1 00:00:00,000 --> 00:00:03,360 Okay, so imagine this AI that can understand language 2 00:00:03,360 --> 00:00:05,600 without like chopping it up into pieces first. 3 00:00:05,600 --> 00:00:07,240 Yeah, it's a pretty wild concept, right? 4 00:00:07,240 --> 00:00:08,440 Yeah, it kind of blows my mind. 5 00:00:08,440 --> 00:00:10,640 And that's exactly what we're gonna deep dive into today. 6 00:00:10,640 --> 00:00:11,480 Exactly. 7 00:00:11,480 --> 00:00:13,480 We have a fascinating paper all about 8 00:00:15,000 --> 00:00:19,000 bite-latent transformers, BLT for short. 9 00:00:19,000 --> 00:00:20,000 T-L-T, yeah. 10 00:00:20,000 --> 00:00:22,080 And how it's kind of shaking up the AI world. 11 00:00:22,080 --> 00:00:23,960 Yeah, so what's really interesting about this 12 00:00:23,960 --> 00:00:27,360 is this model, it actually works directly 13 00:00:27,360 --> 00:00:29,520 with the raw bites of information. 14 00:00:29,520 --> 00:00:30,360 Oh, wow. 15 00:00:30,360 --> 00:00:32,800 So those ones and zeros that make up the digital data. 16 00:00:32,800 --> 00:00:33,640 Okay. 17 00:00:33,640 --> 00:00:35,240 It basically skips over this whole process 18 00:00:35,240 --> 00:00:38,520 of tokenization that's been a standard in AI for ages. 19 00:00:38,520 --> 00:00:42,120 So for those of us who don't speak fluent AI, 20 00:00:42,120 --> 00:00:44,400 what is tokenization and like, 21 00:00:44,400 --> 00:00:46,400 why is ditching it such a big deal? 22 00:00:46,400 --> 00:00:48,400 Yeah, that's a great question. 23 00:00:48,400 --> 00:00:50,080 So think about tokenization, 24 00:00:50,080 --> 00:00:53,640 like breaking down a sentence into individual words. 25 00:00:53,640 --> 00:00:54,480 Okay. 26 00:00:54,480 --> 00:00:56,520 AI models, they traditionally use this method 27 00:00:56,520 --> 00:00:58,000 to try to understand the text. 28 00:00:58,000 --> 00:01:00,240 But imagine you're trying to understand a new language 29 00:01:00,240 --> 00:01:01,760 or maybe a technical document 30 00:01:01,760 --> 00:01:03,320 that's filled with a bunch of jargon. 31 00:01:03,320 --> 00:01:06,440 If your dictionary doesn't have those words, 32 00:01:06,440 --> 00:01:07,640 you're kind of stuck, right? 33 00:01:07,640 --> 00:01:09,200 Yeah, you're gonna have a bad time. 34 00:01:09,200 --> 00:01:12,080 That's exactly the limitation that BLT is trying to solve. 35 00:01:12,080 --> 00:01:15,040 So BLT is kind of like learning a language 36 00:01:15,040 --> 00:01:16,960 by understanding the letters 37 00:01:16,960 --> 00:01:19,840 instead of needing like a massive dictionary of words. 38 00:01:19,840 --> 00:01:20,680 Exactly. 39 00:01:20,680 --> 00:01:22,240 It's like going back to the basics. 40 00:01:22,240 --> 00:01:23,080 Interesting. 41 00:01:23,080 --> 00:01:25,240 And that's what makes it so much more adaptable. 42 00:01:25,240 --> 00:01:28,320 It can handle new words and complex terms. 43 00:01:28,320 --> 00:01:29,160 Wow. 44 00:01:29,160 --> 00:01:31,280 Even those pesky typos without missing a beat. 45 00:01:31,280 --> 00:01:33,240 Okay, cool, but how does it actually work? 46 00:01:33,240 --> 00:01:34,960 That is the question. 47 00:01:34,960 --> 00:01:37,800 And that's where this dynamic patching thing comes in. 48 00:01:37,800 --> 00:01:39,000 Ooh, that sounds fancy. 49 00:01:39,000 --> 00:01:40,320 What is dynamic patching? 50 00:01:40,320 --> 00:01:43,640 So imagine you're packing a suitcase. 51 00:01:43,640 --> 00:01:44,680 Okay. 52 00:01:44,680 --> 00:01:47,400 You wouldn't pack a super delicate vase. 53 00:01:47,400 --> 00:01:48,240 Right. 54 00:01:48,240 --> 00:01:50,080 The same way you would pack like a T-shirt. 55 00:01:50,080 --> 00:01:51,040 Of course not. 56 00:01:51,040 --> 00:01:53,600 BLT does something similar with bytes. 57 00:01:53,600 --> 00:01:55,320 It groups them into these little patches 58 00:01:55,320 --> 00:01:57,560 based on how complex the information is. 59 00:01:57,560 --> 00:01:59,600 So it's like it can tailor its focus 60 00:01:59,600 --> 00:02:00,720 depending on what it's reading. 61 00:02:00,720 --> 00:02:02,320 Yeah, like it's kind of reading the room. 62 00:02:02,320 --> 00:02:05,040 Spending more effort on the parts that are tricky 63 00:02:05,040 --> 00:02:06,320 and then just zipping through the stuff, 64 00:02:06,320 --> 00:02:08,000 that's pretty simple. 65 00:02:08,000 --> 00:02:09,680 It's all about efficiency. 66 00:02:09,680 --> 00:02:12,040 So it's not just like blindly chewing through data. 67 00:02:12,040 --> 00:02:13,600 It's actually being smart about it. 68 00:02:13,600 --> 00:02:14,440 Exactly. 69 00:02:14,440 --> 00:02:15,280 That's so cool. 70 00:02:15,280 --> 00:02:17,040 But how does it actually turn these patches 71 00:02:17,040 --> 00:02:18,840 into something that we can understand? 72 00:02:18,840 --> 00:02:20,200 Well, that's where the magic happened. 73 00:02:20,200 --> 00:02:21,040 The magic. 74 00:02:21,040 --> 00:02:21,880 Yeah. 75 00:02:21,880 --> 00:02:25,000 BLT, it uses this three-part system. 76 00:02:25,000 --> 00:02:26,160 Okay, a three-part system. 77 00:02:26,160 --> 00:02:28,320 Yeah, first you have the local encoder, 78 00:02:28,320 --> 00:02:30,320 which basically takes those raw bytes 79 00:02:30,320 --> 00:02:32,400 and maps them into these dynamic patches. 80 00:02:32,400 --> 00:02:33,240 Okay. 81 00:02:33,240 --> 00:02:35,720 Then we have the latent global transformer. 82 00:02:35,720 --> 00:02:36,560 Okay. 83 00:02:36,560 --> 00:02:38,680 Which is like the main powerhouse of the whole system. 84 00:02:38,680 --> 00:02:41,680 This is where the actual processing happens, 85 00:02:41,680 --> 00:02:44,880 understanding the relationships between those patches 86 00:02:44,880 --> 00:02:46,360 and extracting the meaning. 87 00:02:46,360 --> 00:02:48,480 So it's like the brains of the operation. 88 00:02:48,480 --> 00:02:49,600 That's a great way to put it. 89 00:02:49,600 --> 00:02:51,080 Okay, and then finally there's a decoder 90 00:02:51,080 --> 00:02:52,360 to kind of put it all back together 91 00:02:52,360 --> 00:02:53,320 so we can understand it. 92 00:02:53,320 --> 00:02:54,160 Yeah, you got it. 93 00:02:54,160 --> 00:02:57,400 The local decoder takes those process patches 94 00:02:57,400 --> 00:02:58,880 and turns them back into bytes. 95 00:02:58,880 --> 00:02:59,720 Okay. 96 00:02:59,720 --> 00:03:03,160 But there's this other really cool trick that BLT uses. 97 00:03:03,160 --> 00:03:04,560 Really get the hang of language. 98 00:03:04,560 --> 00:03:07,720 And it's called hash and gram embeddings. 99 00:03:07,720 --> 00:03:09,040 Hash and what? 100 00:03:09,040 --> 00:03:10,800 I know it sounds a little intimidating. 101 00:03:10,800 --> 00:03:11,720 Yeah, a little bit. 102 00:03:11,720 --> 00:03:13,520 But it's not as scary as it sounds. 103 00:03:13,520 --> 00:03:14,360 Okay. 104 00:03:14,360 --> 00:03:16,680 Just think about it like recognizing patterns 105 00:03:16,680 --> 00:03:18,360 in letters sequences. 106 00:03:18,360 --> 00:03:19,200 Okay. 107 00:03:19,200 --> 00:03:22,840 So for example, if you see the letter T in English, 108 00:03:22,840 --> 00:03:23,680 Right. 109 00:03:23,680 --> 00:03:25,360 there's a pretty good chance the next letter 110 00:03:25,360 --> 00:03:26,560 is gonna be an H. 111 00:03:26,560 --> 00:03:27,920 Yeah, A, yeah. 112 00:03:27,920 --> 00:03:28,920 Exactly. 113 00:03:28,920 --> 00:03:31,960 BLT uses these embeddings to kind of pick up 114 00:03:31,960 --> 00:03:34,880 on those same types of patterns, but in the bytes. 115 00:03:34,880 --> 00:03:35,720 Okay. 116 00:03:35,720 --> 00:03:37,760 Which really helps to understand the structure 117 00:03:37,760 --> 00:03:39,480 of the language at a deeper level. 118 00:03:39,480 --> 00:03:41,600 So it's not just seeing individual letters. 119 00:03:41,600 --> 00:03:43,600 It's like understanding how those letters 120 00:03:43,600 --> 00:03:44,440 usually go together. 121 00:03:44,440 --> 00:03:45,960 Exactly, it's all about context. 122 00:03:45,960 --> 00:03:47,720 Okay, that makes a lot of sense. 123 00:03:47,720 --> 00:03:50,480 But how does all this compare to the models 124 00:03:50,480 --> 00:03:51,760 that we're already using? 125 00:03:51,760 --> 00:03:53,400 That's a million dollar question. 126 00:03:53,400 --> 00:03:54,640 Yeah, is it actually better? 127 00:03:54,640 --> 00:03:56,240 Well, the research is pretty promising. 128 00:03:56,240 --> 00:04:00,200 They actually compared BLT to the state of the art model 129 00:04:00,200 --> 00:04:01,560 called LAME3. 130 00:04:01,560 --> 00:04:02,400 Okay. 131 00:04:02,400 --> 00:04:04,240 And they found that BLT could actually match 132 00:04:04,240 --> 00:04:08,600 its performance using significantly less processing power. 133 00:04:08,600 --> 00:04:09,640 Hold on, less power. 134 00:04:09,640 --> 00:04:10,480 Yeah. 135 00:04:10,480 --> 00:04:11,320 Same performance. 136 00:04:11,320 --> 00:04:12,920 Yeah, pretty amazing, right? 137 00:04:12,920 --> 00:04:14,360 That sounds like a win-win. 138 00:04:14,360 --> 00:04:15,200 Yeah. 139 00:04:15,200 --> 00:04:17,400 What does that mean for someone who's like building AI apps 140 00:04:17,400 --> 00:04:18,240 or something? 141 00:04:18,240 --> 00:04:19,960 So that means you could potentially achieve 142 00:04:19,960 --> 00:04:21,720 the same level of sophistication. 143 00:04:21,720 --> 00:04:22,560 Okay. 144 00:04:22,560 --> 00:04:25,320 But with way less computational resources. 145 00:04:25,320 --> 00:04:26,160 I like the sound of that. 146 00:04:26,160 --> 00:04:27,000 Yeah. 147 00:04:27,000 --> 00:04:28,520 It opens up a lot of possibilities, 148 00:04:28,520 --> 00:04:32,360 especially if you're working with limited hardware or budget. 149 00:04:32,360 --> 00:04:34,000 Right, or if you're just trying to be more efficient. 150 00:04:34,000 --> 00:04:35,080 Exactly. 151 00:04:35,080 --> 00:04:35,920 That's huge. 152 00:04:35,920 --> 00:04:38,120 I could really democratize access 153 00:04:38,120 --> 00:04:39,760 to these powerful AI models. 154 00:04:39,760 --> 00:04:41,280 Yeah, exactly. 155 00:04:41,280 --> 00:04:43,200 But hold on, comparing it to LAME3, 156 00:04:43,200 --> 00:04:45,440 isn't that kind of like apples or oranges? 157 00:04:45,440 --> 00:04:46,840 Yeah, that's a good point. 158 00:04:46,840 --> 00:04:50,280 Because LAME3 uses those fixed tokens right. 159 00:04:50,280 --> 00:04:53,000 While BLT is all about these dynamic patches. 160 00:04:53,000 --> 00:04:53,960 You're exactly right. 161 00:04:53,960 --> 00:04:55,600 And that's why the researchers, 162 00:04:55,600 --> 00:04:57,280 they took it a step further. 163 00:04:57,280 --> 00:05:00,960 They wanted to see how well each model could scale. 164 00:05:00,960 --> 00:05:04,800 So basically how they perform as you increase the size 165 00:05:04,800 --> 00:05:06,160 and the complexity. 166 00:05:06,160 --> 00:05:08,720 And that's where that whole patches are better 167 00:05:08,720 --> 00:05:10,560 than tokens thing really comes in. 168 00:05:10,560 --> 00:05:11,480 I am all ears. 169 00:05:11,480 --> 00:05:12,400 What did they find? 170 00:05:12,400 --> 00:05:15,760 So with BLT, you can do something really unique. 171 00:05:15,760 --> 00:05:17,520 You can increase the model size, 172 00:05:17,520 --> 00:05:20,000 A&D, the patch size at the same time, 173 00:05:20,000 --> 00:05:22,480 without needing more computational power. 174 00:05:22,480 --> 00:05:23,760 That sounds like a game changer. 175 00:05:23,760 --> 00:05:25,160 It is, it's a game changer, 176 00:05:25,160 --> 00:05:30,080 because it means BLT can adapt to more complex tasks 177 00:05:30,080 --> 00:05:32,520 and data sets without kind of hitting 178 00:05:32,520 --> 00:05:34,000 that performance bottleneck. 179 00:05:34,000 --> 00:05:36,240 So why can't traditional models do that? 180 00:05:36,240 --> 00:05:38,680 Because they're stuck with that fixed vocabulary 181 00:05:38,680 --> 00:05:40,280 and that fixed token size. 182 00:05:40,280 --> 00:05:42,520 It's like trying to fit more and more clothes 183 00:05:42,520 --> 00:05:44,640 into a suitcase that can't get any bigger. 184 00:05:44,640 --> 00:05:45,480 That's not gonna work. 185 00:05:45,480 --> 00:05:48,240 But BLT is like having this magical suitcase 186 00:05:48,240 --> 00:05:49,760 that expands as you need it. 187 00:05:49,760 --> 00:05:50,920 Okay, I love that analogy. 188 00:05:50,920 --> 00:05:54,480 So more flexibility, less computing power. 189 00:05:54,480 --> 00:05:55,840 What's the catch? 190 00:05:55,840 --> 00:05:57,480 Well, there's always a catch, right? 191 00:05:57,480 --> 00:05:59,880 So the research is still in its early stages. 192 00:05:59,880 --> 00:06:03,280 While BLT shows this incredible potential, 193 00:06:03,280 --> 00:06:05,360 there's a lot more to explore. 194 00:06:05,360 --> 00:06:08,160 The existing tools and the infrastructure for AI, 195 00:06:08,160 --> 00:06:12,480 they're mostly designed for those token-based models. 196 00:06:12,480 --> 00:06:16,440 So we need new approaches to really optimize BLT 197 00:06:16,440 --> 00:06:18,320 and to unlock its full power. 198 00:06:18,320 --> 00:06:20,120 So it's like we have this super fast car, 199 00:06:20,120 --> 00:06:21,720 but there's no race track design for it. 200 00:06:21,720 --> 00:06:23,200 Exactly, that's a great analogy. 201 00:06:23,200 --> 00:06:24,200 It reaches its top speed. 202 00:06:24,200 --> 00:06:26,640 Yeah, but the potential is definitely there. 203 00:06:26,640 --> 00:06:29,680 And the researchers, they're already looking into ways 204 00:06:29,680 --> 00:06:32,800 to improve things like the training process 205 00:06:32,800 --> 00:06:35,320 and make BLT even more efficient. 206 00:06:35,320 --> 00:06:38,880 So BLT might not be ready to take over the world just yet. 207 00:06:38,880 --> 00:06:40,560 But it's definitely a force to be reckoned with. 208 00:06:40,560 --> 00:06:41,400 Absolutely. 209 00:06:41,400 --> 00:06:44,040 This is getting me really excited about the future of AI. 210 00:06:44,040 --> 00:06:46,640 What other cool things did they find out about BLT? 211 00:06:46,640 --> 00:06:48,160 Oh, there's a lot more to unpack. 212 00:06:48,160 --> 00:06:50,040 What kind of tasks is it really good at? 213 00:06:50,040 --> 00:06:51,280 Let's dive into that. 214 00:06:51,280 --> 00:06:53,400 So one of the really impressive things they found 215 00:06:53,400 --> 00:06:59,880 is that BLT can handle messy real-world data, like a champ. 216 00:06:59,880 --> 00:07:03,640 Think about social media posts with all that slang and typos, 217 00:07:03,640 --> 00:07:06,760 or maybe technical documents filled with jargon. 218 00:07:06,760 --> 00:07:07,920 Yeah, that stuff could be tough. 219 00:07:07,920 --> 00:07:09,520 Yeah, those kinds of things can really 220 00:07:09,520 --> 00:07:11,200 trick up those traditional models. 221 00:07:11,200 --> 00:07:13,560 That makes sense if you're relying on a fixed dictionary. 222 00:07:13,560 --> 00:07:15,920 Anything outside of that is going to throw you off. 223 00:07:15,920 --> 00:07:16,800 Exactly. 224 00:07:16,800 --> 00:07:19,720 But BLT's like bite-level approach seems 225 00:07:19,720 --> 00:07:22,440 way more adaptable to those kind of curveballs. 226 00:07:22,440 --> 00:07:22,960 Exactly. 227 00:07:22,960 --> 00:07:25,480 It's like it can roll with the punches. 228 00:07:25,480 --> 00:07:26,160 OK, cool. 229 00:07:26,160 --> 00:07:28,240 So it's not just about efficiency then. 230 00:07:28,240 --> 00:07:30,360 It's about fairness and accuracy too. 231 00:07:30,360 --> 00:07:31,680 It's about robustness. 232 00:07:31,680 --> 00:07:33,120 Yeah, that's a really important point. 233 00:07:33,120 --> 00:07:33,680 It is. 234 00:07:33,680 --> 00:07:35,520 But how did they actually test this? 235 00:07:35,520 --> 00:07:38,920 So to test this, what kind of evidence did they find? 236 00:07:38,920 --> 00:07:42,440 So they used this benchmark called CET. 237 00:07:42,440 --> 00:07:44,000 C-E-T. 238 00:07:44,000 --> 00:07:45,120 C-U-E? 239 00:07:45,120 --> 00:07:47,800 Yeah, it's a cute name for a pretty tough test. 240 00:07:47,800 --> 00:07:48,400 Oh, OK. 241 00:07:48,400 --> 00:07:50,680 It tests how well a model understands 242 00:07:50,680 --> 00:07:54,280 the building blocks of language, like those individual characters 243 00:07:54,280 --> 00:07:54,840 and sounds. 244 00:07:54,840 --> 00:07:55,880 OK, the nitty-gritty. 245 00:07:55,880 --> 00:07:57,040 Yeah, the tiny little pieces. 246 00:07:57,040 --> 00:07:57,600 OK. 247 00:07:57,600 --> 00:08:01,240 And these tasks are surprisingly tricky for those token-based 248 00:08:01,240 --> 00:08:01,960 models. 249 00:08:01,960 --> 00:08:02,320 OK. 250 00:08:02,320 --> 00:08:04,720 But BLT excelled. 251 00:08:04,720 --> 00:08:07,240 So wait, it's not just understanding words. 252 00:08:07,240 --> 00:08:09,760 It's understanding the individual letters. 253 00:08:09,760 --> 00:08:10,400 Yeah. 254 00:08:10,400 --> 00:08:11,960 And how they relate to each other. 255 00:08:11,960 --> 00:08:13,640 Yeah, it's going deep. 256 00:08:13,640 --> 00:08:15,160 That's pretty amazing. 257 00:08:15,160 --> 00:08:17,440 What kind of tasks did they test specifically? 258 00:08:17,440 --> 00:08:20,040 So they look at things like spelling correction. 259 00:08:20,040 --> 00:08:20,760 Oh, wow. 260 00:08:20,760 --> 00:08:23,360 Identifying words with similar spellings, 261 00:08:23,360 --> 00:08:26,400 even figuring out how a word should be pronounced just 262 00:08:26,400 --> 00:08:27,600 by looking at the letters. 263 00:08:27,600 --> 00:08:28,760 That's wild. 264 00:08:28,760 --> 00:08:32,320 Yeah, BLT achieved near perfect accuracy 265 00:08:32,320 --> 00:08:33,400 on some of those tasks. 266 00:08:33,400 --> 00:08:34,120 That's incredible. 267 00:08:34,120 --> 00:08:39,120 It's like BLT is learning to read and understand language 268 00:08:39,120 --> 00:08:40,080 the way humans do. 269 00:08:40,080 --> 00:08:41,840 Yeah, it's learning like a kid. 270 00:08:41,840 --> 00:08:44,040 By paying attention to all the smallest details. 271 00:08:44,040 --> 00:08:45,280 It's really paying attention. 272 00:08:45,280 --> 00:08:47,840 But if it's so good at these fundamental tasks, 273 00:08:47,840 --> 00:08:51,680 what about more complex applications, like translation, 274 00:08:51,680 --> 00:08:52,240 for example? 275 00:08:52,240 --> 00:08:52,640 That's work. 276 00:08:52,640 --> 00:08:53,560 It's really interesting. 277 00:08:53,560 --> 00:08:54,000 OK. 278 00:08:54,000 --> 00:08:56,920 They found that BLT was especially good at translating 279 00:08:56,920 --> 00:08:59,440 languages with very little training data. 280 00:08:59,440 --> 00:09:00,000 Oh, right. 281 00:09:00,000 --> 00:09:02,720 So think about rare indigenous languages. 282 00:09:02,720 --> 00:09:03,080 OK. 283 00:09:03,080 --> 00:09:05,720 Or maybe dialects with limited written resources. 284 00:09:05,720 --> 00:09:06,160 Right. 285 00:09:06,160 --> 00:09:07,760 Where you don't have tons of data to work with. 286 00:09:07,760 --> 00:09:08,200 Exactly. 287 00:09:08,200 --> 00:09:10,960 And BLT was able to outperform those traditional models 288 00:09:10,960 --> 00:09:12,640 by a significant margin. 289 00:09:12,640 --> 00:09:14,080 Wow, that's a game changer. 290 00:09:14,080 --> 00:09:14,920 Yeah, it is. 291 00:09:14,920 --> 00:09:18,000 It could help preserve endangered languages 292 00:09:18,000 --> 00:09:19,800 or like bridge communication gaps 293 00:09:19,800 --> 00:09:20,920 between different cultures. 294 00:09:20,920 --> 00:09:21,880 Yeah, exactly. 295 00:09:21,880 --> 00:09:23,240 It has huge potential. 296 00:09:23,240 --> 00:09:26,240 But how is it able to do that with so little data? 297 00:09:26,240 --> 00:09:27,680 That's the question, right? 298 00:09:27,680 --> 00:09:29,560 Yeah, is it like AI magic or something? 299 00:09:29,560 --> 00:09:32,280 Well, researchers think it has to do with that direct engagement 300 00:09:32,280 --> 00:09:33,360 with the bytes. 301 00:09:33,360 --> 00:09:33,840 OK. 302 00:09:33,840 --> 00:09:37,000 By looking at that raw data, it can pick up 303 00:09:37,000 --> 00:09:40,560 on these subtle patterns and relationships 304 00:09:40,560 --> 00:09:44,320 that those token-based models might miss. 305 00:09:44,320 --> 00:09:44,800 OK. 306 00:09:44,800 --> 00:09:48,080 It's like BLT is a more attentive student. 307 00:09:48,080 --> 00:09:48,720 I like that. 308 00:09:48,720 --> 00:09:49,920 It's really paying attention. 309 00:09:49,920 --> 00:09:51,000 It's soaking it all up. 310 00:09:51,000 --> 00:09:51,720 Exactly. 311 00:09:51,720 --> 00:09:53,000 OK, that makes a lot of sense. 312 00:09:53,000 --> 00:09:54,680 It's like the difference between trying 313 00:09:54,680 --> 00:09:58,360 to learn a language from a phrase book versus like immersing 314 00:09:58,360 --> 00:10:01,080 yourself in the culture and absorbing all the nuances 315 00:10:01,080 --> 00:10:01,760 naturally. 316 00:10:01,760 --> 00:10:03,000 That's a great analogy. 317 00:10:03,000 --> 00:10:05,880 So all this talk about BLT's potential is really exciting. 318 00:10:05,880 --> 00:10:06,400 Yeah. 319 00:10:06,400 --> 00:10:07,440 What are the next steps? 320 00:10:07,440 --> 00:10:11,000 What kind of research needs to be done to make this a reality? 321 00:10:11,000 --> 00:10:14,240 Well, there's a few exciting avenues for future exploration. 322 00:10:14,240 --> 00:10:15,200 Lay it on me. 323 00:10:15,200 --> 00:10:17,640 One is optimizing that training process. 324 00:10:17,640 --> 00:10:18,040 OK. 325 00:10:18,040 --> 00:10:21,000 Right now, it takes a lot of time and computing power 326 00:10:21,000 --> 00:10:23,920 to train these large language models. 327 00:10:23,920 --> 00:10:25,720 But researchers are looking for ways 328 00:10:25,720 --> 00:10:27,520 to make that faster and more efficient. 329 00:10:27,520 --> 00:10:29,480 OK, because that's key. 330 00:10:29,480 --> 00:10:31,920 If we want to make this tech accessible to more people. 331 00:10:31,920 --> 00:10:33,640 Yeah, we got to make it practical. 332 00:10:33,640 --> 00:10:34,160 OK. 333 00:10:34,160 --> 00:10:35,240 What else have they worked on? 334 00:10:35,240 --> 00:10:37,360 Another area is figuring out the best way 335 00:10:37,360 --> 00:10:39,600 to biteify existing models. 336 00:10:39,600 --> 00:10:42,000 OK, we talked about how BLT starts from scratch. 337 00:10:42,000 --> 00:10:42,520 Right. 338 00:10:42,520 --> 00:10:44,960 But what if we could take the knowledge that's 339 00:10:44,960 --> 00:10:49,400 already stored in those massive token based models 340 00:10:49,400 --> 00:10:51,840 and like convert it to a bite level format? 341 00:10:51,840 --> 00:10:54,000 Yeah, what if we could just translate it? 342 00:10:54,000 --> 00:10:55,560 That would be a huge time saver. 343 00:10:55,560 --> 00:10:56,200 Yeah, it would be. 344 00:10:56,200 --> 00:10:57,760 It's like giving BLT a head start. 345 00:10:57,760 --> 00:10:58,600 A little cheat sheet. 346 00:10:58,600 --> 00:10:58,960 Right. 347 00:10:58,960 --> 00:10:59,800 Exactly. 348 00:10:59,800 --> 00:11:01,680 Did the researchers experiment with that at all? 349 00:11:01,680 --> 00:11:02,520 They did. 350 00:11:02,520 --> 00:11:06,840 They took a pre-trained model called LAM-A 3.1. 351 00:11:06,840 --> 00:11:07,360 OK. 352 00:11:07,360 --> 00:11:11,760 And they used its parameters to basically initialize BLT. 353 00:11:11,760 --> 00:11:12,360 Oh, wow. 354 00:11:12,360 --> 00:11:14,320 And the results were pretty promising, 355 00:11:14,320 --> 00:11:16,360 showing that this approach could lead to faster 356 00:11:16,360 --> 00:11:18,320 and more efficient training. 357 00:11:18,320 --> 00:11:21,960 So we could potentially build on the work that's already 358 00:11:21,960 --> 00:11:22,640 been done. 359 00:11:22,640 --> 00:11:23,160 Yeah. 360 00:11:23,160 --> 00:11:25,160 Instead of having to reinvent the wheel every time. 361 00:11:25,160 --> 00:11:25,600 Exactly. 362 00:11:25,600 --> 00:11:27,240 We don't have to start from zero. 363 00:11:27,240 --> 00:11:28,160 That's super cool. 364 00:11:28,160 --> 00:11:28,880 Yeah. 365 00:11:28,880 --> 00:11:30,680 But what about the practical side of things? 366 00:11:30,680 --> 00:11:32,480 Yeah, the real world stuff. 367 00:11:32,480 --> 00:11:36,160 Yeah, we talked earlier about the need for new tools 368 00:11:36,160 --> 00:11:38,040 and infrastructure to support BLT. 369 00:11:38,040 --> 00:11:38,440 Why? 370 00:11:38,440 --> 00:11:39,680 All the behind the scenes stuff. 371 00:11:39,680 --> 00:11:39,920 Yeah. 372 00:11:39,920 --> 00:11:41,200 What does that actually look like? 373 00:11:41,200 --> 00:11:46,680 So right now, most of the libraries and code bases for AI 374 00:11:46,680 --> 00:11:49,640 are optimized for token based models. 375 00:11:49,640 --> 00:11:50,040 OK. 376 00:11:50,040 --> 00:11:53,880 So we need to develop new software and hardware 377 00:11:53,880 --> 00:11:56,960 that can take advantage of BLT's unique capabilities. 378 00:11:56,960 --> 00:11:57,320 OK. 379 00:11:57,320 --> 00:12:00,000 Like that dynamic patching and byte level processing. 380 00:12:00,000 --> 00:12:00,320 OK. 381 00:12:00,320 --> 00:12:01,200 So it's a big challenge. 382 00:12:01,200 --> 00:12:02,400 Yeah, it's a big undertaking. 383 00:12:02,400 --> 00:12:04,000 But if we can overcome it. 384 00:12:04,000 --> 00:12:04,920 I think we can pull it off. 385 00:12:04,920 --> 00:12:06,680 The potential benefits seem huge. 386 00:12:06,680 --> 00:12:07,320 Yeah. 387 00:12:07,320 --> 00:12:10,680 More efficient models, better language understanding, 388 00:12:10,680 --> 00:12:12,920 fairer and more accurate results. 389 00:12:12,920 --> 00:12:13,240 Yeah. 390 00:12:13,240 --> 00:12:14,480 It's a whole new world. 391 00:12:14,480 --> 00:12:17,240 It's like opening up this whole new world of possibilities. 392 00:12:17,240 --> 00:12:17,920 Exactly. 393 00:12:17,920 --> 00:12:21,400 And it's important to remember this research is still going. 394 00:12:21,400 --> 00:12:21,560 Right. 395 00:12:21,560 --> 00:12:24,720 There's still so much more to learn and explore. 396 00:12:24,720 --> 00:12:25,200 OK. 397 00:12:25,200 --> 00:12:28,800 So what's the impact of BLT and its impact for the future of AI? 398 00:12:28,800 --> 00:12:32,400 This whole deep dive into BLT has been absolutely fascinating. 399 00:12:32,400 --> 00:12:33,520 I'm glad you're enjoying it. 400 00:12:33,520 --> 00:12:33,760 Yeah. 401 00:12:33,760 --> 00:12:36,520 It's made me think about AI in a completely different way. 402 00:12:36,520 --> 00:12:36,840 Yeah. 403 00:12:36,840 --> 00:12:38,920 It really challenges the status quo. 404 00:12:38,920 --> 00:12:39,280 Yeah. 405 00:12:39,280 --> 00:12:41,280 It's not just about processing words anymore. 406 00:12:41,280 --> 00:12:45,920 It's about understanding the fundamental building blocks 407 00:12:45,920 --> 00:12:46,600 of language. 408 00:12:46,600 --> 00:12:47,880 A DNA of language. 409 00:12:47,880 --> 00:12:48,560 Right. 410 00:12:48,560 --> 00:12:50,360 But before we wrap up, I'm curious. 411 00:12:50,360 --> 00:12:50,800 Yeah. 412 00:12:50,800 --> 00:12:52,960 What are your thoughts on the bigger picture? 413 00:12:52,960 --> 00:12:54,960 Ooh, the big picture. 414 00:12:54,960 --> 00:12:57,400 What does this research tell us about the direction of AI 415 00:12:57,400 --> 00:12:58,080 as a whole? 416 00:12:58,080 --> 00:12:59,880 Well, I think BLT really highlights 417 00:12:59,880 --> 00:13:02,240 the importance of thinking outside the box. 418 00:13:02,240 --> 00:13:02,800 OK. 419 00:13:02,800 --> 00:13:04,360 You know, for years, tokenization 420 00:13:04,360 --> 00:13:05,840 has been the dominant approach. 421 00:13:05,840 --> 00:13:06,040 Great. 422 00:13:06,040 --> 00:13:06,880 The go-to method. 423 00:13:06,880 --> 00:13:07,440 Yeah. 424 00:13:07,440 --> 00:13:08,560 Everyone was doing it. 425 00:13:08,560 --> 00:13:09,000 OK. 426 00:13:09,000 --> 00:13:10,680 But this research shows that there 427 00:13:10,680 --> 00:13:13,800 are other ways that might be even more effective 428 00:13:13,800 --> 00:13:17,000 and aligned with how humans actually understand language. 429 00:13:17,000 --> 00:13:18,480 That's a really powerful insight. 430 00:13:18,480 --> 00:13:19,040 It is. 431 00:13:19,040 --> 00:13:22,440 It seems like BLT is really pushing the boundaries 432 00:13:22,440 --> 00:13:24,800 of what we thought was possible with AI. 433 00:13:24,800 --> 00:13:26,720 Yeah, it's breaking down those barriers. 434 00:13:26,720 --> 00:13:28,440 But what are the potential downsides? 435 00:13:28,440 --> 00:13:29,040 Yeah. 436 00:13:29,040 --> 00:13:32,880 Are there any ethical considerations or risks 437 00:13:32,880 --> 00:13:34,200 that we need to be aware of? 438 00:13:34,200 --> 00:13:37,440 Oh, that's a great question and a really important one too. 439 00:13:37,440 --> 00:13:39,680 As this technology develops, yeah, 440 00:13:39,680 --> 00:13:41,840 there's always that potential for misuse. 441 00:13:41,840 --> 00:13:43,320 Right, because it's so powerful. 442 00:13:43,320 --> 00:13:43,960 Exactly. 443 00:13:43,960 --> 00:13:46,400 With any powerful technology, you 444 00:13:46,400 --> 00:13:50,560 need to be really careful about how it's used as BLT 445 00:13:50,560 --> 00:13:51,240 and other models. 446 00:13:51,240 --> 00:13:53,200 Like, it get more sophisticated. 447 00:13:53,200 --> 00:13:56,160 We need to think about how they're being used and developed. 448 00:13:56,160 --> 00:13:56,800 OK. 449 00:13:56,800 --> 00:14:00,560 Things like bias in the data that's used to train them 450 00:14:00,560 --> 00:14:03,120 and the possibility that they could be used to generate 451 00:14:03,120 --> 00:14:04,480 harmful content. 452 00:14:04,480 --> 00:14:06,040 Oh, yeah, that's a big one. 453 00:14:06,040 --> 00:14:07,840 And we have to make sure that they're being developed 454 00:14:07,840 --> 00:14:08,800 responsibly. 455 00:14:08,800 --> 00:14:10,360 So it's a lot of responsibility. 456 00:14:10,360 --> 00:14:11,480 Yeah, it's a big responsibility. 457 00:14:11,480 --> 00:14:13,200 Like, we're not just building a better engine. 458 00:14:13,200 --> 00:14:15,240 We're building a whole new vehicle. 459 00:14:15,240 --> 00:14:17,320 Exactly, a whole new way of thinking. 460 00:14:17,320 --> 00:14:21,120 And we need to make sure it's safe and ethical to drive. 461 00:14:21,120 --> 00:14:23,560 It's like giving AI a driver's license. 462 00:14:23,560 --> 00:14:24,520 Right, exactly. 463 00:14:24,520 --> 00:14:27,440 So focusing on the positive for a second. 464 00:14:27,440 --> 00:14:29,160 What are you most excited about when 465 00:14:29,160 --> 00:14:30,960 you think about the future of BLT? 466 00:14:30,960 --> 00:14:31,720 OK. 467 00:14:31,720 --> 00:14:33,800 And bite-level AI in general? 468 00:14:33,800 --> 00:14:36,160 I'm really excited about the potential for this thing 469 00:14:36,160 --> 00:14:38,320 called multimodal AI. 470 00:14:38,320 --> 00:14:39,480 Multimodal AI. 471 00:14:39,480 --> 00:14:40,720 Yeah, it's a mouthful. 472 00:14:40,720 --> 00:14:44,080 But it basically means models that can understand, 473 00:14:44,080 --> 00:14:48,000 not just text, but also images, audio, video. 474 00:14:48,000 --> 00:14:48,600 Oh, wow. 475 00:14:48,600 --> 00:14:50,920 So it's like AI that can experience the world 476 00:14:50,920 --> 00:14:52,160 through all our senses. 477 00:14:52,160 --> 00:14:54,440 Yeah, it's like giving AI all the senses. 478 00:14:54,440 --> 00:14:55,240 That's amazing. 479 00:14:55,240 --> 00:14:55,760 It is. 480 00:14:55,760 --> 00:14:57,160 What would that look like in practice? 481 00:14:57,160 --> 00:15:00,120 So imagine a model that can seamlessly analyze 482 00:15:00,120 --> 00:15:01,520 the text of a news article. 483 00:15:01,520 --> 00:15:02,080 OK. 484 00:15:02,080 --> 00:15:04,280 And then look at the images that go with it. 485 00:15:04,280 --> 00:15:04,600 Very. 486 00:15:04,600 --> 00:15:07,960 And even listen to the tone of voice in a related audio clip. 487 00:15:07,960 --> 00:15:08,440 Oh, wow. 488 00:15:08,440 --> 00:15:12,200 And put all of that together into one unified understanding. 489 00:15:12,200 --> 00:15:13,040 That's mind-blowing. 490 00:15:13,040 --> 00:15:16,000 It would be like having an AI assistant that could really 491 00:15:16,000 --> 00:15:19,760 grasp the world in all its complexity and richness. 492 00:15:19,760 --> 00:15:20,720 That would be incredible. 493 00:15:20,720 --> 00:15:21,880 That would be a game changer. 494 00:15:21,880 --> 00:15:23,760 But it's not just about understanding, right? 495 00:15:23,760 --> 00:15:24,240 No. 496 00:15:24,240 --> 00:15:25,720 What about the creative potential? 497 00:15:25,720 --> 00:15:27,480 Oh, that's the really exciting part. 498 00:15:27,480 --> 00:15:27,760 Yeah. 499 00:15:27,760 --> 00:15:31,880 Imagine being able to generate these brand new forms 500 00:15:31,880 --> 00:15:33,720 of art and media. 501 00:15:33,720 --> 00:15:34,080 Oh, wow. 502 00:15:34,080 --> 00:15:35,960 Blending those different modalities. 503 00:15:35,960 --> 00:15:36,440 OK. 504 00:15:36,440 --> 00:15:38,680 In ways that we haven't even thought of yet. 505 00:15:38,680 --> 00:15:42,200 So like AI-generated art, but on a whole other level? 506 00:15:42,200 --> 00:15:44,360 Yeah, it's taking it to the next dimension. 507 00:15:44,360 --> 00:15:48,600 A world where AI can help us tell stories, compose music, 508 00:15:48,600 --> 00:15:51,160 create art in ways we haven't even imagined. 509 00:15:51,160 --> 00:15:52,760 It's like opening up this whole new dimension 510 00:15:52,760 --> 00:15:54,240 of artistic expression. 511 00:15:54,240 --> 00:15:54,720 It is. 512 00:15:54,720 --> 00:15:55,220 Yes. 513 00:15:55,220 --> 00:15:56,800 Powered by AI. 514 00:15:56,800 --> 00:15:57,320 But hold on. 515 00:15:57,320 --> 00:16:00,960 If we're talking about AI models that can process all this data. 516 00:16:00,960 --> 00:16:01,480 Yeah. 517 00:16:01,480 --> 00:16:02,520 All that data. 518 00:16:02,520 --> 00:16:06,120 Aren't we also talking about massive computational demands? 519 00:16:06,120 --> 00:16:07,040 That's a good point. 520 00:16:07,040 --> 00:16:07,540 Yeah. 521 00:16:07,540 --> 00:16:10,280 Wouldn't that limit who could actually access and use 522 00:16:10,280 --> 00:16:11,160 this technology? 523 00:16:11,160 --> 00:16:11,480 Yeah. 524 00:16:11,480 --> 00:16:12,600 That's a valid concern. 525 00:16:12,600 --> 00:16:15,000 But that's where the efficiency of BLT 526 00:16:15,000 --> 00:16:16,360 becomes super important. 527 00:16:16,360 --> 00:16:16,800 OK. 528 00:16:16,800 --> 00:16:20,360 If we can develop these models that can process information 529 00:16:20,360 --> 00:16:21,480 at the byte level. 530 00:16:21,480 --> 00:16:21,720 Right. 531 00:16:21,720 --> 00:16:24,920 With the same efficiency and scalability as BLT. 532 00:16:24,920 --> 00:16:25,440 OK. 533 00:16:25,440 --> 00:16:30,480 We could potentially make those multimodal dreams a reality 534 00:16:30,480 --> 00:16:33,720 without needing those massive supercomputers. 535 00:16:33,720 --> 00:16:34,440 That's amazing. 536 00:16:34,440 --> 00:16:36,720 It could democratize access to these tools. 537 00:16:36,720 --> 00:16:37,000 Right. 538 00:16:37,000 --> 00:16:38,480 And make them available to anyone. 539 00:16:38,480 --> 00:16:40,840 Enabling innovation across different fields. 540 00:16:40,840 --> 00:16:41,800 That's really cool. 541 00:16:41,800 --> 00:16:45,600 So BLT isn't just pushing the boundaries of what AI can do. 542 00:16:45,600 --> 00:16:46,440 No, it's not. 543 00:16:46,440 --> 00:16:50,160 It's also potentially making it more accessible to everyone. 544 00:16:50,160 --> 00:16:50,840 Exactly. 545 00:16:50,840 --> 00:16:53,600 It's about making AI more inclusive. 546 00:16:53,600 --> 00:16:54,520 That's pretty inspiring. 547 00:16:54,520 --> 00:16:54,920 It is. 548 00:16:54,920 --> 00:16:57,520 It's a really exciting time to be working in AI. 549 00:16:57,520 --> 00:17:00,360 This whole deep dive into BLT has been incredible. 550 00:17:00,360 --> 00:17:01,320 I'm glad you enjoyed it. 551 00:17:01,320 --> 00:17:02,000 It really has. 552 00:17:02,000 --> 00:17:04,040 It's made me realize that we're living 553 00:17:04,040 --> 00:17:07,960 in this time of unprecedented advancement in AI. 554 00:17:07,960 --> 00:17:08,120 Yeah. 555 00:17:08,120 --> 00:17:09,880 Things are moving so fast. 556 00:17:09,880 --> 00:17:12,720 And the potential for positive change is huge. 557 00:17:12,720 --> 00:17:13,800 It's immense. 558 00:17:13,800 --> 00:17:14,560 Yeah. 559 00:17:14,560 --> 00:17:17,920 But as you said, it's also a time for careful consideration 560 00:17:17,920 --> 00:17:19,080 and responsible development. 561 00:17:19,080 --> 00:17:19,720 Absolutely. 562 00:17:19,720 --> 00:17:21,800 We need to make sure that this powerful technology is 563 00:17:21,800 --> 00:17:22,880 used for good. 564 00:17:22,880 --> 00:17:24,360 We don't want to go down the wrong path. 565 00:17:24,360 --> 00:17:24,920 Right. 566 00:17:24,920 --> 00:17:26,920 And that it benefits all of humanity. 567 00:17:26,920 --> 00:17:30,560 It's about using AI to make the world a better place. 568 00:17:30,560 --> 00:17:32,640 That's a perfect note to end on. 569 00:17:32,640 --> 00:17:35,320 Thank you so much for taking us on this deep dive 570 00:17:35,320 --> 00:17:38,160 into the world of byte-latent transformers. 571 00:17:38,160 --> 00:17:39,160 It's been my pleasure. 572 00:17:39,160 --> 00:17:42,080 It's been an eye-opening and inspiring conversation 573 00:17:42,080 --> 00:17:43,120 and to our listeners. 574 00:17:43,120 --> 00:17:44,920 Yeah, to everyone out there. 575 00:17:44,920 --> 00:17:46,920 Kept those imaginations running wild. 576 00:17:46,920 --> 00:17:48,960 Don't be afraid to dream big. 577 00:17:48,960 --> 00:17:51,360 And stay tuned for our next deep dive 578 00:17:51,360 --> 00:17:54,640 into the ever-evolving world of AI. 579 00:17:54,640 --> 00:17:56,920 We'll be back with more cutting-edge research. 580 00:17:56,920 --> 00:18:01,120 Until then, keep exploring and keep questioning. 581 00:18:01,120 --> 00:18:01,960 Keep learning. 582 00:18:01,960 --> 00:18:04,120 Because that's how we unlock the true potential 583 00:18:04,120 --> 00:18:05,360 of artificial intelligence. 584 00:18:05,360 --> 00:18:07,920 That's how we build a better future. 585 00:18:07,920 --> 00:18:09,640 And that's it for today's deep dive. 586 00:18:09,640 --> 00:18:10,280 See you next time. 587 00:18:10,280 --> 00:18:13,280 We'll see you next time.