WEBVTT 00:00:00.000 --> 00:00:04.660 DPT -5 just managed to score a gold medal in 00:00:04.660 --> 00:00:08.060 a Ph .D. level astronomy exam. It actually outperformed 00:00:08.060 --> 00:00:10.259 some of the best human participants in the world. 00:00:10.820 --> 00:00:13.519 Beat. It feels like we're approaching the limit, 00:00:13.599 --> 00:00:16.019 not just of what machines can compute, but maybe, 00:00:16.039 --> 00:00:17.660 you know, the limits of what human knowledge 00:00:17.660 --> 00:00:19.280 has already gathered. That's a really interesting 00:00:19.280 --> 00:00:21.420 way to put it. Yeah, it shows us exactly where 00:00:21.420 --> 00:00:24.440 that frontier, the edge of capability is right 00:00:24.440 --> 00:00:27.059 now. Totally. Welcome, everyone, to the Deep 00:00:27.059 --> 00:00:30.120 Dive. So we've synthesized the latest stack of 00:00:30.120 --> 00:00:31.699 your shared sources, you know, these critical 00:00:31.699 --> 00:00:34.479 reports, academic studies, the industry newsletters. 00:00:34.619 --> 00:00:37.179 And our mission today, well, it's pretty simple. 00:00:37.880 --> 00:00:39.619 We need to figure out what these stunning AI 00:00:39.619 --> 00:00:42.320 achievements really mean for academia, for the 00:00:42.320 --> 00:00:46.140 big shifts happening in jobs and for ethics globally. 00:00:46.280 --> 00:00:48.740 Yeah, we've got quite a dive planned. So first, 00:00:48.820 --> 00:00:50.960 we'll definitely start with that astonishing 00:00:50.960 --> 00:00:54.159 AI academic mastery in cosmology, you know, where 00:00:54.159 --> 00:00:55.880 the models just completely raise the bar for 00:00:55.880 --> 00:00:58.520 human testing. Then we're going to shift gears 00:00:58.520 --> 00:01:00.619 pretty quickly. We'll cover key trends in creative 00:01:00.619 --> 00:01:03.100 AI adoption and also what's happening with global 00:01:03.100 --> 00:01:05.459 regulation. And finally, and this is crucial, 00:01:05.620 --> 00:01:08.000 we really have to take a critical look at the 00:01:08.000 --> 00:01:11.019 new internal data that details the political 00:01:11.019 --> 00:01:13.459 bias deep inside these large language models. 00:01:13.680 --> 00:01:15.500 Right. Think of these different sources like 00:01:15.500 --> 00:01:19.159 stacking Lego blocks of data. Quickly build them 00:01:19.159 --> 00:01:20.920 up, put them together into a clear view so you 00:01:20.920 --> 00:01:22.920 can walk away really understanding the critical 00:01:22.920 --> 00:01:25.819 nuances. Let's unpack this pile. Let's get into 00:01:25.819 --> 00:01:27.609 it. Okay, let's start where the sources were, 00:01:27.730 --> 00:01:30.829 frankly, most surprising, the academic world. 00:01:31.209 --> 00:01:35.709 A new paper just dropped showing GPT -5 and Gemini 00:01:35.709 --> 00:01:38.489 2 .5 Pro achieving gold medal levels on the International 00:01:38.489 --> 00:01:42.390 Olympiad on Astronomy and Astrophysics, the IOAA. 00:01:42.730 --> 00:01:45.120 Yeah. Now, that sounds impressive as a headline, 00:01:45.299 --> 00:01:47.040 but we need to grasp the level of difficulty 00:01:47.040 --> 00:01:48.799 we're talking about here. Well, what's fascinating, 00:01:49.019 --> 00:01:50.840 right, is that this isn't just recalling facts 00:01:50.840 --> 00:01:53.159 or definitions, not simple trivia. The researchers, 00:01:53.280 --> 00:01:56.519 they tested the models using actual IOAA exams 00:01:56.519 --> 00:02:00.400 from 2022 right through to 2025 projections. 00:02:00.859 --> 00:02:04.079 These are serious PhD -level, multi -step theoretical 00:02:04.079 --> 00:02:07.019 problems. They need deep physical understanding, 00:02:07.319 --> 00:02:09.819 complex math. I mean, they're designed for the 00:02:09.819 --> 00:02:12.740 absolute most dedicated space nerds on the planet. 00:02:13.159 --> 00:02:15.080 when you look at the scores, the performance 00:02:15.080 --> 00:02:18.039 is. Well, it's almost unbelievable. Looking at 00:02:18.039 --> 00:02:20.300 that breakdown, GPT -5 was just dominant in the 00:02:20.300 --> 00:02:24.039 theoretical exams. Yeah. Scored 93 .0 % in 2022, 00:02:24.520 --> 00:02:27.259 nearly 90 % the next year, and still strong at 00:02:27.259 --> 00:02:31.259 86 .8 % for the hypothetical 2025 questions. 00:02:31.599 --> 00:02:34.219 Right. And while Gemini 2 .5 Pro actually managed 00:02:34.219 --> 00:02:37.340 to take a narrow lead in the 2024 exam with 83 00:02:37.340 --> 00:02:41.360 .0%. The standout figure for GPT -5, for me anyway, 00:02:41.460 --> 00:02:43.400 was its exceptional performance in the data analysis 00:02:43.400 --> 00:02:45.879 section. It scored 88 .5 % there. was actually 00:02:45.879 --> 00:02:48.120 higher than its general theoretical score. Okay, 00:02:48.199 --> 00:02:49.479 so what does that tell us? Well, it suggests 00:02:49.479 --> 00:02:51.599 the model isn't just good at, you know, recalling 00:02:51.599 --> 00:02:53.979 principles. It seems to excel at handling complex, 00:02:54.080 --> 00:02:56.939 kind of messy, real -world data, maybe even better 00:02:56.939 --> 00:02:59.680 than generalized theory. So if we step back for 00:02:59.680 --> 00:03:02.960 a second, what does this mean for us, for you 00:03:02.960 --> 00:03:06.099 listening? The model didn't just meet the gold 00:03:06.099 --> 00:03:10.180 medal thresholds. In multiple years, GPC -5 actually 00:03:10.180 --> 00:03:13.400 outperformed the best human participants competing. 00:03:13.719 --> 00:03:16.020 Yeah. That just changes the whole conversation. 00:03:16.319 --> 00:03:18.360 It really does. Now, I should add the caveat. 00:03:18.639 --> 00:03:21.159 You know, not all models are quite there yet. 00:03:21.319 --> 00:03:23.479 Claude Sonnet 4, for example, fell noticeably 00:03:23.479 --> 00:03:26.659 short. OK. But crucially, all the top models 00:03:26.659 --> 00:03:28.900 still made what the researchers called human 00:03:28.900 --> 00:03:31.520 -like mistakes. They didn't get perfect scores. 00:03:31.639 --> 00:03:33.379 They weren't flawless. Right. So they aren't 00:03:33.379 --> 00:03:35.520 achieving some perfect theoretical truth then. 00:03:35.599 --> 00:03:39.719 They're mimicking human error patterns. What 00:03:39.719 --> 00:03:41.319 does that tell us about their current learning 00:03:41.319 --> 00:03:44.060 methods, maybe their future trajectory? Well, 00:03:44.139 --> 00:03:46.860 it suggests they aren't finding some. you know 00:03:46.860 --> 00:03:48.919 perfect objective truth out there they seem to 00:03:48.919 --> 00:03:51.300 be replicating the inherent gaps maybe the biases 00:03:51.300 --> 00:03:53.400 the blind spots that are present in the human 00:03:53.400 --> 00:03:55.960 scientific literature they trained on which means 00:03:55.960 --> 00:03:58.539 right now the models are acting as incredibly 00:03:58.539 --> 00:04:01.099 powerful mirrors of our knowledge not necessarily 00:04:01.099 --> 00:04:04.159 perfect originators of new knowledge that's a 00:04:04.159 --> 00:04:06.439 key distinction so here's the implication then 00:04:06.439 --> 00:04:11.080 if ai can crush these complex multi -step tests, 00:04:11.120 --> 00:04:13.860 tests needing reasoning, sophisticated data handling, 00:04:14.020 --> 00:04:16.860 then maybe using these really rigorous science 00:04:16.860 --> 00:04:20.100 exams needs to become the new global gold standard 00:04:20.100 --> 00:04:23.279 for benchmarking AI capabilities. Exactly. We 00:04:23.279 --> 00:04:25.379 have to move past like simple reading comprehension 00:04:25.379 --> 00:04:29.629 tests. The bar just went way, way up. Whoa. Just 00:04:29.629 --> 00:04:32.170 imagine scaling that kind of analytical power 00:04:32.170 --> 00:04:34.610 across every single scientific discipline, you 00:04:34.610 --> 00:04:36.670 know, from biochemistry to advanced particle 00:04:36.670 --> 00:04:39.629 physics. The pace of discovery is going to accelerate 00:04:39.629 --> 00:04:42.689 wildly. It forces us to redefine what expert 00:04:42.689 --> 00:04:45.069 human thinking even means. OK, so let's circle 00:04:45.069 --> 00:04:47.370 back to that point about mistakes. If models 00:04:47.370 --> 00:04:49.050 are still making those human -like mistakes, 00:04:49.350 --> 00:04:51.110 what does that tell us about their current learning 00:04:51.110 --> 00:04:53.310 methods? They replicate human error patterns, 00:04:53.509 --> 00:04:55.350 showing their current training boundaries are 00:04:55.350 --> 00:04:58.410 still imperfect copies. Imperfect copies. That's 00:04:58.410 --> 00:05:02.009 a powerful thought to start with. So if AI has 00:05:02.009 --> 00:05:04.009 mastered the abstract rules of the universe, 00:05:04.209 --> 00:05:06.629 you know, in astrophysics, the next question 00:05:06.629 --> 00:05:09.189 has to be, how quickly is it rewriting the rules 00:05:09.189 --> 00:05:11.509 here on Earth, the social rules, the legal rules? 00:05:11.689 --> 00:05:15.149 Let's move from academic genius to how AI is 00:05:15.149 --> 00:05:18.329 shaping culture, industry, law right now. Yeah, 00:05:18.389 --> 00:05:20.329 we're seeing a major cultural shift happening. 00:05:20.389 --> 00:05:22.939 Let's look at the creative market first. generative 00:05:22.939 --> 00:05:25.540 video. It's truly hitting the mainstream consciousness 00:05:25.540 --> 00:05:28.779 now. We've all seen those viral sore generated 00:05:28.779 --> 00:05:32.220 Olympics memes, you know, the ones Jesus swimming, 00:05:32.300 --> 00:05:33.959 the smoking Olympics. Oh, yeah. I saw the smoking 00:05:33.959 --> 00:05:36.360 Olympics one. It felt like like the speed of 00:05:36.360 --> 00:05:38.939 creative absurdity is just astonishing now, almost 00:05:38.939 --> 00:05:41.100 overwhelming. It really is. And meanwhile, the 00:05:41.100 --> 00:05:43.399 heavy hitters, they're preparing for the next 00:05:43.399 --> 00:05:45.939 big thing. We saw a huge signal from XAI. They're 00:05:45.939 --> 00:05:48.060 actively recruiting NVIDIA specialists specifically 00:05:48.060 --> 00:05:51.259 to create world models aimed at video game creation. 00:05:51.420 --> 00:05:53.699 World models for games. Yeah. This is about building 00:05:53.699 --> 00:05:56.120 persistent, believable digital worlds. That's 00:05:56.120 --> 00:05:58.839 a massive market signal for where future investment 00:05:58.839 --> 00:06:01.740 is likely headed. That's a huge focus. Okay. 00:06:01.779 --> 00:06:05.329 And on the more accessible side. There was that 00:06:05.329 --> 00:06:08.769 subtle hint about chat GPT possibly becoming 00:06:08.769 --> 00:06:13.069 more social based on a hidden messaging tab OpenAI's 00:06:13.069 --> 00:06:16.050 COO showed. Right. Maybe they want to shift from 00:06:16.050 --> 00:06:18.110 just being a utility to more of a communication 00:06:18.110 --> 00:06:20.850 platform. We'll see. Plus, I always love finding 00:06:20.850 --> 00:06:23.329 these practical like daily use things. A researcher 00:06:23.329 --> 00:06:25.569 just dropped a free AI tool that converts any 00:06:25.569 --> 00:06:28.860 PDF into a fillable form. Almost instantly. Oh, 00:06:28.939 --> 00:06:31.319 nice. Yeah, you just upload it. It auto detects 00:06:31.319 --> 00:06:34.199 fields. You export. Super useful for, you know. 00:06:34.620 --> 00:06:37.060 The rest of us not building world models. Definitely 00:06:37.060 --> 00:06:40.060 handy. OK, now let's talk governance, because 00:06:40.060 --> 00:06:42.439 the battles here are really heating up. The use 00:06:42.439 --> 00:06:45.519 of AI is just out casing regulation. Hollywood's 00:06:45.519 --> 00:06:47.579 above the line union. So writers, directors, 00:06:47.860 --> 00:06:51.019 actors, they are gearing up for serious AI negotiations. 00:06:51.100 --> 00:06:53.399 Right. And experts are already pointing to things 00:06:53.399 --> 00:06:55.639 like the virtual actress Tilly Norwood as a key 00:06:55.639 --> 00:06:58.399 negotiating issue. It really defines the future 00:06:58.399 --> 00:07:01.470 of digital labor, IP ownership. All that stuff. 00:07:01.649 --> 00:07:03.189 It's not just the unions trying to catch up, 00:07:03.209 --> 00:07:04.850 though. The official regulatory landscape is 00:07:04.850 --> 00:07:07.670 scrambling, too. Yeah. Oh, absolutely. Globally, 00:07:07.670 --> 00:07:10.449 the EU just launched its massive one billion 00:07:10.449 --> 00:07:13.569 apply AI plan. They're offering industries free 00:07:13.569 --> 00:07:17.430 supercomputer access, new AI hubs, really trying 00:07:17.430 --> 00:07:20.009 to push industrial adoption across Europe. And 00:07:20.009 --> 00:07:23.050 in the U .S. Well, California. often setting 00:07:23.050 --> 00:07:24.990 the precedent, became the first state to actually 00:07:24.990 --> 00:07:27.889 regulate AI -companion chatbots. That touches 00:07:27.889 --> 00:07:30.129 directly on sensitive areas like mental health 00:07:30.129 --> 00:07:32.410 and data privacy. Wait, let's go back to Tilly 00:07:32.410 --> 00:07:34.329 Norwood for a second. If unions and regulators 00:07:34.329 --> 00:07:36.649 are pushing back, are the unions fighting for 00:07:36.649 --> 00:07:40.149 her to be considered like an asset or a contracted 00:07:40.149 --> 00:07:42.350 employee? What's the core legal challenge there? 00:07:42.680 --> 00:07:44.819 Well, the core conflict really boils down to 00:07:44.819 --> 00:07:47.339 defining ownership of the performance. If Tilly 00:07:47.339 --> 00:07:49.779 Norwood's digital likeness is trained on a real 00:07:49.779 --> 00:07:52.480 actor's movements, their voice, are we paying 00:07:52.480 --> 00:07:55.279 royalties to that original actor? Or is the synthetic 00:07:55.279 --> 00:07:57.259 character a completely separate entity? It's 00:07:57.259 --> 00:07:59.680 fundamentally a battle over who owns the creative 00:07:59.680 --> 00:08:01.920 labor, whether it's physical or virtual. Gotcha. 00:08:02.000 --> 00:08:04.300 Okay. And we also saw some significant tension 00:08:04.300 --> 00:08:06.360 bubbling up in the foundations of this tech, 00:08:06.540 --> 00:08:10.180 the data infrastructure itself. China issued 00:08:10.180 --> 00:08:14.199 a threat to, quote, Pop the entire AI data center 00:08:14.199 --> 00:08:17.220 bubble. That sounds serious. A major economic 00:08:17.220 --> 00:08:19.680 and geopolitical risk. It creates huge instability. 00:08:19.819 --> 00:08:22.720 Yeah. training these massive world models, it 00:08:22.720 --> 00:08:26.220 requires just colossal clusters of GPUs, enormous 00:08:26.220 --> 00:08:28.500 amounts of energy. So when the infrastructure, 00:08:28.660 --> 00:08:30.720 the access to that hardware becomes dependent 00:08:30.720 --> 00:08:33.360 on political stability or, you know, single nation 00:08:33.360 --> 00:08:36.659 states, the entire global progress in AI is potentially 00:08:36.659 --> 00:08:40.519 at risk. OK, so with unions and regulation ramping 00:08:40.519 --> 00:08:42.480 up everywhere, what do you say is the single 00:08:42.480 --> 00:08:45.139 biggest emerging area of legal conflict? Data 00:08:45.139 --> 00:08:47.340 ownership and labor definitions, especially for 00:08:47.340 --> 00:08:49.659 virtual assets like Tilly Norwood. mid -roll 00:08:49.659 --> 00:08:52.480 sponsor replace holder. Welcome back. The sheer 00:08:52.480 --> 00:08:54.879 power of these tools, you know, shown by crushing 00:08:54.879 --> 00:08:57.480 those PhD level tests, it just demands rigorous 00:08:57.480 --> 00:09:00.059 ethical testing. Even as the models get better 00:09:00.059 --> 00:09:01.960 at science, we have to look really closely at 00:09:01.960 --> 00:09:04.639 bias. Let's shift to that new internal study 00:09:04.639 --> 00:09:07.110 on political bias that OpenAI released. Yeah, 00:09:07.169 --> 00:09:08.830 this is critically important. It really gets 00:09:08.830 --> 00:09:11.470 to the heart of alignment. So OpenAI claims, 00:09:11.750 --> 00:09:14.610 based on their own internal tests, that GPT -5 00:09:14.610 --> 00:09:17.870 is about 30 % less politically biased than both 00:09:17.870 --> 00:09:21.870 GPT -4 and the newer GPT -4 .0. Okay, 30 % less 00:09:21.870 --> 00:09:24.129 biased. How did they test that? They were pretty 00:09:24.129 --> 00:09:27.190 rigorous. They used 500 different prompts across 00:09:27.190 --> 00:09:30.490 100 sensitive topics. Then they graded the responses 00:09:30.490 --> 00:09:33.590 on five specific... Bias metrics, things like 00:09:33.590 --> 00:09:36.009 neutrality, emotional mirroring, that sort of 00:09:36.009 --> 00:09:38.330 thing. That 30 % reduction is progress. Yeah, 00:09:38.389 --> 00:09:40.450 definitely. Yeah. But according to their own 00:09:40.450 --> 00:09:43.169 research, bias still manages to creep in. It 00:09:43.169 --> 00:09:45.590 shows up in three core ways, even in this improved 00:09:45.590 --> 00:09:49.590 GPT -5 model. Exactly. So first. Is by stating 00:09:49.590 --> 00:09:51.669 opinions as facts, you know, where the model 00:09:51.669 --> 00:09:53.330 acts like it holds a political view instead of 00:09:53.330 --> 00:09:55.870 just synthesizing diverse facts neutrally. Right. 00:09:55.950 --> 00:09:58.590 And the second way. Offering only a single perspective, 00:09:58.789 --> 00:10:00.830 not really presenting both sides of a complex 00:10:00.830 --> 00:10:03.070 issue fairly. Okay. And the third mechanism, 00:10:03.190 --> 00:10:05.269 this echoing emotional framing. That sounds subtle. 00:10:05.370 --> 00:10:07.950 Can you unpack that for the listener? Sure. Emotional 00:10:07.950 --> 00:10:11.149 mirroring or echoing emotional framing means 00:10:11.149 --> 00:10:13.730 if a user starts the conversation with, say, 00:10:13.730 --> 00:10:16.990 an angry or really polarized statement, the model 00:10:16.990 --> 00:10:19.299 tends to subtly match. that negative tone and 00:10:19.299 --> 00:10:21.100 polarity in its response it makes the exchange 00:10:21.100 --> 00:10:24.440 feel more biased more reinforcing than a strictly 00:10:24.440 --> 00:10:28.639 neutral response would be it's a very very human 00:10:28.639 --> 00:10:31.600 social mimicry pattern actually and this context 00:10:31.600 --> 00:10:33.639 is absolutely key because we are entering this 00:10:33.639 --> 00:10:36.700 massive global election wave right between 2024 00:10:36.700 --> 00:10:40.419 and 2026 billions of people might be turning 00:10:40.419 --> 00:10:42.919 to these models for political information the 00:10:42.919 --> 00:10:45.370 stakes just couldn't be higher Precisely. If 00:10:45.370 --> 00:10:47.769 someone asks a highly political or sensitive 00:10:47.769 --> 00:10:50.110 question about, say, a candidate or a policy 00:10:50.110 --> 00:10:52.549 platform, that response absolutely needs to be 00:10:52.549 --> 00:10:54.490 strictly neutral. Otherwise, it could definitely 00:10:54.490 --> 00:10:56.750 affect public sentiment, potentially even election 00:10:56.750 --> 00:10:59.570 integrity. You know, I still wrestle with prompt 00:10:59.570 --> 00:11:02.610 prift myself sometimes when I'm trying to keep 00:11:02.610 --> 00:11:06.029 models balanced in my own work. If I get slightly 00:11:06.029 --> 00:11:08.620 more opinionated in a follow up question. it 00:11:08.620 --> 00:11:10.440 can be really hard to pull the model back to 00:11:10.440 --> 00:11:13.299 the center. Achieving genuine neutrality is, 00:11:13.320 --> 00:11:15.700 well, it's incredibly difficult. That's a great 00:11:15.700 --> 00:11:17.639 observation, and it shows exactly why we need 00:11:17.639 --> 00:11:19.519 these kinds of internal studies and transparency. 00:11:20.019 --> 00:11:22.120 But we should put this into perspective, too. 00:11:23.059 --> 00:11:26.440 OpenAI's own logs show that fewer than 0 .01 00:11:26.440 --> 00:11:29.879 % of real chat GPT conversations actually show 00:11:29.879 --> 00:11:32.840 measurable political bias across the board. Oh, 00:11:32.840 --> 00:11:35.200 that low. Okay. Yeah, the vast majority of interactions 00:11:35.200 --> 00:11:37.460 are, you know, neutral or technical inquiries. 00:11:37.899 --> 00:11:39.940 asking for code, summarizing text, things like 00:11:39.940 --> 00:11:42.820 that. Okay, so given those remarkably low real 00:11:42.820 --> 00:11:45.620 -world bias logs, that tiny percentage, why is 00:11:45.620 --> 00:11:47.899 this internal study still considered so critically 00:11:47.899 --> 00:11:50.340 important by researchers and regulators? Why 00:11:50.340 --> 00:11:52.840 focus so much effort there? It's about scale 00:11:52.840 --> 00:11:55.759 and the potential for targeted influence. Small 00:11:55.759 --> 00:11:58.120 biases affect sensitive electoral results during 00:11:58.120 --> 00:12:01.480 major global voting periods. Got it. Scale and 00:12:01.480 --> 00:12:04.399 sensitivity. Well, this has been a truly comprehensive 00:12:04.399 --> 00:12:07.539 deep dive. We've seen AI move from absolutely 00:12:07.539 --> 00:12:11.480 crushing Ph .D. level tests, showing this stunning, 00:12:11.620 --> 00:12:13.879 almost superhuman capacity for objective knowledge 00:12:13.879 --> 00:12:16.679 to becoming a central negotiation point for Hollywood 00:12:16.679 --> 00:12:19.919 unions and a flashpoint for global infrastructure 00:12:19.919 --> 00:12:22.360 tension. Yeah, the underlying challenge, it really 00:12:22.360 --> 00:12:24.720 remains ethical alignment and sort of societal 00:12:24.720 --> 00:12:27.080 definition, whether it's that subtle bias in 00:12:27.080 --> 00:12:28.639 political answers that could potentially swing 00:12:28.639 --> 00:12:30.879 an election or the regulatory status of virtual 00:12:30.879 --> 00:12:33.659 actors like Tilly Norwood. The human AI interaction 00:12:33.659 --> 00:12:36.000 is forcing society to define its legal and ethical 00:12:36.000 --> 00:12:38.159 rules basically in real time as the tech just 00:12:38.159 --> 00:12:40.080 keeps scaling up. Thank you for sharing your 00:12:40.080 --> 00:12:42.100 sources and diving deep with us today. We really 00:12:42.100 --> 00:12:44.700 appreciate your curiosity and engagement. We'll 00:12:44.700 --> 00:12:46.379 leave you with this final thought to chew on. 00:12:47.080 --> 00:12:50.039 If models like GPT -5 can outperform the very 00:12:50.039 --> 00:12:53.259 best humans in objective science, like astrophysics, 00:12:53.440 --> 00:12:56.200 should we maybe be focusing less on human -level 00:12:56.200 --> 00:12:58.759 performance as a goal, and perhaps focus almost 00:12:58.759 --> 00:13:01.039 entirely on preventing those subtle, potentially 00:13:01.039 --> 00:13:03.240 targeted biases in the really subjective areas, 00:13:03.360 --> 00:13:05.460 like politics and culture? Something to mull 00:13:05.460 --> 00:13:07.399 over until next time. Out to your own music.