WEBVTT 1 00:00:01.199 --> 00:00:06.200 Welcome to the Sentient Code, where intelligence is engineered, autonomy 2 00:00:06.280 --> 00:00:10.439 is emerging, and a line between human and machine grows thinner. 3 00:00:10.800 --> 00:00:15.359 Each episode, we decode the algorithms, explore the robotics, and 4 00:00:15.439 --> 00:00:21.839 examine the ideas shaping the future of artificial minds. 5 00:00:23.839 --> 00:00:25.480 So I want you to just take a second and 6 00:00:25.600 --> 00:00:27.920 look around the room you're in right now, or you 7 00:00:27.960 --> 00:00:29.879 know wherever you happen to be listening to this today. Yeah, 8 00:00:29.920 --> 00:00:32.119 just do a quick scan of your environment, right, and 9 00:00:32.159 --> 00:00:35.000 I want you to try and find a standard, maybe 10 00:00:35.039 --> 00:00:37.280 somewhat dim, incandescent. 11 00:00:36.679 --> 00:00:40.359 Light bulb, like the kind in an old desk clamp maybe. 12 00:00:40.000 --> 00:00:43.159 Exactly, or even just the little bulb inside your refrigerator. 13 00:00:44.119 --> 00:00:47.640 That specific bulb, it takes about twenty watts of power 14 00:00:47.960 --> 00:00:49.359 just to maintain its glow. 15 00:00:49.280 --> 00:00:51.560 Which is I mean, it's barely enough light to read. 16 00:00:51.439 --> 00:00:54.000 A book by, right, it's practically nothing. So I want 17 00:00:54.000 --> 00:00:57.439 you to hold on to that image, that tiny twenty 18 00:00:57.439 --> 00:00:58.920 watt energy budget. 19 00:00:58.719 --> 00:01:01.200 Because we're going to use that as the bits line 20 00:01:01.280 --> 00:01:02.640 for something pretty. 21 00:01:02.320 --> 00:01:05.920 Wild, completely mind bending. Actually yeah, because that exact same 22 00:01:05.920 --> 00:01:08.959 amount of energy, those twenty watts, is what your biological 23 00:01:09.040 --> 00:01:12.760 brain is using right this very second. To do well 24 00:01:13.000 --> 00:01:13.760 literally everything. 25 00:01:13.920 --> 00:01:17.079 It really forces a complete recalibration of how we view 26 00:01:17.120 --> 00:01:18.599 our own biology. 27 00:01:18.040 --> 00:01:20.519 Doesn't it. It totally does we just walk. 28 00:01:20.400 --> 00:01:24.280 Around completely oblivious to the sheer, I mean, the thermodynamic 29 00:01:24.359 --> 00:01:26.519 miracle happening inside our skulls every day. 30 00:01:26.799 --> 00:01:29.680 Yeah, think about what you, as the listener, are doing 31 00:01:29.719 --> 00:01:33.280 on just twenty wants Right now, you are processing the 32 00:01:33.319 --> 00:01:35.120 auditory signals of my voice. 33 00:01:35.239 --> 00:01:38.920 You're converting acoustic waves into electrical impulses. 34 00:01:38.319 --> 00:01:41.560 Exactly, and then you're parsing individual words out of that 35 00:01:41.719 --> 00:01:43.319 continuous stream of sound. 36 00:01:43.519 --> 00:01:45.879 And if you happen to be, say, walking down the 37 00:01:45.920 --> 00:01:47.879 street while listening, it's even crazier. 38 00:01:47.959 --> 00:01:52.239 Oh yeah, you're instantly recognizing familiar faces. You're navigating this 39 00:01:52.359 --> 00:01:55.719 complex three D environment, keeping your balance against. 40 00:01:55.359 --> 00:01:58.040 Gravity, regulating your core body temperature. 41 00:01:58.480 --> 00:02:01.400 Right, and you can learn a complete, letely new abstract 42 00:02:01.439 --> 00:02:05.040 concept from a single example and instantly integrate it with 43 00:02:05.319 --> 00:02:06.640 decades of stored memories. 44 00:02:06.760 --> 00:02:10.360 All of that high level computational work is happening simultaneously, And. 45 00:02:10.319 --> 00:02:13.520 While you're doing all of that, that identical twenty watt 46 00:02:13.520 --> 00:02:20.120 budget is sustaining the profound, completely unresolved mystery of human consciousness. Itself. 47 00:02:20.159 --> 00:02:22.319 It's just staggering all for the energy it takes to 48 00:02:22.360 --> 00:02:23.439 dimly light a closet. 49 00:02:23.639 --> 00:02:26.360 And you know the contrast becomes almost absurd when we 50 00:02:26.400 --> 00:02:28.439 look at our current technological landscape. 51 00:02:28.479 --> 00:02:29.919 Oh absolutely, let's talk about that. 52 00:02:30.240 --> 00:02:33.919 Like, consider the state of the art artificial intelligence models today. 53 00:02:34.639 --> 00:02:39.000 The systems that are performing complex pattern recognition or generating 54 00:02:39.039 --> 00:02:41.039 photorealistic images. 55 00:02:40.800 --> 00:02:42.800 Or simulating human likesuage right. 56 00:02:42.639 --> 00:02:46.840 To achieve even a tiny fraction of human cognitive capabilities, 57 00:02:47.319 --> 00:02:52.919 and usually only within incredibly narrow domains. Current AI requires 58 00:02:53.039 --> 00:02:55.639 physical infrastructure on a planetary scale. 59 00:02:55.680 --> 00:02:58.280 We are definitely not talking about twenty watts anymore. 60 00:02:58.520 --> 00:03:02.439 No, not at all. We're talking about sprawling, warehouse sized 61 00:03:02.520 --> 00:03:06.759 data centers, just packed floor to ceiling with specialized hardware. 62 00:03:06.840 --> 00:03:10.520 I mean, these facilities need dedicated industrial cooling towers just 63 00:03:10.599 --> 00:03:12.680 to prevent the silicon from literally melting. 64 00:03:12.840 --> 00:03:15.560 Exactly. They consume millions of watts. 65 00:03:15.319 --> 00:03:18.039 Millions of wants for a single model. It's basically the 66 00:03:18.159 --> 00:03:20.759 energy equivalent of a small dedicated power plan. 67 00:03:20.919 --> 00:03:24.000 Just to run a specialized software program that can write 68 00:03:24.000 --> 00:03:26.919 a passable email or identify a stop sign and a photograph. 69 00:03:27.120 --> 00:03:30.479 So the ratio of computational capability to energy consumption in 70 00:03:30.520 --> 00:03:33.159 the human brain isn't just like a little bit better 71 00:03:33.159 --> 00:03:33.479 than our. 72 00:03:33.439 --> 00:03:36.759 Ai, No, it is superior by multiple orders of magnitude. 73 00:03:36.919 --> 00:03:39.759 And the most critical realization here is that this isn't 74 00:03:39.800 --> 00:03:42.319 some temporary engineering hurdle, is it not? 75 00:03:42.400 --> 00:03:45.039 At all? This isn't a problem where we can simply 76 00:03:45.319 --> 00:03:47.919 wait for the next generation of silicon microchips and just 77 00:03:47.960 --> 00:03:50.719 assume the efficiency gap will organically close. 78 00:03:50.560 --> 00:03:54.039 Because the foundational architecture of how we build computers is 79 00:03:54.360 --> 00:03:58.599 fundamentally at odds with how biological brains operate physically at odds. Yeah, 80 00:03:58.759 --> 00:04:02.599 so we have this massive multi order of magnitude disparity 81 00:04:02.599 --> 00:04:07.080 and efficiency, and the traditional roadmap of computer science is 82 00:04:07.199 --> 00:04:09.400 essentially a dead end for solving it. 83 00:04:09.520 --> 00:04:12.039 Which brings us to the terrain we're exploring today. 84 00:04:12.120 --> 00:04:16.399 Yes, we are looking at a radical, incredibly ambitious field 85 00:04:16.399 --> 00:04:18.839 of engineering called neuromorphic computing. 86 00:04:19.480 --> 00:04:22.519 The name itself derives from the Greek roots for nerve 87 00:04:22.680 --> 00:04:23.319 and form. 88 00:04:23.600 --> 00:04:25.480 I love that. And just to be clear, this is 89 00:04:25.519 --> 00:04:28.079 not about writing clever software code to run on the 90 00:04:28.120 --> 00:04:30.639 exact same hardware we've been using since the nineteen eighties. 91 00:04:30.720 --> 00:04:34.920 No, this is the engineering quest to build actual physical hardware, 92 00:04:35.000 --> 00:04:39.279 new types of silicon chips that fundamentally think, physically operate, 93 00:04:39.639 --> 00:04:43.079 and are structurally organized exactly like a biological brain. 94 00:04:43.439 --> 00:04:47.079 It requires a complete paradigm shift, It really does, because 95 00:04:47.120 --> 00:04:51.480 the vast majority of artificial intelligence today operates by simulating 96 00:04:51.519 --> 00:04:52.199 a neural. 97 00:04:51.959 --> 00:04:57.319 Network in software, but then executing that simulation on conventional, 98 00:04:57.839 --> 00:04:58.959 non neural hardware. 99 00:04:59.079 --> 00:05:01.079 Right, It's just a similar running on a normal. 100 00:05:00.839 --> 00:05:06.439 Computer exactly, but neuromorphic engineering abandons the simulation. The goal 101 00:05:06.600 --> 00:05:10.519 is to instantiate the physical operating principles of biological neural 102 00:05:10.560 --> 00:05:14.399 computation directly into the atomic structure of the machine itself. 103 00:05:14.560 --> 00:05:17.439 Okay, So to understand why we need to completely reinvent 104 00:05:17.519 --> 00:05:20.439 the physical computer from the ground up, we first need 105 00:05:20.439 --> 00:05:23.240 to talk about the fatal flaw baked into the computers you. 106 00:05:23.199 --> 00:05:25.920 And I are already using, right, the von Neuman architecture. 107 00:05:26.000 --> 00:05:28.160 Yeah, if you're listening to this on the smartphone, that 108 00:05:28.279 --> 00:05:31.839 incredibly advanced device is constrained by a design choice made 109 00:05:31.839 --> 00:05:33.000 back in the nineteen forties. 110 00:05:33.079 --> 00:05:34.879 We have to go back to the mathematician John von 111 00:05:34.959 --> 00:05:35.879 Neuman set. 112 00:05:35.639 --> 00:05:37.839 The stage for us. What was happening in the forties. 113 00:05:37.920 --> 00:05:41.959 Well, during the era of those massive room sized vacuum 114 00:05:42.000 --> 00:05:46.519 tube computers like the Nis, von Neumann formalized a blueprint 115 00:05:46.560 --> 00:05:48.439 for computer architecture. 116 00:05:48.160 --> 00:05:52.120 And that blueprint became the standard for virtually every single 117 00:05:52.120 --> 00:05:54.040 digital device we manufacture today. 118 00:05:53.920 --> 00:05:57.360 Right exactly, and the defining characteristic of this von Noyman 119 00:05:57.480 --> 00:06:00.959 architecture is a strict physical set of duties. 120 00:06:01.040 --> 00:06:02.120 Okay, break that down for me. 121 00:06:02.199 --> 00:06:06.040 You have the processor which performs the actual mathematical computation 122 00:06:06.120 --> 00:06:09.480 and logic, and then physically separated from it, you have 123 00:06:09.560 --> 00:06:12.959 the memory unit which stores the data and the instructions. 124 00:06:13.079 --> 00:06:16.079 So they basically live in two completely different neighborhoods on the. 125 00:06:16.079 --> 00:06:19.839 Motherboard they do, and they're connected by a communication pathway 126 00:06:19.879 --> 00:06:20.720 known as a bus. 127 00:06:20.800 --> 00:06:22.040 A bus got it. 128 00:06:21.920 --> 00:06:26.040 Because they're physically separated. Every single time the processor needs 129 00:06:26.040 --> 00:06:29.639 to execute a task, whether that's adding two numbers or 130 00:06:29.720 --> 00:06:31.920 changing the color of a pixel on your screen, it 131 00:06:32.000 --> 00:06:32.879 can't do it alone. 132 00:06:33.040 --> 00:06:34.680 It has to ask for the data, right. 133 00:06:34.800 --> 00:06:37.600 It must send an electrical request across the bus to 134 00:06:37.639 --> 00:06:39.839 the memory. It has to wait for the data to 135 00:06:39.879 --> 00:06:43.079 be retrieved. The data is pushed back across the bus 136 00:06:43.079 --> 00:06:46.720 to the processor, the operation is performed, and then the 137 00:06:46.800 --> 00:06:48.959 result usually has to be sent back across the bus 138 00:06:48.959 --> 00:06:50.199 to be stored in memory. Again. 139 00:06:50.480 --> 00:06:55.879 Wow, if you really visualize that, it's just endless shuttling back. 140 00:06:55.680 --> 00:06:57.199 And forth, constant shuttling. 141 00:06:57.360 --> 00:07:01.519 Let me train an analogy here. Imagine a chef working 142 00:07:01.560 --> 00:07:05.120 in a high end restaurant kitchen. But there's a massive 143 00:07:05.120 --> 00:07:09.120 design flaw. The pantry isn't in the kitchen. 144 00:07:09.279 --> 00:07:10.199 Okay, where is it. 145 00:07:10.439 --> 00:07:12.279 The pantry is three blocks down the street. 146 00:07:12.480 --> 00:07:14.279 Oh, that sounds awful, right. 147 00:07:14.560 --> 00:07:17.360 So the chef is standing at the stove, water is boiling, 148 00:07:17.959 --> 00:07:20.240 and they realize they need a pinch of salt, so 149 00:07:20.279 --> 00:07:22.079 they have to leave the kitchen. They have to stop 150 00:07:22.079 --> 00:07:24.839 what they're doing, run out the door, sprint three blocks 151 00:07:24.879 --> 00:07:27.600 down the street to the pantry, grab the pitch of salt, 152 00:07:27.759 --> 00:07:30.120 run all the way back to the kitchen and sprinkle 153 00:07:30.160 --> 00:07:30.759 it in the pot. 154 00:07:31.160 --> 00:07:32.920 And then they probably need something else. 155 00:07:32.800 --> 00:07:35.160 Exactly Then they realize they need a chopped on you 156 00:07:35.720 --> 00:07:38.399 Boom out the door again, three blocks down the street, 157 00:07:38.800 --> 00:07:43.399 grab the onion, sprint back over and over millions of 158 00:07:43.439 --> 00:07:45.240 times a day, so by the end of the night. 159 00:07:45.439 --> 00:07:47.639 By the end of the night, the chef isn't tired 160 00:07:47.639 --> 00:07:50.079 from the actual act of cooking. The chef is completely 161 00:07:50.120 --> 00:07:51.240 exhausted from commuting. 162 00:07:51.399 --> 00:07:53.360 That is a perfect way to picture it, and the 163 00:07:53.399 --> 00:07:56.879 physical consequence of that commute is the crux of the problem. 164 00:07:57.079 --> 00:07:58.879 The von Neumann bottleneck exactly. 165 00:07:59.399 --> 00:08:02.600 Historically, in computer science it was viewed as a speed limit. 166 00:08:03.120 --> 00:08:06.560 The processor is incredibly fast, but it spends most of 167 00:08:06.560 --> 00:08:09.720 its time sitting idle waiting for data to travel back 168 00:08:09.800 --> 00:08:10.800 and forth over the bus. 169 00:08:10.959 --> 00:08:13.560 But nowadays it's not just about speed, is it No. 170 00:08:14.000 --> 00:08:17.639 In the context of modern artificial intelligence, the bottleneck is 171 00:08:17.680 --> 00:08:19.920 a massive energy crisis. 172 00:08:19.560 --> 00:08:22.959 Because pushing electrical signals back and forth across a physical 173 00:08:23.000 --> 00:08:27.959 copper wire billions of times a second that generates heat. 174 00:08:27.920 --> 00:08:30.800 Massive amounts of heat. Due to the physical capacitance of the. 175 00:08:30.759 --> 00:08:34.240 Wire, it actually takes physical energy to push electrons down 176 00:08:34.279 --> 00:08:34.679 a wire. 177 00:08:34.919 --> 00:08:38.360 It takes enormous energy. When we run complex AI neural 178 00:08:38.399 --> 00:08:42.519 networks on conventional hardware, we're shuffling massive matrices of data 179 00:08:42.639 --> 00:08:45.320 millions of parameters back and forth continuously. 180 00:08:45.519 --> 00:08:48.440 So the commute is basically bankrupting our energy budget. 181 00:08:48.480 --> 00:08:52.279 Precisely, the sheer thermodynamic cost of moving the data is 182 00:08:52.399 --> 00:08:55.679 astronomically higher than the energy cost of the actual mathematical 183 00:08:55.720 --> 00:08:57.039 operations being performed. 184 00:08:57.200 --> 00:08:59.720 We are burning our energy budget on the commute, not 185 00:08:59.759 --> 00:09:01.080 the computation. 186 00:09:00.840 --> 00:09:03.159 Which naturally leads us to ask a pretty important question. 187 00:09:03.600 --> 00:09:06.759 Right, If that's how our current devices arrange their kitchen, 188 00:09:07.320 --> 00:09:11.080 and it's clearly a disaster for energy efficiency, how does 189 00:09:11.080 --> 00:09:12.879 the human brain around its kitchen? 190 00:09:13.000 --> 00:09:15.120 How is your brain managing to do all of this 191 00:09:15.320 --> 00:09:19.559 high level processing on just twenty watts without completely melting down? 192 00:09:19.639 --> 00:09:21.080 Yeah? What's the biological solution? 193 00:09:21.440 --> 00:09:24.759 The biological solution entirely eliminates the commute? 194 00:09:24.879 --> 00:09:25.919 Wait, entirely. 195 00:09:26.080 --> 00:09:29.720 Yes. In the human brain, computation and memory are colocated. 196 00:09:29.919 --> 00:09:32.320 They occupy the exact same physical space. 197 00:09:32.440 --> 00:09:35.000 Okay, I need to understand how that actually works physically. 198 00:09:35.240 --> 00:09:38.639 Well, the brain contains roughly eighty six billion neurons, and 199 00:09:38.679 --> 00:09:43.120 they communicate with each other through microscopic junctions called synapses. 200 00:09:42.799 --> 00:09:45.039 And we have what trillions of those. 201 00:09:44.919 --> 00:09:48.200 Trillions of synactic connections. Yeah. In the biological paradigm, the 202 00:09:48.200 --> 00:09:51.360 synapse operate simultaneously as the processor and the hard drive. 203 00:09:51.559 --> 00:09:53.840 So the wiring itself is the memory. 204 00:09:53.799 --> 00:09:57.120 The connection itself is the physical medium through which the 205 00:09:57.159 --> 00:10:00.840 computation occurs, and it is also the physical substrate where 206 00:10:00.919 --> 00:10:05.120 learned information is stored. That is wild when you learn 207 00:10:05.200 --> 00:10:08.799 something new, say you're practicing a new language or memorizing 208 00:10:08.840 --> 00:10:12.240 the layout of a new neighborhood. The physical structure of 209 00:10:12.279 --> 00:10:14.440 your brain actually changes like. 210 00:10:14.440 --> 00:10:16.240 It physically alters. Its shape. 211 00:10:16.519 --> 00:10:19.720 The synaptic weight, which determines the strength of the electrical 212 00:10:19.799 --> 00:10:23.320 connection between specific neurons, physically alters. 213 00:10:23.559 --> 00:10:26.559 See if you open up a traditional computer, the memory 214 00:10:26.679 --> 00:10:30.120 is just a file saved in a discrete sector of 215 00:10:30.120 --> 00:10:33.720 a silicon ship. But in the brain, the memory is 216 00:10:33.759 --> 00:10:38.240 the physical shape, density, and chemical strength of the literal wiring. 217 00:10:38.480 --> 00:10:42.600 Computation and memory changed together. They are intrinsically linked in 218 00:10:42.600 --> 00:10:43.720 the same physical space. 219 00:10:43.840 --> 00:10:45.840 So when I see something what happens. 220 00:10:45.759 --> 00:10:48.720 When an electrical signal passes through a network of neurons 221 00:10:48.759 --> 00:10:51.799 in your visual cortex, It doesn't have to pause and 222 00:10:51.840 --> 00:10:54.759 fetch the memory of how to process a straight line 223 00:10:54.799 --> 00:10:56.200 from a different lobe of the brain. 224 00:10:56.440 --> 00:10:58.200 The instructions are just there. 225 00:10:58.360 --> 00:11:01.879 The processing instructions are built directly into the physical pathway 226 00:11:01.919 --> 00:11:04.840 the signal is currently traveling through. There is no shuttling 227 00:11:04.879 --> 00:11:05.240 of data. 228 00:11:05.320 --> 00:11:07.000 There's no bus, no bus at all. 229 00:11:07.279 --> 00:11:09.639 To use your analogy, the kitchen and the pantry are 230 00:11:09.639 --> 00:11:12.240 perfectly integrated at a microscopic level. 231 00:11:12.159 --> 00:11:17.000 And replicating this exact physical colocation in silicon hardware is 232 00:11:17.039 --> 00:11:20.360 the foundational premise of neuromorphic engineering exactly. 233 00:11:20.519 --> 00:11:21.399 That's the core of it. 234 00:11:21.600 --> 00:11:23.919 But you know, the physical layout of the von Normann 235 00:11:23.919 --> 00:11:26.639 bottleneck is really only half the problem, isn't it. 236 00:11:27.039 --> 00:11:27.679 That's true. 237 00:11:27.799 --> 00:11:30.480 It's not just that traditional computers separate the kitchen in 238 00:11:30.519 --> 00:11:34.559 the pantry, it's the very language they use to communicate internally, 239 00:11:35.200 --> 00:11:38.559 the fundamental way we treat the materials we build computers 240 00:11:38.559 --> 00:11:38.840 out of. 241 00:11:39.000 --> 00:11:40.960 Right, we need to look back at the origins of 242 00:11:41.120 --> 00:11:44.080 neuromorphic hardware to really understand this, which. 243 00:11:43.960 --> 00:11:46.480 Shakes us to Carver Mead right at Caltech. 244 00:11:46.679 --> 00:11:50.519 Yes, Carver Mead at the California Institute of Technology, in 245 00:11:50.559 --> 00:11:54.240 a landmark nineteen ninety paper, Mead, who was already a 246 00:11:54.320 --> 00:11:59.279 highly respected pioneer in integrated circuit design, actually coined the 247 00:11:59.360 --> 00:12:00.320 term neural morph. 248 00:12:00.559 --> 00:12:02.279 Okay, what was his big breakthrough? 249 00:12:02.519 --> 00:12:06.399 His most profound realization wasn't just about moving memory closer 250 00:12:06.440 --> 00:12:09.080 to the processor. It was about the fundamental physics of 251 00:12:09.120 --> 00:12:10.480 how we utilize silicon. 252 00:12:10.799 --> 00:12:14.120 Because right now the global tech industry uses silicon to 253 00:12:14.159 --> 00:12:17.919 mass produce transistors, and we treat those transistors in a 254 00:12:18.039 --> 00:12:20.080 highly rigid binary way. 255 00:12:20.279 --> 00:12:22.720 Yes, we treat them like tiny light switches. 256 00:12:22.840 --> 00:12:26.200 They're forced to be either entirely on, representing a one, 257 00:12:26.519 --> 00:12:28.440 or entirely off, representing a zero. 258 00:12:28.720 --> 00:12:33.799 Conventional computing is entirely digital and binary. It operates using discrete, 259 00:12:34.159 --> 00:12:36.679 strictly controlled, clocked operations. 260 00:12:36.799 --> 00:12:39.799 Everything is forced into that binary state of one or zero. 261 00:12:40.120 --> 00:12:44.440 Furthermore, millions or billions of times a second. A rigid 262 00:12:44.480 --> 00:12:49.440 internal clock forces the entire system to step forward synchronously. 263 00:12:48.840 --> 00:12:50.360 Just marching to a beat exactly. 264 00:12:50.639 --> 00:12:54.480 But med looking closely at biology, recognized that biological brains 265 00:12:54.480 --> 00:12:56.600 do not operate like binary calculators. 266 00:12:56.639 --> 00:12:57.960 They don't use ones and zeros. 267 00:12:58.200 --> 00:13:01.639 No, the brain operates in a fundament only analog, continuous 268 00:13:01.679 --> 00:13:02.399 time mode. 269 00:13:02.519 --> 00:13:05.279 So carver Meede looked at these silicon transistors which we're 270 00:13:05.360 --> 00:13:08.159 using to build these rigid digital calculators, and said, what 271 00:13:08.240 --> 00:13:10.279 if we stop forcing them to be strict on and 272 00:13:10.279 --> 00:13:10.879 off switches. 273 00:13:11.200 --> 00:13:14.519 He proposed operating the transistors in what physicists call the 274 00:13:14.639 --> 00:13:15.799 sub threshold regime. 275 00:13:16.000 --> 00:13:19.360 The sub threshold regime, what does that mean? In plain English? 276 00:13:19.639 --> 00:13:22.320 In standard digital logic, a transistor is hit with a 277 00:13:22.399 --> 00:13:26.159 relatively high voltage to snap it cleanly on, allowing current 278 00:13:26.200 --> 00:13:29.720 to flow freely, or the voltage is dropped to snap 279 00:13:29.720 --> 00:13:30.519 it cleanly off. 280 00:13:30.559 --> 00:13:32.240 It's a brute force approach. 281 00:13:32.039 --> 00:13:34.759 It is. But if you apply a very low voltage 282 00:13:34.799 --> 00:13:39.159 below that clean switching threshold, the transistor doesn't just go dead. 283 00:13:39.320 --> 00:13:40.840 Its behavior changes entirely. 284 00:13:41.159 --> 00:13:42.240 It does what happens. 285 00:13:42.480 --> 00:13:46.799 It responds in a graded, continuous analog way. The tiny 286 00:13:46.879 --> 00:13:50.279 trickle of leakage current that digital engineers usually try to 287 00:13:50.320 --> 00:13:52.600 eliminate the stuff they see as a bug, exactly the 288 00:13:52.639 --> 00:13:56.000 bug becomes the core feature. Mead realized that the analog 289 00:13:56.120 --> 00:14:00.679 physics of this sub threshold silicon, how electrical charge slowly accumulates, 290 00:14:00.759 --> 00:14:03.120 how capacitance builds, how ions. 291 00:14:02.759 --> 00:14:04.200 Diffuse, It acts like biology. 292 00:14:04.399 --> 00:14:08.159 It could naturally mimic the continuous physical dynamics of biological 293 00:14:08.200 --> 00:14:10.159 cell membranes and ion channels. 294 00:14:10.240 --> 00:14:11.879 So he wanted to use the raw physics of the 295 00:14:11.919 --> 00:14:15.320 material to just do the math organically, rather than forcing 296 00:14:15.320 --> 00:14:18.679 the material to act like a rigid mathematical abocus. 297 00:14:18.759 --> 00:14:21.159 That's a great way to put it. By utilizing the 298 00:14:21.320 --> 00:14:26.120 natural analog physics of the substrate, you bypass the massive 299 00:14:26.159 --> 00:14:30.559 thermo dynamic overhead of encoding every single piece of sensory 300 00:14:30.639 --> 00:14:35.600 information into long, complex strings of binary digits. 301 00:14:35.240 --> 00:14:37.200 And shoving them through that serial bottleneck we. 302 00:14:37.159 --> 00:14:41.159 Talked about, right, But this analog shift fundamentally changes how 303 00:14:41.279 --> 00:14:44.120 information is transmitted across the network. 304 00:14:44.399 --> 00:14:45.240 How so well? 305 00:14:45.360 --> 00:14:50.480 Modern artificial neural networks in software exchange precise, continuous numerical 306 00:14:50.559 --> 00:14:53.960 values floating point numbers like point seven four to three 307 00:14:54.080 --> 00:14:57.600 fro sheats of numbers basically yeah, But biological neurons do 308 00:14:57.679 --> 00:14:59.519 not exchange spreadsheets of numbers. 309 00:14:59.559 --> 00:15:00.480 They speak in spikes. 310 00:15:00.679 --> 00:15:04.159 They communicate through action potentials, commonly referred to as spikes. 311 00:15:04.279 --> 00:15:06.120 What exactly is a spike? Physically? 312 00:15:06.159 --> 00:15:10.559 These are brief, discrete, essentially identical electrical pulses that travel 313 00:15:10.600 --> 00:15:13.039 down the axon of a neuron and trigger a chemical 314 00:15:13.120 --> 00:15:14.919 or electrical response at the synaps. 315 00:15:15.000 --> 00:15:16.960 Okay, I want to make sure we truly grasp this 316 00:15:17.039 --> 00:15:19.480 because it's a massive departure from how we normally think 317 00:15:19.519 --> 00:15:20.080 about data. 318 00:15:20.200 --> 00:15:21.600 It is a huge mental leap. 319 00:15:21.840 --> 00:15:25.279 If the biological brain isn't sending specific numerical values to 320 00:15:25.320 --> 00:15:30.360 convey information, how does a literal zapp of electricity actually 321 00:15:30.440 --> 00:15:33.480 carry any complex meaning? I mean, a spike is just 322 00:15:33.519 --> 00:15:35.159 a spike, it's a great question. 323 00:15:35.679 --> 00:15:39.039 In biological neural systems, the information is not encoded in 324 00:15:39.080 --> 00:15:42.600 the size, the amplitude, or the shape of the signal. 325 00:15:42.559 --> 00:15:45.000 Because every spike is basically identical, right. 326 00:15:45.480 --> 00:15:48.879 Instead, the information is entirely encoded in the timing and 327 00:15:48.919 --> 00:15:50.080 the rate of those spikes. 328 00:15:50.159 --> 00:15:50.960 The timing and the rate. 329 00:15:51.080 --> 00:15:54.159 Okay, if a sensory neuron in your eye fires rapidly, 330 00:15:54.360 --> 00:15:58.240 producing a dense, high frequency cluster of spikes, it might 331 00:15:58.279 --> 00:16:01.799 be conveying a stronger stimulu, say a very bright light, 332 00:16:01.919 --> 00:16:03.960 compared to a neuron that fires rarely. 333 00:16:04.200 --> 00:16:06.120 That makes sense, more spikes brighter light. 334 00:16:06.200 --> 00:16:08.879 That is called rate coding. But more importantly, there is 335 00:16:08.960 --> 00:16:09.679 temporal code. 336 00:16:09.759 --> 00:16:12.120 Temporal coding like the exact microsecond. 337 00:16:12.200 --> 00:16:16.360 Yes, the precise microsecond timing of a single spike relative 338 00:16:16.399 --> 00:16:18.879 to the spikes of other surrounding neurons, and the network 339 00:16:18.919 --> 00:16:22.360 can carry incredibly complex, high dimensional information. 340 00:16:22.720 --> 00:16:26.279 Let's use an analogy to really visualize why this timing 341 00:16:26.360 --> 00:16:31.960 based spiking communication is such a monumental advantage for saving power. 342 00:16:32.159 --> 00:16:35.720 I'm all ears think about a traditional computer processor like 343 00:16:35.759 --> 00:16:39.799 a massive symphony orchestra, but the conductor is an absolute 344 00:16:39.960 --> 00:16:42.240 tyrant holding a metronome. 345 00:16:41.840 --> 00:16:44.120 A very rigid conductor, very rigid. 346 00:16:44.399 --> 00:16:48.320 The metronome is the computer's internal clock, ticking away billions 347 00:16:48.360 --> 00:16:51.320 of times a second. Every single musician in the orchestra 348 00:16:51.399 --> 00:16:54.240 is forced to play strictly to that beat. Right even 349 00:16:54.240 --> 00:16:56.679 if a musician, let's say the triangle player in the back, 350 00:16:57.000 --> 00:16:59.440 has absolutely no notes to play for a full twenty 351 00:16:59.440 --> 00:17:01.919 minutes of this infanty, they cannot relax. 352 00:17:02.000 --> 00:17:03.080 They have to stay engaged. 353 00:17:03.200 --> 00:17:06.960 The conductor forces them to stand at attention, physically tapping 354 00:17:06.960 --> 00:17:10.680 their foot rigidly keeping time on every single beat, exhausting 355 00:17:10.759 --> 00:17:14.160 themselves just to remain perfectly synchronized with the global metronome. 356 00:17:14.200 --> 00:17:17.200 They are burning massive amounts of energy just to actively 357 00:17:17.240 --> 00:17:17.960 do nothing. 358 00:17:18.160 --> 00:17:22.640 Exactly, and the physical reality of conventional digital chips mirrors 359 00:17:22.640 --> 00:17:24.920 that orchestra exactly doesn't it. 360 00:17:24.559 --> 00:17:28.960 It absolutely does. The global clock cycle forces electrical activity 361 00:17:29.000 --> 00:17:32.480 and power consumption across the entire chip, even. 362 00:17:32.279 --> 00:17:35.640 In sectors of the processor that are momentarily idle, even. 363 00:17:35.400 --> 00:17:38.200 When they have no new data to process. The wires 364 00:17:38.200 --> 00:17:41.720 are constantly charging and discharging just to maintain the rhythm. 365 00:17:41.799 --> 00:17:46.319 But a neuromorphic system, a spiking neural network, is fundamentally different. 366 00:17:46.400 --> 00:17:47.839 How would you describe it. 367 00:17:47.839 --> 00:17:51.680 It's like a cool late night jazz ensemble playing completely 368 00:17:51.759 --> 00:17:54.480 without a set tempo. There is no metronome. 369 00:17:54.599 --> 00:17:55.119 I like that. 370 00:17:55.759 --> 00:17:59.960 A musician in this ensemble will sit in total relaxed silence. 371 00:18:00.680 --> 00:18:04.319 They expend absolutely zero energy. They aren't tapping their foot, 372 00:18:04.319 --> 00:18:06.640 they are actively keeping time. They just wait. They just 373 00:18:06.759 --> 00:18:10.559 wait at rest until exactly the moment they were organically 374 00:18:10.599 --> 00:18:13.559 inspired to play a single note, a spike. They play 375 00:18:13.559 --> 00:18:16.039 their note, and then they immediately return to resting in 376 00:18:16.119 --> 00:18:16.960 total silence. 377 00:18:17.079 --> 00:18:19.720 And the energy savings of that jazz ensemble approach are 378 00:18:19.759 --> 00:18:23.000 staggering a bit, because when those artificial neurons are silent, 379 00:18:23.119 --> 00:18:26.960 the power draw of the chick drops precipitously. This is 380 00:18:27.000 --> 00:18:29.480 the essence of an event driven architecture. 381 00:18:29.759 --> 00:18:32.920 Event driven, so it only reacts when something happens. 382 00:18:33.480 --> 00:18:38.200 Spiking neurons are purely event driven. A neuromorphic chip consumes 383 00:18:38.240 --> 00:18:41.319 practically no power when there is no new data changing 384 00:18:41.319 --> 00:18:42.160 in its environment. 385 00:18:42.319 --> 00:18:43.079 It assists there. 386 00:18:43.160 --> 00:18:46.519 It sits in a quiescent state. Energy is expended only 387 00:18:46.599 --> 00:18:49.559 when and where information is actually flowing through the network. 388 00:18:49.680 --> 00:18:52.279 This means the overall energy cost to the system scales 389 00:18:52.440 --> 00:18:56.319 entirely with the activity of the network, right, not the physical. 390 00:18:55.960 --> 00:18:58.920 Size exactly, not the sheer physical size of the network. 391 00:18:59.319 --> 00:19:02.799 You could build a massive chip containing millions of artificial neurons, 392 00:19:03.079 --> 00:19:05.720 but if only one percent of them are actively spiking 393 00:19:05.720 --> 00:19:10.400 at any given microsecond, your power draw remains astonishingly low. 394 00:19:10.559 --> 00:19:13.440 Which perfectly explains how you get a human brain running 395 00:19:13.480 --> 00:19:14.599 on twenty wants. 396 00:19:14.920 --> 00:19:16.480 It makes perfect sense, now, doesn't it. 397 00:19:16.480 --> 00:19:20.160 It's a massive, dense network of eighty six billion neurons, 398 00:19:20.680 --> 00:19:23.119 but at any given moment, the vast majority of it 399 00:19:23.200 --> 00:19:24.200 is resting. 400 00:19:24.200 --> 00:19:27.920 Just whispering to itself in perfectly timed, sparse spikes. 401 00:19:28.359 --> 00:19:31.599