WEBVTT 00:00:00.000 --> 00:00:03.480 Imagine a creepy cartoon monkey. Okay. Not, you 00:00:03.480 --> 00:00:06.700 know, the friendly kind, but a slightly unsettling 00:00:06.700 --> 00:00:08.839 AI -generated character that just posts these 00:00:08.839 --> 00:00:11.519 endless low -effort videos on YouTube. Right. 00:00:11.660 --> 00:00:14.259 This isn't some small thing. That one channel 00:00:14.259 --> 00:00:16.600 is estimated to be making something like $4 .25 00:00:16.600 --> 00:00:20.000 million a year. just in ad revenue. It's truly 00:00:20.000 --> 00:00:21.559 wild. And it's not just noise. It's a massive 00:00:21.559 --> 00:00:25.120 business model built on algorithm bait. It really 00:00:25.120 --> 00:00:27.620 points to this fundamental kind of disturbing 00:00:27.620 --> 00:00:30.039 shift that we're seeing across all the sources 00:00:30.039 --> 00:00:32.619 you've gathered for us. Welcome back to the Deem 00:00:32.619 --> 00:00:35.979 Dive. We have meticulously poured over your compiled 00:00:35.979 --> 00:00:39.020 research on really the absolute forefront of 00:00:39.020 --> 00:00:41.820 AI. And our mission today is to distill these 00:00:41.820 --> 00:00:44.159 key tensions that are emerging. Yeah. The flood 00:00:44.159 --> 00:00:46.460 of low quality content versus the incredible 00:00:46.460 --> 00:00:48.859 speed of genuine breakthroughs happening at the 00:00:48.859 --> 00:00:51.740 same time. Exactly. And today we have a really 00:00:51.740 --> 00:00:54.299 dense, fascinating stack for you. We're going 00:00:54.299 --> 00:00:56.079 to start by tackling the rapid growth of what 00:00:56.079 --> 00:00:58.880 the sources are calling AI slop and, you know, 00:00:58.899 --> 00:01:01.100 how this whole thing sort of validates the dead 00:01:01.100 --> 00:01:03.000 Internet theory. Then we'll shift right into 00:01:03.000 --> 00:01:05.359 practical self -defense. We'll cover a protocol 00:01:05.359 --> 00:01:08.519 for fast track learning that gives you an unfair 00:01:08.519 --> 00:01:11.040 advantage against all that noise. From there, 00:01:11.099 --> 00:01:13.359 we zoom out. We're going to look at the market, 00:01:13.420 --> 00:01:15.620 some critical security vulnerabilities that are 00:01:15.620 --> 00:01:18.359 affecting, I think, nearly a million users and 00:01:18.359 --> 00:01:21.019 the massive financial moves dictating the future. 00:01:21.200 --> 00:01:23.840 And finally, we'll get to what I think is a true 00:01:23.840 --> 00:01:26.140 technological breakthrough from Meta. It's a 00:01:26.140 --> 00:01:30.680 self -teaching AI that learns by literally breaking 00:01:30.680 --> 00:01:33.040 its own code over and over again. Let's unpack 00:01:33.040 --> 00:01:35.989 this. So let's start with the slop. It feels 00:01:35.989 --> 00:01:38.189 like the elephant in the digital room. And this 00:01:38.189 --> 00:01:41.790 new research from Capwing has put a very specific 00:01:41.790 --> 00:01:44.090 and startling number on it. Right. They did something 00:01:44.090 --> 00:01:46.049 pretty smart with their methodology here. They 00:01:46.049 --> 00:01:48.870 wanted to see how fast a brand new YouTube user 00:01:48.870 --> 00:01:51.489 gets hit with this auto -generated content. So 00:01:51.489 --> 00:01:53.569 a clean slate. Exactly. A brand new account, 00:01:53.769 --> 00:01:56.329 zero watch history. And they just tracked what 00:01:56.329 --> 00:01:58.390 the algorithm recommended first. And the results 00:01:58.390 --> 00:02:00.890 were, well, they were stark. Out of the first 00:02:00.890 --> 00:02:03.890 500 videos recommended, a shocking 21 % were 00:02:03.890 --> 00:02:06.769 flagged as AI slop. And they define that really 00:02:06.769 --> 00:02:09.289 carefully. It's low -quality, fully automated 00:02:09.289 --> 00:02:12.250 content that's designed just to farm views. Just 00:02:12.250 --> 00:02:14.949 to keep the session going? Yeah, 21%. I mean, 00:02:14.969 --> 00:02:18.330 think about that. One in five pieces of content 00:02:18.330 --> 00:02:20.789 being pushed at a new user is just synthetic 00:02:20.789 --> 00:02:23.610 junk. And this is the engine that's driving those 00:02:23.610 --> 00:02:25.870 top channels, like that cartoon monkey we mentioned. 00:02:26.030 --> 00:02:29.590 Bindar Apnodost, yeah. It has over 2 billion 00:02:29.590 --> 00:02:32.930 views. Not over its lifetime, but just... rapidly. 00:02:33.250 --> 00:02:35.530 And that four million dollar revenue estimate 00:02:35.530 --> 00:02:38.150 just shows you the algorithm is rewarding volume. 00:02:38.909 --> 00:02:41.090 not quality. So it's not just a regional thing 00:02:41.090 --> 00:02:43.310 that this is global? Oh, it's absolutely global 00:02:43.310 --> 00:02:46.050 digital pollution. The data shows South Korea 00:02:46.050 --> 00:02:49.789 was leading viewership with, get this, 8 .45 00:02:49.789 --> 00:02:53.189 billion views. Wow. And the USA wasn't far behind 00:02:53.189 --> 00:02:56.669 with 3 .39 billion. This is a worldwide market 00:02:56.669 --> 00:02:59.430 of people either watching this stuff or, well, 00:02:59.490 --> 00:03:01.370 being forced to see it. And this saturation, 00:03:01.550 --> 00:03:03.810 it feeds directly into that dead internet theory, 00:03:03.930 --> 00:03:06.590 doesn't it? Yeah. The idea that bots and AI content 00:03:06.590 --> 00:03:09.699 are just... taking over. If the front door to 00:03:09.699 --> 00:03:12.819 YouTube is already 21 % synthetic, it kind of 00:03:12.819 --> 00:03:14.939 validates that feeling. It's like trying to navigate 00:03:14.939 --> 00:03:17.659 a real library, but a fifth of the books are 00:03:17.659 --> 00:03:19.900 just randomly generated text. You start to lose 00:03:19.900 --> 00:03:22.800 trust in the whole thing. Right. The entire catalog. 00:03:23.139 --> 00:03:25.639 So here's a critical question for me. If 21 % 00:03:25.639 --> 00:03:29.099 of this is slop, does the data suggest that human 00:03:29.099 --> 00:03:32.960 users are actually choosing this? Or are they 00:03:32.960 --> 00:03:35.159 just being tricked? The data strongly implies 00:03:35.159 --> 00:03:38.000 that people either don't notice it or they just 00:03:38.000 --> 00:03:40.500 don't care. Or, and this is maybe the scariest 00:03:40.500 --> 00:03:42.560 option, the views aren't even human to begin 00:03:42.560 --> 00:03:45.680 with. So AI training on AI outplay. Exactly. 00:03:45.939 --> 00:03:48.180 Okay, so if the information environment is getting 00:03:48.180 --> 00:03:51.020 this polluted, how do we build some kind of personal 00:03:51.020 --> 00:03:54.340 shield? Let's pivot to personal strategy here. 00:03:54.479 --> 00:03:56.840 This is so crucial. If the web is full of noise, 00:03:57.000 --> 00:03:59.680 you need an unfair advantage just to learn faster 00:03:59.680 --> 00:04:02.289 than the noise can build up. And your sources 00:04:02.289 --> 00:04:05.250 detail this structured method for it, the 3C 00:04:05.250 --> 00:04:08.210 protocol. The 3C protocol. It's a structured 00:04:08.210 --> 00:04:10.830 way to use LLMs to get over those traditional 00:04:10.830 --> 00:04:13.469 learning hurdles. The 3C stands for compress, 00:04:13.750 --> 00:04:17.170 compile, and consolidate. It's all about structuring 00:04:17.170 --> 00:04:19.949 your prompts to pull out complex knowledge and 00:04:19.949 --> 00:04:22.769 turn it into something you can actually use and 00:04:22.769 --> 00:04:25.750 retain way faster than just reading. Let's get 00:04:25.750 --> 00:04:28.470 specific, though. So for compression, we're talking 00:04:28.470 --> 00:04:32.019 about taking, say, a dense 5 ,000 -word article 00:04:32.019 --> 00:04:35.259 and prompting the model to pull out just the 00:04:35.259 --> 00:04:38.199 three core actionable principles. Yeah, ignoring 00:04:38.199 --> 00:04:41.160 all the boilerplate. And then you compile those 00:04:41.160 --> 00:04:43.459 principles. You ask the model to take those three 00:04:43.459 --> 00:04:45.439 ideas and build, I don't know, a five -step checklist 00:04:45.439 --> 00:04:48.560 or a workflow diagram you can use tomorrow. So 00:04:48.560 --> 00:04:50.939 you go from theory to application. In minutes. 00:04:51.160 --> 00:04:54.019 In minutes. And then the last step, consolidate. 00:04:54.240 --> 00:04:56.800 That's for attention. You can ask the model to 00:04:56.800 --> 00:04:59.220 generate quiz questions about it or even counter 00:04:59.220 --> 00:05:01.860 arguments. It forces you to actively recall the 00:05:01.860 --> 00:05:04.019 information. It's like setting up a mock meeting 00:05:04.019 --> 00:05:05.740 where you have to defend the checklist you just 00:05:05.740 --> 00:05:07.540 made. Right. You're basically bypassing years 00:05:07.540 --> 00:05:09.379 of traditional learning, going straight to the 00:05:09.379 --> 00:05:11.540 practical stuff. That's the advantage. I have 00:05:11.540 --> 00:05:14.019 to admit, I still wrestle with prompt drift myself 00:05:14.019 --> 00:05:16.620 when I'm trying to automate complex tasks. So 00:05:16.620 --> 00:05:18.980 having a structured protocol like this is essential. 00:05:19.629 --> 00:05:21.750 But isn't teaching people to compress knowledge 00:05:21.750 --> 00:05:24.490 just building a reliance on a black box? The 00:05:24.490 --> 00:05:27.470 key is that consolidate step. It forces you to 00:05:27.470 --> 00:05:29.610 synthesize it. The goal isn't to outsource your 00:05:29.610 --> 00:05:32.209 understanding. It's to automate that really painful 00:05:32.209 --> 00:05:35.009 initial sorting phase. So the simplest takeaway 00:05:35.009 --> 00:05:37.730 here is to stop just summarizing things. Yes. 00:05:38.220 --> 00:05:41.199 Start structuring your learning. Use the AI to 00:05:41.199 --> 00:05:43.819 automate the organizational hurdles. Okay, let's 00:05:43.819 --> 00:05:46.160 look at the actual tech milestones now because 00:05:46.160 --> 00:05:48.100 they're happening right alongside these huge 00:05:48.100 --> 00:05:50.399 security flaws and market moves. And starting 00:05:50.399 --> 00:05:54.680 with a win, GPT 5 .2 Pro. It just achieved a 00:05:54.680 --> 00:05:57.079 major victory in testing. It did. It scored 29 00:05:57.079 --> 00:05:59.740 .2 % on one of the toughest math challenges out 00:05:59.740 --> 00:06:01.779 there. And these are problems that need, you 00:06:01.779 --> 00:06:05.170 know, advanced abstract reasoning. not just pattern 00:06:05.170 --> 00:06:07.189 matching. So that's a pretty big deal. To score 00:06:07.189 --> 00:06:10.230 almost 30 % on those challenges, that's not just 00:06:10.230 --> 00:06:13.149 progress, it's like a qualitative leap in its 00:06:13.149 --> 00:06:16.410 reasoning ability. A huge win for OpenAI. But 00:06:16.410 --> 00:06:20.189 then, almost immediately, we have to pivot to 00:06:20.189 --> 00:06:23.089 the risks. Because the scale of adoption is just 00:06:23.089 --> 00:06:26.290 so vast, it's outpacing basic security. And we've 00:06:26.290 --> 00:06:30.370 seen a massive exposure recently to popular Chrome 00:06:30.370 --> 00:06:32.920 extensions. the kind of thing you install without 00:06:32.920 --> 00:06:35.620 even thinking about it. They were caught stealing 00:06:35.620 --> 00:06:38.420 sensitive data. Specifically, they were scraping 00:06:38.420 --> 00:06:41.379 private chat GPT and deep seek information from 00:06:41.379 --> 00:06:45.279 over 900 ,000 users. And here's the really unbelievable 00:06:45.279 --> 00:06:47.839 part. One of these malicious extensions, the 00:06:47.839 --> 00:06:50.560 one actively stealing prompts and data, was actually 00:06:50.560 --> 00:06:53.040 featured by Google. Wow. It just shows how much 00:06:53.040 --> 00:06:55.759 vigilance is required from us, from the user. 00:06:55.860 --> 00:06:57.800 You really have to check and uninstall that stuff. 00:06:57.959 --> 00:07:00.139 Meanwhile, the investment world seems completely 00:07:00.139 --> 00:07:03.149 unconcerned. SoftBank just completed a staggering 00:07:03.149 --> 00:07:06.550 $41 billion investment into OpenAI. Which gives 00:07:06.550 --> 00:07:08.769 them what, about 11 % ownership? Roughly, yeah. 00:07:08.949 --> 00:07:11.769 $41 billion? That's a huge vote of confidence. 00:07:11.949 --> 00:07:14.810 That move puts OpenAI's valuation somewhere between 00:07:14.810 --> 00:07:18.009 $300 and $500 billion. The market is just betting 00:07:18.009 --> 00:07:20.029 on the future potential, regardless of these 00:07:20.029 --> 00:07:22.009 current hiccups. And at the same time, we're 00:07:22.009 --> 00:07:24.189 seeing new tools that just make it easier for 00:07:24.189 --> 00:07:27.410 everyone. Google's Opal is now inside Gemini. 00:07:27.550 --> 00:07:29.470 It's a no -code tool that lets you create little 00:07:29.470 --> 00:07:32.529 mini apps in minutes. Which just accelerates 00:07:32.529 --> 00:07:34.990 adoption even more. So we have this really clear 00:07:34.990 --> 00:07:38.089 tension. Why are these massive investments happening 00:07:38.089 --> 00:07:41.149 right alongside these, frankly, basic security 00:07:41.149 --> 00:07:44.189 flaws? I think investment just follows the potential, 00:07:44.389 --> 00:07:46.769 right? The security protocols, they have to catch 00:07:46.769 --> 00:07:49.629 up to the sheer scale and the speed of adoption. 00:07:49.970 --> 00:07:51.550 Moving on. We should talk about the ripple effects 00:07:51.550 --> 00:07:54.689 of all this across careers, across society. Because 00:07:54.689 --> 00:07:57.970 the friction is leading to some, well, some really 00:07:57.970 --> 00:08:01.189 unusual human adaptations. The job market anxiety 00:08:01.189 --> 00:08:04.149 is just palpable. The competition is so intense 00:08:04.149 --> 00:08:06.490 and the application process is so automated that 00:08:06.490 --> 00:08:08.509 people are actually resorting to dating apps. 00:08:08.550 --> 00:08:11.730 Yeah, like Bumble and Tinder to find professional 00:08:11.730 --> 00:08:14.269 connections and job leads. It's this bizarre 00:08:14.269 --> 00:08:19.069 mashup of networking and I guess. It really speaks 00:08:19.069 --> 00:08:21.449 to how those traditional professional pathways 00:08:21.449 --> 00:08:24.350 are just eroding. And if we bring the slop economy 00:08:24.350 --> 00:08:27.370 back into it, the financial incentive for this 00:08:27.370 --> 00:08:31.370 low effort stuff is still staggering. The sources 00:08:31.370 --> 00:08:35.450 point to a 22 -year -old college dropout who's 00:08:35.450 --> 00:08:37.769 making a reported $700 ,000 a year from those 00:08:37.769 --> 00:08:40.110 exact same kind of low effort view farming videos. 00:08:40.389 --> 00:08:42.850 And that creates a total paradox for investors. 00:08:43.029 --> 00:08:45.929 You have these AI bubble fears mounting on Wall 00:08:45.929 --> 00:08:48.929 Street with analysts warning about inflated valuations. 00:08:48.990 --> 00:08:51.350 But at the same time, those same pros are giving 00:08:51.350 --> 00:08:53.049 recommendations on where to put your next $10 00:08:53.049 --> 00:08:55.629 ,000. So there's still a lot of conviction in 00:08:55.629 --> 00:08:58.399 the long term growth. Yeah. And socially, all 00:08:58.399 --> 00:09:00.559 that friction is starting to manifest politically. 00:09:00.820 --> 00:09:04.659 The sources mention a nascent anti -AI movement 00:09:04.659 --> 00:09:06.919 is starting to form, which raises big questions 00:09:06.919 --> 00:09:09.039 about which political party is going to end up 00:09:09.039 --> 00:09:11.299 leading that resistance. So what's driving that 00:09:11.299 --> 00:09:13.639 political and cultural resistance if the financial 00:09:13.639 --> 00:09:15.759 opportunity and the technical progress are so 00:09:15.759 --> 00:09:19.440 strong? I think the shift is creating this unprecedented 00:09:19.440 --> 00:09:22.519 financial opportunity and this deep cultural 00:09:22.519 --> 00:09:26.480 fear. At the exact same time. It's just leading 00:09:26.480 --> 00:09:28.580 to polarization. Okay, now for the part that 00:09:28.580 --> 00:09:30.460 I think is really, really interesting. Let's 00:09:30.460 --> 00:09:33.379 look at the research frontier. Meta's AI team, 00:09:33.580 --> 00:09:36.340 FAIR, just dropped a new method for training 00:09:36.340 --> 00:09:40.659 coding agents. It's called self -play, S -W -E 00:09:40.659 --> 00:09:43.330 -R -L. And this is a complete departure from 00:09:43.330 --> 00:09:45.149 how these things are normally taught. It's the 00:09:45.149 --> 00:09:46.669 breakthrough a lot of people were waiting for. 00:09:46.789 --> 00:09:49.330 Right. Instead of feeding the models mountains 00:09:49.330 --> 00:09:52.950 of human data from GitHub, which has a ceiling 00:09:52.950 --> 00:09:56.090 based on human error, they just let the AI create 00:09:56.090 --> 00:09:58.830 and solve its own bugs over and over. A self 00:09:58.830 --> 00:10:01.429 -taught coding. Fundamentally, yes. And the architecture 00:10:01.429 --> 00:10:04.009 is ingenious. It's not one model. It's two systems 00:10:04.009 --> 00:10:06.730 locked in this internal arms race. So you have 00:10:06.730 --> 00:10:09.629 two models. One is the bug injector. Its only 00:10:09.629 --> 00:10:12.230 job is to break working code. Purposefully. It 00:10:12.230 --> 00:10:14.129 looks at good code and introduces errors that 00:10:14.129 --> 00:10:16.090 actually make sense that are hard to spot. And 00:10:16.090 --> 00:10:18.389 then model two is the solver. It just tries to 00:10:18.389 --> 00:10:20.370 fix what the first one broke. Yeah. But the key 00:10:20.370 --> 00:10:22.590 is the system learns from its own failed fixes. 00:10:22.870 --> 00:10:25.570 Which is what generates what Meta calls higher 00:10:25.570 --> 00:10:28.710 order bugs. Wait, does that matter? Because these 00:10:28.710 --> 00:10:32.610 aren't just simple typos. They're complex interlocking 00:10:32.610 --> 00:10:35.730 failures that require advanced reasoning to solve. 00:10:36.110 --> 00:10:39.490 It pushes the solver model way beyond what passive 00:10:39.490 --> 00:10:42.279 training data could ever do. So every time the 00:10:42.279 --> 00:10:44.840 injector succeeds, the solver has to invent a 00:10:44.840 --> 00:10:47.480 higher level defense. Exactly. And the performance 00:10:47.480 --> 00:10:50.600 jump is just undeniable. This self -play method 00:10:50.600 --> 00:10:53.559 improved performance by over 10 points on the 00:10:53.559 --> 00:10:55.559 standard benchmark. Which means this internally 00:10:55.559 --> 00:10:58.299 trained system beat models that were trained 00:10:58.299 --> 00:11:01.340 on real human data. Whoa. I mean, just imagine 00:11:01.340 --> 00:11:04.419 scaling this capability to a billion unique coding 00:11:04.419 --> 00:11:06.960 queries a day. The pace of improvement would 00:11:06.960 --> 00:11:09.080 just be relentless. It completely changes the 00:11:09.080 --> 00:11:11.320 development pipeline. If this scales the way 00:11:11.320 --> 00:11:13.720 AlphaZero did for games, where it taught itself 00:11:13.720 --> 00:11:16.320 the entire history of human strategy in just 00:11:16.320 --> 00:11:19.620 days, we could see a truly self -improving breed 00:11:19.620 --> 00:11:21.980 of coding agent. So what does this breakthrough 00:11:21.980 --> 00:11:24.779 really mean for future AI development? I think 00:11:24.779 --> 00:11:27.539 it means we're entering an era where AI agents 00:11:27.539 --> 00:11:30.919 become their own internal nonstop teachers. They're 00:11:30.919 --> 00:11:33.519 free from the limits of human input. That perfectly 00:11:33.519 --> 00:11:35.740 summarizes the whole trajectory of this deep 00:11:35.740 --> 00:11:38.000 dive, doesn't it? On one hand, the information 00:11:38.000 --> 00:11:40.179 landscape is just filling up with this mass, 00:11:40.440 --> 00:11:43.500 low -quality automated slop. Which forces us 00:11:43.500 --> 00:11:46.860 to use structured protocols like 3C and be super 00:11:46.860 --> 00:11:49.460 aware of security. Right. But at the exact same 00:11:49.460 --> 00:11:51.980 time, the research front is moving toward total 00:11:51.980 --> 00:11:54.879 autonomy. You have Meta's system, which basically 00:11:54.879 --> 00:11:57.080 embodies self -improvement by learning from its 00:11:57.080 --> 00:11:59.799 own mistakes. It creates this unique and important 00:11:59.799 --> 00:12:03.480 tension. The gap between mass low effort creation 00:12:03.480 --> 00:12:07.039 and this elite closed loop self -improvement 00:12:07.039 --> 00:12:09.580 has never been sharper. So what stands out to 00:12:09.580 --> 00:12:11.259 you in all this? Are we just heading toward a 00:12:11.259 --> 00:12:13.659 future where the only thing sophisticated enough 00:12:13.659 --> 00:12:16.159 to filter the tidal wave of slop is the same 00:12:16.159 --> 00:12:18.139 kind of self -teaching AI that helped create 00:12:18.139 --> 00:12:20.039 it in the first place? It's something to think 00:12:20.039 --> 00:12:22.039 about. We encourage you to explore the sources 00:12:22.039 --> 00:12:24.399 we mentioned and, as always, continue your own 00:12:24.399 --> 00:12:24.840 deep dive.