WEBVTT 00:00:00.000 --> 00:00:02.580 We always assume that, you know, staffing more 00:00:02.580 --> 00:00:05.500 intelligence, more compute, more AI agents onto 00:00:05.500 --> 00:00:08.359 a single task, that it just guarantees a better 00:00:08.359 --> 00:00:10.839 outcome. Right. More bots have to mean a faster, 00:00:10.900 --> 00:00:13.140 smarter solution, don't they? It just feels so 00:00:13.140 --> 00:00:15.500 intuitive, like you're just stacking these Lego 00:00:15.500 --> 00:00:17.980 blocks of data until you have a skyscraper of 00:00:17.980 --> 00:00:20.339 IQ. But it's not exactly like that. Some breakthrough 00:00:20.339 --> 00:00:24.710 research just confirmed this really... counterintuitive 00:00:24.710 --> 00:00:27.850 truth, which is that sometimes adding more bots 00:00:27.850 --> 00:00:30.230 just burns through your resources, your tokens, 00:00:30.370 --> 00:00:33.350 and actually makes the entire system worse. Dramatically 00:00:33.350 --> 00:00:37.469 worse. Yeah. Less is often truly more, especially 00:00:37.469 --> 00:00:41.189 when you're dealing with high logic tasks. Welcome 00:00:41.189 --> 00:00:44.899 to the deep dive. Today, we're unpacking a really 00:00:44.899 --> 00:00:47.359 fascinating stack of sources that clarify the 00:00:47.359 --> 00:00:49.600 current AI business landscape, and they reveal 00:00:49.600 --> 00:00:51.820 some critical hidden truths about how these complex 00:00:51.820 --> 00:00:54.219 systems actually work under the hood. And we're 00:00:54.219 --> 00:00:56.000 here to help you gain that knowledge quickly, 00:00:56.140 --> 00:00:58.460 but also thoroughly. We have three main areas 00:00:58.460 --> 00:01:00.200 for you today that you could really be informed 00:01:00.200 --> 00:01:02.740 on. First, we're going to look at the big monetization 00:01:02.740 --> 00:01:04.980 risks. If you try to build a business inside, 00:01:05.079 --> 00:01:07.900 you know, the biggest AI marketplace out there. 00:01:08.000 --> 00:01:10.879 Then, an update on new models. Yep. Blazing fast 00:01:10.879 --> 00:01:13.159 new models and practical tools you can use right 00:01:13.159 --> 00:01:15.579 now. We even have a really clever prompting technique 00:01:15.579 --> 00:01:18.140 that pretty much guarantees retention. And finally, 00:01:18.260 --> 00:01:20.680 that breakthrough study we mentioned, showing 00:01:20.680 --> 00:01:23.599 precisely why scaling up AI systems can sometimes 00:01:23.599 --> 00:01:28.219 backfire in a big way. Let's get into it. Okay, 00:01:28.280 --> 00:01:30.099 let's unpack this for you. We'll start with the 00:01:30.099 --> 00:01:31.959 business builders, the entrepreneurs out there 00:01:31.959 --> 00:01:33.540 trying to find a home for their applications. 00:01:34.040 --> 00:01:36.599 OpenAI recently clarified how you can monetize 00:01:36.599 --> 00:01:39.200 your apps within the ChatGPT marketplace. And 00:01:39.200 --> 00:01:41.640 it is a massive distribution channel. I mean, 00:01:41.640 --> 00:01:43.980 you can't deny that. Everyone is rushing toward 00:01:43.980 --> 00:01:46.180 it. Of course. They basically offer two paths 00:01:46.180 --> 00:01:48.359 if you want to make money there. The path that's 00:01:48.359 --> 00:01:50.480 supported right now is called external checkout. 00:01:50.700 --> 00:01:52.459 So that means you handle all the payments yourself 00:01:52.459 --> 00:01:55.379 off platform. You keep control. Exactly. But 00:01:55.379 --> 00:01:57.799 the path everyone really wants, the one that 00:01:57.799 --> 00:02:00.579 gets you right into that user flow, is instant 00:02:00.579 --> 00:02:03.150 checkout. That's the built -in payment system. 00:02:03.370 --> 00:02:05.349 Right, but it's currently in a private beta, 00:02:05.489 --> 00:02:08.889 only for a few select partners. Everyone wants 00:02:08.889 --> 00:02:11.750 to be there, but relying on that single channel, 00:02:11.949 --> 00:02:15.650 it's just really dangerous long -term. And here's 00:02:15.650 --> 00:02:17.409 where it gets interesting, because that massive 00:02:17.409 --> 00:02:19.870 user base comes with some severe trade -offs. 00:02:19.930 --> 00:02:22.669 Oh, yeah. The first major one is what we call 00:02:22.669 --> 00:02:25.810 brand dilution. Buyers start to attribute the 00:02:25.810 --> 00:02:29.599 purchase To chat GPT, not to you, the developer. 00:02:29.699 --> 00:02:31.479 You become invisible. You're just the engine 00:02:31.479 --> 00:02:33.840 behind the scenes. That's the classic marketplace 00:02:33.840 --> 00:02:36.360 squeeze, isn't it? When instant checkout eventually 00:02:36.360 --> 00:02:39.319 expands, you're just another storefront and you're 00:02:39.319 --> 00:02:41.960 probably going to get lost in the noise. You 00:02:41.960 --> 00:02:44.599 get the scale, sure, but you pay for it with 00:02:44.599 --> 00:02:47.780 zero brand loyalty with your actual user. Yeah, 00:02:47.800 --> 00:02:49.810 but hold on. If that channel gets you millions 00:02:49.810 --> 00:02:52.849 of users almost instantly, isn't zero brand loyalty 00:02:52.849 --> 00:02:54.909 worth it at the start? I mean, you're trading 00:02:54.909 --> 00:02:57.849 that long term brand equity for massive, immediate 00:02:57.849 --> 00:03:00.210 scale. Which is the goal for a startup launch, 00:03:00.349 --> 00:03:02.530 right? It is. But that's the Faustian bargain. 00:03:02.689 --> 00:03:05.310 The sources are really clear. The moment you 00:03:05.310 --> 00:03:07.990 rely on their discovery mechanism, you are trapped. 00:03:08.189 --> 00:03:10.789 Just like the early days of Google ads, you know. 00:03:11.419 --> 00:03:12.879 Eventually, you're going to have to pay to get 00:03:12.879 --> 00:03:15.560 seen. Whoever bids the most gets listed first. 00:03:15.819 --> 00:03:18.840 And that just pushes up bidding costs and kills 00:03:18.840 --> 00:03:21.259 any small profits for independent builders. And 00:03:21.259 --> 00:03:24.689 the liability risk? This is massive. And I think 00:03:24.689 --> 00:03:26.949 it's often overlooked by people building on these 00:03:26.949 --> 00:03:29.050 platforms. This is where the cost of convenience 00:03:29.050 --> 00:03:32.530 becomes kind of terrifying. Imagine you built 00:03:32.530 --> 00:03:35.310 a highly customized financial planning bot, right? 00:03:35.389 --> 00:03:38.729 It relies on this complex set of your own internal 00:03:38.729 --> 00:03:41.710 instructions. Okay. And now imagine the AI marketplace 00:03:41.710 --> 00:03:44.370 explains your product is slightly incorrectly. 00:03:45.050 --> 00:03:48.650 A subtle prompt drift happens and it misinterprets 00:03:48.650 --> 00:03:50.969 the volatility of an asset because of that bad 00:03:50.969 --> 00:03:53.330 description. That's not just an unhappy user 00:03:53.330 --> 00:03:56.289 anymore. That is real legal exposure that developers 00:03:56.289 --> 00:03:58.949 now have to insure against. You are liable for 00:03:58.949 --> 00:04:01.370 the AI's bad explanation of your product. Yeah, 00:04:01.449 --> 00:04:04.189 that liability issue. It really resonates. I 00:04:04.189 --> 00:04:05.830 mean, to be honest, I still wrestle with prompt 00:04:05.830 --> 00:04:08.030 drift myself sometimes, just trying to make sure 00:04:08.030 --> 00:04:09.969 my core instructions stay consistent, even on 00:04:09.969 --> 00:04:12.509 simple non -financial tasks. Of course. So imagine 00:04:12.509 --> 00:04:15.469 that struggle when real money and customer trust 00:04:15.469 --> 00:04:17.769 and maybe even legal action are on the line. 00:04:17.810 --> 00:04:19.850 It's a whole other level. And we also saw that 00:04:19.850 --> 00:04:21.629 some products just don't translate well. Anything 00:04:21.629 --> 00:04:24.129 that relies on, say, visual or emotional input. 00:04:24.209 --> 00:04:26.310 The sources mentioned things like fashion, beauty 00:04:26.310 --> 00:04:28.569 or home decor. It just doesn't sell well through 00:04:28.569 --> 00:04:31.870 a chat interface. At the end of the day. ChatGPT 00:04:31.870 --> 00:04:34.670 sells convenience. It doesn't necessarily sell 00:04:34.670 --> 00:04:38.170 your complex, nuanced company. Right. So based 00:04:38.170 --> 00:04:40.709 on all this, what's the single biggest strategic 00:04:40.709 --> 00:04:42.990 shift a developer should be making right now 00:04:42.990 --> 00:04:45.629 to secure some long -term sustainability outside 00:04:45.629 --> 00:04:47.930 of that walled garden? They have to prioritize 00:04:47.930 --> 00:04:50.569 building their own distribution, their own brand 00:04:50.569 --> 00:04:52.949 presence, totally separate from the platform's 00:04:52.949 --> 00:04:55.709 convenience. All right, let's pivot a bit to 00:04:55.709 --> 00:04:57.649 the relentless pace of development, to speed. 00:04:58.110 --> 00:04:59.970 I mean, we're moving so fast now that entire 00:04:59.970 --> 00:05:02.310 model generations are obsolete in just a few 00:05:02.310 --> 00:05:05.389 months. Google just dropped Gemini 3 Flash. And 00:05:05.389 --> 00:05:08.290 as the name suggests, it is blazing fast. It's 00:05:08.290 --> 00:05:10.949 designed for these high volume, low latency tasks. 00:05:11.670 --> 00:05:13.889 What's impressive is that it actually beats the 00:05:13.889 --> 00:05:16.389 higher tier pro version in some coding benchmarks. 00:05:17.040 --> 00:05:19.259 So speed isn't necessarily sacrificing quality 00:05:19.259 --> 00:05:21.740 anymore. Exactly. And it's free to try right 00:05:21.740 --> 00:05:23.860 now, which just lowers that barrier to entry 00:05:23.860 --> 00:05:26.180 for everyone. And the financial velocity around 00:05:26.180 --> 00:05:29.620 this speed race is, it's just staggering. Look 00:05:29.620 --> 00:05:33.339 at Databricks. They just raised $4 billion at 00:05:33.339 --> 00:05:38.519 $134 billion valuation. And that's up 34 % in 00:05:38.519 --> 00:05:40.899 just three months. Three months. That growth 00:05:40.899 --> 00:05:43.839 rate is both terrifying and exhilarating. Whoa! 00:05:44.399 --> 00:05:46.500 I mean, just imagine scaling that infrastructure 00:05:46.500 --> 00:05:49.699 to a billion queries a day. The market is validating 00:05:49.699 --> 00:05:52.040 that speed and integration are now the critical 00:05:52.040 --> 00:05:54.139 infrastructure for this next phase of the Internet. 00:05:54.319 --> 00:05:57.000 Which brings us to how we, the users, should 00:05:57.000 --> 00:05:58.819 be interacting with these tools that are getting 00:05:58.819 --> 00:06:02.079 faster and faster. This velocity highlights a 00:06:02.079 --> 00:06:04.730 crucial point for you, the learner. You can't 00:06:04.730 --> 00:06:07.790 just let AI info dump on you. You've got to maximize 00:06:07.790 --> 00:06:09.790 the knowledge transfer from every interaction. 00:06:10.050 --> 00:06:12.149 Exactly. We found this really practical technique 00:06:12.149 --> 00:06:14.850 called the Feynman loop prompt. This concept 00:06:14.850 --> 00:06:17.889 is brilliant. It forces true understanding, and 00:06:17.889 --> 00:06:19.910 it guarantees you'll retain so much more than 00:06:19.910 --> 00:06:22.670 if you just read a summary. You tell the AI to 00:06:22.670 --> 00:06:25.009 teach you a topic. And then it tests you. It 00:06:25.009 --> 00:06:27.810 continuously tests you and teaches you, iterating 00:06:27.810 --> 00:06:29.949 over and over until you can successfully teach 00:06:29.949 --> 00:06:32.379 the concept back to the AI. So, for example, 00:06:32.480 --> 00:06:35.500 you can start with a prompt like, AI, teach me 00:06:35.500 --> 00:06:38.019 quantum entanglement using the Feynman loop methodology. 00:06:38.439 --> 00:06:41.259 And you add validating my understanding by requiring 00:06:41.259 --> 00:06:43.639 me to teach it back to you, broken down into 00:06:43.639 --> 00:06:46.660 five simple stages. And that shift from just 00:06:46.660 --> 00:06:50.139 passively consuming to actively teaching, it 00:06:50.139 --> 00:06:52.879 fundamentally changes how you learn. You should 00:06:52.879 --> 00:06:55.139 probably save that prompt forever. We're also 00:06:55.139 --> 00:06:57.079 seeing this specialization accelerating across 00:06:57.079 --> 00:06:59.939 practical tools. SEMrush, the marketing intelligence 00:06:59.939 --> 00:07:03.160 giant, is now operating inside ChatGPT. So marketers 00:07:03.160 --> 00:07:05.620 can automate complex reports or analyze competitors 00:07:05.620 --> 00:07:08.379 just by typing simple commands. And others are 00:07:08.379 --> 00:07:11.079 combining models. You see tools like AirOps combining 00:07:11.079 --> 00:07:14.240 40 or more different specialized models to handle 00:07:14.240 --> 00:07:17.439 these niche creative growth tasks. And Alibaba's 00:07:17.439 --> 00:07:21.639 WAN 2 .6 is creating these 15 -second 1080p videos. 00:07:21.959 --> 00:07:24.800 with multiple connected shots. The era of the 00:07:24.800 --> 00:07:27.399 single monolithic AI is fading really fast. It's 00:07:27.399 --> 00:07:29.540 moving toward these highly specific utility tools. 00:07:29.779 --> 00:07:32.319 So given all this speed, the question becomes, 00:07:32.500 --> 00:07:35.819 how do we even measure true intelligence now? 00:07:36.319 --> 00:07:38.800 OpenAI just introduced something called the Frontier 00:07:38.800 --> 00:07:41.199 Science Benchmark. And this goes way beyond just 00:07:41.199 --> 00:07:43.860 synthesizing data or completing some coding tasks. 00:07:44.360 --> 00:07:46.600 This benchmark specifically challenges models 00:07:46.600 --> 00:07:50.240 to, say, hypothesize novel chemical reactions 00:07:50.240 --> 00:07:52.740 or predict entirely new molecular structures. 00:07:52.879 --> 00:07:54.980 Or even suggest revolutionary breakthroughs in 00:07:54.980 --> 00:07:57.720 material science. It's measuring if AI can perform 00:07:57.720 --> 00:08:00.519 actual scientific discovery. The results are 00:08:00.519 --> 00:08:02.819 promising. They show huge potential. But, you 00:08:02.819 --> 00:08:05.620 know, skepticism is still pretty high. It's a 00:08:05.620 --> 00:08:07.800 massive step, though, toward proving that AI 00:08:07.800 --> 00:08:10.160 can innovate, not just summarize old papers. 00:08:10.889 --> 00:08:14.110 So how do these rapid and, let's be honest, complex 00:08:14.110 --> 00:08:16.509 benchmarks like frontier science really help 00:08:16.509 --> 00:08:19.370 the average user or builder who isn't a research 00:08:19.370 --> 00:08:21.870 chemist? Well, benchmarks clearly show the limitations 00:08:21.870 --> 00:08:24.129 and the potential. They guide us on which high 00:08:24.129 --> 00:08:26.230 -stakes tasks are actually safe for an AI to 00:08:26.230 --> 00:08:28.189 execute and which ones still need that human 00:08:28.189 --> 00:08:30.509 oversight. Which brings us directly to the heart 00:08:30.509 --> 00:08:33.289 of the research, those counterintuitive findings 00:08:33.289 --> 00:08:37.049 on multi -agent complexity. This is crucial for 00:08:37.049 --> 00:08:38.909 anyone thinking about building sophisticated 00:08:38.909 --> 00:08:41.730 AI workflows or enterprise solutions. Right. 00:08:41.850 --> 00:08:44.110 Multi -agent systems are incredibly popular right 00:08:44.110 --> 00:08:47.100 now. The idea is simple. It's a team of specialized 00:08:47.100 --> 00:08:50.019 AI bots working together on one complex task. 00:08:50.299 --> 00:08:52.299 And the assumption has always been pretty straightforward. 00:08:52.659 --> 00:08:55.460 More specialized agents means better, faster 00:08:55.460 --> 00:08:58.120 results. But a joint study from Google and MIT 00:08:58.120 --> 00:09:00.899 ran 180 experiments across different major models 00:09:00.899 --> 00:09:03.500 to rigorously test that foundational theory. 00:09:03.740 --> 00:09:06.960 Does adding more AI agents actually inherently 00:09:06.960 --> 00:09:09.750 improve the outcome? And the answer is, well, 00:09:09.850 --> 00:09:12.029 it's absolutely nuanced. For tasks where you 00:09:12.029 --> 00:09:14.470 can easily split data across agents and think 00:09:14.470 --> 00:09:16.970 about financial data analysis or processing massive 00:09:16.970 --> 00:09:19.330 batches of documents, they saw an incredible 00:09:19.330 --> 00:09:21.950 81 % improvement. That's awesome. Great performance 00:09:21.950 --> 00:09:24.769 scaling. Amazing. But for tasks that require 00:09:24.769 --> 00:09:27.629 sequential step -by -step reasoning. I'm thinking 00:09:27.629 --> 00:09:30.009 of it like trying to assemble a complex piece 00:09:30.009 --> 00:09:33.539 of IKEA furniture. If you ask... 10 different 00:09:33.539 --> 00:09:37.559 specialized people to each do one step, but without 00:09:37.559 --> 00:09:40.220 perfectly clear instant communication. You just 00:09:40.220 --> 00:09:43.639 end up with a huge mess. A huge mess. That is 00:09:43.639 --> 00:09:45.419 exactly what happened in the experiments. For 00:09:45.419 --> 00:09:47.539 these high logic planning tasks, performance 00:09:47.539 --> 00:09:51.440 dropped by up to 70%. 70%. That's a catastrophic 00:09:51.440 --> 00:09:54.139 failure rate for system complexity and cost. 00:09:54.379 --> 00:09:57.000 It is. It just completely flips the entire assumption 00:09:57.000 --> 00:10:00.000 we started with on its head. It means blindly 00:10:00.000 --> 00:10:02.659 throwing more compute at a complex problem is 00:10:02.659 --> 00:10:05.340 actually worse than just having one dedicated, 00:10:05.600 --> 00:10:07.759 highly trained specialist. So why did it fail? 00:10:07.820 --> 00:10:09.700 What's the mechanism? It's both economic and 00:10:09.700 --> 00:10:11.840 technical. The multi -agent setups just burn 00:10:11.840 --> 00:10:14.250 through tokens. And we should remember, tokens 00:10:14.250 --> 00:10:17.309 are the fundamental economic unit of AI. They 00:10:17.309 --> 00:10:19.269 represent the computational weight you pay for. 00:10:19.429 --> 00:10:22.950 They duplicated reasoning steps. They overcomplicated 00:10:22.950 --> 00:10:25.629 workflows that really demanded singular, high 00:10:25.629 --> 00:10:29.230 -precision logic. So in essence, they wasted 00:10:29.230 --> 00:10:32.110 huge amounts of computational weight, which translates 00:10:32.110 --> 00:10:34.850 directly into massively increased operating costs 00:10:34.850 --> 00:10:38.970 for absolutely zero benefit. Zero. This confirms 00:10:38.970 --> 00:10:42.559 that finding. More AI does not equal more IQ. 00:10:42.940 --> 00:10:45.759 And this is such a crucial takeaway for any builder 00:10:45.759 --> 00:10:48.299 out there. For high logic or long -term memory 00:10:48.299 --> 00:10:51.940 tasks, a single, finely -tuned agent can just 00:10:51.940 --> 00:10:56.000 crush a fancy, complex, multi -agent team. A 00:10:56.000 --> 00:10:58.139 lot of these agent stacks are just introducing 00:10:58.139 --> 00:11:00.860 fragility and cost instead of actually solving 00:11:00.860 --> 00:11:03.720 hard problems. So based on this, should builders 00:11:03.720 --> 00:11:06.320 just abandon multi -agent systems entirely? Is 00:11:06.320 --> 00:11:08.179 that the takeaway? No, not at all. They still 00:11:08.179 --> 00:11:10.399 excel in that massive parallel data splitting 00:11:10.399 --> 00:11:13.039 we talked about, but they clearly fail in high 00:11:13.039 --> 00:11:15.820 logic precision chains where one weak link can 00:11:15.820 --> 00:11:17.759 just spoil the whole process. Okay, so what does 00:11:17.759 --> 00:11:19.799 all this mean for you as you integrate this knowledge 00:11:19.799 --> 00:11:22.019 into your work or just your curiosity? We saw 00:11:22.019 --> 00:11:24.519 three powerful connected forces at play today. 00:11:24.909 --> 00:11:27.490 First, there's the marketplace risk. That instant 00:11:27.490 --> 00:11:29.509 distribution comes at the cost of your brand 00:11:29.509 --> 00:11:32.830 identity and potentially immense liability. You've 00:11:32.830 --> 00:11:34.690 got to build your brand outside the convenience 00:11:34.690 --> 00:11:37.669 of that walled platform. Second, tool speed and 00:11:37.669 --> 00:11:40.529 utility. Models are faster than ever, driven 00:11:40.529 --> 00:11:44.350 by these massive financial valuations. But that 00:11:44.350 --> 00:11:47.370 speed is useless if you don't maximize the knowledge 00:11:47.370 --> 00:11:49.970 transfer. So, you know, use the Feynman loop 00:11:49.970 --> 00:11:52.529 to guarantee retention. And finally, the paradox 00:11:52.529 --> 00:11:56.399 of scale. Complexity kills precision. Adding 00:11:56.399 --> 00:11:58.820 more agents doesn't guarantee a higher IQ. It 00:11:58.820 --> 00:12:01.659 often just raises your token costs and introduces 00:12:01.659 --> 00:12:04.559 these crippling failure points in logic -driven 00:12:04.559 --> 00:12:07.500 tasks. And this entire landscape is just shifting 00:12:07.500 --> 00:12:10.159 so quickly. Yeah, we even noted the intense debate 00:12:10.159 --> 00:12:13.019 over spotting AI content. The sources mentioned 00:12:13.019 --> 00:12:15.990 five reliable ways besides just... looking for 00:12:15.990 --> 00:12:18.690 the prevalence of emdashes. That's a whole discussion 00:12:18.690 --> 00:12:20.870 that probably deserves its own deep dive. We'll 00:12:20.870 --> 00:12:22.490 save that for end of the day. Yeah. But here's 00:12:22.490 --> 00:12:24.210 our final provocative thought for you to mull 00:12:24.210 --> 00:12:27.250 over. If the Feynman loop prompt is so incredibly 00:12:27.250 --> 00:12:29.850 effective for teaching humans and ensuring validated 00:12:29.850 --> 00:12:32.470 understanding, what would happen if we applied 00:12:32.470 --> 00:12:35.090 that same technique, requiring the AI to prove 00:12:35.090 --> 00:12:36.970 its depth of understanding to the foundational 00:12:36.970 --> 00:12:39.929 models themselves? Could that finally solve the 00:12:39.929 --> 00:12:42.149 issue of AI hallucination and the prompt drift 00:12:42.149 --> 00:12:43.750 we were wrestling with in the first segment? 00:12:44.090 --> 00:12:47.330 I mean, think about that potential. Deeply validated, 00:12:47.509 --> 00:12:50.610 reliable understanding rather than just probabilistic 00:12:50.610 --> 00:12:54.169 synthesis. It could be the key to true AI reliability 00:12:54.169 --> 00:12:57.409 across every single application. We really encourage 00:12:57.409 --> 00:12:59.970 you to explore that profound idea further. We 00:12:59.970 --> 00:13:01.629 appreciate you taking this deep dive with us. 00:13:01.669 --> 00:13:03.330 Until next time, keep exploring the sources. 00:13:03.549 --> 00:13:04.490 Out to your own music.