WEBVTT 00:00:00.000 --> 00:00:03.180 Imagine AI not just doing tasks for you, but 00:00:03.180 --> 00:00:05.599 actually catching cyber attacks live as they 00:00:05.599 --> 00:00:07.719 happen. Or picture it running its own chemistry 00:00:07.719 --> 00:00:10.740 lab, you know, making discoveries in just days 00:00:10.740 --> 00:00:13.839 that used to take years. So today we're digging 00:00:13.839 --> 00:00:16.379 into a source that, well, it really shows AI 00:00:16.379 --> 00:00:18.980 is moving much faster than we thought, making 00:00:18.980 --> 00:00:21.519 real impacts right now. Welcome to the deep dive. 00:00:22.379 --> 00:00:24.420 Yeah, the source we're looking at today is absolutely 00:00:24.420 --> 00:00:27.960 packed with the latest in AI innovation. It's 00:00:27.960 --> 00:00:30.179 a fascinating look, really, at what's happening 00:00:30.179 --> 00:00:32.219 on the ground. We're going to cover AI going 00:00:32.219 --> 00:00:34.340 up against finance heavyweights, some really 00:00:34.340 --> 00:00:36.939 surprising new uses, and just the incredible 00:00:36.939 --> 00:00:39.079 speed this is all moving at. And our goal here, 00:00:39.140 --> 00:00:41.240 as always, is to cut through the noise, pull 00:00:41.240 --> 00:00:43.560 out the key insights, and we'll leave you feeling 00:00:43.560 --> 00:00:45.520 like you really get the picture. Okay, let's 00:00:45.520 --> 00:00:48.450 dive into the first big topic. Anthropix Claude 00:00:48.450 --> 00:00:50.990 for financial services. This thing is taking 00:00:50.990 --> 00:00:53.750 aim directly at the Bloomberg terminal. It is 00:00:53.750 --> 00:00:55.649 a big move. I mean, the Bloomberg terminal, it's 00:00:55.649 --> 00:00:57.590 been the standard in finance for decades, right? 00:00:57.670 --> 00:01:00.670 It costs, what, like $25 ,000 a pop? Exactly. 00:01:00.770 --> 00:01:03.490 So Claude's coming in as this powerful, potentially 00:01:03.490 --> 00:01:05.769 disruptive alternative. So what's actually different 00:01:05.769 --> 00:01:08.030 about it? How does it work? Well, it does the 00:01:08.030 --> 00:01:10.469 expected thing, pulling in external market data. 00:01:10.650 --> 00:01:13.129 Yeah. You know, S &P, Faxet, that kind of stuff. 00:01:13.230 --> 00:01:16.469 Okay. But the key part is... It then merges that 00:01:16.469 --> 00:01:19.650 with a company's own internal data, stuff sitting 00:01:19.650 --> 00:01:22.150 in Databricks or Snowflake, for instance. Ah, 00:01:22.409 --> 00:01:24.530 so it connects the outside view with the inside 00:01:24.530 --> 00:01:26.590 view? Precisely. It's like building with these 00:01:26.590 --> 00:01:28.909 different data Lego blocks, basically. And I 00:01:28.909 --> 00:01:30.930 saw mention of something called Clawed Code. 00:01:31.250 --> 00:01:33.310 Yeah, that's important. It lets analysts run 00:01:33.310 --> 00:01:35.590 their own custom models or simulations right 00:01:35.590 --> 00:01:39.459 there. So no more... Fiddling with complex scripts 00:01:39.459 --> 00:01:42.599 or trying to jury rig different APIs together. 00:01:42.859 --> 00:01:45.200 Right. No more Frankensteining your APIs, as 00:01:45.200 --> 00:01:47.420 they put it. It makes that whole process much 00:01:47.420 --> 00:01:50.480 smoother. Okay. And they've boosted the token 00:01:50.480 --> 00:01:53.780 and query limits, which is huge. We think it 00:01:53.780 --> 00:01:56.239 can handle bigger chunks of information. Exactly. 00:01:56.459 --> 00:01:58.659 Analysts can feed it massive documents like a 00:01:58.659 --> 00:02:01.620 whole 10K filing or complex M &A docs without 00:02:01.620 --> 00:02:03.939 hitting those frustrating token walls where the 00:02:03.939 --> 00:02:06.680 AI just stomps. And how does it actually perform? 00:02:07.159 --> 00:02:08.780 Numbers wise. Pretty impressively, actually. 00:02:08.819 --> 00:02:11.919 On the finance agent benchmark, it scored 44 00:02:11.919 --> 00:02:15.020 .5 percent accuracy. OK. And how does that compare? 00:02:15.360 --> 00:02:17.740 Well, that puts it right up there neck and neck 00:02:17.740 --> 00:02:20.819 with OpenAI's best models. Most of the other 00:02:20.819 --> 00:02:23.580 tools tested were like under 30 percent. Wow. 00:02:23.699 --> 00:02:25.719 That's a significant difference. It really is. 00:02:25.819 --> 00:02:28.419 And the impact on businesses seems real. Breedwater 00:02:28.419 --> 00:02:30.599 apparently saw a 20 percent productivity jump. 00:02:31.840 --> 00:02:34.509 Norway's sovereign wealth fund saved. Get this, 00:02:34.610 --> 00:02:39.650 213 ,000 hours. That's staggering. And AIG cut 00:02:39.650 --> 00:02:42.229 its underwriting review time by five times. And 00:02:42.229 --> 00:02:43.650 we should be clear, these aren't just simple 00:02:43.650 --> 00:02:45.789 lookups, like what's the stock price? No, this 00:02:45.789 --> 00:02:48.509 is complex stuff, like multi -step analysis of 00:02:48.509 --> 00:02:51.090 SEC filings, the kind of thing that, yeah, would 00:02:51.090 --> 00:02:53.729 make a junior analyst break a sweat. That integration 00:02:53.729 --> 00:02:56.009 with internal data, though, it sounds powerful, 00:02:56.090 --> 00:02:59.539 but does the source mention any... security concerns, 00:02:59.699 --> 00:03:02.819 giving an AI access to sensitive company info? 00:03:03.120 --> 00:03:04.639 That's a really good point. The source mainly 00:03:04.639 --> 00:03:08.039 focuses on the efficiency wins. But yeah, it 00:03:08.039 --> 00:03:10.000 definitely implies you need really solid security 00:03:10.000 --> 00:03:12.439 around these systems. Makes sense. And just quickly, 00:03:12.539 --> 00:03:14.680 it also mentioned perplexity is sort of quietly 00:03:14.680 --> 00:03:17.159 emerging as another strong tool for analysts. 00:03:17.460 --> 00:03:20.639 The Coinbase CEO, Brian Armstrong, he called 00:03:20.639 --> 00:03:25.219 their stuff a 10x unlock for AI in finance. Okay, 00:03:25.259 --> 00:03:28.030 so pulling this together. What does this really 00:03:28.030 --> 00:03:31.189 signal for the future of those traditional finance 00:03:31.189 --> 00:03:34.490 tools, maybe even how finance teams work? I think 00:03:34.490 --> 00:03:36.550 it suggests AI is making that high end analysis 00:03:36.550 --> 00:03:40.080 faster and maybe more accessible to beat. All 00:03:40.080 --> 00:03:41.939 right, let's shift gears. The next section was 00:03:41.939 --> 00:03:44.819 called Today in AI. Kind of a rapid fire look 00:03:44.819 --> 00:03:46.620 at different developments. Yeah, this part got 00:03:46.620 --> 00:03:48.280 really interesting. A lot going on simultaneously. 00:03:48.879 --> 00:03:50.900 On one hand, you see this creative expansion, 00:03:51.159 --> 00:03:54.819 like Runway's new AI motion caption model. Oh 00:03:54.819 --> 00:03:57.419 yeah, that sounded cool. It tracks face, hands, 00:03:57.560 --> 00:04:00.539 body, and then lets creators just drop that motion 00:04:00.539 --> 00:04:02.759 into any scene. Pretty amazing for visual stuff. 00:04:03.020 --> 00:04:04.930 Definitely. And then there's that Google business 00:04:04.930 --> 00:04:06.969 calling agent. Right. The one that calls shops 00:04:06.969 --> 00:04:08.629 for you. Yeah. It'll phone up a local store, 00:04:08.729 --> 00:04:10.469 ask if they have something, check the price, 00:04:10.550 --> 00:04:13.169 and then just report back. So no more slightly 00:04:13.169 --> 00:04:15.750 awkward phone calls for basic info. Exactly. 00:04:16.029 --> 00:04:17.750 It could be a real time saver for those little 00:04:17.750 --> 00:04:20.750 tedious tasks. And speaking of innovation, China's 00:04:20.750 --> 00:04:23.389 approach to data centers. Putting them in the 00:04:23.389 --> 00:04:26.250 ocean. Yeah, isn't that wild? Using the seawater 00:04:26.250 --> 00:04:28.750 for natural cooling. It's clever. The first phase 00:04:28.750 --> 00:04:31.750 apparently has enough power, like 198 servers, 00:04:32.029 --> 00:04:36.110 to train GPT -3 .5 in just one day. That's serious 00:04:36.110 --> 00:04:38.689 capacity and a really neat solution to the heat 00:04:38.689 --> 00:04:43.029 problem. And AWS is also pushing forward, putting 00:04:43.029 --> 00:04:45.550 another $100 million into their generative AI 00:04:45.550 --> 00:04:47.689 innovation center. Right, helping more companies 00:04:47.689 --> 00:04:50.790 build actual AI agents. You know, AI that can 00:04:50.790 --> 00:04:53.220 handle... complex workflows, not just single 00:04:53.220 --> 00:04:55.480 tasks. But it's not all smooth sailing and cool 00:04:55.480 --> 00:04:57.600 tech, is it? There's a flip side mentioned. No, 00:04:57.660 --> 00:05:00.100 definitely not. A pretty stark finding was that 00:05:00.100 --> 00:05:04.939 75 % of S &P 500 companies now list AI as a risk 00:05:04.939 --> 00:05:08.000 factor. 75 %? That's huge. What kind of risks? 00:05:08.180 --> 00:05:10.100 Everything, really. Ethical issues, potential 00:05:10.100 --> 00:05:12.639 for operational screw -ups, regulatory worries, 00:05:12.899 --> 00:05:15.750 security vulnerabilities, the whole gamut. And 00:05:15.750 --> 00:05:17.509 just to hammer that point home, there was another 00:05:17.509 --> 00:05:19.930 chat GPT outage mentioned, right? Around July 00:05:19.930 --> 00:05:23.329 16th. Yep. Global users hit with login problems, 00:05:23.490 --> 00:05:27.350 blank screens, lost work. It's a good reminder 00:05:27.350 --> 00:05:30.930 that this tech is still maturing, still has hiccups. 00:05:31.110 --> 00:05:32.990 So taking all these different pieces, the creative 00:05:32.990 --> 00:05:35.329 tools, the efficiency gains, the infrastructure 00:05:35.329 --> 00:05:38.529 challenges, the risks, how does this paint a 00:05:38.529 --> 00:05:40.949 picture of where AI is right now? It feels like 00:05:40.949 --> 00:05:44.029 AI is booming, spreading out fast. Yeah. but 00:05:44.029 --> 00:05:46.350 also hitting some real growing pains, like dealing 00:05:46.350 --> 00:05:49.329 with major risks and needing massive infrastructure. 00:05:49.629 --> 00:05:52.850 Beat. Okay, moving on. The source also touched 00:05:52.850 --> 00:05:55.470 on more practical applications, specific tools 00:05:55.470 --> 00:05:57.449 people are using. Yeah, and it made a point that 00:05:57.449 --> 00:05:59.810 not every AI idea is about, you know, instant 00:05:59.810 --> 00:06:02.170 riches. Right. One actual business model they 00:06:02.170 --> 00:06:04.790 mentioned was using AI to auto -generate videos 00:06:04.790 --> 00:06:07.129 for popular affiliate products. Kind of clever 00:06:07.129 --> 00:06:09.089 niche application. And then there was LeVart, 00:06:09.209 --> 00:06:12.470 described as the first real AI design agent. 00:06:12.860 --> 00:06:14.519 Yeah, that one sounded interesting. It supposedly 00:06:14.519 --> 00:06:17.680 uses models like ChatGPT and Runway to build 00:06:17.680 --> 00:06:20.920 out whole campaigns, websites, branding, all 00:06:20.920 --> 00:06:23.319 from just one prompt. That's quite a leap for 00:06:23.319 --> 00:06:25.920 creative automation, if it works well. For sure. 00:06:26.060 --> 00:06:27.899 And for the builders out there, the tech folks, 00:06:28.079 --> 00:06:31.139 there's a guide mentioned for using N8n templates. 00:06:31.399 --> 00:06:33.339 That's one of those workflow automation tools. 00:06:33.399 --> 00:06:36.040 Okay, what was the guide about? Basically, how 00:06:36.040 --> 00:06:37.740 to use their templates without running into common 00:06:37.740 --> 00:06:40.800 errors, covered risks to watch out for, set of 00:06:40.800 --> 00:06:44.139 steps, debugging. You know, the practical stuff. 00:06:44.279 --> 00:06:46.639 And I have to admit, I still wrestle with prompt 00:06:46.639 --> 00:06:49.379 drift myself sometimes. Oh, yeah. Where the AI 00:06:49.379 --> 00:06:51.560 kind of wanders off. Exactly. It starts strong, 00:06:51.699 --> 00:06:53.920 then veers off course from what you originally 00:06:53.920 --> 00:06:57.060 asked. So guides like that, super helpful, honestly. 00:06:57.279 --> 00:06:59.439 So looking at these examples, automated videos, 00:06:59.740 --> 00:07:02.800 AI design agents, workflow tools, what's the 00:07:02.800 --> 00:07:05.060 common theme here about how AI is being used 00:07:05.060 --> 00:07:07.680 practically? It seems AI is really becoming a 00:07:07.680 --> 00:07:12.329 go -to tool for... creative generation and for 00:07:12.329 --> 00:07:14.970 automating some pretty complex workflows. Now 00:07:14.970 --> 00:07:17.550 this next section, this really grabbed me. AI 00:07:17.550 --> 00:07:20.089 progress might be moving much faster than expected. 00:07:20.350 --> 00:07:22.149 Yeah, this part was eye -opening. The thinking 00:07:22.149 --> 00:07:24.970 was, you know, 2025 would be the year AI agents 00:07:24.970 --> 00:07:27.149 just started handling repetitive tasks well. 00:07:27.310 --> 00:07:29.410 Right, basic automation. But what we're seeing 00:07:29.410 --> 00:07:32.529 now is AI making actual discoveries, catching 00:07:32.529 --> 00:07:35.490 live threats. It almost feels like we've jumped 00:07:35.490 --> 00:07:38.290 ahead, maybe skipped a step in how these intelligent 00:07:38.290 --> 00:07:41.079 systems evolve. The source gave specific examples, 00:07:41.339 --> 00:07:44.879 right? Like Google's big sleep AI agent. Yes. 00:07:45.079 --> 00:07:47.980 That one apparently spotted and stopped a live 00:07:47.980 --> 00:07:51.220 software exploit before it could be used maliciously. 00:07:51.240 --> 00:07:53.399 Before it was even deployed in the wild. That's 00:07:53.399 --> 00:07:55.899 what Google claims. The first real world case, 00:07:56.000 --> 00:07:58.879 they say, of AI intercepting an active cyber 00:07:58.879 --> 00:08:01.420 threat like that. That's a pretty big milestone. 00:08:01.680 --> 00:08:03.990 Huge. And then there's that self -driving chemistry 00:08:03.990 --> 00:08:06.949 lab at NC State. The self -driving lab? Yeah, 00:08:06.990 --> 00:08:09.470 an AI -powered lab that runs its own chemical 00:08:09.470 --> 00:08:11.970 experiments. It collects something like 10 times 00:08:11.970 --> 00:08:14.829 more data than human researchers. Okay. And the 00:08:14.829 --> 00:08:17.230 potential payoff. It could slash the time it 00:08:17.230 --> 00:08:20.370 takes to discover new materials from years, literally 00:08:20.370 --> 00:08:23.550 years down to days. Wow. Whoa. Imagine scaling 00:08:23.550 --> 00:08:25.629 that. A billion queries running experiments. 00:08:25.769 --> 00:08:28.569 That's just incredible potential. And there's 00:08:28.569 --> 00:08:30.569 data backing up this feeling of acceleration. 00:08:31.189 --> 00:08:34.769 Mm -hmm. Data from NTR apparently shows AI is 00:08:34.769 --> 00:08:37.230 doubling the length and complexity of tasks it 00:08:37.230 --> 00:08:39.710 can manage across nine different benchmarks roughly 00:08:39.710 --> 00:08:42.149 every 4 .5 months now. Every four and a half 00:08:42.149 --> 00:08:44.549 months. How does that compare to before? Just 00:08:44.549 --> 00:08:46.490 six months ago, that doubling rate was about 00:08:46.490 --> 00:08:49.029 seven months. So the pace itself is picking up 00:08:49.029 --> 00:08:51.710 speed. So it's accelerating faster. Exactly. 00:08:51.710 --> 00:08:53.789 Which leads to that question they posed. Are 00:08:53.789 --> 00:08:56.090 we maybe moving into a post -agentic AI phase? 00:08:56.409 --> 00:08:59.389 And just to clarify that term post -agentic AI. 00:09:00.570 --> 00:09:03.649 It basically means AI moving beyond just following 00:09:03.649 --> 00:09:05.929 instructions, right? Yeah. Starting to autonomously 00:09:05.929 --> 00:09:08.809 discover things. Yeah, precisely. Not just executing 00:09:08.809 --> 00:09:11.330 tasks we define, but finding new solutions, new 00:09:11.330 --> 00:09:13.929 capabilities on its own, moving beyond just being 00:09:13.929 --> 00:09:16.830 an agent doing our bidding. So if this acceleration 00:09:16.830 --> 00:09:18.769 is real, this almost exponential improvement, 00:09:19.190 --> 00:09:21.929 what could that mean for innovation, for solving 00:09:21.929 --> 00:09:24.169 big problems? It really suggests we might be 00:09:24.169 --> 00:09:26.230 right on the edge of AI fundamentally changing 00:09:26.230 --> 00:09:28.309 how discovery happens, compressing what took. 00:09:28.639 --> 00:09:31.580 research years into maybe just AI weeks in fields 00:09:31.580 --> 00:09:34.740 like science, software, security. Two sec silence. 00:09:34.940 --> 00:09:37.159 Mid -roll sponsor read. Okay, let's try to pull 00:09:37.159 --> 00:09:39.419 this all together. Reflecting on everything we've 00:09:39.419 --> 00:09:42.320 covered from the source today, what's the big 00:09:42.320 --> 00:09:44.620 picture? Well, it paints a picture of incredibly 00:09:44.620 --> 00:09:47.179 rapid AI innovation, doesn't it? Things are moving 00:09:47.179 --> 00:09:50.799 fast. Yeah. From seeing AI challenge huge established 00:09:50.799 --> 00:09:53.580 players like Bloomberg and finance. To actually 00:09:53.580 --> 00:09:56.580 stopping live cyber threats and speeding up fundamental 00:09:56.580 --> 00:09:59.460 scientific research. Feels like AI is shifting 00:09:59.460 --> 00:10:01.899 gears. Shifting from just automating the simple 00:10:01.899 --> 00:10:05.259 stuff. To being a genuine force for like fundamental 00:10:05.259 --> 00:10:07.460 change. It's not just about efficiency anymore. 00:10:07.639 --> 00:10:10.220 And we're seeing these really powerful new tools 00:10:10.220 --> 00:10:13.220 pop up, design agents, better financial analysis. 00:10:13.679 --> 00:10:16.159 But at the same time, there are these really 00:10:16.159 --> 00:10:18.799 crucial conversations happening and needing to 00:10:18.799 --> 00:10:22.620 happen around the risks, the ethics, the sheer 00:10:22.620 --> 00:10:25.019 amount of infrastructure required. Right. It's 00:10:25.019 --> 00:10:27.240 not just capability. It's also about control, 00:10:27.440 --> 00:10:31.120 safety and access. So the main takeaway, I think, 00:10:31.120 --> 00:10:34.279 is crystal clear. AI development isn't just fast, 00:10:34.379 --> 00:10:37.200 it's accelerating. And it's changing what's possible 00:10:37.200 --> 00:10:39.919 fundamentally right now, today. That acceleration 00:10:39.919 --> 00:10:43.019 point really stands out, which leads to a question 00:10:43.019 --> 00:10:46.659 for you listening. Given this rapid, maybe even 00:10:46.659 --> 00:10:50.019 accelerating pace, how do you see AI's progress 00:10:50.019 --> 00:10:52.659 impacting your own field or maybe just your day 00:10:52.659 --> 00:10:54.500 -to -day life, even in the next few months? Yeah, 00:10:54.539 --> 00:10:56.620 that's something worth thinking about. What part 00:10:56.620 --> 00:10:58.820 of this deep dive resonated most with you? What 00:10:58.820 --> 00:11:00.700 caught your attention? Definitely something to 00:11:00.700 --> 00:11:02.220 reflect on. Thank you for joining us on The Deep 00:11:02.220 --> 00:11:04.639 Dive. We really hope this quick tour through 00:11:04.639 --> 00:11:08.240 the latest AI happenings gave you some aha moments, 00:11:08.379 --> 00:11:10.679 maybe some things to chew on. Until next time, 00:11:10.879 --> 00:11:12.559 EOTROMusic.