WEBVTT 00:00:00.000 --> 00:00:02.040 You know, AI agent systems, they represent these 00:00:02.040 --> 00:00:05.000 incredible levels of complexity. You've got coordination, 00:00:05.280 --> 00:00:09.439 reasoning, autonomous critique, often recursion. 00:00:09.740 --> 00:00:13.240 And yet Amazon just dropped this detailed guide 00:00:13.240 --> 00:00:15.279 for their deep agents. And it makes the whole 00:00:15.279 --> 00:00:18.000 thing feel, well, almost unexpectedly simple. 00:00:18.780 --> 00:00:21.539 Yeah, it's the speed, really. That's the disruptive 00:00:21.539 --> 00:00:23.000 thing we're looking at today. We're talking about 00:00:23.000 --> 00:00:25.800 going from like a working prototype on your local 00:00:25.800 --> 00:00:28.899 machine straight to an enterprise -ready production 00:00:28.899 --> 00:00:31.440 system. Okay. And potentially in under five minutes. 00:00:31.620 --> 00:00:33.820 I mean, that just cuts out months of typical 00:00:33.820 --> 00:00:36.380 infrastructure headaches. Welcome to the Deep 00:00:36.380 --> 00:00:38.679 Dive. We're here to unpack this stack of sources 00:00:38.679 --> 00:00:40.820 we've gathered for you. But before we dive in, 00:00:40.939 --> 00:00:43.140 let's just quickly level set on the main idea 00:00:43.140 --> 00:00:47.509 here. What exactly is an AI agent? Okay. Simply 00:00:47.509 --> 00:00:51.109 put, an AI agent is basically a coordinated program. 00:00:51.329 --> 00:00:53.530 It's designed to tackle complex, multi -step 00:00:53.530 --> 00:00:56.689 jobs all by itself, autonomously. Right. Think 00:00:56.689 --> 00:00:58.750 of it like, I don't know, assigning a self -managing 00:00:58.750 --> 00:01:01.090 project manager right inside your computer. That's 00:01:01.090 --> 00:01:03.450 a good way to put it. So our mission for this 00:01:03.450 --> 00:01:06.840 DevDive covers three main areas. First up. We're 00:01:06.840 --> 00:01:09.140 going to analyze Amazon's pretty aggressive agent 00:01:09.140 --> 00:01:11.040 strategy, looking at that big infrastructure 00:01:11.040 --> 00:01:14.239 bet they're making. Second, we'll shift to tracking 00:01:14.239 --> 00:01:17.519 some really crucial global AI adoption trends. 00:01:17.700 --> 00:01:19.879 We're focusing mainly on price and accessibility 00:01:19.879 --> 00:01:22.859 there. And finally, we have to look at this incredible 00:01:22.859 --> 00:01:26.180 breakthrough in materials science. Almost instantaneous 00:01:26.180 --> 00:01:29.640 discovery, really. Powered by AI working with 00:01:29.640 --> 00:01:32.040 human chemists. Pretty wild. Okay, let's start 00:01:32.040 --> 00:01:34.989 there. With Amazon's deep agents, it really feels 00:01:34.989 --> 00:01:36.750 like they're letting you skip the hardest part 00:01:36.750 --> 00:01:39.010 of enterprise AI, doesn't it? You know, the deployment, 00:01:39.090 --> 00:01:41.689 the governance. Yeah. The simplicity seems to 00:01:41.689 --> 00:01:44.209 be the core offering. It's like plug and go autonomous 00:01:44.209 --> 00:01:47.849 research. And what's fascinating is how seamless 00:01:47.849 --> 00:01:52.010 but detailed the workflow is that they've laid 00:01:52.010 --> 00:01:55.150 out. It's this coordinated recursive system structure. 00:01:55.269 --> 00:01:57.549 And somehow it just works without you needing 00:01:57.549 --> 00:02:00.609 to like. touch or manage the underlying infrastructure 00:02:00.609 --> 00:02:03.390 at all. It seems like it's all about having clearly 00:02:03.390 --> 00:02:06.430 defined roles within that system. That seems 00:02:06.430 --> 00:02:09.169 crucial for making it reliable. So you have the 00:02:09.169 --> 00:02:12.599 research agent. Its job is basically scanning 00:02:12.599 --> 00:02:14.520 the Internet, pulling in huge amounts of data, 00:02:14.580 --> 00:02:16.599 doing it efficiently. Right. But then you need 00:02:16.599 --> 00:02:18.860 a filter, obviously. So that's where the critique 00:02:18.860 --> 00:02:21.719 agent comes in. Its whole purpose is to review 00:02:21.719 --> 00:02:25.120 the quality, the rigor of what the research agent 00:02:25.120 --> 00:02:28.479 found. It checks for relevance, source trustworthiness. 00:02:30.060 --> 00:02:31.900 you know, is a massive challenge in AI research 00:02:31.900 --> 00:02:34.840 today. Absolutely. And then crucially, the whole 00:02:34.840 --> 00:02:37.500 thing's managed by this orchestrator agent. This 00:02:37.500 --> 00:02:39.979 agent, it takes the initial big question and 00:02:39.979 --> 00:02:42.840 breaks it down into smaller, manageable subtasks. 00:02:42.900 --> 00:02:46.099 It handles all the file management. And maybe 00:02:46.099 --> 00:02:48.500 most importantly, it controls that feedback loop. 00:02:48.659 --> 00:02:50.580 Exactly. So if the critique agent says, nope, 00:02:50.639 --> 00:02:52.240 this isn't good enough, the orchestrator just 00:02:52.240 --> 00:02:54.819 sends it back, go do more research, critique 00:02:54.819 --> 00:02:57.120 it deeper. And that whole cycle, that questioning 00:02:57.120 --> 00:02:59.199 and re -researching, it happens automatically. 00:02:59.340 --> 00:03:02.379 until the job meets the requirements. The efficiency 00:03:02.379 --> 00:03:06.159 there is just staggering. And the outputs, saved 00:03:06.159 --> 00:03:08.819 as structured reports, usually marked down or 00:03:08.819 --> 00:03:11.500 streamed right back to the user. So you get autonomous 00:03:11.500 --> 00:03:13.719 research that's basically ready for immediate 00:03:13.719 --> 00:03:15.879 use in an enterprise setting. And this really 00:03:15.879 --> 00:03:18.340 shines a light on Amazon's core strategy, doesn't 00:03:18.340 --> 00:03:19.979 it? Yeah. They're not really trying to compete 00:03:19.979 --> 00:03:23.539 head -to -head with, say, OpenAI or Claude on 00:03:23.539 --> 00:03:26.689 just raw model quality. Not right now, anyway. 00:03:26.810 --> 00:03:28.469 No, it doesn't seem like it. They're betting 00:03:28.469 --> 00:03:31.270 on being the absolute best place, the most reliable 00:03:31.270 --> 00:03:34.129 platform to actually run the agency build, but 00:03:34.129 --> 00:03:36.810 run them at massive scale. They're focusing entirely 00:03:36.810 --> 00:03:40.990 on providing those necessary enterprise guardrails, 00:03:41.009 --> 00:03:44.069 the reliable memory systems, deep access to internal 00:03:44.069 --> 00:03:46.710 company tools, all the stuff you need for mass 00:03:46.710 --> 00:03:49.590 adoption. It's a pure infrastructure play. So 00:03:49.590 --> 00:03:51.569 this does raise a key question, though. If Amazon 00:03:51.569 --> 00:03:54.050 makes deploying these potentially mission -critical 00:03:54.050 --> 00:03:56.770 systems this simple, this fast, aren't we just 00:03:56.770 --> 00:03:59.310 accelerating vendor lock -in, getting stuck with 00:03:59.310 --> 00:04:01.550 their specific infrastructure stack? Yeah, I 00:04:01.550 --> 00:04:03.569 think that's a definite risk. We risk rapid vendor 00:04:03.569 --> 00:04:05.889 lock -in when deployment becomes this easy. It's 00:04:05.889 --> 00:04:09.110 just the path of least resistance. Okay, so moving 00:04:09.110 --> 00:04:11.879 from that infrastructure simplicity. Let's broaden 00:04:11.879 --> 00:04:13.419 the view to the global scale. We need to talk 00:04:13.419 --> 00:04:15.379 about adoption patterns. How's the rest of the 00:04:15.379 --> 00:04:18.360 world actually interacting with this tech? And 00:04:18.360 --> 00:04:20.420 the key things seem to be price and accessibility. 00:04:20.560 --> 00:04:23.399 Yeah, absolutely. We're seeing these huge global 00:04:23.399 --> 00:04:26.399 expansion efforts happening like right now simultaneously. 00:04:27.600 --> 00:04:30.079 Both of the major players are rolling out these 00:04:30.079 --> 00:04:32.980 incredibly affordable subscription tiers, and 00:04:32.980 --> 00:04:34.899 they seem specifically aimed at markets with 00:04:34.899 --> 00:04:37.240 really high growth potential. Right. Google, 00:04:37.339 --> 00:04:39.180 for example, they just expanded their AI Plus 00:04:39.180 --> 00:04:42.639 subscription to 40 new countries. Indonesia was 00:04:42.639 --> 00:04:44.399 the first one they named in that big rollout. 00:04:44.420 --> 00:04:48.420 And it costs, what, around $4 .56 a month? Adjusted 00:04:48.420 --> 00:04:50.139 locally, of course. And look what they're packing 00:04:50.139 --> 00:04:52.240 into that. It's surprisingly generous for the 00:04:52.240 --> 00:04:55.319 price. You get access to Nano Banana. That's 00:04:55.319 --> 00:04:58.079 their new small model meant to run right on devices. 00:04:58.360 --> 00:05:01.459 You get VO3, their latest video generation model, 00:05:01.680 --> 00:05:06.160 Gemini 2 .5 Pro, and 200 gigs of storage. That's 00:05:06.160 --> 00:05:09.079 quite a bit. And OpenAI is right there with them, 00:05:09.100 --> 00:05:11.779 matching the pace almost exactly. Their cheapest 00:05:11.779 --> 00:05:14.920 tier, Chad Chippy Tigo. It's also live now in 00:05:14.920 --> 00:05:16.720 places like Indonesia. It launched in India first, 00:05:16.839 --> 00:05:19.079 I think. Costs about the same, around $4 .50 00:05:19.079 --> 00:05:23.860 a month. And again, surprisingly generous usage 00:05:23.860 --> 00:05:26.319 caps for people signing up. That strategy really 00:05:26.319 --> 00:05:28.620 changes the game, doesn't it? Low prices, easy 00:05:28.620 --> 00:05:32.899 access. It makes AI feel less like a luxury tool 00:05:32.899 --> 00:05:35.339 for corporations and more like a utility that's 00:05:35.339 --> 00:05:37.399 widely available globally. Yeah, and the ways 00:05:37.399 --> 00:05:39.040 people are using it are getting more integrated 00:05:39.040 --> 00:05:41.779 into just... Daily life. Like Google integrating 00:05:41.779 --> 00:05:44.819 Gemini AI straight into Google TV. That creates 00:05:44.819 --> 00:05:46.959 these genuinely interactive experiences. You 00:05:46.959 --> 00:05:49.180 mean you can literally talk to your TV now? Yeah. 00:05:49.220 --> 00:05:51.620 About what's on or search requests? Pretty much. 00:05:51.740 --> 00:05:53.680 Yeah. Conversing with your television. Okay. 00:05:53.720 --> 00:05:55.180 Now here's where it gets really interesting or 00:05:55.180 --> 00:05:57.339 maybe just amusing. We found this story about 00:05:57.339 --> 00:06:00.500 a woman who won $150 ,000 using lottery numbers 00:06:00.500 --> 00:06:04.420 generated by ChatGPT. Wow. Seriously? Seriously. 00:06:04.420 --> 00:06:07.279 And the best part. She plans to donate every 00:06:07.279 --> 00:06:10.079 single penny to charity, which is just a fantastic, 00:06:10.399 --> 00:06:13.019 completely unexpected application. That's amazing. 00:06:13.459 --> 00:06:16.000 And kind of random. And speaking of practical 00:06:16.000 --> 00:06:18.339 tools, somebody put together a GitHub repository. 00:06:18.800 --> 00:06:22.689 It's got over 90 really creative prompts. Specifically 00:06:22.689 --> 00:06:24.509 designed for that nano banana model we mentioned 00:06:24.509 --> 00:06:26.269 earlier. Yeah. So people are already building 00:06:26.269 --> 00:06:28.329 communities around these specialized tools. Past. 00:06:28.970 --> 00:06:30.949 Slight pause. You know, I still wrestle with 00:06:30.949 --> 00:06:33.129 prompt drift myself sometimes, trying to keep 00:06:33.129 --> 00:06:35.670 the AI's output consistent over a longer tat. 00:06:36.129 --> 00:06:39.649 It can be tricky. So those kinds of prompt guides, 00:06:39.850 --> 00:06:42.389 the ones that help stop the AI from slowly forgetting 00:06:42.389 --> 00:06:44.730 its original instructions, they're incredibly 00:06:44.730 --> 00:06:47.870 useful, actually. Yeah. Consistency is hard. 00:06:47.930 --> 00:06:50.670 And beyond the sort of fun consumer stuff, the 00:06:50.670 --> 00:06:53.509 real money is flowing fast into vertical AI. 00:06:53.689 --> 00:06:56.350 Yeah. Specific industries like CapitalRx, they 00:06:56.350 --> 00:06:59.350 just secured $400 million in funding. $400 million. 00:06:59.509 --> 00:07:02.589 Wow. For their AI -driven health benefits platform. 00:07:02.689 --> 00:07:05.639 They just rebranded it as Judy Health. That kind 00:07:05.639 --> 00:07:07.759 of massive investment, it shows enterprises really 00:07:07.759 --> 00:07:10.319 trust AI, but when it's wrapped in the right 00:07:10.319 --> 00:07:12.759 regulatory guardrails. We're also seeing tools 00:07:12.759 --> 00:07:14.420 just popping up everywhere for workflow automation. 00:07:14.620 --> 00:07:16.980 Like CX Reports, it personalizes data reports, 00:07:17.240 --> 00:07:20.240 no coding needed. Right. And Lookup, which takes 00:07:20.240 --> 00:07:22.660 raw video footage and turns it into structured 00:07:22.660 --> 00:07:25.579 answers, even verifiable proof clips. And the 00:07:25.579 --> 00:07:27.519 big picture confirms this isn't slowing down. 00:07:27.620 --> 00:07:30.939 The macro trend, OpenAI, Oracle, SoftBank, they're 00:07:30.939 --> 00:07:33.220 all expanding physically. Yeah. Building a combined 00:07:33.220 --> 00:07:37.220 five new... AI data centers. That signals deep, 00:07:37.300 --> 00:07:39.899 long -term infrastructure commitment worldwide. 00:07:40.339 --> 00:07:42.860 Okay. So if this affordability and accessibility 00:07:42.860 --> 00:07:46.560 are driving such rapid global growth, what's 00:07:46.560 --> 00:07:48.639 the sort of unexpected competitive impact? What 00:07:48.639 --> 00:07:52.819 is this low -cost expansion due to, say, local 00:07:52.819 --> 00:07:54.980 software development ecosystems in these new 00:07:54.980 --> 00:07:57.079 markets? Well, that low -cost access instantly 00:07:57.079 --> 00:07:59.600 ramps up the competitive pressure on local software 00:07:59.600 --> 00:08:01.839 companies, especially those not using AI yet. 00:08:01.920 --> 00:08:04.180 They have to adapt fast. All right, let's switch 00:08:04.180 --> 00:08:05.779 gears completely now. Let's talk about the speed 00:08:05.779 --> 00:08:08.079 of scientific discovery. Material science. Yeah. 00:08:08.339 --> 00:08:10.100 It's notoriously slow, right? Developing new 00:08:10.100 --> 00:08:11.600 materials for, you know, everything from car 00:08:11.600 --> 00:08:13.860 tires to medical devices. Yeah, it usually takes 00:08:13.860 --> 00:08:16.720 years, decades sometimes. Exactly. Years of slow, 00:08:16.879 --> 00:08:19.160 really expensive physical testing, tweaking, 00:08:19.160 --> 00:08:22.980 optimizing. But a team from Carnegie Mellon and 00:08:22.980 --> 00:08:26.220 UNC Chapel Hill just showed how AI can fundamentally 00:08:26.220 --> 00:08:30.079 change that whole timeline. They used this sophisticated 00:08:30.079 --> 00:08:32.559 human -in -the -loop model, and they created 00:08:32.559 --> 00:08:34.899 a polymer that's known for being incredibly difficult 00:08:34.899 --> 00:08:37.539 to make. Which one? One that manages to be both 00:08:37.539 --> 00:08:40.200 extremely strong and incredibly flexible at the 00:08:40.200 --> 00:08:43.179 same time. That's usually a trade -off. Ah, the 00:08:43.179 --> 00:08:45.340 classic strength versus flexibility problem. 00:08:45.539 --> 00:08:47.700 Okay. Okay. So this is that essential feedback 00:08:47.700 --> 00:08:50.419 loop working at its best then. The process kicked 00:08:50.419 --> 00:08:52.679 off with the human chemists setting these really 00:08:52.679 --> 00:08:55.299 high -level, ambitious goals. They basically 00:08:55.299 --> 00:08:57.700 asked the system, look, we want something that's 00:08:57.700 --> 00:09:00.159 super strong but also really stretchy. Right. 00:09:00.220 --> 00:09:02.919 And then the AI would suggest new chemical experiments, 00:09:03.179 --> 00:09:05.960 specific reaction conditions, all based on those 00:09:05.960 --> 00:09:08.460 complex targets. And the chemists, they could 00:09:08.460 --> 00:09:10.600 test these suggestions almost instantly using 00:09:10.600 --> 00:09:13.039 automated lab tools. And then comes the crucial 00:09:13.039 --> 00:09:16.019 part, the learning. The results go immediately 00:09:16.019 --> 00:09:18.649 back into the model. It constantly adjusts its 00:09:18.649 --> 00:09:21.309 strategy, learning what works, what fails in 00:09:21.309 --> 00:09:24.070 real time. It's like stacking Lego blocks of 00:09:24.070 --> 00:09:26.730 data, refining the recipe with every single test, 00:09:26.929 --> 00:09:29.529 building knowledge block by block. And the result, 00:09:29.610 --> 00:09:32.590 it's a really remarkable synthesis. This new 00:09:32.590 --> 00:09:35.669 polymer, it behaves like stretchy rubber. But 00:09:35.669 --> 00:09:38.450 at the same time, it holds the toughness you'd 00:09:38.450 --> 00:09:41.450 expect from tire -grade plastic. Wow. It's durable. 00:09:41.629 --> 00:09:44.909 It's highly adaptable. And it's even 3D printable, 00:09:45.070 --> 00:09:47.990 which just unlocks applications across manufacturing 00:09:47.990 --> 00:09:51.190 immediately. The efficiency gain here is just... 00:09:51.190 --> 00:09:53.870 Yeah. It's astounding. It was significantly cheaper, 00:09:53.970 --> 00:09:56.629 too, because the AI model basically skipped all 00:09:56.629 --> 00:09:58.889 the methods and chemical combinations that would 00:09:58.889 --> 00:10:01.889 have just failed inevitably in slow traditional 00:10:01.889 --> 00:10:05.399 testing. It avoids the dead ends. Exactly. Whoa. 00:10:05.600 --> 00:10:07.580 I mean, just imagine scaling this kind of precise 00:10:07.580 --> 00:10:09.639 predictive system. Imagine hitting it with like 00:10:09.639 --> 00:10:12.799 a billion potential material queries and finding 00:10:12.799 --> 00:10:15.139 complex solutions almost instantly. That approach, 00:10:15.240 --> 00:10:18.840 it just bypasses decades of chemists doing agonizing 00:10:18.840 --> 00:10:21.200 trial and error work. Yeah. It completely flips 00:10:21.200 --> 00:10:23.740 the research paradigm on its head. OK, so thinking 00:10:23.740 --> 00:10:26.559 about impact, what do you see is the biggest 00:10:26.559 --> 00:10:30.879 non -obvious impact of them open sourcing? this 00:10:30.879 --> 00:10:33.059 specific chemical model he developed. I think 00:10:33.059 --> 00:10:36.100 making this research tool available to everyone, 00:10:36.279 --> 00:10:39.139 it just democratizes complex material innovation 00:10:39.139 --> 00:10:42.799 for basically all labs globally, big or small. 00:10:43.000 --> 00:10:45.519 So wrapping this up, what does this all mean 00:10:45.519 --> 00:10:49.519 for you listening in? We saw AI making deployment 00:10:49.519 --> 00:10:52.200 incredibly simple and accessible with Amazon's 00:10:52.200 --> 00:10:54.840 agents. We saw it driving global accessibility 00:10:54.840 --> 00:10:57.220 through really aggressive, affordable pricing. 00:10:57.379 --> 00:10:59.879 And then we saw it fundamentally changing the 00:10:59.879 --> 00:11:02.360 sheer speed of scientific discovery with that 00:11:02.360 --> 00:11:04.860 revolutionary polymer breakthrough. Yeah, this 00:11:04.860 --> 00:11:07.059 is powerful convergence, isn't it? Simplicity, 00:11:07.200 --> 00:11:09.580 scale, and speed all hitting in these different 00:11:09.580 --> 00:11:11.360 domains at once. There's a final thought to chew 00:11:11.360 --> 00:11:14.539 on. If AI can instantly hack new polymers for, 00:11:14.620 --> 00:11:17.159 say, next -gen running shoes or vital medical 00:11:17.159 --> 00:11:20.379 devices, what massive core manufacturing industry 00:11:20.379 --> 00:11:22.659 is next? What's next for this kind of radical 00:11:22.659 --> 00:11:24.539 AI -driven redesign? That's something for you 00:11:24.539 --> 00:11:26.539 to consider. Mull it over as you internalize 00:11:26.539 --> 00:11:28.059 this and maybe apply some of this thinking to 00:11:28.059 --> 00:11:30.120 your own field. Thanks for diving deep with us 00:11:30.120 --> 00:11:31.559 today. We'll talk soon.