WEBVTT 00:00:00.000 --> 00:00:03.240 Imagine an AI that can absolutely ace medical 00:00:03.240 --> 00:00:07.919 exams. I mean, outscore real doctors. But then 00:00:07.919 --> 00:00:09.759 when you or I try to use it for something simple 00:00:09.759 --> 00:00:11.960 like a stomachache, it actually makes things 00:00:11.960 --> 00:00:14.580 worse. That's not science fiction. That's the 00:00:14.580 --> 00:00:18.359 intriguing paradox we're unpacking today. Welcome 00:00:18.359 --> 00:00:21.390 to the Deep Dive. Yeah, we're really going to 00:00:21.390 --> 00:00:24.050 explore some fascinating corners of AI in this 00:00:24.050 --> 00:00:27.510 deep dive. We're pulling insights from a really 00:00:27.510 --> 00:00:29.769 rich source, a recent newsletter packed with 00:00:29.769 --> 00:00:31.969 the latest developments. Today, we're looking 00:00:31.969 --> 00:00:34.350 at everything from why these AI tools, even the 00:00:34.350 --> 00:00:36.549 brilliant ones, can be kind of double -edged 00:00:36.549 --> 00:00:39.450 sword in the wrong hands. Yeah. All the way to 00:00:39.450 --> 00:00:41.789 the quiet revolution of AI running locally right 00:00:41.789 --> 00:00:43.670 there on your device. Right on your own machine. 00:00:43.829 --> 00:00:45.969 Exactly. And then there's this flurry of new 00:00:45.969 --> 00:00:47.829 tools changing, well, everything from how we 00:00:47.829 --> 00:00:50.530 see ads to how robots learn. Okay. Oh, and just 00:00:50.530 --> 00:00:52.250 so we're on the same page, when we talk about 00:00:52.250 --> 00:00:54.750 LLMs, we mean large language models. Just think 00:00:54.750 --> 00:00:56.689 of them as like advanced chatbots that learn 00:00:56.689 --> 00:00:58.850 from truly massive amounts of text and data. 00:00:58.950 --> 00:01:01.270 Yeah. They understand, summarize, generate stuff. 00:01:01.670 --> 00:01:03.729 Got it. Okay, so let's really unpack this first 00:01:03.729 --> 00:01:07.969 one. The AI doctor dilemma. If AI is so incredibly 00:01:07.969 --> 00:01:10.510 smart, I mean, demonstrating near perfect knowledge, 00:01:10.750 --> 00:01:12.709 why does it seem to fall apart when an everyday 00:01:12.709 --> 00:01:14.870 person, you know, just needs help with a common 00:01:14.870 --> 00:01:18.969 sickness? It feels counterintuitive. It absolutely 00:01:18.969 --> 00:01:21.390 does. And the research highlighted in our source 00:01:21.390 --> 00:01:23.689 material, it really drills into this. They ran 00:01:23.689 --> 00:01:27.909 this study involved about 1300 people using advanced 00:01:27.909 --> 00:01:31.879 LLMs like GPT -4 -0 trying to navigate. medical 00:01:31.879 --> 00:01:35.260 scenarios. Now, when these LLMs were just doing 00:01:35.260 --> 00:01:37.439 their thing, responding to clinical prompts, 00:01:37.760 --> 00:01:39.760 their accuracy was astonishing. We're talking 00:01:39.760 --> 00:01:43.920 90 % to 99%. Wow. Incredible, right? But here's 00:01:43.920 --> 00:01:46.260 the twist. When humans got involved trying to 00:01:46.260 --> 00:01:48.680 self -diagnose using these same tools, the accuracy 00:01:48.680 --> 00:01:51.480 just plummeted. Plummeted how much? Only 34 .5 00:01:51.480 --> 00:01:53.780 % of the scenarios were correctly assessed. And 00:01:53.780 --> 00:01:56.840 here's the truly wild part. A control group just 00:01:56.840 --> 00:01:59.200 using Google, they actually did better. Better 00:01:59.200 --> 00:02:02.120 than the advanced AI. Yep. 47 % accuracy. The 00:02:02.120 --> 00:02:04.340 data suggests adding this powerful AI actually 00:02:04.340 --> 00:02:06.599 made people worse at self -diagnosing. That's, 00:02:06.599 --> 00:02:08.699 yeah, that's pretty sobering. What's going on 00:02:08.699 --> 00:02:11.180 there? Is the AI giving bad advice or is it how 00:02:11.180 --> 00:02:14.080 people are using it? It's rarely about the AI 00:02:14.080 --> 00:02:16.919 giving flat -out bad advice, not with these professional 00:02:16.919 --> 00:02:19.060 -grade models anyway. The reasons the research 00:02:19.060 --> 00:02:22.039 uncovered were quite clear and, honestly, very 00:02:22.039 --> 00:02:25.460 human. Like what? Well, first, users often gave 00:02:25.460 --> 00:02:26.939 incomplete information. They didn't have the 00:02:26.939 --> 00:02:30.219 diagnostic discipline, you know. Second, frequent 00:02:30.219 --> 00:02:33.699 misinterpretation of what the AI spat out. The 00:02:33.699 --> 00:02:36.240 AI might list possibilities, but people kind 00:02:36.240 --> 00:02:39.219 of latched onto one or just didn't get the nuances. 00:02:39.219 --> 00:02:42.060 And maybe most frustratingly, people just ignored 00:02:42.060 --> 00:02:44.780 good AI advice. Ignored it, even when it was 00:02:44.780 --> 00:02:47.750 right. Even when the LLM correctly flagged relevant 00:02:47.750 --> 00:02:50.189 conditions, only about one in three people actually 00:02:50.189 --> 00:02:52.770 use that information. I have to admit, I still 00:02:52.770 --> 00:02:54.990 wrestle with prompt drift myself sometimes, you 00:02:54.990 --> 00:02:57.550 know, where the AI's answers kind of shift subtly 00:02:57.550 --> 00:03:00.169 over time. So the idea of misinterpreting or 00:03:00.169 --> 00:03:03.530 overlooking AI suggestions, well, it isn't totally 00:03:03.530 --> 00:03:05.610 foreign to me either. So it's less about the 00:03:05.610 --> 00:03:07.870 AI's raw intelligence and more about the human 00:03:07.870 --> 00:03:10.729 interface. And maybe the training needed to use 00:03:10.729 --> 00:03:13.110 that intelligence safely. Because like you said, 00:03:13.250 --> 00:03:15.469 in professional settings, AI seems to be making 00:03:15.469 --> 00:03:17.949 a huge difference. Exactly. It highlights that 00:03:17.949 --> 00:03:21.689 critical gap between raw, super accurate AI data 00:03:21.689 --> 00:03:24.969 and what you might call actionable, trusted guidance. 00:03:25.509 --> 00:03:28.939 A doctor brings years of training, right? deep 00:03:28.939 --> 00:03:31.639 context about a patient, the ability to ask precise 00:03:31.639 --> 00:03:34.460 follow -ups. Right. An average user often lacks 00:03:34.460 --> 00:03:36.500 that. They might not even know what to ask next, 00:03:36.580 --> 00:03:39.419 or they might see a probability as a definite 00:03:39.419 --> 00:03:42.199 diagnosis. Which it isn't. No. Professionals, 00:03:42.199 --> 00:03:44.199 though, they have the framework to interpret 00:03:44.199 --> 00:03:47.060 that data correctly. Take open evidence, for 00:03:47.060 --> 00:03:49.560 example. It's a diagnostic engine trained only 00:03:49.560 --> 00:03:52.000 on peer -reviewed medical literature. Not for 00:03:52.000 --> 00:03:54.439 consumers. It's for pros. Okay. And the numbers. 00:03:54.680 --> 00:03:57.479 One in four U .S. doctors use it. On average, 00:03:57.639 --> 00:04:00.240 10 times a day. Especially valuable in complex 00:04:00.240 --> 00:04:02.659 areas like oncology. It just shows how powerful 00:04:02.659 --> 00:04:04.379 these tools are when they're in the right hand, 00:04:04.520 --> 00:04:07.039 used the right way. So what does this all mean 00:04:07.039 --> 00:04:10.360 for us then? This tension between consumer use 00:04:10.360 --> 00:04:14.039 and professional success. It seems to hinge on 00:04:14.039 --> 00:04:16.399 who's using it and why. It's not about needing 00:04:16.399 --> 00:04:18.240 a robot doctor maybe, but something different. 00:04:18.500 --> 00:04:21.040 Right. It means AI can save lives? Absolutely. 00:04:22.269 --> 00:04:24.850 Only if we design it to genuinely help people 00:04:24.850 --> 00:04:26.930 in a way that respects the complexity of human 00:04:26.930 --> 00:04:29.449 interaction and expertise, not just design it 00:04:29.449 --> 00:04:31.769 to ace a test. So if AI gives correct advice, 00:04:31.990 --> 00:04:34.509 but people don't listen, what's the real barrier 00:04:34.509 --> 00:04:37.769 then? It's about bridging that raw data with 00:04:37.769 --> 00:04:40.930 actionable, trusted guidance. That's the core 00:04:40.930 --> 00:04:42.870 challenge. Okay, shifting gears a bit, we often 00:04:42.870 --> 00:04:45.569 hear about AI in these giant data centers, you 00:04:45.569 --> 00:04:47.230 know, humming away in the cloud, needing constant 00:04:47.230 --> 00:04:49.629 internet. But what if the next big shift is bringing 00:04:49.629 --> 00:04:52.509 that power much closer? Right. Right onto your 00:04:52.509 --> 00:04:54.509 device. That's exactly what's happening. There's 00:04:54.509 --> 00:04:56.490 a really strong push for AI that lives directly 00:04:56.490 --> 00:04:58.949 on your machine, your laptop, your phone, maybe 00:04:58.949 --> 00:05:01.610 even your robot vacuum. Huh. Why the shift? Well, 00:05:01.649 --> 00:05:03.769 a lot of users and developers, too, are getting 00:05:03.769 --> 00:05:06.329 kind of tired of unpredictable cloud AI updates, 00:05:06.529 --> 00:05:10.269 right? And the occasional outages. Local AI offers 00:05:10.269 --> 00:05:13.269 much better stability. Your workflows stay consistent. 00:05:13.529 --> 00:05:15.829 Okay. Stability makes sense. And you get far 00:05:15.829 --> 00:05:17.790 greater control over the AI models themselves. 00:05:18.819 --> 00:05:21.839 And, crucially, enhance privacy for your data. 00:05:22.339 --> 00:05:24.899 Think of it like having your own personal AI 00:05:24.899 --> 00:05:28.079 assistant. Always there, always ready, and your 00:05:28.079 --> 00:05:29.720 info doesn't need to travel across the internet. 00:05:29.920 --> 00:05:31.920 That's a big deal for privacy. It is. It's a 00:05:31.920 --> 00:05:33.980 fundamental shift in how we might interact with 00:05:33.980 --> 00:05:36.819 AI. Moving the processing from some distant server 00:05:36.819 --> 00:05:38.800 farm right into your pocket or on your desk, 00:05:38.860 --> 00:05:41.899 it gives you a level of consistency. Cloud services 00:05:41.899 --> 00:05:45.170 just can't always guarantee. Huge benefit. So 00:05:45.170 --> 00:05:47.889 what's the core advantage then for choosing local 00:05:47.889 --> 00:05:50.610 AI over the cloud? Greater control, consistent 00:05:50.610 --> 00:05:52.850 performance, and stronger data privacy, bringing 00:05:52.850 --> 00:05:55.689 the power to you. Okay. So beyond these fundamental 00:05:55.689 --> 00:05:58.069 shifts in where AI runs, the whole landscape 00:05:58.069 --> 00:06:01.129 is just exploding. New tools, new applications 00:06:01.129 --> 00:06:02.910 everywhere. It's not just helping doctors anymore. 00:06:03.250 --> 00:06:06.500 Oh, absolutely not. It's reshaping creative industries. 00:06:06.740 --> 00:06:09.579 It's influencing public opinion, changing how 00:06:09.579 --> 00:06:13.500 we use tech every single day. The pace is incredible. 00:06:13.720 --> 00:06:15.459 Give us some examples. What's catching your eye? 00:06:15.600 --> 00:06:18.189 Okay, we'll look at voice AI. Eleven Labs, they're 00:06:18.189 --> 00:06:19.870 known for voice stuff, right? They just launched 00:06:19.870 --> 00:06:21.689 an AI assistant you interact with just using 00:06:21.689 --> 00:06:24.350 your voice. And it connects easily to other apps 00:06:24.350 --> 00:06:27.750 like Slack or Perplexity. So just talking to 00:06:27.750 --> 00:06:30.329 your apps. Yeah, imagine just talking naturally 00:06:30.329 --> 00:06:32.889 like they're another person. No typing, no clicking. 00:06:33.310 --> 00:06:35.930 Then there's advertising. Did you see that Kelshi 00:06:35.930 --> 00:06:37.970 ad during the NBA finals? I might have missed 00:06:37.970 --> 00:06:41.110 that one. It was wild. 30 seconds. Made entirely 00:06:41.110 --> 00:06:44.189 with AI. Reportedly cost a tiny fraction of a 00:06:44.189 --> 00:06:46.509 normal ad shoot. It sparked this huge debate 00:06:46.509 --> 00:06:49.350 about the future of advertising. I bet. Disrupting 00:06:49.350 --> 00:06:51.829 traditional industries. Totally. Democratizing 00:06:51.829 --> 00:06:54.709 it, maybe, but also challenging creative workflows. 00:06:55.029 --> 00:06:57.810 Sounds powerful. That kind of power, it often 00:06:57.810 --> 00:06:59.490 brings you challenges, right? Especially around 00:06:59.490 --> 00:07:02.730 information. Absolutely. And this is more concerning. 00:07:02.910 --> 00:07:05.339 We're seeing AI generate... misinformation on 00:07:05.339 --> 00:07:08.060 a really alarming scale. The newsletter points 00:07:08.060 --> 00:07:11.720 to two highly realistic AI generated pro -Iran 00:07:11.720 --> 00:07:14.160 propaganda pieces. Oh, wow. Flooding TikTok, 00:07:14.439 --> 00:07:17.480 Instagram, Facebook, YouTube. Over 30 million 00:07:17.480 --> 00:07:20.839 views on TikTok alone. Posted hundreds of times. 00:07:21.000 --> 00:07:23.300 It raises these really serious questions about 00:07:23.300 --> 00:07:26.139 manipulation, misinformation, figuring out what's 00:07:26.139 --> 00:07:28.579 even real online anymore. Yeah, that's a huge 00:07:28.579 --> 00:07:30.759 ethical challenge. We're just starting to grapple 00:07:30.759 --> 00:07:33.120 with that. We really are. And then on the productivity 00:07:33.120 --> 00:07:36.089 side. AI is making a big push there, too. How 00:07:36.089 --> 00:07:38.449 so? Well, ChatGPT just rolled out new features. 00:07:38.670 --> 00:07:41.649 Real -time document collaboration. Speedy PDF 00:07:41.649 --> 00:07:44.350 export with citations. Ah, so competing with 00:07:44.350 --> 00:07:46.990 Google Docs, Microsoft Word. Exactly, a direct 00:07:46.990 --> 00:07:49.290 play. The productivity sweep battle is definitely 00:07:49.290 --> 00:07:51.129 heating up, and AI is right in the middle of 00:07:51.129 --> 00:07:53.230 it. And connecting back to our first point about 00:07:53.230 --> 00:07:55.589 professional use, look at a bridge. The medical 00:07:55.589 --> 00:07:58.560 AI app. Yeah. It's now valued at a staggering 00:07:58.560 --> 00:08:01.680 $5 .3 billion, just raised $300 million, $800 00:08:01.680 --> 00:08:04.180 million total. It just reinforces the immense 00:08:04.180 --> 00:08:06.860 value when AI is applied professionally. That's 00:08:06.860 --> 00:08:09.259 serious money. It is. And Quick Hits on some 00:08:09.259 --> 00:08:12.160 other tools. RunBear lets you use custom GPTs 00:08:12.160 --> 00:08:15.399 inside Slack or Teams. Dubbing 3 .0 translates 00:08:15.399 --> 00:08:18.959 videos into like 30 languages, one click. Sounds 00:08:18.959 --> 00:08:22.180 natural. Great for creators. Overflow AI turns 00:08:22.180 --> 00:08:25.470 questions into instant charts. makes data easy. 00:08:25.689 --> 00:08:27.930 Lots of tools emerging. And one last thing, a 00:08:27.930 --> 00:08:31.069 big legal and ethical point. Anthropic, using 00:08:31.069 --> 00:08:33.870 copyrighted books to train its AI. It's currently 00:08:33.870 --> 00:08:36.830 being argued as fair use in court. That's huge. 00:08:37.169 --> 00:08:39.009 Yeah, the outcome could have massive implications 00:08:39.009 --> 00:08:41.289 for how all AI models get trained in the future, 00:08:41.409 --> 00:08:43.610 what data they can use. Okay, considering all 00:08:43.610 --> 00:08:45.929 that advertising, misinformation, productivity, 00:08:46.409 --> 00:08:50.309 ethics. What's the most surprising area AI is 00:08:50.309 --> 00:08:52.769 impacting right now for you? Its ability to generate 00:08:52.769 --> 00:08:56.429 media from ads to propaganda at scale and low 00:08:56.429 --> 00:08:59.529 cost. That speed and accessibility is just transformative. 00:08:59.830 --> 00:09:01.769 Okay, let's pivot to the physical world now. 00:09:01.850 --> 00:09:04.429 Let's talk robots. Google, just kind of quietly, 00:09:04.649 --> 00:09:06.370 dropped something pretty big, something that 00:09:06.370 --> 00:09:08.210 could fundamentally shift how robots operate. 00:09:08.409 --> 00:09:10.590 Yeah, this is potentially a huge lead forward, 00:09:10.710 --> 00:09:12.789 a real game changer for robotics. It's called 00:09:12.789 --> 00:09:16.019 Gemini Robotics on Device. On device. So running 00:09:16.019 --> 00:09:17.379 locally, like we were talking about earlier. 00:09:17.559 --> 00:09:21.860 Exactly. Robots running complex AI models locally. 00:09:22.320 --> 00:09:25.340 Think about that. No cloud connection needed 00:09:25.340 --> 00:09:28.840 constantly. No lag. No Wi -Fi dependency. That's 00:09:28.840 --> 00:09:30.519 different. Most robots need that connection, 00:09:30.639 --> 00:09:33.259 right? Until now, yeah. Most AI -powered robots 00:09:33.259 --> 00:09:36.080 had to send signals back and forth to huge servers 00:09:36.080 --> 00:09:38.539 just to, you know, move an arm or pick something 00:09:38.539 --> 00:09:41.299 up. The computation happened off -board. This 00:09:41.299 --> 00:09:44.740 new Gemini model lives inside the robot. So what 00:09:44.740 --> 00:09:46.899 does that mean practically? Faster responses, 00:09:47.100 --> 00:09:49.700 better privacy because the data stays on the 00:09:49.700 --> 00:09:52.399 robot, and way more potential for real -world 00:09:52.399 --> 00:09:54.720 use, especially where internet might be spotty 00:09:54.720 --> 00:09:57.519 or non -existent. Google's saying this on -device 00:09:57.519 --> 00:09:59.639 model performs almost as well as its bigger cloud 00:09:59.639 --> 00:10:02.320 sibling and even beats some unnamed rivals in 00:10:02.320 --> 00:10:04.740 benchmarks. And this isn't just theory. They've 00:10:04.740 --> 00:10:07.590 shown it working. Oh, yeah. The live demos were 00:10:07.590 --> 00:10:09.529 pretty impressive. These weren't just lab tricks. 00:10:09.830 --> 00:10:12.309 Robots using this new Gemini brain were doing 00:10:12.309 --> 00:10:15.570 surprisingly delicate adaptive things in real 00:10:15.570 --> 00:10:18.129 time. Like what kind of things? Things like unzipping 00:10:18.129 --> 00:10:21.269 bags, carefully folding clothes, even tackling 00:10:21.269 --> 00:10:23.470 totally new tasks they hadn't seen before, like 00:10:23.470 --> 00:10:25.690 assembling parts on a moving conveyor belt. It's 00:10:25.690 --> 00:10:29.250 not just about force. It's about nuanced interaction, 00:10:29.570 --> 00:10:32.860 adaptation. Wow. And what's really cool for developers, 00:10:33.000 --> 00:10:35.639 Google's releasing a full Gemini Robotics SDK. 00:10:35.860 --> 00:10:38.379 It's a software development kit. It means devs 00:10:38.379 --> 00:10:41.789 can fine tune robot tasks using just like. 50 00:10:41.789 --> 00:10:44.789 to 100 examples that few yeah you basically just 00:10:44.789 --> 00:10:46.610 show the robot what to do a few times and it 00:10:46.610 --> 00:10:48.629 gets it's like stacking lego blocks of data for 00:10:48.629 --> 00:10:50.710 them teaching them fast okay connecting this 00:10:50.710 --> 00:10:53.330 to the bigger picture yeah it feels like everyone 00:10:53.330 --> 00:10:56.710 wants to build the gp2 of robotics right that 00:10:56.710 --> 00:10:59.470 foundational model for truly smart autonomous 00:10:59.470 --> 00:11:03.629 machines could this local ai be that leap whoa 00:11:03.629 --> 00:11:06.830 yeah imagine scaling this robots that truly understand 00:11:06.830 --> 00:11:09.899 and adapt in our homes our workplaces operating 00:11:09.899 --> 00:11:12.179 reliably without needing that constant internet 00:11:12.179 --> 00:11:14.700 umbilical cord. That's a massive game changer 00:11:14.700 --> 00:11:16.840 for getting robots out into the real world, responding 00:11:16.840 --> 00:11:19.100 instantly to unpredictable stuff. So how does 00:11:19.100 --> 00:11:21.539 on -device AI fundamentally change how we'll 00:11:21.539 --> 00:11:24.200 interact with robots day to day? It enabled true 00:11:24.200 --> 00:11:27.299 autonomy and real -time responsiveness, freeing 00:11:27.299 --> 00:11:30.019 them from network limits. More natural collaboration 00:11:30.019 --> 00:11:33.799 becomes possible. Sponsor. So reflecting on our 00:11:33.799 --> 00:11:36.100 discussion today, we've really explored this 00:11:36.100 --> 00:11:38.720 fascinating tension at the heart of AI right 00:11:38.720 --> 00:11:42.039 now. It's incredibly powerful, yes. Capable of 00:11:42.039 --> 00:11:44.620 passing the hardest tests, mastering complex 00:11:44.620 --> 00:11:49.139 data. Yet its actual value, its true worth, so 00:11:49.139 --> 00:11:51.659 often depends on how we interact with it, how 00:11:51.659 --> 00:11:53.580 we understand its limits, and maybe most importantly, 00:11:53.639 --> 00:11:55.960 how we design it to genuinely help, not just 00:11:55.960 --> 00:11:58.059 impress. Yeah, and we're definitely seeing that 00:11:58.059 --> 00:12:01.100 clear shift towards AI living closer to us on 00:12:01.100 --> 00:12:03.000 our devices or embedded right into robots. That 00:12:03.000 --> 00:12:05.539 gives us more control, better privacy, and just 00:12:05.539 --> 00:12:08.000 incredible stability. Right. And the sheer range 00:12:08.000 --> 00:12:09.940 of new AI tools. I mean, from shaking up creative 00:12:09.940 --> 00:12:12.899 fields and boosting productivity to raising these 00:12:12.899 --> 00:12:15.340 really tough ethical questions around misinformation. 00:12:15.500 --> 00:12:18.460 Yeah. And this shows how fast this field is moving 00:12:18.460 --> 00:12:20.620 and touching, well, pretty much every corner 00:12:20.620 --> 00:12:23.039 of our lives. Which brings up, I think, an important 00:12:23.039 --> 00:12:25.919 question for all of us. As AI gets smarter and 00:12:25.919 --> 00:12:27.799 as it moves physically closer running on our 00:12:27.799 --> 00:12:30.899 phones, in our homes, via robots, what new responsibilities 00:12:30.899 --> 00:12:33.200 do we take on? You know, the users, the developers. 00:12:33.539 --> 00:12:36.120 How do we ensure it truly serves humanity, helps 00:12:36.120 --> 00:12:40.039 us thrive responsibly? Two secs silence. Lots 00:12:40.039 --> 00:12:41.860 to think about there. We hope this deep dive 00:12:41.860 --> 00:12:43.720 gave you some new insights to mull over. Out 00:12:43.720 --> 00:12:44.440 to your own music.