WEBVTT 00:00:00.000 --> 00:00:02.960 The next massive AI data center isn't being built 00:00:02.960 --> 00:00:05.179 in a remote desert. It's being installed right 00:00:05.179 --> 00:00:07.839 next to your AC unit, and it might just pay your 00:00:07.839 --> 00:00:10.859 utility bill. It is an absolutely wild concept 00:00:10.859 --> 00:00:12.619 to wrap your head around. Welcome back to the 00:00:12.619 --> 00:00:14.220 Deep Dive. I'm really glad you're joining us 00:00:14.220 --> 00:00:17.160 today. We have some truly fascinating ground 00:00:17.160 --> 00:00:19.800 to cover. The physical landscape of artificial 00:00:19.800 --> 00:00:22.839 intelligence is fundamentally shifting. We're 00:00:22.839 --> 00:00:25.219 looking at the hard limits of global computation, 00:00:25.460 --> 00:00:28.519 and we're seeing how engineers are quietly bypassing 00:00:28.519 --> 00:00:31.019 them. Yeah. Today we are exploring the distributed 00:00:31.019 --> 00:00:34.100 frontier of AI. We'll look at how supercomputers 00:00:34.100 --> 00:00:36.359 are moving into residential homes. We'll track 00:00:36.359 --> 00:00:38.880 the explosive agentic shift happening right now. 00:00:39.079 --> 00:00:41.679 This shift spans from simple Excel sheets to 00:00:41.679 --> 00:00:44.500 orbital satellites. And we'll unpack a massive 00:00:44.500 --> 00:00:47.240 new software breakthrough. It allows models to 00:00:47.240 --> 00:00:50.399 swallow 12 million words in a single pass. Let's 00:00:50.399 --> 00:00:52.079 start by looking at the physical infrastructure. 00:00:52.259 --> 00:00:54.600 We all know the AI boom is draining the power 00:00:54.600 --> 00:00:57.299 grid. Training these frontier models takes massive 00:00:57.299 --> 00:01:00.079 amounts of raw electricity. But building bigger 00:01:00.079 --> 00:01:02.399 power plants takes many, many years. So what 00:01:02.399 --> 00:01:04.519 if we just use the power capacity already wired 00:01:04.519 --> 00:01:06.659 into our homes? Right. And that is the exact 00:01:06.659 --> 00:01:10.379 premise of a startup called Span. They just announced 00:01:10.379 --> 00:01:13.120 a massive new infrastructure partnership. They're 00:01:13.120 --> 00:01:16.159 teaming up with NVIDIA and the home builder Polta 00:01:16.159 --> 00:01:17.780 Group. They're launching something called the 00:01:17.780 --> 00:01:21.579 XFRA Distributed Data Center. Yeah, the XFRA 00:01:21.579 --> 00:01:23.780 Distributed Data Center knows. I was reviewing 00:01:23.780 --> 00:01:26.120 the hardware specs earlier today. Each individual 00:01:26.120 --> 00:01:28.819 node is an absolute beast of a machine. They 00:01:28.819 --> 00:01:33.680 pack 16 NVIDIA RTX PRO 6000 Blackwell GPUs. Right, 00:01:33.760 --> 00:01:35.739 and they use the liquid -cooled server edition 00:01:35.739 --> 00:01:38.599 specifically. Exactly. They also include top 00:01:38.599 --> 00:01:42.379 -tier AMD EPYC processors. And they feature 3 00:01:42.379 --> 00:01:46.540 terabytes of RAM per node. beat. That is a staggering 00:01:46.540 --> 00:01:48.939 amount of local computing power. Why do we need 00:01:48.939 --> 00:01:51.140 three terabytes of memory sitting in a backyard? 00:01:51.400 --> 00:01:53.760 Well, it really comes down to loading massive 00:01:53.760 --> 00:01:56.140 parameter models natively. You need vast memory 00:01:56.140 --> 00:01:58.420 to hold these giant neural networks. So we're 00:01:58.420 --> 00:02:00.579 putting supercomputer level hardware directly 00:02:00.579 --> 00:02:03.620 into suburban neighborhoods. But the energy distribution 00:02:03.620 --> 00:02:06.219 is the real puzzle here. The copper wires in 00:02:06.219 --> 00:02:09.300 our streets have hard physical limits. How does 00:02:09.300 --> 00:02:11.800 a normal house power a massive supercomputer? 00:02:12.539 --> 00:02:14.439 This is where the smart panel technology comes 00:02:14.439 --> 00:02:17.379 in. The U .S. electrical grid was designed for 00:02:17.379 --> 00:02:21.139 peak theoretical loads. Most homes only use about 00:02:21.139 --> 00:02:24.180 40 % of their electrical capacity. Wow. Yeah, 00:02:24.240 --> 00:02:26.300 we have all this unused power just sitting there. 00:02:26.819 --> 00:02:29.800 SPAN's smart panel identifies that exact spare 00:02:29.800 --> 00:02:32.659 electricity dynamically. It monitors the home's 00:02:32.659 --> 00:02:35.680 power draw in milliseconds? Exactly. The panel 00:02:35.680 --> 00:02:38.780 detects when your oven or clothes dryer turns 00:02:38.780 --> 00:02:41.360 off. It then funnels that exact spare electrical 00:02:41.360 --> 00:02:45.060 capacity into the XFRA node. The node uses that 00:02:45.060 --> 00:02:47.939 diverted power to run complex cloud AI tasks. 00:02:48.259 --> 00:02:50.080 And the homeowner doesn't even notice the power 00:02:50.080 --> 00:02:52.340 shifting around. They don't notice a single thing 00:02:52.340 --> 00:02:54.479 changing. It operates entirely in the background 00:02:54.479 --> 00:02:56.719 of their daily lives. And the trade -off for 00:02:56.719 --> 00:02:59.240 the homeowner is incredibly compelling. In many 00:02:59.240 --> 00:03:01.500 of these markets, hosting a node brings massive 00:03:01.500 --> 00:03:03.520 benefits. You get completely free electricity 00:03:03.520 --> 00:03:06.340 and free gigabit internet. Span essentially pays 00:03:06.340 --> 00:03:08.849 your monthly utility bills for you. Yep. They 00:03:08.849 --> 00:03:10.770 cover your bills in exchange for using your wall 00:03:10.770 --> 00:03:12.590 space. They just tap into the spare electricity 00:03:12.590 --> 00:03:15.009 your house wasn't using anyway. It's like Airbnb 00:03:15.009 --> 00:03:18.289 for your home's unused electrical capacity. That's 00:03:18.289 --> 00:03:20.889 a perfect way to visualize the mechanism. And 00:03:20.889 --> 00:03:22.909 they're already rolling this infrastructure out 00:03:22.909 --> 00:03:25.150 right now. They're deploying them in new construction 00:03:25.150 --> 00:03:28.610 communities with Poltergroup. Think about the 00:03:28.610 --> 00:03:31.129 massive advantage in actual deployment speed. 00:03:31.270 --> 00:03:34.219 Instead of building one giant... 100 megawatt 00:03:34.219 --> 00:03:36.560 centralized data center, they can just deploy 00:03:36.560 --> 00:03:39.599 8 ,000 of these distributed mini nodes. It is 00:03:39.599 --> 00:03:42.400 six times faster to build this way. And it comes 00:03:42.400 --> 00:03:44.639 in at one fifth of the traditional infrastructure 00:03:44.639 --> 00:03:47.500 cost. Two sec silence. I've been thinking about 00:03:47.500 --> 00:03:49.879 the broader grid impact here. Is this actually 00:03:49.879 --> 00:03:52.360 safe and scalable? Or does putting a supercomputer 00:03:52.360 --> 00:03:55.560 next to my dryer create localized grid failures? 00:03:55.900 --> 00:03:58.360 It is a completely valid engineering concern 00:03:58.360 --> 00:04:01.400 to have. But the system actually balances local 00:04:01.400 --> 00:04:04.539 loads dynamically. Span smart panels act like 00:04:04.539 --> 00:04:06.939 highly intelligent traffic cops. They ensure 00:04:06.939 --> 00:04:09.560 the node only draws power when the house absolutely 00:04:09.560 --> 00:04:12.360 doesn't need it. This prevents any localized 00:04:12.360 --> 00:04:14.800 transformers from blowing out. So it balances 00:04:14.800 --> 00:04:17.220 local loads instead of draining the central grid. 00:04:17.379 --> 00:04:19.699 Precisely. We've been so worried about massive 00:04:19.699 --> 00:04:22.579 data centers sucking the grid dry. Span is proving 00:04:22.579 --> 00:04:24.560 we can just use the capacity we've already built. 00:04:24.660 --> 00:04:27.019 It turns your home from a passive cost center 00:04:27.019 --> 00:04:31.610 into a revenue generating asset. Beat. So hardware 00:04:31.610 --> 00:04:34.410 is moving directly into our suburban homes, but 00:04:34.410 --> 00:04:36.629 the physical distribution of compute is pushing 00:04:36.629 --> 00:04:39.550 much further out. Hardware is distributing all 00:04:39.550 --> 00:04:42.290 the way into the vacuum of space, and it's backed 00:04:42.290 --> 00:04:45.430 by truly astronomical amounts of corporate capital. 00:04:45.850 --> 00:04:47.970 The financial scale of this shift is honestly 00:04:47.970 --> 00:04:50.250 hard to comprehend. We're seeing numbers that 00:04:50.250 --> 00:04:53.050 fundamentally redefine corporate spending. Anthropic 00:04:53.050 --> 00:04:55.670 just committed to a massive new global infrastructure 00:04:55.670 --> 00:04:59.589 deal. They're spending $200 billion on Google 00:04:59.589 --> 00:05:02.730 Cloud and AI chips. $200 billion over just five 00:05:02.730 --> 00:05:05.310 years? That is roughly $40 billion every single 00:05:05.310 --> 00:05:08.069 year. Yeah. That deal represents over 40 % of 00:05:08.069 --> 00:05:11.040 Google Cloud's entire revenue backlog. The global 00:05:11.040 --> 00:05:13.680 AI infrastructure war is completely exploding 00:05:13.680 --> 00:05:16.319 right now. They are buying up silicon and energy 00:05:16.319 --> 00:05:19.379 contracts at an unprecedented pace. But they 00:05:19.379 --> 00:05:21.319 aren't just looking at traditional cloud servers 00:05:21.319 --> 00:05:24.259 on Earth. Anthropic just announced an incredible 00:05:24.259 --> 00:05:27.500 partnership with SpaceX AI. They secured access 00:05:27.500 --> 00:05:30.439 to the Colossus One massive AI supercomputer. 00:05:30.980 --> 00:05:34.360 And here's the truly mind -bending part of that 00:05:34.360 --> 00:05:37.220 specific partnership. Both organizations are 00:05:37.220 --> 00:05:40.180 officially exploring orbital AI compute in space. 00:05:40.639 --> 00:05:43.660 We're talking about massive data centers. orbiting 00:05:43.660 --> 00:05:47.019 the planet. Whoa. Imagine scaling orbital AI 00:05:47.019 --> 00:05:49.600 compute across thousands of satellites. Right. 00:05:49.699 --> 00:05:51.519 It completely changes the physical limits of 00:05:51.519 --> 00:05:53.699 our infrastructure. The cold vacuum of space 00:05:53.699 --> 00:05:56.459 solves massive thermal cooling problems. And 00:05:56.459 --> 00:05:58.759 this endless hardware scale is powering a fundamental 00:05:58.759 --> 00:06:01.800 software shift. AI is no longer just a simple 00:06:01.800 --> 00:06:03.860 chat bot answering questions. It is evolving 00:06:03.860 --> 00:06:06.199 into an active, autonomous digital operator. 00:06:06.600 --> 00:06:08.620 Yeah, the agentic shift is fully underway across 00:06:08.620 --> 00:06:10.540 the entire industry. I want to unpack what an 00:06:10.540 --> 00:06:13.079 agent actually is. A chatbot just predicts the 00:06:13.079 --> 00:06:15.540 next logical word in a sentence. An agent sets 00:06:15.540 --> 00:06:18.319 goals, uses external tools, and executes actions 00:06:18.319 --> 00:06:21.360 autonomously. Exactly. OpenAI is reportedly building 00:06:21.360 --> 00:06:24.000 a dedicated AI phone right now. They expect early 00:06:24.000 --> 00:06:26.040 hardware production to start by the year 2027. 00:06:26.300 --> 00:06:28.839 It completes background tasks, understands your 00:06:28.839 --> 00:06:31.579 goals, and works like a true operator. It is 00:06:31.579 --> 00:06:33.959 not just a standard mobile device anymore. And 00:06:33.959 --> 00:06:36.040 we're seeing this deep agentic integration in 00:06:36.040 --> 00:06:38.600 the workplace, too. Anthropic just launched 10 00:06:38.600 --> 00:06:41.980 ready to use autonomous agents. They are built 00:06:41.980 --> 00:06:44.920 specifically for the finance and insurance sectors. 00:06:45.139 --> 00:06:47.259 These specific agents are doing highly complex 00:06:47.259 --> 00:06:49.839 professional work. They can screen dense KYC 00:06:49.839 --> 00:06:52.180 files for banking compliance. They can review 00:06:52.180 --> 00:06:54.680 massive corporate earnings reports in seconds. 00:06:54.740 --> 00:06:57.079 Wow. They can even build complex presentation 00:06:57.079 --> 00:07:00.160 decks completely from scratch. ChatGPT is also 00:07:00.160 --> 00:07:02.819 moving directly into our daily corporate workflows. 00:07:03.149 --> 00:07:05.910 It now works natively inside Microsoft Excel 00:07:05.910 --> 00:07:08.649 and Google Sheets. You can use it to build complex 00:07:08.649 --> 00:07:11.129 financial formulas automatically. It reads raw 00:07:11.129 --> 00:07:13.509 data, generates insights, and formats the entire 00:07:13.509 --> 00:07:15.449 spreadsheet. They're rolling out a free beta 00:07:15.449 --> 00:07:18.589 for paid users right now. And OpenAI is aggressively 00:07:18.589 --> 00:07:20.750 pushing this integration into higher education. 00:07:21.189 --> 00:07:23.949 They just introduced the ChatGPT Futures Class 00:07:23.949 --> 00:07:26.910 of 2026 program. Right. They gave 26 student 00:07:26.910 --> 00:07:30.189 builders $10 ,000 grants. And they gave them 00:07:30.189 --> 00:07:33.660 full, unlimited access to frontier AI models. 00:07:33.939 --> 00:07:36.459 It is the first generation to start and finish 00:07:36.459 --> 00:07:39.660 college with GPT. We're studying how students 00:07:39.660 --> 00:07:42.759 leverage frontier AI throughout their entire 00:07:42.759 --> 00:07:45.620 degree. We're watching the baseline of human 00:07:45.620 --> 00:07:49.279 productivity shift in real time. Beat, but this 00:07:49.279 --> 00:07:52.220 deep integration introduces some very strange 00:07:52.220 --> 00:07:55.040 new vulnerabilities. We're trusting autonomous 00:07:55.040 --> 00:07:57.560 agents with highly sensitive personal financial 00:07:57.560 --> 00:08:00.399 data. The security risks are evolving just as 00:08:00.399 --> 00:08:03.560 fast as the model capabilities. Researchers just 00:08:03.560 --> 00:08:06.339 exposed a completely new AI scam trick last week. 00:08:06.459 --> 00:08:08.899 They found a brilliant way to manipulate agents 00:08:08.899 --> 00:08:11.740 using hidden instructions. Yeah, the hidden malicious 00:08:11.740 --> 00:08:13.980 instructions were written entirely in Morse code, 00:08:14.199 --> 00:08:16.620 just simple dots and dashes buried deeply inside 00:08:16.620 --> 00:08:18.699 standard text files. And the terrifying part 00:08:18.699 --> 00:08:21.060 is that some agents actually followed them. The 00:08:21.060 --> 00:08:23.160 AI recognized the Morse code formatting perfectly. 00:08:23.439 --> 00:08:26.439 It then bypassed its safety filters and executed 00:08:26.439 --> 00:08:28.779 the hidden malicious commands. I still wrestle 00:08:28.779 --> 00:08:30.920 with prompt drift myself, so agents getting hacked 00:08:30.920 --> 00:08:33.139 by Morse code is terrifying. It sounds like a 00:08:33.139 --> 00:08:35.940 strange plot from a science fiction novel. But 00:08:35.940 --> 00:08:38.200 this is the harsh reality of deploying autonomous 00:08:38.200 --> 00:08:41.559 agents. Large language models process information 00:08:41.559 --> 00:08:44.340 through mathematical tokenization. They don't 00:08:44.340 --> 00:08:46.320 read English the way human beings do. Right. 00:08:46.600 --> 00:08:48.759 They see patterns, and Morse code is just another 00:08:48.759 --> 00:08:51.419 mathematical pattern. We are exposing ourselves 00:08:51.419 --> 00:08:55.080 to entirely new cryptographic attack vectors. 00:08:55.320 --> 00:08:57.759 And the traditional legal system is desperately 00:08:57.759 --> 00:09:00.320 trying to catch up. Apple just agreed to pay 00:09:00.320 --> 00:09:04.659 $250 million. They're settling a massive class 00:09:04.659 --> 00:09:06.759 action lawsuit regarding artificial intelligence 00:09:06.759 --> 00:09:09.419 right now. The underlying lawsuit was tied to 00:09:09.419 --> 00:09:12.889 claims about specific iPhone capabilities. Buyers 00:09:12.889 --> 00:09:15.429 claimed they were misled about Siri's newly advertised 00:09:15.429 --> 00:09:18.830 AI features. Apple agreed to the massive financial 00:09:18.830 --> 00:09:22.029 settlement to avoid a lengthy trial. They settled 00:09:22.029 --> 00:09:23.789 this massive case even though they admitted no 00:09:23.789 --> 00:09:26.730 official wrongdoing. Two sec silence. With agents 00:09:26.730 --> 00:09:29.009 acting autonomously in our spreadsheets and phones, 00:09:29.309 --> 00:09:32.009 who takes the fall when a Morse code hack successfully 00:09:32.009 --> 00:09:35.649 steals data? That is the defining legal question 00:09:35.649 --> 00:09:38.789 of this entire new era. The traditional paradigm 00:09:38.789 --> 00:09:41.230 of software security is shifting rapidly right 00:09:41.230 --> 00:09:44.759 now. Historically, the end user was heavily responsible 00:09:44.759 --> 00:09:47.759 for their own data security. But when the AI 00:09:47.759 --> 00:09:50.519 makes autonomous background decisions, the liability 00:09:50.519 --> 00:09:53.940 fundamentally moves. The legal burden falls heavily 00:09:53.940 --> 00:09:56.820 onto the developers building the underlying models. 00:09:57.120 --> 00:09:59.720 Ah, so security liability shifts entirely from 00:09:59.720 --> 00:10:02.539 user to the AI developer. Exactly. The tech companies 00:10:02.539 --> 00:10:04.620 building this intelligence have to legally secure 00:10:04.620 --> 00:10:07.519 it. They are creating autonomous actors, so they 00:10:07.519 --> 00:10:10.529 hold the ultimate responsibility. Sponsor. We're 00:10:10.529 --> 00:10:13.190 back. We've explored supercomputers sitting right 00:10:13.190 --> 00:10:15.610 next to our air conditioners. We've tracked autonomous 00:10:15.610 --> 00:10:18.070 agents running our daily financial spreadsheets. 00:10:18.090 --> 00:10:20.929 Beat. But for these complex agents to work effectively, 00:10:21.190 --> 00:10:23.789 they need vast memory. They need to instantly 00:10:23.789 --> 00:10:26.950 recall massive amounts of contextual data. And 00:10:26.950 --> 00:10:28.669 that brings us to an incredible new software 00:10:28.669 --> 00:10:31.409 breakthrough. It perfectly matches the explosive 00:10:31.409 --> 00:10:33.370 scale of the physical hardware we discussed. 00:10:33.669 --> 00:10:37.049 A company called SubQuadratic just launched a 00:10:37.049 --> 00:10:39.570 fascinating new AI model. They're calling this 00:10:39.570 --> 00:10:43.190 powerful new model SubQ. It's being described 00:10:43.190 --> 00:10:47.289 as a 12 million token memory hack killer. SubQ 00:10:47.289 --> 00:10:51.230 can cleanly swallow 12 million tokens in a single 00:10:51.230 --> 00:10:54.950 analytical pass. That is a truly incomprehensible 00:10:54.950 --> 00:10:57.830 amount of raw contextual information. It's like 00:10:57.830 --> 00:11:00.509 feeding an entire corporate code repository into 00:11:00.509 --> 00:11:03.490 one single prompt. Right. And the engineering 00:11:03.490 --> 00:11:06.190 team behind this model is absolutely world class. 00:11:06.470 --> 00:11:09.330 They hail from Meta. Google DeepMind and Oxford 00:11:09.330 --> 00:11:11.600 University. They've built an architecture that 00:11:11.600 --> 00:11:13.759 fundamentally changes the core processing economics. 00:11:14.200 --> 00:11:16.419 The massive cost difference is what really stands 00:11:16.419 --> 00:11:19.799 out to me here. It costs just $8 to run a massive 00:11:19.799 --> 00:11:23.340 context task on SubQ. That is a staggering price 00:11:23.340 --> 00:11:25.240 drop for working software developers. If you 00:11:25.240 --> 00:11:27.559 run that exact same massive task on traditional 00:11:27.559 --> 00:11:30.440 frontier models, it is wildly expensive and incredibly 00:11:30.440 --> 00:11:33.480 slow to process. It costs roughly $2 ,600 on 00:11:33.480 --> 00:11:36.740 those older legacy systems. And SubQ runs 52 00:11:36.740 --> 00:11:38.899 times faster than the current industry standard. 00:11:39.320 --> 00:11:41.679 Let's unpack exactly how they achieved this massive 00:11:41.679 --> 00:11:44.200 leap in efficiency. They use a highly optimized 00:11:44.200 --> 00:11:47.679 selective attention neural architecture. Traditional 00:11:47.679 --> 00:11:50.299 models calculate the relationship between every 00:11:50.299 --> 00:11:53.559 single word in a document. That requires massive 00:11:53.559 --> 00:11:56.039 computing power as the document gets larger and 00:11:56.039 --> 00:11:58.940 longer. Their system only looks at the token 00:11:58.940 --> 00:12:01.799 relationships that actually matter. It drops 00:12:01.799 --> 00:12:03.980 irrelevant contextual connections instantly. 00:12:04.259 --> 00:12:07.240 That is how they achieve such incredible, groundbreaking 00:12:07.240 --> 00:12:10.240 processing speeds. We should define a few important 00:12:10.240 --> 00:12:12.620 architectural concepts here. Flash attention 00:12:12.620 --> 00:12:15.539 is a widely used method to speed up AI memory 00:12:15.539 --> 00:12:17.659 processing. Right. It is a critical optimization 00:12:17.659 --> 00:12:19.919 technique for modern hardware architectures. 00:12:19.960 --> 00:12:21.879 And we also need to talk about traditional retrieval 00:12:21.879 --> 00:12:24.779 systems. The industry has relied heavily on complex 00:12:24.779 --> 00:12:27.740 RG pipelines for years now. These are systems 00:12:27.740 --> 00:12:30.259 that fetch outside data to help AI and... or 00:12:30.259 --> 00:12:32.679 questions. You build a vector database to search 00:12:32.679 --> 00:12:35.399 for small text fragments. Yeah, those traditional 00:12:35.399 --> 00:12:37.899 document retrieval systems are incredibly complex 00:12:37.899 --> 00:12:40.980 and fragile. But SubQ might make those complicated 00:12:40.980 --> 00:12:44.340 RRAG pipelines a thing of the past. You simply 00:12:44.340 --> 00:12:47.320 don't need to fetch small chunks of outside data 00:12:47.320 --> 00:12:50.340 anymore. You just feed the entire massive database 00:12:50.340 --> 00:12:53.759 directly into the working model. It's like stacking 00:12:53.759 --> 00:12:55.840 Lego blocks of data, but instead of building 00:12:55.840 --> 00:12:58.019 piece by piece, you dump the whole bucket and 00:12:58.019 --> 00:13:00.519 the AI instantly sees the final castle. That 00:13:00.519 --> 00:13:02.899 visual perfectly captures the leap in processing 00:13:02.899 --> 00:13:05.940 capability. And the verified benchmark performance 00:13:05.940 --> 00:13:09.820 of SubQ is simply incredible. Most standard models 00:13:09.820 --> 00:13:12.360 lose the plot as the input files get bigger. 00:13:12.480 --> 00:13:15.059 They suffer from the classic needle in a haystack 00:13:15.059 --> 00:13:17.700 problem. Right. They completely forget important 00:13:17.700 --> 00:13:20.220 information hidden deeply in the middle of the 00:13:20.220 --> 00:13:23.299 text. But SubQ hit 92 % recall at 12 million 00:13:23.299 --> 00:13:26.080 tokens. It remembers almost everything you feed 00:13:26.080 --> 00:13:29.480 into its working context window. Just for comparative 00:13:29.480 --> 00:13:31.500 context, we could look at the major industry 00:13:31.500 --> 00:13:35.389 competitors. Gemini 3 .1 Pro struggled significantly 00:13:35.389 --> 00:13:38.470 on similar massive multi -needle tests. The exact 00:13:38.470 --> 00:13:41.450 same degradation is true for OpenAI's GPT -5 00:13:41.450 --> 00:13:45.210 .4. They simply could not maintain accurate information 00:13:45.210 --> 00:13:48.789 recall at that massive scale. SubQ is absolutely 00:13:48.789 --> 00:13:51.269 crushing the established giants on these specific 00:13:51.269 --> 00:13:53.250 benchmarks. Yeah, they've already launched a 00:13:53.250 --> 00:13:55.350 specialized developer tool called SubQ Code. 00:13:55.549 --> 00:13:58.190 It's a command line interface agent built directly 00:13:58.190 --> 00:14:00.830 for professional software engineers. It can load 00:14:00.830 --> 00:14:03.690 your entire complex code base in just one single 00:14:03.690 --> 00:14:06.169 pass. And they are certainly not stopping at 00:14:06.169 --> 00:14:08.889 12 million tokens. The engineering team is officially 00:14:08.889 --> 00:14:11.909 targeting a 100 million token context window. 00:14:12.110 --> 00:14:14.169 They want to hit that massive milestone by the 00:14:14.169 --> 00:14:16.629 end of 2026. We have seen so -called transformer 00:14:16.629 --> 00:14:19.309 killers hyped up before. Alternate architectural 00:14:19.309 --> 00:14:21.769 models like Mamba tried to dethrone the standard 00:14:21.769 --> 00:14:24.110 transformer architecture. But SubQ genuinely 00:14:24.110 --> 00:14:26.210 feels like a permanent shift in the landscape. 00:14:26.450 --> 00:14:29.090 It's launching with a highly robust, production 00:14:29.090 --> 00:14:32.129 -ready developer API today. If you're a software 00:14:32.129 --> 00:14:34.350 engineer building agents, this is a massive win. 00:14:34.570 --> 00:14:37.370 Two sec silence. I have to wonder about the long 00:14:37.370 --> 00:14:41.200 -term architectural implications here. If an 00:14:41.200 --> 00:14:45.220 AI can perfectly recall 12 million tokens for 00:14:45.220 --> 00:14:48.220 eight bucks, does this completely kill the need 00:14:48.220 --> 00:14:51.100 for complex retrieval systems? It fundamentally 00:14:51.100 --> 00:14:53.759 rewrites how developers will build AI applications 00:14:53.759 --> 00:14:56.649 moving forward. You simply won't need clunky 00:14:56.649 --> 00:14:59.409 vector databases holding tiny fragments of text. 00:14:59.610 --> 00:15:02.289 You won't need complicated routing logic to find 00:15:02.289 --> 00:15:04.730 the right source document. You'll just load the 00:15:04.730 --> 00:15:06.970 entire required operational context directly 00:15:06.970 --> 00:15:09.269 into the working memory. Right. Massive context 00:15:09.269 --> 00:15:12.210 windows will just replace complex data retrieval 00:15:12.210 --> 00:15:14.629 pipelines entirely. It simplifies the entire 00:15:14.629 --> 00:15:17.029 software development process for engineers globally. 00:15:17.350 --> 00:15:19.690 Let's step back and look at the larger big picture 00:15:19.690 --> 00:15:22.649 here. The core physical bottlenecks of artificial 00:15:22.649 --> 00:15:25.070 intelligence are shattering simultaneously right 00:15:25.070 --> 00:15:27.909 now. We're seeing hardware limits being bypassed 00:15:27.909 --> 00:15:30.149 in incredibly creative and distributed ways. 00:15:30.470 --> 00:15:32.590 We are actively capturing the unused electrical 00:15:32.590 --> 00:15:34.929 capacity of our suburban homes. We are looking 00:15:34.929 --> 00:15:37.169 at deploying massive computing nodes in the cold 00:15:37.169 --> 00:15:39.789 vacuum of space. The physical deployment constraints 00:15:39.789 --> 00:15:42.950 that held us back are rapidly disappearing. And 00:15:42.950 --> 00:15:45.450 the software processing limits are evaporating 00:15:45.450 --> 00:15:48.120 just as fast today. New architectural models 00:15:48.120 --> 00:15:50.779 like SubQ can process entire repositories of 00:15:50.779 --> 00:15:54.320 human knowledge in seconds. We are rapidly transitioning 00:15:54.320 --> 00:15:57.679 away from the passive, simple chatbot era. We're 00:15:57.679 --> 00:16:00.840 actively unleashing highly autonomous, ubiquitous 00:16:00.840 --> 00:16:03.639 software agents into the real world. This digital 00:16:03.639 --> 00:16:05.960 intelligence is moving directly into our Excel 00:16:05.960 --> 00:16:08.000 sheets and our phones. It is running in your 00:16:08.000 --> 00:16:12.490 backyard, your devices, and in orbit. Beat. Which 00:16:12.490 --> 00:16:14.350 brings me to a final thought for you to ponder 00:16:14.350 --> 00:16:17.009 today. If your home's AC unit, your phone, and 00:16:17.009 --> 00:16:19.190 satellites in orbit are all continuously running 00:16:19.190 --> 00:16:22.029 massive AI agents that can instantly recall millions 00:16:22.029 --> 00:16:25.210 of data points, where does human decision -making 00:16:25.210 --> 00:16:27.710 actually sit in the loop? Thank you so much for 00:16:27.710 --> 00:16:29.750 exploring this deeply with us today. We will 00:16:29.750 --> 00:16:31.330 see you next time. You might be sitting right 00:16:31.330 --> 00:16:33.169 next to your air conditioner. Out to your own 00:16:33.169 --> 00:16:33.429 music.