1 00:00:00,000 --> 00:00:03,160 All right, ready to dive into some cutting edge AI research. 2 00:00:03,160 --> 00:00:04,000 Always. 3 00:00:04,000 --> 00:00:05,000 Awesome. 4 00:00:05,000 --> 00:00:07,380 Today, we're looking at a brand new paper, 5 00:00:07,380 --> 00:00:11,960 Palo Gema 2, a family of versatile VLMs for transfer. 6 00:00:11,960 --> 00:00:14,400 Oh yeah, this one's generating a lot of buzz. 7 00:00:14,400 --> 00:00:15,800 It's all about vision language models, 8 00:00:15,800 --> 00:00:17,040 or VLMs for short, right? 9 00:00:17,040 --> 00:00:17,880 You got it. 10 00:00:17,880 --> 00:00:20,760 So for our listeners who might not be familiar with VLMs, 11 00:00:20,760 --> 00:00:22,600 what exactly are we talking about here? 12 00:00:22,600 --> 00:00:26,160 Imagine AI that can not only see images, 13 00:00:26,160 --> 00:00:29,160 but actually understand them and connect them with text. 14 00:00:29,160 --> 00:00:31,800 Like they can look at a picture of a cat and know it's a cat, 15 00:00:31,800 --> 00:00:33,840 and then also understand a description of a cat. 16 00:00:33,840 --> 00:00:35,520 So it's like giving AI the ability 17 00:00:35,520 --> 00:00:37,800 to process information the way humans do, 18 00:00:37,800 --> 00:00:39,360 using both sight and language. 19 00:00:39,360 --> 00:00:40,520 Exactly. 20 00:00:40,520 --> 00:00:43,640 And Palo Gema 2 is pushing the boundaries of what's 21 00:00:43,640 --> 00:00:46,240 possible with these types of models. 22 00:00:46,240 --> 00:00:49,080 It's building on the success of the original Palo Gema, 23 00:00:49,080 --> 00:00:50,960 but with some serious upgrades. 24 00:00:50,960 --> 00:00:53,880 OK, so what makes Palo Gema 2 so special? 25 00:00:53,880 --> 00:00:56,920 What sets it apart from the original and other VLMs out 26 00:00:56,920 --> 00:00:57,280 there? 27 00:00:57,280 --> 00:00:59,640 Well, for starters, it's not just one model. 28 00:00:59,640 --> 00:01:01,160 It's a whole family of models. 29 00:01:01,160 --> 00:01:02,200 A family of models. 30 00:01:02,200 --> 00:01:04,400 Like different versions of the same AI. 31 00:01:04,400 --> 00:01:05,520 Precisely. 32 00:01:05,520 --> 00:01:07,800 They come in different sizes with varying numbers 33 00:01:07,800 --> 00:01:08,760 of parameters. 34 00:01:08,760 --> 00:01:11,440 You can think of parameters as a measure of the model's 35 00:01:11,440 --> 00:01:12,480 complexity. 36 00:01:12,480 --> 00:01:15,160 So some are smaller and faster, while others 37 00:01:15,160 --> 00:01:17,400 are larger and more powerful. 38 00:01:17,400 --> 00:01:17,880 Interesting. 39 00:01:17,880 --> 00:01:20,600 So it's like having different tools for different jobs. 40 00:01:20,600 --> 00:01:22,680 You might use a smaller, more efficient model 41 00:01:22,680 --> 00:01:25,160 for a simple task, and then bring out the big guns 42 00:01:25,160 --> 00:01:26,440 for something more complex. 43 00:01:26,440 --> 00:01:27,240 Exactly. 44 00:01:27,240 --> 00:01:28,400 And here's another cool thing. 45 00:01:28,400 --> 00:01:30,600 They also come in different resolutions. 46 00:01:30,600 --> 00:01:34,320 You know, like how you can have an image in standard definition 47 00:01:34,320 --> 00:01:35,400 or high definition. 48 00:01:35,400 --> 00:01:36,120 Yeah, it makes sense. 49 00:01:36,120 --> 00:01:38,120 Well, Palo Gema 2 can process images 50 00:01:38,120 --> 00:01:39,360 at different resolutions, too. 51 00:01:39,360 --> 00:01:42,800 It's like upgrading from standard def to HD, 52 00:01:42,800 --> 00:01:46,040 the higher the resolution, the more detail the model can see. 53 00:01:46,040 --> 00:01:48,520 Ah, so it's not just about having a big brain, 54 00:01:48,520 --> 00:01:49,760 but also sharp eyes. 55 00:01:49,760 --> 00:01:50,440 Exactly. 56 00:01:50,440 --> 00:01:52,880 And the researchers found that this increase in resolution 57 00:01:52,880 --> 00:01:55,160 was just as important as the size of the model 58 00:01:55,160 --> 00:01:57,200 when it came to performance improvement. 59 00:01:57,200 --> 00:01:58,080 That's fascinating. 60 00:01:58,080 --> 00:02:01,000 It really highlights how important visual detail 61 00:02:01,000 --> 00:02:02,720 is for these AI systems. 62 00:02:02,720 --> 00:02:05,080 OK, so we've got these different sizes and resolutions. 63 00:02:05,080 --> 00:02:08,800 How do they actually train these AI models to be so versatile? 64 00:02:08,800 --> 00:02:10,520 What's the secret sauce? 65 00:02:10,520 --> 00:02:12,200 It's a pretty sophisticated process, 66 00:02:12,200 --> 00:02:13,520 but I'll try to break it down. 67 00:02:13,520 --> 00:02:15,920 They use a three-stage training process. 68 00:02:15,920 --> 00:02:16,840 Three stages. 69 00:02:16,840 --> 00:02:17,760 All right, take me through it. 70 00:02:17,760 --> 00:02:18,880 What happens in stage one? 71 00:02:18,880 --> 00:02:21,520 Stage one is all about laying the foundation. 72 00:02:21,520 --> 00:02:23,320 You could say it's like AI boot camp. 73 00:02:23,320 --> 00:02:23,880 Boot camp. 74 00:02:23,880 --> 00:02:24,920 I like that analogy. 75 00:02:24,920 --> 00:02:27,400 So they start with two pre-trained models, 76 00:02:27,400 --> 00:02:30,000 Siglipso, 400 meters for the vision part, 77 00:02:30,000 --> 00:02:32,440 and Gemma 2 for the language part. 78 00:02:32,440 --> 00:02:34,680 And then they combine these two models 79 00:02:34,680 --> 00:02:38,200 and train them on a massive data set of a billion examples. 80 00:02:38,200 --> 00:02:39,520 Wow, a billion examples. 81 00:02:39,520 --> 00:02:40,800 That's a lot of data. 82 00:02:40,800 --> 00:02:41,400 It is. 83 00:02:41,400 --> 00:02:43,240 But that's how these models learn 84 00:02:43,240 --> 00:02:46,280 to make connections between images and text. 85 00:02:46,280 --> 00:02:48,560 They need to see a huge variety of examples 86 00:02:48,560 --> 00:02:50,920 to understand the nuances of the world. 87 00:02:50,920 --> 00:02:53,120 So it's like showing a child thousands of pictures 88 00:02:53,120 --> 00:02:55,760 and explaining what each one is until they start to understand 89 00:02:55,760 --> 00:02:57,360 how to recognize things on their own. 90 00:02:57,360 --> 00:02:58,880 It's similar to that, yeah. 91 00:02:58,880 --> 00:03:00,760 And then in stage two, they move on 92 00:03:00,760 --> 00:03:02,480 to more specialized training. 93 00:03:02,480 --> 00:03:06,000 This stage is all about focusing on tasks that really benefit 94 00:03:06,000 --> 00:03:07,240 from a keen eye for detail. 95 00:03:07,240 --> 00:03:08,640 OK, give me some examples. 96 00:03:08,640 --> 00:03:10,240 What kind of tasks are we talking about here? 97 00:03:10,240 --> 00:03:12,840 Think about things like optical character recognition, 98 00:03:12,840 --> 00:03:14,640 or OCR for short. 99 00:03:14,640 --> 00:03:17,480 That's the ability to read text from an image. 100 00:03:17,480 --> 00:03:19,400 Right, like when your phone can scan a document 101 00:03:19,400 --> 00:03:21,080 and turn it into editable text. 102 00:03:21,080 --> 00:03:23,760 Exactly, or recognizing handwriting, 103 00:03:23,760 --> 00:03:25,800 even if it's messy or stylized. 104 00:03:25,800 --> 00:03:27,720 You know, it's funny how some AI struggled 105 00:03:27,720 --> 00:03:29,720 to decipher handwritten notes. 106 00:03:29,720 --> 00:03:32,000 I wonder how Palo Gemma 2 does with that. 107 00:03:32,000 --> 00:03:34,520 Well, that's what the stage two training is all about. 108 00:03:34,520 --> 00:03:38,240 They crank up the resolution, first to 448 PX SEGAR, 109 00:03:38,240 --> 00:03:41,840 and then to a whopping 896 PX SEG, 110 00:03:41,840 --> 00:03:44,000 really honing in on those fine points. 111 00:03:44,000 --> 00:03:46,920 So they're giving the AI a super-powered magnifying glass 112 00:03:46,920 --> 00:03:48,880 to examine every little detail. 113 00:03:48,880 --> 00:03:49,520 That's pretty cool. 114 00:03:49,520 --> 00:03:50,760 And what about the third stage? 115 00:03:50,760 --> 00:03:51,560 What happens there? 116 00:03:51,560 --> 00:03:53,920 Stage three is all about fine-tuning. 117 00:03:53,920 --> 00:03:56,360 It's like taking those general skills from bootcamp 118 00:03:56,360 --> 00:03:58,920 and applying them to a specialized field. 119 00:03:58,920 --> 00:04:02,600 They take the model and train it further on specific tasks 120 00:04:02,600 --> 00:04:05,600 to make it an expert in that particular area. 121 00:04:05,600 --> 00:04:07,840 So it's like sending the AI to grad school 122 00:04:07,840 --> 00:04:09,360 to become a specialist. 123 00:04:09,360 --> 00:04:10,560 Makes sense. 124 00:04:10,560 --> 00:04:12,520 Now, the big question, 125 00:04:12,520 --> 00:04:15,080 did all that training actually pay off? 126 00:04:15,080 --> 00:04:17,560 Did Palo Gemma 2 live up to the hype? 127 00:04:17,560 --> 00:04:18,560 It absolutely did. 128 00:04:18,560 --> 00:04:20,720 And the results are pretty impressive. 129 00:04:20,720 --> 00:04:22,080 OK, I'm all ears. 130 00:04:22,080 --> 00:04:23,280 Spill the beans. 131 00:04:23,280 --> 00:04:25,680 What kind of amazing things can Palo Gemma 2 do? 132 00:04:25,680 --> 00:04:28,640 Well, first of all, it outperformed the original Palo Gemma 133 00:04:28,640 --> 00:04:30,280 by a significant margin. 134 00:04:30,280 --> 00:04:33,720 Not only that, it even beat some of the current state-of-the-art 135 00:04:33,720 --> 00:04:34,520 models out there. 136 00:04:34,520 --> 00:04:35,000 Wow. 137 00:04:35,000 --> 00:04:37,640 So it's like the valedictorian of the VLM class. 138 00:04:37,640 --> 00:04:38,200 That's a kettle. 139 00:04:38,200 --> 00:04:39,040 It's pretty impressive. 140 00:04:39,040 --> 00:04:40,840 And what's even more exciting is that they 141 00:04:40,840 --> 00:04:43,960 tested it on some unique tasks, going beyond the typical things 142 00:04:43,960 --> 00:04:45,240 you see with VLMs. 143 00:04:45,240 --> 00:04:46,120 OK, I'm intrigued. 144 00:04:46,120 --> 00:04:48,120 What kind of unusual tasks are we talking about here? 145 00:04:48,120 --> 00:04:49,040 Give me some examples. 146 00:04:49,040 --> 00:04:51,120 Well, remember that OCR ability we talked about? 147 00:04:51,120 --> 00:04:54,120 Palo Gemma 2 actually outperformed a specialized OCR 148 00:04:54,120 --> 00:04:55,720 model called HTS. 149 00:04:55,720 --> 00:04:56,120 Hold on. 150 00:04:56,120 --> 00:04:59,040 So it's beating the specialists at their own game. 151 00:04:59,040 --> 00:04:59,960 It seems so. 152 00:04:59,960 --> 00:05:01,360 And that's just the beginning. 153 00:05:01,360 --> 00:05:04,280 It can also understand tables, not just the words in them, 154 00:05:04,280 --> 00:05:06,360 but the actual layout and structure. 155 00:05:06,360 --> 00:05:08,320 You mean it can figure out which rows are headings, 156 00:05:08,320 --> 00:05:10,040 which columns are related, all that? 157 00:05:10,040 --> 00:05:11,280 That's amazing. 158 00:05:11,280 --> 00:05:11,800 Exactly. 159 00:05:11,800 --> 00:05:13,560 It's showing a deep understanding 160 00:05:13,560 --> 00:05:15,400 of the visual information. 161 00:05:15,400 --> 00:05:16,480 And there's more. 162 00:05:16,480 --> 00:05:20,040 Imagine showing it a drawing of a molecule. 163 00:05:20,040 --> 00:05:22,800 And it can actually figure out the chemical structure 164 00:05:22,800 --> 00:05:26,160 represented by that drawing, spitting out something called 165 00:05:26,160 --> 00:05:29,600 a smile string, which is like a code describing 166 00:05:29,600 --> 00:05:30,920 the molecule's makeup. 167 00:05:30,920 --> 00:05:31,840 Wow. 168 00:05:31,840 --> 00:05:33,760 That's getting in some serious science stuff. 169 00:05:33,760 --> 00:05:36,080 Palo Gemma 2 sounds like it could be incredibly valuable 170 00:05:36,080 --> 00:05:37,920 for researchers in chemistry and beyond. 171 00:05:37,920 --> 00:05:38,520 Absolutely. 172 00:05:38,520 --> 00:05:38,920 And get this. 173 00:05:38,920 --> 00:05:41,760 It can actually read sheet music and translate it 174 00:05:41,760 --> 00:05:43,200 into a digital format. 175 00:05:43,200 --> 00:05:44,880 Wait, are you serious? 176 00:05:44,880 --> 00:05:48,200 This AI can hear music just by looking at the sheet music? 177 00:05:48,200 --> 00:05:49,600 That's what the research suggests. 178 00:05:49,600 --> 00:05:52,560 Imagine taking a picture of a handwritten musical score 179 00:05:52,560 --> 00:05:55,480 and having it instantly playable on your computer. 180 00:05:55,480 --> 00:05:56,120 Wow. 181 00:05:56,120 --> 00:05:57,200 That's mind-blowing. 182 00:05:57,200 --> 00:05:59,600 What else can this AI prodigy do? 183 00:05:59,600 --> 00:06:01,440 Well, one thing that really stood out to me 184 00:06:01,440 --> 00:06:05,520 was its ability to generate incredibly detailed captions. 185 00:06:05,520 --> 00:06:06,360 Detailed captions. 186 00:06:06,360 --> 00:06:07,800 You mean like more than just saying, 187 00:06:07,800 --> 00:06:08,920 this is a picture of a cat. 188 00:06:08,920 --> 00:06:09,960 Way more than that. 189 00:06:09,960 --> 00:06:12,960 These captions are long and provide factual information 190 00:06:12,960 --> 00:06:15,480 about the image, almost like a short story. 191 00:06:15,480 --> 00:06:17,680 So it's like having an AI storyteller that 192 00:06:17,680 --> 00:06:21,440 can describe a picture with incredible accuracy and detail. 193 00:06:21,440 --> 00:06:22,000 Exactly. 194 00:06:22,000 --> 00:06:23,600 They tested it on a data set called 195 00:06:23,600 --> 00:06:27,160 DOCCI, which stands for Descriptions of Connected 196 00:06:27,160 --> 00:06:28,800 and Contrasting Images. 197 00:06:28,800 --> 00:06:32,280 And Palo Gemma 2 generated these super detailed captions, 198 00:06:32,280 --> 00:06:34,240 even pointing out the spatial relationships 199 00:06:34,240 --> 00:06:36,960 between objects, the number of objects, 200 00:06:36,960 --> 00:06:39,080 and adding general knowledge about the world. 201 00:06:39,080 --> 00:06:40,000 Wow. 202 00:06:40,000 --> 00:06:41,800 Almost like it's thinking for itself. 203 00:06:41,800 --> 00:06:44,480 Well, it is processing information 204 00:06:44,480 --> 00:06:46,280 and making connections in a way that's 205 00:06:46,280 --> 00:06:47,960 quite similar to human reasoning. 206 00:06:47,960 --> 00:06:49,240 It's pretty remarkable. 207 00:06:49,240 --> 00:06:52,920 So we've got an AI that can read, recognize complex structures, 208 00:06:52,920 --> 00:06:55,640 hear music, and tell detailed stories. 209 00:06:55,640 --> 00:06:57,320 It sounds almost too good to be true. 210 00:06:57,320 --> 00:06:59,560 Is there anything this model can't do? 211 00:06:59,560 --> 00:07:00,560 That's a great question. 212 00:07:00,560 --> 00:07:02,720 And it leads us to some really interesting findings 213 00:07:02,720 --> 00:07:04,880 about Palo Gemma 2's limitations. 214 00:07:04,880 --> 00:07:06,280 Limitations. 215 00:07:06,280 --> 00:07:07,480 OK, now I'm really curious. 216 00:07:07,480 --> 00:07:08,560 What are we talking about here? 217 00:07:08,560 --> 00:07:10,320 Let's delve into those next. 218 00:07:10,320 --> 00:07:13,880 So as amazing as Palo Gemma 2 is, 219 00:07:13,880 --> 00:07:16,120 it does have its limitations. 220 00:07:16,120 --> 00:07:17,760 No AI is perfect, right? 221 00:07:17,760 --> 00:07:18,920 Right, of course. 222 00:07:18,920 --> 00:07:20,680 What kind of limitations did they find? 223 00:07:20,680 --> 00:07:23,480 Well, one thing they noticed was that the largest model, the one 224 00:07:23,480 --> 00:07:26,240 with a whopping 20 billion parameters. 225 00:07:26,240 --> 00:07:27,800 Yeah, the one we were talking about earlier. 226 00:07:27,800 --> 00:07:29,240 Yeah, exactly. 227 00:07:29,240 --> 00:07:33,480 That one didn't always perform as well as they expected. 228 00:07:33,480 --> 00:07:34,120 That's interesting. 229 00:07:34,120 --> 00:07:36,480 I would have thought bigger would always mean better. 230 00:07:36,480 --> 00:07:37,880 Why do you think that was? 231 00:07:37,880 --> 00:07:39,600 Well, the researchers think it might 232 00:07:39,600 --> 00:07:42,160 be because of how it was trained. 233 00:07:42,160 --> 00:07:46,400 This particular model was trained completely from scratch. 234 00:07:46,400 --> 00:07:49,280 But the smaller models benefited from a technique 235 00:07:49,280 --> 00:07:51,200 called distillation. 236 00:07:51,200 --> 00:07:52,280 Distillation. 237 00:07:52,280 --> 00:07:54,080 Now, that's a word I haven't heard in a while. 238 00:07:54,080 --> 00:07:55,640 What does that mean in this context? 239 00:07:55,640 --> 00:07:58,120 OK, imagine a master chef teaching a student 240 00:07:58,120 --> 00:07:59,600 a complex recipe. 241 00:07:59,600 --> 00:08:02,520 Instead of going through every single step from scratch, 242 00:08:02,520 --> 00:08:05,280 they might give the student a simplified version first. 243 00:08:05,280 --> 00:08:06,120 I see. 244 00:08:06,120 --> 00:08:08,160 So they get the gist of it without getting bogged down 245 00:08:08,160 --> 00:08:09,240 in all the details. 246 00:08:09,240 --> 00:08:10,120 Exactly. 247 00:08:10,120 --> 00:08:12,760 And distillation for AI is kind of like that. 248 00:08:12,760 --> 00:08:15,160 You take a larger, more complex model 249 00:08:15,160 --> 00:08:17,840 and use it to teach a smaller model, transferring knowledge 250 00:08:17,840 --> 00:08:18,720 more efficiently. 251 00:08:18,720 --> 00:08:21,000 So it's like giving the smaller models a head start. 252 00:08:21,000 --> 00:08:22,840 Yeah, you could say that. 253 00:08:22,840 --> 00:08:24,480 And it seems like that gave them an edge 254 00:08:24,480 --> 00:08:26,040 when it came to performance. 255 00:08:26,040 --> 00:08:28,520 It highlights that it's not just about the size of the model, 256 00:08:28,520 --> 00:08:30,000 but also how you train it. 257 00:08:30,000 --> 00:08:31,000 So it's not just brawn. 258 00:08:31,000 --> 00:08:32,640 It's about brains, too. 259 00:08:32,640 --> 00:08:33,800 Exactly. 260 00:08:33,800 --> 00:08:36,520 And speaking of brains, another interesting observation 261 00:08:36,520 --> 00:08:38,920 they made was about how Palo Gema II handled 262 00:08:38,920 --> 00:08:41,080 traditional object detection tasks. 263 00:08:41,080 --> 00:08:42,040 Oh, and interesting. 264 00:08:42,040 --> 00:08:43,360 How did it do with that? 265 00:08:43,360 --> 00:08:45,480 I'm assuming not great if you're bringing it up, no? 266 00:08:45,480 --> 00:08:47,360 Well, it wasn't bad. 267 00:08:47,360 --> 00:08:49,680 But it didn't perform as well as they expected, 268 00:08:49,680 --> 00:08:52,560 considering how well it did with those other complex tasks, 269 00:08:52,560 --> 00:08:55,080 like understanding tables and molecules. 270 00:08:55,080 --> 00:08:56,640 Yeah, that is kind of surprising. 271 00:08:56,640 --> 00:08:58,640 You'd think if it can handle those things, 272 00:08:58,640 --> 00:09:01,080 basic object detection would be a piece of cake. 273 00:09:01,080 --> 00:09:01,760 Right. 274 00:09:01,760 --> 00:09:02,640 You'd think so. 275 00:09:02,640 --> 00:09:05,560 But it seems like even though it can see and understand 276 00:09:05,560 --> 00:09:08,440 images, it's not always speaking the same language 277 00:09:08,440 --> 00:09:09,960 as those traditional tests. 278 00:09:09,960 --> 00:09:11,120 That's a great way to put it. 279 00:09:11,120 --> 00:09:14,320 It's like it's learned a whole new way of seeing the world, 280 00:09:14,320 --> 00:09:16,480 but it still needs to figure out how to communicate that 281 00:09:16,480 --> 00:09:17,440 with the old systems. 282 00:09:17,440 --> 00:09:18,280 Exactly. 283 00:09:18,280 --> 00:09:19,800 The researchers believe this might 284 00:09:19,800 --> 00:09:22,720 be because of the language-based training approach they used 285 00:09:22,720 --> 00:09:24,160 for Palo Gema II. 286 00:09:24,160 --> 00:09:26,920 It doesn't quite align with how object detection is usually 287 00:09:26,920 --> 00:09:27,680 evaluated. 288 00:09:27,680 --> 00:09:28,480 Makes sense. 289 00:09:28,480 --> 00:09:31,440 So are there any areas where this unique approach actually 290 00:09:31,440 --> 00:09:33,760 gives Palo Gema II an advantage? 291 00:09:33,760 --> 00:09:35,320 That's a great question. 292 00:09:35,320 --> 00:09:38,240 And it leads us to one of the most exciting applications 293 00:09:38,240 --> 00:09:40,320 of this technology, medicine. 294 00:09:40,320 --> 00:09:40,720 Oh, yeah. 295 00:09:40,720 --> 00:09:42,440 You mentioned earlier that Palo Gema II 296 00:09:42,440 --> 00:09:44,840 had some really impressive results when it came 297 00:09:44,840 --> 00:09:48,000 to analyzing medical images, particularly chest X-rays. 298 00:09:48,000 --> 00:09:48,600 Exactly. 299 00:09:48,600 --> 00:09:53,560 They trained on a massive data set called MIMICXR, 300 00:09:53,560 --> 00:09:57,160 which has over 377,000 chest X-rays 301 00:09:57,160 --> 00:09:59,520 and their corresponding radiologist reports. 302 00:09:59,520 --> 00:10:00,000 Wow. 303 00:10:00,000 --> 00:10:01,280 That's a lot of X-rays. 304 00:10:01,280 --> 00:10:03,760 So they basically show the AI a ton of examples 305 00:10:03,760 --> 00:10:04,960 and taught it what to look for. 306 00:10:04,960 --> 00:10:05,440 Right. 307 00:10:05,440 --> 00:10:07,120 And the results were really remarkable. 308 00:10:07,120 --> 00:10:09,880 Palo Gema II was able to analyze chest X-rays 309 00:10:09,880 --> 00:10:11,800 and generate reports that were similar to what 310 00:10:11,800 --> 00:10:13,560 a human radiologist might write. 311 00:10:13,560 --> 00:10:13,920 Wow. 312 00:10:13,920 --> 00:10:16,000 So it's like having an AI radiologist. 313 00:10:16,000 --> 00:10:19,440 How did they even measure how accurate it was? 314 00:10:19,440 --> 00:10:22,400 They used a special metric called RadGraph, which 315 00:10:22,400 --> 00:10:25,720 compares the AI's reports to the actual radiologist reports 316 00:10:25,720 --> 00:10:26,840 and get this. 317 00:10:26,840 --> 00:10:30,200 Palo Gema II achieved state-of-the-art results. 318 00:10:30,200 --> 00:10:30,800 State-of-the-art. 319 00:10:30,800 --> 00:10:32,280 So it's not just doing a decent job. 320 00:10:32,280 --> 00:10:33,680 It's actually one of the best out there. 321 00:10:33,680 --> 00:10:34,600 That's right. 322 00:10:34,600 --> 00:10:36,640 And just like with those other tasks, 323 00:10:36,640 --> 00:10:38,400 they found that increasing the resolution 324 00:10:38,400 --> 00:10:41,560 and the size of the model led to even better performance. 325 00:10:41,560 --> 00:10:42,280 That's incredible. 326 00:10:42,280 --> 00:10:44,320 It's like giving the AI a better microscope 327 00:10:44,320 --> 00:10:46,320 to examine those X-rays. 328 00:10:46,320 --> 00:10:48,440 This has huge potential for health care. 329 00:10:48,440 --> 00:10:51,760 Imagine being able to use AI to assist radiologists, 330 00:10:51,760 --> 00:10:54,160 especially in areas where there's a shortage of specialists. 331 00:10:54,160 --> 00:10:54,880 Exactly. 332 00:10:54,880 --> 00:10:56,800 It could help streamline diagnoses, 333 00:10:56,800 --> 00:10:59,840 make health care more accessible, and ultimately save lives. 334 00:10:59,840 --> 00:11:00,480 That's amazing. 335 00:11:00,480 --> 00:11:02,320 This is truly groundbreaking stuff. 336 00:11:02,320 --> 00:11:04,600 It sounds like the research on Palo Gema II 337 00:11:04,600 --> 00:11:07,000 isn't just about incremental improvements. 338 00:11:07,000 --> 00:11:08,840 It's about fundamentally changing 339 00:11:08,840 --> 00:11:10,400 what's possible with AI. 340 00:11:10,400 --> 00:11:11,120 You got it. 341 00:11:11,120 --> 00:11:14,360 It's a giant leap towards creating truly versatile AI 342 00:11:14,360 --> 00:11:17,400 that can understand and interact with the world in ways 343 00:11:17,400 --> 00:11:19,200 we never thought possible before. 344 00:11:19,200 --> 00:11:20,400 I'm really blown away by all this. 345 00:11:20,400 --> 00:11:23,000 It makes me wonder, what's next? 346 00:11:23,000 --> 00:11:24,760 Where do we go from here? 347 00:11:24,760 --> 00:11:26,080 So where do we go from here? 348 00:11:26,080 --> 00:11:29,240 What's next for Palo Gema II and VLMs in general? 349 00:11:29,240 --> 00:11:30,560 What's the future look like? 350 00:11:30,560 --> 00:11:33,240 It's hard to say for sure, but I think the possibilities 351 00:11:33,240 --> 00:11:34,520 are incredibly exciting. 352 00:11:34,520 --> 00:11:36,640 The potential applications of this technology 353 00:11:36,640 --> 00:11:38,560 are vast, really. 354 00:11:38,560 --> 00:11:41,360 We've already talked about medicine and scientific research, 355 00:11:41,360 --> 00:11:44,080 but imagine AI that can not only analyze data, 356 00:11:44,080 --> 00:11:45,920 but also generate creative content. 357 00:11:45,920 --> 00:11:47,640 And I mean, based on visual input. 358 00:11:47,640 --> 00:11:50,160 You're talking about AI composing music, maybe? 359 00:11:50,160 --> 00:11:52,600 Or designing molecules, even writing screenplays 360 00:11:52,600 --> 00:11:53,800 based on what it sees. 361 00:11:53,800 --> 00:11:54,440 Exactly. 362 00:11:54,440 --> 00:11:56,160 It really raises some interesting questions 363 00:11:56,160 --> 00:11:57,720 about the future of creativity. 364 00:11:57,720 --> 00:12:00,360 What's the role of AI going to be in our lives? 365 00:12:00,360 --> 00:12:01,120 It's fascinating. 366 00:12:01,120 --> 00:12:01,840 For sure. 367 00:12:01,840 --> 00:12:04,000 But beyond those creative applications, 368 00:12:04,000 --> 00:12:06,480 there's also the potential for AI like this 369 00:12:06,480 --> 00:12:08,920 to help us tackle some of the world's biggest problems. 370 00:12:08,920 --> 00:12:09,560 Absolutely. 371 00:12:09,560 --> 00:12:12,040 I mean, think about climate change, disease diagnosis, 372 00:12:12,040 --> 00:12:13,480 even poverty. 373 00:12:13,480 --> 00:12:16,680 AI that can understand both images and text 374 00:12:16,680 --> 00:12:18,560 could be a game changer. 375 00:12:18,560 --> 00:12:20,440 It's like having this super-powered research 376 00:12:20,440 --> 00:12:23,800 assistant sift through mountains of data, 377 00:12:23,800 --> 00:12:26,520 identify patterns, help us find solutions 378 00:12:26,520 --> 00:12:27,960 that we might miss otherwise. 379 00:12:27,960 --> 00:12:28,600 Precisely. 380 00:12:28,600 --> 00:12:30,760 And maybe even solutions we couldn't have imagined 381 00:12:30,760 --> 00:12:31,720 without AI. 382 00:12:31,720 --> 00:12:32,400 Wow. 383 00:12:32,400 --> 00:12:34,320 So it sounds like the research on Palo Gemma 2 384 00:12:34,320 --> 00:12:36,840 isn't just about making things a little bit better. 385 00:12:36,840 --> 00:12:39,960 It's about fundamentally changing what AI can do. 386 00:12:39,960 --> 00:12:41,440 I think that's a great way to put it. 387 00:12:41,440 --> 00:12:45,720 It's a big step forward in creating truly versatile AI. 388 00:12:45,720 --> 00:12:49,080 AI that can understand the world in a way that, well, 389 00:12:49,080 --> 00:12:51,040 was unimaginable before. 390 00:12:51,040 --> 00:12:53,840 It really feels like we're on the edge of something big here. 391 00:12:53,840 --> 00:12:56,680 It's just been an incredible deep dive into Palo Gemma 2. 392 00:12:56,680 --> 00:12:59,320 I'm definitely walking away with a sense of awe 393 00:12:59,320 --> 00:13:01,800 at what this technology can achieve, that's for sure. 394 00:13:01,800 --> 00:13:03,400 It's been a pleasure exploring this with you. 395 00:13:03,400 --> 00:13:05,200 The future of AI is definitely bright. 396 00:13:05,200 --> 00:13:07,360 Yeah, I'm excited to see what breakthroughs come next. 397 00:13:07,360 --> 00:13:10,920 And to everyone listening out there, stay curious. 398 00:13:10,920 --> 00:13:31,080 Until next time, this is the Deep Dive, signing off.