WEBVTT 00:00:00.000 --> 00:00:03.540 Imagine a world where your next chat GPT subscription 00:00:03.540 --> 00:00:07.019 doesn't just come as software, but with a robot 00:00:07.019 --> 00:00:11.980 in the box. Or where AI completely rewrites human 00:00:11.980 --> 00:00:15.419 knowledge itself, then trains on its own refined 00:00:15.419 --> 00:00:18.640 version. Wow. Sounds like science fiction, doesn't 00:00:18.640 --> 00:00:20.760 it? Yeah. But, well, parts of it are already 00:00:20.760 --> 00:00:23.140 starting to happen. Welcome to the Deep Dive. 00:00:23.160 --> 00:00:25.620 Today, we're really plunging into this fascinating 00:00:25.620 --> 00:00:28.750 landscape where AI is evolving. I mean, just 00:00:28.750 --> 00:00:31.230 at an almost unbelievable pace. It's moving far 00:00:31.230 --> 00:00:33.890 beyond simple chatbots now. We've gathered a 00:00:33.890 --> 00:00:35.729 bunch of sources, mainly looking at a recent 00:00:35.729 --> 00:00:37.750 AI fire newsletter that was just packed with 00:00:37.750 --> 00:00:40.189 crucial nuggets. Our goal is to pull out the 00:00:40.189 --> 00:00:43.570 most important insights for you. Our mission, 00:00:43.649 --> 00:00:45.549 as always, is to give you that shortcut, a shortcut 00:00:45.549 --> 00:00:48.250 to being truly well -informed. So we'll explore 00:00:48.250 --> 00:00:51.829 the frankly bold visions of Sam Altman and Elon 00:00:51.829 --> 00:00:54.469 Musk. Yeah. We'll deep dive into OpenAI's pretty 00:00:54.469 --> 00:00:57.649 ambitious master plan. Consider how AI is fundamentally 00:00:57.649 --> 00:00:59.850 shifting the job market. Which is a big one. 00:00:59.969 --> 00:01:02.530 It is. And then, something truly surprising, 00:01:02.670 --> 00:01:05.010 we'll unpack a new AI architecture that uses, 00:01:05.170 --> 00:01:08.569 get this, 100 times less compute than giants 00:01:08.569 --> 00:01:11.569 like Lama. 100 times less? It's going to be an 00:01:11.569 --> 00:01:13.469 insightful journey, I think. You know, what we're 00:01:13.469 --> 00:01:16.909 witnessing with AI right now, it... Almost feels 00:01:16.909 --> 00:01:19.370 like just the warm up act. Exactly. Just a taste 00:01:19.370 --> 00:01:21.730 of what's really coming. Yeah. Sam Altman, the 00:01:21.730 --> 00:01:24.530 CEO of OpenAI, he's been dropping some really 00:01:24.530 --> 00:01:27.530 intriguing hints about this next phase. He's 00:01:27.530 --> 00:01:30.189 talking about a future where chat GPT isn't just, 00:01:30.230 --> 00:01:32.390 you know, an interface on your screen. He hinted 00:01:32.390 --> 00:01:34.680 recently, didn't he, about. subscriptions that 00:01:34.680 --> 00:01:37.420 might actually ship with physical robots. You 00:01:37.420 --> 00:01:40.700 did. That's a bold vision, right? It's not just 00:01:40.700 --> 00:01:44.200 software anymore. It's AI having a physical presence 00:01:44.200 --> 00:01:47.180 in our lives. That feels like a significant leap. 00:01:47.319 --> 00:01:49.579 It really does. And then you pivot to Elon Musk. 00:01:49.980 --> 00:01:53.140 His idea for Grok, you know, his ex -AI chatbot, 00:01:53.280 --> 00:01:56.500 it's even more radical. He's talking about deleting 00:01:56.500 --> 00:02:00.019 and rewriting humanity's entire knowledge base. 00:02:00.540 --> 00:02:03.840 And then training Grok on its own refined version 00:02:03.840 --> 00:02:07.480 to purge what he calls garbage. It's kind of 00:02:07.480 --> 00:02:09.259 wild when you think about it, purging garbage. 00:02:09.439 --> 00:02:12.439 The sheer scale of these visions, though, from 00:02:12.439 --> 00:02:14.500 both of them, it's immense. They're pushing us 00:02:14.500 --> 00:02:17.599 beyond just thinking of AI as tools. Right. And 00:02:17.599 --> 00:02:19.680 more into these like pervasive, almost fundamental 00:02:19.680 --> 00:02:22.259 shifts in how we deal with information, even 00:02:22.259 --> 00:02:24.879 the physical world. It totally redefines what 00:02:24.879 --> 00:02:28.710 ubiquitous AI could even mean. These ideas from 00:02:28.710 --> 00:02:31.289 Altman and Musk, they're so huge, they almost 00:02:31.289 --> 00:02:34.169 sound like pure sci -fi. Yeah. Is there a realistic 00:02:34.169 --> 00:02:37.229 path to any of this, do you think? I don't think 00:02:37.229 --> 00:02:39.330 it's just sci -fi. It's really about relentlessly 00:02:39.330 --> 00:02:42.550 pushing boundaries and scale. Right, pushing 00:02:42.550 --> 00:02:44.289 boundaries and scale. Okay. Yeah. So speaking 00:02:44.289 --> 00:02:47.490 of ambitious, let's unpack OpenAI's master plan. 00:02:47.629 --> 00:02:49.789 Yeah. Altman recently pulled back the curtain 00:02:49.789 --> 00:02:52.069 a bit on how they went from like literally eight 00:02:52.069 --> 00:02:54.430 people around a whiteboard. Eight people? To 00:02:54.430 --> 00:02:56.110 one of the biggest sites on the internet. Yeah. 00:02:56.349 --> 00:02:59.569 He talks about this AGI or bust conviction. And 00:02:59.569 --> 00:03:02.250 their talent hack, as he put it, was frankly 00:03:02.250 --> 00:03:05.550 brilliant. By aiming for that truly one -of -one 00:03:05.550 --> 00:03:08.770 mission achieving AGI, they basically vacuum 00:03:08.770 --> 00:03:11.889 sealed the smartest 1 % of people. People who 00:03:11.889 --> 00:03:14.210 genuinely cared about that single, incredibly 00:03:14.210 --> 00:03:17.310 difficult goal. That's how you go from... zero 00:03:17.310 --> 00:03:21.370 revenue for years to just rocketing past every 00:03:21.370 --> 00:03:24.430 consumer app launch record ever with ChatGPT. 00:03:24.590 --> 00:03:27.729 It's an almost like cult -like focus on the mission. 00:03:27.930 --> 00:03:30.210 What's fascinating now is what Altman calls the 00:03:30.210 --> 00:03:33.930 product overhang. He says models like GPT -40 00:03:33.930 --> 00:03:36.669 mini and O3 are actually far more capable than 00:03:36.669 --> 00:03:39.789 what today's products expose. API prices are 00:03:39.789 --> 00:03:42.590 collapsing fast and he hinted a powerful open 00:03:42.590 --> 00:03:45.240 source model is coming soon. Said one. Better 00:03:45.240 --> 00:03:46.979 than you're hoping for. Okay. So this product 00:03:46.979 --> 00:03:49.979 overhang, it basically means the AI models themselves 00:03:49.979 --> 00:03:53.039 are much smarter than what we, the users, currently 00:03:53.039 --> 00:03:55.460 get to experience through the apps. Yeah. It's 00:03:55.460 --> 00:03:57.080 like having a supercar but only being allowed 00:03:57.080 --> 00:03:58.919 to drive it in a school zone. That's a great 00:03:58.919 --> 00:04:00.819 analogy. Yeah. Seriously. It's like we're stuck 00:04:00.819 --> 00:04:03.460 in the 1960s of AI product design or something. 00:04:03.520 --> 00:04:06.240 We're stacking these incredibly powerful Lego 00:04:06.240 --> 00:04:08.639 blocks of data, right, these foundational models. 00:04:08.840 --> 00:04:11.860 But we haven't truly built the castle yet. You 00:04:11.860 --> 00:04:14.199 know, the full transformative applications. There's 00:04:14.199 --> 00:04:17.019 just so much untapped potential sitting there 00:04:17.019 --> 00:04:20.680 waiting for the right builders. He also pinpointed 00:04:20.680 --> 00:04:24.160 memory as his favorite 2024 feature. The vision 00:04:24.160 --> 00:04:27.240 here is an always -on AI agent that truly learns 00:04:27.240 --> 00:04:29.800 you. Learns you. It hooks into all your data, 00:04:29.860 --> 00:04:33.120 supposedly, and can proactively act on your behalf 00:04:33.120 --> 00:04:35.949 without spamming you with notifications. Imagine 00:04:35.949 --> 00:04:38.870 an AI that genuinely anticipates your needs. 00:04:39.329 --> 00:04:41.829 That feels like a game changer for personal productivity. 00:04:42.149 --> 00:04:43.810 And here's where it gets really interesting, 00:04:43.889 --> 00:04:46.550 I think. He said hardware is on the table. Mentioned 00:04:46.550 --> 00:04:48.970 a new device coming, co -designed with Joni Ive, 00:04:49.050 --> 00:04:51.389 specifically engineered to make the interface 00:04:51.389 --> 00:04:54.389 just melt away. Melt away. Yeah. The goal is 00:04:54.389 --> 00:04:56.370 for the AI to become seamless, almost invisible 00:04:56.370 --> 00:04:58.910 in your life. No more apps, no more screens necessarily, 00:04:59.170 --> 00:05:02.610 just assistance, pure help. Which brings us right 00:05:02.610 --> 00:05:04.529 back to that initial moment of wonder, doesn't 00:05:04.529 --> 00:05:07.319 it? Long term, Altman genuinely seems to believe 00:05:07.319 --> 00:05:10.240 the high tier chat GPT subscription might actually 00:05:10.240 --> 00:05:13.160 ship with a humanoid robot in the box. Imagine 00:05:13.160 --> 00:05:16.220 that a physical AI companion delivered right 00:05:16.220 --> 00:05:18.959 to your doorstep. I mean, that shifts the entire 00:05:18.959 --> 00:05:21.879 paradigm of human computer interaction. It absolutely 00:05:21.879 --> 00:05:24.120 does. But, you know, there are significant limiting 00:05:24.120 --> 00:05:26.220 factors here. Altman is very clear about this. 00:05:26.279 --> 00:05:29.160 Yeah. Compute and energy are the new oil. He 00:05:29.160 --> 00:05:32.470 basically said that. Open AI is sprinting to 00:05:32.470 --> 00:05:34.410 build the largest, most expensive infrastructure 00:05:34.410 --> 00:05:38.230 in the world. They openly wish our local devices, 00:05:38.430 --> 00:05:40.709 you know, our phones and laptops, would shoulder 00:05:40.709 --> 00:05:44.350 half the workload. The real bottleneck soon won't 00:05:44.350 --> 00:05:46.810 be the algorithms. It's going to be energy. He 00:05:46.810 --> 00:05:49.230 even floated the idea of parking data centers 00:05:49.230 --> 00:05:52.970 in orbit. Orbit. OK. So why does all this matter 00:05:52.970 --> 00:05:56.290 for you listening right now? ground shaking, 00:05:56.550 --> 00:05:58.269 as Altman puts it. I mean, startups especially 00:05:58.269 --> 00:06:01.449 can win big, but they need to stop just cloning 00:06:01.449 --> 00:06:04.350 Chad GPT. Right. They need to find those greenfield 00:06:04.350 --> 00:06:07.069 opportunities, those untouched spaces to really 00:06:07.069 --> 00:06:09.649 differentiate themselves. And defensibility in 00:06:09.649 --> 00:06:12.509 this new era. It's not just about who has access 00:06:12.509 --> 00:06:15.189 to the best model anymore. It's memory. It's 00:06:15.189 --> 00:06:18.029 deep data integrations. And maybe most importantly, 00:06:18.189 --> 00:06:21.250 user trust. Trust. Yeah. Those are the new moats, 00:06:21.310 --> 00:06:23.350 you know, the things that will protect businesses. 00:06:23.730 --> 00:06:25.730 It's not just about building a better model. 00:06:25.810 --> 00:06:27.949 It's about building a trusted relationship with 00:06:27.949 --> 00:06:30.670 the user. It paints this picture of what he calls 00:06:30.670 --> 00:06:33.899 one person leverage. In Altman's 10 -year view, 00:06:34.220 --> 00:06:37.579 a tiny team armed with reasoning agents and abundant 00:06:37.579 --> 00:06:41.720 energy could create an unimaginable superintelligence 00:06:41.720 --> 00:06:45.000 right at human fingertips. Wow. It's not just 00:06:45.000 --> 00:06:48.220 efficiency. It's fundamentally shifting what 00:06:48.220 --> 00:06:50.100 a small group of people can actually achieve. 00:06:50.379 --> 00:06:54.250 So we're talking humanoid robots. Seamless AI 00:06:54.250 --> 00:06:56.589 agents. It sounds incredible. But what's the 00:06:56.589 --> 00:06:58.569 practical limiter here? What's really holding 00:06:58.569 --> 00:07:00.529 us back from having these kinds of companions 00:07:00.529 --> 00:07:03.189 today? Pure compute and energy scale. That's 00:07:03.189 --> 00:07:05.649 the bottom line. Compute and energy. Okay. So 00:07:05.649 --> 00:07:08.769 from OpenAI's internal master plan, let's maybe 00:07:08.769 --> 00:07:11.069 zoom out a bit. Let's see how these visions resonate 00:07:11.069 --> 00:07:14.189 or maybe clash with other major players and thinkers 00:07:14.189 --> 00:07:16.689 in the AI space. Pulling more insights from the 00:07:16.689 --> 00:07:18.730 newsletter. Right. You mentioned Elon Musk's 00:07:18.730 --> 00:07:21.410 desire for Grok to rewrite the entire corpus 00:07:21.410 --> 00:07:23.800 of human knowledge and... purge far too much 00:07:23.800 --> 00:07:27.779 garbage. That's such a powerful, almost godlike 00:07:27.779 --> 00:07:30.180 ambition, isn't it? But that phrase, purging 00:07:30.180 --> 00:07:32.939 garbage, I mean, who defines what's garbage? 00:07:33.139 --> 00:07:35.060 Exactly. That's the question. It opens up a whole 00:07:35.060 --> 00:07:38.139 pan of worms about bias, about what gets included 00:07:38.139 --> 00:07:41.500 or excluded. It's a total double -edged sword. 00:07:42.019 --> 00:07:45.079 On one hand, yeah, the idea of a truly curated, 00:07:45.319 --> 00:07:48.220 maybe unbiased knowledge base sounds appealing. 00:07:48.600 --> 00:07:52.240 But on the other... The potential for one entity 00:07:52.240 --> 00:07:55.759 to dictate truth? That's profoundly unsettling. 00:07:55.800 --> 00:07:57.920 It's a fascinating, almost Orwellian thought 00:07:57.920 --> 00:07:59.759 experiment, really. Then you have someone like 00:07:59.759 --> 00:08:02.259 Reid Hoffman. His view is quite different. He 00:08:02.259 --> 00:08:05.699 says AI will transform jobs, yeah, but not cause 00:08:05.699 --> 00:08:08.079 some white -collar bloodbath. Okay, that's more 00:08:08.079 --> 00:08:10.740 optimistic. Much more. He sees it as person plus 00:08:10.740 --> 00:08:13.620 AI, like an Excel for everything, augmenting 00:08:13.620 --> 00:08:15.480 human capabilities rather than just replacing 00:08:15.480 --> 00:08:17.560 them outright. You know, I still wrestle with 00:08:17.560 --> 00:08:19.620 prompt drift myself sometimes, so I completely 00:08:19.620 --> 00:08:22.759 get that person plus AI concept. It's not about 00:08:22.759 --> 00:08:24.620 the machines doing everything for us. It's about 00:08:24.620 --> 00:08:26.980 a partnership where AI helps us be more efficient, 00:08:27.060 --> 00:08:29.439 maybe more capable, just taking friction out 00:08:29.439 --> 00:08:31.569 of our workflows. But not everyone shares that 00:08:31.569 --> 00:08:34.610 level of optimism. Comic legend Paul Pope, for 00:08:34.610 --> 00:08:37.950 example, he isn't losing sleep over AI plagiarism. 00:08:38.330 --> 00:08:41.210 His worries are more fundamental. Like what? 00:08:41.370 --> 00:08:44.029 Killer robots and mass surveillance. A very different, 00:08:44.110 --> 00:08:46.789 maybe more existential kind of fear brewing there. 00:08:46.970 --> 00:08:50.389 Huh. Yeah. Different focus entirely. Meanwhile, 00:08:50.690 --> 00:08:52.850 you see companies like Amazon consolidating things. 00:08:53.049 --> 00:08:55.700 They're bundling QuickSight. Q business and Q 00:08:55.700 --> 00:08:58.659 apps into a single Q business suite workspace. 00:08:59.100 --> 00:09:02.519 Right. So the Q business suite is basically Amazon's 00:09:02.519 --> 00:09:05.519 combined toolkit for managing data and automating 00:09:05.519 --> 00:09:07.580 workflows from one place. Exactly. It's all about 00:09:07.580 --> 00:09:10.259 streamlining business operations, making it easier 00:09:10.259 --> 00:09:13.120 for companies to actually adopt and use AI effectively. 00:09:13.460 --> 00:09:15.679 And there is that fascinating anecdote from Jeffrey 00:09:15.679 --> 00:09:17.840 Hinton, you know, often called the godfather 00:09:17.840 --> 00:09:20.679 of AI. Oh, yeah. He says Google throttled their 00:09:20.679 --> 00:09:22.480 early chatbots because they were worried about 00:09:22.480 --> 00:09:25.259 reputational risk. Understandable for Google. 00:09:25.440 --> 00:09:28.659 But OpenAI back then had nothing to lose, so 00:09:28.659 --> 00:09:30.860 they just pushed ahead. And apparently when Hinton 00:09:30.860 --> 00:09:33.600 was asked about Sam Altman's moral compass, he 00:09:33.600 --> 00:09:35.899 just shrugged and said, we'll see. We'll see. 00:09:35.980 --> 00:09:38.820 Wow. That's very telling, isn't it? It really 00:09:38.820 --> 00:09:42.159 is. And you also had Mistral AI CEO Arthur Mensch 00:09:42.159 --> 00:09:44.539 raising a really good point about de -skilling. 00:09:44.820 --> 00:09:47.700 De -skilling, meaning AI reducing the need for 00:09:47.700 --> 00:09:50.529 human skills. Yeah, precisely. He argues that's 00:09:50.529 --> 00:09:52.570 the real danger, maybe more than just white collar 00:09:52.570 --> 00:09:55.649 job losses. His concern is we need to keep humans 00:09:55.649 --> 00:09:58.549 in the loop. Otherwise, we risk getting lazy 00:09:58.549 --> 00:10:00.889 minds or losing critical faculties over time. 00:10:01.370 --> 00:10:03.750 That's a concern that resonates, I think. Absolutely. 00:10:04.110 --> 00:10:07.190 So given all these different perspectives from 00:10:07.190 --> 00:10:11.629 Altman's robots to Hoffman's optimism to Hinton's 00:10:11.629 --> 00:10:14.450 caution, what's the immediate impact? How does 00:10:14.450 --> 00:10:16.570 this affect how people work like today or tomorrow? 00:10:17.080 --> 00:10:19.500 I think the consensus leans towards jobs will 00:10:19.500 --> 00:10:22.360 transform, not vanish. Think person plus AI. 00:10:22.600 --> 00:10:25.480 Person plus AI. Okay. Now let's shift gears a 00:10:25.480 --> 00:10:27.879 bit and talk about the bleeding edge. New tools 00:10:27.879 --> 00:10:29.940 and especially these architectural breakthroughs 00:10:29.940 --> 00:10:31.659 that are pushing boundaries beyond just making 00:10:31.659 --> 00:10:33.580 bigger and bigger models. Yeah, this is exciting 00:10:33.580 --> 00:10:35.980 stuff. The newsletter highlighted tools like 00:10:35.980 --> 00:10:38.399 FetchEye, bringing professional AI capabilities 00:10:38.399 --> 00:10:42.200 right onto any Mac app, and Julia's AI Notebooks, 00:10:42.279 --> 00:10:44.940 letting you basically chat with complex data 00:10:44.940 --> 00:10:47.799 to whip up expert insights. These are practical 00:10:47.799 --> 00:10:50.139 applications hitting the market now. Right, tangible 00:10:50.139 --> 00:10:52.399 tools. But then we get to the really deep technical 00:10:52.399 --> 00:10:57.250 dive, Meta's new AUNet. This is, well... It's 00:10:57.250 --> 00:10:59.490 potentially revolutionary. It sounds like a huge 00:10:59.490 --> 00:11:01.509 deal because it's showing incredible efficiency. 00:11:01.809 --> 00:11:04.309 The headline was something like, bite -level 00:11:04.309 --> 00:11:08.149 AUNet chases llama on 1 % of the compute. Exactly. 00:11:08.149 --> 00:11:10.950 So AUNet, to put it simply, it's Meta's new AI 00:11:10.950 --> 00:11:13.830 model architecture. It's designed for incredibly 00:11:13.830 --> 00:11:16.529 efficient language processing. It's a fundamentally 00:11:16.529 --> 00:11:18.769 different way of building these models. Okay, 00:11:18.789 --> 00:11:20.679 different how? Why does it matter so much? Well, 00:11:20.740 --> 00:11:22.799 first, it learns its own way to process language 00:11:22.799 --> 00:11:26.000 on the fly. That means no fixed BPE vocab is 00:11:26.000 --> 00:11:29.059 needed. And a BPE vocab is? That's the standard 00:11:29.059 --> 00:11:31.679 way AI models usually break text down into smaller 00:11:31.679 --> 00:11:33.879 pieces, like words or parts of words, so they 00:11:33.879 --> 00:11:36.159 can understand it. AUNet doesn't need that predefined 00:11:36.159 --> 00:11:38.960 dictionary. Okay. And the second thing? Second, 00:11:39.120 --> 00:11:42.179 it performs like... neck and neck with traditional 00:11:42.179 --> 00:11:45.259 BPE transformers, the standard model type, but 00:11:45.259 --> 00:11:49.159 it burns about 100 times less FLOPs than a massive 00:11:49.159 --> 00:11:54.019 model like LAMA 3 .18b. 100 times, wow. And FLOPs 00:11:54.019 --> 00:11:57.100 are just... Think of FLOPs as the sheer amount 00:11:57.100 --> 00:12:00.080 of computational muscle an AI model needs to 00:12:00.080 --> 00:12:04.240 flex to do its job. So way, way less muscle needed 00:12:04.240 --> 00:12:07.039 here. That is astounding. Efficiency over just 00:12:07.039 --> 00:12:09.840 brute force. It really hints at a coming wave 00:12:09.840 --> 00:12:12.539 of... architectural gains in AI, doesn't it? 00:12:12.559 --> 00:12:14.159 Exactly. Not just throwing bigger and bigger 00:12:14.159 --> 00:12:16.659 GPU clusters at the problem, consuming more energy. 00:12:16.799 --> 00:12:18.879 We're actually getting smarter about how we build 00:12:18.879 --> 00:12:21.320 these models. Smarter, not just bigger. It suggests 00:12:21.320 --> 00:12:23.960 a future where really powerful AI might run on 00:12:23.960 --> 00:12:26.340 much less hardware, maybe even locally more often. 00:12:26.559 --> 00:12:29.740 And there's this cross -lingual freebie, they 00:12:29.740 --> 00:12:31.960 called it, because it works at the byte level, 00:12:32.080 --> 00:12:35.220 like the raw letters and symbols, low resource 00:12:35.220 --> 00:12:37.240 languages get a performance bump automatically. 00:12:37.519 --> 00:12:39.720 Languages that don't have much digital data available. 00:12:39.980 --> 00:12:42.039 for training. Ah, so it helps level the playing 00:12:42.039 --> 00:12:43.720 field for different languages without needing 00:12:43.720 --> 00:12:46.200 tons of extra data. Pretty much. Which is huge 00:12:46.200 --> 00:12:48.679 for global accessibility, for making AI truly 00:12:48.679 --> 00:12:50.720 useful for everyone, not just those speaking 00:12:50.720 --> 00:12:52.919 dominant languages. The big takeaway here seems 00:12:52.919 --> 00:12:55.500 to be that the tokenizer might be the next thing 00:12:55.500 --> 00:12:58.000 to disappear. Yeah. That tokenizer, the part 00:12:58.000 --> 00:13:00.139 of the AI that converts words into numbers for 00:13:00.139 --> 00:13:02.860 the model, this new approach could make that 00:13:02.860 --> 00:13:06.440 complex, often problematic, pre -processing step 00:13:06.440 --> 00:13:10.519 just obsolete. Imagine AI just seeing raw data, 00:13:10.700 --> 00:13:13.320 letters, symbols, everything. Could this new 00:13:13.320 --> 00:13:15.879 architecture, this AUNet, fundamentally change 00:13:15.879 --> 00:13:18.620 how we build AI from the ground up? Is it that 00:13:18.620 --> 00:13:21.179 significant? Yes. I think it really hints at 00:13:21.179 --> 00:13:23.919 a new era of efficiency and much more dynamic 00:13:23.919 --> 00:13:26.039 language handling. Efficiency and dynamic handling. 00:13:26.139 --> 00:13:29.120 Got it. Midroll sponsor Reed Placeholder. Okay, 00:13:29.179 --> 00:13:31.179 let's just take a moment to recap the core themes 00:13:31.179 --> 00:13:33.639 from this Deep Dives. We've explored these radical, 00:13:33.879 --> 00:13:36.860 almost sci -fi visions for AI's future, coming 00:13:36.860 --> 00:13:40.139 from leaders like Sam Altman and Elon Musk. Robots 00:13:40.139 --> 00:13:42.759 and knowledge rewriting. Exactly. We saw OpenAI's 00:13:42.759 --> 00:13:46.960 strategic push towards pervasive AI agents and 00:13:46.960 --> 00:13:49.379 even physical hardware like robots, hinting at 00:13:49.379 --> 00:13:51.879 a whole new era of interaction. Then we broadened 00:13:51.879 --> 00:13:54.460 the view. Looking at the ongoing transformation 00:13:54.460 --> 00:13:57.600 of jobs and skills, weighing those different 00:13:57.600 --> 00:14:00.740 perspectives, optimism versus, you know, some 00:14:00.740 --> 00:14:02.559 real concerns about de -skilling or control. 00:14:02.820 --> 00:14:05.019 Right. And finally, we just touched on those 00:14:05.019 --> 00:14:08.080 incredibly exciting new architectural breakthroughs 00:14:08.080 --> 00:14:10.779 like Meta's AUNet that are making AI potentially 00:14:10.779 --> 00:14:13.220 far, far more efficient. Yeah, the efficiency 00:14:13.220 --> 00:14:15.899 angle is key. It really feels like AI is moving 00:14:15.899 --> 00:14:18.820 rabidly beyond its warm -up act phase now. We're 00:14:18.820 --> 00:14:20.559 heading into something much more integrated, 00:14:20.740 --> 00:14:24.480 more powerful and... maybe surprisingly, sometimes 00:14:24.480 --> 00:14:26.779 achieved with less brute force. It's not just 00:14:26.779 --> 00:14:28.860 about bigger models anymore, it seems. It's about 00:14:28.860 --> 00:14:31.779 smarter models, smarter architectures, and how 00:14:31.779 --> 00:14:33.740 they're all going to reshape our daily lives. 00:14:33.980 --> 00:14:36.879 So as you go about your day, maybe consider how 00:14:36.879 --> 00:14:38.919 these shifts might impact your own work or even 00:14:38.919 --> 00:14:41.419 just your daily life. Are you ready for an AI 00:14:41.419 --> 00:14:44.710 that learns you, that acts proactively? And here's 00:14:44.710 --> 00:14:46.350 maybe a final thought for you to mull over, building 00:14:46.350 --> 00:14:49.710 on that AUNet idea. Consider how a byte -level 00:14:49.710 --> 00:14:52.590 understanding of language could change AI's ability 00:14:52.590 --> 00:14:56.009 to truly understand diverse data. Think about 00:14:56.009 --> 00:14:59.690 it. Code, emoji. maybe even nuanced inflections 00:14:59.690 --> 00:15:02.590 in voice someday, all without needing humans 00:15:02.590 --> 00:15:06.149 to pre -process it perfectly first. What new, 00:15:06.309 --> 00:15:08.710 maybe entirely unexpected applications could 00:15:08.710 --> 00:15:11.190 that unlock for all of us? It's a fascinating 00:15:11.190 --> 00:15:13.330 future unfolding right in front of us. Definitely 00:15:13.330 --> 00:15:15.710 is. Thank you for joining us on this deep dive 00:15:15.710 --> 00:15:17.870 today. Yeah, thanks for listening. If you found 00:15:17.870 --> 00:15:19.509 this insightful, maybe share it with someone 00:15:19.509 --> 00:15:21.090 else who's trying to keep up with everything 00:15:21.090 --> 00:15:22.970 happening in AI. We'll see you next time.