WEBVTT 00:00:00.000 --> 00:00:05.219 Imagine having a team of brilliant experts. Not 00:00:05.219 --> 00:00:07.360 just a chatbot giving facts, but something that 00:00:07.360 --> 00:00:12.119 really thinks, plans, acts for you. Instantly, 00:00:12.140 --> 00:00:14.199 they're in your pocket. This isn't sci -fi anymore. 00:00:14.339 --> 00:00:16.500 It's actually here. It really feels that way. 00:00:16.800 --> 00:00:19.160 Welcome to the Deep Dive. Today, we're plunging 00:00:19.160 --> 00:00:21.660 into what feels like, well, a huge moment in 00:00:21.660 --> 00:00:24.719 AI. Yeah, the GPT -5 launch. Exactly. OpenAI's 00:00:24.719 --> 00:00:28.800 GPT -5. Our mission is to unpack the pretty bold 00:00:28.800 --> 00:00:32.420 claims. Which are surprising. Definitely surprising. 00:00:32.640 --> 00:00:35.939 And what this new unified model really means 00:00:35.939 --> 00:00:38.759 for you. We'll cover its capabilities, its immediate 00:00:38.759 --> 00:01:35.310 impact, and then look at other big A. So it's 00:01:35.310 --> 00:01:37.170 not just about being faster then. It's about 00:01:37.170 --> 00:01:39.310 a different kind of intelligence working together. 00:01:39.590 --> 00:01:41.829 I think that's the idea. It's a significant leap, 00:01:41.930 --> 00:02:08.979 yeah. It allows for way more... Okay, that's 00:02:08.979 --> 00:02:10.639 different. And there's this real -time routing 00:02:10.639 --> 00:02:13.860 thing, too. The model decides for itself. Quick 00:02:13.860 --> 00:02:16.199 answer needed? Or does this need deeper thought? 00:02:16.360 --> 00:02:19.180 And it just does it. Yeah, no user input needed 00:02:19.180 --> 00:02:21.620 for that switch. That's why OpenAI is calling 00:02:21.620 --> 00:02:24.300 it an agent over chatbot. Agent over chatbot. 00:02:24.539 --> 00:02:27.900 Their pitch is basically a team of PhD -level 00:02:27.900 --> 00:02:30.340 experts in your pocket. That's quite the image. 00:02:30.759 --> 00:02:33.180 And they're saying it's the best model in the 00:02:33.180 --> 00:02:35.599 world. Do the numbers back that up? What about 00:02:35.599 --> 00:02:37.479 those benchmarks? The numbers are pretty strong, 00:02:37.560 --> 00:02:41.419 yeah. Coding, for example, 74 .9 % accuracy. 00:02:41.620 --> 00:02:44.379 Okay, 74 .9. How's that compare? It just edges 00:02:44.379 --> 00:02:48.139 out Claude Opus 4 .1, which was at 74 .5%. Very 00:02:48.139 --> 00:02:50.740 close. Very close. Yeah. But it really pulls 00:02:50.740 --> 00:02:53.500 ahead of Google's Gemini 2 .5 Pro, which was 00:02:53.500 --> 00:02:57.969 59. That's a big gap. What about other areas 00:02:57.969 --> 00:03:01.490 like science? Science Q &A, it hit 89 .4%. That 00:03:01.490 --> 00:03:05.430 beats Claude Opus 4 .1 at 80 .9%. And even Grok 00:03:05.430 --> 00:03:09.030 4 Heavy at 88 .9%. Impressive. But I saw something 00:03:09.030 --> 00:03:11.379 about healthcare accuracy. That seemed really 00:03:11.379 --> 00:03:14.439 striking. Yes, that's maybe the biggest deal. 00:03:14.500 --> 00:03:16.599 The hallucination rate, you know, when the AI 00:03:16.599 --> 00:03:18.819 just makes stuff up. Yeah, always a worry. It 00:03:18.819 --> 00:03:23.120 dropped to just 1 .6%. 1 .6%, seriously? Seriously. 00:03:23.479 --> 00:03:28.000 Compare that to GPT -40's 12 .9 % or 03's 15 00:03:28.000 --> 00:03:31.879 .8%. Massive reduction. Wow. What does that mean, 00:03:31.919 --> 00:03:34.580 practically? Well, it's not really for solo diagnosis, 00:03:34.780 --> 00:03:37.460 obviously. But for things like summarizing research 00:03:37.460 --> 00:03:40.099 papers for a doctor, much, much more reliable. 00:03:49.120 --> 00:03:52.259 Humanity's last exam. Super hard reasoning benchmark. 00:03:52.479 --> 00:03:56.340 OK. GPT -5 scored 42 percent, which is just slightly 00:03:56.340 --> 00:03:59.699 under Grok for heavy is 44 .4 percent. So it's. 00:04:20.560 --> 00:04:22.660 700 million people? That's almost 10 % of the 00:04:22.660 --> 00:04:24.740 planet. Getting access overnight? Instantly. 00:04:24.899 --> 00:04:28.199 It's a massive strategic play by OpenAI. Keep 00:04:28.199 --> 00:04:30.879 users in their ecosystem. You know, FightOff, 00:04:31.019 --> 00:04:34.560 Anthropic, XAI, Google. Makes sense. Lock them 00:04:34.560 --> 00:04:36.699 in with the best stuff. Exactly. It's about market 00:04:36.699 --> 00:04:39.240 share and keeping users engaged. You know, it's 00:04:39.240 --> 00:04:41.459 funny. I still wrestle with confidence myself 00:04:41.459 --> 00:04:45.079 sometimes. Getting the AI to consistently understand 00:04:45.079 --> 00:04:47.100 what I mean, especially over several turns on 00:04:47.100 --> 00:04:50.800 a complex task. This is sort of the idea of a 00:04:50.800 --> 00:04:53.339 model that just gets it for those multi -step 00:04:53.339 --> 00:04:55.540 things without me having to constantly tweak 00:04:55.540 --> 00:04:57.899 and rephrase. Right. Less babysitting. Yeah. 00:04:58.139 --> 00:05:00.699 That feels like a genuine game changer for actually 00:05:00.699 --> 00:05:02.939 using it effectively day to day. Definitely. 00:05:03.100 --> 00:05:33.079 So boiling it down. Yeah, the rivalries are heating 00:05:33.079 --> 00:05:36.060 up, aren't they? Like, hours after GPC 5 dropped, 00:05:36.279 --> 00:05:38.939 Elon Musk claimed his Rock 4 Heavy was already 00:05:38.939 --> 00:05:40.899 smarter. Right, said it was smarter two weeks 00:05:40.899 --> 00:05:43.379 before the launch, though, you know, no published 00:05:43.379 --> 00:05:45.720 benchmarks to back that up yet. Still, it shows 00:05:45.720 --> 00:05:49.519 that intense personal and tech rivalry. And speaking 00:05:49.519 --> 00:05:52.699 of competition, OpenAI affordably spent, supposedly, 00:05:52.839 --> 00:05:56.439 $1 .5 million per employee. Yeah, for about 1 00:05:56.439 --> 00:05:59.639 ,000 staff. Over $1 .5 billion total right before 00:05:59.639 --> 00:06:01.519 the launch. Just to keep talent from jumping 00:06:01.519 --> 00:06:04.000 ship. Seems like it. It's a war for talent out 00:06:04.000 --> 00:06:06.519 there. Huge money flying around. And huge... 00:06:35.240 --> 00:06:38.139 They just raised $100 million and a $3 .1 billion 00:06:38.139 --> 00:06:41.120 valuation. Wow, that's a steep climb. Yeah, a 00:06:41.120 --> 00:06:43.759 6x increase in less than a year. And get this, 00:06:43.920 --> 00:06:47.240 they've been profitable since day one. Profitable 00:06:47.240 --> 00:06:51.040 in AI research, that's rare. Very. And they massively 00:06:51.040 --> 00:06:53.439 cut the cost of using their video models from 00:06:53.439 --> 00:06:56.379 like $1 ,400 an hour down to $0 .25 an hour. 00:06:56.420 --> 00:07:00.199 Okay, wait, $1 ,400 down to $0 .25? Yeah, that's 00:07:00.199 --> 00:07:02.000 completely disruptive for anyone working with 00:07:02.000 --> 00:07:03.970 video generation. He makes it accessible. That 00:07:03.970 --> 00:07:07.129 cost reduction alone is revolutionary. And on 00:07:07.129 --> 00:07:09.589 a more relatable note, there is this funny thing 00:07:09.589 --> 00:07:34.230 on Reddit. Yeah. A senior D. And from people 00:07:34.230 --> 00:07:36.589 wanting to learn this stuff. Andrew Ng, you know, 00:07:36.610 --> 00:07:39.449 the AI pioneer. Sure. He announced a big free 00:07:39.449 --> 00:07:42.389 course on Cloud Code partnering with Anthropic. 00:07:42.569 --> 00:07:45.250 Free course from Andrew Ng, nice. Yeah, focused 00:07:45.250 --> 00:07:48.149 on what they call agentic AI, AI that acts with 00:07:48.149 --> 00:07:50.449 more autonomy. Great opportunity to learn from 00:07:50.449 --> 00:07:52.670 one of the best. Okay, so when you put all these 00:07:52.670 --> 00:07:54.829 things together, the rivalries, the new tools 00:07:54.829 --> 00:07:57.790 like 1, 2 .2, the cost drops from Dakar, even 00:07:57.790 --> 00:08:01.990 the funny explanations and the courses. Picture 00:08:01.990 --> 00:08:05.310 the pain of the AI landscape right now. It's 00:08:05.310 --> 00:08:07.910 intense competition driving incredibly rapid 00:08:07.910 --> 00:08:10.750 innovation, leading to practical, accessible 00:08:10.750 --> 00:08:13.329 applications emerging almost constantly. Constant 00:08:13.329 --> 00:08:36.519 mode. and they're already weaving GPT -5 into 00:08:36.519 --> 00:08:39.120 basically everything. Yeah, consumer stuff, developer 00:08:39.120 --> 00:08:42.220 tools, corporate offerings. They're moving fast. 00:08:42.639 --> 00:08:44.820 And Duolingo. Remember the backlash when they 00:08:44.820 --> 00:08:47.960 went AI first? Well, it seems like that's tied 00:08:47.960 --> 00:08:49.679 down. One release isn't hurting their momentum 00:08:49.679 --> 00:08:52.279 much. Interesting. People adapt quickly, maybe? 00:08:52.559 --> 00:08:55.399 Maybe. Okay, here's one that's a bit out there. 00:08:55.559 --> 00:09:00.659 Oh. Go on. Oh. Imagine an AI system discovering 00:09:00.659 --> 00:09:09.740 a new kind of physics. Yeah, that expands what 00:09:09.740 --> 00:09:12.820 discovery even means. Wow. That idea of AI discovering 00:09:12.820 --> 00:09:15.720 new physics, it really is profound. Kind of humbling. 00:09:16.059 --> 00:09:40.799 Absolutely. Also, a quick one. Adapt their reasoning 00:09:40.799 --> 00:09:43.600 in real time. Adapt their reasoning, meaning? 00:09:43.919 --> 00:09:46.440 Meaning they move beyond just using their static 00:09:46.440 --> 00:09:49.200 pre -trained knowledge. They're not just reciting 00:09:49.200 --> 00:09:51.700 facts they learned. Okay, so it's more dynamic 00:09:51.700 --> 00:09:54.700 than just recalling information. Exactly. And 00:09:54.700 --> 00:09:57.460 the results are impressive. It tripled GPT -4 00:09:57.460 --> 00:10:00.679 .1's accuracy on tough biomedical questions. 00:10:01.080 --> 00:10:03.899 Tripled it? Yep. And it even outperformed O3, 00:10:04.120 --> 00:10:06.519 which is already a top reasoning model. How does 00:10:06.519 --> 00:10:08.799 it do that? It generates these internal feedback 00:10:08.799 --> 00:10:13.080 loops while it's working. It can manage its own 00:10:13.080 --> 00:10:16.000 memory, recognize when it's uncertain about its 00:10:16.000 --> 00:10:18.519 own reasoning, and then this is the key part, 00:10:18.679 --> 00:10:21.480 revise its answers mid -process. So it's like 00:10:21.480 --> 00:10:23.440 it's thinking out loud to itself and correcting 00:10:23.440 --> 00:10:25.539 its own path. That's a great way to put it, a 00:10:25.539 --> 00:10:27.379 kind of self -reflection and correction. Like 00:10:27.379 --> 00:10:29.970 an internal monologue almost. Yeah. And what's 00:10:29.970 --> 00:10:31.690 really cool for users, especially scientists 00:10:31.690 --> 00:10:34.450 or devs, is the control it offers. Control how? 00:10:34.669 --> 00:10:37.070 You can set uncertainty thresholds, basically 00:10:37.070 --> 00:10:40.409 telling it how sure it needs to be. You can rerun 00:10:40.409 --> 00:10:43.029 its reasoning paths, even tweak them. So you 00:10:43.029 --> 00:10:46.059 can see how it's thinking and guide it. Exactly. 00:10:46.059 --> 00:10:49.399 You can oversee how it handles tricky, ambiguous 00:10:49.399 --> 00:10:51.960 problems. And it gives you regular, but also 00:10:51.960 --> 00:10:54.480 flexibility. That sounds incredibly useful, not 00:10:54.480 --> 00:10:56.299 just getting the answer, but understanding the 00:10:56.299 --> 00:10:58.899 process. Right. It's like a video GPS. Yeah. 00:10:58.980 --> 00:11:01.360 Instead of one fixed route. Yeah. It's like a 00:11:01.360 --> 00:11:05.120 GPS that can dynamically rethink your path mid 00:11:05.120 --> 00:11:07.960 -drive if it hits traffic or finds a better way. 00:11:08.120 --> 00:11:10.279 That's a fantastic analogy, adapting on the fly. 00:11:10.399 --> 00:11:12.659 It really is. And it raises a big question, doesn't 00:11:12.659 --> 00:11:15.090 it? Which is? If Microsoft can make GPS... P4 00:11:15.090 --> 00:11:18.830 .1, I think. Steer and adapt like this. What 00:11:18.830 --> 00:11:21.990 happens when they apply CLIO to their next generation 00:11:21.990 --> 00:11:25.370 models? Oh, yeah. The potential there seems huge. 00:11:25.529 --> 00:11:52.940 Immense. Right. Powerful AI reasoning, suddenly 00:11:52.940 --> 00:11:55.559 accessible to hundreds of millions. It's a huge 00:11:55.559 --> 00:11:58.240 shift from just chatbots to these dynamic assistants 00:11:58.240 --> 00:12:00.940 that can actually do complex things. And then 00:12:00.940 --> 00:12:04.450 you add things like Microsoft CLIO. And you see 00:12:04.450 --> 00:12:06.990 AI itself becoming more fluid, more adaptive. 00:12:07.210 --> 00:12:09.490 Yeah, not just static knowledge anymore. Exactly. 00:12:09.490 --> 00:12:11.649 They're becoming systems that can think, self 00:12:11.649 --> 00:12:14.669 -correct, learn in real time, explore new possibilities. 00:12:14.669 --> 00:12:17.590 The pace is just wild. The competition's intense. 00:12:18.070 --> 00:12:20.860 But the end result seems to be... AI that's more 00:12:20.860 --> 00:12:23.100 capable, more practical. And just more integrated 00:12:23.100 --> 00:12:25.379 into everything we do. It's an incredible moment 00:12:25.379 --> 00:12:28.639 to watch unfold. It really is. So maybe something 00:12:28.639 --> 00:12:31.279 for you listening to think about. Yeah. How might 00:12:31.279 --> 00:12:33.620 these new capabilities, especially this idea 00:12:33.620 --> 00:12:36.019 of an AI agent in your pocket, actually impact 00:12:36.019 --> 00:12:39.379 your daily life? Yeah. Your work. What specific 00:12:39.379 --> 00:12:41.759 task would you just hand over to this new level 00:12:41.759 --> 00:12:43.860 of AI? Yeah. What's the first thing you'd want 00:12:43.860 --> 00:12:45.919 it to handle for you? And maybe a bigger thought. 00:12:46.490 --> 00:12:49.830 If AI can discover new physics like we talked 00:12:49.830 --> 00:12:53.970 about or adapt its thinking like CLIO, how does 00:12:53.970 --> 00:12:56.490 that change our own understanding of intelligence 00:12:56.490 --> 00:12:58.809 itself in the years ahead? What does it mean 00:12:58.809 --> 00:13:03.009 for us? Deep questions. Lots to ponder. Definitely. 00:13:03.370 --> 00:13:07.149 We hope you found some aha moments in this deep 00:13:07.149 --> 00:13:09.289 dive. And hopefully you'll keep exploring all 00:13:09.289 --> 00:13:11.289 these fascinating developments with us. Thanks 00:13:11.289 --> 00:13:12.970 so much for joining us. Until next time, keep 00:13:12.970 --> 00:13:13.409 exploring.