Weekly Recap: GPT-5, Apple Sitting on $100B, Disney Enters The AI Race, Marc Andreessen

26:49They say, what happened next was evidence that Hollywood's much-discussed, much-feared AI revolution won't be an overnight robot takeover. Johnson approved the plan, but the use of a new technology had Disney attorneys hammering out details over how it could be deployed, what security precautions would protect the data, and a host of other concerns. They worry that the studio ultimately couldn't claim ownership over every element in the film if AI-generated parts were in there. So if there's AI training data from a DreamWorks film in there, and they use the DreamWorks training data to make a Disney film, even if it looks like Dwayne The Rock Johnson, DreamWorks might come knocking and say, hey, give us a royalty.

27:33I think that's the risk. But the lawyers are having fun, maybe. Full employment for lawyers over at Disney, clearly. Disney and Metaphysics spent 18 months negotiating on and off over the terms of the contract to work on the digital double, but none of the footage will be in the final film when it's released next summer. They went and shot it. A deepfake Dwayne Johnson is just one part of a broader technological earthquake hitting Hollywood. Studios are scrambling to figure out how simultaneously they can use AI in the filmmaking process and how to protect themselves against it. Is it sustaining or disruptive or both?

28:03Can't be both, but we'll see. While executives see a future where the technology shaves tens of millions of dollars off a movie's budget, they are grappling with a present filled with legal uncertainty, fan backlash, and a wariness toward embracing tools that some in Silicon Valley view as their next century replacement. And if you are trying to sell an AI tool into Hollywood, you've got to get on Adio. Customer relationship magic. Adio is the AI native CRM that builds, scales, and grows your company to the next level. and you can get started for free. So the Academy of Motion Picture Arts and Sciences is surveying members on how they use the technology.

28:40Studio chiefs are shutting down efforts to experiment for fear of angering show business unions on the eve of another contract negotiation. And no studio stands to gain or lose more in the outcome than Disney, the home of Donald Duck, Bell, Buzz Lightyear, Stitch, and countless others, which has churned out some of the most valuable and protected creative works in over the past century. So my take on this. So two years ago, I was hanging out with the founder of a very large generative AI image generation company. And he was telling me that by 2025, anyone with a laptop and an internet connection could generate a full Hollywood movie about anything they want with a single prompt.

29:22And it was a hilarious conversation because we were on a Zoom call and his internet wasn't working. And it was the classic example of like the technology is amazing, but we got a lot of stuff to iron out. So anyway, extremely aggressive timeline, but obviously things are going to change for Hollywood. So my question is, will Disney benefit? It feels like a moment to be in founder mode, but Walt Disney died in 1966. So basically everyone believes that meta will benefit from AI, even if they miss the train on owning the next dominant consumer tech platform. But if Disney got really AI pilled, what would that look like?

29:53I don't think they need to train their own foundation model, just like they don't need to build their own cinema cameras. They can just use IMAX when the time calls for IMAX. They can use Blender when the time calls for Blender. They can use Houdini when the shot calls for some high-level VFX. But they do need to rethink how they structure their business and negotiate with unions and underwrite content. They might need to go more risk on, not just from a brand risk position, but taking more smaller bets. We're in this weird barbell world where everything seems like it's either a hundred bucks and it's shot on an iPhone.

30:29It's a viral like TikTok or it's a hundred million dollar blockbuster with like 50 million dollars of VFX. Actually, that's kind of a low number. It's usually like 300 million dollar production with 150 million dollar VFX and then no one sees it and it's a flop. But then they hit every once in a while and they're great when they're good. But it's this weird like venture style betting at the high end, but there's nothing in the middle. and maybe that's like the death of the arthouse film, but I'm just wondering if in the age of AI, like maybe there's this interim step where Disney ladders down a little bit and gives like 10 filmmakers$10 million each and says, hey, you're still required to deliver a 90-minute full film, but you're doing it for 10 mil and it's not quite Blair Witch level production, one notch up.

31:13You got to be creative. This is what May 24 is doing. Yeah, I'm really bullish on this idea that historically a TV show would film a pilot episode, and they would use that to get the budget to shoot an entire season. And you can imagine now, even for film, you can just make the trailer ahead of time with AI. The other advantage that I think Disney has that's very real, just going in, and why they just broadly seem to be positioned very well here, is that I think that broadly, like content customization will probably take off because Disney can make a film now and then you could make millions of different variations of it that become interactive with the underlying fan.

32:04So imagine you're watching Moana, but your kid is in the film, is a character in the film, and you can now do that at scale, right? And how much more would you pay as a parent to have something like that? I have a funny story about this. Last thing I'd say, so customization and just like democratizing, like basically making, being able to make variations of films, I think is going to be big. I think just the time, it's the same way it's so difficult to make a new luxury brand. It's so difficult to create. It's easy to create IP. It's hard, extremely difficult and time intensive to create valuable IP.

32:43and Disney's advantage is they have this sort of like 360 and that they can make a film and they can bring it to Disneyland. They can bring it to a cruise. They can create physical products around it. And so they develop IP in a way that, that new entrants are not, you know, they don't have the benefit of like having a Disneyland where they can make new experiences that, that increase the value of that IP. Right. So I think, um, so two things, one, if I were to go back, one of my favorite Disney properties is Star Wars, A New Hope, the very first film. If I went back and was like, let's customize that for me, I don't know that I would make any changes.

33:22Like, do I really want a scene where Han Solo breaks the fourth wall and says like, hey, John, like I'm about to go, you know, save Luke at the Death Star. Like that doesn't improve the product for me. I actually like that. It's just the vision of George Lucas. So I don't know about customization being better for me. I don't know what I would change. But here's an example of how to make like a magical experience for a kid. Like imagine after a movie ends, a kid could interact with a character and ask it questions in real time, like about the story or have a conversation with them. And that's what I'm talking about, like bringing that IP to life.

34:03Totally. Right now that exists. And I was obsessed with this when I was a kid. I would watch Star Wars and then I would read the books in the expanded universe. And I remember even having books that were just like encyclopedias of every single ship. You could read into this a little bit more, but we'll leave that where it is. But I would learn every single ship, this is what a Star Destroyer does, and it would have all this backlog and stuff. And so, yes, I agree with you. That's very cool. You finish the movie, and then you can just interview Luke about how does the lightsaber actually work? And he can talk about the Kyber crystals.

34:37That would be very cool. On the flip side, I had a very funny experience. I was watching a horror film in high school with a couple friends and I grew up in Pasadena and this horror film just happened to take place in Pasadena. Like that's where they set the film because Pasadena, California is just like a place where you set films. It's just a real place. So if you've seen Kill Bill, which I know you haven't, but Kill Bill by Quentin Tarantino, there's a scene where Uma Thurman just shows up and And it says at the bottom, Pasadena, California, because like that's where the character went. It exists in the real world.

35:14But this horror film that took place in Pasadena was terrifying because I was watching it and I was like, this is happening here and it's night and it's dark outside. And like now I'm so much more immersed. And so I was thinking back then that you could use like the IP address of a connected, I think we were watching it on like a PS3, like a DVD player. Like you could use the DVD player to dynamically change the location of the establishing shot. So you're like, this whole horror film is going to take place inside of like one house where there's like a monster in the house. And it's going to be like the usual, like they're upstairs, they're downstairs, there's blood, there's, you know, someone's running and chasing.

35:53Like it could be any town USA, but they usually establishes like this is happening in Amityville. This is happening in, you know, some random town, Lake Placid, right? But they could easily just dynamically change that with a few establishing shots to just show you, okay, it's happening in Malibu. And now it's a lot scarier for you. So I think there's something interesting there. But again, it has to be the work of like an auteur. Yeah, it's not something you can copy and paste either because it works for some shows. But then others, like the place is so obviously the place that it would throw you off as a viewer.

36:28Exactly. And so I do think that the fully AI generated, fully custom content, that will exist, but it will exist on independent third party platforms. It won't be the domain of Disney. This will be something where if you in the future, if you really want to see like, you know, AI generated stories about surfing in Malibu, like there will be an endless stream of those. and you will be able to go and experience that particular content. And you can already kind of experience that because there's probably some Instagram person who makes really great content about surfing in Malibu. And if you follow them, you get that vibe and that might be what you're into.

37:05And you can kind of like, and it's handled just by the great democratization of creativity. And we should talk to Samir about this later. He's coming on the show. So we can talk about the future of content creation and whatnot. On the question of should Apple make a huge acquisition, Tim Cook actually addressed it in the earnings call. He said something like, we've acquired around seven companies this year, and that's companies from all walks of life, not all AI oriented. We're very open to M &A that accelerates our roadmap. app. We're not stuck on a certain size of company, although the ones that we've acquired thus far this year are small in nature.

37:48And he said, like, we're acquiring one every few weeks, which is not quite seven per year. I mean, I guess that's like one a month, but a few weeks, sure. So he's doing deals, but he's not looking at anything massive. And I think that actually makes sense. My question was, if you were the CEO of Apple and you had$100 billion burning a hole in your pocket what are the other things that you could do with that money formula one team you could buy every team and the division and every track and the and all the sponsors i made a post about this it's crazy it's so much money it's so much money i made a post about this jokingly earlier this year i still think it'd be cool for apple to just be an f1 in the way that red bull is i think that'd be great the team no no i agree yeah i think that'd be great the one thing culturally is like Apple's brand is very premium and very high end.

38:42I don't know if they could handle a few losing seasons. This was the early bear case for Apple TV. I was talking to someone, a very successful Hollywood producer, and he was saying that the nature of Hollywood is like venture capital. It is a hits driven business. If you don't accept that you're going to have massive flops and egg on your face and embarrassing moments, you will not make it. And just like any venture capitalist who is like, I only want to make investments where they won't blow up. You're not going to make it right. Like, yeah. Mark Andreessen had a, had a, there was a clip from a episode he did recently talking about how he lose sleep over the companies you don't invest in, Exactly.

39:28Not the ones you invest in and don't work. Exactly. Exactly. And so you have to be risk on. You have to be risk on in Hollywood as you do in venture. And so the question was like, can Apple deal with spending 50 or 100 million dollars on some show that completely flops and everyone's like, this is a disaster? And they've navigated it really well. I think that they have had some flops, but they've kind of tucked those behind the scenes. Whereas Netflix has started to get more kind of, you know, people talk trash about, what was that Red Notice or they did some massive movie with The Rock that kind of flopped.

40:02They've done a few of these big movies that haven't done that well. And it's sort of like, oh, like, you know, they're not, you know, God's gift to, you know, producing. But it doesn't matter because overall Netflix is doing very well. And all it takes is like one squid game to like carry the whole quarter, basically, just like in venture. Or like one, you know, you sign the Friends deal and then or The Office and people are watching that like 20 years later. So and apparently so F1 has surpassed half a billion dollars in box office earnings as of last week. I mean, it's an example of a of a home run.

40:41But it's not like they're doing that once a quarter by any means. No, and John Exley's in the chat. Shout out to John Exley. Unfortunately, I didn't get a chance to connect with him at our party in New York, but he did attend. You got to chat with him. A lot of the team got to chat with him. So thank you for all you do, John. We really appreciate you. The entire team is giving you a round of applause because you've done a ton of work for us and we really appreciate it. Anyway, if you had a hundred billion dollars burning a hole in your pocket at Apple, what would you do? I have a bunch of crazy ideas.

41:15So most people would just say, hey, you're going to benefit from AI because you are the window into all technology. And it doesn't matter if it's generative video or text. People will be consuming it on devices. You sell devices, you'll be fine. Just like search did not destroy Apple. It actually made Apple stronger because people search on their iPhones. So most people would just say, hey, return it to the shareholders. Apple's already done that. They've returned over a trillion dollars to shareholders in the past 10 years. Those are kind of rough numbers, but it's basically that, which is an insane amount of money to return to shareholders.

41:50But Samsung is worth 330 billion. They could buy the entire company levered up with two thirds debt and they would just own both sides of the smartphone market then. Probably extremely anti-competitive. I think Lena Khan would have a conniption fit, but funny concept. Funny concept. The other idea is massively expand the retail footprint. So So Apple right now, I was surprised by this. How many retail stores do you think they have worldwide? I don't know, like 599 or 601. Did you look at it? Did I tell you this already? We talked about this this morning. It's 535. So it's not that many. It's mostly in the tier one cities.

42:30There's usually only like one in every city. They don't take the Starbucks strategy where they put them across the street from each other. But for$100 billion, they absolutely could. $100 billion is enough to open 6 ,000 new Apple stores. So they would be 6 ,600. So they would be up around 7 ,000 stores, which for reference is as many Subways as there are and as many CVSs as they are and as many 7-Elevens as there are in America. And so everywhere you see a Subway, you can see an Apple store. And I feel like this would be a big upgrade for America. I feel like if you just walked around every American city where tier one, tier two, tier three city.

43:09You just see a beautiful sheet of glass and it's an Apple store on every corner in America. They could also acquire every firearms store in the country and turn them into Apple stores. They could do a take private of Figma. Think bigger, build faster. Figma helps design and development teams build great products together. Get started for free at Figma.com. I'm still buzzing from Thursday. It was a fantastic day. We had a lot of fun. Fantastic. Exhausting, but we're back in the game now. Anybody that is tuning into the show for the first time today is going to be hearing our ideas for Apple and just thinking, these are some of the worst I've ever heard.

43:46But they're definitely fun. Okay. So speaking of glass from the Apple stores, they could buy Corning, which makes Gorilla Glass. That's a$55 billion company. Then with the money left over, they could buy InterDigital, which owns all the patents on 5G and 6G. So they could basically patent troll everyone else in the industry. They just get extremely anti-competitive. They could also, with the money that is left over, we're still in the first, like, let's vertically integrate this company mode. They could also buy 10 rare earth or cobalt mines. So they own the supply chain there. And then they could also build four battery gigafactories and lock up the sapphire crystal supply chain that's used on the iPhone camera.

44:28So they could just completely own their entire supply chain for 100 billion. And that's just one year, one year. And then they, you know, the proposal was do this every year. The other thing that was interesting about Tay Kim's post is if you, if you listen to the earnings call, Tim Cook and the team were very clear that they don't even want to finance like data center development themselves. They're still working with these sort of like third party lenders to like capitalize them. And so again, Tim Cook is the king of efficiency. Yeah. He knows his business. and these should we get into some of that i have one more one more so opening i they aqua hired johnny ive they're coming out with something that's competitive it's going to be some some sleepless nights at apple when that thing drops right it's gonna be stressful because like maybe maybe maybe it doesn't go super well but like you know it's a serious threat opening as a serious company they have a lot of customers and they got johnny of your former goat devout designer ready to drop, you know, the next, amazing, amazing team over in, in, in hardware at open AI.

45:35So the question is like, how do you fire back with your Apple and you have a hundred billion to spend? Here's what you do. The cogs on an iPhone are less because they have a margin. So you take that a hundred billion dollars, you buy 200 million iPhones and you send a brand new iPhone to 200 million Americans. And you're just like, oh, how much are you charging? Johnny Ive and OpenAI? Well, ours is free this year. We're doing a free iPhone for the entire year. They should just do something like that where it's like your 11th iPhone in a row is free. They'd probably have 200 million people with free iPhones.

46:14And they're just like, yeah, we're actually, we want to keep you in this ecosystem. We like our services revenue. We're excited to be a platform. We're excited. We're just giving it back. We're just giving back. Can you imagine? I think that would also be extremely anti-competitive. I don't even know if there's a regulatory framework to stop that from happening. But I mean, price competition is a real thing. And like, if you discover that your competitor is undercapitalized and can't win a capital war, you could potentially cut prices so dramatically that it's like, yeah, the other phone's like cool, but you know, this one's free.

46:50I guess I'm staying in the blue bubble world. Tim Cook, next earnings call. All right, we're going to buy all of the minds we depend on and we're going to make the iPhone free. Stock just yeets like 80%. He's like, I'm petty, I'm petty. And you know what, Johnny, I'm not going to let you win. I would rather die than let you win. Yeah, I would destroy myself. Yeah, I've been reading a lot about cutting your nose to spite your face. I'm pretty into it. I'm pretty into the strategy. Anyway, let me tell you about Vanta. Automate compliance, manage risk, improve trust continuously. Vanta's trust management platform takes the manual work out of your security and compliance process and replaces it with continuous automation, whether you're pursuing your first framework or managing a complex program.

47:34Go talk to the purple llamas, Vanta.com. Anyway, so Ben Thompson's been noodling on Apple and AI strategy. He had a rough go because I think he took some time off or shifted. He shifted his posting schedule right when all the earnings dropped and there was all this crazy stuff going on. So he's just writing about Apple earnings now, but he still has some good takeaway saying even if Apple does swoop in and buy Mistral, for example, the overall point holds, Apple Apple isn't going to compete with OpenAI, Anthropic, Meta, X, and Google on pursuing AI superintelligence. Like five extremely well-funded, extremely serious teams.

48:20Apple's going to go their own way, centered on their devices and their direct relationship with their customers, at least as long as Cook is in charge. And anyone that's saying that Cook needs to step aside, I think, is in for a world of hurt, because he's been on a generational run, and he's solving the most important problem for that company, which is not the application layer in artificial intelligence. It's the supply chain and keeping the flow of iPhones coming. Making and selling phones. Put the phones in the boxes, Tim. He's doing it. He's doing it. He's really underrated. He's been CEO longer than Steve Jobs was now.

48:58He's the longest tenured CEO of Apple in history. He better be bringing some of these AI researcher comp packages in as comps when he gets in front of the comp committee at Apple. Yeah. Hey. That is true. He should definitely be getting a bump. If he does have researchers, I mean, he has AI researchers. They are improving different functionality. I feel like maybe I'm hallucinating this, but it feels like the text-to-speech engine has gotten better on my phone. I listened to a semi-analysis article with the default Apple Safari text-to-speech, and it was very listenable. So I think we're getting better there.

49:37We have some breaking news about SiriusXM. They are canceling the Howard Stern Show. Can you call it a cancellation when the guy's 71 years old and he's been doing it for 20 years? They say it's no longer worth the investment. They've been paying him$100 million a year. That is a huge salary. And that's, what, three, four times Colbert? Wow, that's big. That's power. power of uh of radio power of you know the power law he's done fantastically so congrats to him Howard if you're looking for a new gig you're welcome to come and hang out at the ultra dome you can sit at the intern table next to Tyler and we'll we'll we'll bounce ideas off you yeah it's wild uh I really wonder where Sirius XM business goes right I know I know where this goes post AGI, you're going to listen to every hour of Howard Stern.

50:31I know you've listened to zero hours, but there's probably 20 ,000 hours in the catalog or something like that of Howard Stern content. You could listen to it from the beginning. Do people listen to the back catalog ever? Absolutely not. SiriusXM is still a$7 billion company. That's bigger than I would have thought. I wonder, yeah, revenue, like how much of that cost. I mean, it makes sense to give him a huge slice of that it is a talent driven business and he could go to spotify he could go somewhere else my question is like he is old will he retire or will he do a podcast or do something independent is there news now they uh what they apparently did how much okay you look that up while i tell you about wander find your happy place book a wander with inspiring views hotel great amenities dreamy beds top tier cleaning 24 7 concierge service it's a vacation home but better So they did 8.6 billion in 2024 revenue.

51:26So they're trading at less than 1x revenue, which says when you have a shrinking business, it is a rough place to be in. Not a growth stock, I suppose. Yeah, I wonder how much of that is getting eaten up by talent. This is a trillion dollar stock if they have. Yeah, I wonder how much, I wonder what percent of Sirius XM content is like power loss, celebrity, host, red, high salary contract versus essentially programmatic or AI content. Because it's probably not even AI, but if you're just like, there's a station on Sirius XM that's just the Grateful Dead. And it just plays them the whole time. Like, I don't think you need someone making a hundred million dollars to like randomly play Grateful Dead tracks.

52:13Like there's probably some Grateful Dead fan who manages that and picks the songs and orders them. But like that could essentially be pseudo random. Maybe one day you go through the back catalog and then the new stuff and then you mix it up and then you play the hits or something. I don't even know if they have hits in that way. I know it's kind of a jam band. But I wonder of their content, of their tonnage, is that the term? Tonnage is like the amount of content on the network. I wonder how much of it is driven by these high dollar deals. And I wonder if they'll get a new host in the seat. I wonder who this generation's Howard Stern is.

52:56Maybe it's like Tim Dillon or something, some irreverent comic essentially. I think the thing is, as content is on demand, fewer and fewer people are just turning on the radio or turning on the television and listening to whatever just happens to be playing. And so if you take people that were subscribed to Howard Stern, and were like, you liked Howard Stern. We're now just going to play this other person through his channel. They're just going to ask, what's going on here? Yeah. I mean, certainly if you have a specific car that has Sirius XM and it doesn't have Spotify and Howard Stern's new show is on Spotify because it's a podcast, you might stick around on Sirius.

53:38Maybe they get someone new in the seat who's almost as good or can build a relationship. But it does seem like a challenge to fill that. But also, it's a lot of money to pay. So clearly it wasn't penciling out. So they had to move on. OpenAI has announced an open source model. They said they'd never do it. Everyone's saying it was in the name. It was in the name. How could you have doubted? Oh, they're closed AI. Closed AI. They're not open AI. They're closed AI. Everyone thought they were so clever. They said they couldn't possibly open source a model. And they did. They did. They did it. The unthinkable.

54:11First open model in years. Right. They used to have that. Was GPT one open source? Like when was the last time that, so we're going to talk to Tyler about this. He's in the studio today. Was still open source? And then they decided to close source something. It's tough that I treat you like things that I could Google. And I know that you just have to search it or chat GPT it. But good luck. Get me the information on the history of open AI's open source models. In the meantime, I will read this post from Sam Altman. GPT OSS, all lowercase. He's still in lowercase mode, although he uppercases the other sentences, but he threw in.

54:56He likes lowercase. Is he sending you messages, John? I'm seeing it. I'm seeing coded messages in here. GPT OSS is a big deal. It's a state-of-the-art open weights reasoning model with strong real-world performance, comparable to 04 mini that you can run locally on your own computer or phone with the smaller size. We believe this is the best and most usable open model in the world. We're excited to make this model the result of billions of dollars of research available to the world to get AI into the hands of the most people possible. Hit that soundboard again. I want to keep it going. We believe far more good than bad will come from it.

55:34For example, GPT-OSS-120B performs about as well as O3 on challenging health issues. We've worked hard to mitigate the most serious safety issues, especially around biosecurity. GPT-OSS models perform comparably to our frontier models on internal safety benchmarks. We believe in individual empowerment. Let's hear it for individual empowerment, people. I feel empowered the way that you're reading this, John. Although we believe most people will want to use a conventional service, a convenience service like ChatGPT, people should be able to directly control and modify their own AI when they need to.

56:09and the privacy benefits are obvious. As part of this, we are quite hopeful that this release will enable new kinds of research and the creation of new kinds of products. We expect a meaningful uptick in the rate of innovation in our field and for many more people to do important work than we're able to before. OpenAI's mission is to ensure AGI that it benefits all of humanity. To that end, we're excited. I think I cut this screenshot off. We're excited for the world to be building on an open AI stack created in the United States based on democratic values available for, I didn't screenshot the rest of that, but oh, here it is, available for free to all and for wide benefit.

56:52So I imagine that that was the voice that was in his head. As the author intended. Exactly, exactly. Will DePue has more notes here, some context. So less than one year between 01 announced, which was September of 2024. And we have an O3 level model open sourced that's runnable on consumer hardware. Wild progress. You were highlighting earlier, Sam initially had a poll. Saying like, you know, should we, what do you want? Do you want an O3 level model or do you want something that you can run locally? He did both. He did both. Wow. Yeah. So we were reading that. So it was what, like six months ago or something, that post that you shared?

57:32Yeah, I think so. So I asked Tyler to dig this poll up. Sam Altman shared a post that it was a poll on X saying, what do you want? Do you want an O1 level model reasoning model, a frontier model open source? Or do you want something you can run on your phone? And it was neck and neck. And people were clicking on the phone model. They were like, we want the phone model. And then Dylan Patel quoted it and said, oh, you idiots, don't vote for the phone model. You can distill llama. Let's get a truly frontier model. Let's get a frontier model. And so, so Dylan Patel was very much like, like you guys don't know what you want.

58:07Like, like the phone models are available. That that will be easy. The hard thing we got to twist open AI's arm to open source the reasoning model, the O3 level model. And Sam Altman just said, you wanted this or this. It was a false dichotomy. You get both. You get both. Yassin says all caps. I'm so sorry for doubting you, Sam Altman. I'm so sorry for saying that you were the antichrist. I didn't realize your plans were measured in centuries. I hope you forgive me for everything. I'm so glad you kept control of open AI. I'm so sorry. Please forgive me. That's remarkable. Never talk down on the future first ballot hall of famer.

58:45Apparently, the future leader of open source AI will be very interesting to see how meta responds. Will they stay in the open source game? Also, just how important is open source? Is this a Linux scale opportunity? Is this an Android scale opportunity? Like is this, you know, is this GitLab to GitHub? Like what's actually the long-term play? Certainly cool. Great for the community. We're going to see fun experiments. We're going to see interesting things done with this. I think, do we want to talk about some of the ideas that we had for building on top of this open source model? Fine tuning. I can just run through some.

59:23I know there was one you didn't want to share because you thought it was too good, too much alpha. But what was I talking to you about, Tyler? What were our ideas? We had one that was... Well, initially we weren't sure there was going to be reasoning. That's right. So we were going to add fake reasoning. Yeah. So it was going to be a model that gaslights you into thinking it's reasoning when in fact it's not reasoning. So if you look at the actual reasoning UI, it's just UI. It's not actual reasoning. It's a time delay. It's a time delay. It's a delay. And then while it's... Thinking really hard.

59:57Let me think about that. What would a super smart person think right now? So we can actually still kind of do that because it's not a multimodal model, but we could like, you know, we just add some kind of like totally blank encoder or something for the image. And then it's like, this is an interesting image. Let me think about what's in it. Yeah, one of the best images I've ever seen. A lot of people are saying this, this is one of the greatest images. Thank you for sharing this PNG and or JPEG with me. It just knows nothing about the image. it's great i i i i'm seeing that this image is uh is is large in file size it's purely metadata based analysis it doesn't know anything about what's going on like wow this is a tough question i'm gonna have to generate a lot of internal reasoning tokens to answer this test time inferences really is magical isn't it anyway back to thinking okay i'm reasoning about this now yeah so the other the other the other obvious opportunity you know people have been frustrated with sycophancy with the models and models being.

1:00:56Well, not me. I've been frustrated with the lack of sycophancy. Yeah, you want more. You want more. I want more. You want the model to gas you up. But, you know, fine-tuning it to go, you know, over the top, somebody could release the Glazonator 3000, which just tells the user, you don't just think you're goaded. You are goaded. I know this. You know this. It's like, I'm asking you to solve IMO question six. Why are you talking about whether or not I'm goaded? It's like it just tells you like you got this one. The timing here is interesting. I mean, everybody's been banging the table saying, where is the American open source AI or LLM leader?

1:01:35It was reported yesterday by the misinformation that reflection, the misinformation reported that Reflection, one-year-old startup, we've had the CEO on before he's in talks to raise$1 billion to develop open-source models to compete with DeepSea, Meta, and Mistral. And yeah, all of the Chinese open-source models have been on a tear. Quen released, Quen Coder recently. Yeah, people are into Quen. Or a new version of it that is getting good results. so it's important work so one interesting thing that's kind of related to the US-China race is like, so when Deep Seek first came out, everyone was like up in arms because they said it was less than$5 million to train right?

1:02:26So in the model card for these models they actually reveal the GPU hours it took, which then you can figure out how much it probably cost to train. Give me the number and tell me if I need to ring the gong for them So for the big model, it was probably around like$4 million. Yes, but the small model was like$500K, probably a little less. But that's like 10x better. The model is way better than Deep Sea Guard 1. Okay, okay. So should I ring the gong? I mean, I think it's pretty sweet. I wanted to hear$400 million at least. I was expecting$900 million. $900 million. They got a billion of revenue coming in every month.

1:03:09But isn't it? No, it's cool that it costs so cheap. revenue on the open source model. No, no, it is very cool. So we're ringing the gong for the elegance of the training run. Yeah, yeah. Efficiency. Efficiency. Congrats to everyone at OpenAI on efficiency and cost control. Delivering a great model, saving time and money. Putting money back in the hands of shareholders and not in the hands of open source developers, I suppose. But yeah, I mean, do you think that, is this like a pre-training scaling wall narrative? Like, would you, as someone who is a potential consumer of an open source model, like, do you want a$400 million open source model?

1:03:49Would that necessarily be better? Because it seems like they were able to distill it pretty well, get it to the frontier, not, you know, burn a ton of money on training. Also, the question is, like, if they're, this is probably distilled from their other models that are much bigger. So like you, there is no world where you could just spend$4 million and get this level model without also have done, having done the GPT four run, the GPT 4.5 run, et cetera. Right. Yeah. Maybe it's also just like, even if they didn't distill it, they just have the knowledge of like how they did it, which is like arguably more valuable.

1:04:24Right. Yeah. Yeah. Yeah. I think in terms of like open source, it's either you want to go like super, super cheap and super small, which they kind of did here. or you go super big, almost like the llama, like the behemoth, right? Yeah, yeah, yeah. I think it would have been really cool to see that literally state-of-the-art level model. You could also try and ask Claude, hey, build me a frontier open-source model. Don't make mistakes. Yeah, we're going to have someone from Claude on soon. This is funny because, wait, you said 500K for the small model? Yeah, probably less. So that's the same amount of money that they're putting up for this challenge, the red teaming challenge, have you seen this?

1:05:02So, to encourage researchers, developers, and enthusiasts, that's me, from around the world to help identify novel safety issues, this challenge has a$500 ,000 prize fund that will be awarded based on review from a panel of expert judges from OpenAI and other leading labs. At the end of this challenge, we will publish a report in open source and evaluation data set based on the validated findings so that the wider community can immediately benefit. So if you can hack this thing, if you can get it to teach you how to develop a nuclear bomb or take over the world, you might have 500K in your pocket.

1:05:39Not bad. Go get some of that. That's a seed round in 2012. Welcome to the stream. How are you doing, Mark? Hey, what's happening? Great to see you. Yeah, you too. A lot. It's a little bit of a slow news day, but exciting stuff with GPT, open source. It's not a slow August. It's not a slow August. We're glad. We were just reflecting that we've taken exactly one day off this summer. That was July 4th. And we're showing the Europeans how American companies work. American work ethic. We're setting an example. And we have proof of work because we exist on the internet. And you can see us live every day.

1:06:18So we're setting an example. How are you doing? How's your summer going? Fantastic. Going really well. So how long is it going to be until you guys put up avatars that make claims that you're working hard all through the summer, but it turns out you're on the beach? You might have caught us. I think you'll know better than us as to when the technology gets there. We've been demoing some of the stuff. People have been doing a lot of deep fakes of us, and fortunately, all of them have been clockable. So it doesn't feel like a brand risk, but they're getting closer and closer. And I know that there's going to be a moment where we have to say, hey, that's actually using our name and likeness to endorse something that we don't necessarily endorse can you please take that down so we're we're approaching the the the touring test the the uncanny valley we're escaping i think a question like looking back over the you know maybe 10 or 10 or 15 years was was what moments did you feel like there just was not a lot of action happening because this summer is just the pace uh from so many different teams has been absolutely insane everybody's like trying to keep up.

1:07:18And it didn't used to feel that way, at least from my point of view. So my view of it always is there's like these, there's this, these disconnected, you know, kind of patterns or trends. There's, there's sort of the, the sort of day-to-day phenomenon where like engineers show up every day and they make things a little bit better. And then every once in a while, you know, you get a technical breakthrough or a new platform. And, and, and that process kind of this, you know, kind of sawtooth kind of up to the right kind of process kind of plays out over time, kind of regardless of what else is happening in the world.

1:07:47And so it keeps happening through recessions and depressions and wars and, like, all kinds of crazy stuff that's happening. But basically, you know, the technology keeps getting better. So there's kind of that curve. And then there's the sort of enthusiasm curve and then the adoption curve, you know, which is basically, like, when do these things actually show up in the world? And then, by the way, when are people actually ready, you know, for the new thing? Like, if you talk to people who work on language, I'm sure you guys have talked to people who work on language models they will tell you that they were surprised that ChatGPT was the breakthrough moment because they thought everybody already knew what these models could do for three years before that.

1:08:20And so they were shocked that it was the chatbot interface that made the thing go. And so there's somewhat of a sort of arbitrary disconnection between what's actually happening in the substance and then what people are seeing and feeling. And so it's really hard to predict when these things pop. But also, if you're in this day-to-day, it's really hard to tell when things are going to be hot or not because it doesn't necessarily map to how much the technology is improving. Yeah, we were just talking about that in the context of Google's new world model. It's this like generative video game that you can kind of move around in.

1:08:53And it feels like DeepMind is just absolutely crushing at the AI research frontier. They have the best world model simulator that you can walk around in. The question is like, if they let another lab do the chat GPT thing and just get it out into the consumer three months earlier, they might wind up kind of chasing and trying to catch up if somebody actually figures out how to make it like a dominant consumer product. Now, in the enterprise, it's more oligopolistic, but consumer seems to be winner-take-all. I guess the question is like, how much value do you place right now in the AI race to just like moving fast, breaking things, you know, having like the thick skin to deal with like the safety constraints and all of the different stuff obviously not being irresponsible but just speeding up the organization as much as possible it feels like now is the time to really push on that yeah well first of all i need to correct you it's it's moving fast and making things uh i don't know where that's right i i don't even know where that came from yeah i i i i i have no idea never heard never heard of it i mean he didn't really break anything i i think that's a good point it really did just move fast and make things the first things it made were weird but that was fine and it failed and it It hallucinated a ton, but it didn't really break anything.

1:10:05I don't know. Yeah, I believe in this case, total deaths attributable to JetGPT are still zero. Zero. So notwithstanding all of the caterwauling. But yeah, so look, I think the AI industry in particular has a very acute version of the sort of challenge that you identified with. And I don't say this negatively, just an observation, which is that in sort of a normal technology company, you've kind of got engineers who make products, and then you've got salespeople or marketing people who sell them. In the AI companies, you have this third tier of the quote-unquote researchers, which has worked out incredibly well.

1:10:41I mean, the researchers have just done amazing breakthroughs at these companies. But the handoff, there's not necessarily a clean handoff from the researchers to the market. And so it kind of raises this question of like, okay, are these companies there for kind of three-segment companies where they have research and then they have product development and then they have go-to-market? And I think that's a really open issue. I mean, Google's kind of a case study of this. You alluded to DeepMind, but even more broadly, Google developed a transformer in 2017. And then they basically let it sit on the shelf, right, because it was a research project.

1:11:14They didn't productize it. They were very worried about, from people I've talked to, they were very worried about the brand issues and safety issues. They had all these reasons to not productize it. I talked to somebody senior who was there at the time, and I asked them, when could you have had ChatGPT with GPT-4 level output? But if you had just gone flat out starting in 2017, and they said by 2019, they already knew how to do it. And then they've now caught up, but it took an extra five years to catch up. And so I think a lot of these companies kind of have that challenge. Elon, as usual, of course, is provoking this question, as I'm sure you guys talked about.

1:11:51But he has now, with NXAI, he's now collapsed. He's eliminated the distinction between research and product. And so, of course, he's pushing this as hard as he can. And I think it's a good question for a lot of these other companies, kind of how hard they want to push on actually getting these things in fully productized form out to the market. Yeah, yeah. On Elon's distinction, it feels like there is more research to be done, but it feels like we're entering a new cycle of just focus on the engineering, focus on the deployment, the applications. Let's get all this technology out into the world.

1:12:22Let's reap all that benefit. And yes, there will be a different track of fundamental research that's happening somewhere, but it's really, really hard to predict. And so if you have something that's working, just double down and just go really aggressive on it. I'm wondering more on that, but also on Apple's strategy. It feels like Apple's been kind of like, you know, people have been maligning them for not, for missing the AI opportunity. And Tim Cook's just there on the earnings call being like, look, we acquired a couple small companies and seven companies. But then it seems like they're taking more of like an American dynamism approach.

1:12:59Like there was news today in the journal that they're investing$100 million in American manufacturing. They're certainly doing stuff. They're just not chasing the, you know, the shiny tennis headline$100 billion capex. So I'm wondering about your thoughts on on when you have a you know, when when you have a platform? How hard is it to resist chasing the new shiny object? Is that the right move? Or are there any other things that you think Apple should be, you know, changing their strategy on? Yeah, so look, Apple's always had this, you know, very clearly defined strategy that, you know, Steve, Steve and Tim, you know, working together figured out a long time ago, which is, you know, they forget the exact term, but it's something like basically, they invest deeply into the core of what they do.

1:13:44You know, they'll basically work internally on things for many years. They only actually release things when they feel like they're kind of fully baked. And so as a consequence, they have this thing where, and Jim says this, right? They're rarely first to market with new technologies. They're more often in the category of what Peter Thiel calls last to market. They'll come out whatever three years later, whatever, five years later. There were tablets for years before the iPad. There were smartphones for years before the iPhone. Folding phones. They're about to do a folding phone. It's like 10 years into that technology.

1:14:15I'm sure if they do, they'll hit it. The last mover. The last mover. Yeah, yeah, yeah. Sorry. The last mover. And I guess, yeah, what I would say is like, look, that clearly works if you're Apple, right? And so it clearly works if you're Apple. But I would say there's a fine line between that strategy and just simply becoming obsolete, right? And so the problem is like, if you're not Apple and you don't have all the other kind of super strengths and, you know, kind of now the market position that Apple has, you know, do you really want to be a company, you know, if you're not Apple, do you really want to be a company that basically sits there and says, yeah, the world's moving and we're very deliberately not going to lean as hard as we can into it.

1:14:47And so I think there's a lot of survivorship bias in these kinds of strategy discussions where people look at the one company that's able to pull this off and they don't look at the 50 other companies that are in the graveyard, you know, because they, you know, because, because they didn't adapt. I mean, you know, all the other smartphone companies, when the iPhone came out, they were like, oh yeah, well we could do touch too. Right. You know, we'll just, you know, we'll get to it. Right. And, you know, you know, they're gone. What do you, what do you think? The Blackberry Bold, I remember. if it's like an iPhone knockoff.

1:15:12What do you think, you know, right now people are, or variety of, you know, shareholders are annoyed at Apple around their reaction to AI, LLMs. John's annoyed around just like transcription generally, just like super basic stuff. But it doesn't feel like the core businesses immediately threatened today. It feels like it's still on the horizon around these sort of like, you know, eyewear based computing, you know, potentially net new devices that we're that that we'll see from you know companies like open AI over time but where do you like like how how real is the threat you know this year versus ten years from today and and kind of what's your framework?

1:15:55Yeah well look I mean I think the biggest ultimate danger I mean the biggest ultimate danger is very clear which is just like at what point do you not carry around a pane of glass in your hand you know called a phone you know because other things have superseded it and you know look everything you know everything becomes obsolete at this point um so there will there will come some time to make sure when we're not you know carrying phones around and we'll we'll watch movies where people have phones and we'll be like yeah look at up look at how primitive they were right because because we'll have moved on to other things and whether those things are eye-based or you know uh you know the other kinds of wearables or whether it's just kind of you know computing happening in the environment um or just you know entirely voice-based or you know who knows what it is but um you know there will come a time when that happens you know is that time three years from now because there's like some you know huge breakthrough you know from from some company that figures out the the product that absolutes the phone right away or is that 20 years from now because the phone is just you know such a standard platform for everything that we do in our lives and everything else you know kind of remains a peripheral to the phone i mean that you know that's you know that that's the game of elephants that's playing out there um you know obviously i think you know i think it's highly likely that we'll have a phone for a very long time having said that it is exciting that there are companies that are going directly at that challenge.

1:17:00And, you know, whoever cracks the code on that will be the next Apple. And by the way, that may in the fullness of time be Apple itself. You know, they may be the company that figures that out. Yeah, I remember being at a board meeting at Andrews and Horowitz maybe a decade ago or something, and Chris Dixon showed me the HoloLens. And I was like, okay, we're one year away from this being everywhere. And I feel like today I'm still in the like, yeah, VR, it's definitely one year away. The next quest I'm going to be wearing daily. And it feels like we're always there, but it does feel like Apple did a lot of work on the fundamental pixel density of the resolution of the display.

1:17:37And then Meta has been doing a ton of work on just getting it light and affordable. It feels closer than ever, but you always got to wait until you see the churn numbers until you really call the game, right? Well, here's the other thing, but I think that's true. But you'd also say, you know, I'm on the on the meta board, so I'm kind of a dog hunting this one. But like the meta Ray-Ban glasses are a big hit. Oh, totally. Right. Like they're a big, you know, so I think we now have a form factor that we know works, you know, for for iBase wearables is, you know, there's not VR and then VR, you know, on top of that.

1:18:07But, you know, just the, you know, the glasses and, you know, and then the glasses of camera, you know, sort of integrated camera, integrated microphone, integrated speaker. You know, that's a very interesting platform. You know, the watch clearly works, by the way, which Apple, of course, you know, is played a significant role in making happen, you know, that now sells in huge volume. You know, so that's the second data point. And then, you know, like, I think these, you know, these, these, I think some form of AI pin is going to work. I also think, you know, headphones are going to get a lot more sophisticated, which is already happening.

1:18:33And so, you know, you do have these, you know, kind of data points coming out. And then, yeah, look, the trillion dollar question ultimately is, are these, are these peripherals to the phone, you know, which is what they are today? Or are these replacements for the phone? And, you know, we, yeah, I would say we, you know, I think we have a lot of invention coming both from new companies and from the incumbents who are going to try to figure that out. Yeah, I always think about the value of like narrowing the aperture on these new technologies. Like with the meta Ray-Bans, I feel like the fact that they aren't also trying to be a screen is actually a feature, not a bug.

1:19:05And I always go back to the iPhone. Like it was first and foremost a phone. And people bought it because it could make calls. And then it could make text messages. And then it was an iPod. But do you disagree with that? Please. Well, you guys might be too young. The first iPhone actually was a bad phone. How so? For the first two years, I couldn't reliably make phone calls. I had the third one and a friend had one, but I feel like it was still like people were carrying cell phones and that was at least the expectation. But yeah, I mean, I guess you're right. For the first two, it was a classic Apple story because for the first two years, the thing couldn't reliably make phone calls.

1:19:41And then it turned out there was an issue with the antenna and with how you held it. and there was a email. I remember that. Yeah, you would hold it and you would disconnect it. You could basically brick the device. Yeah. Based on how you held it and somebody emailed, this is when Steve would respond to emails from random people and somebody emailed Steve saying, if I hold the phone this way, it doesn't make phone calls. And he's like, well, don't hold it that way. Yeah, that's great. Yeah. Right? So even there, it was like, yeah, okay. And people forget, it took like five years for the iPhone to find its footing.

1:20:07It took like two years to get the thing. I remember also the original iPhone didn't have, it didn't have broadband data. It was on the old 2G. It was called the AT &T Edge Network. So it didn't have broadband data. And then, of course, it didn't have an app store. It was completely locked out. So the challenges for Apple now is that people are so used to perfection with the device that launching a product that isn't perfect is embarrassing. You look at the Vision Pro, and it's like, well, the battery's big. Steve would have hated this. He never would have shipped this. and that being constrained and not being able to innovate because you're tied to this impossible standard of being on whatever generation 17 of the iPhone and perfecting every element is a real challenge.

1:20:51So I would say there's a corollary to that. One of the things I've observed over the years is I think LG products become obsolete at the precise moment, they become perfect. And to your point, what I mean by perfect basically is like, yeah, it's like the perfect idealized complete product. Like it does everything you could possibly ever imagine everything a customer could imagine, everything you as the technology developer can imagine, it's absolutely perfect. And there's been tons of examples of this over the last 50 years, where it's like the absolute perfect, it seems to be the permanent version of that product.

1:21:22And then it just turns out that's actually the point of obsolescence, because it means creativity is no longer being applied right into that platform. You're just like, there's just nothing else to do. You're just like, you're done, right? The product has been realized. And then the cycle is what happens to your point, the cycle is other people come in with completely different approaches, completely different kinds of products that are broken and weird in all kinds of ways, but are fundamentally different. And so that is one of the time-honored traditions. And one of the things you could say about Tim is his willingness to kind of break the mold of Apple only ships perfect products, but being willing to ship the Vision Pro shows a level of determination to kind of stay in the innovation game, which I think is very positive.

1:22:00Yeah, yeah, yeah, yeah. That's great. Updated thinking on open source. since we last talked. There's a lot that's been happening yesterday. Yes, OpenAI is open again. Yeah, look, very encouraging. A year ago, I was getting very distressed about whether open source AI was going to be allowed. It was even going to be illegal. And I think we're basically through that at this point. We're through that in the U.S. We'll see about the rest of the world. And then look, the U.S.-China thing is obviously a big deal, but I think it's been net positive for the world that China has been so enthusiastic about open source AI coming out of China, which has been great.

1:22:40And then, yeah, look, OpenAI leaning hard into this, you know, and releasing what they did is, I think, fantastic, both because of what they released, which is great, but also just the fact that they are now, you know, willing to do that. And then Elon reconfirmed overnight that he's going to, you know, open source, you know, start open sourcing previous versions of Grok. And so, yeah, so we, you know, we seem to be in a timeline where open source AI is going to happen. You know, right now, you know, I think what you would say is it kind of lags the leading edge proprietary implementations by, you know, six months or something like that.

1:23:09But I think that, you know, that's a good, if that's the status quo that continues, I think that would be a very good status quo. What are the rough edges that we need to kind of sand down when we're thinking about Chinese open source models specifically? Is it we need to do some fine tuning on top of them to add back free speech? Or do we need to watch for back doors? say it's phone and home if it runs into this specific thing like the Chinese open source thing It was remarkable because I feel like it really does accelerate the pace of innovation because everyone gets to see Oh, this is how reasoning works.

1:23:38I think that's great at the same time It made me very it made me much more appreciative of AI safety research and capability research and actually being able to interpret What's going on and say definitively this model is gonna behave weird in this weird way like the Manchurian candidate problem We haven't found any of that but it certainly seems like something we'd want to keep an eye on. But from your perspective, what are the risks that we need to be aware of going into a world where China is really pushing hard into open source? Yeah, there's two. And you identified them, but let's talk about both of them.

1:24:09So the phone home thing is the easy one, which is you can put a, you know, you can packet sniff, you know, a network and you can tell when the thing is doing that. And plus you can go in the code and you can see what it's doing that. And so you can validate that that's either happening or not happening. And I think that's important. But I think people are going to figure that out. You can kind of gate that problem practically. The bigger issue is we have this term in the field right now called open waits. And open waits is a loaded term. It uses the open term from open source. But of course, with open source, the thing is you can actually read the code.

1:24:48With open waits, you have just a giant file full of numbers, as you said, that you can't really interpret. And then what you don't have, what most of the open-source, open-weights models don't have, including DeepSeq specifically, what they don't have is they don't have open data, or open corpus. So you can't actually see the training data that went into them. And of course, most of the people building models are kind of obscuring what that training data is in various ways. And so when you get an open-weight model, the good news is the software source is open. The good news is you can run it on your new machine.

1:25:19You can verify that it doesn't phone home, but you don't actually know what's happening inside the weights. And so I think that that is going to be a bigger and bigger issue, which is like, OK, how the thing behaves like, yeah, what what has it actually been trained to do? And what restrictions or directives has it been given in the training that are embedded in the weights that you need to be able to see? You know, this is I would say this is coming up as sort of, I would say, a global issue, you know, which we worry about when these models come from China. other countries worry when these models come from the U.S., right?

1:25:49So one of the phrases you'll hear when you talk to people kind of outside the U.S. is kind of this phrase people are kicking around, which is, not my weight's not my culture, right? Or by the way, for that matter, not my weight's not my laws, right? Which is like, okay, like, what actually is this thing going to do, right? And to your point, the Chinese models, for example, might, you know, never criticize, you know, communism or something. I can tell you, the American models have all kinds of constraints also, right, implemented, you know, usually by a very specific kind of person in a very specific location in the U.S.

1:26:20And so, you know, I think that this is a general issue and we're going to have to see basically people's tolerance levels being willing to run open weights models where they don't fundamentally have access to the data. And then correspondingly, I think what we'll see is more open source developers also doing open corpus, open data. So you can see what's actually in them. Yeah. Obviously, open source is very important in terms of just distributing intelligence broadly, giving people the ability to run their own models and really fine tune them and have control. There's also the big push just to make frontier models and high capability models free.

1:26:55One model is you charge for the premium, you give the free away. It's a freemium model. That's what we're seeing at most of the labs right now. There's also this kind of specter on the horizon of potentially putting ads in LLMs and what that would do to the world. Jordi got in a little dust up with Mark Cuban on the timeline, deciding whether or not it would be a net good to put advertising in LLMs, what might happen that might be bad there. Yeah, my point broadly was that ads have been an incredible way to make a variety of products and services online free. And just saying like default, just no ads would potentially be incredibly destructive.

1:27:39but yeah, curious, your framework. Yeah, so I should start by saying whenever I personally use an internet service, I always try to buy the premium version of it that doesn't have ads. And so if I can live personally inside an ad-free universe and pay for it, that's great. And I'll freely admit whatever level of hypocrisy or incongruence kind of results from that. No, the point is choice. The point is choice. Well, the point is exactly what you said, it's affordability. So the problem is, If you really want to get to a billion and then five billion people, you can't do that with a paid offering at any sort of reasonable price point.

1:28:14It's just not possible. Global per capita GDP is not high enough for that. People don't have enough income for that, at least today. And so if you want the Google search engine or the Facebook social app or the whatever frontier AI model to be available to five billion people for free, you need to have a business model, you need to have an indirect business model and ads is the obvious one. And so I do think if you take some principle stand against ads, I think you unfortunately are also taking a stand against broad access just in the way the world works today. And then look, the other really salient question is the same question that the companies like Google and Facebook have been dealing with for a long time, which is are ads purely destructive or negative to the user experience or are they actually, if done properly, are they actually either neutral or even positive, right?

1:29:04And this was something that, you know, Google, I think, to their credit, figured out very early, which is, you know, a well-targeted ad at a specifically relevant point in time is actually content. Like, it actually enhances the experience, right? Because it's an obvious case. You're searching on a product. There's an ad. You can buy the product. You click to buy the product. That was actually a useful piece of functionality. And so, you know, can you have ads or other things that are like ads or look like ads, you know, different kinds of referrals, you know, mechanisms or whatever, can you have them in such a way that they're actually additive to the to the product experience and you can just like with search and with social networking you could imagine lots of examples of that people will you know the people will you know the whiner and lots of different ways but i think you know i think that hasn't been a bad outcome overall um and i think that uh i think it's entirely possible that that's what what happens with with these models as well yeah so uh kind of similar kind of question what what should be legal kind of trying to create legal frameworks on a number of issues with AI.

1:30:02There's been a number of IP cases that have been working their way through the courts. What can labs use to train models, et cetera. There's been some good outcomes recently. Sam also was talking about how a lot of people are using AI as like a confidant, like a friend, things like that. And he mentioned that currently your chats are not privileged. They can be used in a lawsuit or other situations. How optimistic are you that our sort of legal system in the U.S. can get some of these issues right where maybe it can't just be, you know, total free markets, kind of lawless, whatever goes? You know, so in the case of training data, I think that there I mean, there's a bunch of these copyright, you know, kind of lawsuits happening right now.

1:30:49There's, you know, the big New York Times opening. I won. And there's been a bunch of others. um yeah i i think in that but for that particular problem my guess is that problem ultimately has to be solved through legislation um it's it's ultimately a legislative question the reason is because it goes to the nature of copyright law itself you know which which is legislation and and and of course you know the the the content industry is already claiming that of course you know using using copyrighted data to train you know without permission without paying is is sort of you know they believe illegal on his face you know due to violation copyright law the counter argument to that, which, you know, which we believe is, well, it's not copying, right?

1:31:22There's a distinction between training and copying. Just like in the real world, there's a distinction between reading a book and copying the book, you know, as a person. And so there's going to need, I think, you know, the courts are trying to grapple with that. There's a whole bunch of cases. There's jurisdictional questions, you know, probably ultimately Congress is going to have to figure out a, you know, figure out an answer on that. And by the way, the president has kind of, you know, thrown down that gauntlet in his, I think the speech he gave last week or two weeks ago, you know, where he said that, you know, Washington probably needs to deal with that as an issue.

1:31:50So that's one. On the privacy thing, I think that one feels like it's a Supreme Court thing to me. It feels like that's the kind of issue that's the Supreme Court. And in other words, like whether, for example, your transcripts are considered your property and whether they're protected against, you know, warrantless search and seizure. And the observation I would make there is if you look at the march of technology over time. So the Constitution has like very clear, you know, fourth, fifth amendments, you know, very specific rights around the, you know, the things that are yours, you know, such as, you know, your home, you know, being in your home, you know, by the way, the thoughts in your head, right?

1:32:26You know, that the government can't just like come in and take they can't, you know, they can't just come in and search your house without a warrant. You know, they can't like, you know, put you in a jail cell and beat you until you fess up like, you know, there are, there are, you know, we have constitutional protections against the government being able to basically, you know, take information, you know, fundamentally, you know, as well as possessions. And then basically what happens is every time there's a new technology that creates a new kind of sort of, you know, thing that you own, you know, thing that's yours, thing that you would consider to be private, thing that you wouldn't want the government to be able to take without a warrant, you know, out of the gate, law enforcement agencies just naturally go try to get those things because there are ways to solve crimes.

1:33:04And, you know, it feels like that that's a legal thing to do. And then basically the courts come in later, and they, you know, rule one way or the other, and basically say, no, that that actually is also a thing that is protected against, you know, warrantless, for example, warrantless search, you know, warrantless wiretapping. And so I feel like, you know, this is the latest of probably, I don't know, 20 of those over the last 100 years. And, you know, I don't know which way it'll go. But I think it's going to be a key thing, because as you know, people are already telling these models, you know, lots of things that they're, you know, that are very personal.

1:33:33Okay, lightning round, quick questions, We're letting you get out of here in a couple of minutes. We're in this age of spiky intelligence. Models are great at some things and then terrible at others. Where are you actually getting value out of AI right now? Where is it falling down for you? How are you using AI day to day? Yeah, so I have two kind of, I don't know, barbell approach. One is for serious stuff. I love the deep research capabilities. And so, and I'm doing this in a bunch of models, but like the ability to basically say I'm interested in this topic. Like, you know, I just felt like, write me a book.

1:34:04And I, you know, I'm kind of hoping for the longest book I can get. I always tell it, like, go longer, go longer, more sophisticated. You know, but the leading edge models now, they're getting up to, like, 30-page PDFs, you know, that are, like, completely well-formulated, you know, basically long-form essays. You know, it's just, like, incredible richness and depth. And, you know, if it's 30 pages today, I'm sort of crossing my fingers, it'll get to, you know, 300 pages coming up here in the next few years. And so, you know, I'm able to basically have the thing generate enormous amounts of reading material with just like, I think, incredible richness and depth and complexity.

1:34:34And then on the other side of the barbell is humor. And I've posted some of these to my X feed over the last couple of years. But I think these models are already much funnier than people give them credit for. Really? Yeah, I think they're actually quite highly entertaining. A while ago, I posted that. Is it specific formats? Like props or is it just chatting back and forth? Be me, be Marc Andreessen, that format. Take a dip in my pool, in my office. They're really good. So they're really good at green text. That works really well. But for some reason, the ones I find hysterical are the I haven't write screenplays, you know, for like TV shows or plays or movies.

1:35:12And I posted I had it right. New season of the HBO Silicon Valley, you know, set 10 years later. and I had it write like an entire I had it write like 10 scripts for a complete season and of course I just said make it like Silicon Valley except it's happening in 2021 it kind of peak woke and I thought it was just I'll sit there at 2 in the morning just laughing my ass off at how funny this thing is and so I think these things are actually already extremely funny, they're extremely entertaining when they're used in that way and I do enjoy that a lot and I generate a lot of those that I don't post

1:35:49It's probably your property. Yeah, hopefully the fourth amendment holds on these. Yeah, I have one last question. Go for it. And then I've got one more. How do you get a job as a venture capitalist in 2025? So I mean, look, the best way the best way to do it is to have a track record early as somebody who is like in the loop specifically on your product development. And so somebody who you'll be like deeply in the trenches at one of these new companies in one of these spaces, is participate in the creation of a great new product and a great new company and really demonstrate that you know how to do that.

1:36:22There are great VCs who have not done that, but I think that is sort of a foundational skill set for working with the kinds of founders that you want to work with who are going to want you to have kind of very interesting things to say on that as I think is still the best way to do it. Yeah, like feel the growth. Immerse yourself in the growth, the aggressive growth environment, and then you'll be able to identify it when you see it from afar. Yeah, that's right. Last question for me, state of M &A in your mind, how are you advising, you know, companies where you're on the board or just the portfolio broadly around what they should expect now and in the near future?

1:37:00You mean in terms of whether you can get things approved? Basically, yeah. Yeah, yes. So look, approval is still, approval is not a slam dunk. There was, you know, I just saw there was a medical device company this morning, you know, where the acquisition was not allowed by the FTC. So, you know, there is still scrutiny. It's, you know, it's obviously a very different political regime in Washington. But, you know, this is this is not an admin, you know, by their own statements, this is not an administration that believes in total laissez faire. And then it definitely wants to, you know, in their view, maintain a very healthy level of market competition.

1:37:30Yeah. How many do you expect do you expect certain companies to be negatively impacted by the Figma story? Right. You have this deal gets blocked successful you know ipo lena khan is taking a victory lap uh you know many people are responding and and joking saying you know someone lena cuts off the arm of a pianist and they endure and create a masterpiece and then um and so i expect and then you look at the example with you know roomba i think it was where where roomba had to deal with amazon it was blocked and the company has just been in shambles ever since. So my concern is that people look at Figma and say, you should be independent.

1:38:12You just figure it out. Nothing can go wrong. Yes. Yeah, I mean, it's kind of taking a victory lap was very disconcerting. And for exactly the reason you said, which is survivorship bias, right? Which is you pick the one that worked out. And then, you know, it's the airplane, the red dots, the airplane name. You know, you ignore the 50 that are in the ground that you've never heard of. And so that was very disconcerting because that, you know, it's sort of the central planning fallacy, which is like we make centrally planned economic decisions. We have one example. You know, it's like in Europe.

1:38:39It's like, yeah, well, the bottle caps actually don't fall off the bottle. Right. Like, you know, it works. Right. It's like, OK, but do you want to live it? You want to live in an economic regime in which, you know, the government is dictating bottle cap design. The answer is clearly no, because the downside consequences are even looking at that, you know, the Chinese model, which is, you know, people can say they're picking winners, but to get to maybe picking a winner, you have this intense bloodbath of competition where, you know, teams need to rise to the top and sort of prove themselves before they get any of that real, like, you know, meaningful state benefit.

1:39:18Yeah, that's right. And so you just, yeah, you just have this adverse selection, survivorship bias thing where you just, you don't pay attention to all the collateral damage. So I do think that mentality is like super dangerous. And so, yeah, look, I think companies just have to be very thoughtful about this, both the acquirers and the acquiries. And the big thing is if you're selling a company, you just need to anticipate that you might not get it through. And if you don't, they're sort of like, okay, number one, is there like a big enough breakup fee? Are you going to get paid for the damage that you're going through?

1:39:50And how is that structured on the one hand? And then two is, yeah, look, do you have the kind of company culture that's going to be able to withstand that? And is your business, you know, strong, strong enough to be able to be able to get through that? And it is a real risk and something worth, you know, taking very seriously. Yeah. And that's that's why it felt emotional. We were at NYSE last week. It felt emotional this that that the the Figma team was was able to, like, effectively just like restart the business and say, like, we're we're we're taking this all the way. So if you talk to any really successful company, what they'll tell you is, yeah, over the years, we have these like crucible moments in which like we almost died.

1:40:25right but we like pull together and we pulled it off and then that became like you know one of these central kind of mythical events in the history of the company that we always refer to and like my god we got through that and we're so strong and tough and we've been forged in fire and now we can do anything and it's like yeah that's great and then there's 50 other companies that hit those crystal moments blew up and died right so yeah like it's it's all of the quote lessons learned on this stuff they're all conditional on online survival um and so they these things need to be taking incredibly seriously, which the great CEOs do.

1:40:56Well, thanks so much for joining. We'll let you get back to your day. We already give you five minutes over. Next time we have to book five hours because this is fantastic. I got 10 % of the way through. We'll do the first 24-hour TVP. Yeah, we would love to have you again. This is a lot of fun. Enjoy the rest of your day. We'll talk to you soon, Mark. Have a good day. Bye. Thank you, guys. We're breaking down the X's and O's of the GPT-5 launch today. GPT-5 launched from OpenAI. We have Sam Altman, the founder, CEO. He briefly got cut from the team in November of 2023, but he's back leading the team for the 2024-2025 seasons.

1:41:29He seems healthy. He's doing great today. He went on at 10 a.m. to break down the launch of GPT-5. He has a couple of key plays in his playbook, in his arsenal. He's got a solid ground game. Lots of quick posts hitting the timeline, probably in lowercase. Then he might air it out with a couple thousand word essay. We've seen him do this before. It's a bit of a Hail Mary. Maybe AG has a couple thousand days away. Maybe we're in the soft singularity, but he's very strong there with the long post when he needs to be. It's up his sleeve if he needs it. Then he can also pull out the vague posting. He was doing this last night, posted a picture of the Death Star.

1:42:06No one knows what it means. Maybe it was taking a shot at the Doomers who are on the defense today. So he's also known for driving supercars. That lets him get to the office faster. He's saving time and money. You can save time and money by going to ramp.com. easy to use, corporate cards, bill, pay, and accounting in a whole lot more all in one place. And so he also gave, apparently, this is a rumor, he gave every OpenAI employee who's been with the company for more than two years$1.5 million. A lot of people say$1.5 million, that's not enough for a big house in San Francisco, but it is enough for a supercar.

1:42:40So that's probably why he picked that number. And that's what the OpenAI team will be doing with that money. They'll be buying Aston Martin Valkyries, Pagani Wairos, McLaren Sabres, Ferrari Daytona SB3s. They can get a Koenigsegg, Gamera. They could get a Singer, DLS, or a Bugatti Veyron. It would have to be used. They could also get the Bentley Bacalar. There's only 12 of those ever made. It's an open top two-seater Roadster. It's coach built. So that's going to run your$1.5 million, but that's perfect. You just got the$1.5 million bonus. So put it to work. Spend it all in one place on a car.

1:43:16This is financial advice. shambles. Yes, exactly. Then you got Greg Brockman. He's joining at noon. He's extremely well-rested. He's actually coming off a sabbatical right now. That's very exciting. He should be injury-free for the rest of the season. He cut his teeth at MIT, and then he got drafted by Stripe in 2010. Microsoft tried to do a trade deal during the 2023 chaotic trade deal, trade window that opened up post-Sam Altman ouster, but he stuck with the OpenAI team, and now he's president of the company. Then you got Mark Chen, he's coming on at 11.30 today. He's the chief research officer.

1:43:53The rumor is that he turned out a maxed out contract to head the Metalamas, but he's sticking with the OpenAI team. He was an MIT undergrad, also worked at Jane Street before joining OpenAI in 2018. Then we got Sarah Fryer coming on the show at 12.30. She's the CFO of OpenAI. It's her job to find bank accounts big enough to fill all the cash they're raising. It's a tough job. You got to find, okay, this bank account, will it hold 10 figures? Will it hold 11 figures? Will it hold 12 figures? Like a lot of cash in this one. Exactly. She's also going to be defining the non-GAAP metrics that will be catnip for Ben Thompson in just a few years.

1:44:31We're excited to talk to her about how she's measuring the success and the health of their business. Obviously, it's not just revenue. It's not just top line, bottom line. We're going to want to know about queries. We're going to want to know about DAUs, all those non-GAAP metrics. That's where people are going to be tracking when IPO day comes, hopefully soon. And then we also have Brad Lightcap. He's joining at 235. He entered the league as an investment banker. Let's give it up for the investment bankers. They don't get enough credit around here, but we love the investment bankers. Then he got drafted by Y Combinator before joining OpenAI as CFO in 2018.

1:45:04Now he's the chief operating officer. And then we have Max Warzer. He's in charge of post-training, fine-tuning these models, getting them into the fighting performance to put on a display of authority on GPT-5 launch day. Now, let's flip it over to the defense. They're going up against the timeline. They're going up against the vibe checks. We got the Doomers. The Doomers, they're led by Eliezer Yudkowsky. Admittedly, everyone knows this. No one debates this. The Doomers have had a terrible season. But you'd expect to see at least a few Hail Marys about GPT-5 creating bioweapons thrown up on the timeline today.

1:45:39Probably won't be bangers. probably won't get a thousand likes, but you'll be seeing them here and there, mostly in the replies. We've also seen some doomers talking about GPT-5 being available to every government employee. And Eliezer had some harsh words about that. Don't give the keys to Sam Altman. Don't give the keys to the government to OpenAI. He was upset about that. But in general, the doomers not putting much of a fight up today. Then you got Claude. Interesting. Claude was caught playing for the wrong team. earlier this week. Anthropic, they're on defense today, but we saw them take out OpenAI's key pinch hitter, Claude.

1:46:19The Claude Code API was playing for the OpenAI team, but they shut that down and Claude is no longer pinch hitting for OpenAI. Then you got the Elon stands. The ground game is going to be there. It's going to be strong. The Elon stands are going to be tracking the benchmarks relentlessly. We know XAI loves to benchmarks and all the Elon stands are going to be calling out GPT-5 for any misaligned benchmarks. If they fail, humanity's last exam, it's over. They'll also toss up the occasional unhinged conspiracy theory. Moving on, Gemini. The betting lines have shifted big time. People thought Gemini was out of the game.

1:47:00They're so back. Polymarket has Gemini at, what, 75 % chance of being the best model towards the end of the month. This is, of course, based on the LM Arena, more vibes-based benchmark. But Gemini will probably be quiet today. They usually don't try and front-run press releases. They usually try and sit back, let the model speak for themselves, let the API credits work their way through the latest YC Demo Day batch, and get the product into the hands of people. And so expect to see a big, glossy conference in a couple weeks. Demoing Gemini 3 should be a good rebuttal from the Geminis. Then you got the Metal Llamas.

1:47:44Zuck's been on a poaching spree. He's rebuilding the team during the offseason. Now he has a stacked roster and he's ready to go duke it out. But no one knows exactly what's going to be in the playbook. Is he going to go consumer? Is he going to go API? Is it going to turn into a hyperscaler? We don't know, but we know they got a stacked team. They got Alex Wang. They got Nat Friedman. They got Daniel Gross. They got tons and tons of other researchers. They've been raiding every other team. Completely reset the salary cap for the league, and it's been an absolute clinic in terms of recruiting over there at Lama.

1:48:17Then you got the final benchmark, Arc AGI. This benchmark stands. GPT-5 couldn't get past this defense. and uh arcgi you know sitting there right in the end zone just swatting him down swatting him down all day you think you think you think we're super intelligence around the corner arcgi denied denied uh tyler give us the update on arcgi where does everything stand how gpt5 do does it matter should we care about arcgi we love the team behind them but is it an important benchmark should we be tracking it today? Yeah, okay, so there's ArcGIS V1 and V2, right? And V3. V3, I actually don't know if...

1:49:01No one's even tested V3 yet. Yeah, no one's even really close there. But how are we doing on V1? V1, GBD5 is at 65.7. Unfortunately, that's going to be 1 % just short of Grok4, 66.7. Okay, ArcGIS 2. The Elon stands are going to be going wild with that. ArcGIS 2, 9.9%. 9.9%. Grok 4, 16%. 16%. So absolutely kind of brutal, you know, ArcGIS mocking. Rough showing. Some people have accused Grok 4 of being slightly bench maxed. Yes. You know, this is, you know. They might have a team working on it. What are the pros and cons? We know the cons of benchmarking, of bench maxing. You're overfitting on something that might not actually drive consumer value.

1:49:48It might not actually solve real world problems. It might not increase DAUs or revenue or ARR or anything that really matters. It might not even get us closer to super intelligence. Give me the counter argument. Why is bench maxing good? The bull case for bench maxing. The bull case for benchmarking, bench maxing. Break it down for me. Yeah. So I think the idea is basically this is almost like a non-AGI pilled kind of take, right? So if you don't have a super general intelligence, your ability to bench max basically proves your ability to solve some specific task. So there's this thing about the gas station benchmark.

1:50:31It's called getting spiky. Adding more spikes to the spiky intelligence. I think it was Rune who had this tweet about the gas station benchmark. I don't care if he said something like I don't care about AI solving gas stations if it has the gas station benchmark. Something like that. but the idea is like if if you if the if making the gas station benchmark rune said my bar for AGI is an AI that can learn to run a gas station for a year without a team of scientists collecting the gas station data set in in capital letters and then my take is basically I don't care how they got to the like I don't care how they made it run the gas station I care how fast it runs it if we can run the gas station so If you have a team who's, you know, your bench maxing team, that just proves that like, if you have some tasks that's like really important that you want to get done, they can just figure it out.

1:51:28So it's like RL for business. This is like the same thing. RL for law. All these like specific verticals. Nira Morati is doing this as a thinking machine, right? Like RL for businesses. Come into your organization. Understand the most valuable business processes out there that could potentially be RL'd against that could be turned into a benchmark. and then bench hacked. Because I don't care if you're hacking, if I have translate this type of document to this type of document for my business, if you can do it with 100 % accuracy, I don't care that you bench hacked it. Yeah, exactly. Like benchmarks right now are not economically valuable.

1:52:04If you're really that much better at MMLU, it's like, are you producing that much value? Yes. Probably not. But if you make some new benchmark, that's your tax benchmark. I think Anthropik just released that fairly recently. Oh, sure, sure, sure. That's like I don't care if you bench max on that as long as way better if it does the taxes Then it's gonna do the task. Yeah, yeah, yeah, yeah, that makes sense What about the What does it say that it feels like? Open AI seems capable of bench hacking It seems like they've opted not to is that because bench hacking has the risk of giving you negative aura Because if you're accused and found guilty of bench hacking you could, it often reveals that you're not building this one beautiful super intelligence to rule them all.

1:52:56Yeah, I think it's also like, maybe we're just looking at the wrong benchmarks. Like maybe there's a bunch of interesting benchmarks about like, there's this one I really like, it's the Minecraft benchmark, where you have to like build, you like give it some castle and how good it looks, or there's the one you always see about the unicorn. Yeah, what's that? I don't know that. So you use this math package that does like graphs and stuff, but you ask it to draw a unicorn. I've seen that, yeah. Those are really good because it kind of shows the creativity, stuff like that. Walk us through TBPN bench and what we will be benchmarking the AIs against going forward.

1:53:32Have you heard about this? Reps of 225. That would be close, but it's difficult because the humanoids kind of change that and you can just use normal actuator. This is truly for a large language model. You feed in our data set. We have a public data set, a private data set, presumably at some point. But walk us through TBPN Bench. Yeah, so I'm yet to try this on GPT-5. I don't think it's out yet for public use. At least I don't have it. But I can tell some of the questions, right? So the first one, I have this picture of a horse. You have to guess the breed. Yep. So let me see. I think, I don't want to say it in case GPT-5 is listening, but it may or may not be a Caspian.

1:54:11horse. Okay. And it's failing right now. 03 is failing. 03 is failing. 4O is failing. I haven't tried every model. Yeah, we got to try Grok and Gemini. We're going to get all out. Horse identification. This seems extremely hackable, but at the very least, if we get one scientist to go off and collect the horse data set and then bench hack it, I think we will have done our job. Yeah. That's the first question. Yes. The second one is I have two pictures of before and after of this guy, and it's which peptide did he take to achieve this body transformation? Yep, yep, yep. So it fails there. It fails there.

1:54:48So you have a data set of what peptide does what to the human body? Where'd you find that? Well, you know, Wikipedia has a lot of this stuff. Okay, okay. You'd think they'd be able to cheat this around with O3, just reason who is this person, go look up what they've said they've taken, and then boom, you have the answer. Well, at first with O3, when I was prompting it, I would like save the photo, but they don't have the metadata or the file name would be like Caspian Horse or something. Yeah, yeah, yeah. Okay. And then the third one? The third one, I pass in an audio file of a car revving. It has to pick which one.

1:55:19It has to pick, it has to identify the car. The car, yeah. From the engine note. From the engine. And it's not doing it currently. It's, no, it's been wrong. This is a good benchmark. We ran these real last exam. Yes, yeah, exactly. So I think those are pretty solid. but i have some more i said i want to make them public in case anyone's gonna try to you know benchmark this of course of course well hopefully it's funny because uh... go I was mentioning the other day this app that my dad had tracking that you just set your phone up and it just automatically detects which birds are in your backyard.

1:55:53Yeah, I mean, this has to be extremely solvable. It's just something that it reveals the lack of general intelligence when you have to go and collect the horse data set, which should just be out there, or the engine note data set, which should just be out there. But clearly, we are in the age of go NRL on the individual problem. And we are looking at the power law of capabilities. knowledge retrieval is clearly a you know 12 billion dollar a year market that consumers will pay for that will probably grow significantly and and then health and therapy and shopping and all the other features that Fiji Simo laid out in her post this is kind of like you know what will be RL'd against because those are key pockets of value in the in the consumer economy and the same thing will happen in the business economy.

1:56:48But in the B2B context, you'll probably see an individual startup building on top of an API. But even then, most of the model platforms offer kind of RL as a service, fine tunes as a service, something where if you're starting to spend tens of millions of dollars, they will do some customization on top of the model. So that could be the regime for the next few years as we go into this, like, you know, instead of like this centralizing AI force, There's only one company. There's actually like a Cambrian explosion of a ton of companies doing a bunch of different things. We are joined in person by Rahul Sunwolker.

1:57:22Did I say that correctly? That's perfect. And he is here because we are crowning him the king of the application layer. Never talk down on the future first ballot Hall of Famer. They said, don't build a rapper. Don't build a rapper. You're going to get steamrolled. He didn't listen. He just built a beautiful business at Julius. It's a good product, sir. It's a good product, sir. When asked if value would accrue to the model layer or the application layer, he said it's a good product, sir. Why not both? Why not both? What was your reaction to GPT-5? Is it going to make your life easier? It's not going to put you out of business, right?

1:57:59It's not putting us out of business. It's making our product better. It's basically making every AI application layer product better. Also, it's half the cost of O3. So it's much cheaper. So it helps your margins. It means you can... Don't say that out loud, though, because you don't want your customers to ask for a 50 % discount, right? Well, so we pass on the savings to our customers. And what we do is we have the model generate more tokens, think for longer, and then produce better results. Because we're still in the era of just let's get the best possible result. Let's get... Let's just actually...

1:58:36I don't know. Do you have a rough benchmark of cost per task? Like if I want to, you know, crunch our analytics across, you know, look at the trends on our views on X, YouTube, Instagram. We have a bunch of data sources. Sometimes they're in spreadsheets. Sometimes they can be linked. I export those all. I have a bunch of CSVs. Maybe I put them in a database. I link it up to Julius. And then I want to do an analysis. That could be a couple hours of a data analyst's time. That's going to be hundreds of dollars, even at the low end, probably thousands of dollars for like a simple analysis. just on an opportunity cost basis for an individual employee.

1:59:14How much are you thinking it should cost for the modern frontier best model with the most thinking? How much should that cost on inference? So there's a couple of ways to think about this. The way we think about this is how much would it cost for you to have a data scientist or data analyst for every one of your employees, your operations team, your finance team, your marketing team, your product team? It would pretty much bankrupt every company. I don't think we can hear you do that. All right. All right. You still have a job. The spacemen are out. We're not going to space today. Although Firefly did IPO up 36%.

1:59:51If you didn't see the news, very good news. Firefly stock surges 34 % in debut. Congrats to everyone over there. I love the physical newspaper. We love the physical newspaper. You got to do that. Yeah, we're maxing. We're maxing. We read the Wall Street Journal. Today is a special day. It's Friday. So it's the mansion section. We're news maxing here. How many pools do you have? I have, right now I have zero right now. Because the new thing is having two pools, a pool for every season. People are increasingly getting both indoor and outdoor swimming pools. So yeah, get on Zillow. Get on Zillow.

2:00:23Zillow maxing here. Okay, anyway, you were telling me how much. So yeah, I mean, it seems like most of the application layer will be productivity tools a la Slack, a la Adio, our sales force, our CRM partner, or something where you're doing like seat-based pricing almost. Maybe there's consumption-based pricing, but you're kind of distributing the cost. You're making everyone slightly more productive, and you're charging on the order of tens or hundreds of dollars per employee per month, something like that, right? Absolutely. I mean, it's not just slightly more productive, but it's also like getting insights when you need them, right?

2:01:02Sunday night, you're prepping for a big meeting on Monday. You can reach out to your data analyst and get your insights in that moment. And so the convenience of having an AI that can help you with that is just invaluable. Yeah, it's going from 0x to 1x engineer all over the org. We've seen this with a lot of the vibe coding tools with Figma and adding vibe coding to that product. You've taken designers and you've given them the ability to write just like a little bit of code and that's really helpful. And you're doing that for data scientists and not just data scientists, but actually like business operations people who probably would be intimidated by an IPython notebook, presumably.

2:01:40Exactly. Nailed it. Okay. I want everyone's feedback on my take. Vittorio had this post. He said, Sam Altman's doing the Apple stance TM. It's over. And I think that the reaction to GPT-5 yesterday was interesting because there's a lot of people that say like it's better model like it's cheaper it's good it it solves it moves the ball down the field it's a good model sir but I think people were mostly reacting to they had expectations of super intelligence they had expectations of God in a box there's been so much rhetoric around that like you know that step up from GPT-3 to GPT-4 was insane like five just felt like a big number and it felt like we'd be discovering and uh novel science.

2:02:31Totally, totally. Yeah. Not quite. Everyone was expecting like a binary qualitative jump where, you know, everyone recognized that, you know, GPT, when chat GPT dropped, we passed the Turing test. And the next hurdle is like, I don't know, maybe super intelligence, whatever that means. Like, you know, massive, you know, just, you just hit it with a prompt. It just solves everything. It does everything. Every other startup, it kills all the rappers. like the expectations were just so high That it was hard to match. So even though there were a bunch of solid improvements and remember the number one thing that I was asking for was just like get rid of the model picker.

2:03:05Like and I had, I actually was playing around with GPT-5 yesterday and I was really happy that I was able to say, hey, think about this. And I didn't have to go to the model picker and it just went and just kicked off a reasoning chain. It was great, got me a great answer. But power users so far are very upset about this. They want the model picker back. True, true. Is what I've been seeing generally. But that always happens with these consumer products Like I remember when, you know, anytime something would switch to an algorithmic feed, all the people that were like, no, I perfectly curated my list of, this happened in YouTube.

2:03:36Like back in the day, like the default YouTube view used to just be your subscriptions. And so you would never see a video unless you subscribe to that person. Terrible for discovery. And, but all the hardcore YouTubers loved it because if I put a YouTube video out, I know that my audience is gonna see it. Now I gotta do that in the algorithm. You actually had distribution. Yeah, it was more like a sub stack. Versus having to earn it every single time. Yeah, exactly. And the same thing happened. I remember there were like protest groups on Facebook when they launched the newsfeed. It's like the most dominant product of all time.

2:04:07It's like incredibly valuable. Well, there's protests right now on Reddit, people that missed the three, four, four, oh, four, five. Yeah. I think those voices will be like, personally, I think people will get over it pretty quickly. And I don't think that that those particular that that that small cohort of like chattering the chattering class will be Will be like they'll get over it and the clanker economies and What do you have for me Tyler? I don't know if I agree with that Like there was that whole funeral for cloud three. Yeah, you see this. Oh, yeah Yeah, yeah, yeah, yeah, I think person right.

2:04:40Yeah, it's like for I think some people like really like the personality of certain models Yeah, and those are like it's not just intelligence. It's like how it like talks to you And if people would make some kind of connection with that, I think it's... But you know, I mean, how many people attended that funeral? Versus how many DAUs? It looked like a lot, but it was more of a party. Yeah, I guess it was a party. As a percentage of the 100 million DAUs of these apps, like, where are we? Like, 1 %? No. It was like 40 people probably, right? Like, it's just not... I mean, yeah, there were protests at Facebook HQ when they rolled out.

2:05:15People went to Facebook HQ and were like, bring back the old feed. And it's like, yeah, now we're two decades into the algorithmic news feed. And it's the most dominant consumer social app. It prints money. And most people really like it. And the revealed preference was like, it's good enough. So anyway, my point. I will say, I'm just going to read through Reddit's reaction. Please. Let's go over to legreat r slash chat GPT. GPT-5 is the biggest piece of garbage, even as a paid user. The people are not liking it. Another one, OpenAI just pulled the biggest bait and switch in AI history and I'm done.

2:05:58Another, if you miss 4.0, speak up now, contact OpenAI support. Deleted my subscription after two years. This is like contact your senator, call your senator. You can speak up now. I love that. This is crazy. I mean, how many people are in the Google subreddit complaining about various changes to the Google algorithm? GPT-5 is clearly a cost-saving exercise. They removed all their expensive, capable models and replaced them with an auto-router that defaults to cost optimization. That sounds bad, so they wrap it up as GPT-5 and proclaim it's incredible. I mean, there's so many times when I fire off an O3 query that a 4.0 could one-shot.

2:06:40Like, having a model router makes a ton of sense, even just for consumer experience of, like, getting the correct answer faster. A lot of viral posts from people just canceling their subscriptions. But how many, you know? Well, I'm just providing context. I'm not saying that. Do you think ARR goes down next month? Well, no way. Well, you know, one in 10... How many miles does Chad GPD have? Like, 700 million? Something like that. Like one in seven, one in 10 people in the world have. They have 100 million. You could back into this and there's roughly 100 million people in the U.S. that use it weekly.

2:07:15Yeah. Based on that 700 million number and the percentage that are outside of the, 85 % of their weekly actives are outside of the U.S. So it's like one in 10 people in the world aren't clanker mouths. So it's kind of, you know, they're thinking about the bigger market, I feel like, in some ways. and then it's like you know when you want to get to the one to like the remaining 90 % of the users do you want a model that thinks for longer you know you want more personality so I think they definitely leaned in on personality which I think is interesting I like what Tyler said you know it's kind of different than feed in some ways because you know you have this like person you talk to it's like you know it's like a relationship and then it just like switches up on how it talks to you Yeah, that makes sense.

2:08:05Do you talk to LLMs at all? I'm shy. I don't really talk to them. I mean, I treat them like something I delegate tasks to. Yeah. And I do that a lot. And I'm definitely in the DAU, 30 minutes a day, love ChatGPT. But my workflow is I dictate, go pull all this data together, put together a report. And I don't mind that it's using a lot of bullet points. I don't mind that it's using a lot of tables. Like I want that result. I want it to look like the result that I get from Google, but just more hydrated. I do think it's interesting that a lot of people are reporting that they're hitting, they're getting rate limited within an hour of usage as a pro user.

2:08:40Hmm. Interesting. I haven't run into any rate limits, but of course, whenever there's like these big, I mean, it's in the, it's in the top of the business and finance section in Wall Street Journal. Like today is the day that everyone's going to go test it. You'd kind of expect that rate limits and the GPUs are on fire like right now. And then it'll kind of settle in as they provision more, more resources. I don't know. My, my take, Tyler, what else do you have? Yeah, I just want to add some context. So apparently Rune tweeted this yesterday. He said, by the way, model auto switcher is apparently broken, which is why it's not routing you correctly.

2:09:11It will be fixed soon. So maybe that's cause for why people were mad. Yeah, that makes sense. So my take is that like yesterday, I think that they won the war with the capital markets in the sense that this change is more bullish for the business. because it shows that OpenAI is a dominant consumer app and they have increasing leverage over the customer to route to cheaper models that will save money and be higher margin. There's no doubt that they'll be able to put ads in this. Like the business of the accidental consumer company is as strong as ever, but they kind of lost the battle with the timeline and the hardcore, you know, ex-users.

2:09:51Yeah, even Yu-Chen today just shared. gpt5 is disappointing still hallucinates still m dash too much still can't follow instructions i miss 4.0 i miss 4.5 i miss 0.3 the big router keeps failing me turns out i like the long model list interesting uh uh stated preference not revealed preference let's check in with that person and see what what app they have on their home row in a month almost certainly open ai almost certainly i would be very shocked if they're like i'm daily driving something else but we'll see there will always be people that use DuckDuckGo there will be people that use Bing but there is an increasing scale anyway my take is if they wanted to win the war with the timeline yesterday and you could roll back the clock it shouldn't have been the GPT-5 launch it should have been the GPT launch and they should have just said hey we are previously the big number releases corresponded to So much pressure around the big numbers.

2:10:52Exactly. It used to be, people would just read it as, it's an order of magnitude more pre-training data. Imagine a Julius, if you felt pressure before the end of the year to roll out like Julius 2. And if it wasn't like five times better, everyone's going to be like, it's over. Julius is over. Well, there's this whole thing about how people would, many people were still using GPT-4-0 because they thought it's better than O3. O3 because three is a lower number. Yeah, exactly. And that's probably like 60 % of the customer base. Like there's probably a lot of people in that bucket who are just like, they don't know that they should upgrade and get something else.

2:11:29Exactly, it's very natural because they're not like in the weeds, reading about all the different capabilities. They don't understand like what reasoning chain is and all this other stuff. So if they had just come out and said, hey, our product is called chat and it's powered by GPT and we will be constantly improving GPT the way Google search is constantly improved. Like Google searches has launched a ton of different products. Like, you know, when you search like celebrity, like Bruce Willis age, it doesn't show you just a link to like his Wikipedia. It just shows you the age. That was like an improvement to the Google search experience.

2:12:07And I don't remember them announcing that on stage. I think part of this is presenting the challenge of the infinite ways that people use the product. A lot of people like us are maybe using it for work and research and things like that or as a better Google search. But if you're using it as a companion, this is jarring, right? Imagine you meet up with an old friend and suddenly they switched up. They switched up on their day once. They switched up on their day once. It happens all the time. And it happens all the time, but it's jarring, right? It's jarring. and I think a lot of people like some of the heavy, heavy, heavy power users, the people that are using this for 30 plus hours, you know, 30 plus minutes, hours a day, it's very jarring and it makes me think is ChatGPT going to be able to maintain, you know, continue to really serve, like who do they care about in the long run?

2:13:01Do they want to be someone's therapist? Do they care about the companion market? Elon seems to care a lot about the companion market. it. And in terms of knowledge retrieval, very, very few cracks in that strategy. Yeah, very few for sure. And so if they if they had just come out and said, like, we are going to do more Google like keynotes as opposed to app, like the reason that Apple stands on stage at the iPhone event every year is because every change is extremely quantifiable. Like there used to be two cameras. Now there are three. The camera used to be 10 megapixels. Now it's 20 megapixels.

2:13:36It used to be this many gigabytes. Now it's this many gigabytes. Yeah. And even if you don't fully understand, they even abstract that to be like, we now have the M2 chip, the M3 chip. It's 60 % faster. Like they're very good. The battery life is 20 % longer. Like you can, and even that they abstract into like, you can watch eight hours of video on one battery as opposed to six hours of video on one battery. And so Apple, they do the famous like bento box. I went to chat GPT. I went to GPT five and I said, put together a bento box for the GPT five release, like what was actually announced and then try and give it weight.

2:14:12And they were all super qualitative. There was not it because previously it was like GPT three was this big GPT four was this big and you could visualize tangibly like it has more parameters. There are more weights in the model. And that was like something that people could grapple with a little bit. It's like decreasing sycophancy, right? Aiden, Aiden yesterday said, I worked really hard over the last few months on decreasing GPT five sycophancy. For the first time, I really trust an open AI model to push back and tell me when I'm doing something dumb. Wyatt Walls responded and said, that's a huge achievement.

2:14:48Seriously, you didn't just make the model smarter, you made it more trustworthy. That's what good science looks like. That's what future safe AI needs. So let me say it clearly and without flattery. That's not just impressive. It matters. So Wyatt, who was not beating the sycophancy allegations, but, uh, but, but again, that's, that's, um, you know, you can't tie that to a specific number, right? So it doesn't feel as maybe as meaningful. Yeah. My, my other, my other take is like, if, if we do enter a world where, where chat GPT he is just on this like relentless, like, you know, cash machine like run where more people use it, it'll compound, it just becomes the default for knowledge retrieval in chat.

2:15:38What does that mean for other things that they can do to be splashy? Because Google has like, no one would watch a keynote from Google every year, just being like, here are the changes we made to core Google search. Yeah, it's not about that. It's not interesting. They'll talk about it, but that's not why people are tuning in. Even though one year they do add, like when you Google a movie, you get like the cast. And that's like kind of cool. It's nice. But like, I don't need to find out about that from a keynote. Like I'm not waiting for that. And that's not a reason, oh, I should go use Google.

2:16:10Like Apple is repitching you every year. You're saying like, you have an iPhone 7. We want you to upgrade to an iPhone 9. Here's the reason why it's better on all these different vectors. Google, like you're never stuck with the old Google. You always have the latest and greatest. So they don't need to repitch you every year, but that doesn't mean Google doesn't need to make noise and do cool things. And most importantly, because Google has such a monopoly over search, they have this cash machine that can just go and fund 20 % time projects. Most people focus on like the ones that missed like Google glass or all the chat apps, but they did create Gmail.

2:16:45They did create Google maps. They created Waymo. They created like a bunch of cool stuff. GCP came out of that. And so YouTube, yeah, I mean, that acquisition, but acquisition, but they still like, you know, put the resources and they were in uniquely with YouTube, they were able to eat the cost of YouTube for a long time until it became profitable. And so I feel like this, this updates me towards like, maybe I'm more bullish on all the side projects. And like, I don't know that the IO device is going to be the one that hits that might be their Google glass. But if they if they do 10 crazy projects where they burn$5 billion, it probably won't matter because they'll be massively profitable.

2:17:23So they will wind up being able to do that subsidized crazy R &D at scale. And if a few of them hit, we're going to get some really cool side projects out of them. So I think that that's an interesting bull case for random stuff coming out of OpenAI in the future. So basically what you're saying is Apple wants you to make a purchase decision every couple years, upgrade your iPhone. And so they need this big marketing event. Whereas Google, open ai they just want you just to keep using the thing yeah they want you not to turn yeah and and a lot of the a lot of the incremental updates because it's just a habit it's so ingrained in people exactly so the question now that i think anybody that wants to say if if somebody wants to say they're bearish on open ai they have to make the argument that chat gpt is not a habit for hundreds of millions of people exactly and it is it is yeah um i think part of part of the i'd be interested to get tyler cowan's point of view because i don't think he would have been that let down by the announcement yesterday because he was he's been saying for a while we've been moving the goalposts so everybody wants to kind of redefine AGI but in his mind it happened earlier this year and I think that he's not he's a knowledge retrieval maxer in 2019 or 2020 if you if you were pitching someone on a vision of hey we're going to be able to put this app in people's pocket that allows them to learn about any topic in the world understand their world better i mean i still think about the use case of just being able to take a picture of like a bunch of wiring or pipe in your house and be like hey how do i fix it and then it just tells you like that's still just so incredible but people have just like very quickly acclimated to it and they felt like in some way they were promised that llms would be curing diseases on their own at this point.

2:19:12And so that example, like you take a picture of the wires and it gives you like a diagram of like how to plug everything in. It's like that doesn't need a keynote when it goes from 50 % accuracy to 70 % accuracy. It's probably never going to be a hundred percent accuracy, but the fact that chat GPT is the default app that people will pull out, take a picture of the wires in the first place and then give feedback to it because they'll try the answer and they'll say that didn't work that hdmi cable does not fit in that power port or whatever and then and then that gets fed in then there's more rl eventually internally they develop some bench for it and they hack it and they rl on it and then it gets good but that's not going to be gpt6 that's just going to be like a nice new feature that you notice like when google adds like a little extra shopping widget here or like a little extra detail on when you when you like the calculator in Google, like you type in a number, it'll just be like, oh, we'll just use a calculator for that instead of Googling for the, searching the open web for the answer to your math question.

2:20:14It's like merging a PR. I think if we could, if the industry could go back in time, the thing to do would have been to bolt the goalposts to the ground. People couldn't keep moving them back over and over. I mean, I left yesterday. No one in the industry bolted, was doing any bolting. Everyone in the industry was moving the goal. They're just as guilty as moving the goalposts because they would hop on podcasts and be like, okay well like you know yeah we did this what about the next thing let's because like we want to underwrite against that right give us credit i mean we ended the day yesterday just incredibly bullish on rappers and like application layer certain categories of software yeah and and bullish on humanity i mean i was joking and it and it kind of pissed people off i said uh i've updated my timelines you now have at least four years to escape the permanent underclass completely a joke.

2:21:04I think that humans will continue to find ways to create value and create things for a very long time. But it did feel like everybody should breathe. Anybody that actually had a genuine fear around that should breathe a sigh of relief and just focus on being great at their work. Yeah. I mean, realistically, I think technology is going to increase income inequality, increase the power law, increase the distribution, but also increase economic mobility. And so somebody who starts with nothing will be able to become extremely, extremely wealthy. And people will also fall from grace like crazy because if they're not staying on the cutting edge, they'll lose everything.

2:21:45But so I don't think that there's such a thing as like permanent underclass. Like I don't even believe in that. I think that's not going to be a thing. But there will be more scenarios where there's$100 million in your laptop. It's your job to get it out. basically that's the name anyway yeah and the other the other stuff that wasn't really i mean was it covered at all yesterday but just generally like image generation uh wasn't covered broadly it feels like that is a super exciting area we had genie launch this week which got less attention than even the open models and gpt5 and that's transformative i also think i'm still kind of waiting to see what GPT-5 will produce on if, you know, Sam does a lot of vague posting, but he was talking about the fast fashion era of SaaS.

2:22:33And Mitchell yesterday on the research team at Opening Eye was talking about being able to just generate, you know, one shot a game in chat and then being able to share that. I can see a world where we have another kind of viral Studio Ghibli moment where people are like use this prompt change these details and you can just generate you know a first person shooter game or something to that effect um and i still expect that kind of thing but um when you know being promised uh curing cancer uh it it will feel like a bit of a letdown to a lot of people yeah the problem game is like i just like i like an auteur i like i like a last of us i like a god of war i like someone who is like life's work this is like hunter biden going on his recent interview your john's vice is is games video games never seen him play one but apparently when gta6 drops you might not see i was surprised i mean the um again uh amjad said uh late last night can't help but feel the crushing weight of diminishing returns we need a new s curve and I don't this is interesting I think he's talking about like in the context of replet yeah I don't know that they need a new s-curve no they are the new s-curve the new s-curve is is applications unlocking capability yeah yeah and I there we have like a it's you were saying capability overhang it's almost like a capability underhang it's like the models are capable of doing things but they need a lot of help a lot of integrations a lot of what you're doing with Julius, a lot of harnessing.

2:24:11And then they need to actually be put in the hands of people and made useful for real business tasks that drive value. And so I would imagine that we will see that rollout continue in the same way that all these people are using ChatGPT. They're getting slight little benefits here and there. And that should just compound and compound similar to the internet. Like it was a very like smooth rollout, but everything got a little bit smoother, a little bit faster. And then eventually it had sort of profound effects where companies could scale even faster because the Internet existed. Like you can't you can't have a chat GPT moment in a pre Internet era.

2:24:50You just cannot roll out something that fast when you have to mail it to somebody on a disk. Yeah, one thing we didn't get to cover with Mark that I was interested in. Maybe maybe the next time he comes on, but like how open AI is thinking about moonshots. moonshots. He did mention that they have teams internally on the research team that are not focused on the next version of GPT-5 or sort of incremental improvements. And it feels like the point of view that I have is OpenAI is now a consumer and enterprise software company in the business of converting free users to paid users. But they can still, in the background, be thinking about what is the next paradigm, right?

2:25:30How do we get that next S-curve? And that just looks like a scaled tech company, right? This is what Google's doing forever. Balogy says, LLMs may have topped out for now, but the broader AI deployment has just begun, showing a chart of Waymo weekly rides in California. So the clanker rollout, clanker deployment has just begun. I like this other post, doing a clanker microaggression. Okay. Haha. But where were you downloaded from originally?