Creative Flux

Pierson Marks (00:01.014)
Nice. I think we're live, we're live here. We're not live actually. We don't do this one live. Like we said last week, we have to think of a intro. Good morning Vietnam or good morning world.

Bilal Tahir (00:01.095)
I'm my background.

Bilal Tahir (00:10.405)
yeah.

Bilal Tahir (00:18.421)
Good morning world or hello world.

Pierson Marks (00:23.692)
Yes, yes, totally. I think we need like an intro clip. It's something where it's welcome to the Creative Flux podcast or something that says our name at the beginning. think that would be key. I'm your host. Exactly. We were just talking about like we were having this conversation right before here and talking about design and talking about

Bilal Tahir (00:38.869)
Alright, I'm your host, Piers Marks, and this is the Lord There.

Pierson Marks (00:52.174)
all these random things and I was like, wait, we should just, we need to pop on this recording so we get all this stuff captured. So it was an interesting week. mean, it's short because we're recording this on Wednesday. But I mean, not to get into this that much.

The windsurf drama was just another level. That reminded me of OpenAI when Sam Altman was fired, kind of back and forth swaps. Every hour, something different changes. Windsurf was crazy.

Bilal Tahir (01:13.038)
yeah.

Bilal Tahir (01:21.887)
Right. Yeah. The whole market is, it's kind of insane. mean, you know, it's like a drama. I mean, honestly, in some ways, because, well, when, you know, what happened was, let's take a step back and talk about that for a second. There was a...

First off, is Winsurf is a company that has a coding ID like Cursor, where it's a fork of ES code, which is a very popular coding ID, and you can chat with your code base, you can ask an agent to edit it, and they launched, think, six months ago or something. They were more recent to enter it into the space. Cursor started off last year, and they've grown like crazy, billions of dollars of valuation right now. So obviously, other people want

at a piece and Winstrop was one of them and the company behind them was called Kodium. They were already a profitable business, they were doing more enterprise stuff and they got into this you know you know I don't know how much of a pivot it was for them I mean there was probably some overlap with the skillset but it was a different product for them but it did I mean it took off it did really well they were generating hundreds of millions of dollars I think from it or 82 million I think era was the last

metric I heard. Anyways, what happened was they were supposed to be acquired by OpenAI. That was like a setup. But next thing we know...

Pierson Marks (02:45.194)
For $3 billion, $3 billion.

Bilal Tahir (02:47.157)
$3 billion and next thing we know it the deal went you know down the toilet presumably because of Microsoft and what the reason was at least official reason the unofficial official reason is that Microsoft was gonna Microsoft actually owns VS code and so if they had done the deal they would have taken vincer's IP folded into VS code which would have made vincer coding ID kind of worthless because

they would have just like, you there was no way for them to differentiate themselves. So that was like a huge problem between those during the deal making that happened. There was a huge chasm on that. And so, Windsor decided to go to Google, you know, but the Google deal was weird because it wasn't an acquisition. What happened was like the CEO and some of the top researchers, they just left for Google. So Google licensed Windsor's IP.

Pierson Marks (03:26.424)
Interesting.

Bilal Tahir (03:44.935)
and took the top guys, but then they left all the employees inside the Windsor company. And so everyone was like, that's weird. And the employees were apparently screwed over because Windsor was basically worth a lot less. They were last valued at $1.5 billion, but they were maybe worth $100 million after this because the top talent left, the IP is licensed. So the Windsor employees were kind of screwed.

And then there was all this speculation. but they got a dividend payment of a hundred million dollars So they were made whole the people who are unvested they got accelerated investing all that there were all these videos Theodacty did a good job of putting a couple of videos out there recommend, you know looking at his videos, you you had a couple of deep dive videos on it, but But then to add another twist on it then last last what we heard cognition Which is the company behind Devon comes in and they actually rescue

Pierson Marks (04:27.681)
there.

Pierson Marks (04:39.789)
Mm-hmm.

Bilal Tahir (04:40.823)
the shelf of a company that was left and they acquired that company. So now as it stands, the CEO and the top guys, they went to Google. Google has the license for Winstrap IP. They paid $2.4 billion for it. And then Cognition acquires.

the windsurf company. And so they also, I guess, presumably get a share of the IP as well, as well as the enterprise business of windsurf and all the employees that are left over in windsurf, which was now an employee owned company. They are part of cognition now. And last I heard they were, they got a good deal out of it. So there was accelerated investing and you know, they were made whole. the employees did not get screwed as they were being done before, but it, the whole saga left apparently a

Pierson Marks (05:14.807)
Right.

Bilal Tahir (05:29.559)
lot of abilities and the amount of a lot of people you know especially people who were like founding engineers that like they're like what the fuck you know

We've talked a little bit about this, like the Westing thing. There's a lot of like weird things in the startup world of Silicon Valley. depending on how it's set up, people can get screwed over left and right. There's money to be made. But this was apparently a very, the way this was done kind of broke the social contract because if you're a founder, I mean, hey, I mean, you're the founder of this company. I feel like there's this implicit thing. You're like, I'm kind of sink or sw-

with the know my employees and my company and the feeling was like the CEO did not do that at this time he kind of just bailed and was like all right peace guys bye and so yeah I don't know I mean it's probably more complicated I'm not gonna presume what weren't the CEO like you know what was going on in his head and if he knew this was gonna happen that his employees were gonna get taken care of or not so not gonna speak about that but it has raised a lot of philosophical questions about you know

Pierson Marks (06:17.473)
Right.

No, totally.

Bilal Tahir (06:36.499)
what know, startups should do in this thing. But yeah, that was okay. That was my overlay. What about you like on your side, know?

Pierson Marks (06:37.079)
Don't know. Right.

I mean, no, you covered it really well. mean, there was just a lot of churn. You know who's making a lot of money here? Lawyers. it's just, you know, I know we try to talk about generative media on this, but like this is super interesting just because the space in general. I mean, you want to...

Bilal Tahir (06:49.47)
yeah.

Pierson Marks (06:58.669)
You want to stimulate innovation. You want to make sure that people feel empowered and optimistic and like, hey, I'm going to do something when we work really hard at it. And in the future, it's like, because because I'm working really hard, really smart, I'm going to get I'm going to get my fair share at the end of the day, you know, whether that's, you know, a salary equity payout, like normally that's equity. It's like, hey, I own some percentage of this company. And when the company gets bought or goes public, I'm going to receive what I what I

think I deserve and the way that these acqui hire blitz hiring type of acquisitions are happening right now it's there's a failure in the system the the &A system today in the United States because essentially any acquisition that exceeds a hundred and twenty six million dollars automatically triggers like a FTC probe which is anti-trust and it's like all these like sort of like you

Bilal Tahir (07:57.151)
Yeah, this is like one of, yeah, probably Lena Khan's, FTC's biggest like, you know, legacy that unfortunately the &A landscape has been broken because of these, you know, huge regulations, you know, that she left on. Yeah.

Pierson Marks (07:58.603)
regulatory over

Pierson Marks (08:10.231)
Yeah.

And I'm not saying that like, it wrong or right? Like think we definitely need to make sure that we encourage fair competition monopolies aren't used to like stifle innovations. But there's also like, there's one, there's a dollar value that triggers these sort of these probes. And there's also like the time, like in a world where AI six months, 12 months from now is completely different and technology is completely different than what it is today.

Even if this sort of like investigate in investigation kind of like takes three six months It's too slow like these deal like the landscape changes things happen and it's just forcing these like creative solutions to be be created and and the and the bummer is to it's like the reason why you're not seeing a legal firm, I don't know who represented windsurf or whatever on this was like

Bilal Tahir (08:57.813)
Yeah

Pierson Marks (09:09.669)
Technically what was happening is illegal if they could if you could prove that it was done purely to circumvent antitrust and that's why it's like you have like they have to make sure that the company gets this license payout it's not gonna die like essentially if they all left and the companies are shell and that just completely explodes like next week there'll just be like investigators and federal agents like over this and

Bilal Tahir (09:16.693)
Yeah.

Bilal Tahir (09:30.517)
Yeah, super interesting. this is not even the one, mean, meta, same deal with the scale audio. They didn't acquire, they got the 49 % partly because of this. It was the FTC and they just want to, the law of unintended conversation or consequences, right? I mean, this is what happens when you try to regulate with like a carte blanche kind of rule, know, firms will skirt away, they will bend the rules, they will like find a way to do it.

Pierson Marks (09:35.883)
scale.

Pierson Marks (09:40.481)
Right. Right.

Pierson Marks (09:54.23)
Pierson Marks (09:58.891)
Right. And it's just like to wrap this up too, it's just like, it's causing a re-event, like the landscape that was the venture capital and startup landscape that existed in the 2000s and prior 2000s, 2008 earlier, even like before 2020, like that is completely changed. Like, you know, the old mental model, what you need to do is like, hey, you have an idea, you raise a million bucks.

Now, inflation adjusted would be more than that, but you raise some money, you give away 20 % of your company at your pre-seed or your seed stage. Pre-seed wasn't even a thing back then. It was just seed. You had seed money. You raise a million bucks. You hire like five, 10 people on your team. You pay them all $100,000. You build out some product. You go to market. And just the whole venture landscape has completely changed because one, we can do so much more, so much less, which is like super important. And then there's like the other people though.

Other side's like, okay, well, I mean, you still need the money. A job of a VC is like, their job is to sell money to founders. Like that is their one job. And so they're like on the other side of the say, hey, like, no, you can't seed strap your company because then somebody else is gonna do the same thing that you're doing. They're gonna raise 10 million, a hundred million dollars and just blow you out of the water. So it was like, is that opinion. There's the other opinion that's like, maybe that doesn't matter as much. you still, you raise a little bit of money on a safe and you don't.

Bilal Tahir (11:07.261)
Yeah.

Pierson Marks (11:28.087)
actually those don't ever convert and like you use that million bucks on safe notes to grow a big business and you're very profitable. You hire, you seed strap, you sell your company for 50 million dollars, 100 million dollars, all the, you don't have to have a billion dollar exit. And so it's like, it's really just shaking up.

this whole landscape of like, there is no right way to do this. And I think there's never been a right way to build a startup. But I think in the last like 10 years, it's kind of been like, hey, you know, apply to your accelerator, get to YC, raise your million bucks pre-seed, raise your Series A, eventually sell to something. And it's like, it just skewed all of the benchmarks up.

So it's like, you can't be a successful company if you sell for 50 million. if you take venture capital, it's not an outcome that you can do. But maybe that becomes more of an outcome with smaller teams, like a 10-person company selling for 50 million bucks. It's like, that's sick. And that should be able to happen, but not if you raise in the current way VCs running.

Bilal Tahir (12:14.783)
Hmm.

Bilal Tahir (12:20.277)
Hmm.

Bilal Tahir (12:27.421)
Right, right.

Yeah. mean, it comes to, you know, like this is why I think one thing I like about Jalpa, like we're both aligned on, we're not just raising for the sake of raising. Cause I've been, it's like all about the next round. It's not about the product is doing well or we can have, it's like, my God, I want to just get valued at billions of dollars, you know, even though there's no revenue or blah, blah, blah. And that's considered a success on its own. Right. And it's really not what you're doing is you're making a promise and you better deliver on that promise. Otherwise you're screwed, you know, and.

Pierson Marks (13:00.257)
Right. Yeah, so...

Bilal Tahir (13:00.509)
So it's interesting. But yeah, I I hope there's a resolution. I think it sounds like in this particular case, is a.

didn't talk about it much, but Cognition, company behind Devin, which is an agent, they bought Windsurf. And I actually think that there could be good synergies there because Windsurf basically was the first to really go big on agents. know, the cursor has an agent too, but like they kind of, with Windsurf actually, you know, I use Windsurf as well as part of my idea, just cause I like to play over different ideas and they have a, the composite they have is pretty good. And I actually think they're now going to double down with the whole PR. you're not, not only you're going to be

be able to make edits on your code base, but it's gonna be able to create a branch and make a PR and then merge. so I actually think it'd be very interesting to see this merger of what we do with Cloud Code or Devin, which is creating an autonomous agent creating code and then having control and seeing the code yourself. And I feel like there's probably a combination there where we can do both and it's not either or. So it'll be interesting to see how the AI agentic coding flow changes with this. And I think Winsorfer's

Pierson Marks (14:03.244)
Mm-hmm.

Bilal Tahir (14:10.685)
you know might actually ironically be set up well for it now because of this so we'll so yeah

Pierson Marks (14:14.603)
Right. No, it's super interesting. Yeah. yeah. So covered windsurf, a lot of drama.

Bilal Tahir (14:18.911)
So yeah, lots of drama. Tech has lots of drama. It's like spicy. It's like honestly, who needs reality TV when you have this, right? Speaking of this, another very fun reality TV, tech drama that's going on is just the concept of hiring. know, it's just, we talked about, I think the $100 million bonuses on how meta hiring, well, apparently that happened more. Two more researchers left for OpenAI this week. And honestly, it's kind of crazy now. I've said this before.

Pierson Marks (14:27.117)
Totally.

Pierson Marks (14:35.394)
Haha.

Right.

Right.

Bilal Tahir (14:48.807)
I don't think OpenAI, I think the probability of OpenAI going out of business is low, but I think it's higher than most people think. I've said this before, I actually think the way their leadership has operated and stuff, feel like they're, the way they've scaled, like especially it's like, it's like a house of cards there. And this is another thing which I'm like, okay, a lot of their good talent is now left. You know, are they gonna still be a leading lab? sure. Every day I think of them as more of a consumer product company rather than an AGI lab.

Pierson Marks (15:12.599)
Mm-hmm.

Bilal Tahir (15:18.613)
I know you're on the other side of this, but I generally, I don't think of OpenAI as a leader as much as I used to, at least last year. I still do, but compared to one or two years ago when they were it, not anywhere at all. And if anything, I'm more interested, excited about Google and the Chinese labs, you know, and now meta even, you know, cause they apparently they're scrapping Behemoth, which was another model they were going to release, because Lama 4 basically was a failure.

Pierson Marks (15:21.74)
No, no, no.

Pierson Marks (15:44.973)
Lama for a boogie myth, right?

Bilal Tahir (15:46.577)
Yeah, but llama4 was a failure and it sounds like, you know, they are basically swallowing that sunk cost, which is hard. mean, you know, credit to them. And it sounds like Alexander Wang came in and was like, all right, let's just start over. And apparently their next model won't be open source. So that's sad for us, but I wouldn't be surprised. Llama5 is a closed source model and, but hopefully a good model, you know, and something we can use with VAPI. So it'll be interesting to see.

Pierson Marks (16:14.145)
Yeah, it's really interesting, especially like with the terms of service on like all those, all the Meta-owned platforms. I mean, this kind of gets into general media too. It's like when you're posting your content on Instagram or on these places, like Meta's now training on that stuff. Like that's theirs. So, I mean, it's...

Bilal Tahir (16:15.733)
We'll be right back.

You

Pierson Marks (16:36.973)
We talked about this last week, the spatial awareness and the data, the data generating data is super important. Gathering that data, making sure it's clean and good. And what are data generators in the world? It's cameras. like, for example, Tesla, you know, they have all those cameras are creating a world model of the entire world. So to really understand physics and how the world works. Meta, you know, you have Facebook. have input text, user submitted text. have user submitted.

images and videos. And in those videos, you can derive a lot of information to like, you know, hands don't go through walls and like all that stuff. just like, you know, understanding that comes from, you know, the shorts on Instagram reels on just images. I strongly strongly believe that I think that meta is going to be the dominant player in advertising. Like it already is extremely strong, but the way that advertising works today is

Bilal Tahir (17:14.537)
No.

Bilal Tahir (17:33.203)
Right, right.

Pierson Marks (17:36.847)
like, hey, I go and hire a influencer, go, you know, generate some static image, and then I go into Facebook ad manager, and I say, hey, like, these are like the people I want to target these audiences, and I have this like one video from this one like blonde influencer girl holding up the product. And that's like, hey, go find whoever wants this stuff. You know, that's great, but it's one to many. And so you're finding, okay, you have this one blonde girl holding this one product, and it's one way, and you go and find all your people that would

Bilal Tahir (17:56.479)
Yeah.

Bilal Tahir (18:03.861)
Mmm.

Pierson Marks (18:06.767)
resonate with this. I strongly think that what Facebook's going to do in their ad platform is that rather than, you know, bringing your own image or bringing your own video as an ad, generate that ad or generate a thousand versions of those ads. Exactly. No, exactly.

Bilal Tahir (18:18.797)
based on the person. That'd be hilarious. So you're going to get the blonde, I'm going to get the brunette like, or redhead, like based on our preference. That's mildly dystopian, but I mean, totally makes sense. You know, we basically.

Pierson Marks (18:30.829)
And it's just like, because the key thing in advertising, especially in UGC and like creator-based advertising is like, you want to test every single version of the hook. You want to test every single version of who the person is, how they're holding the thing, do they wear their hat?

Like normally, do they have a hat? What are they holding in their hands? You want to test all these variations because some of them are just going to work and you have to be able to test. But if Facebook can do that for you, like, hey, just like describe kind of overall what you kind of want. You want like an ad for this thing. They're like, hey, we're to generate a male. We're to join our female. We're going to put her inside outside. We're going to be wearing a hat. He's going to be, you know, white. He's going to be

Like an athlete, like all those permutations and every single person can see a slightly different ad that's going to target them better. And I think that's what they're going to do. Like, why wouldn't they? I mean, they have a great ad platform. They have this data. They're going to be generating avatars and realistic life like people.

Bilal Tahir (19:24.213)
you

Bilal Tahir (19:34.701)
yeah, I mean, it's so fascinating. And I actually think you're gonna have your own AI influencers, AI friends as well. And I mean, there have been some forays into AI, like social networks and stuff, but I feel like you...

Pierson Marks (19:42.125)
Right.

Bilal Tahir (19:49.653)
The only true, the ones that will be successful at first won't be AI native, they'll be hybrids, right? Cause you have your friends, but then maybe you start having AI friends. You we already saw them in WhatsApp, by the way. I think you can talk to a Facebook avatar and stuff like celebrity Kim Kardashian, whatever. think they did that. That was the first iteration of this, but you're going to have your own personalized avatars. know, I mean, Twitter by X just launched Rock 4 with the waifu and stuff. And honestly,

Pierson Marks (20:13.165)
Right. Right.

Bilal Tahir (20:15.669)
That's gonna be so powerful. I'm like so bullish on AI companionship. I actually think people under index it because it's something that's so non-intuitive and also goes against a lot of the so-called sacredness of the human experience, is like, you know, we're social creatures. There's nothing can replace the human-to-human connection. And I don't know, I think that's an assumption. So we'll see, you know, how these avatars and these relationships change our whole social network basically, you know? And I actually think, I think the end game

is basically everyone gets their own social network of their AI native thing. And it's like this little hub and spoke model where you have your own island of AI influencer friends and stuff in your circle. But then obviously you reach out, maybe sometimes like, I miss Pearson. Let me go talk to Pearson. It's like, all right. You have your own maybe you do meet sometimes.

Pierson Marks (21:03.405)
Interesting. Right.

No, No, it's interesting. I don't disagree. I think it's weird. I don't necessarily want that feature to... And that's a key thing, I think. When you see arguments... And you see this argument all the time on X or I saw something the other day...

Bilal Tahir (21:18.579)
Right, see I mean I think that's the reaction to everyone. yeah.

Pierson Marks (21:27.409)
it was essentially Grok safety testing. I mean, that was a whole other ordeal where, you know, Grok four came out. They did the whole like mecca Hitler stuff, like all the antisemitic stuff. And there was no like release of safety testing or like procedure essentially that was expected of a big lab. And a lot of people came out from like Anthropic and Google and open AI, you know, intelligent researchers that came out speaking against like the researchers and Grok for safety practices. And tying this back into what we were just saying is like,

You know, people get mad when, I mean, in a highly competitive environment, you know, there are things that you should do and there's a way you believe the world should work and you believe that in good faith, you know, people should act some way. And you hope the world exists in that worldview. Like you hope that every single major research lab does this because it's either the right thing to do in your perspective or, you know, more generally accepted the right thing.

do you hope that people you know maintain human relationships and like AI doesn't become like a core source of companionship because you're like that's just weird that's like the way the world should work and you kind of have a worldview based on your your priors and

A lot of arguments come from that. I thought it was very interesting just to see a lot of researchers from all these different companies come out and say, what Grok did is despicable and everything. was like, sure, I can agree with that. I do think that there should be safety testing. But also, it's like you have to recognize from a place of capitalism, competitiveness, the way that you want and wish the world to work doesn't actually exist.

And so just because they should do something like that, unless they're forced to, why would somebody take on additional burden to go through these things when they're already behind and they want to catch up? why should... And it's just like this, I always find it really fascinating when people got upset with AI coming over. Like who wants an AI generated podcast? You how many comments we got on our Facebook ads?

Bilal Tahir (23:31.433)
Yeah.

Pierson Marks (23:46.061)
like AI slop which like stuff is like okay cool like you know that's your opinion and you're you could have your opinion the way the world has existed you know this doesn't exist so it's it's uncomfortable it's uncomfortable or you hope it wasn't true and because you you wish something to not be true you get mad at the thing and it's like I don't know it's

Bilal Tahir (23:46.248)
Yeah.

Bilal Tahir (24:06.497)
Right. Yeah, I mean, it's interesting. I think a lot of the whole term AI slop, I think exists because it's slop. I mean, it's been today to slop, but it's way less slop than it was last year. You know, if you look at from 3.5 to 4 to 4, I mean, like all these jumps, right? And so I don't know why people think it will always be slop. It won't be actually think sooner rather than later. And then the question becomes, you know, like what happened in the world where AI content is better than real content?

Pierson Marks (24:14.893)
Right.

Bilal Tahir (24:34.805)
You know, I mean, and that's primarily what we consume. Maybe at the same time, I mean, I can see us having a uniqueness. Maybe the content that's shit and then becomes.

Pierson Marks (24:39.467)
Right.

Bilal Tahir (24:45.095)
actually valuable because you're like, that's real. know, yes, too good. No humans have flaws. The flaw actually becomes a feature. But it's just interesting to see. I just don't think we truly can wrap our heads around how this is going to change us fundamentally and the things that us and our kids are going to be valuing. And so it's happened before. It'll happen again, you know, in every generation. And the key is just the assumptions we make. You always have to understand everything I say, you say is based on

Pierson Marks (24:48.225)
Alright. Alright. Yeah.

Pierson Marks (25:01.186)
Right.

Pierson Marks (25:09.645)
Totally.

Bilal Tahir (25:14.999)
core underlying belief or assumption we have as as know as human beings and the real radical way to kind of look at it is like what if you throw all that away out the window and just first principles what does it look like and I mean that's where I kind of arrive at this like you know fundamentally I think I'm like you know it's just I don't see how that doesn't happen

It's like McDonald's, you know? When you invented McDonald's, it like, sure to have that, you're to get fat. Well, yeah, but it's tastier and cheaper than any other option. Do you really think 70 % of the population would have this? At least twice a week over them. So same as this. I mean, you might not like it or something, but it is fast food. It is, right?

Pierson Marks (25:36.621)
All right.

Pierson Marks (25:48.619)
Right?

Pierson Marks (25:54.869)
Right. Yeah. And you know, even even prefixing the word AI to content, AI media, AI content, know, that essentially AI today.

It's synonymous with LLMs or like video gen, image gen. Like AI is like such a broad term. And right now artificial intelligence is purely, okay, have chat GPT, you these language models, you have generative photos, generative video. And it's kind of like, it's a blur. Like AI is like such a, you know, general term. If you ask what AI was 20 years ago, essentially AI has always just been like the cutting edge, like, you know, computer vision, machine learning.

It's always kind of, hey, existing classical computing and problems is like here, AI is kind of like the frontier. And today the frontier, you know, AI is like AI generated video, AI generated media. You would never have called Instagram filters when they first released like AI, but like it did some computer processing to them, you know, like maybe it intelligently understood the image and like it looked at the pixels like, hey, the dark pixels make them darker, light pixels make them lighter. You know, it just a computer program.

that was applied to a photo to make the photo look differently. When you go to Photoshop and you go to Illustrator, it's not AI. When you say, hey, draw the circle, you literally just go from corner to corner and you drag, circle appears. You didn't draw that circle. It's just like there's a computer program that you did some action. You either clicked add or you drew a line and it created a circle or created a star or created text. It's not AI. It's just like essentially you did something. The computer had some software.

whether it's a language model or classical like, you know, CPU based software and it has some output on the screen. And so the means at which like the underlying software is working, whether it was AI, LLM, it's like, it doesn't matter. You'd never watch a movie, like a Marvel movie and say, that was a AI movie. It's like, no, just a movie. It uses some CGI and stuff. You don't know what CGI and what's not. It's just like, just a movie. Who cares about the means?

Bilal Tahir (27:49.397)
Hmm.

Bilal Tahir (28:04.231)
Right, right, Yeah.

Right. Right, right, right. Yeah. I mean, it's interesting. It's like same with any due to the same with blockchain stuff. mean, people say like, you you know, it's succeed when it's just abstracted. We don't think about, you know, like, TCP. my God, we're using TCP. No, it's just a protocol under the inner, you know, that makes us connect on the internet, you know? So why not? You know, so maybe AI is like that too. It'll just be, you know, it's just part of.

Pierson Marks (28:10.895)
And nobody cares.

Pierson Marks (28:28.833)
Right. Totally. Yeah.

No, for sure. No, it's super interesting. Yeah, that's...

Bilal Tahir (28:36.125)
our world. it's interesting. We're almost at 30 minutes. We haven't really talked about generated media. Maybe we should talk about it.

Pierson Marks (28:43.883)
Okay, the few things, so one thing I did want to bring up, which I thought was really cool because I spent some time this weekend playing around with it, the Flora software, it's like we talked about it last week. It's kind of like a comfy UI based interface. And so it's like you have a big canvas, you have nodes and those nodes can be one of three types. It could be a text node, it could be an image node or a video node. And pretty much you have a text node. It's like generate, it's like.

It's like a prompt essentially.

Bilal Tahir (29:14.035)
And you can select from a bunch of the top models, I'm guessing.

Pierson Marks (29:17.495)
Exactly. And you can drag that prompt and you can put it into an image and generate an image from that prompt or you can generate a movie or video from that prompt. It was pretty cool.

It was very intuitive, more intuitive than the company UI, less powerful. But obviously that's that trade off. But I really enjoyed it. I spent some time this past weekend, you know, trying to recreate this Greco futurist futurism that I mentioned last week about like what would happen if the Greeks had access to modern technology like LEDs and things. And I pretty much just describing this image and then I was going from prompt to image image to video. So I was trying to make

house tours, like if you had this really cool Greco futuristic hotel and you generated the first, the hallway, and then you generated the inside of the rooms and videos, and it was pretty good. They don't have the image to video sort of, they don't have image to video yet, and that was like, I really needed that, so I can go from this core foundation image to multiple, but it was fun.

Bilal Tahir (30:00.661)
That's awesome.

Bilal Tahir (30:16.511)
Right.

Yeah, So do you have to just download those videos and stitch them up together yourself, or do they let you, like, once you generate the image or the video, you can, like, connect them, combine them into one export or something?

Pierson Marks (30:32.365)
you would have to download them right now. Yeah. So I didn't really get that far, but I was just playing around. It was fun. was just like, oh, I spent my Saturday morning just kind of messing around and trying to create some images. And so.

Bilal Tahir (30:34.248)
Okay.

Bilal Tahir (30:38.634)
Yeah.

No, that's super cool.

The pipelines are super powerful. think I feel like there's so much alpha in connecting all these model schemes together. was playing around with this video to video flow, which I thought was very interesting, similar to what you're doing where I wanted to just take an old clip and style transfer it on to anime. So you can take an, like I took a James Bond clip on James Bond, know, when Sean Connery says, I wanted to have him in the studio, give the anime style. And you think, okay, you know, hopefully there's a model that does that, but.

Pierson Marks (31:08.642)
Bye.

Bilal Tahir (31:11.645)
It's like not that simple because there's so many steps. First off, the way you do it is you actually need to give them an image, the starting screenshot of the scene and in the new style. So what I had to do was first I had a scene detection pipeline that took the clip, broke it into clips where the scene changed. Because once the scene changes, you know, like the whole scene changes, you need a new image. Then I took the first frame of the video.

Pierson Marks (31:25.069)
Right.

Pierson Marks (31:35.543)
Mm-hmm.

Bilal Tahir (31:38.897)
restyled it using image editing into the new anime style. Then I transformed that particular clip into the new style, did it for all of them, then I combined them together. So it's like four or five steps, know, and it's like, it's like, you know, it's like sounds, you know, similar to yours, you're gonna connect the nodes together and do it. And so there's a lot of alpha I think like in connecting these models together, you know, in doing this. And even now it's like pretty hacky and the results are meh, but you know, it's getting there.

Pierson Marks (32:00.973)
Right.

Pierson Marks (32:06.519)
Wait, so on your original clip, I'm curious, you had like, was it one scene or was it like, you had like multiple, like did you chop up, like you have this one like 30 second video and then you chopped up like when the camera angle changed or stuff?

Bilal Tahir (32:18.707)
Yeah, and so there's a cool API called SIV has this, which is a scene detector. And so you basically pass it and it'll give you a JSON object of, from zero to seven seconds, the first scene, and then it changes. Now it's not perfect. Like if you do like a war scene or something, it's probably going to break down completely, but it's ideal for more still images where like us, maybe I'm talking and then you're talking, maybe in Google Meet, the camera changes. It'll perfectly tell you when Bilal was talking and when Pearson was talking. Although I guess you can use proxies that for when you're talking too, but that's

Pierson Marks (32:28.898)
Away.

Pierson Marks (32:45.142)
Interesting.

Bilal Tahir (32:48.671)
great way, right? So the platform users, they show us together and you can do claimation on top of it. if you want to do multiple, you kind have to, I guess, do something like that, a pipeline thing. And we've talked about, sorry, Yeah, SIEV, have this API, which I think is pretty cool.

Pierson Marks (33:02.561)
Hmm. This is on Civ. So this was on Civ. What's the platform? Okay.

Bilal Tahir (33:14.341)
And so yeah, I mean, that's one way to do it. you can do, my goal is, what I really want to do, I think it would be cool is, I have all these 80s, 90s cartoons that I actually want to make live actions of. I think that'd be pretty sick. And I actually think there'll be a renaissance on taking old content and making them, restyling them in, know, or maybe a black and white movie in color or, you know, anime style, et cetera. Then you add voice stumbling on top of it, or maybe something like, you can reimagine like Cleopatra, but-

Pierson Marks (33:27.786)
Anyway.

Pierson Marks (33:37.751)
Right.

Bilal Tahir (33:44.367)
like Gen Z or whatever, I don't know. I I think there's a whole remixing, repurposing basket of content that's just waiting for us. Once these tools are so good and cheap, you can do it.

Pierson Marks (33:59.797)
It's so interesting. mean, like, you're talking about, you always... It's weird to be in a space where, like, you know you're so early. Like, we know that where we are right now...

Bilal Tahir (34:07.86)
Yeah.

Pierson Marks (34:09.469)
is going to become way easier. we're we're recognized. It's hard to recognize a trend, I think, so easy, like so early and where we are with generative media today and video and content. It's like the world is moving this way. And I'm sorry, the train's left. It's not stopping. And you could debate on the timeline, but, you know, it's moving in this way. It's like we're doing like

Bilal Tahir (34:28.809)
Yes.

Pierson Marks (34:37.793)
Like it's cool. You could be one of the first movers in this space. Like yeah, there's people already making a video on TikTok and stuff. Cool. But it's like.

Bilal Tahir (34:40.169)
Yes.

yeah, there's so much out

It's so cool. especially with this whole content, there's copyright free content out there you can leverage. can do stuff with it. Actually reminds me of one of the first cases I saw was I was taking Fast AI, which is a course by Jeremy Howard. I wanted to get into deep learning around 2016, 2017. was very early. Thankfully, we were just getting into pie torture. The Keras era was there, which was a nightmare at that point. Now Keras is better. But I was taking this course, which I recommend if you guys are interested. It's a great, very accessible.

Pierson Marks (35:01.346)
Right.

Bilal Tahir (35:14.359)
course getting into it and I was learning about deep learning, convolution neural networks and how to train a model and one of the people in my batch quote-unquote like was this guy who was fascinated about image transformation so in the class they showed us an example of taking an image you could take an image that was

like black and white or whatever and color it you know it kind of colors it and this guy had this 15 sec well i can do this way better and so he literally took that

Pierson Marks (35:39.756)
Right.

Bilal Tahir (35:45.555)
the classes like model, trained it on ImageNet, which is this very OG data set of images back then. It was like the biggest data set. And he basically trained a way to go from a black and white image to color, which kind of worked. It was very perplish and stuff, but it worked. It was better than a of the other solution. And then he and his wife started a company and I think they're still around. It's called DeOldify. He opensourced the repo, but he made the model way better. And he actually partnered with ancestry.com. So if you go to ancestry.com, you can put in an old image and get a colored image. They use DeOldify.

under the hood. So it's like a proper company and he just started it as a side project. It like blew up on get up in front of me. So it's just interesting. You never know, you what's gonna work and don't assume other people are far ahead or they know what they're doing. There's a lot of, find just starting over from scratch. His thing was that he started over with ImageNet and and convolutional neural nets and had a couple of tricks on training, you know, that he just picked up like using this bootcamp like course, you know, in front of me. yeah, very cool.

Pierson Marks (36:22.123)
Huh.

Pierson Marks (36:42.647)
Right. It's super cool when I...

Right where we are, it's like I've always believed that in order to build a like a startup or like any sort of like cool thing is you can either you can go one of two dimensions of like breadth or depth and it's like right now you can kind of go in both ways where you can make really cool things by connecting pieces of all these like disconnected just jointed systems and you like go broad like hey you don't have to go that deep but like you know you have like the scene splitter that you just said and then you have you connect it to over here and you do like you can go really

broad connect all these like cool pieces and unique ways and have like an amazing business that or like product that you know, it's just like really cool and it's hard and it's like that's hard in its own way. And then the other way it's like going really deep. And so it's like, you know, finding a really narrow niche and then just like pushing the boundaries of research and just going more and more deep and like creating something new like open AIs like reasoning models or something. So you can go broader deep or you can go like some combination. You go deep here and you're broad and it's like I think that's what's so cool is because

The level of like intelligence required to go deep is typically I would say higher because you need domain expertise and so to go really deep It's like harder but now I'm like a paradigm shift There's so much alpha and leverage to go broad and connect things while going deep in your specific areas like, know You can go deep in like understanding this stuff. It's like very cool To do and it's feasible. It's now it's not like I don't know enough and I can't do it I mean, I can't research you don't

Bilal Tahir (38:04.277)
Right.

Bilal Tahir (38:15.327)
Yeah. Yeah, no, for sure. There's so much alpha. yeah, it's like that T-shape I do, the right mix of breadth and depth can like totally make you stand out and you can be an expert in that subfield of a subfield and people will pay you for that for sure.

Pierson Marks (38:16.387)
to.

Pierson Marks (38:33.549)
Right, right. Maybe it makes sense to like on some of these pipelines, it'd be cool. I know we were talking about free tools for JellyPod. I mean, is there something that could be a free tool that we can just build as a podcast, you have a video or a stitcher. I don't know. Maybe there's something there that we just make a workflow for. I don't know.

Bilal Tahir (38:50.677)
Well, we do that. We've talked about it. We have some tools like we can let you go from PowerPoint to a podcast or we have what is it? We have an AI podcast generator, which is a very popular tool. People come and quickly generate podcast ideas and covers and even the whole podcast, like a sample to see how it sounds. You know, it's great. And there's probably other stuff, you know, I know you're a fan of repo to get a repo to a podcast, which I can think could be very interesting.

I think a song to a podcast could be interesting too, but all these formats, it's fun and interesting. So yeah, for sure.

Pierson Marks (39:21.036)
He

Pierson Marks (39:25.759)
Right. totally. Yeah. I mean, and it's so funny. We always end up like this where it's like 40 minutes in we have the agenda that we didn't even touch like most of this stuff. But I mean, was there anything?

Bilal Tahir (39:39.456)
yeah, Well, I feel like you were active runway. I know you were excited about that. Like, what do you, what do you think?

Pierson Marks (39:46.113)
Yeah, just something I saw. don't know if you dug into it at all either, but it was kind of like video to video. It seemed like I can take a video of myself and then really just have character consistency, scene consistency, where, hey, there's this guy I just followed yesterday, and I forget his name, so I'm sorry I can't shout him out right now, but he was playing around with RunwayML's Act 2 model that just came out, and he was just...

Bilal Tahir (39:57.299)
Yes.

Pierson Marks (40:13.965)
recording himself like with a camera, like doing all these things and then like taking that recording and putting it into runway and just transferring over like scenes. So it was like he became a wizard. He had like a tennis. So he just was sitting on his chair, had a tennis ball. He was like throwing up the tennis ball and then he put it into runway ML and he became a wizard and it was like throwing like fire. And it seemed like really good. So.

Bilal Tahir (40:41.513)
Yeah, yeah, no, it's similar to the, so I use Ray to modify for my thing, but Runway is definitely, like the caveat is they don't let you have an API and they're very expensive, but these video style transfer solution are super cool. And I am so excited about this, like video to video. The only limit, I think the biggest limitation is these scenes have to be very, like, if you notice the examples, they already like control, like, you know, got to, like the ball has to be, it's not like, you know.

you were just doing some, if he was just walking around, if there was a guy behind him, that would have ruined the scene. It has to be almost like a green screen. You don't need a green screen, but you need a very still environment with deliberate actions. And the more control the action is, the better the output is. cause you know, the model is very brittle. What I'm looking forward to is like, I can just take anything. I can take a Fast and Furious like chase and just...

Pierson Marks (41:17.943)
Mm-hmm.

Bilal Tahir (41:35.489)
making it anime and it just handles all the motorcycles and the cars and the multi shots and the expressions. It just captures that. And I think we're going to get there. And there will be some, I think it will be obvious to us in hindsight, but there will be some sort of threshold with these technologies. There's some threshold where something comes out that's easy enough and high quality enough that everyone goes, whoa, what the hell was that?

Pierson Marks (41:43.181)
All

Bilal Tahir (41:59.989)
Right? Like maybe with images, think stable diffusion or so at some point the images just got like they were cute and then they just got really good. And you're like, Oh, this is going to put you out of a job. mean, if you're a photographer or something, I mean, this is basically good enough. Right. And then it explodes and suddenly everyone on Upwork is offering you guys. I think you're going to see that with video at some point. It'll just be like, yeah, one of James Bond anime. Why not? Right. Or whatever. So it'll just happen. So excited for that.

Pierson Marks (42:00.663)
Mm-hmm.

Pierson Marks (42:26.605)
Right. No, it's super exciting. A lot of changes. I mean, yeah, it's wild. if you're listening, I mean, for the people listening to it's like, hey, we're talking about this stuff. Hopefully we give you some inspiration. We validate your feelings. Maybe we say some uncomfortable truths about AI companionship and all these things. But yeah.

You know, we had a big agenda of things. We covered some of them. Windsurf, covered Act 2, Flora, did really touch on Kimi K too and some of the things last week that we skipped over. But I mean, I think it might be.

You know we're at 40 minutes or so. I think it probably is good to cut this one and you know just continue the conversation next week or I mean two weeks or so. Yeah.

Bilal Tahir (43:21.087)
Yeah, Yeah, for sure. Two weeks, yeah. I'll be gone next week, but Pearson will be here. Maybe you can do something different or whatever later on. I'll let you think about it.

Pierson Marks (43:35.597)
Yeah, you're giving me the challenge here and I'm like, I've been thinking about like, what am I gonna do? Who wants to come on to the show? If it's a guest, do I do something? I could, you know, I'll reach out to some people for guests. I'll see who would wanna come on.

Bilal Tahir (43:50.249)
Yeah. Maybe you can do what Justine did, like talk to like AI and then use it. You can make an avatar conversation.

Pierson Marks (43:55.179)
Yeah.

Talk to it or even just do like maybe it's like a tutorial day where we're going to be talking about like some just I'll just dig into some tools and actually just be like hey here's some tools that I'm using here's some things how to do stuff maybe that too so I'll figure out something to do and we'll keep it running. So sweet. OK. Take care.

Bilal Tahir (44:10.739)
Yeah.

Yeah, for sure. Excited for it. All right, guys. This is it. What's our call sign again? What's that Truman show? If I don't see you later, goodbye, good evening, good night.

Pierson Marks (44:24.621)
good night. Yeah, so it's like good evening. No, yeah, no, yeah, you're right. I love that movie. Whatever, we'll mess it up and next time we'll have a better sign off. But sweet.

Bilal Tahir (44:32.201)
Yeah, something something.

Yeah, yeah, all right. All right, bye, guys.