1 00:00:00,000 --> 00:00:02,560 Right now, the smartest artificial intelligence 2 00:00:02,560 --> 00:00:04,540 systems on the planet, I mean, the ones passing 3 00:00:04,540 --> 00:00:07,879 the bar exam and the 90th percentile and writing 4 00:00:07,879 --> 00:00:12,640 incredibly complex code in seconds, they're basically 5 00:00:12,640 --> 00:00:16,399 functionally mindless. Yeah, they really are. 6 00:00:17,129 --> 00:00:19,449 I mean, we are pouring trillions of dollars into 7 00:00:19,449 --> 00:00:21,269 building these synthetic brains, but we're doing 8 00:00:21,269 --> 00:00:25,149 it without actually knowing what a mind fundamentally 9 00:00:25,149 --> 00:00:29,250 is. Exactly. It's wild when you think about it. 10 00:00:29,239 --> 00:00:31,920 Right. We're just staring at the output of these 11 00:00:31,920 --> 00:00:34,119 machines completely mesmerized by the parlor 12 00:00:34,119 --> 00:00:37,439 tricks while sort of ignoring the fact that we 13 00:00:37,439 --> 00:00:40,179 don't have a clear blueprint or even like a functional 14 00:00:40,179 --> 00:00:44,200 x -ray of true cognitive architecture. Well, 15 00:00:44,219 --> 00:00:46,579 the field is suffering from a massive case of 16 00:00:46,579 --> 00:00:48,759 diagnostic money waters, to be honest. Oh, for 17 00:00:48,759 --> 00:00:51,679 sure. We're just completely obsessed with the 18 00:00:51,679 --> 00:00:54,659 performance of the system, but we've largely 19 00:00:54,659 --> 00:00:57,140 abandoned the philosophy of how a system truly 20 00:00:57,140 --> 00:01:00,670 thinks. how it adapts and survives. Yeah, it's 21 00:01:00,670 --> 00:01:03,710 all about benchmarks right now. Right. The current 22 00:01:03,710 --> 00:01:06,469 state of AI is entirely focused on scaling up 23 00:01:06,469 --> 00:01:10,170 what we already have. Just more data, more compute 24 00:01:10,170 --> 00:01:12,750 power, rather than stepping back to figure out 25 00:01:12,750 --> 00:01:15,310 the essential physics of cognition. So today's 26 00:01:15,310 --> 00:01:17,969 deep dive is really about fixing that exact blind 27 00:01:17,969 --> 00:01:20,590 spot. We are unpacking a text that feels like 28 00:01:20,590 --> 00:01:23,349 a totally necessary antidote to the endless hype 29 00:01:23,349 --> 00:01:26,290 cycle of generative AI. It really is. The book 30 00:01:26,290 --> 00:01:28,849 is called On Minds, Reflections of a Cognitive 31 00:01:28,849 --> 00:01:31,810 Architect by Paul S. Rosenblum. And for you listening 32 00:01:31,810 --> 00:01:34,269 right now, whether you are, you know, trying 33 00:01:34,269 --> 00:01:36,370 to understand where the future of technology 34 00:01:36,370 --> 00:01:38,969 is actually heading or you're just fascinated 35 00:01:38,969 --> 00:01:41,629 by human psychology, this discussion is designed 36 00:01:41,629 --> 00:01:44,730 to be your shortcut to clarity. Yeah, and Rosenblum's 37 00:01:44,730 --> 00:01:47,290 perspective is incredibly rare because he isn't 38 00:01:47,290 --> 00:01:50,769 just like a casual commentator or someone who 39 00:01:50,769 --> 00:01:53,980 jumped on the AI bandwagon post... chat GPT. 40 00:01:54,060 --> 00:01:56,379 Right. He's been around. Oh, he has spent over 41 00:01:56,379 --> 00:01:58,840 half a century in the trenches of artificial 42 00:01:58,840 --> 00:02:01,700 intelligence and cognitive science. But more 43 00:02:01,700 --> 00:02:05,219 importantly, he approaches this field as an architectural 44 00:02:05,219 --> 00:02:08,300 explorer. I love that term, architectural explorer. 45 00:02:08,560 --> 00:02:10,360 It's a great way to put it. He isn't satisfied 46 00:02:10,360 --> 00:02:12,759 with just writing a clever algorithm that can 47 00:02:12,759 --> 00:02:15,860 play a game. He wants to know what underlying 48 00:02:15,860 --> 00:02:18,919 universal structures are actually required to 49 00:02:18,919 --> 00:02:22,469 build a complete functioning mind. Which shifts 50 00:02:22,469 --> 00:02:24,629 the framing completely. We aren't just looking 51 00:02:24,629 --> 00:02:27,289 at cold lines of code today. I mean, we are looking 52 00:02:27,289 --> 00:02:30,530 at the foundational philosophy of building minds. 53 00:02:30,530 --> 00:02:34,490 Exactly. Taking these incredibly dense academic 54 00:02:34,490 --> 00:02:37,050 meta -scientific reflections and translating 55 00:02:37,050 --> 00:02:40,129 them into those those aha moments that change 56 00:02:40,129 --> 00:02:41,990 how you look at your own thought processes. That's 57 00:02:41,990 --> 00:02:44,349 our mission today. And the timing of Rosenblum's 58 00:02:44,349 --> 00:02:46,680 work couldn't be more critical. I mean, the tech 59 00:02:46,680 --> 00:02:48,580 world is currently consumed by this white -hot 60 00:02:48,580 --> 00:02:50,939 obsession with deep learning and generative models. 61 00:02:51,300 --> 00:02:53,500 It's literally everywhere. You can't escape it. 62 00:02:53,500 --> 00:02:55,300 You can't escape it. Right. It's a monolithic 63 00:02:55,300 --> 00:02:58,729 trend. But Rosenblum forces us to look past this 64 00:02:58,729 --> 00:03:00,789 current moment. He really demands to what it 65 00:03:00,789 --> 00:03:03,490 actually takes to build a full -spectrum mind, 66 00:03:03,889 --> 00:03:06,729 not just a system that can predict the next word 67 00:03:06,729 --> 00:03:09,449 in a sentence, but a complete cognitive architecture 68 00:03:09,449 --> 00:03:12,889 that experiences, learns, and navigates reality. 69 00:03:13,210 --> 00:03:15,150 We basically have to learn how to think like 70 00:03:15,150 --> 00:03:18,610 him. Right. Early in the text, he lays out his 71 00:03:18,610 --> 00:03:21,789 meta -scientist toolkit. And he starts by categorizing 72 00:03:21,789 --> 00:03:24,430 the entire field into three distinct levels of 73 00:03:24,430 --> 00:03:28,090 subject matter. Yeah. So if you imagine a hierarchy 74 00:03:28,090 --> 00:03:30,919 of understanding, right? At the most concrete 75 00:03:30,919 --> 00:03:33,740 foundational level, you have cognitive architectures. 76 00:03:33,759 --> 00:03:37,479 OK. These are the highly specific fixed structures 77 00:03:37,479 --> 00:03:40,340 and computational processes that theoretically 78 00:03:40,340 --> 00:03:43,800 yield a mind. So if we use a biological analogy, 79 00:03:44,060 --> 00:03:47,039 this is the literal physical layout of neurons, 80 00:03:47,520 --> 00:03:49,580 like this specific wiring of the hippocampus 81 00:03:49,580 --> 00:03:52,219 and the prefrontal cortex. That is exactly it. 82 00:03:52,520 --> 00:03:55,379 It's the hardware design, basically. Then moving 83 00:03:55,379 --> 00:03:58,659 up one level of abstraction, you have the common 84 00:03:58,659 --> 00:04:02,539 model of cognition. Which is what, exactly? This 85 00:04:02,539 --> 00:04:06,020 is an ongoing, massive attempt by the global 86 00:04:06,020 --> 00:04:08,759 scientific community to find consensus. It's 87 00:04:08,759 --> 00:04:10,659 the effort to look at all the different cognitive 88 00:04:10,659 --> 00:04:12,699 architectures researchers have built and say, 89 00:04:13,240 --> 00:04:16,199 OK, what features absolutely must be present 90 00:04:16,199 --> 00:04:19,019 in any architecture capable of human -like cognition? 91 00:04:19,660 --> 00:04:22,399 Oh, I see. So it's like an anatomy textbook that 92 00:04:22,399 --> 00:04:24,699 defines what a mammal is, regardless of whether 93 00:04:24,699 --> 00:04:27,660 it's a mouse or an elephant. That is a perfect 94 00:04:27,660 --> 00:04:31,459 analogy. Yes. And finally, at the top, the most 95 00:04:31,459 --> 00:04:34,560 generic and abstract level, you have minds themselves. 96 00:04:35,060 --> 00:04:37,660 Just minds in general. Right. The living, breathing 97 00:04:37,660 --> 00:04:40,000 entities, whether they are naturally evolved, 98 00:04:40,319 --> 00:04:42,459 like ours, or artificially constructed in the 99 00:04:42,459 --> 00:04:46,139 lab. Okay, so to study these three levels, architectures, 100 00:04:46,300 --> 00:04:48,920 the common model, and minds, he argues that we 101 00:04:48,920 --> 00:04:50,879 can't just use the standard scientific method, 102 00:04:51,079 --> 00:04:52,819 right? No, he thinks it's the wrong tool for 103 00:04:52,819 --> 00:04:55,879 the job. He proposes three... specific forms 104 00:04:55,879 --> 00:04:58,639 of analysis, I think. Essential, dichotomic, 105 00:04:58,759 --> 00:05:00,959 and relational. But the one that really anchors 106 00:05:00,959 --> 00:05:02,800 the first part of this entire journey is what 107 00:05:02,800 --> 00:05:05,459 he calls the reflective method. Yeah, the reflective 108 00:05:05,459 --> 00:05:08,420 method is key. He actually argues that the traditional 109 00:05:08,420 --> 00:05:11,740 scientific method is fundamentally flawed for 110 00:05:11,740 --> 00:05:14,959 this specific type of foundational work. Wait, 111 00:05:14,980 --> 00:05:17,779 really? The scientific method. That's a pretty 112 00:05:17,779 --> 00:05:20,339 bold claim. Well, think about it. The traditional 113 00:05:20,339 --> 00:05:23,569 scientific method. where you observe a phenomenon, 114 00:05:24,110 --> 00:05:26,810 formulate a hypothesis, run a controlled experiment. 115 00:05:27,269 --> 00:05:30,589 It is exquisitely designed for testing specific 116 00:05:30,589 --> 00:05:33,990 observable things in the physical world. Right, 117 00:05:34,089 --> 00:05:37,350 like testing what temperature water boils. Exactly. 118 00:05:37,589 --> 00:05:41,750 Or how a specific chemical reacts. But Rosenblum 119 00:05:41,750 --> 00:05:44,829 is dealing with meta -science, science about 120 00:05:44,829 --> 00:05:48,639 science. Oh, I see. When you are trying to define 121 00:05:48,639 --> 00:05:51,920 massive, historically loaded, abstract concepts 122 00:05:51,920 --> 00:05:56,220 like intelligence or learning or symbols, running 123 00:05:56,220 --> 00:05:59,180 a lab experiment doesn't actually help you. Because 124 00:05:59,180 --> 00:06:02,199 you can't put intelligence in a petri dish. Precisely. 125 00:06:02,439 --> 00:06:04,480 If researchers don't even agree on what a symbol 126 00:06:04,480 --> 00:06:07,259 is, an experiment involving symbols is totally 127 00:06:07,259 --> 00:06:10,459 useless. You need a method that identifies essences 128 00:06:10,459 --> 00:06:14,069 first. Right. So the reflective method is a rigorous, 129 00:06:14,529 --> 00:06:16,949 generalized framework for stepping completely 130 00:06:16,949 --> 00:06:19,089 out of the lab and rethinking the foundational 131 00:06:19,089 --> 00:06:21,769 categories we use to organize reality. Yeah. 132 00:06:21,810 --> 00:06:24,629 And the absolute linchpin of this reflective 133 00:06:24,629 --> 00:06:28,129 method is what he identifies as step three, finding 134 00:06:28,129 --> 00:06:31,209 the absolute essences of these topics. He talks 135 00:06:31,209 --> 00:06:34,629 about lumping before splitting, right? Yes. finding 136 00:06:34,629 --> 00:06:37,290 this incredibly delicate balance of deciding 137 00:06:37,290 --> 00:06:40,009 what a crude historical learning we take with 138 00:06:40,009 --> 00:06:43,329 us into the future and what we completely leave 139 00:06:43,329 --> 00:06:46,009 behind to change our perspective. You know, when 140 00:06:46,009 --> 00:06:48,370 I was trying to visualize this process, the only 141 00:06:48,370 --> 00:06:51,970 analogy that worked for me was Marie Kondo. The 142 00:06:51,970 --> 00:06:54,189 home organization philosophy applied to cognitive 143 00:06:54,189 --> 00:06:56,870 science. Hear me out. I mean, it's like he's 144 00:06:56,870 --> 00:06:59,910 Marie Kondoing decades of scientific dogma. He's 145 00:06:59,910 --> 00:07:02,730 picking up this incredibly complex, historically 146 00:07:02,730 --> 00:07:06,029 burdened AI theory. holding it up, looking at 147 00:07:06,029 --> 00:07:08,089 all its bells and whistles and asking like, what 148 00:07:08,089 --> 00:07:10,970 is the absolute bare minimum essence of this 149 00:07:10,970 --> 00:07:13,230 concept? Oh, I see where you're going with this. 150 00:07:13,629 --> 00:07:15,970 Does this spark functional intelligence? Exactly. 151 00:07:16,410 --> 00:07:18,949 What is mandatory for a mind and what is just 152 00:07:18,949 --> 00:07:22,769 an optional historical elaboration? He's stripping 153 00:07:22,769 --> 00:07:25,550 it down to see what actually sparks joy or in 154 00:07:25,550 --> 00:07:27,910 this case, intelligence. That's actually great. 155 00:07:28,009 --> 00:07:30,029 Taking that Marie Kondo analogy a step further. 156 00:07:30,569 --> 00:07:33,470 In fields that are utterly flooded with jargon, 157 00:07:33,790 --> 00:07:36,269 competing sub -disciplines, and frankly, massive 158 00:07:36,269 --> 00:07:39,870 egos, stripping away those non -essential elaborations 159 00:07:39,870 --> 00:07:41,970 isn't just about making things simple for the 160 00:07:41,970 --> 00:07:44,670 sake of a tidy textbook. Right. It has a real 161 00:07:44,670 --> 00:07:47,230 purpose. It is a vital structural necessity. 162 00:07:47,730 --> 00:07:49,350 I mean, look at the history of science. Isaac 163 00:07:49,350 --> 00:07:52,709 Newton was a massive lumper. Really? How so? 164 00:07:52,860 --> 00:07:55,000 Before Newton, people thought the forces pulling 165 00:07:55,000 --> 00:07:57,379 an apple to the ground were entirely different 166 00:07:57,379 --> 00:07:59,240 from the forces keeping the planets orbiting 167 00:07:59,240 --> 00:08:02,360 the sun. Oh, wow. I didn't realize that. Yeah. 168 00:08:03,100 --> 00:08:05,019 Newton stripped away the non -essential details, 169 00:08:05,379 --> 00:08:07,779 the fact that one is a fruit and the other is 170 00:08:07,779 --> 00:08:10,579 a celestial body, and lumped them together under 171 00:08:10,579 --> 00:08:14,220 the essential concept of gravity. And Rosenblum 172 00:08:14,220 --> 00:08:17,540 is trying to do exactly that for artificial intelligence. 173 00:08:18,019 --> 00:08:21,339 Exactly. By stripping a concept down to its absolute 174 00:08:21,339 --> 00:08:24,430 functional core. You prevent researchers from 175 00:08:24,430 --> 00:08:27,329 spending decades arguing over semantics. You 176 00:08:27,329 --> 00:08:30,089 force everyone to look at the underlying physics 177 00:08:30,089 --> 00:08:33,190 of cognition. Which is exactly why you, listening 178 00:08:33,190 --> 00:08:35,490 right now, should care about this meta -scientist 179 00:08:35,490 --> 00:08:38,629 toolkit. I mean, the next time you open your 180 00:08:38,629 --> 00:08:40,990 news feed and read a breathless headline about 181 00:08:40,990 --> 00:08:43,370 a massive new AI breakthrough that will change 182 00:08:43,370 --> 00:08:46,090 the world, applying this reflective method allows 183 00:08:46,090 --> 00:08:48,029 you to cut straight through the marketing speak. 184 00:08:48,409 --> 00:08:50,509 It really does. It gives you a filter. You can 185 00:08:50,509 --> 00:08:53,610 ask yourself, wait, What core architectural change 186 00:08:53,610 --> 00:08:56,549 actually just occurred here? Or is this just 187 00:08:56,549 --> 00:08:59,629 the same essence dressed up in a new shiny elaboration? 188 00:09:00,409 --> 00:09:02,149 And Rosenblum doesn't waste any time putting 189 00:09:02,149 --> 00:09:05,049 this tool to work. He immediately aims his essential 190 00:09:05,049 --> 00:09:08,090 analysis at the two most debated foundational 191 00:09:08,090 --> 00:09:11,049 and polarizing words in the entire history of 192 00:09:11,049 --> 00:09:14,629 artificial intelligence. Architectures and symbols. 193 00:09:14,830 --> 00:09:17,590 Yes, the big ones. Let's start with the building 194 00:09:17,590 --> 00:09:21,190 blocks. Architectures. He traces the entire lineage 195 00:09:21,190 --> 00:09:23,129 of the concept of an architecture, all the way 196 00:09:23,129 --> 00:09:26,190 from physical buildings, like the Roman Coliseum, 197 00:09:26,429 --> 00:09:28,649 to computer hardware, and finally up into the 198 00:09:28,649 --> 00:09:31,529 realm of cognition. And applying his essentialist 199 00:09:31,529 --> 00:09:34,710 analysis, stripping away all the noise, he determines 200 00:09:34,710 --> 00:09:37,950 that the absolute, irreducible core of an architecture 201 00:09:37,950 --> 00:09:42,029 is fixity. Fixity? Fixity. Meaning it doesn't 202 00:09:42,029 --> 00:09:45,330 change. Right. An architecture provides the fixed 203 00:09:45,330 --> 00:09:48,269 structures and fixed processes that ultimately 204 00:09:48,269 --> 00:09:51,250 yield a system. If you want a mathematical equation 205 00:09:51,250 --> 00:09:53,870 to anchor this, Rosenblum defines it very clearly. 206 00:09:54,470 --> 00:09:57,789 System equals architecture plus variable content. 207 00:09:58,370 --> 00:10:00,330 Okay, let's map that onto the listener's daily 208 00:10:00,330 --> 00:10:03,730 experience. System equals architecture plus variable 209 00:10:03,730 --> 00:10:07,590 content. Let's do it. So if my mind, my personal 210 00:10:07,590 --> 00:10:10,610 consciousness is the system, the architecture 211 00:10:10,610 --> 00:10:13,370 is the fixed biological machinery I was born 212 00:10:13,370 --> 00:10:15,990 with. Correct. It's the structure of my hippocampus, 213 00:10:16,750 --> 00:10:18,970 the neural pathways in my prefrontal cortex, 214 00:10:19,490 --> 00:10:21,690 the biological limits of my short -term memory. 215 00:10:22,330 --> 00:10:25,210 And the variable content is, well, everything 216 00:10:25,210 --> 00:10:27,909 I've experienced. My memory of the route I take 217 00:10:27,909 --> 00:10:30,549 to work, my ability to speak English, my political 218 00:10:30,549 --> 00:10:33,610 beliefs. Precisely those things. The variable 219 00:10:33,610 --> 00:10:36,610 content consists of the constantly changing skills, 220 00:10:36,809 --> 00:10:39,330 knowledge, and memories that your fixed architecture 221 00:10:39,330 --> 00:10:41,830 acquires, stores, and processes over a lifetime. 222 00:10:42,070 --> 00:10:45,110 OK, that makes sense. The architecture is the 223 00:10:45,110 --> 00:10:48,049 permanent stage, right? The variable content 224 00:10:48,049 --> 00:10:50,370 is the infinite number of different plays being 225 00:10:50,370 --> 00:10:52,309 performed on that stage over the course of your 226 00:10:52,309 --> 00:10:55,159 life. That makes perfect sense. But here is where 227 00:10:55,159 --> 00:10:57,679 the text gets highly controversial, especially 228 00:10:57,679 --> 00:10:59,740 for anyone who has studied the history of computer 229 00:10:59,740 --> 00:11:03,100 science. He turns his essential analysis on the 230 00:11:03,100 --> 00:11:07,120 word symbols. Oh yeah, this is a big deal. To 231 00:11:07,120 --> 00:11:09,240 fully grasp the magnitude of this controversy, 232 00:11:09,440 --> 00:11:12,259 we have to look at the historical schism in artificial 233 00:11:12,259 --> 00:11:15,039 intelligence. The Civil War. Exactly. Since the 234 00:11:15,039 --> 00:11:17,299 earliest days of AI, going back to the 1950s 235 00:11:17,299 --> 00:11:20,019 and 60s, there was this guiding philosophy called 236 00:11:20,019 --> 00:11:24,179 the Physical Symbol Systems Hypothesis, or PSSH. 237 00:11:25,250 --> 00:11:27,149 This was developed by the founding fathers of 238 00:11:27,149 --> 00:11:30,110 AI, giants like Alan Newell and Herbert Simon. 239 00:11:30,269 --> 00:11:33,230 This was the bedrock idea that manipulating symbols 240 00:11:33,230 --> 00:11:36,149 based on explicit rules is both necessary and 241 00:11:36,149 --> 00:11:38,570 sufficient to achieve general intelligence. That 242 00:11:38,570 --> 00:11:41,250 was the gospel for decades. This era is sometimes 243 00:11:41,250 --> 00:11:45,629 called good old fashioned AI or Goffi. Goffi. 244 00:11:45,690 --> 00:11:48,669 Goffi. I like that. The idea was that if you 245 00:11:48,669 --> 00:11:51,509 want a computer to understand a dog, you literally 246 00:11:51,509 --> 00:11:54,850 code a discrete symbol, a digital token, for 247 00:11:54,850 --> 00:11:57,940 dog. Right. Then you write explicit, rigid rules 248 00:11:57,940 --> 00:12:00,779 about how that dog symbol interacts with a barking 249 00:12:00,779 --> 00:12:04,259 symbol or a fur symbol. It is highly logical, 250 00:12:04,639 --> 00:12:07,809 highly readable by humans, and incredibly brittle. 251 00:12:08,210 --> 00:12:11,450 Because the real world is messy. Exactly. But 252 00:12:11,450 --> 00:12:15,090 the modern tension, the massive schism, is that 253 00:12:15,090 --> 00:12:17,269 neural networks, the deep learning models driving 254 00:12:17,269 --> 00:12:20,549 today's explosion in AI capabilities, they do 255 00:12:20,549 --> 00:12:22,789 not use traditional symbol manipulation at all. 256 00:12:22,970 --> 00:12:25,929 Not at all. Nope. They use continuous mathematical 257 00:12:25,929 --> 00:12:28,950 equations, adjusting millions of numeric weights 258 00:12:28,950 --> 00:12:32,710 and activations across a massive web of artificial 259 00:12:32,710 --> 00:12:35,529 neurons. So for decades, there has been a bitter 260 00:12:35,529 --> 00:12:38,750 civil war in AI research. The traditional symbolic 261 00:12:38,750 --> 00:12:42,049 rule -based camp on one side and the continuous 262 00:12:42,049 --> 00:12:44,549 mathematical neural network camp on the other. 263 00:12:44,809 --> 00:12:47,450 Exactly. And Rosenblum looks at this decades 264 00:12:47,450 --> 00:12:50,269 -long civil war and basically says, you're both 265 00:12:50,269 --> 00:12:54,139 arguing over elaborations. Let's use essential 266 00:12:54,139 --> 00:12:57,820 analysis. Right, the Marie Kondo approach. Yes. 267 00:12:58,320 --> 00:13:00,740 He takes the concept of a symbol and strips it 268 00:13:00,740 --> 00:13:04,259 down to its absolute bare minimum. He ruthlessly 269 00:13:04,259 --> 00:13:06,379 removes the traditional historical requirements. 270 00:13:06,799 --> 00:13:09,779 In the classical Newell and Simon era, a symbol 271 00:13:09,779 --> 00:13:12,240 had to possess composition, right? Meaning it 272 00:13:12,240 --> 00:13:13,940 could be snapped together with other symbols 273 00:13:13,940 --> 00:13:16,419 like Lego bricks to build complex thoughts. Right. 274 00:13:16,830 --> 00:13:20,669 And it had to possess designation, meaning it 275 00:13:20,669 --> 00:13:23,549 explicitly pointed to a specific identifiable 276 00:13:23,549 --> 00:13:26,690 thing in the real world. But Rosenblum strips 277 00:13:26,690 --> 00:13:29,730 all of that away. He does. He decides the true 278 00:13:29,730 --> 00:13:32,629 irreducible essence of a symbol is simply an 279 00:13:32,629 --> 00:13:35,870 atom or a placeholder. By stripping it down to 280 00:13:35,870 --> 00:13:39,029 a mere placeholder, he is able to propose a massive 281 00:13:39,029 --> 00:13:42,149 unifying umbrella, the hybrid symbol systems 282 00:13:42,149 --> 00:13:46,950 hypothesis, or HSSH. Right. And this new hypothesis 283 00:13:46,950 --> 00:13:50,190 is designed to cover both the traditional, rigid, 284 00:13:50,509 --> 00:13:54,090 rule -based symbolic AI and the fluid, continuous 285 00:13:54,090 --> 00:13:56,789 math of modern neural networks. You know, this 286 00:13:56,789 --> 00:13:59,070 is where I found myself pushing back pretty hard 287 00:13:59,070 --> 00:14:01,850 while reading. Oh, really? Yeah. I mean, if you 288 00:14:01,850 --> 00:14:05,269 strip a word like symbol down to just being a 289 00:14:05,269 --> 00:14:08,509 generic, undefined placeholder, haven't you watered 290 00:14:08,509 --> 00:14:10,929 the definition down so thoroughly that it completely 291 00:14:10,929 --> 00:14:13,779 loses its scientific utility? That's a fair question. 292 00:14:14,000 --> 00:14:16,980 Like, is he just trying to force modern neural 293 00:14:16,980 --> 00:14:21,299 networks into an old 1970s conceptual box so 294 00:14:21,299 --> 00:14:23,720 that the old guard of cognitive architecture 295 00:14:23,720 --> 00:14:26,159 doesn't feel historically left out of the current 296 00:14:26,159 --> 00:14:28,639 AI boom? It is a totally natural skepticism, 297 00:14:28,919 --> 00:14:31,419 and it's a critique frequently leveled at lumpers, 298 00:14:31,460 --> 00:14:34,179 as we talked about earlier. But the defense of 299 00:14:34,179 --> 00:14:36,679 this move is rooted in the search for a universal 300 00:14:36,679 --> 00:14:40,279 physics of mind. OK, unpack that for me. This 301 00:14:40,279 --> 00:14:42,750 isn't watering down the definition. it's finding 302 00:14:42,750 --> 00:14:46,230 the least common core. If you ever want a unified 303 00:14:46,230 --> 00:14:48,629 theory of mind, if you want to build an AI that 304 00:14:48,629 --> 00:14:51,250 can intuitively recognize a face, like a neural 305 00:14:51,250 --> 00:14:53,990 net, but also logically deduce a mathematical 306 00:14:53,990 --> 00:14:57,309 proof like a symbolic system, you cannot have 307 00:14:57,309 --> 00:15:00,090 two completely incompatible scientific vocabularies. 308 00:15:00,450 --> 00:15:02,970 Oh, so they literally need a shared language. 309 00:15:03,289 --> 00:15:06,730 Exactly. By removing the strict rigid rules of 310 00:15:06,730 --> 00:15:09,629 what a traditional symbol must be, Rosenblum 311 00:15:09,629 --> 00:15:12,509 creates a conceptual space where we can actually 312 00:15:12,509 --> 00:15:15,970 talk about how neural networks and symbolic logic 313 00:15:15,970 --> 00:15:19,730 integrate. So he's arguing that even a tiny anonymous 314 00:15:19,730 --> 00:15:22,549 node firing deep inside a massive neural network 315 00:15:22,549 --> 00:15:25,509 operating purely on a mathematical weight of 316 00:15:25,509 --> 00:15:29,330 0 .8 can be viewed as a minimal essential symbol? 317 00:15:29,549 --> 00:15:33,139 Yes. It's a placeholder, representing some abstract 318 00:15:33,139 --> 00:15:36,200 feature of the world, even if we, the human programmers, 319 00:15:36,419 --> 00:15:39,220 have absolutely no idea what it explicitly designates. 320 00:15:40,139 --> 00:15:43,580 That synthesis is the breakthrough. It stops 321 00:15:43,580 --> 00:15:46,399 treating these two camps as fundamentally incompatible 322 00:15:46,399 --> 00:15:49,500 alien species and starts treating them as different 323 00:15:49,500 --> 00:15:51,799 structural expressions under the exact same genus. 324 00:15:52,120 --> 00:15:54,500 We have to agree on the raw materials before 325 00:15:54,500 --> 00:15:57,429 we can build the house. Exactly. But if these 326 00:15:57,429 --> 00:16:00,509 hybrid symbols and fixed architectures are the 327 00:16:00,509 --> 00:16:03,710 raw materials, what exactly are we trying to 328 00:16:03,710 --> 00:16:06,149 build? I mean, the tech world throws the word 329 00:16:06,149 --> 00:16:09,129 intelligence around constantly. Oh constantly. 330 00:16:09,690 --> 00:16:12,750 Everything is smart now. Right. Every startup 331 00:16:12,750 --> 00:16:15,210 claims their software is intelligent. But without 332 00:16:15,210 --> 00:16:18,269 a clear, rigorous definition, we're building 333 00:16:18,269 --> 00:16:20,649 that house without a blueprint. We really need 334 00:16:20,649 --> 00:16:23,059 to define the goal. And Rosenblum is acutely 335 00:16:23,059 --> 00:16:25,799 aware that the cacophony of definitions for intelligence 336 00:16:25,799 --> 00:16:29,700 is just deafening. Biologists have one definition, 337 00:16:30,419 --> 00:16:32,379 software engineers have another, philosophers 338 00:16:32,379 --> 00:16:35,340 have a third. It's a mess. So he attempts to 339 00:16:35,340 --> 00:16:38,039 bring structural clarity by dissecting the concept 340 00:16:38,039 --> 00:16:41,240 of intelligence and organizing it into a nested 341 00:16:41,240 --> 00:16:43,759 three -level structure. This is the Russian nesting 342 00:16:43,759 --> 00:16:46,600 doll model of intelligence. Yes. Think of the 343 00:16:46,600 --> 00:16:49,200 smallest doll, the innermost core. He labels 344 00:16:49,200 --> 00:16:52,139 this level one immediate intelligence. And what 345 00:16:52,139 --> 00:16:54,740 does that entail? This is the raw ability to 346 00:16:54,740 --> 00:16:57,299 process rapid perception, recognize patterns, 347 00:16:57,320 --> 00:16:59,600 and make immediate single step decisions based 348 00:16:59,600 --> 00:17:02,320 on the current context. Which perfectly describes 349 00:17:02,320 --> 00:17:04,680 exactly what large language models, the chat 350 00:17:04,680 --> 00:17:07,519 GBTs and clods of the world currently excel at. 351 00:17:07,700 --> 00:17:10,059 Absolutely. You give them a prompt, they rapidly 352 00:17:10,059 --> 00:17:12,279 calculate the mathematical probabilities across 353 00:17:12,279 --> 00:17:15,180 their massive neural net, and they immediately 354 00:17:15,180 --> 00:17:17,779 generate the next most likely word. It is a highly 355 00:17:17,779 --> 00:17:20,849 sophisticated but immediate reflex. But if you 356 00:17:20,849 --> 00:17:23,750 open that smallest doll, you find it sits inside 357 00:17:23,750 --> 00:17:26,509 a larger, more complex structure, level two, 358 00:17:26,809 --> 00:17:29,109 cumulative intelligence. OK, how is that different? 359 00:17:29,769 --> 00:17:32,269 This represents lifelong pervasive incremental 360 00:17:32,269 --> 00:17:35,049 learning. It's an intelligence that learns online, 361 00:17:35,589 --> 00:17:37,869 constantly updating its internal models, correcting 362 00:17:37,869 --> 00:17:40,009 its assumptions, and integrating new knowledge 363 00:17:40,009 --> 00:17:42,710 across an entire lifetime of experiences. Which 364 00:17:42,710 --> 00:17:44,990 is where current generative AI struggles massively. 365 00:17:45,170 --> 00:17:47,369 Right, because an LLM doesn't actually learn 366 00:17:47,369 --> 00:17:50,009 from you while you're talking to it. Its core 367 00:17:50,009 --> 00:17:51,950 training is permanently frozen. Yeah, I've noticed 368 00:17:51,950 --> 00:17:54,250 that. It has a temporary context window for your 369 00:17:54,250 --> 00:17:56,569 current conversation, sure. But the moment you 370 00:17:56,569 --> 00:17:59,230 close the tab, it basically forgets you. It resets 371 00:17:59,230 --> 00:18:01,849 completely. It isn't fundamentally altering its 372 00:18:01,849 --> 00:18:04,750 base neural architecture with every single interaction. 373 00:18:05,150 --> 00:18:07,769 The way, say, a human toddler does when they 374 00:18:07,769 --> 00:18:10,210 touch a hot stove and permanently rewire their 375 00:18:10,210 --> 00:18:12,109 understanding of the world. That's a great example. 376 00:18:12,269 --> 00:18:14,890 The toddler possesses cumulative intelligence. 377 00:18:15,349 --> 00:18:19,029 The LLM currently does not. And then that level 378 00:18:19,029 --> 00:18:21,549 two doll sits inside the largest all -encompassing 379 00:18:21,549 --> 00:18:24,390 doll, level three full spectrum intelligence. 380 00:18:24,490 --> 00:18:27,410 The whole package. Right. This encompasses absolutely 381 00:18:27,410 --> 00:18:30,589 all aspects of a mind. It includes internal goal 382 00:18:30,589 --> 00:18:33,950 generation, physical sensations, complex temporal 383 00:18:33,950 --> 00:18:36,609 planning, and true holistic awareness of its 384 00:18:36,609 --> 00:18:39,250 environment. And Rosenblum is actually quite 385 00:18:39,250 --> 00:18:41,369 blunt about where our current technology sits, 386 00:18:41,509 --> 00:18:43,890 isn't he? Oh, very blunt. He points out that 387 00:18:43,890 --> 00:18:46,890 today's generative AI systems possess massive 388 00:18:46,890 --> 00:18:49,710 memory banks and incredibly superhuman immediate 389 00:18:49,710 --> 00:18:52,309 intelligence, but architecturally speaking, they 390 00:18:52,309 --> 00:18:55,250 are essentially mindless. It sounds incredibly 391 00:18:55,250 --> 00:18:58,009 harsh, especially given how fluent and convincing 392 00:18:58,009 --> 00:19:01,390 these systems are, but from a cognitive architecture 393 00:19:01,390 --> 00:19:03,890 standpoint, it's objectively accurate. Because 394 00:19:03,890 --> 00:19:06,009 they lack that level two and three stuff? Exactly. 395 00:19:06,220 --> 00:19:08,619 They lack the pervasive incremental learning 396 00:19:08,619 --> 00:19:11,680 loops of a true architecture. They are functionally 397 00:19:11,680 --> 00:19:14,720 static once their training run ends. They are 398 00:19:14,720 --> 00:19:16,839 just processing your input through millions of 399 00:19:16,839 --> 00:19:19,660 frozen mathematical weights. Returning to that 400 00:19:19,660 --> 00:19:22,000 Russian nesting doll analogy for a moment, right 401 00:19:22,000 --> 00:19:24,339 now the entire tech sector, the global stock 402 00:19:24,339 --> 00:19:26,599 market, the media, they are hyper -focused on 403 00:19:26,599 --> 00:19:28,920 the smallest doll, immediate intelligence. Because 404 00:19:28,920 --> 00:19:31,420 it's shiny and new. Right. But because the output 405 00:19:31,420 --> 00:19:33,900 is so convincing, we are desperately trying to 406 00:19:33,900 --> 00:19:36,039 paint that tiny doll to make it look like the 407 00:19:36,039 --> 00:19:38,599 biggest doll, full spectrum intelligence. We 408 00:19:38,599 --> 00:19:40,859 treat chatbots as if they have internal goals, 409 00:19:41,180 --> 00:19:44,339 beliefs, or self -awareness. And Rosenblum highlights 410 00:19:44,339 --> 00:19:46,720 that this projection is exactly what leads to 411 00:19:46,720 --> 00:19:49,059 the dangerous hallucinations we see in AI. We 412 00:19:49,059 --> 00:19:51,380 are demanding full spectrum reliability from 413 00:19:51,380 --> 00:19:54,039 a system that only possesses immediate reflex 414 00:19:54,039 --> 00:19:56,900 level architecture. That framing clarifies so 415 00:19:56,900 --> 00:19:59,640 much of the current AI panic, honestly. And it 416 00:19:59,640 --> 00:20:01,799 ties directly into his exploration of what a 417 00:20:01,799 --> 00:20:04,259 theory actually is. Yeah, because building a 418 00:20:04,259 --> 00:20:06,299 cognitive architecture is ultimately building 419 00:20:06,299 --> 00:20:09,460 a unified theory of mind. He boils the essence 420 00:20:09,460 --> 00:20:11,940 of a theory down to the concept of aboutness 421 00:20:11,940 --> 00:20:15,579 or what he calls meta systems. Aboutness. Right. 422 00:20:16,079 --> 00:20:19,240 A theory is simply a system that is about another 423 00:20:19,240 --> 00:20:21,960 system. A map is a theory about a territory. 424 00:20:22,259 --> 00:20:24,779 And when building these theories, these architectures, 425 00:20:24,900 --> 00:20:27,440 he introduces a contrast that I found deeply 426 00:20:27,440 --> 00:20:30,720 fascinating. The pursuit of functional elegance 427 00:20:30,720 --> 00:20:34,140 versus the trap of exuberance. Functional elegance 428 00:20:34,140 --> 00:20:36,859 is his holy grail. It is the goal of achieving 429 00:20:36,859 --> 00:20:40,380 incredibly broad, complex, human -like capabilities 430 00:20:40,380 --> 00:20:43,539 from a very simple, theoretically elegant base 431 00:20:43,539 --> 00:20:46,599 of fixed mechanisms. So doing more with less? 432 00:20:46,960 --> 00:20:49,000 Exactly. It's the architectural equivalent of 433 00:20:49,000 --> 00:20:52,950 doing a tremendous amount with very little. Exuberance, 434 00:20:52,970 --> 00:20:55,470 on the other hand, is the brute force approach. 435 00:20:55,640 --> 00:20:57,799 Just throwing everything at the wall. Throwing 436 00:20:57,799 --> 00:21:00,279 endless parameters, massive server farms, oceans 437 00:21:00,279 --> 00:21:03,339 of scraped internet data, and sheer computing 438 00:21:03,339 --> 00:21:05,500 power at a problem until it magically starts 439 00:21:05,500 --> 00:21:08,000 working. Which is the exact definition of modern 440 00:21:08,000 --> 00:21:11,099 massive neural networks. Yep. They are pure exuberance. 441 00:21:11,259 --> 00:21:13,460 We don't fully understand how they work. We just 442 00:21:13,460 --> 00:21:16,420 keep adding trillions of parameters and hoping 443 00:21:16,420 --> 00:21:18,640 intelligence emerges from the sheer scale of 444 00:21:18,640 --> 00:21:21,789 it all. And this dichotomy highlights a crucial 445 00:21:21,789 --> 00:21:23,910 difference between architectural understanding 446 00:21:23,910 --> 00:21:27,769 and mere cognitive understanding. LLMs have a 447 00:21:27,769 --> 00:21:29,410 built -in mechanism for learning during their 448 00:21:29,410 --> 00:21:31,769 massive training phase. They predict a word, 449 00:21:32,009 --> 00:21:34,190 check if they are right, and mathematically adjust 450 00:21:34,190 --> 00:21:37,009 their weights. Right. But they entirely lack 451 00:21:37,009 --> 00:21:39,890 what Rosenblum identifies as fundamental, low 452 00:21:39,890 --> 00:21:42,789 -level architectural appraisals. He talks about 453 00:21:42,789 --> 00:21:45,190 architectural emotions, which is just a wild 454 00:21:45,190 --> 00:21:47,769 concept to me. I know. It sounds weird at first. 455 00:21:47,980 --> 00:21:50,200 But he isn't talking about a computer feeling 456 00:21:50,200 --> 00:21:54,299 sadness or joy in the human romantic sense. He's 457 00:21:54,299 --> 00:21:56,640 talking about biological mechanisms translated 458 00:21:56,640 --> 00:22:00,119 into code. Let's look at the architectural appraisal 459 00:22:00,119 --> 00:22:03,720 of surprise. In a true full -spectrum cognitive 460 00:22:03,720 --> 00:22:06,420 architecture like your brain, your mind is constantly 461 00:22:06,420 --> 00:22:08,700 predicting what will happen a fraction of a second 462 00:22:08,700 --> 00:22:11,980 in the future. Constantly. Right. If your mental 463 00:22:11,980 --> 00:22:14,079 model predicts one thing but reality delivers 464 00:22:14,079 --> 00:22:16,390 something completely different, Your underlying 465 00:22:16,390 --> 00:22:18,950 architecture experiences a low -level biological 466 00:22:18,950 --> 00:22:22,230 appraisal of surprise. Okay, think about the 467 00:22:22,230 --> 00:22:24,470 listener's daily commute. You drive the exact 468 00:22:24,470 --> 00:22:26,970 same route to work for five years. Your brain 469 00:22:26,970 --> 00:22:30,390 is essentially on autopilot, highly efficient, 470 00:22:30,630 --> 00:22:32,809 predicting every stoplight and turn. We all do 471 00:22:32,809 --> 00:22:35,049 it. But one morning, you turn a corner and see 472 00:22:35,049 --> 00:22:37,390 a massive sinkhole blocking the road. Your brain 473 00:22:37,390 --> 00:22:40,349 doesn't just crash, it experiences an immediate 474 00:22:40,349 --> 00:22:43,309 spike of architectural surprise. And that surprise 475 00:22:43,309 --> 00:22:46,509 isn't just a feeling. It is a vital, functional 476 00:22:46,509 --> 00:22:50,000 trigger. That spike forces your cognitive architecture 477 00:22:50,000 --> 00:22:52,619 to instantly break out of its habitual cycle, 478 00:22:53,460 --> 00:22:56,180 rapidly shift attention, adapt its internal models, 479 00:22:56,519 --> 00:22:59,400 and begin learning incrementally to resolve the 480 00:22:59,400 --> 00:23:01,759 discrepancy between its prediction and reality. 481 00:23:01,940 --> 00:23:04,259 It forces your variable content to permanently 482 00:23:04,259 --> 00:23:07,980 update. Exactly. Large language models simply 483 00:23:07,980 --> 00:23:11,119 do not have this. They don't experience architectural 484 00:23:11,119 --> 00:23:14,700 surprise. If they predict a totally nonsensical 485 00:23:14,700 --> 00:23:18,000 word, they don't pause, feel a discrepancy and 486 00:23:18,000 --> 00:23:20,619 restructure their worldview. They just plow ahead. 487 00:23:20,839 --> 00:23:23,319 They confidently output the next most probable 488 00:23:23,319 --> 00:23:26,240 token, even if it totally contradicts reality. 489 00:23:26,859 --> 00:23:29,500 They lack the architectural mechanism of surprise 490 00:23:29,500 --> 00:23:32,779 that forces cumulative learning. This shifts 491 00:23:32,779 --> 00:23:35,240 the entire paradigm of how you interact with 492 00:23:35,240 --> 00:23:38,980 a system like ChetGPT. You can now clearly see 493 00:23:38,980 --> 00:23:43,019 it not as an emerging mind, but as an unimaginably 494 00:23:43,019 --> 00:23:46,759 vast ocean of human variable content. Literally 495 00:23:46,759 --> 00:23:50,019 the entire digitized text of human history. All 496 00:23:50,019 --> 00:23:52,619 of it. And it is being processed without any 497 00:23:52,619 --> 00:23:55,680 of the fixed, cumulative architectural structures 498 00:23:55,680 --> 00:23:59,440 required for actual intelligence. It's the world's 499 00:23:59,440 --> 00:24:02,289 largest database of human variable content. being 500 00:24:02,289 --> 00:24:06,029 blindly processed by a very narrow, purely immediate 501 00:24:06,029 --> 00:24:09,069 intelligence architecture. So we have the raw 502 00:24:09,069 --> 00:24:12,230 materials defined, these essential hybrid symbols 503 00:24:12,230 --> 00:24:14,890 and fixed architectures. We have the ultimate 504 00:24:14,890 --> 00:24:17,750 goal defined, full spectrum intelligence achieved 505 00:24:17,750 --> 00:24:20,630 through functional elegance. Right. But if you 506 00:24:20,630 --> 00:24:23,049 look at the actual landscape of AI research today, 507 00:24:23,230 --> 00:24:26,200 it is total chaos. Rosenblum mentions there are 508 00:24:26,200 --> 00:24:29,880 over 195 distinctly different cognitive architectures 509 00:24:29,880 --> 00:24:32,259 out there. It's a very crowded field. There are 510 00:24:32,259 --> 00:24:35,920 thousands of competing papers, endless subfields, 511 00:24:36,119 --> 00:24:39,099 tribal academic wars. How do we navigate this 512 00:24:39,099 --> 00:24:41,319 wilderness without getting completely lost in 513 00:24:41,319 --> 00:24:43,500 the weeds? To navigate the wilderness, you need 514 00:24:43,500 --> 00:24:46,660 a map. This is where Rosenblum shifts gears entirely. 515 00:24:47,059 --> 00:24:49,519 We move from the essential analysis of defining 516 00:24:49,519 --> 00:24:53,039 concepts into dichotomous analysis. He sits down 517 00:24:53,039 --> 00:24:55,619 to build a comprehensive map of the entire artificial 518 00:24:55,619 --> 00:24:58,980 intelligence landscape. Dichotomic analysis. 519 00:24:59,519 --> 00:25:01,759 Meaning analyzing complex systems by breaking 520 00:25:01,759 --> 00:25:04,500 them down into fundamental dichotomies or binary 521 00:25:04,500 --> 00:25:07,339 switches. Yes. Most standard textbooks on AI 522 00:25:07,339 --> 00:25:09,720 try to organize the field by human functions. 523 00:25:10,059 --> 00:25:12,500 You'll find a chapter on computer vision, a chapter 524 00:25:12,500 --> 00:25:15,039 on natural language processing, a chapter on 525 00:25:15,039 --> 00:25:17,319 robotic planning. Which makes sense to us because 526 00:25:17,319 --> 00:25:19,680 that's how we categorize human skills. It feels 527 00:25:19,680 --> 00:25:22,440 intuitive. But Rosenblum argues that categorizing 528 00:25:22,440 --> 00:25:25,420 AI by human skills is like categorizing vehicles 529 00:25:25,420 --> 00:25:27,960 by what cargo they carry, rather than how their 530 00:25:27,960 --> 00:25:30,980 engines actually work. That is a brilliant comparison. 531 00:25:31,279 --> 00:25:33,920 Right. He rejects the functional approach entirely. 532 00:25:34,700 --> 00:25:37,039 Instead, he looks at the underlying mechanical 533 00:25:37,039 --> 00:25:40,410 physics of information processing. He identifies 534 00:25:40,410 --> 00:25:43,089 three essential dichotomies, three binary switches, 535 00:25:43,759 --> 00:25:45,960 And by looking at the cross products of these 536 00:25:45,960 --> 00:25:48,480 switches, he creates an eight region, three dimensional 537 00:25:48,480 --> 00:25:51,859 map that categorizes every single AI system ever 538 00:25:51,859 --> 00:25:54,220 built. I want to walk through these three switches 539 00:25:54,220 --> 00:25:56,480 incredibly carefully because this is where the 540 00:25:56,480 --> 00:25:59,460 actual engineering mechanics of AI get exposed. 541 00:25:59,839 --> 00:26:03,059 The first switch is dichotomy one, asymmetric 542 00:26:03,059 --> 00:26:05,660 versus symmetric processing. How do these two 543 00:26:05,660 --> 00:26:08,259 mechanisms actually work at the base level? This 544 00:26:08,259 --> 00:26:11,119 dichotomy describes the directional flow of processing. 545 00:26:11,359 --> 00:26:15,019 Think about how information moves. Asymmetric 546 00:26:15,019 --> 00:26:17,900 processing is strictly directional. It moves 547 00:26:17,900 --> 00:26:20,460 in a one -way street. Okay, so like a straight 548 00:26:20,460 --> 00:26:22,900 line? The classic example in computer science 549 00:26:22,900 --> 00:26:26,140 is an if -then rule. If the condition is met, 550 00:26:26,859 --> 00:26:30,380 then execute the action. It is a rigid, causal 551 00:26:30,380 --> 00:26:33,099 chain. Symmetric processing, on the other hand, 552 00:26:33,400 --> 00:26:36,299 is omnidirectional. I love the physical metaphors 553 00:26:36,299 --> 00:26:39,230 here. Asymmetric processing is like a Rube Goldberg 554 00:26:39,230 --> 00:26:41,829 machine. A marble rolls down a track, hits a 555 00:26:41,829 --> 00:26:44,269 specific lever, which strictly triggers a specific 556 00:26:44,269 --> 00:26:47,430 pulley, which drops a specific bucket. The action 557 00:26:47,430 --> 00:26:50,309 only flows forward step by step. Exactly. But 558 00:26:50,309 --> 00:26:53,250 symmetric processing is like a spider web. The 559 00:26:53,250 --> 00:26:55,970 spider web is the perfect analogy. In a symmetric 560 00:26:55,970 --> 00:26:58,009 neural network, you have thousands of interconnected 561 00:26:58,009 --> 00:27:00,990 nodes. If you pluck one string on a spider web, 562 00:27:01,109 --> 00:27:03,910 the vibration travels omnidirectionally. It goes 563 00:27:03,910 --> 00:27:07,150 everywhere at once. Every node influences every 564 00:27:07,150 --> 00:27:10,670 other node simultaneously. The entire web adjusts 565 00:27:10,670 --> 00:27:12,849 and resonates until the whole system settles 566 00:27:12,849 --> 00:27:16,289 into a state of equilibrium. There is no strict 567 00:27:16,289 --> 00:27:19,569 A to B to C pathway. There is just a holistic 568 00:27:19,569 --> 00:27:22,849 settling into a pattern. OK, Rube Goldberg versus 569 00:27:22,849 --> 00:27:25,970 Spiderweb. Got it. The second switch is dichotomy 570 00:27:25,970 --> 00:27:28,390 2, controlled versus uncontrolled processing. 571 00:27:28,769 --> 00:27:31,490 This dichotomy is about the mechanism of deliberation. 572 00:27:31,900 --> 00:27:35,480 In a controlled process, a decision explicitly 573 00:27:35,480 --> 00:27:38,539 structurally precedes an action. So it thinks 574 00:27:38,539 --> 00:27:42,339 before it acts? Yes. The system pauses, evaluates 575 00:27:42,339 --> 00:27:44,700 multiple different options, selects the optimal 576 00:27:44,700 --> 00:27:47,039 one based on a goal, and then executes the action. 577 00:27:47,519 --> 00:27:49,720 It requires deliberate architectural control. 578 00:27:50,970 --> 00:27:53,289 An uncontrolled process requires no such pause. 579 00:27:53,470 --> 00:27:55,890 It just acts? It executes automatically, on its 580 00:27:55,890 --> 00:27:58,130 own impetus. The moment the input is received, 581 00:27:58,210 --> 00:28:01,210 it is entirely reactive. To bring this back to 582 00:28:01,210 --> 00:28:03,250 the human experience, it's the difference between 583 00:28:03,250 --> 00:28:06,390 reflex and deliberation. If I accidentally touch 584 00:28:06,390 --> 00:28:09,029 a hot stove, my hand violently pulls back before 585 00:28:09,029 --> 00:28:11,710 I even consciously realize what happened. That 586 00:28:11,710 --> 00:28:14,529 is an uncontrolled, automatic, biological reaction. 587 00:28:14,869 --> 00:28:17,210 Absolutely. But if I am standing in my kitchen 588 00:28:17,210 --> 00:28:19,210 deciding what to cook for dinner... I have to 589 00:28:19,210 --> 00:28:21,490 pause, evaluate what's in the fridge, consider 590 00:28:21,490 --> 00:28:24,170 my dietary goals, and make a deliberate choice 591 00:28:24,170 --> 00:28:27,269 before acting. Rosenblum maps that exact same 592 00:28:27,269 --> 00:28:30,009 biological dichotomy directly onto AI systems. 593 00:28:30,109 --> 00:28:32,670 It does. When a deep neural network recognizes 594 00:28:32,670 --> 00:28:35,109 a human face in a photograph, it doesn't pause 595 00:28:35,109 --> 00:28:37,849 and decide to see a face. It just sees it. The 596 00:28:37,849 --> 00:28:39,569 pixel data enters the input layer and triggers 597 00:28:39,569 --> 00:28:42,509 an automatic, uncontrolled cascade of mathematical 598 00:28:42,509 --> 00:28:44,769 activations through the web until the output 599 00:28:44,769 --> 00:28:47,910 layer fires. It is an incredibly complex reflex. 600 00:28:48,259 --> 00:28:51,119 Which brings us to the final switch, dichotomy 601 00:28:51,119 --> 00:28:54,819 three, data versus metadata. This dichotomy circles 602 00:28:54,819 --> 00:28:57,160 back to our earlier debate about symbols. Does 603 00:28:57,160 --> 00:28:59,940 the core processing of the system operate on 604 00:28:59,940 --> 00:29:03,500 symbolic data, which means discrete categorical 605 00:29:03,500 --> 00:29:06,900 structural elements like a specific word, a logical 606 00:29:06,900 --> 00:29:11,220 concept, or a distinct object? Or? Or does the 607 00:29:11,220 --> 00:29:14,619 system operate on numeric metadata? which means 608 00:29:14,619 --> 00:29:17,660 continuous quantitative values like probabilities, 609 00:29:18,079 --> 00:29:21,259 mathematical weights, or activation thresholds. 610 00:29:21,279 --> 00:29:24,460 So we have three binary switches governing the 611 00:29:24,460 --> 00:29:27,799 physics of AI. The flow of information, asymmetric 612 00:29:27,799 --> 00:29:31,259 versus symmetric. The mechanism of action, controlled 613 00:29:31,259 --> 00:29:34,599 versus uncontrolled. And the nature of the information 614 00:29:34,599 --> 00:29:38,359 itself, data versus metadata. That's the 3D map. 615 00:29:38,740 --> 00:29:40,859 By crossing these three switches, you generate 616 00:29:40,859 --> 00:29:44,019 an eight region 3D map. When I first encountered 617 00:29:44,019 --> 00:29:47,119 this map in the text, I was frankly highly skeptical. 618 00:29:47,500 --> 00:29:49,779 Why's that? Can you really take the sprawling 619 00:29:49,779 --> 00:29:52,480 60 -year history of artificial intelligence from 620 00:29:52,480 --> 00:29:56,180 the clunky 1980s expert systems diagnosing blood 621 00:29:56,180 --> 00:29:59,700 infections, all the way to modern miracles like 622 00:29:59,700 --> 00:30:02,279 AlphaZero mastering the game of Go, and reduce 623 00:30:02,279 --> 00:30:04,880 all of that staggering complexity into just three 624 00:30:04,880 --> 00:30:07,839 binary toggles? I mean, it feels way too neat. 625 00:30:08,400 --> 00:30:12,130 It is a profound level of lumping. which naturally 626 00:30:12,130 --> 00:30:15,470 invites skepticism. But the power of this map 627 00:30:15,470 --> 00:30:17,750 isn't about reducing the field into a simplistic 628 00:30:17,750 --> 00:30:20,970 cartoon. It's about discovering the foundational 629 00:30:20,970 --> 00:30:23,910 physics that govern how these incredibly diverse 630 00:30:23,910 --> 00:30:26,230 systems actually operate. Give me an example. 631 00:30:26,869 --> 00:30:28,690 Let's look closely at the example you just mentioned, 632 00:30:28,990 --> 00:30:32,049 Alpha Zero. Alpha Zero is the Google DeepMind 633 00:30:32,049 --> 00:30:35,569 AI that famously mastered chess, Shogi, and Go 634 00:30:35,569 --> 00:30:37,710 entirely from scratch, playing against itself. 635 00:30:37,849 --> 00:30:40,529 Which is still mind -blowing. It is. If you map 636 00:30:40,529 --> 00:30:42,829 AlphaZero's architecture onto Rosenblum's 3D 637 00:30:42,829 --> 00:30:45,690 grid, you see that it beautifully spans across 638 00:30:45,690 --> 00:30:48,869 multiple dichotomies. And that structural spanning 639 00:30:48,869 --> 00:30:51,950 is the exact mechanical reason it is so devastatingly 640 00:30:51,950 --> 00:30:54,410 powerful. Okay, walk me through the mechanics 641 00:30:54,410 --> 00:30:57,289 of that. How does AlphaZero physically map onto 642 00:30:57,289 --> 00:31:00,609 these switches? When AlphaZero looks at a complex 643 00:31:00,609 --> 00:31:03,289 chessboard, it first uses deep neural networks 644 00:31:03,289 --> 00:31:05,750 to evaluate the state of the board. It looks 645 00:31:05,750 --> 00:31:08,890 at the arrangement of pieces and intuits which 646 00:31:08,890 --> 00:31:11,210 overall positions are structurally strong. Just 647 00:31:11,210 --> 00:31:13,849 a quick glance, basically. Right. That rapid 648 00:31:13,849 --> 00:31:16,490 valuation process lives in one very specific 649 00:31:16,490 --> 00:31:19,990 region of our 3D map. It is symmetric. It uses 650 00:31:19,990 --> 00:31:22,769 the omnidirectional spider web of a neural net. 651 00:31:23,470 --> 00:31:25,990 It is uncontrolled. It happens as an automatic 652 00:31:25,990 --> 00:31:29,150 reflex upon seeing the board. And it relies on 653 00:31:29,150 --> 00:31:31,730 metadata. It calculates numeric probabilities 654 00:31:31,730 --> 00:31:35,069 of winning. So it has this incredibly fast, intuitive, 655 00:31:35,289 --> 00:31:37,210 uncontrolled gut feeling about the board, much 656 00:31:37,210 --> 00:31:39,390 like a human grandmaster looking at a setup and 657 00:31:39,390 --> 00:31:41,890 just feeling the tension on the squares. Exactly. 658 00:31:42,210 --> 00:31:44,190 But it doesn't just play its first gut feeling 659 00:31:44,190 --> 00:31:46,670 immediately. It takes that intuitive evaluation 660 00:31:46,789 --> 00:31:50,329 and feeds it into an entirely different algorithmic 661 00:31:50,329 --> 00:31:53,190 process called Monte Carlo Tree Search. This 662 00:31:53,190 --> 00:31:55,990 is where it explicitly plans out future moves. 663 00:31:56,349 --> 00:31:58,930 And how does Monte Carlo Tree Search work mechanically? 664 00:31:59,490 --> 00:32:02,279 Imagine the AI standing in a maze. Instead of 665 00:32:02,279 --> 00:32:04,660 just running down the path that looks best, it 666 00:32:04,660 --> 00:32:08,259 pauses and sends out thousands of invisible ghosts 667 00:32:08,259 --> 00:32:10,980 down all the different branching paths. Like 668 00:32:10,980 --> 00:32:14,640 scouts. Yeah. These ghosts play out the game, 669 00:32:14,759 --> 00:32:16,519 thousands of moves into the future, reporting 670 00:32:16,519 --> 00:32:19,000 back on which specific sequence of moves actually 671 00:32:19,000 --> 00:32:22,000 leads to a victory. This search process lives 672 00:32:22,000 --> 00:32:24,099 in a completely different region of Rosenblum's 673 00:32:24,099 --> 00:32:27,579 map. Which switches does it hit? It is asymmetric. 674 00:32:27,940 --> 00:32:30,539 It follows directional branching pathways of 675 00:32:30,539 --> 00:32:33,670 cause and effect. And it is controlled. It explicitly 676 00:32:33,670 --> 00:32:36,170 evaluates these future realities before choosing 677 00:32:36,170 --> 00:32:38,670 the optimal path and taking a single action in 678 00:32:38,670 --> 00:32:41,710 the present. Oh, wow. It's like Dr. Strange viewing 679 00:32:41,710 --> 00:32:44,170 14 million alternate futures before making a 680 00:32:44,170 --> 00:32:46,170 single move. That is a brilliant comparison. 681 00:32:46,690 --> 00:32:49,750 Yes. So it leverages the symmetric uncontrolled 682 00:32:49,750 --> 00:32:53,490 side for rapid intuition and the asymmetric controlled 683 00:32:53,490 --> 00:32:56,529 side for rigorous, deliberate planning. It marries 684 00:32:56,529 --> 00:32:58,960 the two beautifully. And this 3D map proves something 685 00:32:58,960 --> 00:33:01,579 vital for the entire future of the field. The 686 00:33:01,579 --> 00:33:04,079 path forward isn't about one side of the academic 687 00:33:04,079 --> 00:33:07,660 war winning. Right. It's not about symbolic logic, 688 00:33:07,900 --> 00:33:10,259 permanently defeating neural networks, or deep 689 00:33:10,259 --> 00:33:13,160 learning proving that symbols are useless. The 690 00:33:13,160 --> 00:33:16,240 future of true full spectrum intelligence requires 691 00:33:16,240 --> 00:33:18,460 spanning these dichotomies to get the absolute 692 00:33:18,460 --> 00:33:21,240 best of both structural worlds. You need both. 693 00:33:21,799 --> 00:33:24,619 desperately need the fast, uncontrolled, symmetric 694 00:33:24,619 --> 00:33:27,460 metadata of neural nets to handle messy real 695 00:33:27,460 --> 00:33:30,299 -world perception, integrated flawlessly with 696 00:33:30,299 --> 00:33:33,019 the controlled, asymmetric symbolic reasoning 697 00:33:33,019 --> 00:33:35,559 of traditional architecture to handle logic, 698 00:33:35,680 --> 00:33:37,779 planning, and goal -setting. The map reveals 699 00:33:37,779 --> 00:33:39,700 that everyone in the field actually needs each 700 00:33:39,700 --> 00:33:42,980 other. It transforms a messy, tribal academic 701 00:33:42,980 --> 00:33:46,700 war into a structured, navigable landscape. It 702 00:33:46,700 --> 00:33:50,250 does. But you know, the creation of a map that 703 00:33:50,250 --> 00:33:53,109 elegant naturally leads to a question about the 704 00:33:53,109 --> 00:33:55,690 map maker. Who builds a framework like this? 705 00:33:56,789 --> 00:33:58,589 This brings us to the final section of the book, 706 00:33:58,910 --> 00:34:02,450 which I found incredibly moving. We transition 707 00:34:02,450 --> 00:34:05,890 from the science to the scientist himself. Part 708 00:34:05,890 --> 00:34:09,829 3, The Architect's Memoir. Yeah, Rosenblum shifts 709 00:34:09,829 --> 00:34:13,309 gears to explore his own deeply personal, triphasic 710 00:34:13,309 --> 00:34:16,150 career as an architectural explorer. And it is 711 00:34:16,150 --> 00:34:18,409 fascinating that a career dedicated to the highest 712 00:34:18,409 --> 00:34:20,750 levels of artificial intelligence began with 713 00:34:20,750 --> 00:34:23,570 a piece of plastic from a toy store. The Hasbro 714 00:34:23,570 --> 00:34:26,530 Thinkatron. The Thinkatron. He describes this 715 00:34:26,530 --> 00:34:30,429 toy from the 1960s. It was advertised as a computer 716 00:34:30,429 --> 00:34:33,369 that thinks like a man. It was basically a hollow 717 00:34:33,369 --> 00:34:35,750 plastic shell. You fed it these little punch 718 00:34:35,750 --> 00:34:38,130 cards with multiple choice trivia questions. 719 00:34:38,630 --> 00:34:41,769 A plastic gear would grind inside and it would 720 00:34:41,769 --> 00:34:44,429 spit out the right answer. Very high -tech. Mechanically, 721 00:34:44,469 --> 00:34:46,730 it was just a simple sorting machine. But to 722 00:34:46,730 --> 00:34:49,469 a child, it sparked the illusion of a synthetic 723 00:34:49,469 --> 00:34:52,909 mind. It planted the seed for a life entirely 724 00:34:52,909 --> 00:34:55,429 dedicated to building the real thing. And that 725 00:34:55,429 --> 00:34:57,530 childhood seed grew into what he categorizes 726 00:34:57,530 --> 00:35:02,079 as his grand challenge. his tier one architectural 727 00:35:02,079 --> 00:35:05,599 exploration. For decades, he was at the absolute 728 00:35:05,599 --> 00:35:07,780 forefront of the field. He really was. He worked 729 00:35:07,780 --> 00:35:10,559 alongside legendary figures like Alan Newell 730 00:35:10,559 --> 00:35:12,960 and John Laird at Carnegie Mellon during the 731 00:35:12,960 --> 00:35:16,079 golden era of AI, helping to create the SOAR 732 00:35:16,079 --> 00:35:18,400 cognitive architecture. For listeners who might 733 00:35:18,400 --> 00:35:21,059 not know, SOAR is one of the most famous, longest 734 00:35:21,059 --> 00:35:23,079 lasting cognitive architectures in the history 735 00:35:23,079 --> 00:35:26,480 of the field. It was the gold standard for symbolic 736 00:35:26,480 --> 00:35:29,510 reasoning. It was a monumental achievement. But 737 00:35:29,510 --> 00:35:32,150 Rosenblum's relentless pursuit of that functional 738 00:35:32,150 --> 00:35:34,469 elegance we discussed earlier eventually led 739 00:35:34,469 --> 00:35:36,610 him to step away from the SOAR project to develop 740 00:35:36,610 --> 00:35:39,309 his own entirely new architecture from scratch, 741 00:35:39,769 --> 00:35:43,750 which he called Sigma. Sigma. Sigma was a massive, 742 00:35:43,989 --> 00:35:46,789 ambitious attempt to unify graphical models, 743 00:35:47,269 --> 00:35:49,690 which is the complex math underlying probabilities 744 00:35:49,690 --> 00:35:52,989 in modern neural networks, with traditional cognitive 745 00:35:52,989 --> 00:35:55,900 architecture. He wanted to build an architecture 746 00:35:55,900 --> 00:35:58,739 that inherently spanned those dichotomies we 747 00:35:58,739 --> 00:36:00,900 just mapped out. But his career wasn't just about 748 00:36:00,900 --> 00:36:03,239 building individual systems. He also talks about 749 00:36:03,239 --> 00:36:06,659 his tier two exploration, the common model of 750 00:36:06,659 --> 00:36:09,139 cognition. Right. This was his effort to herd 751 00:36:09,139 --> 00:36:11,599 the cats of the scientific community and find 752 00:36:11,599 --> 00:36:14,440 actual consensus. And buried in this section 753 00:36:14,440 --> 00:36:16,960 is a scientific nugget that absolutely blew my 754 00:36:16,960 --> 00:36:18,659 mind. I know exactly what you're going to say. 755 00:36:18,920 --> 00:36:22,039 The common model, based on intense study of human 756 00:36:22,039 --> 00:36:25,219 neurobiology and various AI models, relies on 757 00:36:25,219 --> 00:36:28,300 a cognitive cycle that operates at roughly 50 758 00:36:28,300 --> 00:36:31,360 milliseconds. The roughly 50 millisecond cycle. 759 00:36:32,099 --> 00:36:35,099 This is a profound biological reality. In human 760 00:36:35,099 --> 00:36:37,719 brains, cognitive processing is not a smooth, 761 00:36:38,000 --> 00:36:40,440 continuous flow of information. It is a discrete, 762 00:36:40,900 --> 00:36:43,219 rhythmic loop. It's like a heartbeat for thoughts. 763 00:36:43,860 --> 00:36:46,489 The cycle of perceiving sensory input processing 764 00:36:46,489 --> 00:36:48,969 it cognitively and initiating a motor action 765 00:36:48,969 --> 00:36:51,949 bottoms out at a biological speed limit of about 766 00:36:51,949 --> 00:36:55,550 50 milliseconds. The common model adopts this 767 00:36:55,550 --> 00:36:57,849 rhythm as a fundamental requirement. It suggests 768 00:36:57,849 --> 00:37:00,530 that human -like intelligence is inherently cyclical. 769 00:37:01,070 --> 00:37:03,369 It's the biological clock speed of human thought. 770 00:37:03,650 --> 00:37:06,909 Exactly. And the engineering implication is massive. 771 00:37:07,510 --> 00:37:09,389 If you want to build a truly human -like synthetic 772 00:37:09,389 --> 00:37:11,869 mind, you likely cannot just build a continuous 773 00:37:11,869 --> 00:37:14,250 feedforward system. You have to engineer that 774 00:37:14,250 --> 00:37:16,809 rhythmic, cyclical, biological architecture into 775 00:37:16,809 --> 00:37:19,550 the silicon. His ambition just keeps scaling 776 00:37:19,550 --> 00:37:23,010 up. In what he calls Tier 4, the relational model, 777 00:37:23,449 --> 00:37:25,969 he actually takes an entire decade away from 778 00:37:25,969 --> 00:37:27,949 building cognitive architectures altogether. 779 00:37:28,230 --> 00:37:30,510 Just steps away entirely. He steps back from 780 00:37:30,510 --> 00:37:33,010 the lab and writes a book attempting to establish 781 00:37:33,010 --> 00:37:36,250 computing as the fourth great scientific domain, 782 00:37:36,630 --> 00:37:38,630 arguing that it deserves to stand right alongside 783 00:37:38,630 --> 00:37:41,449 the physical sciences, the life sciences, and 784 00:37:41,449 --> 00:37:44,579 the social sciences. He was zooming out as far 785 00:37:44,579 --> 00:37:47,260 as humanly possible, trying to define where the 786 00:37:47,260 --> 00:37:49,960 concept of computing even sits within the entire 787 00:37:49,960 --> 00:37:52,280 history of human knowledge. But the part of this 788 00:37:52,280 --> 00:37:54,760 memoir that really stopped me, the part that 789 00:37:54,760 --> 00:37:57,619 grounds all of this dense meta -scientific theory, 790 00:37:58,219 --> 00:38:01,019 was the personal trials. Oh, absolutely. Deep 791 00:38:01,019 --> 00:38:03,340 in the middle of this academic text, Rosenblum 792 00:38:03,340 --> 00:38:07,099 reveals his brutal battle with AML, acute myeloid 793 00:38:07,099 --> 00:38:10,380 leukemia. And this diagnosis didn't just happen... 794 00:38:10,219 --> 00:38:12,599 at any time. It happened right during the height 795 00:38:12,599 --> 00:38:15,300 of the COVID -19 pandemic. The Jackson position 796 00:38:15,300 --> 00:38:19,760 is incredibly stark. He describes the absolute 797 00:38:19,760 --> 00:38:22,500 terrifying isolation of undergoing cancer treatment 798 00:38:22,500 --> 00:38:25,039 during a global lockdown. The hospitals were 799 00:38:25,039 --> 00:38:27,199 closed to visitors. No one could come see him. 800 00:38:27,380 --> 00:38:30,320 He endured grueling bone marrow transplants completely 801 00:38:30,320 --> 00:38:33,239 alone. He faced the incredibly grim survival 802 00:38:33,239 --> 00:38:36,260 statistics associated with AML. And he makes 803 00:38:36,260 --> 00:38:39,659 it through. There's this visceral deeply human 804 00:38:39,659 --> 00:38:41,719 moment where he describes finally being wheeled 805 00:38:41,719 --> 00:38:44,139 out of the hospital, officially discharged, and 806 00:38:44,139 --> 00:38:46,059 the hospital staff lining the halls to give him 807 00:38:46,059 --> 00:38:47,900 a standing ovation. It's a beautiful moment. 808 00:38:48,420 --> 00:38:50,780 To read about that moment in a book dedicated 809 00:38:50,780 --> 00:38:53,099 to the abstract mathematics of intelligence, 810 00:38:53,480 --> 00:38:56,079 it completely recontextualizes the work. It grounds 811 00:38:56,079 --> 00:38:58,579 the entire intellectual pursuit in human fragility. 812 00:38:59,079 --> 00:39:01,480 It does. But it also throws into sharp relief 813 00:39:01,480 --> 00:39:03,659 an incredible paradox that he writes about quite 814 00:39:03,659 --> 00:39:07,019 openly. Here is a man who was stationed at the 815 00:39:07,019 --> 00:39:10,570 absolute epicenter of AI history, Carnegie Mellon 816 00:39:10,570 --> 00:39:13,550 in its prime. Yep. Decades at the University 817 00:39:13,550 --> 00:39:15,570 of Southern California's Information Sciences 818 00:39:15,570 --> 00:39:18,869 Institute, he worked with the titans of the field. 819 00:39:19,590 --> 00:39:22,110 And yet, despite being at the center of the universe, 820 00:39:22,750 --> 00:39:25,090 he describes feeling like a perpetual outsider 821 00:39:25,090 --> 00:39:27,150 throughout his career. How does someone with 822 00:39:27,150 --> 00:39:29,389 that pedigree feel like an outsider? Well, it 823 00:39:29,389 --> 00:39:31,309 ties directly back to everything we've unpacked 824 00:39:31,309 --> 00:39:34,949 today. Remember his methodology, lumping versus 825 00:39:34,949 --> 00:39:37,579 splitting. Right. You have to look at the economic 826 00:39:37,579 --> 00:39:39,940 and academic incentive structures of modern science. 827 00:39:40,639 --> 00:39:44,119 The entire system aggressively rewards splitters. 828 00:39:44,500 --> 00:39:46,760 Researchers are incentivized to carve out tiny, 829 00:39:47,000 --> 00:39:50,260 highly specialized niches, publish 100 iterative 830 00:39:50,260 --> 00:39:54,019 papers on one minor algorithmic tweak, and ruthlessly 831 00:39:54,019 --> 00:39:56,820 chase whatever the latest white -hot, venture 832 00:39:56,820 --> 00:39:59,460 capital -funded trend is. Which right now is 833 00:39:59,460 --> 00:40:02,139 pure deep learning. Exactly. If you are a lumper 834 00:40:02,139 --> 00:40:03,940 in that environment, you naturally stand apart. 835 00:40:04,230 --> 00:40:06,750 Because you are actively trying to unify the 836 00:40:06,750 --> 00:40:09,269 field rather than carve out a profitable slice 837 00:40:09,269 --> 00:40:11,949 of it. Exactly. When you prioritize structural 838 00:40:11,949 --> 00:40:13,869 functional elegance over whatever is currently 839 00:40:13,869 --> 00:40:16,550 winning benchmark tests, you are going to feel 840 00:40:16,550 --> 00:40:19,699 alienated. You are the architect trying to figure 841 00:40:19,699 --> 00:40:22,340 out how to build the entire watch while everyone 842 00:40:22,340 --> 00:40:24,739 around you is getting famous and wealthy for 843 00:40:24,739 --> 00:40:27,539 slightly optimizing the shape of one single gear. 844 00:40:27,820 --> 00:40:30,719 And his brush with mortality, surviving AML, 845 00:40:30,900 --> 00:40:33,179 seemed to act as a profound catalyst for this 846 00:40:33,179 --> 00:40:35,659 memoir. He describes returning to his work with 847 00:40:35,659 --> 00:40:38,480 a new perspective, something he beautifully calls 848 00:40:38,480 --> 00:40:41,820 the freedom of the superannuated. The privilege 849 00:40:41,820 --> 00:40:45,679 of the superannuated. Having survived a highly 850 00:40:45,679 --> 00:40:48,239 lethal cancer and entering the late stages of 851 00:40:48,239 --> 00:40:50,699 a long career, he realized he was finally free. 852 00:40:50,920 --> 00:40:53,440 Free from the rat race. He no longer had to care 853 00:40:53,440 --> 00:40:56,739 about academic trends or the grueling cycle of 854 00:40:56,739 --> 00:40:59,360 securing grant funding or what the deep learning 855 00:40:59,360 --> 00:41:02,619 purists thought of his unified theories. he had 856 00:41:02,619 --> 00:41:05,739 the ultimate intellectual freedom to just sit 857 00:41:05,739 --> 00:41:09,079 down and explore the deep, philosophical, unifying 858 00:41:09,079 --> 00:41:11,280 theories of mind that truly mattered to him. 859 00:41:11,340 --> 00:41:13,860 It's incredibly freeing. This book is the result 860 00:41:13,860 --> 00:41:16,099 of a scientist who realized he had absolutely 861 00:41:16,099 --> 00:41:19,380 nothing left to lose and an entire lifetime of 862 00:41:19,380 --> 00:41:22,099 knowledge left to synthesize. So we have covered 863 00:41:22,099 --> 00:41:25,019 an immense amount of ground today. Let's bring 864 00:41:25,019 --> 00:41:26,780 this journey full circle for you, the listener. 865 00:41:27,599 --> 00:41:29,900 We started by looking at the murky X -ray of 866 00:41:29,900 --> 00:41:32,500 current AI. recognizing that we are building 867 00:41:32,500 --> 00:41:35,420 minds without a blueprint. Right. We zoomed all 868 00:41:35,420 --> 00:41:37,980 the way in to the microscopic definition of a 869 00:41:37,980 --> 00:41:41,400 single symbol, redefining it as a fluid hybrid 870 00:41:41,400 --> 00:41:44,440 placeholder. We defined the massive goal of full 871 00:41:44,440 --> 00:41:46,920 -spectrum intelligence, recognizing that today's 872 00:41:46,920 --> 00:41:49,960 chatbots, while possessing incredible immediate 873 00:41:49,960 --> 00:41:53,320 intelligence, are architecturally mindless because 874 00:41:53,320 --> 00:41:55,880 they lack cumulative learning and architectural 875 00:41:55,880 --> 00:41:59,059 surprise. And then we zoomed out to a three -dimensional 876 00:41:59,059 --> 00:42:02,000 map that plotted the entire history of AI across 877 00:42:02,000 --> 00:42:05,099 three underlying physical dichotomies. And we 878 00:42:05,099 --> 00:42:07,619 anchored all of that theory in the deeply human 879 00:42:07,619 --> 00:42:10,280 story of an architectural explorer surviving 880 00:42:10,280 --> 00:42:12,960 cancer to write a unified theory of his life's 881 00:42:12,960 --> 00:42:15,199 work. The ultimate goal here is that the listener 882 00:42:15,199 --> 00:42:17,539 is now armed with a robust mental framework. 883 00:42:18,000 --> 00:42:20,659 The next time a tech CEO steps onto a stage and 884 00:42:20,659 --> 00:42:23,800 announces a new AI model that is supposedly approaching 885 00:42:23,800 --> 00:42:26,630 human intelligence, You don't have to just accept 886 00:42:26,630 --> 00:42:29,510 the hype. You have Rosenblum's toolkit. You can 887 00:42:29,510 --> 00:42:32,510 actively ask the structural questions. Is this 888 00:42:32,510 --> 00:42:35,730 new system demonstrating true functional elegance? 889 00:42:35,889 --> 00:42:38,809 Or is it just another display of massive computing 890 00:42:38,809 --> 00:42:41,250 exuberance? Does it possess the architecture 891 00:42:41,250 --> 00:42:44,429 for cumulative lifelong learning? Or is it just 892 00:42:44,429 --> 00:42:46,469 performing incredibly fast, immediate parlor 893 00:42:46,469 --> 00:42:49,070 tricks? Does it successfully span the dichotomies 894 00:42:49,070 --> 00:42:52,190 of intuition and deliberation? Or is it stuck 895 00:42:52,190 --> 00:42:54,679 in one region of the map? You essentially have 896 00:42:54,679 --> 00:42:57,119 the x -ray machine now. You can look past the 897 00:42:57,119 --> 00:42:59,300 impressive conversational output and see the 898 00:42:59,300 --> 00:43:01,760 structural architecture, or the lack thereof, 899 00:43:01,800 --> 00:43:03,920 beneath it. The structural perspective changes 900 00:43:03,920 --> 00:43:06,480 everything. But before we let you go, there is 901 00:43:06,480 --> 00:43:09,199 one final highly provocative thought that Rosenblum 902 00:43:09,199 --> 00:43:12,139 leaves us with, almost as a brief philosophical 903 00:43:12,139 --> 00:43:15,219 detour at the very end of his exploration. And 904 00:43:15,219 --> 00:43:17,880 it has to do with the concept of free will. Oh, 905 00:43:17,920 --> 00:43:20,429 this is fascinating. He directly addresses a 906 00:43:20,429 --> 00:43:22,809 very popular argument made by a prominent Stanford 907 00:43:22,809 --> 00:43:26,389 neurobiologist, Robert Sapolsky. Right. Sapolsky 908 00:43:26,389 --> 00:43:28,530 recently published a comprehensive book arguing 909 00:43:28,530 --> 00:43:31,710 strongly, based on biology and physics, against 910 00:43:31,710 --> 00:43:35,449 the existence of free will. Sapolsky's argument 911 00:43:35,449 --> 00:43:38,010 is that humans are ultimately biological machines. 912 00:43:38,750 --> 00:43:41,090 Every action we take is entirely predetermined 913 00:43:41,090 --> 00:43:43,949 by a complex chain of neurochemistry, genetics, 914 00:43:44,250 --> 00:43:47,230 trauma, and environmental stimuli. So it's all 915 00:43:47,230 --> 00:43:49,800 mapped out already. Therefore, the concept that 916 00:43:49,800 --> 00:43:52,780 we possess independent free will is a complete 917 00:43:52,780 --> 00:43:55,880 neurological illusion. It's a very common mathematically 918 00:43:55,880 --> 00:43:59,159 sound, if somewhat bleak, scientific perspective. 919 00:43:59,880 --> 00:44:02,579 We are all just complex dominoes falling in a 920 00:44:02,579 --> 00:44:04,380 chain reaction that started at the Big Bang. 921 00:44:04,699 --> 00:44:06,960 But Rosenblum tackles this argument not just 922 00:44:06,960 --> 00:44:10,000 as a biologist looking at chemistry, but from 923 00:44:10,000 --> 00:44:12,880 the perspective of a cognitive architect who 924 00:44:12,880 --> 00:44:15,679 has spent 50 years trying to engineer a mind 925 00:44:15,679 --> 00:44:20,199 from scratch. and he introduces a brilliant structural 926 00:44:20,199 --> 00:44:23,340 twist to the debate. What's the twist? He essentially 927 00:44:23,340 --> 00:44:25,639 argues that from an architectural standpoint, 928 00:44:26,119 --> 00:44:28,659 it absolutely does not matter if the universe 929 00:44:28,659 --> 00:44:31,739 is strictly deterministic. Wait, it doesn't matter 930 00:44:31,739 --> 00:44:35,199 if free will is real or not? Functionally, mechanically, 931 00:44:35,380 --> 00:44:38,380 no. Because if you want a mind to operate at 932 00:44:38,380 --> 00:44:41,079 level three, full spectrum intelligence, if you 933 00:44:41,079 --> 00:44:43,719 want a system to function, adapt to unexpected 934 00:44:43,719 --> 00:44:47,420 surprises, evaluate complex moral choices, and 935 00:44:47,420 --> 00:44:49,719 initiate deliberate controlled action across 936 00:44:49,719 --> 00:44:52,829 a lifetime. That architecture must possess the 937 00:44:52,829 --> 00:44:55,630 internal belief that its goals matter and that 938 00:44:55,630 --> 00:44:58,329 it has the active agency to achieve them. So 939 00:44:58,329 --> 00:45:00,989 the system literally has to believe it possesses 940 00:45:00,989 --> 00:45:03,849 free will simply in order for the cognitive software 941 00:45:03,849 --> 00:45:06,590 to successfully run? That is the architectural 942 00:45:06,590 --> 00:45:10,289 requirement. Whether you are a human being trying 943 00:45:10,289 --> 00:45:12,829 to navigate the complexities of society, or a 944 00:45:12,829 --> 00:45:15,289 highly advanced synthetic AI trying to achieve 945 00:45:15,289 --> 00:45:18,969 general intelligence. The illusion of free will 946 00:45:18,969 --> 00:45:21,170 is not just a psychological quirk or a comforting 947 00:45:21,170 --> 00:45:24,769 lie we tell ourselves. It is a strict structural 948 00:45:24,769 --> 00:45:27,630 necessity. Without the internal representation 949 00:45:27,630 --> 00:45:30,230 of agency, without the system believing that 950 00:45:30,230 --> 00:45:33,750 its choices actively alter the future, the architecture 951 00:45:33,750 --> 00:45:36,449 cannot evaluate options and initiate controlled 952 00:45:36,449 --> 00:45:40,460 action. Wow. The underlying engine requires the 953 00:45:40,460 --> 00:45:43,500 belief in choices to function, completely regardless 954 00:45:43,500 --> 00:45:45,360 of whether the laws of physics say those choices 955 00:45:45,360 --> 00:45:49,079 were predetermined 14 billion years ago. That 956 00:45:49,079 --> 00:45:52,480 is a staggering thought to end on. The illusion 957 00:45:52,480 --> 00:45:55,739 of free will is an unavoidable structural requirement 958 00:45:55,739 --> 00:45:58,199 for the existence of intelligence. The architecture 959 00:45:58,199 --> 00:46:00,599 demands it. That is something for you to mull 960 00:46:00,599 --> 00:46:02,659 over the next time you make a conscious choice 961 00:46:02,659 --> 00:46:05,300 today. Whether your decision to listen to this 962 00:46:05,300 --> 00:46:08,090 deep dive was entirely predetermined by a chain 963 00:46:08,090 --> 00:46:11,869 of cosmic dominoes, or whether you truly independently 964 00:46:11,869 --> 00:46:14,789 authored the decision, your cognitive architecture 965 00:46:14,789 --> 00:46:17,010 required you to fully believe in that choice 966 00:46:17,010 --> 00:46:19,469 either way, just so you could get out of bed 967 00:46:19,469 --> 00:46:22,369 this morning and hit play. Keep exploring the 968 00:46:22,369 --> 00:46:24,670 incredible architecture of your own mind, and 969 00:46:24,670 --> 00:46:25,929 we'll see you on the next deep dive.