WEBVTT
1
00:00:00.040 --> 00:00:07.000
There's a level of people where it's just like, "I just wanna pay stuff for something that works." You and I both have coffee machines at home, but sometimes we just go to Starbucks.
2
00:00:07.280 --> 00:00:21.360
[laughs] It's just easier to just pay. Welcome back to Denoised. How's it going, Addy? Long time no see, Joey. Let's go. I know. I know. Finishing up some projects, but, uh, it'll be out.
3
00:00:21.440 --> 00:00:31.300
Actually, by the time this is out, the trailer for it will be out. So- Oh... I'll talk about it next week. Okay. Just stay tuned. Uh, have you been doing any LA Tech Week stuff?
4
00:00:32.380 --> 00:00:39.320
Uh, I went to the AI on the Lot event, but I haven't- Nice... that's about it. What about you? Nah. I still- [laughs]...
5
00:00:39.360 --> 00:00:49.940
cannot fully wrap my head around what exactly Tech Week is besides panels and parties that are loosely organized. You're right. Uh, to me, it feels very VC-driven, like A16Z- Yeah...
6
00:00:50.180 --> 00:00:59.860
and that, that kind of thing, where it's like, uh, get a bunch of founders together, figure out some new ideas, so we can go invest in the thing. Yeah. I don't know.
7
00:00:59.960 --> 00:01:11.340
Everything I've always gone to with these was, like, a zoo, and it's just, like, overpacked too. Yeah. So okay. Speaking of building and whatever the future of software means [laughs]. Oh, God.
8
00:01:11.740 --> 00:01:24.720
[laughs] A couple things dropped in the last, like, week or so. One of the big ones is REA, Reverse Engineer Anything. So crazy, man. Which is- It's giving me the chills just looking at this...
9
00:01:24.760 --> 00:01:36.920
it's a, uh, FCP tool for basically kinda point it at an app, and it can reverse, or I believe a website, and it can reverse engineer how it was built, and then build, vibe code your own version of it. That's right, yeah.
10
00:01:36.980 --> 00:01:51.960
From my understanding of it, um, it gets into the post-compile binary, and then it can reverse engineer that to human-readable code. So, like, imagine all the ones and zeros- Okay... turned back into C++ code.
11
00:01:52.320 --> 00:01:57.200
I don't know how it does it. It's crazy because decompiling is, like, incredibly difficult.
12
00:01:57.600 --> 00:02:07.020
And, uh, it does it, uh, well enough to then recreate the application where you can take over and rebuild it or, you know, modify it to your will. Yeah.
13
00:02:07.420 --> 00:02:13.940
The probably biggest example going around with this, I don't know if they use this or this, the person just vibe coded it, but someone built...
14
00:02:14.120 --> 00:02:26.880
It's called Crafting Apps, and there is a Photo Craft, Film Craft, Light Craft. It is basically a vibe-coded version of Adobe's Creative Suite and other products and other vibe-coded apps. Um- Yeah. Yeah...
15
00:02:26.960 --> 00:02:36.480
this has been probably the, kind of the biggest example going around of just like, oh, well, you know, what are you gonna do if you're Adobe or any other major software- Yeah...
16
00:02:36.500 --> 00:02:53.240
and someone sort of built a free version of it? Now, apps and heavy-duty stuff like this, this feels like the last thing that I would j- jump ship or switch on [laughs] and change my entire workflow to a vibe-coded app.
17
00:02:53.300 --> 00:03:05.120
Also, there's been free video editors and other tools out there forever. There's Photopea, which has been online forever, which is like a web-based version of Photoshop that's ad-supported and, and free to use. If you're
18
00:03:06.020 --> 00:03:14.240
someone in high school, you know, back in the day, I, you know, you, you found the codes to, um, [laughs] unlock Creative Suite or the other apps- I don't know what you're talking about. No.
19
00:03:14.240 --> 00:03:21.260
[laughs] It's actually, this was even before Creative Suite existed. This was just purely when you had to get the code for- The, the CDs and the-... for the software... and the floppies even. Yeah.
20
00:03:21.580 --> 00:03:28.020
Codes floated around the internet, you know, if you were in middle school, high school, and you didn't, um, have any money to pay for, for expensive software.
21
00:03:28.860 --> 00:03:41.500
You know, if you're dabbling, but if you're in an enterprise company or a company where you need stuff to rely on. Yeah. Yeah. I, well, okay. The, going back to the '90s, I think it was called a crack, right?
22
00:03:41.800 --> 00:03:51.200
So when, [laughs] when you would- A crack... download the ISO, the ISO is what this is. You have to download the ISO, which was the- Yeah... disk image... which is the binary code, yeah. Uh-huh.
23
00:03:51.240 --> 00:04:07.720
And then you can then, the, the ISO would mount to your hard drive like a disk, and then you would crack the license activation schema, and off you go. You have Photoshop on your, uh, Windows 95 box or whatnot.
24
00:04:08.000 --> 00:04:19.540
[laughs] I was going for Premiere, Premiere 6.5. [laughs] Oh, okay. Wow. All right. [laughs] You were into video long before me, my friend. [laughs] Yeah, I was doing video first. Yeah.
25
00:04:19.880 --> 00:04:24.420
The, the bigger existential thing is just like, okay, is software dead? Is SaaS dead? If you can just sort of build,
26
00:04:25.340 --> 00:04:34.080
these are clones of other things, but if you could just build whatever you need, you need a video editor, you need it with whatever specific features you need, where does this go? Yeah. I, I am...
27
00:04:34.580 --> 00:04:46.580
Okay, I have two trains of thought on this. So we could sp- speak specifically on the Photo Craft, Vector Craft, like this thing. I, I believe this was REA, um, devised. Like, they did reverse engineer- Mm-hmm...
28
00:04:46.600 --> 00:04:51.100
current things. They also got Word Craft, kind of a, a- Yeah... Word- And that's why they were able to nail the UI- Microsoft Word...
29
00:04:51.160 --> 00:05:03.040
so well, like almost pixel to pixel, 'cause it is just literally the ISO file, the binary file. But from what I am hearing, the performance is severely lacking.
30
00:05:03.160 --> 00:05:16.460
So, like, um, a lot of the, the magic in Premiere, and even in Resolve, a lot of the NLEs, is, like, how well it handles low-level VDU, uh, video computations on the GPU, right?
31
00:05:16.560 --> 00:05:27.860
Like when you're, uh, doing a time remap or when you're putting a transition between two shots, that, that, that's, like, video calculations happening in real time for you to preview it, and then when it renders, it renders, right?
32
00:05:27.960 --> 00:05:37.500
Um, like that stuff is not working correctly. So you can load your timeline, but the preview will be slow and choppy, and then some of the transitions won't work.
33
00:05:37.540 --> 00:05:46.740
And, like, I've seen a couple of videos on YouTube where, um, you know, they use Video Craft, and it's like, oof, oof, that's not quite there yet. Yeah.
34
00:05:46.840 --> 00:05:55.480
I mean, I would, yeah, I, I am not surprised that there's performance lagging issues. It's not there yet, but we can see where this is going, obviously. Exactly. Exactly.
35
00:05:55.500 --> 00:06:07.900
And where regardless if you have to point it at a working app, I mean, this is like the Higgs field jujitsu version of film where they take an existing film and restyle it. Oh, I hate that. I hate that. Yeah.
36
00:06:08.220 --> 00:06:15.168
Like they- This is the same thing where you point it at a- An application that- Yeah... uh, has had a lot of development and works, and then copying it. Right.
37
00:06:15.308 --> 00:06:24.868
But put that aside, and it's just like, "I need an editor, I need a text editor, I need something. Uh, let me start building it." Yeah. We, we tend to forget- And optimize it...
38
00:06:24.948 --> 00:06:37.188
how crazy all of this is because we're just inundated by one development, one progress, one thing after another, right? Like, wait, this was like NSA technology like five years ago.
39
00:06:37.368 --> 00:06:45.948
Only like the highest levels of the government do they have the technology to de-reverse engineer a, a piece of software that China built or whatnot.
40
00:06:45.988 --> 00:06:57.088
Like, this was used for like geopolitical chess, and now anybody has it. Like, we can... Like, think of an application, um, that you use every day, right?
41
00:06:57.168 --> 00:07:08.608
Uh, I mean, of course we use these applications every day, but also stuff like, I don't know, like, a w- like macOS or something crazy like that, and you can reverse engineer that probably.
42
00:07:08.908 --> 00:07:15.288
I, well, I mentioned about the enterprise thing before. Like, I think there's still a level, it's like, okay, so maybe just software development evolves or the companies evolve where...
43
00:07:15.708 --> 00:07:20.868
Like, 'cause you can move so fast and you build things so fast, obviously not everyone's gonna wanna vibe code- No...
44
00:07:20.948 --> 00:07:31.168
their own solutions and have some duct tape thing together, and support that, and keep iterating and building on it. Even as good as agents get, like you still need something to guide it or just- Yeah...
45
00:07:31.208 --> 00:07:42.748
make sure it's all working. And it's like maybe that's what software companies kinda turn into, um, as the just- The maintenance... they can run leaner and faster and push stuff out.
46
00:07:43.288 --> 00:07:53.288
Uh, but yeah, I'm also thinking of like Linux, right? Linux has been around, and Linux is free, but they, I believe it's Linux. I mean, they make a
47
00:07:54.688 --> 00:08:03.948
bunch of money off of supporting and running the servers and like the professional services. I see. I see. Oh, I see- Side... what you're saying. Like, you think that's the business... That is no business model then.
48
00:08:04.108 --> 00:08:13.728
Like, it's very slim pickings at that point. Uh, well, I mean, the other argument, or like, okay, if the software is not a moat anymore and it's what you build- Right...
49
00:08:14.108 --> 00:08:23.008
something else has to be the secret sauce, which is proprietary data or something else that- I think, uh, these are great questions... can't be replicated. Yeah.
50
00:08:23.048 --> 00:08:32.408
Uh, the one thing I will say is, um, we're, we're forgetting to mention that this makes theft really easy. Like, that's no bueno in my book, in your book too I'm sure.
51
00:08:32.448 --> 00:08:42.808
Like, shouldn't be stealing software, just like [laughs] you shouldn't be stealing, uh, data for training models and stuff like that. Yeah. I mean, look, obviously if you're pointed at like,
52
00:08:43.848 --> 00:08:49.648
I don't wanna keep picking on Adobe. If you point it at Resolve, you point it at Final Cut- Yeah... and you're like, "Copy that and make me a version of that"- Dude... that's like, okay, yeah- Yeah...
53
00:08:49.688 --> 00:09:02.588
that's like ripping it off. But if you're just like, "I need a video editor- Yeah... that does X, Y, and Z," it, I know it is, you know, completely capable of starting to build stuff like that, which arguably could be
54
00:09:03.728 --> 00:09:10.848
better because it is like, okay, you're building a custom piece of software built exactly for what you need to do and your specific needs. Mm-hmm. Exactly.
55
00:09:11.168 --> 00:09:21.248
But- Like, for this podcast, we don't need the entire might of Resolve or Premiere or Final Cut. Like, we just need a couple of things to make this podcast edited, right? Yeah.
56
00:09:21.308 --> 00:09:27.668
I mean, obviously there have been gaps in editing this podcast because even with these faster podcast editors, they don't really support...
57
00:09:27.888 --> 00:09:34.388
Like, I, uh, so many times having conversations where it's like, okay, we have your feed, we have my feed. I need to put them like split side by side. Right.
58
00:09:34.888 --> 00:09:42.628
And I need that to be treated as an angle to cut to, and a lot of these auto-cutting things like don't understand that. They don't create a third- And-... third camera that is the two of us.
59
00:09:42.668 --> 00:09:49.908
Yeah, it's like I need a fake third camera- Yeah... that is a two-shot of us. Right. And I have, yeah, on the verge of just being like, "Oh, let's just- I'll, I'll just build this...
60
00:09:50.008 --> 00:10:01.268
just build, [laughs] just build something." [laughs] Well, that, that, that's also the incredible moment in humanity that we're going through is that, uh, the average Joe like you and me can just build something.
61
00:10:01.648 --> 00:10:14.408
We don't have computer science backgrounds, and we don't program every day of our lives, but we're gonna build something that requires thousands and hundreds of thousands of lines of code, and we can do it. Crazy, right?
62
00:10:14.568 --> 00:10:23.868
Yeah. Yeah, I mean, it's been a crazy unlock. Yeah. 2026 has been an insane- 2020-... year. Yeah... yeah. And we have more stuff to cover [laughs] as w- we'll go through. Like, it's just getting started.
63
00:10:23.908 --> 00:10:30.328
Like, the entire, um, self, uh, sort of, um, [clicks tongue] uh, like just being able
64
00:10:31.367 --> 00:10:45.648
to build an entire tech stack yourself, plus make content, plus, you know, do all these other things is, is insane to me, like the h- power we have- Yeah... at our fingertips. Yeah. You know, I mean, we're, we're [sighs]
65
00:10:45.728 --> 00:10:55.128
tinkers, tinklers. Tinkers. [laughs] Tinkerers. Yeah. Like [laughs] I was running through five different versions of that word in my head. I'm like, "That's not, [laughs] that's not the right one." Yeah, yeah.
66
00:10:55.168 --> 00:11:03.248
Your, your NyQuil is hitting real hard right now. [laughs] Yeah. [laughs] And you know, obviously there's a level of people where it's just like, "I just wanna pay so- for something that works.
67
00:11:03.328 --> 00:11:07.208
I don't wanna, I don't wanna deal with- Yeah... all of this- I mean-... even if as easy as it gets." That's the thing.
68
00:11:07.248 --> 00:11:16.968
Like, uh, you know, you and I both have coffee machines at home, but sometimes we just go to Starbucks. It's just like- [laughs]... it's just easier to just pay- [laughs]... ex- for expensive coffee, you know?
69
00:11:17.128 --> 00:11:24.208
[laughs] I thought we were going with that as in we have coffee machines at home, but sometimes we just want to have like a nice pour-over and grind the beans ourself. But you [laughs] went the other direction.
70
00:11:24.288 --> 00:11:34.868
Well, yeah, like the, the option that you're getting now is you're not, you don't just have the pour-over and the grinding, but you can grow the coffee bean, right? Like, that's how crazy- Yeah... things are getting.
71
00:11:34.908 --> 00:11:45.748
Like, you can really go back to the source. But there's nothing wrong with going to Starbucks and just paying, you know, $6 for a giant, uh, venti whatever. Yeah, it's like I need coffee- Yeah... I need it now.
72
00:11:46.468 --> 00:11:48.628
I don't wanna- Yeah... do that- And I think-... headache...
73
00:11:48.688 --> 00:11:59.288
like a lot of the software moat in that sense I think will be undisrupted is because they're low-priced enough to where it's, it's negligible for the everyday person.
74
00:12:00.108 --> 00:12:10.900
Um, you know, if I'm not gonna maybe speak on like Vectorcraft and this stuff, but like Resolve is $300 for the lifetime of the license. Yeah. It's like- Yeah, I-... that's a no-brainer.
75
00:12:10.940 --> 00:12:20.400
No, and all of this dialogue of, of the, the software [laughs] like I didn't see Resolve come up at all because- Yeah, yeah, they're like, "No, no, that, that, leave that alone." [laughs] Yeah. Leave it, yeah.
76
00:12:20.640 --> 00:12:27.080
I mean, well also just basic Resolve is free. It's literally free in download- Yeah... and you just- Right... you, you know, you know it works.
77
00:12:27.280 --> 00:12:40.560
Um- Yeah, I, I think the companies that are in trouble is maybe not even Adobe, but like, you know, if you go up higher in the, in the food chain, like, um, AutoCAD licenses are really expensive, as they should be. Mm.
78
00:12:40.600 --> 00:12:54.940
It's, it's a highly capable tool. Um, you look at, um, Maya. You know, Maya is I think- Mm... something like $10,000 a year on an enterprise fee, if I'm not mistaken. But like, yeah- Yeah...
79
00:12:55.420 --> 00:13:08.800
those, those are gonna be questioned more. Right. But something like AutoCAD is like, okay, are you going to rely on designing a multi-million dollar, hundreds of millions dollar building- No...
80
00:13:08.840 --> 00:13:18.960
infrastructure- Oh my God. Can you imagine? [laughs]... on Vibe Code-it software. [laughs] Yeah. Like building the Burj Khalifa in Dubai with Vibe Code-it software. [laughs] And oh, um, I was off by 10 millimeters. Oh.
81
00:13:19.000 --> 00:13:21.120
And the whole building collapses. You're absolutely right. Oops.
82
00:13:21.140 --> 00:13:35.400
[laughs] I think for some of these, it's more a conversation of how does AutoCAD, um, or Autodesk improve their internal process, where it's like, okay, you can just ship and move stuff faster, but still have the, uh, integrity of, of reliable software.
83
00:13:35.440 --> 00:13:44.700
The other thing the, the SAS companies do that is maybe not public knowledge is that they're willing to build stuff into those tool sets for you for a cost, right?
84
00:13:44.780 --> 00:13:52.780
Like if a studio needs some custom animation tool in Maya, Autodesk will probably build that for you. You know? And I'm just giving, giving example here.
85
00:13:52.960 --> 00:13:58.220
Like in the past, but that would've been like a, like- Like that would've cost you money... enterprise deal. Yeah. That would've been like, yeah.
86
00:13:58.620 --> 00:14:07.720
I mean, I've been wondering too, even just on a user level where it's like, okay, you build the software, you have user preferences where it's like, okay, do you want it like light mode or dark mode?
87
00:14:07.780 --> 00:14:12.320
Do you want, um, your sidebar, you know, what little widgets do you want on your sidebar?
88
00:14:12.700 --> 00:14:40.880
In this future of just whatever generative interfaces look like or something, it's like maybe the back end and everything is still the same of how it operates and works, so it's multi-user, but just maybe the entire interface is built around whatever the specific user needs and how they like to operate and how they like to work, and it's just a whole different paradigm of like what we've been used to of like, okay, you gotta figure out how the software works, and like you're stuck figuring out where the buttons are, and now it's like you, you just- You just- You want, you don't want the buttons there?
89
00:14:41.080 --> 00:14:50.180
Just delete it. Sure. We, we'll move- Yeah, move it... we'll move it. Yeah. Yeah. Build the buttons you need. In defense of traditional SAS, I will also say like, uh, I just thought of SideFX and Houdini, right?
90
00:14:50.220 --> 00:14:57.360
Like Houdini is such a well-respected tool in the industry. It's, it's really magical what they've done over the years, right?
91
00:14:57.400 --> 00:15:09.780
Like pre-, um, pre-real time rendering, that was the only way to get like fluid sims and smoke and particles and collisions, and even outside of that you can do animation rigs and like you can do, really do everything in one package.
92
00:15:10.200 --> 00:15:21.040
And the reason Houdini's successful is because SideFX, who makes it, has scientists and researchers and PhDs like working away, chipping away for years on this stuff.
93
00:15:21.700 --> 00:15:31.920
So like are you gonna Vibe Code something like Houdini? You're not, 'cause that company is specifically built to deliver on something that the normal per- person can't.
94
00:15:32.840 --> 00:15:40.820
Well, I've like sorta wondered about that too, of like, okay, where are these things where they took like really complicated like math breakthroughs or other things?
95
00:15:40.880 --> 00:15:49.800
And, and then I didn't read the whole full, I didn't like dig into it, but I kept seeing the headlines of like the, uh, those like mathematicians at universities- Yeah...
96
00:15:49.880 --> 00:16:00.320
that math problems that have been unsolved for ages, and now it's just like Fable or Astra is just like solving them. Yeah, and, and, uh- And it's just like-... I've seen some [laughs]... it's like, what next?
97
00:16:00.340 --> 00:16:09.420
They're just like, "What the fuck?" Right? Uh, the- [laughs]... the thing that is, uh, even more perplexing is that they're leaving you with the...
98
00:16:09.560 --> 00:16:18.140
The AI models are leaving the mathematicians with the answers and sort of how they deduced it, but then they have to go back and prove it. And so like it's homework on them- Yeah...
99
00:16:18.159 --> 00:16:26.660
to then try to prove this impossible math. Um, so it, it- Homework on the human- On, on the human... or the, on the model? On the human. The model's like, "Yes-" To prove that, uh- "...
100
00:16:26.820 --> 00:16:36.400
X plus Y is this, so then Z plus Y is that." And the human- Yeah... has to be like, "Well..." The me- [laughs] the meaning of life is 42. Yeah, exactly. Don't worry about it. Yeah. Don't worry about it.
101
00:16:36.400 --> 00:16:41.540
That do- just trust me. Yeah, so there, there's a lot of funny stuff happening in the math world right now, for sure. Yeah.
102
00:16:41.600 --> 00:16:48.400
I mean, also I feel like as humans we will always find new unanswered questions for the next thing to figure out.
103
00:16:48.560 --> 00:16:57.640
And so it's like, okay, we'll figure out these math problems, but I'm sure that there will be a new level of things that are just like, okay, we haven't figured out yet, and the AI agents, models can't figure it out.
104
00:16:57.860 --> 00:17:11.000
This is a rumor. This is fun. I think you'll get a laugh out of this, and so would our viewers. Uh, there's a couple of people in- inside Palantir who apparently started smoking now because they think- [laughs]...
105
00:17:11.040 --> 00:17:27.780
we're just gonna cure cancer real soon anyway. [laughs] Like that's the level of, of, uh, trust in AI- Uh... that they have, you know? That's a bit of a, yeah, AI derangement syndrome. But also [laughs]
106
00:17:27.840 --> 00:17:36.940
all right, no diss to smokers, but the thing holding me back from smoking was not the risk of cancer, it was just like I don't, I, [laughs] I didn't like to... I don't wanna smoke. [laughs] It smells bad. Right.
107
00:17:36.980 --> 00:17:45.000
Like yeah. Yeah. [laughs] Next up, sort of, actually, the, I had a lead-in to this next model update- Sure... that also riffs on what we just talked about.
108
00:17:45.040 --> 00:18:00.260
Because going back to SAS, but also the aggregator platform, the AI aggregator platforms that don't necessarily have the original models but are kind of the big tool set, so like Magnific, Higgsfields- OpenArt, Runway.
109
00:18:00.640 --> 00:18:04.220
Well, they have their own model. Oh, sorry. Yeah, the ones that don't. Sorry. Yeah, I'm saying- Yeah...
110
00:18:04.280 --> 00:18:13.304
yeah, everyone that has a model, like they are also trying to be the platform, but the ones that like don't have it are re- like they're not, they don't have like an AI lab- Original model. But what is their defense?
111
00:18:13.344 --> 00:18:21.584
And then in one level we saw Fal, which, uh, you know, was an API aggregator, but then they post-trained Minimax, created their own version of Minimax- Mm-hmm...
112
00:18:21.604 --> 00:18:34.164
that's better, and now it's like, okay, that's a Fal original, so they have something that's original and defensible. Sure. Magnific came out with their own image model. Saw that. Magnific 1. Ish. Magnific 1.
113
00:18:34.704 --> 00:18:43.024
I say ish 'cause actually I, I did some research and we can, we can talk about this. But this definitely seemed s- focused on, I'm guessing, what their core user is of, like, brands. You could do a brand guide.
114
00:18:43.044 --> 00:18:54.284
Um- Yeah, like Instagram commercials... and just create images with... Yeah, images with art direction- Mm-hmm... and, uh, that just gives you better output in the, uh, in the images.
115
00:18:54.584 --> 00:19:07.364
Now, I w- I, I assumed this was, like, a post-trained model on something else, but it seems like in this Venture Beat interview, it's- It's, uh, brand new from scratch, like foundational model? No. Other direction.
116
00:19:07.784 --> 00:19:21.404
It is just a harness around GPT Image 2. Ah. That makes sense. Okay. So that makes me... I'm like, look, no shade to them. I mean, if it works, it works. But, actually, those are hard to defend because it's like if, uh,
117
00:19:22.364 --> 00:19:32.464
OpenAI just adds their own design tool set feature- Yeah... in the app that's already based off GPT Image 2.5, that's, that, that is your- I, I could see from- That is your product...
118
00:19:32.484 --> 00:19:35.084
uh, from a business perspective why they would have their own model.
119
00:19:35.344 --> 00:19:47.324
Like, it would totally make sense from a money standpoint because when you don't have your own model, right, you're basically the middleman, so, you know, uh, the token- Who's... Are you talking about- Magnific.
120
00:19:47.384 --> 00:19:56.384
Yeah, but it's not being, but it's not a cut... Like, it's not what Fal did where they took Minimax, they took the open weights, and then they post-trained and made their own better version of it.
121
00:19:57.144 --> 00:20:00.884
It's like this is just kind of a, a wrapper and harness.
122
00:20:01.064 --> 00:20:09.724
Here, subsequently describe the systems distinguishing technology as prompt-based and consisting of textual instructions shaped by the company's creative special- specialists. Mm-hmm, mm-hmm.
123
00:20:09.744 --> 00:20:18.184
So it's- So they're doing, like, prompt engineering on top of the image model. It's prompt engineering- Yeah. Right... under the hood, which is- It's okay. Yeah... you know, something we've seen in the past.
124
00:20:18.224 --> 00:20:27.284
It works for now until the models just get better and you don't need it anymore. Also, GPT 2.5 is so much better than 2.0. [chuckles] Like, it's a no-brainer. Yeah.
125
00:20:27.324 --> 00:20:32.224
And then it's like that, then you're also stuck in this position 'cause if you built all of these prompts and, like, really, like, worked around- Right...
126
00:20:32.244 --> 00:20:39.744
to figure out how to tweak the model for this one model, and then the new one comes out, and it probably doesn't need as much prompt support as you had before.
127
00:20:40.604 --> 00:20:47.344
But also now you're s- delayed in getting out something new and updated using the latest model.
128
00:20:47.684 --> 00:20:54.224
Secret prompts under the hood I feel like is a tough model because the new, especially the new LLMs, they just keep getting so- So good. Yeah...
129
00:20:54.224 --> 00:20:59.164
much better that they don't need all the skills and instructions and stuff that you had to use before. They're doing it anyway.
130
00:20:59.224 --> 00:21:10.424
So, like, when you prompt into GPT 2.5, I can bet money that there is some kind of LLM, uh, inference right as you prompt that turns it into an expanded prompt, and then that goes into the model.
131
00:21:10.984 --> 00:21:21.604
And, uh, I know you can turn that stuff off in certain, uh, aggregates, but, like, who wants to? Like, it does a really good job. Yeah. You don't need so much of the skill stuff that you built before. Yeah.
132
00:21:22.064 --> 00:21:30.784
Well, what I was gonna say was, like, a couple of speculative things and, [chuckles] uh, man, uh, remember last time Joaquin corrected us when we said some speculative things?
133
00:21:30.824 --> 00:21:36.164
Yeah, you thought they got funding- It's like-... from somewhere else and he was like- Yeah... "No, nothing raised." Yeah. No. Um- Yeah.
134
00:21:36.424 --> 00:21:43.984
I, I'm already saying I could be wrong here, but, like, I think training your own model is a, is a huge expense that could be deducted from taxes.
135
00:21:43.984 --> 00:21:51.144
[laughs] Come, come here for tax advice from- If it's, like, a loss for the business. Yeah. [laughs] This is not tax advice. Yeah.
136
00:21:51.204 --> 00:22:02.604
Like, the infrastructure cost and the, like, just, uh, people and the R&D and everything to build your own model, that, you, that's a nice little write-off, uh, for a company that's really profitable, you know, that needs to show that.
137
00:22:03.144 --> 00:22:17.984
Um, the other thing is if, if GPT 2 weights are really cheap, like if you can just buy them from OpenAI pretty cheap, then why not just make a little bit of extra money from that by just, like, putting a new branding on it, putting a new prompt rewriter on it?
138
00:22:18.044 --> 00:22:30.924
And of course, your power users like you are gonna go to GPT 2.5. You're not gonna touch that. But then somebody who's just getting started on AI will use it, and a lot of that token usage will be in-house.
139
00:22:31.004 --> 00:22:44.564
It'll be like Magnific's token usage. They don't have to pass that on, the profit onto OpenAI. So you're saying they should license, like, the GP2 weights and host it on their own compute to not, to save money?
140
00:22:44.884 --> 00:22:57.124
That's right. Yeah. I think there is a business model, um, where if you have your own weights and you have figured out your own infrastructure, then you can charge the user, let's say, 10% less
141
00:22:58.144 --> 00:23:07.604
for an equivalent model, but then keep a much bigger, uh, portion of the margin because it's your in-house cost. You vertically integrated, right?
142
00:23:07.664 --> 00:23:19.144
Like, it's your own model, your own infrastructure, so you get to keep a bigger chunk of that token usage. Do you know if that's even an option where you could buy or license GPT 2 weights or other- Yeah...
143
00:23:19.204 --> 00:23:19.724
closed source models?
144
00:23:19.784 --> 00:23:36.784
I don't know about GPT 2 specifically, but yeah, you can buy weights from other companies like, um, you know, uh, I don't know, name a- Well, I know there's a lot of companies where you can li- license or they give you licenses if you, uh, do a certain amount, like LTX or Flux, Black Forest Labs.
145
00:23:37.084 --> 00:23:45.604
Yeah. LTX is, like, kinda open source, but then you have to license it. I don't know what the deal is with LTX. It's a license. It was, like, free to use, but if you, like...
146
00:23:45.984 --> 00:23:53.104
It was kinda like the, um, Unreal license thing where it's, like, free to use. If your company makes more than 10 million, then there's, like, uh, you gotta talk to them about licensing. Just imagine this.
147
00:23:53.184 --> 00:24:00.824
I'm just totally making this up, okay? So Nana Banana 2 is a, is a closed model, right? The weights are proprietary to Google. Mm-hmm.
148
00:24:01.104 --> 00:24:16.388
But if a giant like, uh, Amazon or Netflix came to Google and said, "We want, we want to license Nana Banana for $50 million," would they say no? I mean, it's just, like, a matter of business, business practice.
149
00:24:16.388 --> 00:24:27.288
Babe, I mean, Google's, Google's a tricky one because Google would be like, "We already have our own compute." I could see if they went to, you know, some maybe not Anthropic or OpenAI. Well, Anthropic doesn't have this.
150
00:24:27.348 --> 00:24:35.308
But if they had, went to OpenAI, but someone else a little bit lower that needs- Mm-hmm... compute, then maybe that, that kind of thing makes sense.
151
00:24:35.348 --> 00:24:44.208
But look, I mean, if you are like, it is possible to like buy or to, like, pay a bunch of money and- I think so... run, get the weights and run it on your own stuff, I would, would not be surprised. I wouldn't know that.
152
00:24:44.248 --> 00:24:51.868
I mean, I don't, I don't have definitive proof on this stuff, but to me, like, it's a product. It's a thing that you can download and then you can license.
153
00:24:51.948 --> 00:25:03.088
You can attach like a, like a license thing on top of it, and why not make money, extra money on the side, as much as you're doing on, uh, Gemini? Yeah. Anyway. All right. [laughs] Sure.
154
00:25:03.148 --> 00:25:14.368
We'll take Hattie's- Hattie's tinfoil hat theory. Hattie's [laughs] Hat- Hattie tinfoil business plan. [laughs] By the way, get- $200 million... do not come to Denoised for business ideas. This is the worst.
155
00:25:14.528 --> 00:25:24.668
You are gonna go broke, my friend. [laughs] Yeah, just get a bunch of money and license these weights, and then when they come out with the new version, then you gotta give them more money. Yes, exactly.
156
00:25:24.768 --> 00:25:38.148
Oh, isn't that how the AI hype cycle is working right now anyway? It's all funny money. [laughs] Yeah. Yeah. All right, other new models. Muse- Mm-hmm... from Meta came out with a Muse Wonder, which is-
157
00:25:39.208 --> 00:25:44.808
Yeah, can you walk me through-... supposed to be pretty good... what Wonder is? No, I can't, 'cause all their names are Muse, Muse, Muse, Muse.
158
00:25:44.808 --> 00:25:53.948
[laughs] As far as I can tell, Muse Wonder is, like, and you know, according to the graph they have, is, like, their version of Fable and Astra, their, their state-of-the-art- Right... model.
159
00:25:54.208 --> 00:26:06.768
The, is apparently pretty good. If you ask me how would I access Muse Wonder, I don't know. [laughs] Like, I don't know what the equivalent of Claude Desktop is.
160
00:26:06.848 --> 00:26:17.508
I'm sure there's a Muse app, I just have not gone down that rabbit hole yet. Did you see the little, like, puck thing that they debuted with, with the little cute, like, a- animal on it? Yeah, everyone...
161
00:26:17.688 --> 00:26:25.928
Yeah, the, their little little blue thing. Yeah, so, like, you can speak to it, I guess. What does it do? That's what everyone... I mean, I don't know. OpenAI's got one, too. They got the dot thing. Yeah.
162
00:26:25.948 --> 00:26:35.488
Everyone's got little tchotchke AI things. I don't know. Do we need tchotchke AI things? I don't think we n- need more listening devices in our lives. Yeah, um, yeah. I will say
163
00:26:36.428 --> 00:26:45.448
the consumer stuff I have, I don't pay as much attention. Okay, fair enough. I mean, I also don't... Yeah, I don't know if it's really gonna take off. Says the guy with Ray Ban glasses.
164
00:26:45.448 --> 00:26:54.888
[laughs] Those are for filming [laughs] for documentary purposes. Gotcha, man. I got you good. [laughs] You're gonna crap on my business advice? We're gonna go there.
165
00:26:54.888 --> 00:27:09.188
[laughs] Well, okay, also probably part of the reason we didn't see so much about- Mm-hmm... Muse, um, is, uh, this post from Alexander Wang- Mm-hmm... their head of AI, who is also an amazing ship poster on Twitter. Um,
166
00:27:10.408 --> 00:27:18.108
they're basically like, "Yeah, you didn't see a bunch of, um, stuff from us because we built it for coding. We didn't train on 3D stuff. [laughs] We trained on coding, and it's a really good coding agent."
167
00:27:18.168 --> 00:27:26.148
Whole team feels dumb now, but it's true. We did not explicitly train our model on 3D demo games. Obviously, in retrospect, we probably should have done a bit of that, but no regrets.
168
00:27:26.348 --> 00:27:34.208
Please try Muse Wonder on your actual coding work. Oh, so it's really built to be, uh, a software writer. Coder for now. Yeah. Okay. Yeah. Fair enough.
169
00:27:34.628 --> 00:27:45.668
Yeah, he- here are some of his excellent posts of the Muse Labubu thing. Yeah. I know it's not Labubu. I'm sorry I'm infringing on Labubu's trademark, but it's- Yes. Yeah. [laughs] It's not...
170
00:27:45.728 --> 00:27:58.868
It's the, it's a nice-looking Labubu. I thought that Muse was really targeted at marketers. Is that not the case? I am sure under the hood for Facebook Ad Manager, it will be all Muse all the time.
171
00:27:59.088 --> 00:28:08.128
I mean, where is this all going? I'm sure it is to turbocharge making ads between intelligently- Mm-hmm... running them, writing the copy, making the images- Mm-hmm... making the video.
172
00:28:08.528 --> 00:28:16.848
Once they build all that, I'm just gonna run, reverse engineer anything on it, and then I'll have it. [laughs] Just kidding, guys. Just kidding.
173
00:28:16.888 --> 00:28:25.108
[laughs] See, but they got a moat, 'cause they got, they have network effect, so that's one of the other moats, where it's like they got the platform. You can't copy the platform.
174
00:28:25.128 --> 00:28:35.247
Literally everybody on the planet is on their platform, which is crazy. Yeah. Yeah. All right, other new one. We talked about this, but now it's actually out, Gemini 4- Oh, yeah... Argon, and they...
175
00:28:35.308 --> 00:28:46.548
No, they did train it on 3D. Dude, this is so good. So I ran across this video yesterday. Um, I w- uh, Joey, you probably already seen it by the time I sent it to you. I saw it pop up.
176
00:28:46.608 --> 00:28:49.308
I didn't watch the whole thing, so yeah, what, what's, what, what's, what impressed you?
177
00:28:49.448 --> 00:29:00.768
The Gemini team every year, uh, the when- whenever they launch a new Gemini, they have this, like, Japanese pagoda, you know, peaceful zen garden thing that they build, and then they see the quality over time improve.
178
00:29:01.188 --> 00:29:07.948
Mm. And that's, like, a visual indication. So this is their Culver Theater? [laughs] This is their Culver Theater. Yes. [laughs] Okay.
179
00:29:08.008 --> 00:29:29.388
So it's building it from a single prompt, but what's really neat here is, um, that single prompt then goes into an agent that is a blueprint agent, like a architectural blueprint agent, and it'll build a blueprint of this pagoda, and then it'll go into a Nana Banana agent, which is generating images off of that in many different angles for reference.
180
00:29:29.608 --> 00:29:43.808
And then it goes into an adversarial model, or rather, uh, an agent that is, uh, quality controlling the entire build process as multiple agents are building it in Blender. And it's doing this on its own.
181
00:29:43.848 --> 00:29:50.728
You didn't have to, like, give it the instructions to, like, "Hey, do a multi-agent, review your work, and blah, blah, blah." I don't know if you d- So that's the thing.
182
00:29:50.768 --> 00:30:04.468
He said it's off of a single prompt, but how he set up the system may, may have been manual. But regardless, the fact that you can get to this level of quality and fidelity from a single prompt is just nuts.
183
00:30:04.528 --> 00:30:13.268
It's bonkers. It's, this is, like- Yeah... hundreds if not thousands of hours of, um, like, you know, man hours of somebody doing this manually.
184
00:30:13.604 --> 00:30:19.284
By hand Yeah, I mean, that's what has been impressive 'cause, like, the amount of detail work in the single prompts- Right... and not having to- Yeah...
185
00:30:19.364 --> 00:30:27.924
tell it or refine it, and the interior, like, it looks like it built the interior as well. Oh, it's so crazy, and then, yeah, the shaders came in and they- Yeah... they rendered the whole thing in Blender.
186
00:30:28.384 --> 00:30:34.564
Blender seems to be, like, the proving ground, the LM arena of things- [laughs]... nowadays, uh, so yeah.
187
00:30:34.604 --> 00:30:44.884
Well, it's such a good thing too, uh, because it's just also so agent-friendly and easy to control 'cause it's so scriptable. Right. Right So it's just an easy- Yeah... 3D thing to build out and demo. Yeah. Very cool.
188
00:30:45.444 --> 00:30:56.844
You wanna get another one? Yeah, I've got- Florida, Florida take. Um- Yeah, please... this looks like, um, this looks like the Japan land in Epcot. [laughs] Oh. [laughs] This exact angle right here.
189
00:30:56.844 --> 00:31:05.364
[laughs] I've never been. I've never been. I'll take your word for it. I'm sure you've been many times. [laughs] But that just kinda brought the whole thing down a notch.
190
00:31:05.364 --> 00:31:16.564
[laughs] So s- 'Cause I've been to Japan, and I've seen these pagodas in person, and, uh, yeah, when you... It's like, yeah, that, that's in Vegas. It's not a big deal. Like, the Eiffel Tower's in Vegas.
191
00:31:16.624 --> 00:31:25.664
[laughs] I have been to Japan Epcot. [laughs] Check out the world. Uh, can we, can we go to- All right... the next Blender example?
192
00:31:26.344 --> 00:31:46.304
All right, so I guess we don't know what model they used, but this is also done through Blender connected to an LLM, and just incredible the amount of, like, knick-knacks and the real-life randomness that this desk has, and it's all done through an agent.
193
00:31:46.644 --> 00:31:56.184
Yeah, the de- detail here. I saw another one. I think a lot of the ones I've been seeing lately have, they've been saying Opus 5.5- Okay... has been doing it. Yeah.
194
00:31:56.284 --> 00:32:09.664
Um, I mean, also Argon, like, just came out, so I haven't quite seen a showdown yet between, like, Opus 5.5 and Argon. Yeah, here, uh, using Astra and Opus. Astra and Opus, okay. That makes sense.
195
00:32:10.164 --> 00:32:23.864
I'm curious how they handed it off between the two, or if they found one is better at something like research or blocking versus textures, uh, or if it's just purely
196
00:32:25.364 --> 00:32:34.384
maxed out the usage of one, time to switch to the other. [laughs] Uh, also, did you see that it was all from one single image? Like, is that the way, then, just to give it image references?
197
00:32:34.424 --> 00:32:46.464
'Cause that's how the Gemini guy did it, too. Like, Nana Banana agent generated images, and then those images went into the Blender builder. I have found, yeah, images definitely help, so if you give it more stuff,
198
00:32:47.404 --> 00:32:55.304
the better. But also, if you give it just some direction, it will also just start building reference images and making it.
199
00:32:55.364 --> 00:33:06.444
Or if it's also based off something r- real or historical, it'll spin up, like, research agents and then find images online. Yeah Or maps or diagrams or something and then start building off of online research.
200
00:33:06.464 --> 00:33:13.704
Do you think this is why, uh, Tim Sweeney, CEO of Epic, is all of a sudden pro-AI? Like, he's seeing stuff like this. He's like, "Wait a minute."
201
00:33:14.304 --> 00:33:26.964
I mean, I think if you're any type of game or software developer, if you're not thinking of AI first, like, you're just gonna be left in the dust. Yeah. So he's just basically, um- Where the wind is blowing Yeah.
202
00:33:27.004 --> 00:33:40.624
It's that. This is an Argon example, this cathedral build. From an image. From a single image, and then it built and lit. It's doing, like, render passes, yeah. Oh my God, it has, like, volumetric fog and everything.
203
00:33:41.164 --> 00:33:57.444
Uh, the other one I was looking for was this laundromat video. This is Opus 5.5. Which is, uh, one of my faves. I love Opus 5.5, uh, but I have never tried to do this with Blender in it. You know what?
204
00:33:57.704 --> 00:34:07.784
I'm inspired today. I'm gonna try something like this. Okay, cool. Yeah. Not a laundromat, maybe something more visually appealing. [laughs] Visually. [laughs] Were you just about to say something?
205
00:34:07.844 --> 00:34:14.284
No, just, uh, like, Epic had the lead, you know, like, five, six years ago, pre-generative AI boom.
206
00:34:14.924 --> 00:34:27.044
Like, they, they were, and they, I think they still are, they have AI talent in-house because, like, a lot of the game engine stuff needs to be agentic. Like, uh, if you look at NC... you know, NPCs and things like that.
207
00:34:27.584 --> 00:34:38.024
And I don't know why, but they totally missed the boat on the first wave of AI. Like, uh, you know, the, the way that sort of Blender has kinda caught that, Unreal didn't.
208
00:34:38.444 --> 00:34:47.924
Like, in just making- Just being a platform for where AI progress happens and testing happens, Unreal could've been that platform. I think
209
00:34:49.024 --> 00:34:54.704
the ones that were successful in that is, was accidental just because they happened to build a platform that was very scripting-friendly.
210
00:34:54.984 --> 00:35:04.544
Right So, like, Blender is very scripting-friendly, Resolve is scripting-friendly, and that just works really well with an agent. Yeah. As far as I know, like, Unreal is also script-friendly.
211
00:35:04.604 --> 00:35:16.864
Like, it's all C++ Python supported. But maybe, yeah. For building or for blueprints or both? Oh, you're right, yeah, for blueprints. Yeah. Like, I'd expect blueprints- Yeah... yeah, for sure. Yeah. You could...
212
00:35:16.904 --> 00:35:25.004
That one should be scriptable, but, like, the actual built, built model out of world- Right... I don't, I don't think so. Uh, yeah, you're right. Um, I mean, I guess recently they now have an MCP server.
213
00:35:25.044 --> 00:35:37.124
But yeah, it, it's just such a bummer- Yeah... to me, and I know so many people in the, the Unreal Engine world, you know, that could have really benefited from a lot of the AI progress.
214
00:35:37.184 --> 00:35:44.884
Like, it would've been just like rising tides floats all boats kind of a thing, you know, where modeling becomes- Mm-hmm... easier, shading becomes easier.
215
00:35:44.964 --> 00:35:57.564
You can do more, you can build games faster in Unreal, and that would in turn help the, uh, Fortnite ecosystem and all of the stuff that makes money for Epic just be more profitable. Like, it would've been a win-win.
216
00:35:57.704 --> 00:36:08.944
But nevertheless, they're getting into it now. Um, I sent you the tweet from Tim Sweeney the other day. "We have a huge pivot underway here at Epic, moving developers around to retool for maximum AI acceleration."
217
00:36:09.404 --> 00:36:12.944
The aim isn't to do the same stuff with fewer people. AI or SI.
218
00:36:12.944 --> 00:36:26.024
[laughs] I mean, I think you gotta get on board to accelerate your work, uh, not on board to use AI to, uh, replace everything, but use AI to accelerate the human-directed work.
219
00:36:26.184 --> 00:36:40.364
Ooh, look, he replied to somebody who was, uh, dropping some hate. "Your inability to read the room is really, is really astounding, Tim." And then he responded, "Much of the room wants to be in 2016 again, but it's not.
220
00:36:40.464 --> 00:36:48.944
Our job as the provider of tools and technology to much of the industry is to design for the present, and that is, and the future that will soon be.
221
00:36:49.144 --> 00:37:01.784
Personally, I'm optimistic about it and see it as empowering for creators." Dude, that's how we feel. Yeah. This guy knows. Tim Sweeney knows. [laughs] To... No, I met him once in person. He's, he's awesome. All right.
222
00:37:02.104 --> 00:37:13.484
Good place to wrap it up. For those of you who've been listening on Apple Podcasts, give us a like, and, uh, a five-star review would be really nice. You know, it helps with the algos. You know how it goes. It does, yep.
223
00:37:14.044 --> 00:37:20.064
Uh, links to everything we talked about at denoisepodcast.com. Thanks for watching. We'll catch you in the next episode.
1
00:00:00.040 --> 00:00:07.000
There's a level of people where it's just like, "I just wanna pay stuff for something that works." You and I both have coffee machines at home, but sometimes we just go to Starbucks.
2
00:00:07.280 --> 00:00:21.360
[laughs] It's just easier to just pay. Welcome back to Denoised. How's it going, Addy? Long time no see, Joey. Let's go. I know. I know. Finishing up some projects, but, uh, it'll be out.
3
00:00:21.440 --> 00:00:31.300
Actually, by the time this is out, the trailer for it will be out. So- Oh... I'll talk about it next week. Okay. Just stay tuned. Uh, have you been doing any LA Tech Week stuff?
4
00:00:32.380 --> 00:00:39.320
Uh, I went to the AI on the Lot event, but I haven't- Nice... that's about it. What about you? Nah. I still- [laughs]...
5
00:00:39.360 --> 00:00:49.940
cannot fully wrap my head around what exactly Tech Week is besides panels and parties that are loosely organized. You're right. Uh, to me, it feels very VC-driven, like A16Z- Yeah...
6
00:00:50.180 --> 00:00:59.860
and that, that kind of thing, where it's like, uh, get a bunch of founders together, figure out some new ideas, so we can go invest in the thing. Yeah. I don't know.
7
00:00:59.960 --> 00:01:11.340
Everything I've always gone to with these was, like, a zoo, and it's just, like, overpacked too. Yeah. So okay. Speaking of building and whatever the future of software means [laughs]. Oh, God.
8
00:01:11.740 --> 00:01:24.720
[laughs] A couple things dropped in the last, like, week or so. One of the big ones is REA, Reverse Engineer Anything. So crazy, man. Which is- It's giving me the chills just looking at this...
9
00:01:24.760 --> 00:01:36.920
it's a, uh, FCP tool for basically kinda point it at an app, and it can reverse, or I believe a website, and it can reverse engineer how it was built, and then build, vibe code your own version of it. That's right, yeah.
10
00:01:36.980 --> 00:01:51.960
From my understanding of it, um, it gets into the post-compile binary, and then it can reverse engineer that to human-readable code. So, like, imagine all the ones and zeros- Okay... turned back into C++ code.
11
00:01:52.320 --> 00:01:57.200
I don't know how it does it. It's crazy because decompiling is, like, incredibly difficult.
12
00:01:57.600 --> 00:02:07.020
And, uh, it does it, uh, well enough to then recreate the application where you can take over and rebuild it or, you know, modify it to your will. Yeah.
13
00:02:07.420 --> 00:02:13.940
The probably biggest example going around with this, I don't know if they use this or this, the person just vibe coded it, but someone built...
14
00:02:14.120 --> 00:02:26.880
It's called Crafting Apps, and there is a Photo Craft, Film Craft, Light Craft. It is basically a vibe-coded version of Adobe's Creative Suite and other products and other vibe-coded apps. Um- Yeah. Yeah...
15
00:02:26.960 --> 00:02:36.480
this has been probably the, kind of the biggest example going around of just like, oh, well, you know, what are you gonna do if you're Adobe or any other major software- Yeah...
16
00:02:36.500 --> 00:02:53.240
and someone sort of built a free version of it? Now, apps and heavy-duty stuff like this, this feels like the last thing that I would j- jump ship or switch on [laughs] and change my entire workflow to a vibe-coded app.
17
00:02:53.300 --> 00:03:05.120
Also, there's been free video editors and other tools out there forever. There's Photopea, which has been online forever, which is like a web-based version of Photoshop that's ad-supported and, and free to use. If you're
18
00:03:06.020 --> 00:03:14.240
someone in high school, you know, back in the day, I, you know, you, you found the codes to, um, [laughs] unlock Creative Suite or the other apps- I don't know what you're talking about. No.
19
00:03:14.240 --> 00:03:21.260
[laughs] It's actually, this was even before Creative Suite existed. This was just purely when you had to get the code for- The, the CDs and the-... for the software... and the floppies even. Yeah.
20
00:03:21.580 --> 00:03:28.020
Codes floated around the internet, you know, if you were in middle school, high school, and you didn't, um, have any money to pay for, for expensive software.
21
00:03:28.860 --> 00:03:41.500
You know, if you're dabbling, but if you're in an enterprise company or a company where you need stuff to rely on. Yeah. Yeah. I, well, okay. The, going back to the '90s, I think it was called a crack, right?
22
00:03:41.800 --> 00:03:51.200
So when, [laughs] when you would- A crack... download the ISO, the ISO is what this is. You have to download the ISO, which was the- Yeah... disk image... which is the binary code, yeah. Uh-huh.
23
00:03:51.240 --> 00:04:07.720
And then you can then, the, the ISO would mount to your hard drive like a disk, and then you would crack the license activation schema, and off you go. You have Photoshop on your, uh, Windows 95 box or whatnot.
24
00:04:08.000 --> 00:04:19.540
[laughs] I was going for Premiere, Premiere 6.5. [laughs] Oh, okay. Wow. All right. [laughs] You were into video long before me, my friend. [laughs] Yeah, I was doing video first. Yeah.
25
00:04:19.880 --> 00:04:24.420
The, the bigger existential thing is just like, okay, is software dead? Is SaaS dead? If you can just sort of build,
26
00:04:25.340 --> 00:04:34.080
these are clones of other things, but if you could just build whatever you need, you need a video editor, you need it with whatever specific features you need, where does this go? Yeah. I, I am...
27
00:04:34.580 --> 00:04:46.580
Okay, I have two trains of thought on this. So we could sp- speak specifically on the Photo Craft, Vector Craft, like this thing. I, I believe this was REA, um, devised. Like, they did reverse engineer- Mm-hmm...
28
00:04:46.600 --> 00:04:51.100
current things. They also got Word Craft, kind of a, a- Yeah... Word- And that's why they were able to nail the UI- Microsoft Word...
29
00:04:51.160 --> 00:05:03.040
so well, like almost pixel to pixel, 'cause it is just literally the ISO file, the binary file. But from what I am hearing, the performance is severely lacking.
30
00:05:03.160 --> 00:05:16.460
So, like, um, a lot of the, the magic in Premiere, and even in Resolve, a lot of the NLEs, is, like, how well it handles low-level VDU, uh, video computations on the GPU, right?
31
00:05:16.560 --> 00:05:27.860
Like when you're, uh, doing a time remap or when you're putting a transition between two shots, that, that, that's, like, video calculations happening in real time for you to preview it, and then when it renders, it renders, right?
32
00:05:27.960 --> 00:05:37.500
Um, like that stuff is not working correctly. So you can load your timeline, but the preview will be slow and choppy, and then some of the transitions won't work.
33
00:05:37.540 --> 00:05:46.740
And, like, I've seen a couple of videos on YouTube where, um, you know, they use Video Craft, and it's like, oof, oof, that's not quite there yet. Yeah.
34
00:05:46.840 --> 00:05:55.480
I mean, I would, yeah, I, I am not surprised that there's performance lagging issues. It's not there yet, but we can see where this is going, obviously. Exactly. Exactly.
35
00:05:55.500 --> 00:06:07.900
And where regardless if you have to point it at a working app, I mean, this is like the Higgs field jujitsu version of film where they take an existing film and restyle it. Oh, I hate that. I hate that. Yeah.
36
00:06:08.220 --> 00:06:15.168
Like they- This is the same thing where you point it at a- An application that- Yeah... uh, has had a lot of development and works, and then copying it. Right.
37
00:06:15.308 --> 00:06:24.868
But put that aside, and it's just like, "I need an editor, I need a text editor, I need something. Uh, let me start building it." Yeah. We, we tend to forget- And optimize it...
38
00:06:24.948 --> 00:06:37.188
how crazy all of this is because we're just inundated by one development, one progress, one thing after another, right? Like, wait, this was like NSA technology like five years ago.
39
00:06:37.368 --> 00:06:45.948
Only like the highest levels of the government do they have the technology to de-reverse engineer a, a piece of software that China built or whatnot.
40
00:06:45.988 --> 00:06:57.088
Like, this was used for like geopolitical chess, and now anybody has it. Like, we can... Like, think of an application, um, that you use every day, right?
41
00:06:57.168 --> 00:07:08.608
Uh, I mean, of course we use these applications every day, but also stuff like, I don't know, like, a w- like macOS or something crazy like that, and you can reverse engineer that probably.
42
00:07:08.908 --> 00:07:15.288
I, well, I mentioned about the enterprise thing before. Like, I think there's still a level, it's like, okay, so maybe just software development evolves or the companies evolve where...
43
00:07:15.708 --> 00:07:20.868
Like, 'cause you can move so fast and you build things so fast, obviously not everyone's gonna wanna vibe code- No...
44
00:07:20.948 --> 00:07:31.168
their own solutions and have some duct tape thing together, and support that, and keep iterating and building on it. Even as good as agents get, like you still need something to guide it or just- Yeah...
45
00:07:31.208 --> 00:07:42.748
make sure it's all working. And it's like maybe that's what software companies kinda turn into, um, as the just- The maintenance... they can run leaner and faster and push stuff out.
46
00:07:43.288 --> 00:07:53.288
Uh, but yeah, I'm also thinking of like Linux, right? Linux has been around, and Linux is free, but they, I believe it's Linux. I mean, they make a
47
00:07:54.688 --> 00:08:03.948
bunch of money off of supporting and running the servers and like the professional services. I see. I see. Oh, I see- Side... what you're saying. Like, you think that's the business... That is no business model then.
48
00:08:04.108 --> 00:08:13.728
Like, it's very slim pickings at that point. Uh, well, I mean, the other argument, or like, okay, if the software is not a moat anymore and it's what you build- Right...
49
00:08:14.108 --> 00:08:23.008
something else has to be the secret sauce, which is proprietary data or something else that- I think, uh, these are great questions... can't be replicated. Yeah.
50
00:08:23.048 --> 00:08:32.408
Uh, the one thing I will say is, um, we're, we're forgetting to mention that this makes theft really easy. Like, that's no bueno in my book, in your book too I'm sure.
51
00:08:32.448 --> 00:08:42.808
Like, shouldn't be stealing software, just like [laughs] you shouldn't be stealing, uh, data for training models and stuff like that. Yeah. I mean, look, obviously if you're pointed at like,
52
00:08:43.848 --> 00:08:49.648
I don't wanna keep picking on Adobe. If you point it at Resolve, you point it at Final Cut- Yeah... and you're like, "Copy that and make me a version of that"- Dude... that's like, okay, yeah- Yeah...
53
00:08:49.688 --> 00:09:02.588
that's like ripping it off. But if you're just like, "I need a video editor- Yeah... that does X, Y, and Z," it, I know it is, you know, completely capable of starting to build stuff like that, which arguably could be
54
00:09:03.728 --> 00:09:10.848
better because it is like, okay, you're building a custom piece of software built exactly for what you need to do and your specific needs. Mm-hmm. Exactly.
55
00:09:11.168 --> 00:09:21.248
But- Like, for this podcast, we don't need the entire might of Resolve or Premiere or Final Cut. Like, we just need a couple of things to make this podcast edited, right? Yeah.
56
00:09:21.308 --> 00:09:27.668
I mean, obviously there have been gaps in editing this podcast because even with these faster podcast editors, they don't really support...
57
00:09:27.888 --> 00:09:34.388
Like, I, uh, so many times having conversations where it's like, okay, we have your feed, we have my feed. I need to put them like split side by side. Right.
58
00:09:34.888 --> 00:09:42.628
And I need that to be treated as an angle to cut to, and a lot of these auto-cutting things like don't understand that. They don't create a third- And-... third camera that is the two of us.
59
00:09:42.668 --> 00:09:49.908
Yeah, it's like I need a fake third camera- Yeah... that is a two-shot of us. Right. And I have, yeah, on the verge of just being like, "Oh, let's just- I'll, I'll just build this...
60
00:09:50.008 --> 00:10:01.268
just build, [laughs] just build something." [laughs] Well, that, that, that's also the incredible moment in humanity that we're going through is that, uh, the average Joe like you and me can just build something.
61
00:10:01.648 --> 00:10:14.408
We don't have computer science backgrounds, and we don't program every day of our lives, but we're gonna build something that requires thousands and hundreds of thousands of lines of code, and we can do it. Crazy, right?
62
00:10:14.568 --> 00:10:23.868
Yeah. Yeah, I mean, it's been a crazy unlock. Yeah. 2026 has been an insane- 2020-... year. Yeah... yeah. And we have more stuff to cover [laughs] as w- we'll go through. Like, it's just getting started.
63
00:10:23.908 --> 00:10:30.328
Like, the entire, um, self, uh, sort of, um, [clicks tongue] uh, like just being able
64
00:10:31.367 --> 00:10:45.648
to build an entire tech stack yourself, plus make content, plus, you know, do all these other things is, is insane to me, like the h- power we have- Yeah... at our fingertips. Yeah. You know, I mean, we're, we're [sighs]
65
00:10:45.728 --> 00:10:55.128
tinkers, tinklers. Tinkers. [laughs] Tinkerers. Yeah. Like [laughs] I was running through five different versions of that word in my head. I'm like, "That's not, [laughs] that's not the right one." Yeah, yeah.
66
00:10:55.168 --> 00:11:03.248
Your, your NyQuil is hitting real hard right now. [laughs] Yeah. [laughs] And you know, obviously there's a level of people where it's just like, "I just wanna pay so- for something that works.
67
00:11:03.328 --> 00:11:07.208
I don't wanna, I don't wanna deal with- Yeah... all of this- I mean-... even if as easy as it gets." That's the thing.
68
00:11:07.248 --> 00:11:16.968
Like, uh, you know, you and I both have coffee machines at home, but sometimes we just go to Starbucks. It's just like- [laughs]... it's just easier to just pay- [laughs]... ex- for expensive coffee, you know?
69
00:11:17.128 --> 00:11:24.208
[laughs] I thought we were going with that as in we have coffee machines at home, but sometimes we just want to have like a nice pour-over and grind the beans ourself. But you [laughs] went the other direction.
70
00:11:24.288 --> 00:11:34.868
Well, yeah, like the, the option that you're getting now is you're not, you don't just have the pour-over and the grinding, but you can grow the coffee bean, right? Like, that's how crazy- Yeah... things are getting.
71
00:11:34.908 --> 00:11:45.748
Like, you can really go back to the source. But there's nothing wrong with going to Starbucks and just paying, you know, $6 for a giant, uh, venti whatever. Yeah, it's like I need coffee- Yeah... I need it now.
72
00:11:46.468 --> 00:11:48.628
I don't wanna- Yeah... do that- And I think-... headache...
73
00:11:48.688 --> 00:11:59.288
like a lot of the software moat in that sense I think will be undisrupted is because they're low-priced enough to where it's, it's negligible for the everyday person.
74
00:12:00.108 --> 00:12:10.900
Um, you know, if I'm not gonna maybe speak on like Vectorcraft and this stuff, but like Resolve is $300 for the lifetime of the license. Yeah. It's like- Yeah, I-... that's a no-brainer.
75
00:12:10.940 --> 00:12:20.400
No, and all of this dialogue of, of the, the software [laughs] like I didn't see Resolve come up at all because- Yeah, yeah, they're like, "No, no, that, that, leave that alone." [laughs] Yeah. Leave it, yeah.
76
00:12:20.640 --> 00:12:27.080
I mean, well also just basic Resolve is free. It's literally free in download- Yeah... and you just- Right... you, you know, you know it works.
77
00:12:27.280 --> 00:12:40.560
Um- Yeah, I, I think the companies that are in trouble is maybe not even Adobe, but like, you know, if you go up higher in the, in the food chain, like, um, AutoCAD licenses are really expensive, as they should be. Mm.
78
00:12:40.600 --> 00:12:54.940
It's, it's a highly capable tool. Um, you look at, um, Maya. You know, Maya is I think- Mm... something like $10,000 a year on an enterprise fee, if I'm not mistaken. But like, yeah- Yeah...
79
00:12:55.420 --> 00:13:08.800
those, those are gonna be questioned more. Right. But something like AutoCAD is like, okay, are you going to rely on designing a multi-million dollar, hundreds of millions dollar building- No...
80
00:13:08.840 --> 00:13:18.960
infrastructure- Oh my God. Can you imagine? [laughs]... on Vibe Code-it software. [laughs] Yeah. Like building the Burj Khalifa in Dubai with Vibe Code-it software. [laughs] And oh, um, I was off by 10 millimeters. Oh.
81
00:13:19.000 --> 00:13:21.120
And the whole building collapses. You're absolutely right. Oops.
82
00:13:21.140 --> 00:13:35.400
[laughs] I think for some of these, it's more a conversation of how does AutoCAD, um, or Autodesk improve their internal process, where it's like, okay, you can just ship and move stuff faster, but still have the, uh, integrity of, of reliable software.
83
00:13:35.440 --> 00:13:44.700
The other thing the, the SAS companies do that is maybe not public knowledge is that they're willing to build stuff into those tool sets for you for a cost, right?
84
00:13:44.780 --> 00:13:52.780
Like if a studio needs some custom animation tool in Maya, Autodesk will probably build that for you. You know? And I'm just giving, giving example here.
85
00:13:52.960 --> 00:13:58.220
Like in the past, but that would've been like a, like- Like that would've cost you money... enterprise deal. Yeah. That would've been like, yeah.
86
00:13:58.620 --> 00:14:07.720
I mean, I've been wondering too, even just on a user level where it's like, okay, you build the software, you have user preferences where it's like, okay, do you want it like light mode or dark mode?
87
00:14:07.780 --> 00:14:12.320
Do you want, um, your sidebar, you know, what little widgets do you want on your sidebar?
88
00:14:12.700 --> 00:14:40.880
In this future of just whatever generative interfaces look like or something, it's like maybe the back end and everything is still the same of how it operates and works, so it's multi-user, but just maybe the entire interface is built around whatever the specific user needs and how they like to operate and how they like to work, and it's just a whole different paradigm of like what we've been used to of like, okay, you gotta figure out how the software works, and like you're stuck figuring out where the buttons are, and now it's like you, you just- You just- You want, you don't want the buttons there?
89
00:14:41.080 --> 00:14:50.180
Just delete it. Sure. We, we'll move- Yeah, move it... we'll move it. Yeah. Yeah. Build the buttons you need. In defense of traditional SAS, I will also say like, uh, I just thought of SideFX and Houdini, right?
90
00:14:50.220 --> 00:14:57.360
Like Houdini is such a well-respected tool in the industry. It's, it's really magical what they've done over the years, right?
91
00:14:57.400 --> 00:15:09.780
Like pre-, um, pre-real time rendering, that was the only way to get like fluid sims and smoke and particles and collisions, and even outside of that you can do animation rigs and like you can do, really do everything in one package.
92
00:15:10.200 --> 00:15:21.040
And the reason Houdini's successful is because SideFX, who makes it, has scientists and researchers and PhDs like working away, chipping away for years on this stuff.
93
00:15:21.700 --> 00:15:31.920
So like are you gonna Vibe Code something like Houdini? You're not, 'cause that company is specifically built to deliver on something that the normal per- person can't.
94
00:15:32.840 --> 00:15:40.820
Well, I've like sorta wondered about that too, of like, okay, where are these things where they took like really complicated like math breakthroughs or other things?
95
00:15:40.880 --> 00:15:49.800
And, and then I didn't read the whole full, I didn't like dig into it, but I kept seeing the headlines of like the, uh, those like mathematicians at universities- Yeah...
96
00:15:49.880 --> 00:16:00.320
that math problems that have been unsolved for ages, and now it's just like Fable or Astra is just like solving them. Yeah, and, and, uh- And it's just like-... I've seen some [laughs]... it's like, what next?
97
00:16:00.340 --> 00:16:09.420
They're just like, "What the fuck?" Right? Uh, the- [laughs]... the thing that is, uh, even more perplexing is that they're leaving you with the...
98
00:16:09.560 --> 00:16:18.140
The AI models are leaving the mathematicians with the answers and sort of how they deduced it, but then they have to go back and prove it. And so like it's homework on them- Yeah...
99
00:16:18.159 --> 00:16:26.660
to then try to prove this impossible math. Um, so it, it- Homework on the human- On, on the human... or the, on the model? On the human. The model's like, "Yes-" To prove that, uh- "...
100
00:16:26.820 --> 00:16:36.400
X plus Y is this, so then Z plus Y is that." And the human- Yeah... has to be like, "Well..." The me- [laughs] the meaning of life is 42. Yeah, exactly. Don't worry about it. Yeah. Don't worry about it.
101
00:16:36.400 --> 00:16:41.540
That do- just trust me. Yeah, so there, there's a lot of funny stuff happening in the math world right now, for sure. Yeah.
102
00:16:41.600 --> 00:16:48.400
I mean, also I feel like as humans we will always find new unanswered questions for the next thing to figure out.
103
00:16:48.560 --> 00:16:57.640
And so it's like, okay, we'll figure out these math problems, but I'm sure that there will be a new level of things that are just like, okay, we haven't figured out yet, and the AI agents, models can't figure it out.
104
00:16:57.860 --> 00:17:11.000
This is a rumor. This is fun. I think you'll get a laugh out of this, and so would our viewers. Uh, there's a couple of people in- inside Palantir who apparently started smoking now because they think- [laughs]...
105
00:17:11.040 --> 00:17:27.780
we're just gonna cure cancer real soon anyway. [laughs] Like that's the level of, of, uh, trust in AI- Uh... that they have, you know? That's a bit of a, yeah, AI derangement syndrome. But also [laughs]
106
00:17:27.840 --> 00:17:36.940
all right, no diss to smokers, but the thing holding me back from smoking was not the risk of cancer, it was just like I don't, I, [laughs] I didn't like to... I don't wanna smoke. [laughs] It smells bad. Right.
107
00:17:36.980 --> 00:17:45.000
Like yeah. Yeah. [laughs] Next up, sort of, actually, the, I had a lead-in to this next model update- Sure... that also riffs on what we just talked about.
108
00:17:45.040 --> 00:18:00.260
Because going back to SAS, but also the aggregator platform, the AI aggregator platforms that don't necessarily have the original models but are kind of the big tool set, so like Magnific, Higgsfields- OpenArt, Runway.
109
00:18:00.640 --> 00:18:04.220
Well, they have their own model. Oh, sorry. Yeah, the ones that don't. Sorry. Yeah, I'm saying- Yeah...
110
00:18:04.280 --> 00:18:13.304
yeah, everyone that has a model, like they are also trying to be the platform, but the ones that like don't have it are re- like they're not, they don't have like an AI lab- Original model. But what is their defense?
111
00:18:13.344 --> 00:18:21.584
And then in one level we saw Fal, which, uh, you know, was an API aggregator, but then they post-trained Minimax, created their own version of Minimax- Mm-hmm...
112
00:18:21.604 --> 00:18:34.164
that's better, and now it's like, okay, that's a Fal original, so they have something that's original and defensible. Sure. Magnific came out with their own image model. Saw that. Magnific 1. Ish. Magnific 1.
113
00:18:34.704 --> 00:18:43.024
I say ish 'cause actually I, I did some research and we can, we can talk about this. But this definitely seemed s- focused on, I'm guessing, what their core user is of, like, brands. You could do a brand guide.
114
00:18:43.044 --> 00:18:54.284
Um- Yeah, like Instagram commercials... and just create images with... Yeah, images with art direction- Mm-hmm... and, uh, that just gives you better output in the, uh, in the images.
115
00:18:54.584 --> 00:19:07.364
Now, I w- I, I assumed this was, like, a post-trained model on something else, but it seems like in this Venture Beat interview, it's- It's, uh, brand new from scratch, like foundational model? No. Other direction.
116
00:19:07.784 --> 00:19:21.404
It is just a harness around GPT Image 2. Ah. That makes sense. Okay. So that makes me... I'm like, look, no shade to them. I mean, if it works, it works. But, actually, those are hard to defend because it's like if, uh,
117
00:19:22.364 --> 00:19:32.464
OpenAI just adds their own design tool set feature- Yeah... in the app that's already based off GPT Image 2.5, that's, that, that is your- I, I could see from- That is your product...
118
00:19:32.484 --> 00:19:35.084
uh, from a business perspective why they would have their own model.
119
00:19:35.344 --> 00:19:47.324
Like, it would totally make sense from a money standpoint because when you don't have your own model, right, you're basically the middleman, so, you know, uh, the token- Who's... Are you talking about- Magnific.
120
00:19:47.384 --> 00:19:56.384
Yeah, but it's not being, but it's not a cut... Like, it's not what Fal did where they took Minimax, they took the open weights, and then they post-trained and made their own better version of it.
121
00:19:57.144 --> 00:20:00.884
It's like this is just kind of a, a wrapper and harness.
122
00:20:01.064 --> 00:20:09.724
Here, subsequently describe the systems distinguishing technology as prompt-based and consisting of textual instructions shaped by the company's creative special- specialists. Mm-hmm, mm-hmm.
123
00:20:09.744 --> 00:20:18.184
So it's- So they're doing, like, prompt engineering on top of the image model. It's prompt engineering- Yeah. Right... under the hood, which is- It's okay. Yeah... you know, something we've seen in the past.
124
00:20:18.224 --> 00:20:27.284
It works for now until the models just get better and you don't need it anymore. Also, GPT 2.5 is so much better than 2.0. [chuckles] Like, it's a no-brainer. Yeah.
125
00:20:27.324 --> 00:20:32.224
And then it's like that, then you're also stuck in this position 'cause if you built all of these prompts and, like, really, like, worked around- Right...
126
00:20:32.244 --> 00:20:39.744
to figure out how to tweak the model for this one model, and then the new one comes out, and it probably doesn't need as much prompt support as you had before.
127
00:20:40.604 --> 00:20:47.344
But also now you're s- delayed in getting out something new and updated using the latest model.
128
00:20:47.684 --> 00:20:54.224
Secret prompts under the hood I feel like is a tough model because the new, especially the new LLMs, they just keep getting so- So good. Yeah...
129
00:20:54.224 --> 00:20:59.164
much better that they don't need all the skills and instructions and stuff that you had to use before. They're doing it anyway.
130
00:20:59.224 --> 00:21:10.424
So, like, when you prompt into GPT 2.5, I can bet money that there is some kind of LLM, uh, inference right as you prompt that turns it into an expanded prompt, and then that goes into the model.
131
00:21:10.984 --> 00:21:21.604
And, uh, I know you can turn that stuff off in certain, uh, aggregates, but, like, who wants to? Like, it does a really good job. Yeah. You don't need so much of the skill stuff that you built before. Yeah.
132
00:21:22.064 --> 00:21:30.784
Well, what I was gonna say was, like, a couple of speculative things and, [chuckles] uh, man, uh, remember last time Joaquin corrected us when we said some speculative things?
133
00:21:30.824 --> 00:21:36.164
Yeah, you thought they got funding- It's like-... from somewhere else and he was like- Yeah... "No, nothing raised." Yeah. No. Um- Yeah.
134
00:21:36.424 --> 00:21:43.984
I, I'm already saying I could be wrong here, but, like, I think training your own model is a, is a huge expense that could be deducted from taxes.
135
00:21:43.984 --> 00:21:51.144
[laughs] Come, come here for tax advice from- If it's, like, a loss for the business. Yeah. [laughs] This is not tax advice. Yeah.
136
00:21:51.204 --> 00:22:02.604
Like, the infrastructure cost and the, like, just, uh, people and the R&D and everything to build your own model, that, you, that's a nice little write-off, uh, for a company that's really profitable, you know, that needs to show that.
137
00:22:03.144 --> 00:22:17.984
Um, the other thing is if, if GPT 2 weights are really cheap, like if you can just buy them from OpenAI pretty cheap, then why not just make a little bit of extra money from that by just, like, putting a new branding on it, putting a new prompt rewriter on it?
138
00:22:18.044 --> 00:22:30.924
And of course, your power users like you are gonna go to GPT 2.5. You're not gonna touch that. But then somebody who's just getting started on AI will use it, and a lot of that token usage will be in-house.
139
00:22:31.004 --> 00:22:44.564
It'll be like Magnific's token usage. They don't have to pass that on, the profit onto OpenAI. So you're saying they should license, like, the GP2 weights and host it on their own compute to not, to save money?
140
00:22:44.884 --> 00:22:57.124
That's right. Yeah. I think there is a business model, um, where if you have your own weights and you have figured out your own infrastructure, then you can charge the user, let's say, 10% less
141
00:22:58.144 --> 00:23:07.604
for an equivalent model, but then keep a much bigger, uh, portion of the margin because it's your in-house cost. You vertically integrated, right?
142
00:23:07.664 --> 00:23:19.144
Like, it's your own model, your own infrastructure, so you get to keep a bigger chunk of that token usage. Do you know if that's even an option where you could buy or license GPT 2 weights or other- Yeah...
143
00:23:19.204 --> 00:23:19.724
closed source models?
144
00:23:19.784 --> 00:23:36.784
I don't know about GPT 2 specifically, but yeah, you can buy weights from other companies like, um, you know, uh, I don't know, name a- Well, I know there's a lot of companies where you can li- license or they give you licenses if you, uh, do a certain amount, like LTX or Flux, Black Forest Labs.
145
00:23:37.084 --> 00:23:45.604
Yeah. LTX is, like, kinda open source, but then you have to license it. I don't know what the deal is with LTX. It's a license. It was, like, free to use, but if you, like...
146
00:23:45.984 --> 00:23:53.104
It was kinda like the, um, Unreal license thing where it's, like, free to use. If your company makes more than 10 million, then there's, like, uh, you gotta talk to them about licensing. Just imagine this.
147
00:23:53.184 --> 00:24:00.824
I'm just totally making this up, okay? So Nana Banana 2 is a, is a closed model, right? The weights are proprietary to Google. Mm-hmm.
148
00:24:01.104 --> 00:24:16.388
But if a giant like, uh, Amazon or Netflix came to Google and said, "We want, we want to license Nana Banana for $50 million," would they say no? I mean, it's just, like, a matter of business, business practice.
149
00:24:16.388 --> 00:24:27.288
Babe, I mean, Google's, Google's a tricky one because Google would be like, "We already have our own compute." I could see if they went to, you know, some maybe not Anthropic or OpenAI. Well, Anthropic doesn't have this.
150
00:24:27.348 --> 00:24:35.308
But if they had, went to OpenAI, but someone else a little bit lower that needs- Mm-hmm... compute, then maybe that, that kind of thing makes sense.
151
00:24:35.348 --> 00:24:44.208
But look, I mean, if you are like, it is possible to like buy or to, like, pay a bunch of money and- I think so... run, get the weights and run it on your own stuff, I would, would not be surprised. I wouldn't know that.
152
00:24:44.248 --> 00:24:51.868
I mean, I don't, I don't have definitive proof on this stuff, but to me, like, it's a product. It's a thing that you can download and then you can license.
153
00:24:51.948 --> 00:25:03.088
You can attach like a, like a license thing on top of it, and why not make money, extra money on the side, as much as you're doing on, uh, Gemini? Yeah. Anyway. All right. [laughs] Sure.
154
00:25:03.148 --> 00:25:14.368
We'll take Hattie's- Hattie's tinfoil hat theory. Hattie's [laughs] Hat- Hattie tinfoil business plan. [laughs] By the way, get- $200 million... do not come to Denoised for business ideas. This is the worst.
155
00:25:14.528 --> 00:25:24.668
You are gonna go broke, my friend. [laughs] Yeah, just get a bunch of money and license these weights, and then when they come out with the new version, then you gotta give them more money. Yes, exactly.
156
00:25:24.768 --> 00:25:38.148
Oh, isn't that how the AI hype cycle is working right now anyway? It's all funny money. [laughs] Yeah. Yeah. All right, other new models. Muse- Mm-hmm... from Meta came out with a Muse Wonder, which is-
157
00:25:39.208 --> 00:25:44.808
Yeah, can you walk me through-... supposed to be pretty good... what Wonder is? No, I can't, 'cause all their names are Muse, Muse, Muse, Muse.
158
00:25:44.808 --> 00:25:53.948
[laughs] As far as I can tell, Muse Wonder is, like, and you know, according to the graph they have, is, like, their version of Fable and Astra, their, their state-of-the-art- Right... model.
159
00:25:54.208 --> 00:26:06.768
The, is apparently pretty good. If you ask me how would I access Muse Wonder, I don't know. [laughs] Like, I don't know what the equivalent of Claude Desktop is.
160
00:26:06.848 --> 00:26:17.508
I'm sure there's a Muse app, I just have not gone down that rabbit hole yet. Did you see the little, like, puck thing that they debuted with, with the little cute, like, a- animal on it? Yeah, everyone...
161
00:26:17.688 --> 00:26:25.928
Yeah, the, their little little blue thing. Yeah, so, like, you can speak to it, I guess. What does it do? That's what everyone... I mean, I don't know. OpenAI's got one, too. They got the dot thing. Yeah.
162
00:26:25.948 --> 00:26:35.488
Everyone's got little tchotchke AI things. I don't know. Do we need tchotchke AI things? I don't think we n- need more listening devices in our lives. Yeah, um, yeah. I will say
163
00:26:36.428 --> 00:26:45.448
the consumer stuff I have, I don't pay as much attention. Okay, fair enough. I mean, I also don't... Yeah, I don't know if it's really gonna take off. Says the guy with Ray Ban glasses.
164
00:26:45.448 --> 00:26:54.888
[laughs] Those are for filming [laughs] for documentary purposes. Gotcha, man. I got you good. [laughs] You're gonna crap on my business advice? We're gonna go there.
165
00:26:54.888 --> 00:27:09.188
[laughs] Well, okay, also probably part of the reason we didn't see so much about- Mm-hmm... Muse, um, is, uh, this post from Alexander Wang- Mm-hmm... their head of AI, who is also an amazing ship poster on Twitter. Um,
166
00:27:10.408 --> 00:27:18.108
they're basically like, "Yeah, you didn't see a bunch of, um, stuff from us because we built it for coding. We didn't train on 3D stuff. [laughs] We trained on coding, and it's a really good coding agent."
167
00:27:18.168 --> 00:27:26.148
Whole team feels dumb now, but it's true. We did not explicitly train our model on 3D demo games. Obviously, in retrospect, we probably should have done a bit of that, but no regrets.
168
00:27:26.348 --> 00:27:34.208
Please try Muse Wonder on your actual coding work. Oh, so it's really built to be, uh, a software writer. Coder for now. Yeah. Okay. Yeah. Fair enough.
169
00:27:34.628 --> 00:27:45.668
Yeah, he- here are some of his excellent posts of the Muse Labubu thing. Yeah. I know it's not Labubu. I'm sorry I'm infringing on Labubu's trademark, but it's- Yes. Yeah. [laughs] It's not...
170
00:27:45.728 --> 00:27:58.868
It's the, it's a nice-looking Labubu. I thought that Muse was really targeted at marketers. Is that not the case? I am sure under the hood for Facebook Ad Manager, it will be all Muse all the time.
171
00:27:59.088 --> 00:28:08.128
I mean, where is this all going? I'm sure it is to turbocharge making ads between intelligently- Mm-hmm... running them, writing the copy, making the images- Mm-hmm... making the video.
172
00:28:08.528 --> 00:28:16.848
Once they build all that, I'm just gonna run, reverse engineer anything on it, and then I'll have it. [laughs] Just kidding, guys. Just kidding.
173
00:28:16.888 --> 00:28:25.108
[laughs] See, but they got a moat, 'cause they got, they have network effect, so that's one of the other moats, where it's like they got the platform. You can't copy the platform.
174
00:28:25.128 --> 00:28:35.247
Literally everybody on the planet is on their platform, which is crazy. Yeah. Yeah. All right, other new one. We talked about this, but now it's actually out, Gemini 4- Oh, yeah... Argon, and they...
175
00:28:35.308 --> 00:28:46.548
No, they did train it on 3D. Dude, this is so good. So I ran across this video yesterday. Um, I w- uh, Joey, you probably already seen it by the time I sent it to you. I saw it pop up.
176
00:28:46.608 --> 00:28:49.308
I didn't watch the whole thing, so yeah, what, what's, what, what's, what impressed you?
177
00:28:49.448 --> 00:29:00.768
The Gemini team every year, uh, the when- whenever they launch a new Gemini, they have this, like, Japanese pagoda, you know, peaceful zen garden thing that they build, and then they see the quality over time improve.
178
00:29:01.188 --> 00:29:07.948
Mm. And that's, like, a visual indication. So this is their Culver Theater? [laughs] This is their Culver Theater. Yes. [laughs] Okay.
179
00:29:08.008 --> 00:29:29.388
So it's building it from a single prompt, but what's really neat here is, um, that single prompt then goes into an agent that is a blueprint agent, like a architectural blueprint agent, and it'll build a blueprint of this pagoda, and then it'll go into a Nana Banana agent, which is generating images off of that in many different angles for reference.
180
00:29:29.608 --> 00:29:43.808
And then it goes into an adversarial model, or rather, uh, an agent that is, uh, quality controlling the entire build process as multiple agents are building it in Blender. And it's doing this on its own.
181
00:29:43.848 --> 00:29:50.728
You didn't have to, like, give it the instructions to, like, "Hey, do a multi-agent, review your work, and blah, blah, blah." I don't know if you d- So that's the thing.
182
00:29:50.768 --> 00:30:04.468
He said it's off of a single prompt, but how he set up the system may, may have been manual. But regardless, the fact that you can get to this level of quality and fidelity from a single prompt is just nuts.
183
00:30:04.528 --> 00:30:13.268
It's bonkers. It's, this is, like- Yeah... hundreds if not thousands of hours of, um, like, you know, man hours of somebody doing this manually.
184
00:30:13.604 --> 00:30:19.284
By hand Yeah, I mean, that's what has been impressive 'cause, like, the amount of detail work in the single prompts- Right... and not having to- Yeah...
185
00:30:19.364 --> 00:30:27.924
tell it or refine it, and the interior, like, it looks like it built the interior as well. Oh, it's so crazy, and then, yeah, the shaders came in and they- Yeah... they rendered the whole thing in Blender.
186
00:30:28.384 --> 00:30:34.564
Blender seems to be, like, the proving ground, the LM arena of things- [laughs]... nowadays, uh, so yeah.
187
00:30:34.604 --> 00:30:44.884
Well, it's such a good thing too, uh, because it's just also so agent-friendly and easy to control 'cause it's so scriptable. Right. Right So it's just an easy- Yeah... 3D thing to build out and demo. Yeah. Very cool.
188
00:30:45.444 --> 00:30:56.844
You wanna get another one? Yeah, I've got- Florida, Florida take. Um- Yeah, please... this looks like, um, this looks like the Japan land in Epcot. [laughs] Oh. [laughs] This exact angle right here.
189
00:30:56.844 --> 00:31:05.364
[laughs] I've never been. I've never been. I'll take your word for it. I'm sure you've been many times. [laughs] But that just kinda brought the whole thing down a notch.
190
00:31:05.364 --> 00:31:16.564
[laughs] So s- 'Cause I've been to Japan, and I've seen these pagodas in person, and, uh, yeah, when you... It's like, yeah, that, that's in Vegas. It's not a big deal. Like, the Eiffel Tower's in Vegas.
191
00:31:16.624 --> 00:31:25.664
[laughs] I have been to Japan Epcot. [laughs] Check out the world. Uh, can we, can we go to- All right... the next Blender example?
192
00:31:26.344 --> 00:31:46.304
All right, so I guess we don't know what model they used, but this is also done through Blender connected to an LLM, and just incredible the amount of, like, knick-knacks and the real-life randomness that this desk has, and it's all done through an agent.
193
00:31:46.644 --> 00:31:56.184
Yeah, the de- detail here. I saw another one. I think a lot of the ones I've been seeing lately have, they've been saying Opus 5.5- Okay... has been doing it. Yeah.
194
00:31:56.284 --> 00:32:09.664
Um, I mean, also Argon, like, just came out, so I haven't quite seen a showdown yet between, like, Opus 5.5 and Argon. Yeah, here, uh, using Astra and Opus. Astra and Opus, okay. That makes sense.
195
00:32:10.164 --> 00:32:23.864
I'm curious how they handed it off between the two, or if they found one is better at something like research or blocking versus textures, uh, or if it's just purely
196
00:32:25.364 --> 00:32:34.384
maxed out the usage of one, time to switch to the other. [laughs] Uh, also, did you see that it was all from one single image? Like, is that the way, then, just to give it image references?
197
00:32:34.424 --> 00:32:46.464
'Cause that's how the Gemini guy did it, too. Like, Nana Banana agent generated images, and then those images went into the Blender builder. I have found, yeah, images definitely help, so if you give it more stuff,
198
00:32:47.404 --> 00:32:55.304
the better. But also, if you give it just some direction, it will also just start building reference images and making it.
199
00:32:55.364 --> 00:33:06.444
Or if it's also based off something r- real or historical, it'll spin up, like, research agents and then find images online. Yeah Or maps or diagrams or something and then start building off of online research.
200
00:33:06.464 --> 00:33:13.704
Do you think this is why, uh, Tim Sweeney, CEO of Epic, is all of a sudden pro-AI? Like, he's seeing stuff like this. He's like, "Wait a minute."
201
00:33:14.304 --> 00:33:26.964
I mean, I think if you're any type of game or software developer, if you're not thinking of AI first, like, you're just gonna be left in the dust. Yeah. So he's just basically, um- Where the wind is blowing Yeah.
202
00:33:27.004 --> 00:33:40.624
It's that. This is an Argon example, this cathedral build. From an image. From a single image, and then it built and lit. It's doing, like, render passes, yeah. Oh my God, it has, like, volumetric fog and everything.
203
00:33:41.164 --> 00:33:57.444
Uh, the other one I was looking for was this laundromat video. This is Opus 5.5. Which is, uh, one of my faves. I love Opus 5.5, uh, but I have never tried to do this with Blender in it. You know what?
204
00:33:57.704 --> 00:34:07.784
I'm inspired today. I'm gonna try something like this. Okay, cool. Yeah. Not a laundromat, maybe something more visually appealing. [laughs] Visually. [laughs] Were you just about to say something?
205
00:34:07.844 --> 00:34:14.284
No, just, uh, like, Epic had the lead, you know, like, five, six years ago, pre-generative AI boom.
206
00:34:14.924 --> 00:34:27.044
Like, they, they were, and they, I think they still are, they have AI talent in-house because, like, a lot of the game engine stuff needs to be agentic. Like, uh, if you look at NC... you know, NPCs and things like that.
207
00:34:27.584 --> 00:34:38.024
And I don't know why, but they totally missed the boat on the first wave of AI. Like, uh, you know, the, the way that sort of Blender has kinda caught that, Unreal didn't.
208
00:34:38.444 --> 00:34:47.924
Like, in just making- Just being a platform for where AI progress happens and testing happens, Unreal could've been that platform. I think
209
00:34:49.024 --> 00:34:54.704
the ones that were successful in that is, was accidental just because they happened to build a platform that was very scripting-friendly.
210
00:34:54.984 --> 00:35:04.544
Right So, like, Blender is very scripting-friendly, Resolve is scripting-friendly, and that just works really well with an agent. Yeah. As far as I know, like, Unreal is also script-friendly.
211
00:35:04.604 --> 00:35:16.864
Like, it's all C++ Python supported. But maybe, yeah. For building or for blueprints or both? Oh, you're right, yeah, for blueprints. Yeah. Like, I'd expect blueprints- Yeah... yeah, for sure. Yeah. You could...
212
00:35:16.904 --> 00:35:25.004
That one should be scriptable, but, like, the actual built, built model out of world- Right... I don't, I don't think so. Uh, yeah, you're right. Um, I mean, I guess recently they now have an MCP server.
213
00:35:25.044 --> 00:35:37.124
But yeah, it, it's just such a bummer- Yeah... to me, and I know so many people in the, the Unreal Engine world, you know, that could have really benefited from a lot of the AI progress.
214
00:35:37.184 --> 00:35:44.884
Like, it would've been just like rising tides floats all boats kind of a thing, you know, where modeling becomes- Mm-hmm... easier, shading becomes easier.
215
00:35:44.964 --> 00:35:57.564
You can do more, you can build games faster in Unreal, and that would in turn help the, uh, Fortnite ecosystem and all of the stuff that makes money for Epic just be more profitable. Like, it would've been a win-win.
216
00:35:57.704 --> 00:36:08.944
But nevertheless, they're getting into it now. Um, I sent you the tweet from Tim Sweeney the other day. "We have a huge pivot underway here at Epic, moving developers around to retool for maximum AI acceleration."
217
00:36:09.404 --> 00:36:12.944
The aim isn't to do the same stuff with fewer people. AI or SI.
218
00:36:12.944 --> 00:36:26.024
[laughs] I mean, I think you gotta get on board to accelerate your work, uh, not on board to use AI to, uh, replace everything, but use AI to accelerate the human-directed work.
219
00:36:26.184 --> 00:36:40.364
Ooh, look, he replied to somebody who was, uh, dropping some hate. "Your inability to read the room is really, is really astounding, Tim." And then he responded, "Much of the room wants to be in 2016 again, but it's not.
220
00:36:40.464 --> 00:36:48.944
Our job as the provider of tools and technology to much of the industry is to design for the present, and that is, and the future that will soon be.
221
00:36:49.144 --> 00:37:01.784
Personally, I'm optimistic about it and see it as empowering for creators." Dude, that's how we feel. Yeah. This guy knows. Tim Sweeney knows. [laughs] To... No, I met him once in person. He's, he's awesome. All right.
222
00:37:02.104 --> 00:37:13.484
Good place to wrap it up. For those of you who've been listening on Apple Podcasts, give us a like, and, uh, a five-star review would be really nice. You know, it helps with the algos. You know how it goes. It does, yep.
223
00:37:14.044 --> 00:37:20.064
Uh, links to everything we talked about at denoisepodcast.com. Thanks for watching. We'll catch you in the next episode.
