WEBVTT 00:00:00.000 --> 00:00:03.060 We often hear about AI's incredible speed, you 00:00:03.060 --> 00:00:04.839 know, how it promises to accelerate everything 00:00:04.839 --> 00:00:07.839 we do. But what if, maybe in some specific cases, 00:00:08.099 --> 00:00:11.480 it actually slowed things down? Today, we're 00:00:11.480 --> 00:00:13.480 going to explore some surprising turns in the 00:00:13.480 --> 00:00:15.900 AI world and maybe challenge a few assumptions 00:00:15.900 --> 00:00:19.170 along the way. Welcome to the Deep Dive. This 00:00:19.170 --> 00:00:20.949 is where we take a stack of the latest articles, 00:00:21.289 --> 00:00:24.410 research papers, our own notes, and we try to 00:00:24.410 --> 00:00:26.329 pull out the most important nuggets of knowledge 00:00:26.329 --> 00:00:28.809 for you. Think of it as a shortcut, maybe, to 00:00:28.809 --> 00:00:31.510 being truly well -informed without all the information 00:00:31.510 --> 00:00:34.689 overload. Yeah, and today we've got a really 00:00:34.689 --> 00:00:36.590 fascinating journey lined up. We're going to 00:00:36.590 --> 00:00:39.570 explore the current shifts impacting the whole 00:00:39.570 --> 00:00:42.329 AI industry. We'll look at some genuinely unexpected 00:00:42.329 --> 00:00:44.829 applications, and then we'll dive deep into some 00:00:44.829 --> 00:00:47.539 new research on how AI tools are. or perhaps 00:00:47.539 --> 00:00:50.020 aren't really impacted our productivity. Okay, 00:00:50.060 --> 00:00:51.780 so first up, let's talk about OpenAI. I mean, 00:00:51.799 --> 00:00:53.920 they're pretty much synonymous with AI breakthroughs 00:00:53.920 --> 00:00:55.740 these days. They've been riding incredibly high. 00:00:55.899 --> 00:00:58.159 Yeah. A staggering valuation. What is it? $300 00:00:58.159 --> 00:01:04.379 billion. And 500 million weekly users for ChatGPT. 00:01:04.480 --> 00:01:06.920 Well, they're the most hyped AI company on earth, 00:01:07.040 --> 00:01:09.019 basically. That's absolutely right. But, you 00:01:09.019 --> 00:01:11.319 know, what looked like pure dominance just a 00:01:11.319 --> 00:01:15.120 few months back, say March of this year, it's... 00:01:15.519 --> 00:01:18.780 rapidly turning into, well, kind of a messy battle. 00:01:18.920 --> 00:01:22.819 You've got the tech giants, Google, Meta, Amazon, 00:01:23.180 --> 00:01:25.459 even Microsoft, who's their biggest backer, right? 00:01:25.900 --> 00:01:28.599 They're all circling like sharks, applying pressure 00:01:28.599 --> 00:01:30.859 from pretty much every angle. We've seen Meta, 00:01:30.859 --> 00:01:33.439 for instance, go kind of full NBA free agency 00:01:33.439 --> 00:01:36.790 mode. They poached three top open AI researchers. 00:01:37.109 --> 00:01:39.450 Wow. And it doesn't stop there, does it? That 00:01:39.450 --> 00:01:41.689 windsurf deal, the acquisition that completely 00:01:41.689 --> 00:01:44.150 collapsed. And Google apparently picked up the 00:01:44.150 --> 00:01:46.329 talent instead in this. What are they calling 00:01:46.329 --> 00:01:48.569 it? A reverse acqui hire. Yeah, exactly. Plus, 00:01:48.609 --> 00:01:50.930 there's growing tension reportedly between open 00:01:50.930 --> 00:01:53.189 AI and Microsoft. Something about a hundred billion 00:01:53.189 --> 00:01:55.989 dollar AGI feud. And AGI, just quickly, that's 00:01:55.989 --> 00:01:57.629 artificial general intelligence. It means AI 00:01:57.629 --> 00:01:59.890 aiming for like human level thinking. Right. 00:01:59.950 --> 00:02:02.129 Human level cognitive abilities. And their open 00:02:02.129 --> 00:02:05.420 weight model launch. Delayed again, which gives 00:02:05.420 --> 00:02:08.620 XAI's Grok 4 a chance to gain some serious momentum. 00:02:08.979 --> 00:02:11.759 Even that Joanie Ive brand collaboration seems 00:02:11.759 --> 00:02:14.340 to be stuck in legal limbo. And then Amazon's 00:02:14.340 --> 00:02:16.240 apparently making a movie portraying Sam Altman 00:02:16.240 --> 00:02:19.560 as a scheming Zuckerberg 2 .0. That's quite a 00:02:19.560 --> 00:02:22.060 pile on. It's quite the saga, isn't it? Yet, 00:02:22.240 --> 00:02:24.120 despite all these headwinds, you have to say 00:02:24.120 --> 00:02:27.120 OpenAI is still kind of unequivocally number 00:02:27.120 --> 00:02:30.620 one in many ways. ChatGPT is still used by half 00:02:30.620 --> 00:02:33.509 a billion people every single week. That number 00:02:33.509 --> 00:02:35.930 just blows my mind. It's huge. They also landed 00:02:35.930 --> 00:02:39.129 a $200 million U .S. defense contract, building 00:02:39.129 --> 00:02:42.449 battlefield -ready AI with Endural. And get this, 00:02:42.569 --> 00:02:45.669 Mattel is launching AI -powered Barbie toys using 00:02:45.669 --> 00:02:49.210 OpenAI models. AI Barbies? Seriously. Plus, there's 00:02:49.210 --> 00:02:51.509 talk of a chat GPT -powered browser coming, which 00:02:51.509 --> 00:02:54.379 could genuinely threaten Google Chrome. So what's 00:02:54.379 --> 00:02:56.139 this mean for Altman? He's kind of caught, isn't 00:02:56.139 --> 00:02:57.699 he, between being the visionary leader and the 00:02:57.699 --> 00:02:59.780 hands -on business operator. He's got this like 00:02:59.780 --> 00:03:02.860 $300 billion rocket ship to steer while dodging 00:03:02.860 --> 00:03:05.439 lawsuits, keeping partners happy. And constantly 00:03:05.439 --> 00:03:07.759 needing to ship better models than competitors 00:03:07.759 --> 00:03:11.840 like Claude or Grok. It's a lot. So the big question 00:03:11.840 --> 00:03:14.419 is, is this just a temporary wobble for open 00:03:14.419 --> 00:03:17.800 AI? Or are we seeing a true shift in the AI landscape? 00:03:18.139 --> 00:03:19.960 It really seems like being the leader of the 00:03:19.960 --> 00:03:22.340 pack always invites some pretty intense competition. 00:03:22.919 --> 00:03:26.479 Okay. So from that high stakes corporate world, 00:03:26.599 --> 00:03:28.599 let's zoom out a bit. Let's look at some of the 00:03:28.599 --> 00:03:31.360 other fascinating, sometimes quirky, sometimes 00:03:31.360 --> 00:03:33.659 maybe troubling developments happening across 00:03:33.659 --> 00:03:36.199 the wider AI landscape. Absolutely. Okay. So 00:03:36.199 --> 00:03:39.280 first up, apparently someone asked Grok, Elon 00:03:39.280 --> 00:03:42.240 Musk's AI, to create a physical representation 00:03:42.240 --> 00:03:45.219 of itself. And the image it generated, this luminous 00:03:45.219 --> 00:03:47.699 cosmic sort of thing, went totally viral. Over 00:03:47.699 --> 00:03:50.849 10 million views. Wow. It just shows how AI is 00:03:50.849 --> 00:03:53.830 shaping completely new forms of digital art and 00:03:53.830 --> 00:03:56.389 even self -expression. And on the Google side, 00:03:56.689 --> 00:03:58.689 Gemini subscribers can now use something called 00:03:58.689 --> 00:04:01.590 VO3. It transforms your regular photos into these 00:04:01.590 --> 00:04:04.090 AI -generated eight -second videos. Eight seconds, 00:04:04.129 --> 00:04:05.810 yeah. Complete with dialogue, sound effects, 00:04:05.969 --> 00:04:08.729 pretty sharp 720p resolution, too. It's kind 00:04:08.729 --> 00:04:10.909 of incredible how fast these creative tools are 00:04:10.909 --> 00:04:14.080 evolving. Yeah. It is. Though on a darker note, 00:04:14.240 --> 00:04:17.079 Meta's AI culture was actually described as a 00:04:17.079 --> 00:04:20.639 metastatic cancer. That was in a viral exit memo 00:04:20.639 --> 00:04:22.339 from one of his own researchers. Gives you a 00:04:22.339 --> 00:04:24.639 peek into the kind of cultural pressures inside 00:04:24.639 --> 00:04:28.920 these super fast growing AI companies. Then there's 00:04:28.920 --> 00:04:31.560 this thing we're seeing more of, Snapchat dysmorphia. 00:04:32.180 --> 00:04:35.160 It's this strange kind of worrying phenomenon 00:04:35.160 --> 00:04:38.240 where people aren't aspiring to look like celebrities 00:04:38.240 --> 00:04:40.160 anymore. Instead, they want to look like their 00:04:40.160 --> 00:04:43.250 AI filtered selves. Right. I have to admit, I 00:04:43.250 --> 00:04:45.709 still wrestle sometimes with how these AI filters 00:04:45.709 --> 00:04:48.649 can shape our self -perception. It's a really 00:04:48.649 --> 00:04:51.470 complex area. It absolutely is. And, you know, 00:04:51.470 --> 00:04:53.430 related to maybe company culture and loyalty, 00:04:53.790 --> 00:04:56.430 many of the missionaries, that's Sam Orton's 00:04:56.430 --> 00:04:58.810 term for top AI researchers, they actually turned 00:04:58.810 --> 00:05:02.129 down these massive $100 million mercenary signing 00:05:02.129 --> 00:05:05.490 bonuses from Meta. $100 million. Wow. Choosing 00:05:05.490 --> 00:05:07.269 instead to stay at places like Anthropic and 00:05:07.269 --> 00:05:09.649 DeepMind tells you something about where some 00:05:09.649 --> 00:05:12.189 of the top talent feels they belong, maybe. Yeah, 00:05:12.269 --> 00:05:14.170 that's significant. It's such a fast -moving 00:05:14.170 --> 00:05:16.829 space. And on the fundraising front, the Robinhood 00:05:16.829 --> 00:05:20.310 CEO's AI startup, Harmonic, they just raised 00:05:20.310 --> 00:05:24.750 $100 million at an $875 million valuation. They're 00:05:24.750 --> 00:05:27.350 building an AI called Aristotle, and the goal 00:05:27.350 --> 00:05:29.829 is for it to solve complex math problems better 00:05:29.829 --> 00:05:32.730 than any human. Whoa, hang on. Imagine an AI 00:05:32.730 --> 00:05:34.550 solving math problems better than any human. 00:05:34.629 --> 00:05:37.810 That's a truly profound leap. That's changing 00:05:37.810 --> 00:05:40.220 the game entirely. Right. Absolutely mind -boggling 00:05:40.220 --> 00:05:42.360 when you think about it. And just a few more 00:05:42.360 --> 00:05:45.439 quick hits here. Meta's AI glasses. They now 00:05:45.439 --> 00:05:47.519 offer audio descriptions. You just ask and it 00:05:47.519 --> 00:05:49.259 tells you what it sees. Handy. There are lists 00:05:49.259 --> 00:05:51.899 going around of the 17 must -have AI skills for 00:05:51.899 --> 00:05:54.639 your resume in 2025. Shows how the job market 00:05:54.639 --> 00:05:57.740 is shifting fast. Apparently, XAI and Grok had 00:05:57.740 --> 00:06:00.240 to apologize for some horrific behavior recently. 00:06:00.480 --> 00:06:02.980 Details are a bit murky there. And finally, two 00:06:02.980 --> 00:06:06.480 models. GPT -03 and Grok -4. They've apparently 00:06:06.480 --> 00:06:09.379 quietly proved that something called neuro -symbolic 00:06:09.379 --> 00:06:12.439 AI works. Now, neuro -symbolic AI, in simple 00:06:12.439 --> 00:06:14.819 terms, it combines logical reasoning, like traditional 00:06:14.819 --> 00:06:17.480 AI, with pattern recognition from data, like 00:06:17.480 --> 00:06:19.500 deep learning, kind of the best of both worlds. 00:06:19.639 --> 00:06:21.779 Right, blending logic and learning. Oh, and Meta 00:06:21.779 --> 00:06:24.699 also recently acquired Play AI, a startup that 00:06:24.699 --> 00:06:26.819 specializes in generating really human -like 00:06:26.819 --> 00:06:29.939 AI voices. So with all these new tools popping 00:06:29.939 --> 00:06:32.220 up constantly, how should people actually approach 00:06:32.220 --> 00:06:34.579 building with AI now? Well, the trend seems to 00:06:34.579 --> 00:06:37.620 be moving beyond informal vibe coding. towards 00:06:37.620 --> 00:06:40.360 more professional context engineering. Ah, okay. 00:06:40.439 --> 00:06:41.879 That brings us perfectly to our next segment 00:06:41.879 --> 00:06:44.579 then. Decoding AI development and the tool shaping 00:06:44.579 --> 00:06:47.699 it. So what's been called vibe coding, this sort 00:06:47.699 --> 00:06:50.160 of informal, maybe unsystematic way of putting 00:06:50.160 --> 00:06:52.660 AI code together, that's essentially dead, people 00:06:52.660 --> 00:06:54.860 are saying. It just doesn't scale up. Exactly. 00:06:55.000 --> 00:06:57.180 So what's rising in this place is this idea of 00:06:57.180 --> 00:07:00.000 context engineering. Think of it as the more 00:07:00.000 --> 00:07:02.259 professional framework for modern AI development. 00:07:02.639 --> 00:07:05.509 It's really about... precisely designing the 00:07:05.509 --> 00:07:08.029 inputs and the conditions around the AI model 00:07:08.029 --> 00:07:10.050 to make it perform reliably and predictably. 00:07:10.149 --> 00:07:12.129 So being much more intentional. Right. Intentional 00:07:12.129 --> 00:07:13.970 is a good word. We've also seen a lot of advice 00:07:13.970 --> 00:07:16.490 popping up, like articles titled Four Tips to 00:07:16.490 --> 00:07:19.910 Take Your Vibe App Design from Zero to Pro. They 00:07:19.910 --> 00:07:21.769 cover things like using proper UI components, 00:07:22.149 --> 00:07:24.569 remixing professional designs, finding good inspiration, 00:07:24.870 --> 00:07:27.370 that sort of thing. And you see lists everywhere 00:07:27.370 --> 00:07:30.250 of seven game -changing AI tools that promise 00:07:30.250 --> 00:07:33.149 to save you, you know, 10 plus hours every single 00:07:33.149 --> 00:07:35.610 week. Yeah, the promise is always huge time savings. 00:07:35.810 --> 00:07:38.550 For research, presentations, design work, you 00:07:38.550 --> 00:07:40.670 name it. And some of these newer tools are getting 00:07:40.670 --> 00:07:44.009 really specific and, frankly, quite helpful sounding. 00:07:44.269 --> 00:07:47.689 There's one called MCTPDF, converts PDF files 00:07:47.689 --> 00:07:51.170 into over 20 different formats. LLM SEO trends 00:07:51.170 --> 00:07:54.870 monitors, like 2 ,200 live search trends with 00:07:54.870 --> 00:07:57.709 actual search volume. Yeah. Brandthetics claims 00:07:57.709 --> 00:08:00.269 to turn your videos into viral cinematic short 00:08:00.269 --> 00:08:02.930 form content. Oh, ambitious. KissPix. Russell 00:08:02.930 --> 00:08:05.370 says it transforms ideas into stunning visuals 00:08:05.370 --> 00:08:08.689 effortlessly. And Create My Banner helps generate 00:08:08.689 --> 00:08:11.250 banners for all your social media needs. Lots 00:08:11.250 --> 00:08:13.730 of specific tools. Okay, so these tools promise 00:08:13.730 --> 00:08:16.189 these huge time savings, 10 hours a week, whatever 00:08:16.189 --> 00:08:19.480 it is. But do they always actually deliver on 00:08:19.480 --> 00:08:21.899 that promise? Well, funny you should ask. A recent 00:08:21.899 --> 00:08:24.639 study found some surprising, maybe even counterintuitive 00:08:24.639 --> 00:08:27.240 results on that exact question. That brings us 00:08:27.240 --> 00:08:29.139 to this really fascinating piece of research 00:08:29.139 --> 00:08:32.720 from METR. They're a nonprofit AI research group. 00:08:32.840 --> 00:08:35.240 And they took a deep look at the actual productivity 00:08:35.240 --> 00:08:39.220 of AI coding tools. We all know tools like Cursor 00:08:39.220 --> 00:08:41.840 and GitHub Copilot promise big gains, right? 00:08:42.279 --> 00:08:44.840 Autowriting code, fixing bugs, helping with testing. 00:08:45.120 --> 00:08:46.820 Yeah, that's the pitch. And these tools are... 00:08:46.830 --> 00:08:49.230 are powered by the latest AI models from OpenAI, 00:08:49.549 --> 00:08:53.230 Google DeepMind, Anthropic, XAI. And those underlying 00:08:53.230 --> 00:08:55.409 models have improved dramatically, incredibly 00:08:55.409 --> 00:08:57.970 fast. And that's exactly what makes this METR 00:08:57.970 --> 00:09:00.090 study so interesting. They did a randomized controlled 00:09:00.090 --> 00:09:02.570 trial, really rigorous stuff. They recruited 00:09:02.570 --> 00:09:05.629 16 experienced open source developers, people 00:09:05.629 --> 00:09:07.870 who know their stuff. And they had them complete 00:09:07.870 --> 00:09:13.070 246 real tasks on large, complex code repositories 00:09:13.070 --> 00:09:15.450 that these developers actually contribute to 00:09:15.450 --> 00:09:18.580 regularly. Roughly half the tasks were AI allowed, 00:09:18.820 --> 00:09:21.039 meaning they could use top -tier tools like Cursor 00:09:21.039 --> 00:09:24.080 Pro. The other half, strictly no AI allowed. 00:09:24.259 --> 00:09:26.820 Okay, so here's the really surprising part. The 00:09:26.820 --> 00:09:29.399 developers themselves forecasted that using the 00:09:29.399 --> 00:09:32.220 AI tools would cut their completion time by about 00:09:32.220 --> 00:09:35.120 24%. Makes sense. That's what you'd expect. But 00:09:35.120 --> 00:09:37.240 the study found the exact opposite, allowing 00:09:37.240 --> 00:09:40.389 AI actually increase the completion time. By 00:09:40.389 --> 00:09:43.490 19%. Increase. So they were slower with the AI 00:09:43.490 --> 00:09:46.129 tools. Slower, yeah. Developers are slower when 00:09:46.129 --> 00:09:49.350 using AI tooling, is the direct quote. Wow. Okay. 00:09:49.429 --> 00:09:51.970 That is counterintuitive. Did the study suggest 00:09:51.970 --> 00:09:54.220 why that might be? Well, they point to a few 00:09:54.220 --> 00:09:56.399 potential reasons. First, only about half the 00:09:56.399 --> 00:09:58.600 developers had prior experience using cursors 00:09:58.600 --> 00:10:00.720 specifically, even though they were trained for 00:10:00.720 --> 00:10:02.279 this study. So maybe a learning curve issue. 00:10:02.460 --> 00:10:04.679 Could be. They also found developers spent more 00:10:04.679 --> 00:10:07.919 time prompting the AI and then waiting for the 00:10:07.919 --> 00:10:09.799 responses instead of just diving in and coding 00:10:09.799 --> 00:10:12.299 themselves. Ah, the interaction overhead. Exactly. 00:10:12.379 --> 00:10:15.190 And maybe, crucially... AI tends to struggle 00:10:15.190 --> 00:10:18.529 more in those really large, complex code bases, 00:10:18.690 --> 00:10:21.309 which were precisely the kind used in this test. 00:10:21.409 --> 00:10:24.320 The context window problem, maybe. Now, it's 00:10:24.320 --> 00:10:26.759 really important to add the nuance here. The 00:10:26.759 --> 00:10:29.019 study authors themselves are very careful. They 00:10:29.019 --> 00:10:32.580 don't draw strong, sweeping conclusions. They 00:10:32.580 --> 00:10:35.419 explicitly say they don't believe AI systems 00:10:35.419 --> 00:10:38.059 fail to speed up most software developers in 00:10:38.059 --> 00:10:40.159 general. Okay, that's important context. Yeah, 00:10:40.200 --> 00:10:42.899 and other large -scale studies do show productivity 00:10:42.899 --> 00:10:46.919 gains. Plus, AI progress is just so fast, they 00:10:46.919 --> 00:10:49.179 admit these results could be different in even 00:10:49.179 --> 00:10:51.120 three months. True, the goalposts are always 00:10:51.120 --> 00:10:53.279 moving. They also found that AI coding tools 00:10:53.279 --> 00:10:55.539 have actually improved recently at more complex 00:10:55.539 --> 00:10:58.360 long horizon tasks. So it's not all negative. 00:10:58.580 --> 00:11:01.000 Still, this research definitely adds to the skepticism 00:11:01.000 --> 00:11:03.500 about universal immediate gains from these tools. 00:11:03.639 --> 00:11:05.659 And it lines up with other studies we've seen 00:11:05.659 --> 00:11:08.179 showing that AI coding tools can sometimes introduce 00:11:08.179 --> 00:11:11.539 mistakes or even security vulnerabilities. It's 00:11:11.539 --> 00:11:14.100 just a good reminder, isn't it? Not every shiny 00:11:14.100 --> 00:11:16.440 new tool delivers on all its promises right away. 00:11:16.909 --> 00:11:19.070 especially maybe for experienced users working 00:11:19.070 --> 00:11:21.590 on really tough problems. So thinking about the 00:11:21.590 --> 00:11:24.389 everyday user of AI tools, maybe not just coding, 00:11:24.549 --> 00:11:26.970 what's the biggest takeaway from research like 00:11:26.970 --> 00:11:29.649 this? I think it's don't just assume universal 00:11:29.649 --> 00:11:32.789 gains. Critical evaluation of the tools you use 00:11:32.789 --> 00:11:35.529 for your specific tasks is still absolutely essential. 00:11:36.399 --> 00:11:39.399 So as we wrap up this deep dive, the main themes 00:11:39.399 --> 00:11:42.200 that really seem to stand out are, first, the 00:11:42.200 --> 00:11:45.299 intense, almost no holds barred competition happening 00:11:45.299 --> 00:11:48.820 at the very top of the AI industry. Second, just 00:11:48.820 --> 00:11:51.019 the sheer speed of innovation, these rapid fire 00:11:51.019 --> 00:11:53.279 changes that are constantly altering how we work 00:11:53.279 --> 00:11:56.360 and even how we live. And third, maybe most importantly, 00:11:56.559 --> 00:11:59.159 this critical ongoing need to actually question 00:11:59.159 --> 00:12:02.259 our assumptions about AI's true impact, especially 00:12:02.259 --> 00:12:05.990 on things like productivity. Exactly. fascinating 00:12:05.990 --> 00:12:09.730 here is that you know while ai is evolving at 00:12:09.730 --> 00:12:13.110 this absolute breakneck speed its actual integration 00:12:13.110 --> 00:12:15.350 into the real world is proving to be incredibly 00:12:15.350 --> 00:12:18.909 complex it's full of nuances it really requires 00:12:18.909 --> 00:12:21.009 both that genuine excitement for the possibilities 00:12:21.009 --> 00:12:23.350 which is easy to have yeah but also a really 00:12:23.350 --> 00:12:25.909 healthy dose of critical thinking always asking 00:12:25.909 --> 00:12:28.389 yourself you know is this genuinely an improvement 00:12:28.389 --> 00:12:31.110 for me or am i just kind of changing how i work 00:12:31.110 --> 00:12:33.590 to fit the tool so maybe here's a thought to 00:12:33.590 --> 00:12:35.929 take away Next time you find yourself using an 00:12:35.929 --> 00:12:38.250 AI tool, just pause for a second and ask yourself, 00:12:38.389 --> 00:12:40.990 is this genuinely making my process more efficient 00:12:40.990 --> 00:12:43.809 or am I just adapting my workflow to the tool's 00:12:43.809 --> 00:12:45.470 way of doing things? That's a great question 00:12:45.470 --> 00:12:47.750 to ponder. Thank you for joining us on this deep 00:12:47.750 --> 00:12:49.590 dive today. We really hope you'll continue your 00:12:49.590 --> 00:12:51.389 own exploration of these endlessly fascinating 00:12:51.389 --> 00:12:53.029 topics. Out to row music.