WEBVTT 00:00:00.000 --> 00:00:01.679 I want you to just picture a scene for a second. 00:00:02.240 --> 00:00:06.339 A European villa, maybe, midday sun. Okay. There's 00:00:06.339 --> 00:00:09.699 a woman at a table. She looks distressed. A man's 00:00:09.699 --> 00:00:12.320 across from her. And you write down the dialogue. 00:00:12.460 --> 00:00:15.580 She says, these trees will turn yellow in a month. 00:00:15.980 --> 00:00:18.280 And he says, but they'll be green again next 00:00:18.280 --> 00:00:22.870 summer. In the past, and by past I mean literally 00:00:22.870 --> 00:00:26.030 last month, if you fed that into an AI, you'd 00:00:26.030 --> 00:00:29.489 get one continuous, probably pretty awkward shot. 00:00:29.550 --> 00:00:32.429 The camera might just float around. The faces, 00:00:32.770 --> 00:00:35.250 they'd probably blur. You'd feel like surveillance 00:00:35.250 --> 00:00:37.710 camera footage of a soap opera. Exactly. But 00:00:37.710 --> 00:00:40.149 something. Something changed this week. With 00:00:40.149 --> 00:00:43.109 Cling 3 .0, you hit generate on that same prompt, 00:00:43.210 --> 00:00:45.450 and the model doesn't just render pixels anymore. 00:00:45.710 --> 00:00:49.310 It decides to start with a wide shot to set the 00:00:49.310 --> 00:00:52.109 scene. Then right when the woman speaks, it cuts, 00:00:52.229 --> 00:00:55.329 a hard cut to a close -up on her face. Then it 00:00:55.329 --> 00:00:57.469 cuts to the man for his reaction. It's making 00:00:57.469 --> 00:00:59.890 editorial decisions. And that's the moment. That 00:00:59.890 --> 00:01:01.649 right there is the signal that clip generation 00:01:01.649 --> 00:01:04.969 is dead. And actual direction has arrived. It's 00:01:04.969 --> 00:01:07.400 not just rendering. Yeah. It's filmmaking. Welcome 00:01:07.400 --> 00:01:10.400 to the Deep Dive. It is Thursday, February 5th, 00:01:10.400 --> 00:01:15.519 2026. We are looking at Kling 3 .0. And I really 00:01:15.519 --> 00:01:18.879 want to slow down and just process what this 00:01:18.879 --> 00:01:20.659 update means because it really feels like we've 00:01:20.659 --> 00:01:22.799 crossed some kind of threshold. We definitely 00:01:22.799 --> 00:01:24.459 have. You know, for the last year, we've been 00:01:24.459 --> 00:01:27.349 drowning in what I call tech demos. Right. You 00:01:27.349 --> 00:01:30.510 know, cool astronauts, melting clocks, cars on 00:01:30.510 --> 00:01:33.109 fire. It's great for a tweet, but it's totally 00:01:33.109 --> 00:01:35.250 useless for a movie. It all felt very experimental, 00:01:35.489 --> 00:01:38.530 like a toy. It was a toy. But Kling 3 .0 has 00:01:38.530 --> 00:01:41.469 officially replaced 2 .6. And the headline isn't 00:01:41.469 --> 00:01:43.709 just better graphics. It's a fundamental shift 00:01:43.709 --> 00:01:46.469 in the architecture. We're moving from a system 00:01:46.469 --> 00:01:49.530 that generates footage. to a system that generates 00:01:49.530 --> 00:01:52.049 seams. And that distinction is everything. It 00:01:52.049 --> 00:01:54.209 matters a lot. It really does. So let's map this 00:01:54.209 --> 00:01:56.370 out for a bit. We need to talk about this unified 00:01:56.370 --> 00:01:59.250 multimodal architecture, which it sounds like 00:01:59.250 --> 00:02:01.150 marketing speak, but it's actually really important. 00:02:01.290 --> 00:02:03.230 It is. We have to get into the multi -shot feature, 00:02:03.430 --> 00:02:05.189 which is that AI director we were just talking 00:02:05.189 --> 00:02:07.730 about. And we've got to discuss consistency, 00:02:07.829 --> 00:02:10.849 how they finally stopped faces from melting into 00:02:10.849 --> 00:02:13.409 sludge. The sludge is gone. Thankfully. And then 00:02:13.409 --> 00:02:15.469 finally, we have to see where this all fits next 00:02:15.469 --> 00:02:17.729 to Sora and Runway because that whole landscape 00:02:17.729 --> 00:02:20.370 is getting very, very crowded. It is. But let's 00:02:20.370 --> 00:02:23.069 start with the engine, the unified multimodal 00:02:23.069 --> 00:02:26.610 architecture. Yeah. So in the old days, and again, 00:02:26.669 --> 00:02:28.810 we mean like six months ago. Right. The old days. 00:02:28.870 --> 00:02:32.389 We had this Frankenstein workflow. It was a nightmare. 00:02:32.689 --> 00:02:34.930 Oh, it was awful. You'd use one model for the 00:02:34.930 --> 00:02:37.530 video, then you'd have to drag that into a totally 00:02:37.530 --> 00:02:39.770 separate tool for the audio, and then maybe another 00:02:39.770 --> 00:02:41.930 one to upscale it. It felt like you were gluing 00:02:41.930 --> 00:02:44.710 together pieces from different puzzles. It was 00:02:44.710 --> 00:02:47.270 so disjointed. And that's because the models 00:02:47.270 --> 00:02:50.270 themselves were disjointed. The video brain didn't 00:02:50.270 --> 00:02:54.110 talk to the audio brain. But Kling 3 .0 is a 00:02:54.110 --> 00:02:58.409 native, unified system. Think of it like the 00:02:58.409 --> 00:03:01.030 human brain. When you dream, You don't do the 00:03:01.030 --> 00:03:04.710 visuals, then pause to dream the sound and then 00:03:04.710 --> 00:03:06.590 edit them in your head. It all happens at once. 00:03:06.590 --> 00:03:08.990 It's simultaneous. Exactly. It's one experience. 00:03:09.289 --> 00:03:14.110 And Kling 3 .0 is finally trying to dream the 00:03:14.110 --> 00:03:17.250 video natively all at once. That's why the sync 00:03:17.250 --> 00:03:19.430 is so good, because the audio isn't just reacting 00:03:19.430 --> 00:03:21.150 to the video. They're being born at the exact 00:03:21.150 --> 00:03:24.550 same instant from the same data. So because it's 00:03:24.550 --> 00:03:26.870 one system, it's just more efficient. So much 00:03:26.870 --> 00:03:30.060 more efficient. And that efficiency. buys us 00:03:30.060 --> 00:03:32.460 the one thing we've all been desperate for time 00:03:32.460 --> 00:03:35.680 we're jumping from those frantic you know five 00:03:35.680 --> 00:03:39.560 second clips to full 15 second continuous scenes 00:03:39.560 --> 00:03:42.240 okay let's pause on that 15 seconds to someone 00:03:42.240 --> 00:03:44.360 just scrolling tiktok adding 10 seconds probably 00:03:44.360 --> 00:03:47.080 sounds trivial why is that such a heavy lift 00:03:47.080 --> 00:03:49.280 technically why couldn't we just do that before 00:03:49.280 --> 00:03:52.120 it's the memory drift problem imagine you're 00:03:52.120 --> 00:03:54.199 trying to draw a comic strip Frame by frame. 00:03:54.360 --> 00:03:56.180 But by the time you get to panel three, you've 00:03:56.180 --> 00:03:57.780 completely forgotten what the main character's 00:03:57.780 --> 00:04:01.240 nose looks like. The old models had a very short 00:04:01.240 --> 00:04:03.939 attention span. After five seconds of generating 00:04:03.939 --> 00:04:06.439 video, which is a huge amount of data, the model 00:04:06.439 --> 00:04:08.539 would just forget the starting conditions. And 00:04:08.539 --> 00:04:10.699 everything would warp. The background would warp. 00:04:10.740 --> 00:04:13.900 The shirt color changes. It gets chaotic. Extending 00:04:13.900 --> 00:04:17.240 that coherence out to 15 seconds required a massive 00:04:17.240 --> 00:04:19.500 architectural overhaul. And creatively, that 00:04:19.500 --> 00:04:23.660 extra time changes the entire medium. Five seconds 00:04:23.660 --> 00:04:25.779 is basically a gif. It's a reaction, a loop. 00:04:25.959 --> 00:04:28.040 You can't really tell a story in five seconds. 00:04:28.259 --> 00:04:31.660 No, you can't. You can't have a character, say, 00:04:31.720 --> 00:04:34.620 walk into a room, hesitate for a second, realize 00:04:34.620 --> 00:04:36.220 they're in the wrong place, and then turn around. 00:04:36.480 --> 00:04:39.279 That takes time. With 15 seconds, you have room 00:04:39.279 --> 00:04:41.620 to breathe. Characters can finish an action. 00:04:41.680 --> 00:04:43.759 They can have a dialogue exchange that actually 00:04:43.759 --> 00:04:47.410 feels human, not rushed. So, I mean, does adding 00:04:47.410 --> 00:04:49.949 10 seconds really change the fundamental nature 00:04:49.949 --> 00:04:52.490 of the video? It shifts the output from just 00:04:52.490 --> 00:04:55.370 a fleeting moving image to a narratively complete 00:04:55.370 --> 00:04:57.990 scene. A narratively complete scene. Okay, that's 00:04:57.990 --> 00:05:00.490 a perfect segue to the multi -shot feature because 00:05:00.490 --> 00:05:03.370 this is where the AI stops being a tool and starts 00:05:03.370 --> 00:05:05.470 acting like a collaborator. Or a boss, maybe. 00:05:05.709 --> 00:05:08.910 Or a boss, yeah. Let's go back to that European 00:05:08.910 --> 00:05:12.230 villa prompt. The user didn't ask for any cuts. 00:05:12.910 --> 00:05:15.329 They didn't say cut to camera B at four seconds. 00:05:15.449 --> 00:05:18.829 They just described the emotional flow. She speaks, 00:05:19.009 --> 00:05:22.050 he replies. And Kling 3 .0 just understands that 00:05:22.050 --> 00:05:24.529 in the language of cinema, from the millions 00:05:24.529 --> 00:05:27.449 of movies and its training data, when someone 00:05:27.449 --> 00:05:30.269 speaks, we usually want to see them. Right. So 00:05:30.269 --> 00:05:32.329 it automatically creates that shot, reverse shot. 00:05:32.389 --> 00:05:34.730 It plans the sequence for you. But I have to 00:05:34.730 --> 00:05:36.629 play devil's advocate here for a second. If I'm 00:05:36.629 --> 00:05:38.730 a creator and I type a prompt, don't I want control? 00:05:39.180 --> 00:05:41.980 If the AI is deciding to cut, isn't it kind of 00:05:41.980 --> 00:05:44.060 overriding my artistic intent? What if I wanted 00:05:44.060 --> 00:05:46.399 a long, uncomfortable, continuous take? That's 00:05:46.399 --> 00:05:48.579 the tension, right. But think of it this way. 00:05:49.060 --> 00:05:51.899 You are trading micromanagement for high -level 00:05:51.899 --> 00:05:53.860 direction. Okay. Unless you tell it otherwise. 00:05:53.980 --> 00:05:56.860 The AI assumes you want standard cinematic grammar. 00:05:57.160 --> 00:05:59.920 And it's actually, it's freeing for a writer. 00:06:00.040 --> 00:06:01.800 You can just focus on the story, the dialogue, 00:06:01.939 --> 00:06:04.000 the mood, and let the model handle the technical 00:06:04.000 --> 00:06:06.480 stitching. So I become the showrunner and the 00:06:06.480 --> 00:06:09.560 AI is the episode director. Precisely. You're 00:06:09.560 --> 00:06:11.839 not the editor anymore, you know, frame bashing 00:06:11.839 --> 00:06:14.160 in Premiere. You're the producer saying, make 00:06:14.160 --> 00:06:17.220 this scene feel sad. And the AI executes the 00:06:17.220 --> 00:06:19.759 technicalities of sadness, which includes the 00:06:19.759 --> 00:06:22.399 pacing and the cutting. That is such a wild shift 00:06:22.399 --> 00:06:24.839 in agency. But it only works if the characters 00:06:24.839 --> 00:06:27.620 look the same across those cuts. Which brings 00:06:27.620 --> 00:06:31.620 us to the melting problem. Ah, the dreaded drift. 00:06:31.839 --> 00:06:34.699 Anyone who's used 2 .6 knows this. You generate 00:06:34.699 --> 00:06:37.629 a cool character in shot one. By shot three, 00:06:37.870 --> 00:06:40.810 they look like their cousin. By shot five, their 00:06:40.810 --> 00:06:43.110 face starts doing that weird like melting candle 00:06:43.110 --> 00:06:45.430 thing. Yeah. It breaks the immersion immediately. 00:06:45.689 --> 00:06:47.290 Oh, it pulls you right out of the story. It's 00:06:47.290 --> 00:06:50.069 the uncanny valley at its absolute worst. Yeah. 00:06:50.490 --> 00:06:53.110 But Kling 3 .0 has attacked this with something 00:06:53.110 --> 00:06:55.829 called video element references. I read about 00:06:55.829 --> 00:06:58.069 this. It's basically like giving the AI a reference 00:06:58.069 --> 00:06:59.870 sheet, right? It's more than a sheet. It's like 00:06:59.870 --> 00:07:02.269 a memory bank. Instead of just describing your 00:07:02.269 --> 00:07:04.209 character with text and, you know, hoping the 00:07:04.209 --> 00:07:06.790 random noise gets it right twice, you upload 00:07:06.790 --> 00:07:09.730 a three to eight second video clip of that character. 00:07:09.889 --> 00:07:11.730 So not just a photo. It actually sees how they 00:07:11.730 --> 00:07:15.170 move. That's the key. It captures the 3D structure 00:07:15.170 --> 00:07:18.310 of their face in motion. It locks that data in. 00:07:18.490 --> 00:07:21.250 So when you go to generate a new scene, the AI 00:07:21.250 --> 00:07:24.350 says, OK, I know this entity and it preserves 00:07:24.350 --> 00:07:26.490 the structure. No more melting hands, no more 00:07:26.490 --> 00:07:29.649 changing jawlines. That is crucial for any kind 00:07:29.649 --> 00:07:32.110 of serialization. If you want to make a web series, 00:07:32.490 --> 00:07:35.170 you need the same actor to show up in episode 00:07:35.170 --> 00:07:37.930 two. It's the difference between a random generation 00:07:37.930 --> 00:07:41.319 and a cast member. And this ties directly into 00:07:41.319 --> 00:07:43.800 the audio character binding feature. Which is 00:07:43.800 --> 00:07:46.199 fascinating. I saw that example of the cafe scene, 00:07:46.379 --> 00:07:49.439 an English man and a French woman having a bilingual 00:07:49.439 --> 00:07:52.279 conversation. Right. And in the prompt, you use 00:07:52.279 --> 00:07:54.980 these tags. You say an English man says this 00:07:54.980 --> 00:07:57.160 and French woman says that. And it handles the 00:07:57.160 --> 00:07:59.560 lip sync for both languages in the same shot. 00:07:59.800 --> 00:08:01.899 Flawlessly. Yeah. And because of that unified 00:08:01.899 --> 00:08:03.980 architecture we talked about, the lips aren't 00:08:03.980 --> 00:08:06.519 just flapping over a static image. The facial 00:08:06.519 --> 00:08:09.529 muscles are actually moving correctly for. French 00:08:09.529 --> 00:08:11.769 phonetics versus English phonetics. That is. 00:08:11.829 --> 00:08:13.730 Wow. It's approaching reality. It sounds like 00:08:13.730 --> 00:08:15.810 this removes the uncanniness, but does it feel 00:08:15.810 --> 00:08:18.970 human? By binding the voice and the visuals so 00:08:18.970 --> 00:08:22.430 tightly, it bridges that gap between just a simulation 00:08:22.430 --> 00:08:25.589 and a believable performance. A believable performance. 00:08:25.649 --> 00:08:28.709 It's exciting, but also, you know, a little terrifying 00:08:28.709 --> 00:08:31.170 for actors. Just a little bit. Speaking of things 00:08:31.170 --> 00:08:33.620 that need to be reliable. Okay, we're back. So 00:08:33.620 --> 00:08:35.919 we've talked about the automation, the AI directing 00:08:35.919 --> 00:08:39.299 the cuts, the AI syncing the lips. But there's 00:08:39.299 --> 00:08:41.600 a subset of our listeners, and I definitely count 00:08:41.600 --> 00:08:44.220 myself among them, who are control freaks. The 00:08:44.220 --> 00:08:46.820 pixel peepers. Exactly. I don't always want the 00:08:46.820 --> 00:08:49.460 AI to decide when to cut. Sometimes I have a 00:08:49.460 --> 00:08:52.259 very specific vision in my head. And Kling 3 00:08:52.259 --> 00:08:56.179 .0 has a response to that with the Omni Architecture's 00:08:56.179 --> 00:08:58.759 storyboard mode. This is the power user feature. 00:08:58.899 --> 00:09:00.700 This is for when you want to be the director 00:09:00.700 --> 00:09:02.940 and the cinematographer. In storyboard mode, 00:09:03.059 --> 00:09:05.759 you can define every single shot before you hit 00:09:05.759 --> 00:09:07.940 generate. You set the duration, the framing, 00:09:08.100 --> 00:09:10.139 the camera motion, everything. So I can say shot 00:09:10.139 --> 00:09:13.259 one, four seconds, wide angle, slow zoom in. 00:09:13.340 --> 00:09:16.960 Shot two, three seconds, close up, static. Yes. 00:09:17.639 --> 00:09:19.679 There was a great example in the release notes. 00:09:19.840 --> 00:09:23.639 It was a Chinese period drama scene. The prompt 00:09:23.639 --> 00:09:26.600 was oddly specific. It was, who can bully my 00:09:26.600 --> 00:09:29.179 obedient adult? Who can bully my obedient adult? 00:09:29.320 --> 00:09:31.240 I feel like we're missing some context there. 00:09:31.419 --> 00:09:33.299 It's definitely a genre trope. A little bit bad. 00:09:33.559 --> 00:09:35.940 But look at the camera direction. The user specified 00:09:35.940 --> 00:09:39.529 a crash zoom. So the camera rushes forward. It 00:09:39.529 --> 00:09:42.029 cuts past the main character to land right on 00:09:42.029 --> 00:09:44.529 an elderly onlooker's astonished eyes. That's 00:09:44.529 --> 00:09:47.769 a very aggressive cinematic choice. It is. And 00:09:47.769 --> 00:09:50.330 the AI executed it perfectly because the user 00:09:50.330 --> 00:09:52.250 controlled all the camera movement parameters. 00:09:53.009 --> 00:09:55.529 But, and this is a big, but just because you 00:09:55.529 --> 00:09:57.250 can do a crash zoom doesn't mean you should. 00:09:57.409 --> 00:09:59.529 The documentation actually had a whole section 00:09:59.529 --> 00:10:02.009 on what actually improves cinematic results. 00:10:02.330 --> 00:10:04.370 It felt like a mini film school lesson. It was. 00:10:04.529 --> 00:10:07.840 And the biggest tip, composition first. effects 00:10:07.840 --> 00:10:10.100 second stop trying to make the camera do backflips 00:10:10.100 --> 00:10:13.759 exactly the ai creates much higher fidelity images 00:10:13.759 --> 00:10:16.539 when the movement is purposeful a slow dolly 00:10:16.539 --> 00:10:20.220 a gentle orbit when you ask for a fast spin while 00:10:20.220 --> 00:10:22.100 the character runs and the building explodes 00:10:22.100 --> 00:10:25.240 you're just introducing way too much noise the 00:10:25.240 --> 00:10:27.860 ai gets confused and the result looks like a 00:10:27.860 --> 00:10:30.779 video game glitch the rule of one action per 00:10:30.779 --> 00:10:35.240 shot one clear subject one clear movement simplicity 00:10:35.240 --> 00:10:39.009 wins So does the tool really require a filmmaker's 00:10:39.009 --> 00:10:41.289 eye to be used effectively? You need to think 00:10:41.289 --> 00:10:42.950 like a photographer, not just a prompt engineer. 00:10:43.110 --> 00:10:45.669 You have to understand framing. Which is a skill 00:10:45.669 --> 00:10:47.029 I think a lot of people are going to have to 00:10:47.029 --> 00:10:49.110 learn very, very quickly. Okay, we have to look 00:10:49.110 --> 00:10:51.509 at the context now. Kling isn't operating in 00:10:51.509 --> 00:10:55.129 a vacuum. We've got OpenAI Sora. We've got Runway, 00:10:55.169 --> 00:10:57.710 Pika, Vidu. It's a battle royale out there. It's 00:10:57.710 --> 00:11:00.230 getting so crowded. So where does Kling 3 .0 00:11:00.230 --> 00:11:02.629 fit in all this? Is it the best at everything? 00:11:03.120 --> 00:11:05.399 No, and I don't think any single model will be. 00:11:05.480 --> 00:11:07.840 They're all carving out their own niches. Let's 00:11:07.840 --> 00:11:10.340 look at Sora too. The heavyweight Sora is still 00:11:10.340 --> 00:11:12.639 the king of physics. If you need a simulation 00:11:12.639 --> 00:11:15.000 of water crashing into a lighthouse or a glass 00:11:15.000 --> 00:11:17.480 shattering on a floor, Sora just understands 00:11:17.480 --> 00:11:19.759 the physical world better. It simulates gravity 00:11:19.759 --> 00:11:22.820 and fluid dynamics in a way that Kling doesn't 00:11:22.820 --> 00:11:25.620 quite match yet. Okay, so Sora for simulation. 00:11:26.200 --> 00:11:29.759 What about Runway Gen 3? Runway is the tool for 00:11:29.759 --> 00:11:32.879 those control freaks we mentioned. I think Runway 00:11:32.879 --> 00:11:35.519 still holds the crown for granular control. If 00:11:35.519 --> 00:11:37.659 I need to change the color of a car in the background 00:11:37.659 --> 00:11:39.659 without altering the lighting on the protagonist's 00:11:39.659 --> 00:11:42.940 face, Runway's brush tools and local in -painting 00:11:42.940 --> 00:11:45.519 are still superior. Runway lets you be a surgeon. 00:11:45.940 --> 00:11:48.519 Exactly. Kling is painting with a broader brush. 00:11:48.759 --> 00:11:50.840 It wants to give you the whole scene, whereas 00:11:50.840 --> 00:11:53.299 Runway lets you surgically alter individual pixels. 00:11:53.580 --> 00:11:56.299 And Pika and Vidu, where do they fit? Pika is 00:11:56.299 --> 00:11:59.019 for speed and social media effects. It's flashy. 00:11:59.179 --> 00:12:01.980 And Video 2 has really cornered the market on 00:12:01.980 --> 00:12:04.820 Asian aesthetics and anime styles. It just handles 00:12:04.820 --> 00:12:06.700 those textures better than anyone. So where does 00:12:06.700 --> 00:12:09.259 that leave Cling? Cling 3 .0 is the production 00:12:09.259 --> 00:12:11.940 tool. That's its whole identity. It wins on the 00:12:11.940 --> 00:12:14.580 workflow. If you want to create a raw clip to 00:12:14.580 --> 00:12:17.019 edit later, maybe you use Runway. But if you 00:12:17.019 --> 00:12:19.940 want to generate an edited sequence... A scene 00:12:19.940 --> 00:12:22.059 that tells a story right out of the box with 00:12:22.059 --> 00:12:24.100 dialogue and cuts. That's Kling's territory. 00:12:24.399 --> 00:12:26.120 That's Kling's territory. It's not just generating 00:12:26.120 --> 00:12:28.620 footage, it's generating cinema. So is there 00:12:28.620 --> 00:12:30.740 even a clear winner here at this point? It's 00:12:30.740 --> 00:12:33.539 specialized now. Use Kling when you need to tell 00:12:33.539 --> 00:12:36.360 a story with cuts and dialogue. Use the others 00:12:36.360 --> 00:12:39.419 for VFX shots. It's about picking the right hammer 00:12:39.419 --> 00:12:43.000 for the right nail. Let's zoom out then. We started 00:12:43.000 --> 00:12:45.039 this whole conversation with the idea that the 00:12:45.039 --> 00:12:48.960 tech demo era is over. We saw Kling 2 .6, a noble 00:12:48.960 --> 00:12:51.740 effort, but it was flawed. Good visuals, short 00:12:51.740 --> 00:12:54.039 clips, the drifting faces. It's not us interested. 00:12:54.179 --> 00:12:56.480 It's a proof of concept. And now we have 3 .0, 00:12:56.559 --> 00:13:00.340 a unified system, 15 -second scenes, an AI that 00:13:00.340 --> 00:13:03.139 understands film editing logic. The moment of 00:13:03.139 --> 00:13:05.179 wonder for me isn't just the visual quality. 00:13:05.460 --> 00:13:09.220 It's the integration. It's the ability to write 00:13:09.220 --> 00:13:12.600 a prompt that contains complex emotions, like, 00:13:12.679 --> 00:13:16.279 are you always this hopeful? And have the machine. 00:13:16.779 --> 00:13:18.879 Act as the cinematographer, the sound engineer, 00:13:19.080 --> 00:13:21.679 the casting director and the editor all in one 00:13:21.679 --> 00:13:24.460 pass. It's the convergence of all of it. It's 00:13:24.460 --> 00:13:26.940 reliable enough now for agencies and creators 00:13:26.940 --> 00:13:29.559 to actually build businesses on. And that's the 00:13:29.559 --> 00:13:32.360 real shift. It's no longer just look what I made. 00:13:32.399 --> 00:13:34.440 It's look what I sold. So if you're listening 00:13:34.440 --> 00:13:36.639 to this and you have access, I think it's rolling 00:13:36.639 --> 00:13:38.879 out to ultra subscribers first. That's right. 00:13:38.940 --> 00:13:40.820 Paying members get the first bite. Don't wait. 00:13:40.940 --> 00:13:43.960 Start testing those multi -shot features. Push 00:13:43.960 --> 00:13:46.620 the duration to 15 seconds. Try the character 00:13:46.620 --> 00:13:48.879 binding. The barrier to professional storytelling 00:13:48.879 --> 00:13:50.919 has basically just collapsed. You don't need 00:13:50.919 --> 00:13:53.080 a million dollar budget. You just need a good 00:13:53.080 --> 00:13:55.440 idea and the ability to describe it. The production 00:13:55.440 --> 00:13:58.840 era of AI video is here. The real question is, 00:13:58.919 --> 00:14:00.539 what are you going to make with it? That is the 00:14:00.539 --> 00:14:02.299 question. And here's a final thought to leave 00:14:02.299 --> 00:14:05.600 you with. If the AI can now direct the scene, 00:14:05.799 --> 00:14:08.620 edit the scene, and act the scene, and it's trained 00:14:08.620 --> 00:14:12.509 on all of human cinema. How long until we stop 00:14:12.509 --> 00:14:14.669 asking, how do we make this look real? And we 00:14:14.669 --> 00:14:17.490 start asking, whose directorial style is this 00:14:17.490 --> 00:14:22.049 AI actually mimicking? Ooh, that is a can of 00:14:22.049 --> 00:14:24.309 worms. Are we getting a Spielberg cut or a Kubrick 00:14:24.309 --> 00:14:26.750 cut? Exactly. Something to mull over. Thanks 00:14:26.750 --> 00:14:29.009 for diving in with us. Always a pleasure. We'll 00:14:29.009 --> 00:14:29.509 see you in the next one.