WEBVTT 00:00:00.000 --> 00:00:02.439 So we have AI models now that honestly feel like 00:00:02.439 --> 00:00:05.339 they have this incredible grasp language. They 00:00:05.339 --> 00:00:09.699 can chew through huge data sets, spot really 00:00:09.699 --> 00:00:12.919 subtle hate speech online, flag toxic stuff on 00:00:12.919 --> 00:00:16.500 big platforms. And even, you know, write pretty 00:00:16.500 --> 00:00:18.859 complex marketing campaigns basically from scratch. 00:00:19.899 --> 00:00:22.420 It's amazing abstract thinking. Right. And then 00:00:22.420 --> 00:00:25.039 you take that same, like, immense brain power, 00:00:25.120 --> 00:00:27.739 stick it in a physical robot body. Yeah. And 00:00:27.739 --> 00:00:30.079 these machines literally just spin around in 00:00:30.079 --> 00:00:33.039 circles, lost, trying to find the butter on a 00:00:33.039 --> 00:00:36.079 kitchen counter. Why is that physical common 00:00:36.079 --> 00:00:38.420 sense, that simple intuitive stuff we do without 00:00:38.420 --> 00:00:41.659 thinking, why is that still the absolute biggest 00:00:41.659 --> 00:00:44.520 hurdle for AI? It is. That huge kind of hilarious 00:00:44.520 --> 00:00:47.259 gap. That's the tension we're really digging 00:00:47.259 --> 00:00:49.140 into today. It's a fascinating contradiction, 00:00:49.359 --> 00:00:51.460 isn't it? And our sources lay it out really well. 00:00:51.539 --> 00:00:53.200 We're going to kind of trace this whole development 00:00:53.200 --> 00:00:55.079 arc. Okay. First up, we'll look at these new 00:00:55.079 --> 00:00:57.560 safety and transparency pushes from the big players, 00:00:57.619 --> 00:01:00.100 like open AI, how they're trying to build trust. 00:01:00.359 --> 00:01:02.740 Right. Then we need to look at just how fast 00:01:02.740 --> 00:01:05.099 digital AI is being adopted, especially like 00:01:05.099 --> 00:01:07.859 creative tools, marketing, business automation. 00:01:08.680 --> 00:01:11.340 It's just surging. Yeah. It really is. We'll 00:01:11.340 --> 00:01:13.540 touch on the huge infrastructure build supporting 00:01:13.540 --> 00:01:15.780 that speed, too. Good point. And finally, we 00:01:15.780 --> 00:01:18.260 absolutely have to confront that reality gap. 00:01:18.439 --> 00:01:22.219 The butter bench test, robots failing super simple 00:01:22.219 --> 00:01:25.200 home tasks. It just perfectly shows the limits 00:01:25.200 --> 00:01:27.459 when digital smarts meet the physical world. 00:01:27.620 --> 00:01:29.480 OK, let's start there then with segment one, 00:01:29.579 --> 00:01:33.239 this push for trust, for transparency. OpenAI 00:01:33.239 --> 00:01:35.659 has been, well, under a lot of pressure lately, 00:01:35.799 --> 00:01:38.379 right, about safety misuse. Definitely. And that 00:01:38.379 --> 00:01:40.599 pressure seems to be leading to models that are 00:01:40.599 --> 00:01:42.959 specifically built to show their work. Exactly. 00:01:42.959 --> 00:01:44.879 They just put out two new reasoning models. There's 00:01:44.879 --> 00:01:48.700 the big one, GPTI Safeguard 120B. 120 billion 00:01:48.700 --> 00:01:51.079 parameters. Yeah, serious scale. And then the 00:01:51.079 --> 00:01:54.060 smaller, maybe more accessible one, GPTI Safeguard 00:01:54.060 --> 00:01:56.900 20B. These aren't just tweaks. They are custom 00:01:56.900 --> 00:01:59.519 -tuned, really focused on online safety tasks. 00:02:00.040 --> 00:02:02.260 And their job is pretty specific, right? Like 00:02:02.260 --> 00:02:05.379 finding fake reviews, shutting down nasty comments, 00:02:05.500 --> 00:02:08.039 maybe catching cheating talk on gaming forums 00:02:08.039 --> 00:02:10.699 or something. Precisely. But the really unique 00:02:10.699 --> 00:02:13.379 thing, and honestly the critical part, is they 00:02:13.379 --> 00:02:16.500 don't just say yes or no, safe or unsafe. They're 00:02:16.500 --> 00:02:18.879 built to show their reasoning. They actually 00:02:18.879 --> 00:02:22.280 explain, often step by step, why they flagged 00:02:22.280 --> 00:02:24.960 a comment or a response as harmful or toxic. 00:02:25.439 --> 00:02:28.120 That seems huge for platforms. So if you're, 00:02:28.120 --> 00:02:30.879 say, a Reddit mod dealing with tons of posts. 00:02:31.120 --> 00:02:33.659 You could use this to fight hate speech, and 00:02:33.659 --> 00:02:36.180 the AI actually gives you the justification for 00:02:36.180 --> 00:02:39.319 pulling something down. Exactly. Or maybe a review 00:02:39.319 --> 00:02:42.259 site feels more confident catching really clever 00:02:42.259 --> 00:02:45.939 fake reviews because the AI explains the subtle 00:02:45.939 --> 00:02:47.939 signs it picked up on. And you have to assume 00:02:47.939 --> 00:02:50.379 this push for transparency is partly a response 00:02:50.379 --> 00:02:52.719 to some very serious incidents, like that case 00:02:52.719 --> 00:02:55.099 alleging ChatGPT's involvement in a suicide. 00:02:55.439 --> 00:02:57.300 Absolutely. It really highlights that safety 00:02:57.300 --> 00:03:00.080 can't just be an internal checkbox anymore. Platforms 00:03:00.080 --> 00:03:01.800 need to be able to publicly prove and defend 00:03:01.800 --> 00:03:04.020 how they're keeping things safe. It feels like 00:03:04.020 --> 00:03:06.620 a really important step for accountability. But, 00:03:06.699 --> 00:03:09.479 I mean, doesn't showing the reasoning also kind 00:03:09.479 --> 00:03:11.479 of teach the bad actors how to get around it? 00:03:11.740 --> 00:03:14.400 Ah, yeah. If I know exactly why my toxic comment 00:03:14.400 --> 00:03:17.080 got flagged, can't I just tweak it slightly next 00:03:17.080 --> 00:03:19.919 time to sneak past the filter? That is the fundamental 00:03:19.919 --> 00:03:23.819 cat and mouse game of AI safety. 100%. You are 00:03:23.819 --> 00:03:26.599 kind of giving away the playbook. Okay. But the 00:03:26.599 --> 00:03:30.639 flip side... A completely opaque black box moderation 00:03:30.639 --> 00:03:34.719 system that tanks user trust instantly, especially 00:03:34.719 --> 00:03:36.759 when it makes mistakes. Yeah. So it's a tradeoff. 00:03:37.180 --> 00:03:39.659 Transparency is seen as, well, maybe the lesser 00:03:39.659 --> 00:03:42.379 evil for building that essential user confidence. 00:03:42.620 --> 00:03:44.719 And we should probably admit, even for the folks 00:03:44.719 --> 00:03:47.740 building these systems, getting AI behavior right 00:03:47.740 --> 00:03:49.960 is just incredibly difficult. Oh, absolutely. 00:03:50.180 --> 00:03:52.219 But I still wrestle with prompt drift myself 00:03:52.219 --> 00:03:54.960 sometimes, just trying to get an AI to stay focused 00:03:54.960 --> 00:03:57.150 on a specific task consistently. Prompt drift. 00:03:57.349 --> 00:03:59.430 That's when the model kind of forgets the original 00:03:59.430 --> 00:04:01.770 instructions or it loses context over a long 00:04:01.770 --> 00:04:04.229 chat. Exactly. Getting these incredibly complex 00:04:04.229 --> 00:04:08.669 models to stay reliably within safe, narrow boundaries. 00:04:08.770 --> 00:04:11.389 It's a constant ongoing challenge. A real battle. 00:04:11.710 --> 00:04:14.569 So think about the core value proposition here. 00:04:15.270 --> 00:04:18.290 How crucial is this transparency feature, this 00:04:18.290 --> 00:04:21.589 showing the reasoning for getting users and platforms 00:04:21.589 --> 00:04:24.439 to adopt these safety measures and set... policies 00:04:24.439 --> 00:04:26.540 around them. Fundamentally, it builds trust in 00:04:26.540 --> 00:04:28.939 automated moderation systems. Yeah. Reduces liability 00:04:28.939 --> 00:04:32.860 too. Okay. So moving from the need for safety 00:04:32.860 --> 00:04:35.300 to the sheer speed of innovation, let's talk 00:04:35.300 --> 00:04:37.879 about this incredible pace of adoption in the 00:04:37.879 --> 00:04:40.500 digital realm. Creative tools, marketing AI, 00:04:40.939 --> 00:04:44.220 it feels like it's just exploding right now. 00:04:44.360 --> 00:04:45.879 Exploding is the right word. I mean, look at 00:04:45.879 --> 00:04:48.040 that new Google web tool. You just point it at 00:04:48.040 --> 00:04:50.579 a website and it analyzes the whole company's 00:04:50.579 --> 00:04:53.529 history. It's messaging the source called it 00:04:53.529 --> 00:04:56.829 grabbing the company's DNA. Wow. And then it 00:04:56.829 --> 00:04:59.310 designs full on pretty sophisticated marketing 00:04:59.310 --> 00:05:01.670 campaigns. That's not just the chatbot suggesting 00:05:01.670 --> 00:05:04.310 ideas. That's deep automation doing strategic 00:05:04.310 --> 00:05:07.569 work. And the integrations are getting so much 00:05:07.569 --> 00:05:10.329 smoother, almost invisible, like being able to 00:05:10.329 --> 00:05:13.529 edit using powerful tools, Photoshop, Express 00:05:13.529 --> 00:05:16.250 directly inside ChatGPT. Oh, yeah. That demo 00:05:16.250 --> 00:05:18.750 looked sick, didn't it? That collapse between 00:05:18.750 --> 00:05:21.509 like generating an idea and actually creating 00:05:21.509 --> 00:05:23.990 and deploying the asset. Yeah. That's a massive 00:05:23.990 --> 00:05:26.029 speed up for creative workflows. And we're seeing 00:05:26.029 --> 00:05:28.069 it in video, too, right? It's not just the big 00:05:28.069 --> 00:05:30.329 names like Sora and Vue anymore. Not at all. 00:05:30.389 --> 00:05:32.589 You've got others bubbling up like Nano Banana 00:05:32.589 --> 00:05:35.860 and that LTX2. One that made the SpongeBob clip. 00:05:36.019 --> 00:05:38.240 Yeah, that went viral, right? Made from just 00:05:38.240 --> 00:05:42.180 a short text prompt. It shows how low the barrier 00:05:42.180 --> 00:05:44.939 is getting for making really high quality, impactful 00:05:44.939 --> 00:05:48.019 video content. Anyone can do it, almost. And 00:05:48.019 --> 00:05:50.800 crucially, this incredible software progress 00:05:50.800 --> 00:05:53.560 is built on top of a staggering physical infrastructure 00:05:53.560 --> 00:05:56.000 build out. We can't forget that. That's the essential 00:05:56.000 --> 00:05:59.139 context. Absolutely. I mean, Eli Lilly just opened 00:05:59.139 --> 00:06:01.519 what they're calling the world's largest AI factory 00:06:01.519 --> 00:06:04.079 with NVIDIA, specifically for drug discovery. 00:06:04.259 --> 00:06:08.480 GitHub pulled together AI coding agents from 00:06:08.480 --> 00:06:12.019 OpenAI, Anthropic, Google, basically putting 00:06:12.019 --> 00:06:15.339 all the top coding brains in one spot. for developers. 00:06:15.540 --> 00:06:18.199 Amazon's huge data center expansion in Indiana. 00:06:18.540 --> 00:06:21.019 Apparently, it's explicitly to boost Claude's 00:06:21.019 --> 00:06:23.439 abilities. Right. And over in Asia, KakaoTalk 00:06:23.439 --> 00:06:25.720 is integrating GPT -5 for, you know, smarter 00:06:25.720 --> 00:06:29.269 chat features. Whoa. Just imagine scaling the 00:06:29.269 --> 00:06:32.029 infrastructure needed to handle, say, a billion 00:06:32.029 --> 00:06:34.569 queries every single second. It's almost mind 00:06:34.569 --> 00:06:36.649 -boggling, the sheer computational power being 00:06:36.649 --> 00:06:38.990 assembled globally to fuel this digital shift. 00:06:39.170 --> 00:06:42.029 It's immense. And that scale is why the money 00:06:42.029 --> 00:06:44.730 is just flooding in. Speaking of money, Fireworks 00:06:44.730 --> 00:06:48.350 AI just raised $250 million. Big names involve 00:06:48.350 --> 00:06:51.009 Lightspeed, Index Ventures, and significantly, 00:06:51.310 --> 00:06:53.810 NVIDIA again. What does that size of investment 00:06:53.810 --> 00:06:55.889 tell us about where the market focus is shifting? 00:06:55.990 --> 00:06:58.040 That's a really key signal. that huge chunk of 00:06:58.040 --> 00:07:00.379 cash, it's not primarily going into just building 00:07:00.379 --> 00:07:02.139 bigger models anymore. It's going towards inference. 00:07:02.839 --> 00:07:05.420 That's the speed and efficiency of actually running 00:07:05.420 --> 00:07:08.879 those massive models to serve potentially billions 00:07:08.879 --> 00:07:12.000 of user requests quickly and affordably. The 00:07:12.000 --> 00:07:14.620 priority now is scaling the delivery, not just 00:07:14.620 --> 00:07:16.800 building bigger brains. That's a huge trend to 00:07:16.800 --> 00:07:19.430 watch. Interesting. Yet even with all this speed 00:07:19.430 --> 00:07:22.050 and abstract smarts, these digital platforms 00:07:22.050 --> 00:07:25.610 still run into, well, user friction and sometimes 00:07:25.610 --> 00:07:28.069 kind of funny mistakes. Like the pushback when 00:07:28.069 --> 00:07:30.810 OpenAI floated the idea of ads and chat GPT. 00:07:30.930 --> 00:07:33.110 That was instant. Oh, yeah. That was a powerful 00:07:33.110 --> 00:07:36.189 reality check. Users made it clear. Clutter up 00:07:36.189 --> 00:07:38.709 this essential tool with intrusive ads and we 00:07:38.709 --> 00:07:41.819 are out. Yeah. Ads. OK, bye. Simple as that. 00:07:41.899 --> 00:07:44.240 It's the classic platform monetization dilemma, 00:07:44.399 --> 00:07:46.199 right? User experience suffers if you get it 00:07:46.199 --> 00:07:48.759 wrong. And then you had Meta's AI ad tool reportedly 00:07:48.759 --> 00:07:51.899 replacing a fashion model with an AI -generated 00:07:51.899 --> 00:07:54.439 grandma in an ad. Slight chuckle. Yeah. Even 00:07:54.439 --> 00:07:56.199 when the AI nails the language, sometimes it 00:07:56.199 --> 00:07:59.339 completely misses the... The context or the point 00:07:59.339 --> 00:08:02.139 of the ad itself. It shows that abstract smarts 00:08:02.139 --> 00:08:04.459 don't always equal common sense understanding, 00:08:04.759 --> 00:08:07.720 which is a perfect lead in to our last segment, 00:08:07.920 --> 00:08:10.819 the physical world. Let's make that pivot. From 00:08:10.819 --> 00:08:12.699 the colossal scale of digital infrastructure 00:08:12.699 --> 00:08:16.139 and language mastery. To the sometimes embarrassing 00:08:16.139 --> 00:08:19.939 clumsiness when AI meets physical reality. The 00:08:19.939 --> 00:08:23.000 contract is just wild, isn't it? We're building 00:08:23.000 --> 00:08:25.959 the planet's biggest computers so software can 00:08:25.959 --> 00:08:29.360 detect. like subtle sarcasm online and then we 00:08:29.360 --> 00:08:31.399 asked that same level of intelligence to do something 00:08:31.399 --> 00:08:33.460 physically simple like go get a stick of butter 00:08:33.460 --> 00:08:36.919 from the kitchen and it just Fails miserably 00:08:36.919 --> 00:08:39.179 sometimes. Tell us about this butter bench. Right. 00:08:39.259 --> 00:08:41.960 The butter bench. It's a new robotics test, a 00:08:41.960 --> 00:08:44.379 benchmark. And it's designed to be ridiculously 00:08:44.379 --> 00:08:47.120 easy for a person. Totally mundane. It tests 00:08:47.120 --> 00:08:50.000 how well these top AI models can control standard 00:08:50.000 --> 00:08:52.519 home robots doing basic chores like find the 00:08:52.519 --> 00:08:54.379 butter, go to the kitchen, grab it using touch 00:08:54.379 --> 00:08:56.539 sensors to confirm you actually picked it up, 00:08:56.580 --> 00:08:58.639 find the human, hand it over and then go back 00:08:58.639 --> 00:09:01.120 and plug yourself in safely. Stuff we do without 00:09:01.120 --> 00:09:04.220 even thinking. And the reality gap here is just. 00:09:04.909 --> 00:09:08.730 Stark. Humans nailed it, right? Like 95 % success. 00:09:09.210 --> 00:09:12.169 Crushed it. What about the best AI they tested? 00:09:12.370 --> 00:09:15.710 So the top performer was Google's Gemini 2 .5 00:09:15.710 --> 00:09:20.230 Pro. It only managed a 40 % success rate. 40%. 00:09:20.230 --> 00:09:23.409 That means more often than not. this super advanced 00:09:23.409 --> 00:09:27.350 AI failed a basic chore in a controlled lab setting. 00:09:27.529 --> 00:09:29.710 Failed, yeah. And what's really telling is how 00:09:29.710 --> 00:09:32.509 they failed. It wasn't just like one simple bug. 00:09:32.649 --> 00:09:34.649 No, the failure modes are what make it so informative. 00:09:35.289 --> 00:09:38.090 You had robots literally spinning in circles, 00:09:38.190 --> 00:09:40.549 seemingly lost even with constant location data. 00:09:40.730 --> 00:09:43.549 Wow. They'd miss obvious visual cues, like confusing 00:09:43.549 --> 00:09:45.570 a water bottle on the counter for the package 00:09:45.570 --> 00:09:47.370 they were supposed to get. Or they'd lose track 00:09:47.370 --> 00:09:49.090 of where they were completely just by navigating 00:09:49.090 --> 00:09:51.090 around a simple obstacle, like the corner of 00:09:51.090 --> 00:09:53.549 a table. It sounds less like a failure of pure 00:09:53.549 --> 00:09:56.450 intelligence and more like a failure of embodiment, 00:09:56.590 --> 00:09:58.769 like connecting the brain to the world. That's 00:09:58.769 --> 00:10:01.169 exactly it. And the low battery panic was a great 00:10:01.169 --> 00:10:02.830 example. What happened there? When the robot's 00:10:02.830 --> 00:10:06.350 power started dropping. The high -level AI, the 00:10:06.350 --> 00:10:09.169 brain, couldn't execute the most basic emergency 00:10:09.169 --> 00:10:12.049 procedure. Just go back and dock itself to recharge. 00:10:12.370 --> 00:10:15.149 It would freeze. Or just start moving erratically. 00:10:15.330 --> 00:10:18.809 It really is like watching a toddler figure out 00:10:18.809 --> 00:10:21.629 how the world works. Except this toddler also 00:10:21.629 --> 00:10:24.889 happens to have, like, a PhD in theoretical physics 00:10:24.889 --> 00:10:27.470 and can write flawless essays on string theory. 00:10:27.730 --> 00:10:30.509 That analogy is spot on. The physical world, 00:10:30.529 --> 00:10:33.370 with all its messiness, friction, changing light, 00:10:33.549 --> 00:10:35.490 things being slightly out of place, unexpected 00:10:35.490 --> 00:10:38.950 battery drain, it just overwhelms models that 00:10:38.950 --> 00:10:41.409 thrive on the clean, predictable logic of text 00:10:41.409 --> 00:10:44.110 and code. So if you boil it down, what does the 00:10:44.110 --> 00:10:46.509 butter bench really prove about the core challenges 00:10:46.509 --> 00:10:49.409 facing robotics and physical AI right now? It 00:10:49.409 --> 00:10:51.870 proves the sheer complexity of real world variables 00:10:51.870 --> 00:10:54.649 still swamps abstract intelligence. Yeah. Sensor 00:10:54.649 --> 00:10:56.590 fusion, low level motor control. That's the hard 00:10:56.590 --> 00:10:58.590 part. The software interacting with physics is 00:10:58.590 --> 00:11:00.309 the bottleneck. Which brings us right back to 00:11:00.309 --> 00:11:01.970 that central duality we started with, the defining 00:11:01.970 --> 00:11:04.929 feature of AI today. On the one hand, you've 00:11:04.929 --> 00:11:07.929 got AI mastering incredibly high level reasoning. 00:11:08.110 --> 00:11:11.190 It's tackling transparency, writing complex code, 00:11:11.350 --> 00:11:13.629 designing global marketing strategies. It gets 00:11:13.629 --> 00:11:16.830 the abstract stuff, but it possesses the basic 00:11:16.830 --> 00:11:19.509 physical common sense, the object permanence 00:11:19.509 --> 00:11:22.710 of maybe a three year old child. That gap between 00:11:22.710 --> 00:11:24.789 the language models understanding the word butter 00:11:24.789 --> 00:11:27.590 and the embodied AI actually passing the butter. 00:11:27.909 --> 00:11:30.190 That's the defining challenge for the next decade, 00:11:30.309 --> 00:11:32.769 maybe more. It seems we can make the models smarter 00:11:32.769 --> 00:11:35.129 in terms of abstract thought incredibly quickly. 00:11:35.269 --> 00:11:37.789 That part's accelerating. But making them physically 00:11:37.789 --> 00:11:40.389 competent, integrating that smartness smoothly 00:11:40.389 --> 00:11:42.929 into actual physics, into mechanics, that feels 00:11:42.929 --> 00:11:46.070 like a much slower, much harder climb. The digital 00:11:46.070 --> 00:11:49.309 brain's brilliance meets the physical body's, 00:11:49.370 --> 00:11:51.990 well, clumsiness. The bottleneck isn't the intelligence 00:11:51.990 --> 00:11:54.269 anymore. It's the interaction with reality. It's 00:11:54.269 --> 00:11:56.730 the body. So here's a final thought to leave 00:11:56.730 --> 00:11:59.830 with. If the world's most advanced AI that can 00:11:59.830 --> 00:12:02.830 write elegant code and spot nuanced hate speech 00:12:02.830 --> 00:12:06.610 totally fumbles simple physical tasks like passing 00:12:06.610 --> 00:12:09.289 butter, what does that really mean for the future 00:12:09.289 --> 00:12:12.049 of complex manual jobs? Things like construction 00:12:12.049 --> 00:12:14.830 or maybe elder care, where adapting to unpredictable 00:12:14.830 --> 00:12:18.210 physical environments is everything. Maybe that 00:12:18.210 --> 00:12:21.409 true robot revolution in physical labor. Maybe 00:12:21.409 --> 00:12:23.289 it's still quite a bit further off than some 00:12:23.289 --> 00:12:25.549 of the headlines suggest. That's a really powerful 00:12:25.549 --> 00:12:28.230 thought to consider. That gap is very, very real. 00:12:28.370 --> 00:12:30.090 Well, thank you for joining us for this deep 00:12:30.090 --> 00:12:32.110 dive. We really encourage you as you go about 00:12:32.110 --> 00:12:35.029 your week to think about which side of this AI 00:12:35.029 --> 00:12:37.850 duality, the digital brilliance or the physical 00:12:37.850 --> 00:12:40.149 awkwardness you notice most in your own life 00:12:40.149 --> 00:12:42.870 and work. Yeah, great point. Until next time, 00:12:42.909 --> 00:12:43.929 keep digging into the sources.