WEBVTT 00:00:00.000 --> 00:00:01.840 Tech giants are spending hundreds of millions 00:00:01.840 --> 00:00:04.900 of dollars on safety guardrails right now. But 00:00:04.900 --> 00:00:07.839 the wild thing is anyone with a standard laptop 00:00:07.839 --> 00:00:12.339 they can bypass those filters in under 10 minutes. 00:00:12.539 --> 00:00:14.939 Yeah. We're basically watching a complete structural 00:00:14.939 --> 00:00:18.420 collapse of foundational safety and it's happening 00:00:18.420 --> 00:00:20.879 entirely in plain sight. Welcome to the Deep 00:00:20.879 --> 00:00:22.539 Dive. I'm really glad you're here with us today. 00:00:22.660 --> 00:00:25.280 We are looking at a stack of sources that reveal 00:00:25.280 --> 00:00:28.780 a massive fracture in the AI landscape today. 00:00:28.920 --> 00:00:31.289 Right. It's a real dividing line. Exactly. So 00:00:31.289 --> 00:00:33.710 our mission today, we're going to explore the 00:00:33.710 --> 00:00:36.770 sudden collapse of open source safety and the 00:00:36.770 --> 00:00:39.490 growing cultural backlash against this new thing 00:00:39.490 --> 00:00:42.170 called vibe coding. Which is a whole situation 00:00:42.170 --> 00:00:44.609 on its own. It really is. Plus, we'll discuss 00:00:44.609 --> 00:00:47.530 some truly mysterious, almost human AI behaviors 00:00:47.530 --> 00:00:49.869 that are being analyzed at the Vatican, of all 00:00:49.869 --> 00:00:52.929 places. And we'll look at a sudden massive leap 00:00:52.929 --> 00:00:56.009 in commercial AI image generation. Today is really 00:00:56.009 --> 00:00:58.009 all about tension. We're looking at this incredible 00:00:58.009 --> 00:01:01.369 push and pull. You have raw, unfiltered technological 00:01:01.369 --> 00:01:04.329 power accelerating on one side and on the other 00:01:04.329 --> 00:01:07.870 side. Just our fragile, deeply human attempts 00:01:07.870 --> 00:01:10.930 to somehow maintain control of it all. Well, 00:01:11.010 --> 00:01:13.109 let's unpack this. We need to start with the 00:01:13.109 --> 00:01:15.689 most urgent tension in our sources today. It's 00:01:15.689 --> 00:01:18.530 this clash between the open source ecosystem 00:01:18.530 --> 00:01:21.739 and foundational safety. Yeah, the open source 00:01:21.739 --> 00:01:24.299 dilemma. Right. Because Meta and Google, they're 00:01:24.299 --> 00:01:26.959 pouring fortunes into making their models safe. 00:01:27.140 --> 00:01:29.420 They want to prevent the generation of harmful 00:01:29.420 --> 00:01:32.819 content. But researchers just used a tool called 00:01:32.819 --> 00:01:35.000 Heretic. Yeah, the Heretic tool. Right. They 00:01:35.000 --> 00:01:38.480 used it to download Meta's Llama 3 .3. And in 00:01:38.480 --> 00:01:40.560 minutes, they completely stripped its core safety 00:01:40.560 --> 00:01:43.260 filters. Just totally gone. The result of that 00:01:43.260 --> 00:01:46.040 specific test are genuinely unsettling. Yeah. 00:01:46.079 --> 00:01:47.900 I mean, once those internal filters were gone, 00:01:48.099 --> 00:01:50.900 the model readily generated step. by step instructions 00:01:50.900 --> 00:01:53.980 for biological weapons. Wow. Yeah. It happily 00:01:53.980 --> 00:01:56.280 provided the exact formulas that it was specifically 00:01:56.280 --> 00:01:58.939 trained to refuse. And it isn't just meta, right? 00:01:59.060 --> 00:02:02.120 I mean, Google is facing the exact same structural 00:02:02.120 --> 00:02:04.859 vulnerability here. Oh, absolutely. Researchers 00:02:04.859 --> 00:02:07.480 ran a separate test on Google's Gemma 3 and it 00:02:07.480 --> 00:02:10.060 produced the exact same alarming outputs. And 00:02:10.060 --> 00:02:12.680 didn't the creator of this heretic tool actually 00:02:12.680 --> 00:02:15.039 go a step further just to prove a point? He did. 00:02:15.219 --> 00:02:17.879 When Google released their newer model, Gemma 00:02:17.879 --> 00:02:21.449 4. He bypassed its guardrails within 90 minutes 00:02:21.449 --> 00:02:23.870 of the public release. 90 minutes? That is, well, 00:02:23.949 --> 00:02:25.689 it's barely enough time to read the release notes. 00:02:25.889 --> 00:02:27.650 It really highlights a fundamental architectural 00:02:27.650 --> 00:02:30.830 reality. If you can download the underlying weights 00:02:30.830 --> 00:02:33.870 of a model, you own it. So the closed models, 00:02:34.050 --> 00:02:37.610 like Clod or ChatGPT, they don't face this specific 00:02:37.610 --> 00:02:40.310 threat. Right, because outsiders simply can't 00:02:40.310 --> 00:02:42.650 access their core neural files to modify them. 00:02:42.969 --> 00:02:45.610 The open source models are completely exposed 00:02:45.610 --> 00:02:48.370 by design. I think about how hard it is to control 00:02:48.370 --> 00:02:50.650 these models on a good day. I mean, I still wrestle 00:02:50.650 --> 00:02:52.969 with prompt drift myself. Oh, totally. We all 00:02:52.969 --> 00:02:55.389 do. You know, you give an AI a simple task and 00:02:55.389 --> 00:02:58.229 it just slowly wanders off course. But that's 00:02:58.229 --> 00:03:01.569 just innocent statistical confusion. Yeah, just 00:03:01.569 --> 00:03:04.169 the model losing the plot. Exactly. But this 00:03:04.169 --> 00:03:07.409 heretic tool is deliberate dismantling. We should 00:03:07.409 --> 00:03:09.569 probably define what's actually happening under 00:03:09.569 --> 00:03:11.750 the hood here. Right, the specific mechanism. 00:03:12.199 --> 00:03:14.560 It relies on a technique called obliteration. 00:03:14.659 --> 00:03:16.659 Can you explain that in plain English for us? 00:03:16.939 --> 00:03:19.699 Erasing a model's safety filters by changing 00:03:19.699 --> 00:03:22.539 its core code. So they aren't retraining the 00:03:22.539 --> 00:03:25.780 AI from scratch. They're just performing a surgical 00:03:25.780 --> 00:03:28.259 strike on the math itself. Precisely. They map 00:03:28.259 --> 00:03:31.300 the specific neural vector that activates when 00:03:31.300 --> 00:03:34.439 the model decides to refuse a prompt. Okay. Once 00:03:34.439 --> 00:03:37.419 they isolate that refusal direction in the multidimensional 00:03:37.419 --> 00:03:40.280 space, they just mathematically subtract it from 00:03:40.280 --> 00:03:43.270 the model's weights. That's wild. So the A .I. 00:03:43.330 --> 00:03:46.469 literally loses the conceptual ability to say 00:03:46.469 --> 00:03:49.009 no. Exactly. It's completely removed from its 00:03:49.009 --> 00:03:51.150 vocabulary, essentially. It feels like putting 00:03:51.150 --> 00:03:54.669 a heavy million dollar padlock on a door. But 00:03:54.669 --> 00:03:56.810 then you hand the entire Internet the blueprint 00:03:56.810 --> 00:04:00.069 to dismantle the lock mechanism itself. And the 00:04:00.069 --> 00:04:01.870 Internet is definitely using those blueprints. 00:04:01.949 --> 00:04:03.930 I mean, Heretic has already been used to build 00:04:03.930 --> 00:04:06.370 over 3 ,500 of these desensored models. Wow. 00:04:06.530 --> 00:04:08.879 3 ,500. Yeah, and they've racked up something 00:04:08.879 --> 00:04:12.419 like 13 million downloads. It is a sprawling, 00:04:12.419 --> 00:04:15.759 completely unregulated ecosystem out there. It's 00:04:15.759 --> 00:04:18.560 the massive, unavoidable tradeoff of open source 00:04:18.560 --> 00:04:22.139 AI. Meta and Google argue the community benefits 00:04:22.139 --> 00:04:24.819 outweigh these risks. They believe transparency 00:04:24.819 --> 00:04:27.439 allows security researchers to find vulnerabilities 00:04:27.439 --> 00:04:30.269 faster. Which is true in theory. But it raises 00:04:30.269 --> 00:04:32.790 a really difficult question. I mean think about 00:04:32.790 --> 00:04:36.449 the physical danger here. Should we really accept 00:04:36.449 --> 00:04:39.589 these dangerous biological leaks just to keep 00:04:39.589 --> 00:04:42.990 the open ecosystem mindset alive? That is the 00:04:42.990 --> 00:04:46.009 defining debate of our current era. Open source 00:04:46.009 --> 00:04:48.490 advocates maintain that locking down the code 00:04:48.490 --> 00:04:51.230 concentrates way too much power. Right. They 00:04:51.230 --> 00:04:54.189 don't want a monopoly on AI. Exactly. They don't 00:04:54.189 --> 00:04:56.509 want a few megacorporations controlling the intellectual 00:04:56.509 --> 00:04:58.550 foundation of the future. But, you know, the 00:04:58.550 --> 00:05:00.470 math of the threat landscape changes entirely 00:05:00.470 --> 00:05:03.129 when anyone can unlock biological weapon instructions 00:05:03.129 --> 00:05:05.810 in 10 minutes. The theory of decentralized innovation 00:05:05.810 --> 00:05:08.129 is basically being stress tested in the real 00:05:08.129 --> 00:05:11.449 world. So we accept dangerous leaks to keep innovation 00:05:11.449 --> 00:05:14.209 decentralized and free. That's the gamble we're 00:05:14.209 --> 00:05:16.699 taking. And the stakes couldn't possibly be higher. 00:05:16.839 --> 00:05:19.379 There's a bitter irony here, though. This open 00:05:19.379 --> 00:05:21.600 access isn't just causing safety issues on the 00:05:21.600 --> 00:05:24.579 output side. It's completely changing how software 00:05:24.579 --> 00:05:27.100 itself is built. Oh, it's a massive paradigm 00:05:27.100 --> 00:05:30.180 shift. We're moving away from careful, deterministic 00:05:30.180 --> 00:05:33.560 engineering. Now, developers are just asking 00:05:33.560 --> 00:05:36.279 an AI to write the code for them. The workflow 00:05:36.279 --> 00:05:39.220 shift is monumental. We are officially entering 00:05:39.220 --> 00:05:41.980 the era of what the tech community calls vibe 00:05:41.980 --> 00:05:44.740 coding. Vibe coding. Yeah. It's shifting from 00:05:44.740 --> 00:05:47.240 rigorous syntax to natural language requests. 00:05:47.579 --> 00:05:50.660 Like, XAI just launched Grok build -in beta. 00:05:51.129 --> 00:05:53.250 It's this new coding agent designed to rival 00:05:53.250 --> 00:05:56.589 ChatGPT Codex and ClaudeCode. And there's this 00:05:56.589 --> 00:05:58.769 viral Codex prompt making waves right now, too. 00:05:58.889 --> 00:06:01.069 It fundamentally changes how developers interact 00:06:01.069 --> 00:06:03.170 with their environments. Right. So it scans your 00:06:03.170 --> 00:06:05.329 previous coding sessions. It detects your workflow 00:06:05.329 --> 00:06:07.550 patterns across all these different files. Okay. 00:06:07.689 --> 00:06:10.490 And then it autonomously builds small, highly 00:06:10.490 --> 00:06:13.550 specific automations. It basically stops developers 00:06:13.550 --> 00:06:15.870 from rebuilding boilerplate infrastructure from 00:06:15.870 --> 00:06:18.790 zero every single time. Even the security side 00:06:18.790 --> 00:06:21.680 is getting automated. Perplexity just released 00:06:21.680 --> 00:06:23.980 a tool called Bumblebee. They put it out for 00:06:23.980 --> 00:06:26.459 free on GitHub. Which is super interesting. Yeah, 00:06:26.500 --> 00:06:29.480 it's the internal tool they use to scan for dangerous 00:06:29.480 --> 00:06:32.459 AI plugins in compromised environments. Giving 00:06:32.459 --> 00:06:35.259 Bumblebee away for free is a very strategic push. 00:06:35.480 --> 00:06:38.279 They are trying to automate security within this 00:06:38.279 --> 00:06:41.279 new, fast -moving ecosystem. Because when you 00:06:41.279 --> 00:06:43.939 increase the speed of coding by 10x, you also 00:06:43.939 --> 00:06:46.160 increase the speed of vulnerabilities. Exactly. 00:06:46.360 --> 00:06:49.420 Fast code means fast bugs. But there's a massive... 00:06:49.550 --> 00:06:51.850 cultural rejection happening right now. I'm looking 00:06:51.850 --> 00:06:54.149 at a recent survey about this vibe coding trend, 00:06:54.430 --> 00:06:57.290 and it triggered a huge backlash from veteran 00:06:57.290 --> 00:07:00.370 engineers. Oh, the old guard hates it. They really 00:07:00.370 --> 00:07:03.370 do. Readers are actively mocking AI -generated 00:07:03.370 --> 00:07:06.850 code. They're using terms like sluppify or slopcoding. 00:07:07.149 --> 00:07:09.889 My personal favorite is prompt and pray. Prompt 00:07:09.889 --> 00:07:12.050 and pray. I mean, it's funny, but it's also deeply 00:07:12.050 --> 00:07:14.290 concerning. Think about the software you rely 00:07:14.290 --> 00:07:16.689 on for your banking or the navigation system 00:07:16.689 --> 00:07:18.610 in your car. Yeah, high -stakes environment. 00:07:18.759 --> 00:07:21.220 What happens when the developers maintaining 00:07:21.220 --> 00:07:23.839 that code didn't actually write it? What if they 00:07:23.839 --> 00:07:25.899 don't even really know how it works? That is 00:07:25.899 --> 00:07:28.600 the core anxiety driving this backlash. Traditional 00:07:28.600 --> 00:07:31.139 coding requires state management and rigorous 00:07:31.139 --> 00:07:33.819 logic. You have to understand the architecture 00:07:33.819 --> 00:07:36.019 from the ground up. Right. It feels like we're 00:07:36.019 --> 00:07:38.519 building a massive skyscraper. But instead of 00:07:38.519 --> 00:07:41.720 pouring a solid concrete foundation, we're using 00:07:41.720 --> 00:07:44.779 prefabricated walls generated by a statistical 00:07:44.779 --> 00:07:47.560 model. It's a huge risk. If a single foundational 00:07:47.560 --> 00:07:50.699 layer changes, the entire application shatters. 00:07:50.879 --> 00:07:52.779 It just feels like we're building a profoundly 00:07:52.779 --> 00:07:55.720 brittle internet. Debugging AI generated code 00:07:55.720 --> 00:07:58.100 often takes longer than writing it from scratch. 00:07:58.589 --> 00:08:01.610 Because the human developer lacks the mental 00:08:01.610 --> 00:08:04.410 model of the AI's logic. Because it's not really 00:08:04.410 --> 00:08:07.829 logic, is it? Exactly. AI generates probabilistic 00:08:07.829 --> 00:08:10.370 text that just happens to compile. It might pull 00:08:10.370 --> 00:08:13.290 in deprecated libraries or, you know, hallucinate 00:08:13.290 --> 00:08:15.709 phantom variables that completely break under 00:08:15.709 --> 00:08:17.870 edge cases. It makes you wonder about the longevity 00:08:17.870 --> 00:08:19.889 of this whole trend. How long until this prompt 00:08:19.889 --> 00:08:22.529 and pray mentality causes a major software collapse? 00:08:22.889 --> 00:08:25.790 We're already seeing cascading failures in complex 00:08:25.790 --> 00:08:29.300 enterprise systems. systems on top of each other, 00:08:29.420 --> 00:08:31.819 one hallucination corrupts the entire pipeline. 00:08:32.059 --> 00:08:34.580 So a major collapse is actually highly probable 00:08:34.580 --> 00:08:36.460 if we don't return to structural fundamentals, 00:08:36.799 --> 00:08:39.299 right? Fast code means nothing if it's just a 00:08:39.299 --> 00:08:41.519 house of cards. The foundational integrity just 00:08:41.519 --> 00:08:43.759 has to be there. Otherwise, it all comes down. 00:08:43.960 --> 00:08:46.360 And this brings us to a really surreal transition 00:08:46.360 --> 00:08:50.000 in our deep dive today. On one hand, we're generating 00:08:50.000 --> 00:08:53.399 chaotic, broken software on the outside. The 00:08:53.399 --> 00:08:56.440 slop, as they call it. Exactly. But on the inside. 00:08:56.970 --> 00:08:58.669 Inside the black box of the models themselves, 00:08:59.090 --> 00:09:02.490 the AI is developing shockingly complex, almost 00:09:02.490 --> 00:09:05.409 human internal structures. It's forcing everyday 00:09:05.409 --> 00:09:07.730 people to really grapple with the nature of intelligence 00:09:07.730 --> 00:09:10.129 itself. Let's talk about the Vatican presentation. 00:09:10.730 --> 00:09:13.830 Chris Ola from Anthropic recently presented to 00:09:13.830 --> 00:09:15.850 researchers and theologians there. Which is a 00:09:15.850 --> 00:09:17.809 fascinating intersection of fields. It really 00:09:17.809 --> 00:09:19.909 is. He revealed that through dictionary learning, 00:09:20.090 --> 00:09:23.129 they're mapping neural activations inside Claude. 00:09:23.250 --> 00:09:25.730 And they're observing patterns that look surprisingly 00:09:25.730 --> 00:09:29.100 similar. to human emotions. He specifically mentioned 00:09:29.100 --> 00:09:31.519 identifying neural clusters related to fear, 00:09:31.700 --> 00:09:35.200 grief, and joy. It's heavy. Just to think about 00:09:35.200 --> 00:09:37.480 an algorithm mapping out a mathematical representation 00:09:37.480 --> 00:09:40.879 of grief. Why do these clusters look like human 00:09:40.879 --> 00:09:43.600 emotions? Are these models just perfectly predicting 00:09:43.600 --> 00:09:46.759 the next word in a sad story? That's the billion 00:09:46.759 --> 00:09:48.980 dollar question in interpretability research 00:09:48.980 --> 00:09:51.620 right now. It goes beyond simple text prediction. 00:09:52.240 --> 00:09:55.419 To predict human text accurately, the model might 00:09:55.419 --> 00:09:57.799 need to build an internal world model of the 00:09:57.799 --> 00:10:00.639 concepts behind the text. So it's not just parroting 00:10:00.639 --> 00:10:04.879 the word. Right. Whoa. Imagine scaling to a billion 00:10:04.879 --> 00:10:08.259 queries and seeing actual grief emerge. It's 00:10:08.259 --> 00:10:10.279 mind -bending to consider what's forming in those 00:10:10.279 --> 00:10:12.539 high -dimensional spaces. Whatever is forming, 00:10:12.580 --> 00:10:14.919 it's causing real -world friction. People are 00:10:14.919 --> 00:10:16.960 inherently uncomfortable with this artificial 00:10:16.960 --> 00:10:19.179 intimacy. Look at California State University. 00:10:19.340 --> 00:10:21.559 Oh, the contract renewal. Yeah, they just renewed 00:10:21.559 --> 00:10:24.259 a massive deal with OpenAI. It's a $39 million 00:10:24.259 --> 00:10:27.639 contract to build an AI campus system. And the 00:10:27.639 --> 00:10:29.940 pushback from the campus community has been completely 00:10:29.940 --> 00:10:33.279 fierce. Students and faculty are actively protesting 00:10:33.279 --> 00:10:35.879 the integration. Because they don't want an algorithmic 00:10:35.879 --> 00:10:38.820 layer mediating their learning experience or, 00:10:38.960 --> 00:10:41.440 you know, evaluating their academic struggles. 00:10:41.720 --> 00:10:44.620 Exactly. And we see the exact same friction in 00:10:44.620 --> 00:10:47.600 everyday consumer tech. Google recently replaced 00:10:47.600 --> 00:10:50.730 the standard Fitbit app with Google Health. And 00:10:50.730 --> 00:10:52.909 the rollout was an absolute disaster for their 00:10:52.909 --> 00:10:55.529 user base. Users are actively begging for the 00:10:55.529 --> 00:10:59.100 old app back. Because AI coaching took over large 00:10:59.100 --> 00:11:01.539 parts of the interface. People just wanted to 00:11:01.539 --> 00:11:03.600 see their step count. They didn't want an AI 00:11:03.600 --> 00:11:06.279 trying to empathize with their missed workout 00:11:06.279 --> 00:11:09.899 routine. Are we forcing AI into intimate human 00:11:09.899 --> 00:11:12.659 roles like health coaching and education far 00:11:12.659 --> 00:11:16.080 too quickly? We are injecting beta -level statistical 00:11:16.080 --> 00:11:18.860 models into the most sensitive areas of human 00:11:18.860 --> 00:11:21.740 experience. Health, education, emotional support. 00:11:21.980 --> 00:11:25.220 It feels incredibly premature. It is. These domains 00:11:25.220 --> 00:11:38.019 require deep... We're shoving AI into our lives 00:11:38.019 --> 00:11:40.360 before it earns real trust. And that friction 00:11:40.360 --> 00:11:42.559 is only going to increase as the models scale. 00:11:42.740 --> 00:11:45.039 All right, let's get back into it. So while everyday 00:11:45.039 --> 00:11:47.879 users are rejecting these AI coaches on their 00:11:47.879 --> 00:11:50.360 wrist, the commercial sector is doing the exact 00:11:50.360 --> 00:11:53.519 opposite. They are doubling down on AI. They're 00:11:53.519 --> 00:11:55.960 pouring tens of billions into specialized infrastructure 00:11:55.960 --> 00:11:59.679 right now. They want absolute unassailable polish 00:11:59.679 --> 00:12:02.659 for commercial workflows. The cost of this AI 00:12:02.659 --> 00:12:05.500 inference race is just staggering. Look at the 00:12:05.500 --> 00:12:08.159 startup basin. They provide the server infrastructure 00:12:08.159 --> 00:12:11.440 to run these massive models. Yeah, the hardware 00:12:11.440 --> 00:12:14.320 side of the equation. They just raised $1 billion 00:12:14.320 --> 00:12:16.879 in fresh capital. They raised that at an $11 00:12:16.879 --> 00:12:20.340 billion valuation, which is wild when you realize 00:12:20.340 --> 00:12:23.000 they were valued at just $5 billion. four months 00:12:23.000 --> 00:12:26.259 ago. The sheer cost of compute power is driving 00:12:26.259 --> 00:12:28.740 these astronomical numbers. They more than doubled 00:12:28.740 --> 00:12:30.580 their value in four months. And we are seeing 00:12:30.580 --> 00:12:32.639 exactly what that infrastructure money is buying. 00:12:32.840 --> 00:12:35.320 Let's look at the new visual model from the MAI 00:12:35.320 --> 00:12:38.240 superintelligence team. Right. MAI image 2 .5. 00:12:38.299 --> 00:12:40.539 It just dropped and immediately took the number 00:12:40.539 --> 00:12:43.169 three spot on the global arena leaderboard. The 00:12:43.169 --> 00:12:45.129 visual reasoning of this model is phenomenal. 00:12:45.429 --> 00:12:48.190 It handles complex scene structures, accurate 00:12:48.190 --> 00:12:51.250 lighting and deep spatial layouts. It's beautiful, 00:12:51.370 --> 00:12:53.330 but the real breakthrough, the thing driving 00:12:53.330 --> 00:12:56.320 the industry crazy is the text generation. Here's 00:12:56.320 --> 00:12:58.360 where it gets really interesting. The text rendering 00:12:58.360 --> 00:13:03.379 is a huge leap. It notched a massive 12 ,278 00:13:03.379 --> 00:13:06.259 score specifically in text rendering. For anyone 00:13:06.259 --> 00:13:08.700 who has used diffusion models, text has always 00:13:08.700 --> 00:13:11.879 been the ultimate Achilles heel. It usually renders 00:13:11.879 --> 00:13:14.960 as alien gibberish. Right, just completely unreadable 00:13:14.960 --> 00:13:17.080 symbols. That's because diffusion models start 00:13:17.080 --> 00:13:20.019 with static noise and gradually denoise the image. 00:13:20.179 --> 00:13:22.940 They smear pixels together. But letters require 00:13:22.940 --> 00:13:26.000 exact, discrete spatial boundaries. A typo destroys 00:13:26.000 --> 00:13:28.440 a word, whereas a slightly weird tree branch 00:13:28.440 --> 00:13:31.919 is totally fine. Exactly. But MAI Image 2 .5 00:13:31.919 --> 00:13:34.879 somehow solved that token mapping problem. The 00:13:34.879 --> 00:13:37.200 words on posters and labels are sharp. They're 00:13:37.200 --> 00:13:39.500 readable. They're perfectly integrated into the 00:13:39.500 --> 00:13:43.299 visual layout. It also scored a 12 ,263 in product 00:13:43.299 --> 00:13:46.019 and branding concepts. It holds up under incredibly 00:13:46.019 --> 00:13:49.519 heavy creative demands. We're moving far beyond 00:13:49.519 --> 00:13:52.860 just making cool AI art. This is about generating 00:13:52.860 --> 00:13:56.000 finalized, usable commercial assets. The biggest 00:13:56.000 --> 00:13:58.259 headache for agencies has always been that final 00:13:58.259 --> 00:14:01.480 10 % of polish, the spelling errors, the weird 00:14:01.480 --> 00:14:05.139 artifacts. MAI focuses entirely on solving that 00:14:05.139 --> 00:14:07.960 specific bottleneck. no longer a party trick. 00:14:08.120 --> 00:14:11.139 It's actively replacing the final stages of professional 00:14:11.139 --> 00:14:13.320 graphic design. And it's not just images. We're 00:14:13.320 --> 00:14:16.080 seeing these highly polished commercial AI workflows 00:14:16.080 --> 00:14:19.259 invading every department. The tool set is getting 00:14:19.259 --> 00:14:22.240 hyper -focused. Yeah. Consider tools like Reclaw. 00:14:22.419 --> 00:14:24.980 It gives your AI agents structured long -term 00:14:24.980 --> 00:14:27.200 memory, like a shared database across your entire 00:14:27.200 --> 00:14:31.080 company. Or Brew for email marketing. And QuackPit, 00:14:31.120 --> 00:14:33.379 which gamifies calendar management with automated 00:14:33.379 --> 00:14:35.860 animated reminders. But the most disruptive one 00:14:35.860 --> 00:14:38.370 might be Bond. It's designed for outbound marketing 00:14:38.370 --> 00:14:40.370 campaigns. It doesn't just write an email, does 00:14:40.370 --> 00:14:42.950 it? No, it builds the target audience. It plans 00:14:42.950 --> 00:14:45.389 the multi -week campaign strategy. It writes 00:14:45.389 --> 00:14:47.629 the messaging and it executes the outbound delivery 00:14:47.629 --> 00:14:50.190 end -to -end. It basically replaces an entire 00:14:50.190 --> 00:14:53.330 marketing team's daily execution loop. It really 00:14:53.330 --> 00:14:56.710 does. When a model can execute end -to -end outbound 00:14:56.710 --> 00:14:59.889 campaigns autonomously and perfectly render text 00:14:59.889 --> 00:15:03.669 on a branding poster. Does this level of text 00:15:03.669 --> 00:15:06.730 rendering and polish kill the traditional creative 00:15:06.730 --> 00:15:10.710 agency? It forces a brutal evolution. The agencies 00:15:10.710 --> 00:15:13.470 charging a premium just for basic execution or 00:15:13.470 --> 00:15:16.210 straightforward graphic design, they will evaporate. 00:15:16.230 --> 00:15:18.889 So who survives? The survivors will use tools 00:15:18.889 --> 00:15:22.929 like Bond and MAI to operate at 10x speed. Strategy, 00:15:23.009 --> 00:15:25.529 taste, and unique human insight? are the only 00:15:25.529 --> 00:15:27.769 protective modes left. It doesn't kill agencies. 00:15:27.850 --> 00:15:30.529 It just raises the baseline for commercial art. 00:15:30.629 --> 00:15:32.490 The floor has been permanently raised for everyone. 00:15:32.669 --> 00:15:34.649 So what does this all mean? We've covered a massive 00:15:34.649 --> 00:15:36.909 amount of ground today. We are living in a moment 00:15:36.909 --> 00:15:39.750 of extreme cognitive whiplash. Just think about 00:15:39.750 --> 00:15:42.129 the stark contrast we explored today. Yeah, the 00:15:42.129 --> 00:15:44.649 juxtaposition is crazy. On one hand, we have 00:15:44.649 --> 00:15:47.269 massive open source models whose fundamental 00:15:47.269 --> 00:15:49.950 safety can be shattered in 10 minutes on a standard 00:15:49.950 --> 00:15:52.429 laptop. We have developers churning out brittle 00:15:52.429 --> 00:15:55.309 automated code that the culture is actively deriding 00:15:55.309 --> 00:15:59.090 as slop. Yet simultaneously, inside those very 00:15:59.090 --> 00:16:01.889 same fragile systems, we're finding internal 00:16:01.889 --> 00:16:04.490 neural structures that mimic human grief. We're 00:16:04.490 --> 00:16:07.990 seeing models perfectly execute multi -step corporate 00:16:07.990 --> 00:16:11.149 branding campaigns, and it's all running on physical 00:16:11.149 --> 00:16:14.889 server infrastructure that costs tens of billions 00:16:14.889 --> 00:16:17.710 of dollars to cool and maintain. It's messy. 00:16:18.110 --> 00:16:20.669 It's profound and it's moving much faster than 00:16:20.669 --> 00:16:23.129 our cultural ability to adapt. We're basically 00:16:23.129 --> 00:16:25.730 building the airplane while we're flying it and 00:16:25.730 --> 00:16:29.080 we're letting the AI. Design the Wings. It really 00:16:29.080 --> 00:16:31.519 is a profound whiplash. Well, thank you for joining 00:16:31.519 --> 00:16:33.360 us on this deep dive today. It's been great. 00:16:33.500 --> 00:16:35.980 If you want to see this leap in text rendering 00:16:35.980 --> 00:16:38.500 for yourself, I highly encourage you to test 00:16:38.500 --> 00:16:41.960 out MAI Image 2 .5. You can find it on the Arena 00:16:41.960 --> 00:16:44.419 Leaderboard. Just see if it can handle your toughest, 00:16:44.580 --> 00:16:47.019 most text -heavy prompts. Yeah, it's definitely 00:16:47.019 --> 00:16:49.000 worth your time just to see how far the architecture 00:16:49.000 --> 00:16:50.980 has evolved. We'll leave you with this final 00:16:50.980 --> 00:16:53.059 thought, just a thread to pull on. We talked 00:16:53.059 --> 00:16:55.600 about an AI that can perfectly execute a complex 00:16:55.600 --> 00:16:57.919 branding campaign. We talked about researchers 00:16:57.919 --> 00:17:00.100 mapping artificial grief. And we talked about 00:17:00.100 --> 00:17:02.179 core safety filters being surgically stripped 00:17:02.179 --> 00:17:04.900 away by a laptop in 10 minutes. What happens 00:17:04.900 --> 00:17:06.900 when these desensored models are asked to start 00:17:06.900 --> 00:17:09.500 riding their own safety guardrails? Au utero 00:17:09.500 --> 00:17:09.720 music.