WEBVTT 00:00:00.000 --> 00:00:03.020 Okay, so we really need to pause for a second 00:00:03.020 --> 00:00:08.039 here and just absorb this reality. An AI model. 00:00:08.220 --> 00:00:10.460 Well, it recently worked autonomously for 30 00:00:10.460 --> 00:00:13.460 hours straight. Beat. And we're not talking about 00:00:13.460 --> 00:00:15.960 a small script or, you know, a quick fix. This 00:00:15.960 --> 00:00:18.000 thing was building entire software applications, 00:00:18.239 --> 00:00:20.320 setting up databases, handling the backend stuff, 00:00:20.460 --> 00:00:23.000 and even passed its own security audits. Required 00:00:23.000 --> 00:00:24.600 ones. This isn't really just a tool anymore, 00:00:24.719 --> 00:00:27.320 is it? It feels like the new baseline for... 00:00:28.640 --> 00:00:31.039 AI engineering capabilities. Yeah. Welcome to 00:00:31.039 --> 00:00:33.159 the deep dive. We are pulling straight from the 00:00:33.159 --> 00:00:36.619 latest stack of AI breakthroughs and platform 00:00:36.619 --> 00:00:39.140 updates. And honestly, the speed of change right 00:00:39.140 --> 00:00:41.740 now, it's almost overwhelming. So our mission 00:00:41.740 --> 00:00:43.640 today is basically to filter all that noise. 00:00:43.780 --> 00:00:45.960 We want to deliver like the three most critical 00:00:45.960 --> 00:00:48.060 shifts happening right now, the ones that fundamentally 00:00:48.060 --> 00:00:50.219 change how you approach your work, maybe your 00:00:50.219 --> 00:00:51.880 investment decisions, even how you think about 00:00:51.880 --> 00:00:54.460 creativity. Right. So our roadmap. First, we're 00:00:54.460 --> 00:00:56.600 going to unpack the technical stuff behind Anthropic's 00:00:56.600 --> 00:01:00.549 new staff engineer model. Then we'll look at 00:01:00.549 --> 00:01:03.210 how Microsoft is bringing this idea of vibe working, 00:01:03.429 --> 00:01:05.849 you know, just talking to the system into pretty 00:01:05.849 --> 00:01:08.689 much every office app you use. And finally, we'll 00:01:08.689 --> 00:01:11.250 dive into some specific AI prompting strategies 00:01:11.250 --> 00:01:14.489 for hopefully better, faster financial analysis. 00:01:14.890 --> 00:01:17.409 We know you need the clear insights, you know, 00:01:17.409 --> 00:01:19.829 without getting buried in information. Let's 00:01:19.829 --> 00:01:21.609 get into it. Okay. Let's kick off with segment 00:01:21.609 --> 00:01:24.500 one, the autonomous staff engineer. uh claude 00:01:24.500 --> 00:01:28.200 son at 4 .5 the jump here is just staggering 00:01:28.200 --> 00:01:30.900 both in what it can do and for how long i mentioned 00:01:30.900 --> 00:01:32.260 it briefly but yeah this is the breakthrough 00:01:32.260 --> 00:01:34.840 model sonnet 4 .5 feels like the most colleague 00:01:34.840 --> 00:01:37.000 -like model we've seen so far it's a real shift 00:01:37.000 --> 00:01:39.040 isn't it from using a static tool for one -off 00:01:39.040 --> 00:01:41.780 tasks to well engaging with something that feels 00:01:41.780 --> 00:01:44.120 like an active partner And what's really fascinating 00:01:44.120 --> 00:01:45.900 here, and this kind of speaks to the core technical 00:01:45.900 --> 00:01:49.560 challenge, is that jump in endurance. Sonnet 00:01:49.560 --> 00:01:52.500 4 .5 performed autonomous coding for up to 30 00:01:52.500 --> 00:01:55.299 hours. Now, just to give some context, its predecessor, 00:01:55.540 --> 00:01:58.760 Opus 4 .1, it only managed about seven hours 00:01:58.760 --> 00:02:01.879 of sustained unsupervised work. So going from 00:02:01.879 --> 00:02:04.680 seven hours to 30 in what, just a couple of months? 00:02:04.760 --> 00:02:07.980 That's not linear progress. That's serious acceleration. 00:02:08.280 --> 00:02:10.389 Incredible acceleration, yeah. But I think the 00:02:10.389 --> 00:02:12.509 real nugget here is why that duration matters 00:02:12.509 --> 00:02:14.710 so much. It's not just about running code for 00:02:14.710 --> 00:02:16.610 longer, right? It means the model is actually 00:02:16.610 --> 00:02:18.270 overcoming memory limits. It's keeping track 00:02:18.270 --> 00:02:21.050 of complex stuff across many steps. And it's 00:02:21.050 --> 00:02:23.330 preventing what people call plump drift, you 00:02:23.330 --> 00:02:25.389 know, where the AI kind of loses focus on the 00:02:25.389 --> 00:02:28.090 original big goal over time. 30 hours implies 00:02:28.090 --> 00:02:30.610 sustained complex memory management. That's huge. 00:02:31.030 --> 00:02:34.449 Exactly. And get this. In one test run, it generated 00:02:34.449 --> 00:02:37.569 over 11 ,000 lines of code. It handled really 00:02:37.569 --> 00:02:40.689 complex end -to -end engineering. tasks building 00:02:40.689 --> 00:02:43.330 full apps setting up databases even buying domain 00:02:43.330 --> 00:02:46.550 names like the whole product lifecycle and companies 00:02:46.550 --> 00:02:49.349 big ones like canva and cursor they're already 00:02:49.349 --> 00:02:51.930 starting to rely on it for deep research and 00:02:51.930 --> 00:02:54.870 complicated hiring workflows okay so if this 00:02:54.870 --> 00:02:57.669 model is working that long and doing that kind 00:02:57.669 --> 00:03:00.780 of complex work operating like a junior or maybe 00:03:00.780 --> 00:03:02.979 even a staff engineer. Start wondering about 00:03:02.979 --> 00:03:06.180 the cost, sure, but also the failure rate. Did 00:03:06.180 --> 00:03:08.240 it really work for 30 hours straight or was someone 00:03:08.240 --> 00:03:10.300 constantly nudging it back on track? Because 00:03:10.300 --> 00:03:13.060 if the AI is truly passing its own security audits, 00:03:13.099 --> 00:03:15.219 I mean, where does the human oversight actually 00:03:15.219 --> 00:03:17.870 start? What's left? That really hits the critical 00:03:17.870 --> 00:03:20.210 question, doesn't it? The data seems to suggest 00:03:20.210 --> 00:03:22.530 pretty minimal intervention was needed, and that's 00:03:22.530 --> 00:03:24.830 kind of the whole point. This level of autonomy 00:03:24.830 --> 00:03:27.530 means the human engineer's role, well, it fundamentally 00:03:27.530 --> 00:03:30.189 changes. You're not primarily the coder anymore. 00:03:30.270 --> 00:03:32.469 You're becoming more the validation expert, the 00:03:32.469 --> 00:03:35.069 auditor. Your job shifts maybe from creation 00:03:35.069 --> 00:03:37.930 to error management and oversight. So the new 00:03:37.930 --> 00:03:41.289 skill is agent management. Yeah. Learning how 00:03:41.289 --> 00:03:44.289 to manage these AI agents. Precisely, yeah. Okay, 00:03:44.310 --> 00:03:46.349 let's shift gears. Let's transition to another 00:03:46.349 --> 00:03:49.330 major platform update that really embodies this 00:03:49.330 --> 00:03:52.009 idea of the AI colleague, Microsoft's new agent 00:03:52.009 --> 00:03:55.710 mode and office agent, both inside 365 Copilot. 00:03:55.870 --> 00:04:00.009 Ah, right. This is Microsoft bringing that vibe 00:04:00.009 --> 00:04:03.009 -working idea, that really casual, almost conversational 00:04:03.009 --> 00:04:05.830 way of guiding things into the messy complexity 00:04:05.830 --> 00:04:09.039 of Excel and Word. The idea is you can just talk 00:04:09.039 --> 00:04:11.659 to the system, got it casually without needing 00:04:11.659 --> 00:04:13.699 complex commands. That is the core concept. Yeah. 00:04:13.840 --> 00:04:15.860 You might not need to know the exact function 00:04:15.860 --> 00:04:17.740 for a pivot table anymore. You just tell it what 00:04:17.740 --> 00:04:20.259 you need the data to actually show. We should 00:04:20.259 --> 00:04:21.639 look at the specifics, though, because there 00:04:21.639 --> 00:04:24.339 are two distinct agents here. OK, so what's the 00:04:24.339 --> 00:04:25.740 practical difference between them? How do they 00:04:25.740 --> 00:04:28.129 work? Well, first, you've got agent mode. This 00:04:28.129 --> 00:04:30.750 works inside the Office apps themselves, like 00:04:30.750 --> 00:04:33.189 Excel and Word with PowerPoint coming soon, apparently. 00:04:33.410 --> 00:04:36.290 This lets you do continuous iterative refinement 00:04:36.290 --> 00:04:38.069 right there in the document. Think of it like 00:04:38.069 --> 00:04:42.009 this. You prompt Copilot. It spits out an initial 00:04:42.009 --> 00:04:44.730 draft of, say, a sales report. Then you tell 00:04:44.730 --> 00:04:47.050 it, OK, bump this chart's projection up by 10 00:04:47.050 --> 00:04:52.069 % and change the font to Arial. It's all contextual 00:04:52.069 --> 00:04:54.870 guidance. OK, so it's kind of like having a really 00:04:54.870 --> 00:04:56.689 smart intern sitting right there next to you 00:04:56.689 --> 00:04:58.870 straight. sheet, basically ready to take notes 00:04:58.870 --> 00:05:01.709 and make changes instantly. Exactly. A very capable 00:05:01.709 --> 00:05:04.120 intern. Then there's the office agent. This is 00:05:04.120 --> 00:05:06.480 built into the main copilot chat interface. And 00:05:06.480 --> 00:05:08.699 this one's focused specifically on, let's call 00:05:08.699 --> 00:05:11.180 it single shot generation, generating polished 00:05:11.180 --> 00:05:13.540 documents or presentations straight from a simple 00:05:13.540 --> 00:05:16.240 chat prompt. You might say, generate a five slide 00:05:16.240 --> 00:05:18.839 presentation on our Q3 marketing strategy. And 00:05:18.839 --> 00:05:21.019 boom, the output is instantly formatted. It looks 00:05:21.019 --> 00:05:23.699 professional. Early benchmarks are showing something 00:05:23.699 --> 00:05:28.160 like 57 .2 % accuracy on complex multi -step 00:05:28.160 --> 00:05:32.910 tasks like these. 57 % accuracy. It sounds low 00:05:32.910 --> 00:05:35.529 maybe at first glance, but for really complex 00:05:35.529 --> 00:05:38.589 multi -step stuff, that's actually a pretty strong 00:05:38.589 --> 00:05:41.410 starting point. I have to admit, you know, I 00:05:41.410 --> 00:05:43.550 still wrestle with prompt drift myself sometimes 00:05:43.550 --> 00:05:46.970 when guiding these large models through complex 00:05:46.970 --> 00:05:50.589 creative work that takes multiple steps. It's 00:05:50.589 --> 00:05:52.550 tricky. It really is tricky to keep the system 00:05:52.550 --> 00:05:54.189 focused, especially if you're building out a 00:05:54.189 --> 00:05:57.279 long document. Yeah, and the accuracy gap that 00:05:57.279 --> 00:05:59.680 40 -something percent where it might miss the 00:05:59.680 --> 00:06:02.139 mark, that's where human review is still absolutely 00:06:02.139 --> 00:06:05.740 essential. But the goal is pretty clear. Automate 00:06:05.740 --> 00:06:07.699 the creation of polished documents from simple 00:06:07.699 --> 00:06:10.000 chat prompts. Which leads to the question, right, 00:06:10.139 --> 00:06:12.240 does this mean we can realistically stop building 00:06:12.240 --> 00:06:14.800 those complex Excel models and, you know, detailed 00:06:14.800 --> 00:06:17.100 word reports entirely by hand soon? The answer 00:06:17.100 --> 00:06:19.579 seems to be leaning towards yes, doesn't it? 00:06:20.009 --> 00:06:22.610 At least the grunt work, the initial heavy lifting 00:06:22.610 --> 00:06:24.990 of document creation looks like it's rapidly 00:06:24.990 --> 00:06:27.319 being automated. OK, so let's talk about using 00:06:27.319 --> 00:06:30.220 this agent power in maybe more high stakes fields. 00:06:30.339 --> 00:06:32.379 Segment three is all about using advanced AI 00:06:32.379 --> 00:06:34.779 prompts for smarter financial decisions. Right. 00:06:34.879 --> 00:06:37.699 We're seeing new resources pop up offering what 00:06:37.699 --> 00:06:39.819 they call a practical playbook of advanced AI 00:06:39.819 --> 00:06:42.100 prompts. These are designed specifically for 00:06:42.100 --> 00:06:45.259 smarter stock and crypto investing analysis. 00:06:45.540 --> 00:06:47.639 And just to be clear, this isn't about getting 00:06:47.639 --> 00:06:50.240 hot stock tips. It's really about automating 00:06:50.240 --> 00:06:53.019 the deep analysis part. Yeah, the benefit here 00:06:53.019 --> 00:06:55.399 is potentially huge because AI can drastically 00:06:55.399 --> 00:06:57.860 cut down the time you spend on fundamental research. 00:06:58.319 --> 00:07:01.680 Hours and hours of reading filings. This playbook 00:07:01.680 --> 00:07:04.160 apparently shows how to use a specific AI to 00:07:04.160 --> 00:07:07.120 get clear, almost analyst -level answers in, 00:07:07.160 --> 00:07:09.670 like, minutes a day. It's the kind of deep dive 00:07:09.670 --> 00:07:11.750 analysis that, you know, Wall Street firms spend 00:07:11.750 --> 00:07:13.649 thousands of human hours generating. OK, but 00:07:13.649 --> 00:07:15.889 how specific are these prompts? Are we just asking 00:07:15.889 --> 00:07:17.910 the AI like, hey, is Apple a good buy right now? 00:07:18.029 --> 00:07:20.029 Oh, not at all. No, no. We're talking about prompts 00:07:20.029 --> 00:07:22.149 that target really specific, often difficult 00:07:22.149 --> 00:07:24.870 to analyze sections of a company's public filings. 00:07:24.889 --> 00:07:27.970 For instance, you might prompt the AI to analyze 00:07:27.970 --> 00:07:30.329 the risk factors section in the last two quarters, 00:07:30.389 --> 00:07:34.670 10K filings for company X and synthesize a summary 00:07:34.670 --> 00:07:37.790 of potential legal exposures or maybe. Conduct 00:07:37.790 --> 00:07:39.709 sentiment analysis on management's discussion 00:07:39.709 --> 00:07:41.829 of the hiring outlook in the last three earnings 00:07:41.829 --> 00:07:44.850 call transcripts. Very specific data extraction. 00:07:45.149 --> 00:07:48.269 Got it. So you're leveraging the AI as a kind 00:07:48.269 --> 00:07:51.129 of research accelerator, like stacking Lego blocks 00:07:51.129 --> 00:07:52.990 of data much faster than a human could alone. 00:07:53.230 --> 00:07:55.550 But this is finance, right? High stakes territory. 00:07:55.610 --> 00:07:58.009 Is this kind of prompt based analysis really 00:07:58.009 --> 00:08:00.569 secure enough or reliable enough for making truly 00:08:00.569 --> 00:08:02.670 serious investing decisions? I mean, we're talking 00:08:02.670 --> 00:08:05.420 about real money here. and that confidence level 00:08:05.420 --> 00:08:08.019 that's still the key bottleneck absolutely ai 00:08:08.019 --> 00:08:10.680 offers incredibly faster research and synthesis 00:08:10.680 --> 00:08:14.300 no doubt but it absolutely positively requires 00:08:14.300 --> 00:08:17.980 human review for context for confidence checks 00:08:17.980 --> 00:08:20.800 for cross -checking against other sources before 00:08:20.800 --> 00:08:23.300 any serious capital decision is made think of 00:08:23.300 --> 00:08:25.699 it as a powerful co -pilot definitely but you 00:08:25.699 --> 00:08:28.660 are still a pilot in command Makes sense. OK, 00:08:28.759 --> 00:08:31.199 moving on to our final section, then a sort of 00:08:31.199 --> 00:08:34.279 rapid fire roundup of global quick hits and market 00:08:34.279 --> 00:08:36.379 shifts that just show the pace of all this. Right. 00:08:36.500 --> 00:08:39.139 Let's do it. First up, a major creative breakthrough 00:08:39.139 --> 00:08:41.960 out of China. Tencent just dropped Hunyuan Image 00:08:41.960 --> 00:08:45.039 3 .0. Now, this is a free image model that seems 00:08:45.039 --> 00:08:47.100 to exhibit pretty sophisticated reasoning. It 00:08:47.100 --> 00:08:50.240 can handle very long, complex prompts. And this 00:08:50.240 --> 00:08:51.679 is the really interesting part. It can apparently 00:08:51.679 --> 00:08:55.039 draw clean, legible text. inside the images it 00:08:55.039 --> 00:08:57.480 generates. Wait, why is drawing clean text inside 00:08:57.480 --> 00:08:59.519 an image such a big deal for these AI models? 00:08:59.620 --> 00:09:01.259 What's the challenge there? Well, it's because 00:09:01.259 --> 00:09:03.840 these diffusion models, the tech behind most 00:09:03.840 --> 00:09:06.860 image generators, they inherently treat text 00:09:06.860 --> 00:09:08.960 kind of like visual noise when they build an 00:09:08.960 --> 00:09:10.940 image. They see the shapes of letters, sure, 00:09:11.059 --> 00:09:13.879 but not the actual coherence, the meaning. So 00:09:13.879 --> 00:09:16.120 text usually comes out looking like garbled nonsense. 00:09:16.600 --> 00:09:19.700 Hunyuan Image 3 .0 actually solving that. It 00:09:19.700 --> 00:09:22.419 suggests a pretty significant leap in its understanding 00:09:22.419 --> 00:09:25.200 of semantic meaning, not just pixels. Okay, that 00:09:25.200 --> 00:09:27.080 is interesting. And here's something that really 00:09:27.080 --> 00:09:29.759 caught my eye. OpenAI is reportedly prepping 00:09:29.759 --> 00:09:32.049 a... standalone social app, kind of like TikTok 00:09:32.049 --> 00:09:35.330 in style, but powered entirely by Sora too. Everything 00:09:35.330 --> 00:09:37.970 in the feed would be AI generated video. Whoa. 00:09:38.169 --> 00:09:40.809 I mean, just imagine scaling a fully AI generated 00:09:40.809 --> 00:09:43.590 video feed to potentially a billion daily views. 00:09:43.710 --> 00:09:45.570 That's a moment of real wonder and maybe some 00:09:45.570 --> 00:09:48.990 disruption too. Exactly. That potential for pure 00:09:48.990 --> 00:09:52.470 targeted synthetic viral content. That's what 00:09:52.470 --> 00:09:54.090 we're talking about when we discuss these new 00:09:54.090 --> 00:09:58.090 market shifts and consumption patterns. OK, also 00:09:58.090 --> 00:10:00.009 on the policy and platform front, several quick 00:10:00.009 --> 00:10:02.490 updates. California Governor Newsom just signed 00:10:02.490 --> 00:10:06.169 that landmark AI safety bill SB 53 into law. 00:10:06.330 --> 00:10:09.230 So that's official. Chat GPT has new parental 00:10:09.230 --> 00:10:11.509 controls rolling out basically immediately to 00:10:11.509 --> 00:10:14.950 all users. And internally, word is Apple is heavily 00:10:14.950 --> 00:10:17.149 testing its own internal model. It's referred 00:10:17.149 --> 00:10:20.529 to internally as Veritas or sometimes Apple GPT. 00:10:20.590 --> 00:10:22.509 So they're definitely in the game. And for tool 00:10:22.509 --> 00:10:25.090 access, it's worth mentioning App 20X again. 00:10:25.190 --> 00:10:27.070 It's a platform that lets basically anyone. access 00:10:27.070 --> 00:10:29.070 open source alternative models and customize 00:10:29.070 --> 00:10:31.509 them using AI prompts. It kind of eliminates 00:10:31.509 --> 00:10:34.350 those technical hurdles that keep most non -developers 00:10:34.350 --> 00:10:36.289 from using models outside the big players like 00:10:36.289 --> 00:10:38.590 OpenAI or Google. Yeah, and if we connect this 00:10:38.590 --> 00:10:41.389 back to that Sora 2 social feed idea, the implication 00:10:41.389 --> 00:10:43.889 for just human content consumption is massive, 00:10:43.929 --> 00:10:47.110 isn't it? We could be moving from consuming mainly 00:10:47.110 --> 00:10:49.690 human -created content to interacting with purely 00:10:49.690 --> 00:10:52.029 manufactured, maybe hyper -personalized synthetic 00:10:52.029 --> 00:10:54.690 realities. Does that change everything? It feels 00:10:54.690 --> 00:10:57.570 like it could. Okay. let's try to unpack this 00:10:57.570 --> 00:11:00.190 one last time. Recap the big idea. The biggest 00:11:00.190 --> 00:11:02.230 takeaway from our deep dive today, I think, is 00:11:02.230 --> 00:11:04.509 the sheer velocity, but maybe more importantly, 00:11:04.570 --> 00:11:07.490 the direction of this change. We seem to be firmly 00:11:07.490 --> 00:11:10.929 moving past the era of just simple AI tools, 00:11:11.129 --> 00:11:13.950 you know, the basic chatbots, and really entering 00:11:13.950 --> 00:11:17.850 the age of autonomous agents. These are AIs capable 00:11:17.850 --> 00:11:21.230 of sustained, complex work, often without needing 00:11:21.230 --> 00:11:23.809 constant human hand -holding, whether that's 00:11:24.039 --> 00:11:26.419 coding an app for 30 hours or generating a complex 00:11:26.419 --> 00:11:28.539 financial report from a prompt. Yeah, absolutely. 00:11:28.720 --> 00:11:31.279 So we really encourage you, the listener, to 00:11:31.279 --> 00:11:33.460 maybe start testing Anthropic's new approach 00:11:33.460 --> 00:11:35.179 if you can, or try one of those new free image 00:11:35.179 --> 00:11:37.399 models we mentioned. Just to get a feel for the 00:11:37.399 --> 00:11:39.340 power of this kind of autonomy firsthand, it's 00:11:39.340 --> 00:11:42.059 different. But here's the final, maybe provocative 00:11:42.059 --> 00:11:44.200 thought we want to leave you with. If Claude 00:11:44.200 --> 00:11:47.059 Son at 4 .5 can genuinely work autonomously for 00:11:47.059 --> 00:11:49.600 30 hours straight and operate effectively like 00:11:49.600 --> 00:11:52.039 a staff engineer, what is the new competitive 00:11:52.039 --> 00:11:54.399 advantage for a human? working a standard 40 00:11:54.399 --> 00:11:56.940 -hour week. What does being human actually bring 00:11:56.940 --> 00:11:58.919 to the table now that an agent can't replicate 00:11:58.919 --> 00:12:02.620 or won't be able to soon? That's definitely something 00:12:02.620 --> 00:12:04.340 to think about. We'll be pondering that one too. 00:12:04.419 --> 00:12:06.039 Thanks for diving deep with us today.