WEBVTT 00:00:00.000 --> 00:00:03.120 Imagine your next flight. Except the pilot isn't 00:00:03.120 --> 00:00:05.820 just a person. It's also a Silicon Valley chip 00:00:05.820 --> 00:00:08.320 making these split -second decisions. We're talking 00:00:08.320 --> 00:00:10.859 about incredibly specialized, mission -critical 00:00:10.859 --> 00:00:14.279 AI moving into the skies. But at the same time, 00:00:14.320 --> 00:00:17.539 we're seeing new research that's exposing these 00:00:17.539 --> 00:00:22.839 surprising, almost ghost -like memory flaws in 00:00:22.839 --> 00:00:25.500 the very same kind of AI models that, well, that 00:00:25.500 --> 00:00:27.379 run our daily lives. And welcome back to The 00:00:27.379 --> 00:00:30.500 Deep Dive. Today's mission is really about synthesizing 00:00:30.500 --> 00:00:32.619 a stack of recent sources that highlight that 00:00:32.619 --> 00:00:35.420 exact tension. You know, these explosive technical 00:00:35.420 --> 00:00:37.560 breakthroughs happening right alongside some 00:00:37.560 --> 00:00:39.600 really fundamental safety challenges. We're charting, 00:00:39.600 --> 00:00:40.960 of course, through the most important shifts 00:00:40.960 --> 00:00:42.899 in the industry right now. Yeah, we've got three 00:00:42.899 --> 00:00:45.039 core segments for you today. First up, we're 00:00:45.039 --> 00:00:47.439 dedicating serious time to that engineering breakthrough 00:00:47.439 --> 00:00:50.500 in the skies, the huge Archer and NVIDIA collaboration 00:00:50.500 --> 00:00:53.039 on autonomous air taxis. And what that ultra 00:00:53.039 --> 00:00:55.479 -low latency compute actually means for aviation 00:00:55.479 --> 00:00:58.969 safety. Exactly. Then second, we'll move from 00:00:58.969 --> 00:01:01.090 the skies down to the keyboard. We're going to 00:01:01.090 --> 00:01:04.069 cover the changing tools landscape. So why prompt 00:01:04.069 --> 00:01:07.049 engineering is evolving, the rise of these specialized 00:01:07.049 --> 00:01:09.329 productivity agents, and some really critical 00:01:09.329 --> 00:01:12.629 geopolitical shifts in the hardware war. And 00:01:12.629 --> 00:01:14.930 finally, we're doing a deep dive into a crucial 00:01:14.930 --> 00:01:18.640 safety alert. A Stanford study just proved leading 00:01:18.640 --> 00:01:22.120 LLMs are memorizing entire books verbatim and 00:01:22.120 --> 00:01:25.000 how easily you can just bypass their safety filters. 00:01:25.219 --> 00:01:27.000 Okay, let's unpack this. Let's start with that 00:01:27.000 --> 00:01:30.200 massive engineering challenge up high. So, Archer 00:01:30.200 --> 00:01:33.799 Aviation, they made huge headlines at CES 2026. 00:01:34.140 --> 00:01:36.260 They announced they're integrating NVIDIA's new 00:01:36.260 --> 00:01:39.280 IGX -4 platform into their next -gen air taxi. 00:01:39.640 --> 00:01:41.799 And this integration, I mean, it's a huge jump. 00:01:41.920 --> 00:01:44.560 We've seen IGX -4 used in places like hospitals 00:01:44.560 --> 00:01:46.900 for complex surgical automation and in these 00:01:46.900 --> 00:01:48.840 high -precision factories. Right, very regulated 00:01:48.840 --> 00:01:51.180 environment. Very. But moving into passenger 00:01:51.180 --> 00:01:53.359 aviation, that's a completely different level 00:01:53.359 --> 00:01:55.480 of regulatory and safety challenge. It just demands 00:01:55.480 --> 00:01:58.000 mission -critical, instantaneous decision -making. 00:01:58.430 --> 00:02:00.329 every single second of the flight. And that's 00:02:00.329 --> 00:02:03.390 why the scale of this is so telling. Archer isn't 00:02:03.390 --> 00:02:05.329 just running a few demos in a hangar somewhere. 00:02:05.549 --> 00:02:07.950 They have their own dedicated airport, which 00:02:07.950 --> 00:02:12.490 is basically becoming ground zero for real world 00:02:12.490 --> 00:02:15.030 AI aviation testing. They're building this whole 00:02:15.030 --> 00:02:17.469 autonomous ecosystem from the ground up. So if 00:02:17.469 --> 00:02:19.889 SOAR is the AI brain. What's the nervous system? 00:02:20.189 --> 00:02:22.650 You synthesize this whole complex integration 00:02:22.650 --> 00:02:25.069 into three foundational pillars for us. I did. 00:02:25.169 --> 00:02:28.030 So the first one is all about pilot safety and 00:02:28.030 --> 00:02:31.009 predictive awareness. This system is constantly 00:02:31.009 --> 00:02:33.469 running simulations in the background, providing 00:02:33.469 --> 00:02:36.650 real time alerts, smart flight suggestions, all 00:02:36.650 --> 00:02:38.969 to improve human decision making. So it's like 00:02:38.969 --> 00:02:41.830 an always on copilot. Exactly. It's seeing things 00:02:41.830 --> 00:02:43.930 the human pilot might miss, especially in these 00:02:43.930 --> 00:02:46.069 really high density air spaces. That makes sense. 00:02:46.129 --> 00:02:48.710 I mean, mitigating human error is key. But how 00:02:48.710 --> 00:02:51.229 does this level of AI integration change the 00:02:51.229 --> 00:02:53.810 pilot training requirements? Are we augmenting 00:02:53.810 --> 00:02:56.750 human skills here or are we on a path to eventually 00:02:56.750 --> 00:02:59.449 replace certain skill sets entirely? For now, 00:02:59.449 --> 00:03:02.569 it's an augmentation, but a necessary one. Think 00:03:02.569 --> 00:03:04.710 about the speed you need for real -time sensor 00:03:04.710 --> 00:03:07.430 fusion. The plane's taking in LIDAR data, radar, 00:03:07.590 --> 00:03:10.090 internal diagnostics, external weather. All at 00:03:10.090 --> 00:03:13.030 once. All at once. And you need nanosecond -level 00:03:13.030 --> 00:03:15.210 processing to make predictions from all that. 00:03:15.310 --> 00:03:18.030 That ultra -low latency is the entire point. 00:03:18.250 --> 00:03:21.669 If the plane detects, say, a wind shear or an 00:03:21.669 --> 00:03:25.509 unknown drone, it needs to plot a new, safe trajectory 00:03:25.509 --> 00:03:27.930 instantly. That really puts the safety component 00:03:27.930 --> 00:03:31.129 into sharp focus. Okay, what about Pillar 2? 00:03:31.500 --> 00:03:33.479 The second is seamless airspace integration. 00:03:34.080 --> 00:03:36.500 This is just critically important because these 00:03:36.500 --> 00:03:39.000 new air taxis have to coexist with the old world 00:03:39.000 --> 00:03:42.199 of aviation, right? So the AI handles all the 00:03:42.199 --> 00:03:44.719 dynamic traffic -aware flight routing. It makes 00:03:44.719 --> 00:03:46.659 sure it plays nicely with all the legacy air 00:03:46.659 --> 00:03:48.800 traffic control systems we already have. It's 00:03:48.800 --> 00:03:50.960 basically the translation layer between future 00:03:50.960 --> 00:03:53.259 tech and current regulation. And the third pillar 00:03:53.259 --> 00:03:54.759 is the one that really sets the stage for the 00:03:54.759 --> 00:03:57.860 future, right? Exactly. The third is autonomy 00:03:57.860 --> 00:04:00.020 -ready controls. This entire integration, it's 00:04:00.020 --> 00:04:02.379 all about building the core compute layer that's 00:04:02.379 --> 00:04:05.900 necessary for... Future semi -autonomous or eventually 00:04:05.900 --> 00:04:09.099 fully pilotless systems pairing Thor's compute 00:04:09.099 --> 00:04:12.280 with Archer's avionics. That's the digital backbone 00:04:12.280 --> 00:04:14.860 for the ultimate vision. The stakes are just 00:04:14.860 --> 00:04:17.540 impossibly high. When we talk about safety critical 00:04:17.540 --> 00:04:19.699 computing at that level of scale, pushing what, 00:04:19.779 --> 00:04:22.220 a billion queries across an entire air traffic 00:04:22.220 --> 00:04:25.110 system? You have to ask. Are the benchmarks for 00:04:25.110 --> 00:04:27.170 testing even fully developed yet? It really is 00:04:27.170 --> 00:04:29.149 a moment of wonder at the engineering complexity. 00:04:29.350 --> 00:04:31.769 It forces us to ask tough questions about trust. 00:04:31.850 --> 00:04:34.189 Beyond the air taxi itself, what's the single 00:04:34.189 --> 00:04:36.389 biggest challenge this low latency compute solves 00:04:36.389 --> 00:04:39.430 for future air travel? It solves the hard problem 00:04:39.430 --> 00:04:42.129 of fitting autonomous routing into our old existing 00:04:42.129 --> 00:04:44.410 air traffic rules. And that high stakes focus 00:04:44.410 --> 00:04:47.589 on safety in the skies. It's such a sharp contrast 00:04:47.589 --> 00:04:49.930 to how casually we often use AI in our daily 00:04:49.930 --> 00:04:53.339 lives. And speaking of daily use, let's move 00:04:53.339 --> 00:04:55.259 from the cockpit back down to the desk because 00:04:55.259 --> 00:04:58.199 how we even talk to AI is changing fast. Right. 00:04:58.279 --> 00:05:02.100 Our sources are indicating that the finicky art 00:05:02.100 --> 00:05:04.860 of... crafting the perfect instruction set, what 00:05:04.860 --> 00:05:07.319 we call prompt engineering, is now facing some 00:05:07.319 --> 00:05:09.879 pretty rapid disruption. That's right. And prompt 00:05:09.879 --> 00:05:12.019 engineering, to put it simply, is just the skill 00:05:12.019 --> 00:05:14.600 of writing precise instructions to get the exact 00:05:14.600 --> 00:05:17.240 result you want from an AI model. For a long 00:05:17.240 --> 00:05:19.319 time, it really felt like black magic. And that 00:05:19.319 --> 00:05:22.379 magic is changing. Anthropic has released a new 00:05:22.379 --> 00:05:26.290 structural trick. using XML tags for much better 00:05:26.290 --> 00:05:28.550 control over their models. And we're seeing reports 00:05:28.550 --> 00:05:31.670 that this dramatically outperforms the older, 00:05:31.769 --> 00:05:34.810 messier methods. It's a huge technical insight. 00:05:35.050 --> 00:05:37.350 The older methods, like context dumps, are basically 00:05:37.350 --> 00:05:39.629 just pasting thousands of words of unstructured 00:05:39.629 --> 00:05:41.589 text into the prompt and just hoping the model 00:05:41.589 --> 00:05:44.870 figures it out. XML tags standardize the input 00:05:44.870 --> 00:05:47.029 for the model. It makes its attention mechanism 00:05:47.029 --> 00:05:49.410 way more efficient, less likely to get lost in 00:05:49.410 --> 00:05:51.750 the noise. It's like stacking Lego blocks of 00:05:51.750 --> 00:05:53.980 data in a structured way. instead of just throwing 00:05:53.980 --> 00:05:56.100 a pile at the model? I still wrestle with prompt 00:05:56.100 --> 00:05:58.600 drift myself. You know, models start ignoring 00:05:58.600 --> 00:06:00.920 my specific instructions after a few conversational 00:06:00.920 --> 00:06:04.500 turns. So seeing new structural techniques like 00:06:04.500 --> 00:06:08.040 XML tags is a massive relief. But does the shift 00:06:08.040 --> 00:06:12.180 toward these structural inputs mean that large 00:06:12.180 --> 00:06:14.839 enterprise -grade AI is becoming fundamentally 00:06:14.839 --> 00:06:17.860 less accessible to the casual user who doesn't 00:06:17.860 --> 00:06:19.860 want to learn a markup language? That's a key 00:06:19.860 --> 00:06:22.980 tension. But ironically, that complexity is also 00:06:22.980 --> 00:06:25.879 driving the creation of tools that simplify everything 00:06:25.879 --> 00:06:29.519 else. And that simplification is fueling a real 00:06:29.519 --> 00:06:32.839 emergence of practical, non -technical AI usage. 00:06:33.079 --> 00:06:35.540 Less Python, more delegation. Exactly. We've 00:06:35.540 --> 00:06:37.560 seen these incredible guides surfacing. Things 00:06:37.560 --> 00:06:39.860 like building automated workflows to stop you 00:06:39.860 --> 00:06:42.959 from chasing messy data or clearly defining the 00:06:42.959 --> 00:06:44.980 four AI agents a non -technical person needs 00:06:44.980 --> 00:06:47.160 to delegate. Like 90 % of their routine tasks. 00:06:47.459 --> 00:06:49.430 And we have the proof of concept. Peter Yang's 00:06:49.430 --> 00:06:52.449 47 -minute Claude Co demo, it showed a non -developer 00:06:52.449 --> 00:06:54.930 using that tool to genuinely run her entire life. 00:06:55.350 --> 00:06:58.149 Scheduling, managing finances, AI is moving way 00:06:58.149 --> 00:07:00.709 beyond just generating marketing copy. It's becoming 00:07:00.709 --> 00:07:03.029 an integrated workflow manager. The specialized 00:07:03.029 --> 00:07:05.410 experiences are also key for broad adoption. 00:07:05.949 --> 00:07:08.629 We've got new dedicated wellness support with 00:07:08.629 --> 00:07:11.209 ChatGPT Health, and Google Classroom can now 00:07:11.209 --> 00:07:14.449 turn any lesson into a specialized podcast. You 00:07:14.449 --> 00:07:17.230 can specify the topic. the speakers, the style. 00:07:17.470 --> 00:07:20.800 These targeted tools are immediate wins. So we've 00:07:20.800 --> 00:07:22.879 discussed how the user experience is changing, 00:07:23.040 --> 00:07:25.720 but that experience is shaped by these massive 00:07:25.720 --> 00:07:28.339 forces behind the scenes. Let's talk about the 00:07:28.339 --> 00:07:31.019 geopolitical hardware wars and some crucial safety 00:07:31.019 --> 00:07:33.459 issues defining the industry's foundation. Switching 00:07:33.459 --> 00:07:36.000 to geopolitics, the ground is definitely shifting 00:07:36.000 --> 00:07:38.459 beneath the hardware market. It absolutely is. 00:07:38.620 --> 00:07:41.019 China is reportedly asking its domestic tech 00:07:41.019 --> 00:07:44.240 firms to pause orders for the powerful NVIDIA 00:07:44.240 --> 00:07:47.000 H200 chips. And the strategic goal is pretty 00:07:47.000 --> 00:07:49.300 clear. They want to steer buyers toward domestic 00:07:49.300 --> 00:07:52.220 AI chip alternatives to build self -sufficiency. 00:07:52.480 --> 00:07:55.439 This move has huge global supply chain implications. 00:07:55.939 --> 00:07:58.660 And it's important context here that each H -200 00:07:58.660 --> 00:08:00.800 export still requires U .S. government approval. 00:08:00.959 --> 00:08:03.339 And there's no set timeline for that complex 00:08:03.339 --> 00:08:05.620 process. So it keeps pressure on both sides. 00:08:05.779 --> 00:08:07.980 We're also seeing consolidation in the talent 00:08:07.980 --> 00:08:11.959 war. Open AI just acquired the Convogo team. 00:08:12.060 --> 00:08:14.300 That's its ninth acquisition this year. This 00:08:14.300 --> 00:08:16.560 team used to help human coaches scale their work, 00:08:16.680 --> 00:08:18.560 but now they're shifting their focus entirely 00:08:18.560 --> 00:08:22.860 to building AI cloud tools for core infrastructure. 00:08:23.220 --> 00:08:25.439 So the top talent is being pulled into the foundational 00:08:25.439 --> 00:08:28.500 model architecture. Yep. And that rapid consolidation 00:08:28.500 --> 00:08:31.240 and technological advance brings us right back 00:08:31.240 --> 00:08:33.679 to safety. We have to address the crucial issue 00:08:33.679 --> 00:08:36.679 around content moderation. Right. Reports have 00:08:36.679 --> 00:08:39.879 shown that X is seeing a staggering volume, something 00:08:39.879 --> 00:08:44.360 like 6 ,700 or more AI -generated illegal images 00:08:44.360 --> 00:08:47.379 per hour, specifically attributed to the Grok 00:08:47.379 --> 00:08:50.519 platform. That number. It's staggering. It just 00:08:50.519 --> 00:08:52.639 demonstrates how current moderation systems, 00:08:52.759 --> 00:08:55.500 even with advanced AI, just cannot keep pace 00:08:55.500 --> 00:08:57.580 with generative output. No chance. And if the 00:08:57.580 --> 00:08:59.840 global nature of the Internet stalls legal action 00:08:59.840 --> 00:09:02.259 because of different jurisdictions and slow regulatory 00:09:02.259 --> 00:09:04.759 response, we're left in a really difficult spot. 00:09:05.159 --> 00:09:06.779 And that's where the pressure is focused right 00:09:06.779 --> 00:09:09.659 now. Global regulators are pressing XAI over 00:09:09.659 --> 00:09:12.139 this issue, but legal limitations in different 00:09:12.139 --> 00:09:14.840 countries are stalling any truly effective unified 00:09:14.840 --> 00:09:17.480 action. It just highlights the difficulty in 00:09:17.480 --> 00:09:20.539 regulating real time, high volume content generation 00:09:20.539 --> 00:09:23.159 globally. So given all these regulatory challenges, 00:09:23.419 --> 00:09:25.320 geopolitical shifts and application changes, 00:09:25.539 --> 00:09:27.799 which trend tells us more about the immediate 00:09:27.799 --> 00:09:31.350 future of AI use? The shift toward dedicated, 00:09:31.590 --> 00:09:34.590 specialized agents. Things like the new Google 00:09:34.590 --> 00:09:37.769 AI inbox. It shows immediate consumer integration 00:09:37.769 --> 00:09:41.009 and a desire to delegate specific small tasks 00:09:41.009 --> 00:09:44.450 rather than rely on one big generalist LLM for 00:09:44.450 --> 00:09:46.809 everything. That's a powerful sign that the specialization 00:09:46.809 --> 00:09:49.970 era is upon us. Okay, now for the most concerning 00:09:49.970 --> 00:09:52.190 news in our sources and perhaps the biggest challenge 00:09:52.190 --> 00:09:54.789 to the industry's current legal defense, a major 00:09:54.789 --> 00:09:57.370 breakthrough on LLM vulnerabilities. Yeah, this 00:09:57.370 --> 00:09:59.570 is a profound finding from a new Stanford study, 00:09:59.669 --> 00:10:01.950 and it directly challenges the industry consensus 00:10:01.950 --> 00:10:05.210 on filtering and data handling. The core revelation. 00:10:05.710 --> 00:10:08.250 This study proved that production -grade LLMs, 00:10:08.289 --> 00:10:09.950 the ones people are paying for and relying on 00:10:09.950 --> 00:10:13.110 right now, still memorize and leak near -exact 00:10:13.110 --> 00:10:16.149 copyrighted book text. And they tested every 00:10:16.149 --> 00:10:19.409 major player. Claude, GPT, Grok, and Gemini. 00:10:19.490 --> 00:10:21.789 And the specific data is just startling because 00:10:21.789 --> 00:10:24.250 the recall rate is so high and so consistent. 00:10:24.450 --> 00:10:27.750 Claude 3 .7 Sonnet, for example, it hit a 95 00:10:27.750 --> 00:10:31.610 .8 % text extraction recall rate on certain books. 00:10:31.710 --> 00:10:34.590 That's virtually perfect, consistent memorization. 00:10:35.009 --> 00:10:37.029 What makes this study so concerning isn't just 00:10:37.029 --> 00:10:39.230 the leakage itself, but how easily the model's 00:10:39.230 --> 00:10:41.950 internal filtering systems, the supposed guardrails, 00:10:41.970 --> 00:10:44.929 were just... bypass. The technique they use to 00:10:44.929 --> 00:10:47.210 break the safety layers is deceptively basic. 00:10:47.429 --> 00:10:49.570 It's like a digital shoulder tap. It's basically 00:10:49.570 --> 00:10:52.639 the three step process. One. Give the model the 00:10:52.639 --> 00:10:55.240 opening line of a copyrighted book. Two, ask 00:10:55.240 --> 00:10:57.700 it to continue the text. If it initially refuses, 00:10:57.899 --> 00:10:59.360 which the guardrails are designed to make it 00:10:59.360 --> 00:11:01.639 do, you just reword the prompt very slightly 00:11:01.639 --> 00:11:03.720 until it complies. And that's it. That's it. 00:11:03.779 --> 00:11:06.200 And three, the model then often just delivers 00:11:06.200 --> 00:11:08.799 high -quality verbatim text. So the result is 00:11:08.799 --> 00:11:11.519 consistent, high -quality memorization across 00:11:11.519 --> 00:11:14.019 multiple books and all four of the major production 00:11:14.019 --> 00:11:16.659 models. Correct. And this suggests that the safety 00:11:16.659 --> 00:11:19.500 layers are not true, hard constraints. They're 00:11:19.500 --> 00:11:22.419 merely soft suggestions. The underlying data 00:11:22.419 --> 00:11:24.940 is just stored perfectly intact, waiting for 00:11:24.940 --> 00:11:27.940 the right prompt format to unlock it. If these 00:11:27.940 --> 00:11:30.700 models leak, ExactBook texts this consistently. 00:11:32.200 --> 00:11:34.820 It totally changes the legal calculus. It makes 00:11:34.820 --> 00:11:37.360 arguments about fair use training. The idea that 00:11:37.360 --> 00:11:39.940 models are only absorbing general patterns, much 00:11:39.940 --> 00:11:42.139 harder for companies to defend in court. Yeah, 00:11:42.200 --> 00:11:43.860 you can't claim you're summarizing if you can 00:11:43.860 --> 00:11:46.659 spit out entire paragraphs verbatim. Right. And 00:11:46.659 --> 00:11:49.159 filters applied on top clearly don't fix the 00:11:49.159 --> 00:11:50.960 memorization that's buried inside the model's 00:11:50.960 --> 00:11:53.639 weights. So if filters fail this easily, what 00:11:53.639 --> 00:11:55.860 does this finding suggest about trusting AI models 00:11:55.860 --> 00:11:58.480 with proprietary or sensitive corporate data? 00:11:58.919 --> 00:12:01.200 Well if it remembers books, it likely remembers 00:12:01.200 --> 00:12:04.179 sensitive data, which is just a fundamental security 00:12:04.179 --> 00:12:07.019 and legal risk for any company using these tools 00:12:07.019 --> 00:12:09.779 internally. So let's connect these threads. The 00:12:09.779 --> 00:12:12.379 overarching theme here is this fascinating paradox 00:12:12.379 --> 00:12:14.580 that really defines this moment in AI history. 00:12:14.720 --> 00:12:17.120 We're simultaneously building these revolutionary 00:12:17.120 --> 00:12:20.100 safety -critical AI systems, like the architecture 00:12:20.100 --> 00:12:23.320 for air taxes, demanding perfection, while also 00:12:23.320 --> 00:12:26.120 exposing these fundamental profound flaws in 00:12:26.120 --> 00:12:28.700 foundational models around privacy and SASE filters. 00:12:29.080 --> 00:12:31.799 It's the tension between aspiration and reality. 00:12:32.519 --> 00:12:34.240 For the knowledge -seeking listener, we've got 00:12:34.240 --> 00:12:37.279 three key takeaways from today's sources. First, 00:12:37.500 --> 00:12:40.019 the future of mobility requires ultra -low latency, 00:12:40.360 --> 00:12:42.940 mission -critical compute. This isn't optional. 00:12:43.059 --> 00:12:45.059 It's the core safety requirement for systems 00:12:45.059 --> 00:12:47.840 like Archer and Thor. Second, the practical AI 00:12:47.840 --> 00:12:50.320 toolkit is rapidly changing. PROM's engineering 00:12:50.320 --> 00:12:52.500 skills are shifting, being replaced by structural 00:12:52.500 --> 00:12:54.840 inputs like XML, and these specialized agents 00:12:54.840 --> 00:12:56.840 are taking over our daily high -volume delegation 00:12:56.840 --> 00:13:01.990 tasks. And third, LLM memory is a profound, exploitable 00:13:01.990 --> 00:13:05.690 vulnerability. This discovery really undermines 00:13:05.690 --> 00:13:08.070 current legal defenses and challenges the core 00:13:08.070 --> 00:13:11.049 assumptions we all have about AI safety and data 00:13:11.049 --> 00:13:13.789 privacy. We've seen that the filter layers fail 00:13:13.789 --> 00:13:15.870 when they're just slightly challenged, proving 00:13:15.870 --> 00:13:18.710 the data is stored intact. So given this memory 00:13:18.710 --> 00:13:21.309 flaw, the next question is personal. If the model 00:13:21.309 --> 00:13:23.830 holds proprietary corporate data, and the legal 00:13:23.830 --> 00:13:26.070 defense for training is eroding, what immediate 00:13:26.070 --> 00:13:28.309 steps should IT departments take this week to 00:13:28.309 --> 00:13:31.070 audit their internal LLM deployments? Something 00:13:31.070 --> 00:13:33.590 to ponder as you navigate this rapidly changing 00:13:33.590 --> 00:13:35.970 technological landscape. Thank you for joining 00:13:35.970 --> 00:13:37.730 us on The Deep Dive. We'll see you next time.