WEBVTT 00:00:00.000 --> 00:00:02.759 Have you ever asked an AI the same question twice 00:00:02.759 --> 00:00:05.639 with all the settings exactly the same and still 00:00:05.639 --> 00:00:07.679 gotten wildly different answers? Yeah. It's like 00:00:07.679 --> 00:00:09.220 walking into your favorite coffee shop ordering 00:00:09.220 --> 00:00:11.820 the exact same drink, you know? But it tastes 00:00:11.820 --> 00:00:13.580 completely different depending on how busy the 00:00:13.580 --> 00:00:16.039 barista is. Right. That kind of inconsistency, 00:00:16.059 --> 00:00:18.019 well, it can be incredibly frustrating, especially 00:00:18.019 --> 00:00:21.179 when you actually need... reliable outputs. Absolutely. 00:00:21.460 --> 00:00:23.640 That unpredictable quality has been a silent 00:00:23.640 --> 00:00:26.500 bug, really. A pervasive headache for AI models 00:00:26.500 --> 00:00:29.559 for years. It just fundamentally undermined trust 00:00:29.559 --> 00:00:33.820 and scientific rigor. Yeah. But what's truly 00:00:33.820 --> 00:00:36.380 exciting now is that a small, really focused 00:00:36.380 --> 00:00:40.039 team has finally delivered a major fix. And today, 00:00:40.159 --> 00:00:42.100 we're diving deep into what that breakthrough 00:00:42.100 --> 00:00:44.619 means for all of us. Welcome to the Deep Dive. 00:00:45.119 --> 00:00:47.420 Today, we're going to unpack some, well, truly 00:00:47.420 --> 00:00:50.039 incredible breakthroughs and also confront a 00:00:50.039 --> 00:00:52.500 few humbling realities shaping the world of artificial 00:00:52.500 --> 00:00:56.000 intelligence. Our mission here is really to distill 00:00:56.000 --> 00:00:58.020 these insights, giving you a shortcut to being 00:00:58.020 --> 00:01:01.700 well -informed. Yeah, we'll explore that surprising 00:01:01.700 --> 00:01:05.700 fix for an old, persistent AI problem. Then we'll 00:01:05.700 --> 00:01:08.200 shift gears a bit, touch on some exciting new 00:01:08.200 --> 00:01:11.480 tools pushing creative boundaries. Okay. We'll 00:01:11.480 --> 00:01:14.459 also discuss... major industry shifts, the immense 00:01:14.459 --> 00:01:16.859 financial investments pouring into the sector, 00:01:17.000 --> 00:01:19.599 and then look at what AI still can't quite do, 00:01:19.760 --> 00:01:22.780 especially in critical areas like, say, healthcare. 00:01:23.040 --> 00:01:26.579 Right. Get ready for some genuine aha moments, 00:01:26.840 --> 00:01:30.120 you know, things that might reframe how you think 00:01:30.120 --> 00:01:32.980 about AI's future. Okay, let's jump right into 00:01:32.980 --> 00:01:35.280 what might be the most significant recent development 00:01:35.280 --> 00:01:39.079 then. Solving the frustrating inconsistency of 00:01:39.079 --> 00:01:41.459 large language models. Yeah, the big one. For 00:01:41.459 --> 00:01:44.069 years. You could give an AI the same prompt, 00:01:44.209 --> 00:01:46.790 often with its temperature setting at zero, which 00:01:46.790 --> 00:01:48.909 means it should be at its most deterministic, 00:01:48.930 --> 00:01:51.310 right? Most predictable. Should be, yeah. And 00:01:51.310 --> 00:01:53.849 you'd still receive varied outputs. It just felt 00:01:53.849 --> 00:01:55.670 like rolling the dice sometimes. It was exactly 00:01:55.670 --> 00:01:57.750 like your coffee analogy. Even when you tell 00:01:57.750 --> 00:01:59.689 the AI, look, don't be creative. Just give me 00:01:59.689 --> 00:02:03.230 the most likely consistent answer. It still drifted. 00:02:03.409 --> 00:02:06.840 Researchers, developers, businesses. They've 00:02:06.840 --> 00:02:08.979 all been scratching their heads. Because if you 00:02:08.979 --> 00:02:13.020 can't reliably reproduce an AI's output, I mean, 00:02:13.020 --> 00:02:15.919 how can you build critical systems on it? Exactly. 00:02:16.039 --> 00:02:18.759 And the core of the problem, as this brilliant 00:02:18.759 --> 00:02:21.620 team recently discovered, wasn't necessarily 00:02:21.620 --> 00:02:24.719 inside the AI model itself. Right, not the model 00:02:24.719 --> 00:02:26.960 logic per se. But in the server environment, 00:02:27.240 --> 00:02:29.759 your output actually changed depending on how 00:02:29.759 --> 00:02:31.560 many other people were hitting the AI server 00:02:31.560 --> 00:02:34.259 at the exact same time. Huh. So it's like the 00:02:34.259 --> 00:02:36.080 system was getting distracted? Kind of, yeah. 00:02:36.259 --> 00:02:39.099 Its internal state subtly altered by all those 00:02:39.099 --> 00:02:41.819 concurrent requests, leading to these non -deterministic 00:02:41.819 --> 00:02:44.280 results. Wow. Okay, the implications of that 00:02:44.280 --> 00:02:46.500 discovery, they're enormous then. Totally. This 00:02:46.500 --> 00:02:49.020 irreproducibility messed with everything from 00:02:49.020 --> 00:02:51.199 scientific research where consistent results 00:02:51.199 --> 00:02:54.979 are just paramount. Absolutely. business reliability 00:02:54.979 --> 00:02:57.460 for AI -powered apps, and even the efficiency 00:02:57.460 --> 00:02:59.879 of training new models. You'd optimize something, 00:03:00.099 --> 00:03:01.960 run a test, get a great result. And then you 00:03:01.960 --> 00:03:04.020 can't get it again. Exactly. You couldn't replicate 00:03:04.020 --> 00:03:06.419 it consistently. Huge problems for any serious 00:03:06.419 --> 00:03:08.280 application. So what's the fix? What do they 00:03:08.280 --> 00:03:11.500 do for this foundational challenge? Okay, so 00:03:11.500 --> 00:03:14.379 this team at Thinking Machines Lab, impressively 00:03:14.379 --> 00:03:18.240 led by ex -OpenAI CTO Mira Marotti, along with 00:03:18.240 --> 00:03:19.840 a meta researcher, developed something truly 00:03:19.840 --> 00:03:23.020 groundbreaking. It's called Batch Invariant Kernels. 00:03:23.099 --> 00:03:25.460 Batch Invariant Kernels. Okay, what does that 00:03:25.460 --> 00:03:28.340 mean in, like, plain English? Simply put, think 00:03:28.340 --> 00:03:31.000 of them as a new computational method. They ensure 00:03:31.000 --> 00:03:33.719 the AI's internal state for your specific query 00:03:33.719 --> 00:03:36.060 stays completely unaffected by what other users 00:03:36.060 --> 00:03:38.840 are doing simultaneously. Imagine having like 00:03:38.840 --> 00:03:41.240 a dedicated, perfectly soundproofed workspace 00:03:41.240 --> 00:03:45.539 just for your task within the AI. It guarantees 00:03:45.539 --> 00:03:48.400 the exact same outcome every single time, no 00:03:48.400 --> 00:03:50.580 matter how busy the main office gets. Got it. 00:03:50.659 --> 00:03:53.139 It's a foundational stability layer. It just 00:03:53.139 --> 00:03:55.979 eliminates a huge source of irreproducibility. 00:03:56.590 --> 00:03:59.189 the LLMs ignore the noise from other prompts 00:03:59.189 --> 00:04:01.270 being processed in the same batch? Honestly, 00:04:01.330 --> 00:04:04.250 what strikes me as truly remarkable here is that 00:04:04.250 --> 00:04:06.469 this kind of fundamental problem, right? Five 00:04:06.469 --> 00:04:10.069 years old, major players like OpenAI, Meta, Google, 00:04:10.250 --> 00:04:14.069 they seemingly hadn't solved it. Yeah, with all 00:04:14.069 --> 00:04:16.569 their resources. And then this relatively small 00:04:16.569 --> 00:04:19.490 team cracked it. That's... Genuinely inspiring, 00:04:19.730 --> 00:04:21.689 I think, for anyone working in tech. It really 00:04:21.689 --> 00:04:23.930 is. It shows that innovation doesn't always need 00:04:23.930 --> 00:04:28.050 an army of engineers. A focused, agile team with 00:04:28.050 --> 00:04:30.490 a fresh perspective can spot a blind spot that 00:04:30.490 --> 00:04:33.389 maybe larger organizations overlook because of 00:04:33.389 --> 00:04:35.370 existing infrastructure or just different priorities. 00:04:35.750 --> 00:04:39.550 Whoa. I mean, just imagine scaling this newfound 00:04:39.550 --> 00:04:43.459 consistency. Think about a billion queries. All 00:04:43.459 --> 00:04:46.620 reliable. A truly reliable AI could transform 00:04:46.620 --> 00:04:49.620 so many industries. This is a huge step for trust 00:04:49.620 --> 00:04:53.360 and broad adoption. This batch invariant kernels 00:04:53.360 --> 00:04:56.180 fix sounds monumental, clearly. But zooming out 00:04:56.180 --> 00:04:59.540 just a bit, why is this specific kind of AI consistency 00:04:59.540 --> 00:05:02.399 so utterly critical for the whole future trajectory 00:05:02.399 --> 00:05:05.620 of AI development? I mean, beyond just lab reproducibility. 00:05:05.800 --> 00:05:08.519 Well, consistent AI builds trust. That's huge. 00:05:08.740 --> 00:05:11.279 It enables scientific rigor and it makes business 00:05:11.279 --> 00:05:13.839 use cases truly dependable. Right. Dependability. 00:05:13.959 --> 00:05:16.259 That's key. So that foundational fix for consistency 00:05:16.259 --> 00:05:18.480 is a game changer, setting the stage for more 00:05:18.480 --> 00:05:21.019 reliable AI everywhere. But even as we tackle 00:05:21.019 --> 00:05:23.660 core stability, AI is already hurtling forward. 00:05:23.959 --> 00:05:26.360 Incredible new capabilities seem to pop up constantly. 00:05:26.800 --> 00:05:28.639 What's caught your eye recently as a sign of 00:05:28.639 --> 00:05:31.810 where AI is heading next? Well, Google VO3 is 00:05:31.810 --> 00:05:34.350 making massive waves right now. It's a text -to 00:05:34.350 --> 00:05:37.149 -video model, but it can turn any image into 00:05:37.149 --> 00:05:40.529 accurate, highly realistic, lip -synced talking 00:05:40.529 --> 00:05:44.350 videos. Any image. Wow. Yeah. Think viral storytelling 00:05:44.350 --> 00:05:47.029 on a whole new level. Imagine taking a static 00:05:47.029 --> 00:05:49.709 photo, maybe a historical figure, and having 00:05:49.709 --> 00:05:52.310 them deliver a compelling speech perfectly lip 00:05:52.310 --> 00:05:54.949 -synced. Okay, that's pretty wild. For content 00:05:54.949 --> 00:05:57.589 creators. It dramatically cuts down production 00:05:57.589 --> 00:05:59.949 time for things like character animation. or 00:05:59.949 --> 00:06:02.750 explainer videos, the fidelity is just astonishing. 00:06:02.930 --> 00:06:05.110 It really blurs the lines between a static image 00:06:05.110 --> 00:06:07.949 and a dynamic narrative. That opens up so many 00:06:07.949 --> 00:06:10.470 creative avenues and maybe some thought -provoking 00:06:10.470 --> 00:06:13.069 ones too. Okay. And on the more practical, maybe 00:06:13.069 --> 00:06:15.629 operational side, OpenAI just released full support 00:06:15.629 --> 00:06:19.069 for MCP and ChatGPT. Right, MCP. Let's unpack 00:06:19.069 --> 00:06:22.069 that. What exactly does multimodal content processing 00:06:22.069 --> 00:06:24.889 let users do beyond just searching for information? 00:06:25.960 --> 00:06:28.939 MCP, multimodal content processing. It fundamentally 00:06:28.939 --> 00:06:32.079 shifts ChatGPT from being just a conversational 00:06:32.079 --> 00:06:35.160 answer machine to more of an action engine. Instead 00:06:35.160 --> 00:06:37.459 of just telling you how to update a ticket in 00:06:37.459 --> 00:06:39.339 your project management software, for example. 00:06:39.560 --> 00:06:42.730 Okay. MCP could actually... trigger that update 00:06:42.730 --> 00:06:45.110 itself or generate a detailed report or even 00:06:45.110 --> 00:06:48.050 initiate an email sequence all based on your 00:06:48.050 --> 00:06:51.129 natural language commands it acts it acts it 00:06:51.129 --> 00:06:54.689 turns spoken or written intent into direct actionable 00:06:54.689 --> 00:06:57.350 steps within other applications linking different 00:06:57.350 --> 00:07:00.129 workflows together it's moving from conversational 00:07:00.129 --> 00:07:04.139 tool to operational one That's a big shift. And 00:07:04.139 --> 00:07:07.000 speaking of OpenAI, big news also for their structure, 00:07:07.199 --> 00:07:10.819 right? A non -binding deal with Microsoft letting 00:07:10.819 --> 00:07:12.980 them restructure as a for -profit company. Yeah, 00:07:13.019 --> 00:07:14.680 that's significant. So what does this all mean 00:07:14.680 --> 00:07:16.920 for their future direction and maybe the broader 00:07:16.920 --> 00:07:19.670 AI landscape? it really points towards a heavy 00:07:19.670 --> 00:07:22.009 push for commercial viability you know maybe 00:07:22.009 --> 00:07:25.050 even an ipo in the not too distant future this 00:07:25.050 --> 00:07:28.089 move signals a significant acceleration towards 00:07:28.089 --> 00:07:31.709 productization market dominance and we're seeing 00:07:31.709 --> 00:07:34.699 similar financial trends elsewhere too Oracle's 00:07:34.699 --> 00:07:37.199 Larry Ellison, for instance, briefly became the 00:07:37.199 --> 00:07:39.660 world's richest man. Right. I saw that. Thanks 00:07:39.660 --> 00:07:42.579 largely to an AI -driven surge in demand for 00:07:42.579 --> 00:07:45.600 cloud infrastructure. Then you've got Kupang, 00:07:45.800 --> 00:07:48.199 the South Korean e -commerce giant, investing 00:07:48.199 --> 00:07:52.019 about $54 million into a fund supporting 14 AI 00:07:52.019 --> 00:07:54.699 startups just in South Korea. Why? The sheer 00:07:54.699 --> 00:07:56.920 amount of money pouring into this space right 00:07:56.920 --> 00:08:00.389 now. It's just staggering. So all this investment 00:08:00.389 --> 00:08:02.470 is clearly a huge vote of confidence. But, you 00:08:02.470 --> 00:08:05.050 know, for a deep dive, are there any potential 00:08:05.050 --> 00:08:07.230 downsides we should think about? Maybe bubbles, 00:08:07.350 --> 00:08:09.750 given this huge capital influx and rapid expansion? 00:08:10.240 --> 00:08:12.279 Well, yeah, rapid growth always carries risks. 00:08:12.360 --> 00:08:14.439 While the opportunities are vast, we are seeing 00:08:14.439 --> 00:08:16.620 some consolidation. And the valuations, well, 00:08:16.699 --> 00:08:18.480 in some cases, they might be running a bit ahead 00:08:18.480 --> 00:08:21.100 of actual revenue generation. Okay. It's definitely 00:08:21.100 --> 00:08:22.759 something to watch closely, especially in the 00:08:22.759 --> 00:08:25.560 startup ecosystem. But overall, AI is certainly 00:08:25.560 --> 00:08:28.519 moving strongly into video, operational automation, 00:08:28.759 --> 00:08:31.839 and becoming a major economic force. Okay, so 00:08:31.839 --> 00:08:33.980 we've covered some big picture shifts, groundbreaking 00:08:33.980 --> 00:08:38.129 fixes. Let's maybe pivot to some rapid fire updates. 00:08:38.289 --> 00:08:40.629 New tools you might want to know about just showing 00:08:40.629 --> 00:08:42.870 how quickly the landscape is evolving. Absolutely. 00:08:42.950 --> 00:08:45.649 Beyond the big platforms, there's just this constant 00:08:45.649 --> 00:08:49.169 stream of new, often niche AI tools emerging 00:08:49.169 --> 00:08:51.710 daily. Right. On the creative side, you've got 00:08:51.710 --> 00:08:54.049 tools like Quick Deep Thick, pushing boundaries 00:08:54.049 --> 00:08:57.450 with immediate face swapping. AI Figure Generator 00:08:57.450 --> 00:09:00.549 for turning 2D pictures into poseable 3D figures 00:09:00.549 --> 00:09:06.169 for artists, designers. creation from simple 00:09:06.169 --> 00:09:09.269 text, images, or even audio makes content production 00:09:09.269 --> 00:09:12.370 accessible to way more people. And for businesses 00:09:12.370 --> 00:09:14.809 trying to navigate this explosion, there's ThirdEye. 00:09:14.889 --> 00:09:17.370 It helps brands stay discoverable by tracking 00:09:17.370 --> 00:09:19.750 all the various AIs and their capabilities. That's 00:09:19.750 --> 00:09:21.830 a fascinating new challenge, right? Brand management 00:09:21.830 --> 00:09:24.299 in the age of AI. Keeping track of the trackers, 00:09:24.340 --> 00:09:26.340 that's quite a lot of movement in creative and 00:09:26.340 --> 00:09:29.000 business apps. What about broader industry news? 00:09:29.320 --> 00:09:31.679 Well, Grammarly, a tool I think many of us use 00:09:31.679 --> 00:09:34.019 daily, is significantly expanding its reach. 00:09:34.139 --> 00:09:36.740 It now supports 19 languages beyond English. 00:09:36.960 --> 00:09:40.539 Wow, 19. That's a huge leap for global accessibility. 00:09:41.059 --> 00:09:44.759 It really is. And while Google's highly anticipated 00:09:44.759 --> 00:09:46.919 Gemini 3 isn't out this month, they've promised 00:09:46.919 --> 00:09:50.299 it's soon coming after their current 2 .5 Pro 00:09:50.299 --> 00:09:52.980 version. So anticipation is building there. Okay. 00:09:53.059 --> 00:09:55.159 Keeping an eye on that. Right. And the language 00:09:55.159 --> 00:09:57.799 expansion is a big one for global users. We also 00:09:57.799 --> 00:10:00.879 have Stability AI launching Stable Audio 2 .5. 00:10:01.000 --> 00:10:03.740 This lets users create pretty impressive music 00:10:03.740 --> 00:10:06.279 tracks up to three minutes long now with unprecedented 00:10:06.279 --> 00:10:09.500 quality for AI Music Gen. Three minutes. Wow. 00:10:10.139 --> 00:10:12.240 In a really strategic development, both Alibaba 00:10:12.240 --> 00:10:15.360 and Baidu in China are beginning to use their 00:10:15.360 --> 00:10:18.320 own custom chips to train AI models. Ah, interesting. 00:10:18.700 --> 00:10:20.980 Moving away from NVIDIA, maybe. Potentially, 00:10:21.080 --> 00:10:23.539 yeah. It's a significant move, not just for cost 00:10:23.539 --> 00:10:25.460 efficiency, but definitely for national tech 00:10:25.460 --> 00:10:28.240 independence, reducing reliance on external semiconductor 00:10:28.240 --> 00:10:30.820 suppliers. Right. That makes sense. Finally, 00:10:30.879 --> 00:10:32.840 on the education front. Florida State University 00:10:32.840 --> 00:10:35.740 is joining the Google AI for Education Accelerator. 00:10:35.820 --> 00:10:38.960 Okay. So it's truly clear, isn't it? AI is integrating 00:10:38.960 --> 00:10:41.220 into every corner of our lives. Creative work, 00:10:41.279 --> 00:10:44.259 national infrastructure, even how we learn. Honestly, 00:10:44.320 --> 00:10:46.179 sometimes. I still wrestle with prompt drift 00:10:46.179 --> 00:10:48.340 myself, just trying to keep up with all these 00:10:48.340 --> 00:10:51.740 new features and tools. It's a lot to process 00:10:51.740 --> 00:10:53.299 and actually integrate into daily workflows. 00:10:53.539 --> 00:10:55.460 It really is a relentless pace of innovation, 00:10:55.720 --> 00:10:57.860 isn't it? What's the common thread you see weaving 00:10:57.860 --> 00:11:00.139 through all these diverse rapid -fire developments 00:11:00.139 --> 00:11:03.399 we just touched on? I guess. AI is just integrating 00:11:03.399 --> 00:11:07.220 deeper into everyday tasks, creative work, and 00:11:07.220 --> 00:11:09.279 yeah, global tech infrastructure itself. Mineral 00:11:09.279 --> 00:11:12.840 sponsor read. Okay, let's shift gears now to 00:11:12.840 --> 00:11:15.519 a really important application of AI healthcare. 00:11:16.009 --> 00:11:18.169 We're going to discuss a recent study focused 00:11:18.169 --> 00:11:22.409 specifically on AI's role in patient education. 00:11:22.730 --> 00:11:25.509 Yes. A recent study put two prominent large language 00:11:25.509 --> 00:11:29.029 models, ChatGPT -40 and DeepSeq -V3, to the test. 00:11:29.169 --> 00:11:31.710 Okay. They were given identical prompts to generate 00:11:31.710 --> 00:11:34.230 patient education guides for four different chronic 00:11:34.230 --> 00:11:37.090 diseases. The goal was really to see how well 00:11:37.090 --> 00:11:39.330 these AIs could produce clear, understandable, 00:11:39.570 --> 00:11:42.169 and reliable health info for the average person. 00:11:42.370 --> 00:11:44.789 And the results were quite illuminating, weren't 00:11:44.789 --> 00:11:47.220 they? Both models generated guys that typically 00:11:47.220 --> 00:11:49.759 landed at about a 9th, 10th grade reading level. 00:11:49.899 --> 00:11:52.100 Right, which is generally considered pretty accessible 00:11:52.100 --> 00:11:54.899 for adults. And they also scored over 80 % on 00:11:54.899 --> 00:11:57.220 understandability, meaning the content was easy 00:11:57.220 --> 00:12:00.299 enough to grasp. And in terms of reliable quality, 00:12:00.539 --> 00:12:03.299 they scored an average of 47 out of 80 on the 00:12:03.299 --> 00:12:06.539 discern scale, which, for those unfamiliar, is 00:12:06.539 --> 00:12:09.519 a widely recognized tool for assessing health 00:12:09.519 --> 00:12:11.899 info quality. Yeah, it's a standard measure. 00:12:12.120 --> 00:12:14.399 So a score in that range indicates the guides 00:12:14.399 --> 00:12:18.419 provided solid foundational information. Good 00:12:18.419 --> 00:12:20.840 for basic understanding. For basic health info, 00:12:20.980 --> 00:12:23.639 they'll certainly do the trick. DeepSeq v3 did 00:12:23.639 --> 00:12:25.820 show a few standout moves that are worth noting, 00:12:25.919 --> 00:12:28.480 I think. It exhibited more original writing. 00:12:28.720 --> 00:12:31.580 Okay. Its Turnitin similarity score was 32 .5 00:12:31.580 --> 00:12:35.480 % versus ChatGPT's 46%. Now that's quite telling. 00:12:35.659 --> 00:12:38.440 A lower similarity score suggests less boilerplate 00:12:38.440 --> 00:12:40.700 language, more unique phrasing. Which you'd want 00:12:40.700 --> 00:12:43.100 inpatient education. Exactly. Something crucial 00:12:43.100 --> 00:12:45.320 for engaging materials that don't sound like 00:12:45.320 --> 00:12:47.720 they were just copied from a textbook. DeepSeq 00:12:47.720 --> 00:12:50.279 also offered more actionable content. It gave 00:12:50.279 --> 00:12:53.580 clearer next steps for patients 65 % of the time 00:12:53.580 --> 00:12:57.159 compared to 50 % for chat GPT. That distinction, 00:12:57.460 --> 00:13:00.259 originality, and actionability, that really matters 00:13:00.259 --> 00:13:02.539 if you're trying to avoid generic language or 00:13:02.539 --> 00:13:04.480 give patients something they can actually do 00:13:04.480 --> 00:13:06.399 with the information. Right. But what does this 00:13:06.399 --> 00:13:09.279 study teach us about AI's capabilities when it 00:13:09.279 --> 00:13:13.340 comes to critical thinking, especially in a field 00:13:13.340 --> 00:13:15.690 as nuanced as medicine? Well, yeah, this is where 00:13:15.690 --> 00:13:18.289 both models kind of fell flat. When asked for 00:13:18.289 --> 00:13:20.889 more complex tasks like providing financial style 00:13:20.889 --> 00:13:23.490 projections for health outcomes or evaluating 00:13:23.490 --> 00:13:26.669 nuanced personalized care strategies, they really 00:13:26.669 --> 00:13:29.309 struggled. They can articulate existing knowledge, 00:13:29.409 --> 00:13:33.120 structure it well. Sure. But. They're not critical 00:13:33.120 --> 00:13:35.620 thinkers yet. They sound intelligent, but they 00:13:35.620 --> 00:13:38.700 don't reason the way a human expert does. And 00:13:38.700 --> 00:13:40.980 crucially, this study strongly reinforced that 00:13:40.980 --> 00:13:43.299 they cannot and really should not replace doctors 00:13:43.299 --> 00:13:45.919 or human educators for personalized advice. Not 00:13:45.919 --> 00:13:48.600 even close. Their role remains supportive, definitely 00:13:48.600 --> 00:13:51.419 not primary. This study offers a very clear picture 00:13:51.419 --> 00:13:54.620 then. What fundamentally does it teach us about 00:13:54.620 --> 00:13:58.639 AI's current and maybe future role in these complex 00:13:58.639 --> 00:14:01.129 human -centric fields like medicine? Well, it 00:14:01.129 --> 00:14:03.250 shows AI supports basic information delivery 00:14:03.250 --> 00:14:06.210 quite well now. But critical thinking and that 00:14:06.210 --> 00:14:09.250 nuanced care aspect, they absolutely remain human 00:14:09.250 --> 00:14:12.450 domains. What a deep dive indeed from fixing 00:14:12.450 --> 00:14:15.210 a fundamental AI unpredictability bug with those 00:14:15.210 --> 00:14:18.700 batch invariant kernels. I mean, that genuinely 00:14:18.700 --> 00:14:21.820 changes the game for reliability. Huge. To exploring 00:14:21.820 --> 00:14:24.539 new creative tools like Google VO3, and then 00:14:24.539 --> 00:14:27.960 finally understanding AI's very real humbling 00:14:27.960 --> 00:14:30.019 limitations in critical areas like healthcare. 00:14:30.200 --> 00:14:33.360 The pace of innovation is truly staggering. Absolutely. 00:14:33.399 --> 00:14:35.639 We've seen incredible breakthroughs showing how 00:14:35.639 --> 00:14:38.440 even small focus teams can profoundly impact 00:14:38.440 --> 00:14:40.899 the future of AI. That's exciting. Yeah. But 00:14:40.899 --> 00:14:43.700 we also saw a clear, important reminder. While 00:14:43.700 --> 00:14:46.100 AI can process, synthesize, format information 00:14:46.100 --> 00:14:48.700 brilliantly, the capacity for genuine critical 00:14:48.700 --> 00:14:51.720 thinking, nuanced human insight, and truly personalized 00:14:51.720 --> 00:14:54.940 care, that remains uniquely ours. It's a powerful 00:14:54.940 --> 00:14:57.220 balance to understand. So as you continue to 00:14:57.220 --> 00:14:59.500 interact with and apply AI in your own world, 00:14:59.620 --> 00:15:01.600 maybe consider where its strengths truly lie, 00:15:01.679 --> 00:15:03.519 particularly its newfound consistency and where 00:15:03.519 --> 00:15:06.360 that invaluable human touch is absolutely essential. 00:15:06.840 --> 00:15:09.580 Yeah. Think about how we might leverage AI's 00:15:09.580 --> 00:15:12.460 reliable information delivery now that it's less 00:15:12.460 --> 00:15:15.779 like that unpredictable Starbucks, right? To 00:15:15.779 --> 00:15:19.039 free up human experts for what only they can 00:15:19.039 --> 00:15:21.360 truly do. The critical thinking. The empathy, 00:15:21.500 --> 00:15:24.500 the judgment, all the stuff AI simply hasn't 00:15:24.500 --> 00:15:26.299 mastered. Thank you for joining us for this deep 00:15:26.299 --> 00:15:28.059 dive. We hope you gained some valuable insights 00:15:28.059 --> 00:15:30.539 and maybe a clearer perspective on this rapidly 00:15:30.539 --> 00:15:32.700 evolving field. Keep exploring. Keep learning.