1 00:00:00,000 --> 00:00:01,640 All right, strap in everyone. 2 00:00:01,640 --> 00:00:04,480 Today we are going deep on a paper all about, 3 00:00:04,480 --> 00:00:06,360 well, how AI remembers things. 4 00:00:06,360 --> 00:00:08,040 You mean remembers like us? 5 00:00:08,040 --> 00:00:10,960 Well, kind of, but not really. 6 00:00:10,960 --> 00:00:13,120 Specifically, it's about transformers, 7 00:00:13,120 --> 00:00:14,720 those powerful AI models. 8 00:00:14,720 --> 00:00:15,720 Ah. 9 00:00:15,720 --> 00:00:16,560 The paper is called 10 00:00:16,560 --> 00:00:20,080 An Evolved Universal Transformer Memory. 11 00:00:20,080 --> 00:00:21,000 Catchy, right? 12 00:00:21,000 --> 00:00:22,240 Definitely grabs your attention. 13 00:00:22,240 --> 00:00:23,440 And it's a pretty important topic, 14 00:00:23,440 --> 00:00:26,480 how memory impacts an AI's ability to, you know, 15 00:00:26,480 --> 00:00:28,240 learn and solve problems. 16 00:00:28,240 --> 00:00:29,080 It is. 17 00:00:29,080 --> 00:00:31,040 One of the big challenges is figuring out 18 00:00:31,040 --> 00:00:33,480 how to make these transformers smarter, 19 00:00:33,480 --> 00:00:36,000 but without making them super expensive to run, 20 00:00:36,000 --> 00:00:37,080 computationally speaking. 21 00:00:37,080 --> 00:00:38,560 Yeah, that's the balance, right? 22 00:00:38,560 --> 00:00:40,280 Totally, because, you know, 23 00:00:40,280 --> 00:00:42,160 the more a transformer can remember, 24 00:00:42,160 --> 00:00:44,000 like the bigger its context window, 25 00:00:44,000 --> 00:00:45,840 the better it usually performs. 26 00:00:45,840 --> 00:00:48,720 But all that memory requires a ton of computational power. 27 00:00:48,720 --> 00:00:50,280 Exactly, it gets costly. 28 00:00:50,280 --> 00:00:53,440 So this paper, they're proposing a pretty different approach. 29 00:00:53,440 --> 00:00:55,960 Instead of just telling the AI what to remember, 30 00:00:55,960 --> 00:00:57,840 you know, manually programming it. 31 00:00:57,840 --> 00:00:59,000 And what do they do? 32 00:00:59,000 --> 00:01:02,240 Well, they call it neural attention memory models. 33 00:01:02,240 --> 00:01:03,440 Yeah, NAMMs. 34 00:01:03,440 --> 00:01:05,760 It's basically using an evolutionary algorithm 35 00:01:05,760 --> 00:01:08,760 to teach AI how to manage its own memory. 36 00:01:08,760 --> 00:01:12,640 So it's like survival of the fittest, but for AI memories. 37 00:01:12,640 --> 00:01:16,080 The most important info gets to stay, the rest gets tossed. 38 00:01:16,080 --> 00:01:17,440 Uh-huh, you got it. 39 00:01:17,440 --> 00:01:20,920 So these NAMMs analyze what the AI is focusing on, 40 00:01:20,920 --> 00:01:23,840 what it's paying attention to, and then, get this, 41 00:01:23,840 --> 00:01:27,920 they use a technique from audio processing, a spectrogram. 42 00:01:27,920 --> 00:01:30,000 Wait a second, spectrograms? 43 00:01:30,000 --> 00:01:32,000 Like those visualizations of sound waves, 44 00:01:32,000 --> 00:01:34,560 what do those have to do with AI and memory? 45 00:01:34,560 --> 00:01:37,000 I know, it sounds strange, but it's really clever. 46 00:01:37,000 --> 00:01:40,440 By taking the AI's focus and turning that into a spectrogram, 47 00:01:40,440 --> 00:01:42,960 they can train the NAMMs to prune away 48 00:01:42,960 --> 00:01:46,200 all the unimportant stuff, keeping only the key information. 49 00:01:46,200 --> 00:01:48,080 And this works across different tasks too, 50 00:01:48,080 --> 00:01:49,640 not just in one specific area. 51 00:01:49,640 --> 00:01:52,440 So this model, it learns how to manage memory in a way 52 00:01:52,440 --> 00:01:54,040 that works for lots of different things. 53 00:01:54,040 --> 00:01:55,600 Universal, like the title says. 54 00:01:55,600 --> 00:01:57,480 Right, that's what makes it so interesting. 55 00:01:57,480 --> 00:02:00,640 They focused mostly on transformers in the paper, 56 00:02:00,640 --> 00:02:02,040 but they think it could be adapted 57 00:02:02,040 --> 00:02:04,160 to other kinds of AI models too. 58 00:02:04,160 --> 00:02:07,480 Wow, that is like next level. 59 00:02:07,480 --> 00:02:09,520 So it's not just about making transformers better, 60 00:02:09,520 --> 00:02:12,240 it's about giving any AI the ability 61 00:02:12,240 --> 00:02:15,280 to manage its own memory, regardless of the specific task. 62 00:02:15,280 --> 00:02:17,880 Yeah, exactly, it's about equipping AI 63 00:02:17,880 --> 00:02:20,120 with the tools to handle memory independently. 64 00:02:20,120 --> 00:02:22,120 Okay, this is blowing my mind a little. 65 00:02:22,120 --> 00:02:24,120 But let's get down to brass tacks. 66 00:02:24,120 --> 00:02:27,040 Did they actually see any real improvements, 67 00:02:27,040 --> 00:02:29,080 like in how well the AI performed 68 00:02:29,080 --> 00:02:30,880 and how efficiently it used resources? 69 00:02:30,880 --> 00:02:33,560 Oh yeah, they did, they tested it on this benchmark 70 00:02:33,560 --> 00:02:37,080 called Infinite Bench, which uses these super long contexts, 71 00:02:37,080 --> 00:02:38,600 like way more information than usual. 72 00:02:38,600 --> 00:02:39,440 Where'd I have it? 73 00:02:39,440 --> 00:02:41,480 Well, they saw a 10 fold improvement, 74 00:02:41,480 --> 00:02:43,520 like 10 times better performance. 75 00:02:43,520 --> 00:02:44,360 No way. 76 00:02:44,360 --> 00:02:48,320 And get this, when they applied NAMMs to video analysis, 77 00:02:48,320 --> 00:02:50,360 the model figured out how to prioritize 78 00:02:50,360 --> 00:02:52,320 the important language instructions 79 00:02:52,320 --> 00:02:55,080 and just ditch all the redundant video frames. 80 00:02:55,080 --> 00:02:57,520 So it learned to pay attention to the right things 81 00:02:57,520 --> 00:02:58,720 and ignore the noise. 82 00:02:59,600 --> 00:03:01,040 That's seriously impressive. 83 00:03:02,720 --> 00:03:04,400 It sounds like these NAMs are picking up 84 00:03:04,400 --> 00:03:07,080 some pretty clever tricks when it comes to memory. 85 00:03:07,080 --> 00:03:10,040 But did they compare this approach to other methods? 86 00:03:10,040 --> 00:03:12,560 Like how do we know this is better than what we already had? 87 00:03:12,560 --> 00:03:13,440 Good question. 88 00:03:13,440 --> 00:03:15,280 They compared it to three other methods 89 00:03:15,280 --> 00:03:16,640 that people are already using, 90 00:03:16,640 --> 00:03:19,320 and all of those relied on manually designed rules 91 00:03:19,320 --> 00:03:20,720 to manage memory. 92 00:03:20,720 --> 00:03:22,400 Okay, so someone had to program in 93 00:03:22,400 --> 00:03:24,400 exactly what the AI should remember. 94 00:03:24,400 --> 00:03:25,240 Yep, that's right. 95 00:03:25,240 --> 00:03:26,920 And the results were pretty clear. 96 00:03:26,920 --> 00:03:28,880 NAMMs did way better, 97 00:03:28,880 --> 00:03:30,680 both in terms of getting the right answers 98 00:03:30,680 --> 00:03:33,120 and using resources more efficiently. 99 00:03:33,120 --> 00:03:34,920 So this whole idea of letting the AI 100 00:03:34,920 --> 00:03:36,680 learn its own memory management, 101 00:03:36,680 --> 00:03:39,600 instead of forcing it to follow some pre-programmed rules, 102 00:03:39,600 --> 00:03:40,640 it actually works better. 103 00:03:40,640 --> 00:03:41,880 That's a major takeaway. 104 00:03:41,880 --> 00:03:44,880 They're essentially allowing the AI to adapt, 105 00:03:44,880 --> 00:03:46,440 to adjust its memory strategy, 106 00:03:46,440 --> 00:03:47,880 depending on what it's trying to do. 107 00:03:47,880 --> 00:03:50,000 And that leads to better results overall. 108 00:03:50,000 --> 00:03:51,960 Okay, now I'm really curious. 109 00:03:51,960 --> 00:03:53,800 How does all this actually work? 110 00:03:53,800 --> 00:03:56,360 What are the nuts and bolts of these NAMMs? 111 00:03:56,360 --> 00:03:58,680 Like, walk me through it step by step. 112 00:03:58,680 --> 00:04:01,200 Sure, so first we need to understand this thing 113 00:04:01,200 --> 00:04:02,960 called the attention matrix. 114 00:04:02,960 --> 00:04:03,960 Attention matrix. 115 00:04:03,960 --> 00:04:05,240 Sounds kind of sci-fi. 116 00:04:05,240 --> 00:04:06,440 A little bit, yeah. 117 00:04:06,440 --> 00:04:09,040 But it's really just a way of visualizing 118 00:04:09,040 --> 00:04:11,120 what the AI is focusing on. 119 00:04:11,120 --> 00:04:12,680 You know how when we read a sentence, 120 00:04:12,680 --> 00:04:15,640 we pay attention to certain words more than others? 121 00:04:15,640 --> 00:04:16,640 Yeah, the important ones. 122 00:04:16,640 --> 00:04:18,680 Right, and AI models do the same thing. 123 00:04:18,680 --> 00:04:21,840 They learn to focus on the most relevant information. 124 00:04:21,840 --> 00:04:24,600 The attention matrix is like a map of that focus. 125 00:04:24,600 --> 00:04:26,480 So it's like looking into the AI's brain, 126 00:04:26,480 --> 00:04:28,360 seeing what it thinks is important. 127 00:04:28,360 --> 00:04:30,920 Exactly, and that's where the NAMs come in. 128 00:04:30,920 --> 00:04:32,560 They look at this attention matrix, 129 00:04:32,560 --> 00:04:35,160 and then they use a technique called token pruning. 130 00:04:35,160 --> 00:04:36,800 Pruning, like trimming a tree. 131 00:04:36,800 --> 00:04:37,720 Kind of, yeah. 132 00:04:37,720 --> 00:04:40,640 It's about getting rid of the less important information, 133 00:04:40,640 --> 00:04:43,000 keeping only the stuff that really matters. 134 00:04:43,000 --> 00:04:45,640 And that makes the AI much more efficient and effective. 135 00:04:45,640 --> 00:04:47,840 Okay, so we've got an evolutionary algorithm 136 00:04:47,840 --> 00:04:49,720 training these NAMs. 137 00:04:49,720 --> 00:04:53,600 Then they analyze what the AI is focused on, 138 00:04:53,600 --> 00:04:56,160 and they prune away the unimportant stuff. 139 00:04:56,160 --> 00:04:57,280 Makes sense? 140 00:04:57,280 --> 00:04:59,600 But I'm still stuck on the spectrogram thing. 141 00:04:59,600 --> 00:05:03,720 How does analyzing sound waves help with AI memory? 142 00:05:03,720 --> 00:05:05,280 I just don't get that connection. 143 00:05:05,280 --> 00:05:07,280 That's the really cool part of this paper. 144 00:05:07,280 --> 00:05:10,560 They figured out that by representing the AI's attention 145 00:05:10,560 --> 00:05:12,000 as a spectrogram, you know, 146 00:05:12,000 --> 00:05:14,600 that technique used for sound analysis, 147 00:05:14,600 --> 00:05:17,320 they could make it much easier for the NAMs 148 00:05:17,320 --> 00:05:19,440 to learn how to manage memory. 149 00:05:19,440 --> 00:05:23,080 Hold up, they're using sound analysis to study AI memory. 150 00:05:23,080 --> 00:05:23,920 That's wild. 151 00:05:23,920 --> 00:05:24,920 I know, right? 152 00:05:24,920 --> 00:05:26,520 But it seems to work really well. 153 00:05:26,520 --> 00:05:28,160 They found that by using these spectrograms, 154 00:05:28,160 --> 00:05:31,160 the NAMs got way better at figuring out 155 00:05:31,160 --> 00:05:34,120 which information to keep and which to toss out. 156 00:05:34,120 --> 00:05:35,440 This is blowing my mind. 157 00:05:35,440 --> 00:05:36,560 Okay, before we go any further, 158 00:05:36,560 --> 00:05:37,960 let's just do a quick recap. 159 00:05:37,960 --> 00:05:39,960 We started by talking about this problem 160 00:05:39,960 --> 00:05:43,240 of managing memory in transformer models. 161 00:05:43,240 --> 00:05:44,760 Like, bigger isn't always better 162 00:05:44,760 --> 00:05:46,160 when it comes to context windows. 163 00:05:46,160 --> 00:05:47,000 Right. 164 00:05:47,000 --> 00:05:48,800 Then we dove into these NAMs, 165 00:05:48,800 --> 00:05:50,120 these evolved memory models 166 00:05:50,120 --> 00:05:52,560 that learned to pick out the important information 167 00:05:52,560 --> 00:05:54,520 by looking at what the AI focuses on 168 00:05:54,520 --> 00:05:56,960 and strategically forgetting the less important stuff. 169 00:05:56,960 --> 00:05:58,680 And they tested it against those other methods, 170 00:05:58,680 --> 00:06:00,360 the ones with the pre-programmed rules, 171 00:06:00,360 --> 00:06:02,520 and it totally blew them out of the water, 172 00:06:02,520 --> 00:06:05,000 both in terms of performance and efficiency. 173 00:06:05,000 --> 00:06:06,280 You got it. 174 00:06:06,280 --> 00:06:08,040 But what really got me excited 175 00:06:08,040 --> 00:06:10,120 was this whole thing with spectrograms. 176 00:06:10,120 --> 00:06:13,240 Like, they borrowed this tool from sound analysis 177 00:06:13,240 --> 00:06:15,840 to make memory management even more efficient. 178 00:06:15,840 --> 00:06:18,920 Yeah, that's some seriously creative problem solving. 179 00:06:18,920 --> 00:06:21,320 But we've only just scratched the surface of this paper. 180 00:06:21,320 --> 00:06:22,640 I know, right? 181 00:06:22,640 --> 00:06:24,720 There's still so much more to explore. 182 00:06:24,720 --> 00:06:27,680 Like, what about this whole zero-shot transfer thing? 183 00:06:27,680 --> 00:06:30,680 And what are the real-world implications of all this? 184 00:06:30,680 --> 00:06:32,280 We'll have to dive into that next time. 185 00:06:32,280 --> 00:06:33,280 Welcome back. 186 00:06:33,280 --> 00:06:34,760 We're in the middle of exploring 187 00:06:34,760 --> 00:06:38,080 these evolved memory models, NAMs. 188 00:06:38,080 --> 00:06:40,400 It's incredible how they analyze attention 189 00:06:40,400 --> 00:06:42,480 and prune information. 190 00:06:42,480 --> 00:06:43,480 But I wanna know, 191 00:06:43,480 --> 00:06:45,760 where could we actually use this technology? 192 00:06:45,760 --> 00:06:47,400 What kind of impact could it have? 193 00:06:47,400 --> 00:06:48,560 Yeah, exactly. 194 00:06:48,560 --> 00:06:50,640 It's cool to see all these impressive benchmark results, 195 00:06:50,640 --> 00:06:52,360 but let's get real. 196 00:06:52,360 --> 00:06:54,240 Are we talking actual real-world impact 197 00:06:54,240 --> 00:06:56,200 or just theoretical possibilities? 198 00:06:56,200 --> 00:06:57,760 I think this research could really change 199 00:06:57,760 --> 00:06:59,720 a lot of different AI applications. 200 00:06:59,720 --> 00:07:01,560 One area that immediately comes to mind 201 00:07:01,560 --> 00:07:03,800 is natural language processing. 202 00:07:03,800 --> 00:07:05,760 Okay, so like chatbots and stuff. 203 00:07:05,760 --> 00:07:06,760 Exactly. 204 00:07:06,760 --> 00:07:09,200 Imagine chatbots that can actually remember 205 00:07:09,200 --> 00:07:10,800 your past conversations, 206 00:07:10,800 --> 00:07:12,320 or language translation tools 207 00:07:12,320 --> 00:07:14,680 that are way more accurate, more nuanced. 208 00:07:14,680 --> 00:07:18,040 So Siri, but a million times better. 209 00:07:18,040 --> 00:07:18,880 Yeah. 210 00:07:18,880 --> 00:07:20,440 Able to understand all my weird requests. 211 00:07:20,440 --> 00:07:21,960 Ah, something like that. 212 00:07:21,960 --> 00:07:23,720 With this better memory management, 213 00:07:23,720 --> 00:07:26,560 these AI systems could be so much more personalized, 214 00:07:26,560 --> 00:07:27,720 so much more helpful. 215 00:07:27,720 --> 00:07:29,320 Okay, I see where you're going with this. 216 00:07:29,320 --> 00:07:31,200 Like customer service bots 217 00:07:31,200 --> 00:07:32,800 that actually remember who you are, 218 00:07:32,800 --> 00:07:34,320 what you talked about before. 219 00:07:34,320 --> 00:07:36,360 Or imagine educational tools 220 00:07:36,360 --> 00:07:38,080 that adapt to how you learn BAS. 221 00:07:38,080 --> 00:07:38,920 Exactly. 222 00:07:38,920 --> 00:07:40,280 Those are some great examples. 223 00:07:40,280 --> 00:07:42,640 What about stuff beyond just language? 224 00:07:42,640 --> 00:07:46,120 Could these NAMMs be used for other types of AI too? 225 00:07:46,120 --> 00:07:47,080 Oh, absolutely. 226 00:07:47,080 --> 00:07:48,120 The researchers talked about 227 00:07:48,120 --> 00:07:50,120 a lot of potential applications in fields 228 00:07:50,120 --> 00:07:52,880 like robotics and computer vision. 229 00:07:52,880 --> 00:07:54,320 All in, robots with memories. 230 00:07:54,320 --> 00:07:55,240 Well, think about it. 231 00:07:55,240 --> 00:07:58,040 Robots that can learn new tasks way faster 232 00:07:58,040 --> 00:08:00,760 because they remember their past experiences. 233 00:08:00,760 --> 00:08:02,320 Or computer vision systems 234 00:08:02,320 --> 00:08:04,400 that can analyze really complex scenes, 235 00:08:04,400 --> 00:08:06,000 but much more efficiently. 236 00:08:06,000 --> 00:08:08,880 So robots that learn from their mistakes 237 00:08:08,880 --> 00:08:11,160 and adapt to new environments. 238 00:08:11,160 --> 00:08:13,600 That sounds like, I don't know, something out of a movie. 239 00:08:13,600 --> 00:08:15,760 It might be closer to reality than you think. 240 00:08:15,760 --> 00:08:17,360 And this research is definitely 241 00:08:17,360 --> 00:08:19,120 a step in the right direction. 242 00:08:19,120 --> 00:08:22,360 By giving AI the ability to handle its own memory, 243 00:08:22,360 --> 00:08:24,680 we open up so many possibilities 244 00:08:24,680 --> 00:08:27,160 for how these systems learn and interact 245 00:08:27,160 --> 00:08:28,520 with the world around them. 246 00:08:28,520 --> 00:08:30,040 This is all incredibly exciting, 247 00:08:30,040 --> 00:08:31,920 but I have to ask, 248 00:08:31,920 --> 00:08:35,320 were there any limitations to this research? 249 00:08:35,320 --> 00:08:37,600 Or areas where they could improve? 250 00:08:37,600 --> 00:08:39,160 No research is perfect, right? 251 00:08:39,160 --> 00:08:40,000 You're right. 252 00:08:40,000 --> 00:08:41,520 I did point out some limitations. 253 00:08:41,520 --> 00:08:42,920 One of the big ones is that, you know, 254 00:08:42,920 --> 00:08:44,920 they mostly focused on language tasks 255 00:08:44,920 --> 00:08:46,560 for training the NAMMs. 256 00:08:46,560 --> 00:08:48,080 So we need more research to see 257 00:08:48,080 --> 00:08:49,760 how well they perform in other areas. 258 00:08:49,760 --> 00:08:50,600 Yeah, exactly. 259 00:08:50,600 --> 00:08:52,600 It's still early days for this technology. 260 00:08:52,600 --> 00:08:54,120 Another thing they mentioned was the cost 261 00:08:54,120 --> 00:08:55,560 of actually training these NAMs. 262 00:08:55,560 --> 00:08:57,760 It can still be quite resource intensive, you know, 263 00:08:57,760 --> 00:08:59,880 even though they make running the AI models 264 00:08:59,880 --> 00:09:01,480 themselves much cheaper. 265 00:09:01,480 --> 00:09:02,760 Well, that makes sense. 266 00:09:02,760 --> 00:09:04,080 Training any AI model 267 00:09:04,080 --> 00:09:06,000 takes a lot of computing power, right? 268 00:09:06,000 --> 00:09:07,800 Especially when you're dealing with something as complex 269 00:09:07,800 --> 00:09:09,880 as evolutionary algorithms. 270 00:09:09,880 --> 00:09:12,240 But it sounds like the benefits down the line, 271 00:09:12,240 --> 00:09:15,600 in terms of efficiency, could outweigh the initial costs. 272 00:09:15,600 --> 00:09:16,440 That's the hope. 273 00:09:16,440 --> 00:09:17,640 And the researchers believe that 274 00:09:17,640 --> 00:09:20,080 as this technology gets more refined, 275 00:09:20,080 --> 00:09:23,280 the training process will become more efficient too. 276 00:09:23,280 --> 00:09:24,480 So we've got this new way 277 00:09:24,480 --> 00:09:26,440 of approaching AI memory management 278 00:09:26,440 --> 00:09:29,080 with potential uses in all sorts of fields. 279 00:09:29,080 --> 00:09:31,440 But I'm also thinking about the bigger picture. 280 00:09:31,440 --> 00:09:32,960 What does this research mean 281 00:09:32,960 --> 00:09:35,680 for the future of AI in general? 282 00:09:35,680 --> 00:09:38,720 Well, I think this work is really pushing the boundaries 283 00:09:38,720 --> 00:09:40,280 of what AI can do. 284 00:09:40,280 --> 00:09:41,920 One of the most exciting aspects 285 00:09:41,920 --> 00:09:44,440 is this whole concept of zero shot transfer. 286 00:09:44,440 --> 00:09:45,520 Zero shot transfer. 287 00:09:45,520 --> 00:09:47,040 Fresh my memory, what was that again? 288 00:09:47,040 --> 00:09:50,240 It means you can train an AI model on one task 289 00:09:50,240 --> 00:09:52,240 and then use it for a totally different task 290 00:09:52,240 --> 00:09:53,800 without any additional training. 291 00:09:53,800 --> 00:09:55,240 Whoa. 292 00:09:55,240 --> 00:09:57,400 So you're saying like you could train these NAMs 293 00:09:57,400 --> 00:10:00,960 on language data and then use them to improve 294 00:10:00,960 --> 00:10:03,160 memory management in a robot, for example. 295 00:10:03,160 --> 00:10:04,000 Exactly. 296 00:10:04,000 --> 00:10:06,920 And that kind of flexibility, that adaptability, 297 00:10:06,920 --> 00:10:11,400 is a huge step towards creating more versatile AI systems. 298 00:10:11,400 --> 00:10:12,960 It's like moving away from AI 299 00:10:12,960 --> 00:10:15,000 that's good at one specific thing 300 00:10:15,000 --> 00:10:17,080 and toward AI that can learn and adapt 301 00:10:17,080 --> 00:10:18,560 to all sorts of situations. 302 00:10:18,560 --> 00:10:19,600 You got it. 303 00:10:19,600 --> 00:10:22,000 And that shift could completely revolutionize 304 00:10:22,000 --> 00:10:25,000 how we develop and interact with AI in the future. 305 00:10:25,000 --> 00:10:26,640 This is mind blowing stuff. 306 00:10:26,640 --> 00:10:29,120 But I wanna circle back to the specifics of this paper. 307 00:10:29,120 --> 00:10:31,400 They talked about like digging into the why 308 00:10:31,400 --> 00:10:33,800 behind how these NAMs actually work. 309 00:10:33,800 --> 00:10:35,080 Did they uncover anything interesting 310 00:10:35,080 --> 00:10:37,400 about the AI's thought process? 311 00:10:37,400 --> 00:10:38,360 They did. 312 00:10:38,360 --> 00:10:40,840 One of the things they looked at was how memory 313 00:10:40,840 --> 00:10:44,280 is distributed across different layers of the AI network. 314 00:10:44,280 --> 00:10:46,040 Like different parts of the AI's brain. 315 00:10:46,040 --> 00:10:46,880 Exactly. 316 00:10:46,880 --> 00:10:49,200 And they found that some layers tend to hold on 317 00:10:49,200 --> 00:10:51,600 to information from further back in time. 318 00:10:51,600 --> 00:10:55,000 So certain parts of the AI are better at long term memory. 319 00:10:55,000 --> 00:10:56,120 That's what it seems like. 320 00:10:56,120 --> 00:10:58,640 And it suggests that these layers are crucial 321 00:10:58,640 --> 00:11:00,680 for understanding complex relationships, 322 00:11:00,680 --> 00:11:02,400 the ones that span across a lot of data. 323 00:11:02,400 --> 00:11:05,280 It's like the AI is developing its own sense of history 324 00:11:05,280 --> 00:11:07,120 which helps it make better decisions. 325 00:11:07,120 --> 00:11:09,040 Whoa, that's deep. 326 00:11:09,040 --> 00:11:11,520 It's like we're watching this new form of intelligence 327 00:11:11,520 --> 00:11:15,600 emerge shaped by its own unique experiences and memories. 328 00:11:15,600 --> 00:11:17,240 Did they find anything else interesting 329 00:11:17,240 --> 00:11:19,840 about how these NAMs manage memory? 330 00:11:19,840 --> 00:11:22,880 They also noticed that the way NAMs get rid of information, 331 00:11:22,880 --> 00:11:25,200 the way they prune, it actually varies depending 332 00:11:25,200 --> 00:11:26,720 on what tasks they're doing. 333 00:11:26,720 --> 00:11:27,560 Makes sense. 334 00:11:27,560 --> 00:11:29,000 So it adapts its strategy. 335 00:11:29,000 --> 00:11:29,840 Yeah. 336 00:11:29,840 --> 00:11:31,640 For example, with CUD completion tasks, 337 00:11:31,640 --> 00:11:34,960 they found that NAMs removed more of the redundant bits 338 00:11:34,960 --> 00:11:36,960 compared to, say, natural language tasks. 339 00:11:36,960 --> 00:11:39,560 Well, code has to be super precise and efficient. 340 00:11:39,560 --> 00:11:41,040 So it makes sense that the memory management 341 00:11:41,040 --> 00:11:41,880 would be different. 342 00:11:41,880 --> 00:11:44,440 It's like the NAMs are learning to think like programmers. 343 00:11:44,440 --> 00:11:45,920 Ha ha, exactly. 344 00:11:45,920 --> 00:11:49,000 And that's a big part of what makes this research so cool. 345 00:11:49,000 --> 00:11:51,160 These NAMs aren't just remembering everything. 346 00:11:51,160 --> 00:11:53,680 They're learning what's important, what to prioritize, 347 00:11:53,680 --> 00:11:57,200 and they adjust their strategy based on the specific tasks. 348 00:11:57,200 --> 00:11:58,440 Just like we do, right? 349 00:11:58,440 --> 00:12:01,920 As humans, we don't remember every detail of every experience. 350 00:12:01,920 --> 00:12:03,520 We focus on what matters most. 351 00:12:03,520 --> 00:12:04,040 Right. 352 00:12:04,040 --> 00:12:06,160 And that's crucial for intelligent behavior, 353 00:12:06,160 --> 00:12:09,000 whether you're talking about humans or AI. 354 00:12:09,000 --> 00:12:10,800 This is all painting a really interesting picture, 355 00:12:10,800 --> 00:12:15,320 like how these NAMs work and what they could be capable of. 356 00:12:15,320 --> 00:12:17,040 But what about the data they used? 357 00:12:17,040 --> 00:12:19,400 Did they mention anything specific about the data itself? 358 00:12:19,400 --> 00:12:20,400 They did. 359 00:12:20,400 --> 00:12:21,720 One of the things they highlighted 360 00:12:21,720 --> 00:12:26,000 was this new benchmark data set they created called Chobun. 361 00:12:26,000 --> 00:12:26,960 Chobun. 362 00:12:26,960 --> 00:12:27,600 That's a cool name. 363 00:12:27,600 --> 00:12:29,240 What's so special about this data set? 364 00:12:29,240 --> 00:12:31,240 Well, it's specifically designed to test 365 00:12:31,240 --> 00:12:35,680 how well AI can understand long stretches of text in Japanese. 366 00:12:35,680 --> 00:12:36,760 Ah, I see. 367 00:12:36,760 --> 00:12:39,760 So it's about expanding the scope of language understanding 368 00:12:39,760 --> 00:12:40,480 for AI. 369 00:12:40,480 --> 00:12:41,480 Exactly. 370 00:12:41,480 --> 00:12:44,160 Most of the long context language benchmarks out there 371 00:12:44,160 --> 00:12:46,440 focus on English or Chinese. 372 00:12:46,440 --> 00:12:49,960 So Chobun helps to make sure AI development is inclusive, 373 00:12:49,960 --> 00:12:52,320 you know, that benefits speakers of different languages. 374 00:12:52,320 --> 00:12:53,480 That's fantastic. 375 00:12:53,480 --> 00:12:56,000 Did they talk about how the NAMs performed 376 00:12:56,000 --> 00:12:57,240 on this new data set? 377 00:12:57,240 --> 00:12:59,360 They did, and the results were great. 378 00:12:59,360 --> 00:13:02,000 The NAMs showed significant improvements 379 00:13:02,000 --> 00:13:04,360 compared to other memory management techniques. 380 00:13:04,360 --> 00:13:06,760 It shows how well they can adapt to a new language. 381 00:13:06,760 --> 00:13:08,240 This is amazing. 382 00:13:08,240 --> 00:13:10,640 It's incredible to see this research pushing the boundaries 383 00:13:10,640 --> 00:13:14,320 of AI in so many different ways from creating these new memory 384 00:13:14,320 --> 00:13:18,280 management techniques to expanding language understanding 385 00:13:18,280 --> 00:13:20,640 to making sure different languages are included. 386 00:13:20,640 --> 00:13:22,960 It feels like a real turning point for AI. 387 00:13:22,960 --> 00:13:23,840 I agree. 388 00:13:23,840 --> 00:13:26,960 This research has the potential to change how we build and use 389 00:13:26,960 --> 00:13:28,200 AI systems in the future. 390 00:13:28,200 --> 00:13:29,960 This has been an awesome journey so far, 391 00:13:29,960 --> 00:13:31,280 learning about AI memory. 392 00:13:31,280 --> 00:13:33,320 We've covered a lot of ground, but there's still 393 00:13:33,320 --> 00:13:34,440 so much more to this paper. 394 00:13:34,440 --> 00:13:35,360 I know, right? 395 00:13:35,360 --> 00:13:37,080 There's still so much to uncover. 396 00:13:37,080 --> 00:13:40,200 Stay tuned, because we're going to wrap up our deep dive 397 00:13:40,200 --> 00:13:43,720 into an evolved universal transformer memory 398 00:13:43,720 --> 00:13:46,440 with some final thoughts and key takeaways. 399 00:13:46,440 --> 00:13:48,840 And we're back for the final part of our deep dive. 400 00:13:48,840 --> 00:13:51,280 We've been talking all about this amazing paper, 401 00:13:51,280 --> 00:13:54,080 an evolved universal transformer memory, 402 00:13:54,080 --> 00:13:56,360 specifically those evolved memory models. 403 00:13:56,360 --> 00:13:57,840 The NAMs. 404 00:13:57,840 --> 00:13:59,400 What a wild ride this has been. 405 00:13:59,400 --> 00:14:00,480 It really has. 406 00:14:00,480 --> 00:14:03,400 From all those technical details to those mind blowing 407 00:14:03,400 --> 00:14:06,760 potential uses to, I don't know, even touching on some pretty 408 00:14:06,760 --> 00:14:09,280 deep philosophical stuff about AI and memory. 409 00:14:09,280 --> 00:14:10,200 Right. 410 00:14:10,200 --> 00:14:12,800 I think we can both agree this research is impressive. 411 00:14:12,800 --> 00:14:15,880 But before we wrap up, I want to zoom out a little. 412 00:14:15,880 --> 00:14:19,400 What does this all mean for the future of AI? 413 00:14:19,400 --> 00:14:20,280 The big picture. 414 00:14:20,280 --> 00:14:23,640 Well, to me, this paper shows a pretty big shift in how we 415 00:14:23,640 --> 00:14:25,440 think about and develop AI. 416 00:14:25,440 --> 00:14:28,480 We're moving beyond just throwing tons of data at an AI 417 00:14:28,480 --> 00:14:29,720 and hoping for the best. 418 00:14:29,720 --> 00:14:31,560 Yeah, it's not just about brute force anymore. 419 00:14:31,560 --> 00:14:32,120 Right. 420 00:14:32,120 --> 00:14:35,600 Now we're talking about AI systems that can learn how to learn, 421 00:14:35,600 --> 00:14:37,080 how to figure out what's important, 422 00:14:37,080 --> 00:14:38,640 how to adapt to new stuff. 423 00:14:38,640 --> 00:14:42,440 It's almost like we're giving AI the keys to its own mind, 424 00:14:42,440 --> 00:14:45,280 empowering it to be more in control of its own development. 425 00:14:45,280 --> 00:14:46,440 Exactly. 426 00:14:46,440 --> 00:14:48,680 And that has some pretty huge implications 427 00:14:48,680 --> 00:14:51,360 for the future of intelligence itself, 428 00:14:51,360 --> 00:14:53,640 both for AI and for humans. 429 00:14:53,640 --> 00:14:56,720 As these AI systems get better at managing their own memories, 430 00:14:56,720 --> 00:14:58,760 they'll be able to work with us in deeper ways, 431 00:14:58,760 --> 00:15:00,480 solve even more complex problems, 432 00:15:00,480 --> 00:15:03,160 maybe even help us understand ourselves better. 433 00:15:03,160 --> 00:15:06,240 It's like we're on the verge of this whole new partnership, 434 00:15:06,240 --> 00:15:09,000 humans and AI working together to push 435 00:15:09,000 --> 00:15:11,800 the boundaries of what's possible in terms 436 00:15:11,800 --> 00:15:13,200 of knowledge and creativity. 437 00:15:13,200 --> 00:15:14,200 But let's be real first. 438 00:15:14,200 --> 00:15:16,360 I got this research is still pretty new. 439 00:15:16,360 --> 00:15:17,560 What are the next steps? 440 00:15:17,560 --> 00:15:19,160 Where does this all go from here? 441 00:15:19,160 --> 00:15:20,840 Oh, there's so many possibilities. 442 00:15:20,840 --> 00:15:24,480 One thing is to see if these NAMMs work with other types 443 00:15:24,480 --> 00:15:26,640 of AI models, not just transformers. 444 00:15:26,640 --> 00:15:30,040 So expanding their memory magic to other AI species, 445 00:15:30,040 --> 00:15:30,560 so to speak. 446 00:15:30,560 --> 00:15:31,920 Exactly. 447 00:15:31,920 --> 00:15:33,680 We need to know if this approach works 448 00:15:33,680 --> 00:15:36,880 for other AI architectures, other kinds of tasks. 449 00:15:36,880 --> 00:15:38,720 Another exciting direction is to develop 450 00:15:38,720 --> 00:15:41,240 even more advanced evolutionary algorithms, 451 00:15:41,240 --> 00:15:44,400 ones that can train these NAMMs even more effectively. 452 00:15:44,400 --> 00:15:47,840 Survival of the fittest, but for AI memories on steroids. 453 00:15:47,840 --> 00:15:49,640 Uh-huh, that's a good way to put it. 454 00:15:49,640 --> 00:15:51,080 And as we push these boundaries, 455 00:15:51,080 --> 00:15:53,720 we got to be mindful of the impact this has on society. 456 00:15:53,720 --> 00:15:56,320 Make sure these advancements benefit everyone. 457 00:15:56,320 --> 00:15:59,000 AI should be a tool for good, helping us 458 00:15:59,000 --> 00:16:01,720 build a better future, not just for a few, 459 00:16:01,720 --> 00:16:03,280 but for all of humanity. 460 00:16:03,280 --> 00:16:04,080 Definitely. 461 00:16:04,080 --> 00:16:06,760 This research shows us that the future of AI 462 00:16:06,760 --> 00:16:08,800 isn't set in stone. 463 00:16:08,800 --> 00:16:11,160 It's up to us to guide it in a direction 464 00:16:11,160 --> 00:16:14,760 that aligns with our values, what we want to see in the world. 465 00:16:14,760 --> 00:16:17,280 So to bring it all together, this paper, 466 00:16:17,280 --> 00:16:20,160 an evolved universal transformer memory, 467 00:16:20,160 --> 00:16:25,160 gives us a glimpse into a future where AI isn't just powerful, 468 00:16:25,240 --> 00:16:27,960 but also adaptable, efficient, 469 00:16:27,960 --> 00:16:30,600 maybe even more intelligent than we can imagine right now. 470 00:16:30,600 --> 00:16:32,400 It's a future that's both exciting 471 00:16:32,400 --> 00:16:33,960 and a little bit intimidating. 472 00:16:33,960 --> 00:16:34,800 Definitely. 473 00:16:34,800 --> 00:16:36,520 And we have to approach it with a good balance 474 00:16:36,520 --> 00:16:38,560 of curiosity and caution. 475 00:16:38,560 --> 00:16:39,400 Couldn't agree more. 476 00:16:39,400 --> 00:16:41,760 Well, I think we've covered a lot of ground on this paper. 477 00:16:41,760 --> 00:16:42,680 I think so too. 478 00:16:42,680 --> 00:16:44,680 It's been fantastic exploring this with you. 479 00:16:44,680 --> 00:16:45,760 It's been a pleasure. 480 00:16:45,760 --> 00:16:47,240 And to all our listeners out there, 481 00:16:47,240 --> 00:16:49,000 thanks for joining us on this deep dive 482 00:16:49,000 --> 00:16:50,880 into the world of AI memory. 483 00:16:50,880 --> 00:17:16,280 Until next time, keep those brains buzzing.