WEBVTT 00:00:00.000 --> 00:00:02.299 Imagine slipping on a virtual reality headset 00:00:02.299 --> 00:00:05.120 right now. You walk directly into Monica's apartment 00:00:05.120 --> 00:00:07.780 from friends, you pick up a heavy ceramic coffee 00:00:07.780 --> 00:00:10.580 mug, you drop it, and it shatters across the 00:00:10.580 --> 00:00:13.740 floor. It shatters with perfectly accurate physical 00:00:13.740 --> 00:00:16.600 laws. This isn't just some polished video game 00:00:16.600 --> 00:00:19.239 environment, it's a leaked AI building physical 00:00:19.239 --> 00:00:22.359 reality from scratch. We are also looking at 00:00:22.359 --> 00:00:25.379 a brand new vision model. It maps the physical 00:00:25.379 --> 00:00:28.660 world in real time. And it does this using a 00:00:28.660 --> 00:00:34.049 cheap Welcome to the Deep Dive. We have a genuinely 00:00:34.049 --> 00:00:37.509 fascinating roadmap for you today. We're unpacking 00:00:37.509 --> 00:00:40.689 some massive new leaks from OpenAI. We're also 00:00:40.689 --> 00:00:44.469 exploring a staggering $100 billion compute race. 00:00:44.609 --> 00:00:47.350 This massive arms race is happening quietly behind 00:00:47.350 --> 00:00:50.030 closed doors. Finally, we'll break down a major 00:00:50.030 --> 00:00:52.670 breakthrough in real -time 3D vision. Let's start 00:00:52.670 --> 00:00:55.189 by looking closely at the situation inside OpenAI. 00:00:55.679 --> 00:00:58.500 Late last year, Sam Altman issued a massive code 00:00:58.500 --> 00:01:00.979 red. Yeah, he did. That internal warning triggered 00:01:00.979 --> 00:01:03.259 a really aggressive two -pronged release strategy. 00:01:03.759 --> 00:01:06.180 They are feeling immense pressure from their 00:01:06.180 --> 00:01:08.609 competitors right now. The underlying context 00:01:08.609 --> 00:01:12.010 here is totally crucial to understand. They completely 00:01:12.010 --> 00:01:14.730 missed their active user goals for late 2025. 00:01:15.390 --> 00:01:18.650 They originally wanted 1 billion weekly active 00:01:18.650 --> 00:01:21.310 users. Instead, they watched Anthropic experience 00:01:21.310 --> 00:01:24.549 a massive revenue surge. OpenAI also dealt with 00:01:24.549 --> 00:01:27.349 some undeniably bad vibes regarding leadership. 00:01:27.689 --> 00:01:30.010 Right. So they are aggressively striking back 00:01:30.010 --> 00:01:32.530 right now. That brings us to their first major 00:01:32.530 --> 00:01:35.680 leet project. It is currently operating under 00:01:35.680 --> 00:01:39.640 the internal code name Spud. OpenAI is actively 00:01:39.640 --> 00:01:42.040 A -B testing this model in the wild. Yeah, it's 00:01:42.040 --> 00:01:44.540 out there. Some users are randomly seeing it 00:01:44.540 --> 00:01:48.560 inside the GPT 5 .4 Pro interface. Spud apparently 00:01:48.560 --> 00:01:51.140 represents highly advanced spatial reasoning 00:01:51.140 --> 00:01:53.719 capabilities. Right, and spatial reasoning changes 00:01:53.719 --> 00:01:56.500 the entire paradigm completely. A standard language 00:01:56.500 --> 00:01:58.700 model simply predicts the next word. It guesses 00:01:58.700 --> 00:02:00.900 what text logically follows your specific prompt. 00:02:01.159 --> 00:02:03.540 But Spud is predicting the next physical state 00:02:03.540 --> 00:02:06.379 instead. If you drop a virtual mug, it calculates 00:02:06.379 --> 00:02:08.860 gravity, it understands momentum, mass, and physical 00:02:08.860 --> 00:02:11.319 collision in real time. That's how it rebuilt 00:02:11.319 --> 00:02:13.500 a functional 3D version of Monica's apartment. 00:02:13.819 --> 00:02:16.300 You can interact with the incredibly realistic 00:02:16.300 --> 00:02:19.560 physics engine inside. It even generates incredible 00:02:19.560 --> 00:02:22.479 Minecraft -style voxel art from basic prompts. 00:02:22.800 --> 00:02:25.919 What exactly is the mechanism behind voxel art? 00:02:26.159 --> 00:02:29.379 Voxel art means 3D digital models built from... 00:02:29.680 --> 00:02:32.680 tiny cubes. Exactly. You use simple language 00:02:32.680 --> 00:02:36.000 to describe a complex structure. The AI then 00:02:36.000 --> 00:02:38.039 constructs that physical environment perfectly. 00:02:38.300 --> 00:02:40.259 So it is exactly like stacking Lego blocks of 00:02:40.259 --> 00:02:42.990 data. That is the perfect analogy for the specific 00:02:42.990 --> 00:02:45.750 technology. It's a massive shift in underlying 00:02:45.750 --> 00:02:49.310 digital capability. Spud is also producing scalable 00:02:49.310 --> 00:02:52.750 SVG designs for everyday developers. Let's explain 00:02:52.750 --> 00:02:56.150 why that specific detail actually matters. Scalable 00:02:56.150 --> 00:02:58.530 vector graphics are math -based codes for drawing 00:02:58.530 --> 00:03:01.310 crisp images. They don't rely on a fixed grid 00:03:01.310 --> 00:03:03.509 of colored pixels. This means you can scale them 00:03:03.509 --> 00:03:05.780 infinitely without any blurring. Spud generates 00:03:05.780 --> 00:03:08.860 this complex mathematical code incredibly efficiently. 00:03:09.099 --> 00:03:11.319 It accomplishes this with significantly fewer 00:03:11.319 --> 00:03:14.400 lines of code. It creates incredibly clean, professional 00:03:14.400 --> 00:03:17.500 -grade, minimalistic layouts. Reducing the lines 00:03:17.500 --> 00:03:19.699 of code is actually a massive deal. It means 00:03:19.699 --> 00:03:21.620 the model operates with much greater computational 00:03:21.620 --> 00:03:24.680 efficiency. There is simply less room for hidden 00:03:24.680 --> 00:03:27.780 errors to accumulate. It's heavily outperforming 00:03:27.780 --> 00:03:30.379 Claude Opus 4 .7 in technical precision right 00:03:30.379 --> 00:03:32.919 now. This connects directly to the mysterious 00:03:32.919 --> 00:03:36.379 chatbot arena leaks recently. We've seen three 00:03:36.379 --> 00:03:38.860 very strange models testing on the platform. 00:03:39.180 --> 00:03:42.800 They are named Masking Tape Alpha, Gaffer Tape 00:03:42.800 --> 00:03:45.039 Alpha, and Packing Tape Alpha. Classic naming 00:03:45.039 --> 00:03:47.599 scheme. Insiders confirm this is actually a model 00:03:47.599 --> 00:03:51.340 called GPT Image 2. It was built to directly 00:03:51.340 --> 00:03:54.840 rival Google's Nano Banana Pro. They frequently 00:03:54.840 --> 00:03:57.780 use tape names to mask their true identity. But 00:03:57.780 --> 00:03:59.979 the ultimate strategic goal here is remarkably 00:03:59.979 --> 00:04:02.520 clear. They are betting heavily on hyper realistic 00:04:02.520 --> 00:04:05.400 personal avatars right now. These are gorgeous 00:04:05.400 --> 00:04:08.539 Studio Ghibli style personalized digital creations. 00:04:08.860 --> 00:04:10.680 They think these avatars will trigger a viral 00:04:10.680 --> 00:04:13.159 user surge. They desperately want to replicate 00:04:13.159 --> 00:04:15.719 the massive growth of early 2025. The visual 00:04:15.719 --> 00:04:17.560 quality is supposed to be absolutely breathtaking. 00:04:17.899 --> 00:04:20.040 They need this viral hit to reach that billion 00:04:20.040 --> 00:04:23.000 user goal. Whoa. Imagine scaling to a billion. 00:04:23.819 --> 00:04:27.740 Queries. Two sec silence. The digital infrastructure 00:04:27.740 --> 00:04:31.459 required to support that is hard to grasp. How 00:04:31.459 --> 00:04:34.399 does shifting from flat text to 3D spatial environments 00:04:34.399 --> 00:04:38.259 change OpenAI's ultimate endgame? It fundamentally 00:04:38.259 --> 00:04:40.740 transforms their entire core business model. 00:04:41.199 --> 00:04:43.420 They are evolving away from a simple chatbot 00:04:43.420 --> 00:04:46.240 utility. They are becoming an immersive world 00:04:46.240 --> 00:04:49.160 -building platform instead. They want to be the 00:04:49.160 --> 00:04:52.319 primary engine for future virtual realities. 00:04:52.560 --> 00:04:54.399 So they're trading flat text for interactive, 00:04:54.540 --> 00:04:56.839 personalized physical reality. Exactly. It's 00:04:56.839 --> 00:04:58.560 a completely different level of technological 00:04:58.560 --> 00:05:01.300 ambition. OpenAI is trying to win by building 00:05:01.300 --> 00:05:04.160 interactive digital worlds. But Anthropic is 00:05:04.160 --> 00:05:06.399 taking a completely different, highly physical 00:05:06.399 --> 00:05:09.180 path. They are making truly monumental physical 00:05:09.180 --> 00:05:11.040 infrastructure plays right now. Yeah, they are. 00:05:11.160 --> 00:05:13.379 Amazon is quietly investing up to $25 billion 00:05:13.379 --> 00:05:17.019 more. This massively adds to their existing $8 00:05:17.019 --> 00:05:19.699 billion stake. The underlying infrastructure 00:05:19.699 --> 00:05:22.360 math here is genuinely mind -bending. Anthropic 00:05:22.360 --> 00:05:25.579 plans to spend $100 billion on AWS chips. They 00:05:25.579 --> 00:05:27.360 are rolling this out over the next 10 years. 00:05:27.500 --> 00:05:30.860 They are actively securing 5 gigawatts of raw 00:05:30.860 --> 00:05:33.279 compute power. Let's put that 5 gigawatts into 00:05:33.279 --> 00:05:36.709 proper physical perspective. One single gigawatt 00:05:36.709 --> 00:05:39.410 can easily power a medium sized American city. 00:05:39.990 --> 00:05:41.970 Anthropic is securing enough physical energy 00:05:41.970 --> 00:05:45.230 to power five cities and is all feeding into 00:05:45.230 --> 00:05:48.089 a massive digital brain. They desperately need 00:05:48.089 --> 00:05:51.250 this to meet. Surging global cloud demand. Right. 00:05:51.350 --> 00:05:53.649 And we need to examine the actual user friction 00:05:53.649 --> 00:05:56.290 here. The newly released cloud design tool is 00:05:56.290 --> 00:05:58.610 going highly viral. People are creating wild 00:05:58.610 --> 00:06:01.529 visuals and complex mockups from basic prompts. 00:06:01.730 --> 00:06:04.069 Yeah. The design outputs are undeniably spectacular 00:06:04.069 --> 00:06:06.910 and widely shared online. But their flagship 00:06:06.910 --> 00:06:08.810 reasoning model is currently struggling quite 00:06:08.810 --> 00:06:12.610 badly. Claude Opus 4 .7 is facing intense criticism 00:06:12.610 --> 00:06:15.290 from everyday developers. Users have mockingly 00:06:15.290 --> 00:06:18.569 dubbed it Gaslitus 4 .7 online recently. It is 00:06:18.569 --> 00:06:20.910 wildly hallucinating digital files that simply 00:06:20.910 --> 00:06:23.370 do not exist. It delivers incredibly stubborn, 00:06:23.470 --> 00:06:25.870 completely incorrect outputs during complex coding 00:06:25.870 --> 00:06:28.709 tasks. It will literally argue with a developer 00:06:28.709 --> 00:06:31.850 over basic code. It refuses to correct itself, 00:06:32.029 --> 00:06:35.430 even when clearly shown the error. This definitely 00:06:35.430 --> 00:06:37.790 raises some major architectural concerns for 00:06:37.790 --> 00:06:40.149 their immediate future. Does this mean there 00:06:40.149 --> 00:06:43.410 is serious trouble ahead for Opus 4 .8? I still 00:06:43.410 --> 00:06:47.069 wrestle with prompt drift myself. Beat. It is 00:06:47.069 --> 00:06:49.329 incredibly frustrating when the model forgets 00:06:49.329 --> 00:06:52.329 initial instructions. Over a long conversation, 00:06:52.589 --> 00:06:55.250 it simply loses the main logical thread. But 00:06:55.250 --> 00:06:57.449 reliability isn't strictly an anthropic problem 00:06:57.449 --> 00:07:00.420 right now. That is entirely true across the entire 00:07:00.420 --> 00:07:03.540 global tech industry. ChatGPT recently suffered 00:07:03.540 --> 00:07:06.699 a massive, completely worldwide digital outage. 00:07:06.740 --> 00:07:09.120 It totally took down Codex and all their API 00:07:09.120 --> 00:07:11.819 services simultaneously. Thousands of developers 00:07:11.819 --> 00:07:13.920 couldn't work, code, or research anything online. 00:07:14.399 --> 00:07:16.519 OpenAI says the root issue is currently under 00:07:16.519 --> 00:07:19.019 active investigation. It highlights the deep 00:07:19.019 --> 00:07:21.920 fragility of this entire AI ecosystem. We rely 00:07:21.920 --> 00:07:24.000 so heavily on these hidden servers for our daily 00:07:24.000 --> 00:07:27.029 work. With billions spent on compute, Why are 00:07:27.029 --> 00:07:29.209 flagship models still stubbornly hallucinating 00:07:29.209 --> 00:07:32.410 files? Massive processing power doesn't automatically 00:07:32.410 --> 00:07:36.170 fix underlying architecture flaws. Neural networks 00:07:36.170 --> 00:07:39.129 don't actually think in a logical step -by -step 00:07:39.129 --> 00:07:42.149 progression. They analyze massive patterns to 00:07:42.149 --> 00:07:45.350 guess the most likely outcome. They are constantly 00:07:45.350 --> 00:07:47.990 predicting the next most likely token in a sequence. 00:07:48.410 --> 00:07:51.029 If the underlying logic path is fundamentally 00:07:51.029 --> 00:07:55.110 flawed, compute doesn't help. Adding more compute 00:07:55.110 --> 00:07:57.689 just makes the model confidently wrong much faster. 00:07:57.889 --> 00:08:00.889 The fundamental reasoning pathways still desperately 00:08:00.889 --> 00:08:04.189 need radical structural improvement. Throwing 00:08:04.189 --> 00:08:06.569 raw compute at a model won't magically cure its 00:08:06.569 --> 00:08:08.689 fundamental reasoning bugs. That is the harsh 00:08:08.689 --> 00:08:11.449 reality of the current generation. This massive 00:08:11.449 --> 00:08:14.089 scale and unreliability isn't just an abstract 00:08:14.089 --> 00:08:16.730 theoretical problem. It's fundamentally changing 00:08:16.730 --> 00:08:19.290 how everyday businesses are actively operating 00:08:19.290 --> 00:08:21.750 today. Everyday workflows are constantly bending 00:08:21.750 --> 00:08:24.730 around these new AI capabilities. Yeah. Adobe 00:08:24.730 --> 00:08:27.410 recently made a very revealing public statement 00:08:27.410 --> 00:08:30.550 about this shift. They flatly admitted AI could 00:08:30.550 --> 00:08:33.460 deeply disrupt their own massive business. They 00:08:33.460 --> 00:08:34.899 aren't just sitting back and watching it happen, 00:08:34.960 --> 00:08:38.139 though. They immediately released new CX enterprise 00:08:38.139 --> 00:08:41.320 agents for corporate businesses. These digital 00:08:41.320 --> 00:08:44.100 agents completely automate complex marketing 00:08:44.100 --> 00:08:47.440 and sales workflows. Adobe is also actively partnering 00:08:47.440 --> 00:08:50.559 with OpenAI and Anthropic directly. New developer 00:08:50.559 --> 00:08:52.940 tools are rapidly emerging to manage this digital 00:08:52.940 --> 00:08:55.980 chaos. We have to look closely at a tracking 00:08:55.980 --> 00:08:59.309 tool called WayDev. It tracks agent -generated 00:08:59.309 --> 00:09:01.870 code from the IDE straight to production. Right. 00:09:01.990 --> 00:09:04.490 Agent -generated code is software written entirely 00:09:04.490 --> 00:09:08.289 by AI systems. An IDE is basically the digital 00:09:08.289 --> 00:09:11.169 workspace where human coders type. Waitif tracks 00:09:11.169 --> 00:09:13.730 exactly which specific AI agent wrote the final 00:09:13.730 --> 00:09:16.360 code. It thoroughly monitors the total tokens 00:09:16.360 --> 00:09:18.919 consumed during the entire process. Tokens are 00:09:18.919 --> 00:09:20.840 basically the small chunks of words AI reads. 00:09:21.080 --> 00:09:23.500 Wative tracks every single token to calculate 00:09:23.500 --> 00:09:25.899 exact financial costs. It calculates the specific 00:09:25.899 --> 00:09:28.679 cost per individual pull request. A pull request 00:09:28.679 --> 00:09:30.460 is just officially proposing a new code change. 00:09:30.799 --> 00:09:33.960 It also perfectly tracks overall acceptance rates 00:09:33.960 --> 00:09:37.539 and live deployment status. We're seeing a massive 00:09:37.539 --> 00:09:40.220 explosion of these specialized ecosystem tracking 00:09:40.220 --> 00:09:42.879 tools. Look at Gemini Notebooks competing heavily 00:09:42.879 --> 00:09:44.789 with Notebook LM. right now yeah that's a big 00:09:44.789 --> 00:09:47.389 one their new sync loop workflow is a total game 00:09:47.389 --> 00:09:50.250 changer today it perfectly synchronizes your 00:09:50.250 --> 00:09:53.090 complex research across multiple different digital 00:09:53.090 --> 00:09:56.830 platforms we're also seeing fascinating new physical 00:09:56.830 --> 00:09:59.809 hardware integrations emerging right now the 00:09:59.809 --> 00:10:03.190 dune context aware mac keypad is a perfect physical 00:10:03.190 --> 00:10:05.730 example it physically automates your workflows 00:10:05.730 --> 00:10:08.950 and your complex digital meetings wow it automatically 00:10:08.950 --> 00:10:11.370 changes its physical button behavior based on 00:10:11.370 --> 00:10:13.929 your foreground app There's also the fascinating 00:10:13.929 --> 00:10:17.889 new Claude Desktop Buddy project online. It exposes 00:10:17.889 --> 00:10:20.289 a lightweight API directly from the Claude Desktop 00:10:20.289 --> 00:10:22.929 app. This allows you to connect digital workloads 00:10:22.929 --> 00:10:25.370 to physical microcontrollers. Right. You can 00:10:25.370 --> 00:10:27.389 physically bridge the gap between digital code 00:10:27.389 --> 00:10:30.779 and physical reality. But we must violently pivot 00:10:30.779 --> 00:10:33.639 to the macro security risks here. Governments 00:10:33.639 --> 00:10:36.580 are desperately racing to secure their AI leadership 00:10:36.580 --> 00:10:40.279 globally. The UK just launched a massive sovereign 00:10:40.279 --> 00:10:44.860 AI fund recently. It's a 500 million pound domestic 00:10:44.860 --> 00:10:47.120 investment initiative. They're offering massive 00:10:47.120 --> 00:10:50.139 capital, supercomputers and rapid visas for international 00:10:50.139 --> 00:10:53.039 startups. Governments are clearly recognizing 00:10:53.039 --> 00:10:55.919 this as a critical national priority. But we 00:10:55.919 --> 00:10:58.919 have to firmly contrast that with... security 00:10:58.919 --> 00:11:01.899 realities. Yeah, we do. A clever cheat malware 00:11:01.899 --> 00:11:05.440 recently led to a massive Vercel breach. Vercel 00:11:05.440 --> 00:11:07.679 is a highly popular platform for deploying web 00:11:07.679 --> 00:11:10.620 applications. Hackers use this malware to completely 00:11:10.620 --> 00:11:13.159 infiltrate an active workspace. Once inside, 00:11:13.360 --> 00:11:15.399 they didn't just steal a few static text files, 00:11:15.519 --> 00:11:17.759 they gained full access to the underlying automated 00:11:17.759 --> 00:11:20.860 deployment systems. It completely exposed highly 00:11:20.860 --> 00:11:23.059 sensitive Vercel workspace data to malicious 00:11:23.059 --> 00:11:25.759 hackers. The hacker group is now actively demanding 00:11:25.759 --> 00:11:29.019 a $2 million ransom. It's a massive, glaring 00:11:29.019 --> 00:11:31.120 reminder of the hidden dangers lurking here. 00:11:31.279 --> 00:11:34.740 Giving AI agents full workspace access creates 00:11:34.740 --> 00:11:38.500 incredibly serious institutional risks. of AI 00:11:38.500 --> 00:11:41.539 agents total workspace access to automate coding? 00:11:41.740 --> 00:11:44.460 Are we just automating our own security breaches? 00:11:44.700 --> 00:11:47.759 Speed and deep integration currently vastly outpace 00:11:47.759 --> 00:11:50.860 basic security protocols. We are letting AI bypass 00:11:50.860 --> 00:11:53.460 traditional human review processes entirely. 00:11:53.879 --> 00:11:56.539 The automated systems deploy the generated code 00:11:56.539 --> 00:11:58.960 without any friction. Right. We are definitely 00:11:58.960 --> 00:12:01.940 leaving massive vulnerabilities open for global 00:12:01.940 --> 00:12:04.899 exploitation. Deep workspace integration creates 00:12:04.899 --> 00:12:19.200 incredible speed, but opens terrifying Yeah. 00:12:34.139 --> 00:12:35.879 This is a genuinely fascinating breakthrough 00:12:35.879 --> 00:12:38.519 in digital mapping technology. Traditional 3D 00:12:38.519 --> 00:12:41.000 mapping has always had a massive computing bottleneck 00:12:41.000 --> 00:12:43.919 historically. It usually requires feeding a computer 00:12:43.919 --> 00:12:46.840 a massive mountain of digital photos. You have 00:12:46.840 --> 00:12:48.440 to wait until you're completely done recording 00:12:48.440 --> 00:12:51.059 everything physically. Right. Then the computer 00:12:51.059 --> 00:12:54.080 slowly processes that massive visual data into 00:12:54.080 --> 00:12:56.759 a map. That offline processing takes an enormous 00:12:56.759 --> 00:12:59.220 amount of computational time. You can't actually 00:12:59.220 --> 00:13:02.299 see the digital map while you are walking. But 00:13:02.299 --> 00:13:05.299 Lingbot Map... It's a completely new open source 00:13:05.299 --> 00:13:08.700 sees you go visual model. It completely deletes 00:13:08.700 --> 00:13:10.759 that frustrating digital waiting period entirely. 00:13:11.139 --> 00:13:14.159 It's a fully streaming 3D reconstruction foundation 00:13:14.159 --> 00:13:17.639 model. A foundation model is a massive adaptable 00:13:17.639 --> 00:13:21.360 AI trained on vast data. Right. And this specific 00:13:21.360 --> 00:13:24.159 model builds a digital world in real time. It 00:13:24.159 --> 00:13:26.840 processes the complex visual data frame by frame 00:13:26.840 --> 00:13:29.379 as you physically move. Traditional mapping heavily 00:13:29.379 --> 00:13:31.639 relies on incredibly expensive hardware like 00:13:31.639 --> 00:13:34.720 LiDAR sensors. LiDAR shoots precise lasers to 00:13:34.720 --> 00:13:37.220 measure physical distance in a room. But LingbotMap 00:13:37.220 --> 00:13:39.679 achieves the exact same precise result using 00:13:39.679 --> 00:13:42.600 pure software. It analyzes the changing pixels 00:13:42.600 --> 00:13:45.080 from a standard $10 phone camera. The technical 00:13:45.080 --> 00:13:47.659 performance specs on this are genuinely deeply 00:13:47.659 --> 00:13:50.759 impressive. It maintains... a rock -solid 20 00:13:50.759 --> 00:13:53.559 frames per second processing rate it perfectly 00:13:53.559 --> 00:13:56.039 holds this even over marathon sequences of 10 00:13:56.039 --> 00:13:59.399 ,000 frames Wow most importantly it completely 00:13:59.399 --> 00:14:03.200 solves the infamous digital drift problem Drift 00:14:03.200 --> 00:14:06.019 is a notoriously huge issue in traditional digital 00:14:06.019 --> 00:14:09.139 mapping. Imagine walking completely blindfolded 00:14:09.139 --> 00:14:11.600 and trying to map a room in your head. Eventually, 00:14:11.639 --> 00:14:14.460 your internal mental map drifts completely away 00:14:14.460 --> 00:14:17.820 from reality. Early AI vision models suffered 00:14:17.820 --> 00:14:20.200 from this exact same navigational confusion. 00:14:20.620 --> 00:14:22.960 Let's clearly explain that specific tracking 00:14:22.960 --> 00:14:25.720 metric for a moment. Absolute trajectory error 00:14:25.720 --> 00:14:28.279 measures how far digital maps drift from reality. 00:14:28.649 --> 00:14:31.149 Lingbot map drastically cuts this physical error 00:14:31.149 --> 00:14:34.730 by 57 to 75%. That's compared to all the previous 00:14:34.730 --> 00:14:36.529 streaming methods currently available globally. 00:14:37.289 --> 00:14:39.929 On the brutally complex Oxford Spires dataset, 00:14:40.190 --> 00:14:42.750 it performed absolutely beautifully. It hit a 00:14:42.750 --> 00:14:46.470 genuinely staggering 6 .42 meter overall tracking 00:14:46.470 --> 00:14:49.009 error rate. That actually beats high -end offline 00:14:49.009 --> 00:14:51.830 models that see the whole video. And it's officially 00:14:51.830 --> 00:14:54.830 released entirely under the Apache 2 .0 open 00:14:54.830 --> 00:14:57.429 license. That means it is highly open source 00:14:57.429 --> 00:15:00.409 and free for anyone. Anyone can pull the public 00:15:00.409 --> 00:15:03.950 repo and run demo .pi right now. You can easily 00:15:03.950 --> 00:15:06.720 build a 3D viewer in your browser today. If an 00:15:06.720 --> 00:15:08.860 open -source model can map the world flawlessly 00:15:08.860 --> 00:15:11.799 with a $10 camera, what happens to the multi 00:15:11.799 --> 00:15:14.460 -billion -dollar LiDAR and high -end sensor industry? 00:15:14.809 --> 00:15:17.110 Expensive physical sensors will likely become 00:15:17.110 --> 00:15:20.350 highly niche for hypercritical tasks. Mass market 00:15:20.350 --> 00:15:22.809 augmented reality and delivery drones will totally 00:15:22.809 --> 00:15:25.090 rely on cheap vision models. The intelligent 00:15:25.090 --> 00:15:27.450 software will simply replace the complex hardware 00:15:27.450 --> 00:15:30.129 components entirely. Expensive physical sensors 00:15:30.129 --> 00:15:33.389 become obsolete as intelligent software extracts 00:15:33.389 --> 00:15:35.710 perfect data from cheap lenses. It's a truly 00:15:35.710 --> 00:15:38.590 profound shift in how machines understand space 00:15:38.590 --> 00:15:41.350 natively. We've spent years building complex 00:15:41.350 --> 00:15:43.610 robots that have to stop and think. They normally 00:15:43.610 --> 00:15:45.490 rely on... entirely on pre -mapped environments 00:15:45.490 --> 00:15:48.210 to function properly safely. Right. Lingbot map 00:15:48.210 --> 00:15:51.090 fundamentally turns mapping and exploring into 00:15:51.090 --> 00:15:53.669 a single fluid motion. We're reaching the end 00:15:53.669 --> 00:15:56.009 of our deep dive for today. Let's synthesize 00:15:56.009 --> 00:15:58.210 the massive big ideas we've uncovered together 00:15:58.210 --> 00:16:01.970 here. We are watching AI aggressively evolve 00:16:01.970 --> 00:16:05.970 from a stubborn text generator. It's moving from 00:16:05.970 --> 00:16:09.009 hallucinating digital files into an engine of 00:16:09.009 --> 00:16:11.799 spatial reasoning. The technological shift is 00:16:11.799 --> 00:16:14.519 happening across multiple different complex digital 00:16:14.519 --> 00:16:17.580 fronts. It's fundamentally changing how machines 00:16:17.580 --> 00:16:19.820 perceive and interact with physical reality. 00:16:20.100 --> 00:16:23.919 We see OpenAI's Spud recreating sitcom apartments 00:16:23.919 --> 00:16:27.279 with absolutely perfect physics. We see Lingbot 00:16:27.279 --> 00:16:29.759 Map turning environmental mapping into a single 00:16:29.759 --> 00:16:32.639 fluid motion. This directly powers autonomous 00:16:32.639 --> 00:16:35.700 delivery drones and lightweight AR glasses seamlessly. 00:16:36.299 --> 00:16:38.740 The strict boundary between physical reality 00:16:38.740 --> 00:16:41.220 and digital generation is collapsing. This incredible 00:16:41.220 --> 00:16:43.799 evolution is being fueled by astronomical global 00:16:43.799 --> 00:16:46.299 capital investments. Anthropic's $100 billion 00:16:46.299 --> 00:16:49.419 compute plan is a truly perfect example. But 00:16:49.419 --> 00:16:51.879 we also see malicious hackers aggressively exploiting 00:16:51.879 --> 00:16:54.559 this new ecosystem. The invisible digital risks 00:16:54.559 --> 00:16:56.419 are growing just as fast as the capabilities. 00:16:56.919 --> 00:16:59.980 If our AR glasses can perfectly anchor virtual 00:16:59.980 --> 00:17:02.299 furniture in our real -world living rooms in 00:17:02.299 --> 00:17:05.579 real time, and AI can recreate spaces with perfect... 00:17:12.999 --> 00:17:15.059 Thank you so much for joining us on this deep 00:17:15.059 --> 00:17:17.019 dive. It's always a genuine pleasure to explore 00:17:17.019 --> 00:17:19.420 these fascinating frontiers together. We'll catch 00:17:19.420 --> 00:17:21.099 you next time as the future keeps unfolding. 00:17:21.319 --> 00:17:22.140 Out to your music.