WEBVTT 00:00:00.000 --> 00:00:02.459 Imagine you walk into a car dealership. You're 00:00:02.459 --> 00:00:04.759 there to buy the most affordable, sensible car 00:00:04.759 --> 00:00:07.179 on the lot, a basic commuter model, you know, 00:00:07.200 --> 00:00:09.820 something practical. Then someone hands you the 00:00:09.820 --> 00:00:13.419 track results. And this everyday sedan just clocked 00:00:13.419 --> 00:00:15.980 faster times than the limited edition carbon 00:00:15.980 --> 00:00:18.500 fiber flagship sports car, the one that costs 00:00:18.500 --> 00:00:21.280 four times as much. That seemingly impossible 00:00:21.280 --> 00:00:24.239 scenario where budget just flat out beats flagship. 00:00:25.100 --> 00:00:27.140 That is exactly what just happened in the world 00:00:27.140 --> 00:00:30.320 of AI coding. A fast, incredibly cheap model 00:00:30.320 --> 00:00:32.700 just officially upset its own higher -priced 00:00:32.700 --> 00:00:35.240 sibling at solving real -world messy engineering 00:00:35.240 --> 00:00:37.700 problems. Welcome to the Deep Dive. And this 00:00:37.700 --> 00:00:39.899 isn't just an upgrade. It's more like a category 00:00:39.899 --> 00:00:41.939 collapse. We're diving deep into what this all 00:00:41.939 --> 00:00:45.159 means. The topic today is Gemini 3 Flash. It 00:00:45.159 --> 00:00:47.939 was marketed as the fast, cheap option, you know, 00:00:47.939 --> 00:00:50.140 for quick, low -cost tasks. Right, the simple 00:00:50.140 --> 00:00:52.600 stuff. Exactly. But for developers and builders, 00:00:52.799 --> 00:00:55.259 Flash has rapidly become... the new default. 00:00:55.420 --> 00:00:57.780 It's forcing everyone from solo founders to big 00:00:57.780 --> 00:01:00.119 tech teams to completely rethink what budget 00:01:00.119 --> 00:01:02.579 AI actually means and where to spend those precious 00:01:02.579 --> 00:01:05.459 token budgets. Our mission today is to cut through 00:01:05.459 --> 00:01:08.000 all that marketing noise. We're going to analyze 00:01:08.000 --> 00:01:10.599 the performance scores, which are pretty shocking. 00:01:11.069 --> 00:01:13.909 will reveal its secret weapon called dynamic 00:01:13.909 --> 00:01:17.170 thinking and run through four crucial real -world 00:01:17.170 --> 00:01:19.250 code tests. And then we'll get into the playbook, 00:01:19.349 --> 00:01:21.329 how to actually use it, and maybe more importantly, 00:01:21.530 --> 00:01:24.870 what to avoid. Okay, let's unpack this. So DeepMind 00:01:24.870 --> 00:01:27.930 releases Gemini 3 Flash, and they position it 00:01:27.930 --> 00:01:31.290 very clearly beneath their flagship Pro model. 00:01:31.430 --> 00:01:34.129 And historically, that kind of naming implies 00:01:34.129 --> 00:01:39.230 a predictable trade -off. Flash means fast, but 00:01:39.230 --> 00:01:42.920 maybe a little dumb. Pro means smart, but slow 00:01:42.920 --> 00:01:46.079 and expensive. That was the rule. And Flash just 00:01:46.079 --> 00:01:48.719 fundamentally broke it. And in doing so, it really 00:01:48.719 --> 00:01:51.159 disrupted the entire pricing structure of high 00:01:51.159 --> 00:01:53.140 -level AI. And we have the concrete evidence 00:01:53.140 --> 00:01:55.620 for this. To really quantify how big a deal this 00:01:55.620 --> 00:01:57.659 is, we have to look at the benchmark that coders 00:01:57.659 --> 00:02:00.620 actually care about. SW bench verified. This 00:02:00.620 --> 00:02:02.920 isn't just some academic test, right? Not at 00:02:02.920 --> 00:02:05.340 all. It's not multiple choice. It forces the 00:02:05.340 --> 00:02:07.980 AI to solve real world GitHub issues. We're talking 00:02:07.980 --> 00:02:11.319 actual open source bugs, feature requests, messy 00:02:11.319 --> 00:02:15.000 code, stuff pulled from real projects. It's basically 00:02:15.000 --> 00:02:17.229 an honest job interview for the model. And the 00:02:17.229 --> 00:02:19.289 scoreboard from that interview is what shocked 00:02:19.289 --> 00:02:22.250 everyone. It really is. So Gemini 3 Flash, using 00:02:22.250 --> 00:02:26.810 its new thinking configuration, scored a 78 .0%. 00:02:26.810 --> 00:02:29.930 Wow. Yeah, that's a huge figure for real -world 00:02:29.930 --> 00:02:32.990 problem solving. Now get this. Its older, more 00:02:32.990 --> 00:02:36.449 expensive sibling, Gemini 3 Pro, scored 76 .2%. 00:02:36.449 --> 00:02:39.189 So it's actually better. It is measurably better. 00:02:39.330 --> 00:02:43.289 And Claude Sonnet 4 .5 scored 77 .2%. Now, technically, 00:02:43.389 --> 00:02:46.330 GPT -5 .2 did nudge ahead at 80%, but there's 00:02:46.330 --> 00:02:48.370 some critical context here that makes Flash the, 00:02:48.389 --> 00:02:50.800 uh... Best practical model. Hold on a second, 00:02:50.819 --> 00:02:52.379 though. The difference between Flash and Pro, 00:02:52.539 --> 00:02:54.759 that's less than two points. Shouldn't we assume 00:02:54.759 --> 00:02:56.740 that small gap represents the really tricky edge 00:02:56.740 --> 00:02:58.800 cases making Pro still necessary for critical 00:02:58.800 --> 00:03:01.240 systems? That's a really fair point. And yes, 00:03:01.360 --> 00:03:03.240 that small percentage probably does represent 00:03:03.240 --> 00:03:06.280 problems only Pro's depth can handle. But that's 00:03:06.280 --> 00:03:08.419 where the cost comes in. That skepticism just 00:03:08.419 --> 00:03:10.280 evaporates when you look at the price tag. Right. 00:03:10.379 --> 00:03:13.520 We are seeing an economic collapse for high -level 00:03:13.520 --> 00:03:15.360 coding. Right. I mean, the cost difference is 00:03:15.360 --> 00:03:18.780 massive. Pro costs $2 per one. million input 00:03:18.780 --> 00:03:22.560 tokens. Flash costs just 50 cents. So the model 00:03:22.560 --> 00:03:24.960 that is measurably smarter on this key benchmark 00:03:24.960 --> 00:03:28.360 is four times cheaper to run. That's astonishing. 00:03:28.919 --> 00:03:31.340 Precisely. You can run four complete coding iterations, 00:03:31.520 --> 00:03:33.740 four full attempts at solving a problem with 00:03:33.740 --> 00:03:36.280 Flash for the exact same price as a single run 00:03:36.280 --> 00:03:39.180 with Pro. The cost per useful output just dropped 00:03:39.180 --> 00:03:42.110 off a cliff. How does the shift in cost versus 00:03:42.110 --> 00:03:44.289 capability fundamentally change the starting 00:03:44.289 --> 00:03:46.310 point for startups and solo builders who are 00:03:46.310 --> 00:03:48.849 constantly battling their cloud bill? High -level 00:03:48.849 --> 00:03:51.349 AI coding is now affordable for everyone, allowing 00:03:51.349 --> 00:03:53.889 faster, cheaper iteration and experimentation. 00:03:54.370 --> 00:03:57.210 So a nearly flagship model that's four times 00:03:57.210 --> 00:03:59.729 cheaper. How do they do that without killing 00:03:59.729 --> 00:04:02.610 the speed? The sources all point to this new 00:04:02.610 --> 00:04:06.039 feature, dynamic thinking. Which, let's be honest, 00:04:06.139 --> 00:04:08.560 sounds a bit like marketing jargon. It absolutely 00:04:08.560 --> 00:04:11.219 is marketing. But it also happens to describe 00:04:11.219 --> 00:04:14.340 a real architectural change. Oh. Dynamic thinking 00:04:14.340 --> 00:04:20.660 forces the model to pause, plan. and internally 00:04:20.660 --> 00:04:23.920 reason about the task before it generates a single 00:04:23.920 --> 00:04:26.639 line of code. An internal monologue. It's a mandatory 00:04:26.639 --> 00:04:28.959 internal monologue that happens in a single API 00:04:28.959 --> 00:04:31.699 call. And you, the developer, can actually influence 00:04:31.699 --> 00:04:33.759 how much thought it puts in. And that's critical. 00:04:33.899 --> 00:04:35.680 The old way was so frustrating. You'd ask for 00:04:35.680 --> 00:04:37.860 something complex, like a snake game, and the 00:04:37.860 --> 00:04:40.019 AI would just immediately spit out code. Right. 00:04:40.160 --> 00:04:41.819 And it would almost always have that one fundamental 00:04:41.819 --> 00:04:44.459 bug, like the menu doesn't disappear or the collision 00:04:44.459 --> 00:04:46.639 logic is just slightly off. The old model was 00:04:46.639 --> 00:04:49.279 just guessing the next word. The new way. with 00:04:49.279 --> 00:04:51.779 Flash is totally different. You ask for the snake 00:04:51.779 --> 00:04:53.720 game, and internally, before writing anything, 00:04:53.839 --> 00:04:56.279 it's planning. Okay, I need a game loop, a grid, 00:04:56.459 --> 00:04:59.860 handle input, detect wall collisions, use Pygame, 00:05:00.100 --> 00:05:02.540 and then it generates the structured and usually 00:05:02.540 --> 00:05:04.959 bug -free code. And we're seeing that planning 00:05:04.959 --> 00:05:08.040 ability shine in really serious real work scenarios, 00:05:08.360 --> 00:05:10.819 not just little functions. I mean, refactoring 00:05:10.819 --> 00:05:14.360 a messy 800 line legacy file into clean modules 00:05:14.360 --> 00:05:17.439 or debugging a failing off flow where the problem 00:05:17.439 --> 00:05:20.339 isn't obvious. You know, honestly, I still wrestle 00:05:20.339 --> 00:05:22.439 with prompt drift myself sometimes where you 00:05:22.439 --> 00:05:25.399 prompt the AI and by the third turn, it's completely 00:05:25.399 --> 00:05:27.639 forgotten the original constraints. Yeah, that's 00:05:27.639 --> 00:05:29.759 a common struggle. So seeing the model handle 00:05:29.759 --> 00:05:31.779 that internal planning, especially with constraints 00:05:31.779 --> 00:05:34.839 like build rate limiting for 10K requests per 00:05:34.839 --> 00:05:37.560 second and you have to use Redis, it's a massive 00:05:37.560 --> 00:05:39.720 relief. If the model is handling this internal 00:05:39.720 --> 00:05:42.459 structured planning, what is the single biggest 00:05:42.459 --> 00:05:45.120 benefit for a developer's daily workflow? The 00:05:45.120 --> 00:05:47.439 model stops being a simple snippet machine and 00:05:47.439 --> 00:05:49.379 starts acting like an efficient, guided technical 00:05:49.379 --> 00:05:52.000 assistant. Okay, benchmarks are great for headlines, 00:05:52.300 --> 00:05:55.560 but what about production? We need to know if 00:05:55.560 --> 00:05:57.779 Flash is actually stable and useful when the 00:05:57.779 --> 00:06:00.579 pressure is on. So let's review these four gauntlet 00:06:00.579 --> 00:06:03.399 tests. Test number one was all about speed latency 00:06:03.399 --> 00:06:06.040 under pressure. And the prompt was surprisingly 00:06:06.040 --> 00:06:10.399 complex. Create a single file HTML 3 .js scene 00:06:10.399 --> 00:06:14.660 of a cozy, softly lit living room with an animated 00:06:14.660 --> 00:06:18.689 Tom and Jerry SVG loop playing on a 3D TV. Wow. 00:06:18.829 --> 00:06:21.069 So that's testing graphics, animation, library 00:06:21.069 --> 00:06:23.550 knowledge all at once. Yeah. And Flash was faster 00:06:23.550 --> 00:06:26.350 than the last generation. Gemini 2 .5 Pro under 00:06:26.350 --> 00:06:29.930 30 seconds versus 47 seconds. And that low latency 00:06:29.930 --> 00:06:32.029 is so important. It's the difference between 00:06:32.029 --> 00:06:33.949 an assistant that keeps you in the flow and one 00:06:33.949 --> 00:06:35.370 that just, you know, makes you wait and gets 00:06:35.370 --> 00:06:36.790 annoying. Right. You lose your train of thought. 00:06:36.930 --> 00:06:39.810 And here's the surprise. Flash's result was actually 00:06:39.810 --> 00:06:43.529 better than 2 .5 Pro and even 3 Pro. Though, 00:06:43.629 --> 00:06:45.910 I have to point out, the testers did note one 00:06:45.910 --> 00:06:48.720 clear flaw. What was that? There was no TV stand. 00:06:48.879 --> 00:06:50.779 The television was just levitating in the middle 00:06:50.779 --> 00:06:53.699 of the room. Ah, the classic floating television 00:06:53.699 --> 00:06:56.519 problem. That tells you everything, doesn't it? 00:06:56.579 --> 00:06:59.220 It nailed the complex rendering and animation, 00:06:59.420 --> 00:07:02.300 but forgot basic physics. Still needs a human 00:07:02.300 --> 00:07:05.019 editor to remember gravity. Exactly. Then there 00:07:05.019 --> 00:07:08.259 was the stress test, combining complex math with 00:07:08.259 --> 00:07:11.000 animation. Right. This one was a 3D visualization 00:07:11.000 --> 00:07:14.819 of relative scale from a subatomic particle all 00:07:14.819 --> 00:07:18.050 the way up to a galaxy. It demands math, physics, 00:07:18.329 --> 00:07:21.889 animation, JavaScript, all working together perfectly. 00:07:22.110 --> 00:07:23.990 And Flash delivered the result in the shortest 00:07:23.990 --> 00:07:26.750 time. It was great. Though the highest -end model, 00:07:26.949 --> 00:07:29.449 3 Pro, did have a bit more polish on the final 00:07:29.449 --> 00:07:31.430 result. Flash just prioritized being correct 00:07:31.430 --> 00:07:33.709 and fast. And that brings us to test number four, 00:07:33.790 --> 00:07:35.889 which for me is the most impressive one, the 00:07:35.889 --> 00:07:38.889 one -shot voxel art test, the Eagle test. This 00:07:38.889 --> 00:07:41.110 prompt asks the model to write voxel art code 00:07:41.110 --> 00:07:43.290 for an eagle sitting on a branch in a single 00:07:43.290 --> 00:07:45.879 HTML file. And voxel art is really hard because 00:07:45.879 --> 00:07:48.319 you have to manually define the 3D space, coordinate 00:07:48.319 --> 00:07:51.379 by coordinate. Whoa, just imagine scaling a model 00:07:51.379 --> 00:07:54.199 that can handle that level of creativity, abstract 00:07:54.199 --> 00:07:56.779 spatial reasoning, and obscure library knowledge 00:07:56.779 --> 00:08:00.060 in one shot. Correctly defining the relationship 00:08:00.060 --> 00:08:02.360 between the eagle and the branch in 3D space. 00:08:02.660 --> 00:08:06.120 That's a fundamentally new capability for a so 00:08:06.120 --> 00:08:08.480 -called budget model. So the summary from all 00:08:08.480 --> 00:08:10.879 these tests is pretty clear. Flash prioritizes 00:08:10.879 --> 00:08:13.949 speed. correctness and just pure practicality 00:08:13.949 --> 00:08:16.730 pro is still there for depth creative polish 00:08:16.730 --> 00:08:19.550 and complex framing but flash hits that sweet 00:08:19.550 --> 00:08:22.290 spot for just getting things built beyond the 00:08:22.290 --> 00:08:25.009 raw scores and the cool animations what single 00:08:25.009 --> 00:08:27.850 practical outcome makes flash feel truly production 00:08:27.850 --> 00:08:30.050 ready right now It's fast enough to integrate 00:08:30.050 --> 00:08:32.250 directly into your flow while being smart enough 00:08:32.250 --> 00:08:34.990 to confidently handle surprisingly complex multi 00:08:34.990 --> 00:08:37.970 -step tasks. Okay, knowing the models is great 00:08:37.970 --> 00:08:40.129 is one thing. Deploying it without setting your 00:08:40.129 --> 00:08:42.149 credit card on fire requires a real strategy. 00:08:42.409 --> 00:08:44.690 So let's talk playbook. First thing, the API 00:08:44.690 --> 00:08:47.509 parameter. It's vital for cost control. Gemini 00:08:47.509 --> 00:08:50.029 3 models use thinking levels instead of the old 00:08:50.029 --> 00:08:51.889 difficult thinking budget where you had to guess 00:08:51.889 --> 00:08:54.049 how many tokens it needed. Right. Thinking levels 00:08:54.049 --> 00:08:56.529 makes it way simpler. For simple chat, you'd 00:08:56.529 --> 00:08:59.750 use minimal. But for coding, the key is to always 00:08:59.750 --> 00:09:02.730 use thinking level high. This forces it to do 00:09:02.730 --> 00:09:04.889 that internal planning. And this is where we 00:09:04.889 --> 00:09:07.289 have to issue a strong warning. This is the minimal 00:09:07.289 --> 00:09:10.149 trap. You can't actually completely disable the 00:09:10.149 --> 00:09:13.149 thinking. Even on minimal, if the model thinks 00:09:13.149 --> 00:09:16.179 a prompt is tricky. it might still generate those 00:09:16.179 --> 00:09:18.620 reasoning tokens, the thinking trace. So your 00:09:18.620 --> 00:09:20.440 application logic has to be ready for that. If 00:09:20.440 --> 00:09:23.100 your code expects just raw text and it gets the 00:09:23.100 --> 00:09:25.519 model's internal monologue mixed in, your app 00:09:25.519 --> 00:09:27.440 could crash. You have to be ready to parse that. 00:09:27.679 --> 00:09:30.580 The real secret, though, for both cost and efficiency 00:09:30.580 --> 00:09:33.639 is the golden architecture, the manager -worker 00:09:33.639 --> 00:09:36.299 pattern. Right. So you use Pro as the manager, 00:09:36.460 --> 00:09:38.740 the architect, for maybe 10 % of your tasks, 00:09:38.960 --> 00:09:42.120 high -level planning, complex reasoning, the 00:09:42.120 --> 00:09:44.279 big -picture stuff. Then use Flash as the worker. 00:09:44.700 --> 00:09:47.940 the executor for the other 90%. It executes the 00:09:47.940 --> 00:09:50.679 plan, runs the code, processes the data. You 00:09:50.679 --> 00:09:52.340 just don't use the expensive model to change 00:09:52.340 --> 00:09:54.159 a button color when Flash can do it perfectly 00:09:54.159 --> 00:09:56.679 in a fraction of the time. What is the ratio 00:09:56.679 --> 00:09:59.059 we should remember when architecting a new application 00:09:59.059 --> 00:10:02.539 using this pattern? Follow the rule of 90 % Flash 00:10:02.539 --> 00:10:06.519 for execution and 10 % Pro for high -level, complex 00:10:06.519 --> 00:10:09.629 planning and architectural strategy. You know, 00:10:09.649 --> 00:10:11.710 Flash isn't just another update. It's a really 00:10:11.710 --> 00:10:14.350 strong signal about where AI development is headed. 00:10:14.590 --> 00:10:16.529 If you connect the dots, there are like four 00:10:16.529 --> 00:10:18.730 major shifts happening right now because of this. 00:10:18.870 --> 00:10:21.570 Okay, so signal number one. Pro -level intelligence 00:10:21.570 --> 00:10:25.090 is becoming incredibly cheap. That old line between 00:10:25.090 --> 00:10:28.610 fast and basic and expensive and powerful, it's 00:10:28.610 --> 00:10:31.149 just collapsed. And that means low -cost models 00:10:31.149 --> 00:10:33.190 will soon handle things that were just too expensive 00:10:33.190 --> 00:10:36.090 before. Imagine an AI reviewing every single 00:10:36.090 --> 00:10:39.289 pull request or live continuous refactoring happening 00:10:39.289 --> 00:10:41.549 right inside your editor without you even thinking 00:10:41.549 --> 00:10:43.950 about the bill. Signal 2, models are actually 00:10:43.950 --> 00:10:46.039 learning how to think. This dynamic thinking 00:10:46.039 --> 00:10:48.240 is part of a trend where models move beyond just 00:10:48.240 --> 00:10:50.659 guessing the next word. We should expect models 00:10:50.659 --> 00:10:53.039 that debug their own code or agents that can 00:10:53.039 --> 00:10:56.080 plan multi -day development tasks. And that completely 00:10:56.080 --> 00:10:58.740 changes your role as a builder. You spend less 00:10:58.740 --> 00:11:01.200 time on boilerplate and more time deciding what 00:11:01.200 --> 00:11:03.639 should be built and why. You move up the stack 00:11:03.639 --> 00:11:07.580 to strategy. Signal three. Speed is the new battleground. 00:11:08.200 --> 00:11:10.799 Accuracy used to be everything, but now the focus 00:11:10.799 --> 00:11:14.220 is on latency. A slow suggestion breaks your 00:11:14.220 --> 00:11:17.450 flow. The future is near instant responses that 00:11:17.450 --> 00:11:19.730 feel like a native part of your editor. And signal 00:11:19.730 --> 00:11:21.929 four is that development is opening up to everyone. 00:11:22.129 --> 00:11:25.309 When strong coding AI costs almost nothing, the 00:11:25.309 --> 00:11:27.230 friction to build something is basically zero. 00:11:27.809 --> 00:11:30.750 Solo builders can ship complex products. Non 00:11:30.750 --> 00:11:32.649 -technical founders can prototype real ideas. 00:11:32.970 --> 00:11:36.399 Junior devs get senior level support. If the 00:11:36.399 --> 00:11:38.419 quality and cost of the tools are no longer the 00:11:38.419 --> 00:11:40.440 problem, what is the core question that remains 00:11:40.440 --> 00:11:42.679 for developers? The only remaining question is 00:11:42.679 --> 00:11:45.240 what you, the builder, choose to build with this 00:11:45.240 --> 00:11:48.320 new democratized power. Okay, so Flash is incredibly 00:11:48.320 --> 00:11:50.539 powerful, but we have to be really clear. It 00:11:50.539 --> 00:11:52.960 is not a magic bullet. And running into these 00:11:52.960 --> 00:11:54.960 limits will cause some painful, expensive mistakes 00:11:54.960 --> 00:11:57.519 if you treat it as flawless. Limitation number 00:11:57.519 --> 00:12:02.000 one. Very long contexts. Flash still struggles 00:12:02.000 --> 00:12:04.419 when you push it into millions of tokens. It 00:12:04.419 --> 00:12:07.539 works best in that 2 ,000 to 8 ,000 token range. 00:12:07.720 --> 00:12:10.320 Beyond that, it starts missing details. The fix 00:12:10.320 --> 00:12:12.820 is to break up your big card bases into chunks 00:12:12.820 --> 00:12:15.500 and always include a short architecture summary 00:12:15.500 --> 00:12:17.600 in every prompt to keep it on track. Limitation 00:12:17.600 --> 00:12:21.230 two, new or bleeding edge frameworks. Flash only 00:12:21.230 --> 00:12:23.789 knows stuff up to its training cutoff. If you're 00:12:23.789 --> 00:12:25.389 using a brand new framework, it might suggest 00:12:25.389 --> 00:12:27.990 outdated or just wrong patterns. So you have 00:12:27.990 --> 00:12:30.549 to bring the framework to Flash, paste the relevant 00:12:30.549 --> 00:12:32.929 docs right into the context window and be explicit. 00:12:33.110 --> 00:12:36.490 Tell it, use this exact Astro 5 .0 API, not the 00:12:36.490 --> 00:12:39.330 old ones. Limitation 4 is a classic LLM problem. 00:12:39.789 --> 00:12:41.730 Confident answers to vague prompts. If you're 00:12:41.730 --> 00:12:43.690 unclear, it'll just make an assumption and give 00:12:43.690 --> 00:12:45.909 you a confident but potentially incorrect answer. 00:12:46.070 --> 00:12:47.789 The fix is to treat your prompt like a strict 00:12:47.789 --> 00:12:49.909 contract. Define your preconditions and post 00:12:49.909 --> 00:12:51.970 conditions. Don't just say, handle the date. 00:12:52.090 --> 00:12:55.330 Say, if date parsing fails, log a warning and 00:12:55.330 --> 00:12:58.610 return. None. Clear rules, safer output. And 00:12:58.610 --> 00:13:02.230 finally, number five, security and privacy. Flash 00:13:02.230 --> 00:13:05.120 runs on external infrastructure. Never, ever 00:13:05.120 --> 00:13:08.440 send secrets, API keys, credentials, sensitive 00:13:08.440 --> 00:13:11.419 business logic in clear text. The safe ways are 00:13:11.419 --> 00:13:13.759 to anonymize your code, use Google's enterprise 00:13:13.759 --> 00:13:16.299 tier, or run it through something like Vertex 00:13:16.299 --> 00:13:18.740 AI for more control. This is not your private 00:13:18.740 --> 00:13:22.100 notebook. Which limitation, if ignored, poses 00:13:22.100 --> 00:13:24.340 the greatest immediate risk to a real -world 00:13:24.340 --> 00:13:26.879 deployed application? Ignoring security and privacy 00:13:26.879 --> 00:13:29.340 rules or treating external infrastructure like 00:13:29.340 --> 00:13:31.779 a private notebook is without question the biggest 00:13:31.779 --> 00:13:34.779 risk. So Gemini 3 Flash is a rare, genuinely 00:13:34.779 --> 00:13:37.539 paradigm shifting upgrade. It's faster. It's 00:13:37.539 --> 00:13:39.720 demonstrably smarter than its sibling. And critically, 00:13:39.919 --> 00:13:42.179 it's four times cheaper. It breaks that budget 00:13:42.179 --> 00:13:43.960 model tradeoff we've all been living with. You 00:13:43.960 --> 00:13:46.059 get high level intelligence with high speed and 00:13:46.059 --> 00:13:48.159 low cost. And this changes developer behavior 00:13:48.159 --> 00:13:51.019 instantly. Complex tasks like refactoring or 00:13:51.019 --> 00:13:53.320 debugging now feel safe to hand off to an AI 00:13:53.320 --> 00:13:56.110 because the cost is so low you just. The budget 00:13:56.110 --> 00:13:58.889 model really has become the king of code. And 00:13:58.889 --> 00:14:00.750 this leads us to our final provocative thought 00:14:00.750 --> 00:14:03.970 for you. The sources suggest trying to flack. 00:14:04.389 --> 00:14:07.509 Flash out in Google AI Studio for the best control. 00:14:08.090 --> 00:14:11.929 So here's a thought. If AI can now handle complex 00:14:11.929 --> 00:14:14.490 spatial reasoning and code architecture for pennies, 00:14:14.490 --> 00:14:17.129 how long until the cost of computational creativity 00:14:17.129 --> 00:14:20.679 becomes virtually zero? And what is the next 00:14:20.679 --> 00:14:23.200 high -value, uniquely human skill that developers 00:14:23.200 --> 00:14:25.440 will need to cultivate to stay relevant? That 00:14:25.440 --> 00:14:27.539 is the essential question to build on. And if 00:14:27.539 --> 00:14:29.259 you want to experience this dynamic thinking 00:14:29.259 --> 00:14:31.340 firsthand, go try running one of those gauntlet 00:14:31.340 --> 00:14:33.539 tests we talked about, the Voxelar Eagle or the 00:14:33.539 --> 00:14:36.259 complex scale animation. That'll really show 00:14:36.259 --> 00:14:38.399 you what this new architecture can do. Thank 00:14:38.399 --> 00:14:40.179 you for sharing your sources and letting us take 00:14:40.179 --> 00:14:41.899 this deep dive with you. We'll catch you next 00:14:41.899 --> 00:14:42.200 time.