WEBVTT 00:00:00.000 --> 00:00:01.960 You know, if you've spent any real time working 00:00:01.960 --> 00:00:04.259 with these large language models, you know that 00:00:04.259 --> 00:00:06.780 feeling, that specific frustration. You feed 00:00:06.780 --> 00:00:09.560 it this huge document, maybe, I don't know, 500 00:00:09.560 --> 00:00:12.500 pages of tech specs, and you ask this complex 00:00:12.500 --> 00:00:15.970 question about one detail. Buried way deep inside. 00:00:16.269 --> 00:00:18.449 And the model just comes back with, sorry, I 00:00:18.449 --> 00:00:20.809 don't recall that. It's that classic context 00:00:20.809 --> 00:00:23.969 problem, isn't it? The AI just gets overloaded 00:00:23.969 --> 00:00:26.350 and its accuracy kind of drifts off. Exactly. 00:00:26.350 --> 00:00:29.730 Just loses the plot. But what if the AI wasn't 00:00:29.730 --> 00:00:32.409 just passively reading all those tokens? What 00:00:32.409 --> 00:00:35.609 if it could actively debug the document? Imagine 00:00:35.609 --> 00:00:38.289 a system that sort of peeks at different parts, 00:00:38.509 --> 00:00:41.689 asks itself questions, follow -ups, and jumps 00:00:41.689 --> 00:00:43.719 through the data like a... like a really thoughtful 00:00:43.719 --> 00:00:46.259 engineer would. That is exactly the kind of shift 00:00:46.259 --> 00:00:48.600 we're seeing in these sources we looked at. We're 00:00:48.600 --> 00:00:52.679 really diving into a new generation of AI intelligence 00:00:52.679 --> 00:00:54.780 here. It's moving beyond just predicting the 00:00:54.780 --> 00:00:57.219 next word. It's getting into genuine strategic 00:00:57.219 --> 00:01:00.570 reasoning. Welcome to the Deep Dive. Yeah, we've 00:01:00.570 --> 00:01:02.909 got a really fascinating stack of research, some 00:01:02.909 --> 00:01:05.269 current events here, too. Our mission today, 00:01:05.430 --> 00:01:07.390 it's pretty straightforward. We're going to explore 00:01:07.390 --> 00:01:10.650 how AI is getting better at thinking strategically, 00:01:10.810 --> 00:01:13.730 both internally within these massive data sets 00:01:13.730 --> 00:01:15.709 and externally when it's navigating the whole 00:01:15.709 --> 00:01:18.250 web. Yeah, and we've got a pretty clear red map 00:01:18.250 --> 00:01:20.030 laid out. First up, we're going to unpack these 00:01:20.030 --> 00:01:22.469 things called recursive language models, RLMs. 00:01:22.549 --> 00:01:25.310 MIT is using them to basically solve that long 00:01:25.310 --> 00:01:28.010 context blindness. Pretty cool stuff. Right. 00:01:28.519 --> 00:01:31.560 Then we'll hit the current landscape. Quick updates, 00:01:31.840 --> 00:01:33.920 entropics, got some new skills. There's this 00:01:33.920 --> 00:01:37.680 surprising flop of an AI pet. Oh, yeah, the Gen 00:01:37.680 --> 00:01:40.200 Z pet, right? Yeah, that one. And also how these 00:01:40.200 --> 00:01:42.620 conflicts with governments are starting to shape 00:01:42.620 --> 00:01:44.400 the market. And finally, we're going to dig into 00:01:44.400 --> 00:01:48.439 Apple's new search model, DeepMM Search R1. This 00:01:48.439 --> 00:01:50.980 thing is, well, it's basically a self -correcting 00:01:50.980 --> 00:01:53.099 system that learns to research almost like a 00:01:53.099 --> 00:01:55.079 human debugger. So let's jump into the science 00:01:55.079 --> 00:02:00.030 first. Let's do it. So first segment. MIT's breakthrough 00:02:00.030 --> 00:02:04.129 with recursive language models, RLMs, these seem, 00:02:04.269 --> 00:02:06.760 well, Pretty explicitly designed to kill that 00:02:06.760 --> 00:02:09.120 long context problem we mentioned. They absolutely 00:02:09.120 --> 00:02:10.979 are. And what's really fascinating is how they 00:02:10.979 --> 00:02:13.539 do it. So an RLM, think of it like a system where 00:02:13.539 --> 00:02:16.539 the AI takes a giant task, breaks it down into 00:02:16.539 --> 00:02:19.360 smaller, more manageable pieces, and then it 00:02:19.360 --> 00:02:21.979 queries itself to find the answers to those smaller 00:02:21.979 --> 00:02:24.400 pieces. It's basically a model that asks itself 00:02:24.400 --> 00:02:26.280 questions. Okay, that immediately sounds way 00:02:26.280 --> 00:02:28.900 more strategic than just forcing it to read everything 00:02:28.900 --> 00:02:31.099 in one go. But let me push back just a little 00:02:31.099 --> 00:02:33.800 bit. Is this just a fancy new label for a really 00:02:33.800 --> 00:02:36.069 good agent system? Ah, that's a fair question. 00:02:36.430 --> 00:02:38.590 But the distinction, I think, is pretty crucial. 00:02:38.750 --> 00:02:41.009 Instead of trying to cram hundreds of thousands 00:02:41.009 --> 00:02:43.949 of tokens into a single prompt, which, as we 00:02:43.949 --> 00:02:47.569 said, leads to context rot, the RLM kind of adopts 00:02:47.569 --> 00:02:49.710 a developer's mindset. Okay, tell me more about 00:02:49.710 --> 00:02:52.090 that analogy, a developer's mindset. Yeah, imagine 00:02:52.090 --> 00:02:54.789 watching a programmer debugging a huge pile of 00:02:54.789 --> 00:02:56.729 code or data. They don't read every single line 00:02:56.729 --> 00:02:58.849 right. They jump around. So the mechanism is 00:02:58.849 --> 00:03:01.550 kind of elegant. The RLM peeks at chunks of the 00:03:01.550 --> 00:03:03.889 context. Then it does this sort of internal grep. 00:03:04.199 --> 00:03:06.699 searching for specific patterns or keywords, 00:03:06.919 --> 00:03:10.659 maybe like a user ID, user67144 or something, 00:03:10.840 --> 00:03:13.639 it splits the data based on what it finds. And 00:03:13.639 --> 00:03:15.800 then it recursively calls these subqueries to 00:03:15.800 --> 00:03:17.939 focus only on the little segment that's relevant 00:03:17.939 --> 00:03:20.819 for a final answer. Ah, okay. So it's dynamically 00:03:20.819 --> 00:03:23.500 building this chain of thought that's optimized 00:03:23.500 --> 00:03:25.740 for the data structure itself, not just following 00:03:25.740 --> 00:03:28.259 some pre -big instruction list. Precisely. And 00:03:28.259 --> 00:03:30.280 that's why the performance gains are just wow. 00:03:30.479 --> 00:03:33.300 The sources pointed out that an RLM built on... 00:03:33.479 --> 00:03:36.159 this, GPT -5 Mini actually beat the standard 00:03:36.159 --> 00:03:40.060 full -size GPT -5, beat it by 114 % in accuracy 00:03:40.060 --> 00:03:43.120 on complex tasks. That's a huge lift, especially 00:03:43.120 --> 00:03:46.819 for a smaller model. 114%. That's a staggering 00:03:46.819 --> 00:03:49.280 number. And importantly, the sources also note 00:03:49.280 --> 00:03:52.199 it kept that accuracy even when the context ballooned 00:03:52.199 --> 00:03:55.460 to like a thousand documents. That's real world 00:03:55.460 --> 00:03:58.680 robustness. Exactly. That robustness is what 00:03:58.680 --> 00:04:00.939 matters if this stuff is actually going to get 00:04:00.939 --> 00:04:04.039 used widely. And going back to your question 00:04:04.039 --> 00:04:07.099 about agents, RLMs decide how to think. They 00:04:07.099 --> 00:04:08.960 figure out the strategy internally. You know, 00:04:08.960 --> 00:04:10.759 traditional agents, they just follow the fixed 00:04:10.759 --> 00:04:13.180 rules you give them up front. This is different. 00:04:13.460 --> 00:04:15.460 OK, so that difference, the internal strategic 00:04:15.460 --> 00:04:17.980 decision making, that feels like it fundamentally 00:04:17.980 --> 00:04:20.420 changes the potential for long form reasoning. 00:04:20.579 --> 00:04:23.019 It absolutely does. It lets the AI strategically 00:04:23.019 --> 00:04:25.560 manage all that information, basically avoiding 00:04:25.560 --> 00:04:28.230 getting overwhelmed. all right so moving from 00:04:28.230 --> 00:04:31.329 those like lab breakthroughs to what's happening 00:04:31.329 --> 00:04:33.550 out in the market right now we've got some interesting 00:04:33.550 --> 00:04:35.490 quick hits on the tools people are actually using 00:04:35.490 --> 00:04:38.170 day to day yeah it's fascinating how quickly 00:04:38.170 --> 00:04:41.290 these base models are adding specialized skills 00:04:41.290 --> 00:04:44.500 that you know, actually save us real time. Definitely. 00:04:44.939 --> 00:04:47.680 Take Google Notebook LM. It can now handle Arxiv 00:04:47.680 --> 00:04:49.420 papers. So it's kind of like having your own 00:04:49.420 --> 00:04:51.600 personal research professor for academic stuff. 00:04:51.920 --> 00:04:54.040 And Anthropic's back in the mix, letting users 00:04:54.040 --> 00:04:56.339 give Claude specific automation skills. That 00:04:56.339 --> 00:04:58.160 really boosts its usefulness for businesses, 00:04:58.220 --> 00:05:00.199 right? And there was a small tool update, but 00:05:00.199 --> 00:05:03.879 honestly, one I really needed. Chat GPT can now 00:05:03.879 --> 00:05:06.839 automatically manage your saved memories. You 00:05:06.839 --> 00:05:08.660 know, I still wrestle with prompt drift myself 00:05:08.660 --> 00:05:11.500 sometimes. So auto memory management sounds, 00:05:11.680 --> 00:05:14.519 frankly, pretty crucial. Yeah. You can go into 00:05:14.519 --> 00:05:16.660 settings now and prioritize which memories are 00:05:16.660 --> 00:05:18.620 more important. That's a big usability win for 00:05:18.620 --> 00:05:21.699 sure. But shifting gears a bit, let's talk about 00:05:21.699 --> 00:05:25.569 that Gen Z AI pet that just... Kind of flopped. 00:05:25.569 --> 00:05:28.009 Oh, right. The stress relief companion. I remember 00:05:28.009 --> 00:05:29.930 the launch hype. It was all about being this 00:05:29.930 --> 00:05:33.009 nonjudgmental friend. Total flop. And the source 00:05:33.009 --> 00:05:34.750 material mentioned that psychologists basically 00:05:34.750 --> 00:05:36.750 called it. They predicted users would just find 00:05:36.750 --> 00:05:39.410 the interaction awkward, not genuinely soothing 00:05:39.410 --> 00:05:42.550 or comforting. Trying to engineer an emotional 00:05:42.550 --> 00:05:45.589 connection with an algorithm. It just felt hollow 00:05:45.589 --> 00:05:48.629 to people. It seems like what we're really looking 00:05:48.629 --> 00:05:52.079 for from AI is genuine utility. And maybe if 00:05:52.079 --> 00:05:54.399 product tries to lean too hard into that emotional 00:05:54.399 --> 00:05:56.800 side, people just sense the, I don't know, the 00:05:56.800 --> 00:05:58.660 artifice really quickly. Interesting cultural 00:05:58.660 --> 00:06:01.629 read. It really is. But the mood shifts quick 00:06:01.629 --> 00:06:05.470 back to geopolitics. It's not just AI competing 00:06:05.470 --> 00:06:08.209 with AI anymore in terms of capability. The sources 00:06:08.209 --> 00:06:10.490 are really highlighting this AI versus governments 00:06:10.490 --> 00:06:12.589 dynamic now. Yeah, that conflict seems to be 00:06:12.589 --> 00:06:15.170 heating up fast. You had the White House clashing 00:06:15.170 --> 00:06:17.670 directly with Anthropic over AI regulation proposals. 00:06:17.990 --> 00:06:20.889 And some folks in D .C. apparently labeled the 00:06:20.889 --> 00:06:23.209 company's concerns as just fear mongering. Right. 00:06:23.310 --> 00:06:24.990 And this ties straight into the business angle, 00:06:25.050 --> 00:06:27.129 too. Because of these kinds of conflicts and 00:06:27.129 --> 00:06:29.769 the whole push for regulation, AI startups are 00:06:29.769 --> 00:06:32.230 really doubling down on controlling data quality. 00:06:32.410 --> 00:06:35.129 They're starting to see high quality vetted data 00:06:35.129 --> 00:06:38.709 as like the new AI goldmine. It's all about the 00:06:38.709 --> 00:06:41.250 data now. So why do you think this conflict between 00:06:41.250 --> 00:06:43.689 AI companies and government regulation is really 00:06:43.689 --> 00:06:46.589 intensifying right now? Well, the sheer power 00:06:46.589 --> 00:06:49.430 of this new AI to influence society, it just. 00:06:49.870 --> 00:06:52.990 demands immediate and really careful governmental 00:06:52.990 --> 00:06:56.269 oversight. It's becoming unavoidable. Mid -roll 00:06:56.269 --> 00:06:59.529 sponsor, Reed Placeholder. Okay, let's pivot 00:06:59.529 --> 00:07:02.649 to some... practical, actionable strategies we 00:07:02.649 --> 00:07:04.670 pulled from the sources. We often get caught 00:07:04.670 --> 00:07:06.850 up talking about the huge compute needed for 00:07:06.850 --> 00:07:09.269 these giant models. Yeah. But sometimes just 00:07:09.269 --> 00:07:11.370 a really smart prompt can be the secret weapon. 00:07:11.529 --> 00:07:13.470 Oh, that's a massive understatement based on 00:07:13.470 --> 00:07:15.629 this data we saw. There was one novel prompting 00:07:15.629 --> 00:07:19.410 strategy that gave an LLM a crazy 200 % performance, 00:07:19.509 --> 00:07:22.529 left 200 % just in contextual faithfulness, how 00:07:22.529 --> 00:07:24.790 well it stuck to the facts, all from structuring 00:07:24.790 --> 00:07:26.970 the question better. 200 % is just incredible, 00:07:27.129 --> 00:07:29.329 especially because, as the source noted, it actually 00:07:29.329 --> 00:07:32.100 beat out... complex methods like supervised fine 00:07:32.100 --> 00:07:34.620 tuning, SFT, and direct preference optimization, 00:07:35.079 --> 00:07:37.720 DPO. Those usually take a ton of resources and 00:07:37.720 --> 00:07:40.259 complex model tuning. Yeah, it really reinforces 00:07:40.259 --> 00:07:43.800 that idea that how you use the tool matters just 00:07:43.800 --> 00:07:46.819 as much, maybe more sometimes, than the raw power 00:07:46.819 --> 00:07:50.079 of the underlying model. But strategy isn't only 00:07:50.079 --> 00:07:52.319 about prompts, right? This brings us to what 00:07:52.319 --> 00:07:54.959 one source called the silent killer of AR projects. 00:07:55.240 --> 00:07:58.500 The silent killer being. Miss using these incredibly 00:07:58.500 --> 00:08:01.759 powerful AI tools to just accelerate a bad idea. 00:08:01.939 --> 00:08:04.500 You might have the best AI, but if you haven't 00:08:04.500 --> 00:08:06.819 correctly defined the actual business problem 00:08:06.819 --> 00:08:09.000 you're trying to solve first, you're essentially 00:08:09.000 --> 00:08:12.300 just. failing faster maybe more expensively design 00:08:12.300 --> 00:08:14.759 thinking that's presented as the real edge here 00:08:14.759 --> 00:08:17.040 define the problem right right and the sources 00:08:17.040 --> 00:08:19.500 really emphasize that most ai initiatives don't 00:08:19.500 --> 00:08:21.540 collapse because the tech fails it's usually 00:08:21.540 --> 00:08:24.519 human factors organizational friction there was 00:08:24.519 --> 00:08:26.639 mention of needing a framework to handle those 00:08:26.639 --> 00:08:29.699 quote seven workplace personalities during an 00:08:29.699 --> 00:08:31.959 ai shift you know the skeptics the over enthusiasts 00:08:31.959 --> 00:08:34.940 the teams working in silos exactly it's the human 00:08:34.940 --> 00:08:37.000 resistance the poor process definition that actually 00:08:37.000 --> 00:08:39.120 stops the tech from delivering value we're hitting 00:08:39.240 --> 00:08:41.919 people problems, not really coding problems anymore. 00:08:42.200 --> 00:08:44.440 We also saw a quick list of some new tools that 00:08:44.440 --> 00:08:46.919 kind of fit this practical problem solving theme. 00:08:47.100 --> 00:08:50.320 Things like Alphas Fiv converts research papers 00:08:50.320 --> 00:08:53.220 into conversations. Yeah. And Reducto takes documents 00:08:53.220 --> 00:08:56.059 and spits out clean, structured data. Emergent, 00:08:56.059 --> 00:08:58.139 which turns text descriptions into actual working 00:08:58.139 --> 00:09:01.720 apps. And Supercut for auto editing long videos 00:09:01.720 --> 00:09:05.240 into short clips. The focus is clearly on productivity 00:09:05.240 --> 00:09:09.309 gains, making things easier. So, OK, beyond that 00:09:09.309 --> 00:09:12.669 huge prompts lift number, what's the core lesson 00:09:12.669 --> 00:09:15.029 here about using AI effectively? It's got to 00:09:15.029 --> 00:09:16.950 be. Define the right business problem first, 00:09:17.029 --> 00:09:19.950 always, before you throw these powerful AI tools 00:09:19.950 --> 00:09:22.580 at it. Okay. Our final big discovery takes us 00:09:22.580 --> 00:09:25.299 back to strategic AI, but this time focused on 00:09:25.299 --> 00:09:28.000 how AI interacts with the outside world. Apple's 00:09:28.000 --> 00:09:30.679 deep MM search R1. And this is definitely not 00:09:30.679 --> 00:09:32.600 just another simple retrieval model. Right. This 00:09:32.600 --> 00:09:35.019 sounds like a genuinely multimodal LLM. It doesn't 00:09:35.019 --> 00:09:38.039 just search the web. It actually self -corrects 00:09:38.039 --> 00:09:40.080 its own approach in real time if the first results 00:09:40.080 --> 00:09:42.629 aren't good enough. That's exactly it. And its 00:09:42.629 --> 00:09:44.529 abilities are just fascinating because they show 00:09:44.529 --> 00:09:46.909 the kind of strategic thinking we usually associate 00:09:46.909 --> 00:09:50.190 with like highly skilled human researchers. It 00:09:50.190 --> 00:09:52.730 decides when it needs to search and crucially 00:09:52.730 --> 00:09:55.590 what it should search for. It issues actual strategic 00:09:55.590 --> 00:09:58.590 queries to the web. And it handles images smartly 00:09:58.590 --> 00:10:00.970 too, right? If you give it a picture, it apparently 00:10:00.970 --> 00:10:03.389 automatically crops it to zoom in on the important 00:10:03.389 --> 00:10:05.889 part before it searches. So it's prioritizing 00:10:05.889 --> 00:10:08.629 the visual context, not just throwing raw pixels 00:10:08.629 --> 00:10:10.830 at the problem. Yeah, but the self -correction 00:10:10.830 --> 00:10:13.070 loop, that's the real kicker here. It actually 00:10:13.070 --> 00:10:15.169 reflects on the answers it generates. If the 00:10:15.169 --> 00:10:17.789 first batch of web results look kind of weak 00:10:17.789 --> 00:10:20.990 or contradictory, it automatically rewrites its 00:10:20.990 --> 00:10:23.570 own query and searches again. It keeps iterating 00:10:23.570 --> 00:10:26.629 until it finds reliable sources. Wow. That level 00:10:26.629 --> 00:10:30.669 of strategic reflection and... iteration that 00:10:30.669 --> 00:10:33.110 seems incredibly powerful. So how did it actually 00:10:33.110 --> 00:10:35.429 perform compared to the methods we use now, like 00:10:35.429 --> 00:10:38.110 RRAG? Well, it significantly outperformed all 00:10:38.110 --> 00:10:40.070 the open source search baselines they tested 00:10:40.070 --> 00:10:42.850 against. Squared something like 21 points higher 00:10:42.850 --> 00:10:45.450 than common RRAG workflows. Okay, let's pause 00:10:45.450 --> 00:10:47.429 on RRAG for just a second. For anyone listening, 00:10:47.929 --> 00:10:51.370 RRAG is retrieval augmented generation. Basically, 00:10:51.409 --> 00:10:54.230 the AI fetches external documents to add to its 00:10:54.230 --> 00:10:56.309 knowledge before answering. But you're saying 00:10:56.309 --> 00:10:59.399 RRAG workflows often add noise. They often do, 00:10:59.519 --> 00:11:01.940 yeah. RA can sometimes pull in documents that 00:11:01.940 --> 00:11:03.879 aren't truly relevant, maybe just because they 00:11:03.879 --> 00:11:07.080 share some keywords. DeepMM Search R1 seems to 00:11:07.080 --> 00:11:09.440 avoid that by being much more targeted and strategic 00:11:09.440 --> 00:11:12.320 in its search. And the sources also noted it 00:11:12.320 --> 00:11:15.320 nearly matched the performance of GPT -03, despite 00:11:15.320 --> 00:11:18.059 running on a much smaller backbone model, QEN 00:11:18.059 --> 00:11:21.809 2 .5 VL7B. That points to some serious efficiency 00:11:21.809 --> 00:11:25.529 gains. Whoa. Yeah, just imagine scaling a self 00:11:25.529 --> 00:11:27.730 -correcting system like that up to, say, a billion 00:11:27.730 --> 00:11:30.450 queries a day. The efficiency and accuracy improvements 00:11:30.450 --> 00:11:33.929 for any major search platform would be just massive. 00:11:34.269 --> 00:11:36.090 And here's the really crucial bit, the part that 00:11:36.090 --> 00:11:38.289 feels like a paradigm shift. Unlike those big 00:11:38.289 --> 00:11:40.789 retrieval models that need these enormous, constantly 00:11:40.789 --> 00:11:43.370 updated indexes, the system apparently doesn't 00:11:43.370 --> 00:11:45.769 need a huge internal data library. It just learns 00:11:45.769 --> 00:11:48.019 how to use the public web intelligently. like 00:11:48.019 --> 00:11:49.679 a really focused researcher who knows how to 00:11:49.679 --> 00:11:51.919 find things. So if this kind of model doesn't 00:11:51.919 --> 00:11:55.179 need that massive index, how does that change 00:11:55.179 --> 00:11:58.379 the future of AI search, do you think? Well, 00:11:58.419 --> 00:12:00.299 it seems like it shifts the whole focus away 00:12:00.299 --> 00:12:03.419 from just indexing ever more data towards teaching 00:12:03.419 --> 00:12:06.139 the AI how to navigate the web efficiently and 00:12:06.139 --> 00:12:08.539 strategically. It's about skill, not just storage. 00:12:08.700 --> 00:12:11.789 Hashtag tag tag outro. So we started this deep 00:12:11.789 --> 00:12:14.330 dive talking about AI's context blindness, that 00:12:14.330 --> 00:12:16.389 frustration of it forgetting things and long 00:12:16.389 --> 00:12:18.750 documents. And I think our sources today have 00:12:18.750 --> 00:12:22.149 really shown. pretty decisively that ai is rapidly 00:12:22.149 --> 00:12:24.190 becoming much more strategically smart it's not 00:12:24.190 --> 00:12:26.049 just about getting bigger anymore absolutely 00:12:26.049 --> 00:12:29.450 we saw kind of two parallel trends tackling that 00:12:29.450 --> 00:12:31.470 original friction point first you've got better 00:12:31.470 --> 00:12:34.090 internal reasoning that's the rlms handling massive 00:12:34.090 --> 00:12:36.269 context by basically debugging themselves right 00:12:36.269 --> 00:12:38.990 and second better external interaction that's 00:12:38.990 --> 00:12:41.629 the deep mm search model navigating the whole 00:12:41.629 --> 00:12:43.669 web like a strategic debugger correcting its 00:12:43.669 --> 00:12:47.080 own mistakes as it goes yeah so Our combined 00:12:47.080 --> 00:12:49.399 takeaway here feels like the main challenges, 00:12:49.460 --> 00:12:51.000 the friction points, they're actually shifting. 00:12:51.100 --> 00:12:53.600 They seem to be moving away from purely technical 00:12:53.600 --> 00:12:56.259 limits like context, window size or model parameter 00:12:56.259 --> 00:12:59.259 count and moving more towards human limits. Things 00:12:59.259 --> 00:13:02.019 like poor problem definition or just organizational 00:13:02.019 --> 00:13:06.299 resistance to change. So, OK, here's maybe a 00:13:06.299 --> 00:13:08.480 final provocative thought for you, the listener, 00:13:08.559 --> 00:13:12.480 to consider. The clear emerging trend is AI that 00:13:12.480 --> 00:13:14.960 reasons more programmatically, more strategically. 00:13:15.610 --> 00:13:18.330 So if these LLMs can start self -correcting their 00:13:18.330 --> 00:13:20.289 own web queries, if they can debug their own 00:13:20.289 --> 00:13:22.690 context understanding, does that mean the next 00:13:22.690 --> 00:13:24.730 generation of large models might be entirely 00:13:24.730 --> 00:13:27.149 self -auditing? Makes you wonder, you know, how 00:13:27.149 --> 00:13:29.210 quickly human oversight might shift completely 00:13:29.210 --> 00:13:31.470 towards just high -level strategy and problem 00:13:31.470 --> 00:13:33.750 definition rather than getting bogged down in 00:13:33.750 --> 00:13:35.870 the execution details. Something to think about. 00:13:36.139 --> 00:13:37.700 Thank you for sharing your sources with us for 00:13:37.700 --> 00:13:39.440 this deep dive. We definitely encourage you to 00:13:39.440 --> 00:13:41.679 check out the links provided, especially on design 00:13:41.679 --> 00:13:44.240 thinking and navigating that human friction in 00:13:44.240 --> 00:13:47.279 AI adoption. Seems increasingly important. Out 00:13:47.279 --> 00:13:47.779 to you, music.