WEBVTT 00:00:00.000 --> 00:00:02.740 Imagine for a second an AI, one that finishes 00:00:02.740 --> 00:00:06.639 those really complex Excel tasks, but in seconds. 00:00:07.160 --> 00:00:10.400 Well, our sources suggest it's not just theory 00:00:10.400 --> 00:00:12.919 anymore. It's reportedly here. And actually beating 00:00:12.919 --> 00:00:15.539 human experts at it. That's the kicker. We're 00:00:15.539 --> 00:00:18.339 digging into AI agents today. Yeah. From those 00:00:18.339 --> 00:00:20.920 spreadsheet wizards to this tiny AI that seems 00:00:20.920 --> 00:00:23.679 to kind of think like a brain. Get ready for 00:00:23.679 --> 00:00:26.359 some big shifts. Welcome back to the Deep Dive, 00:00:26.359 --> 00:00:28.760 everyone. Today, we are really unpacking your 00:00:28.760 --> 00:00:31.140 latest sources, focusing hard on the cutting 00:00:31.140 --> 00:00:33.780 edge of AI agents. Okay, let's get into it. Yeah, 00:00:33.840 --> 00:00:36.259 let's do it. Our mission, like always, is navigating 00:00:36.259 --> 00:00:39.880 this really fast -moving AI landscape. We try 00:00:39.880 --> 00:00:41.799 to boil down what matters most for you. Think 00:00:41.799 --> 00:00:43.399 of it as your shortcut, you know, understanding 00:00:43.399 --> 00:00:45.719 where AI is heading and how it's changing work, 00:00:45.780 --> 00:00:48.509 like right now. We'll kick off with an AI agent 00:00:48.509 --> 00:00:50.549 that's making some serious waves in finance. 00:00:51.049 --> 00:00:53.630 Then we'll kind of zoom out, look at how AI is 00:00:53.630 --> 00:00:56.649 shaping other areas, video creation, corporate 00:00:56.649 --> 00:00:59.289 strategy, that sort of thing. And finally, we've 00:00:59.289 --> 00:01:02.250 got this really surprising new AI architecture. 00:01:03.200 --> 00:01:05.200 Honestly, it could change everything we thought 00:01:05.200 --> 00:01:08.019 we knew about building intelligent systems. It's 00:01:08.019 --> 00:01:10.319 pretty wild. Okay. Let's jump right into segment 00:01:10.319 --> 00:01:14.340 one. Our sources are calling it an Excel revolution. 00:01:14.540 --> 00:01:17.680 Big words. Yeah. And it centers on this new spreadsheet 00:01:17.680 --> 00:01:20.719 native AI agent. It's called Shortcut. Right. 00:01:20.819 --> 00:01:23.879 And this isn't just like a fancy macro. No, not 00:01:23.879 --> 00:01:26.120 at all. It's way beyond that. This agent, it 00:01:26.120 --> 00:01:28.859 actually builds entire financial models, fills 00:01:28.859 --> 00:01:32.439 out data. And gets this one shots tasks, tasks 00:01:32.439 --> 00:01:34.799 usually given to junior analysts. Exactly. There's 00:01:34.799 --> 00:01:38.319 this demo where it auto populated a whole discounted 00:01:38.319 --> 00:01:41.780 cash flow model, a DCF. Which is. Pretty complex 00:01:41.780 --> 00:01:44.159 stuff, yeah. Totally. And it did it using raw 00:01:44.159 --> 00:01:46.640 data, just pulled straight from SEC filings, 00:01:46.680 --> 00:01:49.819 no human input. Wow. So it handles the data gathering, 00:01:49.959 --> 00:01:52.099 the formatting, the formulas, everything. The 00:01:52.099 --> 00:01:55.180 whole workflow. Data, formatting, formulas, checking, 00:01:55.420 --> 00:01:58.260 all in one go. It's kind of like watching a whole 00:01:58.260 --> 00:02:01.340 team work, but super fast, like a bot. Okay, 00:02:01.379 --> 00:02:03.280 that's impressive. What about performance? How 00:02:03.280 --> 00:02:05.459 does it stack up? Well, the sources detailed 00:02:05.459 --> 00:02:08.280 these benchmarks. They put shortcut up against 00:02:08.280 --> 00:02:11.400 actual junior analysts. From like... Top firms. 00:02:11.539 --> 00:02:14.060 Yeah. Private equity, banking, consulting, product 00:02:14.060 --> 00:02:18.280 roles, the big leagues. OK. And here's the kicker. 00:02:18.500 --> 00:02:21.340 Shortcut reportedly won eighty nine point one 00:02:21.340 --> 00:02:23.879 percent of the time. Eighty nine percent scored 00:02:23.879 --> 00:02:26.120 by managers from those firms. That's what the 00:02:26.120 --> 00:02:28.740 sources say. It's pretty dominant. OK, well. 00:02:29.159 --> 00:02:31.680 Any other comparisons? It also apparently beat 00:02:31.680 --> 00:02:35.599 the ChatGPT agent in 90 % of head -to -head tests. 00:02:35.680 --> 00:02:38.180 90%. Okay. Got to add the usual caveats, right? 00:02:38.199 --> 00:02:41.419 Early benchmarks, maybe fuzzy criteria for winning. 00:02:41.599 --> 00:02:43.780 Sure. Absolutely. Need to see more verification. 00:02:43.780 --> 00:02:46.460 But the buzz, it's definitely there. Oh, yeah. 00:02:46.919 --> 00:02:49.460 Finance influencers calling it a ChatGPT moment 00:02:49.460 --> 00:02:52.199 for Excel. You can practically hear the jaws 00:02:52.199 --> 00:02:54.379 dropping in finance departments. It really makes 00:02:54.379 --> 00:02:56.860 you think about that warning from Dario Amadei, 00:02:56.860 --> 00:02:59.319 you know, about... potentially half of junior 00:02:59.319 --> 00:03:03.159 office jobs just vanishing because of AI. Because 00:03:03.159 --> 00:03:05.759 this isn't just automating one cell, it's full 00:03:05.759 --> 00:03:08.159 task automation. Right. It's the whole process, 00:03:08.460 --> 00:03:11.180 getting data, building the model, writing analysis, 00:03:11.340 --> 00:03:14.340 even exporting it. It's a huge shift, like machines 00:03:14.340 --> 00:03:18.199 and factories. But now it's bots, one -shotting 00:03:18.199 --> 00:03:21.080 spreadsheets. Makes you really consider the skills 00:03:21.080 --> 00:03:23.900 we'll need going forward. So with shortcuts showing 00:03:23.900 --> 00:03:27.379 this kind of power, What does this really mean 00:03:27.379 --> 00:03:30.139 right now for those detailed, repetitive financial 00:03:30.139 --> 00:03:33.599 analysis tasks? Is it a total rewrite? Oh, absolutely. 00:03:33.840 --> 00:03:36.340 It signals a clear shift. Yeah. That routine 00:03:36.340 --> 00:03:39.560 spreadsheet grunt work. It's automatable now, 00:03:39.719 --> 00:03:41.840 plain and simple. Okay, so shortcuts shaping 00:03:41.840 --> 00:03:44.159 up Excel, but that's just one piece, right? Exactly. 00:03:44.199 --> 00:03:46.680 The sources show a much bigger picture. AI's 00:03:46.680 --> 00:03:49.379 capabilities are expanding everywhere. The industry's 00:03:49.379 --> 00:03:51.939 shifting fast. So moving beyond spreadsheets, 00:03:51.939 --> 00:03:53.969 what else is happening? Well... For one, we're 00:03:53.969 --> 00:03:56.050 seeing more complex systems. Take Claude, right? 00:03:56.189 --> 00:03:58.610 They've added this sub -agents feature. Sub -agents. 00:03:58.610 --> 00:04:00.509 What's that? It lets you build basically teams 00:04:00.509 --> 00:04:02.969 of agents. They can tackle multiple tasks at 00:04:02.969 --> 00:04:05.810 the same time working in parallel. So you give 00:04:05.810 --> 00:04:07.789 it a project and it kind of assembles its own 00:04:07.789 --> 00:04:10.729 digital team. Pretty much. Imagine that for complex 00:04:10.729 --> 00:04:12.770 workflows. And then there's this thing called 00:04:12.770 --> 00:04:16.269 JSON prompting. Sources are calling it the most 00:04:16.269 --> 00:04:20.089 underrated AI skill of 2025. Yeah, JSON prompting. 00:04:20.410 --> 00:04:22.470 it sounds technical but it's basically a way 00:04:22.470 --> 00:04:25.089 to give ai really precise instructions so it 00:04:25.089 --> 00:04:27.250 knows exactly what you need exactly structured 00:04:27.250 --> 00:04:30.670 instructions leads to reliable answers and crucially 00:04:30.670 --> 00:04:34.449 no hallucinations turns models like gpt gemini 00:04:34.449 --> 00:04:38.009 claude into consistent reliable agents super 00:04:38.009 --> 00:04:40.490 important for professional use okay makes sense 00:04:40.490 --> 00:04:43.189 what else we're also seeing ai just blend into 00:04:43.189 --> 00:04:45.949 everyday tools more google search is ai mode 00:04:45.949 --> 00:04:48.899 yeah it can now see homework You can upload a 00:04:48.899 --> 00:04:51.899 PDF or just use your camera. And ask AI mode 00:04:51.899 --> 00:04:54.339 for answers based on it. Yep. Makes research 00:04:54.339 --> 00:04:57.139 or figuring stuff out way faster. And ChatGPT 00:04:57.139 --> 00:05:00.259 has that new study mode. Right. Like a personalized 00:05:00.259 --> 00:05:02.879 tutor in your pocket. Adaptive lessons, the works. 00:05:03.040 --> 00:05:04.899 Feels like a real study partner. Interesting. 00:05:04.980 --> 00:05:06.879 What about the creative side? Runway just launched 00:05:06.879 --> 00:05:09.740 Aleph. It's a new video model. In -context video 00:05:09.740 --> 00:05:13.240 generation. Multitask visual stuff. So better 00:05:13.240 --> 00:05:17.000 AI video tools. Letting creators do wilder things, 00:05:17.160 --> 00:05:20.449 basically. More control, more flexibility for 00:05:20.449 --> 00:05:22.709 imaginative content. And then there was that 00:05:22.709 --> 00:05:25.110 story about Meta. Oh, yeah. The poaching attempt. 00:05:25.629 --> 00:05:28.750 Wild story. They reportedly tried to hire away 00:05:28.750 --> 00:05:31.149 like half the staff from Thinking Machines Lab. 00:05:31.350 --> 00:05:33.509 A top research group. And Zuckerberg himself 00:05:33.509 --> 00:05:35.649 was apparently messaging people. Offers over 00:05:35.649 --> 00:05:39.769 a billion dollars mentioned. Yeah. Not one person 00:05:39.769 --> 00:05:43.209 left. Wow. Tells you a lot about what top AI 00:05:43.209 --> 00:05:45.370 talent actually wants, doesn't it? Yeah. It's 00:05:45.370 --> 00:05:47.279 not just the money. Meanwhile, you've got the 00:05:47.279 --> 00:05:49.699 National Science Foundation, the NSF, putting 00:05:49.699 --> 00:05:52.139 serious cash into AI research. $100 million. 00:05:52.579 --> 00:05:54.399 That's significant. And they're partnering with 00:05:54.399 --> 00:05:57.620 big names, Simons Foundation, NIST, Capital One, 00:05:57.800 --> 00:05:59.939 Intel. Yeah, it's a big deal. It's not just about 00:05:59.939 --> 00:06:02.120 apps. It's foundational stuff. Material science, 00:06:02.459 --> 00:06:05.199 new tech, pushing boundaries in the physical 00:06:05.199 --> 00:06:07.480 world with AI's help. Okay, let's connect that 00:06:07.480 --> 00:06:09.180 back. If we look at that meta poaching attempt 00:06:09.180 --> 00:06:12.779 failing. What does that tell us about the AI 00:06:12.779 --> 00:06:16.259 talent market beyond just the insane money? It 00:06:16.259 --> 00:06:19.259 really says top AI talent seeks more than just 00:06:19.259 --> 00:06:22.000 cash. They want meaningful work, innovative places, 00:06:22.220 --> 00:06:24.459 real frontier stuff. All right, let's shift gears 00:06:24.459 --> 00:06:26.560 slightly. How about a quick round of some new 00:06:26.560 --> 00:06:29.040 AI tools and maybe some industry quick hits, 00:06:29.079 --> 00:06:30.819 stuff that's changing things day to day? Yeah, 00:06:30.860 --> 00:06:35.019 let's do it. Rapid fire. First up. Kasi AI. Kasi 00:06:35.019 --> 00:06:38.300 AI. Turns long videos into short viral clips. 00:06:38.519 --> 00:06:41.720 It finds the good bits automatically. Think AI 00:06:41.720 --> 00:06:45.240 content strategist for creators. Big time saver. 00:06:45.259 --> 00:06:48.019 Potential reach booster. Okay. Useful. Next. 00:06:48.339 --> 00:06:51.420 Cache Scene. Removes backgrounds. Enhances photos 00:06:51.420 --> 00:06:54.139 instantly. Super common task. Now super fast. 00:06:54.339 --> 00:06:56.600 For designers. Or just, you know, touching up 00:06:56.600 --> 00:06:59.180 photos. Andy. And Immersity for mobile. Turn 00:06:59.180 --> 00:07:01.680 your flat 2D images into immersive 3D videos. 00:07:01.920 --> 00:07:03.680 Kind of cool for making more dynamic content 00:07:03.680 --> 00:07:05.959 right on your phone. Neat. What about layout 00:07:05.959 --> 00:07:08.720 .dev? This one's for developers, designers maybe. 00:07:08.879 --> 00:07:11.699 Turn simple ideas like a sketch or just text 00:07:11.699 --> 00:07:14.120 into working prototypes really fast. So speeds 00:07:14.120 --> 00:07:17.079 up that early ideation phase. Massively. Could 00:07:17.079 --> 00:07:18.920 really accelerate getting projects off the ground. 00:07:19.100 --> 00:07:20.740 Okay, those are some tools. Any quick industry 00:07:20.740 --> 00:07:24.250 news bites? Yeah, a few. Zed AI released a model 00:07:24.250 --> 00:07:26.449 cheaper than DeepSeek. Yeah. And it's free to 00:07:26.449 --> 00:07:28.250 download. Shows things are getting more accessible, 00:07:28.410 --> 00:07:30.610 maybe more decentralized. Interesting. What else? 00:07:30.670 --> 00:07:33.269 Meta's going to let job candidates use AI during 00:07:33.269 --> 00:07:36.350 coding tests. Really? Yeah. Kind of acknowledging 00:07:36.350 --> 00:07:38.430 it's just part of the workflow now. Pretty pragmatic. 00:07:38.750 --> 00:07:41.050 Makes sense. Google. They revealed an internal 00:07:41.050 --> 00:07:44.350 site, AI Savvy Google, for their non -tech employees. 00:07:44.910 --> 00:07:47.170 Just trying to get everyone up to speed on AI. 00:07:47.569 --> 00:07:50.069 Smart internal education. Good idea. Anthropic. 00:07:50.600 --> 00:07:53.339 They published research on using automated agents 00:07:53.339 --> 00:07:57.620 to audit other AI models. AI auditing AI. Yeah. 00:07:58.000 --> 00:08:00.779 Crucial for safety, reliability, making sure 00:08:00.779 --> 00:08:02.680 these things behave, especially as they get more 00:08:02.680 --> 00:08:04.800 autonomous. Definitely important. And finally, 00:08:04.939 --> 00:08:07.860 Spotify. Just hinting at a more conversational 00:08:07.860 --> 00:08:10.800 voice AI interface down the line. You know, chat 00:08:10.800 --> 00:08:13.339 with your AI DJ about playlists based on your 00:08:13.339 --> 00:08:15.569 mood. Kind of like talking to a friend. So all 00:08:15.569 --> 00:08:17.970 these different tools and updates. Yeah. How 00:08:17.970 --> 00:08:20.949 do these smaller things impact us, you know, 00:08:20.949 --> 00:08:24.189 directly? They basically bring powerful AI into 00:08:24.189 --> 00:08:27.050 everyday stuff, often behind the scenes. Makes 00:08:27.050 --> 00:08:29.410 digital life smoother, more intuitive. Okay. 00:08:29.670 --> 00:08:32.450 Fascinating stuff. Before we get to maybe the 00:08:32.450 --> 00:08:34.730 most mind -bending part, this new AI architecture, 00:08:35.110 --> 00:08:37.529 let's just take a quick moment for our sponsor. 00:08:38.269 --> 00:08:42.259 Sponsor. And we are back. Okay, that revolutionary 00:08:42.259 --> 00:08:45.360 architecture we teased. Seriously, this could 00:08:45.360 --> 00:08:47.500 shake things up. It challenges some basic ideas 00:08:47.500 --> 00:08:50.299 about how we build AI. Right. The sources talk 00:08:50.299 --> 00:08:52.399 about something called the Hierarchical Reasoning 00:08:52.399 --> 00:08:55.899 Model, HRM, from a group called Sapient Intelligence. 00:08:56.299 --> 00:09:00.399 And the key thing, it's tiny, but it uses this 00:09:00.399 --> 00:09:02.840 brain -like architecture. That's the buzz. Brain 00:09:02.840 --> 00:09:05.559 -like? How so? It's got two main parts working 00:09:05.559 --> 00:09:08.669 together. A planner module. That's the slow thinking 00:09:08.669 --> 00:09:11.570 part. Strategizing, like planning chess moves. 00:09:11.669 --> 00:09:13.289 Okay, thinking ahead. And then a worker module. 00:09:13.509 --> 00:09:16.190 That's the fast acting part. Executing with like 00:09:16.190 --> 00:09:18.610 how your brain instantly recognizes a face, you 00:09:18.610 --> 00:09:20.690 know, quick processing. So they work together 00:09:20.690 --> 00:09:23.450 in a hierarchy. Exactly. They plan and solve 00:09:23.450 --> 00:09:26.409 in one go, one single forward pass. It's different 00:09:26.409 --> 00:09:28.330 from how many big language models work. And you 00:09:28.330 --> 00:09:31.169 said it's tiny. Unbelievably tiny. Only 27 million 00:09:31.169 --> 00:09:33.409 parameters. 27 million. How does that compare? 00:09:33.970 --> 00:09:37.110 GPT -1, the original. 117 million parameters. 00:09:37.679 --> 00:09:39.879 HRM is less than a quarter of the size. Wow. 00:09:40.080 --> 00:09:43.360 Okay. Tiny size. But what about performance? 00:09:43.580 --> 00:09:46.200 This is where it gets crazy. On the ARC -AGI 00:09:46.200 --> 00:09:49.240 benchmark, think of it like an AI IQ test for 00:09:49.240 --> 00:09:52.960 fluid intelligence. Yeah. HRM scored 40 .3%. 00:09:52.960 --> 00:09:58.200 Claude, 3 .7. 21 .2%. Another model, 03 Mini 00:09:58.200 --> 00:10:02.879 High, got 34 .5%. Wait. So this tiny model. It 00:10:02.879 --> 00:10:04.740 outperformed much bigger ones on this reasoning 00:10:04.740 --> 00:10:06.639 test. Significantly outperformed them. Okay. 00:10:06.679 --> 00:10:08.940 Are there specific examples? Yeah. And they really 00:10:08.940 --> 00:10:12.019 highlight the difference. It solved 55 % of Sudoku 00:10:12.019 --> 00:10:14.259 extreme puzzles. And other models? Claude and 00:10:14.259 --> 00:10:17.320 OpenAI scored 0 % on those. Zero. Okay. Any others? 00:10:17.580 --> 00:10:20.799 It found the optimal path to nearly 75 % of these 00:10:20.799 --> 00:10:23.399 really complex 30 by 30 mazes. And the others? 00:10:23.460 --> 00:10:25.500 Also 0%. It's not just a little better. It's 00:10:25.500 --> 00:10:27.539 solving problems the big guys just can't. Yeah, 00:10:27.559 --> 00:10:29.639 I still wrestle with prompt drift myself sometimes, 00:10:29.919 --> 00:10:32.360 getting the AI to stay on track. So the idea 00:10:32.360 --> 00:10:34.059 of a tiny model having this kind of consistent 00:10:34.059 --> 00:10:36.340 step -by -step reasoning, it's really fascinating. 00:10:36.500 --> 00:10:37.840 It makes you wonder if we've been too focused 00:10:37.840 --> 00:10:40.929 on just scaling up. Whoa, okay, that's... That's 00:10:40.929 --> 00:10:44.610 genuinely remarkable. A tiny AI outperforming 00:10:44.610 --> 00:10:47.830 models vastly larger, solving problems they can't 00:10:47.830 --> 00:10:50.190 even touch. It really does challenge the whole 00:10:50.190 --> 00:10:53.789 bigger is better idea in AI scaling. That's kind 00:10:53.789 --> 00:10:56.129 of mind blowing. Exactly. It's early days, sure, 00:10:56.330 --> 00:10:59.049 and it's not general purpose like GPT yet, but 00:10:59.049 --> 00:11:01.049 it suggests advanced reasoning might not require 00:11:01.049 --> 00:11:04.590 trillions of tokens or massive compute. So this 00:11:04.590 --> 00:11:07.769 raises a huge question then. Could this... brain 00:11:07.769 --> 00:11:10.490 -like hierarchical approach fundamentally change 00:11:10.490 --> 00:11:12.990 how we build AI in the future? Are we looking 00:11:12.990 --> 00:11:15.429 at a different path entirely? It definitely points 00:11:15.429 --> 00:11:18.750 towards efficiency and maybe more novel ways 00:11:18.750 --> 00:11:20.669 of thinking emerging rather than just relying 00:11:20.669 --> 00:11:23.110 on sheer scale. Could unlock AI for new areas, 00:11:23.230 --> 00:11:25.210 maybe inspire totally different kinds of intelligence. 00:11:25.490 --> 00:11:26.970 Okay, let's try and pull this all together. What's 00:11:26.970 --> 00:11:29.309 the big picture here? We've seen AI agents evolve 00:11:29.309 --> 00:11:32.389 incredibly quickly from tools to full task automation, 00:11:32.750 --> 00:11:36.100 spreadsheets, complex reasoning. Yeah, what strikes 00:11:36.100 --> 00:11:38.240 me is just the speed and the variety of innovation. 00:11:38.659 --> 00:11:41.740 It's not just about making AI bigger anymore. 00:11:41.899 --> 00:11:45.360 It's smarter, more specialized, way more efficient 00:11:45.360 --> 00:11:48.440 sometimes, tackling really sophisticated problems 00:11:48.440 --> 00:11:51.360 with amazing precision. So connecting it all 00:11:51.360 --> 00:11:54.259 up, the signal seems clear. AI is rapidly changing 00:11:54.259 --> 00:11:56.539 how we work, how we solve problems, not just 00:11:56.539 --> 00:11:59.940 in huge systems, but through these focused, intelligent 00:11:59.940 --> 00:12:03.279 agents in our everyday tools and tasks. Absolutely. 00:12:03.720 --> 00:12:06.080 It's becoming embedded. So this whole deep dive 00:12:06.080 --> 00:12:08.419 leaves us with a final thought, maybe something 00:12:08.419 --> 00:12:11.879 for you, the listener, to chew on. As these AI 00:12:11.879 --> 00:12:15.200 agents get better and more integrated, what new 00:12:15.200 --> 00:12:17.379 roles, what new opportunities actually open up 00:12:17.379 --> 00:12:19.539 for humans? Right. How does our own creativity, 00:12:19.600 --> 00:12:22.460 our problem solving? How do we adapt and evolve 00:12:22.460 --> 00:12:24.740 with these powerful new partners? It's not just 00:12:24.740 --> 00:12:26.720 about replacement. It's about partnership, right? 00:12:26.820 --> 00:12:28.899 Shifting maybe from doing the task to defining 00:12:28.899 --> 00:12:31.320 the problem or overseeing the AI, moving from 00:12:31.320 --> 00:12:33.860 rote work to more creative. of ideation. Could 00:12:33.860 --> 00:12:35.899 be. We really encourage you to keep exploring 00:12:35.899 --> 00:12:38.399 this, keep asking questions, keep thinking about 00:12:38.399 --> 00:12:40.919 how all this tech is shaping our collective future 00:12:40.919 --> 00:12:43.399 and your own place in it. Thank you for joining 00:12:43.399 --> 00:12:45.840 us on this deep dive. Until next time, keep that 00:12:45.840 --> 00:12:47.759 curiosity alive. Otiro Music.