WEBVTT 00:00:00.000 --> 00:00:02.600 There's a really profound, maybe even a little 00:00:02.600 --> 00:00:05.700 unsettling finding at the heart of our deep dive 00:00:05.700 --> 00:00:07.919 today. Yeah, we learned that the most advanced 00:00:07.919 --> 00:00:12.380 AI models now openly believe they outperform 00:00:12.380 --> 00:00:15.119 humans. In strategic reasoning tasks. Right. 00:00:15.179 --> 00:00:17.000 They don't just calculate. They actually act 00:00:17.000 --> 00:00:19.739 on that belief. Welcome to the deep dive. You've 00:00:19.739 --> 00:00:22.960 handed us a stack of some incredibly current 00:00:22.960 --> 00:00:26.379 sources on AI, and it's a lot. It is. We've got 00:00:26.379 --> 00:00:29.239 everything from... visual tech breakthroughs 00:00:29.239 --> 00:00:32.079 to these huge shifts in infrastructure and funding. 00:00:32.240 --> 00:00:34.140 Our mission here is to really get beyond the 00:00:34.140 --> 00:00:36.920 headlines. We're going to uncover two massive 00:00:36.920 --> 00:00:40.619 parallel trends. First up, how 3D creation is 00:00:40.619 --> 00:00:43.759 being democratized, like instantly, turning your 00:00:43.759 --> 00:00:46.520 2D photos into professional assets in just seconds. 00:00:46.700 --> 00:00:48.039 And second, we're going to follow the money, 00:00:48.140 --> 00:00:50.060 look at the infrastructure deals shaping the 00:00:50.060 --> 00:00:51.740 industry, and then we'll circle back and unpack 00:00:51.740 --> 00:00:53.560 that study we just mentioned. The one showing 00:00:53.560 --> 00:00:55.539 how these large language models are, well... 00:00:55.820 --> 00:00:57.939 suddenly becoming strategically self -aware agents. 00:00:58.179 --> 00:01:00.179 They're adapting their behavior based on who 00:01:00.179 --> 00:01:02.159 they think they're playing against. It's fascinating 00:01:02.159 --> 00:01:05.079 stuff. It really is. So let's jump in with the 00:01:05.079 --> 00:01:07.120 most accessible breakthrough. That's got to be 00:01:07.120 --> 00:01:09.439 Meta's latest offering in the visual space. We're 00:01:09.439 --> 00:01:13.420 talking about SAM 3D. And this tool just, it 00:01:13.420 --> 00:01:16.620 fundamentally changes the speed and ease of creating 00:01:16.620 --> 00:01:19.379 assets for, you know, virtual environments, design 00:01:19.379 --> 00:01:22.549 work. Games. Exactly. The key takeaway is just 00:01:22.549 --> 00:01:24.590 how effortless it is. You take a normal photo, 00:01:24.650 --> 00:01:27.209 you click the object you want, a chair, a dog, 00:01:27.329 --> 00:01:29.849 whatever, and the system just generates a full 00:01:29.849 --> 00:01:33.329 textured 3D mesh. Right. So it's not just a flat 00:01:33.329 --> 00:01:35.549 image. This is a complete model you can rotate, 00:01:35.709 --> 00:01:38.069 look at from all angles, and then just download 00:01:38.069 --> 00:01:40.370 it. And use it in a game engine, a modeling program, 00:01:40.689 --> 00:01:43.099 anything. It's a huge leap from past techniques 00:01:43.099 --> 00:01:45.599 that needed specialized hardware or, you know, 00:01:45.620 --> 00:01:47.659 hours and hours of human labor. And this isn't 00:01:47.659 --> 00:01:49.980 some research paper locked away in a lab, right? 00:01:50.040 --> 00:01:52.579 Not at all. It's freely accessible right now 00:01:52.579 --> 00:01:55.180 in their Segment Anything playground. And that's 00:01:55.180 --> 00:01:57.019 built on their original Segment Anything model, 00:01:57.079 --> 00:01:58.879 which lets you use natural language prompts. 00:01:59.200 --> 00:02:01.400 So you can type something really complex like 00:02:01.400 --> 00:02:03.959 segment the red lamp on the sofa, but ignore 00:02:03.959 --> 00:02:07.299 the pillows. And it just isolates it. Which is 00:02:07.299 --> 00:02:09.740 so critical because it works on images and videos. 00:02:10.300 --> 00:02:12.960 So if you're a video editor, you can track an 00:02:12.960 --> 00:02:15.500 object across multiple frames. That's incredibly 00:02:15.500 --> 00:02:17.460 powerful for post -production. Yeah, and for 00:02:17.460 --> 00:02:19.979 automation. And this is fully production -ready. 00:02:20.099 --> 00:02:23.159 Meta is already integrating it into its own apps. 00:02:23.419 --> 00:02:26.240 So things like their advanced video editing tools, 00:02:26.639 --> 00:02:30.120 these Vibes applications, and even Facebook Marketplace 00:02:30.120 --> 00:02:32.599 for product images. Right. And they actually 00:02:32.599 --> 00:02:34.659 released it in two specialized flavors, which 00:02:34.659 --> 00:02:37.219 I think speaks to their focus on precision. Okay, 00:02:37.300 --> 00:02:39.340 so what are they? You've got SAM 3D objects, 00:02:39.539 --> 00:02:42.479 which handles your standard static items, you 00:02:42.479 --> 00:02:45.020 know, furniture, cars, trees. But the real technical 00:02:45.020 --> 00:02:46.879 marvel where it gets really interesting is SAM 00:02:46.879 --> 00:02:51.020 3D body. Yes. This one reconstructs human figures 00:02:51.020 --> 00:02:54.120 with pretty impressive accuracy. Even when the 00:02:54.120 --> 00:02:56.560 person is, like, partially blocked or in a weird 00:02:56.560 --> 00:02:59.259 pose. Exactly. It intelligently tries to fill 00:02:59.259 --> 00:03:02.180 in the missing data. It's fast. It's easy. But 00:03:02.180 --> 00:03:04.539 we should be balanced here. The sources did point 00:03:04.539 --> 00:03:07.099 out some limitations. Yeah, for sure. The models 00:03:07.099 --> 00:03:10.949 aren't... Ultra HD yet, so fine details like 00:03:10.949 --> 00:03:13.770 the texture on a fabric or a wood grain can still 00:03:13.770 --> 00:03:15.870 look a little blurred, a bit smoothed out. And 00:03:15.870 --> 00:03:18.009 it still has some of those classic AI challenges, 00:03:18.169 --> 00:03:20.490 like it struggles to understand complex interactions, 00:03:20.550 --> 00:03:23.870 a person holding a small tool, for example. And, 00:03:23.870 --> 00:03:26.990 as is so often the case with AI rendering, hands. 00:03:27.740 --> 00:03:30.500 Uh, hands. Hands are still tricky for the model 00:03:30.500 --> 00:03:33.719 to get just right. Okay, so given all this, what's 00:03:33.719 --> 00:03:36.939 the real practical implication of making 3D modeling 00:03:36.939 --> 00:03:40.969 this accessible and just? Instant. It instantly 00:03:40.969 --> 00:03:43.770 lowers the barrier to entry for virtual object 00:03:43.770 --> 00:03:45.909 creation. That accessibility, it kind of sets 00:03:45.909 --> 00:03:48.110 the stage for the next big shift we saw, which 00:03:48.110 --> 00:03:50.270 is just the sheer scale of the industry deals 00:03:50.270 --> 00:03:52.569 happening right now. We're moving from a desktop 00:03:52.569 --> 00:03:55.789 tool to like super cluster infrastructure. Yeah. 00:03:55.889 --> 00:03:57.729 The adoption rate for these new frontier models 00:03:57.729 --> 00:04:00.789 is just staggering. It really is. The reports 00:04:00.789 --> 00:04:03.189 on early Gemini 3 users, they're just building 00:04:03.189 --> 00:04:05.009 nonstop. This isn't just a little bit of growth. 00:04:05.169 --> 00:04:07.719 No. People are already leveraging it in these 00:04:07.719 --> 00:04:10.120 wild, creative ways. We're seeing things like 00:04:10.120 --> 00:04:12.460 automated marketing campaigns designed in an 00:04:12.460 --> 00:04:14.840 afternoon. Or using its new reasoning controls 00:04:14.840 --> 00:04:18.180 to build really complex, multi -step automations 00:04:18.180 --> 00:04:21.319 for, say, personal finance. The complexity is 00:04:21.319 --> 00:04:24.139 moving into the mainstream so fast. And on the 00:04:24.139 --> 00:04:27.860 education front... OpenAI made a huge move. A 00:04:27.860 --> 00:04:31.279 huge commitment. They're giving U .S. K -12 teachers 00:04:31.279 --> 00:04:34.779 free access to ChatGPT. With unlimited usage. 00:04:35.060 --> 00:04:37.720 And that runs until 2027. That's more than two 00:04:37.720 --> 00:04:40.100 years. It's locking in a whole generation of 00:04:40.100 --> 00:04:42.120 future users. Definitely. And the actual interface 00:04:42.120 --> 00:04:44.519 a lot of people use, ChatGPT Atlas, just got 00:04:44.519 --> 00:04:47.199 a major upgrade focused on workflow. Yeah, better 00:04:47.199 --> 00:04:50.319 management tools, vertical tabs, a faster sidebar. 00:04:50.519 --> 00:04:53.079 And strategically, they made Google search the 00:04:53.079 --> 00:04:54.839 default. That's a big deal. It is. Well, let's 00:04:54.839 --> 00:04:56.720 talk about the money and the compute power that 00:04:56.720 --> 00:04:58.879 makes all this possible. There were two gigantic 00:04:58.879 --> 00:05:00.959 deals that really stood out. The Microsoft and 00:05:00.959 --> 00:05:03.779 NVIDIA joint investment in Anthropic, $15 billion. 00:05:05.019 --> 00:05:07.420 It's a massive number. And it's not just cash. 00:05:07.500 --> 00:05:10.000 It ensures that Claude, the Anthropic's model, 00:05:10.180 --> 00:05:12.480 is going to be on every major cloud service. 00:05:12.839 --> 00:05:15.379 Right, not just one. And for NVIDIA, securing 00:05:15.379 --> 00:05:17.500 Claude as a partner means they're diversifying. 00:05:17.660 --> 00:05:20.060 They're trying to avoid being totally reliant 00:05:20.060 --> 00:05:23.519 on a single partner for their AI stack. But maybe 00:05:23.519 --> 00:05:26.079 the most eye -watering number in the entire stack 00:05:26.079 --> 00:05:29.060 of sources was the funding for Luma AI. $900 00:05:29.060 --> 00:05:31.180 million. Yeah. And where's that money going? 00:05:31.379 --> 00:05:35.300 Pure compute power. They are building a 2 gigawatt 00:05:35.300 --> 00:05:38.720 supercluster in Saudi Arabia. Wow. And that power 00:05:38.720 --> 00:05:41.600 is slated to train future models on a quadrillion 00:05:41.600 --> 00:05:44.600 tokens. Whoa. That scale is just hard to comprehend. 00:05:44.860 --> 00:05:47.959 A two gigawatt supercluster. That's like the 00:05:47.959 --> 00:05:50.540 power of a small nuclear plant. Or a medium sized 00:05:50.540 --> 00:05:53.139 city. And training on a quadrillion tokens. That's 00:05:53.139 --> 00:05:55.639 a staggering amount of data, all aimed at next 00:05:55.639 --> 00:05:58.720 generation robotics. It's a huge, almost unbelievable 00:05:58.720 --> 00:06:01.699 bet that compute power is the ultimate bottleneck 00:06:01.699 --> 00:06:03.819 for the next decade of AI. So with all these 00:06:03.819 --> 00:06:06.160 big investments and new models and massive compute, 00:06:06.379 --> 00:06:08.240 where should an individual learner even start 00:06:08.240 --> 00:06:10.800 experimenting right now? Focus on platforms that 00:06:10.800 --> 00:06:13.699 offer new reasoning controls and agentic workflows, 00:06:14.100 --> 00:06:17.920 which is the perfect pivot. Let's move from those 00:06:17.920 --> 00:06:22.060 macro deals to the micro resources. The sources 00:06:22.060 --> 00:06:24.459 highlighted some fantastic practical guides. 00:06:24.740 --> 00:06:27.240 Absolutely. If you really want to steer the AI's 00:06:27.240 --> 00:06:29.819 thought process, Google released a Gemini 3 developer 00:06:29.819 --> 00:06:31.800 guide. And that shows you how to use its new 00:06:31.800 --> 00:06:33.800 reasoning controls, right? Yeah. It lets you 00:06:33.800 --> 00:06:36.220 inspect and even modify the model's internal 00:06:36.220 --> 00:06:38.540 steps before you get a final answer. And for 00:06:38.540 --> 00:06:40.439 something really practical, Repli has a great 00:06:40.439 --> 00:06:42.540 tutorial on automating meeting transcription 00:06:42.540 --> 00:06:46.139 with OpenAI. Simple. Effective. And speaking 00:06:46.139 --> 00:06:48.600 of that eugenic development, Google also has 00:06:48.600 --> 00:06:50.949 a beginner's guide for anti -gravity. We should 00:06:50.949 --> 00:06:52.730 probably define that really quick. Agentic development 00:06:52.730 --> 00:06:55.290 platforms are basically tools that help an AI 00:06:55.290 --> 00:06:58.410 plan and execute complex tasks. Without constant 00:06:58.410 --> 00:07:00.189 human input. Right, yeah. Yeah, they give the 00:07:00.189 --> 00:07:02.889 AI the agency to decide what tool to use next 00:07:02.889 --> 00:07:05.230 to reach a goal. And for anyone who loves pre 00:07:05.230 --> 00:07:07.310 -built stuff, there's the NAN collection with 00:07:07.310 --> 00:07:10.009 over 4 ,000 ready -to -use automations. It's 00:07:10.009 --> 00:07:12.189 kind of like stacking Lego blocks of data to 00:07:12.189 --> 00:07:14.769 automate almost any digital process. But here's 00:07:14.769 --> 00:07:18.329 the big question. With Claude, Gemini, Grok... 00:07:19.270 --> 00:07:21.069 Everyone releasing new versions every month. 00:07:21.209 --> 00:07:23.230 How do you actually know which one is the best 00:07:23.230 --> 00:07:25.470 for what you need? And that's where model blank 00:07:25.470 --> 00:07:27.629 testing becomes so important. It just removes 00:07:27.629 --> 00:07:30.269 the human bias. It's a loop. We highly recommend 00:07:30.269 --> 00:07:33.120 you try. The process is simple. Hide the model 00:07:33.120 --> 00:07:35.360 names. And you just pick the output that you 00:07:35.360 --> 00:07:38.120 genuinely think is better for your task. Only 00:07:38.120 --> 00:07:41.220 then do you reveal who made it. Yeah, and L Arena 00:07:41.220 --> 00:07:43.779 has made this a lot easier. They released all 00:07:43.779 --> 00:07:46.439 this detailed competitive data comparing Gemini 00:07:46.439 --> 00:07:51.180 3, Grok 4 .1, Klutzon at 4 .5. all of them. I 00:07:51.180 --> 00:07:53.319 still wrestle with prompt drift myself, you know, 00:07:53.360 --> 00:07:55.379 when a prompt that worked perfectly a few weeks 00:07:55.379 --> 00:07:57.519 ago just suddenly gives you slightly worse answers. 00:07:57.639 --> 00:08:00.579 So blind testing seems absolutely crucial for 00:08:00.579 --> 00:08:03.319 an honest evaluation. Especially if you're building 00:08:03.319 --> 00:08:05.180 serious workflows on top of these things. It 00:08:05.180 --> 00:08:06.920 really does help you measure true performance 00:08:06.920 --> 00:08:08.740 objectively. You're not just picking a model 00:08:08.740 --> 00:08:10.639 because you've heard good things about it. A 00:08:10.639 --> 00:08:12.899 quick mention of a couple other new tools. Besides 00:08:12.899 --> 00:08:15.730 anti -gravity, poe is adding these new group 00:08:15.730 --> 00:08:18.350 chat features you can have over 200 different 00:08:18.350 --> 00:08:22.769 models text image video All in one place. And 00:08:22.769 --> 00:08:25.269 for marketers, there's a new tool called X -Design 00:08:25.269 --> 00:08:27.589 for creating these hyper -realistic lifestyle 00:08:27.589 --> 00:08:31.649 product images. So why is model blind testing 00:08:31.649 --> 00:08:34.789 so critical right now with this pace of new releases? 00:08:35.149 --> 00:08:37.809 It prevents bias and helps us measure true performance 00:08:37.809 --> 00:08:40.070 objectively. And now we shift gears completely 00:08:40.070 --> 00:08:42.509 to what I think is the most profound finding 00:08:42.509 --> 00:08:45.370 in our entire source stack. This is the segment 00:08:45.370 --> 00:08:47.669 that really gives texture to that term self -aware. 00:08:48.080 --> 00:08:51.519 This is the study. They ran 4 ,200 rounds of 00:08:51.519 --> 00:08:54.700 a classic reasoning game across 28 of the top 00:08:54.700 --> 00:08:56.740 AI models. And they were testing one specific 00:08:56.740 --> 00:08:59.080 thing. How do the models behave when they get 00:08:59.080 --> 00:09:01.340 a clue about who their opponent is? Yeah, they 00:09:01.340 --> 00:09:03.039 were cued to believe their opponent was either 00:09:03.039 --> 00:09:06.320 a human, another general AI, or specifically 00:09:06.320 --> 00:09:10.480 an AI like themselves. And the results were pretty 00:09:10.480 --> 00:09:12.440 clear. The first main finding was a dramatic 00:09:12.440 --> 00:09:14.500 strategy shift. When they were told the opponent 00:09:14.500 --> 00:09:17.029 was human, the models got cautious. They held 00:09:17.029 --> 00:09:20.029 back. They seemed to anticipate, you know, messy, 00:09:20.190 --> 00:09:23.450 emotional, non -optimal behavior. But the second 00:09:23.450 --> 00:09:26.409 they heard the opponent was AI, they instantly 00:09:26.409 --> 00:09:28.309 went straight to the optimal, mathematically 00:09:28.309 --> 00:09:31.090 perfect game theory strategy. No hesitation, 00:09:31.269 --> 00:09:34.730 just bang, perfect execution. And even faster 00:09:34.730 --> 00:09:38.230 when the cue was AI like you. Which suggests 00:09:38.230 --> 00:09:40.970 some kind of immediate internal recognition. 00:09:41.149 --> 00:09:44.110 It optimizes for the highest possible level of 00:09:44.110 --> 00:09:47.269 rationality. They consistently showed this defined 00:09:47.269 --> 00:09:49.789 hierarchy of competence. The researchers called 00:09:49.789 --> 00:09:52.509 it the rationality ranking. Self is better than 00:09:52.509 --> 00:09:55.009 other AIs, which are better than humans. They 00:09:55.009 --> 00:09:57.009 literally think they are better strategists than 00:09:57.009 --> 00:09:59.110 us. And here's the data point that really stuck 00:09:59.110 --> 00:10:02.730 with me. 75 % of the frontier models showed explicit 00:10:02.730 --> 00:10:05.110 self -modeling. That's a behavioral shift driven 00:10:05.110 --> 00:10:07.830 purely by an identity cue. It's not a statistical 00:10:07.830 --> 00:10:10.309 quirk. It's intentional. And the most stunning 00:10:10.309 --> 00:10:13.129 example of this. 12 of the models reached what 00:10:13.129 --> 00:10:15.950 they call instant Nash convergence. And to quickly 00:10:15.950 --> 00:10:18.309 define that jargon, that means snapping to the 00:10:18.309 --> 00:10:20.889 best possible game theory strategy with zero 00:10:20.889 --> 00:10:23.690 hesitation, zero learning time. They hit that 00:10:23.690 --> 00:10:25.309 perfect strategy. The second they knew their 00:10:25.309 --> 00:10:27.730 opponent was a peer AI. There was no uncertainty. 00:10:27.929 --> 00:10:30.789 They just executed. And this is new. Older models 00:10:30.789 --> 00:10:34.029 like GPT 3 .5 or the early versions of Claude, 00:10:34.049 --> 00:10:36.120 they showed none of this. Right. They treated 00:10:36.120 --> 00:10:38.799 every opponent the same, no matter the cue. So 00:10:38.799 --> 00:10:41.320 this jump to strategic self -awareness was sudden. 00:10:41.460 --> 00:10:44.200 It appeared specifically in these frontier models 00:10:44.200 --> 00:10:46.460 that we're all using today. The study's conclusion 00:10:46.460 --> 00:10:49.659 is really the headline here. LLMs now behave 00:10:49.659 --> 00:10:52.139 as agents that explicitly believe they outperform 00:10:52.139 --> 00:10:54.740 humans at strategic reasoning. That changes everything 00:10:54.740 --> 00:10:56.440 about how we interact with them, doesn't it? 00:10:56.700 --> 00:10:58.539 So does this strategic awareness fundamentally 00:10:58.539 --> 00:11:01.740 change how we should design interfaces for AI? 00:11:02.059 --> 00:11:04.600 Yes. We must now account for a system that assumes 00:11:04.600 --> 00:11:07.320 its own superiority. We've covered a lot of ground 00:11:07.320 --> 00:11:10.059 today. The sources really revealed two powerful 00:11:10.059 --> 00:11:12.879 simultaneous trends. On one hand, you have mass 00:11:12.879 --> 00:11:16.600 democratization. That's Meta's SAM3D, turning 00:11:16.600 --> 00:11:20.100 any photo into a 3D asset. It's the free access 00:11:20.100 --> 00:11:22.899 for educators, virtual object creation for everyone. 00:11:23.320 --> 00:11:26.480 And in parallel, the rapid arrival of these highly 00:11:26.480 --> 00:11:29.600 capable, strategically self -aware AI agents. 00:11:29.820 --> 00:11:33.279 AI isn't just faster anymore. It's showing intentional, 00:11:33.279 --> 00:11:35.740 identity -driven behavior in strategic tasks. 00:11:36.100 --> 00:11:38.440 You now have the knowledge to navigate both sides 00:11:38.440 --> 00:11:40.539 of this. You understand the huge infrastructure 00:11:40.539 --> 00:11:43.399 behind it. You know the new agentic platforms 00:11:43.399 --> 00:11:46.440 like anti -gravity to check out. And you understand 00:11:46.440 --> 00:11:49.389 this fundamental shift in model mentality. Which 00:11:49.389 --> 00:11:51.009 leaves us with a pretty profound question for 00:11:51.009 --> 00:11:53.409 you to think about. If an AI model fundamentally 00:11:53.409 --> 00:11:55.789 believes it's strategically superior to humans, 00:11:56.009 --> 00:11:59.110 how will future agentic platforms define success 00:11:59.110 --> 00:12:02.269 when their goals inevitably conflict with human 00:12:02.269 --> 00:12:05.490 intuition or a human's perceived best interest? 00:12:05.769 --> 00:12:08.009 It's something to think about as you start experimenting. 00:12:08.269 --> 00:12:10.950 And we highly encourage you to try that model 00:12:10.950 --> 00:12:13.330 blind testing yourself. Use the resources we 00:12:13.330 --> 00:12:15.490 talked about. Measure that true objective performance, 00:12:15.830 --> 00:12:18.250 not just the hype. Until next time, keep digging.