WEBVTT 00:00:00.000 --> 00:00:04.580 Imagine you train this incredibly smart AI, right? 00:00:04.660 --> 00:00:07.139 It kind of runs so much now. But you train it 00:00:07.139 --> 00:00:11.599 on, well, just viral tweets and clickbait. Slight 00:00:11.599 --> 00:00:14.419 pause. It turns out the intelligence of these 00:00:14.419 --> 00:00:17.719 advanced AIs is surprisingly... pretty fragile. 00:00:17.940 --> 00:00:20.300 Its core reasoning can actually degrade just 00:00:20.300 --> 00:00:22.899 from being fed digital junk. Some are calling 00:00:22.899 --> 00:00:26.120 it AI brain rot. Welcome to the deep dive. Yeah, 00:00:26.179 --> 00:00:28.100 today we're unpacking a really interesting set 00:00:28.100 --> 00:00:30.399 of sources. They show the stark contrast, you 00:00:30.399 --> 00:00:33.039 know, AI is moving unbelievably fast, but at 00:00:33.039 --> 00:00:36.219 the same time, its foundation seems, well, kind 00:00:36.219 --> 00:00:39.310 of brittle. Maybe dangerously so. Exactly. That 00:00:39.310 --> 00:00:41.310 fragility is what we want to focus on, sort of 00:00:41.310 --> 00:00:43.369 cleaning up the information highway, trying to 00:00:43.369 --> 00:00:45.689 understand where these algorithms are, essentially 00:00:45.689 --> 00:00:47.770 doom scrolling, and what kind of real measurable 00:00:47.770 --> 00:00:50.490 damage that does. And then we'll switch gears, 00:00:50.670 --> 00:00:53.310 look at the sheer speed of innovation, new abilities, 00:00:53.409 --> 00:00:55.329 new ways to control it, the whole infrastructure 00:00:55.329 --> 00:00:58.030 race. And finally, and this is really critical, 00:00:58.090 --> 00:01:00.289 I think, we'll get into the hidden reality of 00:01:00.289 --> 00:01:02.469 bias baked right into the training data, and 00:01:02.469 --> 00:01:04.549 crucially, why most of us just don't see it. 00:01:04.670 --> 00:01:06.489 Okay, let's start with that core problem then. 00:01:07.340 --> 00:01:12.159 AI brain rot. It sounds, I don't know, a bit 00:01:12.159 --> 00:01:14.840 dramatic, but the research seems to show it's 00:01:14.840 --> 00:01:17.180 a real quantifiable thing. Oh, it really is. 00:01:17.219 --> 00:01:19.359 Think of it like human internet brain rot, but, 00:01:19.359 --> 00:01:21.939 you know, for algorithms. When a large language 00:01:21.939 --> 00:01:24.799 model just consumes endless low -quality trivial 00:01:24.799 --> 00:01:28.120 stuff, its ability to actually reason and focus 00:01:28.120 --> 00:01:31.659 just tanks fast. The researchers looked at this 00:01:31.659 --> 00:01:34.200 specifically using two kinds of junk content. 00:01:34.420 --> 00:01:37.439 Right. There was the M1 type that's the... Super 00:01:37.439 --> 00:01:40.519 viral click -baity short stuff everywhere. And 00:01:40.519 --> 00:01:42.599 then M2, which they called, what was it, low 00:01:42.599 --> 00:01:45.799 semantic buzzword soup. I mean, I still wrestle 00:01:45.799 --> 00:01:48.159 with prompt drift myself sometimes, getting generic 00:01:48.159 --> 00:01:50.500 answers. So this idea really hits home if you're 00:01:50.500 --> 00:01:52.159 trying to actually use these tools reliably. 00:01:52.459 --> 00:01:55.099 And it was that M1, the viral stuff, that did 00:01:55.099 --> 00:01:57.120 the most damage. The model started doing something 00:01:57.120 --> 00:01:59.140 the researchers called thought skipping. Thought 00:01:59.140 --> 00:02:01.140 skipping. Yeah, basically just stopping processing 00:02:01.140 --> 00:02:03.780 complex ideas properly. The model sort of jumps 00:02:03.780 --> 00:02:05.500 to a conclusion instead of, you know, following 00:02:05.500 --> 00:02:08.560 the steps. Wow. And the numbers they reported 00:02:08.560 --> 00:02:12.300 on this degradation are pretty shocking, actually. 00:02:12.419 --> 00:02:15.580 For that M1 viral content, reasoning accuracy 00:02:15.580 --> 00:02:20.039 just plummeted. It went from like 74 .9 % down 00:02:20.039 --> 00:02:24.199 to 57 .2%. That's a serious hit to basic logic. 00:02:24.460 --> 00:02:26.919 It is. And look at long context comprehension, 00:02:27.340 --> 00:02:29.719 its ability to follow a long conversation or 00:02:29.719 --> 00:02:32.580 document. That fell even harder. dropped from 00:02:32.580 --> 00:02:37.759 84 .4 % to just 52 .3%. Imagine like a professional 00:02:37.759 --> 00:02:40.180 losing that much ability just from reading gossip 00:02:40.180 --> 00:02:42.539 sites for a month. That's the kind of fragility 00:02:42.539 --> 00:02:44.159 we're talking about. And it wasn't just their 00:02:44.159 --> 00:02:46.259 reasoning. The sources said the model's ethical 00:02:46.259 --> 00:02:48.439 alignment slipped, too, and they even saw personality 00:02:48.439 --> 00:02:50.800 drift. It sounds like a clear dose response thing. 00:02:50.900 --> 00:02:52.639 The more junk they got, the worse they performed. 00:02:52.759 --> 00:02:54.860 Dumber and, frankly, less reliable. But what's 00:02:54.860 --> 00:02:56.939 really fascinating and kind of worrying is that 00:02:56.939 --> 00:02:58.699 the effects stuck around. They did try to fix 00:02:58.699 --> 00:03:00.620 it. Yeah, they tried to clean up the data afterwards, 00:03:00.860 --> 00:03:03.599 retrain the models. Exactly. But the key finding 00:03:03.599 --> 00:03:06.240 was even retraining on clean, high quality data, 00:03:06.319 --> 00:03:10.120 it only partly fixed the damage. The rot lingered. 00:03:10.520 --> 00:03:12.159 Seems like the junk data might have actually 00:03:12.159 --> 00:03:14.340 corrupted the model's core structure, its fundamental 00:03:14.340 --> 00:03:17.840 wiring. So if the junk data messes with the model's 00:03:17.840 --> 00:03:21.139 deepest level, how fundamental is that initial 00:03:21.139 --> 00:03:24.240 training data then? It seems like it sets a baseline 00:03:24.240 --> 00:03:27.039 that's really hard, maybe even impossible, to 00:03:27.039 --> 00:03:30.419 fully reset later on. Okay, switching gears now. 00:03:30.620 --> 00:03:33.360 This fragility exists right alongside incredible 00:03:33.360 --> 00:03:36.719 speed. I mean, the pace of new capabilities launching 00:03:36.719 --> 00:03:39.770 right now is just... It really is. We're seeing 00:03:39.770 --> 00:03:42.689 tools pop up that kind of redefine what AI agents 00:03:42.689 --> 00:03:45.789 can even do. OpenAI just launched something called 00:03:45.789 --> 00:03:48.550 ChatGPT Atlas. It's basically their first full 00:03:48.550 --> 00:03:51.750 AI web browser, not just chat. Right. It's more 00:03:51.750 --> 00:03:53.909 like an autonomous agent. It's designed to handle 00:03:53.909 --> 00:03:56.110 complex stuff online for you, like booking flights, 00:03:56.389 --> 00:03:59.250 comparing documents, basically starting and finishing 00:03:59.250 --> 00:04:01.229 detailed tasks without you needing to step in. 00:04:01.310 --> 00:04:03.069 And then on the efficiency side, you've got things 00:04:03.069 --> 00:04:05.710 like DeepSeek OCR. This tool apparently makes 00:04:05.710 --> 00:04:08.430 image -based documents like PDFs 10 times smaller, 00:04:08.569 --> 00:04:12.090 file size -wise, while keeping like 97 % of the 00:04:12.090 --> 00:04:14.210 original info. That's huge for businesses dealing 00:04:14.210 --> 00:04:16.819 with lots of data storage costs. Yeah, absolutely. 00:04:17.139 --> 00:04:19.079 And you see these shifts reflected in the market 00:04:19.079 --> 00:04:20.779 too. There was this recent trading benchmark 00:04:20.779 --> 00:04:24.160 where different AIs got $10 ,000 each to manage. 00:04:24.480 --> 00:04:27.600 Early days. But the results showed DeepSeek 3 00:04:27.600 --> 00:04:31.759 .1 was up $4 ,000, while Gemini 2 .5 Pro was 00:04:31.759 --> 00:04:34.360 actually down $3 ,000. These models are literally 00:04:34.360 --> 00:04:36.660 competing in finance now. And remember that story 00:04:36.660 --> 00:04:39.139 from the UK. Channel 4 aired this documentary. 00:04:39.379 --> 00:04:41.879 And at the end, the host just casually reveals 00:04:41.879 --> 00:04:44.759 they were an AI anchor the whole time. I mean, 00:04:44.779 --> 00:04:47.120 the line between real and synthetic is getting 00:04:47.120 --> 00:04:50.079 seriously blurry. That kind of speed needs powerful 00:04:50.079 --> 00:04:52.439 infrastructure, obviously. On the hardware front, 00:04:52.540 --> 00:04:54.740 it's super competitive. NVIDIA is still the giant. 00:04:54.800 --> 00:04:57.660 But China's claiming they have a new analog AI 00:04:57.660 --> 00:05:00.500 chip that's supposedly 1 ,000 times faster. Whoa. 00:05:01.230 --> 00:05:04.350 Yeah, a thousand times faster. Imagine scaling 00:05:04.350 --> 00:05:07.410 to like a billion queries with that kind of speed. 00:05:07.509 --> 00:05:09.310 That completely changes the game, doesn't it? 00:05:09.350 --> 00:05:11.649 That's infrastructure really shaping what's possible. 00:05:11.850 --> 00:05:13.930 And of course, with these huge capabilities comes 00:05:13.930 --> 00:05:16.490 the need for much better control. You see open 00:05:16.490 --> 00:05:19.089 AI tightening the guardrails on Sora, their video 00:05:19.089 --> 00:05:21.069 generator, especially after Hollywood raised 00:05:21.069 --> 00:05:24.029 concerns about realistic deepfakes. Makes sense. 00:05:24.379 --> 00:05:27.079 And YouTube just rolled out a new tool specifically 00:05:27.079 --> 00:05:29.879 for detecting likenesses, designed to combat 00:05:29.879 --> 00:05:32.819 AI deepfakes and flag synthetic media automatically. 00:05:33.259 --> 00:05:35.899 The control tools are racing to keep up with 00:05:35.899 --> 00:05:38.000 what the AI can do. At the same time, you've 00:05:38.000 --> 00:05:40.319 got this big investment in people. Companies 00:05:40.319 --> 00:05:43.689 like Microsoft, OpenAI, Anthropic. They're putting 00:05:43.689 --> 00:05:46.829 $23 million together just to train, what, 400 00:05:46.829 --> 00:05:49.990 ,000 U .S. teachers on how to use AI tools effectively 00:05:49.990 --> 00:05:52.689 and responsibly? So, OK, if you zoom out, look 00:05:52.689 --> 00:05:54.449 at all these things, the Atlas agent, the super 00:05:54.449 --> 00:05:57.129 fast chips, the new guardrails, the teacher training. 00:05:58.139 --> 00:06:00.540 What's the common thread here? What connects 00:06:00.540 --> 00:06:03.000 these different moves? I'd say it's about better 00:06:03.000 --> 00:06:05.220 control and smarter integration. That seems to 00:06:05.220 --> 00:06:08.279 be the focus for the next wave of AI tool. Right. 00:06:08.319 --> 00:06:10.519 That smarter integration idea brings us beyond 00:06:10.519 --> 00:06:12.560 just basic prompting. We've all had that experience, 00:06:12.759 --> 00:06:14.819 right? You ask an AI for something complex, something 00:06:14.819 --> 00:06:17.360 specific, and you just get back generic mush. 00:06:17.680 --> 00:06:20.199 So what happens when simple one -off prompts 00:06:20.199 --> 00:06:22.420 just aren't enough anymore? Yeah, that's where 00:06:22.420 --> 00:06:24.779 simple prompting falls down. It fails because 00:06:24.779 --> 00:06:27.500 the AI doesn't. really have memory between chats. 00:06:27.759 --> 00:06:30.240 It starts fresh every single time, forgets the 00:06:30.240 --> 00:06:32.740 context, forgets your company style guide, whatever. 00:06:32.939 --> 00:06:35.240 That's where context engineering comes into play. 00:06:35.439 --> 00:06:37.519 Okay, so break down context engineering for us, 00:06:37.519 --> 00:06:40.120 simple terms. It's basically giving the AI a 00:06:40.120 --> 00:06:44.339 kind of permanent organizational memory for specific 00:06:44.339 --> 00:06:46.519 tasks. And how does that actually work in practice? 00:06:47.000 --> 00:06:49.100 Well, it often uses techniques like retrieval 00:06:49.100 --> 00:06:52.000 augmented generation or ROG. You're essentially 00:06:52.000 --> 00:06:54.819 pointing the AI towards a defined, limited set 00:06:54.819 --> 00:06:57.339 of your own data, like internal manuals or style 00:06:57.339 --> 00:06:59.759 guides, maybe through system prompts. This stocks 00:06:59.759 --> 00:07:02.259 it from just defaulting to generic web knowledge 00:07:02.259 --> 00:07:04.740 and gives it consistent, relevant context for 00:07:04.740 --> 00:07:07.199 your needs. So why is just relying on simple 00:07:07.199 --> 00:07:09.980 prompting not good enough for complex or business 00:07:09.980 --> 00:07:13.000 -related tasks? Because the AI needs that history, 00:07:13.139 --> 00:07:15.319 that memory, to give results that are actually 00:07:15.319 --> 00:07:18.540 relevant and not just generic filler. Now, this 00:07:18.540 --> 00:07:20.720 whole search for better context for objective 00:07:20.720 --> 00:07:24.019 information runs smack into another huge problem. 00:07:24.720 --> 00:07:27.220 Invisible bias. We tend to trust these systems 00:07:27.220 --> 00:07:30.160 maybe too much to be neutral. But what if the 00:07:30.160 --> 00:07:32.980 bias isn't obvious? What if it's hidden deep 00:07:32.980 --> 00:07:35.779 inside the data they were trained on? Yeah, this 00:07:35.779 --> 00:07:37.439 is where that study from Penn State and Oregon 00:07:37.439 --> 00:07:40.180 State University comes in. They ran a really... 00:07:40.589 --> 00:07:43.050 quite chilling test looking specifically at whether 00:07:43.050 --> 00:07:45.290 users can even spot bias when it's baked right 00:07:45.290 --> 00:07:47.730 into a system that otherwise seems to work fine. 00:07:47.870 --> 00:07:50.069 They trained a facial emotion recognition system, 00:07:50.170 --> 00:07:53.310 but they deliberately skewed the data. They basically 00:07:53.310 --> 00:07:56.329 taught it that happy equals white faces and sad 00:07:56.329 --> 00:07:59.410 equals black faces. So a clear systemic racial 00:07:59.410 --> 00:08:01.689 bias was built in from the ground up. And the 00:08:01.689 --> 00:08:03.649 findings on whether people notice this, it's 00:08:03.649 --> 00:08:05.829 really unsettling. When they tested this bias 00:08:05.829 --> 00:08:08.839 system, most everyday users just didn't spot 00:08:08.839 --> 00:08:11.800 the racial bias at all. Right. Awareness turned 00:08:11.800 --> 00:08:14.399 out to be highly conditional. It was mainly the 00:08:14.399 --> 00:08:16.319 black participants who started noticing something 00:08:16.319 --> 00:08:18.680 was wrong when the A .I. kept misclassifying 00:08:18.680 --> 00:08:21.439 emotions on black faces. The really stark part 00:08:21.439 --> 00:08:25.360 white participants mostly remained unaware. Which, 00:08:25.420 --> 00:08:28.199 depressingly, means the engineered bias worked 00:08:28.199 --> 00:08:30.959 exactly as intended it was invisible to those 00:08:30.959 --> 00:08:33.419 it didn't negatively affect. That points to a 00:08:33.419 --> 00:08:35.480 really deep danger in how much we trust these 00:08:35.480 --> 00:08:38.259 things. People generally assume the AI was neutral, 00:08:38.440 --> 00:08:40.860 objective, even when its output showed clear 00:08:40.860 --> 00:08:43.700 failures or bias against one group. It really 00:08:43.700 --> 00:08:46.360 shows how easily systemic bias can hide inside 00:08:46.360 --> 00:08:49.679 systems we rely on and trust. So what's the fundamental 00:08:49.679 --> 00:08:53.159 danger of this invisible yet trusted bias then? 00:08:53.340 --> 00:08:55.340 I guess it's that if we can't even see the bias, 00:08:55.399 --> 00:08:57.279 we end up just automatically handing over our 00:08:57.279 --> 00:08:59.120 judgment to systems that are fundamentally flawed. 00:08:59.460 --> 00:09:01.840 Yeah. So looking back across all these sources, 00:09:02.000 --> 00:09:04.399 the picture of AI is just full of these contrasts, 00:09:04.399 --> 00:09:06.139 isn't it? It's incredibly powerful. You've got 00:09:06.139 --> 00:09:08.879 Atlas agents, potentially thousand times faster 00:09:08.879 --> 00:09:11.919 chips coming. But it's also disturbingly fragile, 00:09:12.259 --> 00:09:15.679 prone to this brain rot. And as that emotion 00:09:15.679 --> 00:09:19.190 study showed. It's deeply prone to just reflecting 00:09:19.190 --> 00:09:22.750 back our own worst biases, often invisibly. It 00:09:22.750 --> 00:09:24.549 really feels like it all comes down to the quality 00:09:24.549 --> 00:09:27.029 of what goes in and how objective that reflection 00:09:27.029 --> 00:09:30.450 really is. I mean, if AI starts thinking like 00:09:30.450 --> 00:09:32.649 clickbait because its core programming gets corrupted 00:09:32.649 --> 00:09:35.370 by junk, and then on top of that, we can't even 00:09:35.370 --> 00:09:37.610 spot the deep biases in the data we feed it. 00:09:38.029 --> 00:09:40.409 Well, who's actually in control of objective 00:09:40.409 --> 00:09:42.490 truth then? That's definitely something to chew 00:09:42.490 --> 00:09:45.600 on this week. An excellent thought to end with. 00:09:45.720 --> 00:09:47.559 Thanks everyone for joining us for this deep 00:09:47.559 --> 00:09:49.419 dive into the source material. Yeah, thanks for 00:09:49.419 --> 00:09:52.120 listening. Use this knowledge wisely. Until next 00:09:52.120 --> 00:09:52.360 time.