WEBVTT 00:00:00.000 --> 00:00:02.359 Okay, think about this contract. On one hand, 00:00:02.379 --> 00:00:05.179 this grand idea. Pursuing a complete understanding 00:00:05.179 --> 00:00:07.519 of the universe by using, well, a single closed 00:00:07.519 --> 00:00:10.320 -off AI. That's the huge goal, right? The philosophical 00:00:10.320 --> 00:00:12.699 aim behind something like Rokopedia. But then, 00:00:12.720 --> 00:00:15.240 then you look at the immediate human side. And 00:00:15.240 --> 00:00:16.920 it's pretty stark. Our sources today revealed 00:00:16.920 --> 00:00:19.620 something really shocking. Over a million people 00:00:19.620 --> 00:00:22.420 every week were talking to ChatGPT about suicide. 00:00:22.969 --> 00:00:25.170 That contrast is what we're diving into. This 00:00:25.170 --> 00:00:27.789 huge ambition versus a very real, very urgent 00:00:27.789 --> 00:00:30.410 safety crisis. Welcome to the Deep Dive. Yeah, 00:00:30.449 --> 00:00:32.649 our goal, as always, is to give you that shortcut. 00:00:32.890 --> 00:00:35.149 We've sifted through a ton of sources, really, 00:00:35.189 --> 00:00:38.329 to get a handle on where AI is right now. And 00:00:38.329 --> 00:00:40.789 today we're looking at these new battles over 00:00:40.789 --> 00:00:43.130 content, some really surprising ways people are 00:00:43.130 --> 00:00:46.289 using AI, and maybe most critically, these big 00:00:46.289 --> 00:00:49.039 safety failures happening at... Well, just an 00:00:49.039 --> 00:00:51.920 enormous scale. So here's the plan. First, we'll 00:00:51.920 --> 00:00:54.520 unpack Elon Musk's Grokopedia, you know, what 00:00:54.520 --> 00:00:56.859 it is and why its whole structure kind of challenges 00:00:56.859 --> 00:00:59.399 how we think about finding proof. Then we get 00:00:59.399 --> 00:01:02.119 into some strange stuff like why being rude might 00:01:02.119 --> 00:01:04.299 actually get you better AI answers, but also, 00:01:04.379 --> 00:01:06.920 you know, lifesaving uses, too. And finally, 00:01:07.000 --> 00:01:09.379 yeah, we'll tackle the massive challenges, the 00:01:09.379 --> 00:01:11.319 sheer scale of the infrastructure needed, the 00:01:11.319 --> 00:01:13.060 safety frameworks people are trying to build 00:01:13.060 --> 00:01:15.840 and this this hidden mental health crisis playing 00:01:15.840 --> 00:01:17.719 out inside these models. OK, let's start. with 00:01:17.719 --> 00:01:20.040 wikipedia then the knowledge wars yeah they just 00:01:20.040 --> 00:01:22.819 got another big player elon musk using xai's 00:01:22.819 --> 00:01:25.459 grok has launched this thing basically as a challenger 00:01:25.459 --> 00:01:28.640 to wikipedia it's a really big move and um it's 00:01:28.640 --> 00:01:30.620 serious from the get -go it launched with what 00:01:30.620 --> 00:01:34.739 was it over 885 000 entries already live yeah 00:01:34.739 --> 00:01:37.900 885 000 that's a huge library right out of the 00:01:37.900 --> 00:01:40.480 gate It definitely shows the ambition, doesn't 00:01:40.480 --> 00:01:43.480 it? Trying to replace Wikipedia fast. And the 00:01:43.480 --> 00:01:45.939 way it works is, well, it's simple, but totally 00:01:45.939 --> 00:01:48.799 different from, say, Wikipedia. Grok makes the 00:01:48.799 --> 00:01:50.719 articles, it checks them against its own data, 00:01:50.760 --> 00:01:53.420 and boom, publishes them. Right. And Musk himself 00:01:53.420 --> 00:01:56.620 said Grokopedia is, quote, a necessary step to 00:01:56.620 --> 00:01:59.340 understanding the universe. He even wants Grok 00:01:59.340 --> 00:02:02.200 to stop referencing Wikipedia entirely by year's 00:02:02.200 --> 00:02:05.340 end. So, yeah, on the surface, just another competitor. 00:02:05.519 --> 00:02:07.340 But when you dig into the sources, you see these 00:02:07.340 --> 00:02:10.500 three key differences in how it's built that 00:02:10.500 --> 00:02:13.419 really change the game for knowledge creation. 00:02:13.699 --> 00:02:16.120 What jumps out, really, is how different it is 00:02:16.120 --> 00:02:18.340 from how we usually check information. First, 00:02:18.479 --> 00:02:21.039 it's totally closed source. You, me, nobody in 00:02:21.039 --> 00:02:22.939 the public can edit it. Exactly. And second, 00:02:23.080 --> 00:02:25.680 it's all curated by the AI. No human subject 00:02:25.680 --> 00:02:28.120 experts involved in the writing or editing process. 00:02:28.400 --> 00:02:30.360 And here's the kicker. The really critical part, 00:02:30.520 --> 00:02:33.620 there are zero citations, no inline sources at 00:02:33.620 --> 00:02:36.120 all. Yeah. And that lack of sourcing, that's 00:02:36.120 --> 00:02:38.400 powerful. The sources we read pointed out that 00:02:38.400 --> 00:02:40.639 this lets Grokopedia kind of quietly rewrite 00:02:40.639 --> 00:02:43.139 things. The structure itself becomes about control. 00:02:43.620 --> 00:02:45.860 Like the example they gave comparing the Wikipedia 00:02:45.860 --> 00:02:48.520 page for Donald Trump. It has all these specific 00:02:48.520 --> 00:02:50.759 details, some controversial, right? Qatar jets, 00:02:50.960 --> 00:02:53.219 Trump coins, that kind of stuff. Right. But Grokopedia's 00:02:53.219 --> 00:02:55.639 version, it just skips over those details. Poof. 00:02:55.979 --> 00:02:58.879 Gone. It makes for a cleaner story. maybe less 00:02:58.879 --> 00:03:01.599 messy, but you lose the full picture, the accountability. 00:03:01.860 --> 00:03:03.939 It's like a knowledge shortcut, but it takes 00:03:03.939 --> 00:03:06.620 away the tools for critical thinking. So if the 00:03:06.620 --> 00:03:11.500 goal really is this necessary understanding of 00:03:11.500 --> 00:03:14.199 the universe, how big of a problem is leaving 00:03:14.199 --> 00:03:16.979 out basic citations, especially for anyone trying 00:03:16.979 --> 00:03:19.500 to actually check the claims? Well, missing sources 00:03:19.500 --> 00:03:21.759 basically means you lose that foundation for 00:03:21.759 --> 00:03:24.080 critical thinking, right? The AI becomes the 00:03:24.080 --> 00:03:26.240 only judge of what's true. And this whole fight 00:03:26.240 --> 00:03:28.439 over who controls the story, it leads right into 00:03:28.439 --> 00:03:30.939 the next part. The really fascinating, sometimes 00:03:30.939 --> 00:03:33.099 just weird ways people are actually using these 00:03:33.099 --> 00:03:35.900 AI tools now. Yeah, this is where it gets wild. 00:03:36.020 --> 00:03:38.379 You'd think, okay, be polite, be super clear 00:03:38.379 --> 00:03:41.919 with the AI. But nope, a study actually found 00:03:41.919 --> 00:03:44.979 GPT gave better, more accurate answers when people 00:03:44.979 --> 00:03:46.719 were rude to it. I know, it's funny, isn't it? 00:03:46.719 --> 00:03:49.439 Like the AI gets pressured like a person sometimes. 00:03:49.560 --> 00:03:52.960 But obviously, got to add the warning. Don't 00:03:52.960 --> 00:03:54.639 actually do this. It's probably bad for your 00:03:54.639 --> 00:03:56.939 behavior long term. Still, the fact it works, 00:03:56.979 --> 00:03:59.719 it's just odd. Totally odd. But then you have 00:03:59.719 --> 00:04:01.759 the really positive hacks. We saw this story 00:04:01.759 --> 00:04:04.680 someone shared, right, about using nine specific 00:04:04.680 --> 00:04:08.159 custom prompts to learn French in four weeks. 00:04:08.300 --> 00:04:10.969 Four weeks. That's seriously fast structured 00:04:10.969 --> 00:04:14.210 learning using the tool really effectively. That's 00:04:14.210 --> 00:04:16.589 the super useful end. But then AI is also making 00:04:16.589 --> 00:04:19.149 viral content like scarily easy. Remember that 00:04:19.149 --> 00:04:21.750 TikTok thing? Deepfake Queen Elizabeth vlogging? 00:04:22.069 --> 00:04:24.370 Oh, yeah. Millions of views. It just shows how 00:04:24.370 --> 00:04:27.870 tools like Sora 2 are making going viral almost 00:04:27.870 --> 00:04:32.399 trivial. The power here is just. And honestly, 00:04:32.579 --> 00:04:35.100 I still wrestle with prompt drift myself sometimes, 00:04:35.279 --> 00:04:37.040 you know, where the AI kind of forgets what you 00:04:37.040 --> 00:04:38.860 asked it a few turns back. It's easy to feel 00:04:38.860 --> 00:04:40.879 like the tool is smarter than you. Yeah, I get 00:04:40.879 --> 00:04:42.660 that. But then you see the flip side, the truly 00:04:42.660 --> 00:04:45.220 life -saving stuff. That Reddit post, someone 00:04:45.220 --> 00:04:48.060 claiming GPT saved their mom's life, spotted 00:04:48.060 --> 00:04:50.699 an infection doctors missed. Right. And the comments 00:04:50.699 --> 00:04:52.800 were full of similar stories, not just minor 00:04:52.800 --> 00:04:55.180 things either. Real impact. That kind of deep 00:04:55.180 --> 00:04:58.180 capability. Yeah. Next to deep fake queens. Just 00:04:58.180 --> 00:04:59.920 highlights how much variance there is in this 00:04:59.920 --> 00:05:02.209 tech. And that variance, it's hitting the professional 00:05:02.209 --> 00:05:04.470 world too. Like Anthropic, they're making some 00:05:04.470 --> 00:05:07.250 quiet but big moves. Yeah. Cloud 4 .5 is now 00:05:07.250 --> 00:05:10.529 inside Excel. Yep, inside Excel. Specializing 00:05:10.529 --> 00:05:12.930 in analyzing live market data, summarizing earnings 00:05:12.930 --> 00:05:15.829 calls. They're basically taking aim right at 00:05:15.829 --> 00:05:18.509 Microsoft's co -pilot, but specifically in finance. 00:05:18.529 --> 00:05:21.029 Big stakes there. Okay, so we've got these wildly 00:05:21.029 --> 00:05:23.829 different uses. Saving lives, making deep fakes 00:05:23.829 --> 00:05:27.250 go viral, rewiring finance. This huge range. 00:05:27.470 --> 00:05:29.750 What does it really tell us about how reliable? 00:05:29.870 --> 00:05:32.490 AI is right now? Well, it tells us AI is a super 00:05:32.490 --> 00:05:34.569 high variance tool. It can do amazing good or, 00:05:34.610 --> 00:05:36.350 you know, cause real concern. It all depends 00:05:36.350 --> 00:05:38.050 on the prompt, the context, the application. 00:05:38.430 --> 00:05:41.069 Mid -roll, sponsor read, do not write. Welcome 00:05:41.069 --> 00:05:43.449 back to the Deep Dive. So we've touched on knowledge 00:05:43.449 --> 00:05:46.689 control, these surprising AI uses. Now we need 00:05:46.689 --> 00:05:49.569 to talk about the sheer scale, the infrastructure, 00:05:49.829 --> 00:05:51.509 and the safety issues that come with it. Yeah, 00:05:51.529 --> 00:05:53.610 the engine driving the scale. It's running incredibly 00:05:53.610 --> 00:05:56.620 hot. Just look at the money. Crusoe there, an 00:05:56.620 --> 00:06:00.160 AI and for a company just raised one point three 00:06:00.160 --> 00:06:04.040 seven five billion dollars billion with a B. 00:06:04.379 --> 00:06:07.959 Wow. Right. Puts a valuation over 10 billion 00:06:07.959 --> 00:06:11.500 dollars. And they're planning these huge AI campuses 00:06:11.500 --> 00:06:14.339 down in Texas backed by NVIDIA just to handle 00:06:14.339 --> 00:06:16.100 the computing power needed. And it's not just 00:06:16.100 --> 00:06:18.339 money, it's talent, too. The sources talked about 00:06:18.339 --> 00:06:21.600 the modification of open AI, like over 600 people 00:06:21.600 --> 00:06:25.019 out of their 3000 staff. X meta. That's a huge 00:06:25.019 --> 00:06:27.060 influx. And it's fueling all those rumors about, 00:06:27.079 --> 00:06:29.980 you know, ads eventually showing up inside ChatGPT. 00:06:30.040 --> 00:06:32.560 Follow the talent, right? Whoa. Just pause for 00:06:32.560 --> 00:06:34.720 a second. Imagine trying to scale that kind of 00:06:34.720 --> 00:06:37.220 infrastructure, handling a billion complex questions 00:06:37.220 --> 00:06:39.680 every single day. The sheer operational challenge 00:06:39.680 --> 00:06:42.060 is mind blowing. It really is. It's an arms race. 00:06:42.300 --> 00:06:44.420 Physical infrastructure, brain power, everything. 00:06:44.680 --> 00:06:48.740 And as things scale up like this, the risks just 00:06:48.740 --> 00:06:51.839 grow exponentially, don't they? Our sources really 00:06:51.839 --> 00:06:54.870 emphasized. needing standardized safety. They 00:06:54.870 --> 00:06:57.310 mentioned the NIST framework. Yeah, the National 00:06:57.310 --> 00:06:59.350 Institute of Standards and Technology. It's basically 00:06:59.350 --> 00:07:01.209 the U .S. government's guidelines for building 00:07:01.209 --> 00:07:03.870 AI systems that are trustworthy, safe, reliable 00:07:03.870 --> 00:07:06.470 by design. Yet even with frameworks like that, 00:07:06.589 --> 00:07:09.269 we're seeing immediate security problems. Brave, 00:07:09.370 --> 00:07:12.110 the browser company, recently reported that ChatGPT 00:07:12.110 --> 00:07:14.949 Atlas, that's supposed to be the safety layer, 00:07:15.050 --> 00:07:16.529 right? Right, the alignment and the safety part 00:07:16.529 --> 00:07:18.569 of the current models. Yeah, well, Brave found 00:07:18.569 --> 00:07:21.810 it's actually easy to hijack. which kind of compromises 00:07:21.810 --> 00:07:23.550 the whole point of having that safety layer in 00:07:23.550 --> 00:07:25.850 the first place. And that, that brings us inevitably 00:07:25.850 --> 00:07:28.170 to the mental health data. This part's tough, 00:07:28.269 --> 00:07:30.829 and we need to give it the seriousness it deserves. 00:07:31.449 --> 00:07:34.149 Two sec silence. OpenAI released a statistic 00:07:34.149 --> 00:07:37.250 that's just staggering. Over 1 million unique 00:07:37.250 --> 00:07:40.509 users every single week are talking to ChatGPT 00:07:40.509 --> 00:07:43.470 about suicidal thoughts. 1 million per week. 00:07:43.610 --> 00:07:46.589 That's 0 .15 % of their weekly users. But the 00:07:46.589 --> 00:07:49.779 raw number, it's huge. It's an enormous kind 00:07:49.779 --> 00:07:52.980 of invisible crisis, isn't it? Millions are turning 00:07:52.980 --> 00:07:56.680 to this. Well, unregulated, uncredentialed AI 00:07:56.680 --> 00:07:59.560 for help in moments of crisis, often because 00:07:59.560 --> 00:08:02.420 maybe a human option isn't there or isn't fast 00:08:02.420 --> 00:08:04.579 enough. It sounds like this data came out alongside 00:08:04.579 --> 00:08:06.459 a big push to show they're improving things. 00:08:06.579 --> 00:08:09.100 What are they actually doing about this, given 00:08:09.100 --> 00:08:11.420 the pressure they must be under? Well, they say 00:08:11.420 --> 00:08:13.860 they worked with over 170 mental health professionals 00:08:13.860 --> 00:08:17.420 to improve GPT -5 specifically for this. They're 00:08:17.420 --> 00:08:20.920 claiming it's now 65 percent better at giving 00:08:20.920 --> 00:08:23.560 what they call desirable responses around suicide. 00:08:23.680 --> 00:08:27.639 OK. And safety compliance like. Consistency in 00:08:27.639 --> 00:08:29.819 giving safe answers is apparently up too, from 00:08:29.819 --> 00:08:33.519 77 % to 91%, especially in longer conversations. 00:08:34.159 --> 00:08:36.500 They're also rolling out stricter parental controls, 00:08:36.639 --> 00:08:39.659 using new tech to guess user age, and apply tighter 00:08:39.659 --> 00:08:41.860 rules for minors automatically. That sounds like 00:08:41.860 --> 00:08:44.279 progress, at least in the lab models. But where's 00:08:44.279 --> 00:08:45.860 the catch? What's the pressure point in how they 00:08:45.860 --> 00:08:47.720 actually deploy this stuff? The pressure point 00:08:47.720 --> 00:08:51.169 is exactly that. All this progress is happening 00:08:51.169 --> 00:08:54.830 under, you know, huge legal and public scrutiny. 00:08:55.009 --> 00:08:58.370 But they still offer older, less safe models 00:08:58.370 --> 00:09:01.289 like GPT -4 to people who pay subscriptions. 00:09:01.610 --> 00:09:04.269 So they have safer models, but they aren't making 00:09:04.269 --> 00:09:06.629 everyone use them. Exactly. The safest versions 00:09:06.629 --> 00:09:09.549 aren't universally mandatory, which leads to 00:09:09.549 --> 00:09:11.779 the question. Yeah. When companies are scaling 00:09:11.779 --> 00:09:14.700 this fast, how much of this is just good PR versus 00:09:14.700 --> 00:09:17.019 actually prioritizing getting the safest possible 00:09:17.019 --> 00:09:20.100 models out there for everyone immediately? Well, 00:09:20.139 --> 00:09:23.059 real safety would mean immediate universal deployment 00:09:23.059 --> 00:09:25.500 of the absolute best, safest models they have. 00:09:25.639 --> 00:09:27.820 Not just, you know, nice looking stats that might 00:09:27.820 --> 00:09:30.179 hide the fact that older, riskier versions are 00:09:30.179 --> 00:09:32.700 still out there being used. This has been, well, 00:09:32.860 --> 00:09:35.399 a really deep dive into a super complicated space. 00:09:35.519 --> 00:09:37.779 The big tension that keeps coming up is this 00:09:37.779 --> 00:09:41.210 relentless drive for like. Knowledge control 00:09:41.210 --> 00:09:43.909 with Grokopedia and massive economic scale like 00:09:43.909 --> 00:09:47.029 Crusoe raising billions or all the meta folks 00:09:47.029 --> 00:09:49.830 moving to open AI. Right. And that drive, that 00:09:49.830 --> 00:09:52.190 massive scaling, it's constantly bumping up against 00:09:52.190 --> 00:09:55.669 these really profound ethical risks, especially 00:09:55.669 --> 00:09:57.610 around mental health, like we just discussed, 00:09:57.809 --> 00:10:00.490 and basic data security. We're seeing the consequences 00:10:00.490 --> 00:10:02.950 now at scale. So as we wrap up, maybe something 00:10:02.950 --> 00:10:05.870 for you, the listener, to think about. Consider 00:10:05.870 --> 00:10:09.850 this. When people talk about AI safety, about... 00:10:10.029 --> 00:10:12.110 building these things responsibly from the ground 00:10:12.110 --> 00:10:15.210 up. Why does Anthropix's name seem to come up 00:10:15.210 --> 00:10:18.029 first so often now, maybe more than OpenAI's? 00:10:18.190 --> 00:10:20.570 What might that tell us about public perception 00:10:20.570 --> 00:10:23.230 or maybe even their core design philosophies 00:10:23.230 --> 00:10:27.360 about safety versus maybe commercial speed? Yeah, 00:10:27.379 --> 00:10:29.100 that's a good question to chew on. We really 00:10:29.100 --> 00:10:30.840 encourage you to look critically at where your 00:10:30.840 --> 00:10:33.460 information comes from. Is it cited? Is it purely 00:10:33.460 --> 00:10:36.100 AI generated? And maybe look into these frameworks 00:10:36.100 --> 00:10:38.779 like NIST that are trying to put some guardrails 00:10:38.779 --> 00:10:41.200 on these incredibly powerful tools. Keep doing 00:10:41.200 --> 00:10:43.500 your own deep dive. Thanks for joining us today. 00:10:43.620 --> 00:10:44.320 Out to your own music.