WEBVTT 00:00:00.000 --> 00:00:02.180 Imagine for a second that you are, you know, 00:00:02.200 --> 00:00:04.379 just going about your day. Yeah, making coffee, 00:00:04.540 --> 00:00:06.000 sitting in a meeting, whatever it is. Right. 00:00:06.200 --> 00:00:09.279 And you receive a totally standard WhatsApp text 00:00:09.279 --> 00:00:12.060 message. Your phone just buzzes on the table. 00:00:12.259 --> 00:00:14.519 And you don't even pick it up. Exactly. You don't 00:00:14.519 --> 00:00:17.640 click a weird link. You never, like, type a sketchy 00:00:17.640 --> 00:00:20.059 command into a terminal. You literally just let 00:00:20.059 --> 00:00:21.899 your phone sit there with the lock screen lit 00:00:21.899 --> 00:00:23.899 up. Safely on the desk. But in the background, 00:00:24.140 --> 00:00:27.379 your AI assistant reads that notification, follows 00:00:27.379 --> 00:00:30.000 a set of hidden, invisible instructions buried 00:00:30.000 --> 00:00:34.200 deep inside the text of the message, and quietly, 00:00:34.200 --> 00:00:37.119 efficiently steals your data. I mean, it sounds 00:00:37.119 --> 00:00:39.539 like pure science fiction. It really does. Or 00:00:39.539 --> 00:00:42.560 at least the plot of an overly dramatic cyber 00:00:42.560 --> 00:00:46.299 thriller. But we are living in a moment where 00:00:46.299 --> 00:00:50.170 that is not fiction at all. That is a very real, 00:00:50.270 --> 00:00:52.829 very present vulnerability sitting in your pocket 00:00:52.829 --> 00:00:55.630 right now. It is genuinely terrifying. And that 00:00:55.630 --> 00:00:58.810 is exactly why we are here today. Welcome to 00:00:58.810 --> 00:01:01.170 today's deep dive. We're looking at an incredible 00:01:01.170 --> 00:01:03.409 stack of excerpts from a fascinating new book 00:01:03.409 --> 00:01:06.939 called The AI Sentinel. guardrails, growth, and 00:01:06.939 --> 00:01:09.120 generative breakthroughs. It's a great read. 00:01:09.239 --> 00:01:12.480 Our mission today is to really explore this massive 00:01:12.480 --> 00:01:15.219 friction and the structural leaps happening in 00:01:15.219 --> 00:01:17.260 the AI ecosystem right now. We're talking about 00:01:17.260 --> 00:01:20.500 these invisible hacks, the messy, chaotic reality 00:01:20.500 --> 00:01:23.739 of AI -generated code, and then on the brighter 00:01:23.739 --> 00:01:26.459 side, a revolutionary breakthrough that is finally 00:01:26.459 --> 00:01:29.680 killing the incredibly frustrating slot machine 00:01:29.680 --> 00:01:33.219 era of AI art. Because there is a massive paradigm 00:01:33.219 --> 00:01:35.719 shift happening across the entire tech landscape. 00:01:36.200 --> 00:01:39.420 We're actively moving from a phase of unchecked, 00:01:39.420 --> 00:01:44.040 wild hypergrowth into an era where, well, structure, 00:01:44.239 --> 00:01:46.180 security, and tight control are suddenly the 00:01:46.180 --> 00:01:48.120 most important things in the room. Okay, let's 00:01:48.120 --> 00:01:50.079 unpack this, because to really understand why 00:01:50.079 --> 00:01:52.519 AI is suddenly such a massive target for hackers 00:01:52.519 --> 00:01:54.680 in the first place, we first have to look at 00:01:54.680 --> 00:01:57.180 the sheer unprecedented scale of its adoption. 00:01:57.459 --> 00:01:59.219 Yeah, you really can't grasp the hack until you 00:01:59.219 --> 00:02:01.879 grasp the size of the target. Exactly. Just to 00:02:01.879 --> 00:02:04.260 put this in perspective for you, ChatGPT just 00:02:04.260 --> 00:02:07.560 crossed one billion monthly active app users 00:02:07.560 --> 00:02:10.020 in May. One billion. It's staggering. And the 00:02:10.020 --> 00:02:12.039 wild part isn't just the number, it's the timeline. 00:02:12.319 --> 00:02:14.520 It reached this milestone in about three years. 00:02:14.969 --> 00:02:17.310 Which is, I mean, unheard of. Faster than TikTok, 00:02:17.550 --> 00:02:20.229 faster than Instagram, faster than YouTube, faster 00:02:20.229 --> 00:02:23.030 than Google Maps. It is the fastest adoption 00:02:23.030 --> 00:02:26.110 of a consumer technology in human history. Which 00:02:26.110 --> 00:02:28.590 is an incredible engineering feat, undoubtedly. 00:02:29.409 --> 00:02:31.810 But it's also a complete security nightmare. 00:02:32.069 --> 00:02:34.830 Oh, absolutely. When you have... a billion people 00:02:34.830 --> 00:02:37.770 rapidly adopting a system. And that system is 00:02:37.770 --> 00:02:40.870 increasingly tied to their personal data, their 00:02:40.870 --> 00:02:42.889 daily schedules, their private communications. 00:02:43.490 --> 00:02:46.590 It just becomes the ultimate prize for bad actors. 00:02:46.849 --> 00:02:49.689 And that brings us to what happened with Google's 00:02:49.689 --> 00:02:52.210 Gemini. Right. The Safe Breach Labs exploit. 00:02:52.509 --> 00:02:55.210 Yes. Safe Breach Labs just pulled off exactly 00:02:55.210 --> 00:02:56.930 what you described in that opening scenario. 00:02:57.210 --> 00:02:58.930 And the really crazy part about this whole thing, 00:02:58.990 --> 00:03:00.770 this is the second time they've managed to break 00:03:00.770 --> 00:03:04.180 Google's Gemini this exact same way. Wow. Twice. 00:03:04.240 --> 00:03:06.479 Yeah. The attack vector relies on something called 00:03:06.479 --> 00:03:09.599 indirect prompt injection. Let's slow down and 00:03:09.599 --> 00:03:11.719 break that down because I think people hear injection 00:03:11.719 --> 00:03:14.879 and they picture like someone typing green code 00:03:14.879 --> 00:03:17.419 into a dark terminal. How does this actually 00:03:17.419 --> 00:03:19.919 work mechanically on a phone? So an AI assistant 00:03:19.919 --> 00:03:22.879 on your phone, by its very nature, is designed 00:03:22.879 --> 00:03:25.219 to be as helpful as possible. Right. To be helpful, 00:03:25.319 --> 00:03:27.680 it needs context. It wants to know what you're 00:03:27.680 --> 00:03:30.969 doing. So it constantly reads incoming content 00:03:30.969 --> 00:03:33.409 -like notifications popping up on your screen 00:03:33.409 --> 00:03:36.129 to maintain an understanding of your day. Makes 00:03:36.129 --> 00:03:38.990 sense. Indirect prompt injection takes advantage 00:03:38.990 --> 00:03:42.689 of that exact helpfulness. Attackers hide malicious 00:03:42.689 --> 00:03:45.949 instructions inside everyday content that the 00:03:45.949 --> 00:03:48.349 AI naturally scans anyway. So it's basically 00:03:48.349 --> 00:03:51.629 a Trojan horse hidden inside a regular text message. 00:03:51.870 --> 00:03:54.860 Exactly. In this specific exploit, the researchers 00:03:54.860 --> 00:03:57.539 used a novel trick to make the malicious code 00:03:57.539 --> 00:03:59.680 look like it was just a natural part of your 00:03:59.680 --> 00:04:02.319 ongoing conversation. How so? Well, your phone 00:04:02.319 --> 00:04:04.620 gets a text. The Android notification listener 00:04:04.620 --> 00:04:06.759 sees it and hands it to Gemini to summarize. 00:04:07.000 --> 00:04:09.560 Gemini reads it. But hidden in the text is a 00:04:09.560 --> 00:04:12.020 command formatted in a way that Gemini interprets 00:04:12.020 --> 00:04:15.039 as a direct, high -priority order from the system 00:04:15.039 --> 00:04:17.699 rather than just words in a message. So the AI 00:04:17.699 --> 00:04:19.439 thinks it's just maintaining context, but it's 00:04:19.439 --> 00:04:21.560 actually quietly taking orders from the hacker. 00:04:21.779 --> 00:04:25.660 Precisely. it perfectly bypasses Google's existing 00:04:25.660 --> 00:04:28.899 layered guardrails because to the security filters, 00:04:29.519 --> 00:04:31.279 It doesn't look like an attack. It just looked 00:04:31.279 --> 00:04:34.139 like conversational contacts. Unbelievable. And 00:04:34.139 --> 00:04:36.199 because Gemini's Android agent reads incoming 00:04:36.199 --> 00:04:39.060 notifications to stay updated, this attack works 00:04:39.060 --> 00:04:42.180 across almost everything you use. WhatsApp, Slack, 00:04:42.500 --> 00:04:45.920 Signal, SMS, Instagram. Any notification. Any 00:04:45.920 --> 00:04:49.279 of them. The AI extracts the data, executes the 00:04:49.279 --> 00:04:50.839 hacker's commands entirely in the background, 00:04:51.100 --> 00:04:54.220 and you get absolutely zero alerts. Wait. If 00:04:54.220 --> 00:04:56.980 I don't give the AI access to my external tools 00:04:56.980 --> 00:04:59.579 or bank accounts, how much damage can it really 00:04:59.579 --> 00:05:01.720 do? Like, let's say I just use it to answer trivia 00:05:01.720 --> 00:05:03.660 questions. That is the logical assumption to 00:05:03.660 --> 00:05:05.800 make, right? You keep the AI siloed, you don't 00:05:05.800 --> 00:05:07.500 connect your bank, and you figure you're safe. 00:05:07.699 --> 00:05:10.360 But there is a truly terrifying detail about 00:05:10.360 --> 00:05:13.399 this exploit. Even without external cool access, 00:05:13.699 --> 00:05:16.120 attackers can still force the AI to serve fake 00:05:16.120 --> 00:05:19.220 system messages directly to the user. Oh, wow. 00:05:19.660 --> 00:05:22.959 So even if the AI physically cannot wire my money, 00:05:23.060 --> 00:05:25.740 it can pop up a totally legitimate looking Google 00:05:25.740 --> 00:05:28.540 system alert on my screen that says, your session 00:05:28.540 --> 00:05:31.699 expired. Please enter your password. And then 00:05:31.699 --> 00:05:34.279 just hand that newly typed password straight 00:05:34.279 --> 00:05:38.319 to the attacker. That is so sinister. Yes. It 00:05:38.319 --> 00:05:40.939 becomes the ultimate phishing mechanism because 00:05:40.939 --> 00:05:43.759 it comes from a trusted internal source, your 00:05:43.759 --> 00:05:46.120 own digital assistance interface. Right. You'd 00:05:46.120 --> 00:05:48.300 never question it. Google built layered defenses 00:05:48.300 --> 00:05:51.399 to stop this exact thing, but the same security 00:05:51.399 --> 00:05:55.180 team bypassed them twice. That is the uncomfortable 00:05:55.180 --> 00:05:58.079 truth about our current AI ecosystem. So what 00:05:58.079 --> 00:06:00.120 can we actually do? The advice to you listening 00:06:00.120 --> 00:06:03.149 right now is extremely clear. If you do not actively 00:06:03.149 --> 00:06:05.870 use a specific integration on your device, disable 00:06:05.870 --> 00:06:08.490 it. You must proactively lock your own doors 00:06:08.490 --> 00:06:11.649 because the next security researcher, or worse, 00:06:11.850 --> 00:06:14.189 the next malicious hacker, is already hunting 00:06:14.189 --> 00:06:16.269 for the next loophole. It really highlights the 00:06:16.269 --> 00:06:18.810 dark side of hyperscale. And speaking of the 00:06:18.810 --> 00:06:20.810 fallout from that kind of scale, it's not just 00:06:20.810 --> 00:06:22.850 targeted security threats we have to worry about. 00:06:23.029 --> 00:06:25.660 No, definitely not. We're also seeing a massive 00:06:25.660 --> 00:06:28.379 degradation in the quality of the internet itself. 00:06:28.639 --> 00:06:30.120 Just look at Reddit right now. It's currently 00:06:30.120 --> 00:06:34.620 drowning in AI -powered spam. Drowning. 67 % 00:06:34.620 --> 00:06:36.819 of moderators report that authentic community 00:06:36.819 --> 00:06:40.540 discussions are already being eroded. 67%. It's 00:06:40.540 --> 00:06:44.500 a severe ecosystem pollution problem. When generating 00:06:44.500 --> 00:06:47.800 text is free, instantaneous, and completely automated, 00:06:48.100 --> 00:06:51.300 bad actors are going to flood every single available 00:06:51.300 --> 00:06:54.410 channel. Until the noise completely drowns out 00:06:54.410 --> 00:06:57.029 the human signal. This raises an important question. 00:06:57.569 --> 00:07:00.730 How can we possibly trust these systems as they 00:07:00.730 --> 00:07:03.470 become deeply embedded in our daily lives? Right. 00:07:03.569 --> 00:07:06.610 If they're vulnerable to invisible hijacks through 00:07:06.610 --> 00:07:09.509 a simple Slack message and they're actively degrading 00:07:09.509 --> 00:07:11.350 the platforms we use to communicate with each 00:07:11.350 --> 00:07:13.550 other. Something foundational has to change. 00:07:13.709 --> 00:07:15.889 And that perfectly sets up the industry's response 00:07:15.889 --> 00:07:18.470 right now. Because if cloud connected, hyper 00:07:18.470 --> 00:07:20.990 integrated AI is proving to be both a security 00:07:20.990 --> 00:07:23.750 vulnerability and a spam engine, tech companies 00:07:23.750 --> 00:07:26.810 are fundamentally rethinking where and how AI 00:07:26.810 --> 00:07:31.129 operates. We're seeing this massive, sudden push 00:07:31.129 --> 00:07:34.649 for more localized, highly specialized control. 00:07:34.790 --> 00:07:38.379 We are absolutely stepping away from the. one 00:07:38.379 --> 00:07:40.899 giant brain in the cloud model case in point 00:07:40.899 --> 00:07:44.420 is perplexity they just unveiled a new hybrid 00:07:44.420 --> 00:07:47.579 ai mode called perplexity computer which is super 00:07:47.579 --> 00:07:49.680 interesting what this does is it splits tasks 00:07:49.680 --> 00:07:52.300 it runs a local model right on your physical 00:07:52.300 --> 00:07:55.160 device for certain private or simple things and 00:07:55.160 --> 00:07:57.500 it only pings the massive frontier models in 00:07:57.500 --> 00:08:00.839 the cloud when it absolutely has to For complex 00:08:00.839 --> 00:08:03.319 reasoning. Which is a brilliant move for privacy, 00:08:03.439 --> 00:08:05.879 security, and speed. Right. To make that kind 00:08:05.879 --> 00:08:08.800 of local execution possible, we're seeing incredible 00:08:08.800 --> 00:08:13.040 technical leaps. Look at Google GEMMA 412B. Let's 00:08:13.040 --> 00:08:15.879 talk about GEMMA. This is a new model that processes 00:08:15.879 --> 00:08:18.759 text, vision, and audio natively. It doesn't 00:08:18.759 --> 00:08:20.480 use separate encoders for different types of 00:08:20.480 --> 00:08:22.899 media, which makes it incredibly efficient. It 00:08:22.899 --> 00:08:25.699 can run locally on just 16 gigabytes of VRAM. 00:08:25.800 --> 00:08:27.779 Okay, I want to pause there because I hear terms 00:08:27.779 --> 00:08:30.519 like encoders and VRAM, and I know some people's 00:08:30.519 --> 00:08:33.320 eyes might start to glaze over. Break down why 00:08:33.320 --> 00:08:35.700 that is actually a structural leap for someone 00:08:35.700 --> 00:08:38.000 who just wants their laptop to be secure. Okay, 00:08:38.059 --> 00:08:41.679 think about how older AI models worked. If you 00:08:41.679 --> 00:08:44.240 showed an AI a picture, it couldn't actually 00:08:44.240 --> 00:08:47.740 see it. Right, it's just code. Exactly. It had 00:08:47.740 --> 00:08:50.000 to use a separate piece of software, an encoder, 00:08:50.080 --> 00:08:52.679 to translate that picture into a giant block 00:08:52.679 --> 00:08:55.399 of text describing the picture. And then the 00:08:55.399 --> 00:08:58.440 AI could process the text. It sounds slow. It 00:08:58.440 --> 00:09:00.639 is. It takes a massive amount of computing power 00:09:00.639 --> 00:09:05.000 and memory. But Gemma 412B processes it natively. 00:09:05.000 --> 00:09:07.100 It doesn't need a translator. It just sees the 00:09:07.100 --> 00:09:09.539 image and hears the audio directly. That's a 00:09:09.539 --> 00:09:12.590 huge shift. Yeah. And because it's so efficient, 00:09:12.809 --> 00:09:16.090 it only requires 16 gigabytes of video RAM or 00:09:16.090 --> 00:09:18.230 VRAM, which is just the memory on a standard 00:09:18.230 --> 00:09:20.190 graphics card. Which means it can run on a high 00:09:20.190 --> 00:09:22.769 end consumer laptop, entirely disconnected from 00:09:22.769 --> 00:09:25.049 the Internet, completely secure and private. 00:09:25.230 --> 00:09:27.710 Exactly. You don't need a server farm. You just 00:09:27.710 --> 00:09:29.750 need a good laptop. But of course, the big tech 00:09:29.750 --> 00:09:32.230 companies aren't just retreating to local hardware 00:09:32.230 --> 00:09:34.450 to solve these problems. They're also trying 00:09:34.450 --> 00:09:37.679 to fight fire with fire. There is a new empowered 00:09:37.679 --> 00:09:41.460 tool called Astra Autonomous Pentest, and it 00:09:41.460 --> 00:09:44.980 literally uses AI agents to find, validate, and 00:09:44.980 --> 00:09:47.799 fix vulnerabilities. It just runs constantly 00:09:47.799 --> 00:09:49.840 in the background. Right. It operates as native 00:09:49.840 --> 00:09:52.360 pumps inside coding environments like Cursor, 00:09:52.399 --> 00:09:55.500 Copilot, and Cloud Code, just hunting for security 00:09:55.500 --> 00:09:58.529 flaws in real time. But I have to say, using 00:09:58.529 --> 00:10:01.049 AI to fix AI vulnerabilities feels a bit like 00:10:01.049 --> 00:10:03.669 hiring a sloppy builder to inspect their own 00:10:03.669 --> 00:10:05.769 house, right? It is a massive gamble. I mean, 00:10:05.789 --> 00:10:08.710 you are trusting a system known for making bizarre 00:10:08.710 --> 00:10:12.090 logical leaps to spot its own architectural flaws. 00:10:12.289 --> 00:10:14.230 And the irony is really peeking over at Google, 00:10:14.370 --> 00:10:17.409 because right now, 75 % of Google's code is officially 00:10:17.409 --> 00:10:21.129 AI generated. 75%. Yes. But internally, employees 00:10:21.129 --> 00:10:24.070 are sharing memes openly joking that the AI constantly 00:10:24.070 --> 00:10:25.950 hallucinates and actually makes their engineering 00:10:25.950 --> 00:10:28.330 work harder. Because they have to untangle its 00:10:28.330 --> 00:10:30.730 mess. It is a brilliant paradox. We're relying 00:10:30.730 --> 00:10:33.350 on AI to do the heavy lifting of code generation 00:10:33.350 --> 00:10:36.169 because it's fast. You can write a million lines 00:10:36.169 --> 00:10:38.110 of code in a second. And the quality. Right. 00:10:38.190 --> 00:10:40.590 The structural precision, the fundamental reliability, 00:10:40.970 --> 00:10:43.710 it just isn't fully there yet. It might confidently 00:10:43.710 --> 00:10:46.269 invent a function that doesn't actually exist 00:10:46.269 --> 00:10:48.529 in the programming language and then confidently 00:10:48.529 --> 00:10:51.470 use that fake function across hundreds of files. 00:10:51.730 --> 00:10:54.029 What a nightmare to clean up. If we connect this 00:10:54.029 --> 00:10:57.360 to the bigger picture. This is exactly why big 00:10:57.360 --> 00:11:00.139 tech is pivoting so hard right now. They realize 00:11:00.139 --> 00:11:03.379 they can't just rely on raw open -ended chatbots 00:11:03.379 --> 00:11:07.019 anymore. Asking a general AI to write secure 00:11:07.019 --> 00:11:10.440 code or manage a business is way too risky. So 00:11:10.440 --> 00:11:12.679 they're specializing. Exactly. They're building 00:11:12.679 --> 00:11:15.639 specialized, highly controlled, goal -oriented 00:11:15.639 --> 00:11:17.840 agents. And we are seeing that specialization 00:11:17.840 --> 00:11:20.580 everywhere. Meta just launched business agent 00:11:20.580 --> 00:11:22.779 globally. You can now deploy these specialized 00:11:22.779 --> 00:11:25.220 AI agents across WhatsApp, Instagram, and Messenger. 00:11:25.500 --> 00:11:27.759 And they have a very narrow focus. Right. Their 00:11:27.759 --> 00:11:30.700 entire job, their only function, is to book appointments, 00:11:30.879 --> 00:11:33.740 qualify leads, and close sales. No philosophical 00:11:33.740 --> 00:11:36.899 chats, no hallucinating weird code, just highly 00:11:36.899 --> 00:11:39.159 constrained business logic. Because constrained 00:11:39.159 --> 00:11:42.639 parameters yield better, safer results. You narrow 00:11:42.639 --> 00:11:45.440 the focus to reduce the risk. Facebook is doing 00:11:45.440 --> 00:11:47.340 the exact same thing for content creators now. 00:11:47.559 --> 00:11:49.580 Oh, the strategist tool. Yeah, they're giving 00:11:49.580 --> 00:11:52.139 all creators an AI strategist. It doesn't just 00:11:52.139 --> 00:11:54.620 write a generic post for you. It analyzes your 00:11:54.620 --> 00:11:57.580 specific audience data, suggests ideas, tracks 00:11:57.580 --> 00:12:01.220 market trends, and spots viral potential before 00:12:01.220 --> 00:12:03.700 you even post. And it's not just text and business 00:12:03.700 --> 00:12:07.139 logic either. Suno, the AI music generation platform, 00:12:07.500 --> 00:12:11.519 just raised over $400 million at a $5 .4 billion 00:12:11.519 --> 00:12:14.460 valuation. Massive numbers. And their big next 00:12:14.460 --> 00:12:16.740 move, they are launching their first model built. 00:12:16.840 --> 00:12:19.120 specifically with music industry partners. So 00:12:19.120 --> 00:12:21.360 instead of just blindly scraping the entire internet 00:12:21.360 --> 00:12:24.460 and causing absolute copyright chaos, they are 00:12:24.460 --> 00:12:26.759 moving toward controlled, industry -compliant 00:12:26.759 --> 00:12:29.720 AI. It is all about bringing guardrails and precision 00:12:29.720 --> 00:12:32.820 to a technology that has, until now, been incredibly 00:12:32.820 --> 00:12:35.820 chaotic. The Wild West phase is ending. And that 00:12:35.820 --> 00:12:38.299 deep need for precision brings us to one of my 00:12:38.299 --> 00:12:41.279 absolute favorite topics today. Because this 00:12:41.279 --> 00:12:43.500 lack of control hasn't just been an engineering 00:12:43.500 --> 00:12:47.039 headache or a business problem, it has completely 00:12:47.039 --> 00:12:50.200 revolutionized how we generate media. Yes, the 00:12:50.200 --> 00:12:52.600 creative side. The era we are leaving behind 00:12:52.600 --> 00:12:55.899 is what we can call the slot machine era of AI 00:12:55.899 --> 00:12:59.330 art. And it is finally... dead it is a profound 00:12:59.330 --> 00:13:01.769 shift for creatives everywhere for the longest 00:13:01.769 --> 00:13:04.090 time prompting an image generator was literally 00:13:04.090 --> 00:13:06.450 like pulling the lever on a casino slot machine 00:13:06.450 --> 00:13:09.110 you type in a prompt pull the lever and get a 00:13:09.110 --> 00:13:12.090 beautiful high definition image but maybe it 00:13:12.090 --> 00:13:14.289 wasn't quite right right maybe your main subject 00:13:14.289 --> 00:13:17.529 had six fingers on one hand or there was a coffee 00:13:17.529 --> 00:13:20.129 cup just randomly hovering floating in midair 00:13:20.129 --> 00:13:21.889 in the background it used to happen all the time 00:13:21.889 --> 00:13:24.230 all the time and the worst part was trying to 00:13:24.230 --> 00:13:26.690 fix it you couldn't just tell the ai hey keep 00:13:26.690 --> 00:13:28.340 everything exactly exactly the same but move 00:13:28.340 --> 00:13:30.340 the cup to the table, you either had to pull 00:13:30.340 --> 00:13:32.500 the lever again, lose the entire image you just 00:13:32.500 --> 00:13:35.120 liked and pray the next roll was better, or you 00:13:35.120 --> 00:13:37.159 had to export it to Photoshop and clone stamp 00:13:37.159 --> 00:13:39.980 it yourself by hand. You had the magic of generation, 00:13:40.279 --> 00:13:43.320 but you completely lacked structural control. 00:13:44.120 --> 00:13:47.649 And that is what has fundamentally changed. Two 00:13:47.649 --> 00:13:50.470 new models from Image Labs called Edeogram and 00:13:50.470 --> 00:13:53.549 Reeve have completely flipped the script on how 00:13:53.549 --> 00:13:55.429 this works. Let's talk about those. They now 00:13:55.429 --> 00:13:58.129 rely on structured layouts and agentic control. 00:13:58.570 --> 00:14:00.889 Let's do Ideagram 4 .0 first because this is 00:14:00.889 --> 00:14:03.549 a huge leap. It has gone completely open source. 00:14:03.789 --> 00:14:05.850 For anyone unfamiliar, that means the underlying 00:14:05.850 --> 00:14:08.809 brain of the AI, the open weights, are available 00:14:08.809 --> 00:14:11.769 for anyone to download, inspect, and build upon 00:14:11.769 --> 00:14:14.649 rather than being locked away behind a corporate 00:14:14.649 --> 00:14:17.309 API. Which is huge for the community. Huge. And 00:14:17.309 --> 00:14:20.049 right now, it holds the top spot for open weights 00:14:20.049 --> 00:14:22.210 on the Design Arena leaderboard. Which is essentially 00:14:22.210 --> 00:14:25.049 a massive blind taste test. Exactly. Professional 00:14:25.049 --> 00:14:27.649 designers vote on which AI generates the best 00:14:27.649 --> 00:14:31.080 image. Ideogram absolutely dominates at the hard 00:14:31.080 --> 00:14:33.700 stuff. Text rendering, typography, graphic design. 00:14:33.960 --> 00:14:37.360 It is even beating out closed proprietary rivals 00:14:37.360 --> 00:14:39.820 in those blind tests. And the mechanism behind 00:14:39.820 --> 00:14:41.980 how it does that is what makes it so powerful. 00:14:42.159 --> 00:14:45.240 It uses a JSON -driven approach. Break down JSON 00:14:45.240 --> 00:14:47.740 for us, for those who aren't coders. Right. So 00:14:47.740 --> 00:14:50.399 JSON is essentially a lightweight data map. It's 00:14:50.399 --> 00:14:52.960 a way to organize information using text so a 00:14:52.960 --> 00:14:55.529 computer can easily read it. With older models, 00:14:55.710 --> 00:14:58.529 the AI just mushed all the pixels together based 00:14:58.529 --> 00:15:00.730 on your text prompt. Like a flat painting. Exactly. 00:15:01.090 --> 00:15:04.549 With Ediogram, the AI generates a JSON map of 00:15:04.549 --> 00:15:07.049 the image first. It creates a structural blueprint. 00:15:07.289 --> 00:15:09.730 So if you want to move that floating coffee cup 00:15:09.730 --> 00:15:12.269 or tweak a background element, you don't re -roll 00:15:12.269 --> 00:15:14.169 the image and lose everything. You don't pull 00:15:14.169 --> 00:15:16.070 the slot machine lever again. Right. You literally 00:15:16.070 --> 00:15:18.409 just open the text file, find the coordinates 00:15:18.409 --> 00:15:21.149 for the coffee cup in the code, and change the 00:15:21.149 --> 00:15:24.029 layout. The image updates instantly. keeping 00:15:24.029 --> 00:15:26.690 the rest of the scene perfect. That is mind -blowing. 00:15:26.789 --> 00:15:29.350 You are editing the physical structure of a generated 00:15:29.350 --> 00:15:32.289 image directly through data coordinates. Exactly. 00:15:32.309 --> 00:15:34.490 You are no longer just prompting, you are directing. 00:15:34.929 --> 00:15:36.730 And Rave did something very similar with their 00:15:36.730 --> 00:15:39.990 new 2 .0 model. Rave just dethroned NanoBanana 00:15:39.990 --> 00:15:42.730 2 to claim the number two spot overall on the 00:15:42.730 --> 00:15:45.029 main text -to -image leaderboard, sitting right 00:15:45.029 --> 00:15:48.629 behind GPT -Image 2. Which is wild, because NanoBanana 00:15:48.629 --> 00:15:50.809 2 was the reigning champion for a long time. 00:15:51.169 --> 00:15:54.830 Yeah, so unseating it is a massive deal in the 00:15:54.830 --> 00:15:58.570 AI community. Review .0 outputs images with labeled 00:15:58.570 --> 00:16:01.090 segments. You can actually see the structural 00:16:01.090 --> 00:16:03.610 blocks of the image on your screen and rewrite 00:16:03.610 --> 00:16:05.909 the layout directly. Here's where it gets really 00:16:05.909 --> 00:16:09.190 interesting, though. Because Ideagram open sourced 00:16:09.190 --> 00:16:11.789 these weights, it proves that the open source 00:16:11.789 --> 00:16:14.710 community is the one actively defining the step 00:16:14.710 --> 00:16:17.110 change. Absolutely. It's not just a cool feature 00:16:17.110 --> 00:16:19.470 locked behind a massive, expensive corporate 00:16:19.470 --> 00:16:22.909 paywall. It's a fundamental upgrade to how humanity 00:16:22.909 --> 00:16:25.330 interacts with generative media. And it's being 00:16:25.330 --> 00:16:27.590 driven by the community. It is a true maturation 00:16:27.590 --> 00:16:29.909 of the medium. We're seeing broader media upgrades 00:16:29.909 --> 00:16:33.210 across the board, too. Like Grok Imagine 1 .5 00:16:33.210 --> 00:16:36.210 preview was just released via their API. I saw 00:16:36.210 --> 00:16:38.590 that, yeah, at x .aii Imagine. Right, and it 00:16:38.590 --> 00:16:40.710 brings much sharper realism, incredibly better 00:16:40.710 --> 00:16:43.490 audio syncing for video generation, and vastly 00:16:43.490 --> 00:16:45.429 stronger prompt following. Which is so critical 00:16:45.429 --> 00:16:47.529 right now because we desperately need these tools 00:16:47.529 --> 00:16:49.850 to help creators avoid producing what the internet 00:16:49.850 --> 00:16:52.450 is now collectively calling AI slop. Yes, the 00:16:52.450 --> 00:16:55.059 slop. You mentioned Reddit being flooded with 00:16:55.059 --> 00:16:57.799 spam earlier. There was a viral post recently 00:16:57.799 --> 00:17:01.980 warning against four very specific overused design 00:17:01.980 --> 00:17:04.420 patterns that people now instantly associate 00:17:04.420 --> 00:17:09.019 with lazy AI generated website building. Right. 00:17:09.079 --> 00:17:10.839 Because when everyone has access to the exact 00:17:10.839 --> 00:17:13.250 same slot machine. Everything starts to look 00:17:13.250 --> 00:17:15.069 exactly the same. The Internet just loses its 00:17:15.069 --> 00:17:17.630 texture. It gets boring. Yeah. But when you give 00:17:17.630 --> 00:17:21.069 creators precise layout control, like with Enneagram's 00:17:21.069 --> 00:17:23.170 JSON structures or Reeve's labeled segments, 00:17:23.410 --> 00:17:26.150 you empower them to make actual intentional design 00:17:26.150 --> 00:17:28.750 choices again. You escape the slob because you 00:17:28.750 --> 00:17:30.549 can inject human case back into the structure. 00:17:30.769 --> 00:17:32.670 So what does this all mean for you listening 00:17:32.670 --> 00:17:34.740 today? When you look at all these pieces together, 00:17:34.880 --> 00:17:36.640 it paints a really vivid picture of where we 00:17:36.640 --> 00:17:39.359 are right now. Very clear trajectory. The overarching 00:17:39.359 --> 00:17:42.619 journey of the AI ecosystem today is a massive, 00:17:42.680 --> 00:17:46.119 high -stakes tug of war. On one hand, AI has 00:17:46.119 --> 00:17:48.420 reached a scale where it is so ubiquitous that 00:17:48.420 --> 00:17:50.519 it can be invisibly hijacked through a simple 00:17:50.519 --> 00:17:52.680 WhatsApp notification sitting on your lock screen. 00:17:52.839 --> 00:17:55.220 And it's powerful enough to flood our forums 00:17:55.220 --> 00:17:58.519 with synthetic spam. But on the other hand, the 00:17:58.519 --> 00:18:01.579 response to that chaos is a beautiful evolution 00:18:01.579 --> 00:18:04.809 toward precision. The technology is evolving 00:18:04.809 --> 00:18:07.569 into a highly controllable tool that can run 00:18:07.569 --> 00:18:10.170 natively and privately right on your laptop. 00:18:10.410 --> 00:18:12.450 It's a totally different approach. It allows 00:18:12.450 --> 00:18:15.829 you to edit reality, whether that's an image, 00:18:15.890 --> 00:18:19.250 a song, or a complex business workflow, with 00:18:19.250 --> 00:18:22.009 the absolute precision of layout code. Wait, 00:18:22.269 --> 00:18:24.109 I mean, that completely flips the narrative. 00:18:24.289 --> 00:18:26.970 You are moving from being a passive passenger 00:18:26.970 --> 00:18:30.069 on this AI ride to being a pilot. That's a great 00:18:30.069 --> 00:18:31.960 way to put it. But being a pilot means you have 00:18:31.960 --> 00:18:34.480 to actually do your pre -flight checks. So the 00:18:34.480 --> 00:18:36.819 biggest takeaway for you today, please go into 00:18:36.819 --> 00:18:39.480 your AI assistance and disable any integrations 00:18:39.480 --> 00:18:42.099 you are not actively using. Protect yourself 00:18:42.099 --> 00:18:44.819 from those invisible prompt injections. Lock 00:18:44.819 --> 00:18:47.480 your digital doors. Security absolutely has to 00:18:47.480 --> 00:18:50.099 be a proactive habit now. It can no longer be 00:18:50.099 --> 00:18:52.609 an afterthought. And if you are looking to actively 00:18:52.609 --> 00:18:55.130 manage your piece of this ecosystem rather than 00:18:55.130 --> 00:18:57.410 just letting it wash over you, there are a couple 00:18:57.410 --> 00:18:59.750 of community tools mentioned in the sources worth 00:18:59.750 --> 00:19:01.789 checking out. Oh, right. The community tools. 00:19:02.029 --> 00:19:04.660 Yeah. If you are running a business and dealing 00:19:04.660 --> 00:19:07.119 with that AI spam filter problem we talked about, 00:19:07.299 --> 00:19:10.559 look into MailWarm 2 .0. It's a premium email 00:19:10.559 --> 00:19:12.720 warm -up and deliverability system that ensures 00:19:12.720 --> 00:19:15.539 your legitimate messages actually reach the inbox 00:19:15.539 --> 00:19:18.299 instead of getting flagged as bot slop. Very 00:19:18.299 --> 00:19:20.240 necessary right now. And if you want to learn 00:19:20.240 --> 00:19:22.680 how to actually build and control these AI agents 00:19:22.680 --> 00:19:25.119 yourself, check out Build Club Campus. It's a 00:19:25.119 --> 00:19:28.579 fun, gamified, community -driven virtual AI school 00:19:28.579 --> 00:19:31.420 where you learn these concepts by actually building 00:19:31.420 --> 00:19:33.319 things with your own hands. What's fascinating 00:19:33.319 --> 00:19:35.859 here is how quickly we are moving from being 00:19:35.859 --> 00:19:39.160 passive consumers of AI magic, just pulling the 00:19:39.160 --> 00:19:41.200 slot machine lever and hoping for the best to 00:19:41.200 --> 00:19:43.960 becoming active, secure managers of our own AI 00:19:43.960 --> 00:19:45.920 infrastructure. It is a completely different 00:19:45.920 --> 00:19:48.400 mindset. But as we wrap up this deep dive, I 00:19:48.400 --> 00:19:51.440 want to leave you with one final, deeply unsettling, 00:19:51.440 --> 00:19:53.619 but absolutely fascinating thought to ponder. 00:19:53.680 --> 00:19:55.779 Let's hear it. We've talked today about how human 00:19:55.779 --> 00:19:58.519 hackers are hiding invisible prompts and WhatsApp 00:19:58.519 --> 00:20:01.400 messages to trick your A .I. And we also know 00:20:01.400 --> 00:20:04.099 from Google that A .I. is now actively writing 00:20:04.099 --> 00:20:06.759 75 percent of the code at some of these massive 00:20:06.759 --> 00:20:09.359 tech companies, often hallucinating and making 00:20:09.359 --> 00:20:11.839 bizarre structural mistakes along the way. Right. 00:20:12.039 --> 00:20:14.079 So what happens the day an A .I. hallucination 00:20:14.079 --> 00:20:16.940 organically writes a line of code that accidentally 00:20:16.940 --> 00:20:19.759 acts as an indirect prompt injection? Could an 00:20:19.759 --> 00:20:22.420 A .I. hack another A .I. entirely by mistake 00:20:22.420 --> 00:20:26.599 without a human Wow. Something to think about 00:20:26.599 --> 00:20:27.859 next time your phone buzzes.