SaaSpocalypse Revisited, Singer x Louis Vuitton, Aman vs Ryan Walker | Igor Babuschkin, Brannin McBee, Garret Langley, Sonya Huang, Sean Cole

32:54Sheep billion. Scale it up. Yeah. The new shrimp farming hustle is sheep farming for sure. The last one here is, I guess, the Supreme 9-11, which that seems not author. Nick in the team chat says, I'm a sheep the same way Odysseus is a sheep. What does that mean? People think he's a sheep, but he's really an absolute dog. Okay, okay, okay. He's a wolf in sheep's clothing. What in the slop is this? I can't even get that out of here. Okay. Anyway, let's do, let's do some other stories. New bombshell reporting from the wall street journal reveals that North Korea has built a secret workforce inside American companies using stolen identities, AI and accomplices in the United States to cheat its way into remote jobs and funnel hundreds of millions of dollars back to kim jong-un's regime it's a fascinating story the wall street journal article is posted uh it's a 30-minute documentary they spent over a year investigating this the fbi says that there are thousands of north korean i.t workers applying for jobs applying for jobs across america after a year-long investigation the journal obtained a trove of leaked browser histories emails calendars and screen recordings from one cell of workers, providing a remarkable look at how the operation actually works.

34:27The North Koreans apply for jobs at enormous scale. One cell tracked by the journal applied to more than 1 ,000 companies over just three months, using AI at nearly every step. And in the documentary, they show a product that is basically, clearly they will leave a voice agent running while they're being interviewed in a technical interview, give the answer when asked about a particular technology, If they are proficient, they will simply have AI look up the answer, give the answer to the interviewer, get the job. And in some cases, they're using real Americans as frontmen. Basically, they will do the interview.

35:03And then they will ship a remote laptop to say, hey, you're our new remote worker. You are qualified to work on our software engineering team at this small company. We'll send you a laptop. and then you just collect a check, send half of it to the North Koreans and you don't have to do any work. So it's like passive income, free money for you, the Americans. So you said they're real Americans, but they're not patriots. Yeah. That's what I'm hearing. The documentary is pretty gut-wrenching. The individual who they talked to who was participating in this obviously is probably going to see some legal consequences for this, But it did seem like he had a very, very tough go and got into a very rough situation to be in that situation.

35:48It was certainly not his first choice. Ryan in the chat has a hot take. Kim Jong-un is such a lovable rascal. It's hard to stay mad at him. Boo. Boo. The operation also relies on help from inside the United States. North Korean IT workers pay American facilitators, they're called facilitators, to host company laptops in the United States, allowing the actual workers overseas to remotely connect to them while appearing to be employees working domestically. Also, the North Korean workers, they go to non-extradition countries. So they'll go to China and Russia and a few other countries. And then from there, they will be remoting into an American laptop.

36:29So it just looks like, okay, the web traffic is coming from, it's not coming from North Korea, but it's coming from a country where if the FBI says, hey, can you send us this person? They're a criminal. That country says, no way. We're not sending them to you. So for one job paying$75 ,000 a year, he said he and the North Koreans split the salary 50-50. The workers also use stolen American identities to pass employment screenings, often juggling multiple identities simultaneously and holding down several jobs under each name. This all boomed during COVID and the remote work boom. And the scale of the operation is enormous.

37:06Some North Korean IT workers earn as much as$300 ,000 a year. The Treasury Department says the North Korean government can seize as much as 90 % of the wages earned by its overseas IT workers. And recently estimated that these operations generated nearly$800 million in 2024 alone. Pretty, pretty big. So the journal's 30-minute documentary titled Infiltrated, North Korea's Secret U.S. Workforce, follows the operation in detail. It's a fascinating watch. And I highly recommend that you check it out. So go take a look. I feel like it's important to ask, like, are they doing a good job? Because presumably they're not just, like, hacking.

37:44Wow, really? You're going to steel man this? Okay. I see where your loyalty is. No, it's like there's like the skit where it's like, oh, we're going to rob the bank and we're going to get a job there. Yeah, yeah, yeah. And over 30 years, they're going to deposit the money straight into our bank account every two weeks. And what happens at the end? Yeah, we walk out. We walk out the front door. It's like, oh, they're like hacking and they're stealing American dollars by working at the companies. Yes. So there are sanctions. And because of those sanctions, America is not allowed to do business with North Korea in any capacity, including this one.

38:13So this is sanctioned. But like seemingly it could be much worse, right? They could just be like hacking the company. That's true. That's true. But they could also be doing both. Because once you let a North Korean IT worker in your systems, they could be planting all sorts of spyware. A very dangerous situation to be in. But, yeah, a wild, wild, wild story. So, everyone, if you're running a small business with some remote head count, make sure to ask everyone to post something negative about Kim Jong-un every day to prove their loyalty, I guess, or something. But it opens with a very, very funny clip of someone doing a remote Zoom interview and asking someone to say something negative about Kim Jong-un.

38:55And the guy's like, oh, I can't hear you. I can't hear you. I can't possibly say that. Anyway, let me tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. Just do it. A new luxury hotel canceled a YouTuber's$4 ,663 stay. Then came the viral feud, reports the Wall Street Journal. A dispute between Amman, which we've talked about a lot on the show, and content creator ryan walker illustrates the growing tension between commercial enterprises and influencers it's a fascinating story and somebody's got to stand up for the amand somebody has i'm ready uh no honestly honestly i don't i don't think the amand actually needs that much standing up for no the wall street journal broke it down in great detail and it's bad for this guy anyway get it Let's go through it.

39:50Should we pull up a little of Ryan's video? Yeah, for sure. For sure. The social media video is titled. He probably should have deleted it by now because it is extremely misleading. It is crazy. It's still up. It's in the timeline. And overly dramatic and not representative of what a normal guest experience would be like. Yeah. It is odd. Let me drop it a few different places, see if we can get this in here. Boom. So the Wall Street Journal reports that it was supposed to be the hottest hotel opening of the year. Let's play a little bit of his video. You don't know me. I am a luxury hotel reviewer here on YouTube.

40:31I specialize in very honest reviews of the world's best properties. I'm not going to give away too much as to what happened. I want you to watch the footage. And this is a cool thesis. Like he pays for the hotels himself. He doesn't get paid by the hotel for the review. the hotel doesn't pay for his stay and consequently he can be more independent this was the doug de miro strategy for years doug de miro said i'm not taking press cars i'm not going to your press event i'm going to borrow the car from just someone who owns it and i'm going to be able to drive it how i want to drive it review it how i want to review it and if you're lamborghini or ferrari or any other company you're not going to be able to have any uh any thumb on the scale of my review.

41:15Very good in concept, but the execution goes a little bit off the rails in this particular video. So you can see him pulling up to a guardhouse that looks woefully unfinished. It looks very, very rough. Your name is Warren Walker? Yes. Yes, I know you. Allow me to verify your information, okay? Okay, thank you. What's going on? Drama. This is such a strange area. It's kind of like... They say cannot enter Apparently our drivers saying in Spanish that they are saying he cannot enter or pass Excuse me, Mr. Walker. Yeah, we don't have your reservation Do you need the number? All right, we can pause let's let's get into let's get into the journals piece baby long story short that video sparked a bunch of yeah There's thousands of comments on there, people trashing the properties, basically fully taking his side.

42:20And his side is basically that he showed up, he had a reservation, they didn't let him in, the place didn't seem together. He had a bad experience and he shares that with his YouTube audience, gets a lot of views for it. But the journal dug in and provided a whole bunch more interesting perspectives. So let's go through it. The travel world was abuzz over the coming August 1st launch of Amanvari, the first Mexican resort from Amman, the multi-billion dollar ultra luxury hospitality group, whose mythos is built on seamless service, total tranquility, and fierce guest confidentiality. The new location is set within Baja California's Baja California sewers, private Costa Palmas community.

43:05And the resort promised 18 beachfront casitas. So only 18 keys, a very small hotel, but very luxurious. So less than a week after its debut, the hotel was the center of a viral controversy in that YouTube video. He said he tried to check in for a one night stay and he was turned away. So an Amman spokesperson told the Wall Street Journal they went on record and they really clarified a lot. They said, the deceptively edited video created by someone unauthorized to be on our property does not reflect the circumstances of the incident accurately. Their dueling accounts reflect a new reality where cameras are always rolling, virality can outpace truth, and public perception is shaped less by what happened than who posts first.

43:49So on July 7th, Walker booked a stay at Amanvari. The YouTuber with more than 150 ,000 followers, pretty solid channel, has built a brand around paying full price for hotel rooms to guarantee honest reviews standing out amid influencers accepting free trips in exchange for positive content. He says, I'm honest and transparent. I'll tell you exactly what to expect and why. He had previously paid$26 ,000 for a four seasons yacht trip that he panned in May. And sorry for the spoiler. He's not honest or transparent. The same month, Walker posted a positive review of Amon Tokyo. Then he traveled to Amonvari on August 3rd to hear Walker tell it in the video he showed the next day.

44:30He was a victim of a hospitality nightmare. And one thing that's, I think, very funny right away is so. So you're going to review this hotel by a one night stay, which which means like, you know, maybe you check in. it doesn't look like he wasn't going to get an early check in either way. Let's say you check in at four. You got to be out of your room by 11. Yeah, that's not. You got a five hour window where you don't even know. How could you how could you accurately review the property? Good point. When you're missing that incredible window from 11 to four. Yeah. Yeah. No, you just. It's a good point.

45:06You just don't have any context. Yeah. How can I trust your review? So I'm already my guard is up. So his video, which rapidly surged to more than 750 ,000 views, shows his vehicle approaching an unfinished wooden guard hut with stud walls still visible on the outside. In the interactions that follow, which appear to take place between two different gates, staff recognize him, inform him that he has no reservation, and according to Walker, eventually call the police to escort him away. Sounds terrible. As Walker drives off, he finds an email he says he missed, sent the day before and apologizes for canceling his stay due to scaled back capacity during opening week.

45:43Not looking closely at the timestamp, he tells viewers the email arrived late the previous night. When later pressed by the journal, Walker revised his timeline, saying it arrived the previous morning. Walker did not amend the timeline in his video or in a subsequent live stream. The internet quickly took Walker's side. Viewers of his YouTube channel, including people who identified as Amman loyalists and travel industry professionals, expressed disbelief. We'll never book another Amman property again, read one of the more than 5 ,000 comments, many of which expressed similar sentiments. Amman's spokesperson said the company does not comment on bookings.

46:20Inspect the footage closely, though, and questions surface, notably around an absence of the police invoked in the video's title. Amon said it didn't call police and showed the journal an incident report that made no mention of law enforcement. In the video, at one point in the encounter, a staff member says, I need to call the police now. Walker sees her on the phone, says off camera, yep, she's calling the police right now, but that might not have been related to him whatsoever. So that's like a very confusing situation. Later, he told the journal he did not see police arrive on site. When the journal told Walker that Amon said it had not called the police, He said he hoped that was true and that the information would make him reconsider how the resort had handled the situation.

47:04There is also the arrival gate that does not look like a typical five-star welcome. And we can pull up a picture of the actual welcome gate because it does look pretty ramshackle. The Amman spokesperson said Walker bypassed the main entrance. He was looking for trouble. My theory is he knows exactly what he was doing. And the other thing is it also came out that they had reached out. They had called him. They had WhatsAppped him. They had made numerous efforts to make contact. And I just think he wanted the drama. Probably. Wyatt in the chat says, I want to have my hotels call the police on YouTubers.

47:50so um the amand spokesperson uh said walker bypassed the main entrance ending up at staff access points instead ambient audio in walker's footage captures someone saying quote this is the employee entrance walker told the journal he had never heard the remark and that his driver followed gps directions so the first entrance showed in the video you can see it there so uh internal documents provided by walker show the resort emailed him five days before his visit telling him it was unable to allow stays involving content coverage until a month-long media blackout and suggested he postpone his travel.

48:25Walker replied that the staff should consider him a standard guest. The property replied that it was prepared to welcome him, noting that he could post content after the media exclusivity window. So I'm sure they're doing a whole bunch of different press elements. So he comes during the first week, books one night, knows that he shouldn't be filming, still decides to film and is effectively looking for trouble the entire time. And the crowd over on the Wall Street Journal loves it. The top-rated comment, I'm on, here I come. Any hotel that bans influencers is doing regular guests a great service.

49:00Yeah, yeah. I mean, that's a big thing is that people go to certain places to not have cameras all over the place, and it's getting rarer and rarer. It's so actually fascinating to look at the difference. YouTube comments are like, wow, I'm never going to go to the Amman. And then Wall Street Journal is like, Walker is a complete tool, says Ray. Edward says, seems Mr. Walker struggles with the truth. Oh, wow. Second most highly rated comment. Walker says he didn't. And then someone else says it takes a special sort of Jack star star star to make me side with the ultra luxury resort with holistic wellness temples.

49:42Yep. But that's the case we're in. A spokesperson for Amon said the resort also sent him a message on WhatsApp. They're telling him to get a real job. The Amon spokesperson said the resort refunded Walker's booking in full, pledged to cover additional travel and cancellation costs and help him rebook his stay. Walker said he received the refund but didn't take Amon up on the other reimbursements. He just showed up anyway. The dispute highlights growing tension between commercial enterprises and content creators looking to record on property. Jack Ezong, CEO of Luxury Travel Advisory Embark Beyond, noted that hotels maintain the right to cancel reservations or deny access.

50:17He added that properties immediately after opening should be approached cautiously. You don't need to be the guinea pig. Give it six to eight months to grow minimum. In this case, nearly every claim has a competing account. Did Amanvari overreact to a content creator's opening week visit to prevent an honest review? Or did Walker turn a misunderstanding at the gate into a viral story? The luxury world may still value discretion. The internet values whoever speaks. Ian, the X-Chats says, imagine staying on a 20-room property and having a YouTuber there. I would be seated. It's so true. I wouldn't be surprised if they update their policies.

50:53We've got to talk about McDonald's before our first guest joins. In Wired, Reese Rogers says, It says, McDonald's built a 515-page dossier on me. It says I'll never stop eating there. It requested a copy of my data from the McDonald's loyalty program and received an extensive personalized report that algorithmically predicts my next purchase. This is amazing. Pulling up the actual article. cool. McDonald's secret sauce is really commercial surveillance, says Jeff Chester, executive director at the Center for Digital Democracy, a group that advocates for consumer protections. Privacy experts I spoke with said this level of detail may feel invasive, but it's fairly standard for how large companies in the U.S.

51:51run their loyalty programs. The report contains specific pieces of personal information about you that were identified by searching McDonald's systems, which contain information about our customers. Tyler, you had the conclusion. What does McDonald's actually know about you? What does McDonald's know about an individual? That they like hamburgers. That's basically the takeaway, right? Like the takeaway is the number one most likely product is a large Diet Coke. Then the spicy snack rack. Then the Grinch McShaker fry large. We should figure out other companies that have this type of data and request.

52:34Full reports? Yeah, we need our own reports. I would love to read. You want your 10 ,000-page report on what you like from Erewhon? Yeah. You're really going to go there? Because I couldn't possibly know by looking at previous hoardings. Glass-bottled water. McDonald's. uh, anthropic is, uh, apparently in talks to buy the cart for$6 billion. Getting image generation. We've had, we've had a bunch of awesome conversations with, with Dean over at Descartes. He's always, he's very, very early. He was willing to do live demos of their product, like in interviews completely on, you know, we didn't even test with him before the show.

53:16He was just ripping them live. So always had a ton of confidence in the product, and they've been just cooking. They were valued at$4 billion a quarter ago. Hemanshu has some extra context here. It says, Descartes could be one of the first AI labs focused on world video models to be acquired by a Frontier AI lab. I think this also could mark an initial phase where Frontier Labs start treating world models as a major strategic capability alongside inference optimization as a major offering from Descartes. Yeah, interesting to imagine how this actually links to the core thesis of B2B and enterprise and coding, but certainly an amazing technology that feels like on the verge of a breakout.

54:06The demos are amazing, but we haven't had the Ghibli moment for world models yet. It's very much a prototype demo video. Go and see what it looks like. But we've talked to a lot of people, Oliver Cameron and Fei-Fei Li, about world models. There's a lot of optimism about how this all plugs into the AGI pursuits of the Frontier Labs. Well, we have our next guest already in the waiting room, so let's bring in Igor Blushkin. But first, let me tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agents to deploy web apps, servers, databases, and more. while Railway automatically takes care of scaling, monitoring, and security.

54:50And we'll bring in our next guest. Here he is. Hey, guys. Thanks for having me. Thanks for hopping on the show. Welcome. Congratulations on the fundraise. But maybe let's go back in time. Tell us a little bit about your history and journey to starting River AI. And Jordy already has the gong ready, so just tell us the fundraising announcement, I guess. How much did you raise? Yeah, we just managed to raise$1.1 billion.

55:23He broke the gaunt. He broke the gaunt. He broke the gaunt. The chosen one. You are the chosen one. Okay. See if it works out. So thank you guys. I started my career as a physicist. I was really interested in understanding the universe. Yeah. But then realized that there was something really big happening, which was AI. Yeah. It started to happen. And I think AlphaGo was really the moment when I started to feel like, wow, I've got to switch and learn how to do AI. So I managed to join DeepMind. It was almost 10 years ago. Wow. I worked there on WaveNet. We trained a StarCraft agent that was really strong.

56:02So I got into reinforcement learning there and then switched to OpenAI at one point. was really interested in reasoning, coding with OLMs, which turns out to be something that works, which is pretty crazy to see. And then ended up co-founding XAI together with Elon. So I thought it was time for another frontier lab. I'm always in favor of more diversity in AI, more companies doing different kinds of things. So I was happy to support him building up XAI. And now with River AI, we're kind of taking that to the extreme. So I feel like the way we've been building AI is maybe not the way of the future.

56:36a few large corporations building these super powerful models everybody has to pay them by the token we want to figure out how we can distribute AI to everybody in a way where you own it you're able to shape your own AI systems maybe you have the inference running in your home or in your office, that would be the best achievement if we can figure out how to do that efficiently and we're running a few different bets on how to help people build up their own AI so we're helping companies build AI with the River API, that product's already out so if you go on river.ai slash API. You can log in and you can start training models based on open weights.

57:11So we support all kinds of very powerful open weight models. And then we're also using that platform to build out personal AI agents that are increasingly personalized to you. So as you're using them, they understand you better and better and should really feel like you're building the AI, you're creating it. Can we go back to your time at DeepMind working on video games? I'm so interested in video games as a benchmark. I saw someone using Codex to play Slay the Spire, and I've played a lot of Slay the Spire. It's pretty difficult. Now, it's not a fast Twitch game, but I'm wondering if you, like, how would you think about the value of video games as a benchmark?

57:55There was another story about the FAA hiring flight traffic controllers who had previously played video games. So there's some sort of transfer where if you're good at video games and maybe SimCity, you might be good as a flight traffic controller. And you could imagine a situation where AI gets really good at playing video games and then becomes more useful in a whole bunch of different work-related tasks. But it feels like the gaming benchmarks are still toys. They're fun. They're people just doing them off on the side. but how do you think about the role of solving video games or testing models on video games in the modern era?

58:36I think it's a great idea, but I might be biased. I used to play a lot of video games growing up. I still do some gaming from time to time, so I think it's awesome. I think the idea here is you want to test your AI on problems that it hasn't necessarily been trained on directly. So you want to have some level of generalization. and games are amazing because they have all kinds of complex things you're going to have to do, all kinds of problem solving you're going to have to develop on the fly and it's kind of measurable how much progress you're making. So if you're getting to the end of the game, you're doing well, you're making progress from one level to the next.

59:10That's measurable. So they have many, many properties that make them pretty ideal for measuring AI capabilities. And the craziest thing today is we have these powerful agents that have been trained mostly on coding tasks or give them some code base and ask them to fix a bug or develop a new feature and they go out and they do all this tool calling to figure out how to do it and rewrite your files. But then you can also hook them up to a game and give them an API like here's how you control the units in the game or here's how you manage your resources or here's how you move around in Pokemon and other games like that.

59:44And I think it's crazy that these models are so capable at playing games and just shows you how much we've how far we've come in terms of generality. Yeah, where are the shortcomings, though? Because I can think of one, which is like with a game, you can basically run an agent, have the agent play a game effectively infinite amount of times. It can fail a lot. It can learn, things like that. One of the challenges is in the real world, like let's say someone was making a sales agent, like an agent that wants to help you get customers. You can't necessarily just let the sales agent run wild in the real world, as a business at least, because it's going to mess up a lot.

1:00:22A bunch of customers are going to have a bad experience. And you could make maybe the agent get slightly better from that experience, but you could have lost a bunch of potential customers or pissed a bunch of people off or things like that. And so how do you think about making the jump from agents that are very effective at playing these games in a generalized way to agents that can be effective at long-running tasks in the real world that involve effectively complex groups that are third parties? That's a good question because most of the training today is done with synthetic environments. So you're building up these RL environments inside of your AI team and you're training the models with reinforcement learning.

1:01:07And so you kind of go for a simulation, you could say, and they're not really interacting with the real world when you're training them. And I think one of the big frontiers right now, one of the big developments you might see is the training moving into an online setting where the models are directly interacting with the users, with the companies that are using them. And as they're solving tasks, as they're figuring out what to do, we update the weights of the model, they get better and better. And it's a big research problem right now. So nobody knows how to pull this off in general. And the best agents, they've all been trained in simulation so far.

1:01:38But it's one of the things that we're working on at River AI. So if any of the viewers are interested in doing some research on this, reach out. Yeah, it feels like we're not that far, at least in the gaming sense, to, you know, in the training step, just create an environment that's just like, here's a Steam account and a credit card, go buy every game and try and, you know, get the platinum trophy or like complete the game and feed that in. But the gaming thing is sort of a pure benchmark at this point because it feels like the labs haven't identified it as something that really they want to focus on.

1:02:13Can you talk about the trade-off between bench hacking for good and bench hacking for bad? Because there's the game that the labs are playing. But then there's also, if you show up with a product and it does the task and it classifies all of my taxes, I don't care if you bench hacked on that as long as it gets the job done, right? And so there's this push and pull between those. How are you thinking about communicating that with your customers, the companies you work with who might be fine with a model that's only good at their specific task. Yeah, exactly. I think that's a big opportunity for any company out there today because you own your own data that you've collected from your customers or from the work that you're doing.

1:02:56And if you eval the systems on that data, this is like the perfect eval for you. If you're able to improve your models based on that, you might end up owning the best model in the world for your particular task. So I think that's actually huge for companies. They should be building these specialized evals and they should be trying to build their own models. and the River API makes it easy to fine-tune your own model, given your eval, given your own training environments. But in general, when it comes to AGI and improving the intelligence of these models, I think we want to hit them with some surprising benchmarks.

1:03:28If we want to measure generality, then throwing in a game that it hasn't been trained on, that it's never seen before, I think that's a very, very interesting measure to see how far out of distribution can they do interesting things we haven't trained them for. Huge, huge fundraising round. AMP is in. We've talked to Ajne a bunch. And he has a very interesting thesis around actually going much deeper in the stack, acquiring compute. How are you thinking about the uses of those funds? Because I could see coming to you as a company and knowing that you have the capital to go and really optimize all the way down to the stack and become a neocloud, build a data center for me, or help me with the more expensive CapEx piece of the puzzle.

1:04:12At the same time, I don't really have a solid frame of, if I come to you and I say I want to fine-tune a near-frontier open-source model, is that actually that expensive? And wouldn't you just ask me to pay for that up front? So that doesn't seem like a huge capital cost to you, but what is the shape of the cost that you are planning on incurring over the next couple of years? Yes, so we're charging the customers. really buy the token. So if you have a fine-tuning run you want to do, if you have an RL run you only pay for what you actually do. Oh, interesting. Choose some kind of model training run size that's perfect for your task.

1:04:50Some customers, they end up training a really, really small, really fast, efficient model because they've got the best data for their task and they end up beating the largest and most expensive models out there. Other customers want something more general or they want to utilize these larger open-weight models like Kimi K3 and others that are coming out. So that's a more expensive training run. But yeah, we obviously need a lot of access to GPUs to make that happen. The funding helps with that. But even if you have the funding today, you still need to get access to GPUs and GPU prices are increasing steadily.

1:05:23So I think what we're going to see is more and more investors collaborating with our portfolio companies around compute, bringing up GPU capacity, distributing it among the portfolio companies. Maybe some of them need a little bit more in one month than others and so on and so forth. And it's a new kind of strategy that we're seeing, I think, to deal with the fact that compute prices are going up with so much scarcity. Yeah, would you say that's one of the biggest challenges for River at this point is just compute planning? I mean, we've seen the full spectrum now. We've seen Sam last year getting really, really, really aggressive.

1:06:03and then we saw earlier this year Anthropic just sort of being caught off guard by the growth and it feels like that is as a CEO. It's a completely new skill set. Yeah, it's like a new... There's never a moment where Mark Benioff was like, I don't have enough servers for Salesforce, I imagine. They were a way different problem set, but now it's demand planning is like a key skill set for a CEO. Exactly. It's a totally new skill set that's super important. And it's so difficult because as a startup, by definition, you have this variance for the future. You don't know if you're going to go by 10x or if you're going to go by 3x over the next 12 months.

1:06:43So you have to play this really, really difficult poker game to figure out what is the right allocation for me or maybe create some deals that are more flexible so you can actually scale up dynamically as the demand is growing. So I think we're going to see more and more of that happening. And yeah, so we're bringing up quite a bit of GPU capacity, and it's been important both for research and also to power the API. So as more people are starting to tune their own models, it's becoming very, very popular now. A lot of companies are reaching out a lot, wanting to build their own custom models that they own and trying to train on their own data using their own evals.

1:07:20So demand is going to keep going up, we expect. How do you think about custom silicon over the next few years? YouTube, I believe, has a custom silicon chip for encoding video very efficiently because you upload one video. They need it in 360p, 480p, 720p, 4K HD. And then we saw Talus bake the weights of Llama, I believe, into their chip and prove that that was exciting enough that AMD acquired the company. And I'm wondering if you imagine a future where a company comes to you. You can imagine the Visa network. We want to run a transformer-based model on every transaction. And it's going to be trillions of prompts, effectively, or more.

1:08:07And so custom silicon might actually make sense. But then the models jump forward and you can do batches. And there's so many different tradeoffs. How do you think custom silicon will play in the diffusion story of AI? Yeah, I think we'll see more and more custom. Silicon, obviously, companies like NVIDIA and AMD are also going to do extremely well, especially on the training side. It's really unmatched what they're able to do. On the inference side, there is some room for optimization because we believe that with personal AI agents coming up, so this is kind of the next evolution of agents after coding agents, which have been super successful.

1:08:46We expect that the demand for tokens will go up even further. so to the point where we don't even know how we're going to serve all these tokens for everyone given limited data center capacity given all the bottlenecks in the data center supply chain so I think we've got to be smart and start to develop some custom silicon specifically for inference of these personal AI agents we could imagine being much more power efficient with these kinds of chips can we maybe bake some of the transformer architecture into the chip to kind of exploit the fact that we know what kinds of models we're going to run so that will make it more specialized.

1:09:20You might not be able to run any kind of model anymore. So Talos takes that to the extreme so you can actually bake in the model weights. But today, you're not able to fit weights of very large models on a single chip that way. So that's the big bottleneck. So you're only able to do maybe 8 billion parameters or something like that, whereas the best models have trillions of params. So hopefully we're going to see some chips that are able to run those top-of-the-line models very, very efficiently. Do we need more Neolabs? Or, you know, basically, is there enough idea space that more people should be spinning up entirely new labs?

1:10:03I imagine a lot of the people that would be candidates to spin up labs themselves, you're probably trying to recruit, but are we past peak Neolab, or are we just getting started? Hopefully we're just getting started because this idea of building powerful coding agents, having APIs and so on, we all had this a few years ago at OpenAI and other places that that would be the future. But now that it's actually arrived, I think a lot of us are feeling like this can be the end of the line. There has to be a different way to own and build AI systems. And so that means there's the opportunities for research, the opportunities for new kinds of business models.

1:10:40Totally new talents can move in. somebody who doesn't have a big name in AI can do something really amazing today because they have to think out of the box, come up with something that the AI expert that's been doing it for 10 years, they might not come up with it because it's such a wild idea. So I think we're going to see a phase of innovation, totally new approaches to how AI systems are built, how you make use of them, and some really cool AI products as well for consumers. So that's what I'm looking for the most. yeah what an exciting time well congratulations and thank you so much for taking the time to come chat with us excited for the next one yeah we'll talk to you soon thank you guys cheers rest of your day goodbye let me tell you about mongo db what's the only thing faster than the ai market your business on mongo db don't just build ai own the data platform that powers it our next guest is the founder of core weave we have brandon mcfee coming back on the show after an insane q2 congratulations give us the headline numbers how'd you do in q2 q2 was phenomenal beat for us right um that that couldn't have come in better um we're extremely excited with the performance of the business uh and balance of the year is just going to keep on improving from here what's the biggest bottleneck now for you is anything changing based on all the data that you've collected over the first half of the year?

1:12:07No, like bottleneck remains supply chain, right? Like that is the hardest part of this business. And it's also what we're best at, right? We have 51 data centers in operation today. We know how to navigate this supply chain. We know how to get this stuff online, deliver the clients do so on time. It's immensely challenging, right? Like this is the most important commodity on the planet right now. And CoreWeave is singular in its ability to deliver the best performing infrastructure out there. You guys are spending a lot of time doing demand planning. Demand is obviously off the charts. What are you, we just had Igor on from River.

1:12:45We were talking with him about this sort of like new skill set that technology founders need to have around demand planning, right? You have a, let's say you have a NeoLab, you're developing products. It's very hard to gauge how much demand you're going to have for them. the right strategy over the last couple of years was just to assume that demand was near infinite and sort of plan against that. But for smaller companies that have smaller balance sheets, it really feels like that is going to make or break a lot of these companies, especially Neolabs that their margin profiles and all these other factors are going to come down to how well they can predict their own demand for their products.

1:13:28Yeah, I completely align with that sentiment, but I'll probably take it a step further. Like, what happens after that demand is planned for us? How do you finance it? Right. And for us, that's been just an absolutely critical skill set that we've assembled team-wise over the last six years. And the latest financing that we did, DDTL5, was a term loan B offering I thought was a Fantastic example of that ability to work with demand planning and get the right financing in place for that demand. That was the first facility that we've done where the contract duration is actually shorter than the amortization period.

1:14:09Right. In other words, it asks the investors in that facility to take on renewal risk on the compute versus all of our prior facilities. investors were fully covered by the original contract value, right? If it was a$5 billion loan, there was$7 billion worth of revenue sitting behind it, right? Now it's opposite, like it flipped. And what that enables us to do is to spend more time in the shorter duration contract market, because we just proved that it's financeable. And for us, that type of client is predominantly enterprise, right? The AI lab cohort, the hyperscale cloud cohort, they like to sit in the five to six year range because they know that they need that compute, they need a scale, they need it for that long duration.

1:14:59Enterprise is at the shorter end of that curve. And as you guys know, historically, we've been really focused on the longer duration contracts, but having this proof point in the market now that we can go finance these shorter duration contracts is really important to us. So I think that goes hand in hand with the demand planning aspect. There's a ton of hyperscalers that have been building data centers for a long time. There's a ton of neoclouds. How are you positioning differentiation? I imagine that demand is so strong that it's not the biggest thorn in your side by any means explaining the value, but it feels like it always has to be in the back of your mind as if demand ever softens, we want to be differentiated.

1:15:43We want to be differentiated even in a market that's tight on demand. So how are you positioning CoreWeave specifically? Look, I think it's widely recognized by our clients, by third-party analysts, by our suppliers even, that we are the best representation of this technology on the planet. So that demonstration of value will continue to be through our technology product. It'll be through our timeliness and deliveries, our ability to scale across all different types of workloads, whether it's training, fine-tuning, inference. It'll be in delivering the best performance-adjusted flops and tokens out there.

1:16:27Jensen was on CNBC recently talking about a$500 billion deal with a lot of big banks. What was your interpretation of what that deal means for the industry overall, where that project is going, and how Corey fits into that story? I think it's wonderful for the space, right? It just continues to show the amount of capital that's willing to underwrite the buildout of intelligence. As you guys know, we've been at the forefront of financing this market for years. I think having more capital in here is just great for the sector. How have you been coaching people through? There's some folks that are like, ah, I'm worried.

1:17:08This is like.com. There's other people that I'm worried about this being like the mortgage, you know, boom. What, what pieces are different this time? What pieces are, okay, we are actually borrowing from this particular build out. A lot of people go to railroads. There was a huge amount of value created. There were booms and busts in various times. How are you dealing with the various critics of the build-out? So one of those points, I feel like we talked about this last time I was on, I believe, was depreciation or useful life of compute. And the variable that we were able to introduce and highlight in our earnings is the A100 SKU.

1:17:49This is a 2020 SKU. and we still have clients coming in asking specifically for that SKU, right? It's not like they're asking for Hopper and like, oh, we only have Ampere or they're asking for Blackwell and we only have Ampere, right? Like they're saying, no, we want Ampere for workloads because that is the most performing platform for their workload. And we signed a contract in the quarter that goes out through 2029, right? That is not an implicit. So it'll be a nine-year-old chip by the end. Nine-year chip, right? And this is a take or pay fixed price contract. And pricing on A100s for us have held solid since early 2025.

1:18:31I think one of those big criticisms was, well, this is two or three year compute and it's useless after that. We've been very consistent that six year depreciable life is accurate for this infrastructure. And I think that there is really strong opportunity for it to have material useful life beyond that. And we see it every day. Right. And it's just in the way that AI workloads are being optimized across different skews. And it's this concept like there isn't one AI model to rule them all. There is one GPU to rule them all. It's just this matrix of different sizes of workloads relative to different sizes of GPUs where it's filling in across.

1:19:11Yeah, my thesis has been that there are AI workloads that get built out, maybe a recommender system, a basic text transformation system, a translation system. And then those will stay in place for a really long time and those will be used by more intelligent models. But then there is the question of just will the chips burn out? Has any thought changed there or is that pretty well understood? that right like this stuff is meant to run the data center um it's meant to run for long duration we're not seeing any acceleration of burnout or like unexpected rates of errors uh the infrastructure is doing great yeah that makes sense uh what what is going on on the on the power side i imagine there's a lot of people over the last year that have realized that uh how how important power is to this whole build out and are trying to front run, you know, players like yourself and get access to that power in hopes that they can resell it.

1:20:12I imagine the utilities are like somewhat sophisticated and at this point, and hopefully for a while, we'll just call around and try to go directly to the operators. But like, what can you share about like the current dynamics around that hunt for power? I would say the power market is very competitive, but there is power out there. And the bottleneck is less electrons. For us, it's more a powered shell or like delivered data center capacity. And that goes all the way down to the components that are in the data center. Think of the backup battery supplies, transformers, et cetera. But it's also the people.

1:20:55Electricians is a skilled trade that is a significant bottleneck for the industry. Has been for some time, will continue to be for some time because that's a trade that takes years to develop the skill sets necessary to be able to work in a data center site. So I believe that that is going to remain more the bottleneck than power is right now. You're absolutely correct. I can't cap the number of emails I get a day from people I've never met before trying to offer us powered land and all across the U.S. But that is less of the bottleneck and it's more on delivered powered shell capacity. Are you hiring electricians directly or does this go through subcontractors for specific projects?

1:21:42This predominantly goes through subcontractors. We have a little bit of self-built ourselves where we're going out and sourcing GCs, contractors, etc. But for the most part, we lease capacity. And that's been the way that we've scaled business. I imagine that there's a lot of folks in the organization that are technologists. They understand the technology and the hardware and the software and how everything pieces together. What are the other areas that you're hiring if somebody wants to join the CoreWeave organization that draws from a different pool of human capital or talent? Going back to the point originally, financing.

1:22:21right it's deeply important to the business um you know we've we've done a fantastic job assembling a team up in new york we've raised uh i think it's north of 40 billion dollars in debt and equity over the last 24 months yeah uh and are pulling people from wall street investment banking hedge funds consulting yes equity like all of the above yeah i'd say predominantly private equity private credit okay investment banking guys like deep financial backgrounds yeah yeah But look, we're hiring across the business, right? It's engineering, it's physical deployment teams. I think we're, it is, our HR organization is doing a great job.

1:23:00Let's put it that way. Yeah. Busy. How domestic is the footprint? There's a lot of nervousness about the build out in America. And my default interpretation would be if there's a lot of pushback in America, like people don't mind waiting 500 milliseconds for LLM responses. So, yeah, you can put them in space, but you can also put them in, you know, another country. How are you thinking about the international opportunity? I think that's right. But I would qualify more on the county level. Right. Like we're getting pushback. The whole industry is getting pushback at the county level in some places.

1:23:33But that ultimately doesn't change the fact that the demand is there. Right. Getting county level pushback, state level pushback. It just means that it's going to be built elsewhere because the demand for AI is the strongest it has ever been. The ROI on AI is the strongest that it ever has been as well. Inference is absolutely profitable for our client base, which means that they're going to keep coming back to core read products. We're in Canada. We're in Europe. We recently announced an expansion into APAC as well. I expect for us to keep moving globally, but keep a focus on domestic deployments.

1:24:16How strong is the correlation between social media pushback and county level pushback at like a city council meeting? Because I feel like there's sometimes a big disconnect where something can go viral. Maybe it has some misinformation, gets 100 ,000 likes, but then the actual county and the residents are fine with what's going on. It's sort of this like getting mad on behalf of someone. but do you see a correlation between a story or a news article that happens in this particular area? And then there is actual movement on the ground at the local politics level. Look, I'd say it's very specific to the sites, right?

1:24:52Like we, we are engaged in conversations where, wherever the local communities would like to understand the value that we're bringing to that region. Yeah, that makes sense. Jordy, anything else? Not for now. Congratulations. And thank you so much for taking the time. Big one. Thanks, guys. Appreciate it. Great to see you. Goodbye. Let me tell you about Shopify. Shopify is the commerce platform that goes with your business and lets you sell in seconds online, in-store, on mobile, on social, on marketplaces, and now with AI agents. It's very interesting. It feels like one of CoreWeave's and some of the other NeoCloud's advantages is to just be able to be completely out of the politics and the drama of AI, right?

1:25:31Oh, yeah, yeah, yeah. All of the labs and the hyperscalers and all these companies, they have financial relationships they have personal relationships some of them some of them hate each other and want each other dead yeah uh core we've can kind of sit there like switzerland and be like anyone need tokens yeah yeah that's true uh when a niche hit when a niche hit tweet fails to bang james heel shares a quote from uh peter j uh when a sub editor when a sub editor complained that the Times columns were too complex. Peter Jay famously replied, I only write this for three people. The editor of the Times, the chancellor of the Exchequer, and the governor of the Bank of England.

1:26:19True, not quite audience of one, but audience of three mentality. I think it's a good way to stay sane if you're posting online. You probably don't want to be too sucked into the algorithm, although this went mega viral. I don't know. Anyway, we have our next guest, Garrett Langley from PhloxA. He's the founder and CEO. He's been on the show before, and we're very excited to talk to him about the latest news. Garrett, how are you doing? He's back. Good. How are you guys? We're good. Welcome back to the show. Thank you so much for joining. Anything been happening since the last time you came on the show?

1:26:51I saw the news in the show. Pretty chill. Pretty chill. I can imagine. Pretty chill. I can imagine. Well, let's kick it off with the report in the journal. Flock adds privacy guardrails after surveillance backlash. What happened? What are you actually implementing? What is the story? Yeah, I mean, there's two big buckets, right? You've kind of got privacy and accountability. And I can start with accountability because I think that's the bigger topic, or I think it's the bigger topic, which is, lo and behold, law enforcement abuses their power at times. And that's horrible, right? I mean, it's really bad.

1:27:27And we built a tool about four months ago as a test to kind of scan the audit logs that we've always had to look for abnormal behavior. You know, we solved a million crimes last year. We know what good investigations look like. We built this tool. And the headlines over the last few weeks show you that we're pretty good at finding abuse. And I thought this was maybe just an isolated to us. But what I found out is when law enforcement searches DMV, they search criminal records, there's no audit logs there. There's no accountability. So we launched that. And I think you're going to see over the coming weeks more headlines of officers being arrested for abusing their position of power.

1:28:08So that's the big story. Where does the oversight board live? Is this a separate product that's sold to a district attorney effectively? Because if the person's like, yeah, I checked the audit logs. because I was behaving poorly. That doesn't really do anything to stop the process, right? No, yeah. So it's a free product that all of our customers are now required to use. Oh, okay. Which we're the only company that I'm aware of in this industry that requires their customers to actually hold themselves accountable. For most of our cities we work with, there's a police administrator and then a city manager that is reviewing the audit log.

1:28:45So in the hopefully unlikely event that the police chief himself is the one committing a crime, the city manager is still there as the double check. Yeah. So the business has grown a ton and obviously Flock products are deployed all over the United States. I'm interested in the backlash that has grown. You see a lot of videos with a lot of social media attention negative towards Flock. Is this actually showing up in your financials at this point? because I imagine you have to replace any camera that is destroyed, but I can't tell if there's just one camera that's destroyed and then a million people view it and like it.

1:29:28Yeah, no, vandalism is real. It's always been real, though. Just like weather has been a problem. We deploy these products in Florida and there's hurricanes in California and there's earthquakes. And so we're used to that. I mean, if you looked at the data, you'd be hard-pressed to say, when did this TikTok trend blow up? Sure. You know, I'd say the sad part to me is you've got these influencers on TikTok and Instagram pushing this and they're convincing these 16, 18, 20 year old guys to go do it. And they're committing felonies. And so I saw this guy that got arrested in New Mexico and he's going to face five to nine years in jail.

1:30:04Wow. And I'm like, that's horrible. This guy's life is over. He's got a felony on his record. He's going to jail to cut a camera down. It just seems really crazy. So one thing that I've been identifying is it feels like there is a gap, particularly between liking a post on TikTok or engaging with anti-flock or the deflock campaign and and maybe even going out in the world and chopping one down versus at least trying to go to your city council meeting. Because I just have to imagine, I've been to city council meetings. If there's 50 people there that are all taking the podium and saying, look, I am a citizen and I don't want flock in my city, city council will usually respond to that.

1:30:53Is that disconnect real? Like what is going on there? Yeah. I mean, look, it's surprising to me too because city councils are normally pretty boring meetings. and there's really only two topics that are being debated right now, which is data centers and flock. And you guys probably have better tools than I do. You can look at social media traffic and they're fully correlated, which makes me wonder, like, why? Because I've been asked this question, why now? And I'm like, I don't know. We've been in business for almost a decade. You know, we went to thousands of city council meetings every year and over like 10 ,000 last year.

1:31:31We're not hiding anything. Yeah, I mean, to be honest, I think a large part of it is like, well, everyone's been talked to about AI safety one way or another, right? It might have been a viral clip of Sam Altman talking about AI in 2015 that gets resurfaced in an Instagram reel today, and people assume that it was said today, and the clip was out of context. um and then but but you have to look at like uh i think i think people are like very scared of of of the technology overall uh which i think is fair because they've been told to be scared in a variety of different ways um and then they're using ai in their in their life or in their work and and that's cool but like data centers present something that's like you know a data center in your backyard doesn't necessarily give you an immediate benefit and it could give it could have some real downsides, right?

1:32:29Like noise or potential, you know, like it, there's all these sort of like possible negative externalities. And then flock, it's like when you, when, um, you know, Dario's talk, talked a lot earlier this year about AI getting, getting so good at looking over sort of like vast quantities of data and maybe, maybe the law is not sort of like staying, you know moving quickly enough to respond to that and so like flock is like a very i i can see exactly why people are fixated on that because the product uh is clearly very very good at fighting crime but then also also if you are a bad actor it's like the dream like basically tool set right it's literally like uh in a video game like god mode right and so i think like there's like very real, I can see why there's a correlation but I don't know if this is where you're going but I don't necessarily think it's some organized nefarious group that is like we want to shut down these two technologies.

1:33:39So anyways. I think the other thing that I think is true is I was asking someone aren't you afraid of all the data brokers that buy your location data and sell it. And they're like, well, I can't see it. And I can see your camera. And I think it's like a valid point, which is you drive by a flat camera and you say, well, I see it and I say, I feel a lot safer. And some people might say, oh, I feel like, you know, this is Big Brother. And I go, it's like, but Big Brother wouldn't have an audit log and wouldn't have transparency and wouldn't do these things. But I do think a lot of the feedback is fair, you know, which is like, how long should this be stored?

1:34:16And I think the thing we got wrong was, expecting local government and state government to follow along fast enough. And I just, I don't think they are in this accountability measure. Like no states require this. They all should. And like we dropped our data retention recommendation from 30 days to seven days, which means the majority of our customers will just simply accept the default. Sure. And state regulators have moved to 21 days, but like we've done the analysis and we think we'll have a 10 % reduction in efficacy with that reduction, but I think it's a fair trade-off. And look, if the government wants to say, like in New Jersey, it's a mandate five years of data retention, that's up to the state of New Jersey to decide.

1:34:58We'll push for what we think's right. But I might agree with Dario there in that case, regulation is a good thing for this type of technology, and it needs to move faster, not slower. Yeah, it seems like the speed of regulation is a huge issue here because the technology moves much faster. And I think folks on the left feel like they could be attacked by people on the right and vice versa. And so if there's a new technology that rolls out like mid-cycle, all of a sudden you're grappling with it and your worst fears on sort of like the extremes of the political spectrum are that they will be used by the opposing side aggressively.

1:35:40against you, even if you are innocent. And I think that's where a lot of this is coming from. What do you think? Yeah, I think that's true. And I think the other thing, and I don't know who gets credit for saying this, as a technology industry, we still feel like we're this bubble, just like a small part of the economy. And the reality is technology is the economy now. And I don't think as founders and CEOs, we've fully grappled with what that means. And I think about someone like Jamie Diamond. I'm like, he has so much influence over our fiscal systems. And I think he probably understands that responsibility because banking has always had this kind of austere posture of their responsibility in the world.

1:36:22And I think as technologists, we're catching up. And I think I'm in that same boat of, we helped a million crimes last year. Last month, we did 1 ,025 missing peoples. That's real human life. There's a real responsibility. uh, with that impact. And I'm glad we're catching up. We're not done. Like we still have a lot of work to do, but I think if I look around the table of my other peers, like in public safety, I think we all have a lot of work to do. Yeah. I think, yeah, the, the, the, there's been a very real positive impact of the product. And I think like people have, um, every reason to also be concerned because even with these guardrails in place and audits, you You have, even if you assume like a very, even a half a percent of police officers in the U.S.

1:37:11are corrupt in some way or not, shouldn't be in these organizations. That's still thousands of bad actors that are in the system. And so I think one of the challenges that I see is like, I feel like this is like too much of an, too big of an issue for just you and the flock exec team to be in charge of. Right. Like, I think you're very smart. I think this company was started for all the right reasons. I think it is having a positive impact in many ways, like you said. But at the same time, like this is a national issue right now. You know, you're not elected. It's not your job to decide, you know, Americans like privacy.

1:37:53And so I think like my question is, like, what is happening at the national level? Obviously, counties and cities and states are trying to figure this out. But I feel like this issue is like beyond. I don't feel like this issue should be decided by you. I think it should be decided by our lawmakers because it comes down to our fundamental rights as American citizens. So I actually I want this. The issue is just blown up on on social media. But I think it's time to like have a much more of a national kind of conversation about this. And then ultimately, probably new laws put into place to make sure it's something that you can point to and just say, like, look, like we are abiding by the law in our country and our lawmakers who were elected.

1:38:41And the correct path to push back on flock is through the democratic process. Exactly. Call your representative. That does feel like the correct outcome. That feels very self-obvious. And it is interesting, you know, I was getting, saw someone push on X, like, you should require a warrant to use your product. And the response I wanted to make was, yes, when it is the law to need a warrant, we will happily follow the law. But 40 courts in the last few years have all deemed this isn't, you know, a violation of the Constitution. And this is tricky, though, because every state, every city has like a very different point of view on what safety looks like.

1:39:21Yeah. I mean, you look at you guys, you know, hometown of San Francisco, San Francisco's own version of safety has also changed a lot in the last decade. Totally. And I think it's a better version today than it was five years ago, but it's still changed and it'll probably change again. And so it is a pretty tricky situation from setting the right laws. I struggle to see how the federal government could establish a law that matches everyone's goals. And so my hope is like states, we had nine state bills get passed this year. I'd love to see 20 or 30 next year that at least start to move this in the right direction.

1:39:56Because I do worry like any federal regulation will just make half the country mad and other half the country just as mad. In theory, like city, even city by city, like there is a like I think people might misunderstand that like Flock doesn't put up the cameras. They sell them to police departments. But the cameras in those police departments have heads that are elected in many cases, right? There's like a sheriff that gets elected. There's people who get to pick that they vote for. But at the same time, I'm sympathetic to someone that basically thinks you put them up because it is an opaque process.

1:40:31Many Americans cannot tell you the names of everyone on their city council and who the mayor is and all these things or when the election is and how solidified that particular mayor is. and whether or not they can actually apply pressure around this issue. It can take years to actually facilitate change in a community. And if you're in the minority, that might never happen for that particular community. But yeah, yeah. I was referencing earlier was like more like friction. Like we, there's historically been security cameras around in a lot of places. You could get access to the footage, but there was a lot of friction to doing that.

1:41:06Sure. Which is good and bad, right? It's good in that a bad actor is going to have a tougher time putting together footage in order to understand someone's movement. But again. Yeah. Yeah. Everyone's comfortable with like there was an assassination attempt on the president. We should look at all the CCTV cameras. But then all of a sudden when it's like, okay, now there's going to be an AI agent that's watching you, Jordy. If you speed at any point during your commute, you're going to get a point on your license. It just feels like an annoying society that a lot of people don't want to live in.

1:41:42And then there's a whole continuum there. And we're grappling with the potential for a way, way, a way higher level of enforcement and way less of a gray area around our rule set and our laws, which is something that the laws might need to change or we might need to adapt to. I don't know. I agree. And I think to your point, like the, I have a lot of sympathy for our elected officials at the local level because they now need to become experts in AI, public safety, data centers, water. Like it's, it's like, and these are, most of these times, these are part-time unpaid positions. They're normally lawyers, doctors, real estate agents in their communities.

1:42:29And so, I mean, they sign up for the job, but it is, it's not straightforward. And we like the way the system works today where every one of our contracts goes to city council for vote. It makes the job harder, but makes it transparent. Like I saw the other day, I was able to find the video from six years ago when we did our first ever city council meeting and how nervous we were because we'd never been to city council. What would that be like? And I've never been to city council. And now it's kind of just a part of our regular day. What's next? because these updates are good, but I don't think they solve.

1:43:05I don't think millions of Americans are still going to have, are basically probably not going to read the updates and are still going to have an issue. So what's on the roadmap? What more are you doing to try to find a solution to the concerns and the issues? Yeah, I mean, I don't think we'll ever be done because I think this balance of privacy and safety is a bit of a moving target. Someone asked me the other day, wouldn't it just be better just to turn off all the cameras? And I said, maybe 30 years ago when every single police department was fully staffed, maybe even overstaffed to today's standards, but today more than 80 % of police departments are understaffed.

1:43:47So I don't think a draconian, no more technology is a viable solution, but I do think as a technologist, we have an opportunity to make better products that deliver on that kind of combined goal. And so for us, we're excited to get these tools out there into the wild. I anxiously await potential headlines of more arrests in the coming days. I think that'll happen. And that'll be tough because I think a lot of trust in communities will be broken. I think that trust was broken many years ago and now the light's being shined on it. But I think for us, we're going to continue building tools that allow law enforcement to both do their job and build trust with the community at the same time.

1:44:28So we've got a lot of stuff coming up. A few more announcements coming up in the coming months, and we'd love to come back and share those when they're ready. Yeah, that'd be great. Thank you so much for taking the time to come chat with us. The entire community in the chat is every possible take is being enumerated. It's very interesting. You want to switch jobs? Yeah, yeah, yeah. I don't think anyone in technology envies your job right now. but the consequence of being consequential is what I tell the team yeah that's accurate well have a great rest of your day thank you so much for the update and explaining everything we'll talk to you soon we've been keeping Sonia Wong from Sequoia Capital waiting too long she's a general partner it'd be great to just talk about venture capital now we get to back to AI back to VC how's it going?

1:45:22We haven't seen each other in a while. Congrats on everything. Congrats on the acquisition. It's been far too long. Thank you. Thank you. And good to see you. Yeah. So, temperature check. What's going on with this AI thing? Is there anything there? Temperature check. Oh, my gosh. There is absolutely something there. The numbers we're seeing from these companies. Yeah. Never seen anything like it before. But interestingly, it's less of a power. I mean, there's power law companies, Anthropic Open AI. There's a lot of companies that are doing really great. But then there's also this diffuse community of like Neo Labs and Neo Clouds and so many different application layer stuff where you're seeing just really solid business fundamentals that previously would take a decade to build up.

1:46:02You're seeing it in two to three years. Totally. Anthropic and OpenAI and XAI are growing at a pace that nobody has ever seen before. But even if you remove them, this next cohort of companies, Harvey, Open Evidence, Glean, Factory, they're growing at rates that we've just never seen before. One of the most interesting things that's happening right now is we used to have this separation in our heads of there's the foundation model companies and then there's the application companies. and I would say like one of the most interesting things that's happening now is that all the application companies are starting to build their own research capabilities, their own labs and it's kind of like this concept of democratized intelligence.

1:46:44So I would say it's not only revenue that's not just accruing only at the top two players, it's the production of intelligence itself. It seems like it's very much democratized. Yeah, we were, I don't know when Dylan was on, maybe it was last week, but Dylan from Figma. We were talking to him. I was like, I want Figma to work on the problem of design intelligence. Basically, Slop has been this like, you know, Slop has been getting better and better. But, you know, it's basically like you have a new breakthrough and then for two months it's like, wow, it's so good now and then you realize that it can only do like one style over and over and over and I feel like there's all these application layer companies that are in such a great position to work on some of these fundamental problems that for better or worse, the labs are not able to focus enough on.

1:47:40Maybe it's only a billion-dollar revenue opportunity. And so if you're a lab and you're adding billions of dollars of revenue a month with your core business, why would you work on a problem like that? And so I think there's so many examples of that from Figma to some of the other ones you mentioned. Totally. And I think if you look at it from the perspective of the startups, there's just this huge wave towards companies wanting to own their intelligence. And I think a smaller set of companies has been beating this drum for a long time. Like I'm on the board of Fireworks. Their tagline is own your intelligence.

1:48:11So they've been advertising this for a long time. Obviously, Cursor went on the journey two years ago of starting to train their own models. But what's happening now is like both the giants and the ecosystem are starting to speak up and the startups are actually getting extremely good at building their own research. So in terms of giants, you have people like Alex Karp talking about sovereign intelligence. He has this phrase, own the means of production, which I freaking love. Satya talking about your proprietary data. Jensen championing open weights. So you have all these giants of the ecosystem speaking up of like, hey, guys, you should own your intelligence.

1:48:43And then if you look at it from the perspective of the little guys, the startups who we are in the business of backing, a couple of years ago, they were primarily looking at moving some of their intelligence towards these open weight models, primarily as a cost rationalization exercise. The thing that's different now is like it is an existential and strategic imperative for them. And so there's this phrase, you guys probably remember this from the crypto days, not your keys, not your crypto. This idea of like, if somebody else is cussing your weights for you, I don't, cussing your keys for you, it's not yours because something could happen to that brokerage, something could happen there.

1:49:22and I think like I think the AI version of this meme is not your weights, not your product because fundamentally if you don't own the weights if you're just making an API call you don't have ownership, you don't have steerability the data flyable doesn't accrue to you and so like people are kind of waking up to this and so if your use case is like you want a coding agent like for that I'd say close model, close agents, that's fantastic. If it is like your core product, I think companies are increasingly waking up to like not my way it's what you sell to your customers it should exactly that makes a lot of sense uh founder office hours coach me through you're on the board of my hypothetical company uh software company i'm using ai 500 million arr let's say and and i and i buy this thesis and i come to you pretty confident example there not bad there we go we're cooking we're cooking no no no i mean like we have we have fully made it through we're you know an ai winner we're accelerating, growing, growth is great.

1:50:22But I come to you and I say, okay, I'm all in, we're going to own our intelligence stack. Is there a moment where we need to have a conversation about what the talent budget will be? I mean, we saw these crazy MSL deals. And for a lot of these companies, they might be unicorns, they might even be decacorns, but they're not going to be able to staff a team of AI researchers and build a full Neolab doing like next generation fundamental research. So what does the team build out actually look like? Is it enough to take your, your best software engineers and have them use coding agents to fine tune models for, for your team?

1:50:59Or are you hiring entirely new disciplines? What is the, what is the correct shape of an internal AI lab at like a successful scaled unicorn, decacorn software company. Yeah. So typically you're not going to be pre-training your own models. There are specific use cases where you actually need pre-trained models. That's a totally different thing. If what you are doing is trying to take off the shelf, open weight models and then adapt them and make them really, really excellent for your domain, that's a much larger talent pool. And what we've seen is there's actually two flavors of talent that are really good at this.

1:51:34One is people that have done post-training before. So the post-training teams at the labs are very, very large at this point. There's plenty of these people floating around. And then the second profile that's actually interesting is just like engineer, or I guess just generally smart person. Because this stuff is actually not that hard. And part of the reason it's even possible for all these companies to have their own labs now is because the actual post-training stack has matured. So it used to be that only OpenAI and Amtopic had the infrastructure in-house. to be able to do things like post-training, reinforcement learning especially.

1:52:10But now you have companies like Fireworks that gives you the post-training infrastructure. You have LangChain that gives you the evals. You have Trajectory that helps with the continual learning. You have Mercor that helps you with the data factory stuff. And so all these components now exist. And so if you as a generally smart person see the menu of opportunities, you can actually cobble together your own research stack in a way that wasn't possible a couple years ago. And so like to give you a sense, very, very small teams can get very far. The Harvey team has put out, I think, pretty extraordinary research.

1:52:46They just put out an entire RL environment last week. Their benchmark is state of the art for legal. And their entire research team is seven people. So you can get very, very far. We're definitely not, you know, you're not competing with Meta or OpenAI for size of talent budget here. What about one click down? I don't want to invest in building the team, owning the stack entirely. There are Neo Labs that will show up and fine tune a model for me. Is that the domain of growth stage startups? Or is that product going to be more consumed by enterprises that maybe don't have the DNA to just move a bunch of amazing engineers over to do a post-training stack and spin up and roll their own?

1:53:32But the NeoLab offers, I'm thinking of like a Thinking Machines tinker, like what happened with Ray Dalio's fund and how they were able to fine tune a model, get really good results. That feels like its own new market of like post-training as a service, fine tuning as a service. But how does that piece in? Is there a world where you don't necessarily have a relationship with Mercor, but your third party does? Yeah. Look, this is a huge ecosystem and a huge market because I think everyone sees the opportunity of, you know, you obviously like the closed model inference market will always be gigantic.

1:54:09Yeah. But I think the open model inference market is becoming very large. And for companies to actually be able to make use of open light models, there is a maturation process. There is a handholding process. There is an entire like, you know, come work with us. We will forward deploy people onto your staff to help you go on that journey. And so there's lots of options. Fireworks has a fantastic team for this. Mercore has a fantastic team for this. It kind of depends on the specific problem you have. So the specific problem you have is, hey, I really want to do reinforcement learning on my online data.

1:54:40Fireworks is fantastic for this. If you're like, hey, I can't train on my customer data. I need to get my model really, really good for this specific domain. Can we create a bunch of synthetic data around this opportunity? Mercore is fantastic for it. So it kind of depends on the use case. generally i think that it's like really important it's an and like companies need to have extremely smart people in charge of this in-house this can't be something you outsource like your lunch menu outsourcing right this is so core and so you need to have smart people in charge deciding on your strategy deciding on your technical roadmap and then making judgment calls of which parts of the stack you want to outsource to others which parts of the stack you want to lean on others for harvey's actually done a really good job of this.

1:55:23They lean on pretty much all of the five or six in your labs I just mentioned as their partners, but you need to intentionally own it. And over time, just like we saw Cursor go on this journey, I think a lot of application companies will go on the same journey that Cursor did. Over time, you become more and more incompetent, more competent in-house and you bring more of that expertise in-house. Sure. Sure. How are you processing the current price war. You have sort of the lagging labs starting to compete more on price. You have leading labs trying to make sure that they have competitive models at every part of the curve.

1:56:00But yeah, anyways, what's your take and where does this go? Look, I think Jevin's Paradox, not to be an annoying VC, but Jevin's Paradox is a freaking wonderful thing because what's happening is like, I see the data for all this stuff, right? Oh, there we go. Bam. Look, our companies, we see the margins going, their gross margins going up because at the same time as like AI usage is going way up, but their gross margins are also going up because they are the beneficiaries of intelligence getting cheaper and cheaper to meter. On the other hand, I see this from fireworks. I see this from the model companies we're in business with.

1:56:34Their cohorts are getting better and better and better. So it's not like they're facing price competition and their businesses are going into the gutter. They have phenomenal businesses where because they're able to provide intelligence at increasingly cheap prices, their businesses actually get better and better. So I actually think this is an everybody wins situation. Yeah, part of the challenge with, I feel like, X trying to process the price wars and how open source is fitting into all of this is that when you have Anthropic and OpenAI still, as private companies, people just don't have the visibility and they don't realize that even with great open source models and even pricing price cuts across different providers, you're still seeing that just massively accelerating revenue.

1:57:19And it's not like the biggest customers aren't aware that open source exists and it's good, and they are using it in a bunch of different ways. So I think it'll be very helpful when more of these players are public just so that everybody has the same access to information. Totally agree. Thank you so much for coming on the show. Let's do it again soon. Great to see you guys. Do it again soon. Great to see you. We'll talk to you again. Goodbye. Let me tell you about public.com. Investing for those who take it seriously. They've got stocks, options, bonds, crypto, treasuries, and more with great customer service.

1:57:58You ready for this next one, Jordy? You're going to love this company. They put a brain in a vat, basically. Really? Yeah. We have Sean Cole from Parasma. He's the founder. This is his first time on the show. Sean, how are you doing? Hi, guys. Thanks for having me. Welcome to the show. Is Brain in the Vat appropriate, or is that derogatory? Should I stay away from characterizing your company that way? I think it's pretty good. I think it's pretty good. Okay. Take us through it. Introduce the company. Introduce yourself. So I'm Sean. I'm the founder and CEO of Prasma. And we are training brain cells for compute.

1:58:35So previously, I think you guys might have seen, you know, we made brain cells play Doom. Yeah. And we just launched, like literally just launched. And we got these cells to do token prediction. So kind of the basis for language modeling. That's amazing. Okay. So how many cells do you need? because many people have said that I have seemingly only a few brain cells. And so how many brain cells do you actually need in order to do X token prediction? I can predict tokens all day with just a few brain cells. It's no problem. So it seems like an easy job for you. Not that many. So we're renting some brain cells out.

1:59:08We're using about 200 ,000 brain cells to do token prediction. And it's good enough. Are these literally like donor brain cells from cadavers or are you growing them? Your brain cells. Are these stem cells? Like where are they coming from? Walk me through the full process. What's your supply chain? I think you got it exactly. So we're using sort of potent stem cells and we're taking these stem cells, differentiating them into different types of human neurons and putting those in a dish that we can stimulate and get responses from basically. Okay. Okay. And then you said you're able to do next token prediction.

1:59:46I imagine that you're not anywhere near the frontier, but how are you actually benchmarking what is capable? Because this is probably all predicated on a very extreme exponential kicking in at some point, but where actually are we in terms of progress? I think somebody gave a pretty funny example, which is it's currently at this current stage. The token prediction is basically, is this a hot dog or is this not a hot dog type of token prediction? It's pretty basic. Well, can you say, say you're alive? And then it says, I'm alive. Yeah, we could get it to do that. You could, okay. I think fundamentally what we've done is we've proved that these types of sequential context-based architectures with electrical stimulation are a viable architecture on these biological substrates.

2:00:36We had a specific example where it was a nonlinear task. So let's say that to refer to context early in the sentence, and if we used a linear decoder, so traditional silicon, it could only maximally get 75 % mathematically. But we got above that, so we got 78%. We beat silicon on this very constrained task because it was able to model these nonlinear dynamics in context. How long do the cells actually last before you need to replenish them? So I think currently they last six months. But of course I think human brain cells last far longer, like 100 years, 80 years. So I think the goal obviously is to extend them for as long as possible.

2:01:23You were kind of getting at this, but not fully. So why do this? I have some ideas of why you might, but it seems like kind of a hassle, so you've got to have a good reason. Is it like... What do you have against silicon, okay? I'm assuming there's energy efficiency. Yeah, play it out. If this goes the way you want it to go, is there actually a benefit over just a huge data center or something in space with a solar panel on it? It feels like the current chip stack, the AI stack, is pretty efficient. We've squeezed out a lot of the inefficiencies of being a human, potentially. Yeah, I think that's where it's really interesting because regardless of how efficient we make silicon, like silicon we have right now is supremely efficient, but we still haven't solved the efficiency aspect of it in terms of power efficiency.

2:02:15I think human brains are extremely power efficient in comparison to silicon. I think there are things that we can harness there on top of other stuff like sample efficiency. Human brains learn very quickly compared to silicon, much fewer examples. and continual learning is free on biological substrates because they keep learning over time. But I think continual learning is something that has to be expanded on in the AI space. We're still figuring out what's the optimal approach to use. But yeah, massive, massive energy efficiency. How do you actually think about that energy efficiency, though?

2:02:44Because I've seen, I mean, there was that news of like$5 ,000 worth of SOL tokens solved a bunch of math problems. When I think about like not even the salary of a mathematician, but I just think about the food that goes into generating the calories, that generates the energy, that generates the theorems from a mathematician. You're way above 5K. So it feels like the models are actually pretty efficient, but what am I getting wrong? I think if you think about the cells as growth, as humans we are expensive, we have to feed ourselves and all that. I think with Doom previously we showed that it's possible for us to inject information, essentially force these cells to learn much faster than a human would learn.

2:03:28So you don't have to learn the basics of language ABCs. We can tell us to just predict whether this is a hot dog, this is not a hot dog, something like that, far quicker. So we can skip that prior human learning building phase that would be very expensive normally. So that's kind of the stuff that we're approaching it with. Is it possible that golden retriever brain cells could be better at long-running tasks like chasing a ball? Oh, I think that human brain cells, for now, empirically, they are the best. Oh, shots fired. Well, you know, we've tested it, but yeah. We've tested it. You put a golden retriever in the van?

2:04:09We tested their rat neurons, right? So they're rat neurons and then they're human brain cells. Whoa, whoa, whoa. Golden retriever and rat. These are not comparable animals. Let's give it to the retrievers and credit. What is the business going to look like over the next decade? Because I imagine at the end of all of this, there's some sort of business model where you're selling intelligence. But in the short term, there's some venture capital that comes in through Y Combinator. Congratulations, by the way. I do. But what's the middle step? Is it partnering with biotech companies? Is it NSF grants or government funding or something like that?

2:04:48Or partnering with a university? How do you keep the lights on and keep the flywheel going? I imagine you can raise more money off of scientific breakthroughs, but I imagine there will also be an economic, a commercial flywheel here even before you're selling the work. Yeah, I think something that we're really looking at right now, which is great because we've just started, is with the exponential increase in AI capabilities, we're going to start building our lab from scratch to be automated. We want to automate the stem cell research to differentiate them into neurons, the optimal compositions, the optimal spacing on the electrodes, let's say.

2:05:29So I think that the automation aspect of our lab can easily be branched out into different things like drug testing. And I think that could be significant revenue in the short term to push this all the way to make sure that we get brain cells to be the fundamental substrate for compute. very cool well congratulations what a fascinating company uh thanks for taking the time and have a great rest we gotta we gotta introduce sean to the guy that we had on yesterday what was uh put him together you got a whole human yeah tissues he's doing he's doing yeah uh vivodyne vivodyne they are doing some interesting stuff well have a great rest of your day great to meet you sean Thank you guys.

2:06:09Thanks for having me. Cheers. Good one. Work day. Do you mind shaving your head for me? Do you mind just letting me take a saw to the top of your head? You think I'm going to put your brain in a vat?

2:06:30You think I'm going to put your brain in a vat? You want to come up to the roof with me? I don't want to go on the roof, John. You think I'm going to throw you off the roof? I like that it's green now. The green's fun. We should get buzz cuts sometime. I think next summer, summer buzz for both of us. Well, it's got to be a bet. It's got to be, oh, you know, Sasspocalypse is over or something. Some prediction that then we can take a victory lap on or shave our head in sadness. Marques Brownlee might have to shave his head, right, because he made a bet. or he said if the cyber cab ships to Elon's schedule, I'll shave my head, something like that.

2:07:09And people were going back and forth. And the Tesla fanboys were saying like, he's going to have to shave his head. And he's like, not yet. Because of course, like, you know, all these projects take a long time. I don't even need much of a reason to shave my head. I actually have gotten a bunch of summer buzz cuts over the years. So you'll be like, if the stock market moves by more than 1 % through the end of the year, I'll shave my head. silver lake is in talks to buy work day sources say that's a big deal right isn't work day huge work day is let's see i'm sure it's popped 50 billion 17 percent wow the take pride uh and so now a 50 billion dollar company it was 43 billion dollar company at least when this article was written uh we'll see i'm curious what that is a bold bold uh bet and but i can imagine there's a bunch of ai transformation that they'd feel like they'd be more easily able to do as a private company so are you a are you a standard crying face emoji guy a tilted crying face emoji guy or a tear streaming down your face emoji guy?

2:08:21I'm normally... Okay, I want to actually pull this up so I can see more closely. Which one are you? Joe Weisenthal has recently transitioned to a tilted crying face emoji. I don't dabble in the tilted. I do the tears. I do the tears. And the straight on. Do you do the streaming down the face? Yeah, yeah, yeah. Oh, I need to mix that one in. It's the most popular emoji of 2025. Yeah, I went through maybe a decade where I wouldn't touch any of the crying. Oh, any of the crying ones. You do a smiley? But I've been laughing a lot in the last couple of years, mainly because we've been doing this show.

2:09:03Yeah, that's great. You've got to throw a laughing emoji. I like that in iMessage you can do the ha-ha, but then you can also throw the crying emoji. But I've got to experiment with the tear streaming emoji. Do you ever go with the cat versions? No, I never LARP as the cat. Never go cat. But expect a cat emoji from me soon, Tyler, in our group chat, because it might be underrated. There might be some alpha there. I don't know. Anyway, there was clearly alpha at this auction. A J.P. Morgan hand-signed mortgage bond from 1886 sold at auction for just$847. Someone got a steal, says Dylan Abrascato.

2:09:43This needs to go in the Museum of Business. There's a bunch of these good ones. I was looking at the... Yeah, why did Dylan find this after the auction had already closed? I don't know. Get it together. Get it together, Dylan. There's another good one. Can I... How do I drop this in the timeline? Microsoft, in 1995, 1996, published a wine guide. This is from the corporation, the hyperscaler, the Mag 7 company, Microsoft They released a wine guide The essential, click to the next image and then the next one There we go, okay The wine guide came on CD The essential reference from vine to glass If you wanted to know about wine, this is where you had to go You had to get the official wine guide from Microsoft, the software company Fascinating know your audience moment a lot of a lot of microsoft fans customers 1996 if you're implementing microsoft in 1996 probably enjoy a glass of wine every once in a while so why not good why not uh well john i'm going to read a buco capital bloke post that i think applies to this exact moment right now he says i think we're very close to the point where caring about ai or talking about it a lot is a bit embarrassing.

2:11:05Move on already. Who cares? Move on already. I mean, there is a point where we don't talk about the Internet anymore. We talk about what's happening on the Internet. We talk about particular Internet companies. This stuff does diffuse. If it diffuses fully, you stop talking about the underlying technology. But there's a horse race on, Buko. There's a lot of money on the line. The entire global economy, potentially. That's right. He's just horsing around. Thank you for tuning in to TBPN. Sign up for our newsletter at tbpn.com. Leave us five stars on Apple Podcasts and Spotify.

2:11:44Great success.