Breaking down Google's BIG AI Updates: Veo 3, Flow, and More!

0:00Real-time AI is going to be such a huge frontier. Yeah, I'm more excited about that than VO3. I agree with you. I agree with you. In this episode of The Noise, we're going to break down all of the new AI updates from Google. Exciting new real-time video generation from Viggo AI. And Johnny Ives joins OpenAI. Let's get into it.

0:22Welcome back. I'm Joey Dowd. I'm Addy Ghani. Good to see you again, Addy. We missed. We missed. Movies. Movies. Movies. Movies. Oh man, I'm so bummed that show ended. Now it ended. Now we have, we're going to run out of studio references to drop in the episode. Yeah, so we missed, earlier this week, we were both traveling out of town. I was down at Tampa at Vue, checking out some new product updates from Vue and sort of the future that they're looking at with the future computer interface and how we create with AI. A lot of interesting stuff, but in a lot of real time, like they're just content at the speed of thought or creating at the speed of thought.

0:56Yeah. And ties into a lot of AI, which also ties into what we're talking about today, but like real-time speed of thought, but also the limitations are not necessarily them, but just like the speed of these models. And that's also kind of a common theme today that we'll talk about. Yeah, I think the Vue team know that latency and the speed issues will eventually solve itself over time. It's, you know, it's not the problem on there, and it's how everything's being generated. Right, how these things work, or if you could do it locally, or how that works. Shout out to Tim and the team. Yeah, yeah, great to see them, and great to, and yeah, I'll have some more updates on the actual cool tech and stuff we checked out, but yeah, I was down there, and then over the weekend And filmed season two of Cinema Synthetica, which is a 48-hour AI filmmaking competition.

1:33Nice. That three teams have to make a film over the weekend, and it's going to play next week at AI on the lot. Nice. And we did this last year. We filmed the three teams doing it. They get a script. They often do the same scripts and interpret it however they want, and then use a variety of AI tools, whatever they want, to make it. How long of a short film is it? The requirements are two to five minutes. Okay. Yeah. And so the interesting things that I saw from that, and again, this also ties into all of the updates that came out today. or this week, because it's my observation. Overall, it's like, yes, all the models are better.

2:03Like, for reference, when they did the competition last year when we were filming, I remember Sora was announced. Yes. No one had access to it. Can you believe that was only a year ago? That was only a year ago. And now everyone was like, oh, man, I wish we could use Sora. And so, you know, it was like runway Gen 3 and some of the older models. And like, the stuff looks pretty good, but had a lot of the issues. Obviously, now the stuff looks a lot better. But the overall creation process was still, hasn't changed much. You know, it's still a lot of text to image, image to video. And then with the video part, it's kind of a crapshoot of like what you're going to get, if it's what you want.

2:36It's just like a bunch of regenerations. Were any of the artists and the filmmakers using like a 3D tool to build the world and then use that as almost like a template? Yes, one team did. They used Cinema 4D to build a rough version of 3D world to frame up their shots and then use that, export the image. And then they had a whole comfy UI workflow to style the blocking of that image to their style. And then it still goes to image to video. And I feel like the image to video part is where you start losing that creative control. Because you get to the image, you're like, okay, this is still awesome.

3:16This is exactly what I wanted. And then the minute it goes into video realm. It's like, what am I going to get? And then they might have to run a second process to get the lip sync and sometimes composite it together or just use the output as is from something like Act 1. So it's still, that part didn't feel like it changed much and there's still a lot of room for development there. Well, transition into our big news with everything Google announced. Yeah, and that's what - Because I think Google has a tool in Flow that fixes that. I don't think so. Okay, so on the demo, I did see that. All right, so the big up.

3:48Well, two big AI-specific updates, Launch of Flow, a new product, and VO3. Yeah. An update to their video model, which VO2 was already impressive. VO3, stuff that looks even more photorealistic. I can't think of a better video generator today. That's photorealistic? No, I think that from the outputs that I've been seeing, extremely impressive. And it can also generate audio, speech, and sound effects that match the video in the same generation. Correct. And this is just like a fraction of announcements from the entire Google I.O. conference. Yeah, we're not going to get every single announcement.

4:22We'll get the AI-focused ones. But yeah, the big one in the AI video generation world, VO3 and Flow, which is a new sort of product that rolls up all of their AI creative tools into a single platform. We talked about it before, but like VO2, it was really hard to actually find and use. You had to like go into Google AI Labs and then call it up or access it through an API, through a third-party tool. now they built flow it is a standalone product you can get an ai studio pretty sleek interface yeah really nice looking interface you can has like a very basic video editor but you can also like extend clips you could do generate clips you can kind of reorder them yeah and you can call up the different models so you can use vo2 which now supports uh first frame and last frame for your video generation or you can also call up vo3 currently vo3 is just text to video so there are no image inputs.

5:13It is all text-based prompts. That'll obviously change in the future. But right in everything you're seeing with VO3, it is a text prompt and then you get the output. Tell me about the pricing. The pricing. This is definitely the most expensive model out there. I believe what I was seeing roughly, so the Google wants, to use Flow to access VO3 if you're not going through an API, you got to get a subscription and the subscription plans. There's Google AI Pro, which is$20 a month and that gets you some of the more. I think it gives you VO2. Yeah, the same flow pro. But then to step up to the big boys.

5:48$250 a month. They have a three-month deal where you can get it for$125 for three months. You get YouTube premium. You do get YouTube premium. That's like$20,$30. And you got 30 terabytes of storage. So you got some... 30 terabytes? They kind of threw in all their extras. Yeah, it says 30 terabytes of storage. That's a lot of storage. On your drive and stuff. Yeah, that's actually pretty good. But this has to be on your personal Gmail account. Because... No, like an enterprise workspace. Yeah, I'm a Google Workspace user and we get Gemini and stuff through that. But when I went to Flow, it was like, you gotta switch accounts.

6:17You can't be on your enterprise, on your workspace account. You gotta be on a personal Gmail account. So then there's gonna be a separate subscription. I'm sure eventually they'll roll out some options, some upsell charts. Can I just say like for a lot of people, it's totally worth it. The 250? Yeah. So how does this break down in the pricing? Looking, I remember seeing someone, I think saying that maybe gets you 80 generations without buying additional credits. It's going off the math off the API. It is about 50 cents a second for VO3 generation. 75 cents a second with audio. And the generations normally make you do at least five seconds minimum.

6:54So that's two to three dollars. But that's a pretty expensive slot machine handle because it might be crap if you get what you get. And then you got to do it again. For comparison, Runway Gen 4, which is, I still think, really, really good. about 12 to 16 cents a second. Oh, wow. Okay. I'm just thinking like the artist or the creative that's behind the console just prompting away. They're like, okay, I really have to get this right because this is 50 cents. Go. What if you like have a typo and you like hit enter back? No. Most expensive typo. Weirdly though, on their pricing page for the API, VO2 is also the same exact price as VO3, which I don't quite understand.

7:38And yeah, maybe, I don't know, maybe that is a typo. So pricey-ish. But yeah, so you can get Flow, you can get the professional, you can get the premium$250 a month package. And yeah, generate images, generate video. The Flow looks like a very sleek product for working with your video. It reminds me a bit of what Sora built, which I think Sora's interface was more interesting than the actual model itself. Yeah. What were you thinking? You were talking about you thought that Flow could do something. I think if Cinema Synthetica folks had flow, one of the features that I even tested it, obviously, I saw on the announcement was you take a clip and it's on the timeline and you can kind of, if you, let's say, like the first half of the clip and the second half of the clip just really goes into a direction you don't want, you can just trim it and then re-prompt it and it'll regenerate.

8:29Yes. You can basically... A clip with the first identical half and then a second new half. Yeah, it's basically like a hack where if you took the frame from the middle video and then use that as your first frame, but just doing it in a very nice, easy interface. And it's quick and yeah, it's sleek. I think one of the biggest problems with video generation is temporal adherence and temporal, like a constant sort of predictability over time. Like if you look at somebody that's walking down a street, you know, the first two seconds, it's exactly what you want. And then he's like, well, no, I don't want him to turn that corner.

9:01No, no, no, he's not. You're like, okay, I need to trim that, restart it. I just want him to walk down. Or you edit it. I mean, a lot of times, like five seconds is a pretty long time. And so like, you're editing your stuff properly. Yeah, most a lot of times, you aren't gonna need the full five seconds anyways. Yeah, I think this is a really primetime ready tool for a lot of YouTubers who just need b-roll. It is a good b-roll generator. That's what I that's that that's a big point I get to like for 250 a month. It's totally worth it. But seeing this, I'm just like, okay, cool ish. It looks more realistic.

9:31But as far as the control and what you could do with this. Like, look at all the Johnny Harris videos. Like, I'm sure some motion graphics person is putting all that together, and it's not$250 total. It's probably in the thousands of dollars. For a motion graphics person. Yeah, and yeah, we're talking about Runway's use case for that, and especially with their first, middle, last frame. Yeah, but to push this tool into making an entirely synthetic film, if you will, I think it's just still a little bit too early. No, I mean, well, currently, definitely right now vo3 is just text to video um there was a kind of a cool demo video uh from dave clark because he had access to this tool and he did the uh some shout out to dave clark yeah and it was like a swat team heist kind of short film and it looked good but you know looking at it if you're like not paying attention it like looks pretty consistent but i'm guessing it's just like very detailed prompts to just kind of get the characters consistent also like swat is like people in black.

10:29It had like a Sicario, like that intro sequence of Sicario, like that kind of feel to it. But yeah, it's just very clever prompting to get it consistent. Obviously this will change and be able to do first image and have a little bit more consistency. But yeah, that's still the problem. And also just going back to the, what I was witnessing with the workflows, it's like the models are better. Stuff looks more realistic, but the workflows themselves of text to image, image to video is not evolving as much as the quality of the models. And I think there's still like tons of room for camera control motion.

10:59You know, Runway sort of started introducing that. In Gen 3, it's not in Gen 4 yet. But that's still like, there's still so much to do there. Yeah, absolutely. I think if you're talking about like a pipeline, a pipeline is the same. It's just the tools are getting better. But somebody needs to just reevaluate the entire pipeline. How do we actually use this? Companies are doing that, but they're just using very complicated workarounds, like very complicated comfy UI processes and you do separating background from characters because if you handle them separately, it works a lot better. Yeah. But yeah, it's a very cumbersome process.

11:33Cumbersome at the moment. Yeah. For sure. And that goes to show where like, I think we talked about this in a couple of episodes back is we're stepping up from a world of foundational models, which was last year. Everybody had their model. This model is better than that model. VO2, VO3, and so on. Now we're at like a workflow layer where it's like, okay, you go from image to video, video to edit, edit to sound, and that's a workflow. I think we'll eventually, in a year or two, get to an application layer where Google has a dedicated application that you download on your desktop or something or run on a browser that has a much sleeker interface, combines multiple workflows, combines storage.

12:17And just think of DaVinci Resolve as a sophisticated application. Like, we need an application at that level for professional filmmakers to engage with. Or someone builds a third party and rolls in their APIs. Totally. And, you know, makes it a tool. There's such a room for somebody to come in. Because, like, I don't know if Google's going to build that outside. Like, I think Flow is probably the biggest product they'll build. Yeah. They're not in the application, M &E application game. But then Autodesk is and Adobe is. Right. I could see Adobe or someone like that incorporating it into their existing pipeline, their existing software to solve a lot of these issues.

12:51I think the most interesting, relevant use case that I saw that of like something that you could use VO3 today and that is like pretty good is a PJ Ace, I believe, posted it. It was a fake like pharmaceutical commercial. Yeah. For a dog. Yeah. For like a dog, like a fake dog medicine. Somebody crying in there? I don't remember that. Was that the one? Yeah. It's like the shots look realistic, but also it's the type of spot where you only need one or two shots from each scene because it's just like a person talking about a thing and like looked really good sounded pretty good and that's the type of use case where it's like you only need one shot of the of the person you're not like looking to make a bunch of consistent character shots so yeah advertising let me ask you just like uh is this really coming for jobs then it's coming for specific markets of jobs like advertising and i mean there's things that this can do that you know you just can't do at scale like if you wanted to create custom ads or very specific to a person there was a whole i don't even have in my notes here but i remember there was also another update from google that was like virtual try on you know which is a huge industry market yeah we've all shifted away from physical malls to online stores and one of the big challenges there is you just can't try stuff on anymore yeah i was looking at me and then you gotta buy it and you know i mean companies got clever like zappos where it's like just buy it and return for free just buy a bunch of stuff and i know on the website it's like you know the i'm sure you buy men's clothes online and it's like the model is 6 '1 wearing size medium i'm like i'm 6 '2 and i never wear size medium i just go i just go to uniqlo and throw my stuff in their little magic bin that automatically tallies everything up they have the best checkout system really you ever go to uniqlo no you just like to throw all your stuff in their bin and the bin just reads the rfid tags and that's it it's just like you put it all in one bin you need low physical stores yeah the physical store really yeah the self-checkout is like a big plastic bin you throw all your clothes in it and it automatically tallies them all up and it's like that's it that's so awesome just like scan your qr code and you're like bye that's that's such a guy way to shop i love it

14:53like if it's too hard i'm not buying new clothes yeah like i just go there yeah as as close to online shopping as you can get being in a physical store that's awesome i love that yeah all right so vo3 flow the biggest updates another interesting update or news announcement uh darren aronofsky yeah in the ai game officially yeah i think he calls it primordial oh the venture is titled primordial soup uh-huh and he's going to use google's uh products to make something entirely synthetic yeah they did i think it's a hybrid they're using ai tools uh but it's a film the first film under this production slate is ancestra which is premiering at tribeca and it's a there's a trailer online but it's a hybrid of ai generated it's a hybrid of synthetic shots and live action shots.

15:40That's exactly how it should be done. Yeah, and you can kind of tell in the trailer, it's like the synthetic stuff looks like, you know, his quick cuts to like microbes and like origin of the universe type shots. And then there's like live action stuff with the person in them. I think a lot of the synthetic shots will be replacing VFX as much as it can without going into traditional VFX and then all of the live traditional shots, quote unquote, will use like 2D motion graphics or something that is relatively easier. Yeah, I mean, I think this looks like it's just more like the hybrid approach is not even like blending the two it's just like in the editing process yeah some shots for it i mean shout out to dan aronofsky for uh postcards from earth like that's doing really well at the sphere yeah and it's like i mean he's been someone who has technologically pushed the boundaries and postcards from earth was the specially developed uh big sky 100 megapixel crazy ass camera that could shoot for the sphere you know a huge technological feat um and this is just another realm in the technological updates yeah for sure and this adds to the list of natasha leon who's making a film with asteria solutions i believe so like now we have two legit hollywood filmmakers that are working on something yeah that's adapting yeah uh ai technology yeah and then yeah other google updates that are not necessarily like video specific but i thought were interesting uh google beam which is i think it's a rebrand of something they've talked about a bit.

17:04Yeah, Project Starline. Yes, which was their 3D video booth. I've been following this innovation closely. So I have a friend, she was on that team like five years ago. So they've been working on it for a long time. And what it essentially is, is a video conference with two people where the TV has built-in stereoscopic cameras, I think six cameras in total. And using the different perspectives, they recreate the remote attendee with the Gaussian splats. That's my guess. Okay. And however, the Gaussian splats are really high quality so there's no splotchiness or, you know. Yeah, the thing that stood out to me was like watching the demo with their new product launch is just like taking 2D images and turning it into 3D-ish but like the looks in real time and looks good.

17:51It looks great. So like if you're sitting in front of this TV kiosk thing and you do this, you sway a little bit, you'll get parallax from the other side. Yeah, you can, Yeah. Kind of feel like you're looking at a hologram version of the person. And yeah, you know, just to clarify, like no headsets, no anything. It is a special booth. Like it is a. It's hardware. It's a dedicated hardware product booth. Yeah. And I'm sure it won't be cheap. No, this is for enterprise. But then if you think about global work, right, like companies now that are completely remote, they have their CEO in one town, CTO in one town.

18:22Like when you put those two executives into a meeting, you want to make sure they're really able to effectively communicate. Yeah. Yeah, you can have more of the same vibe of that in-person thing. This totally makes sense for a world that's post-COVID, post-remote work. Yeah, something where it's like, you want to have a better interaction than a Zoom call or a phone call, but it doesn't maybe necessarily warrant international travel or a full trip. It's still cheaper than a business class ticket for an executive to fly and do something. Yeah, it reminds me of the next gen. Remember the Cisco conference room?

18:56I worked on those a long time ago. You worked on those? Yes. Like the three TVs? Yeah. There was like a semi-circle? Yeah, this is how I got my start at DreamWorks. I came in as an engineer that worked on the original version of that. We called it Halo. Okay. And then HP bought that and licensed it and built like hundreds of those rooms. And then Cisco copied HP and then built a lot of those. So yeah, I worked on the very first iteration of that. and this was right after 9-11 when air travel was difficult but you still wanted that you know close interaction fast forward to like 20 years later we're seeing google beam and i'm just so happy that high-end video conferencing is still a thing yeah do those i feel like i have not seen i'm not that i'm in the corporate world that often but i feel like i haven't did those rooms die out they died out because uh zoom and well blue jeans came in first and zoom and like hangouts and And like the...

19:54It was like good enough and way cheaper. The problem with that was, first of all, it was really expensive. Like I think a Cisco room was like in the millions. Like a pair, you have to at least buy a pair. Because also you had to have their... Right. And did you have their routers and their codecs? Did you have to use their data service? I think they did have a plan for that, but you needed a dedicated line. Like you can just put it on public internet. This was early 2000. So like broadband wasn't as... It was T1. You remember T1? T1 is like$20 ,000 a month. Yeah. So I think we went from like a dedicated room in an office building to anybody's desktop and now to anybody's laptop and anybody's phone.

20:36So it just got more and more decentralized. Yeah. Over time. And then now it looks like maybe we're coming back with Flow. But yeah, I mean, the, or Beam. I mean, it's a cool product. I'm kind of, you know, it's kind of nerdy out of like interesting to see it, but also just the tech behind it and how that could be applied to media entertainment cases in the future. 100%. something that in real time could take 2D and turn it into faux 3D that looks really good. Yeah. So I'm curious to see just adapting the tech from that. Well, even if you just think about meetings between a director and actors or during dailies and editor and the creatives, like the quality of that meeting can go up significantly.

21:10And then the communication that happens there can significantly affect the creative decision making. So just as is, if you plug Google Beam into Hollywood Studios, like I think you'll see an uptick in the level of meetings and level of communication yeah that sounds for us is a good point too because it's like can you use this hardware as a replacement for a headset and can you load up something else like a 3d environment uh you know for like doing virtual tech scouts like could you if you have the gaussian splat scan of a location instead of sending headsets to everyone can you run it on on the on the beam you do a remote location scout yeah i mean which exists but yeah you got everyone has to have a headset to do it so like can you can this replace the headset so you don't have to deal with that set yeah i wonder if that display is a lenticular display.

21:53Are you familiar with those? I've heard this word, but explain it. Yeah, so it's basically a 3D TV without the goggles or the glasses. Like the Sony display. Yeah, it has these tiny prisms in the front, so if you're in a certain cone of viewing, then you're getting stereoscopic, but if you're sitting between the two cones, you're getting a messed up image. Is it like that, you've seen that Sony 3D display? Yeah. It's like at an angle, like a 45 degree angle, and it's like a computer display, but for 3D, but you don't have to have any headsets. Okay, I think that's using a beam splitter, which are typically at 45 degrees.

22:27Okay. A lenticular display will look like a normal display, but it has like a crystal coating on it that breaks the viewing angle into specific cones. Okay. Yeah. Very niche technology. The only reason I know it, because I worked on 3D TVs a long time ago. All right. The Sony one has a built-in camera that tracks your face, so it adjusts the position. Right. So the reason they need to know where people are is because the lenticulars need to focus the viewing angle on each pair of eyes. That's cool. I tried the Sony one out at the AWE and it was cool, but it felt like it worked better when it was like you're looking at a specific product or like an object because the object would sort of float off the display and you can rotate it.

23:08I wasn't quite sure how it would work for an environment. Yeah. I mean, I guess technically it could, but I felt like. You're just limited by GPU, CPU, right? I mean, just the experience, more of the experience of seeing it, like the 3D-ness of it felt better for like a single thing popping out of the screen. Yeah. Less so if I was trying to like navigate or look, be immersed in a 3D space. Like I feel like something like Beam or something that, a headset obviously works really well for that. Yeah. Beam, the display is like very big and like kind of surrounds the front of your face. I think we're finding out with Apple Vision Pro and some of these high-end VR, AR kit, like HoloLens, is people still don't want to wear the glasses.

23:44Yeah, that's a huge issue. Yeah. Although I was on the plane the other day, and it's always in first class. You walk in, and there is a guy who's doing this with an Apple Vision Pro. And like his arms are reaching into the aisle. I'm like, God, just not right now. I mean, you also like were leaving a tech event. So I feel like the likelihood of someone that immersed in tech on that plane would be very high. Absolutely. All right. The last thing that stood on my radar from Google, and this wasn't necessarily new, but I hadn't called it out before, SynthID, which is Google's way of embedding identification inside AI-generated media that you could then run it later and basically identify a way to check in the future if something was AI-generated.

24:33Yeah, so this is almost like a competitor to that Adobe. The authenticity. C-A-P or something like that. content authenticity program. Yeah. I mean, I feel like this is eventually, everyone's going to have to get together and create some sort of open standard or something that everyone can adapt because it's not really going to work out if every model company has their own watermark method and way to check. It's got to be a standard. Yeah. One of the big reasons they do that is for trust and safety. Like every time you roll out a AI solution to a big company, the first thing they're going to ask for is, is it safe?

25:06Is it going to generate something we don't want? And is it going to harm our brand? harmony company so like to track every generation totally makes sense yeah and yeah just for everyone to be able to verify especially for news footage real life footage i mean that was like everything on x has been a feed of like a lot of clips people are developing are real life looking clips like clips from a trade show or an interview clip after a sports game like stuff that looks news word like yeah for a news clip not necessarily cinematic shots or something that looks like something from a movie where you'd be like, oh, was this a Hollywood film or was this AI generated?

25:40It's more like, oh, is this like a real life clip or was this AI generated? Yeah, you could fool somebody pretty easily. Yeah. And then on the flip side, people started posting real clips from like real news events. Be like, look what I made on VO3. More of like a joke, but then it's just like, what is happening? And like, you just start questioning everything. And I even saw today, this morning, people questioning the story we'll talk about in a bit of the video with Sam Altman and Johnny Ives started questioning if that was AI generated. Totally could be. I think that a bunch of Google stuff, I think maybe that wraps that up.

26:13Yeah, it's a really exciting time for everybody in filming TV, I think, once we get our hands on VO3. I mean, it's probably exciting and scary. I know there's a lot of, you know, panic and scary comments around it. For me, I don't know, it's like, the quality does look better, but it doesn't really shift anything if you've been paying attention that we haven't already seen. Like, yes, the models are very good, Still the same limitations of how this could actually be implemented in something. But it's immediately ready to plug into certain use cases. I think we're going to be pretty far away from 100 % synthetically generative films, Hollywood-level films, for a while.

26:47Having said that, every incremental update with generative AI, like Google Flow today, is making the use case and the economic viability of this stuff more and more plausible. and if you kind of track that you almost are forcing upon artists a mandate to learn all this stuff and just retool themselves to use this we all have to kind of evolve into a hybrid of ourselves where we're still experts in the traditional tool sets we've used and also using all the brand new tools and then mixing the two we're the transitional generation right and i think movie making 30 years from now who knows what that'll look like but for us we come from the world of computer generated graphics and vfx and all the tools to now transitioning to these things and then we'll pass the torch to the future generation and i think a good point of like learning the workflow learning how this works is like yes all of these ai updates keep coming out quickly but if you learn the flow behind it yeah the advantage is once you kind of figure out a workflow that works for you.