WEBVTT 1 00:00:00.130 --> 00:00:01.050 Welcome to the debate. 2 00:00:01.470 --> 00:00:06.440 So historically, if a if a systems engineer wanted to 3 00:00:06.500 --> 00:00:09.759 know if a newly designed airplane wing would, you know, 4 00:00:09.859 --> 00:00:14.689 snap under extreme aerodynamic pressure, they had to physically build 5 00:00:14.750 --> 00:00:15.169 the thing. 6 00:00:15.449 --> 00:00:19.489 Right. Yeah. Like an actual scaled thing. Physical model. Exactly. 7 00:00:19.710 --> 00:00:22.079 You build it, you stick it in a massive wind tunnel, 8 00:00:22.219 --> 00:00:24.879 you blast it with air, and you just literally watch 9 00:00:24.920 --> 00:00:26.460 the smoke flow over the curves. 10 00:00:26.699 --> 00:00:28.160 To see where the stress points are. 11 00:00:28.420 --> 00:00:31.179 Yeah, to see exactly where it fails. But today, we're 12 00:00:31.219 --> 00:00:33.920 standing at the edge of this radically different paradigm, right? 13 00:00:34.240 --> 00:00:37.810 Where AI is becoming the wind tunnel. But the industry 14 00:00:37.850 --> 00:00:41.609 is currently just totally fractured over what the fundamental architecture 15 00:00:42.149 --> 00:00:43.689 of that AI should actually be. 16 00:00:44.049 --> 00:00:47.299 And we're seeing this massive divergence in the source material today. 17 00:00:47.380 --> 00:00:48.079 We really are. 18 00:00:48.240 --> 00:00:49.240 It's a huge split. 19 00:00:49.460 --> 00:00:49.679 Right. 20 00:00:50.000 --> 00:00:53.200 On one side, we have this company, Accelerated Understanding. And 21 00:00:53.240 --> 00:00:56.640 they just unveiled a purely physics-based model. And it's capable 22 00:00:56.659 --> 00:01:00.640 of processing this staggering 5 trillion pieces of data at once. 23 00:01:00.780 --> 00:01:02.119 Which is just a crazy number. 24 00:01:02.700 --> 00:01:03.310 It's wild. 25 00:01:03.609 --> 00:01:06.969 But then on the other side, we have Moonshot's massive 2.8 26 00:01:06.969 --> 00:01:11.599 trillion parameter language model, Kemi3. And they're right now negotiating 27 00:01:11.640 --> 00:01:14.920 these unprecedented revenue-sharing deals with major U.S. 28 00:01:15.000 --> 00:01:18.540 Cloud giants. So my position on this is pretty clear. 29 00:01:19.079 --> 00:01:21.859 I argue that AI built on neural operators, which is 30 00:01:21.920 --> 00:01:26.200 a nature-centric, physics-first view of reality, is the absolute only 31 00:01:26.280 --> 00:01:29.900 logical path forward for systems engineering and industrial experts. 32 00:01:30.359 --> 00:01:33.219 Well, and I look at this entirely differently. I mean, 33 00:01:33.379 --> 00:01:38.299 I argue that massive language-centric architectures, you know, exactly like 34 00:01:38.379 --> 00:01:42.150 Moonshot's K3... they remain the superior and just the most 35 00:01:42.209 --> 00:01:42.950 practical baseline. 36 00:01:43.269 --> 00:01:43.909 Okay, but how? 37 00:01:44.209 --> 00:01:49.680 Because language models offer this generalized, multi-step reasoning capability. And 38 00:01:49.719 --> 00:01:54.340 that is simply indispensable for modern enterprise cloud infrastructure, regardless 39 00:01:54.359 --> 00:01:56.359 of what's happening in the physical lab. 40 00:01:56.700 --> 00:01:59.609 I mean, let's start right at the architectural foundation then. 41 00:01:59.969 --> 00:02:02.510 Because to me, trying to use a language model to 42 00:02:02.549 --> 00:02:06.909 solve a physical engineering problem is a foundational error. 43 00:02:07.150 --> 00:02:09.860 A foundational error? That's a strong phrase. 44 00:02:10.280 --> 00:02:11.090 I really think it is. 45 00:02:11.219 --> 00:02:14.340 It's fundamentally applying the wrong tool to the deepest problems 46 00:02:14.379 --> 00:02:17.069 we face. I mean, look at the breakthrough from Caltech 47 00:02:17.110 --> 00:02:21.169 professor Anima Anankumar and Benedict Jenick and Accelerated Understanding. 48 00:02:21.310 --> 00:02:22.069 Right, their new model. 49 00:02:22.449 --> 00:02:26.009 Yeah. What they've done is dispense entirely with the Google-invented 50 00:02:26.069 --> 00:02:27.229 transformer architecture. 51 00:02:27.370 --> 00:02:28.180 They just tossed it. 52 00:02:28.370 --> 00:02:31.169 Well, they're not trying to parse text, so... Exactly. 53 00:02:31.569 --> 00:02:35.889 They aren't parsing text. They are utilizing neural operators to 54 00:02:36.009 --> 00:02:39.990 model reality natively in space and time. It's a completely 55 00:02:40.050 --> 00:02:40.770 different approach. 56 00:02:41.030 --> 00:02:42.909 But I think we really have to clarify what you 57 00:02:42.949 --> 00:02:46.129 mean by applying the wrong tool. Because language isn't just 58 00:02:46.169 --> 00:02:49.289 about human conversation or, you know, writing a clever email 59 00:02:49.349 --> 00:02:49.870 or a poem. 60 00:02:50.389 --> 00:02:50.610 Sure. 61 00:02:51.210 --> 00:02:54.129 It is the ultimate abstraction of logic. When you look 62 00:02:54.150 --> 00:02:57.740 at Moonshot's Kimi K3, you're looking at an open-weight model 63 00:02:57.780 --> 00:02:59.659 with 2.8 trillion parameters. 64 00:03:00.139 --> 00:03:01.759 Which is massive, I'll give you that. 65 00:03:02.189 --> 00:03:04.990 Right. And based on the latest evaluations, it performs on 66 00:03:05.050 --> 00:03:10.319 par with the absolute elite frontier models. We're talking OpenAI's GPT-5.5 67 00:03:10.699 --> 00:03:13.020 and Anthropic's Claude Opus 4.8. 68 00:03:13.020 --> 00:03:15.080 Wait, but reasoning about what exactly? 69 00:03:15.520 --> 00:03:16.500 About complex logic. 70 00:03:16.840 --> 00:03:19.500 If you're reasoning about a text document, sure. But if 71 00:03:19.539 --> 00:03:20.319 you're reasoning about. 72 00:03:20.379 --> 00:03:24.810 Say, thermodynamics on a microchip, a transformer is ultimately just 73 00:03:24.889 --> 00:03:27.030 guessing the next most likely word based on. 74 00:03:27.030 --> 00:03:27.789 Its training data. 75 00:03:28.090 --> 00:03:29.229 I think that's a bit reductive. 76 00:03:29.810 --> 00:03:32.000 But let's unpack the mechanism for a second. Think of 77 00:03:32.020 --> 00:03:35.780 a transformer like a beads on a string model. 78 00:03:36.259 --> 00:03:36.500 Okay. 79 00:03:37.139 --> 00:03:40.310 It takes in a sequence of discrete units, right? Words 80 00:03:40.370 --> 00:03:43.250 or code snippets. And it predicts the next bead in 81 00:03:43.270 --> 00:03:45.270 the sequence based on statistical probability. 82 00:03:45.509 --> 00:03:47.550 Yeah, next token prediction. Right. 83 00:03:48.129 --> 00:03:51.419 It is inherently a human-centric view of intelligence because it's 84 00:03:51.479 --> 00:03:55.979 built on our language. Now, contrast that with a neural operator. 85 00:03:56.240 --> 00:03:57.719 Which works differently. Totally differently. 86 00:03:57.759 --> 00:03:58.400 Totally differently. 87 00:03:58.960 --> 00:04:01.780 A neural operator doesn't calculate point A, then point B, 88 00:04:01.840 --> 00:04:04.919 then point C. It maps the entire continuous flow of 89 00:04:04.960 --> 00:04:07.520 a physical system as a single mathematical object. 90 00:04:07.740 --> 00:04:09.120 As one continuous thing. 91 00:04:09.219 --> 00:04:14.969 Yes. They map what we call infinite dimensional function spaces. 92 00:04:15.590 --> 00:04:19.269 Okay. See, that sounds incredibly elegant in a mathematics paper. 93 00:04:19.949 --> 00:04:22.209 It does. But let's break down what that actually means 94 00:04:22.269 --> 00:04:23.410 for an enterprise on the ground. 95 00:04:24.129 --> 00:04:28.730 Well, it means it understands the entire field of reality simultaneously. 96 00:04:29.300 --> 00:04:30.199 not step by step. 97 00:04:30.399 --> 00:04:31.939 But how does an engineer use that? 98 00:04:32.920 --> 00:04:37.019 Look, expecting a transformer model to optimize next generation semiconductor 99 00:04:37.060 --> 00:04:40.740 materials is, it's like asking a master poet to build 100 00:04:40.779 --> 00:04:43.439 a suspension bridge. Right? 101 00:04:43.639 --> 00:04:44.310 A master poet? 102 00:04:44.500 --> 00:04:47.670 Yeah, based strictly on descriptions of bridges they've read in 103 00:04:47.709 --> 00:04:48.269 a library. 104 00:04:48.550 --> 00:04:49.829 Oh, I see where you're going. Right? 105 00:04:50.529 --> 00:04:52.389 Like, they might know the word tension. 106 00:04:53.089 --> 00:04:54.910 They might be able to write a beautiful, you know, 107 00:04:55.009 --> 00:04:59.639 logically coherent essay on steel. but they don't needively understand 108 00:04:59.680 --> 00:05:02.480 the underlying mathematical equations of load-bearing structures. 109 00:05:02.860 --> 00:05:03.240 Okay, but. 110 00:05:03.680 --> 00:05:08.379 Neural operators do. They don't learn about physics by reading textbooks. 111 00:05:08.860 --> 00:05:13.079 They learn the actual equations of the physical world directly. 112 00:05:13.610 --> 00:05:17.649 You just cannot text-predict your way to optimal thermodynamics. 113 00:05:18.029 --> 00:05:20.089 See, I have to challenge you on that bridge analogy, 114 00:05:20.449 --> 00:05:23.050 because I actually think it perfectly highlights the blind spot 115 00:05:23.069 --> 00:05:24.370 in the physics-centric approach. 116 00:05:24.649 --> 00:05:25.069 How so? 117 00:05:25.649 --> 00:05:28.930 Let's say your master architect, right, your neural operator, figures 118 00:05:28.970 --> 00:05:32.089 out the perfect mathematical state for that suspension bridge or 119 00:05:32.129 --> 00:05:33.129 that semiconductor. 120 00:05:33.449 --> 00:05:35.569 OK, it's mapped the perfect physical state. 121 00:05:35.949 --> 00:05:39.050 Right. But how does the enterprise actually use that? A 122 00:05:39.100 --> 00:05:41.579 bridge isn't built by a single physicist doing math in 123 00:05:41.600 --> 00:05:42.199 a vacuum. 124 00:05:42.360 --> 00:05:43.279 Well, of course not. 125 00:05:43.480 --> 00:05:47.180 It requires a massive team. You're coordinating logistics. You're writing 126 00:05:47.240 --> 00:05:52.829 software integration tests, configuring manufacturing equipment. Parsing complex diagnostic logs. 127 00:05:53.250 --> 00:05:54.860 Sure, there's an implementation layer. 128 00:05:55.160 --> 00:05:59.199 But language is what orchestrates that entire ecosystem. If your 129 00:05:59.279 --> 00:06:03.519 physics model just outputs raw tensor data, infinite dimensional or not, 130 00:06:03.860 --> 00:06:06.879 it's essentially an oracle screaming raw numbers into a void. 131 00:06:07.430 --> 00:06:09.310 I wouldn't call it screaming into a void. 132 00:06:09.649 --> 00:06:12.670 It kind of is. It's completely useless without the language 133 00:06:12.730 --> 00:06:17.129 layer to translate that vast physical complexity into actionable, auditable 134 00:06:17.209 --> 00:06:18.850 steps for the systems engineers. 135 00:06:19.009 --> 00:06:21.050 I mean, I don't think it's screaming into a void 136 00:06:21.069 --> 00:06:21.490 at all. 137 00:06:21.930 --> 00:06:24.769 The output interface is directly with the engineering software. 138 00:06:24.949 --> 00:06:26.029 But who builds that interface? 139 00:06:26.490 --> 00:06:30.050 But more importantly, the bottleneck in industrial engineering isn't writing 140 00:06:30.069 --> 00:06:32.410 the integration test or coordinating the logistics. 141 00:06:33.069 --> 00:06:34.100 That's the easy part. 142 00:06:34.569 --> 00:06:36.040 I don't know if the guys on the factory floor 143 00:06:36.079 --> 00:06:36.740 would call it easy. 144 00:06:37.209 --> 00:06:40.850 Fair but the real bottleneck is the trial and error 145 00:06:40.970 --> 00:06:45.769 in the laboratory it's the months sometimes years spent testing 146 00:06:45.850 --> 00:06:48.449 physical materials that ultimately fail under. 147 00:06:48.269 --> 00:06:50.629 Stress right the prototyping phase. 148 00:06:50.649 --> 00:06:56.399 Exactly accelerated understandings AI fundamentally eliminates that trial and error 149 00:06:56.860 --> 00:07:00.199 because of its context scale and we really need to 150 00:07:00.240 --> 00:07:03.800 talk about this scale because it's mind-blowing. 151 00:07:03.379 --> 00:07:05.060 The five trillion data points. 152 00:07:05.000 --> 00:07:10.360 Yes Their system processes 5 trillion pieces of data in 153 00:07:10.399 --> 00:07:11.160 a single prompt. 154 00:07:11.600 --> 00:07:13.879 And we need to clarify what pieces of data means 155 00:07:14.060 --> 00:07:17.480 in this context, because 5 trillion is just a staggering 156 00:07:17.519 --> 00:07:18.540 number to wrap your head around. 157 00:07:18.860 --> 00:07:19.519 It really is. 158 00:07:20.040 --> 00:07:23.379 To put it into perspective, 5 trillion pieces of data 159 00:07:23.550 --> 00:07:27.810 is roughly 5 million times what Anthropic or Google's flagship 160 00:07:27.850 --> 00:07:29.410 models can typically consume in. 161 00:07:29.389 --> 00:07:30.129 A single query. 162 00:07:30.129 --> 00:07:31.870 5 million times? Yes. 163 00:07:33.240 --> 00:07:36.839 It is the equivalent of reading Tolstoy's War and Peace 164 00:07:37.180 --> 00:07:39.019 five million times in one sitting. 165 00:07:39.500 --> 00:07:41.240 That is, yeah, that's heavy. 166 00:07:41.680 --> 00:07:44.439 But it's not reading text. That's the key. 167 00:07:45.060 --> 00:07:49.459 It is ingesting the complete state of physical systems. We're 168 00:07:49.480 --> 00:07:55.120 talking temperature, velocity, pressure, material density, all across time. 169 00:07:55.560 --> 00:07:57.379 So it's looking at the whole system at once. 170 00:07:57.819 --> 00:08:01.290 Exactly. You ingest all the physical variables of a system 171 00:08:01.449 --> 00:08:06.139 at once. NVIDIA CEO Jensen Huang, he recognized this potential 172 00:08:06.199 --> 00:08:06.740 years ago. 173 00:08:07.079 --> 00:08:08.480 Oh, right, the weather prediction thing. 174 00:08:08.860 --> 00:08:09.300 Yeah. 175 00:08:09.819 --> 00:08:14.300 Back in 2021, when Anand Kankumar's team showed how AI 176 00:08:14.360 --> 00:08:17.689 could speed up weather prediction with the accuracy of those 177 00:08:17.810 --> 00:08:22.230 incredibly complex computational models, Huang didn't just praise it as 178 00:08:22.269 --> 00:08:22.870 a neat trick. 179 00:08:23.189 --> 00:08:23.829 What did he say? 180 00:08:24.529 --> 00:08:28.439 He specifically stated he wanted AI to, quote, Eat all 181 00:08:28.579 --> 00:08:30.240 physics theorists' lunches. 182 00:08:30.420 --> 00:08:31.480 Right, I remember that quote. 183 00:08:31.519 --> 00:08:35.379 Because once you can map that many continuous variables, you 184 00:08:35.419 --> 00:08:39.330 don't need brittle, bespoke math models anymore. 185 00:08:39.710 --> 00:08:42.009 You just simulate reality instantly. 186 00:08:42.330 --> 00:08:44.919 It is a phenomenal vision, I'll give you that. But 187 00:08:44.960 --> 00:08:48.519 this is exactly where the theoretical elegance... crashes into the 188 00:08:48.580 --> 00:08:50.779 reality of enterprise infrastructure. 189 00:08:50.879 --> 00:08:51.220 How? 190 00:08:51.419 --> 00:08:52.639 It solves the court problem. 191 00:08:52.879 --> 00:08:56.279 Let's look at the practical deployment of these massive context windows. 192 00:08:56.580 --> 00:09:00.379 You're talking about ingesting 5 trillion data points of continuous 193 00:09:00.460 --> 00:09:04.590 physical states. What is the infrastructure cost of evaluating that continuously? 194 00:09:04.679 --> 00:09:05.649 It's high, but. 195 00:09:05.730 --> 00:09:10.049 The entire cloud ecosystem, which is the backbone of modern enterprise, 196 00:09:10.230 --> 00:09:12.250 is already built around the language paradigm. 197 00:09:12.470 --> 00:09:16.009 But the infrastructure will adapt to the value of the output. 198 00:09:16.440 --> 00:09:19.320 I mean, if a chip manufacturer can skip six months 199 00:09:19.340 --> 00:09:22.120 of R & D, they will absolutely build the servers. 200 00:09:22.379 --> 00:09:25.559 It's not just about building servers, though. It's about the 201 00:09:25.659 --> 00:09:29.100 fundamental business model of compute. We need to talk about 202 00:09:29.139 --> 00:09:34.070 K3' s 2.8 trillion parameter scale and how it's actually hosted. 203 00:09:34.470 --> 00:09:35.500 Right, the cloud giants. 204 00:09:35.570 --> 00:09:41.830 Yeah, K3 requires major cloud infrastructure, Azure, AWS, Google Cloud, 205 00:09:42.350 --> 00:09:45.779 simply to handle the immense computing costs. And how do 206 00:09:45.799 --> 00:09:50.580 those cloud giants build the enterprise? By tokens. Exactly. They 207 00:09:50.759 --> 00:09:52.720 use token-based usage billing. 208 00:09:52.740 --> 00:09:55.730 Which is a completely arbitrary metric born out of text generation. 209 00:09:55.870 --> 00:09:59.250 It's not arbitrary at all. It's the foundation of cloud 210 00:09:59.309 --> 00:10:03.039 economics right now. now. A token is essentially a chunk 211 00:10:03.080 --> 00:10:05.700 of a word, right? A discrete unit of data. 212 00:10:05.960 --> 00:10:06.159 Yeah. 213 00:10:06.519 --> 00:10:11.000 It is a measurable, highly auditable system for enterprises. When 214 00:10:11.039 --> 00:10:14.679 an engineer uses K3 to run a multi-step diagnostic on 215 00:10:14.720 --> 00:10:18.110 a factory floor, the cloud provider knows exactly how many 216 00:10:18.179 --> 00:10:33.549 tokens were processed. Structure doesn't speak continuous topology. 217 00:10:34.009 --> 00:10:34.529 Well, not yet. 218 00:10:35.169 --> 00:10:38.110 It speaks in discrete billable units. If I run a 219 00:10:38.090 --> 00:10:42.029 5 trillion point continuous physics simulation, how do you audit that? 220 00:10:42.529 --> 00:10:45.870 You audit the output. But the data access and auditing 221 00:10:45.909 --> 00:10:50.049 for the continuous physical variables themselves remain completely unresolved. It 222 00:10:50.110 --> 00:10:53.240 is incredibly costly and opaque compared to the token economy. 223 00:10:53.950 --> 00:10:56.710 See, I feel like you're prioritizing the cash register over 224 00:10:56.750 --> 00:10:58.649 the actual technological breakthrough here. 225 00:10:59.230 --> 00:11:02.220 I'm prioritizing practicality for systems engineers. 226 00:11:02.639 --> 00:11:05.779 But yes, token billing is convenient for Microsoft right now. 227 00:11:06.120 --> 00:11:09.279 But if a system can accurately predict extreme weather patterns 228 00:11:09.639 --> 00:11:12.750 or instantly sift through complex geological data for an offshore 229 00:11:12.799 --> 00:11:15.470 energy company, the enterprise will figure out how to track 230 00:11:15.509 --> 00:11:16.289 the compute costs. 231 00:11:16.590 --> 00:11:17.870 Will they? Of course! 232 00:11:18.610 --> 00:11:23.330 The sheer value of true physics-centric intelligence is proven by 233 00:11:23.370 --> 00:11:26.350 the lengths capital will go to try and capture it. 234 00:11:26.470 --> 00:11:30.110 Capital is certainly flowing everywhere in AI, but where is 235 00:11:30.169 --> 00:11:31.590 it flowing most aggressively today? 236 00:11:31.980 --> 00:11:35.340 Well, look at what Anand Kumar and Yannick just walked 237 00:11:35.379 --> 00:11:39.000 away from. This is crucial context from the source material. 238 00:11:39.019 --> 00:11:39.820 The Amazon thing. 239 00:11:39.919 --> 00:11:44.179 Yeah, they were heavily recruited to lead Jeff Bezos' Project Prometheus. 240 00:11:44.700 --> 00:11:47.419 And the offer they received wasn't just, you know, a 241 00:11:47.460 --> 00:11:49.019 standard tech recruitment package. 242 00:11:49.720 --> 00:11:50.679 It was staggering. 243 00:11:50.840 --> 00:11:51.840 It was a massive number. 244 00:11:52.120 --> 00:11:56.279 We are talking about a $ 2 billion capital outline for 245 00:11:56.379 --> 00:11:58.000 committed financing rounds. 246 00:11:58.139 --> 00:11:59.179 Just to start? Right. 247 00:11:59.710 --> 00:12:03.250 They were offered a 35% equity stake in the company, 248 00:12:03.629 --> 00:12:06.370 and their salaries were set to double to $ 2 million 249 00:12:06.649 --> 00:12:08.009 after just three months. 250 00:12:08.190 --> 00:12:08.970 And they turned it down. 251 00:12:09.309 --> 00:12:11.299 They walked away from all of that. 252 00:12:11.929 --> 00:12:15.750 Just walked away to build their own independent model for 253 00:12:15.850 --> 00:12:19.590 enterprise AI spanning robotics, energy, and geological data. 254 00:12:19.990 --> 00:12:21.190 Which is a huge risk. 255 00:12:21.649 --> 00:12:24.049 But they did it because they knew that a single 256 00:12:24.240 --> 00:12:27.620 generalized physics AI is the actual holy grail of the 257 00:12:27.659 --> 00:12:30.340 industrial world. They didn't want to be locked into someone 258 00:12:30.360 --> 00:12:31.299 else's narrow vision. 259 00:12:31.659 --> 00:12:34.460 Right, and let's not forget the aftermath here. Once they 260 00:12:34.519 --> 00:12:37.019 rejected the offer... Prometheus had to go out and raise 261 00:12:37.039 --> 00:12:40.899 a massive $ 12 billion Series B just to chase that 262 00:12:40.940 --> 00:12:44.720 exact same goal of automating the manufacturing of complex physical systems. 263 00:12:45.100 --> 00:12:46.220 That's a lot of capital. 264 00:12:46.500 --> 00:12:49.299 It proves the underlying thesis. The market knows that whoever 265 00:12:49.379 --> 00:12:51.759 maps the physical world wins the industrial future. 266 00:12:52.230 --> 00:12:54.990 Look, I don't dispute that the pursuit of a generalized 267 00:12:55.049 --> 00:12:59.830 physics model commands mass evaluations. A $ 12 billion Series B 268 00:12:59.909 --> 00:13:02.320 is undeniable proof of investor appetite. 269 00:13:02.600 --> 00:13:03.840 Absolutely. Exactly. 270 00:13:04.299 --> 00:13:07.179 But if we are debating the immediate implications for systems 271 00:13:07.220 --> 00:13:12.309 engineers and industrial enterprise experts, Moonshot's traction proves the absolute 272 00:13:12.470 --> 00:13:14.169 dominance of language models today. 273 00:13:14.860 --> 00:13:18.679 Even with all the massive geopolitical baggage Moonshot carries? 274 00:13:19.120 --> 00:13:22.690 Especially because of it. Honestly, this is what makes the 275 00:13:22.750 --> 00:13:25.269 enterprise adoption of K3 so incredibly telling. 276 00:13:25.830 --> 00:13:26.429 How do you mean? 277 00:13:27.090 --> 00:13:31.120 Let's look at the immense friction Moonshot is currently facing. U.S. 278 00:13:31.169 --> 00:13:34.480 Treasury Secretary Scott Besson has publicly threatened to add them 279 00:13:34.519 --> 00:13:35.659 to a trade blacklist. 280 00:13:36.059 --> 00:13:36.940 Right. U.S. 281 00:13:36.980 --> 00:13:40.139 Officials have actively accused the company of stealing from Anthropic's 282 00:13:40.220 --> 00:13:43.960 Fable model. There are serious accusations about them illegally acquiring 283 00:13:44.019 --> 00:13:44.639 NVIDIA chips. 284 00:13:45.259 --> 00:13:46.940 Those are heavy accusations. 285 00:13:47.139 --> 00:13:51.049 Exactly. These are not minor PR issues. These are massive, 286 00:13:51.250 --> 00:13:55.009 existential red flags for any enterprise supply chain, which should 287 00:13:55.049 --> 00:14:00.750 make any systems engineer incredibly wary of integrating K3 into 288 00:14:00.769 --> 00:14:01.629 their core pipeline. 289 00:14:02.029 --> 00:14:06.429 You would think so. And yet, despite every single one 290 00:14:06.470 --> 00:14:10.460 of those red flags, U.S. cloud giants are still clamoring 291 00:14:10.769 --> 00:14:11.799 to host K3. 292 00:14:12.139 --> 00:14:13.019 Yeah, it's wild. 293 00:14:13.340 --> 00:14:17.840 We are seeing Microsoft Azure, AWS, and Google Cloud willing 294 00:14:17.919 --> 00:14:22.200 to negotiate up to 30% revenue sharing agreements just to 295 00:14:22.259 --> 00:14:24.340 get K3 onto their platforms. 296 00:14:24.720 --> 00:14:27.580 But I think that just speaks to a desperation for 297 00:14:27.700 --> 00:14:31.830 massive parameter counts. not necessarily a superiority of the language 298 00:14:31.889 --> 00:14:32.809 architecture itself. 299 00:14:33.370 --> 00:14:34.789 You think they just want the size? 300 00:14:35.210 --> 00:14:38.669 The cloud providers just want the biggest model to sell compute. 301 00:14:38.929 --> 00:14:41.379 They want to push as many tokens as possible. Now, 302 00:14:41.399 --> 00:14:44.740 I really think it speaks to undeniable utility. When major 303 00:14:44.779 --> 00:14:47.860 industry players, and let's remember, Moonshot is also backed by 304 00:14:47.960 --> 00:14:51.600 Alibaba and raised $ 2 billion in May. When the industry 305 00:14:51.740 --> 00:14:54.799 fights this hard for a model, it proves that language-centric 306 00:14:54.879 --> 00:14:57.240 reasoning is the required foundational layer right now. 307 00:14:57.840 --> 00:14:59.139 Even with the blacklist threat? 308 00:14:59.600 --> 00:15:03.700 Microsoft, Amazon, and Google are willing to navigate an absolute 309 00:15:03.779 --> 00:15:07.659 geopolitical minefield and give away 30% of their revenue. They 310 00:15:07.679 --> 00:15:10.799 wouldn't do that if enterprise clients weren't actively demanding K3 311 00:15:10.820 --> 00:15:12.000 to orchestrate their systems. 312 00:15:12.340 --> 00:15:12.580 I mean. 313 00:15:12.899 --> 00:15:16.049 Moonshot claims their performance gains stem from original changes to 314 00:15:16.070 --> 00:15:19.429 their underlying architecture, and the results back it up. K3 315 00:15:19.490 --> 00:15:23.710 ranks first in Arena.ai's web interface building benchmark. The industrial 316 00:15:23.769 --> 00:15:26.690 market is voting with its infrastructure dollars and it's voting 317 00:15:26.710 --> 00:15:27.320 for language. 318 00:15:27.700 --> 00:15:30.379 But again, we have to ask, what are they actually 319 00:15:30.440 --> 00:15:30.919 voting for? 320 00:15:30.960 --> 00:15:36.860 They are voting for the ability to automate administrative workflows, interface. 321 00:15:36.419 --> 00:15:37.820 Building, and software logic. 322 00:15:38.519 --> 00:15:39.600 Which is core to engineering. 323 00:15:40.120 --> 00:15:43.470 Which, I concede, is highly valuable. But when we look 324 00:15:43.590 --> 00:15:47.090 at the true horizon of systems engineering, the physical world 325 00:15:47.169 --> 00:15:48.389 is the ultimate constraint. 326 00:15:49.070 --> 00:15:49.490 Okay, but. 327 00:15:49.710 --> 00:15:52.809 Think about a next-generation chip design company. They don't just 328 00:15:52.830 --> 00:15:55.009 need a model that can write a Python script or 329 00:15:55.080 --> 00:15:58.370 generate a report on thermodynamics. They need a model that 330 00:15:58.509 --> 00:16:03.769 fundamentally understands heat at the molecular level without running a bespoke, brittle, 331 00:16:03.879 --> 00:16:08.200 computational fluid dynamic simulation that takes four days to render. Right. 332 00:16:08.539 --> 00:16:10.200 They need to move past the transformer. 333 00:16:10.220 --> 00:16:14.019 See, your framing of moving past the transformer implies that 334 00:16:14.059 --> 00:16:16.460 the transformer is a dead end rather than, you know, 335 00:16:16.539 --> 00:16:20.440 an evolving, generalized platform that can eventually integrate physical data. 336 00:16:20.740 --> 00:16:24.100 For native physics, it absolutely is a dead end. Next 337 00:16:24.179 --> 00:16:28.720 token prediction is... simply does not mathematically map to continuous 338 00:16:28.759 --> 00:16:29.730 space and time. 339 00:16:30.519 --> 00:16:32.029 You don't think they'll bridge that gap? 340 00:16:32.389 --> 00:16:36.110 You can scale a transformer to 10 trillion parameters, and 341 00:16:36.129 --> 00:16:39.330 it will still just be a highly sophisticated pattern matcher 342 00:16:39.710 --> 00:16:45.360 of discrete symbols. A physics-centric approach utilizing neural operators represents 343 00:16:45.409 --> 00:16:46.500 the true frontier. 344 00:16:47.039 --> 00:16:48.100 It's a frontier, sure. 345 00:16:48.440 --> 00:16:51.580 When an enterprise wants to push the boundaries of material science, 346 00:16:51.960 --> 00:16:55.710 extreme weather prediction, or robotics, They require an AI that 347 00:16:55.750 --> 00:16:59.820 predicts reality itself, not an AI that just generates text 348 00:16:59.909 --> 00:17:03.679 based on historical patterns of human language describing reality. 349 00:17:04.220 --> 00:17:06.980 And my position remains that the immense parameter scale and 350 00:17:07.059 --> 00:17:12.140 complex task capabilities of models like K3 demonstrate that language-centric 351 00:17:12.180 --> 00:17:15.400 intelligence is the most versatile practical tool available right now. 352 00:17:15.900 --> 00:17:17.740 Practical for the IT department, maybe. 353 00:17:18.160 --> 00:17:21.920 Practical for the whole operation. An industrial enterprise is not 354 00:17:22.009 --> 00:17:25.119 just a physics problem. It is a wildly complex web 355 00:17:25.160 --> 00:17:30.420 of humans, legacy software, physical interfaces, supply chains, and machines. 356 00:17:31.000 --> 00:17:31.460 That's true. 357 00:17:32.119 --> 00:17:36.309 Language is the universal translator across all of those diverse domains. 358 00:17:36.829 --> 00:17:39.690 The sheer demand from the major cloud giants proves this 359 00:17:39.750 --> 00:17:42.210 is the indispensable tool in the ecosystem. 360 00:17:42.589 --> 00:17:43.549 I hear you, but. 361 00:17:43.809 --> 00:17:46.869 You can have the most perfect physical simulation in the world. 362 00:17:47.339 --> 00:17:50.000 But if you can't seamlessly interface it with the rest 363 00:17:50.039 --> 00:17:53.660 of your enterprise stack through generalized reasoning, it stays locked 364 00:17:53.700 --> 00:17:54.200 in the lab. 365 00:17:54.880 --> 00:17:58.069 It is fascinating, though, because despite our deep disagreement on 366 00:17:58.130 --> 00:18:02.509 whether systems engineers should rely on continuous physics architectures or 367 00:18:02.589 --> 00:18:06.589 discrete language architectures, both of these approaches share a striking 368 00:18:06.630 --> 00:18:08.319 reality in the source material. 369 00:18:08.730 --> 00:18:09.450 They really do. 370 00:18:09.779 --> 00:18:16.099 Whether pursuing accelerated understandings neural operators or moonshots language-centric transformers, 371 00:18:16.589 --> 00:18:20.950 The future of this scale of AI is strictly enterprise-focused. 372 00:18:21.430 --> 00:18:24.769 That is a crucial point of convergence here. Neither accelerated 373 00:18:24.809 --> 00:18:29.920 understanding nor moonshot is prioritizing consumer applications right now. 374 00:18:30.279 --> 00:18:31.779 Right, no consumer apps. 375 00:18:32.000 --> 00:18:35.440 We aren't talking about chatbots for retail customers, but computational 376 00:18:35.480 --> 00:18:39.650 demands are simply too vast. They both require massive external 377 00:18:39.710 --> 00:18:44.650 computing partnerships just to function. Yeah. K3 obviously need AWS 378 00:18:44.690 --> 00:18:48.289 and Azure. And even Anand Kumar's startup had to rely 379 00:18:48.390 --> 00:18:52.230 on undisclosed hardware clusters just to develop their initial AI. 380 00:18:52.630 --> 00:18:56.690 The era of a startup building a frontier model entirely in-house, 381 00:18:57.140 --> 00:19:00.599 you know, isolated from the major cloud hardware giants, seems to. 382 00:19:00.539 --> 00:19:01.700 Be completely over. 383 00:19:01.960 --> 00:19:03.079 Oh, totally over. 384 00:19:03.359 --> 00:19:07.339 The staggering computing requirements are entirely dictating the business models 385 00:19:07.500 --> 00:19:08.400 and the partnerships. 386 00:19:08.680 --> 00:19:12.259 Absolutely. And it leaves a profound open question regarding innovation. 387 00:19:12.589 --> 00:19:17.009 how quickly systems engineers will actually adopt these competing paradigms 388 00:19:17.069 --> 00:19:19.990 on the factory floor. Right. Will they find a way 389 00:19:20.029 --> 00:19:24.390 to integrate neural operators directly into their workflow, bypassing the 390 00:19:24.430 --> 00:19:27.670 token economy? Or will they continue to mediate all of 391 00:19:27.710 --> 00:19:32.569 their complex engineering problems through the vast, generalized reasoning capabilities 392 00:19:32.630 --> 00:19:33.430 of a language model? 393 00:19:33.710 --> 00:19:37.299 The material clearly suggests there is much more nuance to explore. 394 00:19:37.789 --> 00:19:42.099 as these divergent technologies move from theoretical capability into heavy 395 00:19:42.160 --> 00:19:43.259 industrial application. 396 00:19:43.640 --> 00:19:44.519 Yeah, it's just the beginning. 397 00:19:44.859 --> 00:19:46.910 It brings us right back to our opening thought. 398 00:19:47.069 --> 00:19:47.529 The wind tunnel. 399 00:19:47.950 --> 00:19:48.470 Exactly. 400 00:19:48.910 --> 00:19:51.619 The systems engineer of tomorrow is standing in front of 401 00:19:51.660 --> 00:19:54.539