WEBVTT 00:00:00.000 --> 00:00:02.279 Close your eyes for a second. I want you to imagine 00:00:02.279 --> 00:00:04.160 you're sitting in a director's chair. You're 00:00:04.160 --> 00:00:06.900 on set. The lighting is moody. The tension is 00:00:06.900 --> 00:00:11.419 palpable. You yell, action. But here's the twist. 00:00:11.839 --> 00:00:15.539 There's no camera crew, no lights, no actors. 00:00:15.759 --> 00:00:18.160 You just type a single sentence into a terminal. 00:00:18.399 --> 00:00:20.460 And for the first time, and this is the part 00:00:20.460 --> 00:00:22.600 that actually matters, the actors remember who 00:00:22.600 --> 00:00:24.719 they are from the first shot to the last. They 00:00:24.719 --> 00:00:26.679 don't morph into different people. They don't 00:00:26.679 --> 00:00:29.649 suddenly grow a third arm. We are looking at 00:00:29.649 --> 00:00:33.890 the death of the cool demo clip and maybe the 00:00:33.890 --> 00:00:37.170 birth of actual AI filmmaking. Welcome back to 00:00:37.170 --> 00:00:39.850 The Deep Drive. It is February 2026. If you've 00:00:39.850 --> 00:00:41.210 been blinking, you might have missed that the 00:00:41.210 --> 00:00:45.729 AI video wars have gone, well, nuclear. And this 00:00:45.729 --> 00:00:47.570 is what we're digging into today. We aren't just 00:00:47.570 --> 00:00:49.390 looking at text -to -video resolution anymore. 00:00:49.530 --> 00:00:51.429 We're looking at something, honestly, much more 00:00:51.429 --> 00:00:55.159 important, workflow. That's the key word. workflow. 00:00:55.159 --> 00:00:57.299 For so long, we were just in that novelty phase. 00:00:57.420 --> 00:00:59.600 Look, a panda riding a skateboard. It's cute, 00:00:59.740 --> 00:01:02.640 technically impressive, but useless for telling 00:01:02.640 --> 00:01:05.200 a story. Now we're finally seeing tools that 00:01:05.200 --> 00:01:07.560 are trying to be actual production suites. So 00:01:07.560 --> 00:01:10.319 today, we're doing a forensic breakdown of a 00:01:10.319 --> 00:01:12.879 major new player that claims to have taken the 00:01:12.879 --> 00:01:16.299 crown. We're talking about Kling 3 .0, and we're 00:01:16.299 --> 00:01:18.079 going to see how it stacks up against the heavyweights. 00:01:18.719 --> 00:01:24.000 OpenAI's Sora 2 and, uh... Google's VO 3 .1. 00:01:24.099 --> 00:01:26.599 And we have a great set of data for this. We're 00:01:26.599 --> 00:01:28.439 looking at a really comprehensive comparison 00:01:28.439 --> 00:01:32.099 by Max Ann over at AI Fire. He didn't just read 00:01:32.099 --> 00:01:34.420 the spec sheets. He actually burned through a 00:01:34.420 --> 00:01:36.599 ton of credits running these head -to -head stress 00:01:36.599 --> 00:01:38.859 tests. Which is great because it saves us the 00:01:38.859 --> 00:01:42.260 money. So here's the roadmap. First, we're going 00:01:42.260 --> 00:01:44.219 to unpack the features that, you know, theoretically 00:01:44.219 --> 00:01:46.980 shift Kling from a toy to a tool, specifically 00:01:46.980 --> 00:01:49.980 Multishot and Omni. Then we'll look at the economics. 00:01:50.079 --> 00:01:52.420 What does this thing actually cost to run? Then 00:01:52.420 --> 00:01:55.079 we enter the arena. We've got five brutal headset 00:01:55.079 --> 00:01:57.439 tests. The fun part. Yeah, dialogue, emotion, 00:01:57.640 --> 00:01:59.379 action. And finally, we'll try to figure out 00:01:59.379 --> 00:02:00.859 who this is actually for. Sounds like a plan. 00:02:00.920 --> 00:02:03.319 Let's start with the core thesis here. Kling 00:02:03.319 --> 00:02:06.859 3 .0 seems to be arguing that we don't need higher 00:02:06.859 --> 00:02:08.919 resolution. You know, we have enough pixels. 00:02:09.280 --> 00:02:11.300 They're saying we need to fix broken workflows. 00:02:11.719 --> 00:02:13.419 What does that actually mean in practice? Because 00:02:13.419 --> 00:02:16.780 workflow sounds a little corporate. It does, 00:02:16.860 --> 00:02:19.099 doesn't it? But think about the pain of making 00:02:19.099 --> 00:02:23.979 AI video back in, say, 2025. You generate a three 00:02:23.979 --> 00:02:26.360 -second clip, then another one. Then you take 00:02:26.360 --> 00:02:28.039 them into Premiere and try to stitch them together. 00:02:28.240 --> 00:02:30.439 But the problem was always consistency. Right. 00:02:30.539 --> 00:02:32.620 In clip one, the character has a blue shirt. 00:02:32.919 --> 00:02:35.659 In clip two, it's navy. Exactly. Or the lighting 00:02:35.659 --> 00:02:38.639 is shifted from sunset to like high noon for 00:02:38.639 --> 00:02:41.099 no reason. You were fighting the tool. So Kling 00:02:41.099 --> 00:02:43.840 3 .0 introduces multi -shot generation. Yeah. 00:02:43.900 --> 00:02:45.560 And it's exactly what it sounds like, but the 00:02:45.560 --> 00:02:47.979 implication is huge. Instead of one isolated 00:02:47.979 --> 00:02:49.919 clip, you're basically directing a mini movie 00:02:49.919 --> 00:02:52.340 in a single prompt. Okay. So you can define shot 00:02:52.340 --> 00:02:54.240 one as a close -up, shot two as an over -the 00:02:54.240 --> 00:02:56.460 -shoulder, that kind of thing. That's it. And 00:02:56.460 --> 00:02:59.360 the AI generates them all in sequence. as one 00:02:59.360 --> 00:03:02.020 cohesive video file. It handles the cuts internally. 00:03:02.240 --> 00:03:04.680 So it handles continuity, the lighting. All of 00:03:04.680 --> 00:03:07.400 it. The flow of movement. You aren't stitching, 00:03:07.580 --> 00:03:10.180 you're sequencing. And crucially, they've bumped 00:03:10.180 --> 00:03:12.840 the generation time up. It used to be standard 00:03:12.840 --> 00:03:16.159 to get, what, 10 seconds of video. Kling 3 .0 00:03:16.159 --> 00:03:18.340 pushes that to 15. I have to play devil's advocate 00:03:18.340 --> 00:03:20.340 here. Five seconds. That doesn't sound like a 00:03:20.340 --> 00:03:22.479 revolutionary leap. Why does that matter so much? 00:03:22.639 --> 00:03:25.000 Oh, in filmmaking terms, five seconds is an eternity. 00:03:25.939 --> 00:03:28.599 It's the difference between like a TikTok edit 00:03:28.599 --> 00:03:31.180 and a scene that actually has room to breathe. 00:03:31.539 --> 00:03:34.020 In a 10 second clip, everything feels rushed. 00:03:34.669 --> 00:03:38.050 You know, start, action, and cut. It creates 00:03:38.050 --> 00:03:41.270 that sort of frantic AI -generated anxiety. With 00:03:41.270 --> 00:03:44.289 15 seconds, a character can hesitate. They can 00:03:44.289 --> 00:03:46.789 look around. They can react. It just reduces 00:03:46.789 --> 00:03:49.969 that subconscious, this feels fake feeling. Okay, 00:03:50.030 --> 00:03:51.990 but there has to be a tradeoff. Computers don't 00:03:51.990 --> 00:03:53.909 just give you more processing power for free. 00:03:53.990 --> 00:03:55.969 What breaks when you go longer? You're right. 00:03:55.990 --> 00:03:57.830 There is a tax. The longer the video, the higher 00:03:57.830 --> 00:04:00.090 the risk of visual drift. Define that for us. 00:04:00.400 --> 00:04:04.759 Visual drift is when the AI starts to forget 00:04:04.759 --> 00:04:07.800 the details of the scene the further it gets 00:04:07.800 --> 00:04:10.780 from the start frame. By second 14, the AI might 00:04:10.780 --> 00:04:12.680 forget there was a lamp in the corner or the 00:04:12.680 --> 00:04:14.699 wallpaper pattern starts to shift. So you might 00:04:14.699 --> 00:04:16.579 have to run it a few more times to get a clean 00:04:16.579 --> 00:04:19.120 take. Exactly. You trade some reliability for 00:04:19.120 --> 00:04:21.519 better pacing. That brings us to the other big 00:04:21.519 --> 00:04:24.160 feature, which, frankly, this sounds like the 00:04:24.160 --> 00:04:26.319 holy grail for anyone trying to tell a story 00:04:26.319 --> 00:04:30.139 with recurring characters. Omni. Kling 3 .0 Omni 00:04:30.139 --> 00:04:33.100 is basically multi -element control. And yeah, 00:04:33.180 --> 00:04:35.040 this addresses the biggest complaint creators 00:04:35.040 --> 00:04:38.360 have had forever. Identity consistency. In the 00:04:38.360 --> 00:04:39.939 past, you'd write a whole paragraph describing 00:04:39.939 --> 00:04:42.660 a guy in a red hat and just hope the AI drew 00:04:42.660 --> 00:04:44.839 the same guy every time. And it never did. It 00:04:44.839 --> 00:04:46.860 would give you generic guy in red hat A and then 00:04:46.860 --> 00:04:49.620 generic guy in red hat B. Exactly. With Omni, 00:04:49.720 --> 00:04:51.639 you can upload up to seven visual references. 00:04:51.939 --> 00:04:55.680 Seven. Seven distinct elements. Seven. So you 00:04:55.680 --> 00:04:58.120 can upload a photo of a specific actor and tag 00:04:58.120 --> 00:05:01.550 it. At Japanese boy. Upload another. Tag it at 00:05:01.550 --> 00:05:04.290 Japanese girl. Upload a photo of a specific park 00:05:04.290 --> 00:05:07.610 bench. Tag it at bench. Then in your prompt, 00:05:07.670 --> 00:05:10.649 you just say at Japanese boy sits on at bench 00:05:10.649 --> 00:05:13.209 next to a Japanese girl. And it actually works. 00:05:13.250 --> 00:05:15.870 It doesn't do that weird face swapping thing 00:05:15.870 --> 00:05:18.730 or blend them into a blob. According to the stress 00:05:18.730 --> 00:05:22.269 tests, it's surprisingly robust. Max, our source, 00:05:22.389 --> 00:05:24.939 described this scene. The two characters, the 00:05:24.939 --> 00:05:27.300 Japanese boy and girl, are sitting on that specific 00:05:27.300 --> 00:05:29.680 bench. They're having a conversation, and they 00:05:29.680 --> 00:05:32.480 even share a pair of headphones. And the AI kept 00:05:32.480 --> 00:05:35.819 their faces consistent the entire time. No morphing, 00:05:35.879 --> 00:05:37.699 no glitches. That's that workflow part you're 00:05:37.699 --> 00:05:39.600 talking about. You aren't just rolling the dice 00:05:39.600 --> 00:05:41.660 on who shows up to set. Precisely. And they've 00:05:41.660 --> 00:05:43.939 updated the audio, too. The lip sync is reportedly 00:05:43.939 --> 00:05:46.779 much better than version 2 .6. Better, or... 00:05:47.400 --> 00:05:50.459 Good is the operative word. We're not at, you 00:05:50.459 --> 00:05:53.160 know, human level dubbing yet. But the characters 00:05:53.160 --> 00:05:55.579 can finally speak without their jaws unhinging 00:05:55.579 --> 00:05:58.300 or the timing being totally off. It doesn't break 00:05:58.300 --> 00:06:00.220 the immersion immediately, which is a big step. 00:06:00.420 --> 00:06:02.519 I want to push on this a bit. If I have to tag 00:06:02.519 --> 00:06:05.199 seven items, upload reference images, describe 00:06:05.199 --> 00:06:08.019 camera angles for three different shots, all 00:06:08.019 --> 00:06:10.660 in one prompt, does increasing the technical 00:06:10.660 --> 00:06:14.279 complexity actually stifle creativity? Or are 00:06:14.279 --> 00:06:17.050 we just becoming like... data entry clerks for 00:06:17.050 --> 00:06:20.250 a GPU. It's a valid fear, but I'd argue it actually 00:06:20.250 --> 00:06:22.050 liberates creativity. I mean, think about it. 00:06:22.209 --> 00:06:24.689 Randomness isn't creativity. Randomness is chaos. 00:06:24.970 --> 00:06:27.730 If you ask for a scary monster and get some random 00:06:27.730 --> 00:06:29.810 result, you aren't an artist, you're a slot machine 00:06:29.810 --> 00:06:33.170 player. Control is creativity. By removing the 00:06:33.170 --> 00:06:35.509 randomness, you're finally free to direct the 00:06:35.509 --> 00:06:37.850 exact scene you see in your head. That's a compelling 00:06:37.850 --> 00:06:40.689 distinction. Control is creativity. Okay, let's 00:06:40.689 --> 00:06:43.430 talk brass tacks. The economics. Usually when 00:06:43.430 --> 00:06:45.810 a pre -version drops, the price doubles. Is the 00:06:45.810 --> 00:06:48.790 King expensive? Surprisingly, no. The math holds 00:06:48.790 --> 00:06:51.550 up. The per second cost is roughly the same as 00:06:51.550 --> 00:06:53.829 Cling 2 .6. It's about two credits per second. 00:06:53.970 --> 00:06:56.329 So you're paying more per generation, but only 00:06:56.329 --> 00:06:58.790 because you're getting more video. Exactly. A 00:06:58.790 --> 00:07:02.410 15 -second clip costs about 30 credits. A 10 00:07:02.410 --> 00:07:05.800 -second clip costs 20. You aren't paying a premium 00:07:05.800 --> 00:07:08.319 for the new features like Multishot or Omni. 00:07:08.639 --> 00:07:10.319 You're just paying for the length. That feels 00:07:10.319 --> 00:07:12.959 fair, assuming the yield rate is good. If I have 00:07:12.959 --> 00:07:15.699 to generate 10 clips to get one usable one, that 00:07:15.699 --> 00:07:18.459 30 credits becomes 300 real fast. And that is 00:07:18.459 --> 00:07:20.279 the hidden cost. We should be clear about that. 00:07:20.339 --> 00:07:22.920 This isn't a magic button. You will burn credits 00:07:22.920 --> 00:07:25.120 on bad generations. Walk me through the actual 00:07:25.120 --> 00:07:27.079 experience. I'm sitting at the computer. How 00:07:27.079 --> 00:07:29.899 do I use this Multishot thing? Is it code? No, 00:07:29.959 --> 00:07:32.439 it's just a UI toggle. You switch on Multishot. 00:07:32.699 --> 00:07:35.279 in the interface and a panel appears with slots 00:07:35.279 --> 00:07:38.740 shot one shot two shot three you treat each slot 00:07:38.740 --> 00:07:41.060 like a separate camera angle and you can set 00:07:41.060 --> 00:07:43.000 the duration for each one and you write a separate 00:07:43.000 --> 00:07:45.740 prompt for each slot you do and this leads to 00:07:45.740 --> 00:07:47.600 a really interesting pro tip from the source 00:07:47.600 --> 00:07:50.220 material that i think is worth dwelling on i'll 00:07:50.220 --> 00:07:52.259 admit i still wrestle with prompt drifts myself 00:07:52.259 --> 00:07:55.860 i tend to write these very robotic stiff instructions 00:07:55.860 --> 00:08:00.009 subject stands at coordinates x y lighting is 00:08:00.009 --> 00:08:02.329 volumetric. We all do. We try to speak computer. 00:08:02.569 --> 00:08:05.310 Right. But the tip here is to stop doing that. 00:08:05.529 --> 00:08:08.970 Use voice to text. Instead of typing, just talk 00:08:08.970 --> 00:08:10.750 to the AI like you're talking to a cameraman. 00:08:11.149 --> 00:08:12.769 Yeah. Okay, give me a close -up on him. He looks 00:08:12.769 --> 00:08:16.269 nervous. Now he opens the door. The source found 00:08:16.269 --> 00:08:18.829 this natural language actually yields better 00:08:18.829 --> 00:08:21.569 results because it captures the vibe of the scene, 00:08:21.689 --> 00:08:24.490 not just the technical specs. That is fascinating. 00:08:24.750 --> 00:08:27.189 It suggests the model understands semantic intent 00:08:27.189 --> 00:08:30.610 better than keyword stuffing. Exactly. So probing 00:08:30.610 --> 00:08:33.009 question here. If the cost is effectively the 00:08:33.009 --> 00:08:35.730 same and the features are better, why wouldn't 00:08:35.730 --> 00:08:37.450 everyone switch immediately? Is there a catch? 00:08:37.710 --> 00:08:39.889 The catch is the learning curve. This is a more 00:08:39.889 --> 00:08:41.590 complex instrument. You aren't just typing cat 00:08:41.590 --> 00:08:44.450 on bike anymore. You are managing shots, references, 00:08:44.750 --> 00:08:47.470 timing. The tax is patience. It requires you 00:08:47.470 --> 00:08:49.750 to actually be a director, not just a prompter. 00:08:49.850 --> 00:08:51.649 Okay, let's move to the main event, the arena. 00:08:52.169 --> 00:08:53.929 No more specs. We're going to look at the brutal 00:08:53.929 --> 00:08:57.110 head -to -head test that Max ran. Kling 3 .0 00:08:57.110 --> 00:09:00.669 versus Sora 2 versus VO 3 .1. Five tests. Test 00:09:00.669 --> 00:09:04.169 number one. Dialogue. The cafe scene. The setup 00:09:04.169 --> 00:09:06.429 is an English man and a French woman at a cafe. 00:09:07.159 --> 00:09:10.379 He speaks English, she replies in French. A classic 00:09:10.379 --> 00:09:13.919 test for video. Who took it? Kling 3 .0 took 00:09:13.919 --> 00:09:16.059 the gold here. The report says it had the most 00:09:16.059 --> 00:09:18.600 natural emotion and the lip sync was solid. It 00:09:18.600 --> 00:09:20.659 felt like a real conversation. And the others. 00:09:20.899 --> 00:09:23.419 VO 3 .1 really struggled. It had that distinct 00:09:23.419 --> 00:09:26.299 AI sound. You know, it sounds a bit metallic, 00:09:26.320 --> 00:09:28.799 a bit hollow. The uncanny valley voice. Right. 00:09:28.879 --> 00:09:31.490 And Sora. Sora wasn't even tested in this specific 00:09:31.490 --> 00:09:33.929 bracket. There were some restrictions on generating 00:09:33.929 --> 00:09:36.429 human faces speaking in that context, which is 00:09:36.429 --> 00:09:38.789 a whole other issue about safety rails versus 00:09:38.789 --> 00:09:41.730 usability. Interesting. Okay, Kling wins on talk. 00:09:41.929 --> 00:09:45.029 Test number two, emotional realism, the angry 00:09:45.029 --> 00:09:47.570 close -up. The prompt, as for a handheld shot 00:09:47.570 --> 00:09:50.929 of a person at home holding back rage, jaw clenched, 00:09:50.970 --> 00:09:53.950 nostrils flaring. This was a tie. Kling 3 .0 00:09:53.950 --> 00:09:56.580 and Sora 2 both nailed it. They captured that 00:09:56.580 --> 00:09:59.120 subtle tension, the darting eyes, the heavy breathing. 00:09:59.279 --> 00:10:02.379 It wasn't cartoon anger. It was a simmer. And 00:10:02.379 --> 00:10:05.519 Vio. Vio went full soap opera. The source called 00:10:05.519 --> 00:10:08.600 it overacting. Unnatural expressions, contorted 00:10:08.600 --> 00:10:11.000 faces. It didn't get subtlety. It thought anger 00:10:11.000 --> 00:10:15.159 meant make a scary face. So Vio is the bad actor 00:10:15.159 --> 00:10:18.500 in the group. Got it. Test three, complex movement. 00:10:18.700 --> 00:10:22.269 This one is tough. The courier run. A cinematic 00:10:22.269 --> 00:10:24.529 tracking shot of a courier sprinting through 00:10:24.529 --> 00:10:26.889 a crowd, locking eyes with another runner, and 00:10:26.889 --> 00:10:29.509 throwing an envelope. Whoa. I mean, just imagine 00:10:29.509 --> 00:10:31.710 the math required here. He had to keep a courier 00:10:31.710 --> 00:10:33.830 in focus, keeping consistent, while generating 00:10:33.830 --> 00:10:36.190 random pedestrians and cyclists crossing the 00:10:36.190 --> 00:10:38.889 frame, all while maintaining lighting and momentum. 00:10:39.149 --> 00:10:41.649 That is a computational nightmare. It sounds 00:10:41.649 --> 00:10:44.710 impossible for a probabilistic model. And yet, 00:10:44.789 --> 00:10:48.029 Kling 3 .0 did it. It followed the prompt. It 00:10:48.029 --> 00:10:50.370 kept the courier in focus. It handled the occlusion 00:10:50.370 --> 00:10:52.350 people walking in front of the camera without 00:10:52.350 --> 00:10:53.990 glitching the main character out of existence. 00:10:54.370 --> 00:10:58.110 What about the big bad wolf, Sora 2? Sora 2 started 00:10:58.110 --> 00:11:01.289 strong, but it stumbled at the finish line. It 00:11:01.289 --> 00:11:03.649 struggled with the ending. The character acting 00:11:03.649 --> 00:11:06.129 got weird when the interaction happened. And 00:11:06.129 --> 00:11:08.950 Vio? Vio failed completely. It basically ignored 00:11:08.950 --> 00:11:10.809 the prompt. It just had random people pointing 00:11:10.809 --> 00:11:13.429 at things. It lost the plot. It sounds like Vio's 00:11:13.429 --> 00:11:16.110 having a really rough day in these tests. Okay, 00:11:16.169 --> 00:11:19.009 test four. This is where it gets funny but also 00:11:19.009 --> 00:11:23.250 revealing. Start and end frames. The prompt was 00:11:23.250 --> 00:11:25.490 to transition from a quiet courthouse at dawn 00:11:25.490 --> 00:11:28.590 to a wrecked hallway inside. This is a classic 00:11:28.590 --> 00:11:30.509 fill -in -the -blank test. Here's the start. 00:11:30.610 --> 00:11:32.370 Here's the end. You figure out how we got there. 00:11:32.470 --> 00:11:36.490 It's cling 3 .0. Hallucinated. Badly. It completely 00:11:36.490 --> 00:11:39.210 misunderstood the logic of the scene. It turned 00:11:39.210 --> 00:11:41.669 the police officers into robbers. It just swapped 00:11:41.669 --> 00:11:43.889 their roles mid -video. It didn't get the narrative 00:11:43.889 --> 00:11:46.909 implication. It just morphed pixel blobs. NVO? 00:11:47.090 --> 00:11:49.509 Broken physics. It smashed the entrance in a 00:11:49.509 --> 00:11:51.629 way that didn't make physical sense. Right. Walls 00:11:51.629 --> 00:11:53.269 just kind of dissolved instead of breaking. So 00:11:53.269 --> 00:11:56.909 who won? Sora 2. This is where OpenAI flexed 00:11:56.909 --> 00:11:59.509 its muscle. It produced what the source called 00:11:59.509 --> 00:12:02.669 an absolute cinema cutscene. It understood the 00:12:02.669 --> 00:12:04.309 physics of destruction, the lighting change, 00:12:04.490 --> 00:12:07.950 the transition perfectly. So Kling wins on logic 00:12:07.950 --> 00:12:11.190 and consistency, but Sora wins on the raw physics 00:12:11.190 --> 00:12:14.610 engine. Exactly. When things need to break realistically, 00:12:14.990 --> 00:12:17.990 Sora understands the world better. Final test. 00:12:18.230 --> 00:12:21.669 Test number five. Multi -shot consistency. The 00:12:21.669 --> 00:12:24.830 rooftop. A romantic dialogue on a European terrace 00:12:24.830 --> 00:12:27.269 about trees turning yellow. This plays right 00:12:27.269 --> 00:12:29.470 to Kling's strength. It won on consistency. The 00:12:29.470 --> 00:12:31.289 characters look exactly the same in every shot. 00:12:31.490 --> 00:12:33.509 Sora 2 was called a worthy enemy, but it was 00:12:33.509 --> 00:12:35.830 hampered by some policy restrictions on video 00:12:35.830 --> 00:12:37.710 quality. It just looked a bit more compressed, 00:12:37.889 --> 00:12:39.690 a bit fuzzier. So looking at these tests, it 00:12:39.690 --> 00:12:41.850 raises a big philosophical question for me. We 00:12:41.850 --> 00:12:43.889 see Kling winning on human elements, how people 00:12:43.889 --> 00:12:46.149 act, how scenes flow. We see Sora winning on 00:12:46.149 --> 00:12:49.509 physics, how glass breaks, how light moves, which 00:12:49.509 --> 00:12:52.629 is harder for an AI to learn. I'd argue logic 00:12:52.629 --> 00:12:55.500 is the harder hurdle. Physics is math. You can 00:12:55.500 --> 00:12:57.899 simulate gravity. You can simulate light bounces. 00:12:58.000 --> 00:13:00.500 We have game engines that do that. But logic, 00:13:00.659 --> 00:13:03.080 logic is human, understanding why a cop shouldn't 00:13:03.080 --> 00:13:05.279 turn into a robber or why a glance is romantic 00:13:05.279 --> 00:13:08.659 versus angry. That requires a model of the world 00:13:08.659 --> 00:13:12.039 that goes beyond just pixels. It requires understanding 00:13:12.039 --> 00:13:15.340 intent. That's a profound thought. Physics is 00:13:15.340 --> 00:13:18.139 math. Logic is cultural. Right. We are back. 00:13:18.379 --> 00:13:20.259 We've looked at the features, the costs, and 00:13:20.259 --> 00:13:23.590 the stress tests. Now we need a verdict. I'm 00:13:23.590 --> 00:13:25.269 a listener. I have a subscription budget for 00:13:25.269 --> 00:13:26.850 one of these. Which one do I actually sign up 00:13:26.850 --> 00:13:29.350 for? It really comes down to your persona. Who 00:13:29.350 --> 00:13:31.490 are you in this ecosystem? Let's break it down. 00:13:31.730 --> 00:13:33.970 Okay. If you are a filmmaker or a marketer, someone 00:13:33.970 --> 00:13:36.230 who needs control, narrative flow, and wants 00:13:36.230 --> 00:13:38.789 to tell a story with Cutscling, 3 .0 is your 00:13:38.789 --> 00:13:41.330 tool. It allows for B -roll and character scenes 00:13:41.330 --> 00:13:43.309 that are actually usable because you can control 00:13:43.309 --> 00:13:46.250 the sequence. The multi -shot feature kills that 00:13:46.250 --> 00:13:48.610 need for external editing of short clips. You 00:13:48.610 --> 00:13:51.429 can build a scene. Okay. What about the perfectionist? 00:13:51.799 --> 00:13:54.240 The visual effects artist. That's Sora, too. 00:13:55.139 --> 00:13:57.360 If you need a specific physical interaction, 00:13:57.620 --> 00:14:00.539 like a glass shattering realistically, or a car 00:14:00.539 --> 00:14:04.139 crash that obeys the laws of physics, Sora's 00:14:04.139 --> 00:14:06.659 the choice. Assuming you have the budget or the 00:14:06.659 --> 00:14:09.330 access, of course. It's the visual effects powerhouse. 00:14:09.350 --> 00:14:11.210 It just understands the physical world better. 00:14:11.309 --> 00:14:14.450 And VO 3 .1, it took a beating in the tests. 00:14:14.649 --> 00:14:17.850 Is it just a write -off? Not necessarily. It 00:14:17.850 --> 00:14:20.870 has a niche. It's for the ecosystem user. If 00:14:20.870 --> 00:14:22.909 you're already locked into Google and Vertex 00:14:22.909 --> 00:14:25.750 AI and you need specific color grading standards 00:14:25.750 --> 00:14:28.929 like Cinema Standard 24 FPS VO fits into that 00:14:28.929 --> 00:14:30.649 pipeline, it plays nice with the other Google 00:14:30.649 --> 00:14:33.950 tools. But for pure creativity, yeah, it's lagging 00:14:33.950 --> 00:14:36.289 behind right now. So are we approaching a point 00:14:36.289 --> 00:14:38.750 where the best model doesn't exist? Are we entering 00:14:38.750 --> 00:14:40.769 a world where we use different models for different 00:14:40.769 --> 00:14:43.610 shots in the same project? 100%. I think we are 00:14:43.610 --> 00:14:45.850 already there. You might use Sora for the explosion 00:14:45.850 --> 00:14:47.710 because it handles particles better. And then 00:14:47.710 --> 00:14:49.990 you switch to Kling for the reaction shot of 00:14:49.990 --> 00:14:52.389 the hero because it handles emotion better. So 00:14:52.389 --> 00:14:54.669 the director of the future is really just a curator 00:14:54.669 --> 00:14:58.029 of models. Precisely. You're casting algorithms, 00:14:58.370 --> 00:15:01.309 not just actors. You pick the model that knows 00:15:01.309 --> 00:15:03.470 how to perform the scene you need. Okay, let's 00:15:03.470 --> 00:15:05.549 wrap this up with the big picture. We have moved 00:15:05.549 --> 00:15:07.889 from the cool clip phase to the director phase. 00:15:08.190 --> 00:15:11.149 That's the headline. Kling 3 .0's multi -shot 00:15:11.149 --> 00:15:13.870 feature is the signal. It shows the industry 00:15:13.870 --> 00:15:16.929 is finally listening to creators who said, stop 00:15:16.929 --> 00:15:19.690 giving us higher resolution pandas and give us 00:15:19.690 --> 00:15:22.289 tools to actually edit. And the omni control, 00:15:22.409 --> 00:15:24.409 that ability to have recurring characters, that 00:15:24.409 --> 00:15:27.029 feels like the barrier to entry for storytelling 00:15:27.029 --> 00:15:30.149 is finally crumbling. It is. While Sora 2 still 00:15:30.149 --> 00:15:32.809 holds the crown for physical realism, Kling 3 00:15:32.809 --> 00:15:36.039 .0 wins on workflow and narrative. And in the 00:15:36.039 --> 00:15:38.659 long run, for professionals, workflow always 00:15:38.659 --> 00:15:40.620 wins. Before we go, we have a little homework 00:15:40.620 --> 00:15:43.259 for you. Yes. Try that voice -to -text prompting 00:15:43.259 --> 00:15:45.399 method we mentioned. Open up Kling or whatever 00:15:45.399 --> 00:15:47.500 tool you use, close your eyes, and just act out 00:15:47.500 --> 00:15:49.899 the scene to the AI. Don't worry about keywords. 00:15:50.019 --> 00:15:52.360 Just describe the movie. See if the result feels 00:15:52.360 --> 00:15:54.580 more human. I bet it will. And I'll leave you 00:15:54.580 --> 00:15:56.700 with this final thought. If AI can now handle 00:15:56.700 --> 00:16:00.440 continuity, emotion, camera angles, if it can 00:16:00.440 --> 00:16:02.879 remember who the actor is from shot to shot, 00:16:03.500 --> 00:16:06.379 How long is it until the prompt make me a sequel 00:16:06.379 --> 00:16:09.460 to my favorite movie actually works? Not just 00:16:09.460 --> 00:16:11.340 a trailer, but the movie. It's closer than we 00:16:11.340 --> 00:16:13.080 think. See you in the deep end. Take care.