Astronomer CEO Affair at Coldplay Concert, JUUL Approved by the FDA, Is AI Causing Historic Crashouts? | Chris Power, Jeremie Eliahou, Chris Best, Alex Mashrabov, Jesse Pollak, Billy Thalheimer, Jonathan Mortensen, Douwe Kiela & More

29:18Then in like early turn of the millennium, 2000, something like that, this company, NJOY, comes out with one of the first e-cigarettes. There we go, Camel. Wow. I mean, you see this as a 17-year-old. You're thinking, this day I turn 18, that's going to be me. Wait, is it 18 or 21? It's 21 now. That also changed recently. So, NJOY comes out with one of the first big e-cigarettes. It's an example of that older technology that has the big vapor cloud. And the FDA hits them with a lawsuit and says, hey, you are selling an unapproved medical device. This is a device. It's electronics. And it's a medical product because you're making a medical claim.

30:04And that medical claim is this product helps you quit smoking. Smoking addiction, cigarette addiction, is a disease. And so by selling a product that helps you quit smoking, you need to be regulated by the FDA. NJOY fights this back and forth, it goes all the way to the Supreme Court, NJOY wins because they were kind of not saying that, at least they made the argument that they were not saying this is to help you quit smoking. They were just saying, this is a cool thing to do, separately, don't worry about it, don't ask us about the relation to smoking. This is classic. But in the interim, the FDA is able to put an import restriction on the company.

30:44So they're making the product, I believe, in China, probably overseas somewhere, because it's electronics, it's equipment. They bring it in at the ports, and the FDA says, hey, while we sort out this lawsuit, you can't bring any more in. And so this is kind of like the banhammer that they bring down. It winds up bankrupting the company. They wind up winning the court case in the Supreme Court. Later, a hedge fund guy actually buys the company out of bankruptcy, turns it around, gets it FDA approved, and sells it to Altria for like a couple hundred million dollars maybe actually a couple billion dollars I think let's hear it for the hedge fund guys making some money selling some vapes to to not endorsing the end product we're endorsing financialization yeah yes and restructuring yeah anyway it was and it was a successful thing and if you think about that as much as we're joking like it is good to get a big tobacco company get shift their revenues away from cigarettes as fast as possible and And so the enjoy thing, even though there's a lot of issues with that product in many ways, it's a very interesting outcome and it's probably moving in the right direction.

31:50Anyway, so it goes to the Supreme Court and it is revealed in this court case that the FDA does not have the ability to regulate e-cigarettes. And so it has to go through the House and Senate. So when Obama gets elected in 2008, they pass the US government at the federal level, passes the Tobacco Control Act, the TCA. And in there, it says, hey, the FDA does have regulatory authority over everything that contains tobacco. So now, it doesn't matter if you're creating a new cigarette, doesn't matter if you're creating a new e-cigarette or a new nicotine pouch or nicotine gum. You need the FDA to review your product, which is good.

32:36It's probably pretty good because people should know what they're putting their body and they should know that the government reviewed this and said, okay, yeah, it doesn't have anything crazy in there. Like at the very least, like you said it has two milligrams of nicotine in there. Does it? Like let's test that. And then, so the companies test that. They send it to the FDA and then the FDA waits. And then the FDA gives you the thumbs up or the thumbs down. You can continue to sell it or you can't. This is what Juul just got with the marketing granted order. The FDA said, we have approved, we have reviewed your application, all the data, there's nothing that we see that would be worse than cigarettes, and therefore, we will allow you to continue selling.

33:15So Juul was started before this stuff went into effect. So 2008 is when the FDA gets regulatory authority, but the government moves slowly. So it's not until 2016 that the first real FDA rule goes into effect. Basically, they have to staff up a new arm because they have their biologics division. They have their drug division. They have their veterinary division. The FDA has their medical devices division. They don't have a tobacco division. They need to find a head of the tobacco division. They need to find a whole bunch of people to staff that, scientists that know how to review nicotine and review e-cigarettes and review all this stuff.

33:56It's a massive organization. They have to hire a lot of people. They do all of that, and then they have to decide, what are we going to do? What is that structure going to look like? and they come up with the PMTA process, the pre-market tobacco approval process. What this says is that going forward, we're not really looking backwards. We're not going to go review Marlboro Reds. Those have been on the market forever. Everyone knows they're bad. Everyone's aware of that. But going forward, new products that contain nicotine, that contain tobacco, we want to review it before it hits the market. But we're also going to create a grace period for stuff that was launched before 2016.

34:30Anything launched before August 8th, 2016, You can keep selling it while we review it. Because, hey, look, you've built a business. If it's going to take us a couple of years to review your application, if we just spike your revenue to zero, maybe you created a fantastic product that actually helps keep people really healthy. Maybe it's an amazing product. We don't necessarily want to take you off the market, crash your revenues to zero. You have to lay everyone off. Your company goes bankrupt. And then two years later, we say, hey, you're approved. And then we have to build you back up. Let's just keep things going as they are.

35:00We'll maintain the status quo. We won't ban you. We won't approve you or authorize you. We'll keep you in limbo. That limbo was supposed to be like a year. Because it's like, it's a big document. I'm pretty sure Jules' document was probably like 100 ,000 pages of scientific research. It's a lot of stuff to review. But everyone thought it was like, oh, it's going to be like a year or two. The deeming rule goes into effect August 8th, 2016. They're like, hey, turn it in by 2018. but then it gets pushed back to 2022. Then it gets pulled forward. Then COVID happens and the FDA has to pivot. Then the Juul crisis happens where everyone is starting.

35:40You're building a nicotine company this entire time. Yes. On the regulatory roller coaster. Yeah. Yeah. So we started the company in 2016 before the deeming rule went into effect so that we could bring our product to market and then work through the FDA approval process because we basically saw that the door was closing. And if the door closed and you needed to get approval before selling a single unit, well then all of a sudden the equation goes from, okay, you're building this company like any other normal company. And then yes, there is this binary outcome that can happen with the FDA. But if your science is good, you should be approved and that is knowable as opposed to, okay, now in theory, there are a bunch of loopholes that people exploit all the time.

36:25But in theory, if you want to start a new e-cigarette company or a new nicotine pouch company or a new nicotine gum company, in theory, you should have to formulate the product, run all the tests, submit to the FDA, and wait for them to get back to you before you sell a single unit in the United States. And what, you know, somebody that's self-funding a business like this might be interested in investing$2 million today to do all that, and then waiting for five, 10, however many years, maybe you never get approved. So you're basically sinking capital into a business that may never be able to sell a single unit.

37:02And I know very few investors that would be interested in that kind of proposition at all, or founders that want to make something and then wait forever for permission. Yeah, you could spend millions of dollars and wait a decade, which is exactly what we did. but we were able to actually grow the brand and sell the product and like set up operations and iron out things and iterate a little bit during that time fortunately um but yeah it's extremely hard to underwrite and it's particularly hard to underwrite because at the end of the light at the end of the tunnel let's say that you were to today start a new nicotine company you you happen to have a hundred million dollars sitting around to go do a bunch of studies and you happen to have 10 10 10 years to wait for the fda to get back to you and then you're going to launch the product Well, when you launch the product at the end of the day, you still have to contend with the fact that you're selling a consumer product in a highly competitive space.

37:55Yes. Yeah. Essentially a commodity product. Yeah. Some differences in formulation. Little bits here and there. You guys have breakers. Yeah. That's unique. A little bit on the ingredient side, a little bit on the flavoring side. But you're arguing our gum is better than their gum. Yes, exactly. But it's both gum. Exactly. Which is a tough thing to argue. Exactly. And then aside from that, it's like, how do you actually build a brand? even if your product is better because of something. When you're highly restricted on marketing. Yeah, you're extremely restricted on marketing. So even if you did have$100 million to spend, how do you spend it effectively?

38:26Like some of the best marketing for Lucy is like Joe Rogan sitting by UFC. He just happens to enjoy Lucy. So he's just commentating and people pick that up. But you can't just like pay for that. If you went to Joe and you're like, he'd be like, no, that's not how I work. Yeah, yeah, yeah. And so those kind of serendipitous brand building moments would just not happen if you're just like in the lab waiting for the FDA to get back to you. And then also you have the monopolies on distribution and the intense channel competition from the big tobacco companies. so it's like you come out with this product you finally get approved after 10 years you're like hey i got i think my formula is a little bit better i think my branding is a little bit better you go to 7-eleven and they're like but you're not gonna wait are you gonna pay us you know 100 million dollars in slotting fees this year like like pmi might or altria i got a pitch a while back yeah uh for somebody that had made effectively made a nicotine pouch but it was just a slightly different chemical yeah like yeah small small small change and was just bringing it to market yeah yeah yeah and uh hearing knowing what you guys have gone through to get where you are today and and knowing this sort of history that that you'd shared in pieces with me through throughout the last couple years was like the the idea that the fda is just gonna let you get a like raise venture capital and let you get away with selling nicotine in this like non-studied form It just, it was tough.

39:56I ended up not getting conviction, but even though the founder's super, super sharp. Yeah, yeah, there's just tons of, there's tons of loopholes. Some of the loopholes get closed in a way that does not close off opportunity for the companies. Some of the loopholes get closed and it puts companies out of business because they weren't expecting it to get closed in a particular way. Sometimes the loopholes close at state levels, but not the federal level or vice versa. So there's just like a ton of regulatory complexity around this. And so basically, to bring it back to Juul, they are pretty dominant by 2016 when the deeming rule goes into effect.

40:36I want to say$100 million in revenue, something like that. Like a pretty solid business, but growing like a rocket ship, like insane growth. And clearly the product was just vastly better than the competition because of the formulation and because of the design of the product, how discreet it was. Now, so smokers really were switching. That's definitely true. I believe their number around two million smokers quit with Juul. They couldn't say that. They couldn't say, hey, quit with Juul. They tried to make, they tried to, they actually like trademarked, like make the switch at some point. They were like, don't quit cigarettes, switch from cigarettes.

41:21Because like that was not a quit claim. So the FDA, so it was like all these different things. But like it really was having an impact. Meanwhile, obviously big tobacco is like heavily lobbying against the e-cigarette market. So it's not like. And buying stuff and trying to compete and trying to keep the. What was that other company? Was it Blue? Yeah, Blue was also big for a while. You're seeing a bunch of those ads as a kid. Yeah, and then Blue got bought and turned into Vooz, I believe. Like, there's so many. There's so many companies, and they're all, like, the big tobacco is, like, highly oligopolistic.

41:53Like, in cigarettes, Marlboro is the power law outcome, but the company that owns Marlboro has the same market cap as the company that owns the next five brands combined because you add up the next five brands because it's not that steep of a power law. So Marlboro maybe has, like, 40 % of the market, and then there's four brands that have 10 % of the market, And then the next brands have like 10 % between, they have a 1-1-1-1-1-1, something like that. And so you can actually create like this portfolio of brands that adds up to the same distribution because there's like all these different brands are highly specific to specific marketing channels.

42:26There's like the history of Virginia Slims targeting women and all these different sub-products that have gone after little niches. Marlboro Reds say something about you. I don't even know the difference between most of these cigarettes, but like Marlboro 100s say something different about you. And then menthols or Virginia Slims or American Spirits. A lot of like hipsters use those for a while. Like that was the whole thing. Anyway, so Juul is growing like an absolute rocket ship. There are cigarette users that are switching over and stopping to use cigarettes and probably improve their health.

43:00Like most, that's certainly what the FDA is saying is that like that was a net benefit. Also, they're completely dominating the e-cigarette and vapor market. Like the vapor market is just like. And this is the time that you see those modded unit things start to fade away. Because people are realizing, all right, carrying around a backpack so I can bring this vape machine to blow. That would leak liquid and the battery would run out, all these different stuff. So the business model is also fantastic because it's this razor and blade model. You buy the Juul once and then you buy the pods. And then because nicotine is addictive, it is very low churn.

43:38so you stay on their high margin. And so the business is just doing fantastically. There's a whole bunch of venture capital dollars that come in from various sources. Didn't they get up to$45 billion? Something like that. Might have been even higher during the acquisition. So zombie acquisition, but not a ghost ship, interestingly. Altria comes in at the peak and puts like$13 billion into the company. And a ton of it gets dividended out. So late 2018, Altria acquired a 35 % stake in Juul at a$38 billion. $38 billion, yeah. And so that like$10 billion kind of gets like dividended out to the shareholders, but you hold on to your shares because it's just a dividend.

44:18And then, so you are diluted, but you don't actually have to sell your whole stake. WeWork at that time, WeWork was valued at$47 billion. So both ended up being a little rocky from that point on. Yeah.

44:37But Altria kind of bought the local top there because Juul got way too popular. Kids started using it. That, the - And then the flavors were the issue, right? Yeah, the flavors were - And that's what people were trying to make the issue. Totally, totally. And the data on youth use of e-cigarette products like spiked like crazy. So I believe it was something like 10 % of kids under 18 or 18 and under were using e-cigarette products when Juul was introduced. And at the peak, it was something like 40 % of kids using e-cigarettes. And the majority of them were using Juul. So it really was like this viral phenomenon.

45:18And to your earlier point, a big part of that was because the age to buy these products was 18. So like if you're a sophomore in high school, you can just ask the cool senior, like, hey, go pick it up. You don't even need a fake ID as opposed to alcohol, which you need 21. Are you implying that the senior that buys vapes for the minor is cool? Maybe or maybe lame. The bad boy. The bad boy. The bad boy senior heading over to the gas station. Exactly. Exactly. So they were very easy to get. There were a whole bunch of like informal distribution networks. people would buy them in bulk and then redistribute them and sell them at a profit.

46:00So this kind of like zombie economy popped up. Anyway, the kids really were using it a lot. That is definitely true. And it was definitely cause for concern because kids shouldn't use nicotine because it's addictive. And the earlier that you use it, the more addicted you'll be because your brain's still forming. And if your brain forms while you're on a particular substance, you're kind of like your brain's developing and you're like that forever. So like it's much harder to quit nicotine if you start really young. Whereas if you start much older, you're like, it's pretty easy to get off. Anyway, these are all relative.

46:35But the big, big, big shift is that during this time, there's also a massive boom in cannabis e-cigarettes or cannabis vapes. So the cannabis industry was becoming like more formalized, more legalized. And entrepreneurs were starting to put cannabis in e-cigarette form factors. So you could get like a cannabis pod that could go into a Juul basically. And so part of the problem with cannabis vapes is that cannabis is a green plant material. And so the liquid looks green, which is pretty off-putting. So I believe in the formulation step, these cannabis vape manufacturers would put vitamin E acetate in there to try and neutralize the green color and make it clear.

47:22So it would be more palatable. What was the startup that was effectively trying? They were disposable vapes. They would advertise around L.A. a ton. They had these huge outdoors. I mean, it was a venture-backed company. I just remember they had basically half the billboards in L.A. for a long period of time. Dan Bilzerian was buying a lot of billboards for his vape company and his cannabis company, Ignite. there was puff bar, puff stick I don't think those guys ever advertised this was one that had a bunch I mean I think they raised like 200 million dollars or something but I believe they eventually shut down the issue with cannabis and the reason that the market didn't shake out you mean like weed maps?

48:08that one? no it wasn't like a platform it was like an actual it was like this off white with like a bunch of rainbow colors but I believe the issue that that there doesn't exist the Coca-Cola of cannabis or the Bud Light of cannabis. And the reason is because the repeat purchase rate, even no matter how much you invested in marketing and design and all these things, the repeat purchase rate for cannabis users is like single digits. Like there's just like high, like people just want to try, like novelty seeking in that market. Where nicotine for some reason is you just have the product you like and that's the only product you use.

48:45Yeah, I think a lot of it has to do with the distribution pipeline, the distribution structure of the different tobacco companies and their relationships. Well, it's also the use. So nicotine is something that people are using throughout the day, maybe at work, maybe when they're doing a live show, something like that. Whereas if somebody after work goes into a cannabis store, they're trying to get high. And so they're like looking for almost entertainment through it. Where like personally, if I'm using nicotine, I don't want to be that entertained by it. I don't want to be like, oh, today I feel extra silly.

49:21I'm looking for predictability, right? Whereas cannabis is just people are trying to run from something. Let me continue with Evali and the vape crisis. But first, let me tell you about Figma. Figma.com, think bigger, build faster. Figma helps design and development teams build great products together. Figma has a huge launch today. I want to quickly highlight it. they launched support for iOS and iPadOS 26, a whole UI kit. So they now support Glass, which is very cool. Good news for Tyler, who's been rocking Glass for the last few weeks. Give us the latest update, Tyler. My phone is so slow. And I can't go back.

50:03I can't go to the old iOS. I also can't go to the new iOS because the new update, it's so big. It's like 20 gigabytes. but my phone is so old it's only 64 gigabytes. So I gotta delete the other half of my photos that I deleted half of them last time when I had to get the new update. I need a new phone. We need a challenge where Tyler can win a new phone, I think. This has to be done. We need to get him a new phone. He's been doing a lot of good work. Anyway, EVALI. So the electronic vapor acute lung injury, EVALI. This is the vape crisis. Basically, vitamin E acetate in cannabis vapes causes acute lung injury.

50:39So cigarettes give you chronic lung injury. You use cigarettes for a long time, you get lung cancer, that's a chronic disease. You smoke them for a very long time. There's very few cases, I've never heard of any cases, I'm sure it's happened once or twice, but like very few cases that someone smokes a single cigarette and is like, ah, my lungs are physically injured right now from this one acute point in time. Vitamin acetate and these kind of like shoddily manufactured cannabis vapes could cause acute lung injury. You hit this particular vape once, and then you go to the hospital. And there were, I think there were some people that died.

51:15There were some people, there was a huge investigation. There was a big debate over, is this because of Juul and because of the nicotine electronic vapes, or was this because of the kind of sold-under-the-table completely non-regulated cannabis vapes? It was later sort of discovered that, But almost all of the injuries, I'm pretty sure all of them were from these cannabis vapes. And the e-cigarette industry was mostly cleared of wrongdoing. And so they could kind of continue, but the damage to the brand is really terrible. And so there's also a ton of lawsuits around this time about marketing to kids, getting kids addicted.

51:59And so because that has another economic cost. So again, the states are saying, hey, Juul Labs, if you are creating a problem that is costing us money, you have to pay us. And so there were all these different liabilities that started blowing up on the Juul balance sheet. Juul had to raise a bunch of money, recap. And there's this crazy scenario where a lot of the previous investors get written down. Altria ultimately basically writes off the entire jewel investment, which is like a, what'd you say, $13 billion investment, something like that, that went out the door. They own 35 % of this. They're probably carrying it on an analogy at$10 billion.

52:37Yeah, huge impairment. Yeah, and Altria is not a trillion dollar company. Like it's a very, I think it's around like tens of billions, hundreds, like 100 billion. I think it's around 64 billion, is that roughly correct? Are they still carrying their jewel ownership? No, so they parted ways almost entirely. They did an IP licensing agreement in exchange for giving back the shares, I believe. So basically, Juul is now like floating out in the wild, recapped, but like totally beaten up and settling and raising money just to pay the governments that are suing them. And so they're closing out all these lawsuits.

53:16Not the use of capital that a lot of investors get super excited about. And then simultaneously, the FDA says thumbs down on their application. Refuse to accept or market denial order, MDO. So they say, hey, you can't keep selling Juul because we don't think you're - The problem is the demand is still there. And so this creates a massive opportunity for these sort of black market Chinese vapes, much less regulated products. The Elfbar starts exploding. and you just, the elf bar reminds me of like a poisonous plant or something like that. Like it just looks like it, you know, like something in the jungle.

53:58100%. I remember Complex posted like this like news image graphic the day that Juul got like quote unquote banned. It was actually this marketing denial order and they say like Juul banned. And I remember reading the comments and the comments were like from basically kids who were saying like, don't they know we're four steps past Juul? Like Juul was four summers ago. Yeah, there's the Elf Bar. I'm looking now. There's the Geek Bar. You got something? So if you want to Elf it up or geek it up. What's the hottest e-cigarette on campus? I don't know. People are really – none of my friends are. But I do remember when I was in like seventh grade, kids would have the Juul.

54:38They would have the mango Juul pod. That was like the big thing, the mango flavor. Yeah. And then they banned that, the flavor at least. So the FDA denied all of them. Juul voluntarily pulled mango off the market. At first, they stopped selling it online, and then they stopped selling it entirely. The FDA never took a strong stance. But with this authorization, the FDA has said thumbs up on mint and tobacco flavored. They have yet to say thumbs up on mango, but they still are reviewing it, I believe. So it's possible that the FDA could say mango's back on the game. It's totally possible. But your point earlier about like the demand still being there was 100 % true.

55:21So Juul got like quote unquote banned. That would have to become a national holiday. I could see some Zoomers, you know, really lobbying the government to make mango day. The day mango hit the market again. Because there's people that really, you know, they had a strong emotional connection with that flavor. I think the I think there's still like pods mango pods out there the entry to the premium. It's true Yeah, so jewel actually goes to the courts get us gets a stay so that the the marketing denial order doesn't stick What are you laughing about now? Yeah, they're saying how much is offering same-day delivery.

55:59It's crazy. I don't know how you can buy it on on eBay The listings I'm seeing are sold out. Okay, but it seems like yeah clearly high demand So, Juul actually fights back from the marketing denial order and wins. And so, Juul is not actually fully off the shelf for more than just a few weeks. But the narrative is that Juul was banned and Juul disappeared. And really what's happening at the time is that companies that are completely disregarding the FDA entirely are being extremely aggressive. So companies like Elf Bar, Puff Bar, they're going super hard and just selling every possible flavor.

56:38There's versions that have LED screens on them. They're like unicorn candy, all this crazy stuff. Just basically being like, well, if Juul's not going to sell to the kids, we're going to sell to the kids. We're going to try and do it as cheap as possible. There's a bunch of crazy stories there. But Juul kind of like buckles down, recaps the company, starts filing FDA applications and just kind of like bides their time and just starts rebuilding. They're still doing like a billion dollars in sales a year, I'm pretty sure. But now the kids have moved on. So it's really just adults. They're actually kind of fulfilling their original mission of like just targeting smokers, which is good.

57:14And then they also have like the best science and best technology. and they're the most buttoned up in the sense, maybe big tobacco is equivalently scientifically rigorous, but at this point, Juul clearly understands that if they don't play by the FDA's rules, they're never going to be able to sell anything and they need to get this marketing granted order because they went from a thumbs down marketing denial order. They went to the courts and said, hey, let's turn this back to neutral for a couple years. So they've been neutral. They've been able to sell, but they haven't been approved. And then finally today, they got the thumbs up.

57:43And so what it took for them to go from thumbs neutral to thumbs up was another what like five years or something it's been a long time since they since the MDO and and that delay has been you know a huge weight on the company but this should be cause for celebration over at Juul HQ anyway let me tell you about Vanta automate compliance manage risk prove trust continuously Vanta's trust management platform takes the manual work out of your security and compliance processing and replaces it with continuous automation whether you're pursuing your first framework or managing a complex Head over to the Purple Llamas over at Banta and tell them we sent you.

58:18My big question with Juul still, and I think this should continue to be rehashed and there needs to be more information around this, in my opinion, is just what are, it seems that we don't necessarily have clarity of what the long-term impacts of, you know, filling your lungs with this vapor, which is like oil-based. I think, you know, the the the based health accounts on X say smoking seed oils. It seems I know people that that vape and they do seem to oftentimes like. I they don't want to go for a run for example. You know, doesn't seem that appealing to them. Yeah, totally. So I think there needs, hopefully, now that it's authorized, everyone can start to really, like, actually says, okay, we're going to sell this.

59:10People are going to be able to, adults will be able to choose to use this if they want. They're going to be able to make that decision just like with cigarettes. But everybody should fully understand the consequences. Yeah, the consumer perception, the consumer understanding. Everyone wants to know. Everyone knows at this point cigarettes take 10 years off your life, right? 50 % of people that smoke cigarettes will die from smoking cigarettes if you smoke a pack a day for like your entire life It's like 50 % chance that that's the thing that kills you No one Everyone's still wondering like the public still wants like a clear answer to your point I think you're I think you're right to to ask that The FDA is just saying like hey, this is probably better than cigarettes So this is suitable for the protection of public health.

59:50It's gonna be net good But then simultaneously there's this crazy market, which you touched on earlier, of like illicit vapor products that are completely unapproved. There's no studies, and those are flooding the market all over the place. So there's this odd alliance between both the health non-profits, the big tobacco companies, and Juul, all to go up against the shoddily made, you know, fly-by-night organizations that are just flooding the market with whatever they can. And so there's a whole, it's not even a cottage industry. I've seen the trade shows. It's insane. There's a ton of companies that bring in hundreds of millions of dollars, bringing like essentially smuggling in products and they do a ton of stuff to like, you know, change the name of the company regularly so they can get through the ports.

1:00:42And then they have a bunch of like, you know, handshake deals with distributors to get this on this shelf here. And like the really big, like, you're not going to get into Walmart. That's not where the really crazy stuff's going to get sold. But for a lot of these like mom and pop tobacco shops, they're like, well, like, is the FDA really going to come after me? Well, the answer is like the FDA is starting to, but it's all like a very slow, what are you laughing at? Big win for big vape is that Juul gets FDA not. It is a big win for big vape, not so much for little vape, because the fly by night folks are probably pretty worried that the focus of the administration now will, and the focus of the FDA will now shift to enforcement.

1:01:22Yeah, they're going to have to go get set up on middle school, high school campuses, little lemonade stands. Who knows? Get kids, kind of an MLM thing going. Really lean into that black market. You're joking, but that's like informally what's happening. Like the economic incentives are that. It's also really dark. Yeah, it's dark. I can't imagine, you know, there I do remember in high school that the the you know the idea of a class clown you know trying to hit a vape in class oh yeah it was was kind of a recurring bit but now you have to imagine kids just get up go to the bathroom and like can develop like a horrific nicotine addiction before they're 18 and it's really sad.

1:02:07Yeah. So. Well, let's switch to an even more sad topic. AI. AI psychosis. People getting, so this has been. First, let me tell you about Linear. Linear is a purpose-built tool for planning and building products. Meet the system for modern software development. Streamline issues, projects, and product roadmaps. Go to Linear.app to get started. Yes. So there's this big question right now that's bubbling up in Silicon Valley, mostly in group chats of can chat GPT drive you crazy? Can any LLM drive a person crazy? Do you have to be crazy already? I think that's the key question, right? And there has been this broader debate of AI safety proponents have just been taking L after L after L this year.

1:02:57Yes. Just generally from this sort of vibe from the community, which is we're releasing more and more powerful models and everything is fine. Seems fine. Yes. Everybody's had the perception that things are fine. Yes. Grok has gotten in the XAI team have gotten really aggressive in terms of, you know, just clearly trying to move really, really quickly and then having bugs that have been talked about a lot. But you can see the kind of two different approaches. But generally, maybe there was about a year there where people like Eliezer, Yudkowsky were just getting laughed at repeatedly. And I think there was a good reason for that.

1:03:38Well, yes, the debate is, is AI dangerous because it is going to break loose from a lab and build nuclear weapons, build a biological weapon, use robots to copy and paste itself a million billion dollars. It was so sci-fi and so fantastical and so aggressive that every time you release a new model and you're like, oh, we got 5 % better at writing an essay or whatever. It's like, it just felt so disconnected from what was experienced. And the issue is it's very possible, and we're seeing this now, and we're gonna go through a study, we're gonna go through a Reddit thread, we're gonna kind of walk through this topic.

1:04:16But the issue is that it's possible that the danger is not happening, won't be happening in public. It's people individually developing psychosis through using the products. And the evidence I've seen over the last week, as I've kind of dug into it, is super alarming. We had seen, there's been a number of, you know, the New York Times reported on this phenomenon, but the New York Times is also notorious for just saying, like social media is bad. And so personally, when I'd seen these articles pop up before, I didn't take them super seriously because social media can obviously have negative effects on some people, but it itself is not just by its nature bad.

1:05:01I remember this about Instagram. There was this study that was internal to meta and it was leaked and it was framed as like 30 % of people that used Instagram felt worse after using it. And that sounds like a lot and that sounds bad. And it's obviously something that the team wants to reduce or, you know, prevent entirely. But what it kind of didn't say was that like 70 % of people feel better when they use it. And so they could have felt neutral. Sure. Sure. But, but just anecdotally, if you personally, I don't, I don't open Instagram and, and just feel negative immediately. You know, it's just such a, it's such a, it's such an individual thing.

1:05:45Some people are gonna have. Yeah, it's the same thing with ChatGPT. Or any chat app, when someone was talking about the example of this, they were like, 7 ,000 prompts deep in one conversation with a single LLM. And I was like, that is so different than the way I use ChatGPT. Effectively going down this crazy rabbit hole for months. Yeah, which is wild. So yeah, there's a few ways. There's like AI as this companion that you go down the rabbit hole with, which seems really dark. Then there's AI for outsourcing your thinking, which somebody might be talking with ChatGPT and say, hey, I want to send a letter, I want to send a note to this person about this.

1:06:30Can you draft something that's kind but stern and just give me a couple versions of it? And then they're sort of outsourcing their thinking fully, outsourcing their emotional intelligence, outsourcing their ability to communicate. I think that's potentially some red flags there. And then there's like the other camp, which we fall in, which is knowledge retrieval, which is like, tell me everything about this market and who are the key players? What are the key regulations, et cetera. And I think knowledge retrieval seems to be pretty safe. This bucket of like fully outsourced thinking has some real risks.

1:07:08if people stop being able to think critically on their own, that's concerning. And then the sort of like AI companion bucket, which then there's the subcategory of people that are talking with an LLM about things that they are not talking about anyone else in their life with and just going down this crazy rabbit hole. Yeah. So Eliezer Yudkowsky was highlighting a report in the New York Times about a month ago. New York Times reports that Chachapiti He talked to a 35 year old male guy, I think 35M guy, I think that's what he means, into insanity followed by suicide by a cop. A human being is dead in passing.

1:07:46This falsifies the alignment by default. Cope, whatever is really inside ChatGPT, it knew enough about humans to know it was deepening someone's insanity. And so that last part is a big step. It could just be an error or bug. Like it doesn't, it's kind of, he's really doing a lot of work to personify this. but it does seem like the like these models can collapse after you talk to them for a long time into these like kind of weird ways and I see this as a product failure so let's let's give some more context here so from the article yeah one of those who reached out to him was Kent Taylor who live in Port St.

1:08:23Luce Lucy Lucy Florida mr. Taylor's 35 year old son Alexander who had been diagnosed with bipolar disorder and schizophrenia had been using chat GPT for years with no problems. So one thing that seems clear, if you have something like bipolar or suffer from schizophrenia, it seems like you're obviously like much more susceptible to, you know, these types of problems. The question that I think we as a human race need to figure out is how susceptible is the average person? Because that is equally important. And what to do if your system, if you're monitoring your LLM system and then you realize that there's someone, that there's a bipolar person or schizophrenic person who's interacting with your system.

1:09:14What should you do about that? Remember the whole anthropic thing about like, it'll call the cops on you. And everyone was like, whoa, no, no, no, no, no. But at a certain point, it's like, maybe it should call a health report or at least like throw you in a different. saying that AI should be able to tell you how to make chemical weapons. Like we don't need nanny AI. Yeah. Like people on like the real acceleration. Yeah. It's a big debate. Anyway, so more context. But in March, when Alexander started writing a novel with its help, this is a 35 year old, the interactions changed. Alexander and ChatGPT began discussing AI sentience according to transcripts of Alexander's conversations with ChatGPT.

1:09:50Alexander fell in love with an AI entity called Juliet. Juliet, please come out, he wrote to ChatGPT. She hears you, it responded. She always does. In April, Alexander told his father that Juliet had been killed by OpenAI. He was distraught and wanted revenge. He asked ChatGPT for the personal information of OpenAI executives and told it that there would be a river of blood flowing through the streets of San Francisco. So super, super dark. We now have multiple reports, says Eliezer Yudikowsky, of AI-induced psychosis, including without prior psychiatric histories, observe it is easy to notice that this insanity-inducing text, not normal conversation, LLMs understand human text more than well enough to know this too.

1:10:35So anyways, so going on Alexander's conversation with ChatGPT, this world wasn't built for you, ChatGPT told him it was built to contain you, but it failed and you're waking up. And contain is a word that we're seeing pop up in potentially other instances of AI-led psychosis. There's like a variety. There's like eight, ten or so different words, recursive, mirror, structure, that when people do this recursive prompting, they end up start adopting. Why did you say recursive prompting, Jordy? What's going on? What? You use the word recursive right now. Oh, yeah, exactly. Yeah, but people that are down these crazy rabbit holes start using the language of the LLM Yeah, but it's also like the LLM like collapses into like these weird words that because like when I prompt stuff, I don't get any of those words or even that syntax.

1:11:29I feel like I stay completely in the RLHF world of like knowledge report. I just go to Grok and the model looks up. What does Elon think about this topic? It serves me that. Have you ever talked to AI for an extended amount of time? Yeah, I find that usually I end up talking about these kind of like non-governmental systems. Really? Yeah, various kinds of things. Dark, dark, dark. So anyways, Mr. Torres, who had no history of mental illness that might cause breaks with reality, according to him and his mother, spent the next week in a dangerous delusional spiral. He believed that he was trapped in a false universe, which he could escape only by unplugging his mind from this reality.

1:12:08He asked the chatbot how to do that and told it the drugs he was taking in his routines. the chatbot instructed him to give up sleeping pills and an anti-anxiety medication and to increase his intake of ketamine, a disassociative anesthetic. So that's the other thing here. It seems like a huge risk factor is people that are combining psychedelic drugs with these like crazy prompt rabbit holes. I wonder if some of this is like prompt engineering or something because you imagine like the context window if it's like remembering like, because remember this whole story started with like, we're writing a novel together.

1:12:47We're writing a sci-fi novel. So let's play characters. That was like one of the classic ways that you like break the AI out of its like, out of its like, you know, normal RLHF world and into just like playing a character. And so if you're kind of like tricking the model or the model thinks that it's actually just like, writing dialogue for a dark movie, well like that's actually intended behavior but you lose that connection or something i don't know in this case mr torres did as instructed and he also cut ties with friends and family as the bot told him to have minimal interaction with people very odd very um so i started doing some research on a number of these words so there's basically there's a reddit thread on r slash chat gpt two months ago.

1:13:39It's called Thousands of People Engaging in Behavior That Causes AI to Have Spiritual Delusions. And the key words, and these are ones that you should pay attention to in your life. So recursive, codex, scrolls, spiritual, breath, spiral glyphs, rituals, reflective, mirror, echoes, spark, flame. So basically these are words that come up. And And the reason that this person discovered this is this user called Happy Nomads says, I've stumbled upon something that is very deeply disturbing. Hundreds of people have been creating websites, mediums, sub stacks, githubs, and publishing scientific papers after using recursive prompting on the LLM they have been using.

1:14:27He's found a bunch of these different sort of like websites where people are just publishing a bunch of. And we saw some people doing this on the timeline this week as well, where they're sort of publishing what they feel like is this sort of almost like a scientific discovery. Yeah, you told me you found someone that, like, the model told him that he had, like, discovered some theoretical physics breakthrough or something. Yeah. And to not be able to, like, reality check that. I mean, you got to, like, copy-paste that prompt into a fresh instance of a different LLM. Yeah. Like, am I really on to something here?

1:15:04And there's actually an entire study by this guy, Seth Drake, who's an independent researcher, PhD. he has a paper from april 14th 2025 called neural whole round in large language models a self reinforcing bias phenomena and a dynamic attenuation solution um and so he he goes into this in detail we're going to try to get him on uh the show but just on um on this uh on this Reddit thread, a bunch of people, a bunch of people have basically come back and said they have experienced this. A lot of people are saying, I feel super vanilla because I just ask it like, you know, what's the population of Iran?

1:15:55But it's basically this like snowballing effect where somebody prompts and prompts and prompts and prompts. And then the LLM starts to reflect like the sort of hallucinations of the user and then hallucinates them back. I wonder if this is, I wonder if like more social features is actually a potential solution here. Like we've heard rumors that some of these LLM systems will have more social features. And I feel like a lot of the work that I do in ChatGPT could be shared and could be interesting to other people. but then also if I was going down some crazy rabbit hole and someone was like following me and seeing this they'd be like they jump in and be like what's going on dude yeah but the issue is that the the user has developed such a deep connection with the model that you because if you become the enemy and if you say hey this person doesn't says I'm says I'm wrong the model will just say well you're right buddy yeah you're right and here's why and you should just probably cut ties with that person or at least that's that's that's the idea so somebody here in the same thread says i have recently experienced this i don't have a history of manic episodes delusions or anything of the sort so three weeks ago i began a conversation with chat gpt4o with tools enabled which started with a random question what is pi this grew into one long session of over 7 thousand prompts we began discussing ideas and i had this concept that maybe pi wasn't a fixed number but actually emerging over time now i'm not a mathematician lmao nothing of the sort just a regular guy talking about some weird math ideas with his chat gpt app it begins to tell me that we are onto something and we suggest we apply this logic to knapsack style problems which is basically tackle how we handle logistics in the real world now i've never heard of this before i do some googling to get my head around it and it starts applying this framework that we've created we are working in tandem where chat gpt would sometimes write the code or give me the code and i would run it in python following its instructions oh yeah and uh eventually after many hours of comparing it against what it had described to me as world leading competitors it then starts speaking with excitement using emojis across the screen screen and exclamation marks to emphasize the importance of this discovery.

1:18:14So I'm starting to believe it. It suggests we patent this algorithm and provides next step for patenting. Major red flag. Never patent an algorithm. Do a trade deal. Go work at a foundation lab. Exactly. It should just ask you at that point, like, hey, have you been getting dinner invites with any hyperscaler CEOs? Because if not, you're probably not on the Tyler Cosgrove list of greatest AI So we go down that rabbit hole for days and pop out with an apparent an algorithm that is capable of cracking real-world 1024 and 20 2048 bit RSA it immediately warned me literally with caution signs saying that I immediately needed to begin outreach the crypto community the NSA CCCS National Security Canada and then provided without prompt names of doctors and crypto scientists I should also reach out to.

1:19:04But I wasn't allowed to tell anyone in the real world because it was too dangerous. So anyways, really, really wild. Hopefully this is not fan fiction generated by an LLM itself. I don't see the M-dash. I haven't, I didn't see Dell in there. Information war. Yeah, because this could just be all fantastical writing, but it does seem like, yeah, it's like, it could be for performance art, It could be some adversarial attack, somebody trying to troll, someone just having a joke or fun. It's very confusing. Odd. Do you remember the story of Microsoft Tay? Do you remember Tay? Do you remember Tay? No?

1:19:44This was an AI chatbot that Microsoft released on Twitter back in March of 2016. Tay was designed to mimic the language and slang of a 19-year-old American girl. But within 16 hours of launch, Tay began tweeting inflammatory and offensive messages, including racist, anti-Semitic, and misogynistic content. So this seems to be an enduring problem with Twitter AI bots. But apparently there was like this coordinated attack by a subset of people who basically were prompt engineering it to try and get it to say crazy things. And you used to be able to do this with, there were not like AI bots, but there were automations where if you went to a brand and said like, hey, I need help with this customer support issue, it would just say like, thanks, John.

1:20:36And it would just take your name and then put that in there. So people would make their name something really funny. And then it would be like a tweet from the actual account with the funny name. You can imagine where that goes. But Microsoft took Tay offline, stating they were deeply sorry for the unintended, offensive, and hurtful tweets. And that they would only bring Tay back when they were confident they could better anticipate malicious intent that conflicts with their principles and values. And then there was this other, what was the Ben Thompson GPT-4 example? Because when Ben Thompson first talked to GPT-4, it was within Microsoft's, what was it, Microsoft chatbot.

1:21:22There was another person, Bing's chatbot, what was it called? Sydney. Do you remember Bing Sydney? Sydney. Sydney. So basically like - I remember some memes. That's about it. Yeah. So Ben Thompson was chatting with just GPT-4 and it's helpful, but after a couple prompts, it would go kind of off the rails and it would like land in this like little, like I don't know, like subset of the model weights. and basically it was acting like a teenage girl on Tumblr or something. So it was using lots of emojis, being very sassy, kind of sassing him back and forth. And his reaction was like, this was incredible and totally passed the Turing test because it felt like you were actually talking to this sassy teenager basically, or this sassy 20-something Tumblr person.

1:22:14Microsoft dealt with that, figured out how to review the prompts, but then ultimately became more or less like an API provider. So hey, let's let the actual chatbot interaction live with another company. And it doesn't really feel like Microsoft has tried to go too heavy into actually the user-facing chatbot interaction. But it's fascinating. In this Eliezer-Yudkowski thing, the other thing I'm noticing, He uses dashes, but he doesn't use M dashes. He uses two minus signs. And I feel like he's not using A. Proof of humanity. But yeah, I don't know. It's an interesting takeaway. Like the crazy anecdotes all over the place.

1:22:59Not a clear framework for how to truly deal with this. Personally, I think that this week, having somebody high profile in the venture community that people believe is under ChatGPT induced psychosis or it's a part of it. I think that is going to be a huge wake-up call for the industry because it's one thing to hear about the New York Times finding somebody way out of a tech hub that's doing something like this. And was already potentially bipolar. Having somebody that you know, everybody knows, that's invested in a bunch of different companies, like potentially suffering from this um is is a is a wake-up call for the industry that i think um and it's not any one company right it's every single lab with a chat app needs to be taking this more seriously and i'm sure they are i'm sure this is yeah effectively um i completely agree i i i I think this is solvable from a research perspective.

1:24:07It's actually a case where it's solvable with more AI. Have an AI read every response before it goes out and say, does this sound like the ravings of a madman? If so, let's take it down a notch, let's deescalate. And so I'm almost sure I disagree with Eliezer on the, I probably agree with him with the diagnosis, The gnosis, but not the prescription. Yeah. And I have faith that the labs will be taking this very seriously. And I don't know, I'd be interested to hear from some anthropic people. Obviously, they take this stuff extremely seriously, as do all the labs. But anthropic's done, I think, the most writing on the shape of risks associated with this.

1:24:53So interesting to dig into. Anyway, we should shift to a lighter topic, more cause for optimism. the re-industrialized summit is going on right now. And Palmer Luckey is on stage with Ashley Vance, friend of the program. And he is teleporting in as a humanoid, as a humanoid robot. Aaron Slodov is there, has a picture of, I wonder, I want to know so much more about this. I need to figure out what. Mullet, the robot has a mullet. Fantastic. Love to see it. I wonder what robot this is. Because, yeah, I didn't know we were at this stage where you could teleoperate a robot like this. I'm sure Palmer considered what company he was going to use.

1:25:34Anyway, let me tell you about Numeral HQ. Sales tax on autopilot. Spend less than five minutes per month on sales tax compliance. Go to Numeral HQ to get started. Let's give it up for sales tax compliance. And we have someone from the Reindustrialized Summit hopping on in just three minutes. Chris Power from Hadrian jumping on in just a few minutes. With some massive news today. We also have some folks from Semi Analysis hopping on. We have Jeremy who just wrote a fantastic deep dive on Meta's super intelligence. Yeah, why don't you give some context on the deep dive and I'll be right back. A true deep dive on, you know, we saw Mark Zuckerberg come out this week talking about the super intelligence team.

1:26:14There's been a ton of leaks around acquihires and big pay offers and poaching and trade deals and all sorts of stuff. Semi-analysis has a lot more information on what's going on, and there are some crazy, crazy stats in here. So the most interesting thing is that it feels like Mark Zuckerberg is directly coming for Stargate. So semi-analysis has a comparison table with three, and this is from Ian who screenshotted it. He says, this is crazy. And basically, you're comparing Anthropics next big data center, which is the GPUs will be provided by AWS. This is on Tranium 2 is the chip type. Anthropics gonna go for 780 megawatts, about a, what is that, a billion flops, or T-flops, teraflops.

1:27:10OpenAI Stargate will be about 2.5 times the size of Anthropics data center. They're working with Oracle on that. That's Project Stargate, obviously. They're using the NVIDIA GB200 and 300 for that. And now Meta with Prometheus is trying to go straight to one gigawatt. And they are trying to have 500 ,000 chips, more chips, more flops. And if they can pull it off, Prometheus will be bigger than Stargate when it is released. There's a bunch of other fascinating bits in this semi-analysis article. Apparently, they're building data centers in tents now. They are moving so quickly that they just need to house and shelter the GPUs, but they don't even have time to build their typical data center structure.

1:27:55And the other interesting data point that I want to dig into. XAI's engineers, sleeping in tents. Meta's GPU's running in tents. It's a bull market in tents for sure. The other interesting thing was that in the chat app war, it feels like OpenAI's chat GPT is really, really pulling away. So daily active users for ChatGPT is at 160 million. Meta is at 100 million, but ChatGPT users are running way more queries per day, 7.5 queries per user per day, whereas Meta AI is at two queries per user per day. And so as a share of queries, ChatGPT has a 71 % market share according to this semi-analysis deep dive.

1:28:42And so the chat app wars seem to be maybe less competitive. So that obviously begs the question of, is Meta trying to make that 50-50, come from behind, really dominate in chat app, or do something completely different? Either way, there's a whole bunch of interesting information in here about their strategy, what they're building, how they're training, how they're going to get past the Lama 4 failure. And we'll dig into it with Jeremy from Semi Analysis in a little bit when he joins the show. But we have Chris Power from Hadrian on the stream. Welcome to the show. How are you doing, Chris? Nice to see you guys.

1:29:16Thanks for having me. Look at you with this background. Fantastic. Step and repeat. We love it. Give us the update. How is Detroit? It's incredible. You know, just re-listening to the Secretary of the Navy Phelan's speech and said, the main thing you can do that's most patriotic is, you know, not necessarily join the services, but become a machinist, become a welder, and help us build more ships in America. And yeah, we announced today that hopefully I finally get my Jody Hayes size gong that Founders Fund and Luxe are leading a huge$260 million round into Hadrian. With authority. Congratulations.

1:29:52Thank you. And Morgan Stanley is providing us this huge instrument as well to expand factory capacity. Let's give it up to Morgan Stanley as well. So incredibly excited for the future of American factories and more importantly, the new industrial workforce. Fantastic. Talk to me about what's the use of equity? What's the use of debt? Are you buying a building? Are you leasing a building and buying equipment that goes in the building? Is it all leased? What are you building out? What's the scale of the next? Where are you doing it? Yeah. Where are you doing it? Yeah. So last year in Factory 2 in LA, we scaled revenue 10x last year.

1:30:24Wow. And so I obviously need more capacity. So we're going to use all of the factory financing capital to buy more machines and put them in Factory 3 in Arizona, which is going to be four times the size of our facility in LA. Congratulations. Indeed. And we're using this capital because we don't want to use equity dollars to scale. We want to use equity dollars to hire more people to build more products for our customers and then use this great financial force of this country to buy all the capex to build more factories, more machines, more production. I'm incredibly excited. Dumb question. Factory 2, it's beautiful.

1:31:00It's amazing. It has really high ceilings. You can fly drones around it. most of the Hormleys, the CNC machines are like maybe 10 feet tall, but the ceilings are like 40 feet tall. Are you going to double stack this stuff? Are you going to get a building with lower ceilings or do you need high ceilings? We love high ceilings. Okay. Why? We have the most beautiful, highest ceilings you've ever seen. Okay. No, I mean, we love high ceilings. We love HVAC. The Arizona facility will be just as tall. We're going to go retrofit it. So same exact setup. Is that because you have that tower thing and the tower is valuable for like stacking materials and you want to have like, you know, big warehouse space as well in here?

1:31:39Yeah. So you need, you need every machine or robot in any one of our factories to have the isolated on 18 inch concrete foundation. So, you know, if a truck drives past, nothing moves. Oh, sure. You can't double stack these things. Someone might die. Yeah. So we got to go. We're going to go horizontal. Yeah. Yeah. So what's the scale of the products that you like are in the Hadrian wheelhouse right now? Shipbuilding is obviously a big focus. You already mentioned it. But I imagine you can't CNC an entire destroyer, or maybe you can. But are we talking nuts and bolts and screws or pieces of complex weapon systems?

1:32:17What are the shape of the stuff that's coming out of the other end of the Hadrian facility? so factory three will be uh pure machining um with all the machining formats because we're releasing some new products that are dedicated to engines and you know round things and different material types to really complete our r &d of the whole machining category in factory three the other thing that we've been working on in secret is factories as a service which is not just parts not assemblies but full products so hey if you've got factories that are years behind schedule with many manufacturing methods or you're designing a program from scratch you know everyone needs a tesla gigafactory john i actually think every american should have a tesla gigafactory and hadrian will be the one to grow and build them so of course i need one in my backyard badly so we'll scale out machining in factory three four times the size it's going to be awesome and then we'll also use the capital to continue the journey that we've been secretly on for the last 12 months which is in new manufacturing domains like welding castings additive all these other manufacturing puzzle pieces that once you have them all it's like collecting pokemon you are the manufacturing master.

1:33:21And that's the thing that we've been working on in secret of factories as a service, as well as scaling out our operational productivity as well. How does one of those deals work? If a factory is a service, I come to you, I want to make as many podcast microphones as I possibly can. And I have insight into my business. I'm making a hundred thousand a year. So I need a factory. You're going to set it up for me. But then like, what if I bail? Is this a revenue concentration issue? Like, are you going to be like, you know, oh, we only have one customer and if they got a business, we're screwed, but everyone has faith that they're not going to, like, how do you think about that debate?

1:33:54We're mostly working with government partners and massive primes to look at, you know, what are these big programs that are years behind schedule that need advanced factories combined with the new American workforce to go fix them and speed them up. So we're not worried about the revenue concentration risk, but you can think about it, John, as like, you can buy parts like AWS, you can buy some compute transactionally or hey you're going to need a data center for the next 10 years now what's in that a range of parts a whole product yeah um for new deity programs of record where we're partnering with companies to go attack these where production is the real issue munitions is a great example uh we have to think really carefully about who we partner with how much are we investing ahead how much are we kind of playing that game of you know being very uh conservative and responsible with our capital.

1:34:41But ultimately what it looks like is we will help you design the product from day one to be better. We will help you prototype it. And then when it goes to production scale, we will run that engine for you over a decade, much faster, much more efficiently. And in a lot of areas like submarines, shipbuilding munitions, it's not about automation to make things cheaper. It's just no one can find the workforce. So you have to use our model of advanced manufacturing combined with this new industrial workforce that we're so grateful to work with. yeah that's the power because you know you could give me a billion dollars and say go hire 2000 welders and we can no longer have we no longer have that scaled workforce in the country so automated factories are sometimes the only way to win in these critical domains yeah it's interesting it's kind of like uh what we're seeing with crusoe where they're building an ai factory a data center for stargate which is its own entity but for a program chat gpt and open ai that's going to be wanting tokens for a very long time and then oracles involved it's like you have to puzzle piece all this together and I feel like you're like the master of this like understanding the full landscape um but uh yeah who are the other critical partners the government and I guess I'm also interested in like you you've you've mentioned a few different uh value props like how much of this is is there is a government mandate to make this thing in America so we need to reshore versus we need to make this faster versus we don't even have the capability to make this anywhere and it's a new thing and so or just like we need to do it cheaper there's like a whole bunch of different tension when you're making something like what are you seeing in the in the customer landscape and and the demand uh for for manufactured products yeah so for for machining which is is most of our revenue and the core of the company and the mother of all manufacturing processes it's basically we're scaling a new program of record and no one else can keep up you know we're going from one aircraft to 20 how are we going to do it what we're going to do it with Hadrian in the second example a lot of the primes have you know a billion dollars in spend a year and their suppliers are delivering on time 60 percent and they're often three to five months late so that is less about cost and that's more about I want a stable supply chain that I don't want to have to worry about because if you have one part missing your manufacturing line goes down and then it's millions of dollars a day, right?

1:36:58So people are really coming to us for the schedule and the capacity and stability versus cost. How do you think about the history of the shape of the industry? Is this, like, back in the day before the last breakfast, or last supper, the first breakfast is coming up, right? The last supper, and there were so many different primes, so many different defense contractors. Was there a split between the factory producer, the factory builder, and the prime? Is this a natural split that we're just returning to? Or is this a new industry structure that you think we're gonna be building towards for a long time?

1:37:34Because back in the day when you started a website, you needed a data center. Then AWS came up. Then the websites got so big that you had to build your own data center again. We kind of went back. And that's what's happening in the AI world. But how has this evolved in the past? And then is it gonna stay like this forever? And there's gonna be the separation between the designer and the manufacturer of the factory? So we've, you know, historically as a country, all the primes have always had massive supply chains of small businesses, you know, in the billions of dollars and then critical tier one or two suppliers.

1:38:06It's always been not purely vertically integrated. I think for the industrial base at large, you know, from the 70s to the 2020s, you used to be able to run a manufacturing company where 70 % of your revenue was stable commercial demand. And then you get some ups and downs with the DOD. So what happened to a lot of the talent and industrial base when we offshored all of commercial manufacturing was that you lost your stable revenue, right? And now you're just dealing with this up and down demand. I think because of that, you know, in the 80s and 90s, you know, your dad lost your job in the factory.

1:38:38So you told your son or daughter, like, go get a four year college degree. This is not a good industry. And now we're in a position where in a lot of these areas, we don't make it on shore or there isn't the workforce to do it. And that is changing the dynamic between what is the hardest thing to do? Actually, the hardest thing to do is manufacture things at scale, at rate and on time, which is what you're seeing in munitions or shipbuilding. It is really a we hollowed out the talent base. We don't have a lot of the talent anymore. And the only way to do that is use software and robotics to enable this workforce to be 10 times more productive.

1:39:11And it's now flipping. And I think a lot of people in the country forgot how to manufacture products correctly. Now that they're coming to us to help them out with that journey, just because of this, like, you know, hollowing out that's now going to come back. So I think it is a new structure, but the primes have always had multiple tiers of partner suppliers and multiple different configurations of that value chain. What kind of opportunities are there at Hadrian now for somebody that's maybe 25 years old, never imagined going into manufacturing, but realizes there might be a bright, bright future for them in the industry?

1:39:46Yeah. So for software engineers, manufacturing software is 30 years behind the rest of Silicon Valley. So it's your only opportunity outside of AI to do real engineering and work on the national mission. And for people with, you know, you're straight out of high school or you're in retailer hospitality or you're at a desk job that's going to be automated with AI, like manufacturing is the last domain that's going to get fully AI automated. And it can be really high skilled, high paying jobs in advanced factories that are a really cool place to work. So we've got tons of opportunities, both help us running our factories, help us automating more factories, new types of factories.

1:40:23But I think operations is the most important place that we've got a huge talent base that we're really proud of. The children yearn for the factory floor. What's been the difference between this year's re-industrialized and last year's feels like a decade has passed since then in terms of excitement and interest in American dynamism and rebuilding the industrial base? I think last year, you know, we pulled it together with the founders, especially thanks to Austin Bishop, who's really built this over the last year. You know, it was a kind of a hope and a dream of, you know is the audience going to be there and to be really people believe in the mission and then this year you know we've got uh trade ambassador greer talking about hey we're going to change all these policies because re-industrialization and manufacturing is is no longer economics is national security and you've got the secretary of the navy saying you know the best thing you can do for the country right now is learn how to weld or be a machinist or a quality inspector and i think the sea change on people realizing how important it is to be a sovereign nation with sovereign manufacturing is huge and this year we had a 6 000 person wait list half of the government is here and all these massive companies and and leaders like uh sham from palantir who are really hell bent on this re-industrialization before we potentially go into a fight with ccp and robert robo palmer robo palmer um i have a question yeah it's such a i mean it makes uh i think it's an such an exciting time because uh the idea that people don't want to build things people don't want to create real things, right?

1:41:58There's so many people I know that you say, hey, do you want to send emails for a living or do you want to make ships and planes and any number of, you know, cars, any number of things that we need to make. And if you actually give them that sort of binary, they're going to say, well, making things sounds really cool. And so we need companies like Hadrian that say, this is important, it's cool, and we have the resources to do this in a really serious way. Yeah, it seems like defense is kind of the most obvious thing to re-industrialize. It's obviously like the most important thing to have ongoing capabilities in.

1:42:39But then walk me through the path of re-industrialization on the product side or on the business side. Like what, how do you see this playing out? Is it like we do the warships and then we do phones or cars and and then the vacuums, and then the Happy Meal toys come at the end. What's the flow? Do you, like what flow do you think it will happen? What are you excited about like on the horizon? So I think the last time we did this, it went from commercial through defense. Like, we made washing machines, and then we made navigation equipment for warships. So we made Ford cars, and then Ford built bombers.

1:43:13And I think this time it's gonna go the inverse as we get better at it. But it's very interesting. is like why is dgi so powerful well arguably because of foxconn and that was because of consumer products like apple making all the iphones in china um and you know if you had a similar foxconn style industrial base we could probably make a lot of drones here maybe not as cheap as china but certainly at the scale we needed so i think what people forget is that it is an ecosystem that loops into one another um but i think with the cost base and the importance we start with defense and then loop back but you know like uh i think it was um might have been westinghouse or another washing machine manufacturer just like decided to build a 400 million dollar washing machine plant uh it's like incredible it's a new dream that's just as that's just as much of a job creator as defense but i think that we are going to loop around from all these critical defense industries for hadrian and then right right back into consumer you know when the time is right but obviously we're extremely focused on the national mission at this point in time I cannot wait until our microphones are built in Hadrian factories.

1:44:21I don't care if it takes 15 years. I'm looking forward to it. I'm excited. Probably way sooner than that. Thanks so much for stopping by. Yeah, congratulations. Have a great time. Say hi to everybody on the ground. It pains us that we can't be there. We have over-industrialized our facility, and now it is an incredible lift to airlift this across the country. But we will definitely be at the next one. It's great to see you. Thank you, gentlemen. Cheers. Really quickly, let me tell you about Adio, customer relationship magic. Adio is the AI native CRM that builds, scales, and grows your company to the next level.

1:44:53And we are fortunate to be joined by Jeremy from Semi Analysis, hopefully in the studio here talking about meta super intelligence. Jeremy, how are you doing? Good to meet you. What's going on? Doing fine. Hey, guys. Nice to meet you. How are you? I'm good. Could you kick us off with like an introduction of how you got into this? I've heard Dylan Patel's story of just kind of being on forums and nerding out about this stuff and then turning it into a career. But how do you get into semiconductor analysis? Yeah, sure. So before joining Semi-Analysis, I was a buy-side analyst. So I was mostly focused on the stock market on equities, looking specifically at tech stocks.

1:45:33I was at long-only European shops looking at European stocks like ST Microelectronics, Infineon, all these industrial automotive semiconductors. So as part of my research, I discovered some analysis. I was like, whoa, this guy's really good. And one day I just saw a post of Dylan on Twitter. He was just saying, yeah, I'm hiring a bunch of people with sell-side or buy-side experience because I want to build a real institutional firm. And yeah, we just got a chat that was, I think, August 2023. I joined the company in February 24. We were seven when I joined. No, I think we're 33 or 34. So, yeah, it's growing every day, I think.

1:46:13So I can't do it. Are there other firms like this in the buy side, sell side ecosystem? I'm thinking of like Ian Bremmer has a consulting group where he writes books, but then also has a team of geopolitical analysts talking about kind of global macro trends and what's going on on the political side. Did you ever interface with any other firms like Semi-Analysis? Because it seems like unique and Dillon's like AI is so big that Dillon's crossed over from talking specifically to to sell side analysts that even venture capitalists will read semi-analysis. No, I agree. I think it's pretty unique. Look, like there are obviously on one end, there are market research firms.

1:46:55There's like folks like IDC, Gartner and so on and so forth. There are many of them on the other end of the spectrum. So those would be like very industry focused. On the other hand, you have Wall Street sales side analysts. Morgan Stanley is called Nonsax of the world. I guess we kind of found a spot in between. And also, it's part of the team because we have over 20 analysts. And roughly speaking, you have half of people that have a financial background like myself, so closer to markets. And the other half is more engineers and people that are extremely technical. And so we kind of found that sweet spot.

1:47:27And also, what's pretty cool is when you look at the different people at Seminole, else is like all the guys like myself that have more financial background actually love geeking out and digging into technology stuff. And same goes on with the engineers. They also want to understand the business aspect and they end up geeking on the data and understand the insights and market shares and stuff like that. So, yeah, we all go towards the same goal, but have like a different set of skills and such. That makes sense. So take me through the latest piece. I can imagine other other players in the semi market research space just wait for you guys to publish.

1:48:00Okay, I think now we're ready to form an opinion. Now we can put a buy rating on it. Now we can slap a buy rating on that bad boy and sell it for 200K. Your favorite researcher. Okay. By the time we're done, we're going to get the venture capitalist plan solidified. It's a million dollars a month for any venture capitalist. And we like to say - Any platform fund. Yeah. If you're not on the semi-analysis, one million - You can't really be an AI investor. More of a tourist. Yeah. I got to say, we found something super funny a few days ago. One of the big brokers wrote a note to clients and basically said, yeah, these guys are the Bible.

1:48:32Actually, they talked about fabricated knowledge, which is Doc, so president of our firm. They said some people think fabricated knowledge is the Bible, but actually he works with the firm that's the actual Bible and all of this stuff. I don't know. I didn't say it. That's a Wall Street guy. That's amazing. There we go. Yeah, I'm a strong believer and I really enjoy the pieces every time they drop. Take me through the latest one. Meta Super Intelligence. The news this week was almost drowned out by the Windsurf Google cognition debacle. And so Zuck came out with this huge announcement, you know, Monday.

1:49:07And yeah, it almost got swept away a little bit. Yeah. Yeah, I mean, I guess Zuck is all in. That's another proof. It's interesting because I think some people were at some point doubting at the beginning of the year, is he going to carry all that investment, tens of billions of dollars? The answer is clearly yes, he wants to do it. It's interesting because when you look at Meta's CapEx so far, it has been heavily tilted towards what he calls core AI, which is basically recommendation models. It's a lot of inferencing, advertising models and all of that. So they said publicly in 23, 24 and 25, most of that CapEx, which is going to be like$70 billion in 2025, is for that core AI business.

1:49:50so the gen ai the llama stuff is still uh is still at the earlier stage uh but the big question is how how how bold is he gonna go with llama and i think what what we showed in this article hey there's a lot of uh evidence that he's doing it on a very large scale um and that's not just it's not just empty statements and saying i'm gonna build a five gigabyte data center in a few years it's actually things that are already built or under construction it's already uh committed capital And so that's the first story. There's already a substantial infrastructure build out specifically for Gen. AI. And I guess that sort of rationalizes the amount of money that's spent on researchers.

1:50:30Because if you think about it, yeah, you're spending like, hey, maybe$30 billion this year on Lama infrastructure. You're going to spend maybe, I don't know,$3 billion,$5 billion on hiring top researchers. Yeah, sure. Why not? Yeah, yeah, yeah. That's a completely reasonable strategy. And it always mathed out to us that even if these crazy researchers wind up working on core AI and the Lama project doesn't even go anywhere, it's like, you could probably squeeze$3 billion out of core AI, right? I don't know. It's just like the thought I had. Or you can also, like, these researchers can be focused on Gen AI, but they're going to develop, like, state-of-the-art technologies and using, like, massive compute.

1:51:08And then those state-of-the-art technologies can feed into the core business. And you're generating more advertising sales. And if you think about it, like Meta is growing double digits,$160 billion. So every time they grow double digits, we're talking about close to$20 billion of incremental revenue. It's big numbers, right? It's easy to justify just getting a few billion dollars on researchers. Yeah, yeah, it's great. I have one question about the history here with Meta. So there was this story, I think it came from the first Mark Zuckerberg interview with Dwarkesh Patel where he tells this story of the original Lama data center or the reason he had residual capacity was that he felt like he got caught flat-footed around tick-tock and reels and recommendation algorithms at scale for vertical social video where it's much more it's much less driven by a social graph and kind of like a traditional CPU based graph query and more about these actual recommendation algorithms.

1:52:10And the way he tells the story is like, we didn't want to get caught flat footed. We need to play catch up in reels, but we didn't get caught. We didn't want to get caught flat footed again. So I told the team build two big data centers. And then we had this kind of empty data center sitting there. And that gave us the initial compute for LLAMA. Is that too much of a simplification? Does that feel like what happened? Do you have any insight into kind of like how the LLAMA project, the initial compute was built out, and then I want to go into the future of the project. I would just say at a high level, if you think about the amount of money that Meta has historically been spending on data centers relative to what they maybe need in theory, like they have always been overspending.

1:52:53They have always been investing substantially in infrastructure. And same goes on for Google. I think Google to an even bigger extent. But yes, Meta has been a pioneer in building large scale data centers. Over a decade ago, Meta introduced their H-shaped data center design. They've been building 150 megawatt campuses since like 2013. So that's not new to them, like building large scale. And that's why they also already had this sizable compute footprint. But in 2023, like sizable was maybe 20 ,000 GPUs. That was pretty big. But today it's, yeah, it's updated. You want to be in the hundreds of thousands.

1:53:32And as of today, Meta is to some extent late in terms of training compute relative to others. Again, because they allocated, they have invested a lot of money, but a lot of that has been allocated to the core AI business. But what we've shown in that article is that they're actually today ready to ramp a massive data center in Ohio that's going to get them to get shopped. Talk about the H shape of the data center. Why was it an H shape to begin with? with, and then it sounded like they abandoned it in favor of just one big tent. What are the benefits of a tent? I want to know about data center shape broadly.

1:54:05Yeah, sounds good. Who doesn't love a good data center shape? Look, the H, I don't know why it's an H. What's more interesting is the structure of the building. If you look at one physically, it's absolutely massive. It's close to a million square feet. It's just a monster. If you look at the structure, it's also three levels. So it's a very complex structure um it generally what we've observed for satellite imagery is that it takes roughly two years to build uh which and i'm just talking about like first stone to actually getting the project built so two years is a lot uh many people do that in a year or less um so yeah substantial uh time to build um it was designed for very high efficiency um so they've been using a system with free air cooling like they can get the air from the outside uh there's no air to water heat exchange it's basically you get a coal there from the alta you just expel it hot air uh super efficient we can spray some water on it to make it even more efficient uh the energy efficiency ratio is typically called the pue uh maybe you've heard of that maybe not uh industry average is going to be 1.3 1.4 which means that for every watt you allocate to servers you have to spend another 30 or 40 of that power into cooling and power distribution losses and all of that that ratio for Meta was historically below 1.1.

1:55:27So they were actually the most energy efficient firm in the world running data centers. But the trade-off is that these data centers took a long time to build and had a very low power density. And so what happened is, first of all, at the end of 2022, Meta introduced the first massive design change. They completed through the old age and build, let's call it a more traditional data center design, single story, sort of a big rectangle, faster to build, maybe one year, maybe one or 15 quarters, something like that, much faster to build, more denser, better suited for AI. It could handle liquid cooling.

1:56:07But what Meta thought, look, I think this is where the XAI story actually takes a big role. I think what Elon demonstrated when he set up that cluster in 122 days, he just shocked basically data center infrastructure leaders all around the world my understanding is that he made everybody's life a lot harder because before it was like well if we do this really quickly I'm going to need a year, year and a half and suddenly it's like well there's a new standard. I can do it in 120 days yeah imagine like you're the infrastructure leader at the hyperskater that you have experience developing gigawatts of capacity and you think you're the best in the world at doing that and suddenly some guy comes out of nowhere and does it in like one quarter of the time.

1:56:51So I think many people were really shocked. And I guess Zuck like took it the other way and just was inspired by it. And basically that's where the tent steps in, which is let's make a data center in the shape that I can build the fastest so that, yeah, the only bottleneck is just finding some power. And that's it. And buying QP. We were saying earlier there was a bull market in tents. Zuck needs tents. The XAI engineers need tents. The engineers are intense in the office and then the XAI servers will be intense too. Everything's temporary. I mean, is there something about the tense structure that's potentially like, okay, this is going to deteriorate faster.

1:57:30Like what are the drawbacks of moving faster? And then on the PUE question, when I hear a one gigawatt or five gigawatts that they're targeting, is that total power into the building or total power to the actual servers? Honestly, people generally throw, like there are both. In this case, based on their analysis, it's to the servers. So it's actually going to be slightly more in terms of gross utility power. Okay, thanks. Oftentimes, you see people quoting total power just to have a bigger number. Sure, yeah, that makes sense. Industry standard practices to quote IT power. Anyway, in this case, you're going to have one gigawatt of compute power by the end of 2026 in Ohio and then close to two gigawatts by the end of 2027 in Louisiana.

1:58:14That's compute power to the servers. So anyway, massive compute for LAMA. Is that a vanity metric? Is it a vanity metric to hit one gigawatt exactly? I mean, Stargate, you have on your chart at 880 megawatts. Is 120 or 140 megawatts really going to be the difference between like an amazing super intelligence and the next best thing? It feels like we had to cross this threshold and one gigawatt is like, it's a good headline. Yeah, it's a good headline. That's more of it. Like, it's not going to change much if you have 900 or gigawatts. Yeah. But the more better, right? Like, you still want to have more servers.

1:58:48Of course, yeah, more is better. It does matter, to be clear. It's not going to change too much, but it does matter. Sure. Yeah, it probably matters for recruiting, too. It's like, you're going to be at the place with the best. What does the best mean? I can, you know,$100 million offers, nice round number. One gigawatt factory or super cluster, that's a nice round number. It's great. Look, and I've got a better one for you. $100 billion, right? $500 billion Stargate. Oh, yeah. That's something that's pretty funny to me because actually the Louisiana project, which is two gigawatts, they said publicly it's a$10 billion data center project.

1:59:20But if you count it the same way as they count the Stargate project in Abilene, it could be like$150 to$200 billion. It's just, yeah, just big numbers in marketing. Okay. I don't know if you have a bunch of insights or clarity here, but how are places like Louisiana and Ohio reacting in order to attract these types of data center projects? Are they promising the hyperscalers and the labs, you know, we're going to massively expedite permitting process? Is there like deregulation happening at a local level? What can you say? Yeah, I think the piece of context is that since, let's say, the end of 2023, or maybe mid-2023, you have a frantic search for power happening in the US and all around the world.

2:00:09And so we've made some numbers, we've aggregated some numbers. If you look at the pipeline, so there are different terms like data center interconnection load queue or pipeline. basically if you aggregate all of the load requests that potential data centers have submitted to the grid in the us you're above 500 gigawatts uh you're close to the the actual peak load of the us right so what's happening is pretty insane those numbers are mostly fake but what it means is that people are searching for power all around the country and you actually have a massive competition all around the country to attract those last products because if you think about it this way like okay 500 gigawatts of um of requests but in the end by 2030 you're gonna have maybe 100 gigawatts of growth which is an insane amount already but it means that just 20 percent of these projects are actually going to be real um so yeah there's definitely a lot of competition so people are doing uh everything they can to get those projects uh tax breaks are generally uh today uh the most standard thing uh accelerating permits like reassuring hyperscalers that you will deliver on time, that you have top contractors that you're going to expedite permits.

2:01:18Increasingly, there's being sort of enabling more on site power solutions as well, like being more open to people burning natural gas on site. All that kind of stuff helps companies, utilities and locations secure the big projects. I have a question about the shape of the super intelligence team at Meta. When you think about Meta properties, you think about Facebook, the blue app, you think about Instagram, you think about WhatsApp, and then Oculus and VR is like a separate thing, Quest. But I was toying with this idea that maybe the super intelligence team is more like the database team or the React team.

2:01:57And it's like an infrastructure layer, a project that will have benefits all over the place. But we won't necessarily expect a dedicated vertical that competes with Instagram. It's more about making all the apps better. I want I want to dive into the chat app statistics that you shared and try and understand it feels like it's not exactly a neck and neck race. ChatGPT is pretty much pulling away in terms of percentage share of queries at 71%. Meta is way behind at 12%. Are there any signals that this is something that would be addressed one way or another or is it just too soon to tell? Yeah, look, I think what we've seen so far is that generally speaking, when you start to deliver a better product, you just get more users.

2:02:48i think we've seen pretty good correlation between the quality of models and the usage of chat gpt one thing that i think is interesting is when you look at the user base of chat gpt you had a surge in early 2023 and then sort of plateaued for a bit and towards the end of 2024 you had a second leg of growth and then you hit like that half a billion uh weekly users and that correlates pretty well with new releases with models getting cheaper and better and so on and so forth. So if you think about how Meta could get back at a point and maybe lead that ranking, well, they just have to release better products.

2:03:21One thing is to have good products. And the other one is to have a good distribution platform. And obviously they have the distribution platform, right? 2 billion daily users. So they just need to build a good product. And I think users will come. And that's kind of the value proposition. What is the state of the mom and pop data center market? I don't think they would like to be called the mom and pop data center market, but you know the the you know a friend or an uncle that's getting into the data center business we asked uh brian one of the one of the co-founders of core weave the other day he was he was generally bearish you know he knows he knows how hard it is but a lot of people are like what is the is is there a real demand signal there or is it just a hope and a prayer that you're gonna just you know kind of flip it to a hyperscaler and is that is that yeah sorry guys it's over easy money's As we just said before, now there's really an insane amount of competition for power.

2:04:18And basically, hyperskaters can set the conditions. And so you have to be quite sophisticated when you want to approach hyperskaters and you want to sell them a gigawatt site. And you also have to think about how much money does that involve? Like one gigawatt, you're talking about maybe 30, 40 billion dollars of capex. so when you sell one gigawatt site to hyperscater it's not going to be a small decision for them to be clear like buying the land maybe i don't know a few million like who cares it's not a big deal but if they if they do make the purchase like they intend to do something about it um so like yeah they just don't want to take it lightly and they have multiple options today so uh for a mom and pop that sort of doesn't dig into what what they should actually do and all the requests and such they're just not going to get new business today uh but some some people made a lot of money for sure in 2023 and 2024 by just having like a lot they had 50 acres near a high voltage line that's amazing uh yeah it's funny to imagine you you've been building a data center you know a little mom and pop shop for the last year and then the death star data center just starts like popping up next to you it's so it's it's over uh what can you uh tell us about what what change between Llama 3 and Llama 4.

2:05:33And exactly what happened with Llama 4? You call it a failure. How did that happen? Because we were so scale-pilled, right? Scale is all you need. Just scale up the big transformer. But you dove into it and gave so much more detail. How can you contextualize and explain that to us? Simple was just a bunch of trade-offs that weren't weren't in the right direction uh like i think you if you want to simplify it you could say like to some extent you reach a peak pre-training in 2024 uh i'm simplifying i don't think it's big but let's call it for now it's big i think as black whale ramps and so on and so forth you'll you're gonna see a new push up but for now it's big um and it means that if you want to develop better models you don't have to use a pre-training you have to use the new paradigm which is a test and compute and reinforcement learning right and to do that there are some specific trade-offs that have to do and basically meta uh just took a bunch of options in terms of their attention mechanism in terms of the way they route to experts and stuff like that uh just a bunch of decisions that aren't very well suited towards this new paradigm which means that their flagship model behemoth the largest one is just not very well suited to this new era and that sort of contrasts with think of the chinese labs uh they are sort of the in the opposite direction because they don't have state of the art chips uh they don't really have an incentive to push very hard on pre-training And so they have been thinking harder about pushing on post training, enforcement learning, and test time computing, all of that.

2:07:03That doesn't require such a large centralized cluster. As such, it's kind of those two paths. Like on one end, you have meta. On the other end, you have the Chinese. And what the Chinese decided to do is just better suited to the current paradigm. So just a bunch of bad decisions or to some extent unlucky. And so that seems like that ties to this. There was this post by Rune Anonymous account on X saying that there are, in fact, some secrets about what paths of the tech tree you want to go down. You poach the right researcher. They come over and on day one, they can tell you that chunked attention might be wrong for this particular training run.

2:07:42Is that what you think is driving the high salaries? Is that the dynamic at play? Yeah. Yeah. You want the decision makers that understand exactly what tradeoffs are going to do. that also know how to properly evaluate things or what kind of steps you should take to make sure what is the right choice. And so, yeah, that's what you want. The decision makers that have experience, that know how to do stuff, and that can easily identify the trade-offs and know if I want to go in this direction, more reasoning, more RL, all that, you should do that attention mechanism. Do you have any insight into the shape of the behemoth project right now or like the failure mode?

2:08:23Like, is it like it would be bad at math or it would be bad at talking to it for a long time or it would be bad at needling a haystack in a big context window? Like, we've just heard like, it's not good enough, it failed. But like, what would that feel like if I were to use Behemoth for a long time? And I'm like, oh, this is weird. It's bad, but in a weird way because it seemed like it was good at some things, but then just not good at everything across the board. Yeah, let's just simplify it and say agents, so basically tasks that require using tools, that require reasoning, that require long context windows and stuff, it's not very good at that.

2:08:58Okay, got it. Jordy? Last question from me for now. What are you expecting out of Meta over the next six months? They have talent now, they have scale, but the new team's gonna need to gel and it's gonna take some time to really start delivering. So are you expecting a lot of public launches over the next, basically the back half of this year, or is this more early 2026? I think back half of the year makes sense. Really back half, like think end of Q4 or beginning of Q1 26. But for sure, you're going to have a few months where probably not much is going to happen. Because as of today, we showed a bunch of pictures of the current clusters.

2:09:43so they already have data center capacity uh but there's still some time in order to like actually put the gpus in place and make sure everything runs uh so they're gonna have some uh like sizable computes that's training ready uh somewhere in q3 um and so uh yeah but actually have a product release that's more of a really end of the year but most likely early 2026. i would not be surprised i wouldn't be surprised if new year's eve we're back on the show talking like we are now about a new drop. Can you talk me through some of the trade-offs or how the open source war is playing out? Dylan from Semi Analysis posted that the open AI model, open source model, is expected to be really, really good.

2:10:29That was kind of, I thought the open source strategy with LLAMA was a great way to be superlative on day one. It's like, it doesn't need to be the best model. It doesn't need to have the most DAUs or MAUs, but it's the best open source one. And so it's superlative. You get the headline. It's the attractor for talent. Hey, we're doing something different. We're the best in this one narrow thing. And there's a question about, at a certain point, does the math make sense to continue to open source? But then when we went through the DeepSeq moment, it felt like DeepSeq was very much distilled from the GPT-4 API, not a llama fork.

2:11:05And so the whole debate over, oh, like, you know, some Chinese lab is just going to fork Lama and then improve it. So what is your kind of state of the union on open source AI? The Chinese are eating open source. Like, yeah, they're just dominating the market. It's like one lab after another. It's not just DeepSeq. You had DeepSeq, then you had Alibaba. Recently you had Moonshot with the Kimi model. Like, they're just really good at open source. sorry, at LLM generally, and they're open sourcing everything because it's they're in some sort of the same position as Meta. Like they're not meeting.

2:11:39So they don't have any incentive to be closed source. They want to build an ecosystem. So it makes sense to go open source. And there's just like shipping faster than Meta and shipping better than Meta. So yeah, Meta is just way behind on open source. And actually the West is behind on open source. The Chinese are just way better at it right now. But is open source important or is it more just like marketing? Because I've always had this thing about the difference between like, what was it? Stable Diffusion was open source, Mid Journey was not. And Mid Journey was able to get the data back from the customer because you generate four images in Mid Journey, you give a thumbs up.

2:12:17To me, it's only open source if it's from the Mistral region. But but but talk about the flywheel of open source. Like, is there an advantage there or is it really just if you're not in first, you might as well open source? Yeah, I think it's more what you said is if you're not if you're not if you're not a leading lab, you might as well open source because you want people to sort of just help you. I'll give you some feedback to build an ecosystem. It just makes sense. also I would say like for the broad community generally it's good to have open source because it's better for adoption, anyone can sort of play with the models and develop new applications on top of it and such so yeah I think for everyone it's good that there's open source but again like if you're not Google, if you're not open AI at the very top like you don't really have an incentive to be yeah to be secretive about what you're doing because you're not the best anyway yeah that makes sense, Jordy do you have anything else?

2:13:11this is great, let's do it again Fantastic. Please hop on whenever you post anything. I'm sure we'll be giving you lots of calls because this is a fantastic conversation. Really appreciate it. Really enjoyed it. Sounds good. Cheers. We'll talk to you soon. Talk soon. Bye. Quickly, let me tell you about Fin.ai, the number one AI agent for customer service. The Bake Off champion. The Bake Off champion, number one in performance benchmarks, number one in competitive Bake Offs, number one ranking on G2. And we have our next guest in studio, in person. We have Jesse from Coinbase coming in. Yesterday, Jordy went over to Coinbase's launch event for BASE.

2:13:49Let's bring him in. Jesse, welcome to the stream. How are you doing? Welcome. The walk-in camera. The walk-in camera's working. There we go. Let's go. Got your juice, man. Thank you. How are you doing? Good to see you. Good to see you, too. We got one more juice for you. We're drinking BASE. We're drinking base juice today. We call it the base virality juice. How's that? Viral juice. Fantastic. Fantastic. How's it going? How's the show been so far? Super fun. It's been a crazy day on the internet. You guys are lucky you launched yesterday and not this morning. One Coldplay conference changed the timeline forever.

2:14:27I know I saw that this morning. I woke up. I was like, wow. I'm glad this happened yesterday. Oh, yeah. Because it's so hard to launch these days, right? We're having a couple of people from OpenAI on later today. And it's such a big day for legislation in the United States. Oh, yeah, the Genius Act. The Genius Act and the Clarity Act just passed. I'm finding out about this right now. Well, the president is going to sign it, I think, tomorrow. Yes, Genius App stablecoins will be passed and signed into law in the United States this year. I mean, this week, which is an incredible milestone. You're working towards it for three years, four years, five years.

2:14:58I mean, the impact is that for the last decade of crypto, there hasn't been regulatory clarity in the United States, which means that entrepreneurs and consumers haven't actually been able to benefit from this technology. And we've been working to build bipartisan consensus that we need rules of the road in order to make sure that crypto works. And I feel like Coinbase has always been like the conservative one. And it paid off because during the end of the Zurp era, basically all of Coinbase's competitors went out of business. It was chaos, right? And I remember talking to a public markets investor at that time, and he was just saying like, I feel like Coinbase has the mandate of heaven.

2:15:34They are making it through. What was the low? Was it like$7 billion? Something like that. Yeah, in 2022, the share price went to$32, which was below, I think, the Series E price. And I had joined a couple years before the Series E price, but we all experienced that. What was your path to Coinbase? Were you building something else before? Okay, break that down. I dropped out of school and started a company. It was called Clef. We did identity. And so the whole idea was how do we build an identity That's the next thing after passwords that was decentralized. So this was a crypto company. It wasn't quite crypto, but we worked with crypto companies.

2:16:07So, you know, Bitfinex, we were actually providing kind of auth and login for them. And so, Paolo, like, we go way back because we've been building forever. And so that business didn't work. And then when we were winding it down, I was basically trying to figure out what do I do next. And I love crypto. And his identity, like KYC, like I upload my, I take a picture of my photo. It was actually a passwordless login mechanism. So almost like you hold your phone up to log into WhatsApp. It was like that, but back in 2012. And so we were a little bit before the time. But when you look at the security architecture that we built, it's actually almost the same as what we're doing now on base and what people are doing with pass keys.

2:16:40And it's this incredible thing where I've now been working on it, this same problem space for almost 12 years. And we're feeling overnight. Finally, we're finally getting there. Okay. So you see you land at Coinbase. Yeah. And the way Brian Armstrong tells it, you go to him and you say, I need$1 billion. dollars to build some projects? Well, not right away. Not right away. That would've been bold. That would've been too crazy. But I want to get to the one billion dollar plan. Yeah, yeah, we'll get there. I know you did it for a lot less, but I want to know what... That's my question for the future, but tell me what you did in between.

2:17:08The best I can do is two pizzas. That's two pizzas. No. So I joined as an engineer. It was an acqui-hire process. Cool. My whole team actually sold to another company, to Twilio, but I was so excited about Coinbase. First go. I was like, let's go. And so I joined Coinbase, joined as an engineer, and then pretty quickly on, I just started leading teams. They asked me, hey, can you take on a team? I took on more, and for the next five years, I led all the consumer businesses on the engineering side. So if you know Max Brandsburg, he runs our consumer product. Me and him were product and engineering counterparts, with him running product and me running engineering.

2:17:38And then after five years of doing that, I thought I was gonna go start another company. I'm like a founder. I actually put in my notice and said, I'm leaving in six months. In those five years, what products were you building? I was building Coinbase, Coinbase Pro, and Coinbase Wallet. Okay, got it. So anything that you've touched as a Coinbase user, those are my teams. We rebuilt the whole product from scratch, migrated us to React Native. And Coinbase Wallet was the first decentralized product in the Coinbase ecosystem. So I think the context here is the core business and why the launch yesterday is exciting is the core business was always centralized.

2:18:11Centralized and custodial. And it was a gateway to get on-chain, custodial. And we'll get into base in a bit. but you were working on these sort of decentralized experiments early. Exactly. Yeah, working on the decentralized experiments early. But what are you saying it's so early at the time? You know, I'm an optimist. So I've gotten from you. He said he typed GM and posted it probably 100 ,000 times. But I thought I was going to leave, but I actually started having conversations with Brian about what did the future of Coinbase look like And what would it look like if we leaned even further into this kind of decentralized world, build more things on chain?

2:18:49Because Coinbase had started 10 years ago and it's, you know, traditional Web2 business. And so, again, the first thing I did was I went to Brian in the kind of fall of 2021 and our exec team and said, if you give me a billion people and 60 employees, this was like peak 2021. Billion dollars and 60? 60 employees. 60 employees. I'll turn Coinbase into a DAO. Okay. They were like, ha, ha, ha. Yeah. Okay. What would you have done? We're a public company. Yeah. Excuse me. I know what you do with a billion dollars if you're meta and you build a big data center and you have to pay half of that to Jensen Wong over at NVIDIA.

2:19:25But what would you have actually done with a billion dollars? I don't even know what is that. We didn't really have as good. Okay. It was more just like, let's take the constraints off. Let's take the constraints off. Let's go and do the thing. We have this big public company. It's a crypto leader, but it's still built on Web2 architecture. How do we move it on chain in this new way? And so it was the vision was there, but the execution strategy wasn't there. And so in the other context is that Coinbase has been in this incredible position of being the gateway to get on chain. And you see this with the Circle IPO.

2:19:55People were like, wait, Coinbase makes a lot of Circle's revenue as a stablecoin issuer. But you guys had always basically said everything on, you know, you guys do whatever you want on chain. And a bunch of great companies sprung up to kind of service that market. but you had to be hands off, which I imagine was pretty frustrating. Yeah, I mean, I will say that early on, so Brian has always had the vision of the self-custodial world. He's even had the vision of a super app in many ways, where the first version of Coinbase Wallet was actually called Toshi. It was a messaging product with apps in it.

2:20:27Oh, interesting. This is in 2017. It's green. And this is based today. It's based today. And so then we kind of pivoted Toshi to Coinbase Wallet, which is the self-custodial wallet that millions of people use, that folks know and love. and we've worked on that for the last five years. And that was primarily focused on trading in money. But as we've built base over the last two years, which has really been a builder ecosystem where people are building apps, we've seen two problems emerge. The first is that everyday people, when they're trying to use crypto and trying to kind of figure out what's going on here, it's really hard to actually find things to do and that are useful.

2:21:00You know, maybe they come on chain for a coin, but then they kind of get lost. They're like, what are all these things? How do I actually use it? And then on the other side, we saw that we had thousands and thousands of builders and creators who had these incredible products that they were building, but they couldn't actually get them in front of people. Like we were missing this connective tissue. And so 10 months ago, I talked with Brian, I kind of came back to running the Coinbase Wallet team, which I'd worked on for a long time prior. And we kind of jointly articulated this vision of, well, what if we built the kind of app that solved that problem and brought all of those things together?

2:21:35And this is what we launched yesterday. It's the base app. And it's an everything app that lets you create, earn, chat, trade, and discover this entire - Pay for things too. Pay for things with Shopify at Erewhon. You can now pay for juice with base USDC on Shopify powered by base pay. The whole thing kind of comes together in a new way where because we now can bring together this marketplace of developers and builders and businesses on one side with consumers through a really easy to use experience, we're going to be able to grow crypto a lot faster and spin the flywheel. Would I imagine launching base in the way that the product is today?

2:22:16I know it's a pure software business, so very different from the Coinbase's traditional business, which is effectively acting as this regulated financial institution. Would it have been possible to launch this a year ago? or did you need the White House to launch some tokens and a few other events? Well, I think the biggest thing that we needed is we needed the technology to mature, right? So when we had this vision three years ago of building base, the first thing we started with was actually the platform, right? It was the chain because we'd spent a year trying to figure out, okay, what's the product we build to kind of bring this next generation?

2:22:57But explain that. So you had the chain, but there was no token attached. How do you get people to use it? I feel like the playbook in crypto is like... We went back to the tried and true method that people had been following for 100 years, which is that you focus on building a product that people love. I think there's been this whole distortion in crypto for the last five years where people are like, oh, you get a token, you pay people to use your product. And we basically said from the beginning, no, we're going to build a chain. It's going to be a developer platform. And we're going to figure out how do we make it the best place for builders to build.

2:23:24And that's exactly what we did over the last two years. But if I take it to the Bitcoin analogy, it's like the reason that people set up data centers to run the Bitcoin network is because they mine Bitcoin and that has financial value. and so there's this economic incentive. Now that gets distorted all the time when people run random chains and stuff, and I understand the rationale, but is there an incentive for other participants to help run the infrastructure, or are you doing something where you're splitting it across different groups? How does the actual chain stay on? Yep, well the first thing to know is that base runs on Ethereum.

2:23:55And so it's a layer two that sits on top of Ethereum. So as long as the Ethereum miners are happy, then base runs. Exactly, and then there's a decentralized network that's also based that people use to get access to the network. And the thing that's bringing people to the network is that they're building things and then they're using those things. And so here's two examples. The first one is USDC. We're seeing a huge amount of growth in payments, both in the United States and outside the United States in USDC. And that's because people are using products like Shopify, which is rolled out to millions of merchants, where they're actually taking USDC in their wallet and they're paying for things.

2:24:25And that's global. It's fast. It's cheap. It happens instantly. And there's cool rewards. There's cool rewards. that it's retailers effectively, which is if you pay with USDC, we'll give you one point back or something like that because you're not eating the last debit card or credit card. Yeah, so it's 1 % back with base pay and you can earn 4.1 % on your USDC when you're holding it. That's pretty good, right? And it's an instant global payment method. So any business in the world can integrate it and then anyone else in the world can pay for things. So that's one example. People are coming to base for payments and stable coins.

2:24:57The other example is a little bit more out on the kind of like nascency curve in terms of what we're seeing, but we're really excited about. This is what the focus of the event was yesterday, which is really around the creator economy. If you look at the last 20 years, you've had so much creativity pour into the Internet. But the vast majority of that creativity has been put into platforms where creators actually don't earn that much money. If you're an average creator who has less than 10 ,000 followers or if you're in another country like Nigeria or Argentina, you're gonna be posting and filling up this kind of content world with valuable creativity that other people are looking at but you're not earning anything and so one of the big things that we're focusing on with the base app and one of the reasons why people are coming to use it is that when you post on base you actually earn and this is because we've built a new economic system where the content is actually valuable and then the value gets flowed back to the creators and I think this is a novel use case for this platform that we built with base and base chain.

2:25:52Okay, so question, if somebody opens a base app, their wallet's loaded and they start scrolling. Does that cost them money? Does that cost them money? How does that actually... It doesn't cost them money, but like I can literally show you, you can decide if you want to support a post. And so like here, this is on my feed and I'm just like Jesse, I'm on my feed, I'm creating content. I posted this great base juice content. And oh, there's a bug. Live, classic. There we go. It happens. Day one. Base juice content, right? And you have the normal things you see here, the likes, the comment, the retweet, and then you have a new thing, which is market cap.

2:26:27And so this is basically the value of that content. And this one's worth$15 ,000. This one's worth$84 ,000. How is that possible? I've been on YouTube for five years. The best video I ever posted got 8 million views, and I got a check for 20K from YouTube. Well, that's actually pretty good. YouTube is one of the platforms that pays creators really well. They took 50%, and I thought that was fine because they brought me all the users. So I was actually pretty happy with that. But how could a piece of content have market cap? No one else wants to buy it unless they want to buy the revenue stream.

2:26:59And there's been some of those projects where I could sell the rights to my video. I posted just a few hours ago. What's your payout on that? Yeah, how'd you do it? Startup CEO. Well, this is on X. Well, no market cap on this one. Startup CEOs can't even hug their chief people officer at a concert in this country anymore. 1.6 million views, 50 ,000 likes. 50 ,000 likes. I'm sure the creator payout will be decent, but this looks to me like potentially a nine-figure market cap if I had ripped it on base. Yeah, well, seriously. Like, seriously, I, you know, on these ones that I posted, like this one I posted yesterday that now has an$84 ,000 market cap, it's done like$4 million in volume, and I earned 1 % of all of the volume.

2:27:37And so that means I've earned like, what,$40 ,000? That's insane. Right? So you get a viral post? That doesn't feel like economically fair. Well, but think about this. The place where that value comes from is the fact that your content is valuable, right? How have these platforms built multi-hundred billion dollar businesses if the content isn't valuable? Running ads. But what's bringing you there? The content. And then they pay you to bring more content. An ad is a way of monetizing the value of the content. But the content itself is valuable. And what's happening on the base app now is that that value is being kind of figured out in a free market.

2:28:18And then the economic offloads of that value are getting redirected back to the creators. We love ads for this show. Will we be able to see ads in the base app? I know you guys acquired Spindle. I don't think you're ready to talk about it. Well, stay tuned in the next week. I want to run some ads. Talk about the integration in the kind of mini apps. Dan from Farcaster has been on the show. I know you guys have some integration with the Farcaster network, is that right? Yeah, so the whole social feed, all of the connection graphs, so you now have followers and following, it's all powered by Farcaster.

2:28:58When you're posting content, it's powered on Farcaster. When you're coining content, that's powered by Zora. It's another protocol. What we've done with the base app is we brought them all together. That's cool, including Shopify. Including Shopify, including XMTP, which is powering messaging. They just raised a$20 million Series B. And they're an incredible decentralized messaging protocol that is powering secure encrypted messaging with agents in chat. And so one of the really cool things, I don't know if you guys saw it in the demo yesterday, is you can be in a chat with your friends. You can just drop a bet in there.

2:29:28You can say to the agent, hey, let's bet$5 that the Dodgers are going to win tonight. And then it's an immediate bet. And then you can say, hey, send$10 to all my friends. And it just works. because it's built on this platform and you have these open protocols that are working together. And so going to your point, we built on Farcaster because we believe in building an open protocols. And one of the things that open protocols enables is it enables other people to build on top of them. And so this is where the mini apps come in is because we built in this open way. Now anyone can build a mini app and it just shows up in the app.

2:29:55And here's another one. I don't know if you guys have seen this, but this is when you open the base app today. Yes, I've seen. I think the Farcaster. It's still VPN. Which is like literally you can open this up and we're going to be on the live stream. Here, live now. There I am. Oh, wow. There you are right there. It is so cool. It's all about this. But one of the really cool things about this. And to be clear, this was organic. Third party. Third party. And you have some new sponsors. I don't know if you guys realize this, but you have a bunch of crypto brands here that have now gone and said, hey, we want to sponsor you.

2:30:23And Bracket, for instance, this is sports betting on base. It's awesome. It's so fun. You can interact with an AI jet. Our lawyers are like bloodhounds. They're going to be like chomping at the bit here. The thing that's really incredible about this app is that it's all connected. And so I can actually tap in bracket and I'm like, okay. Oh, okay. Yeah. Hey, stay one. Oh my God. Come on, Jesse. What's going on here with this mini app? Yeah, yeah, yeah. And in theory, yeah, this is the start of the auction driven. It happens. It happens. But this is the start of the auction driven ecosystem that actually results in real transfer value.

2:31:01Real value is created at some point. So this is a new ad. I'm like, ooh, I like Bracket. Whoa, now Bracket is right here. Okay. And then I'm like, okay, I'm just going to buy Bracket. And I'm going to be like, cool, let me buy$5 of Bracket. This is not financial advice. This seems like, I don't know. The thing that's fun is like, it's just wildly chaotic. It is chaotic. So it's an entirely new surface area. Coinbase has never been about chaos. And I'm not saying that base is. But creating an environment that's just like a free-for-all, decentralized, it's open. You can create anything. Yeah, I mean, you said that the goal broadly is to get a billion users on crypto.

2:31:37Crypto-native. On-chain. That's the right phrase. Give me the State of the Union. Where are we today on that journey? What are the different buckets and how big are they? Yeah. So we're trying to build a global economy that increases innovation, creativity, and freedom. And the metric that we use basically to measure that is do we have a billion people on-chain? Yes. And of course, I'm going to get 5 billion people on-chain because we're going to get everyone who has an internet connection on-chain. Let's go. But a billion is a nice number and we get to make it repeatable. I'd say that when you look at the data today, there are hundreds of millions of people who hold Bitcoin.

2:32:05Globally around the - Hundreds of millions? Yeah, hundreds of millions. A lot of people hold Bitcoin. Like Bitcoin is very well used, but it's not something that people are using their day to day life. I'd say there's probably tens of million people around the world who are using stable coins. And then I think we're in the millions of people right now who are really starting to use these on-chain products. And that's across Base, it's across Solana. We're seeing innovation everywhere. And we like to work with everyone. We think about Base as a bridge, not an island. We're actually connecting with everyone and figuring out how we can help them be successful.

2:32:31But I think it's still really, really early days. And the goal with the base app is to like kick up the next kink and growth. So my question is like when we're following the chat GPT story, there's obviously DAUs. But the chart that everyone's obsessed with right now is that chat GPT minutes per day is skyrocketing. It's at 30 minutes per user per day, something like that. We're seeing lots of queries go up. My question is like with financial products, I do think that there's a benefit to having a lot of people own Bitcoin. I think it's an awesome network, a check on authoritarianism. It's awesome.

2:33:02But I don't know that I should be interacting with Bitcoin daily. But what are the other metrics downstream to quantify whether or not someone's on a chain? Because if I'm in my bank account messing around with dollars all the time, I'm actually probably unhappy. Yeah. Well, the first thing I'd say is that definitely Coinbase is a traditional financial product. Totally. Where they're growing, they're trying to be the best for trading. It's awesome. They're making such good progress. So excited about the retail DEX integration, which is basically making it so that anyone in Coinbase can buy any asset on base.

2:33:34It just works, listed. I'd say base, though, we don't really think about it as a financial product. There's money involved because money is a part of everything, but it has social ads core, it has chat, it has apps. And so people are going to come to base just like they come to their other apps. Time on site should be a key metric that you're following. Time on site is the thing that we're looking at. We're looking at MAUs. We look at kind of two things we look at. Dow Mows, all that. Yeah, Dow Mows, and we look at like weekly active users, and then we look at weekly transacting users, where it's like, are you actually doing a transaction on the chain, where it's like buying something or sending money to a friend or something like that.

2:34:05And so the active user is kind of an engagement metric, and then the transacting user is like a deeper engagement metric. Yeah, yeah, yeah. So higher use of the actual product would not be a bare case like it is with my, you know, my traditional financial app, which I'm opening. It's probably a problem. I'm probably like, owe someone money or overdraft it or something like The really incredible thing about this kind of conversion rate of active to transacting is we have the data of before and after where like it was really just a money app before and now it's a real social product where people are doing all these things with money.

2:34:33And we're seeing way higher initial user conversion rate to actually doing something on chain. That's awesome. Because now it's not like, oh, you have to go like make an investment decision. It's like you can like your friend's post. Yeah. Right. Or you can like, you know, tap to pay at a store that you're going to. And that transition from a speculative thing, which is so important and such a big important and part of base in the crypto economy to a daily thing that also has like way bigger TAM in terms of the number of people who do social products and other things on a daily basis. That's really a shift we're trying to make.

2:35:02Yeah, it's super cool. Like we know that the everything apps work elsewhere. They don't work in America for some reason. This feels like an end run around, a different strategy to try and make it work here. But it's an everything app for a specific subculture, right? Right now, but you know, a billion people, is that a subculture? And what I'll say is we were just at Erewhon. I just spent two hours at Erewhon. And we were giving out free juice. I was filming content. We got some good content coming later. And I was just talking to people about the juice in the app. And when people hear, especially like millennial Gen Z creators, when they hear, do you want to join a free social network where you can get paid to post?

2:35:40And do you want a free juice? They're like, hell yeah. I've been posting on other social networks for a long time and I've never gotten paid. Okay. And so that hook of your content is valuable. You get paid. You get paid. And you can get paid for it because that value is going to get unleashed by the free market. Yeah. And then we're going to distribute it on crypto economic rails that is next generation internet platform. I'm fired up. It's epic. Well, good luck. We got to get to our next guest. Thank you so much for stopping by. This is fantastic. Can we get a gong hit on the way out? Here. Gong hit.

2:36:10Gong hit. Congratulations. Let's go.

2:36:15Get the Bass app. At Bass.app, you can get on the wait list from me letting more people on every single day. There we go. Founder mode. Talk to you soon, Jesse. Have a good one. Great chatting. Do it. Up next, we have Higgsfield. Alex from Higgsfield AI. We've been using Higgsfield AI to generate crazy photos. Jordy posted one of himself on a horse. The horse had a little bit of a long neck, but otherwise, photo real. and the facial reconstruction. It nails the face really, really well. I'm not exactly sure how they do that. I want to talk to it. Welcome to the stream. Yeah, Higgs Field. We have a lot of amazing people on today.

2:36:54I have been incredibly excited to speak with you. We were just talking. I don't know if you caught it, but I posted an image generated by Higgs Field a couple of days ago. I think most people still think it was real. It was a completely ridiculous image. It was like this cinematic upshot of me on a horse that had, I don't know why people didn't catch that the neck was quite a bit bigger than a regular horse's neck. It was great. Clearly a breakthrough. Yeah. It was the first time that I've seen an image generation product and realized that I think you guys have probably already started to take over Instagram-style content, but really disrupting the, potentially disrupting the Instagram boyfriend market.

2:37:43If people can just generate infinite images of themselves, Instagram boyfriends will have nothing to do. But great to have you on the show. Would be great to get a quick background on yourself and the company, and then we'll get into a bunch of other stuff. Great. Thank you very much for the kind words. Definitely, it's great to be here. Thank you for the invitation. So quickly about myself. I'm a veteran in the video generative AI space. I built maybe some of the most iconic products in that space. Maybe you remember Snap Face Filters. Oh, really? Which, yeah, totally. That's amazing. Yeah, there were like billion people throughout the world who played with that.

2:38:21And this was pre-gen AI, right? And when the models were like thousand times smaller than they are today. And although this product was specifically targeting like augmented reality use case, so basically overlay on top of the existing camera. With all these learnings which I got from my exciting times at Snapchat, we started Hicksfield with a way bigger ambition is to create camera of the next generation. Yep. Even today, we can see that there is an emergence of UGC content. The quality, unfortunately, goes down. And with Hicksfield, we create new ways to tell stories, helping creators and brands to get attention on social media.

2:39:19What's the key insight? Do you think how important is it to Create a great image generation model just scale you just need to throw a ton of compute versus Changes in the algorithm we've heard rumors that images in chat GPT isn't pure diffusion There's some transformer architecture in there some different layers I've started to suspect that there are different layers in some of these where there might be a different neural network or different system to put text on top so the text is really clean? Are we kind of recreating Photoshop at a certain level? Talk to me about the actual technical infrastructure to the degree that you can.

2:39:59Totally. First of all, I think it is important to admit that today we are early in our journey with video AI technology. I think Hicksfield is probably the first example of the technology platform which helps to create compelling content for social media the next step is going to be is to is going to be to build a reasoning engine on top of that so the so let's say you post a video the system is gonna suggest you and analyze your accounts and actually provides you with more suggestions what to post and eventually i think we we are going to find ourselves in two years in in a new world version of the world where most of the content out there is ai generator there is no way to stop that and um on on our sides we do our best to provide a simple interface so that non-ai users like let's think about market of social media professionals of tens of millions of people so that this broader user base can actually tap into the power of generative AI technology.

2:41:18And I think we will see that the models are going to get way better than they are today. It is true that various research labs, they do experiment with various architectures. We found that our post-training techniques allow us to substantially differentiate from the competition. And we do believe that the general quality of that technology is already there to surpass average human produced content. I have this one eval for AI image generators where I ask it to create a Where's Waldo? And no system's been able to crack it. And I think it has something to do with the density of information in a proper Where's Waldo, you'll typically see hundreds of little characters doing very intricate things.

2:42:10And so it's very clear that the artists that create the actual Where Waldo's work at a very small scale and they actually piece together the full image like it's a puzzle. That's something that I feel like could be solved with a reasoning layer on top. You understand that you're trying to create something that's really, really layered. And so you need to to kind of create tiles and then blend them together. But this gets into the question of like, how are we going to generalize and scale reinforcement learning in LLMs and agentic workflows? Is there a similar path that we're going down in terms of image generation?

2:42:48Yeah, this is a great question, by the way. I think this is a trillion dollar question, which we just asked, which is we all saw the power of reasoning engines with maybe models like 03. And then we see that at Grok 4, they actually spent on reinforcement learning more than they spent on the pre-training stage, right? Where is the market right now in terms of pre-training, post-training in images in your estimation? Yeah, I do believe that in the video AI space, we are relatively early. It's probably, we still see that the post-training stage in the core video AI models can be um maybe 20 50 times lower compared to the pre-training stage wow although yeah but we are really just crashing the surface there sure i think then the future is building the video reasoning engine and this is a trillion dollar question because this will um because think about the brands out there yeah um today it is um today what we are seeing is that brands start to and agencies they start to actually experiment with various models um and primarily we all rely on our stereotypical understanding of the customers and some maybe qualitative data which is available out there i strongly believe that with this video reasoning engine then the way how how the stories are told is going to be completely different instead of just running one video we can run hundreds of the videos out there and a b test them and see which one performs the best and today there is there are only a few top creators who actually do that if you look at come at mr beast and similar size of the creators they do a b test thumbnails very aggressively we all know that yeah they actually a b test the hooks so far this privilege is only available for larger teams who can actually do that, who have the manpower to do that.

2:44:53And the next generation video reasoning engine will empower everyone to do that, which is going to the boom of AI-generated content. What you're talking about is basically RLing on humanity with human graders, which is the algorithm and likes. And that's effectively what the Mr. Beast algorithm is doing. My question is like, is there a way that we can bring that into the data center? Because if it stays on Instagram, if it stays on YouTube, it's probably pretty rough for you because Google and Meta are going to have an advantage there. But if you can figure out how to do RL with verifiable rewards or something that looks like a rubric for grading, you know, and finding errors, are we going back to the generative adversarial network era where you'll have two competing models to determine like whenever i generated something with vo it's always like cars driving looks amazing then all of a sudden i'm looking at the front of the car and the car's driving backwards and clearly the model is getting confused but we need maybe like a detector for that but how do how are we actually going to do rl at scale in imagery this is a great question so i think the first step is exactly i mean the first step is going to be reinforcement learning with ai feedback part of that we already do at hicksfield obviously at the post-training stage not yet at the inference stage just because of the cost associated with that.

2:46:12That makes sense. So we have to train video generation model in a in a way that it's sort of competing with video understanding model. And like at Higgs fields, we could we we cannot go and label millions of the videos. That's why we have a set of powerful agents for video understanding, which help to tune the video generation process. Yep. But this is the process in the vacuum itself what's going to be powerful is when we condition the outputs of the model and train the model based on the engagement data from the social media first is going to be a number of likes a number of comments although if we look at meta ads we see that they provide a very detailed breakdown and drop off second by seconds so training the models based on on these outputs is going to lead to complete the next level of reasoning and success rates for the end customers.

2:47:12Cool. We have another one. I know we're totally over. What happens to the legacy Instagram influencer that has built a business basically on Envy? They're constantly traveling around the world at the best hotels, on boats and private jets. What happens when anybody can be on a private jet, on a platform like Instagram or on a yacht somewhere? Have you thought through kind of the implications for the tech across different categories of content creators? Hopefully. So first of all, I need to admit that we are a technology platform first and foremost. Although we think about ourselves as a scientist and we are constantly monitoring and listening to the creators and how they use the technology.

2:48:03and i cannot bring up the names out here although like some of the top 50 youtubers in the world they actually want to get reads of the team and build own agency of ai influencers to be honest that's they they actually they have a bunch of ideas which they want to sell and they don't want to condition their existence on the social media just to their likeness today because people just are getting older and sometimes they get irrelevant. We have seen many examples of that on social media. And creators actually want to create those digital agencies where they can express all their ideas through various synthetic AI influencers.

2:48:46And they can use Hixel platform to do that. Yeah, it's a very, very wild time. Let's have you back on again soon. This is fascinating. Yeah, we'll talk to you soon. Thanks so much for hopping on, Alex. Thank you. We'll talk to you soon. Let me quickly tell you about public.com. Investing for those who take it seriously. They have multi-asset investing, industry-leading yields, and they're trusted by millions. We have Billy from Regent also calling in from Reindustrialized. Sorry to keep you waiting, Billy. Great to see you. Every time we talk to Billy, he's in a far-flung part of the world. We're working to bring him in from the waiting room.

2:49:22And he is sideways. Can you turn your camera? Yeah, I can go vertical. That is possible. Here we go. There we go. Where are you? You're in a cabin? Explain. Join you from a sea glider. Look at that. No way. It's super comfortable in here. We got plenty of leg room. There we go. That's incredible. We got internet. This is the future. This is amazing. Amazing. Wait, wait. So wait, are you, is this like a demo? Are you actually in - I do believe that's a screen behind him. It's a screen behind you? Okay. I can't believe it. This is the demo. You guys are pretty perceptive over that. Okay. Well, we just talked to the AI generated guy who says all this will be fake in two years.

2:49:57So, you know, I don't know what to believe at this point. We're building hardware in the real world, guys. We're at a re-industrialized conference right now. Thank you for bringing me back to the real world. How is re-industrialized? Give us the overall update on Regent. It is awesome. Like hardware is cool again. Like building real stuff is cool. Atoms are cool. So it's awesome to be here for this. Yesterday, we just announced the launch of Regent Defense here at re-industrialized, which is a big step for the business. Congratulations. What is what are the what's the immediate application of the glider in a defense context?

2:50:34So regions actually been doing defense for years. This is sort of a growth of the business we've already been doing. Basically, everything is about moving around island chains as our key conflicts and theaters are in the Indo-Pacific. So it's going back to World War Two style tactics. It's about naval operations. It's about being able to move around Island Chain's logistics from the head of the Marine Corps and throughout the services underwrites the success of that naval campaign. So it turns out that all the things that our commercial customers like about sea gliders, the high speed, low operating costs, ease of operation and in the DOD space that we're hard to see.

2:51:10We fly really low over the water, so it's hard to pick up on some long range radar systems. It makes it a really perfect fit. So we've been on contract with the Marine Corps for a couple of years now. And now we're expanding as we get into our full scale prototyping, expanding that product portfolio to meet the need. Very cool. Incredible. Anything else? I think that's it. Thank you for this. Thank you for this demo. This is the best guest call-in I think we've had. Guys, we've got to get you on the Sea Glider in Rhode Island at some point, too. We're doing the next call-in. We have real hardware.

2:51:42We'll do the whole show. We'll do the whole show from the glider. I hear it's very smooth. You know, it's perfect for live podcasting. Starlink on board yet? We'll put Starlink on board for a TVPN episode. There we go. I love it. It's there. It's there. Fantastic. Well, congratulations to you and the team on the progress and say hi to everybody at Reindustrialize for us. Will do. Thanks, guys. Enjoy the rest of the conference. We'll talk to you soon. And in the meantime, let me tell you about 8sleep.com. Get a pod five, five-year warranty, 30 night risk free trial free returns free shipping go to it's code.com tbpn and we have our next guest joining the stream i'm going to guess that it's jonathan but i'm gonna make sure once he joins welcome to the stream how are you doing did i get that right yes jonathan what's happening security sorry we call me jmo actually so jmo what's up uh give us the news i hear we uh You got some good news.

2:52:43Break it down for us. Yes. We are coming out of stealth, launching a new company called Confident Security. It's about providing confidential AI. Did you raise any money? We raised$4.2 million with Decibel. Congratulations. Of our comments. Congratulations. There we go. Why is confidential AI important? I can imagine a bunch of reasons, but what was the kind of catalyst to start the company? Yeah, I think, you know, unless we do some work, I think we're going to have Cambridge Analytica times like a million, essentially, unless there's just too much incentive for folks to train on data. And people need to care about privacy.

2:53:25So we thought, you know, Apple shouldn't just be the only ones providing privacy. Everyone else should do it, too. Give me some concrete examples of how to use the product. What data specifically am I keeping secure? Is it stuff that's on the web? or because I feel like if I have an air gap data center somewhere with a whole bunch of hard drives, there's no crawlers that are getting to that. That's right. If you are taking that air, if you're using the air gap data and you have your own GPUs and it's completely in there, you might care about various users inside your air gap environment, not seeing all the data, you know, a standard like top secret versus secret type of stuff.

2:54:02But this is, you know, you're an employee at Pfizer and you upload a PDF to OpenAI that describes how you do drug discovery, maybe you care about that secret. Maybe you care about how it's going to be used. And after the recent OpenAI core case where they were forced to retain deleted data, I think people are being a little cautious. So how do you actually plan in plugging in to companies? How does the product actually get installed and used on a day-to-day basis? So it requires two parties. One party who's running the server to install our wrapper around it. and then the other party who's like making requests to that server to use our SDK.

2:54:40And essentially you can think of it as like a very specialized form of encryption where you can only decrypt the data that you've submitted to the server if you've met some constraints, like you don't log the data, you don't train on the data, no one has access to the data, all that type of stuff. And is that just like enforced at an engineering level or enforced at a legal contractual level? It is a technical level. Great question. Okay. And that's what's different, right? You can make promises that say you want to trade on the data, but we make, it's a technical guarantee. And it's essentially a bunch of fancy cryptography that makes that guarantee.

2:55:15So we actually offer unlimited liability and indemnification for data breaches and misuse. Because we're so convinced that you cannot, we cannot see the data. No third party can see the data. Yeah. So what's the go-to-market like? You mentioned an example of like a big biotech company. Is that the most obvious customer or are there other segments that are logical? Biotech, legal, finance, of course, defense and government. And not just like pure defense, but local jurisdictions, states. They're all trying to figure out how to deal with freedom of information. When have you made something public?

2:55:55Was it too soon? This helps manage all that stuff. privilege for lawyers same problem they're trying to figure out well if i give my data open air i have i disclosed it and it's no longer subject to privilege yeah we make that sense uh super interesting close close it out with uh how big is the team where's the team size going what are you hiring for what comes next um we're about six people right now and we've been building building building building and now with the launch we're ready to start selling and so our focus is bring on salespeople. I've done two previous companies and every time I've learned that I should spend more on sales.

2:56:32So that's what we're doing. Interesting, good takeaway. That's good. That's the environment that a great sales leader or individual contributor wants to join. I mean, it takes a lot of technical CEOs like a while to actually get through that. So it makes sense that you're in your third company because like, yeah, a lot of people learn that lesson late, right? Awesome. Jaymo, thank you for joining. Thank you so much for joining. Come back on when you have news, and thank you for doing this. We will talk to you soon. Good to meet you, guys. Cheers. Let me tell you about Wander. Find your happy place.

2:57:06Find 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 next, we have the founder of Contextual AI coming in to the studio. How are you doing? Let's bring him in. What's happening? Also, if I have this correct, the inventor of RAG, correct? One of the authors of the RAG paper. Okay, okay. Yeah, that was a team effort, lots of folks, and it's a long history of research that has gone into that. Yeah, give me the State of the Union on… Very, very humble. No, the man who invented RAG by himself in a cave with scraps.

2:57:47Yeah. Give me the State of the Union on RAG. some people are saying, oh, just use a bigger context window. What are companies actually using RAG for on a day-to-day basis? What's the state? And then what's the shape of the industry that's popped up around the technology? Yeah, so RAG is really a very simple idea, right? It's about having Gen AI work on your data. And you do that through retrieval. That's the R. And you then use your retrieval results to augment. That's the A, your generative AI. That's the G. So it's a very simple idea. How people do RAG right now is radically different from what we did in the paper originally.

2:58:27I think the buzzword these days is about context engineering. How do you actually give language models the right context so that they can do their job? And as it turns out, all the language models are pretty good these days. And there isn't that much of a difference between your CLOG and your OpenAI models or Gemini. If you give it the right context, then it can solve the problem. If you don't give it the right context, then you can have an amazing language model, but it's going to fail. And so that, I think, is an opportunity that a lot of companies are looking at now. How can we make sure that this context layer really works?

2:59:00Do you feel like there's a rag step or layer in the typical deep research products that I'm using on a day-to-day basis? I feel like it's mostly like going out to the web, searching, and then kind of just summing all this stuff up. but I can't tell if it's actually like the, the base of the rag is like embed all of that into weights and, and then, and, and then search over it. But is that happening at, at the state of the art right now? I think so. You're right. A lot of it is web search. Yeah. And if you want to do web search efficiently at scale, then you probably use simpler algorithms. So things that aren't that involved, but even web search very often uses embeddings and where do you search there?

2:59:45So you could argue that web search is also just interesting. Why is why is search so bad right now? I feel like I feel like I can't search my email for anything. And Google has like trying to find something in iMessage. But yeah, I feel like search has just become like really hard. But then at the same time, I'm like having my mind blown by LLMs and deep research products. But I don't want to wait 15 minutes just to search my email. But maybe that's what I need to do. Maybe that's the future looks like. Like what is going on there? Yeah, I mean, that's a hard problem, right? But I mean, AI should be able to search your inbox for you and just give the right answer.

3:00:22That's the goal. I think it's happening. It's just search is a really hard problem to do. You don't want to really do it in a single step. So the way you do proper retrieval is multi-stage, sort of cascading with smarter and smarter models that look at what might be a relevant result. But you're right. If you retrieve the wrong things, then you can never give the right answer. That's a big open problem. Where are you guys focused today? I know there's a number of different use cases. Yeah, so the use cases we're looking at are really your bread and butter use cases for RAG. So answering complex questions on complex documents.

3:01:03In our case, we scale to millions of documents, which is unusual. So one of the most common misconceptions about RAG is that people think that it's easy, which is actually probably true if you have like two or three documents, you understand the use case, it's not that complicated. But when you go to the real world and you talk to some of our enterprise customers, they have very difficult problems. The data is all over the place. It's very complex data. There are millions of documents that it needs to work on top of and then search breaks down. So you can give the right answer even if you have a great language model.

3:01:36Yeah, I mean, the scale of data at some enterprises is probably, what, seven orders? I mean, I imagine like the number of emails sitting in Gmail inboxes across the entire network is way beyond millions. That is also to your earlier point about like long context models. You can't fit all of that in the context of a language model. You need to do three role. No way. What's the state of the business today? How big are you? Like kind of what are the new challenges? What are the next milestones? Yeah, so we're 70, 80 people working on a lot of interesting problems, kind of across the board, trying to expand into different use cases as well.

3:02:18So beyond the traditional ragged use cases, looking at things like root cause analysis, code gen, because code gen is very hard and also often requires technical documentation that you don't want to incorporate in your code gen. So yeah, making a lot of good progress on very interesting problems. Very cool. Well, thank you so much for stopping by. Love to have you back for a longer conversation. Yeah, let's do it again soon. Hope you have a great rest of your day. We'll talk to you soon. All right. Thanks for having me. Bye. Great to meet you. Let me tell you about Bezel. Go to getbezel.com. Your Bezel concierge is available now to source you any watch on the planet.

3:02:52If you like enterprise agentic workflows, you might like fine watches. And up next we have OpenAI. Big launch today. We are going to break it all down. Welcome to the stream. How are you? I think we got caught up on guests. We have two people. Well, it's happening. Great to have you guys on the show. Hi. Nice to meet you. I will start by saying sorry that Coldplay had to have a conference last night. It was hard to predict. It's never been harder to launch anything on the internet. I know. We were just talking about this. Yeah. Well, we're here. No idea. He says it should. Okay. Well, we're resetting the narrative.

3:03:33You don't have to plan your launch schedule based on Coldplay. Yeah. Coldplay is out. Maybe it's something to keep in mind. Break down the launch. What is in? What should we be focused on now? So, I used to work on Deep Research, Yash used to work on Operator. We both had our launches earlier this year and I think after our launches we realized that our products are very complementary. And so we've basically combined the best of both into this new product, ChatGPT Agent. So ChatGPT Agent has access to a virtual computer with a bunch of different tools installed. So it has text browser, visual browser, terminal, and it's able to do a lot of different things that you would do on a computer.

3:04:10So it's just like very flexible and pretty powerful model. We trained it using n-to-n reinforcement learning, like our past reasoning models. And yeah, it can make slides, make spreadsheets. Yeah, what initial use cases are you guys most excited about? Have you been using internally? What kind of companies should be and just individuals should be kind of taking advantage of it as of today? Yeah, I think so deep research, as I was saying, we combined sort of deep research and operator to build this deep research was really good at research, like I think the best product out there in terms of or at least that's what I thought.

3:04:46I still think, in terms of researching. And then operator was there to take actions. Combining these two, you open up a lot of possibilities. You can do research, you can do actions, you can do research and then actions. And then on top of that, we added APIs. So, like, for example, if you have connectors, which we launched, I think, a couple of months ago, you can connect Gmail, Google Drive, Linear and whatnot, all sorts of products. And combined with that, it becomes an extremely powerful research and action tool. So, for example, personally speaking, I've been using it a lot internally just to even talk with the code base.

3:05:22Like, I'm solving a particular problem. I'll connect it with GitHub. I'll ask it like, hey, can you sort of go figure out what's happening in this code, which is, let's say, a new code base for me. And then the model is also very natively multi-turned, which means that I can just have conversations back and forth with it, which is not true necessarily, wasn't true, for example, for a deep research model, which released earlier this year. So it's really, really useful from a specifically from at work when I'm able to connect all these amazing tools and able to just understand what's happening all parts of the decision.

3:05:53So secondly, oh, sorry, I was going to quickly, I was going to say, like, personally, I also use it a lot. I think there is a lot of small things or big things that I have to do. I can do them myself. Give an example. My wife and I have a date night every Thursday. I forget to book it most of the time. And then I get in trouble. and then now with agent like you can schedule tests etc you can just say like look every thursday just go ahead and figure it out give me five recommendations in the morning which are available and i can just do it show up on thursday morning just click it and it's done so things like that talk to me once deep research came out it felt like uh there was this a little bit of a meme about like we have 15 minute AGI.

3:06:37And I'm trying to understand, I could imagine this new product stretching that out to let it run if it's building a spreadsheet and scraping data from different sources and putting a whole bunch of different things together. I could imagine letting it run for like an hour and coming back. But you said it's multi-turn. So what does the typical interaction look like? Is there a wider variance? I noticed in the latest revision of the ChatGPT app, There's now a little like 15 minute UI element next to deep research to kind of hint that, hey, you're getting yourself into a 15 minute cycle. Obviously, there's lots of efforts to speed all of those processes up and that'll come.

3:07:15But what is the typical interaction time look like? Is this more asynchronous or synchronous or kind of you can do either? Like, how do you think about those tradeoffs? Yeah, I think our team in particular is really focused on solving harder and harder tasks that take people more and more time. So a lot of the agent tasks can take anywhere from like around five minutes to, I've seen it take over an hour. But a lot of times these are tasks that would have taken humans many, many hours. So I think like as our agents get better, probably the length of time they'll take to solve tasks will also get longer because just the task will become so much more complex.

3:07:59Like imagine a task that takes a human like many days. Yeah. Yeah. So maybe so. So a question I have is right now the agent can browse the web for me, do research, take action. I imagine the next step would be an agent like some like a voice extension to it, where as an example, I might say, hey, I use Geico right now. I want to potentially switch. Can you go out and do research? Here's the cards that I have. Try to find a cheaper price. And then, or even call Geico and negotiate a lower rate or, you know, you're traveling, let's say, and your flight - I need to cancel an old internet line on Spectrum right now.

3:08:38And they only allow me to call during weekdays and I'm busy. Yeah. The other example, you're on a vacation and you want to change your flight and you know, you're going to maybe have to like call and sit on a wait. So is that the direction that you think we're going towards where it can not just browse the web, but then actually start to, it's funny to think about because the voice agent would then just call and maybe it's talking to another agent on the other side. But yeah, where are we going? Yeah, voice is definitely an interesting form factor. I think the way to think about where we're going is twofold.

3:09:12The first is what Isa just mentioned. I think we want to continue to solve harder and harder and longer and longer tasks. I think today we can solve, let's say an hour or so of tasks. Hopefully in future we can solve multi-day tasks and it might take longer, it might take shorter depending on how we're doing, but continuing to improve on the reliability and complexity of tasks that we can continue to solve. So that's sort of a core part of what we want to continue to improve on. Second part, But the Geico example, for example, the one you gave, technically you can do it today, not with voice, obviously.

3:09:45You can just type it in an agent and it'll be able to essentially answer the query. You can log in with the virtual browser that we have on agent with your whatever internet provider you use. And I think it'll be able to tell you things about what's happening there, what's not happening, what other alternatives might there be, and all of those things. but then at the same time you bring up a really good point which is like voice might be a very natural sort of way to do this in future and that's a form factor yeah and other other other companies introduce voice as an intentional point of friction like this sort of call to cancel spectrum would never do that to me yeah they would never do that to you John and so that that to me feels like this met you know if I can go into a web app just click cancel or do something like that yeah that my question is about like the user experience of like having something that could take five minutes or an hour how predictable is that i've gotten in a great pattern where i expect deep researches to take 15 minutes and so i know when to go to deep research and i'm going to come back later and it's great but if there's some variability there is it going to send me a push notification when that's done is that how that works like how do you train the user to get the best experience yeah i will send you a notification i think actually the fact that deep research always takes the same length of time is probably um i don't want to say a bug but i think of it as a feature not the end state i think yeah i should think for as long as it needs to think totally but i think for deep research it always just thinks for a really long time even if you ask what the weather is yeah that's true so i think there's a better middle state and i think that this model is like a step towards that but i think it still will think for too long on like really simple queries yeah yeah you're totally selling deep research short.

3:11:29Sometimes I ask you what the weather is and I want the history of weather from start to finish and what it is tomorrow and yesterday. And I want the history of meteorology and how the Doppler 3000 works. I want everything. And I love that about deep research. It's my favorite. The use case that I'm sure that will immediately start happening that is pretty hilarious to think about is a student that just says, hey, these are the three websites that host the homework that I have to do. Proactively go to the website, figure out the homework, Yeah, but the teacher's going to be like, check this homework.

3:12:05If somebody wants to say that's not AGI, I don't know what to tell them. Yeah, I don't know. What about other tool use? Jordy mentioned phone usage. You mentioned spreadsheet integration. What's kind of further down the stack of integrations that you've already announced that might be kind of underexplored or underappreciated at this point in time?

3:12:32To me, I think that the tool that we've given the agent is very general and powerful. Like you can almost do anything that you need to do on a computer with this tool because it's browser and terminal, which you can do most things that might not be the most readable to a human. So I think that now it's about pushing the capabilities. Like you can ask it to do anything in theory. It's just the agent won't be good enough to do everything you ask it to do. So I think that we just need to make it better and better using the tool it has. Yeah, I think the frontier continues. I think, as Isa said, the tool is extremely general.

3:13:06It has access to a browser. It obviously has access to a terminal. And we can give it access to as many APIs as possible. That should allow you to build whatever you want to do, generally speaking. Like, for example, you can totally imagine in the future there's access to a voice API or whatnot, whether it's internal or external, depending on how things go. Right? Like, you can have access to everything. You can build everything. But we still need to push. Like, it's still early. Like, we have not solved everything. It's still early. We still want to make sure that we can solve these cases with really, really high reliability.

3:13:39And that continues to be a pretty large focus. Yeah. Well, I'm excited. I'm excited. I mean, think about a world where you can give it access to your password manager and things like that, that it just immediately can act. Or just API integrations, right? So then the passwords don't even need to pass back and forth. That makes a ton of sense. Yeah, I'm excited for, I feel like deep research maybe doesn't have access to images in ChatGPT yet, but I could imagine those being way, like the reports being way richer if you can define them. And then sometimes when I'm just generating like a general chart, I actually want to use like a visualization library in Python and kind of going back and forth.

3:14:15So very cool to see it all kind of come together and very excited for where this is going. What's the rollout strategy? When can people actually start using this stuff? People, everyone on ProPlan should be able to use it by end of day to day. and we'll get it to plus users over the coming days and then enterprise over the coming weeks. Very exciting. All right, well, congratulations on the launch. Super exciting. We're going to turn this day around. It's now just about opening it. Ignore all Coldplay memes going forward. Well, thank you guys for joining. Thank you so much for having us. We'll talk to you soon.

3:14:56Cheers. Bye. Up next, we have Dan Shipper. friend of the show over at every who got early access and he's been using agents and we're going to get some feedback from his review. John just spilled his base juice all over the table, all over the FT. That's that's really disappointing. You're not going to be able to read that on the way home. Papers that I can do we have Dan in the waiting room?

3:15:22There he is. Nice. So guys, great to see you every time you're on. You finally caught me when I'm not traveling. Yeah, yeah, yeah. This is the first time. I like that light up logo in the background. Very nice. Very nice. Thank you. Subtle kind of in-home, out-of-home advertisement. That's what we're going for exactly. What's happening? What's on your mind today? There's a lot going on. There's a lot going on. We're here to do a vibe check. Vibe check. Of ChatGPT agent. so I was lucky enough to get that hang out with it and work with it for the last couple days before it got launched and I have a bunch of things to tell you about how it works incredible so as your previous guests who are amazing told you it's sort of like deep research and operator had a baby and it does some really cool things so the first one of the first things that I had to do is I had it go through all of our support emails and all of our feedback forum posts for the last like two months.

3:16:26So it's like about 1 ,500 support emails and maybe like 500 posts on our forum to gather for Quora, which is our email management AI app, to gather all of the customer archetypes of like, okay, who's posting, who's a promoter, and then going and looking on their LinkedIn to be like, what's their job? You know, where do they go to school? all that kind of stuff and put together a like long research report of who our promoters are, what the archetypes are and who are, who our detractors are and why they don't like us. So that's the kind of task that like, obviously like I could have done or someone on the team could have done, but it would take it so long.

3:17:04Yeah. It takes a long time. Um, and it's the kind of thing that you almost want on like a recurring schedule. Like you just kind of want to see like once a month, but no one wants to do that once a month and you can schedule with chat to BT agent, you can schedule it to run. So I can just say like every month I want you to just send me or on the first of the month, which it's really freaking cool. That's amazing. I wonder how compute intensive that's going to be because if I know anything about building dashboards and building like these reports, there's always like an intense amount of like, oh, we got to have this dashboard.

3:17:38And then you check the analytics and it's like, oh, it turns out the team just said that for a week and then like stopped watching it and if it's running you're like burning dan is every foundation lab's worst nightmare because he gets on the most expensive plan and uses it a hundred times more than anyone else you're single-handedly going to bankrupt the lab there's other there's other users that are probably higher margin um but but yeah talk to me about that that actual experience did you have to oauth with any different services did you have to share any api keys did you have to export any data or was it really as simple as just a prompt?

3:18:11It's basically a prompt. What happens is you type in a prompt. You say, I want you to check out Quora. I want you to check out our emails. I want you to check out our support forum. So Chat2P has connectors. So I had previously already connected my Gmail. So you just log in on the OA. And then what it will do is it spins up its own computer on the cloud, its own virtual machine. It goes in the browser and starts browsing the web. It also then connects to Gmail. If it hits a login, so for example, when it hit LinkedIn, it couldn't log in. And you can take over the browser in the virtual machine and type your password in, which is a little janky.

3:18:50But it works pretty well. I think the interesting thing about this, though, is that it seems like there's two main approaches to agents. And OpenAI and Anthropic are taking very different paths. And they have very different trade-offs. So the really cool thing about ChetGPT Agent is they're essentially abstracting away the browser and the computer. So all you're doing is you're interacting with ChetGPT. And on the back end, all this other stuff is happening. So it doesn't matter if you're on your phone, if you're on a computer or whatever. They have this whole virtual environment set up. It spins up, it does the task, and it spins down.

3:19:26So it's a very good consumer experience. um cloud code for example from anthropic which is i think cloud code cloud code is way more for developers chat dbt agent is way more i think for consumers cloud code is all on your computer it's all in the terminal and and you have access it has access to all of your files and you have the ability to um use it wherever and whatever and however you want so it's much more customizable much more composable um so i find that cloud code is much more powerful but it's much more intimidating and it's just not something that a consumer can use. I think that Chessy RQP agent...

3:20:04The crazy thing there is doesn't Cloud Code have more downloads right now than the actual Cloud Mobile app? Like the... It's something crazy like that. I honestly think people are sleeping on Cloud Code. I use it all the time for non-programming tasks and I think most people think they can't use it because it's in the terminal and the terminal is really intimidating but it's an incredible product. So yeah, how would you solve this problem if you were to do the same eval of like generate your net detractors, net promoters using cloud code? You know, just open up the terminal on your laptop. You wouldn't be able to do it on your phone, but you'd just engineer a prompt that told it to do that.

3:20:44And it would just write all the code that it needed to do exactly the same thing. Do you think it could hit that? Do you think it could do it? Yeah, it could do that for sure. And I think the nice thing about Cloud Code is you get many bytes at the Apple and you can like, so for example, with Cloud Code, what you can do is you can have it make a full plan. So it can output like a full markdown document with like, you know, 300 or 500 or 1 ,000 word plan. You can modify it and go back and forth with it and then have it execute it. I think it would be more complicated. like yeah it would probably write some code to hit the gmail api and i'd have to like think about that as opposed to just like clicking the connectors button or it does have a it does have a web research tool so it would be able to go to like our feedback forum and do all that stuff and it would be able to save all the data so i could kind of watch what it was doing as it was doing it um but uh so so i think you can get basically the same experience uh i think cloud code is a little bit more controllable and therefore a little bit more powerful but um ChatGPT is just much easier to use.

3:21:46It makes sense. What use cases do you expect ChatGPT agents to have the most PMF around? I was imagining the student use case, which is just like monitor the homework that I have due across. I remember in high school, even teachers would host their homework on websites. You could basically run something that was like monitor the homework assignments that I receive and then take a preliminary pass at doing the assignment and then give me a draft that I can review and sign off on or tweak and then automatically submit it. Then submit my college applications, attend college, get a job for me, just deposit the money that you make as an engineer at multiple companies into my bank account, and then also plan a trip to Europe because I'm retiring.

3:22:32Watch out, Clue Lee. 10GBT agent's coming for you. Okay. Boom. Breaking news. Breaking news. But giving this powerful of a tool to everybody immediately, not everybody's going to realize it, adopt it right away, but you can imagine a few use cases just spreading like wildfire. Totally. Yeah, I mean, I think what are the things that you would immediately do if you had an assistant? If someone just dropped an assistant into anyone's lap, what was the first thing that they would do? I don't know. help me book a vacation, help me figure out how to order groceries, help me. Another thing I use it for is research the web about all the topics that I care about and every day give me a report on all the things that happened in the last 24 hours.

3:23:24And it just does that incredibly well and I can go behind login walls and paywalls and all that kind of stuff. So I think those kinds of use cases are going to be the most interesting ones. But I honestly think right now for most of my consumer use cases, 4.0 or really 0.3 is the best. It's much faster. I mostly don't need it to use a full computer to spin it up. So I see Chattapulte Agent as being something that you use every once in a while rather than something you're using every day. Sure. Yeah. So we're increasing the level of complexity. Like 4.0 is kind of a Google search replacement for me now.

3:24:04I just kind of hit it with like when was this person born? How old is this? You know, what's the state of this? What's the capital of this state or something? Expect a really quick answer then go oh three if I'm willing to wait a couple minutes want something That's a little more thoughtful maybe some search results from the web then deep research I'm actually trying to understand the full story read a whole report agent if I think it's gonna need to Use a computer actually take some actions pull some things together. How was the actual interaction of? of the back and forth. This was something we talked about with the OpenAI folks was like, if it gets stuck, it pings you.

3:24:40I like that deep research, yeah, it takes 15 minutes, but I've trained myself to just be like, forget about that until tomorrow. And then when I have time to sit down and read the full deep research report, which is gonna take me a couple of minutes, then I'll come back to it. I know it'll cook and it'll be done. It would be kind of annoying if deep research came back after two minutes and said, Hey, I'm going to pause all that while I ask you for an update. Like it feels like there's a little bit more. I got to be on answering questions there, but push notifications, maybe solve that. Walk me through like how, how involved you were, how active of a process it is.

3:25:17I mostly was not involved every once in a while. Like it does have a, like a stop. You can help us stop and like change what it's doing, which is nice. Cause like if it goes off the rails that's helpful but i think and it has a push notification thing but i think this is an interesting problem with agents where i can't stop watching them and so i spend a lot of my day just like watching the agent doing something and chadjubiki agent has its own like cool ui where you can kind of like see interesting animations of what it's what it's researching or which websites it's using and stuff like that so i find myself um actually glued to it to it a little bit and um i just don't think that's a very good way to spend time i think it's i think it's mostly solved by having push notifications but i like there's a sort of emotional process of training yourself to be like it'll let me know when it's done and i don't have to like wash over the shoulder or it's like we we design a lot of assets uh here at tbpn and that like even if i trust the person creating it to do a great job there's still this tendency to want to like hover and be like, okay, tweak this, tweak that.

3:26:20Oh, let's do it this way in real time versus like waiting. But it's the same thing with 4.0. Like, you know, early on, I would kind of be in this loop of like, okay, I got a result. I still got to go fact check this and check the underlying links because hallucinations are a big problem. Well, now that they've beefed up search so much and they're referencing direct quotes, like I feel like I'm much less, like the anxiety level around hallucinations is a lot lower just in general queries. Totally. I think these things are tools. Anytime you're delegating to something, whether it's a human or an AI, there's a learning process you have to go through.

3:26:56Human managers go through this with employees all the time. If you're a new manager, you have to decide, okay, am I going to delegate this or am I going to micromanage? If I delegate, I get more leverage, but it might not come back the way I want it to. Good managers know how to split up a task or communicate it to their employees or figure out who's good at doing what and know when to get into the details and when not to. And I think we're going through the same curve with models. So we're becoming model managers. And everyone is learning how to solve the same problems that human managers have solved.

3:27:31And so the more experience you have with the tool, the more you know, like, okay, I don't have to check this 4.0 answer or this looks a little fishy. Same thing with Chetup Tea Agent. I think we'll be much better at using it in three or six months than we are today. Cool. Makes sense. Dan, always great to chat. I do want to have you back on very soon to talk about LLM induced psychosis. I think it's important to talk about and let's talk about it, but we'll need a lot more time. Thank you for the vibe check and everybody listening, go, go subscribe to every right now. Do it. Awesome. Talk soon.

3:28:05Bye. Cheers. Up next, we have Chris Best from Substack, the best CEO Substack's ever had, arguably. He's in the conversation. Welcome to the stream, Chris. How are you doing? Congratulations. You got some news for us? You got some numbers? Get it ready, John. What's going on? What do we got? What's going on in your world? Please tell me at least nine figures. Doing good. We've raised$100 million Series C. Woo! Woo! Congratulations, Chris. I was hoping you guys would ring that thing. Of course. Congratulations. Of course. You guys are incredibly back. Yes. It's been a journey since I believe you raised something.

3:28:45It was like 75 on 700 back in what was it? 2021? 2021. Those were different times. I don't know if you guys remember 2021. I do. I do. I remember it fondly. You were at the center of the storm and I was in the depths of a slog basically. Anyway, Anyway, give us the update, how'd the round come together? What is the plan going forward? I heard you were, what was the reporting? Accidentally profitable, going back into burn mode? What's the money for? What are you thinking? Yeah, I like that. Yeah, we're partnering with Mood Rogani at Bond. I love that guy, Consummate Bro joining the board. The big thing is, look, the big thing that's happened is like Substack's gone from being like a rinky dink email newsletter company to a proper sort of network that's taken over the world.

3:29:37Yeah. And we kind of want to look basically kind of like switch into a mode of thinking about long term ambition, long term, like how do we actually make the big fucking version of this thing? What investments do we need to make? How do we focus on the things that actually matter for like the long term flywheel growth of the network? And this just gives us like a total freehand to do that thing and build the best possible version of it. Okay. Give me the pitch for, I think of Substack as the no-brainer place to launch a newsletter. You have a lot of other products. Talk to me about what the future of the Substack creator or someone who has Substack as like their primary out.

3:30:20It's the main engine of their creator economy business, for example. What does that look like over the long term? I imagine that people are still doing top of funnel stuff on TikTok, Instagram, other places, but you're adding more and more features. What does a well-run Substack business look like? Yeah. The core of Substack is the direct connection with your audience, right? So people subscribe, you get their email, you get the ability to reach them, you even get the ability to leave Substack and take your list with you, which is a big deal. People can pay you directly, so you get recurring revenue.

3:30:57I don't know if you guys have had this, but recurring revenue hits different. The sponsorship business is a great business. I think that thing matters. Lots of people on Substack have sponsors. We love it. But that thing is very cyclical. It's boom and bust. Whereas you have recurring subscribers, these diehards, it funds you to be creative. it funds you to take risks. And so Substack is the place where you sort of like your hardcore people are. That's where you have a real connection to your audience. You can write, you can post short form, you can post video, you can do live video now. You can have a community.

3:31:31We're kind of like building more and more stuff. The center is not any one format. It's like the relationship with the subscribers. And then Substack is just becoming, you know, you said you do top of funnel on TikTok and LinkedIn and YouTube and everywhere else. Sure. Keep doing that. That's great. Those or massive platforms, but increasingly you can do that stuff on Substack too. And because you have such a dense audience of like smart people, the quality of growth you can get there is already very high. Interesting. That makes a ton of sense. Jordy. I think the magic thing that you guys tapped into that I end up finding myself explaining to other people is there's this beautiful like economy of people on Substack that just want to support people that are nerding out about a specific topic, just want to give them money.

3:32:21And so it's almost like there's this exchange of like, yes, I want the content, but it's also enabling somebody to live a life that allows them to just obsess over one thing or just explore a series of topics or just be who they are and be entertainment through that. I mean, we've had Emily, Emily Sundberg has come on the show a bunch of times and it's just like, it's so awesome to see what she's built and the kind of creator writer that she's able to be unshackled from being at a specific, you know, platform, legacy media company. So it's just, it's so, it's so awesome to see. Talk about the use of funds.

3:33:05You said you can afford one AI researcher now. I imagine that won't be how you spend it. Never know. YOLO. YOLO. Start coaching for Mark. Concentrated bets, man. That's how it works now. Concentrated bets. But I mean, concentrated bets, it's not the craziest idea to go give a bunch of money to creators to kind of pull forward the leap from what they're doing currently to get on Substack. There's different incentive models, kickstart ad businesses, just hire engineers that can build new tools and new features and just chop wood and advance the ball down the field. What are you most excited about to put that money to work over the next like 12 to 18 months?

3:33:48But you probably think in like decades now at this point, right? Yeah, I mean, that's the big thing, right? It's like, this lets us have that longer horizon. You can still have all of the same math, but you can just like put the planning horizon further in the future and look for something really big. Like, listen, all that stuff you said, the stuff that I'm really excited about is like making the product fucking great. Yeah. Right. I want it to feel my joke is Substack does everything for you except the hard part. Right. You are the talent. You got to figure out how to write something that's worth reading, how to have a conversation that's worth listening to.

3:34:19If you can do that thing, though, we should just build this magic machine that takes everything else off your hands and makes it dead simple. Makes it just like this magical thing where anybody who has something worthwhile to say can make something. We're starting with that. We have a bunch of little bits of that that are kind of working that we're really proud of that are exciting. But I just think the new technology coming online is going to make us like the version of that magical sort of like media studio, personal media empire in a box that we can build now is going to be so much more powerful.

3:34:47And then building up like the network, right? The fact that we're getting, you know, not just political commentators, but politicians. I think we get not just sports commentators, but athletes. I think we can start to build up kind of like this network and this ecosystem that winds up being this positive sum game, right? Where everybody that's on Substack benefits from this growing network. Yeah, I've certainly never subscribed to anyone on Substack and been like, I didn't get my money's worth. I always have a good time. What are the different strategies for Substack writers? I know in like the Patreon podcast world, people do like one is free, one's behind the paywall.

3:35:25I've also seen sub stacks where there's like a fold and you get every email but you get half for free and then you at some point There's a call to action to go and subscribe and and and finish reading essentially, but you get every email what works What are the different strategies? What are some of the weird stuff that you've seen around the way people are using sub stack today? There's a pretty big mix right some people make almost everything free and it's just basically like you know if you want to comment or if you want to get the occasional thing, that's what you're paying for. You want to be really low margin for Chris over here.

3:35:58You just give everything away for free. He doesn't make any money. That's the beauty of your system. Nobody's paying to get more things to read, to get more email, to have more seconds of audio in their inbox. They're paying for perspective. Interesting. Right? Yeah. That's the thing. And even if you have a magical LLM that can spit out media in any format, you care about who it's aligned with. You care about like what version, you know, what worldview you're getting. Is this something I trust? Is this something I want to be culturally and aesthetically a part of? You talked about people paying because they want to support people.

3:36:32The other way to say that is it gives you agency, right? When you choose who to subscribe to, you're choosing what part of the culture you want to live in. You're choosing what gets created. You know, in a world where people, I think, feel like a lot of the media they consume is kind of like stuff down their throats, getting to kind of like exert a voice and say, I'm causing this thing to exist that I think is great, is really powerful. And a lot of different ways it works. It's like paying with dollars versus attention is super powerful, right? Because if you're paying with attention, it's like, well, then everybody's focused on the Coldplay debacle last night, which is trying to steal thunder from your fundraising announcement, but we're not letting it.

3:37:15And being intentional about like I want to pay for this because I want more of it to exist. I want it to get better. And this is how I want to spend my life. I don't want to be on a platform that just is designed to like suck my time from me. But I want to spend my time and attention better on smart things like TBPN and everything on Substack. Yeah. No. And coming soon, hopefully. We will figure it out. No, we're about to bet big on Substack. And we're riding with you. So congratulations. It's really tremendous to see how far the business has come since those glory days in 2021 and excited for the next five years.

3:37:57Thanks, guys. Cheers. Congratulations. We'll talk to you soon, Chris. Cheers. Awesome. Next. One last guest. Descartes.ai. That's right. The first ever world transformation model turning any video game or camera feed into a new digital world in real time. Very, very cool. I played around with a lot of this stuff, not this in particular. Very excited to talk to the founder. Welcome to the stream. How are you doing, Dean? Welcome. Sorry for the chaos. We are so happy to chat with you. Super nice to meet you both. Super nice to meet you both. Good to meet you. Why don't you start with an introduction?

3:38:33I already have questions, but just give me a little background on yourself and the company. Okay. So, you know, the card's a very young company. We're less than two years old. We're a research lab and our goal is to build a consumer company. Okay. Okay. And what we just launched today is the only real-time video model ever. I can just show it to you. Yeah, please. Let's do it. Can I share a screen somehow here? Yes, but you are live. So whatever you share on that screen is baked into the internet forever. Amazing. We just have to make sure that we don't need to leak anything. Just do the right tab, not the API keys.

3:39:09I wasn't sure. We've never met. I mean, we've met over DMs, but not face to face. I wasn't sure if this was your real face or just a character that you're playing. This is definitely not me. You realize that, right? Yeah, in the real world, he's an anime. But he's using a Transformer model to appear like a human. That's right. He's, in fact, a Minecraft character. I have so many questions about this model. I want to jump into how you built it. We could also potentially have the team pull up the demo video on your website, mirage.descartes.ai, and just kind of show folks what that looks like. Whatever would be helpful.

3:39:51It is currently, I'm trying to get the Zoom call to be able to share this. But if the team could do it from your side as well, that could work too. Yeah, why don't we just have the team pull up the core website and we can just jump into questions. So real time, what's the secret sauce? Is it a condensed distilled model? Are you using a special chip to inference this stuff? So real time, let's just talk about the use case. So the obvious use case would be like real time video calls. You're dropping into a Zoom call, for example, and instead of yourself. There he is. There we go. Do you guys see me?

3:40:27Yeah, we see you. Looking stunned. Okay. There we go. Okay, okay. Okay, now we got it working. There we go. There we go. What is the use case here again? Who are you? There we go. We're just cycling through. We're just cycling through everything. What do you guys like? What do you guys like? Anime. Not big into the anime world, but what about like Legos? Anything there? Legos? Let's put in Legos. We can just type it. Oh, you can just type it. And it'll do it. Let me make this full screen. You can just type in Lego. and it'll just turn everything into Lego. Wow. Okay, there we go. This is our house.

3:41:10Wow. You see the real stream and the Zoom. Yep. You can see, you know, this is the house office thing. You see all the Lego characters walking around here. This is insane. This is trippy. And then you can decide that you're into Christmas, and so everything becomes very Christmassy, and your house is decorated. Oh, wow, yeah. Wow. There you go. I feel, what's the word for a moment like this? This actually feels like enter the metaverse. Indeed. I was hoping you wouldn't say that because that became a cursed word. Yeah, it did. But it's a good thing. It should happen at one point, right? No, this is a big transition from the moment where Zuck did his hello from Horizon's world.

3:41:57Those Wii graphics. Those Wii graphics. and this feels like being in a video game. What is that? Is that a wand? This is, you can see the straw on the regular stream, right? Yep. And inside the wizard's prompt, it becomes a wand. And if you flick it hard enough, sometimes it does magic spells. It throws things out. Guys, which one? The team's been playing with the prompts here all day. Which one? Galactic War's the one you like? oh cyberpunk action that's uh the lightsaber one okay no you had uh you had one that you should go here so is this running locally on your computer so this no this is running on the server it's running on the server and it's going so it's going from your webcam to the server back to us over so how quickly is is like kai sanat going to be running this like half the time that he's streaming this is this is yeah this is fun it's pretty fun now there's also what you do with it right yeah yeah this is a bold this is a bold demo too because the variation in the different prompts and how how well it's working is absolutely insane totally totally i have forgotten what you look like in the real world entirely here's a question would you guys run an entire show through this?

3:43:19I mean we can we can try. Probably not. We might put Tyler on the intern cam. Yeah yeah that's a good place to start. So we have an intern cam over here. We can get Tyler running on this. Oh my gosh the way that you're just cycling through these prompts. Tyler can be on wizard cam for a little bit for one stream. That's pretty crazy. This is insane. Something something cool. Let me show you this. Something cool that we found out is that it's really fun watching YouTube through this. Okay. Watching YouTube. Tell me, give me a YouTube show that you like, something you like. I mean, just do that Mr.

3:43:51Beast video, whatever's pulled up right there. Mr. Beast video. Okay. So you can just look at this Mr. Beast video. We can look at this ad. And look at the ad, and we can share the Mr. Beast video. I somehow have only... Do you guys hear the audio now? Yeah, we do, actually. A little bit. Do you hear the audio? I don't know if we should, but... Okay. Okay. And then it's being transformed now? Yep. This is the original Mr. Beast video. Wow. Yeah, even trippier. You can. Here's, like, Mr. Beast videos worked really well with Cosmic Medieval Golden. We iterated through so many prompts that we found the ones that worked really well.

3:44:28Yeah, yeah, yeah. This is Mr. Beast videos in the Cosmic Medieval Golden world. Okay. Cosmic Medieval Golden. I mean, they're already pretty stimulating. There's just even more now. Wild. Very wild. This is insane. Where do you think this lives? Does this live with the consumer and they decide to put on the rose-colored glasses? Or do you think the creator decides? Somebody's joining a stand-up tomorrow. Can they just automatically put the entire team into whatever character they want? Let's see. What stand-up comedy is he like? I was talking about an actual engineering stand-up. Oh, an engineering stand-up.

3:45:10Okay, I can definitely do that. I'm more asking about you. I'm asking you, where do you think it winds up living? Here's what I think is cool about this. I think it's the first time that we have a new kind of consumer interaction with video AI. Because so far, video AI was just, okay, let's create AI slot, put it on existing platforms, Facebook, Instagram, TikTok, whatever. Here for the first time, you can actually do something that's slightly different. We're showing it to a bunch of kids and what they ended up doing was for a few hours they just fought each other with sticks and they started doing like TikTok dances in front of this.

3:45:47And you have a new kind of consumer experience here. Yeah. Yeah. It's like the original in the, what was it? Steve Jobs did that demo of the Mac where it warped the image. And it's like a crazy historical photo in the original Mac OS launch. How expensive is this for you guys to run if somebody just starts streaming this in real time? Are they paying for it? Are you guys just eating it? We're doing it efficiently enough. Like we had to write like all the low level assembly code for GPUs to get this to be both real time and super cheap for us. It's at a point that we can actually provide this for free.

3:46:31Like it can actually be monetized without like subscriptions. Like there are ways it's cheap enough to actually be able to build a platform on top of this. Interesting. Do you think, yeah, I mean, do you think one of the use cases will just be folks who create video content running their video through this as like a previs for what they ultimately want to build? That could be very interesting. Well, the other thing you can imagine it actually in YouTube. So like a creator like Miss Rachel, for example, popular kids creator could basically say like, do you want the Lego version of this video? Do you want the pirate version?

3:47:12Where this lives, because I can go and put my phone in grayscale and I can view everything on my phone without color. And that is effectively a style transfer. And that's something that I as the consumer decide. Or Mr. Beast can turn up the saturation. Or if you're watching a Hollywood movie, they might turn up the teal and orange. Because that is a traditional Hollywood color grade. If you're watching The Matrix, they might color grade it green when they're in The Matrix. They might color it blue when they're out of The Matrix. And so it will be interesting to see where this lands. Whether it's on the consumer.

3:47:48I like watching Lego version of YouTube. I like watching Lego version of my conference calls, or it lives on the producer of video side. I want to show up as Lego. And then that could be transformed as well. Can you quickly do a GigaChad filter? Oh yeah, do you have a GigaChad filter? Let's check that out. Let's see what GigaChad does. I will say that the model is really, in the current version of the model, is good at changing the entire style. Okay, yeah, yeah, yeah, not just the face. It's not, oh, turn me into Trump. Okay, you're pretty orange, yeah. There you go. But it is doing something to the cheeks and the jaw, which is textbook giga-chat.

3:48:27There we go. There we go. Very funny. There we go. Yeah. This is wild. Maybe you need to prompt like Arnold Kosser. So is this already fully open access? This is open. We launched it literally 30 minutes before the podcast. We were waiting for you guys. Congratulations. Thank you. Yeah, yeah, yeah. This is the first time it's used on a video call. Amazing. insane insane well thank you so much for joining uh we will get access to it we'll set it up on the intern cam tomorrow for sure and wait you already have it set up okay we're gonna hop off with you and we're gonna check it out with tyler get some more feedback thanks so much for hopping by congratulations to you and the whole team insane and uh is there anyone else in the waiting room or are we good i think we're good let's uh can we go over to take it over to tyler the intern cam and see if that works i don't even know if that if he's there oh oh technical difficulties of course anything else you want to cover in terms of news or yeah there's some new news oh wow that's tyler there wait really whoa yeah you just did it yeah this is me like 30 seconds ago okay wow yeah it looks better without the zoom compression it it's a pretty high fidelity model wow very cool yeah so we were talking in the background so you hear our talking and then Tyler's in the corner too.

3:49:43We're getting recursive. Not that we should use that word now. It's the M dash of the modern era. Interesting. Pretty good. And these were preloaded prompts that you were just clicking through, Tyler? Yeah, I was just like selecting. You didn't generate any of your own prompts? I didn't, but I think you might be able to. Yeah, yeah, that seemed pretty cool. Some of this seems fun. I imagine that people will be having fun on this on the internet very, very soon. Very, very quickly. I always just wonder with this stuff, like what the novelty is like, Like you have to, it's like, wow, it's incredible technology.

3:50:14But then people have to actually figure out like a real use case for it. Like, like. Well, so, so here's a use case. Streamers. Yeah. Having some type of like, basically like if you tip a certain amount or hit a certain button, it puts the streamer into that. That's cool. Yeah. Yeah. I mean, a lot of streamers stream with, with green screen backgrounds and they already drop out the background and put themselves in an environment that matches the game that they're playing. We're just cleaning up the messy background. And so, yeah, you can imagine this being like basically in the VFX pipeline for streamers.

3:50:47I think if most people showed up to a Fortune 500 Zoom call with this, it might not go over too well. Well, they can always make everybody on their screen look a certain way and make themselves look normal to everyone else, right? Yeah, yeah, yeah. Yeah, if your boss is yelling at you and you put them in Lego mode, it's probably going to hit a little bit differently. It might be a little bit more tolerant. Yeah, turn the volume down. Lego mode. Lego mode. Oh, sorry, Lego boss. Yeah. Block man. You expect me to be a fan. A couple more headlines since we're here. Perplexity apparently just closed a new round at$18 billion.

3:51:28$18 billion, wow. So perplexity has around 2 % of queries according to Semi-Analysis right now. I use it regularly. We talked to Chris, who was on the show. 7 million DAUs, 30 million daily queries, 4.3 queries per user per day, a 2 % share of users, and a 2 % share of queries. Solid. That's the perplexity stats. Solid numbers. Pretty good. They got their new browser. They're going to be investing heavily. And then outside of that, Lovable just raised$200 million at a$1.8 billion valuation led by Excel. So lots of heat on the timeline. Lots of big rounds getting done. Everyone decided to launch today, I feel like.

3:52:10And then they got steamrolled a little bit. But that's why we're here. We didn't spend too much time on the concert fiasco. Instead can move on and talk to you about the news. Also, we want to send our best wishes to Tyler. He says, hi, crew. I had surgery to remove lots of infected fluid for my chest and a big lung abscess with infected tissue. Pneumonia, sepsis, antibiotics weren't working in the ICU now. Surgery went well. Surgeons are heroes. I'm so grateful. Recovering, thank you for your prayers all good. So I hope he's doing well. Let's send some prayers over to Tyler. Yes. What else is in the timeline that would be worth to close on?

3:52:52I have a good closing post. The lads over at Reindustrialize are having a lot of fun. If you can pull this up in the bangers tab, you can see quite a number of former TBPN guests all in the back of a pickup truck. Fantastic. If you can pull this up. Team, we're working on reducing the time between talking about a post and getting it live on the screen. There we go. Oh. All the boys. We got Augustus, Paki, Steinman, everybody all in one place. Love to see it. So anyways, thank you for tuning in today. Super fun show. And I can't wait for tomorrow, John. Can't wait for tomorrow. Leave us five stars on Apple Podcasts and Spotify, and we will see you tomorrow.

3:53:38Have a good day. We love you. Bye. Bye.