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 00:00:23.920 --> 00:00:27.600 It is February twenty twenty six, and I actually want 6 00:00:27.640 --> 00:00:29.519 to start by asking you to just take a second. 7 00:00:29.559 --> 00:00:33.840 And you know, exist sure, that sounds easy enough, is it? Though? 8 00:00:33.920 --> 00:00:36.200 I mean, look around you, look at the device you 9 00:00:36.280 --> 00:00:39.399 are listening to this on, or just look out the window, 10 00:00:39.479 --> 00:00:42.560 at the clouds moving, the traffic flowing by. Right. We 11 00:00:42.600 --> 00:00:45.119 are living through a moment right now, mid February twenty 12 00:00:45.159 --> 00:00:49.799 twenty six where the tectonic plates of computing are kind 13 00:00:49.799 --> 00:00:52.439 of shifting right beneath our feet. And I don't mean 14 00:00:53.079 --> 00:00:56.920 a new smartphone release or a slightly faster chatbot. 15 00:00:56.640 --> 00:00:59.560 No, definitely not. We are talking about the invisible engine 16 00:00:59.560 --> 00:01:02.640 that really drives modern civilization exactly. 17 00:01:02.679 --> 00:01:04.400 And the crazy part to me is that most of 18 00:01:04.480 --> 00:01:07.200 us are looking in the completely wrong direction. Like everyone 19 00:01:07.239 --> 00:01:10.359 is watching generative AI, which is huge obviously, don't get 20 00:01:10.359 --> 00:01:12.719 me wrong, but there is something happening in the labs 21 00:01:12.719 --> 00:01:16.120 at Sandia National Laboratories. Right now, that challenges the fundamental 22 00:01:16.120 --> 00:01:17.959 physics of how we compute reality. 23 00:01:18.120 --> 00:01:21.480 It is a subtle shift, but an absolutely seismic one. 24 00:01:21.519 --> 00:01:24.640 It's one of those moments in scientific history where you 25 00:01:24.719 --> 00:01:26.799 might look back ten years later and realize, oh, that 26 00:01:26.920 --> 00:01:28.120 was the day the rules changed. 27 00:01:28.480 --> 00:01:31.040 Yeah. And to understand why it matters so much, we 28 00:01:31.159 --> 00:01:33.840 have to talk about a contradiction, a contradiction that I 29 00:01:33.840 --> 00:01:37.359 think most of us, myself included, honestly, have just accepted 30 00:01:37.400 --> 00:01:40.319 as a fundamental law of the universe. It basically goes 31 00:01:40.359 --> 00:01:44.040 like this. There are creative tasks and there are rigid tasks. 32 00:01:44.200 --> 00:01:46.560 Right. The classic dichotomy it is the split between the 33 00:01:46.680 --> 00:01:47.760 artist and the accountant. 34 00:01:47.959 --> 00:01:52.640 Essentially exactly. On one side, you have art, language, poetry, intuition. 35 00:01:53.519 --> 00:01:55.280 We tend to think of those as things the human 36 00:01:55.280 --> 00:02:00.400 brain does exceptionally well. Biological intelligence is messy, right, It's fuzzy, 37 00:02:00.400 --> 00:02:01.480 it's highly creative. 38 00:02:01.760 --> 00:02:03.879 And then on the exact opposite side you have the 39 00:02:03.959 --> 00:02:09.960 rigid stuff complex mathematics, physics, simulations calculating the trajectory of 40 00:02:10.000 --> 00:02:13.639 a Falcon nine rocket, or modeling how a massive suspension 41 00:02:13.639 --> 00:02:16.879 bridge holds up weight in a category five hurricane, and. 42 00:02:16.840 --> 00:02:19.560 We think of those as things conventional computers. 43 00:02:19.080 --> 00:02:24.840 Do well exactly, logic processors, silicon chips, cold hard binary calculation. 44 00:02:25.280 --> 00:02:28.520 That has been the prevailing wisdom for what seventy years now. 45 00:02:28.719 --> 00:02:31.520 Brains are for thinking, computers are for calculating. 46 00:02:31.800 --> 00:02:33.800 Well, we are here to tell you that as of 47 00:02:33.840 --> 00:02:37.560 this week, that assumption has been flipped completely upside down. 48 00:02:37.960 --> 00:02:41.479 We are doing an analysis today on breaking news coming 49 00:02:41.520 --> 00:02:45.639 out of Sandy and Natural laboratories, specifically a new study 50 00:02:45.680 --> 00:02:48.680 published in Nature Machine Intelligence that basically says we were 51 00:02:48.840 --> 00:02:49.520 entirely wrong. 52 00:02:49.759 --> 00:02:53.639 The headline itself is startling. Researchers have found that brain 53 00:02:53.719 --> 00:02:58.439 inspired computers, what we call neuromorphic hardware, can now solve 54 00:02:58.919 --> 00:03:01.319 super computer life math problems. 55 00:03:01.400 --> 00:03:04.159 And let's be super specific here. We aren't talking about 56 00:03:04.199 --> 00:03:07.360 basic arithmetic. We aren't talking about a chip calculating a 57 00:03:07.400 --> 00:03:10.039 fifteen percent tip at a restaurant. Yeah, we are talking 58 00:03:10.039 --> 00:03:12.039 about partial differential equation. 59 00:03:11.800 --> 00:03:15.560 Pdase, the really really hard stuff, the kind of math 60 00:03:15.599 --> 00:03:18.280 that describe how the universe flows and changes over time. 61 00:03:18.520 --> 00:03:20.599 And the kicker here, the thing that makes this such 62 00:03:20.599 --> 00:03:23.400 a huge deal is that this was previously thought to 63 00:03:23.439 --> 00:03:26.439 be completely impossible for this specific type of hardware. 64 00:03:26.520 --> 00:03:29.080 Impossible is literally the word they used with a widely 65 00:03:29.080 --> 00:03:32.039 accepted fact in the computer science community that you simply 66 00:03:32.080 --> 00:03:34.039 could not do this kind of rigorous math on a 67 00:03:34.080 --> 00:03:34.960 neuromorphic chip. 68 00:03:35.199 --> 00:03:37.719 But they did it. And that breakthrough is what we 69 00:03:37.759 --> 00:03:40.439 are going to explore today, because it is not just 70 00:03:40.479 --> 00:03:43.159 about building a faster calculator, is it not at all? 71 00:03:43.280 --> 00:03:47.800 It is a breakthrough in efficiency, in overall capability, and 72 00:03:47.960 --> 00:03:50.199 perhaps most importantly, and this is the part that genuinely 73 00:03:50.240 --> 00:03:52.840 keeps me of at night, it's a breakthrough in understanding 74 00:03:52.840 --> 00:03:54.479 biological intelligence itself. 75 00:03:54.919 --> 00:03:58.000 So here is our mission for this exploration. We are 76 00:03:58.039 --> 00:04:02.080 going to unpack how a computer modeled physically after biological 77 00:04:02.120 --> 00:04:05.159 gray matter, after the squishy stuff in our skulls, can 78 00:04:05.159 --> 00:04:08.000 outperform traditional logic processors at their own game. 79 00:04:08.240 --> 00:04:10.439 We will be pulling directly from that key study in 80 00:04:10.520 --> 00:04:13.599 Nature Machine intelligence, along with reports from the Department of 81 00:04:13.680 --> 00:04:17.319 Energy and the technical notes provided by Sandy and National Laboratories. 82 00:04:17.439 --> 00:04:20.040 And just so you understand the stakes here as you listen, 83 00:04:20.279 --> 00:04:23.120 this isn't just academic, This isn't just about math geeks 84 00:04:23.120 --> 00:04:25.199 getting hyped about a new algorithm in a lab. 85 00:04:25.360 --> 00:04:27.839 No, the real world applications are massive. 86 00:04:28.079 --> 00:04:31.360 Right, we are talking about national security, We're talking about 87 00:04:31.399 --> 00:04:36.480 global climate modeling and potentially unlocking the underlying secrets of 88 00:04:36.519 --> 00:04:37.839 diseases like Alzheimer's. 89 00:04:38.240 --> 00:04:44.199 It brilliantly connects the very large nuclear stockpile simulations to 90 00:04:44.279 --> 00:04:48.560 the very small down to the individual neurons firing in 91 00:04:48.639 --> 00:04:51.000 your brain right now as you process my voice. 92 00:04:51.079 --> 00:04:54.000 So buckle up. We are going on a deep exploration 93 00:04:54.079 --> 00:04:55.759 of the impossible calculation. 94 00:04:55.959 --> 00:04:56.480 I'm ready. 95 00:04:56.720 --> 00:04:59.360 Let's start with what I like to call the intuition trap, 96 00:04:59.519 --> 00:05:00.959 because I thin I think this is where most of 97 00:05:01.000 --> 00:05:04.439 us get entirely tripped up when we compare computers to brains. 98 00:05:04.560 --> 00:05:07.120 It is a very common trap. It is actually closely 99 00:05:07.199 --> 00:05:10.800 related to something called morvex paradox in early AI research. 100 00:05:10.920 --> 00:05:12.879 Oh right, give us the rundown on the paradox. 101 00:05:13.120 --> 00:05:15.680 Well. In the early days of artificial intelligence back in 102 00:05:15.720 --> 00:05:18.680 the eighties, researchers assume the hard stuff to teach a 103 00:05:18.680 --> 00:05:23.040 computer would be high level reasoning, playing chess, proving logic theorems, 104 00:05:23.040 --> 00:05:25.240 that sort of thing. They thought the easy stuff would 105 00:05:25.240 --> 00:05:29.199 be basic sensor motor tasks, walking across the room, folding 106 00:05:29.199 --> 00:05:31.839 a towel, or just recognizing a face, and. 107 00:05:31.800 --> 00:05:34.519 It turned out to be the exact opposite, exactly. 108 00:05:35.000 --> 00:05:37.759 A nineteen eighties computer could beat a chess grand master, 109 00:05:38.240 --> 00:05:39.879 but you couldn't get a robot to walk over a 110 00:05:39.879 --> 00:05:42.360 pile of laundry without it immediately falling on its face. 111 00:05:43.199 --> 00:05:45.639 The stuff that feels hard to us, like logic and 112 00:05:45.720 --> 00:05:49.360 high level math, is actually computationally simple for a machine. 113 00:05:50.160 --> 00:05:53.360 But the stuff that feels completely easy to us, perception 114 00:05:53.759 --> 00:05:56.480 fluid movement, is a computational nightmare. 115 00:05:56.600 --> 00:05:59.600 So let's apply that directly to this Sandia study. I 116 00:05:59.639 --> 00:06:03.160 want you to picture something. As you listen, Imagine you 117 00:06:03.199 --> 00:06:06.120 are standing on a tennis court. It is a beautiful 118 00:06:06.160 --> 00:06:09.360 sunny day. Someone serves a ball to you. It is 119 00:06:09.439 --> 00:06:11.839 flying across the net at I don't know, one hundred 120 00:06:11.879 --> 00:06:12.360 miles an. 121 00:06:12.240 --> 00:06:13.839 Hour, fast, very fast. 122 00:06:13.879 --> 00:06:16.360 You see it, You step forward, you plant your foot, 123 00:06:16.399 --> 00:06:19.240 you swing your racket, and whack you return the serve 124 00:06:19.319 --> 00:06:20.040 right down the line. 125 00:06:20.120 --> 00:06:22.360 A tech book example of censorimotor integration. 126 00:06:22.480 --> 00:06:24.240 Now, be honest, how does that feel to you, to 127 00:06:24.279 --> 00:06:25.920 the human being actually doing it? 128 00:06:25.920 --> 00:06:29.720 It feels instinctive, It feels purely reactive. You don't stand 129 00:06:29.720 --> 00:06:32.040 there with a notepad and calculate the wind speed. You 130 00:06:32.079 --> 00:06:34.279 don't pull out a protractor to measure the angle of 131 00:06:34.279 --> 00:06:37.920 the sun. You just do it. It feels entirely easy. 132 00:06:37.839 --> 00:06:40.800 Right, It feels effortless. Now, imagine I sit you down 133 00:06:40.839 --> 00:06:42.879 at a sterile desk with a blank sheet of paper 134 00:06:42.879 --> 00:06:45.199 and a pencil, and I say, okay, I need you 135 00:06:45.279 --> 00:06:49.560 to solve this partial differential equation describing the fluid dynamics 136 00:06:49.560 --> 00:06:52.800 of air resistance on a fuzzy sphere traveling at forty 137 00:06:52.839 --> 00:06:55.600 five meters per second with a localized grosswind. 138 00:06:55.759 --> 00:06:59.000 Yeah, most humans, myself certainly included, would break out in 139 00:06:59.040 --> 00:07:01.920 a cold, sweaty meatia. That feels incredibly hard. 140 00:07:02.040 --> 00:07:05.199 Exactly, hitting the ball feels easy. The math on paper 141 00:07:05.240 --> 00:07:08.720 feels hard. But here is where the experts at Sandia, 142 00:07:08.800 --> 00:07:12.839 specifically brad Amone, one of the computational neuroscientists on this project, 143 00:07:12.879 --> 00:07:16.360 say we are totally completely wrong in how we view that. 144 00:07:16.680 --> 00:07:19.240 This is the insight that really flips the script. Em 145 00:07:19.240 --> 00:07:22.199 Moone points out that our intuition about effort is an illusion. 146 00:07:22.560 --> 00:07:25.040 In reality, the motor control required to hit that moving 147 00:07:25.079 --> 00:07:29.319 tennis ball involves what he calls EXAs scale level physics 148 00:07:29.360 --> 00:07:31.319 computations EXAs scale. 149 00:07:31.360 --> 00:07:33.720 That is a word usually reserved for the absolute most 150 00:07:33.720 --> 00:07:37.600 powerful massive supercomputers on Earth. We are talking quintillions of 151 00:07:37.639 --> 00:07:39.120 calculations per single. 152 00:07:38.920 --> 00:07:42.800 Second, precisely. Think about what is actively happening under the 153 00:07:42.839 --> 00:07:45.439 hood when you swing that racket. It is not magic, 154 00:07:45.680 --> 00:07:49.040 It is hard physics. Your brain is actively processing the 155 00:07:49.079 --> 00:07:53.199 trajectory of the ball in three D space using stereoscopic vision. 156 00:07:53.759 --> 00:07:57.120 It is accounting for wind resistance, It is calculating gravity. 157 00:07:57.240 --> 00:07:58.920 And the ball has spin on it too, right, the 158 00:07:58.959 --> 00:07:59.639 Magnus effect. 159 00:07:59.680 --> 00:08:02.959 It curs in the air, right, it has spin, And simultaneously, 160 00:08:03.000 --> 00:08:06.319 while tracking all that, your brain is adjusting this specific 161 00:08:06.399 --> 00:08:09.920 tension in hundreds of individual muscles, your bicep, your triceps, 162 00:08:09.920 --> 00:08:12.920 your lads, your calves. It is constantly balancing your center 163 00:08:12.959 --> 00:08:15.839 of mass so you don't tip over. It is predicting 164 00:08:15.920 --> 00:08:18.759 where the ball will be in exactly zero point five seconds, 165 00:08:19.240 --> 00:08:21.879 all in real time, all within milliseconds. 166 00:08:21.319 --> 00:08:22.959 And it is doing all of that while the ball 167 00:08:23.040 --> 00:08:24.439 is still actively in the air. 168 00:08:24.600 --> 00:08:26.920 If you tried to program a traditional robot to do 169 00:08:26.959 --> 00:08:29.800 that using explicit math, literally writing out the equations for 170 00:08:29.839 --> 00:08:33.240 every single variable, every gust of wind, every tiny muscle twitch, 171 00:08:33.559 --> 00:08:36.679 you would need a staggering amount of processing power. You 172 00:08:36.720 --> 00:08:39.639 would be solving complex physics equations explicitly. 173 00:08:39.919 --> 00:08:43.799 But our brains do it how because I definitely don't 174 00:08:43.799 --> 00:08:46.080 feel like a supercomputer when I'm playing tennis. I just 175 00:08:46.080 --> 00:08:47.399 feel like I'm swinging my arm. 176 00:08:47.600 --> 00:08:50.559 We do it cheaply, we do it incredibly efficiently. The 177 00:08:50.679 --> 00:08:54.159 quote from brad Amon in the material is fantastic. He says, 178 00:08:54.320 --> 00:08:58.480 these are very sophisticated computations. They are exast scale level 179 00:08:58.519 --> 00:09:01.440 problems that our brains are cap of doing very cheaply. 180 00:09:01.639 --> 00:09:04.320 Cheaply meaning low energy, barely. 181 00:09:03.960 --> 00:09:07.039 Any energy at all. Your brain runs on about twenty wants. 182 00:09:06.679 --> 00:09:09.519 Of power, twenty watts like a dim light ball, the very. 183 00:09:09.360 --> 00:09:13.480 Dim light bulb, maybe refrigerator light. Meanwhile, a supercomputer capable 184 00:09:13.480 --> 00:09:17.840 of doing those same exact physics calculations explicitly simulating the air, 185 00:09:18.320 --> 00:09:22.360 the ball the human body would require megawatts of power. 186 00:09:22.399 --> 00:09:24.840 It would literally need its own dedicated power plant. 187 00:09:24.919 --> 00:09:28.080 So the massive revelation here is that our brains are 188 00:09:28.120 --> 00:09:32.519 already solving these complex physics equations. We are solving partial 189 00:09:32.559 --> 00:09:36.440 differential equations constantly, just by existing in the physical world. 190 00:09:36.639 --> 00:09:38.519 We just aren't consciously aware of the math. 191 00:09:38.679 --> 00:09:41.480 We are walking talking physics engines. We just happen to 192 00:09:41.480 --> 00:09:43.759 have a very user friendly interface that hides all the 193 00:09:43.799 --> 00:09:45.480 complex code from our conscious mind. 194 00:09:45.720 --> 00:09:47.440 That is such a cool way to think about it. 195 00:09:47.919 --> 00:09:49.360 We are running the physics code in. 196 00:09:49.360 --> 00:09:52.080 The background constantly. If you casually catch a set of 197 00:09:52.120 --> 00:09:55.039 keys someone throws at you from across the room, you 198 00:09:55.120 --> 00:09:59.240 just organically solved a parabolic arc differential equation. If you 199 00:09:59.320 --> 00:10:03.000 merge your car onto a busy highway, you're rapidly calculating 200 00:10:03.039 --> 00:10:05.600 relative velocities and friction coefficients. 201 00:10:06.080 --> 00:10:09.240 Okay, so we have dropped this term a few times now. 202 00:10:09.960 --> 00:10:13.840 Partial differential equations or PDEs. I want to pause here 203 00:10:13.840 --> 00:10:15.639 for a second. We need to define this because this 204 00:10:15.720 --> 00:10:18.480 is the hard math that the new computer chips are 205 00:10:18.480 --> 00:10:21.320 finally solving. What exactly are these equations? 206 00:10:21.600 --> 00:10:25.360 So simply put, PDEs are the mathematical language of the 207 00:10:25.399 --> 00:10:26.480 physical world. 208 00:10:26.279 --> 00:10:27.960 The language of reality itself. 209 00:10:28.120 --> 00:10:31.399 Essentially, yes, in regular high school algebra, you might solve 210 00:10:31.440 --> 00:10:33.559 for x and x is just a single static number 211 00:10:33.799 --> 00:10:37.120 like ex equals five. In a partial differential equation, the 212 00:10:37.159 --> 00:10:40.360 solution isn't a single number. The solution is a function. 213 00:10:41.159 --> 00:10:44.679 It describes how things continuously change over space and time, 214 00:10:45.000 --> 00:10:49.000 because in the actual universe nothing is perfectly static. Everything 215 00:10:49.080 --> 00:10:54.360 is moving, flowing, heating up, cooling down, stressing, bending. PDEs 216 00:10:54.519 --> 00:10:58.200 are the equations we use to mathematically map those constant changes. 217 00:10:59.000 --> 00:11:00.879 Can you give us some conc create examples from the 218 00:11:00.919 --> 00:11:03.279 source material? What are we normally using these for in 219 00:11:03.279 --> 00:11:03.919 the real world. 220 00:11:04.200 --> 00:11:07.360 The classic example is forecasting weather patterns. That is a 221 00:11:07.399 --> 00:11:11.759 massive PDE problem. You have air, pressure, temperature, moisture, wind speed, 222 00:11:12.120 --> 00:11:16.080 all dynamically interacting over the entire surface of the globe. 223 00:11:15.679 --> 00:11:17.879 And they all directly affect each other. Right If the 224 00:11:17.879 --> 00:11:21.200 temperature drops, the pressure changes, which immediately changes the wind 225 00:11:21.200 --> 00:11:22.320 direction exactly. 226 00:11:22.360 --> 00:11:25.639 It is a tightly coupled system. Another huge example is 227 00:11:25.679 --> 00:11:30.120 fluid dynamics, how water flows through a complex municipal pipe system, 228 00:11:30.519 --> 00:11:32.639 or how air flows over the curved weghing of a 229 00:11:32.639 --> 00:11:37.360 commercial airplane. Or think about structural mechanics calculating exactly how 230 00:11:37.360 --> 00:11:40.399 a steel building material will stress and eventually snap under 231 00:11:40.399 --> 00:11:41.399 a heavy physical. 232 00:11:41.120 --> 00:11:43.519 Load okay that paints a clear picture. 233 00:11:43.559 --> 00:11:48.799 Or modeling invisible electromagnetic fields. These are all strictly governed 234 00:11:48.840 --> 00:11:51.240 by partial differential equations. 235 00:11:50.799 --> 00:11:53.039 And the key thing for everyone to understand here is 236 00:11:53.039 --> 00:11:55.840 that these are not simple arithmetic problems. You can't just 237 00:11:55.919 --> 00:11:59.039 plug them into a standard desktop calculator and hit enter. 238 00:11:59.440 --> 00:12:04.200 No, they are incredibly, incredibly resource intensive. The way we 239 00:12:04.279 --> 00:12:07.639 solve them traditionally in computer science is by breaking the 240 00:12:07.679 --> 00:12:10.879 physical world down into tiny little geometric cubes. We call 241 00:12:10.919 --> 00:12:13.919 it a mesh, like pixelating reality exactly like that, you 242 00:12:13.960 --> 00:12:18.080 pixelate reality into a highly detailed three D grid. Then 243 00:12:18.159 --> 00:12:21.240 you have the computer calculate the physics for every single 244 00:12:21.320 --> 00:12:25.480 little isolated cube and simultaneously calculate how every single cube 245 00:12:25.480 --> 00:12:28.120 interacts with its immediate neighbors. To do that at a 246 00:12:28.159 --> 00:12:31.480 high enough resolution to be useful, like safely simulating a 247 00:12:31.559 --> 00:12:34.759 nuclear explosion or mapping a global climate model, you need 248 00:12:35.039 --> 00:12:40.279 massive energy hungry supercomputers. You need to painfully crunch billions 249 00:12:40.279 --> 00:12:43.279 and billions of numbers just to get a single answer. 250 00:12:43.159 --> 00:12:46.919 Because you're simulating millions of little microscopic interactions happening all 251 00:12:46.960 --> 00:12:49.080 at the exact same time exactly. 252 00:12:49.399 --> 00:12:52.320 And that has always been the fundamental bottleneck in science. 253 00:12:52.879 --> 00:12:55.360 We have the math, We've had the maths for hundreds 254 00:12:55.399 --> 00:12:58.440 of years. Newton and Liivenis started this whole thing, but 255 00:12:58.559 --> 00:13:02.159 actually running the math that costs an absolute fortune in 256 00:13:02.399 --> 00:13:04.799 energy and hardware infrastructure. 257 00:13:04.279 --> 00:13:06.840 Which perfectly brings us to the hardware side of this 258 00:13:06.879 --> 00:13:11.159 breakthrough because the solution Sandy have found wasn't just hey, 259 00:13:11.240 --> 00:13:14.480 let's build a slightly bigger computer. It was let's build 260 00:13:14.519 --> 00:13:15.919 a completely different kind of. 261 00:13:15.879 --> 00:13:18.399 Computer, right, and this is where we finally get into 262 00:13:18.679 --> 00:13:19.960 neuromorphic computing. 263 00:13:20.240 --> 00:13:23.120 Neuromorphic it sounds like something straight out of a futuristic 264 00:13:23.159 --> 00:13:23.879 sci fi novel. 265 00:13:24.000 --> 00:13:25.879 It literally just means taking the form of the brain, 266 00:13:26.240 --> 00:13:28.879 neuro meaning brain morphic meaning form. 267 00:13:29.159 --> 00:13:32.759 So how is this physically different from the laptop I 268 00:13:32.799 --> 00:13:34.879 have sitting right here in front of me, or the 269 00:13:34.919 --> 00:13:37.120 massive server farms running the internet. 270 00:13:37.320 --> 00:13:40.879 Your laptop, your phone, and the massive supercomputers we have 271 00:13:40.960 --> 00:13:43.840 relied on for decades are all based on what we 272 00:13:43.919 --> 00:13:45.559 call the von Neumann architecture. 273 00:13:45.600 --> 00:13:46.799 Okay, let's unpack that. 274 00:13:47.240 --> 00:13:51.480 Von Neuman John von Neuman brilliant mathematician. In the nineteen forties, 275 00:13:52.320 --> 00:13:55.480 he proposed a logical structure for computers that we literally 276 00:13:55.519 --> 00:13:59.200 still use today. In a traditional computer, you have very 277 00:13:59.200 --> 00:14:03.600 clear physical separation. You have the processor, the CPU, the 278 00:14:03.639 --> 00:14:06.360 logic center where the math happens, and you have the 279 00:14:06.399 --> 00:14:08.960 memory the ram where the data is actually stored. 280 00:14:09.080 --> 00:14:10.440 So they live in different houses. 281 00:14:10.559 --> 00:14:13.759 They live in entirely different neighborhoods. On a motherboard, when 282 00:14:13.759 --> 00:14:16.600 the computer wants to do a single calculation, it has 283 00:14:16.679 --> 00:14:19.879 to physically go to the memory, grab a specific piece 284 00:14:19.919 --> 00:14:22.759 of data, drive it all the way over to the processor, 285 00:14:23.080 --> 00:14:25.559 do the math, and then drive the answer all the 286 00:14:25.600 --> 00:14:27.120 way back to memory to store it. 287 00:14:27.240 --> 00:14:30.399 And that driving, that physical moving of electronic data back 288 00:14:30.440 --> 00:14:32.000 and forth is the main problem. 289 00:14:32.159 --> 00:14:36.759 It is the infamous Vunnowmen bottleneck. It takes time, and crucially, 290 00:14:36.799 --> 00:14:39.720 it takes a massive amount of energy. A surprisingly huge 291 00:14:39.720 --> 00:14:42.519 percentage of the energy your computer uses isn't actually doing 292 00:14:42.600 --> 00:14:45.559 any math. It's just moving numbers back and forth across 293 00:14:45.600 --> 00:14:46.360 tiny wires. 294 00:14:46.480 --> 00:14:48.399 It's like having a chef who has to walk three 295 00:14:48.440 --> 00:14:50.879 blocks to the grocery store every single time they need 296 00:14:50.919 --> 00:14:52.440 a pinch AsSalt or a single egg. 297 00:14:52.639 --> 00:14:56.840 That is a perfect analogy. It is incredibly inefficient. If 298 00:14:56.840 --> 00:15:00.480 you are trying to rapidly cook a massive ten course banquet, 299 00:15:01.159 --> 00:15:03.600 you end up spending way more time and energy walking 300 00:15:03.639 --> 00:15:04.879 than you do actually cooking. 301 00:15:05.159 --> 00:15:07.000 So how is a neuromorphic chip. 302 00:15:07.039 --> 00:15:11.200 Different neuromorphic computers are modeled physically and architecturally after the 303 00:15:11.279 --> 00:15:15.080 human brain. Instead of rigid silicon logic gits, they have 304 00:15:15.399 --> 00:15:19.919 artificial neurons and synapses. They process information in parallel, all 305 00:15:19.960 --> 00:15:22.120 at once, rather than one tiny step at a time. 306 00:15:22.440 --> 00:15:24.799 And crucially, and this is the absolute key to the 307 00:15:24.879 --> 00:15:28.639 energy savings, we see the memory and the processing happen 308 00:15:28.759 --> 00:15:30.919 in the exact same physical location. 309 00:15:31.039 --> 00:15:33.840 Just like the biological brain. My memories aren't stored in 310 00:15:33.879 --> 00:15:36.320 the separate biological hard drive in my foot, They are 311 00:15:36.399 --> 00:15:38.480 right there in the network of neurons where the thinking 312 00:15:38.559 --> 00:15:39.600 happens exactly. 313 00:15:39.799 --> 00:15:42.720 This is referred to in the industry as in memory computing. 314 00:15:43.039 --> 00:15:46.000 You don't move the data to the processor. You process 315 00:15:46.039 --> 00:15:47.840 the data directly where it already lives. 316 00:15:47.919 --> 00:15:50.600 That makes total sense intuitively, So why haven't we been 317 00:15:50.720 --> 00:15:53.759 using these brain chips for physics all along? If they 318 00:15:53.799 --> 00:15:57.200 are so incredibly efficient, why are we still building massive, 319 00:15:57.360 --> 00:16:00.000 power hungry GPU clusters, Because. 320 00:15:59.759 --> 00:16:02.200 For a very long time there was a heavy stigma 321 00:16:02.360 --> 00:16:04.960 attached to this kind of hardware in the scientific community. 322 00:16:05.080 --> 00:16:06.679 A stigma really. 323 00:16:06.440 --> 00:16:09.399 Well, maybe limitation is a slightly better word. Because these 324 00:16:09.480 --> 00:16:13.600 chips work organically like brains, they are inherently. 325 00:16:13.120 --> 00:16:15.519 Noisy, noisy what does that mean for a computer? 326 00:16:15.919 --> 00:16:18.799 They use spikes of electricity to communicate, very much like 327 00:16:18.840 --> 00:16:22.480 biological neurons firing. They aren't always wanted to invent rigidly 328 00:16:22.519 --> 00:16:25.879 precise in the way a traditional digital calculator is. A 329 00:16:25.960 --> 00:16:29.240 standard calculator gives you exactly five point zero zero zero 330 00:16:29.360 --> 00:16:34.279 zero zero zero zero zero every single time. Neuromorphic chip 331 00:16:34.519 --> 00:16:37.679 might give you an output that is statistically five, but 332 00:16:37.799 --> 00:16:40.840 it fluctuates slightly. It has inherent mathematical randomness. 333 00:16:40.840 --> 00:16:43.320 Oh I see. So the broader scientific community thought they 334 00:16:43.320 --> 00:16:45.000 were only really good for fuzzy. 335 00:16:44.679 --> 00:16:49.000 Tasks, right, task like pattern recognition showing an artificial intelligence 336 00:16:49.000 --> 00:16:50.960 a thousand photos and asking is this a cat or 337 00:16:50.960 --> 00:16:55.080 a dog? Neuromorphic chips are absolutely fantastic at that. They 338 00:16:55.080 --> 00:16:57.679 could look at the gestalt, the whole messy picture and 339 00:16:57.720 --> 00:16:59.480 make a fast, efficient judgment call. 340 00:16:59.759 --> 00:17:02.919 But you definitely wouldn't trust a fuzzy judgment call to 341 00:17:03.000 --> 00:17:06.640 calculate the structural integrity of a suspension bridge you're driving over. 342 00:17:06.759 --> 00:17:11.440 Precisely. That was the core assumption. The skepticism was incredibly deep. 343 00:17:11.720 --> 00:17:15.079 It was widely assumed that because they lacked that rigid 344 00:17:15.240 --> 00:17:18.799 linear decimal point precision, because they were so called noisy. 345 00:17:19.680 --> 00:17:23.559 They simply couldn't handle rigorous math like PDEs. You don't 346 00:17:23.559 --> 00:17:26.519 want a fuzzy ballpark answer when you are dealing with 347 00:17:26.640 --> 00:17:30.200 critical nuclear physics. You want the exact precise answer. 348 00:17:30.359 --> 00:17:33.599 So everyone just naturally assumed neuromorphic chips are for AI 349 00:17:33.720 --> 00:17:36.839 looking at cat photos and giant traditional supercomputers are for 350 00:17:37.079 --> 00:17:38.160 serious match. 351 00:17:38.000 --> 00:17:41.319 Until right now, until this publication in February twenty twenty six. 352 00:17:41.599 --> 00:17:44.400 Enter Brad Siloman and Brad Amone, the two. 353 00:17:44.200 --> 00:17:46.759 Brads, the dynamic duo at Sandy National Labs. 354 00:17:46.920 --> 00:17:49.200 So what did they actually do to break this assumption? 355 00:17:49.240 --> 00:17:51.759 Did they have to invent and manufacture a completely new 356 00:17:51.880 --> 00:17:52.920 kind of physical chip. 357 00:17:53.480 --> 00:17:57.920 Interestingly, no, the hardware itself wasn't the primary novelty hear, 358 00:17:58.240 --> 00:18:02.119 They actually used existing state of the art neuromorphic architecture. 359 00:18:02.319 --> 00:18:05.680 The massive breakthrough was the algorithm. They created a fundamentally 360 00:18:05.720 --> 00:18:07.559 new way of operating the hardware. 361 00:18:07.799 --> 00:18:11.960 Okay, walk us through this. How do you force a noisy, 362 00:18:12.160 --> 00:18:15.559 fuzzy brain chip to do rigorous hard math. 363 00:18:16.119 --> 00:18:19.920 They essentially realized that the physical structure of these neuromorphic chips, 364 00:18:19.920 --> 00:18:23.680 the specific way the artificial neurons connect and fire, could 365 00:18:23.720 --> 00:18:28.039 be mathematically mapped directly to the structure of partial differential equations. 366 00:18:28.079 --> 00:18:31.119 Wait, so the math itself physically looks like the brain 367 00:18:31.160 --> 00:18:31.839 structure in. 368 00:18:31.839 --> 00:18:35.440 A highly abstract way. Yes, this is the fascinating part 369 00:18:35.440 --> 00:18:38.160 of the paper. They took a circuit model that was 370 00:18:38.200 --> 00:18:41.440 already very well known in neuroscience. It is a model 371 00:18:41.480 --> 00:18:43.640 of how biological cortical networks work. 372 00:18:43.799 --> 00:18:46.160 Cortical networks being the outer layer of the brain right. 373 00:18:46.119 --> 00:18:48.599 Right, the cerebral cortex, the advanced thinking part of the brain. 374 00:18:48.839 --> 00:18:51.960 This specific circuit model had been kicking around in academia 375 00:18:51.960 --> 00:18:55.920 for about twelve years. Neuroscientists used it strictly to understand 376 00:18:55.920 --> 00:18:59.240 how our biological neurons inhibit and excite each other to 377 00:18:59.279 --> 00:19:03.200 process visual or sense reinformation. Okay, but nobody had ever 378 00:19:03.240 --> 00:19:05.359 deeply looked at that biological model and said, hey, wait 379 00:19:05.359 --> 00:19:10.200 a minute, this exact pattern of inhibition and excitation looks 380 00:19:10.240 --> 00:19:13.559 exactly like the mathematical operations we used to solve diffusion 381 00:19:13.599 --> 00:19:14.559 equations on a computer. 382 00:19:14.880 --> 00:19:17.720 Ah. So that is the twelve year gap mentioned in 383 00:19:17.759 --> 00:19:18.720 the Sander report. 384 00:19:18.839 --> 00:19:22.359 Exactly, the information was already there. The biological circuit was 385 00:19:22.359 --> 00:19:26.519 heavily documented, but the direct link to applied mathematics wasn't 386 00:19:26.559 --> 00:19:30.039 made until Thielman and Amone finally connected the dots. 387 00:19:30.200 --> 00:19:32.920 It is a classic case of academic silos, isn't it. 388 00:19:32.960 --> 00:19:36.519 The neuroscientists are in one building studying biology, the applied 389 00:19:36.559 --> 00:19:40.839 mathematicians are in another building studying equations, and rarely. 390 00:19:40.519 --> 00:19:43.119 Do they ever sit down over coffee and realize they 391 00:19:43.119 --> 00:19:46.200 were looking at the exact same map from two different angles. 392 00:19:46.920 --> 00:19:49.480 The neuroscientists were looking at it strictly as a model 393 00:19:49.519 --> 00:19:53.039 of wet biology. The mathematicians were blindly looking for a 394 00:19:53.039 --> 00:19:56.519 faster way to solve dry equations. Fielman and Emohen realized 395 00:19:56.559 --> 00:19:58.880 it was fundamentally the same mechanism. 396 00:19:58.680 --> 00:20:01.839 So they applied this new outme algorithm. They essentially program 397 00:20:01.880 --> 00:20:06.119 the neuromorphic chip to behave exactly like this biological cortical network. 398 00:20:06.400 --> 00:20:07.519 And what actually happened. 399 00:20:07.559 --> 00:20:11.000 They successfully proved that these brain like systems can rapidly 400 00:20:11.039 --> 00:20:15.279 solve sparse finite element problems that is the highly technical 401 00:20:15.359 --> 00:20:18.920