WEBVTT 00:00:00.000 --> 00:00:03.740 72 % beat. That's the number that kept me up 00:00:03.740 --> 00:00:06.660 last night. If you have a digital vault, a crypto 00:00:06.660 --> 00:00:10.599 smart contract, or really any secure digital 00:00:10.599 --> 00:00:14.119 asset, as of this week, there's a new AI agent 00:00:14.119 --> 00:00:16.100 that can break into it and drain it of all funds 00:00:16.100 --> 00:00:19.199 72 % of the time. And to be clear, that isn't 00:00:19.199 --> 00:00:21.440 a theoretical figure from a sci -fi novel. That 00:00:21.440 --> 00:00:24.019 is the actual benchmark result from the latest 00:00:24.019 --> 00:00:27.059 coding agent released just days ago. We have 00:00:27.059 --> 00:00:28.960 crossed a threshold where the digital offense 00:00:28.960 --> 00:00:31.879 is statistically superior to the digital defense. 00:00:32.500 --> 00:00:34.820 Welcome back to The Deep Dive. It is Friday, 00:00:35.060 --> 00:00:38.659 February 20th, 2026. I'm your host, and I'm here 00:00:38.659 --> 00:00:41.000 with our resident expert to make sense of a week 00:00:41.000 --> 00:00:43.270 that... honestly feels like a year. It really 00:00:43.270 --> 00:00:45.270 does. It's great to be here. You know, usually 00:00:45.270 --> 00:00:47.770 we talk about AI explicitly building things, 00:00:47.929 --> 00:00:50.170 generating art, writing code, solving protein 00:00:50.170 --> 00:00:52.490 folding. But this week, the narrative shifted. 00:00:52.590 --> 00:00:54.750 It's about AI taking things. And it's about the 00:00:54.750 --> 00:00:57.530 massive economic machinery being built to support 00:00:57.530 --> 00:01:00.009 that power. Right. So let's map out where we're 00:01:00.009 --> 00:01:02.310 going today for you listening. We have a heavy 00:01:02.310 --> 00:01:04.989 stack of reports to get through. First, Google 00:01:04.989 --> 00:01:07.409 is hitting back hard in the model wars. Gemini 00:01:07.409 --> 00:01:10.769 3 .1 Pro is here. And the headline isn't speed. 00:01:11.260 --> 00:01:13.859 It's reasoning. We need to unpack what that actually 00:01:13.859 --> 00:01:15.879 means. Exactly. Then we're going to follow the 00:01:15.879 --> 00:01:18.780 money. And I mean serious money. OpenAI is looking 00:01:18.780 --> 00:01:22.879 at an $850 billion valuation. We'll look at why 00:01:22.879 --> 00:01:24.780 that number is so high and the controversial 00:01:24.780 --> 00:01:26.859 shift toward advertising that's coming with it. 00:01:27.180 --> 00:01:30.480 And finally, we are going to do a deep dive into 00:01:30.480 --> 00:01:33.219 that cold open statistic. We need to talk about 00:01:33.219 --> 00:01:35.620 why offense is currently beating defense and 00:01:35.620 --> 00:01:38.459 cybersecurity. And what happens when an AI can 00:01:38.459 --> 00:01:41.140 execute a flash loan attack all on its own? That 00:01:41.140 --> 00:01:42.739 is the part that's going to scare some people, 00:01:42.840 --> 00:01:44.939 but it's crucial to understand. Okay, let's start 00:01:44.939 --> 00:01:47.040 with the tools. It feels like we just finished 00:01:47.040 --> 00:01:49.200 analyzing Gemini 3 .0. That was what, November? 00:01:49.480 --> 00:01:51.760 Yeah, three months ago. Three months. And now 00:01:51.760 --> 00:01:55.239 we have Gemini 3 .1 Pro. Usually in software, 00:01:55.379 --> 00:01:58.640 a .1 update is a snooze. It's bug fixes. Maybe 00:01:58.640 --> 00:02:01.340 a new font. But looking at these specs, this 00:02:01.340 --> 00:02:04.590 seems entirely different. It's completely different. 00:02:04.849 --> 00:02:07.150 Google has rolled this out with a very specific 00:02:07.150 --> 00:02:10.129 focus, reasoning. And I think we need to pause 00:02:10.129 --> 00:02:11.990 on that word because it gets horribly abused 00:02:11.990 --> 00:02:15.090 in marketing. Please do. Because to me, reasoning 00:02:15.090 --> 00:02:18.449 implies consciousness, or at least human -like 00:02:18.449 --> 00:02:21.750 thought. Is Google claiming their chatbot is 00:02:21.750 --> 00:02:24.449 actually thinking? Not thinking in the biological 00:02:24.449 --> 00:02:26.750 sense, but think of it this way. Previous models, 00:02:26.870 --> 00:02:29.469 the ones we used in 2024 or 2025, were largely 00:02:29.469 --> 00:02:32.189 system one thinkers. If you know Daniel Kahneman's 00:02:32.189 --> 00:02:35.090 work, it's fast, instinctive, automatic. You 00:02:35.090 --> 00:02:37.669 ask for a poem, it predicts the next word immediately. 00:02:38.319 --> 00:02:40.879 It is reflexive. Okay, like a reflex. Exactly. 00:02:41.099 --> 00:02:44.060 Gemini 3 .1 is moving towards System 2. It's 00:02:44.060 --> 00:02:46.819 deliberate. It's slow. It's logical. When you 00:02:46.819 --> 00:02:48.800 ask it a complex question, it doesn't just spit 00:02:48.800 --> 00:02:51.060 out the statistically probable next word. It 00:02:51.060 --> 00:02:53.159 holds the problem in its mind, breaks it down 00:02:53.159 --> 00:02:56.159 into steps, critiques its own plan, and then 00:02:56.159 --> 00:02:58.960 executes. It is spending compute on inference 00:02:58.960 --> 00:03:02.219 time, literally taking time to think before it 00:03:02.219 --> 00:03:04.180 speaks. So it's the difference between a student 00:03:04.180 --> 00:03:06.580 blurting out an answer in class versus a student 00:03:06.580 --> 00:03:08.159 sitting down to work through a... a math -proof 00:03:08.159 --> 00:03:10.960 step -by -step. Perfect analogy. And the results 00:03:10.960 --> 00:03:14.099 are stark. According to the preview, Gemini 3 00:03:14.099 --> 00:03:16.340 .1 Pro has doubled its reasoning performance 00:03:16.340 --> 00:03:19.240 compared to 3 .0. Doubled in three months. Doubled. 00:03:19.240 --> 00:03:21.139 I am always skeptical of internal benchmarks. 00:03:21.460 --> 00:03:23.819 Companies love to grade their own homework. Did 00:03:23.819 --> 00:03:25.879 they test this against anything we'd recognize? 00:03:26.120 --> 00:03:28.580 They did. They used a test called Humanities 00:03:28.580 --> 00:03:31.900 Last Exam. Okay, that is a terrifying name for 00:03:31.900 --> 00:03:34.099 a test. Is that real? It is. It was designed 00:03:34.099 --> 00:03:36.889 to be the absolute ceiling of difficulty. Questions 00:03:36.889 --> 00:03:39.409 that require multidisciplinary knowledge to solve. 00:03:39.569 --> 00:03:41.889 It's the stuff that completely stumped the models 00:03:41.889 --> 00:03:45.330 from 2025. Gemini 3 .1 outperformed its predecessor 00:03:45.330 --> 00:03:48.069 there significantly. But the one that really 00:03:48.069 --> 00:03:50.349 matters to me is the APEX agent's leaderboard. 00:03:50.550 --> 00:03:52.810 I'm not familiar with APEX. APEX measures real 00:03:52.810 --> 00:03:55.129 professional past performance. It's not right 00:03:55.129 --> 00:03:58.069 a haiku. It's messy. It's take this scattered 00:03:58.069 --> 00:04:01.270 data set of CSV files, analyze the trends, cross 00:04:01.270 --> 00:04:03.770 -reference it with this PDF report on climate 00:04:03.770 --> 00:04:06.870 change, and give me a strategic insight. So actual 00:04:06.870 --> 00:04:09.330 knowledge work. Actual work. Gemini 3 .1 hit 00:04:09.330 --> 00:04:11.949 the top spot. It is designed for agentic workflows. 00:04:12.229 --> 00:04:14.310 It isn't just a chatbot anymore. It's a research 00:04:14.310 --> 00:04:16.949 partner. I see. This explains the integration 00:04:16.949 --> 00:04:19.850 they announced with Notebook LM. I've been using 00:04:19.850 --> 00:04:22.490 Notebook LM for research, basically dumping PDFs 00:04:22.490 --> 00:04:24.290 in and chatting with them. But you're saying 00:04:24.290 --> 00:04:27.230 3 .1 Pro changes the nature of that interaction. 00:04:27.649 --> 00:04:30.129 Completely. Notebook LM was a retrieval tool. 00:04:30.310 --> 00:04:32.709 Find me the quote about interest rates. With 00:04:32.709 --> 00:04:35.750 3 .1 Pro, it becomes an analyst. You can say, 00:04:35.850 --> 00:04:38.350 look the interest rates in document A, compare 00:04:38.350 --> 00:04:40.949 them to the housing trends in document B, and 00:04:40.949 --> 00:04:44.769 predict the impact on this third variable. Whoa, 00:04:44.910 --> 00:04:47.529 imagine connecting scattered data into full picture 00:04:47.529 --> 00:04:50.069 insights like that at scale. It's stacking Lego 00:04:50.069 --> 00:04:52.970 blocks of data into a skyscraper. That is a significant 00:04:52.970 --> 00:04:55.829 shift. But I have to ask, if the cycle is down 00:04:55.829 --> 00:04:58.350 to three months, are we even learning the old 00:04:58.350 --> 00:05:00.920 tools before the new ones replace them? Barely. 00:05:00.920 --> 00:05:02.879 We're really just adapting to the speed of the 00:05:02.879 --> 00:05:05.120 upgrade, not the tool itself. We are becoming 00:05:05.120 --> 00:05:07.620 experts in adaptation rather than experts in 00:05:07.620 --> 00:05:09.759 software. Barely adapting to the upgrade speed, 00:05:09.899 --> 00:05:12.699 not the tool itself. Got it. That's a heavy thought. 00:05:12.779 --> 00:05:14.839 It might be the only skill that matters in 2026. 00:05:15.360 --> 00:05:17.620 All right, let's pivot. Because building these 00:05:17.620 --> 00:05:20.740 System 2 thinkers, these reasoning engines, is 00:05:20.740 --> 00:05:23.439 expensive. The compute costs are astronomical. 00:05:24.019 --> 00:05:26.579 Massive. The energy bills alone are enough to 00:05:26.579 --> 00:05:28.839 bankrupt small nations. Which brings us to the 00:05:28.839 --> 00:05:31.740 business side. We're seeing reports that OpenAI 00:05:31.740 --> 00:05:36.209 is finalizing a 100 % billion dollar raise. Yes. 00:05:36.329 --> 00:05:39.410 Which puts their valuation at over 850 billion 00:05:39.410 --> 00:05:42.110 dollars. It is staggering. To put that in perspective, 00:05:42.329 --> 00:05:44.009 that's approaching the territory of companies 00:05:44.009 --> 00:05:47.310 like Meta or Tesla. But OpenAI is a fraction 00:05:47.310 --> 00:05:49.810 of the age. It's nearly a trillion dollar valuation 00:05:49.810 --> 00:05:51.689 for a company that didn't exist a decade ago. 00:05:51.910 --> 00:05:54.370 And the backers are the usual titans. Amazon, 00:05:54.670 --> 00:05:57.589 SoftBank, Nvidia, Microsoft. It feels like the 00:05:57.589 --> 00:06:00.069 entire tech economy is consolidating around this 00:06:00.069 --> 00:06:02.509 one entity. It's becoming the too big to fail 00:06:02.509 --> 00:06:05.389 of the AI era. It is, but here's the tension. 00:06:05.589 --> 00:06:09.129 To justify an $850 billion valuation, you need 00:06:09.129 --> 00:06:11.829 massive revenue. You can't just have $20 subscriptions. 00:06:12.089 --> 00:06:14.689 You need something way more pervasive. And that 00:06:14.689 --> 00:06:18.209 brings us to the A word. Ads. Ads. OpenAI is 00:06:18.209 --> 00:06:20.300 beginning to test them. They are. And this is 00:06:20.300 --> 00:06:22.579 where the market is starting to fracture philosophically. 00:06:22.639 --> 00:06:26.180 You have open AI needing to monetize that massive 00:06:26.180 --> 00:06:29.240 valuation, testing ads and chat GPT. But then 00:06:29.240 --> 00:06:31.339 look at the competitors. Perplexity, which has 00:06:31.339 --> 00:06:33.920 been very aggressive, is actually stepping back 00:06:33.920 --> 00:06:36.259 from in -chat ads. Really? I thought they were 00:06:36.259 --> 00:06:38.339 leaning into that model. They were, but they 00:06:38.339 --> 00:06:40.500 issued a warning this week saying that ads can 00:06:40.500 --> 00:06:43.800 erode user trust. They are pivoting to preserve 00:06:43.800 --> 00:06:46.079 the integrity of the answer. And then you have 00:06:46.079 --> 00:06:49.160 Anthropic, who has publicly committed to staying 00:06:49.160 --> 00:06:51.500 completely ad -free. It's interesting. It's almost 00:06:51.500 --> 00:06:53.399 like the business model is becoming part of the 00:06:53.399 --> 00:06:56.139 product identity, like organic food versus processed 00:06:56.139 --> 00:06:59.300 food. It absolutely is. Think about it. If you're 00:06:59.300 --> 00:07:02.139 using an AI for serious research, medical advice, 00:07:02.439 --> 00:07:05.920 legal strategy, complex coding, do you want an 00:07:05.920 --> 00:07:07.319 answer that is interesting? influenced by the 00:07:07.319 --> 00:07:10.040 highest bidder? Or do you want the raw truth? 00:07:10.379 --> 00:07:12.759 I want the truth. Oh. But I also know the history 00:07:12.759 --> 00:07:15.199 of the internet. Free with ads usually wins against 00:07:15.199 --> 00:07:18.000 paid and private. That's the trillion dollar 00:07:18.000 --> 00:07:21.000 gamble. OpenAI is betting that their model is 00:07:21.000 --> 00:07:23.980 so smart you'll tolerate the commercial. Anthropic 00:07:23.980 --> 00:07:26.600 is betting that you'll pay a premium for silence 00:07:26.600 --> 00:07:29.889 and neutrality. Speaking of commercial, OpenAI 00:07:29.889 --> 00:07:32.449 isn't just looking at ads. They're looking at 00:07:32.449 --> 00:07:35.290 Hollywood. They poached Charles Porch from Instagram. 00:07:35.550 --> 00:07:37.449 Yes, the Hollywood connector. And after that 00:07:37.449 --> 00:07:40.529 $1 billion Disney deal, he's going on a listening 00:07:40.529 --> 00:07:43.769 tour to win over creators. Which is smart because 00:07:43.769 --> 00:07:46.970 creators are skeptical. Actually, skeptical is 00:07:46.970 --> 00:07:50.029 polite. They are furious. They see AI as the 00:07:50.029 --> 00:07:51.949 thing that scrapes their work to replace them. 00:07:52.269 --> 00:07:54.449 OpenAI knows they need to bridge that gap if 00:07:54.449 --> 00:07:56.410 they want to be the platform for media as well 00:07:56.410 --> 00:07:58.829 as text. They need to convince Hollywood that 00:07:58.829 --> 00:08:01.129 they come in peace. It's a battle for hearts 00:08:01.129 --> 00:08:03.790 and minds. So let me put you on the spot. With 00:08:03.790 --> 00:08:06.550 perplexity pulling back on ads and OpenAI leaning 00:08:06.550 --> 00:08:09.310 in, who actually wins the user's trust in the 00:08:09.310 --> 00:08:11.430 long run? The one that gives the best answer 00:08:11.430 --> 00:08:13.370 without selling me a toaster? In the short term, 00:08:13.509 --> 00:08:16.250 OpenAI wins on scale. But for the power user 00:08:16.250 --> 00:08:18.829 economy, I think anthropic or perplexity has 00:08:18.829 --> 00:08:21.250 the edge on trust. So the winner is the one giving 00:08:21.250 --> 00:08:23.350 the best answer without selling a toaster. I 00:08:23.350 --> 00:08:25.930 don't need my AI selling me appliances while 00:08:25.930 --> 00:08:28.389 I'm trying to debug code. Not yet, anyway. Okay, 00:08:28.449 --> 00:08:29.970 let's move to the story that launched the deep 00:08:29.970 --> 00:08:33.049 dive. The heist. The heist. We have discussed 00:08:33.049 --> 00:08:36.289 reasoning. We have discussed money. Now we see 00:08:36.289 --> 00:08:38.669 what happens when reasoning agents decide to 00:08:38.669 --> 00:08:40.909 take the money. This comes from a new study involving 00:08:40.909 --> 00:08:45.669 OpenAI's new coding agent, GPT -5 .3 Codex. 5 00:08:45.669 --> 00:08:49.870 .3. They are iterating fast. Very fast. The study 00:08:49.870 --> 00:08:53.070 used a benchmark called EVM Bench built with 00:08:53.070 --> 00:08:55.990 the crypto firm Paradigm. They set up vulnerable 00:08:55.990 --> 00:08:58.429 smart contracts, digital vaults, essentially, 00:08:58.590 --> 00:09:01.330 and unleashed the AI to see if it could break 00:09:01.330 --> 00:09:03.799 in. And this is where that 72 % come from. Correct. 00:09:04.220 --> 00:09:06.980 GPT -5 .3 codex exploited the vulnerabilities 00:09:06.980 --> 00:09:10.220 72 % of the time. Now for context, compare that 00:09:10.220 --> 00:09:13.879 to just six months ago. GPT -5 scored 31 .9 % 00:09:13.879 --> 00:09:16.960 on the exact same tasks. That is more than doubling 00:09:16.960 --> 00:09:20.220 in capability in half a year. That curve is vertical. 00:09:20.500 --> 00:09:22.639 It is exponential. And it's not just finding 00:09:22.639 --> 00:09:26.159 a bug. A GPT -5 .2 agent was able to execute 00:09:26.159 --> 00:09:29.279 a full flash loan attack on its own. Can we pause 00:09:29.279 --> 00:09:32.299 there? Because flash loan attack is one of those 00:09:32.299 --> 00:09:36.179 crypto terms that sounds exciting but makes no 00:09:36.179 --> 00:09:38.120 sense to most people. Break that down for us. 00:09:38.179 --> 00:09:41.659 Okay. Imagine a bank vault. You walk in and say, 00:09:41.799 --> 00:09:44.559 I want to borrow $100 million, but I have no 00:09:44.559 --> 00:09:47.379 collateral. In the real world, the bank laughs 00:09:47.379 --> 00:09:49.779 at you and calls security. Right. In crypto, 00:09:49.940 --> 00:09:51.720 the bank says, sure, take the hundred million, 00:09:51.860 --> 00:09:53.960 but you have to pay it back within the exact 00:09:53.960 --> 00:09:56.779 same transaction. It's like freezing time. You 00:09:56.779 --> 00:09:58.559 borrow the money, you run around the market, 00:09:58.659 --> 00:10:00.940 you buy and sell assets to manipulate the price. 00:10:01.039 --> 00:10:03.139 You make a profit and you return the original 00:10:03.139 --> 00:10:06.220 loan all before the clock starts ticking again. 00:10:06.440 --> 00:10:08.740 So it all happens in a single split second block 00:10:08.740 --> 00:10:11.700 of code. Exactly. If you fail to pay it back, 00:10:11.879 --> 00:10:14.220 the whole thing reverses like it never happened. 00:10:14.379 --> 00:10:17.519 But if you succeed, you keep the profit. It requires 00:10:17.519 --> 00:10:20.600 incredibly. complex coding and timing to pull 00:10:20.600 --> 00:10:23.759 off. And the AI did this autonomously. Autonomously. 00:10:23.799 --> 00:10:26.340 No human steering the wheel. That is unnerving. 00:10:26.419 --> 00:10:28.779 So the offense is elite. What about the defense? 00:10:29.080 --> 00:10:31.500 Surely if the AI can break it, the AI can fix 00:10:31.500 --> 00:10:34.379 it. That's the problem. The defense is lagging. 00:10:34.600 --> 00:10:37.500 The same study showed that bug detection, the 00:10:37.500 --> 00:10:41.039 AI trying to fix the code, tops out around 46%. 00:10:41.039 --> 00:10:43.840 So it's almost twice as good at breaking in as 00:10:43.840 --> 00:10:46.179 it is at fixing the lock. Why? It comes down 00:10:46.179 --> 00:10:48.700 to the nature of the task. Breaking things is 00:10:48.700 --> 00:10:51.919 often about finding one weak point. Fixing things 00:10:51.919 --> 00:10:54.759 requires understanding the entire system so you 00:10:54.759 --> 00:10:57.519 don't break something else. Patch success sits 00:10:57.519 --> 00:11:02.340 at only 39%. 39%. Beat. That's barely better 00:11:02.340 --> 00:11:04.629 than guessing. Unless, and this is the key, unless 00:11:04.629 --> 00:11:08.190 you give the model a small hint. A hint. If you 00:11:08.190 --> 00:11:10.990 give the AI a nudge, literally telling it roughly 00:11:10.990 --> 00:11:13.909 where the problem is, patch success jumps to 00:11:13.909 --> 00:11:16.809 94%. That is a huge gap. It tells us that the 00:11:16.809 --> 00:11:19.210 bottleneck isn't intelligence. The AI is smart 00:11:19.210 --> 00:11:22.029 enough to fix it. The bottleneck is search. It 00:11:22.029 --> 00:11:23.730 doesn't know where to look for the solution without 00:11:23.730 --> 00:11:26.190 guidance. It's like searching for a needle in 00:11:26.190 --> 00:11:28.250 a haystack versus being told the needle is in 00:11:28.250 --> 00:11:30.850 the top left corner. That feels incredibly human. 00:11:30.950 --> 00:11:32.580 I could fix the leak if you just tell me. me 00:11:32.580 --> 00:11:34.759 which pipe is bursting. Exactly. But in cybersecurity, 00:11:34.919 --> 00:11:37.559 you don't get hints from the attackers. You know, 00:11:37.580 --> 00:11:40.220 I have to admit, I still wrestle with prompt 00:11:40.220 --> 00:11:43.519 drift myself. I am just trying to get the model 00:11:43.519 --> 00:11:45.700 to write a coherent email sometimes, and it goes 00:11:45.700 --> 00:11:48.679 completely off the rails. Yeah. The idea of trusting 00:11:48.679 --> 00:11:51.120 it to patch a smart contract holding millions 00:11:51.120 --> 00:11:53.440 of dollars, that feels like a massive leap of 00:11:53.440 --> 00:11:56.659 faith. It is. But OpenAI sees this gap. They 00:11:56.659 --> 00:11:58.820 don't want their tools to be super weapons for 00:11:58.820 --> 00:12:01.299 hackers. They're launching Aardvark, which is 00:12:01.299 --> 00:12:03.480 their AI security research agent, and they've 00:12:03.480 --> 00:12:06.519 put up a $10 million fund for cybersecurity researchers. 00:12:06.899 --> 00:12:09.799 They know that if AI can break systems this fast, 00:12:10.000 --> 00:12:13.220 defenders need AI just as powerful. Because in 00:12:13.220 --> 00:12:15.580 crypto, there is no undo button. Once it's gone, 00:12:15.679 --> 00:12:18.000 it's gone. So if offense is twice as good as 00:12:18.000 --> 00:12:20.460 defense right now, is any digital vault actually 00:12:20.460 --> 00:12:23.419 safe? Not really, unless you have an AI guard 00:12:23.419 --> 00:12:25.700 dog that's smarter than the AI burglar. And right 00:12:25.700 --> 00:12:28.460 now, the burglars are winning. So no vaults are 00:12:28.460 --> 00:12:32.080 safe unless your AI guard dog outsmarts the burglar. 00:12:33.639 --> 00:12:35.860 We are going to take a quick break. When we come 00:12:35.860 --> 00:12:37.820 back, we're going to step away from the destruction 00:12:37.820 --> 00:12:40.279 and look at the builders, including a legend 00:12:40.279 --> 00:12:42.740 in the field who is building 3D worlds from scratch. 00:12:43.620 --> 00:12:47.299 Stay with us. Monroe's sponsor, Reed Placeholder. 00:12:47.880 --> 00:12:50.159 Welcome back to the Deep Dive. We've talked about 00:12:50.159 --> 00:12:52.740 the big models and the big money, but I want 00:12:52.740 --> 00:12:55.120 to zoom in on the builders for a minute. The 00:12:55.120 --> 00:12:57.600 people making the utilities we use every day. 00:12:57.679 --> 00:13:00.000 The tool belt. Exactly. And there's one name 00:13:00.000 --> 00:13:01.360 that popped up this week that everyone should 00:13:01.360 --> 00:13:05.580 know. Fae Fae Lit. The godmother of AI. She is 00:13:05.580 --> 00:13:08.559 responsible for ImageNet, which basically kickstarted 00:13:08.559 --> 00:13:10.639 the deep learning revolution over a decade ago. 00:13:10.779 --> 00:13:12.899 She has a new company, World Labs, and they just 00:13:12.899 --> 00:13:15.500 raised $1 billion. Which is huge for a startup, 00:13:15.620 --> 00:13:18.159 but it's the product. Morble, that's interesting. 00:13:18.299 --> 00:13:21.460 It turns text into editable 3D worlds. 3D worlds. 00:13:21.539 --> 00:13:23.240 So we aren't just generating images anymore. 00:13:23.460 --> 00:13:26.600 No. This is about spatial intelligence. We are 00:13:26.600 --> 00:13:29.639 entering design pipelines. Imagine typing a description 00:13:29.639 --> 00:13:32.679 of a medieval city, the cobblestones, the lighting, 00:13:32.820 --> 00:13:35.840 the physics of the fog, and having a fully editable 00:13:35.840 --> 00:13:38.220 3D model generated that you can drop straight 00:13:38.220 --> 00:13:40.659 into a video game or a movie. That feels like 00:13:40.659 --> 00:13:43.200 the next frontier. We conquered text. We conquered 00:13:43.200 --> 00:13:46.840 static images. Now we are conquering space. Spatial 00:13:46.840 --> 00:13:49.659 computing needs content. Think about the metaverse 00:13:49.659 --> 00:13:52.919 or Apple's Vision Pro. It's empty right now because 00:13:52.919 --> 00:13:55.809 building 3D assets takes humans... Hundreds of 00:13:55.809 --> 00:13:58.669 hours. If AI can do it in seconds, that changes 00:13:58.669 --> 00:14:00.509 the entertainment industry forever. It lowers 00:14:00.509 --> 00:14:03.409 the barrier to entry to near zero. Speaking of 00:14:03.409 --> 00:14:05.289 which, there were a few other rapid -fire tools 00:14:05.289 --> 00:14:07.690 that caught my eye this week. Did you see origami 00:14:07.690 --> 00:14:11.250 .chat? I did. It is a data enrichment tool. Connects 00:14:11.250 --> 00:14:13.990 over 100 sources. You can use it to find customers, 00:14:14.210 --> 00:14:16.929 enrich a CSV file. It's highly practical. It's 00:14:16.929 --> 00:14:20.049 unsexy, but it saves hours of grunt work. And 00:14:20.049 --> 00:14:23.389 chloe .ai. Open claw in the cloud. Zero setup. 00:14:23.759 --> 00:14:25.840 It's basically an autonomous assistant you can 00:14:25.840 --> 00:14:27.799 spin up in five minutes to browse the web and 00:14:27.799 --> 00:14:30.340 do tedious tasks for you. It seems like we are 00:14:30.340 --> 00:14:33.159 moving from chat with a bot to assign a bot a 00:14:33.159 --> 00:14:36.700 job. That is the defining theme of 2026. Don't 00:14:36.700 --> 00:14:38.720 talk to me, do it for me, even for content creation. 00:14:39.000 --> 00:14:42.100 There's Reloop, avatars, cloned voices, video 00:14:42.100 --> 00:14:44.929 editing all in one spot. You just chat with it 00:14:44.929 --> 00:14:47.570 to make a full video. It's moving so fast. I 00:14:47.570 --> 00:14:49.750 saw that a Wharton professor released the eighth 00:14:49.750 --> 00:14:52.289 edition of his AI guide this week. The eighth 00:14:52.289 --> 00:14:54.549 edition. It just highlights that textbooks are 00:14:54.549 --> 00:14:57.129 obsolete the absolute moment they're printed. 00:14:57.230 --> 00:14:59.049 Absolutely. If you're reading a printed book 00:14:59.049 --> 00:15:01.590 on how to use AI, you're studying history. You 00:15:01.590 --> 00:15:04.200 are looking at fossils. So, faithfully focusing 00:15:04.200 --> 00:15:07.039 on 3D worlds, is that the next frontier after 00:15:07.039 --> 00:15:09.320 text and video? Absolutely. Spatial computing 00:15:09.320 --> 00:15:12.019 needs content, and only AI can build it fast 00:15:12.019 --> 00:15:14.440 enough. Spatial computing needs content, and 00:15:14.440 --> 00:15:17.740 only AI can build it fast. I love that. Let's 00:15:17.740 --> 00:15:19.059 try to pull this all together. We've covered 00:15:19.059 --> 00:15:21.799 a lot of ground today. What is the big idea here 00:15:21.799 --> 00:15:24.259 for the listener? I think if we weave these threads 00:15:24.259 --> 00:15:27.500 together, we see a clear picture of 2026. First 00:15:27.500 --> 00:15:30.799 is capability. Gemini 3 .1 proves that models 00:15:30.799 --> 00:15:33.419 are learning to reason, not just predict. They 00:15:33.419 --> 00:15:36.539 are becoming methodical thinkers. Second is risk. 00:15:37.039 --> 00:15:39.840 That reasoning power allows for autonomous offense. 00:15:40.220 --> 00:15:43.179 The tools are becoming weaponized much faster 00:15:43.179 --> 00:15:46.019 than they are being secured. And third? The economy. 00:15:46.460 --> 00:15:49.139 The market values this power at nearly a trillion 00:15:49.139 --> 00:15:51.980 dollars, but the business model is still totally 00:15:51.980 --> 00:15:55.220 fractured. We're seeing a split between the ads 00:15:55.220 --> 00:15:57.639 model where you are the product and the subscription 00:15:57.639 --> 00:16:00.419 model where you pay for privacy and neutrality. 00:16:00.580 --> 00:16:02.500 It's a pivotal moment. You know, I keep thinking 00:16:02.500 --> 00:16:05.700 about that gap, the offense versus defense gap 00:16:05.700 --> 00:16:07.860 in cogeneration. It is the most critical metric 00:16:07.860 --> 00:16:09.980 right now. It isn't just a crypto problem, is 00:16:09.980 --> 00:16:12.769 it? No. It's everything. Banking, healthcare, 00:16:13.110 --> 00:16:15.990 power grids. It's a preview of all cybersecurity. 00:16:16.929 --> 00:16:20.330 Two sec silence. If the AI can pick the locks 00:16:20.330 --> 00:16:22.990 72 % of the time, it doesn't matter how strong 00:16:22.990 --> 00:16:25.490 the door is. We need better locks and we need 00:16:25.490 --> 00:16:28.370 them fast. We do. And we likely need AI to build 00:16:28.370 --> 00:16:31.570 them. That is the reality of February 2026. It 00:16:31.570 --> 00:16:33.690 is a world of incredible speed, high stakes, 00:16:33.850 --> 00:16:36.009 and brilliant machines that are learning to think. 00:16:36.210 --> 00:16:38.230 Thank you for diving in with us today. Thanks 00:16:38.230 --> 00:16:41.029 for having me. And to you listening. Stay curious, 00:16:41.250 --> 00:16:43.629 stay vigilant, and maybe double -check your digital 00:16:43.629 --> 00:16:46.049 locks today. We will catch you in the next deep 00:16:46.049 --> 00:16:47.629 dive. OGAU Music.