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 --> 00:00:19.000 examine the ideas shaping the future of artificial minds. 5 00:00:23.879 --> 00:00:27.399 Welcome back today we are we're tackling a subject that 6 00:00:27.440 --> 00:00:29.559 honestly feels a little bit like we're stepping into the 7 00:00:29.600 --> 00:00:32.240 middle of a massive theoretical physics debate. 8 00:00:32.719 --> 00:00:33.600 Yeah, it really does. 9 00:00:33.640 --> 00:00:36.679 But the twist here is that it's actually happening right now, 10 00:00:37.119 --> 00:00:39.359 Like in engineering labs all over the world. 11 00:00:39.640 --> 00:00:43.079 It is. And it's one of those topics where if 12 00:00:43.119 --> 00:00:46.280 you just pay attention to the headlines, you might think 13 00:00:46.320 --> 00:00:48.840 you know the whole story. You see the press releases 14 00:00:48.880 --> 00:00:53.079 and the hype cycles, But the real story, the actual 15 00:00:53.320 --> 00:00:56.920 engineering reality of it all, is hiding deep in the details. 16 00:00:57.039 --> 00:01:00.159 Exactly so for everyone listening, let's set the scene for 17 00:01:00.200 --> 00:01:02.799 this deep dive. If you were to ask someone to 18 00:01:02.920 --> 00:01:06.920 name the two most you know, buzzworthy, world changing technologies 19 00:01:06.959 --> 00:01:09.560 of the twenty first century, I think pretty much everyone 20 00:01:09.599 --> 00:01:10.760 would give you the same two. 21 00:01:10.560 --> 00:01:12.959 Answers, Artificial intelligence and quantum computing. 22 00:01:13.120 --> 00:01:17.000 Right, AI and quantum and usually when we think about them, 23 00:01:17.000 --> 00:01:20.120 we picture them as these entirely parallel tracks. 24 00:01:19.799 --> 00:01:21.239 Two totally separate disciplines. 25 00:01:21.519 --> 00:01:24.200 Yeah, exactly that. You know, over here in this one 26 00:01:24.239 --> 00:01:28.159 building you have the AI researchers building these massive, large 27 00:01:28.239 --> 00:01:32.920 language models. They're obsessing overweights and biases and neural nets. Sure, 28 00:01:32.959 --> 00:01:35.959 and then way over there, maybe in some super cooled 29 00:01:36.000 --> 00:01:38.959 underground lab, you have the physicists and they're just trying 30 00:01:38.959 --> 00:01:42.319 to isolate a quibit. They feel like entirely separate worlds. 31 00:01:42.560 --> 00:01:45.560 That is definitely the common perception, right, It's almost like 32 00:01:45.599 --> 00:01:49.760 two different genres of science, right. One is all about software, right, 33 00:01:49.799 --> 00:01:54.159 pattern recognition, statistics, massive data sets, and the other is 34 00:01:54.519 --> 00:01:59.120 deeply rooted in hardware, coherent states and fundamental physics. But 35 00:01:59.159 --> 00:02:01.079 what we found in the en Else's for today is 36 00:02:01.120 --> 00:02:04.159 that this separation is actually a huge misconception. 37 00:02:04.400 --> 00:02:06.599 That was a big aha moment for me reading through 38 00:02:06.640 --> 00:02:10.159 these sources. They aren't just parallel lines. They are quite 39 00:02:10.199 --> 00:02:11.719 literally entangled. 40 00:02:11.759 --> 00:02:14.159 They are growing up together. Yeah, and it's not just 41 00:02:14.199 --> 00:02:16.080 that they happen to exist at the same time in history. 42 00:02:16.120 --> 00:02:17.840 The relationship is much more intimate than that. 43 00:02:17.879 --> 00:02:19.039 They're accelerating each. 44 00:02:18.919 --> 00:02:23.759 Other precisely, they are accelerating each other. In fact, the 45 00:02:23.800 --> 00:02:26.039 sources we're looking at make a pretty strong case that 46 00:02:26.080 --> 00:02:28.280 they are becoming intimately dependent on one another. 47 00:02:28.639 --> 00:02:30.879 So that's our mission today for you guys listening. We 48 00:02:30.960 --> 00:02:34.400 are going to explore this reciprocal partnership. We really want 49 00:02:34.400 --> 00:02:39.599 to understand how AI is essentially acting as the mechanic 50 00:02:39.639 --> 00:02:42.719 that is building quantum computers right now, yes, and then 51 00:02:42.960 --> 00:02:44.960 we're going to flip the script and look at how 52 00:02:45.080 --> 00:02:48.919 quantum computers might eventually completely revolutionize AI. 53 00:02:49.240 --> 00:02:53.240 It's a fascinating story of a feedback loop. But before 54 00:02:53.280 --> 00:02:55.560 we get into that feedback loop, we have to understand 55 00:02:55.560 --> 00:02:56.680 the players, right. 56 00:02:56.719 --> 00:02:58.400 We need to set the stage exactly. 57 00:02:58.439 --> 00:03:00.520 We need to set the stage with the fun fundamental 58 00:03:00.520 --> 00:03:03.800 problem that both of these technologies are desperately trying to solve. 59 00:03:04.240 --> 00:03:06.080 And I really don't want to start with this standard 60 00:03:06.400 --> 00:03:07.719 what is a bit stuff? 61 00:03:07.800 --> 00:03:07.919 Oh? 62 00:03:08.039 --> 00:03:10.759 Please know right, our listeners know what a binary system is. 63 00:03:10.919 --> 00:03:13.039 Yeah, we definitely don't need to do the whole imaginal 64 00:03:13.120 --> 00:03:16.039 light switch analogy. Let's talk about complexity. Let's talk about 65 00:03:16.080 --> 00:03:17.319 the classical wall. 66 00:03:17.439 --> 00:03:20.719 The classical wall. I love that term because this isn't 67 00:03:20.800 --> 00:03:22.000 just about speed. 68 00:03:21.960 --> 00:03:23.800 It's not about making your laptop faster. 69 00:03:24.159 --> 00:03:26.719 No, it's not about how fast you can render a 70 00:03:26.759 --> 00:03:32.680 four K video or query a massive SQL database. Classical computers, 71 00:03:32.680 --> 00:03:36.520 whether it's your phone or a massive supercomputer like Frontier, 72 00:03:37.120 --> 00:03:40.520 they operate within a very specific complexity class. 73 00:03:40.199 --> 00:03:43.120 And this brings us right to P versus NP, right, 74 00:03:43.240 --> 00:03:47.120 and specifically the simulation of nature itself. 75 00:03:46.919 --> 00:03:49.759 That is the absolute crux of it. The sources highlight 76 00:03:49.879 --> 00:03:54.199 over and over that classical computers fundamentally struggle with problems 77 00:03:54.400 --> 00:03:57.080 that involve massive combinatorial spaces. 78 00:03:57.120 --> 00:03:59.360 And the specific example that keeps coming up in the 79 00:03:59.400 --> 00:04:03.319 literature is simulating quantum systems, like trying to model the 80 00:04:03.360 --> 00:04:05.560 Hamiltonian of a molecule. 81 00:04:05.159 --> 00:04:07.840 Right, And just to clarify, the Hamiltonian being the operator 82 00:04:08.000 --> 00:04:11.039 corresponding to the total energy of that system exactly. 83 00:04:11.120 --> 00:04:13.199 So, if you want to simulate how a molecule behaves, 84 00:04:13.280 --> 00:04:15.240 let's say you're trying to design a new drug or 85 00:04:15.240 --> 00:04:17.959 maybe a better catalyst for carbon capture, you have to 86 00:04:18.000 --> 00:04:19.879 track the interactions of all its electrons. 87 00:04:19.920 --> 00:04:23.800 But electrons are quantum objects. They exist in superposition, they're 88 00:04:24.000 --> 00:04:25.480 entangled with one another. 89 00:04:25.199 --> 00:04:26.959 And the math there gets brutal. 90 00:04:26.720 --> 00:04:31.160 So fast it's exponential. It is brutally exponentially. If you 91 00:04:31.199 --> 00:04:35.199 have a system with n electrons the state space, you 92 00:04:35.319 --> 00:04:39.480 need to simulate scales as two to the power of en. 93 00:04:39.920 --> 00:04:42.240 So, just to put that in perspective, if I add 94 00:04:42.279 --> 00:04:45.439 just one single electron to my simulation, I literally double 95 00:04:45.480 --> 00:04:46.800 the amount of memory required. 96 00:04:46.839 --> 00:04:49.920 You double it every single time, and that gets entirely 97 00:04:49.920 --> 00:04:53.160 out of hand almost instantly. To simulate a relatively simple 98 00:04:53.240 --> 00:04:57.040 molecule like say, caffeine, with perfect quantum fidelity on a 99 00:04:57.040 --> 00:05:00.959 classical machine, you would need a computer memory larger than 100 00:05:00.959 --> 00:05:04.000 the number of atoms in the observable universe. 101 00:05:03.759 --> 00:05:06.160 Which is just wild to think about. That is a 102 00:05:06.199 --> 00:05:08.800 hard limit. That's not something Moore's law is ever going 103 00:05:08.879 --> 00:05:11.160 to solve for us. You can't just throw more GPUs 104 00:05:11.199 --> 00:05:12.560 at that problem and hope for the best. 105 00:05:12.639 --> 00:05:16.319 You physically cannot build a classical computer big enough, right period. 106 00:05:16.560 --> 00:05:19.639 That's the classical wall. And that is exactly why Richard Feynman, 107 00:05:19.800 --> 00:05:22.399 way back in the nineteen eighties said that if you 108 00:05:22.439 --> 00:05:24.879 want to simulate nature, you'd better make it a quantum 109 00:05:24.920 --> 00:05:25.920 mechanical simulation. 110 00:05:26.199 --> 00:05:28.759 Enter the quibbot and again let's skip the basic it's 111 00:05:28.759 --> 00:05:31.600 a coin spinning in the air analogy. Because the real 112 00:05:31.680 --> 00:05:34.720 power here isn't just the superposition itself, right, it's the 113 00:05:34.959 --> 00:05:36.519 state space it opens up. 114 00:05:36.560 --> 00:05:41.560 It's the ability to manipulate that enormous exponential state space directly. 115 00:05:42.199 --> 00:05:46.079 A quantum computer with roughly three hundred perfectly functioning quibods 116 00:05:46.240 --> 00:05:49.959 could represent more states simultaneously than there are atoms in 117 00:05:50.000 --> 00:05:50.560 the universe. 118 00:05:50.600 --> 00:05:53.600 It doesn't just simulate the physics, no, it essentially is 119 00:05:53.639 --> 00:05:57.240 the physics. And for very specific problems like factoring large 120 00:05:57.240 --> 00:06:01.360 integers for breaking cryptography, or searching map of unstructured databases, 121 00:06:01.480 --> 00:06:05.600 or simulating these complex Hamiltonians, the speed up isn't just 122 00:06:05.639 --> 00:06:07.600 like a ten x or one hundred x improvement. 123 00:06:07.720 --> 00:06:11.279 It's asymptotic. Yeah, it completely changes the math. It turns 124 00:06:11.319 --> 00:06:15.800 a practically impossible exponential problem into a solvable polynomial one. 125 00:06:16.040 --> 00:06:18.399 We are literally talking about the difference between finding an 126 00:06:18.399 --> 00:06:20.680 answer in a few minutes versus finding it in the 127 00:06:20.720 --> 00:06:21.560 age of the universe. 128 00:06:21.680 --> 00:06:25.160 Okay, So if the math is so clearly undeniably superior, 129 00:06:25.199 --> 00:06:28.680 why are we still recording this deep dive on traditional silicon? 130 00:06:29.319 --> 00:06:30.360 What is the catch here? 131 00:06:30.439 --> 00:06:33.680 The catch is what the industry calls the hardware crisis. Right, 132 00:06:33.920 --> 00:06:36.519 We can write the beautiful math on a chalkboard all 133 00:06:36.639 --> 00:06:41.120 day long, but building the physical machine is an absolute nightmare. 134 00:06:41.639 --> 00:06:44.240 Quibbits are incredibly fragile. 135 00:06:43.959 --> 00:06:46.519 And when we say fragile, we aren't talking about, you know, 136 00:06:46.600 --> 00:06:47.839 dropping the chip on the floor. 137 00:06:48.040 --> 00:06:52.800 No, we're talking about information fragility decoherence. To maintain that 138 00:06:52.959 --> 00:06:56.759 delicate state of superposition, the quibbit has to be perfectly 139 00:06:56.879 --> 00:06:59.360 isolated from the outside environment. 140 00:06:59.079 --> 00:07:01.839 Because the universe is fundamentally noisy, very. 141 00:07:01.680 --> 00:07:08.120 Noisy heat, electromagnetic fluctuations, vibration, even a stray cosmic ray, 142 00:07:08.800 --> 00:07:12.000 literally everything around it is trying to couple with that quibbit. 143 00:07:11.800 --> 00:07:14.040 And the moment the environment interacts. 144 00:07:13.560 --> 00:07:16.199 With it, the wave function collapses. 145 00:07:15.759 --> 00:07:18.800 The quantum information just leaks out into the environment exactly. 146 00:07:18.839 --> 00:07:23.839 That is decoherence. Currently, our absolute best superconducting quibots can 147 00:07:23.920 --> 00:07:28.120 only hold their state for milliseconds, sometimes just microseconds, before 148 00:07:28.160 --> 00:07:29.920 they just degrade into random noise. 149 00:07:30.120 --> 00:07:32.680 So we have a theoretical computer that is more powerful 150 00:07:32.720 --> 00:07:36.319 than the universe itself, but it essentially crashes every thousandth 151 00:07:36.319 --> 00:07:36.879 of a second. 152 00:07:37.120 --> 00:07:39.279 That is a very fair summary of where we are 153 00:07:39.360 --> 00:07:42.560 right now. This is the NIIC Q era, the noisy 154 00:07:42.839 --> 00:07:48.000 intermediate scale quantum era. The hardware is just noisy. 155 00:07:47.680 --> 00:07:49.319 And this brings us to the first half of our 156 00:07:49.360 --> 00:07:53.720 partnership because this extreme fragility is exactly where artificial intelligence 157 00:07:53.720 --> 00:07:55.879 steps into the picture. Yes, this is the part of 158 00:07:55.920 --> 00:07:58.480 the sources I found so fascinating. We usually think of 159 00:07:58.519 --> 00:08:01.199 AI as the software thing that runs on the computer 160 00:08:01.319 --> 00:08:04.720 once it's built, but here the literature describes AI as 161 00:08:04.759 --> 00:08:08.319 the crucial tool used to actually build and fix the 162 00:08:08.360 --> 00:08:09.240 computer itself. 163 00:08:09.360 --> 00:08:13.399 It acts as the mechanic. Building a quantum computer requires 164 00:08:13.600 --> 00:08:18.680 solving an array of control problems that are individually enormously complex. 165 00:08:19.040 --> 00:08:22.399 You have to calibrate these tiny devices, route the microwave signals, 166 00:08:22.399 --> 00:08:26.639 perfectly correct errors on the fly. Human engineers simply cannot 167 00:08:26.720 --> 00:08:27.240 keep up with it. 168 00:08:27.519 --> 00:08:29.839 Let's look at calibration first, because the sources spend a 169 00:08:29.879 --> 00:08:31.600 lot of time on this. They mentioned that a quantum 170 00:08:31.639 --> 00:08:33.799 processor isn't just a plug and play device. You don't 171 00:08:33.840 --> 00:08:34.799 just hit a power button. 172 00:08:34.960 --> 00:08:39.519 Far from it. Imagine you have a superconducting processor. Let's 173 00:08:39.519 --> 00:08:42.519 say it has one hundred and twenty seven quibits. Each 174 00:08:42.679 --> 00:08:47.080 one of those individual quibits has its own unique resonant frequency, 175 00:08:47.159 --> 00:08:50.559 It has a specific in harmonicity, it has a highly 176 00:08:50.600 --> 00:08:54.080 specific coupling strength to the quibbots sitting right next to it. 177 00:08:54.159 --> 00:08:56.559 And the frustrating part is those properties aren't static. 178 00:08:56.679 --> 00:09:01.559 They constantly drift. They are notoriously unstable. As the system ages, 179 00:09:01.919 --> 00:09:05.320 or even if the temperature and the massive dilution refrigerator 180 00:09:05.519 --> 00:09:09.519 fluctuates by just a fraction of a millikelvin, those parameters shift. 181 00:09:09.679 --> 00:09:11.519 It's like trying to play a grand piano where the 182 00:09:11.519 --> 00:09:14.320 strings are just constantly loosening and tightening on their own 183 00:09:14.559 --> 00:09:15.879 while you're trying to play a concerto. 184 00:09:16.080 --> 00:09:19.799 That is a perfect analogy. You have to retune it constantly. 185 00:09:20.120 --> 00:09:22.720 And the old method for doing this was manual calibration 186 00:09:23.200 --> 00:09:26.720 or using very traditional linear graph based models. 187 00:09:26.360 --> 00:09:28.600 Which means you'd have to stop everything you're doing. 188 00:09:28.399 --> 00:09:30.840 Right, You stop the computation and you run a long 189 00:09:30.879 --> 00:09:34.120 series of traditional Ramsey experiments just to figure out what 190 00:09:34.159 --> 00:09:35.679 the new frequency of each quibit is. 191 00:09:35.799 --> 00:09:36.919 And that takes a lot of time. 192 00:09:37.320 --> 00:09:40.519 It takes hours. You're looking at hours of downtime just 193 00:09:40.519 --> 00:09:43.840 to get maybe a few minutes of actual reliable compute time. 194 00:09:44.159 --> 00:09:45.399 It's wildly inefficient. 195 00:09:45.480 --> 00:09:49.639 So how exactly does AI change that equation? How does 196 00:09:49.679 --> 00:09:51.399 it fix the tuning problem? 197 00:09:51.600 --> 00:09:54.840 Engineering teams are now using machine learning models, very often 198 00:09:54.879 --> 00:09:58.679 graph neural networks to predict these drifts before they ruin 199 00:09:58.720 --> 00:10:03.679 the calculation. The AI continuously monitors the system's background diagnostics. 200 00:10:03.759 --> 00:10:06.039 So it's watching the temperature or the ambient noise. 201 00:10:06.200 --> 00:10:09.360 Yes, it sees a tiny temperature fluctuate, and it thinks ah. 202 00:10:09.519 --> 00:10:12.320 Based on the last ten thousand hours of operation data, 203 00:10:12.399 --> 00:10:14.440 I know that quibit forty three is about to drict 204 00:10:14.519 --> 00:10:16.120 by two megahotes. 205 00:10:15.720 --> 00:10:18.679 So it predicts the drift proactively exactly. 206 00:10:18.720 --> 00:10:21.000 You can identify the anomalous quivots, the ones that are 207 00:10:21.039 --> 00:10:24.440 starting to act up instantly. This reduces calibration time from 208 00:10:24.480 --> 00:10:27.080 hours down to just minutes. It keeps the machine operating 209 00:10:27.080 --> 00:10:28.600 in that sweet spot much much longer. 210 00:10:28.679 --> 00:10:32.000 Okay, so AI is acting as this active real time stabilizer. 211 00:10:32.200 --> 00:10:34.840 But let's go a layer deeper, because the real bottleneck, 212 00:10:34.879 --> 00:10:37.399 according to all the papers were reviewed, is error correction 213 00:10:37.799 --> 00:10:38.679 fault tolerance. 214 00:10:38.919 --> 00:10:43.080 That is the holy grail of quantum computing. We fundamentally 215 00:10:43.159 --> 00:10:45.679 need to be able to fix errors faster than they 216 00:10:45.759 --> 00:10:47.320 naturally occur, and. 217 00:10:47.240 --> 00:10:49.679 This is where the geometry of it all gets really interesting. 218 00:10:49.840 --> 00:10:53.399 The sources talk extensively about something called the surface code. 219 00:10:53.960 --> 00:10:57.320 The surface code is currently the leading architecture for achieving this. 220 00:10:58.120 --> 00:11:00.759 The core idea is to create what we call a 221 00:11:00.840 --> 00:11:02.440 logical quibit, which is. 222 00:11:02.399 --> 00:11:04.399 The actual data bit you care about. 223 00:11:04.320 --> 00:11:07.360 Right, But because physical equibits are so fragile, you don't 224 00:11:07.399 --> 00:11:10.759 store that logical bit on one single physical wire. You 225 00:11:10.799 --> 00:11:14.879 smear the information across a massive checkerboard pattern of many 226 00:11:14.919 --> 00:11:19.320 many physical quibbitts. Safety in numbers, pure redundancy. But here's 227 00:11:19.360 --> 00:11:22.159 the trick, and it's a deeply counterintuitive one. If you're 228 00:11:22.200 --> 00:11:25.120 used to classical computing, you cannot just look at the 229 00:11:25.159 --> 00:11:27.279 quibbitts to check if an error happened. 230 00:11:27.000 --> 00:11:29.440 Because if you measure the data equibits. 231 00:11:28.960 --> 00:11:32.799 You destroy the entanglement exactly, you collapse the superposition and 232 00:11:32.840 --> 00:11:34.720 you instantly lose the entire calculation. 233 00:11:34.919 --> 00:11:36.399 So if you can't look at them, how do you 234 00:11:36.519 --> 00:11:37.919 know if an error occurred? 235 00:11:38.240 --> 00:11:41.360 You use what are called ancill equipments. Think of them 236 00:11:41.360 --> 00:11:44.480 as help equibits. You interleave them with your data equibits, 237 00:11:44.919 --> 00:11:48.159 and you carefully entangle these helpers with the data equibits 238 00:11:48.399 --> 00:11:49.960 to perform a parity check. 239 00:11:50.120 --> 00:11:52.000 A parity check. Let's break that down. 240 00:11:52.200 --> 00:11:54.759 So you aren't asking the data equibit are you currently 241 00:11:54.879 --> 00:11:58.120 zero or one? You are asking a small group of them. 242 00:11:58.360 --> 00:12:02.200 Did any of you suddenly relative to your neighbors got it? 243 00:12:02.639 --> 00:12:05.840 So if they usually agree and suddenly the math shows 244 00:12:05.879 --> 00:12:08.799 they disagree, the Helper and Sill equibit lights up. 245 00:12:08.960 --> 00:12:12.039 Exactly, and that pattern of Ancilla measurements the ones that 246 00:12:12.159 --> 00:12:14.120 loy up is called the syndrome. 247 00:12:14.320 --> 00:12:17.279 The syndrome, it's like the shadow of the error rather 248 00:12:17.360 --> 00:12:18.440 than the error itself. 249 00:12:18.519 --> 00:12:20.000 It's a great way to think of it. It's the 250 00:12:20.080 --> 00:12:21.039 diagnostic symptom. 251 00:12:21.120 --> 00:12:24.039 Okay, so you have this continuous stream of syndrome data 252 00:12:24.080 --> 00:12:26.679 coming out of the fridge. Why do we need an 253 00:12:26.679 --> 00:12:30.279 advanced AI for that? Can't a classical computer just look 254 00:12:30.360 --> 00:12:32.559 up the error pattern in a pre computed table and 255 00:12:32.639 --> 00:12:34.080 apply effects. 256 00:12:33.799 --> 00:12:37.200 In an idealized, purely theoretical world. Yes, you use a 257 00:12:37.200 --> 00:12:40.360 standard algorithm. The most famous one is called minimum weight 258 00:12:40.399 --> 00:12:43.440 perfect matching. It basically looks at the syndrome grap and 259 00:12:43.480 --> 00:12:47.240 finds the simplest, shortest set of errors that explains the 260 00:12:47.279 --> 00:12:48.200 lights turning. 261 00:12:47.879 --> 00:12:51.159 On Oakham's razor. The simplest explanation is usually the right 262 00:12:51.159 --> 00:12:52.200 one precisely. 263 00:12:52.840 --> 00:12:56.639 But here is the massive rub. The physical hardware is 264 00:12:56.720 --> 00:13:00.720 not ideal, it's messy. The source is heavily emphasize the 265 00:13:00.759 --> 00:13:02.360 problem of correlated noise. 266 00:13:03.000 --> 00:13:06.240 What does correlated noise actually look like in a quantum chip. 267 00:13:06.519 --> 00:13:09.639 It means errors don't happen in isolation. If quibbit A 268 00:13:09.799 --> 00:13:13.240 suffers an error, it physically affects quibbitt B right next 269 00:13:13.240 --> 00:13:15.480 to it. Maybe a high energy cosmic ray hits the 270 00:13:15.519 --> 00:13:18.039 silicon substrate and wipes out a whole localized patch of 271 00:13:18.080 --> 00:13:21.039 quibbits at once. Or there's cross stock right crosstock between 272 00:13:21.080 --> 00:13:24.200 the microwave control lines. The errors cluster together in weird, 273 00:13:24.440 --> 00:13:26.000 highly non random ways. 274 00:13:25.720 --> 00:13:29.000 And the standard algorithms, those traditional graph matters you mentioned, 275 00:13:29.120 --> 00:13:31.440 they assume errors are completely independent. 276 00:13:31.879 --> 00:13:35.080 Usually yes, they assume a simple noise model, so when 277 00:13:35.120 --> 00:13:38.360 they encounter complex correlated noise in the real world, they 278 00:13:38.360 --> 00:13:42.240 get completely confused. They essentially hallucinate the wrong correction. 279 00:13:42.399 --> 00:13:45.320 And if you apply the wrong correction to a quantum state, you. 280 00:13:45.399 --> 00:13:48.320 Actively inject more errors into the system. You kill the 281 00:13:48.360 --> 00:13:49.639 logical quibot yourself. 282 00:13:49.720 --> 00:13:51.080 But a neural network a. 283 00:13:51.039 --> 00:13:55.639 Neural network absolutely loves correlations. That is quite literally what 284 00:13:55.720 --> 00:13:57.679 deep learning is built to find. You can train a 285 00:13:57.720 --> 00:14:02.000 neural network on the highly specific, messy noise fingerprint of 286 00:14:02.080 --> 00:14:06.000 one individual chip. It learns through observation that oh on 287 00:14:06.159 --> 00:14:10.519 this specific processor, when Quibot five flips, Quibot six almost 288 00:14:10.559 --> 00:14:12.679 always rotates by a few degrees. 289 00:14:12.519 --> 00:14:14.799 It learns the unique personality of the hardware. 290 00:14:14.919 --> 00:14:17.399 It really does, and the empirical data shows that these 291 00:14:17.399 --> 00:14:21.039 neural network decoders can interpret these complex syndromes much faster 292 00:14:21.399 --> 00:14:24.919 and far more accurately than the standard algorithms. It's a 293 00:14:24.919 --> 00:14:27.840 difference between a general practitioner doctor who just strictly follows 294 00:14:27.840 --> 00:14:31.000 a textbook checklist and a season specialist who has seen 295 00:14:31.000 --> 00:14:34.360 this highly specific, rare disease pattern a thousand times in 296 00:14:34.399 --> 00:14:34.879 the clinic. 297 00:14:35.159 --> 00:14:37.879 That is a massive leap forward. It's taking us from 298 00:14:37.960 --> 00:14:43.480 purely theoretical error correction on paper to practical hardware aware 299 00:14:43.720 --> 00:14:45.360 error correction in the real world. 300 00:14:45.679 --> 00:14:48.600 It's the bridge that makes fault tolerance actually achievable. 301 00:14:48.960 --> 00:14:50.879 Let's move up the stack a little bit. We have 302 00:14:51.000 --> 00:14:54.799 AI calibrating the machine. We have AI correcting the localized errors. 303 00:14:55.240 --> 00:14:58.240 Now we actually have to run a program. We have 304 00:14:58.279 --> 00:15:01.200 to run code. This brings us to the section the 305 00:15:01.240 --> 00:15:03.919 sources call optimization and compilation. 306 00:15:04.320 --> 00:15:06.320 Yes, the tetris phase of. 307 00:15:06.360 --> 00:15:08.720 Quantum computing, but incredibly high stakes. 308 00:15:08.759 --> 00:15:12.600 Tetris extremely high stakes. You see, when a programmer writes 309 00:15:12.639 --> 00:15:15.240 a quantum algorithm, they write it in a very abstract way. 310 00:15:15.559 --> 00:15:17.879 They might write a line of code that says, apply 311 00:15:17.960 --> 00:15:20.759 an entangling c and Ot gate between quibut one and 312 00:15:20.840 --> 00:15:21.440 quibit ten. 313 00:15:22.279 --> 00:15:25.759 But on the actual physical silicon chip down in the fridge, 314 00:15:25.919 --> 00:15:27.600 quibuit one might be nowhere near. 315 00:15:27.519 --> 00:15:30.720 Quibut ten exactly. Physical chips have a specific topology. They 316 00:15:30.720 --> 00:15:32.879 have a layout. Maybe it's a square grid, or maybe 317 00:15:32.919 --> 00:15:35.200 it's a heavy X lattice, which is what IBM tens 318 00:15:35.200 --> 00:15:38.159 to use in these layouts. Physical qubits can only talk 319 00:15:38.200 --> 00:15:40.600 directly to their immediate physical neighbors. 320 00:15:40.960 --> 00:15:43.320 So if I want to entangle quibut one and quibit 321 00:15:43.360 --> 00:15:45.639 ten and they aren't physically touching, you. 322 00:15:45.600 --> 00:15:48.639 Have to manually move the quantum information across the chip. 323 00:15:48.679 --> 00:15:51.080 You have to use what are called swapgates. You swap 324 00:15:51.120 --> 00:15:54.120 the quantum state of quibit one into quibit two, and 325 00:15:54.159 --> 00:15:56.480 then from two to three and so on, cascading it 326 00:15:56.559 --> 00:15:59.080 until it is physically sitting right next to quibut ten. 327 00:15:59.200 --> 00:16:02.799 But every single swapgate takes physical time to execute. 328 00:16:02.879 --> 00:16:06.519 Time is the enemy. Every nanosecond adds exposure to noise. 329 00:16:06.879 --> 00:16:10.120 Every extra gate you add lowers the overall fidelity of 330 00:16:10.159 --> 00:16:12.480 your entire circuit. So you end up with a massive 331 00:16:12.600 --> 00:16:16.720 complex optimization problem. How do I map this highly abstract 332 00:16:16.799 --> 00:16:21.159 software circuit onto this rigid physical graph using the absolute 333 00:16:21.200 --> 00:16:22.720 fewest possible swapgates. 334 00:16:22.759 --> 00:16:25.720 That sounds suspiciously like the traveling salesman problem. 335 00:16:25.799 --> 00:16:28.440 It essentially is. It is an NP hard routing problem 336 00:16:28.480 --> 00:16:32.480 for large complex quantum circuits. Finding the mathematically perfect mapping 337 00:16:32.759 --> 00:16:36.480 is computationally impossible. For traditional classical solvers, they just choke 338 00:16:36.519 --> 00:16:38.120 on the complexity they take too long. 339 00:16:38.360 --> 00:16:41.000 So enter the AI architect reinforcement learning. 340 00:16:41.039 --> 00:16:44.600 Specifically, researchers are deploying agents very similar to alpha zero. 341 00:16:44.919 --> 00:16:47.600 You mean the AI that famously mastered the games of 342 00:16:47.679 --> 00:16:48.799 Go and chess. 343 00:16:48.759 --> 00:16:50.200 The very same architecture. 344 00:16:50.240 --> 00:16:52.799 How does that apply to routing quantum circuits. 345 00:16:53.000 --> 00:16:55.720 Well, think of the quibbit topology. The layout of the 346 00:16:55.799 --> 00:16:59.039 chip as the board. The quantum states are the pieces, 347 00:16:59.559 --> 00:17:03.320 the plays the game of routing the circuit millions and 348 00:17:03.399 --> 00:17:06.559 millions of times. In simulation, it gets a reward signal 349 00:17:06.559 --> 00:17:09.480 for minimizing the overall circuit depth, and it gets heavy 350 00:17:09.519 --> 00:17:12.160 penalties for adding unnecessary swap gates. 351 00:17:12.680 --> 00:17:16.079 And through playing this game, it discovers routing strategies that 352 00:17:16.200 --> 00:17:17.519 human engineers miss. 353 00:17:17.720 --> 00:17:22.720 It absolutely does. It consistently finds non intuitive, brilliant paths. 354 00:17:23.119 --> 00:17:26.400 It recognizes deep patterns in the circuit structure. It might 355 00:17:26.440 --> 00:17:28.880 look at the code and realize, oh, this specific block 356 00:17:28.920 --> 00:17:32.519 of operations mathematically resembles a quantum foury eight transform, so 357 00:17:32.559 --> 00:17:34.480 I can fold it this particular way on the heavy 358 00:17:34.480 --> 00:17:38.480 hex lattice to save twenty gates. It actively squeezes every 359 00:17:38.559 --> 00:17:41.559 drop of efficiency out of the imperfect hardware in ways 360 00:17:41.559 --> 00:17:43.440 static compilers just cannot match. 361 00:17:43.640 --> 00:17:45.559 So, just to recap this whole first half of the 362 00:17:45.599 --> 00:17:49.240 discussion for you listening, we have AI tuning the physical instrument. 363 00:17:49.359 --> 00:17:52.559 We have AI actively diagnosing and fixing the errors, and 364 00:17:52.640 --> 00:17:55.319 we have AI rewriting the sheet music so it perfectly 365 00:17:55.359 --> 00:17:56.880 fits the quarks of the instrument. 366 00:17:57.160 --> 00:18:00.799 That is a perfect summary. AI is the critical infrastructure 367 00:18:00.839 --> 00:18:05.240 layer that is making quantum computing plausible. Without it, the 368 00:18:05.279 --> 00:18:09.799 sheer crushing complexity of managing the system entirely overwhelms us. 369 00:18:09.920 --> 00:18:11.920 Okay, so now I want to flip the script entirely. 370 00:18:12.000 --> 00:18:15.839 We've solidly established that quantum computing absolutely needs AI to function, 371 00:18:16.480 --> 00:18:19.799 But does AI actually need quantum This is. 372 00:18:19.720 --> 00:18:22.680 Where we have to tread very carefully. We are entering 373 00:18:22.720 --> 00:18:26.680 the realm of quantum machine learning or QML, and I 374 00:18:26.759 --> 00:18:29.319 really want to be clear upfront, there is an enormous 375 00:18:29.319 --> 00:18:31.519 amount of hype in this specific area that the sources 376 00:18:31.599 --> 00:18:33.839 explicitly caution us against, right. 377 00:18:33.640 --> 00:18:35.839 Because you see these wild headlines all the time, things 378 00:18:35.880 --> 00:18:38.720 like new quantum computer will train GPT five in three. 379 00:18:38.559 --> 00:18:41.759 Seconds, and that is at least currently pure science fiction. 380 00:18:41.960 --> 00:18:44.279 The expert analysis in all of our sources is actually 381 00:18:44.359 --> 00:18:46.839 quite skeptical of those broad sweeping claims. 382 00:18:47.079 --> 00:18:50.559 Let's unpack the theoretical advantage, though. Why do people even 383 00:18:50.599 --> 00:18:53.319 think a quantum computer would help AI in the first place. 384 00:18:53.759 --> 00:18:57.440 It all fundamentally comes down to linear algebra. Modern AI 385 00:18:57.720 --> 00:19:02.000 at its absolute core is just mass matrix multiplication. It's 386 00:19:02.119 --> 00:19:04.480 navigating high dimensional vector spaces. 387 00:19:04.119 --> 00:19:08.119 And quantum mechanics is also at its core linear algebra exactly. 388 00:19:08.519 --> 00:19:13.240 A quantum computer naturally manipulates vectors within a vast Hilbert space. 389 00:19:13.720 --> 00:19:17.400 There is a famous theoretical algorithm, the HHL algorithm, that 390 00:19:17.559 --> 00:19:20.960 proves a quantum computer can solve massive systems of linear 391 00:19:21.000 --> 00:19:24.839 equations exponentially faster than any classical computer ever could. 392 00:19:24.880 --> 00:19:28.319 So the industry logic basically goes AI is linear algebra, 393 00:19:28.519 --> 00:19:32.039 Quantum is extremely fast at linear algebra. Therefore quantum will 394 00:19:32.079 --> 00:19:33.079 supercharge AI. 395 00:19:33.559 --> 00:19:37.240 That's the compelling syllogism. Yes, but there are major caveats 396 00:19:37.279 --> 00:19:39.960 hiding under the surface. The biggest one, the sources point out, 397 00:19:40.000 --> 00:19:41.319 is what we call the input problem. 398 00:19:41.359 --> 00:19:43.920 You mean actually loading the data into the machine. 399 00:19:44.119 --> 00:19:48.279 Yes, we do not currently have quantum RAM. If you 400 00:19:48.400 --> 00:19:52.000 have a massive classical data set, like say the text 401 00:19:52.000 --> 00:19:54.039