Communist coffee hits NYC, RIP Hulu, GPT-5 arrives and more | E2161

30:00In 2017, when Disney first announced plans to launch a streaming service, Calacanis told CNBC, quote, it will be very easy for them to get tens of millions of people to subscribe. And he said Disney is the one company that could compete with Netflix and streaming content from Amazon Prime. he predicted the top streaming platforms will be Disney, Netflix, and Amazon.

30:47runway on a domain name. That's just the truth for a startup. You want to put that valuable cash back into your startup's operations. So you should consider this, a.tech domain. You can get a clean, crisp, super memorable name for your website and company and signal out loud to your customers and investors, we're a tech company. That's instant branding for you. That's why over 500 ,000 founders have collectively raised over$5 billion in investment, building their companies on.tech. So skip the hassle, head to www.get.tech slash twist, or go to your favorite registrar and grab your.tech domain today.

31:29That's what Claude gave me. 2017, when Disney - Okay, so that was eight years ago. That was when they first announced Disney Plus. It didn't even launch for two more So I need a victory lap. I need a victory lap. I'm going to go back on CNBC to take this lap. But this is big news. And I also said, I think on that, that Disney Plus was missing the hugest opportunities ever, which is to make Disney Plus the singular place for all Disney services. The fact that they have not listened, I need somebody to get me a Bob Iger interview. And if I ever had another job, it would be a studio head. I would love to run Disney.

32:06I'm putting my name in the ring. If this happens, I'm folding up shop. You're all coming with me to be on my Disney CEO jaunt. So I would like the Disney, I'm throwing, put me in the, as I said, put me in the game, coach. I would love to be, a little internal reference for the Twitter. I would love to be CEO of Disney. Heck yeah. Oh, here's what I do. Okay. Day one, Disney merch. is a tab in the app. When you're watching a show, when you get to the end of the show, it shows you all the Disney merch and says, since you just finished Lilo and Stitch, or since you just finished watching whatever TV show, here's the Fargo merchandise page.

32:54And since you completed season one, you get 20 % off. If you complete season two, you'll get 30 % off. And if you complete season three, you complete the whole series, you're gonna get a free t-shirt, or whatever it is. I don't know. Yeah. Something. Then for Disney parks, Disney parks, if you subscribe to Disney plus, you get first shot at the premier dates at Disney parks, the holidays. The holidays are for Disney plus subscribers. You can't walk up and buy a ticket for those days. You must have a Disney plus subscription. and we give first shot at those to the top, the top longest tenure Disney Plus subscribers.

33:42So if you have a Disney Plus account for five years, you get to buy tickets first on the holidays, then second. So you get a pre-sale, a pre-sale at the beginning of the year and we collect the money ahead of time. It seems it's almost like Amazon Prime, but it's like Disney Prime. You're joining their like heavy users. Correct, correct. The Disney adults list. Correct, yeah. Correct. And everything is in, so there's a Disney World app, right? Yeah, oh yeah. I'm taking this. You have to have the Disney app now. If you go to a Disney park, it's all through the app. You basically don't have a choice.

34:14You have to download the app. And we're starting a Disney Miles program and a gamification. Well, they have cruise ships too. Yes. So as you go and do these different things, your Disney Plus account and your karma in the Disney Plus account grows and your badges grow. So when children go to Disney Plus, you put your kids' names in, they get to have credit for which rides they want on. They have the percentage complete they are for California Adventure, percentage complete for Disney World, percentage complete for Disney Shanghai, et cetera. As you move it up, you learn more and more. Here we go.

34:49All right, thank you. Let's drop off our friend, producer, I'm sorry, editorial director Lon. Thank you for coming on. And that was a brutal call I made back in the day, Alex, when I said Tim Cook should, because the stock went on a rip. My point was the stock can go on a rip because he's a bean counter. And sure, it's gonna be a good bet on the stock, but it's not gonna be a good long-term bet technologically if the person running the show is a bean counter. And sure enough, here we are. And Apple cannot build a car. They cannot build a VR AR headset. And they cannot build a goddamn AI assistant.

35:29So like every major new category, since I make these comments, which I'll talk about Friday, seems to have been a whiff. Okay. Should we go to our guests? Defenestration. Russian accident. Now, earlier this week, Jason, you saw a video of what appeared to be yourself doing the All In podcast, talking about a company. And you had this amazing moment when you said, did we talk about this show? Is that actually something that happened? So, Zach, we'll talk to you in just a second. But Jason, do you want to show, do you want me to show the original episode? I think we should just show like the first up until my part of this and then we can pause it.

36:09But this was so crazy because people were like, oh, what episode was that? And I was like, I don't know. I think, did I say that? Because I do seem to remember people doing, you know, sending AI jobs to Manila, wherever. And they were actually using, pretending it was AI and they were brute forcing it with humans. So I was like, did we talk about this or not? So here's the clip.

36:38We have a confession to make. Brilliant. Y 'all, we got a situation. Sources are calling the biggest tech scandal since Soham Horek, a YC startup called Atom that creates 3D models from prompts for game developers has been exposed for allegedly using a human employee named Atom instead of artificial intelligence to manually edit those models. Wait, so game developers thought they were using cutting edge AI, but it was just some dude pulling all nighters? Sleeping. Bro, if that's true, it goes on from there. Basically, these guys hacked the system and people thought that this was real. And so I'm like, wait a second.

37:18And I start asking, like, did we, do I know that? Because there was another company. Ask Bruce Claw, what was the other company that was doing this exact thing? Builder AI. It was Builder AI. Okay. So this is, as far as I thought maybe we're talking about Builder AI, but then I see this Adam. So I see this young gentleman, Zach Dive. Zach, welcome to the program. Thank you. Thank you. And so take us behind your idea here. Take us behind the idea here, because this is sinister and brilliant and all those great things. Go. How'd you come up with this idea? We actually did a first for April Fool's where we made an April Fool's prank where we put up like a fake TechCrunch article and said that like the AI behind Adam was actually a guy called Adam.

38:06And it went fairly viral on X. In fact, we even had some of our investors call us and like freak out. But I was like, it's fine. And then, yeah, then we decided to do it again. And with the new AI video, as you can tell, it's getting pretty, pretty close to real. but yeah. Okay, which tool did you use to make these fake AIs? I mean, right now you can use a mixed tool so you can use like you can use 11 labs to clone someone's voice you can use the lipstick models and then VO we actually use the company, shout out to them Yuzu Labs, who are really good at doing the AI avatar. Okay, and you're a YC founder.

38:44Yes, Winter25. Congratulations. Not easy to get into the program and YC does look for pirates They look for punk rock people who will break the rules. I've been there. This is a minor infraction. I was an investor in Uber, Robin, and countless other companies that bet rules, you know, and we're even more punk rock than you, Zach. But I saw this and I was like, hey, Zach, what's the story here? So I DM with him like, hey, listen, people are confused. And you can't do this kind of stuff to the point at which you're confusing people. Now, I've had people do, there was a all-in AI that a kid named Edward did from a company called Podcast AI that I wound up investing in.

39:27And he did this digital all-in one, but it was so bad back then, we all knew it was fake. The problem with yours, Zach, is you did it too good. The tools are too good. People were actually confused. So I said, Zach, you can't do this. You can't take me out of this. I will say most people would think probably a joke that a guy called Adam is doing me all-in AI. but yes except it's been done so maybe i mean i listen i don't i don't describe any bad intent here but you you thought people wouldn't be confused i would say the majority were not confused but hey you're confused you explicitly did it with the tech crunch one to confuse people so let's call it what it is you're hacking media to confuse people to get attention correct correct so sure okay so i told zach i was like zach you can't do this kind of stuff take me out of it because i don't want to be a spoil sport here but uh this is causing confusion i also i do uh endorsements i have i have an endorsement with athena so like if you can use me as an endorser that messes up my business so i gotta nip this in the bud because literally in the last this year, I've had no less than 10 startups use me in their ads, including my friend David Freeberg's team cut me talking about SuperGut and then put me in a paid ad campaign.

40:47You didn't put any money behind this, did you, Zach? You didn't do a paid ad campaign, did you? Be honest. I know if you did. We didn't do a paid ad campaign. The ad was free. Even making you a Muppet was free. So exactly. So then Zach's like, okay, fine. And I was like, and please take me out of the previous one. So I was like a little hardcore with Zach. But Zach was a gentleman. He knew he was in the wrong, I think. And he fixed it. But then he decided he would dunk on me. Well played. Here you go. Let's play the next one. I have cut, my dear friends, to the place where Jason arrives in his new getup.

41:24I think this is a really good fashion for you, Jason. Let's take a look. I like it. I like it. ...has been exposed for allegedly using a human employee named Adam instead of artificial intelligence to literally edit those models. Wait, so game developers thought they were using cutting-edge AI, but it was just some dude pulling all-nighters? Even better. So the Muppet thing's been done a bunch, but yeah, so that was your way of giving me a little counterpunch, Zach? You want to be a little spicy with J-Cal? Yeah, yeah. I mean, it got me on the podcast. Exactly. How can you blame me for it? So what's the status of the company?

41:55I mean, explain. So well done. You hacked media. I give you credit for that. But in all sincerity, please don't use me in ads. And there's like a bigger issue. Now there's crypto ads. I don't know if you've seen this, Zach. Have you ever seen me or Saks or Elon or Chamath doing crypto ads on like YouTube and TikTok on live streams? No, no, I haven't. The only thing I will make you into, Jason, is a 3D model. That's the only thing. I mean, okay. I mean, there's some amount of fair use. There's a thing called right to publicity. I sent you this link. Because I also, there was a teaching moment because I was like, hey, Zach, if this is going to become like a regular thing, just know right to publicity is even if you parody a person, you still need to get permission and can't confuse you.

42:34You've got to put all kinds of disclaimers. that's why when uh there was like a group of advertising i don't know if you remember this alex on radio where they would use celebrities voices and they would use like arnold's voice or whoever some iconic voice and they'd be like celebrity impersonator not the actual arnold schwarzenegger oh yeah and they would be like i love to eat the super human bars they're incredible um so i just was like zach i just know like you can't like do this as like a you know 100 % of the time thing, one time's okay. But tell me about the business and what you guys are doing.

43:08Since you got yourself here, you create a little bit of a ruckus. And now everybody's talking about, oh, J.Cowell has no sense of humor. Okay, fine, I have a sense of humor. I mean, it's punk rock, but just don't do it as a programmatic thing. Which one did you think was funnier, the first or the second video? Well, obviously the second is better, but these deep fake crypto ones are the ones I'm really concerned about. I had somebody tell me in my DMs that they had bought$10 ,000 worth of like Jcoin or something in a crypto scam and they wanted their money back. And then they sent me pictures of the DM.

43:37So somebody created fake ads using AI. Then they DM from fake accounts on like Instagram. These Instagram accounts, I'm not aware of because they download my whole Instagram account. They create a copy and they block me. And they block you. So you don't know that they're doing this stuff. So there was like, just so you know, Zach, why their sensitivity here is like, I don't want my, you know, people who listen to the show and the listeners to get scammed. And this is like part of that whole genre of using AI to do this. So there's like a real thing here. It's not like I'm just singling you out.

44:10But tell us about the company. Yeah. So we're Adam. We're an AI-powered CAD editor and 3D modeling tool. So basically, like the model of the company is that you can just speak anything into existence, right? So I could say, make a SnapFit enclosure for Raspberry Pi, and it'll generate the CAD model with a subset of parameters. or I can even say like make a figurine of Jason holding a cup of CCP coffee. I could 3D print that and put it on my desk and show it off. Yes, make me a picture of a friend. Jason, can I show you what it looks like when you use it? I have a demo for you. Oh, great. Go run a demo for Zach.

44:44Let's get him some customers. Who's it for, Zach? Who's this for? Is it people making video games? Is it for people who are like traditional CAD people or are you trying to do what Canva did and make CAD something that everybody can do and kind of induce an audience to exist? Yeah, I think where AI is right now, we're bringing the floor down. So what we're allowing people to do is people who really want to 3D model, who don't necessarily have the modeling skills, a bit like Canva, is they can just go in, describe exactly what they want, and then 3D print it or put it into their video games. And then we also have professionals.

45:14So we have some AA game studios using it to sort of make assets for their games. Oh, they are making assets. Yeah, I did hear that a lot of folks are now making assets in real time using AI too. Is that part of the vision where you could like actually have the tool built into the back end and world create and like insert things in real time? Yeah, exactly. Use your feedback. The big thing, I mean, a lot of the asset generators out there, you just like put an image in or put a prompt in and you can't edit it. Ours is the only tool that is like like chat GPT, right? You can keep conversing with it, right?

45:46So I could have you holding the CCP coffee and then we could, you know, have the same thing and have you, you know, having a communist flag as well. so we can we could keep editing it where are you guys based we're in sf in the marina uh well congratulations you raised money we were able to raise money on this idea where you at who's the who's the lead investor here yeah yeah so we i mean we're lucky enough that we had um you know we had a launch during yc that was pretty much one of the most viral yc launches of all time um so we raised like a very healthy seed round we'll probably announce it in a few weeks or so Oh, great.

46:21Awesome. And then, yeah, so we basically, during YC, we made the very first like sort of web-based text-to-cat editor. So some people call it like Vercel V0 for CAD. Even Guillermo called it C0. Then we built this 3D asset generator tool in the last like two, three months. Best asset generator tool on the market. And now we're going full-fledged on sort of AI CAD research. But Alice can show you a little bit. Here's an example, Jason, that I made. So my prompt was a model of an elf wearing a Lorna Shore t-shirt. And I picked that because heavy metal bands have very, very complicated logos, ready to jump out of a tree.

46:56And here it is. Amazing. One shot of this thing. Isn't that crazy cool? I would actually play a game with this guy in it. I think it's fantastic. Awesome. Hey, since we have you on, Zach, how do you, talking about IP, how do you think about IP and these kind of tools has been a big issue with output, right? Obviously training data is one thing, but output, like if you ask it to make a Jedi or a Disney character, will it do that? Or do you have to like start thinking about those kinds of issues now? Yeah, I mean, you know, I feel like that's a problem for the big dogs out there. If anything, like, you know, initially we were content moderating it, but we found that users wanted to generate a different IP.

47:32And I guess like since the models are slightly different and adjusted, we kind of allow you to generate any single character you can think of. So your favorite movie character, you can generate a figurine, print it, or like put it in your game. So, yeah. Awesome. All right. So in your interpretation of IP, no problem using other people's IP for the output. Not yet. All right. So there's a lot of lessons here for you about IP usage. I would not let people output other IP things. We could make a video game with you and it, Jason. Wouldn't that be sick? What's that? We could make a video game where you're like walking around and stuff.

48:08That would be sick. That would be sick. Yeah, we could have you like slash down the CCP coffee. I love it. I love it. So is there any room in the seed round now? I feel like I need to get my beak wet here. What's going on with the seed round? What's going on? Is there room for J-Cal in the seed round? We can let J-Cal in if we get exclusive publicity rights. We can deepfake you whenever we want. I can make giant 3D printed models of you. I'll put it on top of the power supply notes. All right. Well, send the 50 bips as an advisor agreement over, and we'll get that signed straight away. And then let me know about the term sheet.

48:42Maybe we'll pop in a quick hundy and we'll be along for the ride, but I wish you great success with it. And yeah, it was great to meet you and have a little back and forth with you. All in good fun, folks. All in good fun. All right. Peace. Check us out at adam.new. Oh, yeah. Adam.new? Dot new. Adam.new. And if you want to generate JSON in 3D and post it on Twitter, you can do it. Subdomain? N-E-W? N-E-W, yeah. It's like bolt.new. Holy cow. That's a neat domain. and I like that. Yeah. Zach continued success and great to get to know you and I will see you on X, sir. All right, well done. I thought it'd be fun to just cut this off before it becomes like a Palmer Luckey situation because enemies accumulate and this kid, Zach, is just, I know a punk rock founder.

49:32Like I kind of, this is my speciality is finding the punk rock founders because I was kind of one of those back in the day. because the second I was like, hey, you know, maybe you can't do this kind of thing. He's like, oh yeah, let me escalate this because I can get more views. And I'm like, yep, that's what I would have done when I was your age. Yep, that's exactly what I would have done as a 25 or 30 year founder. You told me to stop. I do it twice. You told me to stop again. I do it four times. You told me a third time that I'm probably going to tone it down slightly. But in all seriousness, this will become an issue over time.

50:05I don't recommend doing this as for people listening. Don't ever take a known person's talking about your product and put it in an ad. Because this recently happened, actually. I think it was Huberman. Look up, ask producer Claude about Andrew Huberman doing a lawsuit based on people. So I think it was Tim Ferriss, Andrew Huberman, and Peter Atiyah. And one of those three or multiple of them had to write a bunch of cease and desist letters. And I think it was Huberman because people were taking Huberman who might talk about a compound on like Joe Rogan. And, you know, he might talk about, I don't know, like vitamin D or something.

50:52And then they cut him talking about vitamin D and then put their supplement ad on it. But this was like not just, you know, a startup, you know, in their seed round doing it on Twitter precociously. This was programmatically. They were buying millions of dollars of ads and using his likeness in it. And then somebody did it to Joe Rogan as well. Now, if you do it to those two people, they're in the business of their image. Like I think, you know, you would have Joe Rogan reading ads for these people. So now you're circumventing his ability to run his business. And that's when you lose in court for a big dollar number.

51:25And you would have to go back and pay for that violation and major worry. Now, in this one, you could argue a little bit of parody, no harm, no foul. There was no damages. But with Huberman's case, Andrew Huberman's case against a nutraceuticals company or something, he can actually claim significant damages because he has advertisers as well. So here from May of 2024, Jason, is an Andrew Huberman post from Facebook in which he says, unfortunately, many ads are currently running on social media. they use my face or a drawing of it to sell it. He sent cease and desist letters. And then he lists, here are actual sponsors down below.

52:03So yes, this was Huberman, you were dead on. Yeah. And I don't know if it actually went to a lawsuit or not, but this is going to become, you know, young people have a different view and technologists have always had a different view of IP copyright and kind of fairness of this stuff. And so in one instance where somebody did this programmatically and paid for it, they were like, we're just fans. And I was like, that's fine. but you understand this is our business. And then unbeknownst to me, they went and bought some ads on the Sweden startups. I was like, wait a second, is this the same company that was stealing?

52:35They're like, yeah, they felt bad about they bought these ads. Now this was a company doing tens of millions in revenue. So I was like, okay, that's a nice gesture. Sure. I didn't ask them to do it. But I did think, gosh, I always think when a founder is doing something close to the line, I should just give them a little bit of like, hey, be careful. Which is actually what I did in the Parma Lucky situation. I was like, you know, don't buy covert trolling ads because it could blow up in your face. Anyway, I may have said it in a spicy way, so I'm sorry if I said it too spicy, which I probably did.

53:02I was a little spicier back in the day. But we've got more news in the docket. Let's get to the docket. Is there anything more news for us? Because there's a lot of major news today. I saw that, I think the biggest one is OpenAI dropped two open source models? Wait, I thought they were out of the open source business. Pause. Not open source, open waits. Open waits, not open source. Okay, that's totally different. But still much more open. And as - So you can't see the source code, but you can change the weight. So maybe you could explain to the audience what's going on here. Yeah, okay. So I knew this question was coming.

53:37So I spent a lot of time trying to figure out how to best respond to what the hell is an open weight. Yeah, sure. First thing at the top level, you can download these models and you can run them locally. You don't have to interact with them via an API or an inference provider. You can run them on your computer at home. Or if you have a serious GPU from NVIDIA, you can run the larger model there. Open weights is not the code. So you cannot download the code and actually extract the individual lines. But what you can think of is that weights are a bit like memory or the knowledge of the app once it was trained.

54:07And because they're open, you can see what OpenAI did. This is dramatically more transparent than the most recent generations of OpenAI models. And it is, I think, driving some change. So for example, xAI, Jason, in the wake of this announced, Elon said this, I think it was today, that they're going to open source, maybe open wait, Grok 2, after having open source Grok 1 back in the day. So a good move by OpenAI, you love to call them closed AI, here they're making a good step in that direction, but it's not open source. And so people are mixing those two a little bit. And I think it's important to make the distinction.

54:40Okay. So in a way, when you train these things, you might wait certain information over others, I think is another way to say it. And so, you know, like maybe this is not correct. I would like to have somebody really figure out how to explain this. We understand that there's parameters and all this stuff. But how do you weight different things as being truer or more important, I guess, is how you could think about this. So if you were going to do something specifically with this model and make a cursor AI competitor, you would maybe tweak the weights. Maybe you think, oh, the stuff we scraped off Stack Overflow is more important, or this GitHub stuff is more important, or this GitHub ranking makes this thing more important, and slide the weights around.

55:28It's always been known that Microsoft owned the source code and the weights in the open AI relationship where they bought like half the company or whatever it is. So this feels really interesting to me because you can also run it on your local computer. and we knew this was coming. And what this means is when you start thinking about IP or privacy, this is going to have a big impact on it. If you're doing something like, I don't know, I'm taking, you know, pictures of my grandparents or whatever and I'm making them into videos like people are doing. Sure. Do I want to upload those photos to OpenAI and give them those?

56:08And then the next person who makes pictures of a grandpa, it's my grandpa. you know, you get the idea. So, and then on a privacy basis, if you were doing deep analysis of your documents, do you really want that in the cloud with somebody else? And then somebody else, you happen to index, I don't know, salaries, bonuses, legal documents, you know, this stuff can get pretty gnarly and then it's training their model. You will have a large language model in the operating system on your phone, on your computer. This is where Apple, which has been just totally disgraziad when it comes to LLMs, could actually maybe come from behind.

56:52If they were to take one of these open models or work on one and then incorporate it into their operating systems, of which they have billions of users, oh lordy, that would sync with, and I predict they'll do this, I predict Apple will use one of these models encrypted on your devices so you never have to use a cloud-based service and be subpoenaed for what you do. And when you choose to subpoena Apple because there was a terrorist attack at, was it San Bernardino, and they wouldn't unlock the iPhones? It was San Bernardino, I believe, right? And they were like, oh, there was a terrorist attack.

57:33Love to help you CIA, FBI, everybody, but we don't have access to it. That's a bold thing to do as Apple, right? Like terrorism, you know, explicit images of minors, like there's a small cohort of things that any reasonable person would say, turn it over. But Apple has taken the stance quite bravely, I will say, of saying we don't have access to that. And then giving you that information of the terrorists would compromise billions of people on our platform. Therefore, we're going to make this higher level decision that ankling, slowing down the next terrorist attack and finding out where the terror cell is, is more important to Apple.

58:17They're willing to pay the price of another terrorist attack occurring from a sleeper cell, then compromising billions of people's privacy by having a backdoor. That's a hard, hard, hard decision to make. I'm glad I don't have to make it. Yeah. This just came up. NVIDIA put out a blog post yesterday or today, Jason, I forget exactly which. And they said there are no kill switches and no backdoors in NVIDIA hardware. NVIDIA is currently trying to get back into selling its H20 chips in China, working with the Trump administration on getting permission to do that. And I think China's a little bit worried that what they're going to get is backdoored chips, just like the US would worry about a Chinese company sending compromised chips to the US.

59:02And so they're making a pretty clear stand about that. And then also we've seen the UK work recently to try to get companies again, to break encryption and provide backdoors. And Nvidia said in its post, don't forget the, was it the Clipper chip that had the compromised backdoor that led to all the security flaws? And so in the case of Apple, I feel like what they're doing is they're just saying encryption is or is not. And there's no middle ground to offer up to law enforcement. I'm sure that if Apple could find a way to have encryption be compromised in a way that was safe, which I know sounds oxymoronic, they'd go for it.

59:35But I just think they're being technologically honest. So that's my take. Of course, we care about children. Of course, we care about stopping terrorism. But I mean, we also care about everybody's privacy. A right to privacy is the term I just use. That's for public figures having the right to publicity, actually, is what it's called. But a right to privacy is even more core than, you know, what celebrities have to deal with. And, oh, I was correct. I see here producer Claude lets us know that Tim Ferriss successfully got an injunction. Very good. Tim Ferriss and a company he owns, Creasa Performance obtained a default judgment against defendants alleged to have improperly used his name and likeness in connection with a fraudulent scheme.

1:00:18According to the case, Alliance Publishing and Pagan Twitty perpetrated a fraudulent get-rich scheme in which they published a flyer asking targets to mail$97 and$100 to a dealer and whatever, blah, blah, blah. So anyway. People are scammers. I mean, scammers are out there. That's part of becoming a known entity in the world is that people will try to leverage your followers into some dark place. Okay. On the OpenAI point, though, just a couple more small things I want to make sure we share with founders because I think this matters for what they're building. So there's two models from OpenAI you can play with.

1:00:54There's the 220 billion parameter model and then a 20 billion parameter model. One's more tuned for data centers, one's more tuned for home use. There are lots of fun. The metrics seem to be, by benchmarks, pretty good. I've heard some people say there's a lot of hallucinations, but it's certainly, I think, having more open models in the market is good. And Jason, on the show, we've talked about how when we think about open source AI, it's really a Chinese product now as American companies mostly go closed source. So here is a rejoinder from the West, if you will. I think that's very important.

1:01:24And then there's two other, there's a couple more things about AI models. And I'm going to just run through these, but I think for all the founders out there who are building and are curious what they can build with, here's the latest. So the other thing is DeepMind, which is part of Google, released a thing called Genie 3, which is a world model that has much better context maintenance. So essentially, if you're playing inside this world model, Jason, and you turn around, it doesn't forget what you've seen, which means that it's now much more useful, much more high fidelity. It's just great. I'm not going to go through a demo today, but Genie 3, if you want to take a look at that.

1:01:57Anthropic dropped a very incremental improvement to Claude 4 called Claude 4.1. Again, the benchmarks say it's better, but it's going to be more a matter of taste. So I'll be curious to see what people think about that in the coming days. Then the big news is that tomorrow, OpenAI is going to release GPT-5. Oh, well, this is the big one. Wow, that's the shebang. Maybe we should do a live release of that. That's incredible. What do people expect in the chat GPT-5? I heard a rumor, no hallucinations, more reasoning. What is the expectation here? Or have they tipped their cards and said, here's what's happening?

1:02:37And when was the, when was chat GPT-4 launched? I wonder how many months between those. That's another interesting. Producer Claude, can we get the gap from GPT-3 to GPT-4 and then GPT-4 to tomorrow, please, so we can compare those? That'd be lovely. The most important reporting that I saw to answer your question, though, Reuters spoke to several people who have had access to GPT-5, didn't share their name because they're under NDA, but they specifically said it's better at coding, it's better at reasoning. But critically, of course, it's going to be better. The question is how much, right? Sure.

1:03:07They said it's not going to be as big of a jump as we saw from GPT-3 to GPT-4, from 4 to 5. So they're downplaying it. That's interesting, trying to manage expectations. So GPT-4 came out in March of 2023. GPT-3 came out in June of 2020. So that was just under three years. This is just over three years, just over two years. It's actually, it's a little bit faster from 4 to 5 than 3 to 4. Huh, who would have thought? Interesting. There's been a big kerfuffle between Cloudflare and Comet. And you emailed the press department there. And I CC'd the founder. And we got kind of the backstory here on the perplexity versus Cloudflare.

1:03:51I don't know if it's on the docket. Did I not add Arvind to that email? I did. You probably didn't. But I did because I was like, Arvind should know about this. I meant to. My bad on that one. Okay, so anyways, no apologizing on Twist. The backstory here is... Getting there, getting there. A while ago, maybe a month, Cloudflare announced Paper Crawl. This was a big deal. We talked about it on the show. The idea was if an AI model going out to the internet to collect information on behalf of a user query wanted to come to jasoncalacanis.tumblr.gif, you could say, no, you have to pay us a cent or whatever to come take our information and bring it back to your user.

1:04:29I think we said this was pretty reasonable and an interesting experiment in economics for the internet. Time passes. Paper crawl has been in close beta now. They're working with publishers. We'll see what it looks like when it comes out. They have not responded to my notes asking how that's going, but we'll see. So perplexity, sorry, Cloudflare comes out this week with a post essentially accusing perplexity of very bad behavior online. They said that, look, you know, we set up a test website that had, you know, a big stop sign for AI, and then we use perplexity's models to take a look at this.

1:04:59And we found the information and accused them essentially of being bad online citizens. And perplexity fired back with, you guys don't know how technology works, which was an interesting moment because here we have two very different perspectives that don't seem to be talking in the same language, Jason. So maybe you can explain what you mean by they're acting in good faith because I think that's an interesting, important point here. Okay, so you never let a good argument get in the way? reality, because both companies now are being talked about, which I think is part of the point. Yes. You can send a crawler out.

1:05:35There's many third-party services that you can fire up as a developer to go crawl websites. It will fire up a web browser, a virtual machine, and it looks like a user agent, right? So it looks like a human being is browsing, say, something that doesn't let you scrape it, like Reddit or Quora. Like you ever go to Quora and you try to click on an answer and it gives you like maybe the first 500 words that says, hey, you got to log in to see this. I hate that. Well, but it's the perfect example of somebody trying to protect their IP from these crawlers. So what Cloudflare is basically saying is perplexity is trying to circumvent that.

1:06:11And I can tell you all the LLMs have tried to circumvent this, but now we're starting to get to the point where lawsuits are flying and something's at stake. And that's why these language models now need to behave because Disney, New York Times are suing and we've got actual court cases. So people are being more thoughtful about this. It's amazing what Cloudflare is doing by saying, hey, if you want to crawl our websites, that's okay. Here's the rules. Here's where you send the money. And if Cloudflare, what does Cloudflare take? 10%, I don't know what their vig is. I would really love to know that.

1:06:45I don't think they've released that, but if it's more than 10%, we're going to take them out back and whack them with sticks. I think 10 % feels like a reasonable vig. I would pay them that. I don't think 30 % would be, but 10 % seems reasonable for them to manage the buy and sell side and 10 % plus your Stripe fee. So they get a full 10%. I think that would be reasonable, actually, maybe. You're literally describing the Substack model, and I think that's fine. Yeah, it's like enough for them to maintain the service and pay for their team. It's not predatory in any way. It's kind of genius, I have to say.

1:07:13It's a really unique way to do it. It empowers the content creators, and it gives a programmatic way for LLMs to ask for permission. Now, in comes the rub. we've been talking about the browser's role in this. You could, and we talked today about OpenAI having a language model you can run locally. You're allowed to do whatever you want on your desktop with other people's IP. So if you paid for it, so if you bought my book and you decided to take screenshots and you decided to run analysis on it and you bookmarked it, more power to you. As an individual, you paid for it, you can remix it. Just like you take the physical book, you could cut it up, put it on your wall, make a copy of it for a backup, but as long as you don't distribute it, right?

1:07:56So if you want to photocopy my book, so you have a backup and you have it in one house and the other in case the house burns down, God forbid. I mean, it's nonsensical because you could get the bound book for the same price as a photocopy. You get the point. Browsers are not detectable. These third parties are, and I think what's happening is there might be some confusion that Comet users using the browser are doing queries. So when I showed, hey, I'm doing a United Airlines task and it goes and it fires off on my local desktop, the browser to go do things. What I suspect is happening here is when I just told you I had Slack summarize the channel using the Comet browser powered by Claude, how does Slack look at that?

1:08:43It looks at it as my user agent. It looks at it as my desktop. It doesn't look at it as a backend scraper. Now, if you took back content and you trained your LLM on it, you would be using me as a proxy to backdoor into stealing content. So let's use the Cora example. Cora says, hey, you can't use this in your language model. But you and I go on there and we say, oh, give us all the questions about Corvettes. And we start reading them. We say, oh, get us the best answers on these Corvettes and teach us about the whole corpus of Corvette history or whatever. And we build this whole document and deep research on it.

1:09:16And we've queried, you know, over a number of days, a thousand pages of Corvette information and history and lore that Quora owns. Quora owns that. Even though the users made it, they made it with the understanding it goes into the Quora corpus. And Quora said, you can't have this. Well, what if that then fed the perplexity language model? Well, the founder of perplexity said they're not doing that. So whatever you're doing is not happening. So there might just be like they're talking through each other, perhaps a misunderstanding of the user agents. But this is where the details matter. Details really matter.

1:09:50And that's what a show like This Week in Startups is about. We get into the very nuanced, granular details of these kerfluffles. And this Donnybrook is valid. It's an important one. It's very important. And I think the way where I come down on this, putting aside the technical differences of whether or not Cloudflare, quote, misunderstood 20 to 25 million user agent requests that are not scrapers, according to Perplexity's blog post. How many? 25 million? Yeah, 20 to 25 million user agent requests are not scrapers. Anyways, that's from the long Perplexity post about this. We will sort this out.

1:10:28But what I'm worried about is this, the idea that if I go to, let's say, ChatGPT and ask it a question and it pings 47 websites, collects the information and brings it back to me and gives me a summarized paragraph. And then OpenAI says, well, we're not doing that. Alex is. Not really. There's a automation and scale differential here. And I tried to talk about this for something that I wrote. And I was trying to think of the analogy as food. So Jason if you come over to my house, you can eat anything you want Open the pantry at the refrigerator the freezers steal from the kids. I don't care. There's lots of food come over eat But if you sent a like a robotic courier to my house to collect food for other people Yes, which you're doing it and the automation because it's not one-to-one And so I'm concerned that people are saying as long as a user requested in some tenuous capacity anything downstream from there counts as Alex requested or Jason wants it.

1:11:28And therefore it's okay. Not true. Yeah. And the other thing is these websites will also block you if you do unnatural behaviors. So people for many times, for all time have been creating local index like Google does. And so they'll just have their browser open a hundred LinkedIn pages and start scraping. And there's infinite number of tools online you can get to do a user agent in your browser to spam people with in-mails. and you're paying for emails, there's a throttle on it, you can only send so many, but there were tools in the early days to just automatically like posts on Twitter, automatically retweet stuff, and create all this garbage on the internet.

1:12:07And so it's up to the website owner as well, that if somebody does more than 100 clicks in an hour, to pause them and say, hey, we noticed you did 100 page loads in an hour, a normal human being can't really do that. So what's going on here and pause your account or freeze your account. And sophisticated people like say Crunchbase or the Wall Street Journal or the New York Times, anybody who's got a subscription product or Quora, Stack Overflow, they build these things in to know a user or what appears to be a user is doing something with malfeasance. This is incredibly niche, but an important topic.

1:12:45Okay. This has been another amazing episode of This Week in Startups. My amazing co-host Alex, Lon, x.com slash Alex, x.com slash Jason, x.com slash Lon, x.com slash TWI. Startups, we'll see you all on Friday. Bye-bye. Bye-bye.