OpenAI's New CEO of Applications Shares Optimistic Memo, Stargate Hits Early Struggles, P&G Pays Unilever $10M in Spying Case | Anatoly Yakovenko, Aaron Levie, Neil Parikh & Daniel Cahn, Jan Sramek, Nathaniel Manning, Chris Andrew

26:56Yeah, for sure. AI coaches, on the other hand, can be available throughout every day. She says, this isn't about replacing human connection, but filling a gap that often goes unfilled. Many people don't feel comfortable opening up to family or friends, and most people don't have access to a therapist or coach that can call regularly. Even people who do have access often spend an hour a week or less with these professionals. At the core of philosophy and religion is the idea of self-knowledge. To become who we want to be, we have to understand who we are. So very interesting. Anyway, let me tell you about Linear.

27:29Linear.app. Linear is a purpose-built tool for planning and building products, meet the system for modern software development, streamline issues, projects, and product roadmaps. It is what OpenAI uses to build their products. That's true. So yeah, I think it's interesting to see how useful is it to have these six ideas in mind because in many ways, like ChatGPT, the magic of it is that it's just a blank text box and you can do anything with it and then these properties emerge. Like the, yeah, I mean like Google over time had to figure out, okay, there are people that are searching for products.

28:09let's surface them the Google shopping experience. And that is a specific UI that is triggered when you search for a certain product. If you ask for best headphones over the ear, under$200, it will just put you in a different product more or less, instead of just the 10 blue links. Same thing for if you're asking for a movie. It will pull up the pictures of this cast and the year and show you the Wikipedia page or news. It'll take you to the news portal. And I'm wondering how much of this will actually bifurcate into different trees with different rules and different features. Because for something like support, I don't know that if I'm having a conversation with a coach about a strategy that I want it to wait 15 minutes to get back to me.

29:01That might be something I want it to do separately. but then if it's health and I'm uploading my health records or some documents, I'm like, really, really, really don't hallucinate. So like take as much time as possible. Again, that's something that right now I would decide as the user, but I'm wondering if this buy for kids at all. Well, yeah, obviously OpenAI, ChatGPT as a business, when anybody wants to do anything, they want people to go there, right? When you want companionship, you want to talk with somebody about something, you go there. When you want to, they do the book flight demo that's overplayed, but they want to own that.

29:34When somebody's shopping, they want to own that. When somebody's trying to diagnose a medical issue, they're trying to own that. So I just look at this article, this essay, whatever you want to call it, as a way for people internally at OpenAI to sort themselves into what area. These are all different verticals that I can imagine people working on. Like I want to go to Chachi Beauty and work on health, right? Or giving people their time back, et cetera. I listened to Ben Thompson talk about the flight booking agent demo, and he was like, I have a personal assistant. I book my own flights. And I think that's very, very common, where it's like he's so specific about what he wants that he's fine to just do it.

30:16And it doesn't take that much time. And so I do wonder how stick it out. I have a flight booking AGI. Yeah, in your brain? First name ramp, last name travel. That's true, yeah. I mean, still, when I use that, it's like I'm in control. It just does all, it harnesses everything to bring all the data together. Yeah, the interesting thing, I've had EAs over the years. I've had remote EAs. I've had EAs that are embodied, real people that are absolute killers. And I find that the most, the internet already makes it so easy to do so many things that the greatest, some of the greatest value that you can get from an EA is just having somebody in your city that can help you with a variety of different things.

31:00And so I think no matter how good ChatGPT agent gets, there's still, I mean, there's just still so much that I think people expect out of an EA. Yeah, yeah, it's very particular. Like people have all sorts of preferences and I don't know, it might be more sticky, who knows. We'll see. It'll certainly continue to be a good eval. Well, I want to just look at what is, you know, it feels like OpenAI has a really strong lead in consumer, the product, they're thinking about the applications and actually like the manifestation of like these six different categories. Polymarket has OpenAI climbing now in which company has the best AI model at the end of 2025, the December 31st rankings.

31:56OpenAI is up to 29%. Just a week ago or so, July 12th, XAI was crossing, about to beat them. But Google's still in the lead. 48 % chance that Google winds up with the best model. But that doesn't necessarily mean that... I mean, this is based on LM Arena. Exactly. It's not necessarily the most valuable model to users. It's who's dominating benchmarks. Yeah, I actually pulled up some interesting, so do you know how there are movies that are loved by critics but disliked by audiences? So the classic example is Star Wars The Last Jedi, 91 % on the tomato meter, but audiences gave it a 41 % score.

32:43And then Dave Chappelle's comedy special had the opposite, where the critics hated it, but the fans loved it. And so these things are like controversial. And I was wondering if the same thing exists in foundation models. And it does. So Gemini 2.5 Pro has, which one is this? Gemini 2.5 Pro has an MMLU score of 86%. And it's the top rank model on LM Chatbot Arena. So if you summarize these, let's see. Kimi K2 illustrates how a model can excel on benchmarks and still be less preferred by users. Possible reasons include limited availability. Similarly, Claude Opus 4 and GPT 4.5 have high MMLU scores yet rank lower in human preference because they may focus on careful reasoning rather than conversational ease.

33:40So you can think about it like you don't necessarily always want to be talking to someone who's like the most rigorous. It's almost like they're too smart, maybe. And then Gemini 2.5 Pro 03 and Gemini 2.5 Flash show the opposite pattern. Users love them despite modest MMLU scores. So these models have long context windows. We need to get Tyler to go to a random place in America and do a man on the street interviews and just ask people about different models. I bet you 99 % of respondents probably like 990 out of 1 ,000 are going to tell you, I just go to chat.com and I just open a new chat. There's multiple models?

34:28Yeah, totally. But it's interesting that if you want to win with people, Gemini 2.5 Pro, 03, these models have long context windows, so they're staying focused a little bit longer if you're having a conversation. multimodal features, the images and audio, kind of just a better user experience. You can talk to it in different ways, faster latency, more personable style, makes them more enjoyable. Even if they don't top academic tests, people are happy with it. So I thought that was an interesting little anecdote. So anyway, let me tell you about numeralhq.com, sales tax on autopilot, spend less than five minutes per month on sales tax compliance.

35:09you had a post today from the Wall Street Journal this was wild, this stood out to me we were commuting into the studio this morning and this stood out so OpenAI has committed, the journal reported that OpenAI is committed to paying Oracle$30 billion per year within three years for four and a half gigawatts of data center capacity which will be the equivalent of two Hoover dams worth of energy so a good time to be uh uh uh if you if you happen to build a hoover dam mega project you know years and years ago yeah this would be a good time if you're squatting on a hoover dam yeah you could get 15 million a year from that well that's obviously not not not the entirety of of the cost but um but yeah it's it's it's wild to think uh opening eyes latest like run rate right They're reporting somewhere around$10 billion run rate, or they want to end there by the end of the year.

36:08And so they're expecting to be able to have$30 billion worth of that. And that's just with this partnership with Oracle. Yeah, there was a, so Tane, who's been on the show, says, ChatGPT is a truly iconic product. Consumer engagement at scale, over 500 million weekly active users, an estimated 7 billion ARR. and it looks like an annual chart and it's a monthly chart. So August 2024, they're at a 2.6 billion estimated consumer ARR. Then in September, they're at a 3.3. December, they're at a 4 billion ARR. Then they're accelerating. February, they're at 5.6 billion. March, they're at 7.2 billion.

36:53And of course, we're all the way in July now. So they could potentially already be at 10 billion ARR. And so if you just chart this out, it 30 billion in a few years doesn't seem that crazy. And we were talking about like, how bad is that? How painful is that to pay a 30 billion dollar bill? But it is kind of the only cost is like to serve the models. And so I don't know. It feels like it could work out. I don't know. I mean, it's certainly not the only cost. I mean, especially if you look at stock-based comp, they're spending more than 119 % of revenue on stock-based comp over the last year. So very, very significant.

37:41But this is what betting on yourself looks like. If you want to level up in life, commit to$30 billion of fixed costs three years from now. You've got to make it work. So for sure. So this all comes from a Wall Street Journal report saying that OpenAI's Project Stargate is struggling, which is an interesting term. We can kind of debate what constitutes a struggle. So for some context here, all the hyperscalers are generally in the high tens of billions in terms of CapEx. So Google, Microsoft, Amazon, AWS, they're all in like the 60, 70, 80 billion dollar range. But the majority of that is on core AI or core data center build out.

38:30So traditional data center build out because they still need to serve just, you know, relational databases. Not everything is cutting edge NVIDIA chips serving transformer based language models. Ghiblis. Ghiblis. Not all of it's Ghiblis. But a lot of it is. and Mark Zuckerberg recently said that about half of the spend maybe going forward, something like that is going towards Gen AI versus Core AI. So he builds out a big AI data center. A lot of that's going to just serving reels, picking what adds to show you all of that as AI as well, but it's not in this generative AI, not training these large models.

39:09So the majority of that spend is not on generative AI-focused data centers, but there's still big numbers. you can think about all the hyperscalers. Zuck has his own Manhattan project. Yes. That's basically what, if you show that your new data center is going to be the size of Manhattan, that's one way of saying, I am doing a Manhattan project for AI. And that, like, you know, the size of Manhattan, that particular data center, he's targeting five gigawatts, which is right here at four and a half gigawatts in three years. And when Mark Zuckerberg was talking about that five gigawatt data center, he was saying, this will be in a couple years.

39:49Right now, they're focused on the one gigawatt data center. And when we talked to the co-founder of CoreWeave, he was saying, like, you don't understand how big five gigawatts is. Like, that's insane. It's such a big data center. Like, this is a huge project. These are mega projects. And so when I hear something like, oh, it's struggling, like, a couple months in, I'm kind of like, you know, we're building something so big. Like, did we ever expect there not to be a hiccup? I'm not ready to say like, oh, this is all doom and gloom. It seems like, you know, they shot so big with 500 billion. So, yeah.

40:23So, one, got to give, regardless of what someone thinks about Sam, right? He's got plenty of critics. But the fact that he was able to finesse his way to getting the president to announce this project with Masa. with a number so large that people immediately just started looking at SoftBank's balance sheet and saying, okay, cool, how is this happening? Elon quickly fired in the comments saying they don't have the money. It was always unclear where the money was going to come from. But one of those things, you shoot for$500 billion, you land at$100 billion, you're still doing pretty well. Yeah, and it feels like it's more about the corporate structure and the timing.

41:14Because if you go to Google, Microsoft, or Amazon, and you say, we're so in on AI, we're going to invest$500 billion over the next 10 years. All of a sudden, it's like, okay, yeah, well, you're on track to do that. That totally maths out. So the question was always, in my mind, okay,$500 billion is the goal, but what is the time frame? And that's the one that people are pulling on right now because the immediate number was$100 billion. They said, we're going to do$100 billion right now. That was in January when Trump, Massa, Sam Altman, and Larry Ellison were at the White House. They announced Stargate.

41:52They say, we're targeting$500 billion, but we're doing$100 billion immediately. Now, that's maybe been scaled back. But it's still so huge. Like SoftBank's putting 30 billion in open AI and then open AI is paying Oracle 30 billion per year starting within three years. So struggles is a little odd of a characterization. Oracle's projecting 25, over 25 billion for 20, 20, 20, 26 capex. Everybody's just throwing money in the pot. Yeah. And I mean, it makes sense because all the reporting is the definition of Stargate is just being expanded. Sure. So they're counting the Oracle, this sort of commitment that they're making at Oracle as part of it.

42:37Yep. And so I think there are two questions. One is like the super intelligence AGI pill, like we gotta build the next greatest thing and whoever builds that at 10x the scale, whoever gets to the 50 billion dollar, 500 billion dollar data center, they just train and then they win because they'll have a product that's so much better than everything else that everyone will shift over and ChatGPT will be obsolete. The flip side is that if we're in a little bit of an S-curve, if we're in a little bit of a plateau, well then, in order to hold on to that 71 % market share, ChatGPT just needs to keep delivering inference tokens reliably at a level that doesn't cause churn.

43:20And so if I go to ChatGPT and I'm getting a ton of errors and I just cannot get my prompts done and like my Ghibli's aren't working, I'm gonna open up the Gemini app and I'm gonna go over there and see if they have any inference tokens. And we're hearing this from the model routing companies that a lot of businesses that rely on LLMs and AI inference are seeing brownouts, intelligent brownouts in different models at different times. And so they need to be able to reroute from - Yeah, somebody was looking at Anthropics uptime and it's just constantly down. Like there's so much demand and they're obviously doing everything in their power to deliver against it.

43:56Yeah. But they don't have enough GPUs yet. And so that feels like everyone's got to build bigger data centers. Maybe there is a point where we build too big, but everything looks like, okay, it's at least worth building the one gigawatt data centers, probably worth building the five gigawatt data centers. We might need some more research breakthroughs to justify the bigger data centers, the 10 gigawatt, the, I don't know, 20, 50 gigawatt, something crazy. But in general, I think the main goal for OpenAI right now is just to continue serving inference at a level that avoids churn. So when someone shows up, get them the answer.

44:34I think new products like Agent are cool, but it takes months for people to really adopt these new tools en masse and seriously change inference demand. We saw this like Agent went out Thursday. We talked to the team. And then over the weekend, I was talking to people. And smart people in tech were not using it yet. And I kind of had to like step up and be like, it's my job to test this thing. And now I'm using it. And now I would say I'm probably a weekly active user of agent. I'm definitely a DAU of 4.0. I'm a DAU of 03 Pro. I'm probably even a DAU of deep research. But I'm not a DAU of agent yet, but that will come.

45:11But I'm clearly on the early adopter side. It will be a while until they're doing so much agent, so many agent workloads that they're really - Now it just default takes 15 minutes. So deep research takes 15 minutes almost every single time. I thought chat GPT agent always. No. It can be shorter. It can be longer. That's part of the beauty of it is that it sizes the inference relative to the task. And so it can spend a really long time. It can come back to you. It can interact with you. Although I haven't actually gotten a push notification. Maybe that's just because of the nature of my prompts.

45:44But I haven't had to step in, which has been surprising because when we talked to Dan Shipper from every, he said that he did have to interact, but I think what he was prompting was a little bit more robust than what I was asking for. So, um, yeah, what he, what, what he's doing that I think will, uh, will get a lot of adoption, but a bit slower is he sort of like batched tasks that I think are really interesting, but that is a expert mode for now. So, um, that's the story in the wall street journal. Uh, there's flip over, uh, maybe cover this, this AI, Agent for customer service, number one in performance benchmarks, number one in competitive bake-offs, number one in ranking on G2.

46:24Bake-off champion. What do you want to flip over to? I was going to say, maybe we cover this XAI story quickly. Yeah, totally. And then we have actually Tyler, our very own Tyler over and Michael are at the Tesla diners. Oh, cool. So we'll flip over to them and get a review. So Sam Altman, Larry Ellison, Trump in some ways, Masa are all working to build Stargate simultaneously. Elon Musk is trying to ramp up things at XAI. He's aiming to raise up to$12 billion for XAI chips as the startup burns through cash. And so just weeks after Musk's XAI raised$10 billion through sales of stock and debt, the startup is working with a trusted financier to secure up to$12 billion more for its ambitious expansion plans.

47:13Valor Equity Partners, Antonio Gracios, is in talks with lenders to raise the capital. The money would be used to buy a massive supply of advanced NVIDIA chips. And so the number was kind of all over the place. I saw a million chips floated. That would be much bigger. So Elon shared this morning 230 ,000 GPUs, including 30 ,000 GB200s, are operational for training Grok XAI in a single super cluster called Colossus I. Colossus I. And at Colossus II, the first batch of 550 ,000 GB200s and GB300s also for training start going online in a few weeks. Yeah. So if you compare that to what the other clusters are, Prometheus from Meta is around 500 ,000 chips.

48:02Stargate's supposed to be around 400 ,000. And I believe Anthropik's in that range too. And so it's a knockout, dragout fight. Everyone's trying to build a massive cluster, do the next big training run. And XAI has been moving extremely quickly. 122 days for XAI to build its first giant data center in Memphis, Tennessee, Colossus. It originally housed 100 ,000 NVIDIA GPUs. It was among the world's largest clusters. Just 92 days later, XAI doubled Colossus' size to 200 ,000 GPUs. and now he's going even bigger. So good luck to everyone. Pulling in capital from everywhere. SpaceX had to get into the round.

48:44It's quite the party round. Yep. It's all over the place. Five billion in corporate debt, included bonds in loans secured by data centers, which makes sense if you're building something CapEx intensive, you would use debt for that. Anyway, in other news, I wanted to talk about Adio, customer relationship magic. Adio is the AI native CRM that builds scales and grows your company to the next level. Get started for free. No, I want to talk about Procter & Gamble and the idea of this Michelin tires question that we have been toying with. Been toying with, to be clear, since the original deal and rippling debacle.

49:26Yes, yes. Since the timeline was in turmoil. Yes. Our reaction was that on a long enough time horizon, people will not care that a payroll provider may have gotten a little too competitive in the past. and the comp that we discussed was if you went to your mother or a family member and said, hey, I saw you're driving using Bridgestone tires. Did you know that the CEO of Bridgestone was actually spying on Michelin? She would say, that's very sweet, honey. I'll keep that in mind when I replace my tires in two years and then probably forget about it immediately. And it also relates to the astronomer story where the CEO was caught having an affair at a Coldplay concert.

50:19That is something that people were saying he has to go. It reflects poorly on the company. But if you look at Apache Airflow Management, are you going to rip out astronomer over this, especially when they replace the CEO so quickly? And so this question of what is existential? what puts you in the FTX category or the Theranos category versus a company that can come back and rebuild from what is controversial. So there's this old story from 2001. This was in the New York Times, and I think Fortune broke the story. So I'll kind of give you the high level. So late on a spring afternoon in April of 2001, a red flag memo landed on the desk of Procter & Gamble chairman John Pepper in Cincinnati.

51:07Internal security investigators had discovered that a handful of employees working with an outside outfit of ex-US intelligence officers, I guess like CIA guys, had been rifling through rubbish bins outside Unilever's offices and posing as market analysts to extract secrets about forthcoming shampoo launches. The covert snooping, management learned, had been going on for months under the radar of P &G's famously buttoned down compliance system. The operation, codenamed The Ranch, by the company's competitive intelligence unit, first took shape in late 2000 when P &G hired Alabama-based Phoenix Consulting Group, a firm led by Vietnam-era clandestined veterans.

51:54So like CIA guys in Vietnam come back from the war. They're probably like 50 at this point. And they're like, yeah, let's do some corporate espionage as a service. So by the time the plot was exposed, the operatives had spirited away roughly 80 internal Unilever documents. So the rival consumer packaged goods company marketing calendars, packaging mockups, even margin analyses, along with samples lifted straight from the trash dock. So they're literally digging through the trash. And if it's so funny to think why this was necessary. Oh, it's incredibly valuable. So if you know someone's marketing calendar, you know that they're going to launch this product on Tuesday.

52:34You can launch a competitive product on Monday, steal all the wind from that. You can launch a counter position product. There's so much that you can do. Yeah, I know, but... And margin analysis is really key because if I know that you're selling a product at low margin and I can undercut you, then you're backed into a corner. You can't lower your margin without losing money on that. As opposed to if you have a high margin product, I might not want to attack that. There's a whole bunch of different things that you can do if you have good competitive intelligence. and even packaging mock-ups, like understanding that, okay, they're going with a bigger size.

53:08They're gonna be on this shelf in Walmart. We should go and try and lock up Walmart, overpay to block them out of that slot. There's a lot that you can do with this stuff. So Pepper, who is the chairman of Procter & Gamble at the time, his immediate instinct was containment by candor. He dismissed three staffers, seized the files, and telephoned the Unilever co-chairman, Neil Fitzgerald, to confess. Within days, P &G couriers arrived at Unilever Chicago headquarters with bankers boxes of paperwork and a pledge that no one who had seen the material would touch hair care strategy again. This was an unfortunate incident, Pepper would later say, and we are ensuring the information has not been and will never be used.

53:49That might have been the end of it until Fortune Magazine landed the scoop on August 30th, 2001, under the headline, P &G's Dirty Little Secret. The story, which read like a corporate John le Carre novel, who, of course, does Tinker, Taylor, Soldier, Spy, and a bunch of other great spy novels you should definitely read, detailed dumpster dives, safe houses, and fake business cards, and triggered a global media pile-on. Within a week, the Los Angeles Times, The Guardian, and newspapers from Dublin to Delhi were splashing shampoo spy puns across their business pages. Facing a reputational firestorm on the eve of the crucial back-to-school selling season.

54:26I didn't know that that was that big for P &G, but I guess it is. Going back to school, you need some soap. I guess you need new soap. The two consumer goods giants struck a confidential settlement, announced September 7, 2001, published reports put the payment at about$10 million. More painful for P &G, Unilever secured an independent monitor empowered to audit new product timetables and the reassignment of any P &G employee tainted by the files. This agreement ensures our confidential information is protected, Unilever's U.S. Chief Charles Strauss told reporters pointedly. Regulators never filed criminal charges.

55:07Dumpster diving on public property skirts US theft statutes, but the incident singed P &G's carefully cultivated Midwestern morals brand. The company rolled out mandatory ethics and competitive intelligence training worldwide, while industry groups tightened codes now that flat out ban clandestine trash raids and pretext phone calls. Competitive intelligence textbooks and MBA ethics courses still open their corporate espionage chapters with the P &G Unilever saga as the moment when lawful information gathering fell headfirst into the dumpster of public opinion. I mean, this probably created a bull market in paper shredding.

55:48For sure. Right. Yeah. Yeah. You don't want people diving into your dumpster. We got to shred all of our materials when we're done. All of our bangers. All of our printed posts. Can't find these anywhere on the internet. More than two decades later, both Pantene and Suave remain grocery aisle staples, proof that the scandal left no lasting dent in shampoo sales. What endured is the cautionary tale. The world's dullest products can tempt even the bluest chip companies into spy novel theatrics. And that in the age-old war for market share, the most corrosive thing you can pour over a brand isn't a rival's conditioner, but the acid of public embarrassment.

56:24Very fun. Wild story. Well, we should jump over to Tyler, who is at the Tesla Diner. here he is looking sharp on the ground. How are you doing? Yeah, I can hear you. Okay, sweet. Yeah, I'm here at the Tesla Diner. I'm on the roof. We couldn't get food. The line was like extremely long. But up here on the roof, you can see it's kind of like a retro futurist, like very Americana style diner. You know, they have burgers, fries. I think there's apple pie. But here, I can walk around. Yeah, yeah, sure. We heard it's farm to table too. Is that true? or they're partnering with local farms? It's unclear, I would have to ask some of the cooks here, but you can see there's these two massive screens they have.

57:08Let's see if you guys see this. And they're playing what looks like, I don't know, a reality show or something. There's like these massive, I think there's 80 superchargers. Can you guys see this? Okay, yeah. Are they full? People are charging their cars? I think they're almost all full, yeah. So I think the way it works is you can order from your car if you have a Tesla. If you don't have a Tesla, then you've got to wait in the line. It's like 45 minutes long, probably. You can see, let's see if I can find a box. The burgers come in these little Cybertruck boxes. Let's see if I can pick it up.

57:45Oh, wow. Very cute. Yeah. And then I look over here. Did you just pick up a random stranger's box. Do you know that person? Julia. Oh, that's Julia. Hey, Julia. How are you doing? If I walk over here, there's actually an Optimus robot serving popcorn. Okay. Let's see it. Let's see here if I can walk over. Can you guys see this? Oh, wow. Yeah, there you go. Any indication if this is teleoperated or end-to-end? It's unclear, but there is someone standing right next to it. Okay. I think maybe for security. Yeah. but yeah overall it's very cool right on Sunset Boulevard okay Tyler we're going to give you we're going to give you a ramp bump on your card we're going to make it so that you guys can start going offering 10x the retail price of the burger you guys have to try a burger you can't leave without a burger so go around and just offer start offering people start bidding on people's burgers right at their table.

58:49Just say, you got that for$10. I'll give you$100 right now. And figure out a way to get a burger. You can't leave without a taste test. No, you should. Yeah, you should wait in line and talk to people and actually try the food. Give us your review at the end of the show. Okay. Yeah, I'll try to find someone. Awesome. Awesome. Thanks for checking in. Very cool. Do you think this is a meaningful business line or a marketing stunt? Great marketing activation. I think that I can see them putting these in tier one cities. I mean, the charging stations are interesting, right? Gas station convenience stores are great businesses because you have this captive audience.

59:41and people that are charging their cars typically have to hang out for quite a bit longer than somebody that's filling up gas for a few minutes. So I can see them having a handful of these or even developing their own kind of convenience store concept over time. Yeah, it kind of harkens back to like when the soap companies, Unilever P &G, going back, started like the soap operas to like promote soap. And it's kind of like this adjacent business that actually could grow into something. but there's a big question about like, should this also be on the Tesla team's plate to manage and grow a chain of restaurants?

1:00:17It's like a lot of stuff to do. But it is a very, very cool like view into the future. It's a showcase. It feels like the World's Fair. That's what it feels like. It feels like going to a World's Fair. Very cool. Well, we have Tully in the Restream waiting room. Let's bring him in. Very excited for this conversation. There he is. How you doing? Welcome. Good to meet you. You are live. How are you doing? Can you hear us? No sound yet. Let's ping him. Hopefully bring him in from the Restream waiting room. In the meantime, let me tell you about adquick.com. Out-of-home advertising, make easy and measurable.

1:00:58Say goodbye to the headaches of out-of-home advertising. Only adquick combines technology, out-of-home expertise, and data to enable efficient, seamless ad buying across the globe. and if we are ready, we will go back to Anatoly from Solana. Get the full story. I'm obsessed with the story of Solana. I made a whole YouTube video about it. Fantastic founder story, really deeply technical. I don't know. It's just a fascinating story. And he's been at it for a very long time. So very excited to talk to him about it. And he made the careers of a lot of different venture investors. For sure. For sure.

1:01:35Angel investors. Yeah. Yeah. And just fascinating trade-offs and how you design. I mean, pretty, I don't know, it just feels like such a massive challenge to start an L1 after Bitcoin, Ethereum are out there. And you're like, I'm going to do the hardest thing in the world, probably. And then I'm going to make it through some of the most tumultuous markets ever. and kind of be at peak awareness right during the interest rate rise and massive sell-off in crypto, stuck with it, kept building, and is here today to tell us more about the story of Solana. So hopefully we can bring him in. We'll see.

1:02:13Okay. In the meantime, we should talk to you about MKBHD. Did you hear this story? MKBHD in 2020 he had a video that he put on YouTube exposing this fake phone Escobar fold you saw this I saw this video when it went out so it turned out some guy was scamming people by sending phones to influencers but not to buyers so basically what he would do is he would buy a Samsung phone give it a wrap basically rebrand it and then he would sell it and send it to to influencers so If you bought on the website, he would look you up and say, okay, you're an influencer. You're going to review. I'm going to send you the actual product.

1:02:53But then anyone else, he would just be like, oh, you're on the backlog. You're delayed. So in February or in June of 2020, just three months after Marquez Brownlee launched that video, the FBI came to his studio and collected his phone as evidence. So MKBHD keeps all the phones that he reviews. And they say, hey, we'll give it back to you. And then five years later, that guy who was scamming people pleaded guilty to fraud. He faces millions of dollars in fines and 20 years in federal prison. So what a crazy story. I wonder how many of these - Citizen journalism. Did they figure out how many of these fake phones he actually sold?

1:03:32I don't know. I mean, I imagine if it's millions in fines, he probably sold millions of dollars worth of phones. It was a very expensive phone because it was wrapped in gold foil. And so it was marked up, I believe, pretty heavily. Or maybe the strategy was to underprice it or something. The CEO said they sold over 50 ,000 telephones. Anyway, let me tell you about public.com. Investing for those who take it seriously. They got multi-asset investing, industry-leading yields, and they're trusted by millions. Go to public.com. So just to put this into context, that means they did potentially over$17 million in sales of a phone that was never delivered.

1:04:13Yes. Yeah, pretty crazy. And this was Pablo Escobar's brother? I believe he was like riding on the Escobar name and kind of a distant relative, which is a funny brand to align with because it's like, trust me, I'm related to a criminal. And he wound up being a criminal as well. Pablo Escobar objectively is one of the greatest CPG entrepreneurs of all time, right? Bootstrapped to tens of billions of revenue. Yeah. global empire just chose a product that happened to not be very legal. Yeah, very rough. Mert from Helios, friend of the show, says, in eight hours, Solana blocks will become 20 % bigger.

1:04:55Put differently, you can now do more transactions per second on the Solana blockchain. Very exciting news. And hopefully we'll be talking to Anatoly in just a few minutes. He is waiting to come into the show. We are sorting out some audio stuff. In other news, Buko Capital Bloke says, the people who create HR slash security slash compliance training videos and platforms without an option to 2x video speed should be sent to Bukele's Dungeons in El Salvador. I think that's a legal thing, right? Like you can't – You should be able to do 2x speed though. Yeah. What's funny is Aaron Erikson here says, startup idea.

1:05:37AI agent that watches this stuff for you. Controversial post because people have done that and got in a lot of trouble. If you have to legally watch a security compliance video and you don't watch it and you have a computer do that for you, you're in a lot of trouble. We got them ready. Oh, we got Antoli. Welcome to the stream. Sorry for the technical difficulties. Great to have you here. Thank you so much for joining. How are you? Great. Thanks for having me. Fantastic. Great to have you. I would love to kick off with just a little bit of the State of the Union. I know a little bit of the history, but how do you tell the story of Solana and what you built?

1:06:14And then I want to go into the State of the Union of all the different projects and all the different applications, and then we can kind of weave through some of those. But could you kick us off with just a little bit of an introduction? Sure. So I'm an Atoli. I'm a co-founder of Solana Labs, CEO of, and I'm kind of the brain child of the Solana protocol. It's my baby. I started around 2017, 2018 with the dream of building a high performance blockchain network. So it took us about two years to launch it. I don't know if folks remember 2020, but we launched right after the double black swan. COVID was just effectively became a thing.

1:06:56S &P dropped 70%, Bitcoin dropped 70 % and our launch date was announced to be three days later. We launched into the abyss, but I think it was like a lucky timing, I would say, because we were the only next generation network out. And we're still, I think, processing more transactions than every other major blockchain combined. So not just more than every other one, but more than all of them combined, which is a pretty fun stat to have. Yeah. What gave you confidence that you needed to build a new L1? There was a narrative at the time of just like, hey, we have existing networks, Bitcoin and Ethereum.

1:07:36Let's just speed up those. And that feels like, I mean, you can even abstract that all the way to, hey, like the fiat transfer rails, the Swift system is slow. Let's just speed it up. We don't need an entirely new network. Like what gave you confidence that you would have like an enduring advantage in speed and the other platforms, whether in fiat or crypto, wouldn't just catch up? Why do founders do things? That's the question I'm asking. Sorry. The real reason is because I could. I spent most of my career doing high performance like optimizations at Qualcomm. So I kind of knew how silicon moves data.

1:08:17like as an act that that's my expertise and i knew the five to ten people that i needed to hire most of them came out of qualcomm and we were just laser focused on building something that is optimized um to like its full potential um i don't know if it's gonna like how it's gonna revolutionize finance or anything like that we we did have the idea that these systems are useful for trading because if you kind of take a step back and try to look at like what does it mean to synchronize the state and agree in a ledger globally um it's a similar problem to like how nasdaq or nasa do price discovery they're trying to figure out what's the price of an asset given all of this crazy information that's happening around the world that all the market makers and traders are trying to send to their um machines that through their trades to to this exchange just running in New York or whatever.

1:09:13But with a blockchain and this kind of unique technology that was developed really like over the last two decades on the city of Byzantine fault tolerance and agreeing in a piece of state globally without a trusted third party, you can do the same thing, but you don't need to trust Nizia or Nasdaq to do it. So I kind of had the sense that the purpose of these systems is to do price discovery. I'm not an exchange person. and I've never built a financial product before, but I knew how to move data through these systems, how to synchronize them. So we looked at it as a physics problem, and this is how we tackled it.

1:09:51And we thought the things that are going to take advantage of this would be like real financial applications, stocks, all these things. But what happened was NFTs and meme coins were like the first big booms. But it's the same technology and the same infrastructure is used For those, you can now leverage for real equities and real global price discovery without a third party in the middle. Yeah. Do you think that the that like the I see like this tech tree developing with all these different applications and every few years a check in with crypto and there's different there's different narratives, whether it's NFTs and meme coins.

1:10:30Now it's prediction markets and stable coins. And for a while there was gaming and metaverse. And it feels like the detractors of crypto point to that as like a failure. But I'm not sure that it's any different than what's happening in enterprise SaaS with ideas that get tried and fail or space exploration with ideas that get tried and failed. Do you think there's something like fundamentally different about the way the crypto community explores the possibilities and the applications in the tech tree? Or do you think it's just done more transparently because everyone can see it or there's more tension on it because of the financial element?

1:11:06Like, what do you think is going on? I think, you know, I grew up in the 90s. So I was there like in AOL chat rooms and RC chat rooms. And the same kind of like almost religious ideas about what the internet can accomplish were around then. You know, people were arguing about what file system they'd pick in their Linux kernel configuration. What religious zeal on forums, like to the death, right? So I think when you have a technology that has this unforeseen changes that people can imagine could be unlocked by this, you can almost have this faith-based argument to it. The difference now, I think, from the 90s is that we have the internet.

1:11:48Everything is global. So we're seeing the iteration cycles just to be insanely fast. And while a bunch of people tried ideas two cycles ago, that was 2020. and now we're starting to see those ideas get traction i think um the fact that there's 250 billion in stable coins that have been issued before congress took up legislation is insane like there's no way i would have predicted this in 2020 like it's just incomprehensible to me and it seems like very realistic that we're going to see over a trillion dollars in stable coins by the end of you know like in the next four years and people telling me that i'm bearish when I say only one trillion.

1:12:33Is that stablecoin rollout? I feel like for a long time, the whole idea of like Bitcoin was actually in some ways a threat to the US dollar. And now it feels like with the latest White House announcement, like crypto is actually strengthening the dollar in America. What's actually going on there? It's like saying that gold is a threat to US dollar. I actually think that's kind of the wrong way to think about it. But I'm not a finance person, but I learned through osmosis. So Bitcoin can't function as a U.S. dollar because for you to use something for trading, like trading oil or whatever, you need futures.

1:13:11You need a stable denominator to price those futures in. And if you're using Bitcoin, you have to pay for shorting Bitcoin for the duration of that contract. Imagine taking out a mortgage in Bitcoin and having to pay for 30-year short position on that Bitcoin. right this just doesn't work yeah and two like banks can you know if when when i have like an oil field and i want to buy you know whatever a farm with it i can use the oil field as collateral the bank will create dollars up to the value of my collateral to facilitate facilitate that transaction you can't do that with bitcoin you can't do that with gold which is why these things don't function as like real world trading currencies but they were great as a store of value as a hedge against, you know, sovereign events, a whole bunch of other stuff.

1:14:01You know, fiat is managed by people, which is the great part and also the scary part. And gold is not managed by people, it's managed by physics. Bitcoin is managed by the rules that create, you know, enshrined in this protocol, which are bounded by physics, which is what gives it these really nice properties. But I don't see a world where Bitcoin weakens the US dollar. If anything, the growth in stablecoins has been as a spot. The vast majority of the growth in stablecoins has been used to support spot markets for Bitcoin. So when those grow, you have more stablecoins demand, more treasuries are bought.

1:14:39It's actually, if anything, has been working to benefit the U.S. dollar. Do you think there's, do you see any potential needs for new stablecoins? Obviously, with the new legislation, a lot of people, big institutions, seeing dollar signs, thinking, oh, look how much tether's printing, look how much circles making. We should get in the game. They should. Competition is the best way to grow products to iterate quickly. They 100 % should, right? Like, why wouldn't you? This is, I think, you know, a lot of the U.S. finance was built after World War II. But when you look at India and China that has been built after the Internet, their financial infrastructure and rails are just much, much faster and more robust.

1:15:27Which is crazy to say that like what we consider to be the West, the advanced economies are actually sitting on rails that were built before the Internet. So on that note, what about the other side of this is institutions saying, let's say somebody wants to tokenize securities and they think, okay, well, if we want to do this, we should make our own L2 or we should make our own chain. That feels like potentially a distraction from delivering the end value, which is trustless, fast transactions globally. So, you know, I have my own horse that I picked, obviously. Yeah, the goal is to not meet. Yeah, so I guess what I'm pushing on a little bit is like, you know, competition for you, you know, or competition.

1:16:20But not for me. But at the same time, you guys have faced so many competitors over the years. You launched into a competitive market, so you're not necessarily afraid of it. You know, the worst thing would be for the government to stop innovation. Totally. Like that. So like if those folks think that they can outcompete us by launching an L2, let them try. My goal is to show that like, hey, everything that you want to express in those smart contracts running in your L2, you can do it much cheaper and faster on Solana and you'll actually grow revenues faster without needing all this info. This is what, you know, my pitch to those companies, but let them innovate and try.

1:16:57like happy to yeah it's also it's a if you if you choose to build on solana you just don't have to worry about onboarding developers you don't have to worry about liquidity all these things that you would have to really be uh would have to be top of mind if you want to go compete and potentially that's just a huge distraction from delivering the end product experience that that that's that's the dream we want to you know we want to convince people there's no point to build their own private intronets just use the one public giant internet that can host everything what about uh what about developer activity there was i i remember during 2021 and 2022 there was billions of of value creation billions of of uh of um you know at least in usd terms uh on chain and there was still at some point i remember there was like single digit thousands like true blockchain engineers?

1:17:55How are you sort of tracking developer growth internally? I think if you look at the reports that have been published online and developer growth, Solana is the second largest at this point and the fastest growing network. Ethereum is the largest, but I think we've been outpacing in terms of growth, the number of developers onboarded. So like everything else, We saw in the early days, there was a lot of demand for smart contract developers because there was a lot of innovation to be had in the smart contract land. But I think of them more as drivers. Like Windows has excellent drivers, but you don't see innovation or that being a differentiator anymore because there's only so many smart contracts that actually get built.

1:18:43A stablecoin is a token. So all the innovation to launch a stablecoin is happening, not in the smart contract space, but the integration with the bank, with how you handle the mint and the draws and that stuff. So a lot of the work has shifted away from L1 smart contract development to the higher levels. And this is where I think Solana has really excelled because a lot of our focus has been on building reusable contracts that actually get deployed once and reused everywhere. There's really only two token standards in Solana. It's not an interface. There's one implementation called SPL token, one implementation called token 22, and everyone uses them.

1:19:22So they've been effectively, through Linde, verified to be correct and bug free. And you as an organization don't have to take smart contract risks when you deploy these. Is there any hope that, I don't know if there'll be a new bull market, new smart contract development, but it felt like developing a secure smart contract with super high stakes. You needed an insanely good engineering team, insanely good engineers, hold all this crazy context in their head to understand all the different risk factors. You needed to reject 50 North Korean engineers. Yeah, it was not something that you just wanted to like, you know, like hack together on a weekend.

1:20:01But I was wondering if like the AI code tools would be an accelerant there. But at the same time, I've heard a lot of stuff about AI code gen tools being like kind of falling short when you need a huge context window, you need to hold a bunch of things. So like, is it smart contract development, like the last thing that AI will be able to do in terms of code gen, or is it actually an accelerant, but we don't really need it because we have all the smart contracts we need? I think the way that I've seen AI used effectively is augmenting an expert. So you have somebody that has deep expertise, they can just do more work.

1:20:39And where smart contracts could excel is like, you're right, where AI can accelerate smart contract development is, you built a complicated smart contract, you need to formally verify it. The formal verification stuff is a huge pain for, and you need a lot of expertise, but you can automate a lot of the boilerplate and generation if you have an expert guiding at AI to go do that work. But you still need, I think, your core five or six, super really great smart contract engineers, they can just accomplish more with that small team using AI tools. Did you read anything into the OpenAI and Google news about the IMO gold medal and the use of Lean to do complex math that feels somewhat adjacent?

1:21:22Is that relevant at all? Yeah, this has been like, you know, I've had 15 plus years of experience as an engineer. You know, I love being an IC. I love churning out code. but now my workflow is using Claude and O3 and Super Heavy to go validate each other, come up with like a robust plan, robust testing plan. And the jump in reasoning abilities just year over year, just from last year has been astronomical. Like it went from being kind of a, I'm going to try this as a gimmick to now it's really an embedded part of my workflow. I can't imagine myself building software without having these tools now.

1:22:05Do you think those tools are, yeah. Yeah, I still don't trust the code that it generates. I think this is where an expert that can have this big context in their area of expertise can quickly scan stuff over and guide the tools to go do more testing and target more things. You still need those people. But their reasoning capabilities are getting stronger. Yeah, it's scary. I don't know what happens in five years. Yeah, it feels like there's kind of two failure modes in crypto projects. One is technical. There's a bug and someone just drains the wallet or whatever and hacks it. And then there's another, which is like economics.

1:22:45And people didn't realize that there was no real solid foundation. They kind of wound up with like a Ponzi and it just kind of fails that way. Well, then there's the other. I mean, social engineering is the other big one. Sure, sure. Sure. And I'm wondering, do you think that these reasoning models are getting close to superhuman, close to amazing at economic reasoning? Yeah, you're right to point at that as being a really hard problem because this is where you really need to have that expert that can understand the economic attacks to guide these tools to go build a simulation and actually test it.

1:23:22And I wouldn't trust myself to be able to go find those given all these tools. Even if I'm aware of the problems, me just asking, oh, three, give me all the top 10 economic attacks and then simulate them. That's the best I can do. And it's going to suck. Like I can already tell you. It's going to do maybe as good of a job as like a new grad. right but like i think you really need to have an economic expert guide they'll still similarly like that economic health expert if they asked all three to go generate a bunch of bug free code i'm gonna go look at it and find a bunch of bugs so i think you still need people in the loop there i don't know what happens in five years given the the pace of progress of ai but um it's remarkable of how good it is but you still need people for now.

1:24:15You guys have been really good at kind of building brand through short strings of words. So internet capital markets, the team has really rallied around that. You hear it, it makes sense, and you can think about all the ways that that evolves. With that phrase, internet capital markets, what are you excited about on a one-year timeline? And maybe what are you excited about on like a five-year timeline? I think getting fully approved, regulated, boring trad fight stuff on chain that has plans on future cash flows, like all that stuff that you read in the intelligent investor. Having those things be tokens, I think, is really, really important.

1:25:02And this is kind of my dream for internet capital markets, getting all the, you know, jumping through all the regulatory hoops, crossing out all the T's and dotting the I's, making sure this is as compliant as the government wants it to be and like minimize the fraud. because I think once you have those assets on chain, you can then use the best of DeFi, which is the purpose of it is to create transparency, right? And to run the risk engine every 400 milliseconds instead of once a week or whatever happens in a bank. Always know exactly what are the positions at risk, what collateral that particular contract is managing so everyone can see the actual risks involved.

1:25:43And you don't end up with the kind of tail end failures that you see in traditional finance. That's the beauty of it. But we need real assets to do that because if you do that with just meme coins, they don't have any value except price. And eventually that just doesn't work. All the risks are correlated and you end up with a tail of months. That's to me the short term. The five-year kind of like dream for me is I would want like a startup founder that wants to do this on their own to be able to take all this open source software and to run their own IPO, which means do all the disclosures, do the price discovery on chain, do use everything in an immutable open source smart contract and kind of do the Linux from scratch IPO and have the SEC approve it.

1:26:32And that doesn't mean that every founder is going to do that. You're going to have a bunch of services spring up on top of it that can take advantage of all this technology. But you should see the fees of doing that basically compressed down to the actual value that they provide and no longer kind of be based on their ability to control the markets or the regulatory capture and stuff like that yeah that makes no sense last question from my side how how did you were you uh conscious of trying to make solana lindy in the early days because i know it was like you know like any crypto project it was the price you know massive amount of volatility and then you'd go through these periods of insane activity.

1:27:16And I think we've seen this with various companies at the product level, you know, OpenAI, PumpFun, Axiom is the other one that's like done what, like 170 million of revenue in like six months or something crazy like that. And anytime you see these like massive spikes in activity or revenue, the question is like, how sustainable is it? And you guys were able to kind of like navigate through the volatility and get to a point where I think everybody would have real conviction that Solana will have more activity a year from today, five years, 10 years, et cetera. But I feel like that's a number for crypto founders out there.

1:27:57It's like, how do you go from, you know, massive activity to being durable and sort of Lindy? I don't know. That's a great answer. Well, and it's somewhat fundamental to the technology that when you reduce friction, when you make finance frictionless, that reduces moats, right? It makes it so that things can leave as fast as they come. I can rationally give you all those arguments, but the reality is that you're just worried about surviving the next eight weeks. This is how we operated from that launch because we simply didn't have any other option. We had limited runway for the first three, four years.

1:28:41And trying to get as many developers on board, trying to increase the bandwidth of the network, reliability, all this stuff. You're just thinking in eight-week terms. And you read every book on startups and stuff. That's actually what you're supposed to do. You're incrementally fighting for life. and hopefully all those all those sprints all those efforts all those impulses accumulate into something sustainable um and i think we're starting to see that sustainability actually play out through you know the last cycle but it's really hard to predict like if i knew what was going to be really important 12 months you know i i don't think i'd need to work

1:29:27Can you update on just the progress of decentralization? What's important to decentralize? How you increase decentralization? I remember when in maybe 2022, there was this critique, Solana's not as decentralized as the other platforms. Just judging by the performance of the company, feels like that was not a bear case that stuck around. But what's actually happened with the company? What plans did you put in place? And what do you feel like you've executed on properly to kind of satisfy the level of market demand for decentralization? Because I know that there's trade-offs here. I mean, I look at it as a kind of key product, like feature of the network.

1:30:16I grew up in the 90s. I loved Linux. This is how I learned how to code, is hacking through Linux code. Huge believer of open source. Every part of the network, you can participate permissionlessly. This is kind of the key part. And to me, that's really important because when you think about disrupting finance, if you disrupt settlement, you've disrupted DTCC. The number of people that can even know what that company does, you can fit them all in one elevator. But everyone can name the exchanges, the brokerages that they use. and this is where all the money is made in finance. So if you cannot permissionlessly participate in the formation of order and price discovery, I think you haven't really decentralized the important part of finance.

1:31:03So Solana has always been about decentralizing how blocks are made, how transactions are ordered, the sequencer, which is in contrast to, I think, a lot of the direction that Ethereum went, where they have done an amazing job decentralizing settlement, but they've offloaded all the work of actually running the financial applications to layer twos where you now have like a single sequencer that runs the vast majority of layer twos. And that sequencer has the same functionality that Robinhood does. So to me, you haven't really decentralized the important, interesting part where money is made or revenues are made.

1:31:44so from my perspective Solana has been more decentralized than a lot of the networks where it counts over the last three years where we have taken a trade-off is that the cost to run Solana validators is higher because we are focusing on increasing throughput making sure that anyone can participate in being a block producer and ordering transactions and all this stuff the cost of those validators is higher but over the last four years we haven't had to increase the cost to run the boxes because of the software performance improvements so if you actually look at the network how it's run, how it's operated I can't fire anyone that actually works on it like Labs doesn't actually employ anyone that builds the protocol anymore we're primarily focused on building a phone yeah, that makes sense I wish we had time to talk about the phone but we'll have to have you back on again very soon.

1:32:41This is great. Thank you so much for joining. Really enjoyed this. Yeah, appreciate it. We'll talk to you soon. Thank you, guys. Have a good one. Bye. And let me tell you about 8sleep.com. Five-year warranty. 30-night risk-free trial. Free returns. Free shipping. Get a pod five over at 8sleep.com. Well, we have Tyler and Michael have secured burgers. Oh, let's go back to Tyler. Let's go back to them. How are we doing? How's it going, guys? How are you doing? Yeah, so they're playing Star Trek behind me. It's really loud. Okay Got the burgers here. Okay I Think there's a couple ways you can rate burgers right to this price point.

1:33:21Yeah, so it's sitting at I think it was like 1450 around It's a little pricey so that means like you got to rate it against five guys shake jack, right? Sure It's kind of in that tier. Yeah, so I think it's solid. It's very similar to shake jack. I would say okay Am I going to be going here? Is this my go-to burger? Probably not. I'm more of an in-and-out kind of guy. I'm not very wealthy, so I've got to go cheaper sometimes. Not yet. But there's some cool stuff. Apparently, the fries and the hash browns are beef tallow. Okay. Wow, let's give it up for tallow. Very good. Generational run. So, yeah, definitely a feel there.

1:33:58Apparently, they're really good. I think I showed you guys this earlier, but there's the little boxes, right? Yeah, yeah, yeah. They're cute. Are there any protests going on? Is this a controversial space? I have not seen any protests, actually, surprisingly. Seems like any Grok integrations. Oh, yeah. Any Grok integration. The only AI I've seen here is the Optimist. Talk to me about the actual ordering experience. Did you order from a tablet? Did you order from a human? So, yeah, I think we ordered from a human. It was like totally normal. There was no kiosk or anything. But you can order from your Tesla.

1:34:33We didn't try that. because I think we had to kind of expedite the process. But apparently you can do that. But yeah, honestly, I would have liked to see some more optimists here. Just doing popcorn, you know, it's not super hard to do. And then there's like a nice bar over here. Hey, got to start somewhere. Popcorn serving. Yeah, overall pretty good though. I'm not sure I would come back. It's like specifically for the burger. For the burger. But it's definitely a cool place. For a charge? For a charge though? If you need to charge. I wonder if the charging is free if you get the burger. I think the charging is free.

1:35:09Okay. That makes sense. That would be a big draw. Yeah. That's very exciting. But, yeah, there's like 80 cars here. But, yeah, overall, solid experience. Very cool. Very cool. All right. Well, enjoy your burgers, and we'll see you back in the studio. See you back in the studio, Tyler. Thanks for doing the on-the-ground reporting. Awesome. Well, next up, I believe we have Aaron from Fox in the Restream waiting room, so we can bring him in when he's ready. And in the meantime, I'll tell you about Aaron Levy from Box. Welcome to the stream. I was getting near and had to read, but you are here live.

1:35:43Thanks for hopping on. How are you doing? Hey, good. How are you guys doing? It's good. Your office always looks so cool. You were one of our first guests a couple of months. Wow. I mean, I'll take all the credit for your guys' success. There you go. Take it. Take it. I want to talk about AI, obviously. I want to talk about, quickly, I mean, let's start with the consumer side, and then I want to go into enterprise. Did you have a chance to read Fiji Simo's outline of where OpenAI is going in consumer? Kind of six categories, health, knowledge retrieval, creativity, what are you using personally?

1:36:25What's your daily driver? Are you using it with your family? have used it for health. What's just been your general experience with consumer AI? I think I've hit every one of those use cases. It's, I mean, it's, yeah, therapist, doctor, branding assistant. What about executive coach? I thought that was an interesting one. She says that she has a coach. Obviously, you're at a level where you can afford a full-time human coach, but what goes into a good coaching experience at the executive level? Yeah, I mean, that's a fantastic use case. Probably actually I should be using it for that. I'm sure anybody that works with me wishes I did.

1:37:03But often you're bringing a coach, a set of problems that you're running into. I've got this org design issue. I've got these people that aren't collaborating well. And oftentimes it's most actually just helpful, just the purpose of talking about the issue and then laying it out that gets you thinking about it. and then seeing a set of options. So I think Chachamiti would obviously make a fantastic coach for that because just the very active of just getting you to be able to describe the issue and then seeing a set of options would be helpful in like 90 % of situations. So I think those use cases are great.

1:37:41I mean, the health care stuff alone has been extremely powerful. I seem to run into one new ailment per week. And I used to go into urgent care way too frequently. Like, oh, I have some sensitivity in my finger, and then I'd go to urgent care. And so now just Chachabuti just solves all those issues. You are a hypochondriac. Anyway, on the org structure issue, what you would go to an executive coach about, Do you feel like AI should be more of like a product team or more of like a service team? Like, do you see it like it's my infrastructure organization or it's my front end engineering packages versus it is a new product that will sit alongside other products and maybe be sold as a different skew by the sales team?

1:38:37Because I've been going back and forth on this, but I'd love your take before I kind of wrestle with it. Is this in a B2B context? Yeah, yeah, in your business. I imagine that you're thinking right now there are LLMs, there are tools, there are APIs, there's a whole bunch of stuff. There's coding agents and IDEs and stuff. And you want to get this stuff into the hands of everyone in your organization. The question is, I bet a lot of companies, I imagine, are thinking about, do we need an AI organization? Do we need an AI product? And you can think about, like Salesforce is launching, and Copilot is like a separate SKU almost that sits alongside.

1:39:20Gemini is a separate subscription. It's a separate product, separate app, versus just, hey, we do X, Y, and Z. And all of that is made better by AI in the same way that, you know, memory caching, like Redis, the database made things better across everything. But it was not its own product. Yeah, I think in our case, and I think so much depends on the type of product that you were in before AI happened. Yeah, totally. In our case, what we did was our ChatGPT moment was, you know, same time as everybody else. But we did, you know, fortunately step back and we said, OK, if we were to found our company in a post-AI world, how would we how would we treat the very kind of business model, the product experience, the actual value proposition?

1:40:08And we started kind of laying that foundation. Now, it took at least a year and a half to like execute in the way that even looked like what you could do from a blank sheet of paper. But the idea was if we were to start Box in 2025, let's say we ran this exercise now, you know, we wouldn't think about the AI version and the not AI version. We wouldn't be like you go over here to do AI and everything else is sort of like dumb. And so we decided to bake AI into the core value proposition of the product. And that shows up in the kind of obvious ways you'd expect. So let you talk to your data, ask questions of documents, take a collection of content and be able to query it and ask questions of it.

1:40:45And then now in a world of agents, that's also showing up as agents being able to kind of run around and execute automated tasks for you. So a very natural one for us is a lot of customers have all of this unstructured data. It could be contracts or invoices or marketing assets. And they want to be able to read what's in that data and then basically apply structure to unstructured content. So take all of the variables from a contractor, from an invoice or from a digital asset, and then let me have effectively intelligence around that. I can ask any question of that data now. I can automate any workflow around it.

1:41:18I could have a dashboard that shows me exactly when the different contracts are up for renewal or which ones have risky clauses in them. And so this is fundamentally changing the entire value proposition of our product. It gets us more in line with what our holy grail was to begin with, which is like let you tap into all of this information that you have in the enterprise. But yeah, so this is built directly into the core fabric of the platform. And most product managers at Box at this point are working on something AI related. There's sort of nobody that is unaffected by AI at this point. How do you think about AI surfacing itself in the product in a, I don't know, like above the UI or below the UI?

1:42:01I don't know, front end or back end. Basically, like I feel like I love when a company I use is saying, hey, we're going all in on AI. But I don't necessarily want a chat box with every product. What I want is, hey, you guys did the hard work. You guys did the prompting. And all of a sudden, search just got better. Or if I want to just click a button and see it as a CSV, boom, it's there. Or it's already been transformed. It's already in a database. It's already built for me instead of me having to actually say that. But it's really hard to define those things. So how do you think about, like, is this going to be the prompt is in the hands of the user or the prompt lives within the Box engineering team?

1:42:46Yep. You know, it's entirely case by case probably across all of software. Even in our product, we have variants of both of those. So we have a product experience where you can chat with your documents, which means it's another chat box. We have a product that is, for all intents and purposes, you never have to even think about it as AI. It's data extraction from a document. You press, you know, extract data. You don't chat with it. And then the agent is, you know, a set of context, a set of tools that we've effectively tuned to that workflow. but you don't have to think about it as being AI or not.

1:43:21And it's not obvious which one wins out. I think there's a variety of ways that people will expect to use these use cases. The thing I'm probably most fascinated with and the thing that will have the most GDP impact is effectively these background agent use cases. When AI first started emerging really just due to Chachubiti, we thought about AI as, okay, I go into a prompt window and ask a question, I get an answer back. or maybe a year prior to that, you had GitHub Copilot, which was really kind of type ahead prediction functionality. In both of those cases, the AI can only work as fast as you can.

1:43:57And so that means our productivity gain was still capped by how fast we could work. With background agents, you're now like there's no, you're untethered in terms of the ultimate upside of the amount of productivity gain, because it's just, you know, if you have a well-described prompt that could, you know, basically translate, you know, days or weeks worth of work into a couple minutes for the agent, then that's, you know, obviously like that could be 100x productivity gain. And so we have a number of features we're working on that more relate to background agent type use cases. And that's, I think, what's most exciting.

1:44:33So when you think about Cursor's background agent, or you think about Codex, or what Replit does with its agent, that's way more of the future. And that's, you know, probably more what we're putting our efforts into. Yeah, I feel like it's like it's less flashy and it doesn't get as much attention for some reason, but and it's harder It's harder to make AI just a button and it works But when it when it when somebody can boil that down to click button get result like it's just that's the original magic of software and I think well But I you know, I think it's it's still probably click button but give some instructions to get what you want But it's not as open-ended as it was before which is just here's an empty chat box do anything in the world Yeah, this will be an interesting tension between the let's say the most horizontal frontier players versus the applied, more vertical specific use cases, the vertical players and it could be industry vertical or some other kind of vertically integrated product.

1:45:27The vertical or domain specific players will will be able to pre pack in the use case. And that will actually be in for many enterprises, the easier to adopt capability. Instead of like the super horizontal, we can do everything, super intelligence, understanding the use case in the domain will likely improve the efficacy and ability to drive the rollout within the enterprise. How are you thinking about Box as a, is it a garden that you welcome people to? Is it a walled garden? Is it my data? Would you be okay with me taking my data and sharing with us? Well, yeah, it's just like you guys are building a bunch of workflows on top of your data.

1:46:08But then I imagine you're happy to, in some cases, let other companies come and make the data more valuable. What's your stance on letting the fox in the hen house? We've been going back and forth on that. Sometimes people say, it's a friendly fox. Let it in the hen house. Yeah, we're a public park. Anybody can build their garden here. That's great. No, I mean, I don't want to read between the lines too much, but I assume this is about some of the closed off data access, you know, pipe data walls. Yeah, yeah. Tricky topic. So for us, I'll just say our general biases. Well, so first of all, to be unbelievably clear, any data within boxes are customers' data.

1:46:53So just, you know, get that out of the way early. And in general, we bias toward a very heavy degree of API access into that data. Cool. There's a couple kind of complicated issues that are the emerging topics in the agent world. And this is obviously probably germane to the origin of the question. There are some systems that do these kind of heavy indexing of data tasks. And in those areas, you have a couple issues. One, you just have a cost that the kind of originating vendor then incurs based on that indexing. But then also you have these open questions around security and permissions and data governance because you're trying to keep permission models in sync between multiple systems.

1:47:35And so all of a sudden, the brand promise of one platform can possibly be sort of mutated by another platform if there's a data leak event or whatnot because your data was indexed by that other system. So these are going to be more, I'd consider them more open questions that need to be figured out by the industry. I'm more a fan of this idea that you'll have any system that today sort of wants to index all of the information from another platform, a sort of more ideal architecture would be that that system just talks to the system with all the data and says, hey, I'm looking for this particular piece of information.

1:48:13Can your agent go run off and find that for me and then pull that into that other agent? So the problem is that everybody's agents today operate differently. We all rank the information differently. There's different latency. So that's why you kind of need indexing right now. But I think in five years from now, I think it would be a little bit of an architectural anomaly if you had to index all information like 20 times across the Internet. That seems like that would be very inefficient. And then it would lead to lots of security vulnerabilities. Because let's say I share a set of documents or a communication channel with you, and then that gets indexed by some AI agent system.

1:48:52But then I unshare that with you. how quickly does that propagate to the agent system that was indexing that data? And now we have a security risk that that first system has no control over. So these are the very real kind of issues that you run into in enterprise software, which is why I also think that you're going to continue to see so much power in the systems of record, even in a world of AI agents. I have two questions. When you speak speaking of agents specifically, there's billions of dollars being poured into roll ups to acquire maybe more kind of Main Street style businesses, let's say an accounting firm.

1:49:30And the promise there is basically that we're going to that that the investor group is going to be able to an operator is going to be able to dramatically reduce headcount and increase the earnings potential of the business. Do you think that there's enough excitement around agents and agentic workflows that these companies have been able to raise billions of dollars? Do you think that agents are already at a point where they are going to be able to kind of deliver on that on that thesis? Or will companies like there's two paths like companies can leverage off the shelf agents, maybe coding agents, for example, or they can build their own agents or workflows.

1:50:13But how how how ready are these businesses to be able to strip out, let's say, like 50 % of headcount? I think that depends on on do you want to grow revenue or are you fine with the existing install base of that company? And and then, you know, likely some kind of terminal value that you can extract out of that. There's no question you could cost you could cut costs with with AI agents in a number of spaces. But, you know, if you had a very competitive industry where the product roadmap really mattered, you know, I think you'd probably be more in the camp of using AI agents to accelerate the work that you're doing as opposed to cutting headcount.

1:50:54So I think all depends on the space. I haven't picked up on which industry people are most attracted to. Is it like health care, office, you know, back office firms? Is it law firms? Is it accounting firms? I think it'd be hard. You'd have to be really running a business very inefficiently if you could cut 50 percent of the cost with AI and then and then keep the business sort of maintaining its customer base. But, you know, private equity has found those opportunities in the past. So I think it's a just a more it's the evolutionary kind of update to probably how private equity will think about some of these efficiency gains.

1:51:29But but it and it has a place in the economy. It's not something that is sort of to me the most exciting part of AI. but it could work and it could produce some degree of profit. Yeah, makes sense. We have a friend who is running a startup and recently a hyperscaler launched a competitor. What advice would you give him for competing with someone that basically has unlimited resources and how do you think about getting through that as an entrepreneur? It must be an emotional rollercoaster. We've talked to him a little bit. And I've been thinking about different case studies and mental models and companies that have made it through.

1:52:13And it's fascinating. But what would you say to a founder in that space? Well capitalized, products loved, but now facing fierce competition from hyperscaler. I think there's a therapist mode in Chachapiti that apparently that you can trigger. So, sorry, is this like a known company? Can you say the company or not? Yeah, browser base. We've been talking about it on the show for a while. Okay, okay. It's a computer use agent, or it helps you, if you have an agent that needs to go browse the web, it's an API that lets you harness the computer and actually go click on things, and AWS launched something.

1:52:48And obviously that's competitive, and it's something that, we had him on the show a couple times, we asked, hey, this feels like something that AWS might launch, they wound up doing it. You know, like what would your advice be to him, I guess? Yeah. So I think actually the – first of all, if you – somebody, maybe you guys could – do you have like an analyst team yet? Does that exist? It's coming. It's coming together. We have some interns. I think you should – Deep research. Yeah. If you could run some analysis, I bet you in the history of hyperscalers launching a competing capability and then you look at that market, I bet in all cases – sorry, not in all cases, but on average, the market size kind of grows for even the independent players.

1:53:33I mean, think about how many times Snowflake or Databricks probably shouldn't exist based on some of the data services of the hyperscalers or Cloudflare for Matthew, probably like 10 years running would show the market share graph of Cloudflare versus all of the other CDNs. And you would have been like, why would Amazon or Google not wipe out Cloudflare? It's like they already have the network. It's not that hard. In fact, in many cases, Cloudflare probably runs on their networks. And yet it turns out that like hyper focused teams and platforms in big markets, you know, are usually able to carve out enough of the space because you're just going to be just more wired into the customer base, the use cases, the better APIs.

1:54:16You're going to jump on the next trend more quickly. In the case of Amazon, you know, I'm guessing that Amazon won't work as well with OpenAI where browser based can work with all of the model providers. So there tends to be, with very few exceptions, plenty of moves that you can make in these kinds of competitive situations. And in a lot of cases, it just actually, again, lifts up the awareness of the market, makes it more exciting. You tend to not want to be in a market by yourself unless you're like NVIDIA or something. Like for the most part, being in a dynamic market where people are comparing options and looking at SDKs and choosing things and talking about stuff, that usually actually grows the market as opposed to makes it more zero sum.

1:54:59So I would be bullish in that scenario more than anything. Do you think like 10, 15 years ago, you or the chattering class was saying that about cloud storage broadly? Oh, saying that we would be crushed? Yeah, saying that you would be crushed or did anyone clock it correctly? Because what actually happened was not, you know, oh, so one wins, one doesn't, one in, like there's multiple independent companies that are doing well. There's also hyperscalers that have grown that like, it's been this weird, like oligopolistic positive sum market. It's been fantastic for all parties involved, basically.

1:55:38Yeah, no, no, nobody, nobody got it right. I mean, our first funding round was, I think our valuation was$240 ,000. Let's hit the gun for that. Congratulations.

1:55:58So, I mean, basically the idea was like storage was a commodity, but what people didn't realize was, no, it's the software above the storage that matters. And so assuming that you're in a market that is going to grow, again, zero-sum markets are very different. But it's hard to find zero-sum markets in software right now unless you're, like, deeply legacy. If you're in a fast-growing market, there's usually a move. There's usually some range of motion you have to differentiate, to specialize, to go into a vertical, to build more software on top of the commodity infrastructure. Sure. So I think we're in a period right now where you can kind of be generically bullish on on just like categories broadly.

1:56:37Now, maybe we don't need the 20th kind of language model. You know, that is there's some asterisks to this. But if you're if you're one or two in a category right now and a big player enters, I think you just ride it out because because it's just too early at the moment on what's what's happening. Yeah, it seems like a wildly different market dynamic in the enterprise in B2B than in B2C, where it's very much winner take all. And if you gave me$100 million to start a direct competitor to ChatGPT, no, it just doesn't matter at this point. Because there's so much energy and aggregation and network effects and all sorts of different things.

1:57:16And yet at the same time, I'd say every vertical in the enterprise is up for grabs. I think there's no vertical where you can just completely call a winner. I think there's some cases, maybe these like healthcare providers that do transcription and like maybe you can call a winner in that space. But for the most part, you know, when I talk to CIOs, they're in the very earliest phases of picking their first set of platforms for these major use cases. And so we have a window right now where you'll have many, many names that emerge that will be, you know, a hundred times bigger in five years from now.

1:57:48Totally. Yeah, and to the earlier point we were talking about where like AI can be not just a new product for a company that's existing, but just little improvements to everything that exists. And that winds up being a bunch of different vendors, a bunch of different internal projects, a bunch of different APIs, a bunch of different code revisions. There's so many different areas to just squeeze out percentages of improvements. So exciting stuff. Last question I have. You've seen a number of different cycles. you started the company in 2005 so you've you've uh i i do i do look that tired yes but uh but i'm curious we were joking about a week ago that there's a lot i mean there's so many different like signals blaring that that uh obviously there's a lot of excitement a lot of new capability coming online new intelligence being birthed at the same time we have ethereum treasuries and SPACs for companies that maybe shouldn't be SPACed.

1:58:47And NFT profile pictures coming back. How do you do? I'm not asking you to like call the top or call the bottom or anything, but but do you have do you have any type of outlook? Do you have any kind of concerns about about how overheated things could be right now? um weirdly i'm i'm i'm pretty um i tend to be pretty much on the conservative side in general uh and it's lost me lots of money um by by just you know not believing in like 80 of things that have worked out um uh but for ai and then maybe this is a counter signal actually so maybe this is like the jim kramer effect but but uh in ai i i i don't think we have reached anything close to close to any any peak of anything.

1:59:36Now, sure, you might be able to point to specific categories that are that are a little inflated. But when you look at I mean, this isn't this is a an economy altering technology. And if you just look at it as any percentage of labor, you know, as as as one way to model the potential size of these markets, we're in still the very earliest phase of what this looks like. Yeah, in 1999, if you were like, you know, these internet companies are really potentially overvalued, you maybe looked smart for a little bit, but over a decade, you looked pretty dumb. But if you even took 99 business, if you did a 99, you know, kind of like extrapolation from any of these valuations, OpenAI would probably be worth$5 trillion.

2:00:26Like these are not, I mean, if you look at like those companies were valued on, you know, 100, 200, 300x revenue for the mature companies. So like we are not even in 99 territory on the amount of kind of, you know, hype and excitement that there is. So I'm pretty bullish on where things go from here. If you look at something like the enterprise software market, it's a few hundred billion dollar market in just like, let's say, the US and total spend on enterprise software. There's no reason why this isn't a doubling or a tripling, even in the conservative scenarios, when you can have AI agents go and do real work that is tied to the software, because you just have so many more categories that the enterprise software can sell into.

2:01:11do. And there's just examples everywhere. Like the enterprise software size, the category size of enterprise software for the legal industry, like five years ago, was about a couple billion dollars max. Like that is all that the legal industry spent on enterprise software. And in a world of AI agents, I think you could easily underwrite 10, 20, 30 billion dollars being spent on just AI agent automated work in the legal space. So you could have a lot of these spaces that become five or 10 times larger than they were in the pre-AI era. And I think you could do that across every category of knowledge work for the most part.

2:01:47Makes total sense. Last question I have is, how tolerant do you think CIOs and executives broadly are going to be of the, like, you know, if you're thinking about, you know, Apple, for example, thinking about implementing an LLM, it's not an app, Apple's nature to have imperfection, right? Like they're not looking for something that even one out of a thousand times is going to say something that's offensive or anything like that. And we were thinking, you know, we were talking about Grok, Grok obviously is on a wild trajectory. They're closing, you know, they closed a massive partnership with the DOD last week.

2:02:33And that's partly, I imagine, like the Elon effect. But how tolerant do you think CIOs will be of just the kind of inconsistent nature of AI agents, you know, saying, yeah, we're going to make a big bet on a lab, even though one out of, yeah, one out of 10 ,000 times it's going to say something. Yeah, probabilistic computing is here. And one out of every 10 ,000 times you use the product, it might say something that 10 years ago you would have been fired for if you had implemented that software. Yeah, you can't make Hitler jokes in a company setting. So I would say increasingly the tolerance is going up because the understanding and the shared understanding of what these models are doing is going up.

2:03:22And so what's happening is companies are putting in sort of informal or formal guardrails on the fact that these models are probabilistic. And that will sometimes take the form of just an internal policy that says, okay, the AI agent is going to give you your first draft, but your job and your responsibility and your liability is to ensure that what you're putting into production or the legal brief that you're writing or the marketing asset that you're creating is fully reviewed by a human. And that is the switch from maybe two years ago, we would have looked at a model's kind of inaccuracy as a flaw in the AI.

2:03:59And now we have sort of accepted that it's just a part of the workflow. I walked by an engineer's desk a couple weeks ago, and I just said, hey, what's your now new common practice of doing work? And he kicks off an agent to go write a bunch of code. And then his job is to go review the code and fix its errors and clean up syntax and whatnot that might be different. and that's just an accepted reality. Whereas maybe two years ago, you would say, hey, this model's producing a bunch of, you know, kind of garbage code. It probably was garbage because it was just where the models were at, but you would have seen that as a flaw as opposed to that is just an accepted reality of how these systems work now.

2:04:39And so with that, I think, evolution of the understanding, then you start to flip the mindset, which is, okay, I can now use AI to give me the first draft of the thing that I'm doing. It can make the first draft of the blog post. It can auto transcribe the doctor note. It can generate the first draft of your board deck. Yes. And then we can go through it and review it and clean it up. And that still saves you 10 hours. And so that recognition is now kind of rippling through most organizations. This is why this idea of going AI first is kind of picking up steam. I talked to a lot of CEOs and CIOs in the Fortune 500 kind of universe and this is the most easily the most exciting and animated You know kind of topic that has existed in 20 years in technology Yeah, they know they're much more likely to be fired for not Like being like at the forefront of this wave then like having a model that goes off the rails and offends a user Yeah One out of every hundred there is absolutely more room for ask for permission in this phase of the kind of excitement, you know, kind of adoption curve that we're seeing.

2:05:52Totally, totally. Well, we've kept you too long. Thank you so much for joining. Great hanging, as always. Ask for forgiveness, not permission. Yeah, yeah, yeah. Ask for forgiveness. That makes sense. And thank you for catalyzing our early success. Yes, thank you. We appreciate it. That was very helpful. You're the man. We'll talk to you soon. Cheers. Bye. Really quickly, let me tell you about Bezel, getbezel.com. Your Bezel Concierge is available now to source you any watch on the planet. seriously any watch and we will bring in Neil Parikh from Slingshot AI announcing Ash. Let's bring him into the studio from the Restream waiting room.

2:06:25How are you guys doing? What's going on, guys? Good to meet you both. Kick us off with an introduction. Hey, I'm Neil, one of the co-founders of Slingshot. And I'm Daniel. Good to meet you guys. Introduce the company, introduce the product. Absolutely. Yeah. So Neil and I met about 18 months ago. We were both chasing the same vision, AI for mental health. We're finally coming out of stealth. We've built the first AI designed for therapy. You probably know this mental health crisis. Product launch. Wait, wait, also, did you raise some money? We're announcing our fundraise. We've raised$93 million to date.

2:07:03Let's go. Congratulations. There we go. Almost jumped the gun there. That's a big number. Fantastic. I was just going to share, I mean, you probably know this mental health crisis. You might not know that for every 10 ,000 people out there who need help, there's one person to offer it. Wow. There's kind of no way that we can get past that except for AI. So that's why we've worked quietly, carefully with clinicians in the loop to build the first foundation model for psychology and decided to actually help people with their mental health. Yeah. What does that mean? First foundation model, fine tuning, pre-training?

2:07:42What's the actual tech stack? building? Yeah, we train in three steps. So we do continued pre-training, not from scratch, but from a checkpoint order of magnitude. It's about 1 % the size of a data set that a foundation model lab would use. So not going to cost us$10 billion, but enough to substantially change the behavior of the model and introduce completely new knowledge. We then do alignment, which is where our clinical team adapts from the modality of human therapy to AI therapy and learns how this completely new kind of app works. And then finally, we do reinforcement learning, learning from what works at scale.

2:08:15That makes sense. What concerns you about people using traditional LLMs models for therapy today, you know, just basically going into their favorite app and, you know, you can't, you effectively can't stop somebody from starting to use an LLM for therapy, but we've seen over the last two weeks specifically that seems to be some very real concerns around um some of the people kind of it feels like they're almost accidentally jailbreaking the model winding up in sci-fi territory role playing like what are the best case scenarios it sounds like those three uh those three uh strategies that you outlined are part of it but what else is going on that that can make sure that there's always a good outcome i think the first thing with your knowledge is like chat kpt and clod are not made for therapy right like they're not trained on clinical data people can end up down these very weird rabbit holes like we've seen, right?

2:09:08Where people end up in essentially a mode collapse or somewhere that's in a very weird part of the distribution where, you know, it's bringing out psychosis and other issues and stuff. You know, they'll say that it's useful for support, but at the end of the day, what we're building is something that's very different, right? The first way to think about it is we're not trying to build an instruction following model. Like if you go to ChatGPT, it's an assistant. So if you ask it like, hey, give me a recipe for a piece of cake and ChatGPT were to ask you like, so why are you hungry? You'd be like, what the, That's so annoying.

2:09:34Yeah. It's the alignment by default question, right? Sorry? Yeah. Yeah. Alignment by default. Like the model is supposed to always help you answer the question. So if you ask it, you know, like prove that I'm a genius, it will just say, okay, of course, let's do that. No problem. And that's not what a therapist does. Exactly. If you're going to a therapist and you're like, hey, I feel angry. Help me. You know, they're going to say, so why is anger a bad thing? you know, and you're gonna be like, why is anger a bad thing? Like, I don't know. And then you get into some interesting deep territory and that's where change happens.

2:10:09But yeah, chat GPT is not therapy. They're not, I actually just saw that Sam Altman posted about this and had said, uh, yeah, chat GPT, a lot of people are using it for therapy, but it's not what it's built for. And we're looking forward to seeing a startup emerge in this space. So interesting. Um, how, uh, how excited are you guys to just like i mean effectively you're going to be running um uh just by getting this out to more people you're going to be running a massive trial i think the debate uh did the debate among um uh you know high functioning people a lot of people that listen to the show people in silicon valley in general is that like like there's almost like this question of like does therapy even work is it is it good does it put you in these um kind of doom doom loops um how how are you guys like looking to measure results right you're actually wanting to change lives but what does that look like yeah uh super uh true by the way on the controversy it's funny uh on on x uh if you neil posted today for our launch and like if you read through it it's fascinating uh how how much people find this controversial and you know it's interesting to see what's what's going on there what people are really feeling when they're uh thinking about yeah And it's, I feel like it's super common where I, I, I, my parents, when I was an angry teenager in high school, they were like, you should see a therapist.

2:11:34And they would ask me, how was the therapist? And I would say, well, I don't know how the therapist was, but it was helpful to kind of just talk through the things that were top of mind for me. So I feel like there's this question of like, can you make something that's valuable and helps people better understand themselves and the decisions they're making. But then there's all, like, how do you actually be more beneficial than just having them sort of talk through a problem and actually implement change? 100%. So we built with a beta group of 50 ,000 people to make sure that we were building this right, that we were able to show efficacy, safety.

2:12:12We've not yet published our results. It hasn't gone through peer review, but very exciting results so far. We've been able to show clinically effective results for symptoms of anxiety, depression, loneliness, increasing a sense of hope, achieving your goals. And I think the one that was coolest for us was actually increasing the number of people that you have in your support system in the real world. Are all those metrics self-reported? I fill out a survey and say, I'm feeling anxious. And then at the end of my therapy session, I say, okay, I'm feeling less anxious on a scale of one to 10. is that basically the official metric so uh phq9 and gads7 are the main ones those are the official ones for depression and anxiety but we use a number of others others like the ucla metric but essentially we track people over time um from before they start using the app through using the app and uh we measure how these symptoms change over time to see whether or not it sustains the interesting thing is i mean these are used widely across every area of psychology they're also it It has been shown, I think interestingly, that with normal chatbots, on average, after using it for a while, you actually reduce your connection to people in the real world.

2:13:18Not too surprising. It was amazing for us to see that with Ash, people actually did build stronger connections in the real world. And that just has to do with building something that's purpose built to help people with their mental health. Yeah. How excited are you guys around frequency? I think one of the challenges with therapy and actually driving change is if somebody's seeing a therapist once a week or every couple weeks, it's like very easy to kind of revert back into old patterns or forget what you had just talked about and we're trying to change. Is part of the potential here is like somebody can effectively have this like constant contact with a therapist or do you think we're the equivalent of a therapist?

2:14:00Yeah. So one thing that's important to know is that I think we're inventing a whole new modality of therapy, right? It is something like this hasn't really existed before, which means that with that comes, you know, we can basically throw out, you know, the whole nature of like, we have to have a 45 minute session. It happens once a week, right? Prices or, or like getting into a tough situation might happen at two o 'clock in the morning, right? Your therapist is going to answer the phone at that time. Now that said, we also don't want people talking for six hours on one day. Like we feel about you and it's like important for you to like work on something, probably sleep it out.

2:14:34right? Think about it, process it, continue to work on it over some amount of time because like growth takes some amount of time. Like very few people just snap and then overnight, you know, are completely different. So I would say this has been like a design evolution for us as we've been with building like early on, we noticed people wanted to use it, you know, until they solve their problem, even if that takes 10 hours and that really doesn't work. You need to at minimum like sleep, like most change happens when you sleep, when you engage in the world. So like one thing that RA.AI does that's unique compared to others is at some point it does gently push you out.

2:15:04It will say things like, you know, okay, you know, let me just ask you one more question. And then maybe this would be a good time to end for its day. And what that does is it builds a more healthy habit where you're able to use this as something that you turn to. And people do usually engage a lot more frequently than they would with therapy, but they'll engage so that they can actually get through their problem and not create some sense of dependency. Instead, it's something that you've come back to something that you're working on in a style that makes sense and a frequency that makes sense for you and not just for your insurance company.

2:15:35Very cool. I have a bunch more questions about business model and stuff, but we are out of time. Thank you so much for hopping on. Congratulations on the launch. Last question, why Slingshot? Ah, it's a great question. Our product is called Ash. Ash AI Therapy, highly recommend. Check it out. Slingshot, I think makes sense. Great as a tech company, communicate the idea of the massive change that you go through. Ash was a much more gentle, friendly, you know not too human sounding not not human sounding um and uh so it's it's been if you're still depressed tomorrow i will take a slingshot and hit you in the head hit me in the head with it all right well try ash we love your feedback yes thanks so much for hopping on we will talk to you soon have a good talk soon up next we have uh yance rimec from california forever he has some exciting news he shared it re-industrialized last week we were slammed we got him on the show today to break it down for us.

2:16:30Sorry for keeping you waiting. Jan, good to see you. How are you? Beautiful backdrop. Is that a real photo or are we looking at a virtual background? He's on a hot air balloon right now. Exactly. No, that's a virtual background. But it will hopefully be a real photo. What is it? What is it describing or what is it showing? It's showing the Solano Foundry. It's showing this advanced manufacturing park that we announced last week, which is at the edge of the new city. Very cool. Yeah, announced it re-industrialized. We were confused because we were like, wait, California or forever is building in Detroit, but you're building in California.

2:17:07We're Californians. We love to see this. Take us through a little bit more of the announcement. What exactly is coming together? I imagine there's a piece of public support, some government stuff, some financial capital. There's a lot of different moving pieces to this. What is in the announcement specifically? Yeah, a few different components. So the Solano Foundry is in the new city and it has been in the plan. We've had a big area for basically industry and technology in the plan since the beginning. That hasn't changed. What's changed with the announcement is we've put a lot more details behind it.

2:17:43And so we've put forward specific master plans, the industries we would focus on. And we've worked with JLL, which is probably the leader in this domain. I mean, JLL works for majority of the companies in the Valley when they decide that they need to go from 50 employees to 500. They call JLL and they say, hey, where we should go. And JLL has spent, as they've told us, they've spent the last 10 years moving companies out of California. And they're kind of sick of doing that. And you're going to give them another option to do it closer to home. So we're putting up JLL. They wrote a whole 40-page white paper about why they felt that actually you could do manufacturing in California if you had a scale that we have and if you are an hour outside of Silicon Valley.

2:18:31And so they published the white paper. They've joined us as the exclusive leasing agent for the 2 ,100 acres. It's about 40 million square feet of space. So it would be the biggest advanced manufacturing park in California. And then it would work as an integrated ecosystem with both the new city that we're building and with the Solano shipyard, which is about seven miles south of there. Okay. Tell me some historical examples that you're learning from. I want to hear about Apple's Fremont factory. That was an example of California manufacturing in advanced computing. Didn't go so well. Then we're seeing a bunch of promising stuff happen in Arizona with TSMC.

2:19:12what are the various lessons that you're pulling from what are the anecdotes yeah real real quick i just wanted to get so so the foundry and the shipyard i imagine we'll share like have have like be heavily connected in terms of a company that's at the ship shipyard might also have a facility at the foundry and they would obviously have uh talent living in the city like how do the How did the three kind of core areas? It's a great question. So Citi provides a place for all of these people to live, whether they work at the shipyard or whether they work at the foundry. Obviously, employee housing, huge issue in California, particularly if your people have to commute from far away, right?

2:19:53That's one of the biggest drivers that makes people quit their job is they just get fed up with doing a two-hour commute. And so we can basically provide a 10-minute commute next to the foundry, 15 minutes from the shipyard. people can walk to work people could bike to work i mean there's very few places very few places in america where you could work in a factory and bike to work um everyone keeps talking about we're gonna re-industrialize the country but nobody's talking about the reality that nobody wants to work off some random freeway exit in the middle of in the middle of texas hours in the desert exactly you drive to the factory an hour and a half from your house and then you work off a freeway exit next to a Burger King and a McDonald's.

2:20:33And somehow there's still traffic even though these places are in the middle of nowhere. Yeah, exactly. Exactly, exactly right. And so I think if we're going to actually try to get people to work in manufacturing again, you have to provide a quality of life that I think our generation has come to expect, which means you need to be able to go work in a factory and then you want to have lunch. So you meet up with your friends and you sit at an outdoor plaza and you hang out and you can do the same thing after work. and there's basically nowhere in America where you can get that other than this place.

2:21:02And so that's a component of it. And then, yes, the shipyard and the foundry are very similar. The shipyard is basically the foundry on the water. I mean, the foundry is an advanced manufacturing park and shipyard, if you really think about what that is, that's a manufacturing park on the water that just happens to have really large parcels. And so you would have companies that build ships in the shipyard and then their suppliers are in the foundry. And then we actually have a rail connection between the two. And so you can ship the components from the foundry to the shipyard. And so the whole thing is basically this integrated advanced manufacturing ecosystem where, in particular, we've learned a ton from how China does it.

2:21:45I mean, you look at the evolution of this whole thing and we started it. I mean, we perfected it in America during the Second World War. We built the global manufacturing industrial superpower. We discovered that it made a lot of sense to create these ecosystems. Detroit was powerful not because we made cars in Detroit, but because we made mufflers and brakes and tires and all of the components that go in the car, right? It's the whole supply chain. And then we did the same thing in Silicon Valley. One of the things that I learned in the course of working on this, in 1965, Lockheed Martin employed 28 ,000 people in Sunnyvale.

2:22:24in Sunnyvale 28 ,000 people can you imagine um and so um you have to have the whole ecosystem and then what happened in the 90s is China learned the lesson they basically took all of the best lessons from America and they perfected them and they built these tightly integrated hardware ecosystems in China where you have whole cities that focus on drones or robotics or shipbuilding or whatever it is and that's how they've been able to outcompete us to the point where what is it China builds 10 million drones a year, and we built about 100 ,000 in the US. China built more ships in one shipyard last year than we have in the entire country since the end of the Second World War.

2:23:06We have these entire cities that specialize in production. And our view is that if we are going to compete, particularly in the high-tech areas, we're going to have to replicate it. What was the number of Lockheed employees in Sunnyvale? 28 ,000. Today, it's at 3 ,500. Wow. Huge fall off. Incredible. Should we do a little bit of the history question I had? Yeah, yeah, yeah. I mean, he covered a little bit. Yeah, I'm just interested to hear your take on, you know, what SpaceX is doing in Texas, what TSMC is doing in Arizona. There's a bunch of interesting projects that you're probably pulling lessons from.

2:23:44What does it take to get this right? Yeah, I mean, I think we are so far behind that it's going to require kind of an all of the above approach. It's not here or Arizona or Texas. It's all of the above. I think there's an interesting question about specialization. And one of the other things that I noticed is if you look at all of the breakthrough products that we've built in America, you can go look at Boeing 747 in the 60s. you can look at Tesla Model 3 10 years ago, or you could even look at some of the stuff that Andrew is doing in Southern California. Basically, all of the breakthrough products that we've built, they scaled between 10 and 20 miles away from where they were invented, right?

2:24:29The Boeing factory at Everett Field was about 20 miles from R &D. Tesla Fremont was what, 25 miles from the HQ in Palo Alto where they designed it. I think you could make a, I mean imagine Tesla ramping up Model 3 production if the factory was in Texas. You could argue that it would have gone very differently, right? They would never have been able to figure it out if you didn't have the tight integration between the engineers and the production line. And you can actually drive there and be there in an hour. And so the way that we think about the foundry is it's probably not the place where you put your factory that has 10 ,000 employees.

2:25:07It's probably not the place where you put your gigafactory, but it is the place where you go from having 50 people in Menlo Park to having 200 people in your factory number one to having 1 ,000 people in your factory number two, maybe 4 ,000 people in your factory number three. And then if you really, really hit it big and you're operating at Tesla scale and you need a factory that employs 10 ,000 people, then yeah, go put it in Texas or Reno or somewhere else. But it's basically a place for all of these companies to scale from prototype to a multi-billion dollar company. And then also keep some of the cutting edge development close to the R &D talent.

2:25:48I mean, even if you look at a lot of what Tesla and SpaceX are doing, a lot of the stuff remained in California. I mean, the engines that go on Starship are still made in California. Yeah. What's the timeline for the foundry? Are you guys already working on? And, you know, I'm imagining you're talking with a bunch of tenants already about potential leases. How quickly do you want to see people with actual their employees boots on the ground standing up these factories, facilities, etc.? Yeah, I mean, the current timeline is to break ground in 28. And that's basically on track. So there's a local city called Susun that basically came forward and said, hey, we want to look at approving the entirety of the new city by annexing it into our city.

2:26:35That happened a few months ago. And so we are now going through the environmental work. I think they're expecting to publish the draft environmental report in Q1 of 26. But since we've announced the foundry and actually same for the shipyard, we've had a bunch of companies come to us and say, hey, we would love to build here. but we really would like to break ground in 26 or 27. We have lots of orders. We need to make a decision. And so we're going to look at, can we work with the stakeholders around the table to accelerate it? Because it would be a shame to send another 20 ,000 jobs to Texas.

2:27:10By the way, very topical here. California just put out the numbers for unemployment last month. We just lost another 7 ,000 jobs in the Bay Area. California now has the highest unemployment rate in the whole country. So you have this crazy juxtaposition of all of the greatest economies, one of the greatest economies in world history, period. World history, yeah, it's bigger than most countries. And still this unemployment crisis. Wild, wild. Well, thank you for everything you do. Good luck. It's a huge challenge, but it's been fun following along and exciting to see the latest development. We're going to have to set up our own studio.

2:27:48I was looking at, I'm looking in the background. I'm seeing TVPN Studios North right there. Right behind me. Right there. Right in the center. Exactly. Good talking to you guys. Awesome. Congratulations on all the progress. We'll talk to you soon. Thanks. And if you're looking to visit California, find your happy place at Wander.com. Find your happy place. Find your happy place. Book a Wander with inspiring views, hotel-grade amenities, dreamy beds, top-tier cleaning, and 24-7 concierge service. It's a vacation home, but better, folks. and we are ready to bring in our next guest, Nathaniel Manning.

2:28:20Let's bring him into the studio from the Restream Waiting Room, if we can do that. Nathaniel, how are you doing? I'm doing well. Can you hear me? Yeah, yeah, can hear you fine. Good to meet you. Why don't you kick us off with the introduction on yourself and the company? Yeah, hey, I'm Nat Manning. Everyone calls me Nat. And the CEO of Legend, Legend AI. What we do is essentially make Earth searchable. We want to be able to query Earth over space and time. So how do you do that? You have all these large language models out there. They're trained on language. And so if you were to want to find all the fire breaks in the Pacific Palisades, for instance, you can't really ask ChatGPT that because it's not in language today.

2:29:04It's not been marked. Right. But the answer is that does exist. It exists in pixels that have been collected over the last 20 years. in satellite imagery or drone imagery or low earth flying air. And you can train a model, a large earth observation model on that data set and be able to begin to query the pixels themselves. So that's what we do. Yeah. What are the data sources? I imagine Planet Labs and some other stuff. Are you buying data? Are you just using stuff that's out there on the open web and fine tuning or kind of putting it all in a database? What does the market for satellite imagery look like these days?

2:29:47The special part about this industry is a lot of it's made free and available. Depends on how recent you want it or how high resolution. Then you have to go private. But you can train it. We train on Sentinel, which is from the European space. And then there's a a bunch of NASA out there as well. And we can train on all that. So we get to train on open data. And then when someone wants something that's either more high resolution or more recent, we can run our models and easily on top of that purchase data as well. Who's the ideal customer for this? I mean, I can imagine it being something I'm using personally like every once in a while, but I imagine like it's more of a B2B use case.

2:30:35So walk me through the customer workflow. Yeah, there's really three products in a stack that we're building. So the first is at the API level. And I would imagine the kind of similarity there is like LiveKit, right? If you want to put audio into your work today and you're building an AI product, you can work with a LiveKit or another one like that. It's similar that way. If you want to be able to bring the intelligence that's in this data, which is arguably also one of the largest data sets ever collected, right? So language is under one petabyte of data that all these models are trained on. There's almost 200 petabytes now of Earth imagery.

2:31:10Now, the amount of imagery doesn't necessarily equal value, but it does kind of create complexity and possibility, I think. And so, you know, typical Earth observation market has been mostly government and defense and intelligence with agriculture and all sorts of kind of environmental climate, supply chain, logistics, finance and insurance. Insurance is a world I come from. So those are those are some of the key ones. and then as I said on the stack it's kind of API level which would be thinking like Google Maps Mapbox kind of kind of tools uh or live strategy here it feels like they have a lot of data in Google Earth and Google Maps and there's a search box there but uh it's certainly not powered by Gemini yet but do you think it's on the roadmap for them or do you think that they they're just focused more on like the consumer use case here yeah I mean Google's the big um 800 big bear in room yeah big big girl in the room yeah look they they have they have uh you know they have all the ingredients to make this stew they also have like a lot of competition a lot of focus um it's one of the reasons some of the investors we took on or were the founding team of keyhole which became google earth and google maps you know people who really arguably have made more uh from this in this space than like anybody else um one of my favorite stats actually was that uh and talking them learning that of the 200 billion in search revenue that google makes one third of it triages their location geospatial database so that's arguably like the most successful you know tool in this space now that's location as well and and not just uh what's called geospatial but um yeah the there's so api level and then we're building an enterprise product as well and then we'll we'll put out uh kind of a fun consumer application as well um kind of draw people's attention to what is now really possible that everyone's kind of dreamed about for years, this idea of being able to kind of query the earth over space and time, but the technology wasn't there until a few years ago.

2:33:20Jordy, you have anything? I got another question. Yeah, no, I was curious, like, where you expect the most pull from on the enterprise side? Is this like, on the site, you have a bunch of different use cases that feel really tied to insurance underwriting? Is that an area that you're excited about? Yeah, that's where I came from. I started a company called Kettle that focused on, this is kind of how I came about to this, which was building, we trained CNNs on all sorts of imagery and weather data and built our own risk models and ended up going fully verticalized and sold insurance in California.

2:33:59We built kind of our own models around wildfire risk and arguably, I think, built some of the best prediction models for wildfire and then wrote insurance. But the hard part with insurance, you need a giant balance sheet to be able to write it. And so it was hard to be in a position where you have product market fit so clearly, like so clearly saw that there was going to be a breakage in the California market, that you could do a better job using AI to predict this risk and then just be kind of hampered by conservative mindsets. And so, um, well, wasn't there also, what, what wasn't, uh, we, John, John and I live in two areas that were impacted by the LA fire.

2:34:37Fortunately, our homes were okay, but I had read that, that California had effectively had price controls by not allowing you to price based on risk, but only price based on historical loss rate. and then that changed in December. I don't want to go on a tangent, but I'm curious. No, no, I can talk about insurance all day long. And yes, it's definitely one of the main verticals we're focused on because it's one of the most kind of important data sets for underwriting, property underwriting. But in short, yes, the admitted market in California using Prop 103, because Prop 103 basically says you cannot, yeah, you have to kind of use historical risk, which in the world of climate change or any other kind of how fast everything's moving right now, the last 100 years does not predict the next five years nearly as well as it used to.

2:35:31And so, but there is this thing called the excess and surplus market. So once you have three denials from the admitted market, you can go to the excess and surplus market and those rules don't apply. And so that market has grown at, I think, a 17 % CAGR year over year for the last eight years in California. Wow. Because essentially the mid-market's on strike, saying like, hey, and fairly so. They're like, I need to be able to raise my rates. I need to be able to use models that actually work. And they struggled to do so. So I got a lot of sympathy for them. It's hard to be in that business. How much demand are you seeing from people that just want to use AI tools to get cleaner and more data into a traditional geospatial database versus more of like an end-to-end AI solution where you're training on like the raw satellite images and then trying to kind of query that directly?

2:36:29Yeah, a fair number. I mean, what were the first customers, the people kind of doing this today, what you're essentially offering, what we offer is the ability to create um aaron actually was kind of talking about this earlier on your show uh the ability to create information out of or structured data out of unstructured data so imagery is unstructured data yeah and kind of historically you'd have a you know stage one is you have an analyst look at it and you know looking human eyes are pretty good at looking at this picture and saying ah that's a fire break and this is a neighborhood and and that's a fire on the other side of et cetera, or a forest on the other side.

2:37:08And then you would train, kind of stage two is you'd train a specific model, a convolutional neural net. Detect all the trees, detect all the roads, and then that's how Google Maps. And you give it hundreds and thousands of images and teach it how to do that specific thing. And we spent millions of dollars a kettle doing that. And that's kind of the norm today. There are tons of teams and multiple data scientists and an ML ops person and, you know, a quarter or two to do one of these. So you're spending like half a million bucks kind of per query. And because, again, it's just like convoluted data.

2:37:44And it's across all those kind of industries we talked about. But what happened when you have the attention is all you need, you know, a paper that came out, the transformer model. When you apply transformer model architecture to data, the output is embeddings. So you have these vector embeddings for language. And what happened over the last 25 years, for 25 years, keywords were the first order data object for making sense of language. And in the last two and a half years, we've seen the process of it changing from that to language embeddings. And it's one of the biggest stories sort of in the world.

2:38:18And our take is the same thing is going to happen in geospatial data. For 25 years, it was map tiles or the pure raster data, the pixels themselves. And that we're in the process of that changing to geo embeddings. And so our whole thesis, what legend really is, is the stack for managing embeddings, creating, tuning and storing them, which is the output of these models on the raw data itself. And just that's the wave we think is going to happen for the same reason that because it's happening in language. Very exciting. Thank you so much for stopping by. Quick hit. Happy to be here. Do you have any news for us?

2:38:56Yeah, we raised a$9 million round. There we go Very exciting We have some great guys on it Kareem, your buddy Ryan Delk was kind of our first check-in Fantastic I've known Ryan for years Amazing Congratulations to you and the team We'll have you back on soon We'll see you soon We have to pull a video up And react to it Apparently some guys Were made a video this morning that I will drop into for the production team to pull up. Here we go. They made a launch video. I think their company called Composio. And apparently we were featured in it. Oh, really? Well, they violated the derivative works clause.

2:39:54Anyway, so let's watch it. Let's watch this. This is the first time I'm seeing this. We're reacting to this thing. Oh, there we go. Hello. My name is Jonathan Liu. Today I'm launching Cupidly, the cursor for dating. Let me show you how it works. So when you visit the site, you first enter your type. So let's enter my type. Oh, this is HingeBench. We've been talking about this. This is not the video that I pulled up. Then you go ahead and connect your Hinge profile. This is AI Agent for Hinge dates. Okay. It scrapes your hinge profile and it tries to do meta-burner matches. Not the video that I was referencing.

2:40:30So can you guys look in the tab getting some audio? Some weird sound effects in that video. Yeah, here it is. No, not this one. Not this one. The second, the first one, there's two videos. Okay. So first they violate the derivative works clause and then I guess they had to tap out as well. Okay, okay, okay. Let's see. So let's pull up the first one. We're going to have to find out what that other video was. Oh, yeah. that's a big viral launch so someone launched AI agents for guests. We're not getting video guys. I mean how could it go wrong we got into fucking YC. Yeah that was useful and eventually becoming nothing just like the agents we built.

2:41:09Fuck Luli. I wish our agents just worked. Beautiful cinematography. I sometimes wonder what it'd be like if we had agents that actually were useful. You're watching There we go! Crazy founders on our show, they literally ran from almost going homeless to an AI agent unicom. Over the past three months our customer recently said something. No fucking way! Our agents keep breaking, how'd you fix that? We started using this new tool. I saw this agent AI thing on Facebook, but it made no sense to me. Your new agents are amazing, they just look. So how did you grow so fast? Honestly, our agents sorta sucked before.

2:41:48We tried to build everything ourselves. Then we started using the friend of the show Composio. They're fucking great. They allowed us to build these incredible agents. Last year we got reliable agents. Ones you could trust to code, test, and ship. This year the shift is bigger. Agents! Agents! Agents! Teams aren't just working with agents, they're managing swarms. Yes. Dozens of agents, including work. Handing off context. Self-correcting. Agents! Agents! Agents!

2:42:30This is amazing. Very drawn out. Okay. Fix your agents before it's too late. So Composio. Very cool. I'm pulling it up. The skill layer of AI. There was so much in that video. Yeah, there was a lot going on in there. But we'll pull up the second video now because it sounds like they've respected our terms of use. They respected it in the video. I saw Tapwater in the video. They don't even need to do a second video, but they did. So thank you to them for respecting the derivative works clause and paying homage. I love how the team is able to deliver 20 hours of cinematic content a week. But if you get them, try to get them to pull up a video.

2:43:12Here we go. Let's see this.

2:43:16Look at this cinematography. So many cuts. Wow. This is fantastic editing. Congratulations. I can't fucking believe it. You already tapped out. There we go. You love it. Great job, guys. Great job. Fantastic. Fantastic. That's a fantastic match. Full gold return. Congratulations to them. Love it. We have to have one of the show. Yeah, we got a month. They raised$29 million to build skills that evolve with your agent. Okay, tomorrow we're having them on. We have slots. We will reach out. We'll make it happen. We have one more guest. I think Chris is still in the Restream waiting room. Is that correct?

2:43:54Is that right? Chris, sorry to keep you waiting. There was a viral video. We had to react to it. No worries. Thanks so much for hopping on. Thanks so much for the patience. Kick us off with the introduction. How do you know TJ Parker? Yeah, TJ Parker. Man, we went to Sunday school together. I think so. 25, 30 years ago. That's a good start. That's great. I know TJ's a dear friend. We're both based out in Park City, Utah. Oh, amazing, amazing. And give us the introduction on the company, the news. What else is going on in your world? Yeah, so I'm Chris Andrew, co-founder and CEO for Scrunch AI.

2:44:28We've built a customer experience platform for the AI buyer journey. Today, we're announcing a$15 million Series A. Congratulations. Led by Decibel. Thanks so much. I had to hit the gun. And I mean, very simply, our belief 18 months ago was that all of search was moving to be AI search. Really, like we outsourced browsing without talking about the fact that we never wanted to browse the Internet. It was inefficient. We outsourced it. AI does it for us. And it means that AI visits websites instead of humans. And so we're building a company around that shift. Okay. Talk to me about the AI shopping journey recently.

2:45:08We're looking for gear for our studio basically always. I wanted to build a mobile camera rig. I needed certain features. There are HDMI ports, this much battery life, all the camera. Kind of describe what I want. It puts together a little shopping list. I send it to the production team. We kind of go back and forth. There are some links there. I don't think those are affiliate links yet. I feel like in the future there will be ads. But what else is important in terms of actually positioning if you're selling a particular camera? how do you make sure that you show up and you beat it out? Because I kind of knew I wanted a Sony camera, but they could have sold me on Blackmagic there.

2:45:45Blackmagic should have been hand-wring me. They should have been paying OpenAI to be like, hey, have you considered this Blackmagic one? And the team likes Blackmagic too. So talk to me about all the different nuances that go into AI shopping. Yeah, I think really ultimately all that top of funnel discovery, kind of the 90 % of the buyer journey, which used to be you curating research from Reddit, reviews, sites based on pricing. We've trusted OpenAI and Perplexity to do that for us. So they're basically running a search over a search index, which is increasingly GPT's own search index. Prior, that was based on Bing or Google or Brave.

2:46:24But they're looking at a list of websites, identifying title, description, snippet, visiting the website, deciding if it's relevant based on the prompt that you asked. So it's basically trying to match the intent of a human prompt to content on websites, to cite the sources that it deems relevant. So what we're seeing in the end is buyer journeys that are massively reduced. People are arriving on a website and buying the first time they're there. A lot of our customers are seeing 2 to 4x conversion from AI search visitors compared to organic search visitors. And I think search and shopping will keep changing.

2:46:58I just heard that GPT is embedding Shopify checkout in their feed. They're taking a cut of that checkout. I think it's 4 % or something in terms of like another cut off the top. But today what's happening is AI is doing the research, citing sources, and we as humans are clicking out to purchase and engage. So how important is it if you run an e-commerce website to go into Cloudflare and turn off that AI default because that seemed like Matthew Prince came on the show, a great entrepreneur. He's certainly working in favor of content producers who might not want to be scraped and have the harder work to blogs that they've developed automatically generated in AI search results.

2:47:42But if I'm just selling a product, I definitely want to be in search. So how is that landscape shaping up these days? Yeah. Matthew Prince, another great Park City local, we've all got our fingers crossed he can buy the mountain. We're pulling for him and also pulling for him that, you know, content creators can get paid in AI crawling. I think that model makes a lot of sense for Cloudflare. Our customers, on the other hand, I think the 50 to 60 % of the internet that has a product or service to sell is desperate to bring content online to help AI understand their products and services. So we still come across Fortune 500 brands who have their site blocked.

2:48:19And so when When I ask a question about that brand, what's the price? What's the product description? Compare it to XYZ. It's going out to the open web to represent your product. So our customers are like, my product and service is no longer what I say it is. It's what ChatGPT says it is. And if GPT can't access or can't effectively utilize your content on your website, it's going to pull from every other source that it can find on the internet. So that's part of why we've launched what we call our agent experience platform today, which is essentially an optimized version of your website. So we take your website, we scrunch it, called our brand Scrunch, because we literally want to strip away everything on a site that is confusing to AI.

2:48:58These models don't read JavaScript. They struggle with carousels. They're not converting your multimedia into text. Large language models want language. So it's critical for businesses to be enabling the crawling on their site and making that content rich and accurate for the crawlers to access and pull through. How is that actually different from just generic SEO? I feel like if I had a video, I'd want subtitles and text and like an HTML tag or something that just Google search could pull up anyway. There was kind of like that keyword stuffing era, dump a bunch of words in white text on a white background at the bottom, one point font or whatever.

2:49:35Are we going back to that? No, I mean, I think we're actually moving more towards human language and intent. I mean, the keyword stuffing is a great example. If you pull up a recipe site, go Google chicken noodle soup, you're going to find the word chicken 60 times on that page. It's absolutely manic. AI wants the intent of the query. So like if you ask GPT for a recipe, it gives you the answer, right? It kind of strips everything away and reduces it to what you want. So I do think we're actually returning to a lot of core SEO principles. You need language. Large language models want language.

2:50:07They're struggling with carousels and multimedia. They need that text. And so I think best practices in SEO, ultimately my hope is we just call this search. We get rid of GEO, AEO, AI. I ultimately think all of search is going to be AI search and it's agents retrieving content for us. Anything else, Jordy? No, this is great. Thank you for joining. Congratulations on the raise. We will talk to you soon. Thank you very much. Thanks, guys. Appreciate it. Have fun out there in Park City. Give our best to TJ. Yeah, we'll talk to you soon. Have a good one. Bye. Let's rip through some timeline. Did you see the medieval knight raising a seed round of 500 gold Ducats at 3 ,000 Ducats post?

2:50:47And they already have their first LOI signed with the Holy Roman Empire. Hard not to be bullish. We must beat the Ming Dynasty in the gunpowder race. What is this video, by the way? We've got to pull up this video. Pull up the video. I don't know if we can. This is going to be hard. This is going to be a big challenge. It's an image right now, but it is in the tabs. You might have to scroll way back. loosely modified after the Tannenberg Castle gun. Mine is loaded with around 50 to 60 grains of 2F black powder and fires a.75 caliber lead ball. Do we have the video? Can we pull it up? One second.

2:51:21This is an OG 2A enthusiast right here. Yeah, indeed. And I love that some founders alt. Edward is just like completely transposing this to the defense tech meme of like, we must beat the Ming dynasty in the gunpowder race. Like, you know, I've invented defense tech and this guy is firing some sort of a castle gun. I've never heard of this before. Apparently from 1400. We should get one for the studio. They had guns 600 years ago? Wow. I thought that was more modern. I thought that was like maybe 300 years ago. When was the first gun created? Let's figure it out. Appeared in China. The earliest firearms known as fire lances appeared in China around the 10th century.

2:52:07They were essentially tubes, often bamboo, filled with gunpowder and shrapnel mounted on spears. So they would just, made a lot of sense. You got a spear, you want to point it, you want some range, throw on some gunpowder. Fantastic. More news, Nudge raised a$100 million. Oh, here we go. We got the video. I think we got it. What are these guys doing, by the way? This is incredible. Like, what is this operation? Who are these guys? legends that is so crazy it is bamboo it looks like it's bamboo i don't know if it is but insane insane ben get one of these for the studio i need one yeah yeah yeah for an ex confetti cannon we'll load up yeah with real confetti cannon with which just did yc demo day boom shaking the whole building every time someone announces a closed seat round uh yes nudge the brain computer interface company co-founded by Fred Ursem, the co-founder of Coinbase, says half the world will experience a brain disorder in their lifetime.

2:53:07At Nudge, we're building brain interfaces that are safe, precise, and non-invasive to solve that problem. They raised$100 million in Series A, led by Thrive Capital and Green Oaks to go faster. We're hiring Join Us. Very exciting. Nudge, kind of the non-invasive Neuralink. There's other news in here in the deck. Like Neuralink just did two implantations in the same day. So they're scaling the actual implantation of the Neuralink device. That, of course, requires invasive surgery. Nudge does not. It's a kind of a helmet that you put on, but then can stimulate the brain in different ways and solve a whole bunch of different brain disorders.

2:53:43This very likely could be the key to scaling from three to six hours a day. I think so. Live. Put one of these on. I mean, you can actually turn off sleepiness, basically. I mean, like once you get the different, it's literally like a tinfoil hat. You put it on and it can just be like, turn off fear, up-recolate risk-taking, go turbo long. Yeah, mitigate. You want capital alligators to be super risk-on but still get the weight loss benefits of GOP. Yes, yes, yes, exactly. Something there. This is a wild leaked memo from Anthropics CEO Dario. Dario is basically justifying to the team why they're going to go directly to the Gulf to raise capital.

2:54:31And this is not something that probably should have ever been in a memo. He says, unfortunately, I think no bad person should ever benefit from our success. It's a pretty difficult principle to run a business on. Just blanket putting the Gulf states in the bad. So it's such a weird thing because you'd think you'd just be like, I met them. I dug in. I don't think they're bad people. And so I'm doing business with them. And let's debate that. And I think it's a Pareto improvement. Everyone is made it better off by Anthropic working with the Gulf. That's the argument he should be making, but he's not, I guess.

2:55:03And so Dylan Patel says Anthropic has been a series of ideological decisions later defeated by business realities. We've talked about this before where, you know, it was like all the all the idealistic AI scientists were like, let's try doing this without capitalism. And then and then the AI God was just like, I actually enjoy capitalism. It's thrilling to me. Thrilling. And I like five hundred billion dollars. You're exactly right. The right way to do something like this is say, I went, we visited, we spent a few days together. We have a shared vision for how AI can increase human prosperity.

2:55:37Totally. Whatever the critiques are, I don't even know what the critiques are of the UAE. I'm sure there's a bunch of both the UAE and Qatar. Like there are all governments and every company and every person ever. But you dig in, you figure out like, oh yeah, how can we work together to resolve any lingering problems? Should we have a different perception? You don't just be like, I'm going into business with someone I think is bad. Wild, wild choice. Anyway. Choice of words. In better news, Blake Scholl says, at Mach 1.7. Overture out flies Earth's rotation. That's cool. So you can have breakfast in London, fly to New York for breakfast and then to San Francisco just in time for breakfast.

2:56:15This is going to be huge for the breakfast for dinner folks. You ever been a breakfast for dinner guy? Pancake dinner. Yeah. Pancake dinners. Let's do it. Kind of insane. As a kid, that was mind-blowing. We'd do that maybe once a month. Cereal for breakfast. Waffles. My dad would make waffles. Dying laughing at the concept of eating cereal for breakfast. it really is oh it was the funniest thing it's so imagine breakfast is so good it is and when it's not my favorite really yeah i'm not a big breakfast guy no i think uh i think it's my even even my default breakfast is like a breakfast burrito which is really like lunch or dinner masquerading as breakfast you know yeah it's kind of like it's true they've just kind of like swapped out i just remember as a kid my dad making waffles for dinner was so awesome yeah yeah oh it was amazing uh do we have the tyler cam working can we can we check in with the final review of the Tesla cafe.

2:57:08How was the experience? How far, how far, how long did it take you to get there from here? It's like five minutes away. Oh, it's really close. Super close. Oh, wow. Okay. Like we could probably walk there. Okay. Um, yeah. Did they, I wonder if they did that on purpose. We didn't get the invite to the opening. There's some other influencers there yesterday. We were, we're behind the ball. We had to go. They know we're always live. Yeah. Um, yeah, it's definitely cool. Like the building is super cool. Uh, there's a ton of stuff on the menu. It's just like a diner. So that's cool. I mean, there was like, I don't know, there's like hundreds of non-hundreds.

2:57:37There's a lot of things on the menu. Yeah. Like every kind of meal. Yeah. So I think it's open 24 hours, 24 hours. That's cool. So maybe, you know, we go back later. Do they have steak? Try something else. Do they have steak? Do they have dinner? I don't know if they have steak. I guess you just get a burger. We didn't look. Yeah. But so we took a Tesla like to get there. And then so when we parked, you could see on the menu. How did you take a Tesla to get there? Adam. Adam has a Tesla. Oh, okay. He just owns one. Yeah. Got it. Okay. It wasn't a, uh, it wasn't like a robot taxi because the robot taxi is rolling out really aggressively, but it's not in LA yet, but this feels like kind of tied.

2:58:09They're teasing. They're, they're, they're, they're testing the waters. Um, anyway, yeah, I think the burger was solid setting. Are you feeling still good or any upset? Tommy, I feel pretty good right now. Um, okay. Yeah, it was cool. I, I'm not, I wouldn't go back just for the burger, but maybe I would go back. Would you, would you go back and do a muck bang and eat every item on the menu in exchange for an iPhone. Have an exchange for an iPhone. Yeah, I would do that. You would eat every item. You would try and eat every item on the menu during our stream. I think I could probably do that. Three hours, every item.

2:58:39You have to eat them all. Yeah. It's like$7 ,000. What do you do by every item? Is it a burger and a burger with bacon? I saw a graphic that had like 12 different items on it and so it would be 12 different meals basically. Yeah. Three hours? He's a growing boy. He's a growing boy. I've never been in the competitive eating world. That's a good challenge. I think it's a little vulgar. I think we'll come up with a better, better challenge for you. You will win the iPhone within the next few weeks, but we got to come up with something. It's going to happen. I wouldn't want to put you through the mukbang.

2:59:10I think it's probably a little bit too aggressive, too hardcore on your body to try and eat all that food. But three hours though, three hours, three hour show. Maybe you can do it. You do the mukbang. I got to be here. Yeah, exactly. Anyways, we have a post here from Jira Tickets. His has been an absolute regular recently. Yes, yes. It feels like he's on every other day. He says, ah, yes, good morning. Let's open up the Infinite Information River for a nice 20 minutes before getting out of bed or thinking any original thoughts at all. I love it. So real. I, because I wake up so early and I race over to your place and then get in the car, I haven't been opening the Infinite Information River first thing in the morning.

2:59:50I haven't been doing anything other than just get, actually, I go outside. I go outside and I get sun. Who were we talking about that with? Maybe Huberman or something? but Rob. So I've been doing that and I don't know if it's really been that much better. Sometimes the light that blasts your eyes from your phone is the best alarm clock. And when you open up your phone and you're greeted with a post like this, it just kicks you off to like a great start of the day. You're just laughing. If Elon could pin this to the top of the feed, that would be great. You're having a great time. You're having a great time.

3:00:24What else is in here? Oh yeah, I mean we do have the post about somebody built cursive or dating. We got to talk to this fellow. Spent his life - Can we pull up the video now? I was super confused earlier, but we may as well - Spent my life savings on this cinematic video because I have no VC money to burn. Did one shot, very funny, and then digs into AI agents for Hinge. The AI agent goes and scrolls your Hinge profile, does all the swiping, all the talking, and then surfaces matches that it thinks you're actually interested in matching with. So kind of like a layer on top of hinge, I suppose.

3:00:58Yeah, let's pull it up. Let's see it. I mean, the cinematography, fantastic, for the first couple seconds, and then it gets a little silly. Hello, my name is Jonathan Liu. It's a cool shot, cool shot. It says he ran out of body, here's a loom demo. The cursor for dating. Let me show you how it works. So when you visit the site, you first enter your type. So let's enter my type. What does he say? I can't read that. Then you would go ahead and connect your profile. It's funny that he's censoring Hinge because it's obviously Hinge. After doing that, you would click this big black button to start searching for people your type.

3:01:33And this launches an AI agent that swipes for you on Hinge, filtering for people who think your type of people. Why is he censoring it? Everyone knows. And then also, it seems like he's harnessing Hinge in some weird way because I don't think Hinge has an API. Tyler, do you know if Hinge has an API? Not that I've looked, but I don't think there is. Okay. Because I imagine that they would not want this to happen. They don't want bots on their network whatsoever. He thinks you're a good match. And, oh, this picture's pretty cute. Let's go ahead and send a message. The LY is really coming back, isn't it?

3:01:59Either send a message using AI. Yeah, it was a real throwback. Yeah, and then you send a message using AI. That's wild.

3:02:34I don't know what this is done for. it out with Mike Krieger, co-founder of Instagram at Anthropic. He says, today I'm excited to share that I'm joining Figma's board of directors. Having known Dylan for years and watched him build the platform that's become essential for design teams, I'm looking forward to partnering with him and the Figma team on what's next. Fantastic pickup. Heading into the IPO. Couldn't have found a better person to join the board. So congratulations to everyone involved in that. Last one, Close So Investments says, not calling a top here because of AI, yada, yada, yada.

3:03:07But White House eliminates day trading regulation. Poly market returns to the US. Robinhood enables 24-7 EU gambling. The architects of the 2022 SPAC bubble are posting. CNBC guys have a crypto treasury play. Wall Street bets. Meme crashes 40 % and almost closes negative. I don't understand that one. If we blow up later this summer, can't say there weren't signs. I mean, the benefit of like, you know. And what he says is all I'm saying is what is going on is pretty crazy. I mean, the benefit is that if no one's saying this, then you need to be really worried. Because at least people are saying this and then they're thinking it.

3:03:44And then hopefully people are pulling back before you get to a true precipice. I feel like in 2021, 2020, people were saying this is crazy. Yeah. I see saying all I'm saying is what is going on pretty crazy. Pretty crazy. Pretty crazy. Hyper capitalism. We'll see. We'll see. Gotta love it. Anyway. Thank you so much for watching. Last but not least. What you got? Hope. Oh yeah. Hope's Revenge. One of the top posters in the world. Hope's Revenge. Topping the power ranking. Originally said my GF is upset because her favorite albino alligator was bought by Anthropic and ruined. And then he actually has a picture here and it says the care of Claude the alligator is generously supported by Anthropic.

3:04:25I think this is great. I don't see why it's ruined. I love to see big AI supporting alligators. Yes. And I actually want to get I'm really excited for BCI's brain computer interfaces to get into the alligator category because they're pretty wild. You ever I mean, just pull up a picture of an alligator. These things are insane. The fact that they're just cruising around. No one's working on humanoid alligators like robotic alligators. We've seen that's maybe that's the killer for everybody's saying, Oh, you guys could do incubations. I feel like a robotic alligator would be a good defense tech play, too.

3:05:00Yeah. I mean, there's a lot of investment in bears. Yeah. Robotic bears. Yeah. You let the robotic alligator can sneak under stuff. Swim. Jump up. Run. Attack the terrorists. Yeah. Attack the enemy. Seems like. An alligator, a robotic alligator with drone-based wings. So it could actually fly while snapping. Could reshape the modern battlefield. Could be looking at a billion dollars. Home defense play too. If you had some robo alligators on the roof and you had somebody breaking, I'm big into moats. I'm big into moats. What am I going to put in the moats? That's right. That's right. Yeah. And there's going to be some upkeep costs with maintaining a fleet of robotic alligators, but probably less than how many, how many chickens or cows you got to throw in the moat if they're real alligators.

3:05:47So yeah, we're trying to, I think more single family homes should have moats, right? We're always talking about businesses, talking to founders. what's your moat what's your moat yeah meanwhile we're sitting in houses that don't have moats yeah and i think that should change so lots of lots of opportunity and alpha here at the end of the show fantastic have a great day we will see you tomorrow cannot wait have a great evening