WEBVTT 00:00:00.000 --> 00:00:02.240 I was sitting at my desk this morning, just staring 00:00:02.240 --> 00:00:04.839 at my computer case. It's nothing special, you 00:00:04.839 --> 00:00:06.919 know, the standard mid -range tower, the kind 00:00:06.919 --> 00:00:09.640 you buy off the shelf. Right. And on the screen, 00:00:09.679 --> 00:00:12.060 there was a progress bar, just moving steadily 00:00:12.060 --> 00:00:14.980 to the right. And I'd pulled the Ethernet cable 00:00:14.980 --> 00:00:17.059 out, just to be sure. So it was totally offline. 00:00:17.280 --> 00:00:19.679 Completely offline. And in under two minutes, 00:00:19.940 --> 00:00:24.800 that machine spat out a 10 -second video. Wow. 00:00:25.179 --> 00:00:27.780 But it wasn't just, you know, a silent, glitchy 00:00:27.780 --> 00:00:30.989 mess. It had dialogue. No way. Synchronized lip 00:00:30.989 --> 00:00:33.149 movements, background noise. It was a complete 00:00:33.149 --> 00:00:37.210 scene. And it just, it hit me. This isn't science 00:00:37.210 --> 00:00:40.409 fiction anymore. It's 2026. And that computer, 00:00:40.530 --> 00:00:43.469 it isn't just a computer. It's a persistent, 00:00:43.609 --> 00:00:46.590 autonomous production house sitting right on 00:00:46.590 --> 00:00:48.770 my desk. It is a striking realization, isn't 00:00:48.770 --> 00:00:50.810 it? We've spent the last few years assuming that 00:00:50.810 --> 00:00:53.509 level of creation required, I don't know, a massive 00:00:53.509 --> 00:00:55.689 server farm in California. Exactly. But the center 00:00:55.689 --> 00:00:57.869 of gravity has really shifted. It really has, 00:00:58.070 --> 00:00:59.789 and that's what we're unpacking today. Welcome 00:00:59.789 --> 00:01:02.090 to the Deep Dive. We are looking at a stack that 00:01:02.090 --> 00:01:06.590 is effectively ending the era of the $30 a month 00:01:06.590 --> 00:01:08.989 subscription. For sure. If you are still paying 00:01:08.989 --> 00:01:14.069 for cloud services like Sora or Veo or... nervously 00:01:14.069 --> 00:01:16.090 counting your credits every month, you might 00:01:16.090 --> 00:01:18.250 want to cancel that after you hear this. We're 00:01:18.250 --> 00:01:21.530 talking about the Pinocchio plus WAN 2GP stack. 00:01:21.829 --> 00:01:24.989 It's a bit of a mouthful to say WAN 2GP, but 00:01:24.989 --> 00:01:28.469 the implications are just, they're massive. We're 00:01:28.469 --> 00:01:30.150 going to walk through how to take a standard 00:01:30.150 --> 00:01:33.469 PC and turn it into a studio running LTX 2 and 00:01:33.469 --> 00:01:36.090 WAN 2 .2. And these are open source models that 00:01:36.090 --> 00:01:37.909 have finally caught up. They've reached parity 00:01:37.909 --> 00:01:39.569 with the corporate giants, yeah. And the best 00:01:39.569 --> 00:01:42.230 part. It's free. No credits, no corporate filters, 00:01:42.430 --> 00:01:45.090 just you and the hardware. We've got a roadmap 00:01:45.090 --> 00:01:46.890 for this because it gets a little technical, 00:01:46.989 --> 00:01:48.609 but we're going to keep it grounded. First, we 00:01:48.609 --> 00:01:50.590 need to demystify the hardware because I know 00:01:50.590 --> 00:01:52.530 that scares people. Then we'll look at the setup, 00:01:52.670 --> 00:01:54.870 the models, especially LTX2, which is kind of 00:01:54.870 --> 00:01:57.010 mind blowing. And finally, the philosophy of 00:01:57.010 --> 00:02:00.010 why local privacy matters. Sounds good. Let's 00:02:00.010 --> 00:02:01.810 start with the metal, the hardware barrier. Right. 00:02:02.069 --> 00:02:04.189 Because I think there is this pervasive fear. 00:02:05.049 --> 00:02:08.250 I look at these AI videos with complex physics, 00:02:08.509 --> 00:02:11.710 lighting, water, and I just assume I need a $5 00:02:11.710 --> 00:02:14.129 ,000 rig. That is the most common misconception 00:02:14.129 --> 00:02:17.569 we see. And honestly, it stops people before 00:02:17.569 --> 00:02:20.310 they even start. It's intimidating. It is. The 00:02:20.310 --> 00:02:22.530 whole credit system model of the last few years 00:02:22.530 --> 00:02:25.830 trained us to think computation is scarce and 00:02:25.830 --> 00:02:28.129 expensive. We've been conditioned to think, oh, 00:02:28.169 --> 00:02:29.409 I can't do that. I don't have a supercomputer. 00:02:30.509 --> 00:02:33.270 But if you look at the specs required for ATX2 00:02:33.270 --> 00:02:37.050 or 1 .2 today, they are surprisingly accessible. 00:02:37.389 --> 00:02:40.030 You do not need a monster machine. So let's get 00:02:40.030 --> 00:02:43.069 specific. I'm at my computer right now. If I 00:02:43.069 --> 00:02:45.710 pull up my task manager, what numbers am I actually 00:02:45.710 --> 00:02:48.210 looking for? What's the make or break stat? Okay, 00:02:48.250 --> 00:02:50.689 so you're looking for the GPU, specifically the 00:02:50.689 --> 00:02:53.069 VRAM, the video RAM. That's the tank of gas for 00:02:53.069 --> 00:02:54.990 your AI engine. Right. If you're on Windows, 00:02:55.270 --> 00:02:58.719 you hit Control Shift Escape. Go to performance 00:02:58.719 --> 00:03:00.939 and you just look at that number. The magic number 00:03:00.939 --> 00:03:05.060 for 720p video, which, let's be honest, is perfectly 00:03:05.060 --> 00:03:08.500 fine for most creative work or drafting, is 8GB. 00:03:08.759 --> 00:03:13.099 Wait, really? 8? That seems... That seems low. 00:03:13.340 --> 00:03:15.259 I mean, most mid -range cards from three or four 00:03:15.259 --> 00:03:17.020 years ago have that. Are you sure that's enough 00:03:17.020 --> 00:03:19.419 for video generation? That's the floor. Eight 00:03:19.419 --> 00:03:21.979 gigs gets you in the door. It lets you run the 00:03:21.979 --> 00:03:25.000 distilled models. Okay. Now, if you want 1080p, 00:03:25.139 --> 00:03:27.259 you're probably looking at 12 GB+. If you want 00:03:27.259 --> 00:03:30.759 native 4K, sure, then you need 24 GB, and that 00:03:30.759 --> 00:03:33.370 gets expensive. Right. For the vast majority 00:03:33.370 --> 00:03:35.870 of people, just trying to create an NVIDIA RTX 00:03:35.870 --> 00:03:39.289 3040 or even the new 50 series with 8GB of VRAM 00:03:39.289 --> 00:03:41.669 is enough. I have to admit, I have a bit of that 00:03:41.669 --> 00:03:43.729 hardware fear myself. I've hesitated to even 00:03:43.729 --> 00:03:45.729 try installing these things because I didn't 00:03:45.729 --> 00:03:47.650 want my computer to, you know, melt. Yeah, or 00:03:47.650 --> 00:03:49.310 just freeze up so bad you have to hard reboot. 00:03:49.509 --> 00:03:52.210 Exactly. But you're saying 6 to 8GB is actually 00:03:52.210 --> 00:03:54.949 viable. It is, especially with the distilled 00:03:54.949 --> 00:03:56.569 versions of these models, which we'll get to. 00:03:56.689 --> 00:03:59.469 You also need about 16GB of system RAM. That's 00:03:59.469 --> 00:04:03.560 your normal memory. Roughly 20 to 40 gigs of 00:04:03.560 --> 00:04:06.719 disk space. Which, in the context of modern computing, 00:04:06.960 --> 00:04:09.520 that's not a supercomputer. Not at all. That's 00:04:09.520 --> 00:04:11.879 a mid -range gaming setup. If you can play a 00:04:11.879 --> 00:04:14.180 modern video game on medium settings, you can 00:04:14.180 --> 00:04:16.600 probably run a movie studio. That's a comforting 00:04:16.600 --> 00:04:20.290 thought. So... If the hardware isn't the gatekeeper 00:04:20.290 --> 00:04:23.089 anymore, what is? Why isn't everyone doing this? 00:04:23.329 --> 00:04:25.689 It's just the willingness to install the software. 00:04:25.810 --> 00:04:29.310 The barrier is psychological, not physical. Psychological. 00:04:29.370 --> 00:04:31.569 That's interesting. It's that feeling of this 00:04:31.569 --> 00:04:34.470 is too technical for me. Exactly. Which leads 00:04:34.470 --> 00:04:36.870 us perfectly into the solution because historically 00:04:36.870 --> 00:04:40.769 installing local AI was, well, it was a nightmare. 00:04:40.990 --> 00:04:43.420 Oh, it was dependency hell. Total nightmare. 00:04:43.839 --> 00:04:46.060 Dependency hell. I love that term. Explain what 00:04:46.060 --> 00:04:47.600 that actually felt like for people who missed 00:04:47.600 --> 00:04:50.160 that era. So imagine you want to run a program. 00:04:50.660 --> 00:04:53.279 But to run it, you need Python. Okay. But not 00:04:53.279 --> 00:04:56.560 just any Python. You need version 3 .1 .0 .6 00:04:56.560 --> 00:04:59.800 specifically. And then you need CA drivers, but 00:04:59.800 --> 00:05:02.459 only version 11 .8. And then you install some 00:05:02.459 --> 00:05:04.920 other library. and it updates your Python to 00:05:04.920 --> 00:05:07.819 3 .11 and suddenly where everything breaks. You'd 00:05:07.819 --> 00:05:10.339 spend three hours just staring at a command prompt 00:05:10.339 --> 00:05:13.160 with red text scrolling by. And you haven't even 00:05:13.160 --> 00:05:15.959 generated a single pixel. Not a single one. Yeah, 00:05:16.000 --> 00:05:17.660 that's usually the moment I close the laptop 00:05:17.660 --> 00:05:21.139 and go watch TV. But the source material highlights 00:05:21.139 --> 00:05:23.980 this tool called Pinocchio. They call it the 00:05:23.980 --> 00:05:27.000 Steam for AI. That's a big claim. It is, but 00:05:27.000 --> 00:05:29.439 it's a very apt comparison. I mean, think about 00:05:29.439 --> 00:05:33.290 Steam for video games. You don't manually install 00:05:33.290 --> 00:05:36.009 DirectX or audio drivers for every game. No, 00:05:36.050 --> 00:05:38.250 you just click play. Pinocchio does that for 00:05:38.250 --> 00:05:40.649 AI. It is essentially browser or a launcher. 00:05:40.810 --> 00:05:43.129 It manages everything under the hood. So it handles 00:05:43.129 --> 00:05:45.750 all that messy stuff. Completely. You download 00:05:45.750 --> 00:05:48.449 the installer, you run it, and it creates a contained 00:05:48.449 --> 00:05:51.430 environment, like a sandbox. So it can't mess 00:05:51.430 --> 00:05:53.889 up my actual computer. It builds a wall around 00:05:53.889 --> 00:05:56.069 the software. It handles the Python scripts, 00:05:56.250 --> 00:05:58.050 the Torch installations, the virtual environments. 00:05:58.170 --> 00:06:00.290 You don't see any of that. So it's effectively 00:06:00.290 --> 00:06:04.019 one click. Effectively, yes. You install Pinocchio, 00:06:04.180 --> 00:06:06.839 you search for WAN2GP in its discovery tab, it 00:06:06.839 --> 00:06:08.720 looks just like an app store, and you click install. 00:06:08.980 --> 00:06:11.959 And WAN2GP is the interface, the control panel. 00:06:12.160 --> 00:06:15.100 Exactly. Think of WAN2GP as the studio control 00:06:15.100 --> 00:06:18.060 panel that runs inside Pinocchio. It unifies 00:06:18.060 --> 00:06:20.300 all the different models, like WAN and LTX2, 00:06:20.500 --> 00:06:24.290 into one clean web UI. You're not typing code. 00:06:24.389 --> 00:06:26.310 You're clicking buttons in a browser window. 00:06:26.449 --> 00:06:28.629 That shifts the dynamic completely. It goes from 00:06:28.629 --> 00:06:31.629 an engineering task to just an installation task. 00:06:31.810 --> 00:06:35.110 Installing Word or Excel. Precisely. I have to 00:06:35.110 --> 00:06:37.350 ask, though, does this tool actually democratize 00:06:37.350 --> 00:06:39.490 the technology or does it just simplify the interface? 00:06:39.629 --> 00:06:41.889 Is there a difference? It bridges the gap between 00:06:41.889 --> 00:06:44.550 coder and creator, making the tech invisible 00:06:44.550 --> 00:06:47.089 so the art can happen. I love that. Making the 00:06:47.089 --> 00:06:49.089 tech invisible. So once you have that invisible 00:06:49.089 --> 00:06:51.449 tech running. Yeah. You have to choose a brain 00:06:51.449 --> 00:06:53.810 for your studio. The sources talk about two main 00:06:53.810 --> 00:06:56.990 heavy hitters, one 2 .2 and LTX2. Right. As a 00:06:56.990 --> 00:06:59.410 newbie, how do I choose? They serve different 00:06:59.410 --> 00:07:01.750 purposes. Think of them like different camera 00:07:01.750 --> 00:07:04.550 lenses or different directors. If you're looking 00:07:04.550 --> 00:07:08.290 for pure visual fidelity, Hollywood -level textures, 00:07:08.550 --> 00:07:11.550 complex physics, lighting that just behaves correctly, 00:07:11.910 --> 00:07:15.230 one 2 .2. is your benchmark so it's the powerhouse 00:07:15.230 --> 00:07:17.949 for visuals it's the powerhouse if you want a 00:07:17.949 --> 00:07:20.449 shot of a futuristic city with rain reflecting 00:07:20.449 --> 00:07:23.810 off the pavement 1 2 .2 is your go -to but then 00:07:23.810 --> 00:07:26.029 there's ltx2 and this is the one that really 00:07:26.029 --> 00:07:28.089 caught my attention it was released back in january 00:07:28.089 --> 00:07:32.649 2026 and the source calls it the star why is 00:07:32.649 --> 00:07:35.009 this one getting so much hype because it solves 00:07:35.009 --> 00:07:37.610 the synchronization problem This has been the 00:07:37.610 --> 00:07:40.649 holy grail for AI video for so long. Right. Previous 00:07:40.649 --> 00:07:42.689 models generated video, then you had to go to 00:07:42.689 --> 00:07:44.550 a separate tool for sound, another tool for dialogue, 00:07:44.750 --> 00:07:46.470 and then try to stitch it all together. And the 00:07:46.470 --> 00:07:48.389 lips were always a little off, like a bad dubbed 00:07:48.389 --> 00:07:51.750 movie. Exactly. LTX2 changes that. It's a 19 00:07:51.750 --> 00:07:54.389 billion parameter model that generates video, 00:07:54.689 --> 00:07:57.930 audio, and lip sync narration in a single pass. 00:07:58.149 --> 00:08:00.449 A single pass. It understands the scene acoustically 00:08:00.449 --> 00:08:03.170 as well as visually. Wait, so it handles the 00:08:03.170 --> 00:08:06.490 foley work? Like footsteps and wind? Yes. If 00:08:06.490 --> 00:08:08.430 the character is walking on gravel, you hear 00:08:08.430 --> 00:08:11.449 gravel. If they're speaking, their lips move 00:08:11.449 --> 00:08:14.189 in time with the audio. It's multimodal by default. 00:08:14.449 --> 00:08:16.370 So you're not prompting for a video, you're prompting 00:08:16.370 --> 00:08:19.689 for a moment. A moment. That's a great way to 00:08:19.689 --> 00:08:22.790 put it. That is a massive leap in workflow. But 00:08:22.790 --> 00:08:26.069 there's a catch, right? A 19 billion parameter 00:08:26.069 --> 00:08:29.949 model sounds... Heavy. It is. You mentioned earlier 00:08:29.949 --> 00:08:33.190 that 8GB of VRAM is the floor, but 19 billion 00:08:33.190 --> 00:08:35.490 parameters usually requires way more than that. 00:08:35.590 --> 00:08:37.929 It's huge. The full model is around 40 gigabytes. 00:08:38.110 --> 00:08:40.149 For most people, that is going to choke their 00:08:40.149 --> 00:08:42.809 system, which is why the guide strongly recommends 00:08:42.809 --> 00:08:45.289 the distilled version. Distilled. It sounds like 00:08:45.289 --> 00:08:48.039 whiskey. Or specialized chemistry. What does 00:08:48.039 --> 00:08:50.860 that actually mean in AI terms? Well, in a way, 00:08:50.879 --> 00:08:52.899 it is like whiskey. It's a concentrated version. 00:08:53.179 --> 00:08:56.179 They've managed to compress the model size from 00:08:56.179 --> 00:08:59.240 40 gigs down to about 20. How do they do that? 00:08:59.419 --> 00:09:01.879 They essentially teach a smaller student model 00:09:01.879 --> 00:09:05.759 to mimic the big teacher model. You lose a tiny 00:09:05.759 --> 00:09:08.379 fraction of visual fidelity. It might be slightly 00:09:08.379 --> 00:09:10.559 less crisp. Maybe the background is a little 00:09:10.559 --> 00:09:13.279 softer. But the tradeoff is that it runs smoothly 00:09:13.279 --> 00:09:16.220 on consumer GPUs without crashing. So you're 00:09:16.220 --> 00:09:18.120 trading a little bit of pixel perfection for 00:09:18.120 --> 00:09:20.519 the ability to actually run the thing. Exactly. 00:09:20.720 --> 00:09:23.139 So why choose the distilled version over the 00:09:23.139 --> 00:09:25.360 full power model? Performance over perfection. 00:09:25.580 --> 00:09:28.779 It cuts storage in half and stops your home PC 00:09:28.779 --> 00:09:31.240 from melting. I'll take a working PC over a melted 00:09:31.240 --> 00:09:36.000 one any day. Okay, so we've got Pinocchio installed, 00:09:36.240 --> 00:09:38.580 we've selected the distilled LTX2 model, now 00:09:38.580 --> 00:09:40.740 we have to configure it. The source mentions 00:09:40.740 --> 00:09:43.240 a critical do -this -once step about memory profiles. 00:09:43.740 --> 00:09:45.720 This seems like the place where people get tripped 00:09:45.720 --> 00:09:47.639 up. Yes, this is where you prevent the headaches. 00:09:47.799 --> 00:09:50.100 Inside the WAN2GP interface, there's a configuration 00:09:50.100 --> 00:09:52.720 tab. Under performance, you can select a memory 00:09:52.720 --> 00:09:55.059 profile. And these profiles correspond to your 00:09:55.059 --> 00:09:58.289 hardware tier. Correct. Profile 1 is for lower 00:09:58.289 --> 00:10:00.730 -end systems. Maybe you're scraping by on 6GB. 00:10:01.169 --> 00:10:04.710 Profile 3 is for the high -end beasts with 24GB 00:10:04.710 --> 00:10:08.289 or more. But... The sweet spot for that standard 00:10:08.289 --> 00:10:10.870 16 GB setup we talked about. Is profile 2. Is 00:10:10.870 --> 00:10:13.769 profile 2. Yeah. You set that once and the system 00:10:13.769 --> 00:10:16.389 knows how to manage its resources. It tells the 00:10:16.389 --> 00:10:19.009 AI, hey, don't eat more than this amount of memory 00:10:19.009 --> 00:10:21.129 so it doesn't crash mid -render. Okay. Profile 00:10:21.129 --> 00:10:24.169 2. Locked in. Now let's generate something. The 00:10:24.169 --> 00:10:26.970 guide suggests starting with text prompt only. 00:10:27.190 --> 00:10:29.500 Mm -hmm. And they give the specific example that 00:10:29.500 --> 00:10:31.320 I found really funny. It's a business meeting. 00:10:31.519 --> 00:10:34.600 Ah, yes. The famous cookie prompt. Right. The 00:10:34.600 --> 00:10:37.679 prompt is, three people sit around a table. One 00:10:37.679 --> 00:10:40.039 woman says, the cookies will tell us when the 00:10:40.039 --> 00:10:43.299 time is right. It's so weirdly specific. It's 00:10:43.299 --> 00:10:45.480 a great stress test, though. It's surreal. But 00:10:45.480 --> 00:10:47.419 what happens when you run it? Walk me through 00:10:47.419 --> 00:10:50.240 the actual experience of hitting generate. Well, 00:10:50.279 --> 00:10:51.960 first, you have to be patient. It takes about 00:10:51.960 --> 00:10:53.899 two minutes for the first run because the model 00:10:53.899 --> 00:10:56.519 has to load from your hard drive into that VRAM. 00:10:56.600 --> 00:10:58.799 We call that the cold start. Okay, so don't panic 00:10:58.799 --> 00:11:00.759 and nothing happens for a minute. Exactly. Don't 00:11:00.759 --> 00:11:03.559 start clicking wildly. But once it's loaded, 00:11:03.759 --> 00:11:06.700 subsequent runs drop to about 30 seconds. Wow. 00:11:07.340 --> 00:11:09.980 And the result of that cookie prompt, it's a 00:11:09.980 --> 00:11:11.879 realistic office scene. Professional lighting, 00:11:12.039 --> 00:11:15.220 suits, ties. But crucially, the woman actually 00:11:15.220 --> 00:11:18.610 says the line. Her lips match the words, and 00:11:18.610 --> 00:11:21.289 the context, that bizarre tension of the cookie 00:11:21.289 --> 00:11:24.289 prophecy, is maintained in the visuals. That's 00:11:24.289 --> 00:11:26.730 wild. It's doing the visual rendering and the 00:11:26.730 --> 00:11:28.750 audio synthesis at the same time. And it costs 00:11:28.750 --> 00:11:31.590 zero dollars. That's the kicker. So what is the 00:11:31.590 --> 00:11:35.929 significance of that specific cookie prompt working 00:11:35.929 --> 00:11:38.190 correctly, aside from being funny? It proves 00:11:38.190 --> 00:11:41.049 the model understands context, dialogue, and 00:11:41.049 --> 00:11:44.360 surrealism simultaneously, all locally. Understanding 00:11:44.360 --> 00:11:47.659 surrealism is a high bar for AI. Or maybe for 00:11:47.659 --> 00:11:49.759 humans, too. For sure. But let's say I don't 00:11:49.759 --> 00:11:51.179 want to start from scratch. I have a bunch of 00:11:51.179 --> 00:11:53.919 images from Midjourney or Nanobanana Pro. I want 00:11:53.919 --> 00:11:56.039 to animate them. This is the image -to -video 00:11:56.039 --> 00:11:58.779 workflow. And honestly, for storytellers, this 00:11:58.779 --> 00:12:01.360 is incredibly powerful. You're not rolling the 00:12:01.360 --> 00:12:03.580 dice on a new character every time. You're taking 00:12:03.580 --> 00:12:06.659 an existing asset and giving it motion. The example 00:12:06.659 --> 00:12:09.519 prompt in the source is evocative. The person 00:12:09.519 --> 00:12:12.220 keeps walking toward the castle. Snow falls. 00:12:12.840 --> 00:12:16.000 The camera slowly moves forward. Simple directional 00:12:16.000 --> 00:12:18.740 commands. You upload the image, type that prompt, 00:12:18.860 --> 00:12:21.539 and the AI calculates the physics. It knows how 00:12:21.539 --> 00:12:23.620 snow falls. It knows how a camera dolly works. 00:12:23.899 --> 00:12:26.159 It extrapolates the movement. And again, where 00:12:26.159 --> 00:12:28.779 does this file go? In the cloud era, it went 00:12:28.779 --> 00:12:32.100 to some library on a website somewhere. Here, 00:12:32.120 --> 00:12:34.019 it goes straight to your hard drive. Specifically, 00:12:34.220 --> 00:12:38.899 the folder path is one, git, then app. then outputs. 00:12:39.299 --> 00:12:42.000 It's all organized by date. So it's just a file, 00:12:42.039 --> 00:12:44.600 like a Word doc or a JPEG. Exactly. You own it. 00:12:44.659 --> 00:12:46.559 There's no middleman holding your IP. There's 00:12:46.559 --> 00:12:48.240 no download button you have to click before your 00:12:48.240 --> 00:12:51.559 subscription expires. It's just yours. How does 00:12:51.559 --> 00:12:53.860 local storage change the creative relationship 00:12:53.860 --> 00:12:56.740 with the work? It shifts ownership. Your IP never 00:12:56.740 --> 00:12:58.860 leaves your drive, granting total privacy and 00:12:58.860 --> 00:13:01.299 control. Okay, we're back. We've covered the 00:13:01.299 --> 00:13:03.340 how -to, but I want to talk about the what if 00:13:03.340 --> 00:13:05.899 it breaks. Because let's be real, technology 00:13:05.899 --> 00:13:08.679 breaks. Oh, it does. And when you're your own 00:13:08.679 --> 00:13:12.129 IT department... That can be scary. The source 00:13:12.129 --> 00:13:14.570 mentions that most bugs aren't actually code 00:13:14.570 --> 00:13:17.049 errors. Right. Beginners often panic when the 00:13:17.049 --> 00:13:19.409 generation fails or the app crashes. They assume 00:13:19.409 --> 00:13:21.850 the software is broken or they installed it wrong. 00:13:22.269 --> 00:13:24.909 99 % of the time, it's simply a resource limit. 00:13:24.990 --> 00:13:26.830 You ran out of VRAM? Yeah, I ran out of VRAM. 00:13:26.850 --> 00:13:29.009 So the solution isn't to reinstall Windows or 00:13:29.009 --> 00:13:31.330 format your hard drive? No, no, definitely not. 00:13:31.429 --> 00:13:34.000 The solution is usually simple. Switch to the 00:13:34.000 --> 00:13:35.980 distilled model if you haven't already, or maybe 00:13:35.980 --> 00:13:39.159 lower the resolution from 1080p to 720p. It's 00:13:39.159 --> 00:13:41.440 about managing the size of your gas tank, essentially. 00:13:41.639 --> 00:13:44.000 Right. If you try to drive 500 miles on a gallon 00:13:44.000 --> 00:13:46.179 of gas, the car will stop. It's not broken. It's 00:13:46.179 --> 00:13:49.000 just empty. And what about compatibility? We've 00:13:49.000 --> 00:13:51.159 talked a lot about NVIDIA. Is this a Windows 00:13:51.159 --> 00:13:54.080 -only club? No. Pinocchio handles Windows, Mac, 00:13:54.379 --> 00:13:58.399 and Linux. Now, I have to be honest. NVIDIA GPUs 00:13:58.399 --> 00:14:00.639 are the gold standard here. That's just the reality 00:14:00.639 --> 00:14:03.919 of the CUDA ecosystem right now. But there is 00:14:03.919 --> 00:14:06.919 documentation for AMD support. It's becoming 00:14:06.919 --> 00:14:09.500 more universal. If you're on a Mac with Apple 00:14:09.500 --> 00:14:12.639 Silicon, Pinocchio handles that translation layer 00:14:12.639 --> 00:14:15.360 too, though it might be a bit slower. This brings 00:14:15.360 --> 00:14:17.720 us to the philosophy of it all. Why does this 00:14:17.720 --> 00:14:19.759 matter? Why go through the trouble of installing 00:14:19.759 --> 00:14:22.279 Pinocchio and managing memory profiles when I 00:14:22.279 --> 00:14:24.779 could just pay the 30 bucks and use a cloud website? 00:14:25.059 --> 00:14:27.809 It comes down to three things. Privacy, quotas, 00:14:27.809 --> 00:14:30.769 and speed. Let's unpack privacy first because 00:14:30.769 --> 00:14:32.570 I think people shrug this off sometimes. I have 00:14:32.570 --> 00:14:34.669 nothing to hide. Sure, but it's not just about 00:14:34.669 --> 00:14:37.610 hiding things. It's about creative freedom. When 00:14:37.610 --> 00:14:40.169 you type a prompt into a corporate cloud model, 00:14:40.350 --> 00:14:44.149 that prompt is data. It's being analyzed. It's 00:14:44.149 --> 00:14:46.269 potentially being used to train future models. 00:14:46.470 --> 00:14:49.830 Your creative process is being profiled. Locally, 00:14:49.830 --> 00:14:52.940 that data never leaves your Ethernet port. You 00:14:52.940 --> 00:14:55.519 can explore controversial topics, personal ideas, 00:14:55.679 --> 00:14:58.179 or proprietary business concepts without leaking 00:14:58.179 --> 00:15:00.679 IP. That's huge for businesses especially. You 00:15:00.679 --> 00:15:02.399 don't want your confidential marketing strategy 00:15:02.399 --> 00:15:05.960 training a public AI model. Exactly. And then 00:15:05.960 --> 00:15:08.830 you have quotas, the credit anxiety. You know 00:15:08.830 --> 00:15:11.269 the feeling. I have 50 credits left this month. 00:15:11.370 --> 00:15:13.649 Do I really want to waste one on this experimental 00:15:13.649 --> 00:15:16.490 idea? Oh, absolutely. I hoard them. I end up 00:15:16.490 --> 00:15:18.309 not making anything because I'm saving them for 00:15:18.309 --> 00:15:21.070 the perfect idea. Exactly. That kills creativity. 00:15:21.370 --> 00:15:24.250 It makes you risk averse. When the cost is zero, 00:15:24.429 --> 00:15:27.649 you can generate 100 bad variations to find the 00:15:27.649 --> 00:15:30.049 one good one. You can make mistakes. You can 00:15:30.049 --> 00:15:32.190 afford to fail. You can afford to play. And that 00:15:32.190 --> 00:15:34.250 is where innovation happens. When you aren't 00:15:34.250 --> 00:15:36.750 watching a meter tick down, you try things you 00:15:36.750 --> 00:15:38.470 never would have tried otherwise. And speed. 00:15:38.850 --> 00:15:40.950 You mentioned earlier that the first run takes 00:15:40.950 --> 00:15:43.429 two minutes, which sounds slow compared to some 00:15:43.429 --> 00:15:46.250 websites. The cold start is slower, yes. But 00:15:46.250 --> 00:15:49.360 once that model is loaded into your VRAM... It's 00:15:49.360 --> 00:15:51.919 blazing fast. You aren't waiting in a server 00:15:51.919 --> 00:15:54.399 queue behind 10 ,000 other users. You aren't 00:15:54.399 --> 00:15:56.879 dealing with internet latency. It's instantaneous 00:15:56.879 --> 00:15:59.059 responsiveness. It feels like an instrument, 00:15:59.179 --> 00:16:01.860 not a service. It feels like we are moving toward 00:16:01.860 --> 00:16:04.860 a world where local AI isn't the alternative. 00:16:05.039 --> 00:16:07.600 It's becoming the standard for serious creators. 00:16:08.080 --> 00:16:11.100 I think so. The gap between what a cloud cluster 00:16:11.100 --> 00:16:14.379 can do and what a home GPU can do is shrinking 00:16:14.379 --> 00:16:16.960 every single month. If the gap is shrinking, 00:16:17.259 --> 00:16:19.720 what is the only remaining advantage of the cloud? 00:16:19.879 --> 00:16:22.639 Just raw scale. But for personal storytelling, 00:16:23.000 --> 00:16:26.200 local offers freedom the cloud can't match. Freedom. 00:16:26.200 --> 00:16:28.460 That's the big idea here. We've moved from a 00:16:28.460 --> 00:16:31.320 credit -based economy of creativity where you 00:16:31.320 --> 00:16:34.919 pay per thought to an era of ownership. It's 00:16:34.919 --> 00:16:37.269 a fundamental shift. With a mid -range card, 00:16:37.429 --> 00:16:39.590 some open source software, and a bit of patience, 00:16:39.870 --> 00:16:42.629 your home computer is a studio. And the barrier 00:16:42.629 --> 00:16:45.610 to entry isn't $5 ,000. It isn't a degree in 00:16:45.610 --> 00:16:47.830 computer science. It is simply the initiative 00:16:47.830 --> 00:16:50.779 to download and install it. That is it. The source 00:16:50.779 --> 00:16:52.279 ends with a question that I think is worth chewing 00:16:52.279 --> 00:16:54.960 on. Will you be ready when this becomes normal? 00:16:55.080 --> 00:16:57.340 Yeah. Because it is becoming normal fast. It 00:16:57.340 --> 00:16:59.379 is. And the people who learn to manage these 00:16:59.379 --> 00:17:02.519 local workflows now, who learn how to prompt, 00:17:02.659 --> 00:17:05.319 how to manage VRAM, how to iterate without cost, 00:17:05.559 --> 00:17:08.119 they are going to have a massive advantage over 00:17:08.119 --> 00:17:10.039 the people still waiting for their monthly credits 00:17:10.039 --> 00:17:13.240 to refresh. So here's your homework. Right now, 00:17:13.400 --> 00:17:16.700 hit Control -Shift -Esc, check your performance 00:17:16.700 --> 00:17:19.880 tab. Look at your GPU memory. If you see an 8 00:17:19.880 --> 00:17:23.920 or a 12 or anything higher, give it a shot. Go 00:17:23.920 --> 00:17:27.240 to pinocchio .co. Try the installation. It costs 00:17:27.240 --> 00:17:28.920 you nothing but a little bit of disk space to 00:17:28.920 --> 00:17:31.740 see the future. And honestly, it's pretty fun 00:17:31.740 --> 00:17:33.460 to watch the cookies tell you when the time is 00:17:33.460 --> 00:17:35.519 right. It certainly is. Thanks for diving in 00:17:35.519 --> 00:17:36.880 with us. We'll catch you in the next one. Bye.