LIVE CLIPS
EpisodeĀ 1-28-2026
New perk. That's fantastic. Okay, so, yeah, take us through the status of the company. Like, how far along are you? How big is the team? What are you thinking in terms of release time? Are you planning on sort of going heads down, age of research mode, SSI mode, and releasing something when you've hit some major milestone, or do you want to be more iterative with it? Okay, so we're about two months old. The team is now 11. We're super excited. We've got people we really love who are both brilliant but also just wonderful people. Excited about it. I think. I think we are sort of in the middle. We're definitely age of research mode. Like, I think the goal is not to commercialize. Not because we're not commercial people. Like my background, even Aiden's background, Ben's background. All reasonably commercial in some sense, as opposed in addition to deep research. Sure. It's just that when you, like when you get revenue, you have to focus on it. Like you have to focus on providing for customers. And that makes it harder to build, you know, deep technology. Yeah. So, you know, our goal is to try to find the biggest market we can to solve the most important problem we think we can solve, which is the data efficiency problem, before doing anything of that. At the same time, our approach is probably to be building a little bit more in public. We'll release some research artifacts, at least that I think will be cool reasonably soon. But who knows exactly when the runs will finish or how many times they'll crash before they work. Our biggest training run today. So bad time. Next launch for our maintenance. There we go.
Any agent. Yes. So Gary Weiss, everything she says internally immediately becomes external messaging. Tough, tough job. But she put in, she was talking with the team and saying she wants to put a huge emphasis on scoops. This is about cbs. Cbs, yes. Not so. She's the editor in chief at cbs but she was acquired in through the Free Press. So the Free Press is operating, the CBS is operating and there's an effort to sort of merge the two. Free Press, obviously very fast moving startup culture. CBS has been around for decades. So the quote that has hit the timeline is Bari Weiss also said that the network CBS will now put a huge emphasis on scoops, but not scoops that expire minutes later, but investigative scoops. And crucially, scoops of ideas, scoops of explanation. This is where we can soar and where we'll be investing. She continued. Scoops of ideas. People are roasting this, but I think one fires me up. Yeah. Also you just look at the modern media landscape, there are a lot of scoops of ideas that are. They create these. Like everyone watch that podcast, everyone is repeating that idea from that person. They set the tone. They introduce this concept, this philosophy, this thesis. They unpacked it for two hours exclusively in this one place. You got to go there to get it. Yeah. And it starts a conversation that wasn't happening. It's not just like here's some facts that then immediately get copy and pasted everywhere. Totally. That's what you're talking about. Scoops that are expiring. It's not about. You have to start a conversation that wasn't. Yeah. I mean I go to Dwarkash Patel, the Leopold Aschenbrenner. It wasn't a scoop, I guess, but situational awareness was posted as PDF but also made available in a multi hour sit down podcast where he unpacks those ideas that flew around that started a conversation about where AI is going, its impact on the financial markets. Andrej Karpathy Same thing on Doorkesh. There's been a number of things where certain ideas, certain scoops, certain explanations have recontextualized things. And I think that CBS has the, the production horsepower to bring a really polished product together. But if there's something that's been already revealed, the facts are out and then you just do. You're reinstantiating on cbs. That's not like must watch necessarily for the modern media consumer. So I don't know, I think it's a reasonable thing to say, even if it's sort of a funny phrase. Keep running.
And then obviously OpenAI, that is a real firestorm for Apple right now. Like OpenAI, obviously they're working on these AirPods competitors the ChatGPT built. Sweet pea. Sweet pea. Jony I've is, you know, running the show and design there. That's a big frickin deal. Okay. He raided the whole design team, all of the Apple designers that were there under Jony. I've not all of them. 95% of them are gone. You know, a ton of them are now working at love from and OpenAI and other companies. Some of them have retired. But the crux of it is how do you partner with a company that's trying to put you out of business? Right. So there's no way that they were going to work with OpenAI, no matter how good the thing. What do you think about some of the.
Just insane amounts of money on cloud code, which maybe they're losing some amounts of money. It's like. Well. You look back at the. I forget which interview Dario was doing where he was trying to get people to think about. Like you think of each model as a company where you spend all this money on training, which is capex, and then when you're actually running the model, it's very profitable. But if you look at the business as a whole, you have massive, massive losses from training and stock based comp and hiring 1,400 of the best engineers in the world. So if you actually look at it on a company, company wide, you have continued scaling, massive losses. But the important thing is effectively at the product level, when they're selling the product, they actually are making money. So in the S1, good to you and to me, looks like training clearly broken out as capex. Solid gross margins above 60%, something like that on inference. And then all of the AI copium can probably subside. If OpenAI or Anthropic go out with an S1 that shows really solid inference margins and if it comes out that it's like, oh, their inference margin is like 10% or something, then people are going to be panicking. To be clear, I think these will be some of the most special S1s that have ever graced the capital markets. It'll be great. Let me tell you about second.
One way. One way. The reason that I feel like that is such a powerful moment, and I'm glad that he shared it, is if you give somebody. If you're talking with a model and you give it a task, and then it just hits a dead end. It's just incredibly like. That's sort of like. People are very used to that right now. Yeah. And it's not that it needs to be that way, but it's just kind of like the steady state. Yeah. People are used to. Okay. I know. That's effectively the agent having real agency that makes it an agent. Yeah. It's saying, like, well, I didn't know how to do this, or I was confused, and I tried a number of things until I did what you wanted. Right. And that's like, what you want out of. That's what you want out of a team member. Right. If you're working with somebody on a project and they have a task, they don't just try one thing and come back or just say, like, I actually can't handle this because the file type's wrong. Convert it. Figure it out. What was your reaction, Tyler? Yeah. I mean, it's, like, pretty insane. Definitely. Definitely raised my chance of permanent underclass. Oh, no. It's making me a little worried. Yeah. Will says it's over. It's over. We need to move. And there are a lot of people quoting this G. Fodor has the Eliezer Yudakowski meme. And lots of people here. Lots of people were, you know, interesting that he uses Codex here.
Walk me through why data efficiency and why the data efficiency? A problem is important. It sounds all good. Oh, you get a really smart model and you don't have to ingest the whole Internet, but everyone can just ingest the whole Internet. Like you can download it, you can scrape it, there's plenty of models that are trained on it, so break it down. So. Exactly. I think the goal is not necessarily in the long term to not train on the entire Internet. I mean, it's research. I don't exactly know. I think the idea is that this is not needed. Right. And, and the fact that it's not needed suggests that we're actually missing something. Because currently for the, for the existing technology that we have, it, it is necessary. Yeah. So why do I think it's an important problem? You know, to the extent that AI has been hard to integrate into the economy and you know, we always see, you know, these Bloomberg articles that are like, you know, oh, like chat and search are working and coding is working, but like, what else is really doing? For me, to the extent that's true, I really think it's because models are much less data efficient than humans. But like, if you wanted to learn a new task or put it in a new vertical, it like it takes thousands of times more effort than it does to just tell a human what to do. So I think if you can make a model a million times more data efficient, it's like a million times easier to put into the economy. I. Also think there's just like tons of cool stuff that, that you can do in really data constrained regimes if you, if you can learn to learn with less data, for example, whether it's robotics or scientific discovery or even something like trading, which we have to acknowledge is like the most valuable next token prediction problem in the world from a pure economic perspective, these problems have very limited data and existing AI systems aren't quite as good as them as they are at other things. I think that learning to learn with less data is just tremendously valuable in all these spaces.
Because the app developers, they know Siri is so terrible. You know, here's a question. So what if I told you that you could call an Uber with Siri, right? Like you guys probably know that but like does anyone use it? I don't have the data but I can tell you that they announced support for that and they actually rolled out 10 years ago. No, 10 years ago. 10 years ago. Okay. It doesn't work. Because nobody uses it because it's too. Cumbersome. Totally. Okay, Like I'm very tech forward as anyone watching this probably knows. You know it's right up my alley to have Siri call my Uber for me but like I've never done it because you know what? I don't freaking trust it. Yep. It's. Not, I don't think it's gonna work and it's just going to be a time suck. And people didn't trust ChatGPT with three.
Talk to me about fragmentation and steerability. If you achieve sample efficiency like you're planning to do, you envision a world where you're sort of creating some sort of base model that is so sample efficient that just with like a basic prompt or a few examples, it becomes incredibly effective at a specific task and sort of replaces like the heavy duty training data runs and the RL environments and these massive data collection efforts to do some sort of fine grained task at a very, very high level to sort of like five nines of efficiency. Is that what it looks like or is it more like you will wind up vending a sample efficient model that's for specific use cases? Yeah. So one way to think about this is that it seems like the reinforcement paradigms of today are actually just shockingly inefficient. You don't really get much generalization across tasks. You teach a model through one kind of of learning and then you teach to the next one. It's kind of like whack a mole or something. And we look at this, we think this is kind of crazy. I mean, the first time I really saw RL scale, what it brought me back to was actually good old fashioned AI back in the dawn, you know, this primordial age of AI when people were kind of hand designing these convolutional filters for eyes and noses and things. Yeah. And then they were like, wait, just throw data at it, just throw scale at it. What are you doing? And in this really weird way we're kind of looping back onto that where it's like, oh, just make another environment, bro, just make another environment. Just one more, just one more. The next genome AI will not be just, you know, environment slot. I mean, I think, I think, I think it is a piece of it. Like I do think that there's a long tail of tasks in the world, right. And like there will always be a place for people to produce custom data. It's just like a question of how much operational difficulty it takes. And like one route is just to slog through the operational difficulty constantly. You incur variable cost. Another route is to do a bunch of fixed cost investment into trying to make that variable cost lower. So the last thing I'll just say, to answer your question, I don't really think we know at the end goal looks like entirely like, you know, AI is a big space. Like, you know, human level intelligence is not the ceiling, it is merely the floor on what is possible. Like if you can train models with vastly less data and possibly more compute in very different ways. Like, what is going to happen? Like, we actually don't know. Like, I think. I think it's actually unlikely that they will uniformly dominate frontier models. Even in the very best case scenario for us. They're, like, going to know less facts, they're going to be different. They're not going to have memorized all of Harry Potter. Like, that's actually a useful skill in some ways. But I do think they'll be different and weird, and they'll have interesting capabilities that we'll find really valuable ways to use. So, yeah, I think it's an experiment we're really excited to run. Yeah. What do you expect the.
Walk me through why data efficiency and why the data efficiency? A problem is important. It sounds all good. Oh, you get a really smart model and you don't have to ingest the whole Internet, but everyone can just ingest the whole Internet. Like you can download it, you can scrape it, there's plenty of models that are trained on it, so break it down. So. Exactly. I think the goal is not necessarily in the long term to not train on the entire Internet. I mean, it's research. I don't exactly know. I think the idea is that this is not needed. Right. And, and the fact that it's not needed suggests that we're actually missing something. Because currently for the, for the existing technology that we have, it, it is necessary. Yeah. So why do I think it's an important problem? You know, to the extent that AI has been hard to integrate into the economy and you know, we always see, you know, these Bloomberg articles that are like, you know, oh, like chat and search are working and coding is working, but like, what else is really doing? For me, to the extent that's true, I really think it's because models are much less data efficient than humans. But like, if you wanted to learn a new task or put it in a new vertical, it like it takes thousands of times more effort than it does to just tell a human what to do. So I think if you can make a model a million times more data efficient, it's like a million times easier to put into the economy. I. Also think there's just like tons of cool stuff that, that you can do in really data constrained regimes if you, if you can learn to learn with less data, for example, whether it's robotics or scientific discovery or even something like trading, which we have to acknowledge is like the most valuable next token prediction problem in the world from a pure economic perspective, these problems have very limited data and existing AI systems aren't quite as good as them as they are at other things. I think that learning to learn with less data is just tremendously valuable. In all these. Talk to me about fragmentation and steerability.
On that, by the way. The smart glasses, was that just completely reactionary to Meta? Oh yeah. They started toying with the smart glasses in terms of like this non AR smart glasses. Right. First of all, AR smart glasses, that has been the vision from day one for a decade plus. But in terms of like this non display smart glasses, that is a concept that Meta has really popularized. And you know, when these things started to gain a little bit of Steam in 2223 is when they started taking a very hard look at it. And they're gonna do it. And I think they're gonna destroy Meta with them. I'll be honest. Yeah. Is it the styling, the pricing, the features, the integration? The integration. To me, to me. The challenge is if you can't deliver me imessage. Yeah. Apple has the ingredients because of their login to destroy any company in any hardware. It's just about them, through figuring out how to do it and get it done and not waiting too long. The biggest problem with them is they just take too long and over engineer everything. Like the Vision Pro is the most over engineered device ever. Right. They could have. Got that out three years earlier with a little bit less fit and finish and maybe it would be. More successful today. Yeah. It does feel like there was.
They don't have an App Store. They don't have an App Store on day one. Equivalence of apps on the fly. Well, if you ask me, we're already in territory where iOS and the app Store are legacy features. The App Store is a legacy world. The iOS user experience, the Mac OS user experience, where you're jumping between applications, where you're going into something to get information, where you're going back to your home screen and launch another app. Apps are the past. AI agents are already here, and that's the move forward. And whenever OpenAI comes out with a phone, and I do eventually anticipate them coming out with a phone, and when I say phone, I'm not talking necessarily about something you put to your ear, like a classic phone, like an iPhone. I'm talking about. I feel like people are always going to have some sort of slab in their pocket. Right. Because it's convenient. You get the display, you get the sensors, you get the battery, you get the cameras. It's not a replicable experience, no matter how many different gadgets they can put on your body. Like a banana, like a plastic banana. Smart banana. That would be fun. Yeah. So, yeah, AI agents are where the world is going. And I totally expect Apple to move in this direction. This new Siri Campo that they're launching at the end of this year, that is a huge step towards the AI agent Ification of. Of different features on the phone and being able to, for instance, tell your phone, pull up this photo that I took at the studio, find photos where I have bottles of water, and remove the bottle of water from the photo and email or text it to Mark. Right. Like the AI agent ification of iOS is happening. Sure. Yeah. And theoretically, Siri should have killed like the weather app a long time ago.
Of course, subscribe to Power on in our tech bundle. It's great value. Okay. Ternus, he's 50. Everyone else on the Apple executive team, late 50s through their mid-60s, turning 66 this year. In the case of Tim Cook, you're Apple's board. You like continuity. You like an insider. You like people who know what they're doing, have been there for a while, they know where the bodies are buried. Okay? These guys all have hundreds of millions of dollars, if not more. Yeah. At 50, he's the only one who is. If let's say Tim Cook hangs out another three to five years, you're not going to appoint another CEO who's 65, 70 years old. He's the only guy Apple, they get vast majority of the revenue from hardware. He's the hardware guy. Have they screwed up any hardware since he's been in charge? No, he's a steady hand. He knows what he's doing. He's really the only choice. You know, there was this New York Times report a few weeks ago, basically saying that could be Greg Joswiak, could be Eddie Q, could be Deirdre o', Brien, could be Craig Federighi, for sure. Not going to be Craig, it's not going to be Deirdre, it's not going to be Eddie, it's not going to be Jaws. The only category that makes sense is an operations person, because you look at the current CEO, Tim obviously comes out of the ops world. You look at the guy who would have been CEO if Tim Cook didn't stay so long. I'm not saying he shouldn't have stayed so long. He's done, obviously, a fantastic job for shareholders and the employees and what have you. Would have been Jeff Williams, he was the coo. So Sabi Khan, he was named COO a few months ago, but he's really been in that job for the last half a decade, I would say. So anyways, it'll be Turnus or Sabi or someone completely out of left field. I don't think this is imminent, so we'll see what ultimately happens. But all signs are turning towards Ternus. Everyone has an opinion that Ternus is gonna be the next CEO. Fine. I've been shouting this from rooftops for the last two years, but no one has given evidence. Like, what is this based on? Right. Has there ever been a baton handoff? Is he getting more responsibility? Do they have a big baton? They've got one of these. You know what you have? You have white smoke coming out of. Well, actually, no. There's no smoke. No smoke. Very environmentally friendly environment. Friendly. So, you know, maybe out of the. Reflection out of the sun, solar panels or something. A glint off the solar panels. Exactly. So what, you want evidence? Yes. You want to hear that he's been getting more responsibility? Okay. One, a few months ago, took full control of the Apple Watch engineering team that was co run with the old COO until he retired. Okay, there's something for you. When they started soft firing the head of AI, the guy we were talking about earlier, they took the robotic stuff away from him. They gave that to Ternus. And then the real news, as I broke last week, is that even though on paper Tim Cook is running the Apple design teams, it's not. It's Turnus. He took over at the end of last year for a variety of reasons. But you look at who's run design at Apple over the course of history. Well, like Steve Jobs, Tim cook himself between 15 and 17, Jeff Williams, who was the number two in heir apparent for a long time and then obviously all heard of Jony. I've. Now you had John Ternus to that list. It's a big sign. It's a big indicator because you look at what Apple is known for as a company and its design. Right, that design function, it's not just hardware, it's hardware and software that he's overseeing as the manager of both of those teams. So I would say that is your first piece of evidence that he's getting some more material. What does he have to do since he's the hardware guy, to communicate that he has a steady hand on the tiller as.
Activate go golden retriever mode. Market clearing order inbound. 5. I see multiple journalists on the horizon. Stand by, Founder. You're watching TVPN. We're bigger today and it's Wednesday, January 28, 2026. We are live from the TVP in Ultradome. The temple of technology, the fortress of finance, the capital. Ramp.com, baby. Time is money save. Both these use corporate cards, bill pay, accounting and a whole lot more all in one place. That's right, Will. I am wearing a sample today of this new Ramp quarter zip. And you'll notice here, I included a little pocket here for your ramp card. It's just right here. Always close. Right over your heart. Right over your heart. Excited to get this one out in the world very soon. Yeah, this will be fun. I wrote about Netflix. There was a funny, very brief interaction between Ben Thompson of Stratecheri and Netflix co CEO Greg Peters on last Thursday's Tratecheri interview. And they go back. The only mention of AI in like an hour long interview or something. It's just two little exchanges. Ben Thompson says, is AI slop going to save you if it overwhelms the UGC platforms? And basically it's like you're a refuge. So this is actual, this is real. And Greg Peters just says, I think it's credible. I don't know if that's the reality. So I can't say with certainty that's where we're going to land. But it's a credible possibility. It's a, like, maybe that's a bull case. Maybe that's. It is interesting. I mean, Netflix has been trading down over the last couple of months, but in general it's up. I think it's 4x up since the launch of ChatGPT and is generally like near all time highs. Like the business is doing very well. But every CEO needs to contend with the AI question. The AI issue, how will AI change their platform? And AI has already been changing Hollywood. I mean I was, I was reflecting on the Avengers. When did the, when did Infinity War come out? Infinity War, that was what, 2018? I just remember seeing maybe he was even in one of the first ones. But the whole CGI process for Thanos, he has this like very distinct large chin. So Josh Brolin is the actor that plays Thanos. Is he a mogger? He is a mogger. He has this huge chin. It's actually like he's kind of like the og. I don't know, he looks like chin implants. It's kind of crazy, but it has these, like, cracks in it and it has this, like, very distinct look. Thanos. And normally the way the VFX pipeline works is that you go and you put these black dots all over your face and then you wear a helmet that has a camera pointing at your face. I think it's a. I don't know what type of camera, but it tracks all the points. So when you smile, like, it sees that the actor that's driving the performance capture is smiling. And then that facial movement is transferred. So they're recording the lines, they're acting it out, they're giving their facial performance. And then that's transferred all the little subtleties of how their eyebrows move. All of that is transferred to the CGI character. It can look a little flat, though. So what they did with this is they still have all the points on the face, but then they interpolate from the small points that are on the face into a higher res model. Yes, don't read that, don't read that, don't read that. But it is a good point. I wasn't even reading the chat. I didn't even see that you said that. Of course. I was just looking at this absurd picture. Oh, yeah, it is an absurd. So all of those are tracking markers. And then the question is like, you have a much higher resolution CGI model. If you just transfer with 50 points or 20 points, you're not getting all the detail of what a human face actually looks like and the way it moves. And so Digital Domain, which was one of the many VFX studios that worked on the Marvel series, they built a straight up machine learning pipeline. Like, they used AI. It wasn't a diffusion model, it wasn't an LLM, but they used a machine learning model to basically translate from the low resolution, just a few dots to a much higher resolution mesh. That then became the performance of Thanos on the screen. And I don't know if you remember 2018, the movies. Obviously you didn't see any of these movies, but I don't remember like, AI backwards. My prefrontal cortex wasn't fully developed. But truly, I mean, people did make the, oh, it's too cgi, the explosions are too crazy, it's too over the top. But in general, people weren't up in arms about like a use of AI or use of to. Everyone was just like, this is a CGI epic. This is a crazy, you know, Marvel movie. Like, we're fine with all this. And there wasn't backlash to that. And I don't think that There would be backlash to this type of, like, AI tool. Now, obviously, Marvel's Avengers, that's Disney property, but the same VFX pipeline is being used all over the industry, and it will continue to be used. Interestingly, I talked to Jason Carmen about the Carminator. The Carminator, The Carminator about. About using AI tools in filmmaking because he's obviously making movies and doing VFX and stuff. And. And I was like, certainly if you need to rotoscope out a background. So rotoscoping is where you are basically using like, you're. You're cutting out like a subject from the background and then just doing like a background replacement. That's an example of rotoscoping. But it's over motion, so it's moving. So. So you need to track the hand here, move it over here, track it again, track it again. And it can be very, very time consuming. Typically this is offshore to like a bpo and then they have a whole team of people that are all aiming it and they have some software that's used. But I was like, this feels like something AI could. Just one shot. He said that AI was not there at the level that he wanted to deliver. He wanted to deliver in 4K. And so he went to a team. I think he paid a fortune. They did it. And when they rotoscope ahead, they actually draw new hairs on to kind of create this. It's very like, artisanal still. But obviously AI can rotoscope. You see it in the cap cut edits. Where's the rotoscoping? Neolab. It's actually Runway. We've had Cristobal on the show. And pre chatgpt Runway had a fantastic AI rotoscoping tool where you could basically load up a video, put a couple dots on what you wanted to keep, and then add and then flip over to the red dots and put those in the background and be like, cancel all this out. And it would sort of use that as like an intuition for the model to drop out the background. And you could do a really, really clear, like, cut this person out of this video just in Runway. And now that's available in capcut and edits. And that's where you're seeing all those crazy hype reels where, you know, you'll see the F1 driver, like, standing up, and then it'll drop out the background and cut in a different background and then the F1 driver drops out. And it's like this very, like, schizo. A really cool technique. Yeah, you can see is this runways edit tool. So it basically draws a mask around it. You just highlight like what do you want to actually rotoscope. This is an example of like motion adding motion to video but like you're putting in a little bit of your own aesthetic taste quickly. Let me tell you about Restream 1 livestream 30 plus destinations. If you want to multi stream go to restream.com so goat. So like I don't think Netflix should take a hard line stance on AI broadly because they want to use AI tools. Obviously they've been using AI for recommendations forever. The original collaborative filtering algorithms were machine learning models and that's how you open up Netflix and says we think this would be good for you based on what you watched. And it is much more nuanced than just if you like K Pop Demon Hunters, we're going to recommend Squid game next. It is machine learning and so many of these rote tasks will be AI enabled and they already are. And there's not going to be. I don't think there'll be a crazy pushback here, although it's possible that there's some sort of comms mishap. Especially if a director comes out and is like we didn't use any AI in this film because they don't think they used any AI. But there's a VFX house that when the motion capture stage did use AI to up res motion capture data. Like you could see AI being used in matte painting for the background and the director doesn't even know because they just said like yeah, the background just make the forest a little bit bushier. And they think that they're hand painting it. After that they were using CGI and 3D modeling it. Now they're using AI and that sneaks in and then all of a sudden they face some backlash and But I don't think that's. I think that's manageable. I don't think that's that big of a deal. The bigger question is like how does Netflix position itself against YouTube and the UGC platforms is what we're talking about. So Neil Mohan, the CEO of YouTube has taken a very open stance on AI and I liked his stance. He was like, we're not going to throw an AI tag on everything. We're going to let you use AI tools. Right in the shorts creator we have VO3. We're great at this. Like we're going to lean into this and the algorithm will sort out if you like it and if you think it's slop. And you don't want to see slop. The algorithm will learn that and not show you that stuff. But for the people that like that, they will be served it. But it is sort of a balancing act. And there's definitely this stated preference for like, I don't want any AI on my platform now. How real is that? We'll see. Yeah. And what we were getting into earlier before the show started is Netflix decision. The bigger, like, the decision that's bigger than just like, are we going to sort of lean into AI or not? Will just really be. Do we have. Does Netflix ever lean into ugc? Right. They've been signing some bigger podcasts and they obviously work with independent media companies. But there's a very big difference between just allowing people to like, will they ever have an upload button? It feels like no. And I feel like that could end up being their advantage. Where there is part of what has made YouTube magical since the very beginning was that anybody could go and put YouTube, you know, upload a video. Anybody could be a creator. And I think like, as the amount of content that gets created, ten hundred X's, thousand X's because of AI, it's going to be, yeah, Netflix could be this, like, refuge where you're like, okay, at least if I go here, I know that there was some filtering process. I know that this isn't just a total free for all. Yeah, yeah. The upload button is probably a bigger deal than like AI on Netflix. Exactly. I think that's a very good thesis. Quickly, let me tell you about console consul builds AI agents that automate 70% of it. HR and finance support, giving employees instant resolution for access requests and password resets. Well said. And someone was asking in the chat about the linear lineup. Of course, linear is the system for modern software development. 70% of enterprise workspaces on linear are using agents. And we have a great, great show. Mark Gurman's coming in person. Miles in person. And then we have an amazing lightning round coming up. And I think we got a surprise. Guest fund over at Voyager. And then Mitchell and Gabriel, Gabe from Rogo. Super excited for that one. That's fun. So back to Netflix and YouTube. So YouTube has been on an absolute terror in terms of watch time on TVs. According to Nielsen, YouTube has been number one in streaming watch time in the US for nearly three years. And so this has been the backbone of the case for like, let Netflix buy Warner Brothers, even though they'll get HBO Max. Like you're merging two seemingly big Streaming platforms. But the combined watch time will still be lower than YouTube, so should be fine from a regulatory perspective. But the bigger question is, like, there's this gap between Netflix and YouTube. And at the start, back in. I mean, Netflix is almost 30 years old. YouTube's over 20 at this point. Back in 2005, like, these were seen as, like, wildly different platforms. One was DVDs in the mail, and the other one was, like, a video of a guy going to the zoo on his, like, VHS camera. They felt extremely separate, and they felt extremely separate for years and years and years. Now they are starting to converge, especially around video podcasts. I feel like. I feel like YouTube really drove a big boom in video podcasting because the podcast was, I think, invented by Apple. Yeah. When I started doing any work with YouTube channels and podcasts back in the day, there wasn't a lot of overlap. It was very clearly like this. These were just different types of creators. Totally. And then there was a big shift. Yeah. It was like, wait, I'm leaving a ton of attention on the table. I'm not uploading to YouTube. And that force a lot of creators to actually get into video. Totally, totally. Yeah. There was the pivot to video. And then Spotify went really big into video podcasting. They went on. I didn't realize how big of a push Daniel Ek really did around podcasting. So seven years ago, in 2019, they acquired three companies. Gimlet Media, Anchor and Parcast. Anchors Mike. Yeah. From Lightspeed's company. Yeah. I mean, all three very interesting. Anchor's more of like a product. Parcast had a bunch of. Yeah. Anchor was like, we'll make it easy for you to create a podcast because it's still like, a lot of people were there. It was. It wasn't like it was impossible to figure out, but there was quite a bit of friction. Yeah. And they knew that everyone who wanted to distribute a podcast wanted to distribute it everywhere, but they could sort of default you to getting into Spotify as well. So that did very well. And then they signed exclusive deals with Joe Rogan and a bunch of other people. Even, like, some of the royals also signed a deal. I forget who they are, but they really spent a ton of money trying to get into podcasting. I think they were successful. I mean, we see a ton of audience on Spotify. Interesting. The timing. Apparently there was. Somebody sent me a chart. There was over a million podcasts launched in 2020. Oh, yeah. You saw that chart. Jeremy shared this with us. That's a crazy chart. It actually Fully retraced to now there's like sub, sub 50,000 a year. That feels like you would really have to. Maybe they're being extremely explicit about defining what a podcast is, but either way there was. If you look at the chart, it looks like there was like an insane bull market in podcasts and then a huge correction. Yeah, it does feel like there was a 2020 boom during COVID I mean we talked to the folks that acquired about that and they said that during COVID they saw a huge spike in people like just going for walks, throwing on AirPods. All of that sort of hit around the same time AirPods were getting to like mass adoption. So we were just throwing on podcasts constantly and it really, really grew. Anyway, graphite code review for the age of AI graphite helps teams on GitHub ship higher quality software faster. So there is this chance that with YouTube, with the trough of YouTube getting sloppier by the day, that Netflix carves out more of a unique value prop and you wind up seeing more space between the two. So I think of Netflix and YouTube as starting out extremely separate, then sort of like coalescing with like Joe Rogan experience. And I mean Portnoy and Bill SIMMONS Both are YouTube dominant and now have Netflix deals, right? Yeah. Isn't there deal set up so that the video can only be on Netflix? Exactly. Audio is still elsewhere. Why do you remember that? Because Portnoy drilled it into your head. He said, if you want video, Netflix, Netflix, Netflix, video, Netflix, video, Netflix, barstool video, Netflix. And it's true like that. And so I would think of Barstool, like if I'm trying to watch a barstool YouTube video or podcast, I would just go to YouTube. Right now. Netflix is getting into that. So the gap between Netflix and YouTube is getting like pretty narrow. Like they're becoming competitors. But there's this question on the AI issue. Do they diverge more and is that valuable? The pure AI feeds like Sora and Meta, they haven't really been able to hang on to the top spots in the charts. I think Soar is around 60, 70 point ranking. But like it's still too early to call that because the quality will get better and better. The audio is still very like clockable when you, when you hear it. And there's a lot to be done there. But the struggle will be to create like unifying conversations around particular pieces of content that are AI generated. I still feel like the K Pop Demon Hunters moment, the Squid Game moment, the Alex Honnold Taipei 101 Moment. These live events, these key things that everyone talks about are really, really valuable. And that's a lot of what's driving the Warner Brothers acquisition is people still dress up as Batman around Halloween. And if you can be the place there, that's way easier than. Well, I did. I've been getting AI generated content about a superhero, but my superhero is different than your superhero. And so if I wear a T shirt with that superhero on it, you're like, what is that AI slot man. I'm not a fan of that. I like that there are some things that will remain true. So, yeah, I mean, about the upload button, it's almost better to think about Netflix less as being AI free and more about being UGC free. They're paying for curation and a quality bar that's backed up by a brand that's going on 30 years in the business. The AI tools will come, some fully AI generated content will come, but true slop will be filtered out by the Netflix team. And I think that that's important for a lot of viewers. Anyway, let me tell you about TurboPuffer, serverless vector in full text search, built from first principles on object storage. Fast 10x Che and extremely scalable. Should we talk about Saudi Arabia? Yes, let's talk about Saudi Arabia. What's going on in the Middle East? According to Bloomberg, Saudi Arabia is widening its search for capital, turning to some of the kingdom's wealthiest families as the government looks to ease pressure on public finances and fund the next phase of the crown Prince's economic overhaul. I saw this headline and was deeply concerned. I was like, why aren't you guys supposed to be funding the whole build out? We were kind of counting on you guys to be quite liquid while the rest of us over here in America are levering up so seemingly somewhat of a liquidity squeeze. This had been reported since back in October, and we can get into it a little bit. They're also raising from. They've been raising, tried to raise money from the Qataris. They apparently asked for something like 10 billion from the Qataris in the UAE. Qataris threw in 10B. Okay? Allegedly, the UAE did not. And there was frustration around that. Thought you were going to say the Qataris threw in 10B and Saudi Arabia was like, and we deployed it. We need another 10B. Time to re up. Time to re up. Can we have another time? As part of these efforts, the PIF gathered about a dozen prominent families on the Red Sea last month to assess, assess their appetite for participating in future opportunities at the summit, which also included others from the private sector. The one trillion dollar wealth fund called for more collaboration on deals, the people said, asking not to be identified. Government entities, including the Ministry of Investment, have also stepped up outreach to family offices, wealth managers, domestic businesses, according to some of the people. Local families are being sought after to play a bigger role in partnering with global investors to draw more money to the kingdom, they added. So years of excess expenditure and subdued oil revenues alongside a tighter lending environment have challenged the Gulf nation's ability to bankroll expense expansive projects planned under the $2 trillion Vision 2030 agenda. Officials this week said they would postpone the 2029 Asian Winter Games, and the government had previously pared back spending on other elements of Saudi Arabia's economic rejig. It seems hard to host a winter game that seems extremely expensive to host a winter Games in Saudi Arabia if that's what's going on. They have. Don't they have some mountains? They have mountains there. I don't know. Against that backdrop, Riyadh has been stepping efforts to. Yeah, I know that they have some stuff indoors and I guess you can just do everything indoors. But again, that feels expensive. Yeah. So as opposed to like just go to Russia and walk anywhere and you can ski. Yeah. So they had been developing something with neom. Oh yeah. Which is also, I guess, in the process of pivoting. Before we move on, let me tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. Against that backdrop, Riyadh has been stepping up efforts to look for alternative sources of financing, including a rare loan deal. Now a range of local entities have begun to sharpen their focus on Saudi Arabia's family offices and businesses, which collectively control assets worth hundreds of billions of dollars. So the number of family offices in the Middle east in 2019 250, 2024, 290. Now we're up to 310. And the projection for 2030 is 350 family offices. There are big portfolios. The wealth is sizable, said the chief executive officer of the national center for Family Business in Riyadh. These entities have long dominated the Saudi economy, and close to 95% of private businesses in the kingdom owned Interesting. They're not doing a lot of IPOs over there. Of these, many groups are only just starting to form family offices as they grow in size and look to formalize strategies to help spread the wealth across multiple generations. That makes both established and new family offices a prime target for more investment. They're naturally looking to diversify and want to contribute in areas where we've just scratched the surface. In addition, more complex areas of finance are also beginning to emerge, drawing the attention of family offices. That includes private credit and industry in its infancy in the kingdom as overstretched banks struggle to meet more explosive needs for financing. I like that the financing needs are getting explosive at this moment. Not expansive, not expanding. Exploding lenders for years have been the primary financiers for individuals, businesses and government entities looking to drive investment into Saudi diversification agenda. But they are starting to pull back as liquidity tightens, leaving many local firms scrambling to find new sources of financing. Well, we got to bring somebody on. To learn more about this and I will tell you about public.com investing for those who take it seriously. Stocks, options, bonds, cryptos, treasuries and more with damn great customer service. Shaco says you're bearish on the US Dollar. The token used to buy AI products. Bold. That's really good. Yes, the US dollar is all over the place. Tether is shaking up the Gold market with massive metal Hoard I did not see this. This is interesting. There are roughly 370,000 nuclear bunkers in Switzerland. That's so many. I've seen a video about one of them, but I didn't realize there were so many. The legacy of the Cold War that are now rarely used. One of them though, is a hive of activity. Every week more than a ton of gold is hauled into the high security vault owned by crypto giant Tether holdings sa, which is now the world's largest known hoard of bullion outside of banks and nation states. Over the past year, Tether has quietly become one of the biggest players in the global gold market, the embodiment of a meeting of the crypto and gold worlds whose shared distrust in government debt is a major factor behind the surge in prices to never before seen highs above 5200 now. It was 5000 yesterday on the COVID of the Journal. Gold is on an absolute tear and yet relatively little is known about its inner workings or its gold strategy. When two of the most senior gold traders quit leading bullion bank HSBC holdings last year, the industry was abuzz about gossip about where they would head next. Few guests, few guests that the answer was Tether. In an interview with Bloomberg, Chief Executive Paolo Adorno described the company's role in the gold markets as similar to that of a central bank and predicted that Washington's geopolitical rivals would launch a gold backed alternative to the dollar. Interesting. He revealed that it plans to keep plowing its enormous profits into gold while also beginning to compete with banks in trading the metal. We are soon becoming basically one of the biggest, let's say gold central banks in the world. Interesting. This is like the original crypto narrative, right? E Gold, even before bitcoin. Well, was e gold actually gold backed? I think that was the whole pitch was, yeah, it was trying to be digital gold and I don't think it ever really got adoption. It was very early, late 90s. Tether has a scale now. They also launched their US focused stablecoin this week to compete with USDC. The new token is known by its ticker. USAT is being issued by Anchorage. I think we had the CEO of Anchorage on at one point. Maybe it was really quick. Cantor Fitzgerald, which already manages the reserves of Tether's mainstay 186 billion USDT stablecoin will do the same for the new coin as its designated reserve custodian and preferred primary dealer. So USAT is already available. They're trading as of yesterday. So we'll see. We'll see. It'd be interesting if, like, at what point does USDT depeg upward if gold keeps ripping? Right? I mean, does USDT have a claim on the overall assets of Tether? I think that's not what the product is. I think like the Tether stock would own the Treasury. Sure. But historically, when you saw a stablecoin like depeg, it was because people were turning around the reserves. Yeah, but I think that the contract is that they'll never give you more than a dollar, so. No, I know, I know, but I'm just saying, like, stranger things have happened. In crypto where, no, it's always been a very profitable company, so certainly bullish for them. Banta Automate Compliance and Security. Vanta is the leading AI trust management platform. There's one funny quote in here. So Tether makes his money from its dollar, stablecoin, that is the giant of the sector with 186 billion in circulation. The company takes in real dollars in exchange for that USDT token and invests them in treasuries or other assets such as gold, raking in billions in interest and trading profits. Processing the physical metal is crucial, Adorno said. So much so that the company has taken the unusual step of storing the bullion itself in the former nuclear bunker in Switzerland, guarded by multiple layers of thick steel doors. And he says it's a James Bond kind of place. It's crazy. That's a great quote to Give Bloomberg. It's just like James Bond. The secretive nature of another CEO. This is a positive, this is a positive reference. It's okay if you're building a secret bunker to hold all of your gold. I think you can safely use the James Bond analogy. The secretive nature of the gold market means that while its easy to describe broad drivers of investment, it can be hard to pinpoint who exactly is behind the buying. China, for example, officially disclosed just 27 tons of purchases last year, but many traders believe it bought much more. Such is the scale of Tether's disclosed purchases that some market watchers have pointed to their role in shifting global prices. The purchases likely Contributed to Gold's 65% rally last year, Jefferies said, describing Tether as a significant new buyer, which could drive sustained gold demand. Still, Tether is only a small part of a much larger rush from investors into gold, with central banks and ETF investors collectively buying more than 151,500 tons of metal. I wonder if this is moving the gold watch market, do you think? You know, the Texas Timex is booming on the back of gold spiking. We'll see. I think a lot of prices are still down pretty dramatically 2020, 2021 era. But I definitely, I mean you got to be a little bit scared right now if you're in the business of manufacturing totally old watches and other precious metals. Just given that your input costs are going, you know, I'm sure they have like one or you could imagine one or two years worth of, of supply. So they're not, they can be somewhat insulated, but price prices will go up or at least costs will go up. Yeah, well, FIN AI, the number one AI agent for customer service. If you want AI to handle your customer support, go to FIN AI. I heard a very funny, very funny interview with an actor who was talking about why he always wears a gold Rolex and he was calling it a helicopter watch. Did I send this to you? Yeah. And he's saying that like this was army. I think it's army Hammer. And he's saying it's like obscure. It's such a situation, such a funny situation. But he's like, if you're ever in a crisis, you're on the top of a building, the zombie apocalypse is upon you, someone shows up with a helicopter and they're gonna save a few people. Everyone knows what a gold Rolex is and they know that that's valuable. But if you're, you're not gonna have time to be like, no, it's an FP Journe. It's a Patek. You know, like let me explain high horology. Here's the tourbillon. Aftermarket prices. Yeah, yeah, exactly. It's like gold. Rolex is a store of value. It's always going to trade and now it's probably going to trade even higher Anyway. Cisco. On February 3rd, the Cisco AI Summit brings together leaders from Nvidia, OpenAI, AWS and more to discuss the future of AI and the economy. See you there. The whole thing will be live streamed and we'll be there for a gig. Paula says SF escape room called the permanent underclass and it's just a room with a laptop and Claude code installed. How do you get. There's something here that's funny. So good. Have you ever done an escape room? No, I have never once in my life thought that that would be fun, nor thought that I would. This is how I want to kill time. Yeah, it's like an hour. Well, if you're good, you get out faster, right? Yeah. Have you. Are you. I'm not into them, but I've done them and they're fun. Sometimes it depends on like if it's a well structured one. But I like puzzles. Like, it's fun. It's like fun puzzle, like figure out. But you can very quickly get caught in like just like overthinking what's going on and being like, oh, it must be some like complex math thing. And it's like, no, you actually just needed to like press this lever instead of like analyze the situation or something. Are some people in there using their phones to try to figure stuff out? Because I imagine you could just say, I imagine there's only. Take a picture. There's probably only a series of rooms. Yeah, well, usually you start in one room. They're all different. But oftentimes you'll start in one room and then there'll be a series of puzzles, locks and keys and whatnot. And then oftentimes you'll unlock something and then you'll progress to another room and then you'll progress to a third room and then you'll finally get out of a series room because it's a lot of. They're pretty easy to set up in sort of like a defunct office space or like, you know, storefront. That's just kind of like, you know, going in between. It's like the spirit Halloween of commercial real estate. Like you just like anyone can come in and just say like, oh, we'll be in there for like a couple months. It's not super permanent. The build out's pretty simple. It's mostly just like some walls and decorations and like some creativity on the puzzle side. Trey says can you guys put Tyler in an escape room and see if he gets out before the show ends? I mean, it'd be very hard to find a three hour long escape room. I think most of them aim for like 45 minutes. Well, maybe we need to make one. One of our guests had to leave the ultra dome and go straight to an escape room. So there are escape room fans among the TVPN army all over the place. Have you done one, Tyler? I've not. Are you interested in it? You're a speedcuber, so, you know. Yeah. I mean, it could be fun. I don't know. I'm kind of in the Geordi camp though. I've just like never. Yeah, it's very like, I don't know, maybe it's great with. It's probably great with like kids who are maybe like 8 to 10 who can do the puzzles. Like, family event would be fun. And then it was a common thing. Like when they came out, there was definitely like a boom in like in escape rooms. Bobby in the chat says escape room would be a good benchmark for AI, certainly humanoid. I mean, you can go in there, take pictures and be like, help me solve this, help me solve this, help me solve this. I know, but I want to see a humanoid run through it. Let's call 1X, see what they can do. Call any of them, Hit them up. Figma. Figma make isn't your average Vicod tool. It lives in figma so outputs look good, feel real and stay connected to how teams build, create code back prototypes and apps. Fast moving on. Lots of reaction to claudebot. Very much enjoyed our interview with the creator of claudebot now Moltbot. Yesterday one of. Yeah, yeah. Incredibly, he had some wild lines, but incredibly refreshing kind of conversation viewpoint. Totally, just totally counter to the entire philosophy that I feel like a lot of people, at least on the west coast in the way that they're approaching AI right now in the way that America is approaching AI. He's like, we asked, you know, you think somebody will like, you know, fork what you're doing or clone it? And he's like, yeah, I'm sure they will. Like, I don't care. Yeah, I'm building this for myself. Yeah. And he's like, I have enough, I have enough money. Yeah. And I'm super. It's going to be really fun to follow along. I Have a feeling that he's just going to keep launching a bunch of random projects because he clearly is just in it for the love of the game. But I do hope that he can get a lot more resources and really scale up. I don't know, I don't know that he needs more resources. He has more agents and whatnot. The interesting thing about Claudebot as a product is you download it from GitHub, it installs and it has all these different integrations and it does something that's very complex and it has all this safety text and different. There's a website and there's community page and all of that feels like okay, yeah, this is like a 10 person startup. They probably worked on this for a year but it's like no, it's one person and it's like three months because the guy is you know, using agents. Yeah, you would think that he would build his. His big complaint was the security inbound security researchers asking himself, yeah, it's like okay, bot bought it. I think he will, I think he should. But yeah, he said he posted 12-26- he had his Codex dashboard up. So he said he's done 250 billion tokens which is probably top 10 Codex users from. So I don't know that he needs way more resources. I mean he should be able to get donations if he needs them, get credits or something. I don't know. The financial strain on that business does not seem, it seems like it's more constrained by his ideas, how he's thinking about designing the system, integrating things, rolling it out. A buddy of mine, Michael, watched the interview and he said that he's had claudebot set up. He set it up right at the beginning of January and was initially just kind of got a little frustrated using it, but has now got it. He's got it set up so that it's able to make phone calls on his behalf. Specifically wants to get Hailstone reservations. So he just like basically wait, what's the restaurant? Oh, okay, yeah, yeah, yeah. So he's just like going to use it to start getting like reservations at different restaurants. It's so funny. Google has a product for that that does an AI phone call but for some reason it just hasn't really rolled out or it just hasn't gotten to like adoption. I don't know. There's some sort of, there's some sort of like memetic like I think people seeing the clip where he's like exploring, explaining how he's feeling. The AGI really, really hit People, obviously. The security thing, he has it make him a daily brief that feeds into an RSS on a podcast. So in the morning he can just listen to a five minute podcast on what his day's like, things that he should be responding to. Kind of like it seems like you. Could have this podcast ever your enemy texted you or something and you will be deeply upset when you see what happens. Everyone's preying on your downfall. No, no. Security is important with Claudebot now Moltbot. CrowdStrike is also important. Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches. A random 10 person team in Paris just dropped what looks like. It looks like something superior to Cloudbot. According to Chubby on X, it's called Twin. The AI company builder, Hugo Mercier. Yeah, they raised a $10 million seed round. They have over 100,000 agents deployed. Okay, so Tyler, give it a spin. Check it out. Twin Lives. There's a community note on here. This is undisclosed advertisement, but Doug over at Semianalysis Fabricated Knowledge says, I think that claudebot is going to be a moment and yeah, someone's going to do this. And I'm still wondering about how quickly a business can actually scale when you're getting constantly hammered with TOS violations from every Mag 7 legal department. Constantly. Like, hey, yeah, we noticed that you built a CLI for the web interface on WhatsApp and and we don't want you to do that. And it's one thing if it's like an open source repo that people are running themselves and can change and can edit and fix and tweak versus like hey, you're a corporation that we could potentially sue or tie up in litigation or go on a press tour around. It's just a very different dynamic. Yeah. And the prompt injection risk as well. Yeah, totally. It's like does that liability fall back on the company that. I'm excited to talk to Mark Gurman about how he's processing this with what will happen with Siri, the timelines there, what will be integrated because Apple and the iOS ecosystem should have a lot of the same functional hooks into these products. But we talked about this Monday, there was a team that built a product which sold to Apple and became shortcuts. And then they built software applications incorporated in a product called sky and they were pre launch and OpenAI acquired it. It was focused on integrating AI, the operating system. And so you can imagine OpenAI behind the scenes has been cooking on a lot of stuff like this, but this isn't the kind of thing that you can just at OpenAI scale just say, hey, we're going to just ship this and see what happens. Yeah. Whereas a startup or an open source project. Yeah. Let me tell you about Lambda Lambda is the super intelligence cloud building AI supercomputers for training and inference that scale from one GPU to hundreds of thousands. There are a lot of people building stuff on Multbot around Multbot around the concept. Another there's actually someone in the chat that's working on a claudebot for the cloud. He says he has a name. I'm very excited to learn more about his project. I think his name is Jeff. In the chat Brexton says yes, that was fast. Yup. Software is cooked because Kalyn Y says introducing Multbot for teams. One click connections to all the apps client side encrypted and builds your team's memory. Get your momo is the product. I wonder. This feels like almost something where if you've been building a product that's like this, you should potentially reintroduce it on the back of all the Multbot hype. And. If you're like a multi tentacled agent that can touch a lot of different systems, we'll put it in terms that people understand and say, hey, like this is, you know, this is claudebot or Mult Bot for the enterprise and it has these, you know, functions and these benefits and this security approach. But clearly a lot of people will be focused on this. Let's click over to re watching the interview with the claudebot creator, Peter Steinberger, because I want to hear about his AGI moment. The AGI achieved moment. First, I'm going to tell you about Gusto, the unified platform for payroll, benefits and HR built to evolve with modern small and medium sized businesses. And then we will play the video from claudebot creator. I want to see. So in November. Yeah. I don't know, you know, I wake up every day, I'm like, okay, what do I want to work on now? What would be cool? And then they was like, okay, I want to chat with my computer on WhatsApp because if my agents are not running and then I go to the kitchen, I want to check up on them or I want to do little prompts. So I just hacked together some WhatsApp integration that literally receives a message, calls cloud code and then returns what cloud code returns one shot and it took like one hour and it worked. Okay, that's kind of cool. But I usually use prompts like a little text and an image. Because images are like, they often give you so much context and you don't have to type so much. So I feel like this is like one of the hacks where you can prompt faster, just like, make a screenshot. So the agents are really good at figuring out what you want. So I hacked together images, and then I was on a trip in Marrakesh with like a weekend birthday trip, and I. I found myself using this, like, way more than I. Than I thought. But not for. Not for programming. It's more like, hey, we're like, there's like restaurants. Because it. It had Google in it and it. It could figure out stuff. And it's like, especially when you're on the go, it is like, super useful. And then. And I wasn't thinking, I was just sending it a voice message, you know, But I didn't build that. It wasn't a support for voice messages in there. So the reading indicator came and I'm like, I'm really curious what's happening now. And then after 10 seconds, my agent replied as if nothing happened. I'm like, how the f did you do that? And it replied, yeah, you sent me a message, but there was only a link to a file. There's no file ending. So I looked at the file header. I found out that it's opus. So I used FFMPEG on your Mac to convert it to wave. And then I wanted to use Vispa but didn't have it installed and there was an install error. But then I looked around and found the OpenAI key in your environment. So I sent it via curl to OpenAI, got the translation back, and then I unresponded. And that was like the moment where, like, wow. Yeah. You know, it's like that's where it clicked. These things are like damn smart, resourceful beasts if you actually give them the power. Sure. App. Love it at the end. I love it. Yeah. People were reacting to this. People are not. Are genuinely not ready, says Vittorio Lots. Well, that. That's the one way. One way. The reason that I feel like that is such a powerful moment and I'm glad that he shared it, is if you give somebody. If you give. If you're talking with a model and you give it a task and then it just hits a dead end. It's just incredibly like. That's sort of like people are very used to that right now. Yeah. And it's not that it needs to be that way, but it's just kind of like the steady state. Yeah. People are used to. Okay. I know what the models can do. That's effectively the agent having real agency that makes it an agent. Yeah. It's saying like, well, I didn't know how to do this, or I was confused and I tried a number of things until I did. I did what you wanted. Right. And that's like what you want out of. That's what you want out of a team member. Right. If you're working with somebody on a project and they have a task, they don't just try one thing and come back or just say like, I actually can't handle this because the file type is wrong. Convert it, figure it out. Yeah. What was your reaction to that? Yeah, I mean, it's like, pretty insane. Definitely. It definitely raised my chance of permanent underclass. Oh, no. You know, it's making me a little worried. Yeah. Will says it's over. It's over. We need to move. And there are a lot of people quoting this. G. Fodor has the Eliezer Yudakovsky meme. And lots of people here. Lots of people were, you know, interesting that he uses Codex here. Rune said Codex 5.2 is really amazing, but using it from my personal and not work account over the weekend taught me some user empathy. Lol. It's a bit slow. And Yu Chin Jin says, every time I ask my OpenAI friend, when will you beat Claude at coding? They say, we already beat them. I think Sam is realizing speed, not intelligence, is Codex's blocker to a Claude code moment. Cerebrus chips, Cerebras chips might unlock a Codex moment. Yeah, that would be very interesting. And that's one of the. The models are so powerful now that there's so many moments where you know it's going to deliver. It can deliver because it can work around all these problems. But if you're like, I need this done in five minutes, I'll just do it myself. Yeah. Like 5.2 Pro is like an incredible model, but it just takes like 5, 6, 10 minutes every time you prompt it. Totally. It's kind of like thinking or, you know, deep research. Type model. Well, let me tell you about Label Box. Reinforcement learning environments, voice robotics evals, and expert human data. Label Box is the data factory behind the world's leading AI teams. I like the fireworks. I'll do some fireworks for that. I like the fireworks. What else is in here? So you already mentioned this, Tyler, but Peter's probably in the top 10 users of Codex at the moment. Over 250 billion tokens in a few months is a lot. So how much does that actually cost? That feels like you're up in the hundreds of thousands of dollars. No, it says it here. Yeah. How much does it say? $51,000. That's not much. Wow. Yeah. Usage cost. And he has a streak going of 74 days. No days off. Yeah. So $51,000 for this type of result. Pretty remarkable. So where do we move the goalposts now? Yeah, I'm. So what do you want? I know there's going to be something you're unhappy with about AI, like you can't be satisfied. Me? Yeah, me. Yeah. Permanently dissatisfied. Well, part of it's a bit. Oh, yeah, I know. Obviously, because we got to move the goalposts, I have to provide like an altered opinion. But of course we have the goalposts that have been sitting in one place. But I guess what I'd be interested in is Peter and his team built a hit product. This novel experience that has taken the world by storm. And how much did we know? They only spent a few months on it. How many people actually worked on it and what was their total. How much did they spend in total on tokens? Right. Because they were spending some with codec. You sort of see that from the GitHub commits, right? Probably for some sort of average. Yeah, yeah, I'm sure you can do that. I mean, just based off that, it's like, okay, $51,000 on 250 billion tokens. And a lot of those tokens are probably using claudebot to do things right, not just building it. Right. Yeah. I mean, that was over how many months those tokens were not just on cloudbot. Yeah, that was just all of his, like hundreds of. And he says he talks to Opus 4:5 all the time too. So I wonder what numbers he's putting up there. But still, it's interesting that it's not in the millions. It's pretty accessible. I don't know. We need to find a new AGI benchmark. Okay. So I think maybe it's something like this. Right. So I think one of the main takeaways I have from this is just that, like, big model companies can't really release an equivalent product just because of the integration thing. Right. Mag7 companies are not going to talk to each other. You're not going to get OpenAI in WhatsApp stuff like this. This. So it really. But do you think they would go so far as to say, we are not going to allow you to navigate our applications using. Using something like. I mean, that's literally happening with like, The New York Times does not allow OpenAI to browse it with a bot. Like, you can go, yeah, but. But it's different having your local device being effectively used by. Yeah, no, it is a computer just using itself at the same time. If OpenAI has the ChatGPT app right now, if you go to the ChatGPT website and you say, hey, I found this, I'm a subscriber to the New York Times and there's an article in the business section that I want you to summarize for me. Here's a link. Can you open the link and turn into some bullet points for me? It'll just say, like, no, no, no, I can't go over there. Like they said, don't go. I'm not going right now. If I download the ChatGPT app and it's able to run that and maybe the New York Times can't block it, the New York Times still has the ability to sue OpenAI and, like, be a headache. So then they can just say, hey. Force the same thing about a situation where the user. Let's say they have a Mac mini, they're running Moldbot, they on the computer, they've given Molebot their New York Times login, so it's fully logged in and they're paying New York Times subscriber. I would be. I think users would be upset that you're saying if the New York Times says we're not going to allow. We're going to figure out a way to try to. I don't even know how you would go about trying to actually stop that activity. But I'm just saying, even I can imagine users are basically going to demand if my data. I don't care if my data is stored in the cloud. I want to be able to access it on any device that I'm logged into, even if I'm not physically present with the device. I just think it's going to be a really, really, really tough argument for some of these larger companies to say you can only access your data if you are physically moving the mouse yourself. Yes. I want to get Tyler's response, but first I'm going to tell you about Cognition. They're the makers of Devin, the AI software engineer. Crush your backlog with your personal AI engineering team. Tyler, what do you think? Okay, yeah, so I agree. I think basically what you're describing is that all these companies need incredible deal guys, right? This is the whole thesis, like with open source, with this hacker culture, you can kind of always get around these rules. Is it like maybe breaking terms, service like def. Yes, definitely. These big companies can't do it. They need more deal guys. Right. But it's like hard to find these deal guys. You need the AI to become the deal guy. So you need a claudebot deal guy. That's the new benchmark. That's the new benchmark. You're gonna be good at making deals. Okay. Claudebot AI that can do a deal between two Mag 7 companies. There we go. There we go. Move it to the other side. See ya. See ya. Tyler, why do you. I kind of want to take this a little bit further just because, I don't know, like, how do you actually enforce this? Is Google Drive going to say you need to have your camera on and we need to see that you're sitting in front of your computer actually using it? I mean you can always do these like bot detection stuff with like whatever it's coming from the same ip, these things. I think it's like kind of a constant race, but like, I think the open source stuff will, if you like really grind it out, you can kind of always get around this stuff. And so like the only, only real solution is to have like actual, I think, deals between these like different companies. Yeah, I think that if you, if you have an app from a big company that enables you to do something like, like if Chrome started shipping with a torrent browser, something that would allow you to manage torrents which are not illegal, that's just software. You can just download torrents that are not piracy. I can just be like, hey, I put up a torrent of TVPN content. Anyone can download it. But just by virtue of the fact that it's so prevalently used to pirate media, they would face pressure from media owners Warner Brothers and Disney and they would get angry letters and they would strike a deal and say, okay, yes, like we're not going to enable this. And so even though, so if any of the big tech companies launch an app that enables behavior that other apps don't like, they will have leverage and they all have deals together and they all have legal teams and there's a bunch of different pressures that they can apply and it's this like constant negotiation. Oh, well, you know, maybe we won't renew this deal or this contract. Maybe we'll go with someone else. If you're the bad actor, you'll get pushed out. Right. So I don't know. I think it's going to be a big debate. I would be surprised if any of the big labs just launch something that can actually just go and do anywhere and do anything because the pushback. But do you think, do you think this is something that will like, why would this just not happen at the OS layer? Yeah. I don't know. I mean then you need Apple to like actually lock in. You do, which is like. Yeah. Or Satya. Yeah. And that's why we're having Mark Gurman on the show to break down Apple's strategy and what they're doing anyway. Interesting. MongoDB choose a database built for flexibility and scale with best in class embedding models and re rankers. MongoDB has what you, you need to build. What's next? Kimi 2.5 is out and Alex Chima is running it on his desk. It runs at 24 tokens per second with two M3 Ultra Mac studios connected via Thunderbolt. He went much bigger than the Mac Mini and Joe Weissensk. By the way. Apparently Kimi likes to be refers to itself as Claude. Yeah. What happened here? Tyler, can you break down what's going on with Kimi? Yeah, I mean it's always kind of unclear with the Chinese open source models, like what, what they're actually doing to train the models. So there's like. Yeah, when you ask the model to like introduce yourself, it's like, hello, I'm Claude. Yeah. So there's probably like two scenarios. One is that they trained on the outputs of like OPUS or entropic model. Right. You'd think that they would just find and replace in the training data though. That would be pretty easy to just be like, okay, we're going to scrape a ton of responses from Claude. Let's make sure we do a find and replace on Claude so that if it says, hi, I'm Claude, I'm a helpful assistant, it just changes that. That's just like a. Yeah. I mean you would expect that Kimmy knows that it's gonna call itself Claude. Yeah. So it's like, why would they allow that? Maybe they're just basically like mogging anthropic like, oh yeah, we have Claude. What are you gonna do about it? Yeah, like what are you gonna do? We stole Claude. Yeah. And then there are also some rumors, like, very unclear if this is just like completely fake headlines. But like maybe there was some leaked checkpoint that somehow got out to China and then they did a fine tune on it or something. Yeah, that would be pretty, pretty crazy if. Yeah, people have said that before for other models. I'm not really sure how reliable they are, but it seems like fairly very Likely that they just trained on opus output. Yeah, yeah. Because you kind of people said Deep Seq did that originally with ChatGPT. Yeah, yeah. I talked to one technologist about what it takes to reverse engineer like the gpu and it's a surprisingly low amount of outputs to sort of interpolate all the weights or something that approximates it. So it's clearly sort of a game back and forth, just like a constant war. There's somebody. I think it is quite. I think it's probably a good sign for us AI labs that the Chinese models are essentially just like completely based off the US ones. Like they're not actually training these models from scratch. Yeah. They're not doing these incredible training runs for much cheaper. Like people say they're actually just. They're basically just like copying off the us. Do you think that's because of just chip restrictions or actual architectural hurdles? Yeah, it's probably all of them. It's hard to figure out the architecture. You don't have the chips to try these different science projects. Training runs. Wow. The. The Chinese models lag by exactly how long it takes to scrape the API basically and do a fine tune. What a coincidence. Dean Ball has been on a tear the last 24 hours. He said people significantly underrate the current margins of AI labs. Yet another way in which pattern matching to the technology and business Trends of the 2010s has become a key ingredient in the manufacturing of AI copium. And he says this Derek Moeller post. Was very informative here. Just look at the market clearing prices on inference from open source models and you can tell the big labs pricing has plenty of margin. Deep Infra has GLM 4.7 at 43 cents in, $1.75 out. Sonnet is at $3 in, $15 out. How could anyone think Anthropic isn't printing money per marginal token? And Dean Ball says the reason they think labs lose money is because 10 years ago some companies in an entirely unrelated part lost money on office rentals. We work and taxis, Uber. And everyone thought they would go bankrupt because at that time another company that made overhyped blood tests, Theranos did go back, did go bankrupt. That is literally the level of ape like pattern matching going on here. The machines must look at our chattering classes and feel great appetite. Yeah, you could also have like seen that the margins from the whole open code thing with Anthropic. Oh yeah, right. Where there was open code. So Anthropic, they have Claude code. Right. You can get the Claude subscription and you get like free Claude code tokens and then you could use those to auth for open code and they removed that. And the whole reason you would want to do that is because the Claude code tokens versus actual Claude API is like 10x difference. So there's like a massive markup to the actual API. So you'd assume that unless Anthropic is losing just insane amounts of money on which maybe they're losing some amounts of. Money, it's like, well, you look back at the. I forget which interview Dario was doing where he was trying to get people to think about like you think of each model as a company where you spend all this money on training, which is capex. And then when you're actually running the model it's very profitable. But if you look at the business as a whole, you have massive, massive losses from training and stock based comp and hiring 1400 of the best engineers in the world. So if you actually look at it on a company company wide, you have continued scaling massive losses. But the important thing is effectively at the product level, when they're selling the product, they actually are making money. So in the S1, good to you and to me, looks like training clearly broken out as capex, solid gross margins above 60%, something like that on inference. And then all of the AI copium can probably subside. If OpenAI or Anthropic go out with an S1 that shows really solid inference margins and if it comes out that it's like, oh, their inference margin is like 10% or something, then people are. Going to be panicking. To be clear, I think these will be some of the most special S1s that have ever graced the financial markets. It'll be great. Let me tell you about Sentry. Sentry shows developers what's broken and help them fix it fast. That's why 150 thousand organizations use it to keep their apps working. And without further ado, we have Mark German, the German manager, managing administration.
I don't think they want to use Llama. I don't think they want to work. On open source models. So there aren't that many games in town. Well, let's take a step back here. So they're using Google in the us right? But they're gonna have to do something in China. Okay. And so you'll see them use a combination of Alibaba and I think it's Weibo or Tencent or one of those different AIs for different features. So they'll use them. And just because they're partnering with Google. Gemini on Siri in this chatbot doesn't mean they're not using other players. There's a lot of OpenAI and a lot of applications. They launched their Creative Cloud competitor today. And a lot of those AI features are powered by ChatGPT. Like some image generation stuff. I still think on an image generation. Side you're getting a little bit better. On OpenAI than you're getting from Google. And then a lot of stuff internally, like Apple runs on Anthropic at this point. Anthropic is powering a lot of the. Stuff Apple's doing internally in terms of product development, a lot of their internal tools. So that's a big one to watch. They have custom versions of Claude running. On their own servers internally too. Because this Google deal, this just came. Together a few months ago. They were not going to use Google. Apple actually was going to rebuild Siri around Claude. But Anthropic, they were holding them over a barrel. They wanted a crap ton of money from them, several billion dollars a year and at a price that doubled on. An annual basis as well for the. Next three years or so. From what I understand at the time, Google was really an afterthought because they. Were in the middle of the trial. With the Department of Justice and then. For some reason the judge ruled that Apple and Google's deal was kosher, even though everyone knows it's a huge issue and a monopoly and all that. I'm not here to be a judge. I write, I don't make judgments for legal proceedings. But anyways, Apple and Google get off scot free. They can do whatever they want now and you know, they're not being held back at all. And then obviously OpenAI, that is a. Real firestorm for Apple right now. Like OpenAI obviously.
Well, let me just take a step, please. Yeah, yeah. You know, with Apple and AI. Yeah. So in 20, was it 2018, they hired John Gene, Andrea, he was this high flyer at Google, he ran AI in search and Apple thought they had a coup here. Apple thought they would hire this guy and really just hit the ground running and be at the forefront of artificial intelligence. Just seven years earlier, they announced Siri in 2011. There was absolutely nothing like it. It was breakthrough, but then it just became utter junk. Google Assistant lapped it, Alexa lapped it. So they thought they were going to bring this guy in and it'd be a game changer, turns out. And maybe this will be Tim Cook's fait accompli. But this was the biggest mistake, this hire of Tim Cook's tenure. I think it's easy to say Apple is so behind in AI. There has been so much ink spilled on this and so many conversations on this. And I written about and talked about it half a million times. I think it does. You haven't even scratched the surface about how big of a problem this is for Apple. Right. They've completely screwed up AI in every which way. And it comes down to just hiring the wrong people and entrusting the wrong. People. But is it a do nothing win scenario? Because I think we're seeing a situation now where the Mac Mini might end up selling out. Boxing is. Fine.
And we'll be there for a gig. Paula says SF escape room called the permanent underclass. And it's just a room with a laptop and Claude code installed. How do you get. There's something here that's funny. So good. Have you ever done an escape room? No, I have never once in my life thought that that would be fun, nor thought that I would. This is how I want to kill time. Yeah, it's like an hour. Well, if you're good, you get out faster, right? Yeah. Have you. Are you. I'm not into them, but I've done them and they're fun. Sometimes it depends on, like, if it's a well structured one. But I like puzzles. Like, it's fun. It's like fun puzzle, like, figure out. But you can very quickly get caught in, like, just like, overthinking what's going on and being like, oh, it must be some, like, complex math thing. And it's like, no, you actually just needed to, like, like, press this lever instead of, like, analyze the situation or something. Are some people in there using their phones to try to figure stuff out? Because I imagine you could just say, I imagine there's only. Take a picture. Probably only solve. This is like a series of rooms. Yeah. Well, usually you. You start in one room. They're all different. But oftentimes you'll start in one room and then you'll. There'll be a series of puzzles, locks and keys and whatnot. And then oftentimes, like, you'll unlock something and then you'll progress to another room, and then you'll progress to a third room, and then you'll finally get out of, like, a series room. Because it's a lot of, like, they're pretty easy to set up in sort of like a defunct office space or like, you know, storefront. That's just kind of like, you know, going in between. It's like the spirit Halloween of commercial real estate. Like, you just like, anyone can come in and just say, like, oh, we'll be in there for like a couple months. It's not super permanent. The build out's pretty simple. It's mostly just like, some walls and decorations and some creativity on the puzzle side. Trey says, can you guys put Tyler in an escape room and see if he gets out before the show ends? I mean, it'd be very hard to find a three hour long escape room. I think most of them aim for, like, 45 minutes. Well, maybe we need to make one. One of our guests had to leave the Ultradome and go straight to an escape room. So there are escape room fans among the TVPN army all over the place. Have you done one, Tyler? I've not. Are you interested in it? You're a speedcuber, so, you know. Yeah. I mean, it could be fun. I don't know. I'm kind of in the Geordie camp, though. I've just like, never been. It's very, like. I don't know, maybe it's great with. It's probably great with, like, kids who are maybe, like, 8 to 10, who can do the puzzles. Like, family event would be fun. And then it was a common thing. Like, when they came out, there was definitely a boom in, like, in escape rooms. Bobby in the chat says escape room would be a good benchmark for AI, Certainly a humanoid. I mean, you could go in there, take pictures and be like, help me solve this, help me solve this, help me solve this. I know, but I want to see a humanoid run through it. Let's call 1X, see what they can do. Call any of them. Hit them up. Figma. Figma make isn't your average vite coding tool. It lives in figma, so outputs, look.
We'll see. Yeah. And what. We were getting into earlier. Before the. Show started. Is. Is Netflix's decision, the bigger, like, the decision that's bigger than just like, are we going to sort of lean into AI or not? Will just really be. Do we have. Does Netflix ever lean into ugc? Right. They've been signing some bigger podcasts and they obviously work with independent media companies. But there's a very big difference between just allowing people to, like, will they ever have an upload button? It feels like no. And I feel like that could end up being their advantage. Where there is. Part of what has made YouTube magical since the very beginning was that anybody could go and put YouTube, you know, upload a video. Anybody could be a creator. And I think, like, as the amount of content that gets created, ten hundred X's, thousand X's because of AI, it's going to be, yeah, Netflix could be this, like, refuge where you're like, okay, at least if I go here, I know that there was some filtering process. I know that this isn't just a total free for all. Yeah, yeah. The upload button is probably a bigger deal than like, AI on Netflix. Exactly. I think that's a very good thing.