Scott Belsky on AI and creativity: 'Personalization effects are the new network effects'

6 years old. Wait, I'm on like 16. He's got a vintage behind. So, it's either that or I think also usually in the beta versions it's it's a bit like slower. I have a 16 Pro. Yeah, I'm going to light years ahead of you. Oh my god. But yeah, it looks pretty good. I'll bring it up there. Okay, cool. Bring it up.

Yeah, bring it up. Let's see it. We're putting on the printer cam. We'll see it. Uh and uh yeah, here we go. Okay, cool. Uh can we see I don't see much it. It actually looks pretty similar. I I expect Okay, there's the there's the there translation. Yeah, this doesn't look as it's it's mostly just the home screen.

Like if you go I I went into a bunch of the like alphabet opening the calculator. No, no, no, I'm kidding. I'm kidding. Yeah, I mean I mean you can see the glass effect uh like right here up at the top. Um yeah, it's nice. It's and look so so look at the search bar as as I scroll the apps here.

You can see you know that that that glass effect. It's not too bad. Um you bring down the the settings. I'm bullish. This is groundbreaking. I think this is I think this is groundbreaking. I think this works. This is this is pretty readable. Anyway, thank you for volunteering.

Uh you'll have to we'll have to check in with you over the next couple days. We're sorry if it destroys your productivity, but uh you're cracked. So, you're cracked. You'll figure it out. Anyway, thank you, Tyler. Um, you know, Apple's going to be pushing this pretty hard. They got to do some out of home advertising.

They got get they got to get on adqu. com. Out ofome advertising made easy and measurable. Drop the Genmoji ads. Market by every single Yep. billboard in San Francisco. Say goodbye to the headaches of out of home advertising.

Only ad quick combines technology out of home expertise and data to enable efficient seam by seamless ad buying across the globe. I'll be right back. Um, we have our next guest, Scott Bellski, coming into the studio in just a minute. In the meantime, we will react to some posts on the timeline.

Uh Kendall says, "Trying to quickly change the subject on the Zoom call when everyone else is cracking jokes about the Whimos and I can feel I can feel myself choking up. " Lots of sympathy for the Whimos.

Uh, lots of coverage in the Wall Street Journal today about the about the burning whimos, talking about how this played directly into the hands of of the of the Trump administration by uh making the tech company clearly align with um, you know, against the burning of Whimos. Very sad to see.

Um, hopefully things resolve. It does feel like the report from LA I've been driving around today. Uh, it feels like things have calmed down a lot.

Um, in general, I was able to drive all the way across LA and without any I was in downtown, didn't really notice anything, and so it does seem like things are are calming down, thankfully. Um, anyway, um, I like this post from Tyler. Uh, I'm going to mispronounce this name. Well, I'll just say Tyler.

Um, he says, "Unlike the rest of you cowards, I think the W the Y Combinator valuations are too low. " Uh we will find out what the valuations are coming in at tomorrow when we're at YC Demo Day live in San Francisco. We're bringing a huge swath of the crew. We're going up right after the show wraps. The team is stoked.

We'll be giving out hats. We have a whole bunch of different hats. Some TBPN hats, some ramp hats, and the ramp hats have TVPN on them. So they're all limited editions. I think people really like them.

So, if you're a founder at YC Demo Day or you're an investor at YC Demo Day, stop by our table, come chat with us for a couple minutes and we will uh get your take and hear your pitch and hopefully turn uh your demo day into a massive party round because we will be live from the Palace of Party rounds.

Um Mike Salana uh reposted the video of the Whimos leaving Los Angeles. Apparently, Whimo is moving their cars out of LA. I don't know how real this is, but there it was a video of several Whimos uh driving through LA kind of all together. Maybe they're leaving.

Doesn't quite make sense because you think you could just put them wherever they store them and charge them and clean them, but yeah, I don't necessarily believe that this was a response. This seems like fake news, but uh Salana has a great post as always. The elves are leaving Middle Earth. Very funny.

Um anyway, uh RAMP has a post uh in 2012. In 2024, RAMP placed 32 on CNBC's disruptors 50. This year, number six. They're grateful to their customers for betting on a better way and allowing us to fix the parts of finance people hate so they can do more of what they love.

And shout out to our friends at Ander for making number one. What a fantastic lineup over there. Um what else is in the news? Oh, uh Warner Brothers. Oh, wait. We have Scott Bellski in the studio. Let's bring him in. How are you doing, Scott? Hey, what's up? There he is. Finally. Long overdue.

I I I'm sorry it took me so long to invite you on. I'm so glad we could make this happen. Uh it's fantastic to find. We were we were sending invites constantly. Mentally, we were like, "This is a dream guest. " So, thank you so much for taking the time to uh join us.

Uh would you mind kind of setting the stage for us because you've done a lot in your career. What are you focused on right now? What's interesting to you? and then we can kind of dig into all the hot topics of the day.

Well, first of all, it's been fun to watch you guys at work and uh congrats on just, you know, creating conversation, which I think is uh is appreciated by all of us. What am I up to these days? I mean, my obsession has always been the intersection of creativity and technology. Yeah.

Behance is probably what 58 million 60 million creatives online right now showcasing their work. that brought me into Adobe. Had a long stint at Adobe um leading emerging products and design and various other new things there.

Um and uh you know now I'm excited to uh kind of explore the world of storytelling and the role that technology plays there.

um obsessed with the implications of new technology, you know, that emerge and uh I love engaging in conversation about I think you guys were talking about the implications kind of writing I do and uh and riffing off of I don't know like what's going to happen because of what's going to happen is the question. Yeah.

Yeah. It's a big question. It's a big question. Uh maybe we should start with uh some of the news that's coming out of like this idea of collective memory. What's going on in the enterprise? We were talking about Glean, OpenAI, deep research. There's so many new products. We're using stuff. We're wiring up stuff.

We have an intern in the corner vibe coding different stuff for us. We're we're essentially a podcast or a live show, a media company, and we're we're writing software. Everything about how we're building this media business is different than I think people would have done uh even just a few years ago.

Um, but yeah, it's actually crazy. The um if if somebody told me three years ago, I we have a a show, a podcast, and we're building a bunch of internal software, just be so bearish, and yet we get so much value out of our internal software we call Newsmax. Um, but anyways, um, back back to you.

No, I mean I think the uh well, first of all, it's it's when you're building a new team from scratch these days, it's almost like you're you're able to build with another new generation of tools as opposed to retrofit something built with old tools. Totally.

So I do think there's this advantage that a lot of new companies are realizing now or new teams within companies, you know, but you asked about this this concept of collective memory which I've been obsessed with, you know, this this notion that the context window of these AI tools that we're using.

Of course, those contact windows are context windows are growing, which means that their memory of us and their understanding of us is growing and persisting across all the inquiries we have.

and uh you know and this is becoming almost like you know an extension of our knowledge and opinions and identity and everything else and and it's great when these LMS get to know us and give us better and better answers to our questions.

um what happens when in the enterprise we access each other's memory right so if if the if the context window of me working you know as one of your interns for years you know is suddenly accessible to you even when I leave you know are you able to keep saying hey what does Scott think or you know what what what did Scott do when we had this situation and is that a is that does that belong to the enterprise and you do see by the way like every company right now trying to build this memory for each of its products users.

And in some cases, a lot of companies are building connectors to get the data from other products that that customer uses to enrich that memory and understanding of their customer. And it'll be interesting when that kind of memory that's for each of us becomes a collective memory.

And then what happens in the in the social like in the consumer world like you know does your girlfriend say like I want access to your context window, your memory, you know, and is that kind of weird? I think there's all kinds of weird questions and implications that arise from that.

Do you think memory can be the lock in that I think certain companies would like it to be. It doesn't feel like it is yet, but I think there's still an idea that it could be. Sometimes it actually goes the opposite way.

Like I I'll ask Shhat BT to tell me a joke and it'll be like hyper specific about something I talked about a year ago and I'm like that actually makes the joke worse. Like I'd prefer if you ignored all my questions about trains or something. Well, that could be in the prompt, right? Yeah, exactly.

I feel like uh you know, you see OpenAI launch um the ability to log into tools with Open AI. Yep. And you know, when I saw that, I was like instantly thinking, gosh, that's a way to enrich the personalization. Um I I do think personalization effects are the new network effects.

You know, we used to talk about in terms of the the moat that companies would have with a network effect. today.

I do think they will have with personalization and that is a direct, you know, a direct uh out outcome of the memory and extended context window and all of the data, you know, that they can get to give us a personalized experience that no one else could ever give us.

But you're right, like there are other ramifications of that, you know, if you if you're too wellknown. I asked OpenAI the other day, you know, what I didn't know about myself and it told me that you're a real hypochondric. And I was like, well, I didn't know that about myself. So that's funny.

Um I I I I want to know more about um these the like this idea of collective memory in the enterprise. What sorry sorry to interrupt but it's an interesting thing where like the model is effectively trying to do pattern recognition and it just identifies something that is like you ask a lot of questions.

It's like yeah because you're a question answering machine. Yeah. Like that's not me that's you.

But uh but on on the concept of like creating a collective memory in the enterprise, I've noticed that even though we are this new company that's starting from scratch, total green field, we have still not been able to go all in on one walled garden. We have Gmail, we we send a lot of iMessage groups.

We also have a Slack that, you know, we are interfacing with different people on that's a Salesforce product. And so I'm wondering if uh you know there's been uh Eric Mikovsky was trying to build Beeper, this product to let you create one unified messaging interface.

That was a real struggle because every major company was like we will not let you break our walled garden. We're not really there yet. And and it seems like most of the companies are playing ball um with MCP servers and the like.

But what do you think the big dynamics of the big tech companies will be like once they realize that, hey, maybe if I don't want to let the fox in the hen house?

Well, I think you're, you know, you're you're you're forecasting a bit of a call it a data war, you know, in the years ahead where everyone's going to leverage each other's APIs and start start to pull all the data, you know, in and and they're going to probably be some backroom board level bilateral negotiations where it's like, well, let's not cut off Slack because we don't want them to cut off us.

So we're going to let them and may the you know may the best may the best model win right with the data provided to it.

So um but that means that there if if if the data becomes sort of s you know readily available across the board from all of the connectors that are made for all of the products we use and to your point if you've built the network your your company with all these different products that's fine because the data is all stored probably in one of three or four clouds and you can just instantly uh access all that data use AI to see it as structured data and then be able to have AI apps built on top of it and you don't even need to care what actual products create the data as long as you have access to all the data.

But then the question is what are the modes, right? I think one of them is the personalization layer that we talked about just now and uh and that context window becoming ever more important uh in the product.

You know, I think another one is permissioning like in the enterprise especially permissioning is freaking hard, right? To know exactly who's about who's allowed to ask what question and get what answer or what column of information that might relate itself to that answer.

that is a a very very hard thing to crack and I do believe that the enterprise products that have really incredible permissioning layers and products uh that allow people to control and access permissioning you know will also will also have an advantage.

So I mean and then and the operating systems um we talk about operating systems in the consumer world with you know your iOS or Android person or whatever and that's the operating system.

Those those operating systems like if they can't figure it out well shame on them like they literally own the uh the the top layer of everything underneath and they should be able to you know manage our attention and our needs with AI at that level. But what are the operating systems of work? Mhm.

You know, I think there's going to be another war played around those operating systems. One might be the browser. I think one might be the way that projects are managed.

You know, different functions of the enterprise might have different operating systems and people are going to be fighting to own the top layer of those. Yeah.

I wonder if that's like a almost a bullcase for data lakes becoming more of a mid-market or early stage product because like I I I I I take your point about like the the the deals between the large tech companies uh just kind of happening behind the scenes but at the same time I'm just like there's going to be friction and I just know these companies like yes I can export all my data from Facebook if I click 12 buttons and do it every single day.

like the API is not fully there for a lot of these products. And so I wonder if more and more people are going to be saying like even for my personal life, I want to be able to funnel everything into a data lake that then I can drop my own AI on and I don't necessarily want lock into one AI tool right now.

I'm wondering how that how that will evolve. I don't know. Yeah, I mean we're going to see this in consumer too. It'd be really nice right now if you could describe in chat GBT, hey, go find this image and then make it a studio Gibli, right? or or make it like an impressionist painting, right?

And and obviously there there's that kind of barrier wall. It just feels like the the big tech companies want they're fine having lock in and like you see this with WWDC like the the Siri button cannot be remapped still.

And it's like that would have been an easy thing to do that would have been something that a lot of consumers demanded. uh what was your reaction to WWDC and kind of like how how the the AI Apple uh like uh decisions are playing are playing out.

Yeah, I mean I think it's really it's a really fascinating time, you know, in that company's history. I think from the outside we're questioning like where's the vision, you know, where are we all going? And I think inside it's probably some sense of let's wait until we're really ready. We've had a few false starts.

We can't get it wrong again. And so, we're probably going to get something that's more fully baked, which is hopefully great for Apple. You know, I never I never discount Apple because of its DNA and its rigor and the quality of talent that's there.

You know, on the other hand, like they're late to the game at this point, like super late. And I also wonder the things we discussed just a minute ago around personalization and some of the moes that are that are going to play.

Also, by the way, remember this is a new era of experiences that is all about data and Apple's always been all about privacy. So, Apple in some ways as a policy has done everything to make it impossible for others to get others data.

Y and here we are in a world where you know that's what's going to enrich the experience we have from AI tools. So, I think it's a it's a different Do you think that Apple do you think that Apple will end up looking really smart on privacy?

Like we're in a very chaotic time where people will let a new software, you know, company just have full access to visual screen recording. Record my screen. That's like a okay thing now.

And I'm even surprised at this point that we haven't I I can't remember a major incident of a model producing private data in response to sort of like a a general query. I've heard about little anecdotes here and there, but nothing really.

You can imagine a world some days where you know people get their kind of totally um lines crossed and you ask for you ask about some some you know acquisition for example and somebody at some point dumped in.

We even tried to get ChachiPT to guess a picture that was from your own house and it was and it was smart enough to say, "Hey, you're maybe trying to spy on someone, so I'm not going to do that.

" Even though Yeah, but I could I could it would it would basically gave me like the exact description of it without giving me effectively the street. So, it knew it knew not to step up the line. I I don't know. Anyway, on privacy, I mean, it's a great question.

You know, I always go back to first principles on this one and I say to myself, you know, first of all, what do we know about the next generation? They seem to care less about privacy than the generation before them. Yeah.

Um, you know, number two is we're we're willing to trade a lot for less friction in our lives and it seems like people are willing to always click accept or yes or yes or yes, you know, again, just to be able to have a frictionless experience in a product.

Now, I think you're surfacing the point that very, you know, we we don't oftentimes realize like how much we're giving over um in those instances. And I think the world of AI will make that very apparent to us because we'll start getting these personalized experiences.

We're like, how the hell do they know that about me, right? Um but I don't know, like there's also part of me that says that the best technology takes us back to the way things once were, but with more scale and efficiency. Well, the way things once were hundreds of years ago is we were known.

Like everywhere we went in our small towns, we were remembered. They called us by name. They knew our kids' names. They knew our favorite cut of the butcher. Like you were kind of known. And in some weird way, we're being brought back to a world where we're going to be known again.

Um the question is we want to know how we're known. And and also we don't want to be known by people that we don't trust. Yeah. Yeah.

there there is kind of this interesting economic uh dynamic where there's almost like a de facto bug bounty for privacy in the fact that if a company like Apple has a privacy fi fiasco that could wipe off hundreds of billions of dollars of market cap and so they have this massive incentive to invest to not have these privacy fiascos the data to leak and so you have kind of this massive economic capital cannon towards let's secure the public clouds and So uh that my one of my former colleagues was making the point that um in terms of open- source AI versus closed source AI um oftent times it's actually better to trust a hyperscaler with your data that because it's harder to break into an AWS data center than it is to the on-rem server that yes you're running it but you're the only security guard and what happens if someone breaks in and just takes the rack like I'll tell you what though like here's a bullish case for Apple would be that models are all going to be local you know x number of years the most power like if you look at the performance of of LMS right there are these small LMS and there's these ones that run locally on on devices and the the the the delta between those and top LMS is of course decreasing um and so as chips get better and LM run locally you can imagine that all of your data can actually be stored locally in a super encrypted way all these local models can operate to answer all of the questions you have can search the internet when it when is necessary.

And maybe we all opt to this like super private AI world.

Maybe for all we know, Apple's optimizing for that flag that they know the puck is going to, you know, and they're going to launch who knows how many years from now like 100% local AI experience that is world class with zero privacy considerations or or risk rather. Yeah. Yeah. Yeah. Yeah. Yeah. That makes no sense.

Uh can you talk a little bit about knowledge arbitrage? like these models, we've gone from uh you know, massive databases of facts and so there wasn't as much incentive to just be a the the the person that memorizes all the facts because you can just Google it on your phone.

Now we're in this age of even more significant knowledge retrieval intelligence potentially to chief meter if we can call it intelligence. Um what is knowledge arbitrage? like what? Yeah.

I mean, it's funny like when I think about the term knowledge arbitrage, like I'm actually brought back to um and I think this is very relevant for like people who are listening to us or kind of earlier in their careers right now and figuring out like how to plot their path forward.

You know, you go back to the early 2000s and the people stood Twitter and social media and things like that were getting hired by like CMOs of Fortune 500 companies to school them on what the heck this thing was and like how to have their brand participate.

And it was really looked upon as like this complex crazy thing that every leader in in in in corporate America felt they needed to understand.

And so there's this moment of knowledge arbitrage where people who were like, you know, in their late teens, early 20s who were deeply natively familiar with this stuff called social media could just like school the people who had no clue. Here we are again.

You know, you have the leaders of all these Fortune 500 companies saying, "Oh my gosh, like we have to refactor how we work and every function is going to operate differently with AI, but we don't even know like when to use chat GPT. We don't even know what a prompt is and like how does this stuff even work?

" And so once again, these like very young people who got through college using chat GPT and talked to chat GPT about their boyfriend or girlfriend issues or whatever else, like it's native to them. And so I do think this knowledge arbitrage moment, you know, is here and the windows open right now.

Yeah, it's funny that there's this massive fear from from young people around how the job market is evolving. Are jobs getting killed or whatever. Um, and at the same time, yeah, it it feels like social media didn't reinvent every part of a company.

Like it it sort of massively changed marketing and how you should focus marketing dollars and how you should think about building brands and all these things, but it wasn't something that that changed internal comms like you're talking about with like collective memory.

It wasn't something that changed a certain aspect of the business like compliance. It it didn't change, you know, it wasn't the sort of like full stack transformation.

And and for that reason, it creates more opportunity than the social media wave to just get good at using LLMs and and agents and whatever whatever other tools you have and then use that.

I mean, we the the the last person that we hired was somebody who who vibe coded um uh like a guest directory for us and they did it last like Friday. Last Thursday and then we hired them that and and then we saw it like Friday. We they started on Monday. came on.

Yeah, they came they did on Thursday, they came out on Friday and we had them start and just by demonstrating the ability to leverage these tools and create a product that was a V1, but it was a solid V1 and it was just a super easy decision to be like, "Yeah, you should join the team. " So, shout out shout out to Adam.

48 hours work though. That's like incredible. Um, so and by the way, that's changing the way product is even done.

Like I I'm so used to a world where you concept, you do like a sort of you know product scoping session, then you have the designers work in uh a product like Figma and make prototypes and then engineers don't even see it until the prototypes are locked and then it's like redlinined and engineering and start to like you know now it's a different world, right?

Y it's like we need we need we need this to be done. You kind of vibe code it out in parallel. A designer says okay let's how to finesse it. here's how we can maybe change that, make that more accessible to more people. And then you just deploy and test and iterate.

It's a different it's a whole different product stack to some degree.

And that's why I always say the best talent these days is like a collapsed stack talent strategy where you hire people that have exposure to different parts of the stack, whether it's engineering and design or design and copy or product and design because you're just trying to collapse the stack with the types of people you have.

So you can leverage these tools, you know, efficiently. Can you talk a little bit about agency? Um, it's it's an interesting concept because we don't have like a eval for it. Uh, in terms of the LLMs, uh, we're we're we're you know destroying all the benchmarks in math and IQ.

Um, but the models don't seem to have valition and yet the trend is around agentic applications. We're trying to give them agency. it still seems like uh some there's some sort of combination of uniquely human traits that uh seems to be increasingly valuable. How are you defining it right now?

Well, I think that when you talk to when you talk about sort of agentic workflows and the the role, you know, the role of agents in the future, you know, and I think about it every company, you know, I'm calling companies these days in my mind cognos like cognition driven businesses where it's ultimately inference in the center, right?

And there are these the functions of a company which were once accounting and legal and design and like all these different functions are going to be AI tools that are running you know AI agents basically that are running that function or collection of agents doing different aspects of the function which then begs the question of what's the role of the humans.

I think we are the orchestration designers and orchestration engineers. We're like orchestrating how these agents work and how they work together. And we're also developing enforcing the rules by which they work because remember these agents are told what to do and they're rewarded based on doing it.

They have to be working within a subset of rules that um that need to be administered and enforced to some degree by humans. So I you know so I I start there and then I start to think okay so what are the functions of every organization going to look like?

you know what and then that dictates like what kinds of tools or agents are empowered to uh quoteunquote take agency in those functions.

But I also want to say that when you say agency at first I think of like the human, you know, the the human getting a got it like we gota dude we got to we got to exercise our agency and our tastes more than ever before these days.

Um and the the the years the days of us kind of getting a prescribed job function and sticking to it that is like the deathbell of a career these days. Yeah, totally. As soon as it can be defined, it can be reinforcement learned against.

Were you were you surprised at all about this the the concept of taste coming into the the the uh discourse the Silicon Valley discourse? It feels like one of those things.

It actually reminded me of when venture capitalists discovered the creator economy in 2020 and 2019, but it was a trend that had been happening for like a decade at that point where people were creating content online and turning it into a business.

And then all of a sudden, people I think saw a statistic that was like, you know, look at this, you look how fast this category is growing. We should invest a bunch of dollars there. But in many ways, I feel like taste has always been a key part of um our industry.

And and if you look at a lot of the best companies, they don't all look the same, but they have people that run them that have um taste that that sort of gets applied in in many different ways, whether it's hiring all the way through obviously the the visual side.

you know I am brought back to a uh just earlier in my career you know it was all about what skills you had right and and the resume would always were you good at PHP or JavaScript or could you do wrote code in Python or whatever the case might be come from a world where skills like were the hardest things to achieve and the most differentiating things for humans right in in a project now as a lot of the skills have been offloaded to compute you know, now it's like, well, what's left?

Like taste is actually one of those things that is definitely left, right? And so I think that the fact that everyone's sort of focused on on on taste these days is more of a commentary on how less focused they are on skill. Yeah.

And how much confidence they have in all these models and coding tools and everything else. Can taste be taught though? That's the question. Skills can be taught. Taste taste be taught. Like let's you know can taste be taught is like a thing I think about and debate all the time with with friends.

You know what is taste? Gosh it's like the human experience. It's childhood traumas. It's mistake of the eyes. It's judgment.

It's knowing how many things to see before you make a decision you know but it's also knowing how to choose something that isn't so on the nose but like leaves something for the imagination, you know. It's knowing where people are going as opposed to where people are.

Like I think it's all of these factors that are so critical whether you're making a product or making a marketing campaign or a film or whatever the case might be. And I just, you know, I I I think that hopefully we'll just overindex on human experience in order to achieve taste.

Yeah, there's Yeah, it almost feels like just independent thinking um confidence like you can maybe develop No, but it's independent independent thinking that elicits a collective response that you want in some ways, right? I mean, I I think of this in the context of how companies get attention.

There's a lot of different ways to get attention. Some some some ways are are ta, you know, some ways are tasteful and elicit, you know, a positive response.

other others get, you know, a lot of attention, but but a sort of visceral negative reaction and and taste in the context of social media algorithms and timelines is like such an important um question because it's really easy right now to do to instantiate something but not to Yeah, it's fascinating. Yeah.

Taste taste too in the context of AI uh slop, right? There's tasteful, there's tasteful. I always keep coming back to the Harry Potter Balenciaga video.

I feel like that was, you know, it it had the aesthetics of AI slop because it was AI generated with it and yes, but it it had the it had the touch of a human to think of combining those two kind of completely unrelated disperate ideas and when you put them together it was something that you had to watch.

And by the way, that's perfect evidence the fact that you're still citing that