WEBVTT 00:00:00.000 --> 00:00:03.379 Welcome back to the Deep Dive. For years, the 00:00:03.379 --> 00:00:06.019 big leaps in AI, they always seem to happen on 00:00:06.019 --> 00:00:07.900 a certain rhythm. We even had a name for it, 00:00:07.940 --> 00:00:11.019 right? Thursday AI. Well, you can go ahead and 00:00:11.019 --> 00:00:13.039 scratch that off the calendar. Google just completely 00:00:13.039 --> 00:00:15.720 reset the clock with Tuesday AI. Their Gemini 00:00:15.720 --> 00:00:19.219 3 launch was massive. And the metrics they dropped, 00:00:19.359 --> 00:00:21.140 I mean, they just changed the entire scoreboard. 00:00:21.420 --> 00:00:26.170 This thing scored a 1501 ELO on El Marina. And 00:00:26.170 --> 00:00:28.929 maybe more importantly, it crushed GPT -5 Pro 00:00:28.929 --> 00:00:31.710 on the humanities last exam. We're talking 37 00:00:31.710 --> 00:00:35.810 .4 % to 31 .6%. That is a huge margin, over six 00:00:35.810 --> 00:00:38.009 points. This tells me it's not just a smarter 00:00:38.009 --> 00:00:40.990 chatbot. The real signal here is that we're looking 00:00:40.990 --> 00:00:44.890 at an entirely new class of AI teammate, something 00:00:44.890 --> 00:00:47.020 that takes action. And that's exactly where we're 00:00:47.020 --> 00:00:48.979 focusing today. So for everyone joining us on 00:00:48.979 --> 00:00:50.759 this deep dive, our whole mission is to cut through 00:00:50.759 --> 00:00:53.119 all the noise, all the data, and just pull out 00:00:53.119 --> 00:00:55.780 the most important signals from this whirlwind 00:00:55.780 --> 00:00:58.079 of AI updates. We do have a packed roadmap. We're 00:00:58.079 --> 00:01:00.560 going to unpack the raw power of Gemini 3, this 00:01:00.560 --> 00:01:02.780 new agent model, and really look at what those 00:01:02.780 --> 00:01:04.879 big benchmark scores mean for how you can actually 00:01:04.879 --> 00:01:08.409 use it. Then we're going to pivot to its unexpected 00:01:08.409 --> 00:01:11.930 competitor, Grok 4 .1, the whole Vive first model 00:01:11.930 --> 00:01:14.170 that tried to steal the spotlight. And we'll 00:01:14.170 --> 00:01:16.430 finish up with the really practical stuff, the 00:01:16.430 --> 00:01:18.989 new agents coming from industry and the just 00:01:18.989 --> 00:01:21.329 colossal funding that proves this shift is here 00:01:21.329 --> 00:01:23.430 to stay. Right. The focus is understanding that 00:01:23.430 --> 00:01:28.150 shift from a simple chatbot to a true multi -step 00:01:28.150 --> 00:01:31.730 agent that can plan and build things. Okay. Let's 00:01:31.730 --> 00:01:34.099 get into it. We have to start with Gemini 3. 00:01:34.280 --> 00:01:36.299 I mean, the performance numbers alone basically 00:01:36.299 --> 00:01:38.219 set a new state of the art. It didn't just win. 00:01:38.420 --> 00:01:40.659 It sort of redefined the whole arena. That's 00:01:40.659 --> 00:01:43.579 completely right. And when we say 1501 ELO on 00:01:43.579 --> 00:01:46.140 El Marina, for you listening, if you don't follow 00:01:46.140 --> 00:01:48.519 the leaderboards, El Marina is where these top 00:01:48.519 --> 00:01:51.400 models go head to head on really complex real 00:01:51.400 --> 00:01:54.260 world stuff. That score, it's the new high watermark 00:01:54.260 --> 00:01:57.239 for general AI intelligence. But the win on humanity's 00:01:57.239 --> 00:01:59.700 last exam. That feels more consequential to me. 00:01:59.719 --> 00:02:01.599 That's not a trivia contest. It's a test designed 00:02:01.599 --> 00:02:04.280 to measure really deep cross -disciplinary reasoning. 00:02:04.459 --> 00:02:06.900 Winning there signals a huge leap in cognitive 00:02:06.900 --> 00:02:09.159 ability. And it's built on tech that's pushing 00:02:09.159 --> 00:02:11.639 way beyond what we have now. They even mentioned 00:02:11.639 --> 00:02:13.800 a separate model, the DeepThink version, hitting 00:02:13.800 --> 00:02:19.219 45 .1 % on ARC -AGI2. And that test, that measures 00:02:19.219 --> 00:02:21.759 fluid intelligence, you know, abstract reasoning. 00:02:22.099 --> 00:02:23.599 That's the score that got everyone's attention. 00:02:23.819 --> 00:02:26.449 I mean, you had Elon Musk and Sam Altman. both 00:02:26.449 --> 00:02:28.770 publicly congratulating them on the launch. Oh, 00:02:28.789 --> 00:02:30.990 yeah. When your biggest competitors stop to acknowledge 00:02:30.990 --> 00:02:32.990 the jump you just made, you know it's a big deal. 00:02:33.129 --> 00:02:36.330 It absolutely is. But the real story here, it 00:02:36.330 --> 00:02:39.750 isn't just the score, it's the role. We have 00:02:39.750 --> 00:02:42.669 to talk about this agent upgrade. An agent is 00:02:42.669 --> 00:02:45.090 an AI that stops just answering your questions 00:02:45.090 --> 00:02:47.789 and becomes like a planning teammate. Yeah, it's 00:02:47.789 --> 00:02:50.370 about owning the outcome, not just finding the 00:02:50.370 --> 00:02:53.370 information. An agent can handle multi -step 00:02:53.370 --> 00:02:56.349 tasks. It plans out its actions on its own. And 00:02:56.349 --> 00:02:58.349 it can simulate the results of those actions 00:02:58.349 --> 00:03:00.610 before it even does them. It's like it's stacking 00:03:00.610 --> 00:03:03.250 Lego blocks of action and data, not just handing 00:03:03.250 --> 00:03:05.389 you one block. And that all starts with what 00:03:05.389 --> 00:03:08.469 you give it. The input. Gemini 3 is natively 00:03:08.469 --> 00:03:10.770 multimodal. And this isn't just a theory. You 00:03:10.770 --> 00:03:14.389 can give it anything. A complex PDF, a photo, 00:03:14.569 --> 00:03:17.409 a technical diagram, even a rough scribble on 00:03:17.409 --> 00:03:20.539 a napkin. It uses all of it as context for its 00:03:20.539 --> 00:03:24.860 plan. The examples the Google CEO gave were just 00:03:24.860 --> 00:03:28.180 stunning. Imagine sketching a website idea on 00:03:28.180 --> 00:03:30.960 a napkin, you hand it to the AI, and it turns 00:03:30.960 --> 00:03:34.219 that scribble into a full working website with 00:03:34.219 --> 00:03:37.509 all the code. That's utility in seconds. Or think 00:03:37.509 --> 00:03:39.789 about education. You take a complicated physics 00:03:39.789 --> 00:03:42.270 diagram and it turns it into an interactive lesson 00:03:42.270 --> 00:03:44.090 where you can actually move things around, run 00:03:44.090 --> 00:03:46.509 simulations in real time. And it scales up, too, 00:03:46.610 --> 00:03:49.090 for big business tasks like it can analyze a 00:03:49.090 --> 00:03:51.169 long video of a golf swing, find a technical 00:03:51.169 --> 00:03:53.310 flaw and then suggest the exact drills you need 00:03:53.310 --> 00:03:55.210 to do to fix it. And the output isn't just text 00:03:55.210 --> 00:03:57.710 anymore. It can generate dynamic layouts with 00:03:57.710 --> 00:03:59.669 interactive tools built right in, you know, like 00:03:59.669 --> 00:04:01.990 actual data sliders that appear in your search 00:04:01.990 --> 00:04:04.030 results. Right. That's a perfect example of it 00:04:04.030 --> 00:04:06.469 moving from just giving you text to actively 00:04:06.469 --> 00:04:08.569 building tools for you. And we should circle 00:04:08.569 --> 00:04:10.659 back to DeepThink for a second. that version 00:04:10.659 --> 00:04:12.759 is built for the really heavy lifting for multi 00:04:12.759 --> 00:04:15.659 -hop logic what that means is the model can link 00:04:15.659 --> 00:04:18.360 distant non -obvious bits of information together 00:04:18.360 --> 00:04:21.180 from totally different data sets that's a sign 00:04:21.180 --> 00:04:23.939 of real abstract reasoning not just better memory 00:04:23.939 --> 00:04:26.579 it's still in safety testing before it hits the 00:04:26.579 --> 00:04:28.680 ultra tier which kind of tells you how powerful 00:04:28.680 --> 00:04:31.439 it is so if this model can build interactive 00:04:31.439 --> 00:04:34.740 uis and simulate outcomes and plan things out 00:04:35.399 --> 00:04:37.480 What does that actually mean for the future of 00:04:37.480 --> 00:04:39.740 professional workflows, for creative work, technical 00:04:39.740 --> 00:04:42.240 work, all of it? It means AI is moving from being 00:04:42.240 --> 00:04:45.220 a passive source of information to actively building 00:04:45.220 --> 00:04:48.100 your tools and running complex operations. Now, 00:04:48.160 --> 00:04:50.240 shifting gears a bit, we have to talk about the 00:04:50.240 --> 00:04:52.220 surprise competitor that dropped right before 00:04:52.220 --> 00:04:54.120 Gemini 3 sucked all the air out of the room. 00:04:54.199 --> 00:04:57.740 I'm talking about Grok 4 .1 from XAI. That was 00:04:57.740 --> 00:04:59.860 such a brilliant counter move. They presented 00:04:59.860 --> 00:05:02.680 it as vibe first, you know, focused on a specific 00:05:02.680 --> 00:05:05.680 tone fast, witty, a little edgy, but then it 00:05:05.680 --> 00:05:07.540 also delivered some really impressive intelligence 00:05:07.540 --> 00:05:10.040 scores. It's that classic EQ versus IQ battle. 00:05:10.800 --> 00:05:13.839 Precisely. And Grok 4 .1 gives you two modes 00:05:13.839 --> 00:05:16.199 to work with. You've got the standard version, 00:05:16.379 --> 00:05:19.319 which is super snappy, instant replies. And then 00:05:19.319 --> 00:05:22.720 there's Grok 4 .1 thinking, which it takes its 00:05:22.720 --> 00:05:24.579 time to process, but it's aiming for a deeper, 00:05:24.620 --> 00:05:26.879 more analytical answer. And that thinking version 00:05:26.879 --> 00:05:29.459 actually, for a moment, had its day in the sun. 00:05:29.660 --> 00:05:31.720 It hit number one on the El Reno leaderboard, 00:05:31.860 --> 00:05:35.079 scored 1510 ELO, just for a few hours before 00:05:35.079 --> 00:05:36.920 Gemini came along and bumped it to number two. 00:05:37.220 --> 00:05:39.480 But it proves this isn't just a personality. 00:05:40.009 --> 00:05:42.430 It's a serious contender. Oh, it's very capable, 00:05:42.569 --> 00:05:44.509 especially in creative writing. It's incredibly 00:05:44.509 --> 00:05:47.949 strong. Only the next gen GPT model scored higher. 00:05:48.329 --> 00:05:50.410 But for me, the most important signal wasn't 00:05:50.410 --> 00:05:52.670 the score. It was the massive improvement in 00:05:52.670 --> 00:05:55.470 just. basic reliability. They really needed to 00:05:55.470 --> 00:05:57.250 fix their accuracy problem, didn't they? They 00:05:57.250 --> 00:05:59.350 absolutely did. So they slashed the hallucination 00:05:59.350 --> 00:06:02.550 rate from 12 % all the way down to 4 .2%. And 00:06:02.550 --> 00:06:04.910 their internal test showed factual accuracy went 00:06:04.910 --> 00:06:07.829 up by 66%. And look, I'll be honest, I still 00:06:07.829 --> 00:06:09.589 wrestle with prompt drift myself when I'm trying 00:06:09.589 --> 00:06:11.649 to build these complex, multi -layered queries. 00:06:11.970 --> 00:06:14.790 So a reliability jump like that, it's huge for 00:06:14.790 --> 00:06:17.699 user trust. Yeah. And that subjective quality 00:06:17.699 --> 00:06:21.420 matters, too. People describe the tone as friendlier, 00:06:21.519 --> 00:06:24.000 more empathetic, like the model just gets you 00:06:24.000 --> 00:06:26.699 a little bit better. That vibe is so important 00:06:26.699 --> 00:06:28.860 for getting people to actually use it. But here's 00:06:28.860 --> 00:06:31.720 the huge catch, the big limiting factor. Grok 00:06:31.720 --> 00:06:35.259 4 .1 is completely platform locked. You can only 00:06:35.259 --> 00:06:38.759 use it on X, the social platform. or the Grok 00:06:38.759 --> 00:06:41.819 website and app. You can't plug it into external 00:06:41.819 --> 00:06:44.019 tools. You can't build it into your company's 00:06:44.019 --> 00:06:46.699 workflow with an API. It is worth mentioning, 00:06:46.740 --> 00:06:49.300 though, that XAI did release a white paper with 00:06:49.300 --> 00:06:51.379 some of their training info. That's pretty rare 00:06:51.379 --> 00:06:53.480 for them, so it shows they're serious about backing 00:06:53.480 --> 00:06:56.360 up these performance claims. So why is that platform 00:06:56.360 --> 00:06:58.639 restriction such a serious limiting factor right 00:06:58.639 --> 00:07:00.519 now, especially when you compare it to Gemini 00:07:00.519 --> 00:07:02.740 3, which is obviously being built for maximum 00:07:02.740 --> 00:07:04.759 utility everywhere? The tradeoff is control. 00:07:05.199 --> 00:07:07.699 Grok's usefulness is limited because it can't 00:07:07.699 --> 00:07:10.240 integrate into existing business tools or external 00:07:10.240 --> 00:07:13.139 APIs. Okay, as we move into the wider industry 00:07:13.139 --> 00:07:14.839 context, I think it's really important that we 00:07:14.839 --> 00:07:18.500 just pause here and reinforce a point that always 00:07:18.500 --> 00:07:20.740 gets lost in the hype. That warning from the 00:07:20.740 --> 00:07:23.160 Google CEO. It really does need to be repeated. 00:07:23.319 --> 00:07:25.060 With all this excitement about these powerful 00:07:25.060 --> 00:07:27.420 new agents, we have to remember they are still, 00:07:27.500 --> 00:07:30.680 quote, prone to errors. You can't just blindly 00:07:30.680 --> 00:07:32.980 trust the output. Critical thinking is still 00:07:32.980 --> 00:07:35.970 step zero. That being said, the scale of the 00:07:35.970 --> 00:07:38.110 infrastructure being built to support these agents 00:07:38.110 --> 00:07:41.610 is it's just staggering. Just look at the funding. 00:07:41.970 --> 00:07:45.250 Lambda, an AI cloud company, just raised over 00:07:45.250 --> 00:07:48.290 one and a half billion dollars with Microsoft 00:07:48.290 --> 00:07:51.269 heavily involved. And it's specifically to expand 00:07:51.269 --> 00:07:53.870 the infrastructure needed for this stuff. A billion 00:07:53.870 --> 00:07:56.910 and a half dollars. That signals a complete arms 00:07:56.910 --> 00:07:59.899 race for computing power. These models need a 00:07:59.899 --> 00:08:02.660 ton of silicon, a ton of energy to serve up these 00:08:02.660 --> 00:08:05.139 complex multi -step agent requests at a global 00:08:05.139 --> 00:08:07.060 scale. Exactly. I mean, when you think about 00:08:07.060 --> 00:08:09.319 the multi -hop reasoning, the simulations these 00:08:09.319 --> 00:08:11.459 agents are running, the demand for computation 00:08:11.459 --> 00:08:13.819 is just exponential compared to a simple text 00:08:13.819 --> 00:08:18.100 query. Whoa. Yeah, imagine scaling that. Agents 00:08:18.100 --> 00:08:20.910 simulating outcomes and building UIs for... a 00:08:20.910 --> 00:08:24.310 billion queries a day. $1 .5 billion is what 00:08:24.310 --> 00:08:27.050 it takes just to get ready for that future. That's 00:08:27.050 --> 00:08:29.480 a real moment of wonder. Thinking about that 00:08:29.480 --> 00:08:31.500 scale. And that infrastructure is already being 00:08:31.500 --> 00:08:34.179 put to work. Microsoft just confirmed this industry 00:08:34.179 --> 00:08:37.320 -wide pivot to agents at their Ignite 2025 event. 00:08:37.639 --> 00:08:41.379 They announced Agent 365 and 12 new co -pilot 00:08:41.379 --> 00:08:43.740 agents, all designed for specific professional 00:08:43.740 --> 00:08:46.419 jobs like marketing or customer service. The 00:08:46.419 --> 00:08:49.600 agent race is officially on. We also got that 00:08:49.600 --> 00:08:52.899 funny little reminder of the ethical gray areas 00:08:52.899 --> 00:08:54.779 that open up when these models get so good at 00:08:54.779 --> 00:08:56.840 manipulation. Right. The story about the DoorDash 00:08:56.840 --> 00:08:59.320 customer who got a ref... fund by using AI to 00:08:59.320 --> 00:09:01.799 fake a photo of a raw burger. Right. It's a small 00:09:01.799 --> 00:09:03.799 thing, but it points to the immediate potential 00:09:03.799 --> 00:09:05.840 for just, you know, everyday misuse and fraud. 00:09:06.080 --> 00:09:08.639 And looking ahead, this competition is only going 00:09:08.639 --> 00:09:11.220 to get more intense. OpenAI is definitely not 00:09:11.220 --> 00:09:13.720 sitting still. Their VP of research mentioned 00:09:13.720 --> 00:09:15.799 they have a stronger version of their IMO winning 00:09:15.799 --> 00:09:19.039 model coming in 2025. And that's their math champion 00:09:19.039 --> 00:09:21.460 model. So they're pushing pure mathematical reasoning 00:09:21.460 --> 00:09:24.639 just as hard. So given all of this, the new power, 00:09:24.759 --> 00:09:27.259 the acknowledged risks, what's the single most 00:09:27.370 --> 00:09:29.370 important rule for the average user right now 00:09:29.370 --> 00:09:32.370 to be both effective and safe. Critical thinking 00:09:32.370 --> 00:09:35.509 remains essential. Always use AI results as a 00:09:35.509 --> 00:09:37.950 starting point, not the final unquestioned fact. 00:09:38.309 --> 00:09:40.850 So to wrap up this deep dive, I think the core 00:09:40.850 --> 00:09:42.809 takeaway for you, for the listener, is this. 00:09:42.929 --> 00:09:45.090 The whole AI race has fundamentally shifted. 00:09:45.429 --> 00:09:48.169 We are not arguing anymore about who can write 00:09:48.169 --> 00:09:50.230 slightly better text. Right. The new battleground 00:09:50.230 --> 00:09:54.110 is complex, multimodal. action planning. We're 00:09:54.110 --> 00:09:56.669 seeing the rise of true agents, not just smarter 00:09:56.669 --> 00:09:59.450 chatbots. They can take in all kinds of data 00:09:59.450 --> 00:10:01.330 and then actively build things for you. And the 00:10:01.330 --> 00:10:03.990 competition is now on two clear paths. On one 00:10:03.990 --> 00:10:05.970 side, you have pure, raw performance and deep 00:10:05.970 --> 00:10:08.490 reasoning, which is Gemini 3. On the other, you 00:10:08.490 --> 00:10:11.169 have specialized tone, speed, and empathy, which 00:10:11.169 --> 00:10:13.929 is what Grok 4 .1 is going for. And both are 00:10:13.929 --> 00:10:16.120 fighting for your attention. Thank you for joining 00:10:16.120 --> 00:10:18.559 us on this deep dive into the new era of autonomous 00:10:18.559 --> 00:10:21.419 AI agents. The changes are happening so fast 00:10:21.419 --> 00:10:23.620 and the capabilities are becoming so much more 00:10:23.620 --> 00:10:26.100 powerful, which makes these conversations just 00:10:26.100 --> 00:10:28.860 absolutely necessary. And as you think about 00:10:28.860 --> 00:10:31.059 the complexity of these new systems, here's a 00:10:31.059 --> 00:10:33.320 final thought for you to explore. Considering 00:10:33.320 --> 00:10:35.639 that safety tests are what's slowing down the 00:10:35.639 --> 00:10:38.500 release of the most powerful versions, how soon 00:10:38.500 --> 00:10:41.059 do you think the high reasoning Gemini 3 DeepThink 00:10:41.059 --> 00:10:43.799 model will actually clear safety? And what new 00:10:43.799 --> 00:10:46.639 really complex ethical questions will its multi 00:10:46.639 --> 00:10:49.100 -hop logic raise when it finally gets out there? 00:10:49.179 --> 00:10:49.840 Something to think about.