SIGGRAPH 2026: Film Production and AI Research Meet

16:22Addy:trying to get in the first place.

16:23Joey:Yeah, I didn't take a picture of the output of the prompt, but yeah, it was just a long, long prompt. And I was thinking, because we were going back, we had that debate of if JSON prompting was useful, and I felt like it might have died down. But he found a very, I think, practical way to use it. And also, a human is not writing out the JSON structure.

16:48Addy:Yeah, I think for day-to-day, if you're hand-typing anything, don't even worry about JSON. But if a Vibe-coded app is dynamically generating the prompt, absolutely need JSON. It's going to be millions of characters and stuff.

17:04Joey:Yeah, so that was super cool. Really interesting, and I'm excited to see when that comes out. Oh, right, there was also a really good VES panel. It was future forward.

17:16Addy:Yeah. Yeah. So yeah, Paul DeBevick, legend. So I caught a little bit of it Monday morning. I walked into it a little bit late, but it felt, I was expecting Paul to just kind of go through the latest Eyeline research paper, which I think we're going to cover soon. But it wasn't that. It was more of like a very interesting take on the history of VFX, homage to VFX, if you will, and where we're going to go into the future with AI. So yeah, I thought it was presented well.

17:52Joey:The Perpetual Pipeline, 50 years of symbiosis and computer graphics. And it was interesting talking about where is AI going to fit into the pipeline, and then still the issues of like, yeah, it's also good at solving some problems, but you still have to know, you still have to have that base knowledge to know when the advice is wrong or when it's missing stuff. Because I forgot there were some examples that someone gave of asking it to solve like some lighting issues in Unreal or Maya, and the advice I was giving was just completely incorrect. Still having that base level of knowledge where you can speed up your process, you can speed up reviewing papers, you can speed up brainstorming, but you still need to know if it's accurate or if it's right and can get you to that point.

18:34Addy:Yeah, you know, the most confusing thing about AI in general is it just doesn't say, I don't know. It'll just keep computing and giving you the wrong answer and then just so confidently say, like, this is it. Here you go.

18:49Joey:Yeah, you have to double check. But also, I found, like, it's good if you ask it to check its work because then it'll sometimes find things, especially with the newer stuff like Fable and Opus 5 just came out. I don't know if you saw that. But now there's an open file. Yeah, yeah.

19:04Addy:I've been messing with Opus 5.

19:06Joey:Yeah, that's how it was good. What else have we got? All right, we had the NVIDIA. They had that cool party on Sunday at Datalab. Datalab? This new event space down in downtown LA.

19:19Addy:Yeah, I got to see you in person. That was nice, even though we live in the same city. So DataLand is this new venue supposedly just opened last month. And it is a really interesting take on what a visual and oral experience could be. And AI-driven, data-driven.

19:41Joey:Because you wore these smell necklaces. would get triggered and emit smells based on the digital environments that you were looking at.

19:53Addy:Yeah, so you walk in and it's very unassuming, right? Like you're just looking at LED walls and stuff. And like somebody like you and me, we've looked at plenty of LED walls. And then you walk in, you got to put on this wrist fitness tracker thing, this olfactory neck thing. and then off you go into different rooms with different experiences and media. And the media is dynamically generated based on your data. The whole thing is just really unique and very new as a concept.

20:24Joey:Yeah, it was one of those kind of immersive LED experiences. But the projections were cool. The art was great. And yeah, like you said, it's tracking you. It's kind of making it personalized.

20:34Addy:Yeah, Rafiq Anadol's art is in there. He's that famous AI artist. His work is at a lot of places. And then the closest comparison I can make to Dataland is if any of you have ever been to Team Labs in Tokyo. So Team Labs is along the lines of turning giant LED cubes and volumes into really entertaining experiences. And also Luminarium. You mentioned Luminarium.

21:02Joey:We did also have a joke that this is basically collecting all our data to clone us somewhere in the background. They're not only collecting your data, but they're charging you money to collect your own data. What else? Recap. Oh, yeah. Sunday, I went to Open Source Day, which we talked about last week. That was a great event and really interesting and just kind of cool to see different projects using various open source tools from the Academy Software Foundation. There were some projects that I didn't even know about. So they have different tiers of some open source standards are considered, I forgot the terminology, but established.

21:34Joey:Some are still in the assessment phase, even though they're kind of widely established, like ACES. Color is still technically in this middle assessment stage. Still being worked on, yeah. And then there's the, I forgot what they call it, sandbox or just experimental phase, where it's just very new projects that are still getting worked on and seeing if there's more use for them. And there are a couple interesting ones that I want to try out later, which was like some was like a dailies review tool, Another one was, I don't think it was fully for like provenance tracking for AI. I think it was more around VFX, but it seemed like it could be adapted for kind of this AI work future of like a standard that can kind of.

22:18Addy:Yeah, like governance. Exactly.

22:21Joey:And that was definitely a theme that came up a lot throughout the conference was I feel like we haven't we don't talk as much about the legal concerns.

22:30Addy:We don't talk, yeah, but that is a big concern for studios is like, how do I monitor everybody using the models? What are they prompting for? Where are the like IP blockers and things like that, some safety controls, and when are they being invoked? So like just having a bird's eye view of what the AI system is actually doing across like hundreds or thousands of people using it at the same time. They're interested in that because, you know, it can be used wrongly.

22:59Joey:Yeah, that, and I think it's still, everything's still getting sorted, or there is no clarity yet on, like, what is okay and not okay, you know, with usage. And so I think a lot of it's just save everything possible so that if you do need to prove stuff for now, later you have the paper trail. You can track, like, in this video shot, like, what was every single input, and then for those images, what was every single input into that. Very important to track that. if it will come into play and if that'll be like the standard going forward, if it'll be needed, if like more focus will be on the outputs rather than the inputs, you know, that might play out differently in a few years.

23:39Joey:But for now, everyone is just like if they're using it, everything's got to be tracked and logged.

23:43Addy:Yeah. For professional use, I mean, yeah, when you're going to make a movie and that movie is going to go out to distribution, you're going to have to show a receipt for every shot and every origin, right, of everything that's in the movie in the final pixel. So it's pretty easy to do that with computer graphics. We have assets, right, literally like FBX files, USD files, you know, Premiere projects or whatnot. But it's difficult with AI because everything is either an image or video. And if the metadata doesn't have enough stuff in it, then it's kind of a mystery as to where it came from. Yeah, you need to know.

24:16Addy:What else we got?

24:16Joey:Oh, yeah, Beeble also gave a presentation. That was a packed room. And that's Hoon, yes. That's Hoon, right? Hoon from Beeble? Yeah, yeah. Big fan of Hoon. And they gave a couple demos of some projects that I hadn't seen. One was sort of this Korean production, and they're basically kind of shooting this film in their office with office-y backgrounds, not even clean plate backgrounds. Yeah, like cubicles and stuff. And the video-to-video looked great, you know, on a small projector frame. But the stuff that they're doing looked great. Another project. Actually, this was an interesting, something that I think ties into what we were talking about with Codex last week.

24:57Joey:They did a 3D scan. I think it might have been with Polycam. They did a 3D scan of like a historic space. And then they turned it into a mesh, but the mesh was like super blobby and messy from the scan because it was like, I'm going to guess it was probably like an iPhone scan. It wasn't the best quality. and then they either used Codex or Clod and just said clean up the mesh and then it just cleaned it up. And I was like, oh, that was a great use of, also a great use of this, of like if you have an existing scan and you just want to clean it up. So that was a good tip.

25:35Addy:That's so interesting. I mean, your Codex Blender thing still blows my mind even now, two weeks later. But also, how does Codex know to clean up meshes? That's such a human-driven task. I don't know. That's crazy.

25:49Joey:Does this look good? No. Okay, keep fixing it. I don't know. Right.

25:53Addy:Like, who's supervising that, right? I'm a fan of Switch X. I told you a couple of weeks ago that I was testing it around. Finally, I got around to getting a Beeple subscription and testing it out. The driving plate stuff, oh, my God, it's so good. And that's me coming from the LED volume driving plate world where I've seen car processes unfold in front of my eyes. looking at it on the onset monitor and things like that and thinking, wow, this is the future. Like the LED wall is putting all those reflections down on the windshield and everything that's so realistic. How can you replicate this with anything else?

26:32Addy:And here we are just a year or two later. Switch X is definitely, you know, challenging that, everything that I just told you.

26:40Joey:Yeah, and the thing with the one limit with SwitchX was like you're kind of getting these, you know, 8-bit 1080 kind of web output. So that was like a limit. Right. They teased or talked about SwitchX 2.0, which I got up on the screen right now. And so this is coming soon-ish. And this is, you know, a step in the right, in the progressive direction. We're like, what if you could just do all of the effects and your output is SwitchX? And it's at a high enough quality that you can still work within a professional pipeline. So the next version is 4K outputs, 10-bit, and then some improved quality and stuff.

27:15Joey:And so that's definitely, you know, obviously in the right direction to, you know, if this could just handle the entire or give you an output that can fit back into the professional pipeline.

27:26Addy:Yeah, everybody knows what the gap is. It's like no matter how good of an AI video-to-video task you do, end of the day, just can't go back out into a professional pipeline just because, as you said, it's web quality video. I know Topaz is working on some kind of I'm going to say up dynamic ranging if you will probably, so it's good to know Beeble's working on it, as I'm sure Magnifique or all the upscalers out there are working on. Beeble has a separate

27:55Joey:HDR model that I've heard people really like as well, it's like Switch HDR, I think it's called

28:03Addy:Yeah, it's not enough to take the 8-bit dynamic range and then just stretch it to 10-bit because that's incorrect. And a lot of upscalers now will do that. You actually have to go into the scene with the VLM and detect the light sources. For example, if it's the sun, it's not going to be just multiplied by 2 or multiplied by 4. It's going to be like a thousand times brighter than like a light bulb. and so to have that accurate dynamic range i think that's the hard part that a vlm has to solve and solve it for every frame and then reinterpret it as the light source disappears and another light source appears into the video like it's a very complex task and i'll be i'll be really curious

28:51Joey:to see what people comes up yeah uh maybe we should do maybe in the future one we could do a HDR off because, yeah, the Switch HDR exists. And I've heard it's good because it gives you slider controls over the highlights and how much you want to increase stuff. Yeah, we could try that between that and Topaz. And LTX has an HDR up res as well. There's a couple out there we could do a bake-off.

29:18Addy:Yeah, just a real quick side note on the camera side of things. You and I are big camera nerds. Like the Sony FX5 just came out, which is like this$5 ,000 Venice, basically. It's insane. The thing has something called dual gain triple ISO. Okay, triple ISO. What it means, I think, yeah, so it has three different native ISO settings. Yeah. And the dual gain is, I think it reads the sensor twice every frame or something. Like supposedly an Aria-Lexa does that. So on one pass, it'll read the stuff that's really bright. And on the second pass, it'll read the stuff that's really low. Bracketing exposure in real time, sort of.

30:01Joey:On a frame-by-frame basis.

30:04Addy:Isn't that insane?

30:04Joey:Yeah, that's wild.

30:06Addy:On a$5 ,000 camera. So Sony claims that it has 16-plus stops of dynamic range. I mean, I just say that because, like, AI video generations are at best maybe 10 stops of dynamic range, something like that, right? Like, there is so much catching up to do where the camera technology is today, what we're accustomed to with dynamic range capture, and where AI video is today with what it can, you know, produce.

30:37Joey:Yeah. Speaking of Sony, the other thing Sony had at their booth, I didn't take any photos because they had a lot of no photo signs, so I honored that. But they had a HDR prototype monitor on display that would blind you. It's like 1 ,000 nits or something? No, it was 10 ,000. It was 10 ,000 nits.

30:58Addy:I think that's an IMAX projector, if I'm not mistaken.

31:01Joey:Basically, it was their tiles shrunk down into a monitor format. Right. But, yeah, you do need that.

31:11Addy:You do need that. As you go up from 10-bit to 12-bit, like, you know, your dynamic range expands, you're getting closer and closer to reality, what our eyes are custom to. And, you know, we're custom to daylight. Like, daylight is really bright. Yeah.

31:28Joey:I just remember Conrad was like, go check this out. And then I'll stand in front of it, and then it's like a driving, like Gran Turismo demo, I think. And it was like, go, then the sun comes out. And I was just like.

31:42Addy:Did you wear your Ray-Ban glasses?

31:44Joey:I was not then, but they had sunglasses on the table, chained to the table, so you didn't steal them. But they had sunglasses you could borrow to watch it.

31:52Addy:You walked out with a tan, and you're like, oh, I guess.

31:56Joey:It's like the scene from Oppenheimer.

31:59Addy:Yeah.

32:02Joey:So, yeah, it's coming.

32:04Addy:I'm glad it's coming. Yeah, like I was saying, the camera, the display world is pushing ahead. They have never slowed down. The AI models need to really catch up on that sense for sure.

32:17Joey:Yeah, last cool thing, there was a good kind of 4D Gaussian splatting demo, and I know a lot of people are excited about that. So that was at the Dell booth, 4DV, which I think we might have talked about a while ago. I think they did that demo video of the dude, the guy sitting in the chair, and then you're able to reposition the camera around him. But yeah, they shot their rig. I mean, it depends what you're shooting, but it was like 20-ish cameras. But yeah, a lot of interest in 4D Gaussian Splatty. And then, of course, back at the Kino event and stuff, There's a lot about the Kronos group, and they're trying to standardize, kind of create format standards for volumetric content.

33:01Joey:And then, yeah, there's also Skyrim, which was doing volumetric capture with a bunch of 8K cameras of sporting events and stuff. I know volumetric capture has kind of been an attempt at a thing for a while, but I think between 4D gauging splatting as a format coming up, And then also just, even if you're not capturing it, to create some immersive viewing experience, because I'm still, I don't know, maybe for sporting events it would work, but I feel like overall we tried that with headsets and it's kind of cumbersome. But I think being able to capture stuff in 4D and have this volumetric video that you can then use as a base knowledge to send it to other pipelines and outputs, I think there's something there for future workflows.

33:43Addy:The pipe dream was having the Cosm experience at home, right? Imagine if you had a sports broadcast where you can pick the perspective where you want to witness it at. So you can turn an NBA game into a courtside experience on your TV at home.

34:04Joey:I get the idea of that. I feel like that would work if there was still someone mixing the Cosm experience. Like, Cosm works because you have these really cool immersive angles, but you still have someone.

34:21Addy:Cosm works because it's Cosm. I mean, it's just a massive investment. No, the venue and the camera angles and the stuff.

34:27Joey:But I'm saying, like, I feel like when you have to do work and you're like, I have to navigate my joystick and then I've got to kind of follow the ball around. You need someone who is technical directing your immersive experience for you. I feel like when it's too much work, then it's not as fun, if that makes sense. But if you can volumetrically capture a basketball game, but then you can recreate those dynamic Cosm angles, and it's dynamically switching. If they're on one end of the court, you see it from here. and then if they're on the other end, you see it from here.

35:03Addy:Like when someone's about to go up for a dunk, the camera switches to behind. Yeah, I don't have to take my joystick. And then switches back, yeah.

35:13Joey:I don't have to take my joystick and then try to navigate to get the good angle. You're right. Someone else is doing that for me or someone else is using this volumetric data to create that Cosm experience for me as a viewer. I think that would work. That would be a sweet spot.

35:29Addy:I think if you had one human camera operator sort of just like on a video mixer, you know, just press buttons and then your broadcast would essentially follow those button presses, camera angles that switch, that could work. I mean, other than like an AI agent, but then how would you train that agent and how, you know, that goes into a whole different thing.

35:54Joey:I'm sure you could, man. Out of all the stuff, that seems the easiest to follow the action. Follow the action, sure.

36:01Addy:But it's a real-time task is what I'm saying. Highly real-time. It has to be within milliseconds. Like something that would be able to switch. And that's why I broadcast. You have these giant video matrices where somebody's actively switching the camera with physical buttons. Because that stuff has to be done in milliseconds.

36:21Joey:Ross, Grass Valley. All that, yeah. Blackmagic. yeah that's fun that stuff's crazy that stuff's crazy that's the world for the one switching they got like hundreds of cameras exactly that stuff blows my mind exactly they're able to cut you know how hard it is to cut the action and so like it logically makes sense to you but all this stuff is happening split second yeah it's crazy stuff technical director is like crazy yeah

36:43Addy:and that's what i'm saying like i don't think an ai agent will be that sophisticated like even the best of us can't even do it somebody needs years of training and experience to be able to switch

36:52Joey:cameras that well clip that add it to the archive come back to this yeah we'll come back to this we're at it nab and it's like this is the ross is introducing your ai switching yeah right exactly

37:05Addy:because uh i'm still hung up on the fact that that codex can um clean up noisy meshes like that

37:12Joey:was an art build a 3d environment right it's crazy so yeah maybe i will eat my own words very soon

37:19Addy:Yeah, we have to make a clip archive to come back to this.

37:24Joey:Okay, I think that's it for SIGGRAPH-ish.

37:26Addy:All right, real quick, real quick, real quick. Neil Blokamp movie, have you seen it?

37:33Joey:I saw that it popped up. I did not actually watch the thing.

37:36Addy:I want to give a shout out to Matt Workman

37:37Joey:because I follow his podcast and he had a really good episode.

37:41Addy:He had all this information. I was like, Matt, where did you get this stuff?

37:45Joey:Okay, what did he just ask? I know Neil Blokkamp did an AI, like he's posted like an AI short film he did. And he's like, I like AI. It makes me, I can make things.

37:53Addy:So the entire thing was made with C-Dense 2. And I guess ByteDense had a big ceremony or event in Singapore where Neil Blokkamp came up on stage and just kind of announced the project or whatnot. So when you first start watching this movie, the first thing that went through my mind was like, is this real with like insert shots in AI? because that shot looks incredibly real. And then the talking head stuff, and I know you do so many talking heads, right, as a documentary filmmaker, like that one, if you pause on that. The lighting on him and his performance and everything, I was like, wait, so they shot talking head, they shot some B-roll, and then they probably cut it in with AI action sequences because those would be the VFX shots.

38:41Addy:No, dude, the entire thing was synthetic, and I couldn't believe it.

38:46Joey:but didn't they use i mean didn't he use performance capture right so film actors i

38:51Addy:didn't know how many like i didn't know if this guy was real or not so i watched it to the end the credits actually have i don't know something like 20 people on so like there was quite a few

39:02Joey:actors that were captured maybe we should also like have given some context the old block camp

39:06Addy:who did district nine and chappy district nine um the movie with matt damon one of my faves um district no no no matt damon was he wasn't he was the other one yeah the one with where he has like a cyborg attachment behind him gosh i'm blanking on the name i know it was like the image of him

39:27Joey:that you're talking about yeah so i'm here it says featuring yes so these are i'm guessing these are the actors right and it's like what did he do performance capture are these also just voice

39:36Addy:actors i'm gonna i'm just gonna i'm just gonna guess here i'm just gonna guess here i think uh these the people here were scanned or photographed to create the digital ai version of and then they definitely did the voice acting because the voice sounded really authentic to me but i don't know if it was fully performance capture driven like i don't think the body and the motion really came from them right yeah like this stuff uh i thought that was warner

40:04Joey:Herzog by the way. Sea Dance is good because like it can take audio references and so and use that to drive. Also who knows maybe I don't know if they talked about it. I mean they probably would have hyped it if they was but like maybe he maybe he was using Sea Dance 2.5.

40:19Addy:Yeah and the thing that is always a telltale sign of AI generated movies is like the cuts don't 100 % line up like continuity is is really difficult with shot to shot to shot. But here as I was watching this there was like that distraction went away just like a real movie you don't have to worry about shot continuity unless they you know intentionally put it in there i think that also just comes down to

40:44Joey:like an actual filmmaker and editor and professional team knowing how to use this like that is more when i see some of the eye stuff and it's just like shot cut like shot dialogue cut shot dialogue cut and it's like that's just lack of editing yeah experience and style the other

41:03Addy:thing i always complain about the the lens thing right like how um ai generations are just not

41:09Joey:quite they're like there's an uncanny valley of optics if you will right like it's just had a

41:14Addy:nice subtle like it has but like the back wall is maybe a little too close and this this entire cabin can't be this tight like in a and then that roof probably can't be that low off in this yeah so like spatial things like my eye really picks up but then some of the outdoor shots like um when there flying in the Apache and stuff, it looks so real. Like I couldn't believe. The motion jitter in this feels a little off. Like here, as they're walking, the truck in the back should be a little bit more out of focus. It depends on the lens, I guess. But yeah, it does give away a little bit. But overall, I thought this was one of the highest quality, if not the highest quality AI film that I've ever seen.

41:56Joey:All right. I watched the whole thing. I haven't watched the full thing in its entirety it it's funny now we're nitpicking at like uh the height of the ceiling and uh the motion blur but like this came out like a year ago everyone would have like everyone's brains would have melted exactly no i was like well you know it looks good but the emotions the motion blur is still a little bit you're right yeah let me let me let me recalibrate this is the worst it'll ever

42:22Addy:this is the worst it'll ever be i i think this if this was a 90 minute film and it was in a movie theater i would 100 pay a ticket and go watch it i love neil blow camp so like for me that's a no-brainer and i couldn't tell if it was ai or not for the most part that's probably a better test

42:40Joey:when you if you just show this to someone you're just like hey watch this and you like to take out all the you know any ai mention and then see if they like notice something no yeah not just someone

42:50Addy:but someone who's accustomed to his style of filmmaking, right? Like that handheld stuff that he does, the talking hand, that's all from like Chappie or, you know, District 9. Like that's his style. Yeah, because somebody was like, I like District 9.

43:03Joey:You're like, oh, hey, the director that made a new short film, like go watch it. And then, you know, I think as soon as you preface anything that just says, you know, that AI was involved in, it just completely warps their viewing experience and assessment of it. But I think Netflix has been doing the smart move of just putting the stuff out and then after the fact being like, I was using it, by the way.

43:25Addy:Yeah, by the way, 300 films. 300 movies.

43:28Joey:And it's like, you will never be able to pick, identify what had what.

43:32Addy:Yeah, by the way, that name of the movie is Elysium. One of the greatest sci-fi movies of all time. Yeah, for sure.

43:38Joey:All right, real quick. Flux 3. Flux 3. Is it the video world? So they released Flux 3 model. Flux has just been images. but now Flux 3 can also do video, and the demos I've seen look really good.

43:54Addy:All right, let's look at some examples.

43:57Joey:And now Flux is interesting because this is Black Forest Labs. They split off from Stable, right? Were the founders officially from Stability?

44:08Addy:Yeah, they didn't officially. Yeah, so they're in Germany and not too far from London. And so during the early days of stability AI, like 1.0 when Imad was the CEO, some of the founders of Black Forest Labs worked there and then they just spun off and did their own thing. I don't think it had anything to do with Imad or the leadership, but once you have AI research knowledge, I guess you can do what you want with it.

44:39Joey:they're also doing some of the coolest stuff for being a relatively independent small AI lab

44:47Addy:they're relatively small, relatively independent but they punch way above their weight

44:52Joey:this is not backed by a massive company and the stuff they put out is really impressive

44:57Addy:like FluxDev and FluxSchnell is probably the number two and number three open source image model from the early days of all of this, right? You had SDXL LSD 1.5, FlexDev, and FlexSchnell were the four image models that you can mess around in comfy before all the other ones were available.

45:17Joey:With the video model, up to 20 second duration, with audio, you could do text video, image to video, video to video, keyframe to video, which is nice, multilingual dialogue, and And then a high style diversity, high style diversity and strong typography.

45:38Addy:Yeah, look, hey, they have some Omni and C Dance comparisons, of course, and Happy Horse, which we didn't cover on the pod, but it is one of the better video models.

45:48Joey:You like Happy Horse? I think we mentioned it. That's Alibaba's video model. Yes. Have you been using it or experimenting with it?

45:57Addy:Just to hear it. I mean, just to see what it can do, but not regularly. Like I would probably use Google Omni more than anything else on this list. Yeah. And actually, no, I take that back. I use runway as well.

46:09Joey:Yeah. Okay. Runway. I mean, Omni I found was like very good at a modification of video.

46:17Addy:Yes. Most of my work, you know, that I'm testing with is generally video to video rather than novel video generation. So I would already have something. I just want to modify it. Yeah. Yeah, Omni's really good at that.

46:30Joey:I need to give... I have heard good things about Luma Ray 3.2. I found like Ray 3 was a little fuzzy, warpy, but I've heard some pretty good things about 3.2, so I've got to give that another revisit.

46:44Addy:Yeah. I also have to test out Luma Agent a little bit more. I've heard good things about that as well.

46:53Joey:Yeah, so another solid model. I'm going to give this one a shot once it's out.

46:58Addy:We'll give this one a shot. You know, I say this every time as we see a new video model or an image model. Like, the work begins years before. Like, to figure out what the architecture is, get the data, label the data, put it through training, you know, put IP blockers, all that stuff. Like, this stuff is really labor intensive to make a video model. By the time you complete it, you put it out on the market, then see dense 2.5 drops. It's like, oh my God, the timing couldn't be worse. So this is the unfortunate thing about AI work is you can never win. But if you have a good model and if you have the customer base that loves it, hopefully there is a business case for it.

47:44Yeah, yeah, yeah. And I believe they are going to open the weights so you can just run this.

47:53Joey:I think I saw that.

47:57Addy:okay so there's going to be an open source model did you see jensen um tweeted uh fighting words yeah like the world is a place for open models and closed models he's like

48:13Joey:he's like the world ai president did you see who also got into the uh open open source AI model debate and back Jensen?

48:23Addy:No.

48:23Joey:Another company that you would 100 %... Denny's? Denny's? Sells chips. Van Dyke. Knows the importance to stay in open. Well, isn't that also because it wasn't part of NVIDIA's origin story? Didn't they... Doesn't he love Denny's?

48:40Addy:Jensen loves Denny's? That sounds very Jensen. He's like an everyday man, dude. He's like the least non-billionaire billionaire. he's the most commoner billionaire i say that like i really like the guy and uh yeah that

48:54Joey:totally tracks yeah he's just focused on on work and the mission he's just like trying to yeah make the best shit yeah yeah uh all right good place to wrap it up on denny's yeah hey uh if

49:06Addy:you're listening to the podcast on apple or spotify do us a favor just hit the subscribe button it won't hurt it'll only take a second thanks for watching thanks for everything we talked about

49:16Joey:at denoisepodcast.com. We'll catch you in the next episode.