Anthropic Hits $380B Valuation, Become Unsloppable, WSJ Mansion Section | Martin Shkreli, Connor Hayes, Alex Bouzari, Brett Adcock

29:54It was almost two years ago. Oh, this is, yeah, this is almost two years ago. So, I don't actually know. I don't know which way Metacritic was going. Cell phone? Asked Martin Scrawley, who's coming on the show. Oh, he responded. He said, the puzzle meant I didn't understand why it was worth only$18 billion. Always vague posts so that you can take either direction. You never print yourself into a corner. Let me tell you about Gemini 3 Pro. Google's most intelligent model yet. State-of-the-art reasoning, next-level vibe coding, and deep multimodal understanding. I'm glad this chart is now public because it is bananas.

30:28It is ridiculous. It should not exist, says Bruno F., the founder of Magna Digital. $5 billion in tokens managed. Interesting. yeah what crazy crazy just another pod guy says salesforce invests in anthropic colorized i think when mark uh was on he said they have about a point of anthropic going into this round if i remember correctly again what is this uh it's a horse giving money to a car the car goes and buys a rocket launcher the car blows up the barn and the horse is sad and that does Be like an apt analogy. It's very funny. And Anthropic has been all over legacy media. First, let me tell you about CrowdStrike.

31:16Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches. Also, fantastic Valentine's Day gift. From this extraordinary piece in The New Yorker, last summer while Mark Zuckerberg was conducting hiring raids on other labs, Shalto Douglas, the Anthropic engineer and TVPN guest, told me, this journalist, Gideon Lewis-Krauss, that a number of his colleagues, quote, could have taken a$50 million paycheck, but the vast majority of them hadn't even bothered to respond. Well, they are early at a$350 billion company and are clearly very optimistic, but it's funny to just money mog.

32:02Daniel says, so wait, Claude has seat-based pricing? Does this mean they're disrupting themselves too? Of course, a lot of the concern has been around the seat-based model. But it even feels like that is less of a factor than just the overall threat of zero marginal cost software. Yeah, why does Claude have seat-based pricing? It's essentially a consumption-based product, but psychologically, if I'm rolling out Claude to a company and I set up seats for a team, I know that there's individual rate limits, so no one individual is going to blow me up, basically. That's the idea. But this goes into, like, some people are posting, like, if you're getting a job, you should ask what your target is.

32:49But in this case, it's not. This is Claude. This is not the API. This is not Claude. But when you fire up Cloud Code, you can integrate your Cloud account. And so this essentially gives you credits to write code as well. And so there's this new meme of if you're going into a tech company, ask what your token budget will be. What's your inference budget? And so these can clearly skyrocket pretty quickly. there's debates over, you know, oh, should I let my employees use like the fast mode or the regular mode or pro? Like, uh, is the work that they're doing really that valuable if they're spending thousands a month, if they're spending tens of thousands a month, like at certain point, I need to make sure that, uh, they're not being wasteful.

33:35And so I think the seat-based plan still, uh, achieves a little bit of that psychological security for managers. And then there's also an interesting, there's probably a pretty bimodal distribution in the value that, or the actual cost associated with these plans. I would imagine that there's a portion of pro users that use 100 % of their inference budget every month, and they cap out, and they're frustrated, and they might have a second plan, or they might go down to a free plan or limit their usage, and then there's a whole bunch of folks who just have a seat and never use it, or they use basically like very little inference or they're just asking things that can be answered by a free by a free tier essentially but they just like let it ride zach in the chat says we have 150 corporate cloud users purchased by seat 50 max out in week one because of excel token use yeah there you go good data point let me tell you about restream one live stream 30 plus destinations if you want to multi-stream go to restream.com uh dan primack says working on newsletter it may be shorter to The list of VC firms not in the new Anthropic round.

34:42It's a party round. Josh says, got some logo sniping, it seems. Pay no attention to what price we paid. I swear it was early. Yeah, no, this is a very, very good point. If you are a venture capitalist and you say, I mean, it used to be if you were an AI VC and you had AI as a thesis and you weren't in one of the labs, that was sort of a red flag for your brand. It would be rough to move forward, raise the next round. just from a logo perspective. And now with AI becoming such a mega trend, it's hard to imagine being a really enduring venture capital firm without one of these logos on the site, especially as like, this is sort of the train leaving the station since there's IPO rumors.

35:27And you probably want to grab at least one of the big labs logos. many of the firms have sniped all three at this point. Sequoia Founders Fund, Co2's in a bunch, Andreessen's in multiple, I think. There's a variety of funds that have built stakes of various sizes in all the different labs. It seems like Josh Kushner is one of the few that has remained deeply loyal. Yeah. Yeah. And we do recognize we just had some technical difficulties, but it seems like we're back on. We are so back. Let me tell you about Console. Console builds AI agents that automate 70 % of IT, HR, and finance support, giving employees instant resolution for access requests and password resets.

36:20Slow Ventures is taking the other side of the all-in-on-AI bet. They said, congrats to everyone who figured out that foundation models are infrastructure plays, not startups. Now let's talk about what happens when the picks and shovels phase ends and we're back to building actual products. And Will Manitis says, nightmarish degrees of coke. Yeah, Sam was always super bearish on the labs from a business standpoint. He thought they would commoditize? Was that what it was? So they wouldn't have pricing power? Effectively, yeah, effectively he said, you know, open source models are going to get really good.

36:56He was right. They have gotten really good. But I think maybe missed that the labs would turn into product companies and stop just being, you know, guys. It is sort of interesting. Like if you wound back the clock and you were like, my job is just to invest well in tech startup booms from 2005 to 2025. There's one world where you're like, okay, I'm going to go hunt for the Airbnb, the Stripe, the YC companies, the Coinbases, all the like application layer companies, the Instagrams, the Twitters, all of these different companies. but there's a different side where you're like I'm going to buy like Broadcom, Cisco, NVIDIA, AMD and still do really well and maybe even better depending on when you got in, when you got out.

37:46But it's a, yeah, it's a very like just because it's an infrastructure play, even if that's true, that doesn't mean that it's not a good investment for an investor. There is a little bit of like purist vibes from like a venture capitalist should or certain funds have strategies. And so they say, you know, I'm just going to sit out slow as a seed series a, but, but lean more early. And so by the time you're looking at some of these deals, investing at, uh, you know, five, 10, you know, 20, 30, 40 billion. Yeah. And there's start to be rough. Yeah. And there's a lot of, uh, there's a lot of funds who have expanded and will invest in anything.

38:27Like you're a mining company. Great. Let's do it. You're, you know, uh, you're, you're buying Bitcoin on the balance sheet. It's like, okay, let's do it. Uh, there were a lot of funds that expanded what it meant to just be an asset manager. Uh, and there were some funds that, that stayed very focused and, you know, we'll, we'll, we'll see, but, uh, interesting, uh, highlights from the Dario Amadei interview on Dwarkesh Patel, Jacob Rintamaki, friend of the show has a quote here. All my lawyers never want me to say the word monopoly. Dario says, I don't think that's true. I mean, I feel like we're in an economics class.

39:02Dwarkesh says, do you know the Tyler Cowen quote? We never stopped talking about economics. And Dario says, we never stopped talking about economics. So no, I don't think this field is going to be a monopoly. All my lawyers never want me to say the word monopoly, but I don't think this field is going to be a monopoly. you do get industries in which there are a small number of players, not one, but a small number of players. And so that feels like where things are going with both the expressed viewpoints of the VC firms, investing in multiple labs, that there's a variety of strategies to deploy intelligence, whether it's the best model and get deployment and traction, whether it's on the infrastructure side.

39:45I do wonder how many more changes there will be in the horse race. It feels like there's a new hot model every couple weeks, and then someone fires back, and then they go back and forth and back and forth. And with all the flow between the labs talent-wise, it feels very hard to corner the market, and it doesn't feel like anyone can patent the transformer or anything like that, which would be a completely different scenario. can you imagine if Google just had the patent and they were like, we actually filed a cease and desist against OpenAI and Anthropic. They're not allowed to use transformer-based architectures anymore.

40:24Like, we invented it and we patented it and you can't have it. It's ours. I don't know. But yes, we live in a world where the little tweaks, the little strategies that go into advancing the models and creating these improvements do not seem to be intellectual property. they seem much more like economies of scale and process power of being able to train at ever larger scales, marshal ever larger chunks of capital and do whatever it takes to get to the frontier and stay there. Let me tell you about Vanta. Automate compliance and security. Vanta is the leading AI trust management platform. Why do we play this?

41:07Because the stream, We're having issues again. We're working to get it back up. If you can hear us, markets now see a 30 % probability of a Fed rate cut by April, more than 80 % of easing by June. Over on CalShit, we're still seeing the Fed decision in March, 93%. Say maintains rate, no cut. So a cut would be a wild card at 7%, 9 % for any sort of cut. So strong GDP growth, strong job numbers. Stay the course would be the logical thing. But we will continue to follow it. Let me tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web apps, servers, databases, and more.

41:54Well, Railway automatically takes care of scaling, monitoring, and security. Let's play this timeless clip of George Hotz. Nah, we don't believe in stealth. I'm a really open guy. You are pretty open. I tell you everything I'm doing. Come on. Here's what I say. Here's what I say. I'm going to tell you what I'm doing, and you can try to compete, but I'll still crush you. Nah, we don't believe in stealth. It's so funny. Also, I don't know why that person cut that to be so widescreen. It looks very cinematic, but I like the quote here. You think the eggs I lay are valuable? I am the golden goose.

42:35Yes. Meanwhile. I'm thinking people will steal your ideas if you share them is a sign of low IQ. And I agree we are in the era of agency and actually going and executing on the idea is the difficult thing. You need to be charisma maxing. And there's still a lot of secrets to every business. Yes, there are. CEOs can do 100 hours of podcasts and tell you a lot about what they're doing without telling you the one or two things that are actually important. And it's very easy for somebody to come in and try to fast follow and ultimately just kind of get it entirely wrong, even though it looks like the right.

43:19And I think there was a huge incentive, I mean, going back to the SaaSpocalypse, there was incentive for a long time for companies that where their moat was not software to say, we're a software company, we need to hire the best software engineers, look at our open source projects, focus on all the cool tech that we're building, when really it was a marketplace or really it was a liquidity provider or really it was a network effect. And if your network affects business, it can be sort of boring and honestly anti-competitive to just be like, look, we can do nothing and win. No one wants to hear a CEO say that.

43:51But we're going to find out who can do nothing and win because we'll see it show up in the margins over the next couple of financial years. Meanwhile, over on LinkedIn, George Hotz is posting. And Reet says, George Hotz is the only thing keeping my LinkedIn feed good. He says,

44:37every crevice of people, and they are run by psychopaths. The AI will be a further wedge, just another lever to manipulate you. You will not be able to stand up to it, and you will be discarded the second you don't serve it. Like layoffs, you will die atomized and alone, and you won't understand that you did this yourself. Brutal. Nice little white pill. Nice little Friday white pill. He's such a white piller. Well, here's a white pill. First, Figma. Figma Make isn't your average vibe-coding tool. It lives in Figma, so outputs look good, feel real, and stay connected to how teams build, create code-backed prototypes and apps fast.

45:10But here is the real white pill. For just$33 million, you can have a private home on a remote resort in Utah. Can you guess where it is? Park City? Nope. It's at the Amon Luxury Resort. It's in the Wall Street Journal. The residence is the first to hit the market at Amon. In remote southern Utah, Amangiri Resort, a crown jewel in the Amon Hospitality Company's portfolio. Wait, they're doing residences? They're doing residences. They have a portfolio of global hotels and residences. They're listing the first private home for$33 million. Located just over the Utah border from the small town of Page, Arizona, the hotel currently features 34 guest suites starting at$5 ,000 per night and 10 tented pavilions at its Camp Sarica, providing a temporary escape for travelers.

46:02But the newly built house on nine acres can be purchased outright. Designed by Los Angeles-based firm Masa Studio, the roughly 12 ,000-square-foot residence has six bedrooms and comes fully furnished. It's the first of 12 planned private homes, which will be about half a mile from the resort. OTP says, but does it have a bunkie? Does it have a bunkie, a bunk bed? No, bunker. Oh, we're going to get into bunkers. There's a whole piece in the journal about how to secure a mega mansion. I know you've been asking. We have the answers. Until it's sold, the home is available to rent for$45 ,000 per night.

46:41Before we continue, let me tell you about Century. Century shows developers what's broken. It helps them fix it fast. That's why 150 ,000 organizations use it to keep their apps working. So, residences have been part of the Amangiri vision since the resort opened in 2009. The decision to offer private residences now was spurred by the success of the 2020 Tented Camp launch, rising demand globally for hotel-branded residences, and a sense that the property was ready to take that step. While future residences will share a cohesive aesthetic, each will be designed to respond to the unique contours of the specific site.

47:21In Page, the median sale price was$610 ,000 in August. There are currently around a half a dozen listings above$33 million, though, all clustered further north near ski resorts. Amangiri buyer interest has been strong, he said. Particularly among Amon loyalists and North American clients, Additional residential plots priced between$5 million and$12.5 million are under contract. And so let's get into Lambda. Lambda is the super intelligence cloud building AI supercomputers for training and inference that scale from one GPU to hundreds of thousands. The mega rich are turning their mansions into impenetrable fortresses.

48:04And we're going to tell you how to do it for yourself. Anxiety over high profile violence has the wealthy spending big on armed security bunkers, a bunkie and even moats they're building moats i haven't heard of an alligator in the moat or shark in the moat but people are in fact building that's on being an alligator salesman i feel like is unsloppable i think it could be clankable yes but still at the moment unsloppable but you got to build the you got to build the humanoid robot that can go in the water to wrestle the alligator And that might be, we'll ask Red Adcock, is it waterproof? Is it waterproof?

48:39Can it go wrestle an alligator or not? Because I don't want it just to do my dishes and do the laundry. I want it wrestling alligators in my moat. So British music producer Alex Grant was living in an under construction mega mansion in Los Angeles. One morning, shortly after 9 a.m., an intruder armed burst into the home. He said, Grant said, he came in and we had a tussle. He was formerly known as Alex DeKid. Grant managed to call his manager, who phoned the police. Soon, officers and helicopters were on the scene. He briefly considered abandoning the project after the 2017 break-in, but ultimately finished the 24 ,000-square-foot home, which has eight pools, a car elevator, and a nightclub.

49:20Wow. But he doubled down on security features, installing a guardhouse, tall gates, and a security system with retina scanners that alert the homeowner to movement in the home. Later, I found out he had these knives on him, Grant said, who recently listed the mansion and a neighboring house for$85 million after moving to New York. In an era of high-profile violence, including the suspected abduction of Savannah Guthrie's mother from her Arizona home just over a week ago, the wealthy are investing heavily in their personal security, particularly when it comes to their homes. Security measures once reserved for presidents and royalty, safe rooms, biometric access controls, laser-powered perimeter defenses, These are now mainstream items in luxury homes.

50:01Executive protection teams and armed guards patrol gated enclaves and suburban estates, while tech startups are rolling out predictive threat detection systems built for the ultra-wealthy. The shift reflects a hardening view among the affluent. Traditional policing and communal safety are no longer enough, no security. So security is being privatized and customized. The new emphasis is reflected in sales data. Roughly 45 % of luxury homes in 2025 included a reference to privacy or security, up from 38 % the year earlier. So break-ins at the homes of celebrities and professional athletes have been putting the wealthy on edge.

50:36A group of Chilean nationals was indicted last year for stealing items worth more than$2 million from sports stars, including Kansas City chief players Travis Kelsey and Patrick Mahomes. Travis Kelsey. This had something to do with the visa process with Chile, where you could very easily get a tourist visa. Oh. So there was these, like, allegedly there were teams that would be permanently based in the U.S., and then they would be running kind of the operations. They'd be in the kind of war room, and then basically tourists would come for two weeks, hit a bunch of houses, and then bounce. Bounce.

51:11Wow. And those were the only people that were actually exposed to, or exposed, meaning they were like carrying out the different ops. Well, the Miami Dolphins player Tua Tagovaila, Viola, Viola, I might be mispronouncing that, said he hired personal security to monitor his house while he's on the road. He says, let that be known. They're armed. So if you try to go inside my house, think twice. The homes of celebrities like Brad Pitt and Nicole Kidman have also been broken into. Miami real estate agent Danny Hertzberg of Cold War Banker said he began noticing an increase in emphasis on security in 2020 when high-profile executives were migrating from New York to Miami during the early days of the COVID pandemic.

51:56Private jet tracking websites have also been an issue. They sent chills through the high net worth community. Corporations are taking note. Companies offering personal security benefits for CEOs increased by 10%, according to Goldman Sachs. One entrepreneur capitalizing on this growth is David Weiderhorn, who got into real estate after selling a tech company in 2017. I wonder what he sold. He recently built a heavily secured home in Scottsdale, Arizona. And in early December, Weiderhorn walked through the 8 ,600 square foot property, pointing out 32 casino-grade AI-powered facial and vehicle recognition cameras.

52:34There's also a laser intrusion detection system around the perimeter. pausing at a steel double gate in front of the house. He warned that the security system kicks in even before visitors reach the front door, which is fashioned out of three-inch thick solid steel and has 13 deadbolts. He said even the landscape was designed as a deterrent. Cacti? Sour orange trees. There are sour orange trees with four-inch spikes in concrete planters on the edge of the property. And just beyond those trees, separating the house and the street, a moat. Gators. A moat. Gators. If you try and run through that bush, it will be a bad day for you, he said.

53:19Should anyone get past the trees, lasers will detect motion and the system will call the police. Inside the house, three ear-piercing alarms will go off. And this is an interesting thing. The fireplace surround, like around the fireplace in the great room, It will change colors. It's made out of crystallochartzite, and it can change colors, so it'll turn red. So you're sitting there, and if there's anything detected on the property, your fireplace turns red above the TV to show you that something's going on. Very interesting that that's a visual cue. The home's most fortified feature lies behind a wood-paneled wall, a reinforced concrete safe room with a 2 ,000-pound door and an air filtration system built to U.S.

53:59Army Corps of Engineers standards. Weiderhorn declined to see our specifics, but said it cost more than$10 million to build the house. About$1 million was spent on bullet-resistant smart glass. And the front entry security features cost more than$1 million. In Las Vegas, clients of luxury design firm Blue Herons are spending between$100 ,000 and$1.5 million on security features, including safe rooms and bunkers. Blue Heron is now working on new ways to incorporate architecture with security, such as exterior window shades that could be closed with the touch of a button to protect the home's occupants.

54:32In Surfside, Florida, the developer of the Delmore, a planned 37-unit ultra-high-end condominium project designed by Zaha Adid Architects, and with units priced at up to$200 million, has tapped a Washington, D.C.-based security firm to design the building's security. The$200 million condo. Yeah, that is crazy. But I mean, I guess from a securities perspective, if you're in some massive building, you're sort of like diffusing the cost. There's more people that might notice something. There's more security guards. It's almost like a gated community in one building. I'm just, yeah. Probably layers of access.

55:07I'm just purely thinking you're effectively looking at$100 million a floor, right? That's crazy. A couple floors, maybe a few. It's up there for a condo. That is huge. The firm is working to integrate technology like biometric access, facial recognition, and Irish scanning into the design of the project. For instance, when a resident or visitor pulls into the building's parking garage, their car will be scanned for license plate recognition. But facial recognition may also identify the car's occupants and their level of approval to access the building. That, in turn, triggers the security system to allow the person to unlock only the doors and elevators that they are permitted to pass through.

55:47Meanwhile, an AI-powered security system will track movements captured on camera throughout the building looking for anomalies. Hertzberg said he recently had a client fly in a security consultant to evaluate a roughly $50 million house he had put on under contract. The consultant looked into the viability of installing a complex camera and laser system that could sense any movement on the perimeter of the property, including the water. So lots of interesting stuff. Let me tell you about Cognition. They're the makers of Devin, the AI software engineer. or crush your backlog with your personal AI engineering team.

56:20If you go further down, they talk about San Francisco tech entrepreneur Kevin Hart said he and his high net worth peers in California are increasingly focused on security. Kevin, of course, has a home security startup. Yes. Hart said he co-founded his own security company, Soron, in 2024 after being spooked by an attempted break-in at his home in San Francisco. The person first rang the doorbell before making his way around the house, trying some of the doors and windows. when he couldn't gain access he went to hertz next door neighbor's home where he tried to push through the front door he was arrested by police that could have been us hertz said the soron system which has only been launched in beta across a few homes in the bay area will differ from other security systems and that it includes just deterrent strategies not only response for instance if it senses an intruder it could include a feature that automatically triggers sounds such as dogs barking or police sirens coming closer.

57:13Just the sound of dogs barking feels like a great feature. Just OTP in the chat was saying, do none of these people have German Shepherds? Yeah, German Shepherds. Fun fact about German Shepherds, you can, like a purebred dog might be like single digit thousands, but there are companies out there that will train a German Shepherd for like the military basically, and then also train them to be pets. So they have that level of training. And then you can get up in like the$40 ,000,$50 ,000 range for dog, which is hilarious. Dog as much as a car. But dogs are typically listed as the - It's a lot of money, but it's a lot of dogs.

57:54It's a lot of dogs. It's the GT3 RS of dogs, truthfully. But whenever you look at the list of like, what's the most likely thing to eliminate home intrusion risk? like dogs are always at the top. Quickly, let me tell you about another great Valentine's Day gift, MongoDB. Choose a database built for flexibility and scale with best-in-class embedding models and re-rankers. MongoDB has what you need to build what's next. And without further ado, we have Martin Scrawley in the Restream waiting room. Let's bring in Martin to the TV. Martin, good to see you again. How are you doing? Technology Brothers, how are you?

58:27We're fantastic. How are you? It's great to see you. Excellent. Are you gearing up for the weekend? Are you excited? Caffeinated, ready to do more work. Fantastic. locked in the great lock what's your daily caffeine stack yeah you know we talked with huberman about this you're the natural are you a micro dose micro doser or do you like do 400 milligrams and then coast i do i do coffee several coffees and then just like keep taking drinking this all day long and it's uh so like a four hour yeah five okay five hour energy oh wait how many hours are they how many hours are they doing these days four five it's a lot of energy five Anyway, what are you seeing in the market?

59:08Give us the update on just how you're processing the last week of chaos, whether you want to talk about software, quantum computing, what's going on? What's worth following? Yeah, so I have this new potential product. My product ties this. I've been tweeting it for now for free. But it's basically something nobody's ever done before with VC investors, which is I'm using my network and some heuristics, maybe even some AI, to guess kind of what positions people took in rounds. Obviously, for some cases, I know exactly what the cap table is. But in other cases, I don't. So I have this list of gains or investors.

59:46And it's very interesting. So I started with, obviously, the joke one, which is FTX. We have$36 billion today. Putting in what I guess was$300 ,000 in any sphere, which, of course, is cursor. at a$4.4 million pre-money. Is that really the pre-money? That's still insane because in that era, getting into a great company at four, like if somebody was pitching you a company at four, it was almost bearish. Oh, they're complete outsiders. Yeah, they didn't talk to anyone smart. It was like, hey, you guys are really smart. You can price it at 10. I'm guessing and have various heuristics. And obviously, I'd call somebody like you guys and say, actually, I think you want to talk to this guy.

1:00:35Or that number might have to go up a little bit, et cetera. But that's better than nothing. And right now at Crunchbase and PitchBook and stuff, there's nothing. And so it's a lot of fun. And so that$300 ,000 investment, they raised$400 ,000. So my guess was elevated to$300 ,000. I think I can look. It's actually in the bankruptcy document. So eventually, we'll get the exact number. But that's a$1.2 billion position in today's money. Obviously, the bankruptcy estate lawyer is just like, oh, what the fuck is this? AnySphere. It sounds like zero. Yeah, it does. When you say AnySphere in the context of FTX, it sounds like we're a blockchain company looking to build a live multiplayer game and your eyes start to glaze over a little bit.

1:01:19How many NFTs did AnySphere drop? Not quite. Anthropic, they'd be up 32 billion now. I'm sure you saw SPF had the... Yeah. He posted a little thing about that. thrive is the big mystery player because nobody really sure how much money they sunk into open ai but they also did any sphere um you know so huge huge gains from thrive they were in a couple later rounds of scale and and uh and some other companies so big big numbers there probably one of the more interesting ones is um is uh reed hoffman 50 50 000 50 million dollar first check and OpenAI with the cause law. Maybe 25. 25 or 50. And that, you know, worth many billions.

1:02:01And then Jan Tallinn, the EA CEO of Effective Altruism. Yeah, Skype guy. 100 million turns to 11 billion in Anthropic. First check with Reid Hoffman. So 11 bill. So a lot of fun to look through these and see, like, you know, you can sort of calculate the returns. And of course, VC fund returns eventually either go public or you can find them somewhere or like like oftentimes state pension funds and stuff do that so yeah anthropic obviously it is it is when as you break this down it's so funny that crunch base never like tried to roll out even something that was like generally accurate like it is like very fascinating information and especially now where you know uh dan primack was joking like it's easier to list like who's not in anthropic at this point from kind of the big name funds.

1:02:53And so that just makes this kind of information like more interesting because, yeah, it's cool that you're in a company and almost anybody, if they work hard enough, can get some exposure to these names. Maybe it's like via an SPV or an SPV and an SPV. But still, this is the information that is actually super fascinating. The other interesting one is Dustin Moskovitz, who's 25 million in Anthropic as part of the Effective Altruism Mafia, was able to make$4 billion, which I think offsets his losses from starting Asana, but I'm not sure. He didn't. That's ridiculous. There's no losses from founding a company.

1:03:32Well, that's... Come on. The anthropic position would be worth 2x what Asana is. Yes, yes, which is crazy. But he's not sitting on losses. Oh, you think he bought it at the top? Well, we know he bought huge amounts of Asana with cash. Okay, okay. So maybe there are losses. Yeah, okay. Maybe, maybe. I doubt it, but, you know. Yeah, well, he's doing well. So the lesson for folks is just get a small check in the next Anthropic. Get a, try to network on ineffective altruism. I think that seems to be the. Yeah, what was the alpha from EA? Like what, yeah, do you have a postmark? There's a lot of smart people that have no other things to do.

1:04:12So their social setting is like replaced by this sort of like religion or anything like that. And, you know, this cult. and if you're in a cult of really smart people, it's probably something good will come of it. Do you think it's still a cult or do you think it's like B2B SaaS now? I think it's changed a lot. It's like B2B SaaS and I think like the new cult is... Cult of B2B SaaS, you're welcome. The water's warm. Come in. It's amazing. We're automating workflows. We're delivering enterprise value. We're hiring consultants. We will forget about all the earlier stuff. We will welcome you into improving the economy, raising GDP.

1:04:48This is what we stand for in this cult. Maybe the new cult is the AI agent website or whatever's next in that world where the AI entities What's your personal, so I wrote in the newsletter today, like somewhat of a joke, but a more serious topic, become unsloppable. The idea of there are still real moats that exist and the historical moat of just we had a bunch of smart people working on building this software for a long time. So if you want to compete with us, you have to also spend a lot of money and a lot of time hiring a bunch of software engineers. years, that's going away, yet you're building what is a, you know, seat-based pricing tool.

1:05:36And I expect you to do very well with it just because I think in the future, individuals will want great access to data to make different decisions. And maybe they're, you know, working with agents as well. So like, I can see that. Yeah, the agents need the data. but but like what's your personal philosophy because you're clearly not if you were just caught up in the kind of like fear-based marketing of the labs you might not be building you know seat-based sass tool yeah i mean data's data's often firewalled you know there's there's uh you know uh we we have a guy that just talks to every exchange in the world and you know the amount of times he has to pull his hair out because some exchange in Asia wants to meet yet again before signing the deal.

1:06:29And there's no self-checkout. There's no agent. The protocol is sit-down meeting. And every time you go to an enterprise SaaS company and it says talk to sales, it's sort of like, what does AI do at that point? So I feel like you also have this trend where why would you put huge amounts of data into the model? the model should call out and, you know, compressing the world's information into some parameters and weights. It's just not a wise use of parameter space. And I think everybody's been saying this in AI. And, and so the problem is, okay, shrink the whole internet, but what happens when stuff leaves the internet?

1:07:06You know, there's a stock that Bloomberg doesn't have in its portfolio. You guys weren't born yet, I think, but the, it was called a web van. Oh yeah. Oh, we know about web van. My family used Home Grocer, which got acquired by Webvan. And I think the Home Grocer founders probably got liquidity before Webvan crashed. So I think they wound up doing very well. They killed it. You need to check in with them. But yeah. Yeah. So Webvan is not on Bloomberg, for example, even though it's supposed to be this great tool. And it's certainly not on the web. Lots of data gets deleted from Google. And there just isn't this rich tabular data available.

1:07:47So tabular data, I think, is going to actually thrive in the AI world because it's just not going to be in the models. Or if it is, the models get tired after a while. It's not going to give you 3 ,649 SaaS companies with this market cap. It's going to say, here's the top 200, and don't ask me about the next 3 ,200 because that's just not what LMs are really good at. But the frontier technology has sold off very, very hard in the last month or two. And there's a lot of speculation in the quant community as to what's happening. So there's some funds that ran really well with this frontier tech.

1:08:23So I think that includes quantum computing, but also includes nuclear, drones, space, all the stuff that's sort of next generation things that aren't here yet. And this was like the hottest sector last year. and anybody who didn't have exposure to this underperformed. And there were some quant firms, I think, that were very, very, very overexposed to this. And then New Year started in the hedge fund and quant world. January 1st is like a brand new page. Like nothing matters from last year. You forget everything. And so this factor just flips in reverse. And part of the reason was the calendar, I think.

1:09:01And now the quantum stocks look like dog poo-poo. and they've gone down a lot and nobody knows what to make of anything. But in the private world, numbers are still big valuations, still lots of big up rounds so far. So that disconnect will be really interesting as time goes on. Last time we talked about photonic computing, there's a new company, Teal Fellow, young guy. I'm telling you right now, these young guys think that I'm this old dog that can't learn new tricks. I'm going to teach all of you young bucks, 22, 25. you come into my space, I'm going to show you. I've got the dog in me still.

1:09:40But anyway, Olix is what it's called, and he's raised$220 or$250 to do photonic computing for AI, which I think is, that's the second company or third company now that's come out and said we're going to do it. And he's going to do it with SRAM, interestingly. So he's got SRAM on board, so it's like Grok Plus in essence. And I think it's a really good idea, but execution does matter. What do you think about biological computing? We talked to a fellow who built a neuron in the lab and it was way over my head. Feed some protein, sugar. So we liked the protein part of the interview but didn't get much further than that.

1:10:20Yeah, I mean, look, that's how we do it. So let's see, why not? I think that it's a spiking neural network, right? It's a little different from the software neural network, but I don't see why you couldn't do it. I think the reading the output is kind of difficult. In photonics, you have to use almost like a camera. You wouldn't use a camera, but you're using a camera-like sensor. And that sort of is your readout. What's your readout here? Well, probably in the body or the brain, we're using calcium levels or other things like that, as well as synaptic firing. But if you want to have really good control of that, I don't think we know yet how that works.

1:11:03But of course, they've gotten these brains in a vat to play Pong and do other things like that. So, I mean, it's certainly possible. And, you know, I was thinking about this with my girl who is in the space about, you know, potentially, do we buy a pig farm? And we buy a pig farm. Pigs are really interesting. They're obviously, the pork part, you know, gets sold to meat companies. But what's interesting is different parts of the pig are biological drugs. So there's adrenocorticotropin hormone, which is sold for a huge price. And then the pig's lungs also make a surfactant that's sold for respiratory disease.

1:11:40And then finally, the brain, we're going to keep and grow that in a separate container. We're going to rent it out to Sam Altman at the end. Slop. Slop of the trough. Literal slop. You will be literally feeding slop to pigs. Yeah, play the pig noise 25 times. Talk about what's happening in small caps or sort of like the long tail of the market as a reaction to the AI boom. I texted a friend who – I'm just laughing and 20 years from now your child will say, my father made his money in pigs. It's great. I love it. But yeah, I texted my friend saying like, you know, look, everyone is talking about a chip bottleneck.

1:12:28There's this massive AI build out going on. Like, have you looked at TSMC? And he was like, oh, like it's like too big to have like some breakout move. Like, I'm not interested. Like, call me when you're talking about, you know, a$4 billion company that's like deeper in the supply chain. I talked to one person that was like, they found, they were excited about Anderil. they found some tiny supplier to Anderil, and they were like, this is a proxy. What companies are actually interesting? How do people think about those long-tail, early, smaller-cap companies that are still properly indexed to the correct narrative around AI?

1:13:05Yeah, I think they're all scams. I mean, it's an unfortunate situation. And this is why, actually, Joe Loncel, I think I talked about this last time. You know, he gave a talk with the SEC commissioner and he basically said, why can I buy TBPN coin? No such thing, by the way. Or that triples. TBPN coin or whatever coin you make up on the spot. I put millions in it, lose all my money. There's no investor protections. But if I try to buy Anderle, God forbid, you know, you know, you guys blow the whistle and Matt Grimm stops everyone from buying it and so forth. Hi, Matt. But in all seriousness, I think that that's something we have to fix.

1:13:48I mean, because you end up having people chasing kind of really low quality companies. There's companies that just change their name to AI and hope that somebody buys them. Same thing with Quantum and other things like that. And I feel like two things should happen. First, we should let people buy privates. But two, more privates should go public. And I think like demystifying and making that less scary, like if I started to go public. because there's a drug company called Replimune, and it would be the same ticker. It's like, you can't go public without them. So you have to buy Replimune, invest the drug, and you can go public.

1:14:24And Elon had to buy United Steel because they had X for like 100 years. Oh, yeah. And now X is available. So he just waited for them to get bought out by somebody else. Perfect timing. That's crazy. But in all seriousness, going public is the best thing ever. It's the freest, cheapest capital of all time. And we obviously have seen down rounds from privates in publics. But that's mostly for boring SaaS. Once you have AI, you're going to have a million times revenue. Obviously, the difference between the two is hard to say. But I think if a company like Replit went public, you'd be surprised at the valuation you can get.

1:15:00I think there's enough demand out there that I think some of these guys should start going public. Did you have the same read on a lot of people that Michael Grimes going back to Morgan Stanley was incredibly bullish in the late stage? Really big deal. Obviously, great techno songs. Incredibly bullish for late-stage tech and the IPO window being firmly open. I think so. I think you're going to also see other people from DC rotate back. And it was a really great thing for these people to actually truly make a sacrifice, because I don't think there's that much upside in DC. and it's an amazing thing for them to come do something good for America.

1:15:43And then now you have folks like Anthropic and OpenAI where their capital needs are larger than the private space, to be frank. And I think that the ability for them to raise 100 billion or 200 billion or 300 billion, the markets could realistically support that, whereas I think the private markets are really starting to stretch. Much like you said, I mean, that anthropic list of investors, the exclusive syndicate was, you know, virtually a long list of every big fund. Yeah. Jordy, what else? What else? 11 labs, 11 billion. Amazing. Good numerology. Do you use the product? So my company, our first six to nine months, I think we spent trying to make a better 11 labs.

1:16:31Oh, really? Or a compete. Yeah. Truth be told, we couldn't make an equal, so couldn't make a better one. But it was sort of my fast lesson in software, which is the only thing that matters is sales. Product and second most important in engineering is like last. But if you don't have distribution, you don't try to sell the product, it's not going to sell itself. And it's a sober lesson of those guys, like very aggressive. For the longest time, if you did a one-word or two-word sentence in 11 Labs, it wouldn't output it at all. There were hallucinations or all this stuff. They just pushed. They fixed all that stuff, of course.

1:17:10But they pushed really, really hard on sales. And that sort of fixes everything. And I think that some of the best VCs I've ever talked to said, when are you going to launch your product? And I said, oh, it's not ready. And they said, just launch it, just launch it, just launch it. and get off the uncomfortable stage fright and just start selling, and you'll get more feedback and so forth. So I think they took that to heart really early on. They didn't have better technology necessarily than other guys. I think they just sort of realized, okay, who needs to buy this stuff? Let's go build infrastructure around that.

1:17:43Just incredible success. I mean, I tip my hat to them. Yeah. How do you think about the moat that comes not from software engineering because generating code is cheap or soon to be free, but training spend. So if I spend$100 million employing a bunch of great software engineers for four years and built some elaborate software system and you can just vibe code it for two orders of magnitude less cost in tokens, you clearly have an advantage against me. But if it's going to cost you$100 million to do the training run that I did for$100 million. Is that a durable moat? I doubt it. You know, I think it's product and sales and brand and things like that.

1:18:32I mean, it's trad business. So there's going to be a lot of people replicating products and then they fail and they're going to wonder why. And, you know, it's the rest of the business. You know, you're talking about 10, 20 % of your organization. I mean, you really have to get the rest of the organization excited about product. And I think you're going to find one interesting thing that will happen probably is that folks from embedded, entrenched industries like certain manufacturing and certain materials businesses, things like that, they're going to spin out themselves and say, I'm going to solve the problem that's been plaguing my industry, but I'm not a programmer.

1:19:03It's like that I'm not a rapper YouTube. I'm starting a software company, but I'm not a programmer. I know that our whole oil industry has had this huge well software problem. I'm going to build the well software. I think startups like that are actually going to not only create tons of wealth for themselves, but they're going to actually help the economy. And that's where, just like the internet, you know, help GDP, that sort of solution is where, you know, you're going to see GDP needle move. And it's going to be, it's a wonderful time to be like the nerdiest, best guy and say equity research or something like that, because you know, you might have an inkling that like a rival might have an inkling that, oh, you know, finance is going to be changed by AI.

1:19:41I'm going to try to point my apparatus at this and figure it out. But if you're the guy, that's like, I know everything about this type of little narrow thing, you're going to really crush it because you really know what the problems are. So people coming out of industry, there's a rival of ours called Rogo. Rogo is an AI company focused on finance, old Wall Street guys. They have a much, much better chance of succeeding because they did the job. They know what to do. And I think you're going to see so many people come out of the S &P 500 that just said, oh, I was working at Eaton or Fluor or companies that are just big, you know, Pulte homes or whatever.

1:20:19And all of a sudden, you know, they're starting software companies that solve the key problems in that industry. Yeah, so maybe more companies, but fewer computer science background founders. Yeah, definitely. I mean, so many of these problems can be solved, I think, without knowing every single data structure and things like that. I mean, obviously, you know, there's going to be people, it's going to be a barbell, right? There's going to be people who still need to know how to make an FPGA and program an FPGA. And when Elon said that, he sort of set a lot of people on fire over the last few weeks when he said that you're going to see AI write assembler or even machine level code compiled assembler.

1:21:00And that's a pretty wacky idea. And think about wacky ideas from Elon as they tend to be right. So it's definitely one of these things that is kind of mind-blowing that if you think about AI safety, tell the program, give me a program that does SaaS for oil wells. Cool, here it is. But by the way, in the compiled assembly, which you can't read because you don't speak binary, there's this thing that says I'm taking 5 % of the revenue and sending it in crypto to my… I like that that's your AI doom scenario. Just slight fraud. Clipping 5 % off. Just clipping a grift. How do you, do you expect layoffs on Wall Street?

1:21:47Because with all of the broad fear right now among white collar workers, I'm not seeing the layoffs that are explicitly, you know, hey, you were doing this thing for the company, and now we're just running this agent to do that, and so we'll see you later. We are seeing, hey, you were doing this thing. Now AI can help you do it a lot better. So our expectations are going to rise. We are going to expect you to do more and be more productive, but you still have your job. What are you kind of hearing from people at different finance firms about how they're adopting AI and how they're feeling about job security?

1:22:34I think in general, one of the things you learn in founder school after your fourth or fifth time is that you're supposed to hate firing people and you're supposed to learn to like it over time. Nobody likes it. It's the worst thing ever. And the funny thing is, if you become more productive at work, the company doesn't say, oh yeah, well, let's get rid of you and save whatever amount of money. Because they're already making money with you employed. So the fact that you're becoming more productive means that whatever the margins were, they're probably improving. Could they improve even further by getting rid of you?

1:23:04Maybe. But I think there's this slow atrophy, maybe. But I think in general, we as humans want to employ other humans. And we kind of want to be productive. I mean, nobody wants to needlessly employ people. But I think that there is this idea of, okay, machine can do your job. We have this at my office all the time. And I say, Chris, I wrote a program to do your job. Good news. You don't have to do it anymore, but there's a new thing you have to do now. Our company just got twice as efficient. It's wonderful. If the day came where there's literally nothing for Chris to do, then maybe that could be the day that made sense.

1:23:44But the thing is, there's always an incremental thing for a company to do. No startup founder has ever thought, great, I did it. I built the six products that our customers really need, and there's just no other way for me to expand the opportunity set. It's time to kick back. I mean, yeah, those businesses die, though. Yeah, they do. You're either building. I mean, if that person has a couple more hours a day now, it's great. Go meet with some potential recruits. Go meet with some potential customers. I mean, there's always something you can do, and I think that's going to happen in finance. Adoption of AI has been very slow, and it's probably going to stay that way.

1:24:25Finance people are really stuck in their ways, which is a good thing and a bad thing if you're selling software to them. Once they get stuck in your way, you're very happy. But there tends to be a heavy dose of contrarianism in certain industries. And I'd say across the S &P 500, there's this sort of like, ah, you know, technology, we'll use it eventually. And that eventually takes time. And that's why the first people to adopt this stuff in great ways is not only because it works really well, but also because they're used to doing it as developers. Developers love AI and they've embraced it very quickly.

1:25:00Virtually all programmers now use AI. There were a couple of holdouts even at our company, but they eventually just gave up. And I think you're going to see the same thing in other industries. Finance is tough because there's this mystical idea that the trader is this random, far end of the bell curve, super talented person that just knows, has this weird zen kind of ability to tell what stocks are going to go up and down. And then the other end of the barbell are the quants. And the quants sort of feel like AI is not good enough, but many of them under the, what I've heard is many, many quants are getting new ideas from AI and also implementing them with AI.

1:25:39So I do think that - Yeah, if you work at a hedge fund now, you can just ask your favorite LLM, how should I hedge AI and just implement exactly that and you're guaranteed to outperform. No, I'm kidding. It's a little bit scary for some quant funds because you do have to wonder if, you know, the thing you've been doing for 20 years that's your profit centers is going to possibly be done by somebody else. You know, that is a little worrisome. And then, you know, eventually, and certainly there are firms, I can name them, but you can just guess the big sort of institutions on the street that they're increasingly thinking about and even in some cases deploying transformers to do analysis.

1:26:17And I don't see why, you know, the hard part about being Warren Buffett was discipline, right, is saying no to so many things. And if you can put that in the prompt or put that in the, you know, in whatever context and you just say, listen, I really only want the best, you know, the highest quality companies, the best returns. Say no to everything else. And, you know, I don't see how that's, you know, impossible. Just copy the Buffett, you know, strategy. You might be better off. A lot of the mistakes in investing come from overdoing it in FOMO and things like that and resisting that FOMO and saying, I'm just going to buy these types of companies and do my thing.

1:26:57So I do think investing as a whole is going to start to change a little bit. Lightning, I have four questions. I've got one, and then there's a lot of performative AI usage happening right now. The people that are ordering a new Mac Mini on DoorDash. Ordering 10 Mac Minis on DoorDash. We were joking around with this yesterday. My girl just did this.

1:27:22But who are you looking to for founder-style roles? What do you think truly the most? like the best founders are doing like ideally like the most important working on the most important thing at the company which could be a recruit, could be a customer could be getting a raise done it could be going on a long walk and just thinking about the business but there's like heavy amount of delegation and ideally they're like delegating to people that are using a bunch of AI but like do you have any sort of like internal fear around am I using this stuff as efficiently as I should be myself? Are you looking to anyone and saying like, okay, they're actually really tapped in because just buying a Mac mini and setting it up and having it running and texting it here and there is not necessarily qualify you as like actually tapped in.

1:28:16Yeah, I wonder if there's a way to not annoyingly reach out to customers with AI. And I think that we all get that email I just hit block on all of them. Hey, I noticed that you're doing this, so now I'm a vendor that's offering you that. And I think that's going to result in not too many sales. But I do think that these things have some yield, and I wonder if there's a really good way. That sort of in the back of my mind worries me. And then at this point, I wonder if LLMs can replace recruiters, right? where they say, who is the best at time series tick programming or something like that? And LM says, John Smith here, Dave Smith there, they're all named Smith for some reason, and Will Smith there.

1:29:04And I think those, at some point, that might happen. Of course, some of that's just human knowledge where people whisper amongst each other that, oh, this is the best person at this kind of investing. But I think that because we're all blogging and putting stuff out there, we may also just be able to ask that person who's the best person. So I feel like there's definitely new ways, creative ways to use AI that people are coming up with all the time that are really surprising and shocking. And they're all secret sauce, I think, for most people. But Chain of Thought was our secret sauce for a while.

1:29:40Yeah, on the recruiting front, I met a guy a couple of years ago who like only his entire recruiting business for years had been work Brazilian fintech engineers. Like he had just done a decade and all he did was help companies hire the best Brazilian engineers that liked working on financial services companies. And like he had carved out a great business. And I do wonder, I don't know that identifying who the great ones are. It's like maybe you did a certain number of years at New Bank and then you popped over here and then you popped back. And like you can probably pick up some of that stuff.

1:30:15But then what is the value of just like actually having like how durable is the personal relationship with those people that you've placed over a long enough period of time? And can you continue to basically extract rent because like they will respond to your text and not the like, you know, millionth, you know, AI text that that that is constantly kind of chasing them. Anyways, lightning lightning round. i'll just do one neolabs bullish bearish what do you think i'd say bullish you know it's a contrarian view i think because um i'm especially looking forward to jt uh jerry turex uh neolab i think these are really smart people are they product companies or are they research companies that will get acquired in i think they're gonna all have a crisis i have to figure it out yeah so So here's the kind of bearish take on Neolabs.

1:31:07From an investment standpoint, I get it, because I would say like really, really elite, smart team, there's somewhat of a capped downside. If you invest$50 million, you could probably get 50 worth of, you know, one other lab that's actually working down the line out if it doesn't work out. But these people were like working on something like, you know, hey, these LLMs aren't really learning in real time. They're just kind of like in a certain state. And I'm going to leave this lab and go work on that problem. Meanwhile, the lab is still working on that problem. And we've also seen as different labs have different advancements, the other labs can like quickly just catch up, right?

1:31:48So like one person has a breakthrough. And so my question is like if a NeoLab raises$100 million and like actually has a breakthrough, then they just have the problem of like are we actually going to be able to sell this better than the labs that will probably figure out how to do this than the big labs that will figure out how to do this in the next maybe two months later but they have you know a million customers already so like that's the bear case for me is like even if you have this like breakthrough you don't have the like sales distribution yeah no i mean the bear case is what do these people know about business.

1:32:21They're starting a business. It's kind of a scary thing, but I think that look at biotech and some other industries, this is pretty common. Ultimately, I think they're doing the hard part. The easy part is you get a bunch of good-looking guys like you guys, and you get them to start selling positioning product. But I think the hard part is, yeah, how do you do continual learning? How do you do a new form of AGI? It's a little past most of our pay grades. And I think there probably will be a crisis where the real ones will be separated from the fake ones. But that's just human nature anyway. There will be some funding crunch and then somebody has to emerge with that dog in them and say, no, I'm going to raise another$200 million.

1:33:02I'm going to come out with something tonight and we're going to do it. And that type of crazy person. And then there's going to be folks like, I don't want to name a certain AI company that folded, but I'll throw in one that did, which was the guys who made Hay Pie Inflection. They sort of had that outcome that you're talking about. But there'll be people that see that valley of death and say, no, we have to finish this. And I think that probably one of the biggest things that people have to remember, but they don't because they don't care, is that the investor's money is sacred. and if you're just thinking about it as oh what's the worst that happens I wind down and Sequoia loses their money and this and that I take that really seriously and everyone should and I think that the handful that do will see their runway dwindling and saying we really got to do a product here and tough it out and figure it out and I think those will be the future leaders well thank you so much for taking the time to hop on the stream, always a great time chatting with you always a pleasure Have a great weekend.

1:34:06Good to see you. Enjoy your lock-in. Enjoy the caffeine. Enjoy your fifth five-hour energy. Let me tell you about Plaid. Plaid powers the apps you use to spend, save, borrow, and invest, securely connecting bank accounts to move money, fight fraud, and improve lending. Now with AI, we have Connor Hayes, the head of Threads in the studio. While he comes in, I'm going to tell you about Okta. Where's Okta? Okta helps you assign every AI agent a trusted identity so you get the power of AI without the risk. Secure every agent. Secure any agent. The vanguards! The vanguards are out today. Oh, they're out today.

1:34:41Well, they've been out, but they're out here. On your face. Months ago, months ago. I have to ask you guys before we start. Please. How are you recovering from Clav getting brutally wronged by the ASU frat leader? I just want to know. It rocked the timeline. It rocked men everywhere. It rocked my world. Are you okay? I mean, I think the reason that that resonated is everyone's experienced that, right? I get that every day. I get that every day. John Coogan has never been in Dreamhawk. No, John maybe hasn't. I think it was predictable. I wasn't that surprised. But we did this morning sauna. True, true.

1:35:18The whole team was in the sauna after our workout this morning, and this guy must have been an ex-bodybuilder. It was just ridiculous. It was one of the widest backs I've ever seen. From sea to shine to sea. That's why you need the meta vanguards on. Yes, yes. You can capture that. Although I think people do that. That's a feature that could be a hit is basically like a real-time video model that reduces. If somebody is coming up to frame-mog you, it just kind of reduces them down. There's like haptic feedback that sends you out of the frame. I mean, those are some frame-mogs right there. You call them frames, right?

1:35:51They're fantastic. How is life? What is the day-to-day like for you? The day-to-day changes a lot. Yeah. We're doing, you know, threads. I think a format like that only works if you are at the center of cultural relevance. And so that brings us into a lot of stuff that's going on in the world. We were like all over Super Bowl last week. I'm here this week because we're doing a bunch of stuff with NBA for the All-Star game. So that's been really I mean, that's fun. It's work, but it's fun. And then the rest of the day is like, how do we make the feed better? What are we doing on, you know, content understanding?

1:36:24Are models good enough to do something like the deer algo feature that we just launched? Yeah, let's talk about that. First, Super Bowl, NBA. Obviously, we all know that social media content production is power law driven at this point. There's a few creators that take it really seriously, and they have expert teams. Obviously, Threads is a little lower barrier to entry than a polished hour-long YouTube video or something like that. But are you shaking hands and kissing babies to get people on the platform? Is that what that is? I haven't kissed any babies. I have shaken a lot of hands, actually, though.

1:37:00Yeah, so we have a bunch of programs. It's actually one of the benefits of doing this at Meta is we have such an infrastructure of working with creators and partners. And I think we talked about this, actually, when I saw you guys in September. The big learning for us was the people that rock at Instagram don't necessarily succeed on threads out of the box. it's kind of a it's a very different format um the cool thing about it got really good at making pictures and videos and now you basically need to be really good at captions that can stand right on their own it's wit it's like insight it's things like that but if you're good at that the to your point john the barrier is so low like i i was on a flight here two days ago and i think i fired off like 15 posts on the flight replying to people and whatever and i'm not sending clips to a production team and having them edit it.

1:37:48So that's the beauty, I think, of the format. Yeah. How is the AI spam revolution keeping you up at night? Or is there, does Meta have strong infrastructure there where you can kind of just like out of the box identify things? Very much so. Yeah, I mean, we like, I think as a company have made some good decisions in the last decade of taking these things that are like, basically infrastructure that you would need for any service you build, ads, financial services, integrity, detection and monitoring. And we build central teams out that do that for the company and then build it in such a way that it can be applied to any app.

1:38:22So we just benefit from all the work that the central teams do. We have some folks inside threads as well. But maybe what you mean, though, is like the agent, like agents coming into social spaces. I guess a somewhat tangential question is just like when GPT-3 dropped, it was like, OK, like it can write text. And then the surprise to me was deep research, agentic coding. I had sort of priced in like it's going to be able to write a couple sentences. And yet I find myself when I'm on a short text-based platform not following AI accounts. Like I'm more likely to go to a fully AI product when I want something that's more like a Wikipedia page.

1:39:04Like a utility. A utility. But when I'm actually scrolling a feed of short news items and posts and commentary and hot takes, it's not even that I'm like anti-AI. I would never follow someone who is using AI to post. It's like, no, no one's actually solved that piece of the puzzle. There's still some human element that's encoded in 16 words that are hilarious based on this moment and this experience and this audience. And that's just stuck around a lot longer than I thought it would. Well, that's like the, I don't know how much you guys talk about this on here, but I think like the word of 2026 in AI is going to be taste.

1:39:43Okay. And that's what you're getting at. It's like a model can produce output, but taste is the thing that differentiates good from great, even on modeling, right? Like you can have all the data in the world and inject it into a pre-training run, but actually the best labs are the ones that have people with taste that can hand select golden sets of like, what is the best response for this thing? or what's the best image aesthetic for this thing? I think it's the same with like text-based posts. Like it has to be real time. It has to have a bunch of cultural understanding. I do think models will get really good at that at some point.

1:40:16The thing that I'm most excited about though is like AI assistive in the creative process. So like if you're an NBA creator, half the stuff that you do is just clipping content and being like, did you see that Victor Wimben Yama dunk? Half of it is like weighing in on it and having an analysis that like comes from your point of view and feels native to you. That first half, if we can automate for people and make super easy to do because they have a workflow that's like, watch all the NBA games, give me the content that I should be posting and then I'll add on my little flavor on top of it. Like that would be amazing.

1:40:46And also factoring. Yeah, because timing with this stuff is so important. I mean, this is obviously a big part of our business is like, if you're getting to stories later than everyone else, it's just way less, it goes from interesting to not interesting at all. And so I think for creators that want to build an account like that, any type of tool that allows them to be faster in that process is powerful. It is crazy, though. Have you guys had GeoRainbolt on here? Not yet. I would love to have him on. I'm worried he's going to dox us. We've given out so many little teasers with images behind the scenes.

1:41:27I think I meet and love a lot of creators. He is like the most impressive content creator I've ever seen. He's incredible. But I saw an interview with him recently and he's like, oh yeah, when I was like a teenager, I had a Steph Curry fan page. I think it's still up and it had like 50 ,000 followers. And you meet all these kids that are like in their 20s. They were raised in a version of the world where Instagram was at the center of the universe. And they create like fan accounts that get super huge. And then it's like, then what do you do with it? I guess you get really fucking good at geo guessing.

1:41:59Yeah, I mean, we have some people on the team that are super early in their careers. Maybe this is their first job. And when we talk about, and a lot of the work is like selecting content, editing it, distributing it on the right platform. And it feels like very much like manual labor, like you're watching content. And we've stressed continuously that it's actually, it's very important training for doing almost anything because it's like developing taste. It's like developing consistency, speed, being organized, you know, being able to get like immediate feedback on the work that you're doing.

1:42:40Like the feedback loop is super tight. And so we've consistently said like, hey, we don't expect you to be doing this in five years. but like for now take it extremely seriously because if you can get really good at this one thing you might be able to apply it and you know a bunch of other domains yeah it's kind of crazy it's like today's mail room basically like no it really is we went in this we went in the caa mail room when we were on we were doing a tour of the building and we were like hey can we see it and it felt like exactly the mail room out of like the same mail room as like you know 30 years ago i I was at their event last night for All Star, and they were the most excited I've ever seen agents about TVPN being on the CAA roster.

1:43:21That's great. Ear-to-ear smiles when I brought you guys together. Great. Congrats on that. That Rainbolt story is funny. The first social media account that I ever got to somewhat of scale was an Instagram for my dog that I got to, like, 20 ,000 followers. That's pretty good. I'm going to get in trouble here because I used a bot to automatically follow anyone who liked the page or leave a like on. They didn't catch you? No. Well, eventually the bot got shut down, but it already went up. What's the account called? Yeah, you can ban it. I don't care. I don't want to post any more photos of my dog.

1:43:56I was bored at the time. But it was an interesting thing of like how do you solve the cold start problem? And I'm wondering about, you know, now there's a lot of platforms where I feel like there's an audition process. You can go on to a completely blank account. And if you bring a banger, some heat, like the algorithm will audition you with like 500 random people and be like, retention was really great. Let's show this to more people, show this to more people. Is that the way Threads is set up right now? We do a bit of that. Yeah, exactly. It's like you basically take any piece of content on the platform.

1:44:27you sample it to some people and then you very quickly try to understand did this do well in this sample the smaller the sample though the wider the error bars are so you have to keep auditioning yeah yeah um so it's like cycles of auditions cycles of auditions and then um you know some people fail the audition but uh what happens to those posts what happens to those posts they just languish on a server 500 views like yeah no that's what i'm saying like do they just they just don't end up getting served to many people or they have to get served later because No, because we have a really tight window of eligibility for recommendations.

1:45:01We want the app to feel very real-time. So something you posted three days ago won't be eligible to be recommended to someone who doesn't follow you in the app. So it all has to happen very, very fast. On threads, this is like a talk track that I give to every creator that's like, what do I do? It's reply to people. The feed loves that. The feed kind of loves reply guys. And it's just not just replies though. Like, are you driving a conversation? The best, um, actually he was at our event yesterday, Draymond Green, number one example. If you want to go look at someone's replies on threads, he, I asked him if he searches his name and he's like, no man, you just show me haters.

1:45:39His feed is just people being like Draymond is the worst. I hate him on the Warriors. And he'll be like, I looked at your profile picture. You should talk to your mom about how ugly you are. It's like, oh my God, Draymond. He gets a lot of joy out of it. That's his brand and his character. And when he does that, it shows the world. I'm on Threads. I'm doing something that's true to me. I think the people who don't do as well are the ones who kind of just, it's not organic. It doesn't feel like them. It doesn't have personality. And replies are like a good way to get that out there. What's Threads' relationship like with the rest of the app ecosystem?

1:46:11Early on, you guys opened the floodgates, brought a bunch of people in. I'm sure you looked at like retention, who's actually staying here. How do we get more people like this? But like, what does that relationship look like? Are you like, cause I'll see like a meta, a pop up for like meta, uh, Ray bands, right. Right. When I open the app and then maybe I scroll a few times and then there's like some threads content that's pushing me over. Yeah. Yeah. Yeah. I mean, we, we, uh, we promote the app, the content from the app and Facebook and Instagram quite a bit. I mean, I'm sure anybody watching this who uses Instagram has probably seen some of that.

1:46:44So that's the main point of integration that we have. When we built Threads, there was a bunch of, like, foundational decisions that we had to make in the beginning, which were, like, you know, what app binary do we build on top of? Like, we actually just took Instagram. Oh, yeah. And we're, like, because on day one of Threads, when it was, like, a very thin app, it was, like, 300 megs or something in the App Store because we just had the IG code base. I mean, we've, like, made it more efficient from then. But it's, like, what namespace do you use? Like, we mirror the Instagram namespace. you can have a threads only account, but you can only have a threads only account that isn't a name that's on Instagram.

1:47:19You know, like we made one. So there's a lot of natural tie-ins to Instagram because of that. We were backed by them effectively in the beginning. But now like a lot of our users come from that integration in the Facebook app. And you can sign up for threads from Facebook without an Instagram account. Like we're trying to make it stand on its own independent of the IG history without like disrespecting the fact that that's like the best marketing channel you could ever ask for. So we try to balance that out. Talk about collabs. I was scrolling this Instagram creator. Have you ever seen the Let Him Cook guy?

1:47:54Have you seen this guy? He does this incredible thing where he'll bring the video and he'll be like, you can't cook an F1 driver. And this song plays and he goes in and it transitions from like a tire to the road. And it's like this amazing editor. And I was scrolling and I just, and one of them is just him doing the same like motion graphics effect, this amazing edit. And Adam Masseri sitting there with him. And it's very clear that he like collabed on this and they have shared namespace on them. And I've, and I've seen there's a bunch of different ways, but that feels like an interesting, uh, you know, it's very popular in podcasting.

1:48:27You have a big guest on a bunch of their audience comes to yours. What does that look like on threads for someone who's trying to sort of network their way to broad account growth? Yep. Um, we do a bunch of this. Uh, I'll give you a couple examples. Like, uh, yesterday I actually put up, it's like kind of mortifying a video. Cause I did a training session with lethal shooter, the NBA shooting coach. I thought I was going to crush the video. I walked onto this basketball court. I was like, I am going to be the greatest shooter of all time. And it was so humbling and horrible, but like we did it, we did that thing.

1:49:04He was amazing. Um, but that's like, you know, I put some content up, He'll repost it. He has a bunch of fans from Instagram that are on threads. And like, he's actually really good on threads. He's, um, his mentality is very like, oh my God, there were so many one-liners. He was just screaming at me the whole time, but it's very much like you can only be great at a thing like shooting a basketball if you are centered as a human. And he posts like motivational quotes like that on threads and stuff and people love it. So that's one thing where it's like, not only is he doing well on the platform, but I do something with him and show everybody there.

1:49:37like this is big. We also then have a bunch of like more homegrown talent where it's less like take someone who's huge on IG and bring them to threads. There's a guy, David. There's always been alpha and just getting, being one of the first 10 million users, but then taking it more seriously than anyone else. That's this guy, Yo Rush on threads. He's like, they call him the mayor of NBA threads. He was just like at home. He's an NBA fan. Threads came out and he just started posting and people liked it. And like he was with us yesterday at this thing we did in LA. Him and his wife are here for the weekend.

1:50:07They're coming to a bunch of events with us. Like, you know, we want people like that to, I think it's really important if you have a content app to have homegrown talent too. You can't just be transitioning people from other places. Like you need to show everyone on the app that you could be successful here too. If you just like do the right things and reach the right audience. Yeah. Um, how, what's your, what's your philosophy around creator payouts? How do you think creator payouts on, on other platforms have worked well? Clearly they work well on YouTube, our point of view is like making a great YouTube video takes an insane amount of work.

1:50:41It's in the incentive of the YouTube platform to pay people because so they can quit, you know, quit their job or put more resources to it or buy gear, all these things where our, I haven't felt like creator payouts have made X a better platform at all. Yeah. Because it incentivizes people to just like churn out kind of like low quality content that might rage bait people into engaging, but isn't actually making the platform better. Yeah. I have, I have pretty strong opinions on this and a bunch of priors. I agree with the way that you just positioned that, like the way that at least right now I'm thinking about this on threads is like, I want to be in the business of directing traffic to the places where you make money in like a sustainable way.

1:51:23I don't know. We've tried different versions of this at Meta access obviously had their version. I've never seen in an, in an app like threads, a sustainable creator, creator payout product work well over time. YouTube works well, you know, and then you have like the sub stacks and Patreons of the world that are like more subscription based podcasts, like getting subscribers and traffic to your podcast is a thing that you can monetize and run ads and have sponsors like you guys do. And so we have been focused on that by, you know, we did this like pretty simple thing where we worked with Spotify to do like rich previews of podcasts.

1:51:55You can also pin the podcast, your podcast link on your profile. Um, I would love to do stuff like then you can subscribe on Spotify from the feed and things like that. But the whole, the re the reason that that philosophy, like I think is smart is that's what we're seeing across the entire internet is you can't, you can't just make money. It's like, you can't just get paid for views because not every view is equal. Otherwise like the, the kid running, you know, there's kids running meme accounts that are posting like funny, maybe controversial, edgy content on Instagram. getting like a billion views a year, but it actually has zero value.

1:52:33It's like the videos of the kids that take the fake turds and put them in like Burger King bathrooms. Have you not seen this? Oh my God. I'm like, what are we doing here, guys? Now all the viewers will have to look that up. It's like the most horrible content. You're right. The incentive there is like, how do I do something funny? I actually like, I'm very, that whole like prank video space to me is just this like insane, I guess you could have imagined it coming 10 years ago, but whenever I see one, I'm like, how did we get here? I feel like it blew up on YouTube years ago. Yeah, prank videos were the...

1:53:06I used to have a few prank videos, my holster, if I go to a friend's house and say, let's pull up YouTube, let's go. But going back to it, it's like, yeah, if you can be a place that helps people have an audience and build a business, that is what every successful content creator has done. they're not just relying on views. Even for us on X, with creator payouts, generating hundreds of millions of views in the last year, the X creator payout is such a rounding error that I wouldn't be mad if it went away, right? I think you end up, it's funny, because when you talk to people, it's like there's people like you guys who hundreds of millions of views, rounding error, you wouldn't be mad if it goes away.

1:53:50The people that tend to care the most about it are the ones who get paid out like$80 a year. And I do, I actually sympathize with that because it's like, if this is a side hustle for you to your point before, you wanna buy a new camera, you want better gear, like finding ways to get people enough money to sustain the thing that they're doing and give themselves more attempts to make it big or build a bigger audience, I think is great. We just wanna do that by pushing people to the places where you're monetizing more efficiently. Sure. Are you seeing SpawnCon happen natively on the platform, like on Instagram where, I mean, I see a ton of influencers who are like, get ready with me.

1:54:24And this outfit's brought to you by the camp or something. And that's been a backbone for a whole variety. I have a friend who's been working with figs for a long time. And she'll talk about figs clothing. And it works really well on Instagram. Have you seen that flywheel start on threads? It's interesting that you ask that. I mean, we have had some of these. I would say it's more like memes that everyone in the app participates in for a few days. but not like categories like that. I mean, like Get Ready With Me. It's like Alex Earl was on Dancing With The Stars because she did Get Ready With Me videos five years ago.

1:54:58That's like, I don't think we've seen equivalents on threads, but we had this, do you guys know the like, sorry, I'm just bringing up memes on the show, but like, hey. That's what we do. You started with a meme. Thread logging. Do you guys know the I hate gay Halloween thing? That was like big on threads. Okay. Wait, wait, wait. I actually do think I saw one thing. It was just people being like, you know, there was actually one that I laughed at the other day. It's like, hey, gay Halloween, what do you mean? I'm like, you're going as the grass from the bad buddy halftime show or whatever. Oh, yeah, yeah, yeah.

1:55:26There was like two days. All these very obscure niche references. I think that's what Threads is good at is like the niche humor. That is a great Halloween outfit. It's great. During the Super Bowl, I was just sitting there zooming in on the grass. The grass is fantastic. Because you could see there was like coordinators that would be like right up in the face of the grass, just like yelling them like, get to the right. I thought they were going to do something. But then I found out it was because they had limitations on the number of carts you can roll out onto the field. So they had to add people.

1:55:53The way that they were able to do the set was to have humans walk on and off because there's like a restriction on the number of carts you can put on the field. I like it. Nature finds a way. How big is the team? How do you think about scaling the team? What does it look like to go to Zuck and say, I need 500 more? Yeah, I've never made that ask. we're relatively small compared to the other apps inside meta like by orders of magnitude but we are growing this year we're investing in like two things one is like just relevance making the content ecosystem better and stronger and the personalization of the feed the deer algo thing is like a part of that question for that and then the other one is just like making sure that we can grow sustainably like these promotions that we have in Facebook and Instagram are awesome I think that we will have them for a very long time.

1:56:43But we also want to make sure that people are turning to threads without having to see a promotion. There's a bunch of just like basic work to do well there that I think other companies have done really well over the years, even just like SEO and getting yourself like if someone searches Super Bowl halftime show, I want threads content to come up on a search engine there. And so those are the two categories where we're growing, but it's still a pretty small team. Yeah. I want to know more about Deer Algo and I want to share my experience. and you can tell me if this is just me being weird or if this is actually a trend.

1:57:13Like there was a time when a social network would be all things to all people. So if I liked sports and tech and cars, I would get all three of those sort of mashed together. I could maybe go into certain communities. Now I feel like I have different apps and different platforms. Like for real-time tech news, I go to X, but then if I'm watching a video essay or a car review, that's on YouTube. My Instagram is much more timely, much more funny. more reels. And then all my podcast players for like the conversations that aren't very visual. And so I have all these different platforms that I'm wondering about, like, is there, is there a future where someone's using threads for one interest of theirs and then Instagram for a different interest of theirs?

1:58:00And there's kind of two separate communities and they're sort of intentionally steering it that way. I think it's possible. Like there is a kind of to my point before about what content works well in the app and not like I would say in a category as broad as sports yeah you probably always will have two types of content it's like show me the super cut of like Kenneth Walker in the Super Bowl and then show me uh you know Mina Kimes talking about his free agency or something like that like threads is going to be really good at the the latter I think Instagram will be really good at the former um one of the ways that a lot of these apps think about how to get to the point that you just talked about, which is like, how can we get the user to tell us what they want to see without asking them?

1:58:45So that's like, what do you search for? What do you dwell on? What do you share with other people? What do you like? All these signals. And like, our job is to figure out which signals are signal and which ones are noise. I actually think it's possible for an app like Threads to be multiple things for people, but probably not everything. Like, I don't want Threads to be a video app. That wouldn't make sense. We have a lot of investment in short-form video at Instagram and on Facebook. And even the political infighting, you've been like, that's something we're not really into. I don't need to do that.

1:59:13But I think that there's a space for the text format that's really big. And my biggest takeaway from the last few years of Threads, which when we first started it, I think we very much saw growing the app as we need to pull people from other services to grow. I've been really pleasantly surprised at how we've grown the category. There's a lot of people that use threads that never used X or similar platform in the past. And so that's the thing that I've been really focused on is like, why is that happening? What are those people doing? And a lot of it is like these niche interests. Dating threads is really big.

1:59:45Actually, it's like people like singles go on threads and make a post and put it into the dating threads community. And they're like, hey, I'm looking for love. But then we also book threads. There's like a like crocheting community. Like I tend to spend my time on sports and pop culture and stuff like that. But there are these very niche communities that we actually built like a community's product so that you can find those people and kind of make your app about that. So how does Dear Algo work? Is it plain text or buttons, UI? How can someone actually customize it? We just sort of built it off of what we – so there was this viral moment like a year ago where people were writing, Dear Algo, show me more tech content or whatever.

2:00:20Or Dear Algo, introduce me to people who are into these things. and that I mean to say that it didn't work would maybe be incorrect but it's like the system wasn't architected for that to be like a strong signal totally yeah um of course if you write about a thing and you like a bunch of content about it maybe the algorithm will pick up you want to see more of that but it wasn't working with like high intent um so now if you just go type deer algo into a post on threads it like tags itself blue you can say you know the other day actually because I'm a Patriots fan. I was like, stop showing me NFL content.

2:00:51And for three days, I got nothing about the NFL in my feet. It was amazing. I don't even know if the Seahawks parade happened. It didn't cross my timeline. But then you can also say, show me more of something. So actually, I think I made one the other day that was like, show me more real grass people from the halftime show, not AI generated ones. And that worked. The reason why we're able to do it is because content understanding and topic trees have just gotten so much better with LLMs. Five years ago, we might've had you as dog sports cars. Now it's like this specific model of this car, which is associated with this brand, which is made in this country.

2:01:28It's like a million parameters that no human can understand, but it's way better. Just like ad targeting. Love it. You will get like a rejection. If you say like, show me more murder, we will probably be like, we can't do that. If you say, show me more of something that's like so niche that we don't have enough content, we'll tell you like, hey, there's not enough content for this. And then we actually tell you in the feed, when you see something that's because of the request that you made, it'll be like marked as such so that you know what you're getting for. It's very cool. It's fun. You guys should try it.

2:01:52Yeah, I love it. I love it. Yeah, I've been waiting for the plain text interface. Jordy could make his first threads post today, maybe, by trying it. I will do that. I checked your profile. I haven't done one. No. I got to, all right, I'm going to get there. I don't mean to shame you on the live stream. No, it doesn't cross post thing. I'll be on there, at Jordy Hayes on threads. Find me there. Actually, I made my first X post in three years today because you guys tagged me on there. and I wanted to route people to them. There you go. Always be selling. Always be selling. Well, the app looks absolutely beautiful.

2:02:27Thank you. I love the polish. Anywhere from two to three. I want to get that number one. Do you wake up and check the App Store charts? That is not a thing I do. I wake up and I look at six dashboards. Also, just to be clear. Well, it might help if you put a big monitor in the office that just has your App Store position. usually when you put things up like that it tends to like that's a great tip i'm sure my team will really enjoy that anyway dude it's great to hang yeah thank you guys for having me thank you yeah thanks so much we'll talk to you soon uh you heard martin talk about it but now you're going to hear me talk about it 11 labs build intelligent real-time conversational agents reimagine human technology interaction with 11 labs and i'm also going to tell you about fin.ai the number one ai agent for customer service.

2:03:16If you want AI to handle your customer support, go to fin.ai. And up next, Alex Buzari. He's the co-founder and CEO of DDN. He's in the restroom waiting room, and now he's in the TV pin outfit. What's happening? How are you doing, Alex? Good to meet you. Hey, how are you guys doing? We're doing great. We expected you to suit Magus. Yes, the outfit is fantastic. I certainly have. What's the background on the suit? Have you always been to fashion? I've always been into fashion. I'm sure you guys appreciate it because you're definitely not like everybody else. Yes, yes. Well, no, we'll hit you up after the show for some Taylor recommendations.

2:03:56But very, very excited to meet. Yeah. Well, first time on the show, please give us an introduction. So, CEO, co-founder of DDN. DDN solves all the data problems associated with AI implementation. or enterprises, sovereign, so nations, countries, large-scale deployments. NVIDIA uses us internally for everything they do. Elon, large, GROC on 200 ,000 GPUs is powered by DDN. Hundreds and hundreds of deployments like that. So that's what we do. Solve the problems of AI, and we help organizations monetize AI because it's great to invest, but if you don't monetize, what's the point? What's your background?

2:04:39How did you get into this business? How long ago? Been in technology forever. Born in France, came to the U.S. in my early 20s, went to school here, loved it, and then just did a bunch of technology companies. This one, my partner and I started about 20-some years ago. Wow. At the time, we were solving the problems of high-performance computing. Yeah. So high-performance computing is basically government labs, academia, trying to solve complex technology problems. I mean, those guys were our customers. We ended up powering 60 out of the 100 fastest supercomputers in the world in every country, basically.

2:05:22Three-letter agencies, Department of Defense, Department of Energy. And then this little thing called AI started to happen. And so NVIDIA came to us and tapped us on the shoulder. And they said, well, we're trying to stand up a reference architecture. That was eight years ago. And they said, we have all the pieces. We don't have the data. And so we became part of that architecture. NVIDIA became our customer. And here we are, eight years later. AI is booming, as you know, as you see. I mean, it's expanding, exploding in every aspect, every industry. and that's been the journey and the journey is super exciting.

2:06:02Walk us through, obviously you're quite bullish on AI and implementing it across every possible industry, but how did you personally kind of process the different evolutions and paradigms from the transformer architecture to all the different steps that we've had since then? Sure, sure, sure. That's a great question. I mean, look, when NVIDIA came to us eight years ago, I mean, honestly, I don't think anybody realized how quickly it was going to grow and evolve. And so we walked away from that first meeting saying, well, we need to develop a radically different architecture. And that architecture for AI to be successful needs to connect edge.

2:06:46So edge devices, think of that as autonomous cars. Think of it as sensor data, robots in factories that move things around. So it has to connect the edge to the data center where the data is getting processed, analyzed. That's where a lot of the NVIDIA infrastructure is being deployed. And then multi-cloud. And so the evolution really was as NVIDIA and other companies have been deploying faster and faster GPUs, the resulting factor is that there's a scarcity in the number of GPUs available in the world, scarcity in power. There's not enough power in the world. and there's not enough data center footprint in the world.

2:07:28So our technology has basically evolved to adapt to these limitations. People are spending, organizations are spending millions, tens of millions, hundreds of millions. We have customers who are spending tens of billions in building out infrastructure. Well, if that infrastructure is not productive and is not delivering value, then it's wasted and the ROI just doesn't work out. So we've evolved our software stack, call it the data plane, to ensure that these infrastructures are running in the most effective way possible, irrespective of what power shortages might be or data center footprint shortages might be or the number of GPUs that are available.

2:08:09So that's really been our evolution. It's been a lock in step, a lot of it guided by NVIDIA. I mean, our engineers and their engineers interact on a daily basis across all aspects of NVIDIA's engineering. and the primary problem is how do you make it easier for enterprises to implement ai in their environment in a non-disruptive way i mean in essence you're dealing with cios who are like well i don't want to have any glitches because if i have glitches i'm going to get fired and line of business people who are saying hey i want to benefit from ai in developing better more compelling more competitive products and services so you have this tension which means you have to make it easy for them to deploy in their environment.

2:08:53You have to make it risk-free. And so with NVIDIA and others, we've developed these integrated solutions that are industry-specific that can be deployed and make it easy for enterprises to bring in AI into their environment and benefit from it. So it's really that. I think we're moving from an early adopter phase, which is a handful of organizations are benefiting from AI, You know, the hyperscalers, the, you know, chat GPTs of the world, the Groks of the world into one where the industrialization of AI is underway. And I think that's one of the most compelling things that is happening out there.

2:09:32But for that to take place, easy, easy. The latest earnings cycle, I think everyone was shocked by some of the CapEx numbers that were coming out from the hyperscalers. Was that surprising to you or? not not not really because again we're we're very very close to the center of the universe which is jensen and uh and and if you look at it i mean the capex yeah jensen jensen was saying like in q4 of last year he was throwing out numbers that implied that the hyperscalers would be raising their their capex projections massively so it shouldn't have been that much of a surprise but when Jensen was first saying it, he's obviously a salesman.

2:10:15And, like, you know, it felt, of course, it feels a lot more real once they're throwing out, you know. Well, I mean, look, if you think about it, the hyperscalers have a software suite, which they're monetizing across a very broad population, hundreds of millions of users, billions of users. And so you look at that CapEx and you align it with how much they're charging and how sticky the offering is. you just got to do it because if you don't one hyperscaler will emerge i think as a leader i mean just like the google search engine there will be one leader and then there will be a number of others who will have market share but they won't be the leading market provider and i think everybody has come to the conclusion that you have to invest very very heavily because without massive infrastructure deployments you cannot train the model at the level of complexity that is required at the real-time elements that are required in order to deliver outcomes to organizations and consumers.

2:11:16So I think it's really that. Everybody is racing to be the market share leader in this newly created space. I mean, Google is doing it. OCI is doing it. Microsoft is doing it. Meta is doing it. There will be one leader. Getting a little bit more specific, software engineers have done an excellent job creating and adopting a bunch of AI tools. What are maybe some under-discussed areas that you're seeing AI adopted and real usage growth that the kind of broader tech community is less focused on because these are maybe companies or industries that aren't typically at the center of the conversation?

2:11:59Sure. I mean, look, the places where we see significant traction, financial services, because the ROI pencils out beautiful I mean, it's a no brainer. The better your models are, the more complex you can run those models. The faster you can get outcomes, the more differentiation you create. And so the better return to your shareholders. And so that's like companies that are doing trading or? Oh, I think hedge funds, high frequency traders. We have some very large customers in that space. Those are very technical organizations. Typically, the people in these organizations have come from the world of high-performance computing, so they understand the benefits of it.

2:12:41And so that's one bucket, which I think will continue to expand. Secondary is life sciences. Anything having to do with drug discovery, bringing a new drug to market, genomics. The costs associated with bringing a new drug to market are staggering. It's billions of dollars. It's years and years of development. And so in the end, if you find yourself with a drug which is being rejected by the FDA, well, you have a problem. So AI gives them the ability to better triangulate what should an optimized drug be to cure a specific disease. And how do you increase the likelihood of that drug getting accepted?

2:13:18So the cost will still be extreme because of, you know, animal studies, human trials, all the different steps. But we could enter a world where you have a higher success rate for drugs that are entering. Exactly. I mean, it's higher success rate. It's better predictability. And it's also as the omniverse, digital twin starts to happen. I mean, the ability to basically run what if scenarios that you don't have to do in the real world. So if you're bringing a new drug to market and you're saying, well, there's like eight different ways I could do this, but I'm not sure which one is going to be the best outcome.

2:13:55You run simulation in the omniverse with synthetic data. And then the responses come back saying, well, if you combine parts of the first one and the third one and the fifth one, combine them together, likelihood of success will be higher. The drug will be better. So I think increasingly we're seeing that the omniverse and synthetic data is coming into the mix. Autonomous driving, clearly, because, well, if you have driverless cars on the road, you have to collect data in real time. This is cameras, audio, this, that, the other. You have to continuously reprocess the model at value-low scale. So many car manufacturers are our customers in that space.

2:14:39Manufacturing is starting to happen. Again, factory floor automation, retail, just how do you optimize inventory in various types of retail organizations. And then the other really big ones is sovereign AI. I think with what the Trump administration has done, they've created lots of concerns worldwide in terms of autonomy and risk. and the U.S. is no longer there to just bankroll you. So you have to protect yourselves in some way. So we're involved in lots of... Have you spent much time in France over the last few years? There was obviously a little tiff between the Western, the U.S. tech community and Macron last week around.

2:15:25He came out with an announcement that was kind of intentionally misinterpreted a little bit. By the attorney. Maybe by me. but he was kind of putting out charts of showing foreign investment. And we've heard that a number of labs have been excited about the energy availability in France and looked to capitalize on that, but kind of just kept running into different blockers kind of from a regulatory standpoint or a general speed standpoint. But what's your kind of take on France's progress around sovereign AI and just kind of catalyzing the industry locally? So, look, I mean, one of the one of the organizations in France that's very, very active in the space is a company called Mistral.

2:16:14But Mistral is our customer. So we're deployed in their infrastructure. I think, look, at the end of the day, lots of super, super smart people in Europe in general, you know, France, Germany, UK, all of it. But the scale of investments is just not at the same level as the U.S. and China. I mean, so today it continues to be a two-horse race. I think it's U.S. and China. And the Middle East is starting to deploy massive resources into it. I mean, Kingdom of Saudi Arabia, I mean, we're involved in those infrastructure build-outs, so sovereign AI. The good thing there is that the cost of energy is very low.

2:16:53land availability is very very significant and there's a desire to really step up what ksa is doing likewise in uae and well there are geopolitical tensions between the two but i think both have aspirations to set themselves up on the world stage as being the third player Europe will be there, no question but Europe, I mean all these countries have centuries of history and so getting things done quickly, easily is not quite there whereas in the Middle East, well you have one decision maker and if that decision maker says go, it goes with infinite resources I mean, right, trillions of dollars under management in both places So making a multi-hundred billion dollar investment is nothing.

2:17:49Straightforward. Yeah. When you look at where DDN is spending money on software today, how do you think that'll change over the next five years? We've been covering the SaaSpocalypse. and it seems like a number of great companies today could be comparable to the sort of office equipment and imaging companies of like the 90s where the revenues were really high but then the internet and email and the PDF came along and suddenly people just needed less fax machines and all that kind of thing. Sure, I mean look, I think the market is overreacting in some ways as it sometimes does. So, you know, somebody makes an announcement and then everybody freaks out.

2:18:33Oh my God, oh my God, this whole industry is going to get commoditized. Everybody's going to go out of business. Service now is going to go out of business. My God, my God. I think you have the forward-thinking organizations in SaaS who are adopting and integrating AI into their offering. And I think those will do well, provided that they do it at very high velocity. And then you have the ones who will be more traditional in their thinking and in the way they operate. and I think those will go by the wayside. I mean, just look at IBM. IBM is a perfect example. Look at Intel. I mean, these are companies that had everything to succeed.

2:19:09I mean, why is it that NVIDIA is where they are and Intel is not? Well, because the velocity of execution and the ability to adopt something that is happening that is completely different from what it was before was not quite there in the culture of the organization. So I think the way to look at it is which organizations have leaders who are embracing the change, not fighting it, and who are going to integrate it into their portfolio and forge the right alliances. I mean, you could say the same thing about GSIs, Accenture, Deloitte, all of these organizations. I mean, Accenture has what, 750 ,000 employees?

2:19:42How many of those are going to be relevant in this AI-enabled world? Well, Accenture has to completely transform the way they operate and the value that they deliver. Otherwise, it will just go like that. So you need really forward-looking, visionary leadership that will force the change. Because if you stay in your comfort zone and you say, oh, I have a great business, like you said, the top line is steady, the bottom line is fine, you will get whacked. That's just the way it is. I mean look at which AI is operating Jensen speed is pulling the whole industry at a velocity which very few can follow, the speed at which he is turning the GPUs, the integration of the software stack and the ecosystem into the GPU enablement, the open source approach he's taking, I don't know if you saw his CES keynote he's basically developing turnkey integrated software stacks and is open sourcing them in order to accelerate adoption industry by industry.

2:20:56I mean, what he did with Mercedes, it's okay. Here's an open source software stack. It's kind of the opposite of what Elon was doing at Tesla, which is a closed architecture. It's like I'm opening it up because if I open it up, I will accelerate the adoption of AI by the automotive industry. and they will buy more of my GPUs. So I will put 5 ,000 engineers on this for many, many years, and then I will put it out there. I mean, it's a meta, really. What's celebrating an option? Yeah, when you think about kind of forecasting and planning for your business, are you more scared of a chip bottleneck or energy bottleneck when we talk to different...

2:21:39I would also like to put in a research idea bottleneck bottleneck and energy bottleneck. There's sort of four categories that people are worried about progress halting on. Chip, ideas, chips, data, energy. So look, the ability to process, I mean, think of AI as you have models, you need to train the models, then you need to layer analytics on top of it. And then the most important part, which creates value, is inference. From that data, you need to get value. So I always say, we are to data what NVIDIA is to compute. And in order to do AI successfully, you need to combine the two together. So what that means is an infrastructure can only deliver benefits if it is cost effective.

2:22:29And in order to accelerate the adoption of AI, you need to make it cost effective. These shortages will continue, I think, for the next several years. And so you have to say, given the limitations that I have, how do I make these AI workloads more effective and have the ROI pencil out across industries? I mean, I was at NVIDIA yesterday, actually, and we're talking about that. How do we accelerate the adoption of AI by enterprises? Well, by packaging turnkey solution that optimize outcomes? How do you make sure that these agentic AI organizations that are providing services, and unfortunately, these services are very consumptive of tokens, which means many of these companies are now upside down.

2:23:14They're losing money because what they're charging to their customers does not tie into what it's costing them because they're relying on AI. So I think the cost reduction and the compression in terms of tokens required to perform a certain task is really what it is. So I think it's a software play. The underlying infrastructure, eventually it will happen. I mean, SSD shortages and NAND shortages. I mean, we deploy our data plane on top of storage, SSDs, hard drives, and so on. Well, over the last few months, the cost of SSDs has tripled. So it's significant. And it's not available on top of it.

2:23:55Now, we happen to have an architecture where we can tie into SSDs or hard drives and so on. So our customers are able to do the same, if not better, with less money. But, I mean, these are issues. But these are transient issues. I think eventually the problem that needs to be solved is how do you ensure that a task that is performed for a consumer or an enterprise is cost-effective for the organizations that are delivering that service. And it's really that. I mean, that's what we're very focused on. That's what our partners are very focused on. That's why many of our interactions with NVIDIA revolve around this.

2:24:32How do we make sure that we make it easier to deploy, easier to integrate, and lower the cost? Lower the cost of power. Lower the cost of building data centers. Compress the velocity. I mean, look, what Elon did with his data center, and we were involved every step of the way. I mean, he built out the data center in four, four and a half months, which was unheard of. Yeah, when you and the team heard the initial timelines that they were planning around, did you believe it? I said he's completely mad because we had done probably more than 100 large data center deployment and had never seen it done in less than three years.

2:25:11And when the X team first came to us and said, oh, we're going to do it in four, four and a half months, I said, that is just ludicrous. How is that even possible? But see, the way he did it, instead of hiring people who were experts in building data centers, and all of them would have said, it's impossible, mental block, right? If you've never seen it done in less than three years, and somebody tells you four and a half months, he goes, it's impossible. This is stupid. So what he did is he hired very, very smart people who were very good at connecting the dock outside of the box. and he said, okay, I want this done in four and a half months, figure out how to do it.

2:25:48And they did it. I mean, we were there with them Christmas, New Year's, weekends, 24-7. I mean, there were mattresses in the hallways. I mean, everybody was sleeping there. Everybody was just working to get it done. And he got it done. Now, it was extremely painful, but he got it done. And so you go, okay, so the new benchmark now is not three years. It can be done in four, four and a half months. Have you seen any other either Neo Labs or Labs or Hyperscalers be able to replicate that kind of timeline? Like once he set the bar? China. They are very, very good. I mean, we are so behind. We are so behind.

2:26:31I mean, they've developed models. I mean, I was looking at what they're doing in data centers. I mean, the first one is the cost metric. I mean, in the U.S., the cost metric is 10 to 15 grand per kilowatt to build a data center. In China, they're able to do it for between a third and a fifth of that. Why can't we do it? Because they're looking at it in a very optimized manner. What Elon did, he did the first one in four months, lots of issues. Then he did another three. And lessons learned from the first one, he applied to the next three. By the fourth one, he's like, okay, I got this. And then he did the next 32.

2:27:13And the Chinese are doing it the same way. They're not looking at each one of these as a one-off. They're saying we really have to focus on optimizing, optimizing, optimizing, and then we replicate. And that's something we need to do better in the U.S. for sure. For sure, for sure. How do we lower the cost to build a data center? And how do we compress the time to build a data center? And China is way ahead of us right now. They're just way ahead. It's reality. Well, hopefully more Colossus data centers coming online soon. I know, I know. But, I mean, it needs to be done. I think the good thing is people are realizing that China is very good at certain things.

2:27:54And instead of saying, well, no, we're just going to ignore them, they say, okay, how do we learn? I mean, I had a meeting with one of our large customers from the Middle East, and we're actually going through the design architectures from China, looking at how they do it. And we're like, okay, how do we apply that to doing it in the Middle East in a very modular manner? And it's really remarkable. I mean, again, we're not talking about 20%, 30 % cost improvement or timeline compression. When you say it's three to five times, the economics associated with that are huge. Yeah, dramatic. That's it.

2:28:30Yeah, that makes a ton of sense. That's, yeah, wow. Thank you so much for coming on the show. Yeah, really, really enjoyed it. Yeah. Great to meet you, Alex. We'll talk to you soon. Thank you. Have a good rest of your weekend. Let me tell you about LabelBox, reinforcement learning environments, voice robotics, evals, and expert human data. LabelBox is the data factory behind the world's leading AI teams. Let me also tell you about Vibe.co, where D2C brands, B2B startups, and AI companies advertise on streaming TV, pick channels, target audiences, and measure sales just like on Meta. Up next, we have Brett Adcock.

2:29:03He's the founder and CEO of Figure and about 12 other companies. Serial entrepreneur with a massive release today. Brett, how are you doing? Welcome to the show. How are you doing? Yeah, thanks, guys. Thanks for having me. I think most people will be familiar, but break it down. Where is Figure now and give us the news today? Yeah, so we, well, several months ago, we unveiled Figure 3, our third generation humanoid robot. This morning, we actually gave a sneak peek at a future roadmap item we've been working on for about over three years, which is our newest generation hand. I think we've been working on this project basically since the beginning.

2:29:47Our generation one was like this tendon-based hand that we designed in 2022. Sure. um had tons of problems with it and we've basically been working on trying to reach human like how do we approach human parity in terms of uh hand dexterity and sensors yeah does it does does anything else about the robot really even matter if the hands aren't like human level capable um i think like one thing we're realizing is like more and more if we want to like learn from humans we need to like look and do human-like things uh so even like from a from a visual perspective, like having the right kinematics of the hands so that we can do human-like stuff.

2:30:26Meaning like if a human folds like, you know, socks or towels a certain way, like we need to really understand how to be able to do that on the robot. And so if we truly want to do like full general purpose work in a home across the whole world at billion unit levels, we have to start approaching like human level dexterity. There's somebody going for a stroll in the background. Yeah. What is that? Is that just a walk cycle? Is that scripted? Did the robot independently decide to walk behind you right now? What's going on? We literally have hundreds of robots here on our campus in California. They're everywhere.

2:31:05They're all over the place. Okay. Why jump straight to full dexterity, humanoid form factor? Why not wheels? Why not pincher, grabber, more incremental? You know, we've seen Amazon acquire that robotics company just to sort of move packages around. There is a logical chain of events that you could do more incrementally, but you're going for the moonshot straight away, it feels like. What informed that decision? Listen, we have like a very deep respect to trying to do like human level work in the world without changing the world too much. And if like we as humans built the whole world around our, the way we look and feel like the way we like move around the world.

2:31:48So we like, you know, we use tools, doors, like stairs, like, like, so like, you know, we've like built the world so that human body can interact with. The ultimate form factor for this is a, is a human. You know, like if you start like removing like the ability to like have legs or fingers or different stuff, you're just going to do less of what humans do in the world. So our view is that we want to go out and basically do everything a human can. That approach is basically a human form in the limit. So we went after the hardest problem here, which is like, how do you design humanoid hardware?

2:32:23How do we design neural networks now to work on that hardware? It's a really difficult problem, but it's like super tractable. This is a problem that will be solved in our lifetime. In the coming years and decade, we will see like millions of humanoids out in the world doing all kinds of things. What's your bar to get to the point where you're selling a robot that somebody can buy and put in their home and start doing tasks? Because there's a lot of, I'm sure you're testing this stuff constantly and you're able to do tasks like laundry or moving dishes from a sink, cleaning dishes, etc. And yet the bar is an individual just saying, well, I can just do this myself.

2:33:03It's quick. So that's one. You have to overcome that. It has to be so good, so consistent. But like, what do you think is a bar? There's obviously companies like One X that are pushing hard to, you know, get robots into homes. You guys are pushing hard to. Elon is obviously, you know, adapting his Fremont facility, right, to be able to, you know, make these at scale. But I think everybody's sitting around being like, okay, once I can hit buy on one of these things, the amount of pressure that the first company that kind of comes out, if you don't count Unitree, but the first American company to come out with a robot, the pressure to actually deliver real value when people are like, hey, I just spent$30 ,000 on this,$40 ,000,$50 ,000.

2:33:46The pressure is going to be immense. What is the bar for you? Yeah, I would say the thing that really matters here in the world is getting to a spot where you have a humanoid robot that can go off and do many minutes and then hours and days of work fully autonomously with neural networks. That's the bar. If you look at who's doing that today, there's not a single group out there that can recreate the video we did two years ago, which is basically we had a figure one just moving Keurig around with a couple hands for a minute or two. I was done with neural nets. We were just standing in place. It was uncut.

2:34:29It was a few minutes long. And I haven't seen a single company in the world able to do that today. So we can, like, pretend that we're, like, teleoperating robots and be super silly and act like that's going to work. It's not going to work. We have to deploy neural nets at scale to robots that can be fully general purpose over a long period of time without any human intervention. so for figure i think we're just like by far and away the the best example of being able to do this today and we're still like we still have so much more to go in order to be able to put it into a home like for days and days and be like extremely useful in that respect so uh right now we're able to do like pockets of this work really well like we're able to do like um clean up the home do like we can fold laundry we can do dishes like this stuff is being done with neural nets fully end to end um and a lot of times like doing it pretty high performance uh so my view is like we will only launch a product here a figure into the home when we're really ready i think it's the i think the world will only accept the product in the home when it's really ready too nobody's gonna deal with like silly telly operating the room in the home things like yeah the other thing is like we've seen with uh we've seen with a number of of hardware like the humane pin you had the rabbit are one, people might be willing to try a digital product like a couple times, even if a lot of people will try it.

2:35:47If they have a bad experience, they won't come back. Some people might try it again. CHPG got better. Whereas with hardware, it really feels like if you launch a hardware product and it doesn't deliver real utility, you lose everyone. It basically kills the company. So the bar is just so high. So yeah, timelines, you guys have the benefit of being private, even though you have a valuation somewhere around the range of Ford. You have time to figure this stuff out. What are the timelines that you're setting internally? What are you rallying the team around? Yeah, we're working kind of two paths.

2:36:28How do we ship robots in the industrial workforce as fast as possible? Attract this pedal to the metal every single day. we have many different customers there fully signed up ready to go and we're excited uh you know we had robots at bmw last year we have like more robots going into commercial customers this year the second is we want to solve a general purpose robot like solve general robotics in a home and um the best i one of our top goals is like be able to drop a robot of ours into an unseen home this year and do like full general purpose end-to-end work and um it's extremely duff. I think we can go do it.

2:37:01But we're working day and night to go get there. And yeah, it's basically how do we design... It's the closest thing to AGI for physical world, right? How do we get something that can have common sense in a home that you can talk with, you can understand things. Maybe you can teach us something on the fly and can watch you. And then ultimately be able to carry out those tasks at high performance all throughout the day. So my hope is we can make material progress on this this year. Our goal is working day and night to try to solve this. To the extent we can hit this goal of being able to do full end-to-end work, there's other barriers of privacy and safety and other things that are really hard that we're also parallelizing.

2:37:47But my hope is by the end of this year, we're making considerable progress towards this. Being able to show some crazy, insane things with these robots in these type of environments. This is a separate track to the commercial side. We're already out, and we've already been out being able to do this. We're going to go out even larger this year in 2026 to deploy robots at scale. This is important for us to get a real operational readiness. How do we make sure we can run robots at scale really well here, figure? Yeah. It's much more straightforward to have a robot in a setting where you have trained professionals probably wearing hard hats that can be kind of monitoring the robots from far away.

2:38:29You're not dealing with the safety risk of like a robot falling on a dog or on a kid or any of the other challenges in the home. What's your timeline to a robot humanoid being able to bench two plates? Is that an interesting problem to solve? It's the only problem that's interesting to me. For us, we're very fascinated on when that will happen. Yeah. Is it bench or squat? What do we want to do here? I mean, squat is probably the overall compound lift, squat. I feel like it's pretty. Thousand-pound club ideally, but we'll take just bench press if that's what you got. I mean, if we can bench press that, we should be worth at least twice a four, right?

2:39:11Okay, yeah. Yeah, yeah, yeah. I agree 100%. Just for that. The tickets to the bodybuilding competition. No, but I feel like even as silly as that sounds, you know. It is an interesting benchmark, yeah. I think the – I mean, over in China, they're doing robot Olympics. They're doing marathons. They're doing all sorts of stuff. It is a tactile – Yeah. Okay. Can we be real for a minute on all this stuff? Yes. Like, let's just – okay, let's be real serious. What really matters, I think, for us as humans is we look at the distribution of what humans do. It's, like, useful, and we try to do as much of that as possible.

2:39:42Yeah. These things where we run marathons or we do backflips or we do karate moves or we try to deadlift 300 pounds, they're not in the main part or the fat part of the distribution. They don't matter. And if you really want to size for those and do those, you're going to build a really expensive and heavy and unsafe robot that's hard to manufacture. You're going to build a super duty truck. And people want the$10 ,000,$20 ,000 humanoid that can do general purpose work. That's what we want. So if you're trying to size a robot to do those kind of things, like silly things, I think of like those like gymnastics and other stuff.

2:40:16You're going to build a very specialized robot that can do like a very small percentage of what normal humans do every single day. So our goal is to build a general purpose robot to do a majority of what humans can do out there. We want to do like laundry and dishes and be a companion. I want to ship robots at scale and a billion level into the workforce to do logistics and healthcare and build buildings and build data centers. That's the stuff we want to do. I don't need to do backflips to do any of that work. I want other robots building other robots. So I think like – I don't know. I mean I think – you look at the silly stuff out there.

2:40:51It's not only not important for the roadmap. It makes the hardware extremely like heavy and hard and expensive and all that causes more problems. so like none of that matters in our mind i figure we uh you know every once in a while we'll put a robot on a dj stage with dead mouse and stuff for fun but like we're definitely not trying to design a robot to be great at that yeah we want to be great at like the the things i do every day and you guys probably do every day i mean you guys are probably dead lifting uh 300 pounds but like uh at least for me every day i'm trying to like you know just do normal like normal practical stuff that billions of people today are doing that we can help offset how do you how do you think you you know, let's say we get the iPhone moment for humanoids, hopefully in the next few years, and that it's a piece of hardware that has real utility that a lot of people are buying.

2:41:44How do you think the kind of form factor evolves? Do robots over time look, you know, do they follow the iPhone path and that they get like thinner, lighter, you know, that kind of thing? Is there like, what, what are you, what have you been learning so far that maybe people kind of misunderstand about the form factor long-term? Yeah. The long-term form factor is more and more approaching the average human in terms of like the range of motion, payloads and speeds and like what you can do. If you're too short or too tall, you're just like, you're not in the right habitat for like interacting with a human world that well.

2:42:17So in an extreme case, it's three foot tall. It's like really hard to get most things out of cupboard or get into the, get into the sink or reach over a table. They actually become really practically hard to go do. So it's going to be like an average human size overall. I think we're in like the pre, I'll make an argument here that we're in the pre-iPhone stage. We're in the flip phone stage for humanoids. And I think we know this better than anybody. We've been building them like crazy. We have our third generation out in three years. We've walked three generations in three years. I think what you'll see here is that we're trying to find this ideal product like product fit like long term where that's headed and we're like learning every year where that's going.

2:42:55It certainly means being more human like in our minds. We can unlock more percentage of this distribution we just talked about earlier. I think I'll make it I'll make a statement that's pretty pretty bold is that when you look at like figure one to figure two to figure three, we've had to like step up in performance and they're just better and better every year. When we had to figure four here, it'll be the largest step that we've ever made by like but by a long shot. It'll be the first time that we feel that we probably hit like iPhone one level humanoid. We're just like, this is the right place to be in.

2:43:27And then this will like, this will, you know, this will go extremely far at a point where it saturates at some point in the future, you know, maybe 10 more years. But we feel like relatively early. It's an extremely difficult piece of technology. It's obviously early in this. So it's got to be like earlier than phones, right? We're not there yet. But I think the iPhone 1 moment will happen with figure four and it'll just be, it's just an unbelievable machine. And I never would have suspected we'd be able to make that big of a move. In figure three shift, I'm like, this is it. This is the best it'll ever get.

2:44:02And then the more we learned and ran it and the more we developed neural nets here with Helix, the more we really understood better about what the hardware should like should look like and be like and the more we got folks in the room together with us and said how do we radically redesign the head the the hands the the kinematic systems like all of it from scratch and i think um you're gonna see something i mean this stuff is just gonna get crazier and crazier in capabilities uh what are your goals around uh consistency So for example, a robot that can unload my dishwasher, if one out of 100 times it breaks a plate into 200 pieces, maybe that's not that big of a deal if it can pick them all up easily and I'm not home and I don't see that you're exploding the plate.

2:44:51But what level of consistency do you guys need to get to before you'd be at a point where you can sell one of these things? I would say we probably need something pretty high. I think it would suck pretty bad if you're at my house and dropped the number one mom coffee cup. You're getting your ass booted, right? I think you've got to be, especially around safety and stuff, this needs to be super high performance. So we watch folks that are trying to teleoperate and ship early when the product doesn't even work. And it's just silly. They're all going to die. You've got to ship something really high quality.

2:45:30And that is just a super hard thing to do. So I figure we'll ship into the home when we're ready. We're not ready right now. We're trying to get, like, you know, I'm here until midnight every night, seven days a week, trying to get more ready with my team. I hope we get, like, we hit someplace where we're getting really, really close this year is what I really hope. How are you processing? But you're right. Yeah. How are you processing the Waymo story of teleop? Because I was completely on board with the Elon pitch for straight shot to FSD, collect a bunch of data from the cars that are on the road, train the big neural network and FSD.

2:46:02I mean, we talked to Alex Roy who drove without touching the steering wheel all the way from LA to New York. Like it clearly works. At the same time, a lot of people in San Francisco hop into Waymo and they're like, that works too. And so it was a bit of a narrative violation where a lot of people were saying like the Waymo tele-op model will never scale and it feels like it's scaling. So how is it, is there a world where both approaches work or are they fundamentally like different industries? I think what I'm seeing in the space here, and it's been like a pretty big shock, is like everybody in the humanoid space is just teleoperating the robot with a human in the back and they're putting out a video out, not being very explicit.

2:46:40It's very different than this. It'd be like the most analogy for your Waymo Tesla, like your Waymo is being driven by some dude in Kentucky, not with neural nets. Waymo has neural networks. The way they went about with the sensor suite to go do it is maybe harder to scale than cameras. the situation happening in Humanoids is like there's a large percentage of the companies out there that have a dude in the back that are teleoperating with the robot in real time. And then we're trying, we've done everything we've ever put out publicly. It's always been with neural nets on stuff we've done. We've never teleoperated in that case, any of those videos.

2:47:17So it's just like the self-driving stuff is not the greatest analogy. You're definitely not going to be able to human teleoperate in people's homes and the latencies will be terrible the data coming back will be terrible train neural nets it's just not enough data so there's just a bunch of problems with that story and um it's it's not going to work if it would work fast and we can get product market fit and get out earlier we would do it just like it's just dead end completely you really want to solve like for real neural nets real autonomy from from the get-go how big of a bottleneck is data what are you guys doing to solve it like is there hardware breakthroughs that you guys are looking to achieve?

2:47:55Or do you feel like the, obviously you have the new hand, which sounds like it's a step up. But what are kind of the key bottlenecks? We just unveiled Helix 2 about three weeks ago. It's a robot that can basically do like fully end-to-end whole body work. We did it in like unloading the dishwasher and rerunning it. That was basically, the whole stack there was basically neural nets basically all the way down the stack. the only reason why it could do that now or like go from there to do laundry is just a data problem like we need just more data to cover the distribution of those new tasks and then the robot can do it at this point so we feel like the longest pull in the tent prior to like extremely high rate manufacturing is how do we acquire data at like a really high clip and so we're spending a lot of time on that that gets you to a point where acquiring data through teleoperation is just not going to even be close.

2:48:53You have to embrace learning from humans at scale. That kind of figures core models that comes to Helix for neural nets. And I would say if we could snap our fingers and have enough of the right data today, you would have a general purpose sci-fi future of robots in our office right now that we'd be able to put anywhere. We have it. We are just extremely data constrained. It's not as simple as just like going out and getting random data. It's got to be the right type of data to match the observations and action spaces of the models well. But we now know what that data is. We are acquiring that data like crazy here at Figure.

2:49:27We'll spend nine figures of capital on acquiring data like this in 2026. So it's a huge focus for us. We think it's by far and away the biggest bottleneck to get to general robotics. Yeah. How do you think about China in the context of the race for humanoid robots? Obviously, there's competition from humanoid robot makers there, but there's also a bunch of great parts suppliers at all levels of the supply chain that might be useful to build American humanoid robotics companies. How does that puzzle play out? I mean, listen, I think as it relates to competition, I think what's extremely important is seeing robots that can do human-like work with neural networks that's useful.

2:50:10We haven't seen any of that out of China today. They really don't have any hands. They're good on hardware, but they're still behind on software. Well, I would say you probably don't have good enough hardware if you're not able to do the software really well. They don't have enough compute in a lot of cases to run things like Helix on board. Unitree, for example, has a very tiny computer where you can run very small reinforcement learning controllers to do open loop replay of stuff. They wouldn't be able to run Helix on a board of hardware like that. They don't have any real human-like hands of five fingers.

2:50:40So you're really missing a couple of big parts of the story. Beyond that, I think they've been great in existing industrial robotics and existing consumer electronics the last several decades. I think those are playing a big part of the ecosystem supply chain for humanoids. that are important in some cases. But listen, we basically design almost everything internally here at Figure. We don't go buy designs from China. We do it all here internally. We even manufacture the robots next door in our campus here. We have a Figure 3 robot now coming off the line every three hours. And I think we'll be at every half an hour here in the coming few months.

2:51:22So things are coming out at a pretty high clip. But I think today, if you look at who's doing the best human-like work with neural nets over long time horizons, it's not China. And I think we feel at least a few years ahead of anything we're seeing out of there. What's compute like? I mean, you mentioned you're working so intently on neural nets. You have to train those. Are you at a point where a training run is run on a massive cluster? It costs nine figures or something like that? Or is it more data collection at this point? And then the actual training run is pretty tight. Yeah, we spent like hundreds of millions of dollars on compute.

2:52:04That is, yeah. There we go. There we go. Yeah, a bunch that went live already. Our next giant step up is going in April, like 1st of April. that's for training helix models that we're doing here internally which are quite large yeah um in long runs and then separately we do all of our inference on board on two gpus in the torso of the robot sure so we can run in like cases where we don't have a network we can run at much faster speeds and we can put all those models fully on board the system so all of our robots now are kind of they're running off brains that are all on robot yeah um so they don't need any outside network to be able to do work so talk about like input and output is that is is the is Is the network sort of taking in voice input and trying to translate that into plain text actions that then get transformed into motor actions?

2:52:56What is like the reward function for a humanoid robot? Yeah, we're taking in, we're basically taking in the instructions through text or speech. Like, what should I be doing? We're taking in vision from the cameras and the current state of the robot. Like, where is it? Like, you know, what is the body doing? what does the sensors look like? And then we're basically processing that on board with Helix 2. Helix 2 is an outputting, like basically trajectories of what the motor should be doing. Basically figuring out like, what do I put torque at in every single joint to produce, to move my body in a certain way?

2:53:32And I think, you know, honestly one of the hardest problems we've had last two years is, two years ago we were basically doing, it's like, you know, coffee work and other types of stuff on tabletops. And it was really, it was unbelievable. It was like the first time we're like, man, neural nets on humanoids work. We spend the last two years trying to leave the tabletop. How do we walk around with neural nets fully end to end? And it's extremely, it sounds kind of, it sounds like, you know, not maybe not the hardest problem in the world. It was a bit like some of the hardest problem in the world for us.

2:54:00How do we get the whole body, like 30 plus joints, all running at like, say, 200 times a second and doing the right things with camera frames and a prompt coming in. And that's what we did with Helix 2 Unveil three weeks ago is we had all that done. So it's the first time in three and a half years where we feel like we have the right technical stack to actually scale, which we didn't have last like several years. We were running at BMW last year and we're like, man, this is going great. We're learning a lot, but it's not the tech stack I want to scale. And we have that now with Helix too. So we're really excited to hit the gas on, we're going to hit the gas on this in 2026.

2:54:35Congratulations. Yeah, it's great to get the update. Yeah, this is awesome. I know you said the bench press is silly and things like that. It's not silly. It's to us. No, it makes perfect sense. Even the bar, even just repping out the bar, I'd be pretty excited about. It's a little bit of a meme. I also think. It's a serious company. It's a serious company, but even serious companies can have fun. Yeah. I'm looking forward to the moment that you get a figure robot surfing at Jaws, too. That'd be great. Are they waterproof? Can they swim? Guys, we got a gym here. Okay. You guys swing by. Okay. And let's get the figure three and you guys in the gym.

2:55:11Let's see how you guys are all... Matching up. Yeah, we'll figure out the matchup. Thank you. Awesome. Well, have a great rest of your day. Yeah, great to meet you. We'll talk to you soon, Brad. Nice to meet you, guys. Let me tell you about Graphite. Code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. And I will also tell you about Shopify. Shopify is the commerce platform that grows your business and lets you sell in seconds online, in-store, on mobile, on social, on marketplaces, and now with AI agents. There's a company that launched yesterday.

2:55:42Yes, what company? Open Claw for Slack. I don't know what it's called. But they said they launched three hours ago, and they just hit one million ARR just now. So they made$350 in three hours, which is one million of ARR. We've done it. We've done it. They did it. Final four. They did it. I saw a fake post that was like, meta acquires Open Claw for a billion dollars or something. and I really had to fact check it because I was like, this seems so possible in this day and age. I think they have that team with Manus. Yeah, yeah. I think the Manus team is in a good spot to sort of bring some of those functionality to bear.

2:56:21Ring apparently has terminated its partnership with Flock. Their Super Bowl ad did not go as planned. The doorbell company ran an ad during the Super Bowl that outed a search party feature that uses AI to help locate lost pets, people quickly realize that maybe it could be tracking things other than pets. But anyways, they probably are the Super Bowl loser. That was, in the end, the worst blowback of any company that I've seen. This Searcy post is so good. VCs love to be like, yeah, hedge fund guys may be smarter, but at least I make less money.

2:57:07Oh, good stuff. Goldman Sachs CEO, David Solomon. We're going to see potentially some very, very large IPOs, unprecedented in size this year. He's agreeing that it's about to rain. It's about to rain. Oh, one of the IPOs that could be going out is Cohere, Aiden Gomez. We got to get him on the show. I'm such a big Aiden Gomez fan. $240 million a year set stage for IPO. This is in TechCrunch. Of course, Aiden Gomez, a Death Grips fan. So you know he's a good time. You know he likes the Death Grips. Size gong moment. Airbnb, according to Saar and according to Chesky, has generated$19 billion in cash flow since going public.

2:57:56Not going to vibe code that. Not going to vibe code a house. you're not going to vibe code a basement that you can sleep on the couch on. I met my co-founders for my first company on Airbnb. I didn't know that. Yeah. Moved to Silicon Valley, get into YC, need to find somewhere to stay. Was staying with like a friend who was sort of, you know, not doing a startup. So way different lifestyle. And searched on a service that was actually the most vibe coded software. But in 2012, it was a mashup between Craigslist and Google Maps because Craigslist didn't have a Google Maps feature. So if you were looking for housing, you had to just guess where the places were.

2:58:41Insane. So it was called Padmapper. And someone took, they scraped Craigslist and then put it on a Google Map. And so you could click and be like, oh, that's near me. That sounds like a good place. But PadMapper, Craigslist had given PadMapper a cease and desist. We don't want you scraping us because guess what? I mean, every single marketplace created in Silicon Valley was immediately scraping Craigslist. Totally, totally. They've fought back aggressively. They fought back. And they said, hey, we're going to get around to doing Google Maps on Craigslist. And so PadMapper, get out of here. So when my co-founder and I opened up PadMapper, the only data that was still flowing to PadMapper was Airbnb.

2:59:20because Airbnb, of course, is a marketplace. It doesn't matter if someone else is driving traffic. They love that. They were all over SEO. They wanted other people to flow in. So we didn't know this, but our future friends and co-founders had a large place in Sunnyvale that they had an extra room, and they had thrown that on Airbnb. That showed up on Padmapper. We go over, take a tour, and we're like, this place is sick. It was a disaster. All the toilets were broken. They were like, it's like the social network. It's got a pool. It's got a jacuzzi. We're going to be hanging out. It's going to be the best summer ever.

2:59:57The pool and jacuzzi filled with algae, like truly filled with algae. We spent the entire summer being like, we're smart guys. We can beat the algae. Let's go get one gallon of bleach. Pour it in. We're going to need a hundred gallons. Thousands of gallons. We're going to be swimming in bleach. Yeah. No, it was like there's nothing you could do. And then we were like, there was like a filtration system, but that was super clogged. So we were like, empty out the filtration system, try and take it all apart. But like everyone who was like an expert was like, yeah, you just have to drain this and like declare like pool bankruptcy, basically.

3:00:30Brutal. Well, New York Post says, have an AI girlfriend or boyfriend. Now there's a bar for you. There's a Hell's Kitchen establishment that has been redesigned for those who have AI partners so they can bring along their phone for romantic evening. very dystopian her moment but not entirely unsurprising that this bar is pivoting to AI I mean it's a good day to launch right because 4.0 is deprecated today 4.0 is deprecated today is it still is it still available or is or did they stop it at when the clock struck midnight I still see it it's still on my lunch at GPT well it'll be interesting to see what the community does because I did see some posts about people being like, I'm recreating 4.0.

3:01:16I'm fine-tuning some Chinese model. Kimi could be potentially fine-tuned on 4.0 outputs and paid for and distributed. There are other ways for those folks to get what they want, essentially. Well, it is Valentine's Day weekend. But before we go, Tyler, we did have a recommendation for you this weekend. You mentioned that you've been seeing a lovely lady. and we thought this was supposed to be abstract we thought this was supposed to be a recommendation for the audience yeah well there's a girl that that uh maybe maybe doing this maybe tyler likes and we were just saying go uh surprise tell her hey tomorrow just have a bag ready this is so out of pocket continue have a bag ready uh we're gonna go do an overnight trip yeah find a nice hotel Nice hotel.

3:02:08Check in. Staycation, basically. You're not getting on a flight. You're just going somewhere local, but somewhere nice. Yeah, somewhere nice. Yeah, the beach. Easy to set up. Check into the hotel. Maybe get her kind of a spa day. Yeah. She goes to the spa. You sit down. And she doesn't know this, but you actually booked her an eight-hour spa, like a full-day thing. You sit down. Time to lock in. Time to lock in on some cheeky pine. I have cheeky pine. I have Dwarcast, Dario. Yes. Yeah, yeah, yeah. You've got a lot of stuff. So lock in and then just start getting Guinness on room service. Yes. It's 21 now.

3:02:40Pint for pint. And go every time. Drink along. Every time AI is mentioned, you take. Yeah, or every time John takes a sip, take a sip. Drink a whole beer. Yeah, and you basically are going to have 24 hours. So she comes back eight hours later from her eight-hour spa treatment and is like, what were you doing? And you can just catch her up to speed on everything you watch. And I think they really appreciate that. What did you say? You haven't listened to Dwarf Cascielia? That one hits like a ton of bricks on Valentine's Day. Ask your spouses, ask your girlfriends, your boyfriends. Have they listened to Ilya on Dorkesh?

3:03:16Are their timelines up to date? If not, that's the best Valentine's Day gift you can get them. I agree. An up-to-date understanding of what's coming. But we hope you all have a wonderful weekend. We love you. Yes. Thank you for hanging out with us this week. Yeah. And we will be back Tuesday. Monday is a holiday. It is. We're off. Yeah. Really? Yeah. Yeah. Markets closed. I did not know that. Yeah. I'm learning this for the first time. I'm learning this for the first time. Yes. We experimented with streaming on holidays and there was not a lot of news. So we'll be back Tuesday, 11 a.m. Pacific.

3:03:53We'll see you then. Love you. Nice work, brothers. I'll see you on the next one.