1. Podcasts
  2. Denoised
  3. Is Ben Affleck the Next James Cameron of Tech?
Denoised

4 Oct 2026 · 32m · 13 chapters

Episode artwork

Is Ben Affleck the Next James Cameron of Tech?

In short

Ben Affleck’s tech background and an interview about Interpositive’s AI video post-training; then a broader roundup of current AI model trends for VFX, image/video editing, and enterprise workflows.

Guests

Addy and Joey (podcast hosts). No other named guests; discussion references Ben Affleck (filmmaker/technologist) and companies/models (Interpositive, Netflix, OpenAI, Minimax, FAL, LTX, Utipi, Ideogram, Flux, etc.).

Key claims

Interpositive (Netflix acquired for ~$600M) reportedly post-trains open-weight video models by “unfreezing weights,” using months of footage shot for a film as training data (dailies), stopping around ~8,000 epochs to avoid overfitting. The model outputs were described as 720p, with Netflix-grade results attributed to upscaling. Post-training is framed as useful for bespoke transforms (e.g., wire removal) but not a magic “unlock.”

Notable examples

wire removal/VFX, custom “cinema language” training; Blender gray-box to final render models; LTX AlphaGen keying; Ideogram 4.5 image editing; Flux 3 “precision edits” with keyframe adherence.

Written by AI. May contain mistakes.

Chapters

0:15

Ben Affleck: Filmmaker and Technologist

Discussion on Ben Affleck's skills in technology and his AI company.

“My entire feed has been filled with Ben Affleck.”
3:14

The Interpositive Model Explained

Exploration of the Interpositive AI model and its training methods.

“By the way, also in this interview, I did not realize he, in the 90s, tried to create an NLE competitor to Avid because he thought Avid's were too expensive.”
5:38

The Ethics of AI Training

Discussion on the ethical implications of AI training on copyrighted material.

“It just is like, do you need, is this enough data?”
7:25

Challenges in AI Model Training

Insight into the complexities of post-training AI models.

“Or what if we give the new Batman footage and then, you know, somehow becomes in this training data?”
9:29

Netflix's Investment in AI

Speculation on Netflix's reasons for acquiring Interpositive and its potential.

“I think you can post-train it to do certain things like what Allura does, you know, is like very special things.”
13:14

Conclusion and Future of AI in Film

Final thoughts on the implications of AI in film and upcoming projects.

“Those two things combined, you would not buy an inferior product.”
14:00

Ben Affleck's AI Influence

Explore Ben Affleck's impact on AI technology and filmmaking.

“I've seen them pop up a few times of demos of like an AI studio or an AI tool.”
18:32

New AI Models in VFX

Discuss the latest AI models and their applications in VFX workflows.

“Both kind of do the same thing, and this also ties into a trend of what we've been talking about and seeing.”
21:01

Ideogram 4.5 and Image Editing

Learn about the new capabilities of Ideogram 4.5 for image editing tasks.

“Nano Banana Pro, I mean, that's an ancient model by these standards.”
24:16

Advances in AI and Complex Problems

Investigate how AI is tackling complex problems like fusion and supersonic flight.

“I'm curious what the pricing is going to compare like.”
Show all 13 chapters
28:00

Exploring Flux 3's New Image Model

Learn about the innovative features of Flux 3's latest image model release.

“with Blender and with scripting, with Previs, right?”
29:38

Precision in Image Editing

Discover the importance of precision and keyframes in modern image editing.

“Yeah, so it might just not be on the API yet.”
30:43

Audience Engagement and Journey

Understand the connection with the audience and the ongoing journey of technology.

“One thing Flux 3 and Runway LF2 does really well is adhere to keyframes temporally.”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On Denoised, what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00Joey: Ben Affleck is like, I think, a secret James Cameron. Yeah, right, right. Being very technologically savvy. He's like, yeah, I used to build computers for gaming. I know Python and stuff, and it's like, damn, Ben Affleck. All right.

0:14Joey: Welcome back to Denoised. Addy, how's it going? Joey, let's do it. Let's do it. All right. My entire feed has been filled with Ben Affleck. Has your feed been filled with Ben Affleck?

0:26Addy: Not really, no. Okay, well, it's an X thing. I'm a huge fan of Argo, Gone Girl, and all of his movies. So, like, yeah, let's bring it. I love Ben Affleck as a filmmaker. Yeah.

0:37Joey: But also am very impressed and entertained by Ben Affleck, the technologist. Yes. Now, he definitely had quotes and stuff about AI. And we've talked about Interpositive when he came out. And it was just like, oh, hey, I have secretly been building an AI company that Netflix bought for, like, 600 million bucks. So he was doing an interview at Bloomberg Screen Time with Lucas Shaw and went into more detail about what they were doing, how he was thinking about AI, but also like very impressively detailed about specifically training models from Ben Affleck. because open models, just like open language models, open video models are available to be trained on.

1:17Joey: And so you just unfreeze the weights and you say, okay, this stopped at step 8 ,631. Now here's step 8 ,632. And he's describing how they basically took open weight models and then spent like eight months shooting their own production data from their own footage and then use that to post-train an open weight model. He didn't say which one. I'm guessing at the time, this was what, a year or two ago, it would have probably been Juan 2, maybe before 2. I don't think 2.2.

1:45Addy: There are a ton of video models that are open from China, right? You have the Juan family.

1:52Joey: Yeah. You have the Quinn. Maybe an LTX, an early LTX. I'm just thinking because this would have been two-ish years ago and he was like, oh, hey, these models are kind of trained on, they kind of suck, but if we give it this cinema language and the cinema footage, we can post-train and get better stuff out of it. Interesting. I mean, if this was happening now, I think it would be clear that like Minimax would be the one to post-train on. Because I got some new updates, but like that seems to be what everyone is grabbing onto and post-training to build their own custom flavor.

2:22Addy: Yeah, that's like the new hot flavor of the day now. That's the new hot thing. Yeah, that's like. Two years ago, not as many good options. That's the new Ben and Jerry's flavor right now. But let me tell you, that confirms one thing though. Because Interpositive is a mystery to me, as I'm sure it is to you. Ben Affleck was so cryptic about it pre-acquisition, right? Like he didn't even mention that he had an AI company until we found out Netflix bought it, right? And then it's all, he's like, yeah, this is a great model. I think Interpositive is not a net new model. It feels like it is a post-trained model or perhaps a full-weight tune of an existing open-source model.

3:05What do you think? Yeah, well, he could dig into some of that detail. So, okay, so let's back up.

3:09Joey: So he, Ben Affleck, the filmmaker, Ben Affleck, the technologist guy. By the way, also in this interview, I did not realize he, in the 90s, tried to create an NLE competitor to Avid because he thought Avid's were too expensive. And he started his own company to build his own nonlinear editing system. And it just didn't work because the discs were not fast enough and they had drop frame issues.

3:30Addy: You sure you're not talking about Jeffrey Katzenberg? I mean, it's like the same vein of thought. It's like filmmakers that are just dipping their hand into really complex tech.

3:40Joey: Don't discount Ben Affleck. It's like, oh, Ben Affleck is like, I think, a secret James Cameron. Yeah, right, right. Being very technologically savvy. So, and he's like, yeah, I used to like build computers for gaming. I like, I know Python and stuff. And it's like, damn, Ben Affleck. All right.

3:57Addy: And he's Batman, like at the same time. And he's Batman.

4:02Joey: when like open AI was coming out and he was like, yeah, well, I use this, you know, privilege that I have where right or wrong, I could just call up places and be like, Hey, I'm Ben Affleck. Can I like check out what you're doing? And they're normally like, yeah, sure. And so he went over to open AI when like Dolly was in its first version. And then he said he had a panic attack and called Matt Damon and said, we're like screwed. And then he was like, well, as I kept messing with it, it was clear. Like these weren't, this is, this is not a threat.

4:29Addy: Yeah.

4:30Joey: This isn't the way filmmakers work but that was the idea of like okay can we fine-tune this to make something that can work with filmmakers but also um you know he didn't want to work with models that were trained on you know what he called his peers work so it seems like if he trains on open weight models he was fine if it was trained on like publicly available web data but not hollywood

4:52Addy: movies yeah that's his life yeah and yeah you can't really do that right unless you have a Netflix screen recorder hack where you're just going through Netflix movies and you're just grabbing them illegally to just push it into your... I've heard of people talk about that. I'm sure that's possible.

5:12Joey: I think that's where the light is of what you want to trade on. So he took one of these open models and then they shot a bunch of footage and they trained it. But the inner positive pitch is you shoot stuff for your movie, your movie, the footage you shoot with the movie is the quote training data, the dailies. You feed the dailies into their system and then you have this basically like bespoke custom model. It's not enough. It's not enough. On the footage. It's not enough. That's what seems kind of weird. It just is like, do you need, is this enough data? You need thousands of hours of video footage to have a full weight impact on the model.

5:49I mean, that's my understanding.

5:51Addy: Look, I'm not a... That's my first reaction too,

5:54Joey: but it's like, well, I don't know what they pulled off or what it's capable of. And when they had that demo video on the acquisition, their pitch was wire removal, a different camera angle, some VFX work.

6:07Addy: But dude, like C-Dance can do wire removals. You just put the keyframe in without the wire and then –

6:14Joey: Look, let's say – we could probably put C-Dance as like a different bucket because like, you know, one, it's okay. It depends where you're at of like training data, Chinese models, and then also – Oh, okay. Okay, so you're putting it in a different ethical bucket. I'll put it in an ethical usage bucket, geopolitical bucket. Like, yes, we know we can do that with CDance today. But let's say you wanted something that was local, walled off, secure, custom to your workflow or to your project.

6:45Addy: Yeah, movie studios are obsessed with local buying closed doors models. Like models work best at scale. They need the cloud, man.

6:53Joey: That's this balance of like, yeah, do you want protection or Soda, the best model possible for the job? Yeah, sure. Yeah, that is this balance back and forth. And also stuff that's still not clear yet of like rights usage. And I mean, look, if you're on any enterprise level, pretty much every enterprise contract, it says like they're not going to train on your data, you know, even if you send it to their API or their cloud. So that's pretty standard. That's like not it's not going to be some secret training on your data. I think there's still this fear of just like, what if it leaks? Or what if we give the new Batman footage and then, you know, somehow becomes in this training data?

7:32Addy: Can I go back to the post-training thing? Yeah. Okay. I want to make one note on the post-training thing. And viewers, please correct me here. Is Addy going off again? Okay. As far as I know, the reason they run the epochs, right? So basically, the training data set will run through the neural network once. It'll run again, a second epoch, and it'll do like, I don't know, a thousand epochs. Like it's insane the amount of times the same data runs through the network to strengthen those weights. The reason they stop at like the 8 ,000th epoch, as Ben Affleck said, and not the 8 ,001, is because after that, the model tends to overfit.

8:10Addy: So like any more training will actually harm the model. It'll produce results that are not as nuanced. it'll like blend Addie in with Joey like we're going to look like the same man you know like so there's a reason why training stops where it stops and post-training may not be the most elegant new thing that Ben Affleck is saying it is here it sounds good to people that don't know AI and how to train AI but to like somebody that's been training models you know they're like wait a minute, it's like taking a Ferrari and it's like, oh, we're going to change out those factory tires and put in these way wider tires.

8:53Addy: Well, the reason Ferrari has those tires is because that is what was engineered to go with the engine and the chassis and the weight of the car and everything else.

9:01Joey: You're saying that it's like these models, they were built for whatever they were trained on or however that was.

9:05Addy: They perform the best at that training stage is because that's where the AI researchers feel that that is where the mark has been hit. And they're constantly evaluating the model performance as the training happens. So like you can't come in after the fact and post-train and get this incredible result, like unlock something brand new. I don't think it works like that. I think you can post-train it to do certain things like what Allura does, you know, is like very special things. Like did they make Allura for the movie? Exactly. So like maybe that's what they did. I don't know enough about, again, it depends on the model.

9:44Joey: Well, how does something like what FAL has been doing with Minimax, where they built their own post-trained Minimax H3 Max, they took Minimax and they brought their own data, made what they're saying is a better model. And then I'll jump ahead to it, but Utipi, one of these other AI companies, they released their own, quote, model, but it's also a post-trained Minimax. Oh, Utipi is back. Dang, dude. They're back on the radar. I thought it was vaporware.

10:11Addy: They came out with the model. That's awesome.

10:14Joey: They came out with the model, but it's based on Minimax H3. So we got two kind of big models.

10:18Addy: Specifically Minimax, we'll pin the LTX stuff, LTX post-training stuff. I heard, I could be wrong, that Minimax is a distilled version of C-Dense. So it's not really a parent model. Like it's not an actual model, but it's just a neural network that mimics how C-Dense behaves. so you can give it further data to keep mimicking other models or other sort of video transforms. Does that make sense? Sort of. But you're saying that it's like different than if I was trying to post-train like WAN 2.12. Yeah, like the PR - An LTX open weight model. The prime models are the actual models and then the fast models are, I think the distilled versions of those.

11:01Addy: Minimax is not a prime model as far as I know. Again, correct me here. I'm not high confident on this. If we have any AI scientists in the comments, please let us know, because this is where I'm very curious about how this actually comes to be. Totally. So in that sense, I think somebody like Interpositive can further post-train Minimax specifically because it's not a real model to begin with. It's just mimicking model behavior, you know, if you take my assumptions with a grain of salt. But like if you take WAN 3.0, like a fully built out actual prime model, post-training it further, it's like a law of diminishing returns.

11:43Addy: You can only get so much more squeeze out of it.

11:47Joey: The other thing that also stood out to me is, and especially because this was probably a year or two years ago, is that model is definitely like 720p quality and trained on 720p and they're getting 720p outputs. How are they getting that to Netflix standards?

12:03Addy: Upscale.

12:04Joey: Upscale, upscale. But also the upscale models have only gotten, and the HDR models have only really flourished in the last three to six months.

12:13Addy: Yeah, like Topaz, Starlight, and Beeble has one, right? And Runway has one. Pretty much everyone has one now. Magnifique. But these are all relatively new. Yeah, you were saying like two years ago, how did 720p work for actual? Or he's like, we're shooting the cinema footage and then we're post-training it.

12:34Joey: But I know whatever they're going to get out of this thing is getting the shooting, I'm sure, in 4K. But whatever they're going to get out of this thing is like some 728-bit. The recurring thing we keep talking about. But it's like you're going to get this very low-fi web quality out of your stuff. Unless they also figured out some secret sauce back then to also up-res or get 4K. Yeah.

12:56Addy: I mean, this is like chalking up to you don't know what you do. don't know kind of thing because these are my reasonings. Follow along here. Netflix is full of AI ML talent and really smart people. Yes. $600 million is a lot of money. Yes. Those two things combined, you would not buy an inferior product. So I think they've invested in a positive because they do see some crazy potential and some secret sauce that we just don't know about.

13:27Joey: Yeah, I'm sure there is some secret sauce that's working. Also, he said it was used a bunch on his latest film, which was coming out on Netflix. He had a good analogy. I think Netflix has played this way well of being like, just telling you AI was used after the fact. And he was like, yeah, it was used. I don't want to tell you where and what. Because if a magician tells you, yeah, I got a second rabbit hidden in the table. Ruins the magic. Then you're looking at that. So you could watch the film and see if you could figure out where they used it. but yeah I mean he said it was used a bunch in his latest film so you can see it in action which one the one that takes place in Hialeah no no not the rip he's got a new one I think literally coming out oh okay

14:07Addy: alright I will watch every frame of that movie Animals oh that's right yeah it's called Animals the trailer looks good for that Hialeah that was gosh I forget The Run what was it called The Rip The Rip it was The Rip yeah you're right The Rip

14:22Joey: yeah The Runs hopefully you'll find out more about even more about Interposit, but this was a good interview and you should check it out he does know his stuff, man, yeah, you're right he is a James Cameron Jr. in that sense he's a secret James Cameron I was going to say Jeffrey Katzenberg but yeah, he's more like a James Cameron okay, so Unipi which I think we actually talked about them on the pod, we've always been a bit like what do they do? I've seen them pop up a few times of demos of like an AI studio or an AI tool. I thought they had some like avatar stuff, chatbot stuff earlier a couple years back.

15:00Joey: They do have an agent thing. Pi, production assistive intelligence. But anyways, they released a model Utipi X. And so this is similar post-trained on Minimax, same as FALS version. I have not used it. Also, this is saying no API available. So I'm not quite sure if it's going to just be in their walled off thing. But yeah, they built a post-trained model, going back to Ben Affleck. And even the Doug Liman Bitcoin stuff, where these platforms or these systems were developed two years ago, when they had to really fine-tune or really try to build out, come up with these technological achievements to get some progress.

15:41Joey: And then you look at today, and I am just curious how even more accessible doing something like this is now. we're basically we have a minimax model out and it's like clearly that is like best model if you want to post something and uh what would the ben affleck quality be yeah and then six months ago

15:59Addy: that was ltx 2.3 now l takes 2.5 is out and yeah there is always like an open weight model that is highly configurable tunable um you're saying like because it's a moving target like yeah i'm saying

16:13Joey: Well, I think it's also a combo of Claude and OpenAI have also gotten so much better where like maybe before when you would have had to have been an AI research scientist to figure out how to post-train this stuff. What can you get away with now where you just have Astra or Fable start figuring this out and doing it? Basically, like, do you need to be an interpositive today or is building out your own post-trade model custom to your workflow more doable by just having a conversation with Fable or Astra? I think you could do a lot.

16:42Addy: The limit would come down to capital because you would need a lot of money for the compute infrastructure to train a model. So obviously Ben Affleck and probably I think he's invested by Redbird Capital. So like there's probably hundreds of millions of dollars invested in Interpositive to train the model and then all that. So like average Joe couldn't come up with that kind of money. So they wouldn't. Right. Yeah, that hard limit of compute. You need compute. But I think. Train this. With a couple of, I think he has 16 people on his team. And he shot his own data set. So that's also big. Again, like, I don't think that's enough.

17:19Addy: A person can't shoot enough data in their lifetime.

17:23Joey: I think if you're trying to do a specific thing. Look, I'm curious, and I think this is something we could dig into more, but it's like, look, I mean, Beeble launched because they built their own. They took Paul DeBevick's paper. They built their own little light box, and they shot their own training data. And it was enough to build SwitchLite. SwitchLite. SwitchLite. SwitchLite. the original local process model for relighting. But, you know, that's like very specific to like, here's footage of a person, create the different lighting passes.

17:48Addy: Then Beeble probably tuned an existing model to get that performance. They didn't build it from scratch, right? But they built their own. They got their own training data. Yeah, right. You need your own training data for the fine tune or the deep tune. I mean, you could also, you know,

18:02Joey: synthetically make your own training data. That's what Corridor Key did. Yeah, right. So, you know, that's how you can, an easier way to get to scale for having a lot of options.

18:11Addy: It's that, yeah. Again, you're now requiring more compute and money before the actual model training starts because you're generating synthetic data. I'm done with Ben Affleck, I promise. Ben, if you're listening, come on the pod. Clarify everything. We have questions.

18:28Joey: We're a tech-friendly, filmmaker-friendly space. Yes. Okay, two new models came out. Both kind of do the same thing, and this also ties into a trend of what we've been talking about and seeing. Using Blender, block out your shots. There are two models specifically for this now. FAL has just been cranking out new post-trained versions of H3MX. So now they have one that is specifically for the input is supposed to be like a Blender gray box.

18:54Addy: Yeah, Blender play blast in and then final render out kind of thing. Yeah, that's great. That's awesome.

19:01Joey: Then same, LTX has been dropping some VFX tools this week and they have their own model that does the same exact thing.

19:06Addy: Yeah, I heard that LTX now supports ASUS, like ASUS Color Workflow.

19:13Joey: Okay, that seems to make sense because it seems like they're really dialing in on like VFX-specific pipelines. Now they also have their own key model, AlphaGen in beta. So yeah, this one, you can just give it a video clip and it'll key out and give you a nice composite. Those keys look clean. All right, other new model, Ideogram 4.5. Yeah. precision seems like now precision precision is the uh yeah table stakes exactly uh for most uh

19:40Addy: brand agency work marketing work you're coming into it with a preconceived notion or even some imagery and you're just wanting to do photoshop tasks on it right like in this case make that toaster red relight it and so on so nanobanana has been doing it for a couple years now but But now I believe Ideogram is open source or will be. Ideogram 4.5. So you can possibly run this on Comfy on your own. Their older models were open source. Were they?

20:10Joey: They put open weights out?

20:12Addy: I thought so.

20:13Joey: You know, you mentioned that. That does sound familiar. Or maybe, are we thinking of Flux?

20:16Addy: Maybe, yeah.

20:18Joey: No, no, no. You're right. Ideogram 4 is an open weight model.

20:20Addy: Yeah. I remember back in the day, and by that I mean like 18 months ago, So Ideogram was the jam for anything text-based, like if you wanted to make a product label. Yes, there's like the best graphic design model. Absolutely. So that was their strength. And now they've sort of opened it up to most image editing tasks, looks like, which is awesome. Yeah, and the quality has gone up there. Yeah, Ideogram 4.5, open wait soon. Yeah, oh, look, on their Twitter, they're saying they're comparing versus Sunburst and Nano Banana Pro. So, like, they're bumping heads with the big boys over here. That's interesting.

21:00Joey: Yeah. I mean, I think, look, Sunburst is really good. Nano Banana Pro, I mean, that's an ancient model by these standards. Right. But I would expect, you know, we're probably going to get a Nano Banana update. Yeah. Gemini 4, the LLM, just bumped up, right? Jumping ahead, man. Oh, my bad. Yeah, Gemini 4 Argon. Argon. Yeah, so, um... Noble? Obviously... We're in Noble Gases now.

21:25Addy: Xenon's next.

21:27Joey: Gemini has sort of been in a backseat of the conversation between all the updates with Anthropic and OpenAI. And they've just been releasing all these, like, little flash model updates. Right. But it seems like Gemini 4, according to their benchmarks, is new... New game. Will be. It's not released yet. It's still going through safety reviews and stuff. But once it's publicly out...

21:49Addy: Oh, look at that. It's ahead of GPT-6 Astra and 5.5, Opus 5.5, and Fable. God dang, dude. Yeah. I mean, I don't know what any of these numbers mean, but that seems impressive to me. I know. The Gemini column is blue, so it must be better. And it's ahead. It's like in front of all those guys. I'll have to find it.

22:11Joey: There was a really funny chart.

22:13Addy: Did you see that Demis Hassabis left DeepMind? Or he stepped down? Yeah. He did? I mean, double check me here, but I thought that was the news. I was kind of shocked at that because he's like a Nobel Prize winning scientist guy. Like he's extremely smart. I'll check that in a second. This was the joke graph that I saw.

22:33Joey: OpenAI model numbers are consistently higher than leading competitors, purely based off the model number of 6.1 versus 5.5. They're just graphing the numbers. They're just graphing the model numbers. That's awesome. Yeah, because sometimes when I look at these benchmark reports, it feels like that, where it's just like we're higher than all the other models, but I don't know what that means.

22:56Addy: Oh, you know what, Joey? The other day I used Opus 5. After our Opus 5.5 talk, I was so inspired. I went back and connected Blender to it, and then I wanted to do something different. I wanted to make a character rig of myself and then just put different outfits on myself and things like that. And it was terrible. It didn't work at all. So like all the Blender demos that you see, it's mostly environments and buildings because that stuff actually works. But when it comes to like a deformable rig that with like skeleton and things like that, it's not, I mean, maybe I'm just instructing it wrong, but yeah, it's not there.

23:32So you were trying to have it rig up a skeleton for a model of this?

23:36Addy: Yes, yeah, like a very basic animation rig. Like in that sense, it didn't work. What model were you trying to do this with? Opus 5.5 connected to Blender. Oh, okay. To Blender, okay. Okay, all right. I have not tried skeleton rigging, so.

23:51Joey: But buildings all day, all day long, baby. Okay, yeah, it does seem like this was back in August.

23:58Addy: Oh, okay. That was a while ago.

23:59Joey: He stepped down as CEO of DeepMind, but he is still the chair of DeepMind. Yeah. It just seemed like some slight reshuffling in executive leadership in DeepMind and Alphabet AI. Okay, yeah. You were correct. I'm in touch, dude. I keep track. I did not realize that. Yeah. Okay, so yeah, the Gemini 4 Argon. I'm curious what the pricing is going to compare like. The one thing I will say is Opus 5.5, as extremely capable as it's been compared to like Fable and Astra, the usage has been excellent. Because like Astra, Fable would just burn through my weak usage. Oh, you're saying affordability-wise.

24:41affordability and just how efficient it is with using its

24:44Joey: quote usage, which I mean would equate to using tokens and spending and stuff. My like weekly anthropic usage allotments It's like you walking into

24:54Addy: Taco Bell with$20. You're like, I can buy everything here. I can do anything. Yeah. Yeah, it's like that, huh?

25:02Joey: Yes. $20 goes a lot further with Opus 5.5 than any other models, but while also being like very capable at the stuff like I'm trying to do. I'm not a research scientist. I'm sure the problems we're throwing at it are way easier for the things that it can do or what it's capable of. Yeah.

25:21Addy: I mean, it's right now curing cancer, right? Curing aging. Not doing that. Like to make like a blender scene of Miami is like so low. And we're blown away by that. We're like, oh my God. please edit my podcast video it's out there trying to like uh so you know you know we're

25:43Joey: really close to nuclear fusion like i'm mapping the human genome oh yeah so there's like been a

25:49Addy: resurgence of a lot of startups so two two things non-mne that ai is helping with one is uh supersonic flight uh there's a boom supersonic and there's another one called overture i think that are now trying to bring the Concorde back, but obviously for today's world with like fuel efficiency and safety and all that stuff. So I think AI is helping with like the parts design and things like that, like modeling. And then the - I mean, I know these companies have been around for a while. Yeah, right, right. Like pre the AI boom, they've been doing these startups that have been trying to - Yeah, these things take like decades to get government approval to fly and things like that.

26:28Addy: The other thing is nuclear fusion is back in a big way because there has been some new breakthroughs, new designs, like physical designs of how the fusion can happen. And then AI has been helping with a lot of those computations, like really complex math.

26:45Joey: Yeah, that makes sense. And that's where like, I think these models really shine. Not so much for our, please make me a black model.

26:53Addy: Yeah, the other thing, like we always worried about like losing our jobs in M &E, right? Like VFX artists or da, da, da, da. the people that you would not expect are worried are like very high-end mathematicians like a Demis Hassabis type like PhD, dual PhD types because supposedly the new models like Opus and GPT-6 Astra are really good at theoretical math that hasn't been proven yet like extremely good that's crazy yeah like there's a graph that it started to tackle those problems in like 2024 like very recently. Before that, it wasn't even part of the training data. Like it wasn't built for that.

27:35Addy: And then there's like a spike on how good it is. It's like from 2024 to 26, it just shot up beyond human capability. That's awesome. Yeah, I mean, it's crazy what like real things

27:45Joey: it unlocks and makes possible. For good, for the good of society and not for, you know, destroying us.

27:51Addy: Well, yeah, I think you don't know which is which until it's too late. You know what I mean? The crazy thing is, is like that same model that they're using is the same model we're using with Blender and with scripting, with Previs, right? It's like one thing that could do that and this.

28:07Joey: Yeah, yeah. And then you don't have to have a specialty trained model. You could just have this one super model. And then we're having like three separate, but all like equally powerful and good models that are sort of like neck and neck. It's also kind of crazy. And yeah, last one this week, Flux 3 has a new image model.

28:25Addy: Yay, let's go Black Forest Labs. I'm a fan of Black Forest Labs. They're like the little guy against all these big giants, you know, and they always come out and deliver something amazing.

28:37Joey: This demo looks cool. This is like kind of just showing descriptive text boxes on a layout and then having the image build around. This kind of reminds me of CREA.

28:48Addy: Oh, like annotations and things like that?

28:51Joey: Yeah, annotations and just having it. There you go. drawing out and typing in what you want layout-wise and then...

28:58Addy: That's a new novel approach to composition. Rather than hand sketching, you just draw the box and then describe what should go in the box.

29:07Joey: Yeah, you could probably do something similar with the other models if you very roughly blocked out the image and then described it in the text prompt but sort of like drawing and prompting on the image and having it understand it. I don't know

29:20Addy: if I'm wrong or right about this, but I think... I tried to look for the Flux 3 image model, and I went to FAL, and I got Flux 3 video model instead.

29:30Joey: Yeah, that's the thing. Flux 3 video has been out. At the same time? The 3 video has been out before. The image model is the new release. Oh, got it. Okay. Release. Yeah, this came out yesterday. Okay. Control every pixel. Precise. Edits. Yeah, so precision is the game. Yeah. Yeah, so it might just not be on the API yet. I think it's rolling out because this came out yesterday. Okay, cool. OpenWeight's version of Flux 3 Image is launching in the next few weeks. Precision Edit is the name of the game now with image models.

30:02Addy: Like, I don't think anybody's even doing text to image anymore. No, unless you're like ideating. Yeah, but yeah, you have a set of images you already have and you're just kind of bringing it in. But yeah, if you're making a frame, it's like, yeah, I'm going to have a reference.

30:15Joey: I'm going to have like location references, character references, and I'm going to need to arrange that in a specific way.

30:21Addy: Yeah, and in our world, I always equate image reference as camera screenshots, like frame grabs from the camera. Like bring that in, do your task on it, and then feed it into a video model as keyframes.

30:35Joey: Yeah, and also keyframes, reference frames. I feel like I rarely rely on like starter end frame anymore. It's more like reference as default.

30:43Addy: One thing Flux 3 and Runway LF2 does really well is adhere to keyframes temporally. So you have, let's say, a keyframe for a 100-frame shot. You say frame zero goes at zero, keyframe two goes at the 23rd frame and so on. And it'll stick to it pretty well, which I think is really usable.

31:05Joey: Yeah, that's a good, yeah, if you need to hit very specific moments in a shot. I know Luma is also very big on keyframes. That's one of their differentiators. Right. Cool. Anything else?

31:16Addy: No, that's it. Just been really enjoying the support from our viewers and our audience. So the last video was a reaction from the TIFF video, and a bunch of us just kind of jumped in and was like, look, we know this stuff is not ready, but we're just following the journey. And that really resonated with me.

31:34Joey: Yeah, I think we've got a smart audience. The ones that understand it and get it, they see where this is going, and we're all kind of talking on the same page. Yeah. Hey, everybody, we're making the show because this is just a fun ride. We don't know where this is going to end up, So follow us, follow us for the journey. Follow for the adventure. Yeah. Thanks for everything we talked about at denoisepodcast.com. Thanks for watching. We'll catch you in the next episode.

From the publisher

Ben Affleck’s AI company InterPositive was recently acquired by Netflix, raising new questions about how filmmakers are post-training open-weight video models. Joey and Addy break down the technical details behind Affleck’s workflow, the ethics of mode...

More from Denoised

All episodes