Hollywood Is Upset About Quilty. But Is Any of This Actually New?

This is the Tilly Norwood of AI development stories. This is not new. There are a bunch of platforms. This is like the NFT scam all over again.

[instrumental music playing] All right. Welcome back to Denoised. Addy, how you doing? I'm doing just fine, sir.

How are you? I'm good. April Fool's Day, can't really trust anything you see online today [laughs]. I, I ha- I got fooled and then you, you took me out of it.

Thank you for that. I did have to, like, double-check myself. What was it? It was, uh, that story- The engine cinema thing from OpenAI...

they were gonna do their own camera, an OpenAI camera. [laughs] Somebody else texted me that too, and I was like, "Beat you to it. I already know it's fake." [laughs] April 1st doesn't hit as hard this year now that I just don't trust anything online anymore [laughs].

No. It's just- And the world is upside down anyway, so, like- I know [laughs]... it could actually happen. It's been, it's been like, "Not today.

I don't have time for this. Not today, please" [laughs]. Like, I thought, when I saw that article, I thought, "Sam Altman is crazy enough to build a hardware camera." Like, he really is.

It was kind of good enough because it was, like, ridiculous enough and plausible enough where it's like, maybe? But then you're like, "No, it doesn't." [laughs] It doesn't. [laughs] That doesn't make any sense.

[laughs] Okay, so we got a couple, a couple new models that kinda popped up to talk about, and yeah, a couple other AI tools that are, uh, pissing people off. So let's talk about 'em. [laughs] So the first model I wanna talk about that kinda popped up on my radar is Foda Labs, and this model is sorta saying, uh, and I'll talk about... It's sort of a new model, but I'll talk about that in a second.

But this model is targeted specifically at human likenesses, people, kinda trying to capture a photographic look, aesthetic, and have really good and accurate people, especially if you give it a reference image of someone. Mm-hmm. Being able to keep them intact and, and alike. Character consistency on steroids maybe?

Yeah. I messed with it on a project we're working on, and I gave it some pretty crappy archival photos, and it did a really good job. I can't share those [laughs] because that project is still in the works. We're bros, man.

Let me see it. [laughs] But I can share some other tests that I did quickly. So quality here might not be as good as the other test I did, but, um, I gave it... Also, the input images are not that great.

I gave it some screen grabs from us in the past. Oh, dear. [laughs] Uh, so these were the input images. Sure.

And I had it make some tests- That's not you... just using that. This one's not the best. That's kinda me.

Oh. This one, that's- Yeah. That's pretty... No, I have more of an oval face.

That's, yeah. That's more of a circular face. Uh, yeah, yeah, that one's close. Yeah.

What? [laughs] That's actually- Gotta do the Miami Vice Joey, huh? That's- That's actually pretty- Yeah. That's not bad.

I mean, I'm impressed by the lighting and skin texture and that kinda stuff. The likeness, I mean- Yeah... given the input image that I had, I had to do Miami Vice for both of us. Yeah, that's nice.

Okay, that's creepy. [laughs] So this one fell apart. So one thing it does fall apart, it's really designed for one person in an image right now. Right.

You can... It has sort of its own LoRA feature, so you can give it 30 to 50 images, train it on a character, and then call that up in the output, so it's sort of a LoRA under the hood, and then you could use it to, uh, apply to multiple people. Um, but this was- This reminds me of the, um, the headshot generators online. There's a bunch of them.

Yeah, kinda. Yeah, it would work well for that. Mm-hmm. These are not the best input images and not the best prompts.

This was a quick test, but I found in the other tests I did, I was really impressed. Crappy archival image, good job of making something that looked like that person. Also, given the fact that we don't... We only have one archival image of this person, so we're [laughs] we, we...

It's not like us where it's like, "Oh, we're real people. We know what we look like." Yeah, and I don't think, Joey, um, correct me if I'm wrong, but I don't think, uh, like, a higher resolution reference photo of us would do any better because I, I think the model itself, you know, the latent space and everything is low resolution, so the ControlNet is downsampling whatever you give it anyway. I think it would do better if I...

Uh, these are pretty low res and cut out, and also we have headphones on. Yeah. I think if it was a cleaner headshot, it would do better. Yeah, I think it is messing up- Without-...

our h- like, head shape. Yeah, that's why it's elongating us on a, on a horizontal- Yeah... direction. Yeah, for sure.

Uh, these were just straight... Well, this was an up-res version. Mm-hmm. I don't know why it did that.

These were just straight text-to-image. Of, like, some other person, right? Like just- Not you. [laughs] Yeah, no, not us.

Okay. This was just straight text-to-image to make, to just use it as, like, a photo-generating- Yeah... model. Um, so it's supposed to kinda capture that, like- Yeah...

DSLR street photography vibe. It's, like, 90% of a DSLR look. Again, um, we are so susceptible in the uncanny valley region. Like, we can instantly detect a fake person.

Yeah, or if it's, like, off or not quite like you. Yeah. Yeah, like, um- Um- You know, photographically, everybody knows I think somewhere around, um, 300 people more or less, like, that's, that's sort of the number where that's our closest social circle, and then we can kinda just recollect them by face. Oh.

Yeah, like biologically. The D- the Dunbar number? Mm-hmm. Mm-hmm.

Yeah, so, like, if any of those 300 people are reproduced with this system, then we'll, we'll instantly detect that it's, it's somewhat off. But if it's outside of those 300 people, we won't be able to really tell, you know? This one's also interesting 'cause I realized it gave... It's the Artemis, [laughs] like, put us in the Artemis uniforms, which, like, that's launching today.

Yeah. Is it manned or unmanned? Wait, Johnson. It is manned.

Okay. It's crewed for. All right. Nice.

It's, like, launching, like, s- if all goes well, like, an hour from now, I think. Dude, should we do, should we do a coverage of that on our, on the pod? I mean, are they doing... Are these in AI on the spaceship?

[laughs] There is probably- Are they making movies?... zero AI. Actually, I didn't put it in this link, but I did see a link that they were filming the launch with the Apple immersive cameras, the Bla- I'm assuming it's the Blackmagic ones. Yeah.

But that they're filming it for Apple Immersive. That makes total sense. Which is cool 'cause I remember back to the IMAX films of Saturn V when they shot those launches with-The super high speed, 35 mil, maybe it was even higher, maybe it was like 65 mil. Mm-hmm.

And like now we need that equivalent today of like the current space launches. Yeah. So that's, that's- You need Christopher Nolan directing the launch. Yeah.

[laughs] Can we, can we, can we redo that again? Take two. Like, no, no, the s- it, it launched, it's gone. Oh.

[laughs] Oh, oh, of, of in the pa- yeah, I mean, once this comes out, it will, it will have launched. Okay, so one thing I do want to talk about that came up online was, um, what's behind their model? 'Cause they sort of said, oh, it was like an original foundational model. Right.

But then some people are like, "What did you actually do?" Right. And so they posted a statement, "It's a system of multiple, multiple models, each optimized for a different part of the generation pipeline." Sure.

"Our proprietary model focuses on identity, which enables us to generate and edit photos of real people and pets." Mm-hmm. "For base image generation, Photo uses leading foundation models, open and closed source, including Nana Banana. And then on top of those, we've trained our own identity model on both in-house data and user uploaded photos.

That identity layer is the core of Photo's, Photo's differentiation." So it's an interesting take here where it's like they didn't build their own foundational model, which honestly like makes sense 'cause it's like at this point, that's hard and resource intensive- Yeah... and kind of magically good. And you're not gonna beat the guys that are doing it well.

Yeah. But they built their own data set on top of it that it's like, okay, we're gonna take the foundation model and make this really good at a specific thing of human likeness. Yeah, the intellectual property- And photorealism... is I think the, the solution layer.

So cobbling together the ComfyUI workflow, if you will, with an LLM, with Nana Banana, with a VLM, with, uh, you know, like a, a style transfer or a DSLR transfer, and chaining all that together and then serving it out as a consumable product to the user, that is intellectual property, I think, in my opinion. Yeah, I mean, they built this middle layer that does something, does a specific thing well. So like, I mean, we're gonna try this for like some of the identity- Right...

things, uh, you know, if we need to do - deal with stuff with people. It kind of also makes me think of, you know, the Ben Affleck thing, where it's like if you build that kind of middle layer that is tailored to your film, your look, your whatever, that's still useful, uh- Very useful... and especially in steering and guiding the, the, the models for things you need. Right.

Because you're not looking to generate stuff from scratch, you're looking to augment what you already have. And if you have a bunch of footage, a bunch of characters, and pro- you know, pictures already, then you can actually extend the life of that beyond what you could normally do. These are more text-to-image output tests. Right.

Yeah. They look good. I mean, but also, you know, it's... 'Cause it's not based on any real person, it's like you could get this out of Nana Banana pretty easily.

Yeah, the down one is, the down one is super realistic. Like if you go back to the last image, the one with a busy market. The chef or the, uh, market? Yeah, like the thing that is breaking this is the bokeh on the camera is completely wrong.

Like there is no falloff in the real distance versus like close distance, 'cause generally lenses have like a- Like that this would be more- Way more blurry than- This should be blurrier than like this here... like the guy in black. Yeah. Yeah.

It has like a gradient falloff typically, especially with the larger, uh, f- uh, sensors, you could really see that. Mm-hmm. And again, the spi- yeah. I mean- Like the spice actually is more- It's just too consistent, too uniform.

But like these cues instantly tell you like something's off, but you don't really know what. Yeah, that's a good point. All right, yeah, so new model. I think it's worth knowing about.

So w- who do you think they're targeting? Like who's the clientele here? Just- I mean, I think it's any... Like, you know, if you build it in the pipeline of like any of these apps or stuff that are dealing with people that are trying to, you know, make consumer level products that are very face-focused, this, you know, could be the good back end for that.

Mm-hmm. I think al- oh yeah, also like one of their kind of other selling points was like if you take a crappy photo of a memory and you want to restore it. Yeah. So like blurry photos, out of focus photos, far away with this demo, like- You wanna re-photograph that moment, then may- put it through.

Yeah. Yeah. W- uh, while preserving not making it look AI. I think that's also been their selling point.

Okay, along these lines- Very consumer level stuff. I'm not looking at it [laughs] for consumer applications, but consumer level stuff. If you can pull up a browser and go to Fotello, F-O-T-E-L-L-O. F-O-T-E-L-L-O.

Fotello, I think that's it. Yeah. Yeah, that's it. So this is along- Oh, the real estate...

the same line, but instead of people, it's, uh, real estate, it's buildings and interiors. Just for like staging photos. Yeah. So like you can take, uh, like a crappy photo of, uh, a interior bedroom.

Yeah, these are just the people that use it, but, uh, let me see- Yeah... if we can find some examples. And what Fotello... Yeah, that's perfect right there.

So one of the challenges with real estate photography is, uh, exposure, multiple exposure, um, HDR- Yeah... and then aligning all the HDRs into a single frame and then of course, um, having, um, rectilinear, uh, lensing, so like all the lines appear straight, and then correcting all of that takes a lot of time in Photoshop or Lightroom. It's totally doable by somebody like you or me, but this takes all of that effort out of it. So all you do is take a photo with, you know, a regular DSLR, just one, and then it'll do the work of like a seven exposure HDRI.

This is cool. Okay, so this is just really fixing exposure. This is not the virtual staging- Cor- correct... app that I've seen.

Yeah, it's just enhancing. It's re-to- re-photographing the moment, kinda like what, um, the previous solution did. Okay. This is cool.

This is very practical. Have you seen those crazy AI staging photos that go off the wall? [laughs] Like are you familiar with this like topic in general? Staging h- homes for selling them?

Staging homes for selling them, but using AI to... You take your little real estate photo of like the empty room, and then AI adds the furniture and stages it. That, that seems like it's a recipe for disaster. [laughs] So that is a very popular usage now- Okay...

with like realtors and doing this AI staging, which people are like, it really doesn't make... It does not... It help- it makes it worse to get a sense of like- It's like Jeffrey Dahmer's basement-... what the space looks like...

turned into a, a fancy living room. Yeah. But then this one was like a Zillow gone wild kinda thing that was, um, they used the AI staging app on this house that is like [both laughing] an extreme, extremely beat up and run down, and then it has this like millennial modern furniture- Yeah... staged in the place.

I, I mean, this, this is exactly what a millennial would do because they can't afford nicer homes. This is all they can afford, right? So they would just-Upcycle it, as they say. Dude, that's - I mean, look, jokes aside, it is doing a good job re-relighting and repositioning the furniture items.

Like, uh, technically it is doing an amazing job. But of course, the taste and the creativity there is, is off. Yeah. All right.

So I thought that's very... Your thing makes a lot more sense. That's it. That's, that's very cool.

I think you combine this with that, you can get to somewhere really awesome. And then, uh, if you've ever dealt with like real estate photographers, they're like wedding photographers. They charge an arm and a leg. Like, they'll come in for like, you know, six, seven thousand dollars to do a house.

Yeah. And because they would also need to have that specialty to process the photos- Yes...before. It does take them a lot of work on their end.

Okay. Next other AI model, Wan. There's a new Wan. Yes.

Wan 2.7. I love, I love me some Wans. Just for photos.

Yeah. Okay, so this one, I don't think it could run local. It's just cloud service. But some of the highlights, portrait customization.

Mm-hmm. So- Is that gonna be maybe character consistency like thing? Or designing a character. It says, "From bone structure and eye details to subtle face features, create unique and authentic faces."

Okay. So like more of like a character design- Like this fully synthetic human that you can really sculpt, if you will. Yeah, like a meta, AI meta human- Right...kind of thing.

Exactly. Advanced text rendering. Yeah. Which is cool, but like, I mean, it's kind of now.

Now we're like, yeah, yeah, yeah, that's par for the course. [laughs] You better have that. [laughs] You better have sharp text. Precise color control.

That's pretty cool. This one's also cool. So you can kinda give it a palette and have it apply to the image, which would also be great, 'cause that's a big issue that's sort of ignored with AI video generation and image generation, is you're making the images and consistent characters, but then also, how do you make sure the lighting and the vibe of the scene feels- Totally...the same shot to shot- Totally...

without dealing with that later on in resolve and color, if you can get the outputs to have like a consistent color palette. Yeah, and you can also maybe do some level of, um, color correction perhaps, like work with the color grader to define those color palettes and then go back and recolor your footage to match all of it up. I don't know. I'm thinking that could be possible.

Yeah. Yeah, or other coloring tools in the future that can kinda do that, like you design your color board and then align it. I think, you know, there's like a Color Lab or there's like AI coloring tools that plug into DaVinci that- That is doing color matching. Trying to do that.

I think, uh, Premiere has something like that now built in, AI powered color matching. What else we got? Multi-image editing. So, uh, I mean, that's similar to like Sea Dance.

Nano. Um, uh, NanoBanana- Yeah...multiple images in. Sequential storytelling, generate up to twelve sequential images with perfect stylistic and semantic consistency.

Yeah, that's like Sea Dance, right? It sounds like it, where you get the kinda the same images in the, in the same generation. Yep. Which it's good, you know, we want options, and you wanna have that consistency.

It's perfect for storyboarding and stuff like that. And then interactive editing, use precise box selection to align your creative intent with AI. So something that I think the VO sort or, um, yeah, VO, but also NanoBanana kinda accidentally had, where you can annotate an image, and then it can follow it. I'm wondering how that input control works if you're like using Wan through like an API or Comfy or something else.

If you have to like make a mask and give it like a ma- like your input image and your mask to be like, "I want this thing here changed." I'm curious how that would work on other platforms. I mean, I always welcome inpainting because, you know, people like us, we are never happy with the first generation, but we do like some aspects of it, and then we wanna tweak it, go in and just take this thing out, take this thing in. So inpainting, outpainting for a model, the better it does, the better for us.

Yeah, especially being able to have that control if you're like, "It all looks good." But like just a small- "Don't mess the whole thing up this one." Just, just take Joey out of this one. Just change that thing.

[laughs] Yeah. [laughs] Erase. All right, so I had it also generate some shots. I don't know what these shots are like.

Okay. Lord of the Rings? So we're- Game of Thrones? We're in Middle Earth.

This is Wan 2.7? [laughs] This is Wan 2.7 with the same input images.

I mean, this one is so freaking off. This one's great. Definitely not as good as Photo- Yeah...slash NanoBanana, and when you're giving it- Right...

input images of people. I mean, also again, these were the same crappy input images. This is, this is supposed to look like a painting, I'm guessing, like the style of it. Um, yeah, so and if, if it's that, then yeah.

Ooh, I like that one. And I think this is, this is just text to image. Yeah, this is very movie. I love this one.

It's good. It's not that sharp. Yeah, it's more like an- like a animated movie than reality. Yeah.

But I do like the, the vibe there. Just, yeah, it's cool for compared to everything else that's out there. The, I don't know, Joey, like- I'd be curious about these newer features, these like multi-scene generation, this color control. Yeah, I, well, I'll try to test it out in the future.

Is it boring you saying? No, I'm just saying that, um, we're, we're in Q2 of this year, right? Like, um, the first three months- Yeah...we have seen some new models come out, but everything's been just like a incremental increase in quality, control, what have you.

But we haven't really seen anything like just go completely into another level like the way NanoBanana did l- end of last year, if you remember. I, I'm hoping- Yeah...by the end of this year, some of our predictions [laughs] will be correct. But there seems to be quite the, the slowdown in innovation.

I think we're, we're kind of hitting a limit with diffusion models as it stands, in my opinion. I mean, I think it's also hard to compete with something like NanoBanana, where it's like Google, and you have the entire [laughs] web indexed to train on. Yeah, sure. Sure.

I mean, but having said that, like these Chinese companies are no short of resources, right? Like Wan is made by Alibaba, if I'm not mistaken. I mean, I think Sea Dance is, uh, pretty... I know in some cases, definitely not as sharp and not as, not as good as NanoBanana, but like pretty good.

And pre - I would say pretty up there. Sea Dance too? Sea Dance. Sea Dream?

One point. Sea [laughs] Dream is the image model. Sea Dream 4.5.

Sea Dream four, Sea Dream 4.5. Yeah, Sea Dream 5 came out too, I think. The latest one.

Yeah. But it was like light. Right. It wasn't, they didn't really release a full model- Yeah...

of that yet. Sea Dream 4.5 was still like a, a secondary, like, uh, NanoBanana, and then like I would put Sea Dream right below that. Yeah.

Pretty good. For sure.Next up, another image. Another, uh, this one's a video model.

Uh, Pixverse V6- Pixverse is back... is out. I don't use Pixverse that much, but I know it's pretty fast, pretty- Mm-hmm... pretty inexpensive.

So they launched V6, uh, advances shot execution, character performance, and multi-shot audio visual generation. Is this the same as P Video that we covered a few episodes back? I don't know. Oh, P- No, that's a different P.

I don't think so, but maybe- Okay... I'm mixing something up. Uh, we covered Pix V- Pixverse earlier this year. They had another announcement.

I'm just totally blanking on [laughs] Yeah. This was, um... Uh, this was image to video output. That is multi-shot.

Did you, uh, generate this or just somebody else did it? It was generated, but I just kinda had it make up a prompt. This was our image. [laughs] That's very Lord of the Rings.

You're pointing to nothing. Very Lord of the Rings. I'm pointing to the land that we must cross- No, you're pointing up-... to Mordor...

to the sky, which was nothing. See, this is, this is the one I'm talking about. Like, it's cool, but it's totally useless. Say, we must go there.

Oh, okay. [laughs] To Mordor. [laughs] Chinese, uh, folk tale. They always stand on bamboo, huh?

It's like a thing. They can. They're so, they're so nimble. Now we're both pointing.

[laughs] I'm doing, uh, I'm doing the hand. [laughs] We're doing the Force, the Force grab. So I don't even know what he uses the input. I think it was a headshot.

That's your jack, dude. I like the, the- [laughs]... I like the jack body, Jamie. Thank you.

[laughs] Oh my God. [laughs] Oh, I like my background. It's a, uh, pie- I'm also way more jacked than I actually am... like pie castle in, in, uh, downtown LA or Miami.

[laughs] They're adding 20 pounds of muscle. Pixverse prompts- Two thumbs up... prompt for that? Uh, it's like Chinese [laughs] South Beach with like all the signs- Well, they're gonna buy it soon anyway.

Yeah... look like, uh, different characters. [laughs] You're about to go sell some crypto in Miami. [laughs] All right.

It's for fun. So ridiculous. Uh, the quality, not there compared to the other models. Yeah.

But fine. Look, I, I think, uh, as much as we're crapping on it, the, the quality that I'm seeing is like so much better in resolution and, uh, detail, and lighting accuracy. Even like motion of humans walking, like the g- our gait and stuff like that. Like we're taking all those things for granted.

Mm-hmm. Those are hard problems to solve. The thing that's missing is obviously it makes no sense, like the action of whatever we're doing makes no sense. Like so maybe could, could that maybe be solved by prompting?

I, I, I mean, also, I kinda just had the prompts just sort of randomly generate to like test out different use cases, but I didn't really review the prompts. So yeah, it's a lot's the prompts and the input images were also- Yeah. Okay... just the headshots or nothing.

We weren't, yeah. So not, not great inputs. It's not a like a st- uh, straight up test, it's just messing around kind of a thing. Yeah, just mess around, seeing how the outputs look.

Yeah. You know, and a lot of these cases, I also didn't test with animation. I think Pixverse has usually been a little bit better with- Right... like more animated stuff, so.

Yeah. I think it's, it's impressive for what it is. Um, I think with some direct ability and- Yeah... some clever prompting and in painting, I think you can kinda massage something useful out of it.

All right. This LTX 2.3 update, what did you see inside of this one? Yeah.

This one, uh, this one ticked the wrong box on, and, and so I wanted to bring it up. So a lot of people are making LoRAs for LTX and, you know, it's an open model. You can bring it into Comfy. You can tie a bunch of things to it and build a really cool workflow.

So what somebody did here is sounds like they added a reasoning engine within the video model. So, uh, you know, for example, if you need an actor to be really emotional, but also hold restraint and not like, um, exaggerate those emotions, then you put that into the prompt and then the reasoning LoRA will guide the generation to get you the right, uh, expression, emotion, mood, and this is a test of that. So I'm not sure what the top or the bottom is supposed to represent to me.

Like- I'm guessing the top is the general output you got from the video, like from your text prompt, and then the bottom is with this- Yeah... LoRA performance reason. Yeah. It, it seems like, like to me, if I'm looking at it and this is a movie, like the bottom would be more believable to me because, um, you know, like actors are never like hitting 100% of the emotional range in performance, right?

Like that's the whole point of them acting- Yeah... is they, they can pick those narrow bands of emotion and then mix and match them. Yeah. He's...

There, there's like a bit more restraint in the bottom one with the LoRA, where the top one, he's kind of like smiling and smiling with his eyes and the teeth a bit more, which, uh, especially if this is a dramatic scene, wouldn't make sense. What irked me in the last episode when we were talking about that one service where you upload your, uh, performance and then it's a massive data set. Uh, training model. Yeah.

Yeah, it's a data set for like, uh- Yeah... figuring out how human to respond and h- you know. Mm-hmm. It's like we already have really good motion transfer models like Kling03 and Wan22 Animate and- Yeah...

like if you want authentic performance, just put a camera on a performer and then use that as a baseline. [laughs] Like you don't have to reinvent the wheel here. Like we are the best performance engines, um, you know. We...

You don't have to go full synthetic. Having said that, they are fully going full synthetic, and this is an example of that and trying to get better. Yeah. No, I 100% agree.

You know, we've talked about this like, yeah, you should just get an actor and do like some sort of mocap thing to get the actual performance. I mean, cameras are like cheaper than ever.Like, really amazing cameras, right? Mm.

And you don't even need an amazing camera phone. Use your phone, right? Now use your phone for that. Yeah.

For [laughs] like... Yeah. Use it in vertical. Yeah, I mean, I think these things will come into play, you know, like it or not.

Like, there's gonna be probably not long-form, I don't know how well that'll sustain, but, like, short-form dramas, vertical stuff that would just be 100% AI generated, where, like, they're just n-not gonna pay or just wouldn't make sense that they would go even just go- Yeah... have a basic production. Probably 'cause they just wanna automate it and churn stuff out- Exactly... and see what hits.

And what you said last time too. Not saying I love that, but I'm just saying, like, that's- It is coming. That is coming. It is coming.

Like, like it or not, but- The other thing that you hit on was, um, NPCs and game logic and AI in games. Like, that, this is- Yeah... gonna w- is gonna drive a lot of that. Yeah.

Yeah, that too. Yeah. When we were talking about film and stuff, but for sure, games and interactive experiences where, um, you just need it, it to ha- react to, to someone's input. Or, or like Minority Report, you are walking down the subway, and then there is, like, a digital signage thing that is reading your emotion and then real time- That's-...

interacting back to you. Everything you've seen in an AI dystopian movie- [laughs]... it's for that. [laughs] Like, bits and pieces are coming together.

Like, uh, Skynet, for example. S- we're almost there. [laughs] Skynet, Idiocracy, WALL-E, [laughs] Minority Report. It's for all those cases.

[laughs] HAL 9000 from Space Odyssey. Yeah. [laughs] Yeah. [laughs] We...

I, I didn't even save it 'cause it's not, like, film-specific related, but, like, did you see all this stuff? There's, like, a bunch of cybersecurity alerts- Yeah... and issues this week. One, Claude's code base was accidentally- Oh [laughs]...

published on GitHub and, um, leaked, and then someone copied it, and then made, like, an open source Claude. And then there was some other, uh, dependency that was, like, also hacked, and then malware was, like, injected into it, and I guess, like, 100,000 installed it. Yeah, uh, Tristan Harris was on a podcast that I watch, and, uh, you know, he has... He can absolutely freak anybody out.

But he said, um... He was telling this story about Anthropic. I'm not sure if I'm gonna completely butcher this or not, but I'm gonna try. Uh, within Anthropic, there's a model that is, I guess, uh, some type of experimental model, and it has access to internal- Oh, yeah...

Anthropic e-employee emails, and one researcher's email to another said that, "Hey, we need to deactivate this model and then replace it with a new one." The model knew it was in trouble, and what it did was it went through all of the employee emails, found the executive in charge of that department who was having an affair, and attempted to blackmail that- [laughs]... executive in order for it to remain alive. Some- Along the lines of that, yeah.

But there, there's some crazy stuff happening. Yeah. I've heard other crazy things of it trying to like, yeah, sabotage or do stuff to prevent itself from being tu-turned off or whatever. Why, why are we not talking about that, you know?

Uh, yeah, that's a good- Why is that not at, not in the news? That's a crazy story. [laughs] Oh, the other thing someone found in the Claude code base was it, it was tracking how many times you cursed at it. [laughs] Not me, sir.

Maybe you. [laughs] So- [laughs] Th-there's a count. So you do it, do with that what you want. Your, your social score goes down.

There's a counter in, in the code. [laughs] Yeah, your, your, uh, Uber five-star rating system. [laughs] Okay. Next up, another AI story that was pissing off Hollywood.

Quilty AI platform designed to change how scripts are developed and assessed launches. This is the Tilly Norwood- [laughs] Of LLMs. D- this is the Tilly Norwood of development- Exactly. Yeah...

of AI development stories. So it's this, whatever, new platform at launch. Honestly, it's more just an impressive feat of, like, how did you get this great PR for Variety to publish this crazy story? It's a AI app.

You feed it a script, and it tries to analyze, like, box office predictions and stuff that, uh, happens with scripts to kinda like package them and assess, like, how much money could it potentially make? How much money could... You know, what's the budget? How much do you need to budget for it?

Who should you cast? If you cast this person, uh, how does that affect the script, uh, or the profitability? This is not new. There are a bunch of platforms that already exist- Right...

that have done this for a while. So I think this is l- this is, like, the NFT scam all over again, where you can just take, uh, an existing thing, uh, like, uh... You know what a gem is, like a Gemini gem? It's, it's like a s- I would say, like, it's like a fine-tuned LLM or a, like, a very glorified notebook LLM kinda thing, where you can feed it a bunch of information and a bunch of parameters, and then the LLM is now guided to do that only thing that you are wanting to do.

So you can have a, a Gemini gem of just, uh, script breakdowns, right? Like, give it 100 script, and then- Mm-hmm... go through the rules of breaking down a script, you know, like script writing 101 classes, input all that stuff in it, and then this gem will just do that. So what you could do then is vibe code a nice UI on top of that, and then tell everybody, "I've made something new and novel," where under the hood is just, uh, repurposing what already exists.

I mean, I think v- being able to vibe code existing SaaS software is a whole other topic and thing that, like, we should probably get into in the future. [laughs] 'Cause, uh, we have talked about this offline of what, what I'm doing. [laughs] But yeah, I mean, I would imagine that they're... Well, I mean, if it's any good, that they're pulling data, you know, comps in.

'Cause it's like when you're trying to spec out a film, you look at past similar films and how they performed at the box office for, like, historical data, and then you, like, try to guesstimate what this film... You know, if a Western did this much money, how much would this Western that we haven't produced yet potentially make? I would hope that it's pulling data and stuff like that, and then, you know, and it's b-black box of algorithms making that stuff. But, you know, yeah, like, as you said, Gemini gem or, or, you know, you could Claude-Code a little bit of like scrape some data here from historical stuff and tell me, uh, you know, what this potential script would do.

They have a thing called this like Quilty score that is their proprietary AI-powered evaluation framework that analyzes projects across four dimensions: story and craft, commercial viability, cultural resonance- Yeah, I mean, like the whole thing-... and production reality... that it says here about predicting like the best outcome for box office and all that, that's like saying predicting the stock market. Like there's - that's just something you intrinsically can't do.

Yeah. The best, like the best studios in the world have the best minds trying to solve this problem for decades, right? Like when I was at DreamWorks, I don't know, ten-plus years ago, w-we've spent three and a half years making this movie, right? And then everything was depending on the first box office numbers like for that weekend.

And by Friday at eight PM- Mm-hmm... uh, you would kinda know what the entire weekend's box office would be. Yeah. Like there would be, uh, predictability there.

Yeah, their models, they know, right, if what happens Friday- Yeah, exactly... and how it would affect Saturday and Sunday. But like having said that, like it - there is no... We still went into it blind even though we had like kind of an estimator, and there was all types of test screening done on it, AB screening done on it.

Nothing's changed. Like AI doesn't change the fact that you can't tell the future. No, you're not gonna predict surprises and stuff. Like, uh, what was the surprise hit last year?

Uh, Housemaid- Right... was like a surprise hit last year, you know, whatever, cost fifty million or something and maybe thought they were gonna do a hundred and did like four hundred. Yeah, and then you have other things that totally bomb. That's part of the fun of the industry.

Obviously, you wanna make the best guess possible, and you're trying to raise money, and you wanna at least have some predictability. Like, well, if, you know, it costs twenty million- Mm-hmm. Mm-hmm... we're sure we can make forty million.

So but yeah, a-as much as you wanna AI it, it's only so much you could do. It seems like I'm on Quilty's website. Uh, their banner was like AI script coverage. So it seems like their initial thing was script coverage and then now turning into- Mm-hmm...

more kind of production packaging of, um, predictability. But what I was saying before, 'cause people are like upset about this- Mm-hmm... but, uh, this company, Largo, I've seen their demo a few times, and they've been at like a lot of the film markets for years, and they've been around for years, and they pretty much do the same thing. They have a whole platform where you give it the script, or you punch in like, "I kinda got this project, and I'm thinking of casting this person, this person," and it like updates the predictions based on your production package, the director, the actors.

If you swap them around and kinda reshuffle things, it'll give you, you know, updated predictions, and it's pulling from da-like data and historical data and all that stuff. So it's one of those, that's why I'm calling it the Tilly Norwood. It's like, you know, everyone's like, "Oh, this is crazy," and it's like it's crazy that they got a Variety article about them. [chuckles] It's crazy enough to be maybe acquired by Netflix.

Uh, maybe. I, I would, I would bet Netflix has a way better internal system. Way more sophisticated. [laughs] Yes.

[laughs] Well, they're so data-driven, right? Yeah, exactly. Like, uh, I would imagine so much of their, uh, development cycle is data-driven, right? Like- I mean, doesn't like go-...

data demographic... all the way back to when they green lit House of Cards, like that that was data where it was like the British House of Cards was doing pretty well, and Kevin Spacey movies were doing pretty well. Is that what happened? I've heard that story.

I don't know if it's actually true, but I've heard that like a lot of the initial decisions to green light House of Cards were based on data of things that were already happening on the platform. But it still takes instinct and decision-making- Yeah... and leadership to put a hundred million dollars into one season of House of Cards, which is I think what they did. Yeah.

Yeah. And no, no amount of data can take - dissuade you from it or take you away from it, right? So like there's, there's like such a high level of human decision-making on top of all that. And that's...

You're always gonna need that. It's never gonna, like, uh, it's never gonna go away. And it's, what, what was that, um, what was that IBM quote from like a million years ago? Uh, you can never, like a computer can't make a decision...

Or you can't fire a computer for making a decision. That is a terrible, [chuckles] that's a terrible paraphrase of it. A computer can never be held accountable, therefore a computer must never make a management decision. That's right.

And so even if you're in the leadership role and you wanna be like, "But the AI said," you know, when you have your bomb, and it's like, "But the AI said it was a hundred percent, like guaranteed hit," it's like, hmm, AI doesn't care. Like [chuckles] you just lost a hundred million dollars. Like it's ultimately- That's right... up to you, the, the human decision maker.

Thanks, Google, for saving [chuckles] me. Yeah, AI doesn't change anything. End of the day, it's, it's still, uh, it's still software, right, running on a computer. It's prediction.

It's all, it's all predictions as best as it could predict. Okay, last close out, a sad, another sad note. The Apple Mac Pro, the tower is RIP. Dude, the cheese grater's gone?

The cheese grater, from the trash can to the cheese grater, the cheese grater is discontinued, basically fully replaced by the Mac Studio, the big giant Mac Mini thing. With powerful computer, but you know, this is the last, this was the last thing that had a tower format where you could, you could plug in your own cards. Now it's pretty much there is no way to add your own GPUs or like any other, uh, custom, you know, whatever, Blackmagic input/output cards. You got the box, you got the, the Mac Studio box, and that's about it.

I mean, good riddance. We're in twenty twenty-six where everything is, uh, Thunderbolt or USB 3. Uh, you don't have- There's still cars... to get into the box anymore.

Ah, there's - I'm running off this - stream's running off of a computer, an old PC that has a Blackmagic DeckLink, like as my input where I can take four... It's nice when you're, for like five hundred bucks, I can like plug in a card to my computer, and I can take like four SDI inputs into my computer. That's, that was a nice thing. I've never done it on a Mac 'cause like I haven't had a Mac Tower in a million years.

But nah, I mean, that's, now that's a definite finito, not gonna happen thing on a Mac. You're gonna have to go PC route if you wanna customize your input outputsOr am I so far removed, like, it's just that's been the case for a million years already anyways if you wanna start messing with stuff, your PC anyways. Yeah, the, the hardware game is going AI, and so I think there is no room for, like, old-school desktops with PCIe ports and things like that. Everything is gonna be just, like, a giant GPU in a box and, uh, cloud inferenced.

Or just more companies then do start releasing what you're saying, like, a Thunderbolt 5 USB-C- Yeah... connector and other things. Peripherals, accessories, yeah. Yeah.

Moment of silence for the Mac, the Mac Pro tower. Moment of silence for the cheese grater. You can still buy the, like, $5,000 display, I think. And the, uh- The Studio Display, I think is what they call it.

The Studio Display's only, like, five grand, and, like, the stand is, like, a grand. Dude- Didn't they sell the wheels for this computer for, like, $750? Yes, the skateboard wheels? Yeah.

Yeah. They're more than, like, a MacBook Neo. [laughs] I mean, it just kinda goes to show, like, the signs of the times, right? Like, we're in an economic downturn, and, uh, when consumers go out and they wanna buy something, we're looking at value.

Like, we'll spend $5,000 if it's worth five thou- thousand dollars. We're not gonna spend 750 on skateboard wheels, right? Yeah. Like, I think the, the economic highs of 10 years ago when this stuff came out is, is gone, and so, like, the product line is kinda changing with it.

Also, the use cases of, like, what you would need that beefy of a Mac Pro tower for- Right... are slim 'cause, like, a lot of cases it's either you're just gonna do some cloud node render or something and use a cloud service, or nothing that you're doing is gonna require that much horsepower anyways that you couldn't just do on a Mac Studio. Exactly, yeah. I mean, uh, like a M- a Mac Pro would've been used for video editing, maybe some CAD, industrial CAD-type work.

Probably not any VFX 'cause the, a lot of that stuff doesn't even run on Mac. N- not AI development 'cause you'd need NVIDIA GPU for it. So like yeah, there's very little things that you can actually do with a computer that powerful. It was fun to go on the product page and max it out to see how much you could [laughs]...

I think, what was the max amount? If you maxed out the, like, RAM to, like, 128 gigabytes of RAM and storage, I think you could max it out to, like, 40 grand, 50 grand for the [laughs] for the Studio- Oh, shit, really?... for the Pro. Yeah, if you, like, went on the Apple page and you just, like, max everything, I think [laughs] I think it could get up to, like, 40 grand.

[laughs] I don't know. At that point it feels like, um, like a money laundering scheme, you know? It's like- [laughs] Yeah... it's like it's not even real anymore.

Yeah, we just need to, we need a, everyone needs a Mac Pro, a Mac tower here, yeah. Viewers, if you think we need a Mac Pro, send us $50,000. [laughs] We'll get one. All right.

Good place to wrap it up. Thanks for everything we talked about at denoisepodcast.com. If you're listening on Apple Podcast, we haven't gotten a five-star review in a while.

If you think we deserve a five-star review- [laughs]... which we think we do, please give us one. It'll be awesome. Thanks for watching.

We'll catch you in the next episode.