WEBVTT 1 00:00:03.439 --> 00:00:07.719 Welcome to Bedtime Astronomy. Explore the wonders of the cosmos 2 00:00:07.759 --> 00:00:12.279 with our soothing Bedtime Astronomy podcast. Each episode offers a 3 00:00:12.359 --> 00:00:16.320 gentle journey through the stars, planets, and beyond, perfect for 4 00:00:16.399 --> 00:00:20.239 unwinding after a long day. Let's travel through the mysteries 5 00:00:20.239 --> 00:00:22.440 of the universe as you drift off into a peaceful 6 00:00:22.480 --> 00:00:23.800 slumber under the night sky. 7 00:00:26.920 --> 00:00:30.160 Okay, let's unpack this. We are basically building the smartest 8 00:00:30.199 --> 00:00:33.840 machines in human history, right to unlock the secrets. 9 00:00:33.399 --> 00:00:35.000 Of the universe, right, that is the goal. 10 00:00:35.320 --> 00:00:39.359 But it turns out we're accidentally teaching them to be 11 00:00:39.479 --> 00:00:40.399 completely blind. 12 00:00:40.679 --> 00:00:43.840 Yeah. It is really the ultimate irony of modern science. Honestly. 13 00:00:43.880 --> 00:00:46.479 I mean, we are not dealing with a hardware failure here, right, 14 00:00:46.759 --> 00:00:48.960 you know, a glitch in the code. What we are 15 00:00:49.000 --> 00:00:52.240 looking at is a fundamental flaw in the actual philosophy 16 00:00:52.280 --> 00:00:55.759 of how a machine learns to understand reality. 17 00:00:55.280 --> 00:00:58.359 Which is wild to think about, because I mean, think 18 00:00:58.359 --> 00:01:01.039 about the expectation of precision. We carry into something like 19 00:01:02.560 --> 00:01:05.359 a medical diagnosis. Oh, absolutely, Like you fall, you break 20 00:01:05.359 --> 00:01:07.680 your arm, You go to the hospital, the doctor takes 21 00:01:07.680 --> 00:01:10.400 an X ray, clips it to the lightboard, and there 22 00:01:10.400 --> 00:01:10.680 it is. 23 00:01:10.799 --> 00:01:12.920 It's right in front of you exactly, just. 24 00:01:12.879 --> 00:01:16.640 A jagged white line cutting across the dark bone. It's binary, 25 00:01:16.719 --> 00:01:18.359 you know, broken or not broken. 26 00:01:18.719 --> 00:01:22.439 And that stark visual evidence is so comforting because it 27 00:01:22.480 --> 00:01:26.840 maps perfectly onto our desire for reality to be easily categorized. 28 00:01:26.879 --> 00:01:29.599 Yes, we want the universe to be just as clean 29 00:01:29.640 --> 00:01:30.359 as that X ray. 30 00:01:30.400 --> 00:01:31.000 We really do. 31 00:01:31.359 --> 00:01:35.519 We want to point these massive, powerful telescopes into the dark, 32 00:01:36.159 --> 00:01:39.200 take a cosmic X ray, and just have an AI 33 00:01:39.400 --> 00:01:42.519 point to a jagged white line and say, boom, there 34 00:01:42.560 --> 00:01:43.719 is the new law of physics. 35 00:01:44.439 --> 00:01:48.519 But the cosmos violently refuses to be categorized that easily. 36 00:01:48.680 --> 00:01:52.120 It really does the phenomena that cosmologists are hunting for 37 00:01:52.280 --> 00:01:55.719 right now. They do not leave clean jagged lines. They 38 00:01:55.799 --> 00:01:56.879 leave whispers. 39 00:01:56.920 --> 00:01:57.879 Whispers. I like that. 40 00:01:58.000 --> 00:02:01.480 Yeah, they leave incredibly subtle statistic variations and the way 41 00:02:01.799 --> 00:02:05.959 galaxies cluster together over billions of years. And to find 42 00:02:06.000 --> 00:02:09.080 those whispers, we have essentially handed the mathematical keys to 43 00:02:09.120 --> 00:02:11.680 the universe over to artificial neural. 44 00:02:11.439 --> 00:02:13.960 Networks, which makes sense on paper, right because they can 45 00:02:14.000 --> 00:02:14.879 process so much data. 46 00:02:15.000 --> 00:02:18.199 It makes total sense. But the massive often overlook challenge 47 00:02:18.199 --> 00:02:21.599 here is that the prior knowledge we feed into those networks. 48 00:02:21.240 --> 00:02:23.919 The baseline rules we teach them exactly. 49 00:02:23.479 --> 00:02:26.120 The foundational rules we give them just so they can function. 50 00:02:26.879 --> 00:02:29.759 That can actually become the very obstacle that prevents them 51 00:02:29.759 --> 00:02:32.159 from recognizing a revolutionary discovery. 52 00:02:32.199 --> 00:02:34.840 Okay, So imagine someone hands you a pair of highly 53 00:02:34.840 --> 00:02:38.479 advanced glasses, right, Okay, and they tell you these glasses 54 00:02:38.479 --> 00:02:42.560 will let you see the true, unvarnished architecture of a 55 00:02:42.599 --> 00:02:44.000 completely alien planet. 56 00:02:44.360 --> 00:02:45.520 Sounds great so far. 57 00:02:45.680 --> 00:02:49.479 Right, But there is a hidden catch. The lenses have 58 00:02:49.479 --> 00:02:53.159 been perfectly tinted, perfectly calibrated by your guide to only 59 00:02:53.240 --> 00:02:56.560 highlight shapes and colors that already exist in your own backyard. 60 00:02:56.759 --> 00:02:58.479 Ah. I see where you're going with this. 61 00:02:58.800 --> 00:03:02.039 So you step onto this completely alien landscape, a place 62 00:03:02.039 --> 00:03:06.039 with entirely new physics and weird new structures. But everywhere 63 00:03:06.080 --> 00:03:09.159 you look, the glass is aggressively filled to the light. 64 00:03:09.280 --> 00:03:12.080 So all you see are what familiar oak trees and 65 00:03:12.560 --> 00:03:13.400 picket fences. 66 00:03:13.439 --> 00:03:17.240 Exactly, you are staring directly at the unknown, but your 67 00:03:17.280 --> 00:03:18.960 tools are forcing you to see the known. 68 00:03:19.240 --> 00:03:22.240 That is precisely the trap we are inadvertently building for 69 00:03:22.280 --> 00:03:23.479 our most advanced AI. 70 00:03:24.080 --> 00:03:26.719 It's terrifying when you would like that. Yeah, So to 71 00:03:26.759 --> 00:03:29.400 grasp the sheer scale of this trap, I feel like 72 00:03:29.400 --> 00:03:30.039 we have to do so. 73 00:03:30.599 --> 00:03:33.639 Right now, the blueprint of the universe is dominated by 74 00:03:33.680 --> 00:03:34.919 the Standard Model of. 75 00:03:34.919 --> 00:03:37.439 Cosmology, which has a very catchy name, right Oh. 76 00:03:37.360 --> 00:03:40.960 Yeah, LAMB to CDM, very clunky LAMB to CDM. 77 00:03:41.439 --> 00:03:44.319 But it is the raining heavyweight champion of physics right now. 78 00:03:44.360 --> 00:03:47.080 It absolutely is. It tells the mathematical story of how 79 00:03:47.120 --> 00:03:50.520 we got from the literal dawn of time to the vast, 80 00:03:50.680 --> 00:03:53.919 complex cosmic web we observe in the night sky today. 81 00:03:54.120 --> 00:03:56.120 And it does a really good job at certain things, 82 00:03:56.159 --> 00:03:56.639 doesn't it. 83 00:03:56.639 --> 00:03:59.479 It does a brilliant job. Actually. It explains the rate 84 00:03:59.520 --> 00:04:02.680 at which the universes expanding, for one, right, and it 85 00:04:02.719 --> 00:04:05.319 accounts for the existence of cold dark matter, which is 86 00:04:05.400 --> 00:04:08.719 basically the invisible scaffolding that holds galaxies together. 87 00:04:08.840 --> 00:04:13.960 Okay, so lamba CDM is this monumental achievement of human intellect. 88 00:04:14.120 --> 00:04:17.480 It works, It provides a highly accurate description of the 89 00:04:17.560 --> 00:04:18.959 large scale structure of the. 90 00:04:18.920 --> 00:04:21.399 Cosmos, and there's always a button always. 91 00:04:21.480 --> 00:04:24.639 Every physicist knows it is an incomplete story, ye it is. 92 00:04:25.240 --> 00:04:27.439 It's a brilliant rough. 93 00:04:27.279 --> 00:04:30.240 Draft, right because the closer we look the glorier that 94 00:04:30.279 --> 00:04:32.319 initial cosmic X ray becomes. 95 00:04:32.360 --> 00:04:36.759 Exactly, observations have started pointing toward anomalies that simply do 96 00:04:36.839 --> 00:04:40.319 not fit neatly into that lambda CDM framework. 97 00:04:40.480 --> 00:04:43.879 So scientists are desperate to find what they broadly just 98 00:04:43.920 --> 00:04:45.079 call new physics. 99 00:04:45.199 --> 00:04:50.160 Right, new physics. We are talking about universe altering variables. 100 00:04:49.519 --> 00:04:51.199 Here, Like, what's an example of that? 101 00:04:51.639 --> 00:04:55.199 Well, for instance, the effects of massive neutrinos okay, or 102 00:04:55.360 --> 00:04:59.600 modified theories of gravity that suggest our fundamental understanding of 103 00:04:59.600 --> 00:05:03.480 how attracts mass might need a really massive overhaul. 104 00:05:03.639 --> 00:05:05.560 Wow, So changing gravity itself. 105 00:05:05.759 --> 00:05:09.199 Yeah. And there is even the tantalizing possibility of an 106 00:05:09.240 --> 00:05:10.519 evolving dark. 107 00:05:10.399 --> 00:05:12.600 Energy meaning it changes, right, a. 108 00:05:12.560 --> 00:05:15.879 Repulsive force that doesn't just stay constant, but actually changes 109 00:05:15.920 --> 00:05:19.360 its behavior and strength over a vast cosmic time scale. 110 00:05:19.399 --> 00:05:23.319 That is mind bending. But exploring those possibilities requires a 111 00:05:23.319 --> 00:05:25.519 completely different approach to science, doesn't it. 112 00:05:25.519 --> 00:05:28.040 It really does. I mean, think about how normal science works. 113 00:05:28.040 --> 00:05:30.600 If a biologist wants to test a new variable, they 114 00:05:30.600 --> 00:05:33.000 just set up a controlled experiment in a Petrie dish. 115 00:05:32.839 --> 00:05:35.879 Right, they had a chemical see what the bacteria do exactly? 116 00:05:36.120 --> 00:05:38.920 Or if an engineer wants to test a new aerodynamic shape. 117 00:05:39.079 --> 00:05:40.720 They build a model and they put it in a 118 00:05:40.720 --> 00:05:41.279 wind tunnel. 119 00:05:41.720 --> 00:05:44.360 But you can't exactly put the universe in a Petrie dish. 120 00:05:44.480 --> 00:05:46.720 No you cannot. You can't just tweak the dial on 121 00:05:46.759 --> 00:05:49.519 the gravitational constant of the real universe and observe what 122 00:05:49.560 --> 00:05:51.000 happens to a galaxy cluster. 123 00:05:51.199 --> 00:05:51.959 I mean, I hope not. 124 00:05:52.040 --> 00:05:55.000 That sounds dangerous, right, So the only option left is 125 00:05:55.040 --> 00:05:56.279 to build a fake universe. 126 00:05:56.360 --> 00:05:59.759 You literally have to construct a digital cosmos from scratch. 127 00:05:59.680 --> 00:06:04.319 Which it brings us to the reliance on massive, extraordinarily 128 00:06:04.399 --> 00:06:07.360 detailed computer simulations. 129 00:06:06.519 --> 00:06:07.920 Like the Quixote simulations. 130 00:06:08.000 --> 00:06:10.800 Yes, the Keyote simulations are a huge cornerstone of this 131 00:06:10.879 --> 00:06:14.319 kind of research. The methodology here is just fascinating. 132 00:06:14.560 --> 00:06:15.800 How does it actually work? 133 00:06:16.040 --> 00:06:20.079 Well? Researchers take a perfectly identical region of a virtual universe, 134 00:06:20.199 --> 00:06:24.040 just a massive cube of simulated space, and they run 135 00:06:24.079 --> 00:06:26.680 it through time over and over again. Okay, but each 136 00:06:26.720 --> 00:06:29.839 time they run it they change the underlying mathematical laws 137 00:06:29.839 --> 00:06:31.040 of physics. 138 00:06:30.600 --> 00:06:33.360 So they are essentially running an ab test on reality. 139 00:06:33.639 --> 00:06:36.399 That is exactly what it is, an ab test on 140 00:06:36.439 --> 00:06:40.639 the universe. They take the standard lambda CDM model and 141 00:06:40.680 --> 00:06:41.199 they hit. 142 00:06:41.319 --> 00:06:43.120 Play and they just watch what happens. 143 00:06:43.279 --> 00:06:45.759 Yeah, they watch how the dark matter and the gas 144 00:06:45.839 --> 00:06:50.000 coalesce into galaxies over billions of simulated years. 145 00:06:50.120 --> 00:06:51.639 Okay, that's the a test, right. 146 00:06:52.000 --> 00:06:55.600 Then they take that exact same starting configuration, that same 147 00:06:55.639 --> 00:07:00.360 digital cube, but this time they inject the new physics. 148 00:07:00.160 --> 00:07:03.120 So they sprinkle in some massive neutrinos, or. 149 00:07:03.120 --> 00:07:05.639 They tweak the laws of gravity exactly, and then they 150 00:07:05.720 --> 00:07:06.399 hit play. 151 00:07:06.199 --> 00:07:07.879 Again and then comes the comparison. 152 00:07:08.079 --> 00:07:11.040 Yes, yes, they lay the final snapshots of these two 153 00:07:11.199 --> 00:07:12.920 virtual universes side by. 154 00:07:12.800 --> 00:07:16.199 Side, and I imagine they aren't looking for like giant 155 00:07:16.240 --> 00:07:18.519 neon signs saying gravity changed here. 156 00:07:18.720 --> 00:07:20.639 No, not at all. What they're looking for are not 157 00:07:20.839 --> 00:07:25.360 massive glaring differences, but incredibly subtle structural shift. 158 00:07:25.600 --> 00:07:26.519 What give me a. 159 00:07:26.519 --> 00:07:29.319 Visual like, did a massive dark matter halo stretch may 160 00:07:29.360 --> 00:07:31.920 be a fraction of a percent more in the modified universe? 161 00:07:31.959 --> 00:07:33.040 Wow, fraction of a percent? 162 00:07:33.160 --> 00:07:35.839 Or did the cosmic web, you know, the massive filaments 163 00:07:35.879 --> 00:07:39.079 of matter that connect galaxy clusters, did they clump together differently? 164 00:07:39.319 --> 00:07:41.839 Okay, So if you are listening to this, you might 165 00:07:41.879 --> 00:07:44.399 be thinking what I'm thinking right now, which is wait 166 00:07:44.439 --> 00:07:49.079 a second, mike smartphone can process billions of calculations a second. Right, 167 00:07:49.560 --> 00:07:53.199 we have server farms that span entire city blocks. Why 168 00:07:53.319 --> 00:07:57.000 is running a computer simulation such a bottlemex. 169 00:07:56.399 --> 00:07:57.279 It's a fair question. 170 00:07:57.439 --> 00:07:59.920 Why can't we just generate a million of these virtual 171 00:08:00.120 --> 00:08:03.079 universes by next Tuesday and just comb through the. 172 00:08:03.079 --> 00:08:09.519 Data, Because that intuition completely underestimates the sheer computational violence 173 00:08:09.680 --> 00:08:11.680 of an n body universe simulation. 174 00:08:11.920 --> 00:08:13.959 Computational violence. I love that phrase. 175 00:08:14.079 --> 00:08:17.560 It really is. You are not rendering a static image. 176 00:08:17.920 --> 00:08:21.759 You are calculating the gravitational pull of billions upon billions 177 00:08:21.759 --> 00:08:23.600 of individual particles of matter. 178 00:08:23.759 --> 00:08:26.399 Let's try to ground this in a visual for everyone. Okay, yeah, 179 00:08:26.480 --> 00:08:31.040 imagine rendering a highly detailed four K open world video 180 00:08:31.120 --> 00:08:31.879 game environment. 181 00:08:32.000 --> 00:08:34.000 Oh, like a massive RPG exactly. 182 00:08:34.519 --> 00:08:37.200 The computer has to calculate the lighting, the shadows, the 183 00:08:37.200 --> 00:08:39.960 physics of the wind moving through the leaves on a tree. 184 00:08:39.679 --> 00:08:41.639 The reflections in a puddle on the ground. 185 00:08:41.919 --> 00:08:44.200 Right, and if you blow up a building in the game, 186 00:08:44.519 --> 00:08:47.799 the physics engine has to track the trajectory of every 187 00:08:47.799 --> 00:08:49.159 single piece of debris. 188 00:08:48.799 --> 00:08:50.679 Which takes a lot of processing power. 189 00:08:50.480 --> 00:08:53.279 Even with the best graphics cards on Earth. If there's 190 00:08:53.360 --> 00:08:57.120 too much happening on screen, the game lags. The processor 191 00:08:57.159 --> 00:08:59.039 simply cannot do the math fast enough. 192 00:08:59.159 --> 00:09:00.000 It drops frames. 193 00:09:00.320 --> 00:09:03.919 Right now, take that video game environment, but instead of 194 00:09:03.919 --> 00:09:06.679 a forest or a city block, you have to render 195 00:09:06.759 --> 00:09:10.559 every single massive structure, every galaxy, every dark matter halo 196 00:09:11.240 --> 00:09:15.200 in a swath of the cosmos, millions of light years across. 197 00:09:15.559 --> 00:09:18.480 And crucially, and this is the part that breaks supercomputers, 198 00:09:18.639 --> 00:09:21.679 every single particle in that simulation is interacting with every 199 00:09:21.679 --> 00:09:23.200 other particle simultaneously. 200 00:09:23.480 --> 00:09:25.200 Wait, really, every single one. 201 00:09:25.320 --> 00:09:28.320 Yes, because gravity has infinite range. If you move one 202 00:09:28.360 --> 00:09:32.200 particle a millimeter, it's gravitational pull on a particle a 203 00:09:32.240 --> 00:09:35.120 billion light years away technically changes Oh my god. And 204 00:09:35.200 --> 00:09:37.919 so you have to recalculate the entire system. 205 00:09:37.600 --> 00:09:38.639 Every time anything moves. 206 00:09:38.720 --> 00:09:41.080 And you have to do this frame by frame, step 207 00:09:41.080 --> 00:09:45.200 by microscopic timestep over a timeline of fourteen billion years. 208 00:09:45.639 --> 00:09:48.279 That is just I can't even wrap my head around 209 00:09:48.320 --> 00:09:48.679 that math. 210 00:09:48.840 --> 00:09:53.720 The sheer volume of mathematical integration required is unfathomable. It 211 00:09:53.840 --> 00:09:57.320 literally brings the most powerful supercomputers on the planet to 212 00:09:57.399 --> 00:09:58.039 their knees. 213 00:09:58.399 --> 00:10:01.240 But those simulations are non negoti right, Like we have 214 00:10:01.320 --> 00:10:01.799 to do them. 215 00:10:01.919 --> 00:10:04.759 They are the only way forward. Those tiny variations in 216 00:10:04.799 --> 00:10:07.639 how madd clusters in a Quixote simulation compared to a 217 00:10:07.679 --> 00:10:11.879 standard model simulation, those are the exact breadcrumbs scientists are 218 00:10:11.960 --> 00:10:12.960 desperately looking for. 219 00:10:13.240 --> 00:10:15.840 Right if we want to find the faint whispers of 220 00:10:15.960 --> 00:10:19.240 new physics hidden in the grand structure of the cosmos, 221 00:10:19.679 --> 00:10:22.799 we cannot afford to cut corners on the physics engine itself. 222 00:10:22.919 --> 00:10:26.120 We can't. But the computational cost in terms of time, 223 00:10:26.200 --> 00:10:30.480 electricity and just hardware or wear and tearor is simply unsustainable. 224 00:10:30.759 --> 00:10:32.480 So researchers hit a wall. 225 00:10:32.440 --> 00:10:36.240 A massive wall brute force computing. Just building a bigger, hotter, 226 00:10:36.360 --> 00:10:39.960 more expensive supercomputer was no longer a viable. 227 00:10:39.600 --> 00:10:41.440 Path because it just scales too poorly. 228 00:10:41.600 --> 00:10:44.039 Exactly. They had to find a radically smarter way to 229 00:10:44.080 --> 00:10:47.360 process the data. They couldn't change the complexity of the universe, 230 00:10:47.399 --> 00:10:49.480 so they had to change the efficiency of the observer. 231 00:10:49.919 --> 00:10:53.000 And this is where we see this brilliant pivot. 232 00:10:52.799 --> 00:10:54.720 Yes, a total paradigm shift. 233 00:10:54.879 --> 00:10:58.919 Researchers like Vienna Krishnaraj and Adrian Bayer from Princeton University 234 00:10:58.919 --> 00:11:02.200 in the Flatirn Institute, they turned to a specific machine 235 00:11:02.240 --> 00:11:04.919 learning technique to basically bypass this. 236 00:11:04.960 --> 00:11:06.840 Bottleneck, and it's called transfer learning. 237 00:11:06.919 --> 00:11:09.960 Transfer learning. Oh okay, So instead of taking a completely 238 00:11:10.000 --> 00:11:13.480 blank artificial neural network and just forcing it to learn 239 00:11:14.000 --> 00:11:18.559 from the most complex, expensive new physics simulations right out. 240 00:11:18.480 --> 00:11:20.879 Of the gate, which would take forever, right. 241 00:11:20.919 --> 00:11:22.960 They completely changed the curriculum. 242 00:11:22.600 --> 00:11:26.120 They utilized pre training. They took the AI and they 243 00:11:26.120 --> 00:11:30.519 first trained it exclusively on the simpler, vastly cheaper simulations, the. 244 00:11:30.440 --> 00:11:33.440 Ones based entirely on the standard LAMB to CDM model. 245 00:11:33.559 --> 00:11:37.440 Exactly. They allowed the AI to build a baseline mathematical 246 00:11:37.519 --> 00:11:41.519 understanding of how a standard unexceptional universe operates. 247 00:11:41.679 --> 00:11:43.519 So they taught at the basics first right. 248 00:11:43.840 --> 00:11:46.919 Only after the neural network had established its foundational weights 249 00:11:46.919 --> 00:11:50.320 and biases. Only after it had mastered the basic rules 250 00:11:50.360 --> 00:11:55.200 of gravity and clustering, did they introduce the sophisticated, expensive models, the. 251 00:11:55.200 --> 00:11:58.279 Ones containing the modified gravity or the massive neutrinos. 252 00:11:58.399 --> 00:12:01.279 Yes, Adrian Bayer had perfect analogy for this. 253 00:12:01.600 --> 00:12:03.559 Oh, I love a good analogy, let's hear it. 254 00:12:03.639 --> 00:12:06.720 He compares this approach to the way humans learn from textbooks. 255 00:12:07.440 --> 00:12:10.200 You do not hand a first year medical student the 256 00:12:10.240 --> 00:12:15.080 most advanced cutting Edge Journal on Experimental Neurosurgery on their 257 00:12:15.120 --> 00:12:16.000 first day of class. 258 00:12:16.080 --> 00:12:18.960 No, they would just cry. It would be entirely incomprehensible. 259 00:12:19.080 --> 00:12:22.240 Right, You give them a basic anatomy textbook. You have 260 00:12:22.360 --> 00:12:26.639 the memorize the standard placement of bones, organs and arteries, 261 00:12:26.840 --> 00:12:29.639 the foundation. They learn the basic book to grasp the 262 00:12:29.679 --> 00:12:33.840 foundational concepts, and once that architecture is mapped in their mind, 263 00:12:34.080 --> 00:12:36.600 they can move on to the really complicated book that. 264 00:12:36.600 --> 00:12:40.120 Makes total sense. Vina Krishniaj, the lead author of the study. 265 00:12:40.320 --> 00:12:44.080 She pointed out that this phased strategy prevents the AI 266 00:12:44.159 --> 00:12:45.879 from having to digest everything. 267 00:12:45.519 --> 00:12:48.480 All at once, exactly, because when a neural network starts 268 00:12:48.480 --> 00:12:51.440 from scratch, it doesn't know what a galaxy is. 269 00:12:51.679 --> 00:12:53.159 It doesn't know what empty space is. 270 00:12:53.519 --> 00:12:56.120 Right, It is just looking at raw numbers. But by 271 00:12:56.159 --> 00:12:58.600 pre training it on the cheaper simulations, it learns to 272 00:12:58.639 --> 00:13:02.559 recognize the broad, easily calculable rules of the standard model. 273 00:13:02.840 --> 00:13:06.879 So then when it's finally confronted with the complex new physics, 274 00:13:07.360 --> 00:13:11.240 it doesn't have to relearn what gravity is from scratch exactly. 275 00:13:11.519 --> 00:13:14.559 It only has to fine tune its existing knowledge to 276 00:13:14.639 --> 00:13:16.879 account for the subtle anomalies. 277 00:13:16.360 --> 00:13:19.080 Here's where it gets really interesting to me, because it's 278 00:13:19.159 --> 00:13:21.919 exactly like teaching a kid to ride a bike. Oh 279 00:13:22.039 --> 00:13:25.320 how so, Well, you don't take a six year old, 280 00:13:25.879 --> 00:13:28.720 put them on a twenty one speed carbon fiber mountain 281 00:13:28.720 --> 00:13:31.039 bike and just shove them down a steep, rocky trail. 282 00:13:31.159 --> 00:13:33.159 I mean, I wouldn't you start them on a simple 283 00:13:33.200 --> 00:13:36.759 tricycle on a flat driveway. Right. You let their brain 284 00:13:37.200 --> 00:13:41.039 build the fundamental muscle memory of balance, of pedaling, of 285 00:13:41.279 --> 00:13:45.120 how turning the handlebars correlates with changing direction, the basics. 286 00:13:45.200 --> 00:13:48.559 The tricycle is the simple LAMB to CDM simulation, and 287 00:13:48.639 --> 00:13:51.919 once their nervous system has mapped those basic physical relationships, 288 00:13:52.120 --> 00:13:55.159 transferring those skills to the complex mountain bike is infinitely 289 00:13:55.200 --> 00:13:57.399 faster and easier than if they had started on the 290 00:13:57.399 --> 00:13:58.440 mountain bike on day one. 291 00:13:58.639 --> 00:14:00.360 That is a brilliant way to put it, and the 292 00:14:00.399 --> 00:14:06.120 results of applying this textbook strategy to cosmology were absolutely staggering. 293 00:14:06.480 --> 00:14:10.200 By employing transfer learning, the researchers were able to reduce 294 00:14:10.200 --> 00:14:14.320 the need for those incredibly expensive complex simulations by a 295 00:14:14.320 --> 00:14:15.559 factor of more than ten. 296 00:14:16.000 --> 00:14:18.480 A factor of ten. We really need to sit with 297 00:14:18.480 --> 00:14:19.120 that for a second. 298 00:14:19.279 --> 00:14:21.080 Yeah, that's not just a small improvement, It. 299 00:14:21.080 --> 00:14:23.799 Is a revolution and workflow. A project that would have 300 00:14:23.840 --> 00:14:28.639 required a decade of constant supercomputer time, consuming millions of 301 00:14:28.679 --> 00:14:33.000 dollars in electricity and hardware, wear and tear, can suddenly 302 00:14:33.000 --> 00:14:34.919 be completed in a single year. 303 00:14:35.200 --> 00:14:38.480 It completely shifts the paradigm of what is actually possible 304 00:14:38.559 --> 00:14:39.840 in astronomical research. 305 00:14:39.919 --> 00:14:43.639 It does, and the underlying mechanism here mirrors the exact 306 00:14:43.679 --> 00:14:47.240 techniques driving the current explosion in generative AI and large 307 00:14:47.320 --> 00:14:48.320 language models. 308 00:14:48.080 --> 00:14:49.960 Right like the chatbots everyone uses. 309 00:14:49.720 --> 00:14:54.320 Now exactly The entire modern artificial intelligence industry relies on 310 00:14:54.519 --> 00:14:55.960 foundation models. 311 00:14:55.600 --> 00:14:57.879 Where you pre train a massive model on billions of 312 00:14:57.879 --> 00:15:01.799 pages of basic human text so fundament understands grammar and context, 313 00:15:02.080 --> 00:15:02.720 and then. 314 00:15:03.080 --> 00:15:06.200 You fine tune it for a specific task like writing 315 00:15:06.279 --> 00:15:09.360 code or drafting legal documents. 316 00:15:08.799 --> 00:15:12.639 Which dramatically speeds up inference and slashes training costs. 317 00:15:13.000 --> 00:15:17.440 The researcher successfully took this cornerstone of computer science and 318 00:15:17.480 --> 00:15:20.960 applied it to the structural analysis of the universe itself. 319 00:15:21.320 --> 00:15:24.120 So saving ten times the computing power sounds like the 320 00:15:24.240 --> 00:15:25.879 absolute ultimate triumph. 321 00:15:26.360 --> 00:15:27.159 It really does. 322 00:15:27.240 --> 00:15:29.399 It feels like the moment in the movie where the 323 00:15:29.440 --> 00:15:31.840 scientists cracked the code. They popped the champagne and the 324 00:15:31.840 --> 00:15:33.399 credits role it would be. 325 00:15:33.320 --> 00:15:38.039 If the story ended with computational efficiency. Ah, but speed 326 00:15:38.120 --> 00:15:39.879 comes at a psychological cost. 327 00:15:39.720 --> 00:15:43.200 Because learning from a basic textbook creates a very specific, 328 00:15:43.279 --> 00:15:46.720 incredibly dangerous kind of blindness. Yes, this is where the 329 00:15:46.720 --> 00:15:49.799 triumph of transfer learning takes a really sharp turn into 330 00:15:49.879 --> 00:15:53.080 the psychological pitfalls of artificial intelligence. 331 00:15:53.159 --> 00:15:56.440 The researchers uncovered a severe downside to their brilliant shortcut. 332 00:15:56.480 --> 00:15:58.600 It's a phenomenon known as negative transfer. 333 00:15:58.639 --> 00:16:01.200 Negative transfer it sounds what exactly is it? 334 00:16:01.480 --> 00:16:04.600 Negative transfer is exactly what it sounds like. It happens 335 00:16:04.679 --> 00:16:08.159 when an AI relies so heavily on the familiar foundational 336 00:16:08.159 --> 00:16:11.000 patterns that learned during its pre training phase that it 337 00:16:11.039 --> 00:16:14.840 completely misses or misinterprets the evidence of something genuinely new. 338 00:16:15.039 --> 00:16:19.200 So the AI becomes heavily biased by its own baseline education. 339 00:16:19.480 --> 00:16:22.559 Yes, the pre trained neural network is no longer a 340 00:16:22.600 --> 00:16:26.919 blank slate. Its internal parameters have been deeply, deeply shaped 341 00:16:26.960 --> 00:16:28.080 by the standard model. 342 00:16:28.360 --> 00:16:31.279 It is internalized the assumption that LAMB to CDM is 343 00:16:31.320 --> 00:16:32.320 the definitive truth. 344 00:16:32.480 --> 00:16:36.759 Exactly, So, when the researchers finally introduce the complex simulations 345 00:16:36.799 --> 00:16:40.720 containing the new physics, the massive neutrinos or the modified gravity. 346 00:16:41.000 --> 00:16:43.799 The AI does not view this new data objectively. 347 00:16:44.120 --> 00:16:48.399 It interprets the unfamiliar information entirely through the rigid, tinted 348 00:16:48.519 --> 00:16:50.159 lens of what it already knows. 349 00:16:50.679 --> 00:16:53.360 To really understand the mechanics of this danger, we need 350 00:16:53.399 --> 00:16:56.159 to extend that medical diagnosis analogy you brought up earlier. 351 00:16:56.200 --> 00:16:56.919 Okay, let's do it. 352 00:16:57.039 --> 00:17:00.399 Imagine that medical student, learning from the introductory textbook. They 353 00:17:00.399 --> 00:17:03.320 spend months memorizing the symptoms of the common cold, A 354 00:17:03.399 --> 00:17:05.720 running nose, a slight fever, lethargy. 355 00:17:06.240 --> 00:17:07.480 Very standard stuff. 356 00:17:07.359 --> 00:17:11.640 Right, Those patterns are neurologically ingrained. The heuristic is set 357 00:17:11.680 --> 00:17:14.519 in their brain. So years later they are a practicing 358 00:17:14.519 --> 00:17:17.480 physician and a patient walks into the clinic with a 359 00:17:17.559 --> 00:17:20.880 running nose, a slight fever, and lethargy, and. 360 00:17:20.880 --> 00:17:25.000 The doctor's brain, seeking efficiency, immediately fires off the established 361 00:17:25.000 --> 00:17:26.799 pattern common cold. 362 00:17:26.680 --> 00:17:29.759 The cognitive path of least resistance exactly. But what the 363 00:17:29.839 --> 00:17:33.720 doctor fails to realize because they are so blinded by 364 00:17:33.720 --> 00:17:36.720 the familiarity of the symptoms, is that the patient actually 365 00:17:36.720 --> 00:17:41.359 has a vanishingly rare, highly lethal autoimmune disease that just 366 00:17:41.440 --> 00:17:43.799 happens to share those exact early symptoms. 367 00:17:44.200 --> 00:17:48.480 Wow. So in this scenario, the doctor's existing knowledge, the 368 00:17:48.480 --> 00:17:52.480 foundational education from the introductory textbook, actively causes harm. 369 00:17:52.640 --> 00:17:55.319 It acts as a set of cognitive blinders. 370 00:17:54.960 --> 00:17:58.680 Because they have seen this symptom cluster a thousand times before, 371 00:17:59.319 --> 00:18:03.200 their brain forcefully encourages the wrong, familiar conclusion. 372 00:18:03.279 --> 00:18:06.279 Yes, if that doctor had been a completely blank slate, 373 00:18:06.640 --> 00:18:08.960 or if they hadn't relied on the heuristic shortcut, they 374 00:18:09.039 --> 00:18:10.000 might have looked closer. 375 00:18:10.279 --> 00:18:12.079 They might have ordered bloodwork, they might have. 376 00:18:12.160 --> 00:18:15.039 Noticed a subtle anomaly in the white blood cell count 377 00:18:15.119 --> 00:18:16.720 and made the correct diagnosis. 378 00:18:17.000 --> 00:18:20.400 But their prior knowledge actively pushed them away from the truth. 379 00:18:20.640 --> 00:18:24.319 And that cognitive trap is negative transfer, and the AI 380 00:18:24.440 --> 00:18:28.160 systems observing the simulated universes fell right into it. 381 00:18:28.240 --> 00:18:29.160 That is wild. 382 00:18:29.359 --> 00:18:32.920 In some of the simulations, the mathematical signatures left behind 383 00:18:32.960 --> 00:18:36.640 by the new physics produced patterns that superficially resembled the 384 00:18:36.680 --> 00:18:40.359 patterns the AI had already associated with the standard cosmological model. 385 00:18:40.640 --> 00:18:45.519 So is the AI actually failing in these moments or 386 00:18:45.640 --> 00:18:47.960 is it just doing exactly what we trained it to do? 387 00:18:48.240 --> 00:18:49.799 That is the big question. 388 00:18:49.960 --> 00:18:53.000 I mean, we built the neural network to find patterns, 389 00:18:53.000 --> 00:18:56.240 to optimize for efficiency, and to rely on its prior 390 00:18:56.319 --> 00:18:59.599 experience to minimize its error rate. Can we really blame 391 00:18:59.599 --> 00:19:03.160 the machine for being an exceptional student of a flawed curriculum. 392 00:19:03.279 --> 00:19:06.200 No, we can't. It is entirely a philosophical failure on 393 00:19:06.240 --> 00:19:07.759 our part, not a computational one. 394 00:19:07.799 --> 00:19:11.680 The algorithm is executing its mathematical gradient dissent flawlessly. 395 00:19:11.880 --> 00:19:15.720 Flawlessly. The mistake belongs entirely to the architects of the 396 00:19:15.799 --> 00:19:19.480 learning process. By pre training the AI heavily on the 397 00:19:19.559 --> 00:19:23.960 lambda CDM model, we implicitly taught the machine that the 398 00:19:24.000 --> 00:19:26.920 standard universe is the default state of reality. 399 00:19:27.000 --> 00:19:30.000 We hard coded a conservative bias into its worldview. 400 00:19:30.079 --> 00:19:33.640 We did so. When the AI encounters a slight deviation 401 00:19:33.759 --> 00:19:36.759