 00:13:31,560 Anyone who will listen, share what you're learning, 448 00:13:31,560 --> 00:13:35,200 ask questions, and be open to different perspectives. 449 00:13:35,200 --> 00:13:37,920 A lot of times those casual conversations 450 00:13:37,920 --> 00:13:40,040 are what really lead to some big change. 451 00:13:40,040 --> 00:13:41,360 Yeah, I think so. 452 00:13:41,360 --> 00:13:41,840 OK. 453 00:13:41,840 --> 00:13:44,960 So as we wrap up this deep dive into Beastar 454 00:13:44,960 --> 00:13:48,360 and this constantly changing world of AI, 455 00:13:48,360 --> 00:13:50,800 I want to leave our listeners with a question. 456 00:13:50,800 --> 00:13:51,840 OK. 457 00:13:51,840 --> 00:13:54,600 What kind of world do you want to create with AI? 458 00:13:54,600 --> 00:13:55,440 Wow. 459 00:13:55,440 --> 00:13:56,760 That's a powerful question. 460 00:13:56,760 --> 00:13:57,480 It is. 461 00:13:57,480 --> 00:13:59,320 I don't think there's an easy answer, 462 00:13:59,320 --> 00:14:03,400 but I think by staying engaged, informed, and curious, 463 00:14:03,400 --> 00:14:05,520 we can all play a part in making sure 464 00:14:05,520 --> 00:14:07,840 that AI is used for good in the world. 465 00:14:07,840 --> 00:14:08,320 Well said. 466 00:14:08,320 --> 00:14:10,720 And for those of you who want to learn more about Beastar 467 00:14:10,720 --> 00:14:12,520 and all this research that we discussed, 468 00:14:12,520 --> 00:14:14,520 we got links in the show notes to the original paper 469 00:14:14,520 --> 00:14:16,120 and some other resources as well. 470 00:14:16,120 --> 00:14:16,640 Awesome. 471 00:14:16,640 --> 00:14:17,840 Thanks for listening, everybody. 472 00:14:17,840 --> 00:14:19,720 Until next time, keep exploring. 473 00:14:19,720 --> 00:14:22,040 Keep learning and keep asking those big questions. 474 00:14:22,040 --> 00:14:22,840 Absolutely. 475 00:14:22,840 --> 00:14:25,880 And they could also look at different ways 476 00:14:25,880 --> 00:14:29,200 of how they select the outputs that are used for training. 477 00:14:29,200 --> 00:14:32,760 So there are more sophisticated decoding techniques 478 00:14:32,760 --> 00:14:35,400 that could give the model even finer control 479 00:14:35,400 --> 00:14:38,200 over its exploration and allow it to come up 480 00:14:38,200 --> 00:14:42,920 with even more diverse and creative solutions. 481 00:14:42,920 --> 00:14:46,720 So it was like we're giving it a bigger toolbox to play with. 482 00:14:46,720 --> 00:14:47,080 Yeah. 483 00:14:47,080 --> 00:14:49,080 And then on the exploitation side, 484 00:14:49,080 --> 00:14:51,840 they could look at using dynamic reward models 485 00:14:51,840 --> 00:14:55,120 that actually adapt and change over time 486 00:14:55,120 --> 00:14:58,040 so that as the model learns, the feedback 487 00:14:58,040 --> 00:14:59,920 is even more nuanced and relevant. 488 00:14:59,920 --> 00:15:02,040 So it's all about just making it more adaptive. 489 00:15:02,040 --> 00:15:02,880 Yeah. 490 00:15:02,880 --> 00:15:05,360 OK, so we've talked a lot about the technical side of things, 491 00:15:05,360 --> 00:15:08,920 but I'm curious about the real world applications. 492 00:15:08,920 --> 00:15:12,880 Like what fields could be transformed by an AI 493 00:15:12,880 --> 00:15:14,520 that can do all of this stuff? 494 00:15:14,520 --> 00:15:16,680 I think there are so many possibilities. 495 00:15:16,680 --> 00:15:18,240 It's really exciting to think about. 496 00:15:18,240 --> 00:15:21,280 So one area that comes to mind is robotics. 497 00:15:21,280 --> 00:15:24,160 Where machines have to adapt to these very complex 498 00:15:24,160 --> 00:15:26,240 and unpredictable environments. 499 00:15:26,240 --> 00:15:29,440 And right now, robots mostly rely on these pre-programmed 500 00:15:29,440 --> 00:15:30,720 instructions. 501 00:15:30,720 --> 00:15:33,760 But with Beastar, we could potentially 502 00:15:33,760 --> 00:15:36,680 see a future where robots are constantly learning 503 00:15:36,680 --> 00:15:39,280 from their experiences and getting better and better 504 00:15:39,280 --> 00:15:40,200 at what they do. 505 00:15:40,200 --> 00:15:41,960 It's not just them being a tool anymore. 506 00:15:41,960 --> 00:15:43,840 It's them being more like a partner. 507 00:15:43,840 --> 00:15:44,800 Exactly. 508 00:15:44,800 --> 00:15:48,760 Yeah, so imagine a robot surgeon that 509 00:15:48,760 --> 00:15:52,600 could learn from each surgery that it performs 510 00:15:52,600 --> 00:15:54,760 and continually refine its techniques 511 00:15:54,760 --> 00:15:56,440 and reduce the risk of error. 512 00:15:56,440 --> 00:15:57,440 That would be incredible. 513 00:15:57,440 --> 00:16:00,160 Yeah, or think about a search and rescue robot 514 00:16:00,160 --> 00:16:02,040 that could adapt to different environments 515 00:16:02,040 --> 00:16:03,920 and navigate challenging terrain. 516 00:16:03,920 --> 00:16:06,400 Yeah, just in health care and disaster response alone. 517 00:16:06,400 --> 00:16:06,920 Right. 518 00:16:06,920 --> 00:16:08,160 The applications are huge. 519 00:16:08,160 --> 00:16:09,600 Yeah, they're really significant. 520 00:16:09,600 --> 00:16:14,520 So we tend to think of AI as being good at logic and reasoning. 521 00:16:14,520 --> 00:16:15,280 Right. 522 00:16:15,280 --> 00:16:16,520 What about creativity? 523 00:16:16,520 --> 00:16:17,720 That's a great question. 524 00:16:17,720 --> 00:16:20,680 Like, could Beastar help AI become more artistic? 525 00:16:20,680 --> 00:16:23,440 Yeah, so we've already seen AI create some pretty impressive 526 00:16:23,440 --> 00:16:24,480 artwork and music. 527 00:16:24,480 --> 00:16:25,080 Right. 528 00:16:25,080 --> 00:16:27,240 But it often feels kind of mechanical. 529 00:16:27,240 --> 00:16:27,560 Right. 530 00:16:27,560 --> 00:16:29,960 Because it's relying on these existing data sets 531 00:16:29,960 --> 00:16:30,720 and patterns. 532 00:16:30,720 --> 00:16:31,640 Yeah. 533 00:16:31,640 --> 00:16:34,560 It's essentially remixing what's already out there. 534 00:16:34,560 --> 00:16:37,240 But with Beastar, AI could potentially 535 00:16:37,240 --> 00:16:40,760 develop its own unique style, its own creative voice. 536 00:16:40,760 --> 00:16:42,480 So instead of just imitating us. 537 00:16:42,480 --> 00:16:43,040 Exactly. 538 00:16:43,040 --> 00:16:44,760 It could become an artist in its own right. 539 00:16:44,760 --> 00:16:45,520 That's right. 540 00:16:45,520 --> 00:16:46,560 That's a game changer. 541 00:16:46,560 --> 00:16:47,600 It really is, yeah. 542 00:16:47,600 --> 00:16:51,800 So we could have AI composing original music or writing 543 00:16:51,800 --> 00:16:54,200 stories or even designing new products. 544 00:16:54,200 --> 00:16:54,800 Exactly. 545 00:16:54,800 --> 00:16:59,440 And because Beastar is so good at this balance between exploration 546 00:16:59,440 --> 00:17:02,840 and exploitation, it could potentially 547 00:17:02,840 --> 00:17:07,040 find that sweet spot between pushing creative boundaries 548 00:17:07,040 --> 00:17:09,760 but still producing work that really resonates with people. 549 00:17:09,760 --> 00:17:09,960 Right. 550 00:17:09,960 --> 00:17:11,680 So it's not just totally out there. 551 00:17:11,680 --> 00:17:12,080 Right. 552 00:17:12,080 --> 00:17:13,680 And the stuff that we can connect with. 553 00:17:13,680 --> 00:17:14,120 Yeah. 554 00:17:14,120 --> 00:17:16,480 And it could help us discover entirely new forms 555 00:17:16,480 --> 00:17:17,800 of art and expression. 556 00:17:17,800 --> 00:17:18,360 Wow. 557 00:17:18,360 --> 00:17:20,560 OK, so this is making me think about education. 558 00:17:20,560 --> 00:17:23,600 Like what if we could use Beastar to create personalized 559 00:17:23,600 --> 00:17:24,720 learning experiences? 560 00:17:24,720 --> 00:17:25,760 Oh, that's a cool idea. 561 00:17:25,760 --> 00:17:28,520 That adapt to each student's individual needs and learning 562 00:17:28,520 --> 00:17:28,920 style. 563 00:17:28,920 --> 00:17:29,240 Yeah. 564 00:17:29,240 --> 00:17:31,800 So imagine a tutoring system that 565 00:17:31,800 --> 00:17:35,560 could really pinpoint a student's strengths and weaknesses. 566 00:17:35,560 --> 00:17:35,960 Yeah. 567 00:17:35,960 --> 00:17:38,680 And then tailor the curriculum based on that. 568 00:17:38,680 --> 00:17:39,120 Right. 569 00:17:39,120 --> 00:17:41,360 It could give customized feedback, 570 00:17:41,360 --> 00:17:44,200 adjust the pace of learning, even recommend new topics 571 00:17:44,200 --> 00:17:45,280 for them to explore. 572 00:17:45,280 --> 00:17:47,320 It could be like a personal tutor for every student. 573 00:17:47,320 --> 00:17:48,480 Exactly. 574 00:17:48,480 --> 00:17:48,960 Yeah. 575 00:17:48,960 --> 00:17:49,800 That's incredible. 576 00:17:49,800 --> 00:17:51,960 OK, so this conversation has got me really excited 577 00:17:51,960 --> 00:17:53,320 about the future of AI. 578 00:17:53,320 --> 00:17:53,920 Yeah. 579 00:17:53,920 --> 00:17:57,000 But all this talk about self-improvement and these 580 00:17:57,000 --> 00:18:02,000 systems constantly evolving, it does make me wonder, 581 00:18:02,000 --> 00:18:04,880 where does human control fit into all of this? 582 00:18:04,880 --> 00:18:05,680 Yeah. 583 00:18:05,680 --> 00:18:07,360 That is a really important question. 584 00:18:07,360 --> 00:18:09,760 Like are we just going to let the machines take over? 585 00:18:09,760 --> 00:18:11,120 I don't think so. 586 00:18:11,120 --> 00:18:14,920 But I think as AI systems become more and more autonomous, 587 00:18:14,920 --> 00:18:18,240 it's really critical to ensure that we maintain 588 00:18:18,240 --> 00:18:19,720 some level of oversight. 589 00:18:19,720 --> 00:18:20,040 Right. 590 00:18:20,040 --> 00:18:22,320 We don't want to create something that we can't understand 591 00:18:22,320 --> 00:18:23,080 or control. 592 00:18:23,080 --> 00:18:24,400 So it's a balancing act again. 593 00:18:24,400 --> 00:18:25,080 It is, yeah. 594 00:18:25,080 --> 00:18:27,480 Between letting the AI learn and evolve. 595 00:18:27,480 --> 00:18:28,080 Yeah. 596 00:18:28,080 --> 00:18:30,000 And also making sure that we're steering it. 597 00:18:30,000 --> 00:18:30,480 Right. 598 00:18:30,480 --> 00:18:31,280 In the right direction. 599 00:18:31,280 --> 00:18:32,000 Exactly. 600 00:18:32,000 --> 00:18:33,840 So it sounds like we need to be thinking not just 601 00:18:33,840 --> 00:18:36,080 about the technical side of AI, but also 602 00:18:36,080 --> 00:18:39,720 about the social and the ethical and even 603 00:18:39,720 --> 00:18:41,600 the philosophical implications. 604 00:18:41,600 --> 00:18:43,600 I think that's absolutely right. 605 00:18:43,600 --> 00:18:46,840 AI is not just a tool, it's a reflection of ourselves 606 00:18:46,840 --> 00:18:47,760 and our values. 607 00:18:47,760 --> 00:18:51,160 And as we develop these really advanced systems like BSTAR, 608 00:18:51,160 --> 00:18:54,120 it's so important that we do it responsible and thoughtfully. 609 00:18:54,120 --> 00:18:54,520 Yeah. 610 00:18:54,520 --> 00:18:56,040 We need to ask those tough questions 611 00:18:56,040 --> 00:18:59,920 and have those difficult conversations about what kind 612 00:18:59,920 --> 00:19:01,760 of future we want to build. 613 00:19:01,760 --> 00:19:03,040 I think that's really well said. 614 00:19:03,040 --> 00:19:04,760 So it's not just building a better AI, 615 00:19:04,760 --> 00:19:06,680 it's building a better future with AI. 616 00:19:06,680 --> 00:19:07,080 Yeah. 617 00:19:07,080 --> 00:19:08,320 I really like that. 618 00:19:08,320 --> 00:19:08,600 OK. 619 00:19:08,600 --> 00:19:11,240 So this has been a really fascinating deep dive. 620 00:19:11,240 --> 00:19:12,040 It has been. 621 00:19:12,040 --> 00:19:15,200 We started by talking about BSTAR and its ability 622 00:19:15,200 --> 00:19:18,480 to help AI models learn more efficiently. 623 00:19:18,480 --> 00:19:19,320 Right. 624 00:19:19,320 --> 00:19:21,920 But this conversation has really expanded my thinking 625 00:19:21,920 --> 00:19:23,080 about the whole field. 626 00:19:23,080 --> 00:19:26,320 I think that's what's so great about exploring these cutting 627 00:19:26,320 --> 00:19:28,440 edge advancements is that they often lead us 628 00:19:28,440 --> 00:19:31,800 to these unexpected insights and really challenge 629 00:19:31,800 --> 00:19:33,760 our assumptions about what's possible. 630 00:19:33,760 --> 00:19:34,520 Absolutely. 631 00:19:34,520 --> 00:19:36,360 And I think that's something our listeners should 632 00:19:36,360 --> 00:19:37,400 take away from this episode. 633 00:19:37,400 --> 00:19:42,000 That BSTAR is just one example of the incredible 634 00:19:42,000 --> 00:19:44,240 innovation that's happening in AI right now. 635 00:19:44,240 --> 00:19:44,880 Absolutely. 636 00:19:44,880 --> 00:19:47,680 It's a field that is just constantly evolving. 637 00:19:47,680 --> 00:19:48,080 Yeah. 638 00:19:48,080 --> 00:19:50,640 And who knows what breakthroughs are just around the corner. 639 00:19:50,640 --> 00:19:52,520 That brings up a really important point. 640 00:19:52,520 --> 00:19:55,560 How do we keep up with all of this as individuals? 641 00:19:55,560 --> 00:19:56,040 Right. 642 00:19:56,040 --> 00:19:56,360 Yeah. 643 00:19:56,360 --> 00:19:58,520 It can feel like a tidal wave of new developments 644 00:19:58,520 --> 00:19:59,600 coming out all the time. 645 00:19:59,600 --> 00:20:01,280 It can feel really overwhelming. 646 00:20:01,280 --> 00:20:01,720 Yeah. 647 00:20:01,720 --> 00:20:04,560 So how can we even begin to keep up with it all? 648 00:20:04,560 --> 00:20:06,360 I think it starts with curiosity. 649 00:20:06,360 --> 00:20:06,840 OK. 650 00:20:06,840 --> 00:20:09,760 So read articles, listen to podcasts like this one, 651 00:20:09,760 --> 00:20:10,280 of course. 652 00:20:10,280 --> 00:20:11,800 Before. 653 00:20:11,800 --> 00:20:15,640 You know, seek out events that are related to AI. 654 00:20:15,640 --> 00:20:18,760 There are so many resources available now, both online 655 00:20:18,760 --> 00:20:22,040 and in your local community, that it's easier than ever 656 00:20:22,040 --> 00:20:23,440 to tap into this world. 657 00:20:23,440 --> 00:20:28,280 And you don't need to be like a PhD in computer science. 658 00:20:28,280 --> 00:20:28,560 Yeah. 659 00:20:28,560 --> 00:20:29,120 You don't. 660 00:20:29,120 --> 00:20:33,040 To understand the basics or to have an opinion about how 661 00:20:33,040 --> 00:20:34,320 it's going to impact our lives. 662 00:20:34,320 --> 00:20:35,880 I think the more we understand about AI, 663 00:20:35,880 --> 00:20:39,160 the better equipped will be to make those informed decisions 664 00:20:39,160 --> 00:20:40,560 about its development and use. 665 00:20:40,560 --> 00:20:41,000 Right. 666 00:20:41,000 --> 00:20:42,960 It's like anything else, the more you learn, 667 00:20:42,960 --> 00:20:44,160 the more empowered you are. 668 00:20:44,160 --> 00:20:44,880 Absolutely. 669 00:20:44,880 --> 00:20:46,320 To participate in the conversation. 670 00:20:46,320 --> 00:20:46,680 Yeah. 671 00:20:46,680 --> 00:20:49,680 And don't underestimate the power of just conversation 672 00:20:49,680 --> 00:20:50,200 itself. 673 00:20:50,200 --> 00:20:50,680 Right. 674 00:20:50,680 --> 00:20:52,000 So talk to your friends. 675 00:20:52,000 --> 00:20:53,800 Talk to your family. 676 00:20:53,800 --> 00:20:55,000 Talk to your colleagues. 677 00:20:55,000 --> 00:20:57,720 Anyone who will listen, share what you're learning, 678 00:20:57,720 --> 00:21:00,560 ask questions, be open to different perspectives. 679 00:21:00,560 --> 00:21:02,400 Sometimes just those casual conversations 680 00:21:02,400 --> 00:21:03,760 can really spark some change. 681 00:21:03,760 --> 00:21:04,600 I think so too. 682 00:21:04,600 --> 00:21:05,080 OK. 683 00:21:05,080 --> 00:21:08,160 So as we wrap up this deep dive into B-Star 684 00:21:08,160 --> 00:21:11,240 and this constantly evolving world of AI, 685 00:21:11,240 --> 00:21:13,600 I want to leave our listeners with one final thought. 686 00:21:13,600 --> 00:21:14,280 OK. 687 00:21:14,280 --> 00:21:16,840 What kind of world do you want to create with AI? 688 00:21:16,840 --> 00:21:18,200 That's a really great question. 689 00:21:18,200 --> 00:21:20,640 And I don't know that there's a simple answer. 690 00:21:20,640 --> 00:21:24,720 But I think by staying engaged and informed and curious, 691 00:21:24,720 --> 00:21:26,520 we can all play a part in making sure 692 00:21:26,520 --> 00:21:28,200 that AI is used for good. 693 00:21:28,200 --> 00:21:28,880 Well said. 694 00:21:28,880 --> 00:21:30,200 And for those of you listening who 695 00:21:30,200 --> 00:21:32,640 want to learn more about B-Star and the research 696 00:21:32,640 --> 00:21:34,360 that we talked about today, you can find links 697 00:21:34,360 --> 00:21:37,280 to the original paper and other resources in the show notes. 698 00:21:37,280 --> 00:21:38,920 Thanks for joining us, everyone. 699 00:21:38,920 --> 00:21:41,960 Until next time, keep exploring, keep learning, 700 00:21:41,960 --> 00:22:08,040 and keep asking those big questions.