TML Inkling, California Forever?, TSMC Capex | David Baszucki, Everett Randle, Eric Glyman, Jordan Black

You're watching TVPN. Today is Thursday, July 16th, and we are live from the TVPN UltraDome, the temple of technology, the fortress of finance, the capital of capital. Let me tell you about Ramp.com. Time is money. Save both. Easy to use corporate cards, bill pay, accounting, and a whole lot more all in one place. We have a very special show for you today because we have Tyler Cosgrove guest hosting. He's here in our chair. Great to be here. I know I'm going to make mistakes today because I always throw it over to Jordi. I got to remember this, Tyler. It reminds me of this video from Warren Buffett. We got to play it to show what I'm going through emotionally today without Jordi in the TVP and Ultradome. Let's pull up this video of Warren Buffett throughout the years at the Berkshire Hathaway shareholder meetings. Charlie, how do you feel about that?

Charlie, Charlie, me?

Are you Charlie?

Charlie, I'm so used to that.

Oh, you got to play from the beginning.

I strike myself a couple of times.

Years and years, Warren Buffett always goes to Charlie after he gives his comment. He, he, he, he gives his speech and then he kicks it over to Charlie. Charlie, through the years. Charlie, Charlie, Charlie.

Tearjerker. Yeah. Makes me want to cry.

It's emotional. Charlie, how do you feel about that? Charlie, Charlie, me, Charlie, Charlie. That's Greg Abel. So if that happens today, I apologize. But there is a ton of news we're going to be going through today. We have a packed lineup as well. We have Randall, Eric Gleiman, Jordi Hays, Dave Buzuki from Roblox coming on. We have a great show and there's a lot of news. The first big story is, of course, Thinking Machine's new model has released. Mira Murati's AI startup released its first model in bid to loosen AI giants' grip. We're going to be talking about open source, closed source, where the frontier is, national geopolitical model moves.

And some people had an idea this was going to happen. Some inkling.

Oh, they had an inkling that—

yeah, that they're going to release it.

I didn't have an inkling that they were going to jump into the open source space.

I think it makes a lot of sense given the Tinker API, right?

Okay.

The whole business is, you know, you're doing fine-tuning on open-source models.

Yeah.

It makes a lot of sense that they're going to have their own.

Yeah.

That, you know, you can easily substitute.

It's sort of like they are set up as a business to launch an open-source model without it degrading any other piece of their business because the Tinker API, that fine-tuning that they do, that integration with the customers that they have actually benefits from open source.

Yeah.

And then they can go to their clients and say, look, you know, it's the Red Hat model at any time, you can leave because we are giving you the weights of the model open source. You can do whatever you want with them, but keep working with us because we're helping you a bunch and we're making money in the process. So Thinking Machines Lab, first, the first model is an open weights model designed to chip away at the lead of OpenAI and Anthropic, says the Wall Street Journal. Former OpenAI technology chief Mira Murati is betting on more customizable artificial intelligence models to chip away at the lead that the frontier labs such as her former employer hold over the technology. TML, a company led by Murati, released its his first AI model Wednesday and did it with open weights, meaning that others can modify it with their data, called Inkling. The model has 975 billion total parameters, making it far smaller than estimates of the most advanced closed-source models.

Yeah, but I want to remember, so it's a mixture of—

Yeah, mixture of experts.

So only, I think the number is 41 billion of those are actually like active at any moment. Yeah, so this is definitely on the bigger side of open-source models. Sure, yeah. But like that number, it's not, These aren't like dense models, like what you would traditionally think of, of the models like, you know, 4 years ago.

Yeah, yeah. Murati told the journal, we trained it to be a broad, balanced foundational— foundation model, strong across many domains, flexible enough to adapt. Inkling is not the strongest overall model available today, open or closed, which is a different frame of reference for many of these model launches. There's— it's been— everyone's been jockeying for the frontier, even if they're not world-class at everything. Usually when they launch, they say, oh, well, we're best at something or we're best at this, but a different tone, different communication strategy. And I think it's being well received. I think people are having fun with it. The reaction.

Yeah, I mean, I think the main pitch here is that this model is like uniquely set up for the Tinker API. It's built to be fine-tuned.

Sure, sure.

That's the whole point.

Got it. DD Das says, "Thinking Machines just dropped the best open-weight AI model outside of China." And obviously that is a big topic of conversation as business leaders in the United States have some policies and some reticence about using Chinese open source models, even if they're not worried about the dystopian, you know, Manchurian candidate hidden inside the weights. Maybe they just want to be aligned with a US-based company for a variety of reasons. Inkling beats NemoTron 3 Ultra and benchmarks put it between KIMI K2.5 and 2.6. Of course, there's also news today that KIMI K3 will be launching and is another jump forward. But there's back and forth between some AI researchers around what's going on there, how long that strategy will continue. So DD says many were contending to this throne, but Thinky has come out on top. Really solid release and will pair well with Tinker. So there are some benchmarks that you can go and dig into if that's your thing. There's a, there's another very bullish take from Jack Morris of Engram Labs. He says people are underestimating what a big deal this is. This is the only open weight model that's trained without distilling from OpenAI or Anthropic. KIMI distills, GLM distills, Quen distills, Nemotron distills, KIMI and DeepSeq, which counts. Basically a fully different tech stack. The first pure open frontier coding model. Very exciting. And there's a community note on this. Can you break down exactly like where are they standing on the shoulders of giants? Where are they not?

Yeah.

So yeah, so I think this tweet is not exactly true. In the blog post, they say, to bootstrap post-training, we ran an initial supervised fine-tuning on synthetic data generated by open-world models, including KIMI-K 2.5.

Okay.

So I think that's generally how people think of distillation, that they mean something related to this.

Sure.

So I think this actually is not that different than what people at NVIDIA with Neematron did.

Sure.

So this is not very new, I think.

But it's sort of like the lightest touch of distillation that could happen because it's just one piece of the pipeline, one small amount of data. It's not one of these scenarios where we're like, why is it identifying as Claude? Or why is it just saying that it's ChatGPT?

But it is funny because you can kind of say like, oh, well, if this is like kind of distilled on KIMI and KIMI is kind of distilled on closed source, well, then maybe you get some kind of two-layer distillation.

This sort of round trip loop. But at the same time, there's probably something to be said for the more layers of abstraction, the more ingredients you pour in, like the distillation becomes weaker and weaker.

Yeah. And I think There's also an important question of like, well, okay, they're doing some level of distillation, like why? Because you can either be like, well, they're just doing it to save time, whatever. Like obviously they have these capabilities, but there's no point in, you know, doing everything over again. You might as well just use what's out there already. Or is it because actually like these capabilities that they get from this, you know, distillation light, whatever it is, are those actually super imperative to the model like being good?

There was a reaction from Ingram, company founded by Jack Morris. First, let me tell you about Consul. Consul builds agents that automate 70% of IT, HR, and finance support, giving employees instant resolution for access requests and password resets. So Ingram says our founder Jack Morris recently issued some unfounded claims that got community noted. We deeply apologize for the confusion caused by his original post, the follow-up post, and the follow-up to the follow-up post. Nevertheless, we stand by his conviction in his own takes and in strong open-source models like Inkling. And there is a question of like, distillation is a vague term where it's not a binary thing. And if it's not in the pre-training data, does it count?

I think it's also very much this. I mean, people love to talk about ONNX. Yeah, they like to kind of, you know, scapegoat, oh, you know, it's all distillation. That's the only reason Chinese models are good. Yeah. Is that actually true?

I mean, Anthropic's head of national security policy, Taran Chhabra, accused Zhipu, the Z.AI, of distilling both Claude and OpenAI models for GLM 5.2 at the Aspen Security Forum. Earlier this week. This is from Vincent Chow, senior AI reporter at SCMP. He said it's the first time that they've named Zhipu specifically after previously calling out Deepseek, Alibaba, Moonshot, and Minimax. Join the, join the club at this point.

Yeah.

Also accused— they also accused Deepseek of continuing its adversarial campaign of distillation. Anthropic is now shutting down distillation accounts on the order of millions accounts per week. That is crazy scale. You have to— I mean, you always think about it as like, oh, there's like shut down that one company or shut down that one block of IP addresses. But when there's a really, really distributed attack, we've even heard about whole companies that just like resell Claude tokens or GPT-5.6 tokens. And that looks like a reasonable business because it's just a wrapper company. Of course you want to work with them, but then you don't realize that but on the other side, who are their customers? Why did they get to 100 million run rate so quickly? Well, maybe it's a lab that's trying to distill through this pass-through entity. And of course it's hard to watermark the tokens once they go out the API and they get passed through some other system and they can go through other countries, all sorts of things. So millions per week, that is crazy. That's gotta be really difficult to, it's a game of whack-a-mole. They say GLM is quote, probably the most advanced Chinese model on the market now, which poses significant cyber, security challenges. They hinted that Anthropic will expand access to Mythos to ensure a fair fight for cyber defenders. And they said that distillation challenge is real in shrinking US lead in AI, suggesting that the US government could do more to clamp down on Chinese model adoption globally by working with allies similar to trusted telecom efforts like Huawei and ZTE. So obviously a hot topic and people will be debating how exactly how heavy of a hand the government should be putting.

Yeah, I think this release is also, makes a lot of sense in the, I think it was a week ago, there was that article about like Beijing is looking at curbing overseas access to Chinese top AI models.

Yeah.

Right, so if you're not gonna be able to access the Chinese open source, right, it makes a lot of sense to start doing American open source, Western open source.

Yeah, it really does feel like there's a pretty wide.

It feels very well timed.

Yeah, it seems like there's a pretty wide gap between at least what's reported preferences from Beijing, from the actual government and the companies. The companies are like, yeah, send us all the Nvidia chips, let's distill everything, let's— and then let's open source these models and compete internationally. And Beijing's like, hey, maybe we need like, you know, an indigenous supply chain here. Maybe we need to, you know, lock down these models, keep our lead over here, go work internally. I don't know. But if you're worried about security, head over to CrowdStrike. Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches. This was an interesting post from Grace Lee. She asked the question, how did OpenAI's SOL finally learn design taste? She projected 1,000 websites by GPT-5.6 SOL into a design manifold and discovered big holes. These holes were where GPT-5.5 previously generated outputs with, quote, bad AI smell. So there were, you know, there's these tells in any AI model that it's not this, it's that, the em dash. Once people start identifying those as, ah, we don't like that, it's too AI, it's too generic. One way it appears to actually sort of beat that out of the model is to actively avoid those specific things. And then she calls out 3 particular areas that have been avoided as anti-patterns. One, the bento box layout in dashboards. Two, large typefaces and hero images. I did realize that sometimes you would ask for a website and you would just get a massive block of huge text and that's just not the way you— when you land on a beautiful website, it's usually there's more wordsmithing, there's more—

Yeah, yeah.

—curse language.

Well, you know, you make your first website with 5, 6 code or whatever and it looks really good. Yeah. And then you make 10 and they're like, "Oh, okay. There's actually a lot of patterns I'm seeing." Totally, totally. And you can start clocking them like everywhere you see— Yeah. —you see a lot of like Claudisms, whatever on— Yeah. —general design. Especially— You see them everywhere.

Yeah. Especially if you don't come with any—

You know, it's like the— The, you know, high border radius on the edges.

There's a little color on the side. Yeah, yeah, yeah. Especially if you don't come with like an opinion. If you come like we made a whole vibe-coded website in Codex for just the latest episode of Nick Bostrom on Joe Rogan. And I wanted it to look like a UFC fight card and a fight promotional website. And it doesn't look like any normal AI slop. I mean, there's still like AI-generated images. It looks like AI, but it doesn't look like, oh, yes, Yes, that's the bento box layout or that's the offset layout or that's the purple or it's stealing from linear. It's a completely different style. So if you at least inject like one reference point, you'll usually land somewhere. Yeah.

I mean, it is interesting though. This makes it seem like, you know, the new model is not necessarily— it doesn't have like higher variance with outputs it gives, but we basically just found like, oh, there's certain examples that people really don't like. Let's just remove those. But you're not necessarily like making the model more creative by, by removing these like patterns it always comes to.

Yeah, well, you're giving like the, the flavor of creativity and maybe that's—

Yeah, but you can imagine if we kind of keep the same model for 6 months, we'll just notice new patterns. Totally. And you'll have this kind of—

But at the same time, like Midjourney had like a very distinct look and people like that look, at least some people. And so if you can if you can quickly personalize and customize and land in a place where someone whose job is designing dashboards is happy every time with the layout. Like, there is somewhat of a platonic ideal for some of these design patterns. And at the same time, if you're working on certain— like, there's certain designs that are just like solved. Like, you know, make the call to action green or blue, not red, right? And so some of those like do need to be consistent. And then also I imagine that many folks who are using these tools like in enterprises are doing, even if it's not a fine-tune, they're uploading a reference for everything that they're designing so it's consistent with the brand that they've designed, right? Anyway, let me tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. Just do it. Just do it. Stop making excuses. California Forever lost a $3.2 billion shipyard project from defense startup Seronic after the company chose the Port of Brownsville, Texas, over Solano County. I was— oh, no. You're not supposed to clap for that. We got a Texan in the studio who's happy about that. This is bad news for California. We want California to have a whole bunch of amazing stuff. Brandon Corral, who wrote the newsletter tppn.com today, was very disappointed about this. The automated shipyard shipyard known as Point Alpha— Port Alpha is expected to create roughly 10,000 permanent jobs along with thousands of union construction jobs. Supporters say California's lengthy approval process ultimately cost the state one of the first marquee tenants that California Forever had pointed to as evidence its planned city could anchor a new era of American shipbuilding. Joshua, executive director for the California Alliance for Jobs, said California failed to move with the urgency the project project required, quote, while Texas moved quickly and aggressively. Thank you, Jackson. California could not provide clear expedited approval process needed, he said, calling the decision an enormous loss for Solano County, California workers, and our state's manufacturing economy. Earlier this year, California Forever signed a 40-year construction labor agreement covering 70,000 acres, and labor groups later backed legislation to fast-track environmental review and permitting for the proposed shipyard. The legislation has yet to advance. Instead, Texas approved a $211 million tax abatement package in June to secure Serenics' investment at Brownsville, roughly 20 miles from Starbase. Labor leaders said they warned that without expedited approvals, the project would leave the state. And that is exactly what happened, a project insider told the San Francisco Chronicle. The California Forever itself remains on track, but acknowledged that losing a major defense contractor sends a powerful signal about the state's ability to compete for large industrial investments. Very disappointing. But I like Yan, I like the California Freeway Project, and I'm excited for where he takes it next. I'm sure he's on the hunt for the next major tenant. But we have our next guest soon in the waiting room. We'll bring in Everett Randall from Benchmark in just a minute. Got to talk about TSMC. Yes, TSMC. Where is this in the stack? TSMC, both beat earnings and raised their CapEx guide.

They're spending a lot more money and pledged to invest an additional $100 billion in the US. Yes. Plans to spend a record amount cementing its position atop the global semiconductor supply chain.

Yes. But and yeah, and they're investing another $100 billion in Arizona fabs. But people are worried about overspending. The news is that the Nasdaq dropped 1% on TSMC's spending plans offset by strong results. Very, very, very, very odd story that in a time when even TSMC, which was not a particularly AGI-pilled company for a long time since they've been through the smartphone boom, the so many booms and busts, so many cyclical build-out cycles, that when they are finally like, yes, now is the time, people are skeptical. But we have Ev Randall, in the waiting room. Let's bring him in to the TVP and Ultradome. How are you doing? Hey, gentlemen, how are we doing? Welcome to the show. Jordi traveling. Jordi, we have Tyler Cosgrove. Jordi's traveling on our team. A guest host. Very excited to have you. How are you doing? How is, how is the year going? I'm interested in just like your general state of the markets. You recently said— I've— you said it's an incredibly disorienting time to be investing. What's disorienting? You just put money in every company and they all go up.

That's, that's certainly what it feels like. Feels like for the last 18 months, which is, which is scary in and of its own. Yeah, I did. I was on, you know, my partner Jack Altman's podcast Uncapped with, with Trey and Dylan over at Founders Fund. And I was saying that one of the scariest things of today, and then we can go to the disorienting part, but one of the scariest parts is this feeling of inevitability. Like, it kind of feels like sometimes, oftentimes, it does feel like as venture investors, we are monkeys throwing darts at a dartboard. Um, and you know, you never really know when, like, what, what, what number you're going to hit. Um, but imagine you're that monkey and you just keep hitting like the triple 20, like, every time you throw it, and you're like, this is weird. Um, and, and, you know, your response to that usually is just like throwing a lot more darts. So the last 18 months, last 24 months Everyone, I think, is feeling, feeling, you know, very, very confident in, in the market or just like how all their companies are going. And all these companies are growing extremely fast. And so everyone is investing a ton of money. It's all getting marked up very, very quickly. We obviously had the SpaceX IPO, which was huge for several firms. Yeah, we're probably going to have the OpenAI and Anthropic IPOs that are going to be huge for a bunch of firms. Yeah. So you're like, there's just this sense that like we all know that AI is going to change the world in so many ways. Is we all know that space is big, defense is big, nuclear energy is big. All these things are big, and therefore, more and more money into these things, and they just keep getting marked up. And the last time I felt like investors had this much of a sense of inevitability, like, yes, it's expensive, but it's going to get marked up in 6 months, so therefore, we should do it, was the summer and fall of 2021. And we all know how that ended up.

Yeah, I want to talk about what's similar to 2021, what's different. But first, there was something— I don't remember exactly who said it on that podcast that you did with Jack, Dylan, and Trey, but there was this concept of when a firm has one of these huge power law wins like a SpaceX and an Anduril and Anthropic and OpenAI, then the liquidity is coming. And there— the phrase that was used on the podcast was playing with house money. Does that mean like literally recycling or just your LPs are more excited to back you, write bigger checks, like the purse strings are looser? Like, what does playing with house money mean? How does it feel? What are the risks?

Yeah, a useful analogy here is maybe like in sports. So like, let's say you are, you know, up to bat in baseball in the World Series. And you hit a grand slam. And then the next, the next at-bat, you hit like another grand slam. Um, and so you're like, okay, I've hit two grand slams in a row. Basically, whatever else I do for this game, like, I did my job. Like, I'm good. Like, I'm great. And, um, counterintuitively, honestly, that might make you bat better because you're just swinging free, you're swinging away, you're, you're taking a lot of risk. You're like, screw it, I've already done my job. I've already, you know, contributed to my team and scored a bunch of, a bunch of runs for my team. So like, I'm just going to swing away. I think that very much is the case around the industry where if you have a lot of exposure, especially in Anthropic, but also in OpenAI and a few other companies, a lot of these firms have sprinkled their exposure across several different funds. So it's not like you invested in Anthropic in one fund. Yeah, like, I know funds that have Anthropic in like 8 of their funds because they're like, this is like, this is the golden goose, the Masa golden goose, you know, all the golden eggs. So we're going to put it in every fund. And so then now they're like, look, all of our funds look great. They're going to look even better when Anthropic IPOs. So like, let's take some risk. Let's like, you know, let's swing away. And honestly, it might do them well. But that's the whole concept of playing with house money is like, Anthropic's going to buoy so many fund returns across their portfolio of funds that you might, you know, you're like, look, look, we don't need to play super conservatively from here on out.

And, and that feels like something that is uniquely different than 2021. 2021 felt like much more of a broad-based bubble in the sense that there were power law winners at the time, but you didn't have the same effect of if you've spread this one company across all of your— maybe I'm just not remembering, but I don't remember it that way. I remember it much more like there are, there are so many SaaS companies that are going to go from $100 million ARR to $2 billion, $1 billion that we can underwrite them all even though they are competing. And then there's this AI wave that's coming. But what do you remember about 2020, 2021 that was similar or different? It feels like it was a little bit more driven by spreadsheets and actual growth math as opposed to a major technological shift. What was that era like for you? It's funny.

Yeah, you still had like, obviously, you didn't have the extent of IPOs of like a SpaceX and OpenAI and Anthropic. But you had like DoorDash, you had Airbnb, you had Palantir. So you did have some liquidity, Nubank, a few others. So you had like, there was a lot of money that was made. But you're right that it wasn't to the extent of Anthropic where you're like, oh, we had Nubank, so therefore, like, we can like go to the beach and just like take a bunch of risk.

And Palantir in particular was, I mean, what, $10 billion IPO and not really sprinkled across many funds, very concentrated and sort of a black sheep of venture for a long time. I don't know if that's the right term, but yeah, you're totally right.

Like the actual IPO, like, you know, one, it didn't do well. Yeah. I think the share price was like stuck at like $8 for like a year or two, or 3 years. It took a while for it to take off. So you didn't have nearly as much money. It's funny, I actually went back. I remember in, I think in like 2024, yeah, '24, so a couple of years ago, I was like, I want to go back and like read, like, what were we doing? Like, were we all high? Like, what was going on? Like, were we just drunk?

Yeah, there was this guy who wrote this whole piece. About aggressive crossover funds coming in. That was a crazy moment. That was a crazy moment.

But I was like, okay, this is— if I was to go back and read investment memos and put myself in my mind then, and not me, but just the industry's mind, was this all rational? And I think that the really tough part about 2020 and 2021 was that from the last 2 decades before then, all of these trends that we were investing behind were very secular. So if you, like, look up the e-commerce penetration rate as a percentage of total commerce in the US, it's like the most straight linear line you've ever seen. But then 2020 happens, we all get locked up, and there's like this insane acceleration. And so tech investors were always, you know, taught and trained on the fact that, like, growth kind of goes one way. Like, there's— it's not cyclical. It's not like up and down. And so you look back at a lot of these investment memos and a lot of what people were thinking, and the companies were doing unbelievably well. Like all of these SaaS companies were growing like, you know, 200%, 300%. They had really good fundamentals. The cohorts looked good, like customers were expanding, they were staying like that. All the fundamentals were really, really good. And then of course, you had an overlay on top of that, which was like the public markets were pricing SaaS companies at like 40 to 50 times sales. Um, that's the craziest thing to me. It's like you look back at, at like, you know, the Snowflake IPO. Yeah. And you're like, I, I was, I was reading the S-1 and I was like, the thing was at like less than $500 of revenue, went out at like $80 billion. Yeah. And I was like, okay, there's some crazy stuff that's happening right now, but we are like, we have not seen anything, anything like the tops of the SaaS 2021 bubble. When like Snowflake was going out at like 120 times revenue on the public markets.

How much— so there are— how much multiple compression that's happened? Is venture capitalists actually learning their lesson, the market learning their lesson, or just purely interest rate effects?

I think it's actually just the compression in multiples that we see now is people are just more scared of the business models. Like, obviously, you know, everyone's saying like, oh, SaaS is basically dead, like terminal value concerns, all those things.. But when you think about all the most popular business models and the most popular businesses that are getting funded right now, like there's no, there's no precedence on public markets of like an AI app company. Like we don't know what like an AI app company will trade at. We don't know if it'll trade much better than SaaS, moderately better than SaaS, or the same as SaaS. We don't really know how like a space company besides, you know, with the exception of SpaceX, which is like an exception, not the rule. We don't know how like a space company will trade. We don't know how—

with that one, what about like ASTS, Rocket Lab? Like, there's a couple of public comps in like the pure-play space area. No, that's true. And less on launch.

I more mean like, you know, satellite companies. Sure, sure. Or like— and yes, you could say the same thing about like, yeah, with Anduril, there's obviously a lot of primes. But like, God, I hope it doesn't trade like that.

Yeah, yeah, yeah. Much more like Paladin.

Yeah, that's the thing where it's like, a lot of the things that people have been investing in either don't have precedence. Sure. They have something about them, like lower gross margins or capital intensity that's scary, or the incumbents trade really poorly. And so, you're like, well, this is kind of like when Affirm went out, I think in 2020. No one knew if it would trade like a bank or like a payments company. Yeah. And it sort of traded as a hybrid of the two, but no one really knew how to think about the company for a few quarters until it's public. So I think lower multiples, mostly because people are just discounting because they don't really know how these things are going to trade.

Sure. I want to talk about closed source, open source, just token maxing and the economics of AI right now. But I want to go back to your January 31st, 2024 post. You said making a real effort to not take for granted the, quote, $3 Uber, Uber across town era of AI. And I hope you are too. That feels remarkably prescient, extremely true. I didn't realize— I think the last time we talked about it was around just the models getting expensive, around reasoning and the gross margin changing. But now it feels like this is like directly targeted at CFOs of like large companies who are showing up with— I mean, we have Eric Lyman coming out on the show. They had one— their AI spend hit $1.5 million in a single week. And so, yeah, we are out of that era. But as you reflect on that, was that what you were predicting? How were you seeing opportunities across new companies in the era of like cost control and ROI maxing, how are you processing that idea of like the end of the Uber era? Because in many ways, there's still a knockout drag-out fight between Codex and Claude Code, and they're resetting limits every 12 hours, 6 hours, they're fighting it out. Like we're still sort of in the capital fight, but at the same time, we're also in the, some enterprises are really controlling costs now.

Yeah, there's, oh God, there's so much to this topic, and there's just so many different things. Things we could talk about. So I'll try to hit the best portions of it. Yeah, the tweet, the tweet, I think there's like a consumer part of the tweet and then an enterprise part of the tweet. And like the consumer part of the tweet was like, there's no ads, you know, the LLMs don't want anything of us. They're still super raw. It's just this like silly little service that like clearly is going to need to mature into an adult, you know, cash flowing product one day. But right now it's not like you could also compare it to like the old days of Instagram. Instagram. Like, when Instagram didn't want anything from you, um, it was amazing. It was like the best app ever. And now it wants money from you in the form of like you, uh, you know, being an ad unit. And so all you see is short-form video, and it's like, you know, rot, you know, and like brain rot. Um, and TVP and videos, those are, those are the ones that save. And then on the enterprise side, there was all this, um, you know, that there was, um, reports that sometimes it was like, man, cursors, it's really hard to compete against against something like Cloud Code because they would compare the basically amount of cloud usage you could get through a subscription plan with, with, with Anthropic versus API via Cursor. And you'd get like 20x the usage with Anthropic. Yeah. Do you see insane subsidization? Yeah. On Anthropic. And I think both just the incremental adoption of AI by enterprises and that starting to like move down a bit. I still think there's like a fair amount of subsidization., but it's like starting to go away little by little as, as these, as just like the token hungriness of the models gets much higher. Yeah. Like now that we've done like, you know, we've moved from like, again, just ChatGPT to long-running agents, that are like way token hungrier because they're reasoning models. Um, and all, and like you, that use subagents and all these things. Yeah. The costs have just like absolutely ballooned. Um, and, and so I like, on one hand, On the one hand, it's very clear that we've exited the $3 Uber stage. On the other hand, there's folks like Dylan Patel at Semi Analysis, and there's some of these really forward-leaning companies that are like, not only are token costs reaching our human labor costs, but we hope they go over. We hope all of our competitors use open source models, and we hope all of our competitors use dumb models because our employees will be using the frontier intelligence models and like that's our competitive advantage. And so I don't think there's like there's, there's a lack of clarity around all of this that I think is again, part of the disorienting thing about even investing in AI. I will say we're seeing an immense amount of app companies. I think if an app company like app companies are either already at 80% plus open source usage, for like their own apps that they're, that they're serving to, to their customers, or like 90% of them are trying to get to that ratio plus. Um, so a lot across our portfolio and beyond, a lot of people are working with folks like Fireworks. Yep. Um, to, you know, custom train and fine-tune. Yeah. Um, these increasingly good open source models. Mm-hmm. Um, because, and like Jesse at, at Decagon actually had a really good post on this where he talked about just the fact that like like for mature use cases that can be— they're well known and can be fine-tuned on. You just don't need frontier intelligence anymore. And oftentimes it's best for these app companies to fine-tune open source models around those specific use cases.

Yeah. Are you finding with the application layer companies internally they're adopting the semi-analysis approach? They're using the frontier to build their tools, but then if they're vending out tokens, they want those to be very efficient because that scales with their user base, not with their employee count.

Yes, I think that, that is probably completely correct. I think it's like an extremely good take. Yeah, where it's like, yeah, like for— and, but the only reason I think of that is because what are they using it internally for? You know, like, like they're using it for coding and like they're using it for— and like these are, these are high-velocity startups where like the number one thing that matters is high-quality shipping velocity of software. And so, um, if they can get a competitive different, uh, if they can be differentiated competitively by shipping faster by using Fable and just like spamming, you know, Fable the whole time, they're gonna do that. Um, but then the thing that they're selling their customers is, is a very different use case. Yeah. So if it's like a support, you know, agent, totally, um, you can use open source for that. But like the frontier intelligence on coding is still, you're seeing a lot of incremental gains for that.

Yeah. Yeah. The frontier labs are the spikiest on coding, like that Thinking Machines example from Bridgewater felt like a unique spike in research and news analysis that might not show up at a Frontier Lab because it's maybe a smaller market. But Thinking Machines, they'd like bring that to bear, fine-tune something that outperforms everything at a lower cost. What else are you sort of retreating to in the age of the application layer, are you more likely to look at two-sided marketplaces, network effects, something with like the other sources of power from zero to one? If just like big pile of code is not defensible in the long term, like it was maybe a decade ago. Yeah. Yeah, what are you actually looking for? And then are you actually seeing entrepreneurs acknowledge that and then go and build that. Yeah. Yeah.

I mean, I will say, like, seeing a nice at-scale marketplace with clear network effects, it's like a, it's like a drink of water in the desert. It's like, oh my God, you know, like, I don't have to worry about AI labs or, you know, business model quality or any of these things. And so it is a breath of fresh air because I feel like, you know, those are always in style, you know, these like super 7 Powers, Modi businesses that somehow avoid exposure because maybe they're consumer marketplaces or something. So I absolutely think that's the case. I'm also— I've always been relatively less or like not really in the bearish camp around lab risk for a lot of these app companies. You know, one, like, one, one, like, silly example could be like, you know, some people like, oh, like, you know, Cloud for legal is coming out, like Anthropic's going into legal, like, oh, legal space. And my response has always been like, look, like Anthropic in like 3 months of 2026 probably added the amount of like near, like even medium or long-term TAM that exists in AI legal, which is still like a ton of revenue. Like it's still an immense amount of revenue. But like, you're telling me that they're going to put like their, like their SWAT team to like grind out what is a top-down sale market over like 7 years to maybe add some portion of the ARR that they added like in Q1. Yeah, like it just doesn't— like it just doesn't make any sense at all in terms of like the highest and best use of like the lab's time. Yeah. And so I think all of these markets, like, again, like people need to think about like, well, what are like— what are the theoretical competitive risks and what are like the the practical competitive risks. And like throughout SaaS, I'm sure like if, you know, Microsoft at one point was like, we're going to win this extremely niche market for X at any given point, and it became like the number one project for the business, then they could go do that. But they didn't do that because they had like Microsoft Office, you know, doing tens of billions of ARR and like their security business doing tens of billions of ARR. And so it's all these things around prioritization that I think are— that I think people don't think about. And so I'm not that bearish on the app layer stuff. And then the last thing I'll say is that the other really interesting thing that I'm seeing is that there are all these like short-term things that I totally get why companies are doing them. And I think they like kind of have to do them, that in the long term, we're going to look back and be like, this clearly made no sense to do long term. The biggest one that I see is like like every app or not every app company, but like a lot of these app companies now feel like they need to have like a labs team where it's like every single app company, you know, they'll hire like, you know, a few researchers from Meta or something. And then all of a sudden it's like, well, you know, we can defend ourselves from the labs because like we also have a research team and we're doing like, we're doing like AI research. And it's something that I think we'll look back at and like in some cases it will have been valuable. But in many cases, I think it's just like a way for founders to be like, oh no, look, like we also have researchers, like we don't have risk from the labs, like we're doing our own research. And I just think that like there's not, there's not that many opportunities for like an in-house research team to be doing that much groundbreaking work, you know, relative to like what, what the actual best researchers are doing within all the places that have the most GPUs, which is the frontier labs.

Yeah, it's interesting because labs Labs can mean a few things. It could mean AI research or it could mean experiments. And if you're an eBay— This isn't Ramp Labs, by the way, because Ramp Labs is doing like ultra experimentation. Yeah, very different case. But there's a world where you're eBay and you realize that the labs aren't really going to steamroll you because you have this liquidity and this market and this network effect. But, you know, just figuring out the right way to integrate AI and maybe it's not just stuffing a chatbot in the corner, maybe it's something a little bit more polished and having a team that can go out and look at the full product surface area without needing to be like the top-down AI mandate of like every feature needs to be AI enabled is probably the wrong pattern. But letting a team go around and say, well, yeah, actually we don't We don't need to add AI to the checkout flow because we want addresses to be deterministically verified. But in terms of descriptions, if somebody's asking about this product, throwing an AI summary there might make sense. And we're going to use an open source model for that because it's going to run on millions and millions of product descriptions or something like that. Tyler, I want to give you a chance to ask a question if you have anything.

Yeah, yeah. I was curious, how worried are you about about like the massive dependence that open source has on China. There was this article, I think a week ago, and it was something to the extent of like Beijing is looking at curbing overseas access to Chinese top AI models. Almost all of the Western open source models seem to be like quite reliant on the Chinese models, which in some sense seem to be reliant on American closed source. Maybe it's just a circle, but it seems like that's like If a lot of these app layers are just training their own models, they're doing fine-tunings on open-source models from China. Yeah, from China. If those go away or they stop accelerating because China says no.

Yeah, yeah, yeah. 100%. It's funny, I am on a podcast recently. I just for some reason can't prevent myself from saying spicy takes that get people mad at me. And we were talking talking about open source models. I think it was on Harry Stebbings' pod. And I was like, where are the good Western models? Like, why? Like, all— like, and some— and I'm like, we don't have any good open source models. Then everyone was like, we have so many good open source models.

Jensen is done.

This company and this company and this company is like good. And then I didn't because I restrained myself, but I just wanted to tweet back the OpenRouter token rankings. Yeah. Then where are If they're so good, why don't people use them? Yeah. Why do you— like, look at the top 10 on OpenRouter. Yeah. Like, all of them are Chinese. All of them. And so I think that— I think that the— but like, a good pushback to what I said on that podcast was like, well, they're coming. And I've always been like, well, where are they? But now we actually are starting to see some American teams like actually really go hardcore at the opportunity. And so like, one, NVIDIA obviously cares deeply about there being a strong open source ecosystem. They've done more than any single company across the entire stack to fund and support and like bring into existence an awesome open source ecosystem. And so I think we all owe NVIDIA and Jensen a debt of gratitude for pushing so hard for a healthy open source ecosystem. So they have NemoTron now. NemoTron, by all appearances, is like a great start. And I know a lot a lot of people in our portfolio that are actually quite bullish on the future of Nimotron. Um, Tinky, just Thinking Machines Labs, just came out with Inkling, um, which again, a lot of people are very excited about. Um, they, they don't claim that it's, you know, at the frontier even of Chinese open source yet. Um, but again, it's, it's extremely customizable, which people love, enterprises and app companies love. Um, and, and again, it's, it's a commit, you know, seemingly a bit of a commitment to continue to develop open source models. Reflection AI has always been, you know, that they now have their strategy is they— I don't think they've released one yet, but like their strategy is clearly to, you know, have the American open source model. And so I think relative to even 3 or 4 months ago, there's more that you can point to around like, okay, like the West is actually trying to do these things in terms of like the nested contingencies of like who's distilling who and like where are all the model capabilities coming from. I think, like, I think the whole distilling thing is like, I think it's oversimplified. Like, I think it's like a little too simplistic. I think it's also like a little, like maybe there's some like xenophobia in there where it's like, um, in the public where it's like, well, the only way that the Chinese models are good is if they're just like distilling. Sure. And they're doing nothing else.

Yeah. It's like, well, they clearly cook some stuff up.

Like they clearly are doing some really good creative things. And like the DeepSeq paper was like truly groundbreaking. Yeah. Like it was awesome.

And you can see it because like open source architecture is get ported back into foundation models and closed-source labs all the time. And also you just see that there's a ton of successful AI researchers who come from China and stuff. But yeah, I think that's all a good point. Yeah, I have another question related to the application layer. I'm interested, like, venture's been through these like sort of experiments, whether it was like biotech for a little bit, D2C e-commerce was sort of enabled by Facebook and then eventually D2C didn't die, it just sort of returned to the, the better fit was like CPG private equity firms know how to do it, they still do it. There's still some amazing outcomes usually in like the $500 million to $1 billion category, you're not seeing trillion-dollar CPG companies anytime soon. But is there, are there Are there any areas in the application layer that you're seeing, oh, this category is now potentially investible as a venture opportunity because of AI? I'm thinking like game studios or something else. Or is there a pocket of previously venture-backable companies that maybe should be moved more over into, hey, just bootstrap that, get it to scale, do some private equity, secondary, run it cash flow positive, positive, more of like a lifestyle business?

Yeah, yeah, I think there's a lot of both. I mean, even like if you think about it, even, even in some of the more obvious verticals where we've seen early winners, like even in legal, right? Like legal before AI was not seen as like a venture category.

There was Atrium and Clearspire, two like really solid runs at that, the tech-enabled law firm. And, you know, rough goes on both, on both accounts.. And now it's like, oh, the money's flowing. The business looks like a normal tech company. Yeah, there's still like margins and whatnot. They got to pay for tokens. But like in general, it looks a lot more like a tech company than a law firm for sure.

And even the SaaS companies that sold into law, it was always just like, oh, super constrained TAM. It's a slog. You can't get law firms to pay a lot. And now, you know, Legora and Harvey have just like absolutely eye-watering numbers that, that like, you know, for the last 3 years now and clearly are among like the very—

and Harvey just acquired Benchmark, right?

That was a jump scare. What? What?

Yeah, that was probably a good strategic move, but also just a hilarious troll on the timeline.

But on the PE, like on the, on the PE, on the, on the other side, there's like obviously we're in record and they become this is a huge, awesome platform for the task economy and finding the data that actually now moves these models forward. There's so many— if you think about at the limit, if the limit to getting AI agents to be able to do everything is to find all the data in the world and feed it to them in a really high-quality way, there's so many people I know that are finding— going out in the world and finding extremely niche or just data that no one would think of, and then making it high-quality servable to labs or anyone that wants to buy that data. So, there's one, for example, that's basically instrumenting a medical clinic. So, every conversation is recorded, everything that they're doing is video recorded. They're almost doing like a— yeah, like putting telemetry throughout about every portion of a medical clinic, and then like making that a dataset. And so I feel like, like, again, like, is that venture scalable trillion dollar? Probably not doing it if you're like, just like the medical clinic data company. But I think there's gonna be a lot of like entrepreneurial people that make a lot of money, just bootstrapping these things and building them to, you know, $50 to $100 million revenue businesses for like 8 to 10 years. Yeah. And like, that's, that's sort of all you need to do. Yeah.

You can kind of like entrepreneurship broad Absolutely. Like, it's just, it's just an exciting time to be building. Well, thank you so much for taking the time to come chat with us. Have a great rest of your day. Have a great weekend, and we'll talk to you soon. Thanks, guys. Goodbye. Yeah. Let me tell you about Public.com. Investing for those that take it seriously. They got stocks, options, bonds, crypto, treasuries, and more with great customer service. Our next guest is TPN royalty. We got Eric Lyman from Ramp. He's the co-founder and now the co-CEO. He's in the waiting room and we'll bring him in in to the TVP UltraDome. First time we've talked to him since he's co-CEO. How is it? How does it feel?

I feel so good to be back.

I feel so good to be back. I missed you too. First time chatting with Tyler directly. I think he's popped in a few times. You've, you've chatted with us in every different permutation, the yellow suits, in person, in the New York Stock Exchange, all over the place. But how are you settling into the new title, co-CEO?

It feels good. It's what's been so fun. I've worked basically this way together for 15 years.

Yeah, I was about to say everyone's like trying to do like hot takes around it and I'm like, have you actually met these guys? They've been co-CEO the entire journey, even going back before Ramp.

It's, it's been fun. Like I like, like for me, like part of what makes this fun is like for we've known for years, like Kareem is the secret weapon of the company. Yeah, he's driving so much of what's going on now. It's like more obvious to people of like we got at least two of us. There's, there's actually way more interesting people at the company. But no, we're, it's, we're moving fast, growing faster this year and it's just fun, you know, now it's he can swap and take some events off my hands too, which is great. Yeah, I love it.

Well, businesses on-ramp are moving tokens fast through their systems. It's showing up in their books. You built a website, token-spend.fm. I love the .fm. I think that's a very fun TLD. But what inspired this? I imagine it was from like direct conversations with your customers. Did this come internal? What were the findings? What were the goals of the project?

You nailed it. Like, over the last year, the last 4 months alone, ramp customer spend on tokens has grown by 21 times. Wow. 21x. Like, is that a gong or is that a wah wah?

It's a little bit of both.

It's some of every— depends who's making the money or spending the money. And look, by the way, like, the crazy part is it's not like people are like, turn it off. It's like, no, I actually want to spend more but on the right things. And so like, you know, so we saw this from our customers, we saw this from ourselves. Um, you know, a few years ago AI spend was a routing error, like error to— yeah, I think in May it hit almost 10% of our payroll spend. The equivalent was on tokens on payroll. And look, there's great things you can say about it. We're launching products faster than ever. We're more efficient than ever. We're growing faster. And yet, you know, you— I hate to say it, there are people at the company who like use Fable to like look up the weather. That happened. Yeah, we heard about that at other companies.

I can't believe I'm teasing you.

But no, it's a— there's this whole thing where people know and get in the abstract of models even from a year ago were amazing. Are great at doing tasks and we can be more efficient. Two, accounting teams need to be able to monitor this. Like it's, you know, I think of our CFO who would get a bill for hundreds of thousands of dollars and then the work began of like, you have to allocate some to engineering, some to sales and marketing, some to you name it. And so there's just so many products that weren't built by the labs. And so today's launch around token spend management is a place where any company, whether or not you tried Ramp, can, can, can link up your API keys and then you're good to go. And, you know, we're helping people really within minutes start cutting their spend by several percentage points. And so it's been a very fun launch. Yeah.

What was the, what was the first sort of generative AI application in RAMP. Was that the GPT API for understanding receipt data?

So this was actually— so that was our first ML model for sure. But the generative use case, um, this— I think it was 6 months before ChatGPT came out, we hopped on GPT-3 and we started using this for, for, uh, um, go, go team!

Yeah, classifying different receipts, putting things in the right expense category, things that could be done deterministically. Fuzzy logic could apply, but LLMs were uniquely suited. But at the time, that was a rounding error. Then you go forward to May and you get a bill for $1.5 million in a single week. What's the actual process for, for untangling what's happening? Are you looking at prompts or are you just going to Slack and saying, hey, you know, you were one of the top 10 drivers of token spend. Can you flesh out a little bit of what you got done this week? That type of thing. Like, what is the correct way for an organization to interrogate their spend, maybe qualitatively, after they're done with the quantitative side? So, great.

Most organizations that are even at a place to be thinking about this, they're using lots of tools, right? They're spending on OpenAI models, they're spending on Anthropic, they're using Gemini. They might be dabbling in open source, they're using Cursor. And so the first step is actually just seeing it. It's being able to link up your keys so you can understand and start to break up basic questions of like what's happening today, not in a month when you go get your bills connecting and tagging that. And that's something you can do out of the box through the product. We come back within minutes to help you understand it. And so you can start to go and see, okay, not just this person normally spends $1,000 a month, but they've already ran through $800 in an hour. And so you can set up notifications. We show you unusual spikes. And so again, think back to the early days of ramp and corporate cards. A lot of this was alerting and visibility. It's these types of insights. And you also see things like, you know, we know how the most efficient companies are running. And so If you aren't cashing, you are overspending. Some people will leave fast mode on, which can be multiple times more expensive. And so we'll just highlight that for you. Sure. And then finally, at the end, you get to controlling it, you know, acting on it. And, you know, I think for today there's so much to do on the analytics itself. But I do think there's more sophisticated opportunities to be had, whether that's in in small model training, in routing, and much more. And we're excited at the whole space. I think there's so much to do to help companies save.

Okay, sure.

Yeah. So I see the average company, 59% of their token spend is on frontier models. How do you— how should companies be thinking about allocating between frontier models and open source? Is it almost a thing where like you're in an explore phase using the frontier models models, you're doing these kind of net new coding tasks, whatever, and then once you find this repetitive thing, you're running the same process every day, every week. Is that when you allocate it towards open source? Maybe you're even doing a fine-tune. How should people be thinking about that kind of stuff?

It's a perfect question. And I think for companies out there, if you're listening to this and haven't used these models, I would say using frontier models just to get a feel is good. It is surprising the capabilities that models have. And, you know, if you, if you don't have a multi-thousand-dollar-a-month bill, like, to start there, I think is, is reasonable. But then you start getting into optimization, and there's two— there's a few sets of interesting questions. One, um, there are cases when using frontier models can in fact be cheaper. Like, for example, in our own benchmarks, um, on our software engineering benchmark, people think of Sonnet as an older model in, let's say, the Anthropic world. And it is cheaper per call, but it needs to think a lot harder and call more agents working in collaboration. And actually the smarter models, like it's kind of like you met someone smart, like they don't need to go and do like long division. They can just like do division in their head. There are cases where using— for example, Scott Wu calling Scott Wu is much faster. His hourly rate is much higher, but there might be certain things, you know, that Scott Wu per second might be cheaper than, you know, hiring a team of—

certainly the bad venture capitalists have made. I am like, I want one Scott Wu instead of 1,000. This is right.

So that's part one. But then the really interesting part is when you get down into benchmarking of what is the nature of work that you're doing, you know, you can use these trillion parameter models to answer really tough questions. But when you start seeing very high production use cases, you'll see, you know, if you know all of the input tokens you're getting are around customer service, you know, you'll see companies like Sierra having their own small models around what makes great efficient. And so they've lobotomied, you know, just the part of the brain that is really useful for those types types of services to if you can dynamically start to route based on the complexity of the task. You can say a small model just for accounting, it maybe is all we need. We don't need to go ask, you know, a model that can allow us to cure cancer and do quantum physics and that kind of a thing. And so in some sense it's on both and you can start to get really interesting answers the more that you can benchmark your own business. And the more that you have high fidelity about the nature of the inputs you're getting, the outputs you're seeing, and then the efficacy per cent on getting to that. And there's so much to build around this area.

Talk a little bit about the shape of AI product development, the AI work you're doing, Ramp Labs, sort of the surface area, the spikes there, because because there's, you know, just using AI to improve the product that is deterministic, write better software. Then there's also AI integrations. Like, I mean, it's such an anodyne feature, but I love the fact that you can open up the Ramp app and just ask a model like, how much did we spend on camera equipment last month? And it'll just tell me. And that's amazing. And I'm sure I could like wire it up to some other system, but like I love just having it there. And then there's also, AI research and harness development, all sorts of work that's happening there, and that can be expensive. But how are you thinking about all the different trade-offs and all the different work within the AI/Labs umbrella?

So on using this, I mean, this is just an incredible technology, as you know. Like, I think about like part of what— like, let's talk about like B2B SaaS for a second. You know, if we must. Yes, please.

You know, you know, come on. I was waiting. I was waiting to—

Yes, we're finally doing it, you know, boys. Here we go. All right. So the problem why most B2B SaaS is awful is companies are complicated and people want different things. And so the accounts payable clerk wants a view. Your accountant wants a view. The CFO wants a different view. One person wants a button. You want to make it simpler for these people, more advanced for another. And How do you deal with this? Well, turns out generative interfaces where based off of who you are, how using your product, it can show you the interfaces you need with the views, the graphs that you like to do your work can be intuitive. And so it's, it's very interesting in making products that are very powerful but feel simple and relevant for the products. And so on one side of it, like forget ramp, I just think organizations in spending time around dynamic and generative interfaces, there's I think you can just make better computers. You can make better tools for people. Why do we care as a provider of services? What is so different structurally about spend on models and on tokens and on software is you could just hammer Salesforce all day. You're not going to get like a bigger bill from, from Salesforce. Sorry, Colin. Colin, I'll call him back.

Conference, man.

Yeah, you're on. I'll save it for a minute. But to go a bit deeper on it, if you start going and using, as folks know, in the token maxing era, lots of tokens, like, your bill can go from $1,000 to $10,000, $100,000 to $1 million, like, very quickly if you start going.. And it's not like payroll spend that companies are able to manage in some way. It's not like normal vendor spend. In some sense, it's like an untapped corporate card that you can spend as you go. And by the way, the meter is not running. Yeah. And you don't see it. And this starts to feel a lot like what we've obsessed over for years of can you help people manage every incremental dollar far better? And so it's pulled us deep into answering the question of if you want to know return on investment. We should be thinking about like, what is return? What did you buy? Can you understand just the semantics of it? So what was the output to the efficiency as there's more types? So it is— we're obsessing over this from every floor of the building.

One of the, one of the interesting things about Ramp is, you know, everyone who has a Ramp card in an organization has, you know, they can open up and see all their transactions. They can see their budgets, their limits. The policy that applies to them. And they, you know, they can implicitly understand that, you know, if they spent $10,000 on their card or $1,000, like, did they deliver that much value to the organization? And I'm wondering if you see either Ramp or just generally businesses need to push more of that data to the end user, the token consumer, and how will that evolve? Instantiate?

I'll give you an example from like the just card world. And we're seeing this already in, in, in spend management. So in the card world, there's this concept of just like out-of-policy spending. Like let's say that you— let's say you never got a notice and you spend on Uber Eats because you forgot to switch over the card on the weekend. Sure. You might keep running it up and never really realize it. And it turns out if you just tell people once, hey, that was out of policy, You see spend and out-of-policy spend just drop. You tell people like you weren't supposed to do that. They don't do that. You know, people actually want to be good and do the right thing.

And this is the checking the weather with Fable 5.

Yeah, it's like I didn't realize I spent $100 to check the weather. Yeah. And, you know, sometimes some subtle UI like this and feedback goes a long way. And so in the product we've already seen Actually just exposing people. Here's your AI spend. Like, like, like John Tyler. I don't know if you guys know what you spent on token spends yesterday, but on Ramp you can know. And it turns out when you see it, you start getting more efficient. And so even just the act of exposing it to you is driving down these savings for people. So you nailed it.

There's so much there. Yeah, we probably spent a lot on tokens yesterday. We vibe-coded a bunch of really, really jokey sites that provided a lot of, a lot of laughs and a belly laugh. We had Sagar and Jetty on the show who, who does not like prediction markets, and we built him an entire prediction market for his entire life. And the belly laugh that he got from that was priceless. It's all play money, but I think it was worth it. But I actually haven't seen the token mill, so I don't know. It might have been, it might have been rough. I want to talk, do you have another couple minutes? Of course. Okay. I guess last question for me. The Ramp Econ Lab, that is a separate lab. I'm very interested in the knock-on benefits of Ara Kherazian's excellent work studying the economic impacts all over. I mean, the research has gone to the front page of the Financial Times, the Wall Street Journal, so many other places. What has that unlocked for you as, as CEO in conversations with customers? What have been like the knock-on effects of that, that project?

At first, it's just you kind of just understand the world better. It's very obvious in X. It's been this, I guess to paraphrase Elon, like a supersonic tsunami where it's gone from didn't really exist years ago to perhaps 1% of the United States GDP in the next 12 months, like it looks quite likely at this point. And this data is going so fast, like you look at most economic indicators and you get things a quarter later. It's not measured precisely, whereas Ramp data, you know, we're, we're tracking about 1% of all corporate spend in the United States. And you can just see it and you can understand it and you can adjust your strategies faster. So I think it actually helps people better run their business and not get left behind. So I just find it useful for finance teams and technologists, people building businesses. Beyond it, it's just grown awareness. You know, we're competing against some of the best-known brands ever created. And, you know, if we're going in and trying to win you over and say, you know, try our tools, you should trust us to move money, store money, help you get more more from every dollar an hour and you've never heard of us, it's just a much harder sell to, oh yeah, I saw Ramp in the journal on TVPN, you know, on the Founders Podcast or, you know, more and more it starts unlocking it. And when it shows up in a way where it's actually already providing value to you before we have a conversation, it's— you can get deeper much faster. And, you know, I think over the long run that that leads to more growth.

Growth. That makes a lot of sense. Well, thank you for taking the time to come chat with us. Congrats on the launch of the website token-spend.fm. Go check it out. Optimize your token spend today with Ramp's latest project. Have a great rest of your Thursday. Have a great weekend. We'll talk to you soon. Goodbye. Let me tell you about Figma. Agents meet the canvas. Your AI agents can now create and modify your Figma files with design system context. We have a couple interesting op-eds I want to take a cruise through in The Wall Street Journal. The Bear Case for Malibu. This is interesting. For decades, this 21-mile stretch of coastline served as one of America's most durable expressions of wealth, operating— offering residents ocean views, private beach access, and the opportunity to spend large portions of the day on Pacific Coast Highway. That proposition is beginning to face scrutiny. Rising insurance costs, wildfire exposure, limited restaurant options, and the difficulty of completing basic errands have weakened the case for full-time residents, according to property advisors who are familiar with the West Side. Quote, "You are paying $18 million to live somewhere that makes buying toothpaste feel like regional travel," said Graham Pelt, a partner at Royal Property Research. Several homeowners have recently shifted their primary residences to Brentwood, Montecito, and Pasadena, retaining Malibu properties for occasional use. The beach remains attractive, but the emerging bear case is that visiting it may be sufficient. Interesting. Interesting. This is— this is the wrong— I feel like this is the wrong day for Jordi to step away from the show because, yeah, this is, uh, this is not good news. I feel like he would— if you were here, he would be defending Malibu, but yeah. Unfortunately, he's not.

I have another— I guess we just can't really— I think I mostly agree with what I'm reading here.

Yeah, yeah, it makes some good points. There's another interesting op-ed. This one's in the Financial Times. They're calling it the end of Bottega Veneta. Bottega Veneta's position as the preferred label of consumers seeking to display wealth without displaying a logo is showing signs of strain. The Italian fashion house's woven leather bags, oversized footwear, and muted branding helped to defined the quiet luxury era, but their growing visibility has made the products less useful as signals of discretion. Quote, the customer bought Bottega because only certain people recognize it, said Isabel Marchand, an analyst at Mode Capital. The problem is that now everybody recognizes it. That's not good. Stylists say some younger shoppers are moving towards vintage accessories, smaller Japanese labels, and tap-out t-shirts— familiar with those— uh, whose provenance cannot be immediately identified from across a restaurant. Bottega remains a significant luxury business, but its cultural difficulty is that the absence of a logo has effectively become one. Interesting. I feel like Jordi would push back on that too. But there's one more we should go through. Apparently, the G-Wagon is losing its grip. The Mercedes-Benz G-Class, long the default vehicle of musicians, professional athletes, and men who describe routine commercial activity is motion, is losing ground among a small influential group of tastemakers. In recent polling, Future, Jacob Elordi, and Gunna each identified the 2014 Jeep Grand Cherokee as the emerging vehicle. Yeah, I've been hearing this. Yeah, yeah, it is on the come-up, uh, citing its restrained styling, limited social media exposure, and availability with a factory-installed CD player. That's a nice feature. Uh, quote, the G-Wagon communicates the 'has money,' said one person briefed on the survey. 'The Grand Cherokee communicates that the owner has somewhere to be.' That's a good point. Dealers have reported increased interest in low-mileage examples with dark paint, tinted windows, and minimal modifications. Mercedes remains, remains dominant among conventional luxury buyers, but among chads with motion, according to the study, the status hierarchy is shifting. Authenticity now requires cloth seats, a loose headlight, liner and at least one dashboard warning light.

So yeah, and I think we found this, this graph of different tastes in cars, how they've—

yes, can we pull this up? How overall taste in cars has shifted from 2025 to 2026. Looks a little bit like polling numbers, and we'll see the Mercedes G-Wagon has dropped from 95% approval to just 45%. Meanwhile, the G— the 2014 2014 Jeep Grand Cherokee here has jumped from 33% to 97%.

Yeah, I mean, certainly among my friends, I think the 2014 Jeep Grand Cherokee has almost become kind of like the aspirational car. Sure. This is like, you really know you've kind of made it once you're riding in one of these. Yes, yes.

Other big movers, the Nissan Murano Crossover Cabriolet, of course, went from 65% to 83%. The Dodge Challenger Hellcat Scat Pack has moved from 45% to 74%, eclipsing the G-Wagon as the must-have vehicle. And the 2016 Honda Accord, a strong showing, going from 22% to 54%.

Yeah, I mean, over the course of a single year, that's, that's a pretty big jump.

Yeah, it seems like if you, if you, if you got a G-Wagon, you could get a 2014 Jeep Grand Cherokee and a Honda Accord from 2016. Yeah, actually, that might be the move. Well, we'll keep following the story and we'll see what Jordi has to say about all of this in the future when he's back in the TVP and UltraDome. We had to take a shot at him. We had to have fun. Um, no, of course we're, we're joking around with Jordi. Uh, let's go back to the timeline because there are some—

you're going to talk about the, the DoorDash CLI.

Yes, DoorDash launched a CLI.

Yeah, which people have been asking for this.

Have they? Have they been? Um, DoorDash, I mean, it There's that funny meme of like, "Robot, push the order food button." This makes it like one step easier. Are they imagining that people are locked in terminals vibe coding and that they will want the DoorDash CLI to order them food? Or are they expecting developers to build whole applications on top of DoorDash?

So, yeah, I mean, I think the idea is mostly that you'll have coding agents use the CLI to then order. So presumably they're super locked in. You know, you've had your Codex goal running for 6 hours. It's going to ping something.

Is this bad news for MCP? Like, because they have an API, they have just a web front end that a computer use agent can go and use. Sure. I wonder why the move to CLI. Maybe it's faster, maybe more token efficient. People are trying to be efficient. A CLI means you can drop DoorDash into whatever you already run wired into your internal tools and office catering orders itself. I could imagine office catering if you're starting to build or vibe code sort of like an ERP for your office organizing, you could potentially do that. But I mean, DoorDash shared orders if in an office is pretty seamless. Usually someone who's like quarterbacking the order would just drop a link in Slack Everyone picks what they want in the DoorDash app. It's all linked together and paid.

I mean, I think it's just one of those, like, last mile problems where you've automated so much, but you still have to manually do the DoorDash order. Might as well just—

People are having fun with it. People are having fun with it. Lawrence Jang says, my biggest fantasy is becoming a reality. Jarvis, order $57 worth of Shake Shack on DoorDash. No tip. That's a lot of— that's a lot of Shake Shack. That's potentially too much.

That's actually not that much Shake Shack. Steak Shack is actually pretty expensive.

Yeah? $54, $57? Sure, maybe for a steak. That's got to be like 2 huge burgers and a shake. Yeah, that's true. Well, ModRetro CEO Torin was interviewed by Take Him, and Dylan Abruscato on our team shared his favorite moment from the interview. Tay's Take says, I'm extremely bullish on ModRetro. They have Steve Jobs-like product taste. That will serve them well beyond retro video game consoles. That's interesting. What would they do outside of retro video game consoles? Something new, something to compete with the Xbox. Jobs' original DNA was combining beautifully designed, high-quality products with a simplified, friendly customer experience. After using the Chromatic, I believe, mod, Retro can deliver that type of premium user experience with less frustration and fewer data privacy issues. I I wouldn't be surprised to see them expand into televisions, headphones, and other high-volume consumer electronics. Maybe a dumb TV, which Palmer was talking about when he came on the show. He's sick of smart TVs and their advertisements and sign-ups and all the QR codes and all that.

iPod-like device, music.

I'm very interested in— do you think the M64 will be somewhat hackable in the sense that you could get a sort of a dummy card cartridge and then you could potentially vibe code a ROM that loads onto the ModRetro, gives you the full experience of the controllers because we've been building a lot of web-based games in WebGL. And those are fun, but it just doesn't feel the same when you're on, you know, a MacBook keyboard that's not really, you know, mechanical keyboard. You don't have a mouse. You don't have a controller. I bet— I mean, I imagine with, it's also pretty easy to wire up an Xbox controller as an input to a web-based thing. Yeah, yeah.

I mean, you can just connect Bluetooth.

It's like fairly— Yeah. But there's something— there's something magical about the original sticks of the N64. Yeah. And actually seeing it— like, the constraints actually breed the innovation. You can create some mashup between Mario Kart and GoldenEye or whatever you want pretty easily with vibe coding. It will be interesting to see where the hacker take it, how moddable the Mod Retro is. Uh, anyway, we have our next guest in the waiting room. Let me tell you about Cisco first. Critical infrastructure for the AI era unlocks seamless real-time experiences and new value with Cisco. And we have Jordi Hays from Senra Systems. He's the co-founder and CEO with a huge funding announcement. How are you doing? Welcome to the show.

Doing great. Thanks for having me.

Long overdue. I mean, I heard about this company because because you've been working with Founders Fund for a while, right? But take me through a little bit of the history, introduce yourself and the company.

Absolutely. Jordan Black, CEO, co-founder. Started this company a little over 3 years ago right out of my apartment building wire harnesses on my carpet floor. Yeah. Founders Fund been great investor in us since the pre-seed all the way to Series B, which we just did.

How much was the Series B? Tyler's gonna hit that gob. How much did you raise?

$65 million.

Okay, so explain like I'm 5 the, the, the product, the customer base, what's fueled the growth, what's, uh, what's in the supply chain, what do you— what's your key value add?

Perfect. Perfect. Uh, I brought some props on the show, but, uh, Senra is solving the wire harness problem, uh, in the US for aerospace defense. And you're probably wondering what a wire harness is, or anyone watching this, and this is what it is. It's just like a bundle of wires, connectors, cables. Think of it like your iPhone charger, but a lot more complex. And this whole thing I'm holding in my hands is designed in Excel spreadsheets and PowerPoint slides, and it's It's all done by hand today as well, by really skilled workforce putting this together. And think of it like a Cheesecake Factory, but trying to scale it. It's really complex and there's no culinary schools that exist and no recipes, and you just gotta figure it out and make it happen too.

How on earth are people designing these in PowerPoint? That seems like a crime. That's like, is it— they're drawing out out little diagrams on PowerPoint slides with like the draw a line tool?

They use that. They use Microsoft Paint. Visio is a big one.

People design things on.

This is anything from the large aerospace defense companies to startups too. But it's the input into wire harnessing is not standardized. The output of who builds this, how you build it, what's the right way to build, it's not standardized. And like SendRAS, solving this $165 billion market, uh, by being the ones that says like the bullshit's over and we're just going to take over and standardize the entire thing.

So why not a pure software solution, like just wire harnessing SaaS? There are big companies that have been grappling with vertical integration— the Anderl, SpaceX— like I imagine that you could sell this as a SaaS product, but why, why work on the actual production? That's where the problem is.

I— we started the company even with the design tool, and we got some traction with that too. Okay. But in the day, every company just wants the wire harness, and they want to come faster than normally is. They want to plug in and work, uh, and it's the quality, it's the speed, it's the cost, but they care about the physical product. And I think if you want to make a generational company where it's like, where were we before Google Maps, you know, came out. It's like, where— how are we building harnesses in the US today? It's like we're with Senra 10 years ago. And that's kind of what the goal we want to have. But to fix the problem, we have to vertically integrate the entire thing.

And what's driving the customers to Senra? Is it speed because you're local? Is it forward-deployed engineers that you've sent into organizations to sort of co-design wire harnesses before you make them? Is it the made in America or price or speed? Like, what, what, what are the key factors that jump out?

Yeah, I think just being better is not, is not how we're going to win and how we're going to win every time. But it's like quality is our number one thing. Like, it's really hard to get a high-quality harness because it's all done by hand. So you don't know it works until you plug it in, turn on the rocket, and the wires fit in the right spot, or the wires aren't crossed and the whole thing blows it up. Like, they just recall like a million Jeep vehicles vehicles because like the wire harnessing was bad. Like, whoa, this is a really big problem just in any industry. So quality is our number one thing. And we're over 99% first pass yield with all our customers. And like, you just will never go out of style and this is why you keep coming back to us. The second is going to be speed. It usually takes months to quote something, takes, you know, even over 6 months to even get a wire harness from a customer to our company today. We're just going to be faster. For the customer we're growing with. And the last one's like that forward deployed engineering portion of it. So we do everything prototype to production. So we will partner with the neo-primes, the primes, just any of these companies say, let's be your engineering partner. Let's be your expert and tell you how you should design this thing, how you should be thinking about it, and let's go build you the prototypes and let's go scale the production sense of it. Like we're not just your contract manufacturer. Like here's the PowerPoint slide, go build it. Build it. It's like, here's the PowerPoint slide, let's go design this perfectly, build the prototypes perfectly, and then make this at least your problems in the long term too.

What, what subcategories of electronics or, or products are actually growing their wire harness footprint? I'm thinking of like the smart fridge boom where there was a move to put a screen and a Wi-Fi router in every device in your kitchen, and it feels like that sort of peaked and maybe that's declining. What are you seeing outside of the obvious automotive, space, defense categories that are sort of like sneakily interesting markets or maybe on a growth trajectory that you might not have expected unless you were on the ground in the industry?

Yeah, that's a great question. I think it's, it's, it's maybe obvious, but like the data center. Oh, sure. One. But what's really interesting, it's not just the wiring that goes in the data centers, but the surrounding, uh, infrastructure around it. So like, we need more generators, we need more chillers, we need more HVAC equipment. So you're seeing a really big, uh, inflection point in that sector too. Um, aerospace, on just the actual commercial aerospace, like more satellites. You have obviously defense with everything growing into, um, but anything that requires electricity. So now you're seeing seeing all these new energy companies. So I think the one that's going to start taking off is gonna be like micro reactors or nuclear reactors. Like those are really difficult wire harnesses to build. Mm-hmm. And there's not a lot of people who know how to build that stuff as well. So as people, the new incumbents and the Westinghouses of the world start like, uh, scaling up, it's like we're ready to be kind of part, uh, part of that journey as well too.

What does your path to automation look like? How, how, like, I— how did you make the first wire harness? How are you making one today? What does it look like in a decade?

Uh, it's scaled from me building in my apartment 3 years ago and just going to investors saying, look, I have revenue, I can build wire harnesses.

Um, so you like literally like ordered the parts on Amazon and like put it together basically and just put together, carpet floor.

I got, you know, this, this PO, let's, uh, revenue generating, uh, within the first few weeks. Um, I think the first year was, uh, got a really great team of like, because of my background is from SpaceX, of like really skilled technicians who've been doing this for 10, 20 years and they were building harnesses. Now probably over 90% of our workforce is people who we've trained to go build the harness in 4 weeks versus the number of years it takes to go build that. Uh, so I think that's like the inflection point now of like, how do you scale the unscalable? It's like build a training program, build an operating system for them to go build these harnesses on, and then build as much— integrate as much semi-automation to the factory too. And then what's going to happen the next like 1 to 5 year range is like we have a dedicated team working on AI and robotics of like talking to all the new companies, talking to the existing incumbents of like We are ready to go bring in as much robotics and much automation into this factory. We're tracking all the data from the design input into telling technicians exactly how to do with work instructions to the physical automation data points of like how are they doing things with their hands. And the really goal, I think everyone from the technicians building it to the executive level suite is how do we automate this as much as possible? And I think that's the really exciting part, but we have a really good foundational like we're building this, we're collecting the data, we're working with our customers, we're seeing what their cost points are and trying to grow as much as possible.

A couple of years ago, Daniel Gross wrote a blog post called AGI Bets, and I think the very first one was, is copper underpriced, which I think was a very funny one. He was 100% right. Copper has, has risen in cost a lot. Do you think about raw material input costs as an important lever on your business? Do you hedge? Is this material or are you able to pass it through to customers and not have it affect the core business? I think it's a little bit of both.

But overall, like, we're buying this and we're, we're, we're not making the copper wire or not making the connectors. We assemble it. Sure. That's kind of our point of value add. But I think the difference is because 'cause we're this engineering partner, we want to be a cost-effective supplier for these customers too. So we can go, we have actually our own software that we built out in an AI tool that ends up like looking at the BOM or the bill of material cost and says, this is how much it should be. And then suggesting to the customer, you should go with X, Y, Z other component 'cause it's cheaper, readily available, it'll work for your application. But nobody's telling them this is how you should do it. And that's why I say it's much like a Cheesecake Factory of like, why are we using this really exquisite beef when we can go maybe this over here, which like has the same flavor, or you can't even taste it once you put 1,000 calories of cheese on it or something like that too. So, um, that's really, I think, the value add too is like if the price of copper is going up, nothing we can necessarily change in our control to do that besides having better suppliers and distributors we work with and working with our end customers of choosing the right products for them too.

Where are you based and who are you hiring?

We have two factories, one in Redondo Beach and then one in Cypress, California. And that's— we just opened that up. It's 82,000 square feet of manufacturing space. And we're hiring all across the board. We're hiring in operations. So technicians, engineers, industrial engineers as well, too. Manufacturing people just beyond the the ground floor, building the factory from the ground up. And it's really exciting. It's like there's not a lot of greenfield factories that you say, this is how things should be built, this is how they should be done. This is— we're building the ship as we're driving it too. We're hiring people on the engineering side from the software perspective and automation of like, how do we go at the lever of making this more efficient and more automated in the future? And then we're also hiring us like business development in sales of like partnering with these customers and the program management side of like, you get to go fly these customers, see what the wire harness needs are, see what your product goes into, and like really get integrated with their end, uh, their end goals too.

Very cool. Well, congratulations on the new round. Thank you for everything you're doing and thank you for coming on the show. Have a great Thursday. Have a great weekend. Let me tell you about Railway. Railway is the all-in-one intelligent cloud provider or use your favorite agent to deploy web apps, servers, databases, and more. Well, Railway automatically takes care of scaling, monitoring, and security. I'm very glad that the dog bark, or the, uh, yeah, the bark made it through. Let me also tell you about Shopify. Shopify is the commerce platform that grows with your business and lets you sell in seconds online, in-store, on mobile, on social, on marketplaces, and now with AI agents. And without further ado, we I have David Baszucki from Roblox. He's the founder and CEO. He's been on the show before. Welcome back, David. How are you doing?

Hey, it is great to be here.

Thank you. Thank you so much for taking the time to come chat with us. Uh, I'm very excited about Build. I'm very excited about a lot of other things going on in the Roblox world. But, uh, take us through the announcement today. Uh, it makes so much sense. I think a lot of people could have predicted this happening, but I'm very interested in the shape of the project and how you actually see it changing what Roblox is.

Yeah, the shape is ambitious and at the same time it goes all the way back to our roots. You know, 20 years ago when we launched, we had this mission and vision of removing limits of gaming and having everyone be a creator. The notion of a UGC platform was new. It's interesting that 20 years ago our first slogan was, "You make the game." And so time has an interesting way of going full circle. Fast forward to where we are right now. Roblox Studio has a bunch of models working to accelerate creation. We have the Studio Assistant. People are using it with MCP. And so we're gathering all of those models and bringing the Studio Assistant right to a tab in our mobile app, both for one-shot game creation and iterative gaming creation. It's really an elegant unification of what has traditionally been Roblox Studio all the way to the, the mobile app tab as well.

So right now, I've seen tons of models that are capable of designing incredible games when prompted appropriately in a variety of, of systems. But they're all sort of expensive right now. Some of them are subsidized on plans. But I think everyone's optimistic about the cost coming down by orders of magnitude very quickly. But how are you thinking about doing that dance right now of giving creators the most powerful AI and not putting too many limits on it, but setting yourself up for a really solid financial footprint over time as the product gets more popular?

The hope is, you know, our well over 130 million DAUs, more and more of them can be creators. And I'll highlight that the economics aren't changing. You know, we expect more and more money to flow to our creator ecosystem as even part of this announcement. There's a, there's a neat thing underneath what people see on the user-facing side of Roblox, and that we're a huge infrastructure company. We have data centers. All around the world. We have bandwidth, we have CPU, we have inference. And we started this fortunately well over 10 years ago to keep costs down and performance high. A lot of the inference we run today, we run all on our own data centers. We optimize that in a way. The philosophy and the optimism of costs coming down goes straight to Build in that just as Roblox Studio was the way people made games 20 years ago, that's still going to be live. 5. But we think, uh, as games get more complex, as they get multiplayer, as they have interesting economies, as we want to save the infrastructure, all of that running behind it, that promise comes back with people being able to prompt and create games. Our vision is that everyone will have access to this for a certain amount. We will initially have the ability to buy more token usage if you want, but the beautiful thing is this is eventually consistent And as inference gets cheaper and cheaper and cheaper, just as some other things were expensive 20 years ago, we think more and more people are going to be able to use this more and more.

How, as you reflect on the decision to build your own infrastructure, how bad would it be if you hadn't done that right now? Would you— do you— would you be feeling like a memory crunch? Would you be— are you feeling any of that? Like what about what's going on in public cloud or the AI buildout is well-positioning you or, or maybe even affecting you?

Yeah, I think we've said publicly before the cost benefits of our infrastructure may be roughly 3x. Now, I don't know. I don't want to quote that on inference and other things, but there's an enormous advantage to that and an enormous advantage of taking all that money, pushing it to creators, um, rather than cloud providers. We, we are great partners with some of the cloud providers, especially in burst mode. There's a mathematical optimization where if we can run on our own infra most of the time and burst to cloud when we need to, that's kind of like the optimal cost structure. But, but, but I do think the vision is we want all 130 million of those creators and users to participate in making games.

Take me through some of the other AI initiatives that are rolling out across Roblox right now. I was fascinated by the AI up-resing technology. I'd been very optimistic about that all over gaming. We've seen it, early signs of it with DLSS, but clearly this is next generation technology. What's the, what's What's the feedback been? What's the process like? How—

I think there's a little bit of a fun boiled toad thing going on here. We're all used to what a game looks like. We've been all playing games for 10 or 20 years. We know what they look like and we have an expectation of what we look like. I don't think any of us have the expectation that a game should be photorealistic at 4K 60Hz like a movie, like, like that is something that's the province of movies right now, offline video models that create super high res. And in our minds, we haven't just connected like, what could that be like? Our vision and our dream and our hope is multiplayer gaming at photorealism, 4K 60Hz. And we believe the best strategy for this is not pure 3D traditional gaming, not pure video world model, you know, action systems and all of that, but a hybrid system that is both optimizing the coordination of the 1,000 people in the multiplayer world with local super-upressing for each user. And so that's, you know, that's what we published a blog post on. It's why we're working on this super-upressing video model that we're going to incorporate We showed some early demos, more demos to come, but we do believe it's almost going to be like going from black and white movies to color movies. No one has— we just aren't used to a purely photorealistic feeling game.

So if the successful— that like the keys to success as a UGC game creator, sometimes it can be asset driven. They can create create something really high fidelity. Sometimes it can be they are able to actually instantiate the game quickly. All of that's becoming easier with the two products you mentioned. What remains the secret sauce to a breakout UGC creator on Roblox?

Yeah, I think it's— there's a couple of interesting factors here. The overarching thing, if we take a step back that gaming provides us in society is really our belief. Like, we think the world needs to be playing a little bit more, like hanging out together, using their imagination. You know, it's how we've evolved as humans, and, and play is just a good thing. We think that as far as the creation of these experiences, we're going to see more and more interesting ways for people to play together. I, I feel it's a little bit like we've moved from oil painting to Photoshop. We're all gonna get better at making digital assets and things like that, but there's gonna be enormous room for creativity around what type of environments we play in. Um, we're all gonna feel more powerful, um, creating those environments. We're gonna have the same opportunities, I think, for many of the large teams on Roblox blocks that are big studios now, but we will also have opportunities for first-time creators, just as sometimes first-time video producers get a breakout hit. I do think we're going to make that possible as well.

Yeah, I'm curious how you're thinking about using AI not just to build the games, but also within the games. So I think everyone who's used these models has had this idea of, oh, this would be an incredible NPC. NPC, right, in a game, you can talk to it. It has this incredible ability to respond to what you're doing. I'm curious how Roblox is thinking about using models for—

I would say yes, yes, yes, yes. Both NPC creation as well as dynamic world creation as well. One can imagine a gaming world where, like in the movie Inception, there is a dream master who can modify the whole world for other people in real time. That's a very complicated computational challenge. You know, folding the road over in that dream sequence in Inception. But dynamic world generation, I think we're going to see in real time. Also NPCs, you know, both not just doppelgängers and digital twins in every single sci-fi movie we've ever seen. There's another really interesting aspect for NPCs, and that it's very hard. Um, I, I think it's much harder in the gaming industry to figure out, is that thing going to work or is it going to be interesting? We, we've seen so much success in code acceleration because we can define it, we can write a test plan, we can do iterative loops, we can let it go all night and get better and better. A game is a lot more difficult to define what is good. I think we're going to see more and more NPCs used as testing agents. And the interesting thing about NPCs is rather than trying to write a test plan, well, like, let's put 100 random NPCs in a game, give them the same mission as a human, and see how they perform. The fun thing about really extrapolating to the future with a lot of compute and a lot of inference is those NPCs could conceivably run 100 times faster than humans. So it's feasible to imagine a 100-player test for 100 hours with NPCs being compressed down to an hour, that starts to introduce the notion of Wiggins loop programming for game creators initially with NPCs.

You mentioned Inception, the whole team's nodding because they all just saw it in 70mm in theaters because it was re-released. They love that. What about Roblox the movie or IP that is developed developed in Roblox going to the big screen. We saw Backrooms, a Blender project, very successful summer blockbuster. What's the future of that pipeline?

We are looking into it. We've been really conservative because what we've always imagined Roblox as an IP factory where new IPs that are not coming from the movie angle are coming from the, you know, dress to impress and grow a garden angle. I just went and bought some Dress to Impress physical toys. So I think we're looking at how maybe in the future we would accelerate less of Roblox and more of those creators. It's fun to imagine a future where the brands of those creators supports property. So I would say a lot of opportunity there.

How is advertising developing in Roblox? It feels I feel like AI unlocks, it just lowers the cost of actually doing a branded integration. Someone who's a non-technical marketer can potentially come to you and say, I want a physical instantiation of my store. I want to do some sort of stunt. And you can say, yeah, like this will just be a couple tokens. It's going to be a couple hundred bucks.

I think you're exactly right. What we're experimenting as well, even the descriptions of experiences, the thumbnails that creators make, all of that is starting to become AI and dynamically generated. I would also say as our discovery tool Moments more and more starts to become front and center in the product, that's a much more advertising format a lot of young people are used to. Also our, our on-platform advertising business, which is creators in addition In addition to using our recommendation algorithm, boosting by buying advertising on our homepage is going very well. And then as you mentioned, we're seeing continued growth in both brand integration as well as video advertising as well. Sorry.

I did it. I did it. Sorry. It's very funny. How, how are you thinking about new platforms, new consoles, just the general strategy of, of the Roblox footprint? Are we past peak new platform or is there— are there more surfaces that you want to expand onto? I know 3 surfaces right now.

We're doing early experiments with Android TV, which is exactly the same Roblox client. Yeah, many experiences on Roblox that are not, you know, super high Twitch related are super interesting there. I think there's a huge opportunity to be on all TV platforms. Of course, we really respect Nintendo and Switch, continuing to talk with them. We do think Roblox should be on that device. And then as we start to see all this collection of AR glasses and VR headsets starting to pop out, yeah, you know, we're, we're gonna see more and more maturation of what is kind of the standardized AR, AR form factor, which is heads-up display overlaid on glasses, ultimately projection. That's a real interesting platform for 3D communication as well. So we're trying, tracking that. And of course, we're big fans of Meta Quest and Roblox is live there and a lot of people play in VR right now.

Yeah, in general, the glasses format, I mean, there's been some, some exciting launches, Snap Spectacles and the Meta Ray-Ban displays, but it feels like we're in a little bit of like a, like an AR/VR winter in the sense of like it's just wait for the next major iteration. We're just sort of in between cycles. Is that what you feel?

I think there's a huge AR glasses opportunity once we get speaker, microphone, camera, projection, lightweight, all working together. Everyone's dancing around the edge of it. Arguably, the Google Glass from 10 years ago, I feel, was on the right track. What's the lightest wearable everyday option? So I feel we may be in an AR winter. I'm not sure, but I am very optimistic about that form factor.

Yeah, yeah. It's, yeah, it's interesting. I'm such a VR bull, but at the same time, when I see just rumors about, oh, this next great display was canceled or something like that, I'm not—

Yeah, I mean, VR and AR are super different. I had the privilege of trying on a VPL data glove literally when Jaron Lanier ran the first, you know, VR system on SGIs. So I saw it super early. So I'm super optimistic about AR.

What about like screenless experiences? I mean, there's a lot of like smart speakers and the Meta Ray-Bans that don't have the displays have an AI that you can talk to. And I'm wondering, like, is there, is there, is there any opportunity for a Roblox creator to create a game that— I mean, there are games on Amazon Alexa, and I'm wondering if there's any Roblox creator that could potentially create a game that doesn't require a screen, at least for part of the experience.

I think the more Roblox is is not thought of just as a game or a platform, but incredibly high-performance, low-cost infrastructure that's available to everyone. That infrastructure is both single-player and multiplayer. It's 2D and it's 3D. It can be on low-end devices and high-end. And, and as that infrastructure more and more incorporates NPCs and audio and those types of things, you can imagine crossover type games that involve NPC interaction jumping from those with a visual display to pure NPC interaction. So I would say we're, you know, very focused on providing high-quality, low-cost infrastructure throughout the platform. And I think also focused on where it makes sense from performance and privacy, whether it's our voice safety models, our text filter models. Yeah. Our super upsampler model that we've announced, NPC model, game creation coding models, scene gen model, single part creation model. There's many of these where we feel we have the data to build world-class and kind of put it all together in our build harness.

What, what is your like overarching philosophy on where certain pieces of safety and KYC infrastructure should live? Because I mean, I have kids, they're not playing games yet, but I imagine they will at some point. I'm probably gonna have an opinion about the amount of screen time. I'm probably gonna do some research on what the right amount is. Some of that can live at the device level, some of it can live at the application level, some of it can live just at the parenting paying attention level. How do you think about the different pieces of the puzzle, where identification lives, where screen time lives, where different filters live? Is this an important trade-off?

This is a top big discussion in DC, right? Sure. What responsibility do device manufacturers have? Do they have to tag the age of every device? I would roll back all the way to Roblox values. One of the four is we are responsible, and that value has has led us to, in a way, uh, shutting off a phone call on my Apple Macintosh. Sorry about that. I hope you don't hear that. Um, so that has led me to, um, sorry about that. You're good. Problem having integrated phone on your PC. That has led us to, um, taking responsibility and doing a few big moves. No, no sharing image or video, age checking now everyone on the platform. So both AI and biometric signals and really being unique in leading that charge. We ultimately will take every age signal we can get, whether DC mandates it from a device, whether we collect it, and we just get higher and higher resolution. But we're not waiting for a law or for device manufacturers to do that. Yeah, like we're, we're, we're already done. Like we're already age checking everyone and using that to ban communication and monitor content. The other thing I would highlight is there's no silver bullet there in that even with device age check, so many parents are so busy that so many devices get handed around the the household. Just go take my phone, go take my tablet. That it— that we will have young people, we will have 9-year-olds on devices that are age-checked for 18. That there is some reliable, um, reliability and constantly doing continuous age-checking like we do.

Yeah, I mean, the funniest example I can remember is, uh, there's this old Twitter exchange where, uh, presumably young woman got banned from using Twitter by her parents, and so she logged into her Samsung smart fridge and sent a tweet. It was like, I'm back online. And it just shows you, like, kids are so ingenious, like, they will get around things. So you have to have layers of protections at every level, and there's no, there's no one-size-fits-all approach, which I think is your—

that's exactly right, right?

Yeah. I'm curious how you're thinking about safety, especially in regards parts to AI. I mean, every time a new frontier model comes out, you'll see someone within like an hour who's jailbroken it. Yeah. They have this crazy prompt and you can get the model to say all these like crazy stuff. So I'm wondering how you're thinking about that, especially in terms of these, these like large language models.

Oh, like red teaming? Yeah, sort of.

Yeah, I would say we're going further than that in that now for everyone on our platform under 16 right now, they're— they have a huge corpus of content. It's called Kids select content. It's well over 20,000 games, but these are going through a fairly excruciating both user funnel as well as moderation funnel. And in a way, it's working really wonderfully. That's an enormous amount of content. At the same time, it's not the full UGC catalog of millions and millions of things. And because we have age check, we can push it there. The vision with Build is we have the ability in Build build that we don't fully have in Studio, where everything's being created by a prompt, where we can see a history of all the prompts that have been used to create a game, as well as ultimately the images that have been uploaded, the models, all of those kind of things, to really make a similar determination on the quality of that content. Whereas when you use a wide-open IDE, whether it's Roblox Studio or VS Code or whatever, you probably can make anything. And so we, we are optimistic we will keep the same safety gauntlet for experiences created by Build. We're, we're also really optimistic that because our discovery system is so unique to gaming, it's, it's very much based on estimated long-term retention based on direct measurement that our discovery team's pretty optimistic skeptic. Even if there's 20 times more games being created, one could have the fear, oh my gosh, we're going to get a lot of AI slop. But, you know, we already have millions of people making games. Some of them are amazing. They're not all amazing. They need to make it through a high retention gauntlet before they start really being surfaced. So the beauty of this is we think our current discovery systems are already perfectly poised for the expansion of creation from build.

Yeah. This is the aggregator thesis. The— you're already filtering everything and you're set up for it. Um, how are you thinking about, uh, intellectual property on Roblox in this where it gets even easier to make games and experiences? Um, there have been some artists and IP owners that have been more lean forward on, yeah, remix my song. Or use my likeness. Sora had this weird cameo feature that I thought was pretty elegantly implemented where someone could go in and say, "Anyone," like Jake Paul was like, "Anyone can make— can remix my image. I want to be all over the internet." Other people said, "Don't put me in any videos that you generate." And I'm wondering if there's like what the long-term relationship with different pieces of IP. Some filmmakers might want a bunch of free UGC games that promote their movie, and some might not want any of their IP to leak into Roblox. How are you treating that?

Yeah, so we, you know, that relationship between IP holders and current creators and game developers is really complicated. There's multiple countries, there's Hollywood contracts, like having, imagining a small 1 or 2 person shop shop, going and getting a license to some big IP is somewhat unfathomable. We do believe the future of this is just like Roblox, solving it in a systems-oriented way. We have a platform right now that is somewhat new called IP Manager, where we want— and we are running the gamut all the way from some pieces of IP which, hey, we're open for licensing. We want multiple creators to come and use our IP. The IP from the movie Saw is in our IP manager. And so a creator who's interested in using that IP can directly contact, you know, do an online contract and start producing something from that. But that IP manager can also act as an IP controller as well. IP can go up there more. And the more that AI gets good at auto scanning and auto detecting can also be some companies that just say no one anywhere should be using our IP. So I think we're viewing this as a systems opportunity. And just like the democratization of gaming, we want more creators making interesting games. Democratization of the licensing process.

Yeah, I mean, I've been really optimistic about AI solving this because I've watched firsthand how YouTube has dealt with it, where, you know, you put a song in, immediately the royalties just go over there. You don't make money, but that artist does. And most artists are fine with that. And then they always have the option to click that button and say, actually, I don't want that at all. And it's not perfect. Everyone has approvals at the right time, but it's like this stable equilibrium where I think everyone gets what's economically viable. I'm very optimistic, like, this will ultimately be a solved problem. Yeah, it's exciting. Well, congratulations on Build. Congratulations on all the progress. Thank you so much for taking the time to come chat with us.

July 28th, Build goes into alpha. July 28th. Thank you.

Thank you so much for coming on the show. Have a great Thursday. Have a great rest of your week. We'll talk to you soon. Goodbye. Let me tell you about MongoDB. What's the only thing faster than the AI market. Your business on MongoDB. Don't just build AI, own the data platform that powers it. In creator world, there is some news from Colin and Samir. Lexus is now the official car of Colin and Samir. What does that actually mean? They made 4 ads for them that roll across the rollout across YouTube. They're sponsoring 4 videos on their channel. It's the first of its kind deal that represents a broader shift taking place in media. The aperture of what it means for a brand to work with the creator is changing quickly. It's very cool to see because obviously they've been on YouTube for a long time. They've done a lot of like host read ads, mid-roll ads. But this is a much deeper integration and something that I think will be hopefully replicated all over YouTube and be a new source of revenue for for creators of all kinds. So I was excited to see this. In other entertainment news, Jake from Econ C-M-Pick, Econome Pick, says, this is almost hard to believe. Disney spent $129 billion acquiring Marvel, Star Wars, Pixar, ESPN, and Fox, which is 100 $82 billion in today's dollars. Throw in all their legacy assets and the entire company's market cap today is $169 billion. Wow. Do you know what this picture is missing?

Uh, which— what do you mean?

So, so they're saying they acquired all these assets and the company is only worth $169 billion. What's missing from this analysis? The cash that's been returned to shareholders. Disney across dividends and buybacks has returned like $70 billion, maybe more to shareholders, which is— I don't know. I just— I'm giving up for sure. And it is, it is, it is an interesting angle because they have spent a lot acquiring and the company is not worth more than what they acquired. So there's this question of like, were those, were those acquisitions accretive or destructive or dilutive? But there is a whole separate picture, which is that a lot a lot of cash has been returned to shareholders throughout this journey.

Yeah, also, I mean, that's the mechanism with, with which those acquisitions were funded also. Yeah, should—

yeah, matters a lot. I don't know, it was sort of interesting. Uh, Sean Frank has a pitch. He says you should move to New York City, bro. You got to move to NYC. The weather, horrible, 100 degrees easy. AC, eff that. Taxes, so high. Rent, highest in the country. Air quality, some of the worst in America, tech bro, we banned— wait, do they really ban Waymo in New York? Uh, I believe so.

Yeah, they know Waymos. Wow, that's very wild.

Uh, yeah, if you can make it here, you can make it anywhere. Uh, so I don't know, do you ever have aspirations to move to New York City?

Uh, at some point.

You've been to New York City? Yeah, yeah, it's a nice city. That's the thing is that all of this is true and it's still a great city to hang out in. It's so fun, so dense, you can see so many people walk around. Beautiful. It's just like, I don't know, it's unlike anything else.

Still great, but, uh, yeah, you never lived in New York City.

I've never lived in New York City, but I've spent like a lot of time there, so I've had a good time. Anyway, let me tell you about Codex. Codex is a powerful workspace for getting work done with AI agents, whether writing code, analyzing data, creating content, or automating business workflows. Codex helps you move projects forward from start to finish. And there's some, there's some other news in the timeline,, but we can go over it more next, next week. Eli Lilly is buying psychedelics firm Atai Beckley for an initial $2.8 billion. Atai was founded by Christian Angermayer. I've actually interviewed him years ago. Fascinating deal working on psychedelics and a bunch of other things, therapeutics for mental health conditions. Huge deal going on there. And there's a few other stories, but we can get to them tomorrow. We're off tomorrow. There's some travel going on, but we'll be back Monday at 11 AM Pacific sharp. Thank you to everyone who tuned in in the chat. Thank you for positive reviews of Tyler. Let us know what you think of Tyler. Leave us a review on Apple Podcasts and Spotify. Write us an email. Tell us how he did. I think he did fantastic.

Have the best Thursday of your life. Yes.

There you go. That's a good impression. That's what you say, right? That's a good impression. But thank you. Sign up for our newsletter at tbpn.com, and we will see you on Monday. See ya. Goodbye. Rolling flashbang. Oh, we got the flashbang. There we go.