B (2:12)

Focus on the game, focus on the game.

C (2:14)

I feel like there's room to grow. But yeah, I guess what I'm trying to get out of it is just to like meet as many people as possible, make as many connections as possible because this trip probably won't, a trip like this probably won't happen again for a while. So yeah, that's sort of my goal.

A (2:30)

What's the status of The YC application. You've submitted it?

C (2:33)

Yeah.

A (2:33)

Have you heard back yet?

C (2:34)

No, it hasn't been. Still like I'll need to.

A (2:36)

Recommendation.

B (2:37)

Yeah, we can, we got, we got a lot.

A (2:39)

Your YC alum watching this. Please go leave a recommendation. Yeah, but congratulations. Thanks so much for coming by. What is the stage of development of the actual application? The product itself. Are you live?

C (2:54)

Can people go down demo? Right now we're getting like beta testers, but the beta should be launching soon, probably by like the end of this year.

A (3:03)

Do you have a wait list? Are you doing email capture yet?

C (3:05)

Yeah, like wait list beta testers. We've got like a couple hundred, but yeah, very cool.

A (3:10)

Incredible.

B (3:11)

Well, congratulations on all the attention. I'm sure you'll convert it into a lot of opportunity and have a great trip.

A (3:18)

Yes. Great to have you and good luck with the YC application.

C (3:20)

Thanks so much.

B (3:22)

Looking sharp in the suit.

A (3:23)

Looking sharp in the suit.

B (3:25)

Amazing.

A (3:26)

Have a good restream.

C (3:27)

Thank you.

A (3:27)

Thanks for stopping by. Before we move on to the rest of the show, let me tell you about Restream One livestream. 30 plus destinations. If you want to multi stream, go to restream.com calm and it's been three years since ChatGPT launched. I wanted to reflect a little bit. Everything changed or maybe nothing changed or maybe some amount of change in between everything and nothing. You're more on the nothing changed camp. I sort of agree with you. I was sort of. I was sort of reflecting on like, okay, Thanksgiving's happened. It was Thanksgiving over the weekend. You know, how different is my world? Like, there's not a humanoid robot that's cooking for. And also, even if we had a humanoid robot, I think Thanksgiving would be the day we let the robot sit in the closet. Because we enjoy. No, no, let us cook. We enjoy cooking. Cooking is a fun family experience. And so of all the things, let us cook. Thanksgiving is like the track day of cooking. Like, even if you have the robot that does it, you still want to do it on Thanksgiving. You don't want to cook on a random Tuesday when you're busy. You got lunch, you know, all this other stuff. Thanksgiving is the Nurburgring.

B (4:32)

And I was doing some dishes after Thanksgiving and I felt like it was a good way to kind of like, it felt like walking off the pie.

A (4:38)

In a little way.

B (4:39)

It wasn't walking. Wasn't walking very far, kind of back and forth.

A (4:42)

And so, yeah, so that, that, that hasn't really changed that much for me. I was thinking I was reflecting more on the agentic commerce thing, it feels like ChatGPT and OpenAI, they really are pushing to make revenue from agentic commerce, like before in this holiday season. And incredible speed of execution. Like, clearly it's a big opportunity. If you can figure out how to run ads, commerce, convert, take a cut of that. That's big. My experience actually demoing it, it was kind of interesting. Like, the actual product in ChatGPT is pretty good, but you can see that the walled gardens are already going up. So one place that I like to go to for reviews of products specifically around the holidays is the wirecutter. Now, the wirecutter, their whole twist was they wouldn't rate each product, which. What they would do is they would pick a category and then they would just tell you what their best product was in that category. Sort of like a cluster max of vacuums. So they would give you the platinum tier vacuum and then a budget pick. And so I've always liked the wirecutter. I think they do a very rigorous job. They were acquired by the New York Times. The New York Times is currently in a lawsuit with OpenAI. And so if you go to ChatGPT.

B (5:51)

And say, hey, and I think they're about to be in a lawsuit with.

A (5:53)

David Sacks, maybe, maybe. Which we will talk about on the show in a little bit. But if, but, but, but if you go. So I went to ChatGPT and I was like, hey, okay, pull a deep research report. Just pull everything from the wire cutter and, and tell me every category and every product that's top ranked. Because then I can just scan it really quickly and be like, oh, yeah, I didn't even remember that that category existed. That would be a great gift. I'll get it. And I'll go through the. I'll go through the wire cutter link. I'm fine with that. I'm paying ChatGPT. I'm happy to go and use their affiliate link on the wirecutter. That's how the wirecutter monetizes. But it couldn't do it. It couldn't do it. It said, hey, we can't touch the wire cutter. Like, it's off limits. You got to head over there yourself. Pop open a Chrome tab, brother, if you want to get over there, like, that's on you. Or maybe an Atlas tab, I don't know. But. So that had not really changed that much for me. But the one thing that did really change on Thanksgiving was the discourse. The AI narrative has fully arrived to just family and friends.

B (6:55)

You mean family in the Home.

A (6:57)

Yes, yes. In people that don't work, in technology that don't. Their job is not showing off their favorite trough. Not that more talking about. Is it a bubble? Where do you think all this stuff goes? The stuff that we've been talking about?

B (7:10)

I'm sure you're not living in a bubble. You think the average family in America.

A (7:14)

Is not living in a bubble? I saw multiple newsletters where the whole conceit of the newsletter going into the holidays was how to talk to your family about the AI bubble and how to talk to your family about AI generally. And I think it's real because if you've been watching your 401k over the last year, you've seen a massive spike and then a recent sell off. And if you've turned on any news or opened up any newspaper, you've been hearing about $1 trillion and you're like, what? A trillion dollars? That ChatGPT app, they need a trillion dollars to make that thing work.

B (7:45)

Chat.

A (7:46)

Right, chat. And so it is a really big narrative. And so I wanted to reflect on what has actually changed over the last three years. And specifically in the Mag 7. The Mag 7 has been on absolute tear. Just over the last three years, the value as a whole has basically tripled. It was a little under $8 trillion. Now it's over $21 trillion. It's a lot of value created in the last three years. Nvidia was second to last in the MAG7. When ChatGPT launched, it was worth just $420 billion. Something on there today. The stock is up over 10 for $4.36 trillion.

B (8:29)

And up today, and up today. Despite all the chaos, Dylan Patel was.

A (8:32)

Trying so hard to bring that stock down, but he couldn't do it. He's coming on the show at noon. We're going to confront him about his bear posting and whether or not the.

B (8:41)

Market is funny enough, Broadcom is down today.

A (8:43)

Okay, why is that?

B (8:46)

The maker of the tpu.

A (8:47)

Oh, yeah. I mean, a lot of these things, it's like it's already been priced in. I mean, even when you read that semianalysis piece, you know, a lot of it's like, we've been writing about this for months. People have already put this trade on, et cetera, et cetera. But I do think that the, The Nvidia, the 10X that's happened has really created some crazy zealots and just an entire industrial complex because there are so many people who put, who heard AI. They tried the ChatGPT thing and they Were like, this is big. How do I get it on this? I can't buy OpenAI. OpenAI is running away with it. Oh, they need Nvidia chips. That's the logical next step. They went in Nvidia and they got a 10x and they could have gotten a 10x on like $1 million. $10 million. Like there's no amount of money because it was, it was already a $420 billion company. So you could be, you could put your entire retirement savings in it, no problem. Complete liquidity. Right. It's not, oh, you got to get some spv. It was really easy.

B (9:45)

Sikh Chen from Runway was saying that back. I think it was 20, 20, 2021. He, he said he put an uncomfortable amount of his net worth into Nvidia.

A (9:53)

And obviously and near cyan. Same story, right?

B (9:59)

Still underappreciated the Nvidia 10 year fund. All it does is buy Nvidia just by investing in it. You can't possibly sell God's chosen company. Yes, that's what the, I think title.

D (10:10)

Of the fund was.

A (10:11)

Oh, really?

B (10:11)

Yeah.

A (10:12)

That's hilarious. And so, I mean, yeah, there's been a ton of zealots. We're going to talk to Dylan Patel at noon about some of the zealots that have been attacking him previously. The world's largest company in November of 2022 was Apple. And at the time they had a sizable lead over Microsoft, Amazon and Google. Now that gap has closed a bit as the hyperscalers have grown more over the last three years on the back of the AI boom. And it's interesting, I mean, you can sort the mag7 by market cap and today you get the following ranking. Tesla, then Meta, then, then Amazon, Microsoft Alphabet, Apple, and then Nvidia at the top. And the big question, I think that's on everyone's mind and kind of underpins the horse race that we cover every day on the show is what will that ranking look like in the next three years? Is Nvidia really a monopoly? Is it impervious to attacks from different suppliers?

B (11:08)

What does Broadcom have to do to get into the Mag 7?

A (11:12)

I don't know.

B (11:13)

Tesla's sitting at 10 on the market cap companiesmarketcap.com, which we are not affiliated with, which is a fantastic website. Broadcom is sitting at number six above Meta currently.

A (11:27)

I don't know. I mean, I think several years in the one trillion dollar club, just being undeniable at that scale. There's also just a bit of branding. Some of the companies that made it into the Mag 7 were I feel like the Mag 7 leaned understandable, like not that deep in the supply chain. Even Nvidia was the deepest. Nvidia had the least of like a consumer brand. But still a lot of people used the gaming graphics cards. Broadcom is really tricky because there's no consumer angle whatsoever. Consumers can buy Tesla, they can use Meta products, they can buy on Amazon, have a Microsoft operating system, they can use Google, they can have an iPhone and they can have an Nvidia gaming graphics card.

B (12:15)

The top right now, Tesla sitting at 10 TSMC at 98 is Saudi Aramco 7 Metta and then 6.

A (12:23)

I also think you have to be an American company to be in this like mag 7 or whatever the hot ranking is. Like Fang Fang did not include, never included oil companies, never included international companies because if you go there then you could be like oh well let's include like the Chinese tobacco company that's worth a trillion dollars or something that like there are some crazy, there are some crazy like foreign owned companies that are, if they were independent might be worth a TR because they just have so much of assets. Yeah, exactly. But it doesn't really count because it's just sitting there out in the ether. Well, let me tell you about Gemini 3 Pro. Google's most intelligent model yet. State of the art reasoning, next level vibe coding and deep multimodal understanding. And speaking of that, Buco Capital bloke has a post here. Gemini app downloads are catching up to ChatGPT and Gemini users now spend more time in the app than ChatGPT users. People are going back and forth on can Gemini catch up. You know the model clearly very good. The big bombshell in the semi analysis piece over the weekend was this idea which I think has been bandied about before. This idea that OpenAI has not done a proper pre train since 4.0 and the 4.5pre train kind of got mothballed and so but there was this question about like is pre training dead? Seems like the Google folks said no it's not. And then they went and did a Pre train and Gemini 3 outperformed anthropic also pre trained. Pre trained Pelled.

B (13:55)

We asked Sholto about this and he said oh yeah, we're still bullish on scaling. I think actually Sholto kind of like in the subtext said the reason opus 4.

E (14:06)

5 was good is not because it.

D (14:07)

Was a new pre chain, it's because it was rl.

A (14:09)

That's what I read. That was your reading?

F (14:11)

Yeah.

A (14:11)

I feel like there's still juice in the lemon of pre training, but it's not scale. Like, we only have one Internet. Ilya was correct about that. It's not scaling the size of the pre train, which is what happened with 4.5 from GPT. 4.5, that was just bigger, I guess. But it does seem like there's little optimizations that you can do on the pre training side. But I don't know, we'll have to dig into it. But I think the thing that no one is debating is the fact that the Gemini 3 as a model with Nano Banana Pro with VO3 is just like the actual foundational intelligence is plenty good to be dominant in the consumer AI category. The question is, can you actually get people to install the app, use it, can they enjoy it? Do they not churn and go back to ChatGPT? I've been fighting back and forth left and right going into one app and the other. I was getting a ton of disconnects errors with the Gemini app, even though the model's great and there's some really cool features.

B (15:13)

Yeah, they need to catch up on the product side.

A (15:15)

Exactly. Yeah, yeah, the product side. And so a lot of people are saying like, oh, Gemini team should just. The app team should just go and, you know, copy ChatGPT homework and. And, you know, copy all these little features. I've put out a post that the folks over at the Gemini team actually, you know, did turn into bug reports and I think are working on. But it really does seem like it's. It's a really. It's a sprint to actually create an app that is as sticky as ChatGPT because ChatGPT, the app is fantastic and very, very, very well designed. And so the.

B (15:45)

Yeah, and there's some reporting from similar web is what the FT is using to track average user minutes.

A (15:54)

I always find those hard to. I mean, it must be like Nielsen ratings where they're like polling people or something because.

B (15:59)

Yeah.

D (15:59)

And I don't know, you can't get.

A (16:00)

A pixel in OpenAI. Like you can't get a pixel into the Gemini app.

B (16:04)

And are they counting user minutes if a tab is open but I'm not actually in.

A (16:08)

And is this just desktop? Because that's like completely separate from mobile.

B (16:11)

Use desktop and mobile web, which again, I don't.

A (16:14)

No one's using a mobile app that.

B (16:15)

Are using mobile web.

A (16:16)

I don't know. I wouldn't read too many too much into this data specifically. I would much more look at like, what are the structural advantages that we Know, exist. And I mean with Gemini, one of them is to that point about the wire cutter. You know, you know where the wire cutter shows up Google search results. You know what, you know what company has one bottom for scraping everything? Google. So the Google bot is identifies as one entity. So you can either say I'm allowing Google or not. And it's a big, it's a tall order to be like, yeah, I don't want to be in Google results. And so a lot of companies are saying, yeah, I'm good with Google showing up in Google results, but that also shows up in AI search results. And they can, and there are things that companies can do to say, hey, don't put me in the Gemini, you know, like training data set necessarily. But in terms of just, just actually showing up, you've seen it in the Gemini app, it says using Google search. And so if I go to Gemini and I say, hey, head over to the wire cutter, find me the best vacuum cleaner. Google probably can do that. Gemini can probably do that. Yeah, test it. Whereas OpenAI is in a fight with the New York Times. Whereas Google and the New York Times, they might not love each other, but they definitely have an uncomfortable truce. Right.

B (17:32)

A funny Gemini integration that I've used is that you've land in a, in a hangout and you just say, who is this person?

A (17:40)

Is this real? You can actually do that?

B (17:42)

It is, it pulls up a sidebar. You can just ask like, who, who am I meeting with right now? And it'll give you like a, it's.

A (17:49)

Clearly who am I meeting with? What should I say? What should I say to them?

D (17:52)

Should I ask them what is my name?

A (17:54)

What, what do they want to know about me? What should I tell them about me?

B (18:01)

Okay, so Gemini was able to pull Wirecutter recommendations. I don't know. This is interesting. I feel like, I wonder if Wirecutter is actually benefiting from this in any way yet I'm assuming for sure because.

A (18:20)

Google hasn't, Gemini hasn't rolled out the agentic commerce stuff that would actually scrape out the referral, the referral token. And so if I'm in Gemini and I'm saying I'm going to do some agentic shopping or whatever and I say, pull me the best vacuum cleaner from the wire cutter. It goes over and does that and then I land on the wire cutter and then I click that link, that should give the wire cutter the credit. Now if I as a follow up prompt go in Gemini and say, okay, great, the wire cutter told me the Best vacuum cleaner is from James Dyson, of course, is the Dyson find me the Amazon link. Well, Gemini's probably not given the wire cutter the attribution at that point. It might even be taking its own attribution. I don't know exactly how it's. How it's functioning right now, but I would imagine that that link does not get reinstantiated as the. As the wirecutter affiliate link. And so we could see. I mean, these are all, like, going to be pretty existential questions for the SEO crowd. Anyone who's monetizing off of SEO. We saw some. Some screenshot that apparently site traffic to Vox properties is down 50%. And I don't know how much of that is just the shift to social media versus the shift to.

B (19:33)

Yeah, how much is it their business strategy just being like, hey, we want to do more video and that'll be distributed off our site for the most part.

A (19:41)

I think a lot of people generally do not. They consume more and more content on social media platforms. They go from YouTube to their RSS player to audiobooks to Twitter to Instagram, and they kind of bounce around from one and the other. And then every once in a while, they will go in and actually land on a particular site. Like you can. If you go to tbpn.com, you can get our newsletter in your inbox every morning. And you can also sign up for Cognition. They're the makers of Devin, the AI software Engineer. Crush your backlog with your personal AI engineering team. Well, speaking of the New York Times, David Sacks is going to war with the New York Times. He says inside the NYT's hoax factory, calls it a hoax factory because the New York Times posted a piece about David Sacks saying that the headline was, Silicon Valley's man in the White House is benefiting himself and his friends. And Ryan Mack was going back and forth with Sean McGuire or Sean. Yeah, Sean McGuire. Ryan. Max says today has been a good example of what X has become. Complaints from a subset of wealthy tech folks about a story that circulates more widely than the actual story itself. Musk bought the platform to control the message, and he and his friends are getting just that. And Sean McGuire says, you don't get to run this headline, then write an article that doesn't validate the claim, and then get away with playing the victim. We see through the ruse. And so David Sachs has responded in full to the NYT's hoax factory. He says five months ago, the five New York Times reporters were dispatched to create A story about my supposed conflicts of interest working as the White House AI and crypto czar. Through a series of fact checks, they revealed their accusations, which we debunked in detail. Not surprisingly, the published article included only bits and pieces of our responses. Their accusations ranged from a fabricated dinner with a leading tech CEO to non existent promises of access to the President, to baseless claims of influencing defense contracts. Every time we would prove an accusation false, NYT pivoted to the next allegation. This is why the story has dragged on for five months. Today. They evidently just threw up their hands and published this nothing burger. Anyone who reads the story carefully can see that they strung together a bunch of anecdotes that don't support the headline. And of course, that was the whole point. At no point in their constant goalpost shifting was NYT willing to update the premise of their story to accept it. I have no conflicts of interest to uncover no conflicts of interest. As, as it became clear that NYT wasn't interested in writing a fair story, I hired the law firm Claire Locke, which specializes in defamation law. I'm attaching Claire Locke's letter to the NYT so readers have full context on our interactions with NYT reporters over the past several months. Once you read the letter, it becomes very clear how NYT willfully miscre, characterized or ignored the facts to support their bogus narrative.

B (22:41)

So Will says hiring Claire Locke for this is sick cruise missile to blow up a straw hut.

A (22:48)

He's a big fan of litigation. He loves litigation. Well, people have been supportive of this broadly in tech. Let's go through some of the reaction. Sam Altman says David Sachs really understands AI and cares about the US leading in innovation. I'm grateful we have him. Brian Armstrong.

B (23:07)

Yeah, here's my takeaway. If you believe that AI and crypto are industries that we should support in the United States, then you want to have a czar focused on those things that generally feels positively about those things and wants to create the best possible environment for those industries to thrive in the US I think that there's actually a debate on both fronts, right? Like there's people on the left that think AI and crypto are just default bad, they want less of them. And there's people on, there's people on the right that believe that too. But I think that ultimately there's arguments for why the US should lead in stablecoins, which, you know, is part of the, part of why the genius act is important. And a lot of, you know, the AI action plan there's going to be debates on individual points in that. But in general, I think creating an environment in the US where we can continue to lead in AI is important. So I think there wasn't, I didn't see any sort of like smoking gun in any of this stuff. There were some allegations around, around the altar.

A (24:28)

I don't think they smoke very much at all. I think it's mostly tequila drinking.

B (24:31)

That's true. All in tequila.

A (24:35)

Although, although JCAL does tote a gun regularly.

B (24:37)

Oh, yeah.

A (24:38)

So maybe that's the smoking gun.

B (24:39)

He's a Texan.

A (24:40)

Yeah. No, I didn't see anything very specific. I mean, it's, it's, it's all in like they're, they are super connected. If you partner with them in some ways, like you would expect to get more of a read on where they're spending time in D.C. what they're seeing. That seems like there are clear lines on what you can share, like what, what turns you into a lobbying firm and what doesn't. I think they've stayed out of becoming a lobbying firm and so they have clear, clear rules on that. Yeah, I think Boz distilled it pretty well before we read his post. Let me tell you about adeo, the AI native CRM. Adeo builds scales and grows your company to the next level. Boss said, I don't know David Sacks, but I want more expertise in government. Experts tend to have made money in their area of expertise, in their area of expertise. If people can't have history or friends in a field before leading it, then our leaders won't know anything. And I thought this was a good distillation of like the core debate about, like, should you have someone who has never participated in an industry overseeing it, or should you like someone who's purely academic, purely outside of it?

B (25:51)

And I believe there's some readers and probably people at the New York Times that would like somebody that hasn't participated in either industry to be running in a role like that and just blanket against both industries and sort of like hold them back.

A (26:07)

So the reaction is interesting in the comments. I mean, first the top comment is somebody beefing with Boz over how he ran the Quest Store. It's like, clearly a VR aficionado who has an axe to grind over niche VR policies. But the second post is what I want to get to because it actually addresses the core claim here. And Alex says the construct you're thinking of is called a council. It's been used for a long time to allow the elected with limited knowledge on a domain to get a consensus of options from a range of experts. This minimizes conflicts and prevents kleptocracy. But, like, isn't that what a czar is? I thought Sachs was a counsel. Like, he's not an elected official. Like, the elected official is Donald Trump, the president, and like, there's a variety of folks there. And then Sacks is, like, appointed to this czar role that is just to give his, like, his, like, he doesn't have the right. He doesn't have the ability to just like, create legislation out of thin air. Right.

G (27:08)

Yeah.

A (27:09)

He is very much a California.

B (27:10)

I was trying to look up the history of czars. Right.

A (27:13)

It is weird. Like, have we always had czars? I know there was a whole thing.

B (27:15)

About border czar czar was Bernard Barak, appointed by President Woodrow Wilson to head the War Industries Board in 1918. The press dubbed him that industry czar because he had sweeping powers to coordinate wartime production. During World War II, President Franklin D. Roosevelt appointed several czars to manage the massive wartime economy, including a shipping czar and a synthetic rubber czar. These roles were synthetic rubber czar, one of the most iconic people are stoked for that. People don't talk about the need for our ongoing need for synthetic rubber czar.

A (27:50)

No.

B (27:50)

These roles were essential because existing government bureaucracies were too slow to handle the urgent demands of total war during the Nixon era. The modern concept of the czar, a policy specialist with a specific portfolio, solidified under Nixon. During the 1973 oil crisis, Nixon appointed William Simon as the energy czar to manage fuel shortages. He also had a drug czar during the sort of like, beginnings of the war on drugs. So anyways, again, I think unless you're just blanket against these industries, it's hard to argue that you want somebody that doesn't have any expertise in said industries.

A (28:41)

Yeah, some of these claims here, here's one it's sort of hard to track. Like, so he says, free from those. This is from the New York Times from the actual article for the screenshot. Free of those restrictions, Mr. Sachs flew the Middle east in May and struck a deal to send 500,000American AI chips, mostly from Nvidia, to the UAE, the United Arab Emirates. The large number alarmed some White House officials who feared that China, an ally of the Emirates, would gain access to the technology, these people said. But the deal was a win for Nvidia. Analysts estimated that it could make as much as 200 billion from the chip sales. And so, like I, we've covered the debate around export controls and should Nvidia, where should Nvidia be able to sell things? But it's never been an open and shut case in my mind. It's never been like, oh, it's so obvious that the UAE is completely off the table. Yeah. I don't know.

B (29:39)

I mean, it was also just like painting the friendship between Sachs and Jensen as like something that felt wrong. Was.

A (29:48)

Yeah.

B (29:49)

Was a little bit rough considering it's the most valuable company in the world.

A (29:53)

Yeah.

B (29:53)

One of the most important AI companies. Potentially the most important AI company if you just go by weight in the. In various indexes.

A (30:05)

Yeah. I don't know. I mean, it's like it's clear that he doesn't have Nvidia bags directly. Like, that's completely debunked. So you have to do these like 25 different steps to get to some sort of conflict. It's a lot of like.

B (30:17)

You know, I read this and I think like this is if you're. If you're the average New York Times subscriber.

A (30:23)

Yeah.

B (30:24)

This is probably that you were. They were probably like very excited by this story, right?

A (30:30)

Yeah, I mean a lot of. I think a lot of people are definitely like, yeah. Just riled up by the all in podcast.

B (30:37)

Charlie in the chat says all in pod about to be an all timer after this article. Do you think it's possible that David and Jason coordinated to get this hit piece done to grow all in even further? They said we're at such an insane stage. Oh, yeah, that was a great thing.

A (30:54)

Yeah. Jason. Jason said a bunch of. I mean, Jason made a lot of good arguments about this, but one thing was he was like, we would be smaller if we. What was it? He was like, we would be bigger if we didn't talk about politics. And that seems crazy to me. I feel like politics is like the ultimate TAM expander in the history of podcasting and media. Broadway.

B (31:14)

The audience for political content is like 10 times.

A (31:17)

I would think so. I do believe that Jason loves talking about tack and I think he's an og. He's an og. He's said that multiple times. But I would be shocked if politics was not a TAM expander for podcasts broadly. And then the other thing is that he said they lost money on the all in events. I don't know how that's possible. And those events, obviously they're big budgets, but I would imagine that the sponsors and the ticket sales, they're not cheap tickets. Right. I would imagine that they'd be making money off that. I certainly hope so. I mean, they've been running this thing for five years. It's incredibly valuable in the ecosystem. They should be able to capture some value there.

B (31:57)

Maybe they set up their own data center to sort of manage.

A (32:02)

Yeah, we decided to bring it back. Podcast production on Prem and we ordered a lot of 100,000.

F (32:07)

Black.

A (32:09)

Really has us by the balls.

H (32:11)

It's rough.

A (32:12)

Martin Shkreli here says the Sacks piece illustrates the exact problem with the New York Times. Voters specifically want this type of person, not a bureaucrat who has never worked a real job. Lina Khan, Case Street.

B (32:22)

Yeah, that's so the issue and the reason I think this article was written is that New York Times subscribers specifically want this type of article.

I (32:32)

Mm.

A (32:34)

Yeah. Yeah. Whiskey Titans going back and forth here. Did you miss the entire point of the article? This isn't a quote. We can't have businessmen in government. This is a. We can't have the government officials who host government summits and sell access to the President for $1 million via their podcast business. And Martin Shkreli says, I doubt it was Sachs who wanted to sell $1 million passes. And whiskey Titan says, I agree with you. I'm sure it wasn't. But letting Jason run rampant until Susie Wiles steps in isn't a great look. I happen to think Sax is doing fine at this particular role, but I also understand the general public feelings, like there's a lot of graft. The New York Times isn't the right conduit for that argument, though, and they're going back and forth. The timeline truly is in turmoil over this. Dan Premack had a good take. He had a whole breakdown of this, which I think was interesting. He said, let's kick this off. But first let me tell you about fall build and deploy AI video and image models. They're trusted by millions to power generative media at scale. So Dan Premax said, lots of people are sending me the New York Times story on David Sacks. Outside of the all in sponsorship proposal, which feels oblivious at best, corrupt at worst, I'm not seeing much in there that's new, at least to those who've been following. Dan Pramack says, as an aside, it's true that Saks Kraft still have a ton of AI investments. Thing is, all tech investments at this point are AI investments. It's kind of like Internet investments at this point. If you invest in tech startups, you de facto invest in AI startups. And Jason says, we lost money on the event. The NYT knew this and deliberately published false information. And Dan Primack says they included statement that you lost money on it. What did they print that was false? They were somehow that we are somehow making money in this or some gain. And Dan Primack says just reread. Just reread doesn't claim that all in made money said you tried to generate revenue via $1 million sponsorships, including for VIP reception that didn't end up happening. But ads that you don't know what but ads that you don't know what sponsors ultimately paid or that it doesn't know what sponsors ultimately paid included the statement that you lost money. Am I missing something? And Jason says Mr. Sachs has raised the profile of his weekly podcast all in through his government role and expanded its business. Confused? I thought you were talking specifically about the White House AI Summit pieces. Dan Primek talking in general. Don't know how you would would not quantify SAC's role in white House def raised all in profile at least among normies. As for role in biz expansion, guess you could stake your claim there. I completely disagree with this. I feel like the all in podcast put the White House on the map. I feel like a lot of people were like they found out about the White House and about the US Government.

B (35:36)

Which house?

A (35:37)

Exactly, Exactly. Because of the all in podcast. They were listening all in podcast and they were like, wait, wait, wait, you're.

B (35:43)

Telling me there's people in Washington D.C.

A (35:45)

They run this whole country. They create, they're in charge of the rules.

D (35:49)

Yeah.

B (35:49)

They create sort of laws and framework for how our country should operate, which industries we want to support and grow.

A (35:59)

You're telling me, you're telling me that there's a group of people and one of my besties is one of the best. This is amazing. I got to learn more about this. I got to figure out what a bill is. I got to tell how a bill turns into a law chat.

B (36:12)

What is a bill?

A (36:14)

Jason says if anything going deep into politics has been a net negative for all in at least in my opinion we would, we would be growing faster and wouldn't have lost some percentage of our left leaning audience if we'd stuck to tech, markets, science, VC, etc. That's an interesting take. I, I still think that politics made it so important. It made it so big.

B (36:34)

Well, yeah, and it made the content polarizing. But I think that polarizing in media is good. You actually get more attention. Not necessarily good from all points of view, but good from a pure, just like reach.

A (36:49)

I mean, yeah, I was looking at the I Think the ratings or like the amount of viewers for, for like CNBC is bigger than Bloomberg by like a pretty significant margin because Bloomberg's like extra wonky and CNBC's a little bit. I mean, it's literally called consumer business News. Like that's what the C stands for, I believe. And then you have Fox, which is even more like Fox News is political and it's much bigger ratings than CNBC or Bloomberg. And then ESPN is like by far the biggest because it's like sports. Everyone loves sports. And so like maybe that's the final form. They should go full poker and then full sports become sports center competitor.

B (37:34)

I could see it. It might be the way AB says, I only learned about Trump because Chamath endorsed him.

A (37:40)

Yes, exactly. I had never heard of this guy. Who?

B (37:46)

The vodka and social media entrepreneur.

A (37:49)

He's running for president. Okay, so Dan Primack is weighing in again. Concluding it. He says the New York Times story was mostly a nothing, bur at least for those familiar with the situation. As for hoax, the story itself as published isn't being disputed. Obviously the New York Times had info questions that Sachs lawyers answered and disproven info wasn't included. That's how journalism works. The real complaint seems to be about the headline quote, silicon Valley's man in the White House is benefiting himself and his friends. I get the complaint, but that's not. But it's really a matter of interpretation, not true, false hoax. Imagine if you had a friend and they went to the White House and they didn't try and benefit you.

B (38:32)

You'd feel like you might not be friends with them.

A (38:34)

You might not be friends with them anymore. Sachs and the Trump White House are pursuing let them cook AI policy. I like that. That they believe will help us win the AI race and that the rewards outweigh the risks. Others disagree. Yeah, this is so true. It's like there is no, like, oh, we now know the correct way to win the AI war. Like, we know that there's a correct way. It's very obvious. No, everyone's debating this constantly, even inside of tech. And Sachs has one view that I think has actually played out pretty well considering that he's been anti doomer, anti fast takeoff, more industrial capacity, more opportunity to grow gdp. There are some elements of his takes that are a little bit more like TBDs, like what actually happens to jobs over the long term. How does it manifest in GDP growth over the long term? But so far I think he's been correct and I think that's What Dan Primek's saying here, he says only time will tell if Sachs is correct. What we know for sure though is that his deregulatory policies should help VC funds his those runs by his friends, those run by strangers, et cetera. Thus the headline is defensible, albeit pushing an agenda. And that's the timeline in turmoil folks. Let me tell you about graphite.dev, code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. I can't read this. AI Amblicus this name, this Ilia interview will be compulsory viewing for any future student trying to understand what misallocation of capital looks like in real life. See, I completely disagree with this take. People were going back and forth on this. We talked about this a little bit over the holidays, but fleeting in bits says. Can you say more? Just that he doesn't have any business direction or something else and the original poster says these are my intuitions but for what it's worth on the micro level, he just seems to drift in a sea of possibility and not the kind of person.

B (40:30)

See, I originally read this as. The misallocation of capital that I've seen is like the, the 10th, 11th, 12th foundation model lab that has like 100 to a billion dollars that is just like kind of iterating on what Ilya already worked on, already developed. Right. And doesn't necessarily like if they just do, if they just create a model that's like not state of the art, like I don't know that there's going to be incredible value in that. Meanwhile I'm like okay, you take the guy that whose work led to ChatGPT and you give him a few billion dollars and let him continue to iterate and he's not just firing a single multi billion dollar cannon and hoping he hits a target. It's like this incremental research that I think is still one of the best shots at developing the next paradigm. Whatever comes after LLM. So I read this, I think that you're. Your reading of this was right but I initially read it the other way and I was like yeah, I do think this is, you know, somewhat bearish on the, on the incremental large language model lab.

A (41:42)

Yeah, I don't know, I mean I can kind of steel man both like we're going to have Julian on the show in when is he coming on at? Or we're having Vincent from Prime Intellect come on the show and I was talking to him. We'll get more information from him, he's going to be on at 140. But Vincent was explaining that more and more companies and different business processes, they do need specific training runs, they do need the skill sets of a foundation model lab. But there's a lot of business to be done that's not purely AGI seeking, not purely paradigm shifting. So I do think that there's some value there if the business can be run well, which is a big if. But there is a path where a thinking machines or one of these companies is going to go and do specific reinforcement learning, specific model development for a specific company and task that can work out. It's a very different business than searching for the next paradigm doing science. And maybe you shouldn't even call it a lab because you're not really even trying to do foundational science necessarily. You're more productizing company. Yeah, it's a business, which is great. We love that. What's interesting about Ilya is that but when we talked about this, it is a venture style bet. Let the scientist go, experiment, maybe it will work out. It's extremely high risk, probably a zero, but if it works, it's huge. The expected value is still high. What's crazy is that we're doing a venture style bet at growth scale and it's just massive amount of capital for something that I think the consensus here is that it's either he solves it and it's incredibly valuable and leapfrogs everything and is just amazing and or it's just you do get lost and you get lost in the sea of research and ideas and you never really produce anything. So I love the high risk bets. I just understand why people are saying, well at that scale, that's a lot of money. That's a lot of money. But that has been happening internally at Google for a long time. They probably burned a lot of money on research projects. Hasn't been that big of a deal because they had the engine for it and if the investors are significantly diversified, they should be fine. Anyway, what else is in the timeline today? Fin AI, the AI that handles your customer support. The number one AI agent for customer service. We did get a good meme. We got a couple good memes.

B (44:10)

Cody says when my wife asks what we should eat for dinner but says no to my first two suggestions.

A (44:16)

We are back to the age of research. I like it. And then when she asks what I want for dinner from Bazelord, the answer to that question will reveal itself. I think there will be lots of possible answers. Very true. It's a great new meme template. I like it. When my husband asks how many Amazon packages are still on the way, the answer to that question will reveal itself. I think there will be lots of possible answers, but I think that's actually true. If he creates some new AI, there's a bunch of different ways to monetize it. We know this is. We know this is a fact. But of course Ilya is now joining the ranks of Yann Lecun and Rich Sutton and Andre Karpathy of sort of industry legends that are more or less saying that scaling is over and LLMs are dead. You know, on the other side, Sholto is saying scaling maybe not over. So we'll see.

B (45:13)

This is what this post is. Yeah, this post is great. Scaling is over and LLMs are a dead end. Aw, you're sweet. Scaling is over and LLMs are a dead end. Hello, Human resources.

A (45:25)

Human resources. I love this meme template because it's like, yeah, Jan Lecun has been saying the same thing.

B (45:31)

Jan says, for the record, my current BMI is 24.

A (45:36)

This guy rocks. Very funny.

B (45:39)

I thought he would have dropped the meta tag on X by now, but I guess he's still.

A (45:46)

Oh, he still used to wrap and them, didn't he leave? Okay, he's like on his way out, more or less. Another 1 billion to SSI. There's a bunch of this in the SSI bucket. Let me tell you about Profound. Get your brand mentioned in ChatGPT. Reach millions of consumers who use AI to discover new products and brands. Of course, we are having Dylan Patel on the show in 12 minutes and we should do a little bit of a run through of the drama on the timeline. The timeline was in turmoil. Lots of people very upset with semianalysis Latest Post how dare you. How dare you take a video. They took a swing at the King, which was the name of their article. They said tpuv7 google takes a swing at the King. The King is of course Nvidia and they are asking is this potentially the end of the Cuda moat? Anthropics. They're talking about anthropics. 1 gigawatt TPU purchase the more TPU meta ssi xai OpenAI anthropic buy the more GPU capex you save next generation TPU V8 and they're going into what the battle between TPU and the next generation GPU out of Nvidia will look like. And this, this upsets some people. There's a lot of folks who are long Nvidia either they have invested in Nvidia, they made a lot of money in Nvidia, or their their whole business is tied to Nvidia or AMD even and so or they bought the local top a month or potentially there's a whole bunch of reasons. You could also just disagree with this and you could just think that you know that semi analysis their takeaways are wrong. But I think it's a thought provoking article. I think there's a lot of data, they're extremely thorough and I think that they do leave you with a lot of new information that you can, you know, do with what you want. And I think in general the response to this article was very positive, but there were some folks who were very upset by it and went and went.

B (47:56)

All over the place and on accounts that put a noun and then capital as their name.

A (48:02)

Yes.

B (48:03)

And suddenly they're an expert on everything.

A (48:05)

Yes, yes, yes. Yeah, it was a little odd seeing the credentialism come out from the anons because like I don't think we should get in the two can play that game camp. It's a little bit rough, but there's a little bit of interesting stuff in here. I want to read through some of this. Let's kick it off with the opening of the semianalysis article. The two best mod the world Anthropic's Claude 4.5 opus and Google's Gemini 3 have the majority of their training and inference infrastructure on Google TPUs and Amazon's Trainium. Now Google is selling TPUs physically to multiple firms. Is this the end of Nvidia dominance? The dawn of the AI era is here and it's crucial to understand that cost structure of AI driven software deviates considerably from traditional software. Chip microarchitecture and system architecture play a vital role in the development and scalability of these innovative new forms of software. The hardware infrastructure on which AI software runs has a notably larger impact on CapEx and OpEx and subsequently the gross margins, in contrast to earlier generations of software where developer costs were relatively larger. Consequently, it is even more crucial to devote considerable attention to optimizing your AI infrastructure to be able to deploy software. Firms that have an advantage in infrastructure will also have an advantage in the ability to deploy and scale applications with AI. And they say we've long believed that the TPU is among the world's best systems for AI training and inference. Neck and neck with King of the jungle Nvidia 2.5 years ago we wrote about TPU supremacy and this thesis has proven to be very correct, TPU's results speak for themselves. Gemini 3 is one of the best models in the world. And there's a very funny bit in here I need to find it Saving. Oh yeah, here. So this is a very spicy line in here. He says OpenAI hasn't even deployed TPUs yet and they've already saved 30% on their entire lab wide Nvidia fleet. This demonstrates how the perf per TCO advantage of TPUs is so strong that you already get the gains from adopting TPUs even before turning one on. And so basically what he's explaining is that because of the competitive dynamic between Nvidia and Google with CPU now you can use TPU as a stalking horse and say, hey, if you don't cut your prices, Nvidia, we know that you.

B (50:38)

Have really high margins or not even cut prices but encourage an investment.

A (50:41)

Exactly. And so that's what they're.

B (50:45)

Nvidia would rather invest back into your business instead of cutting prices.

A (50:49)

Yes, and so says we think the more realistic explanation is that Nvidia aims to protect its dominant position at the Foundation Labs by offering equity investment rather than cutting prices, which would lower gross margins and cause widespread investor panic. Below we outlined the OpenAI and anthropic arrangements to show how Frontier Labs can lower GPU total cost of ownership by buying or threatening to buy TPUs. And so, so OpenAI Nvidia, it was $22 billion per gigawatt the rest of the system. So it's a $34 billion billion dollar per gigawatt expense to Nvidia. But Nvidia is doing effectively an equity rebate of $10 billion per gigawatt in investment. And so how that works out is a 29% partner discount. Anthropic has similar math, but a little bit higher at 44% partner discount because Microsoft is paying for a P. And so it's an interesting thesis and it's unclear exactly like, well, if the claim that the investors will panic if it was actually just lower gross margins. Well, if you say the quiet part out loud like this and you do the math to show that there is basically a discount that margins might be coming down because of competitive dynamics, does that wind up resulting in investor panic? I mean, certainly it didn't today. Isn't Nvidia up today, Right? Yeah, Nvidia is up 1% adding a casual, you know what, 10 trillion or 100 billion or something. 10 billion quadrillion. Yeah, gigajillion dollars.

B (52:35)

Yeah, just, I mean, and again, we said this earlier on the show, but Broadcom is down almost 4% today, which I would have expected it to be the other direction given that to actually buy TPUs physically you need to go through Broadcom.

A (52:52)

Yeah, yeah. So a lot of people are going back and forth on, you know, can semianalysis be trusted? Because they're writing about, you know, Nvidia and Dylan, I think some people didn't understand that he was joking. Zephyr here has a post. Dylan is being tongue in cheek, but he's not wrong. Nvidia was extremely dominant for the last three years as we saw in the stock. It's up 10x over the last three years. New competitors will cause a reduction in market share and margin compression. But TAM is big, so revenue profits won't go down. 75% of GM is just unsustainable. Hyperscalers will also use the cheap TPUs threat to extract better deals from Jensen. Priority access for Rubin, Feynman or discounts on GPUs. Jensen called Altman and initiated the $10 billion deal after he saw the information about the information article about OpenAI testing TPUs. And so this is in reaction to that, that that point about OpenAI hasn't even deployed TPUs yet and they've already saved 30%.

B (53:59)

There's a decent post here from just another pod guy. They say Dylan speedrunning through all the learnings of sell side research, industry capture, pissing off IR execs, gatekeeping info based on client tier difficulty scaling beyond single star analysts, distorted MSN representation of your notes, eventually spending too much time marketing versus researching amazing biz content. Though obviously Dylan would push back on a lot of this stuff. If you actually read through the entire article, there's nothing in the article should actually in this article should be that surprising because so much of the article is just referencing old semianalysis research. Some of which they did sort of before the paywall, some of which they did under the paywall. But it felt like a kind of a culmination of everything that they've been saying for a really long time. And I think that part of. I think the surprise here is just how much faster this conversation has really come to a head than people may have expected. I think that at least like surface level on the timeline, I think people felt like the TPU threat was maybe like a 2026, 2027 conversation versus being like it's a part of these buying discussions right now and negotiations.

A (55:19)

Yeah, yeah. The other buried lead in the article was of course about pre training. So there's A snippet in here. OpenAI's leading researchers have not completed a successful full scale pre training run that was broadly for a New frontier model since GPT 4.0 in May of 2024. And you know this, this is, it's so interesting that this, like if this was wrong, you would imagine that there would be a whole bunch of reaction from OpenAI people or like proxies or surrogates, right? People quote trading and be like that's just not true. Wow, something else is cooked. But the fact that I haven't seen anyone respond to this and say like, oh this is wrong, like we actually did not that like that's the North Star for what the business is like. The business's job is to create profits, it's not to complete successful full scale pre training runs. That's not the goal. That's just something that they might do in service of making a better model, making a better product. But ultimately it's whatever the customers want. And if the customers are happy with 4.0 level base pre train and a bunch of reasoning on top, that's fine. So what else is in the back and forth? People are also, I mean really, it does make me happy that we didn't go deeper into ranking people because it does feel like when you create a list of tiers and rank a bunch of people, you're just creating a big bucket of enemies down at the bottom of like people who want you dead because you rank them low. But I'm sure we'll get into the discussion of Cluster Max and what how people are interpreting clustermax because there's a whole bunch of ways to read it. Like one way to read it is like which stock should you buy, right? But like that's not necessarily the read. The other read is like which product is the best to work with as a customer. But it's like what customer are you? There are some that are in the lower tiers that are fantastic for very specific use cases like this is the nature of every business. Like one of the, one of the Neo clouds that was particularly upset with Dylan is in a very niche market. But if you're in that niche market, it's probably a great product. It's probably great for you. If you satisfy this specific list of criteria and you don't need these features, you're probably fine then. But it's a lot of fun. People are going back and forth. They're also debating whether or not Dylan is independent, given that he lives with Sholto from Anthropic and we gotta ask.

B (57:59)

Him why he has roommates. I'm not even concerned about a conflict.

A (58:03)

It's roommate gate.

B (58:04)

Yeah, it's roommate gate.

A (58:06)

But what about this other tinfoil hat post from Jukon? My theory is that Meta deliberately leaked the story to the information about Google's about acquiring Google's TPUs for meta. It's a classic risk free power play. The moment Jensen Huang reaches wind, catches wind of Meta using Google Silicon, Nvidia is likely to rush in with an investment. They might even be negotiating as we speak. This allows Meta to secure capital and shift from burning their own cash to potentially getting discounts or effectively buying Nvidia chips with Nvidia's own money. Plus, if they actually do secure Google TPUs, they solve their compute shortage. It covers all bases. I wonder when other hyperscalers will catch on to this magic wand. All you have to do is hint at using TPUs.

B (58:50)

Okay, but the issue, the issue is, is how many red flags would be waving if Jensen was like, yeah, we're investing $20 billion in Meta. We're very excited about, we're very excited about Meta and owning a piece of.

A (59:07)

Yeah, that seems very, very odd.

B (59:08)

So he's in a position where. I don't know what kind of leverage Jensen has around in those conversations with Meta because. Because he doesn't want a discount. And it's not like OpenAI where you can just announce an investment or an anthropic, et cetera. So how does any type of rebate actually happen is the question.

A (59:30)

Yeah. Well, before we bring in our next guest, let me tell you about TurboPuffer, serverless vector in full text search, built from first principles and object storage. Fast 10x cheaper and extremely scalable. Let's read through some more TPU stuff to set the table. So Clive Chan says, I keep seeing stuff about tpu. Has anything materially new happened? There's no evidence Google has ever trained Gemini on non TPU hardware going back to pre GPT models like BERT. TPUs predate Nvidia's own tensor cores, anthropic and character and SSI and midjourney have long used TPUs. I'd be surprised if Meta weren't looking at them. Nvidia's moat has never been deep for the big labs. See OpenAI deciding it could do better than Cuda and investing in Triton instead regularly edging out C U D N N on benchmarks. There's nothing magical or structural about any of this, just good engineers doing good work. TPUs are not that much more efficient than GPUs. And small performance per watt difference are dwarfed by whether Meta has the right kernels and systems engineering talent to pull it off. Both Nvidia's and Google's moats are small. And we are still at the point where individual good engineers can flip the entire balance. Why was this not priced in? This is all super old public info. I have a feeling that this Clive Chan, who I guess is over at was it Tesla and then OpenAI is a little bit of like, first time in the public markets, first time realizing that the people who trade this stuff are not necessarily like, on the super inside of the labs, actually understanding the decisions that are being made inside the labs. Like it's a completely separate ecosystem. And that's why organizations like Semianalysis exist. And I believe we have Dylan Patel from Semianalysis in the Restream waiting room. Let's bring him in. Dylan, how are you doing?

F (61:24)

I'm doing fantastic. How about yourself? You know, I saw. I saw the meme image that you guys put out there for me, so I had to wear. Let's go.

A (61:33)

Let's go.

B (61:34)

Dude, we need a bigger, bigger screen for that bicep. We'll work on it.

A (61:40)

Let's go.

B (61:42)

Where in the world are you?

F (61:44)

I'm in Florida spending Thanksgiving with my family here. I'm trying to chill out a little bit. It's nice to have the family pamper me a little bit because I broke my foot a couple weeks ago.

A (61:54)

I'm sorry, How'd you break your foot?

B (61:56)

Tripped over a TPU family.

F (62:00)

Family reunion playing football in Texas. We're American, as we can get.

A (62:03)

There you go. Well, we were just running through a little bit of the TPU article. Can you actually set the table for me on, like, what do you think is new about it versus what has Semianalysis already been saying? And this is more just like tying everything in a bow.

B (62:23)

Yeah. Half of the article is just referencing recent.

A (62:25)

We've been saying this for two years. We've been saying this for one year.

B (62:28)

And even referencing Google's own content about the tpu dating back. Back even further.

F (62:37)

Yeah, I would say the majority of this piece was if you're a client, it's already been pretty much all published. But it hasn't been tied together. It hasn't had a narrative around it. Right. Because when we think about what we put out on the paid side versus what we put out on the newsletter, our clients sort of get what changed, what happened. Here's the numbers. That's about it. We don't explain the technology that much because our clients are sophisticated. They're either in the industry or they're finance bros who don't give a shit about the technical stuff. And so it's either of those two. Right. And so we're just explaining, here's what's happening, here's the change, here's the numbers, right? So for months we've been saying Google's selling TPUs. For months we've been saying, hey, here's TPUv7 versus Blackwell. We've even put out updates on here's what we think TPUV 8 is versus what we think Rubin is. And so generally it was making it into a narrative and explaining the technology and the corporate, I would say politics or dynamicism around it. Right. So that's, you know, I think, I think there has been bits and pieces put out by other, other folks. Right. I think the information has done great reporting on some of the stuff after we did. But in the public space, I think, you know, for, as an example. Right. Like so, so other people have put out bits and pieces surrounding this, but they haven't put out the full picture. So, so as far as like what's new, it depends on where you sit in the stack. But anthropic and meta and folks like that have been talking to Google about buying TPUs for many months. Right. Whereas people externally are. Last week, when Gemini 3 was launched or two weeks ago, people were just learning that TPUs are trained, training Google's models, right? So it's where are you in the information spectrum?

A (64:17)

Right, yeah, totally. So on that information spectrum, the finance bros, they can probably just like if they read into this, oh, bullish Google or bearish Nvidia or whatever, they can kind of trade in, in and out as they please. But on the more technical side, like are people using semianalysis research to understand, like, okay, I'm a NEO cloud, what do I want to rack for next year? Maybe I need to be putting in a TPU order. Is that how people interpret your research? Like what happens on the technical side of the house?

F (64:50)

Yeah. So as far as like some of the paid stuff we do, we have one model called the TCO model, right. Which is calculating the TCO of all these different hardware performance, building up the entire cluster cost, breaking it down to like a dozen plus different things, whether it's storage or networking, and breaking down the cost of everything. So there we put out research on TPUs because as soon as NEO cloud started getting offered, hey, you want to buy TPUs we're like, okay, we need our own ground up model. So when you're negotiating a big contract, what you do is called it should cost. Right. You go and calculate what it costs for the company versus what it costs for me to deploy. And then you think about like, oh, what's the margins they have? What is ridiculous to offer them, what is not? Right? Because everyone always wants to know like, hey, what margin are they making off of me? Can I push that down a little bit? What is ridiculous to demand in a negotiation versus what's not? So we've already been working through this TCO model. We've put out four different updates on the TCO of TPUs V7 and V8 because there are Neo clouds out there as well as labs who are purchasing TPUs that are using that to understand what's the cost. Now, you know, anthropic, I will say, just already knew and figured it out because they've hired so many Google people. But other labs are also looking at it, right?

A (66:00)

Yeah.

F (66:01)

And so you know, when you say, hey, on the cost side of things, on the technical side of things, right. There's a lot of network engineers now out there who have never deployed Google hardware that are now like, okay, I need to figure out how to do this techs. Right. Like, you know, so there's people who have DMed me that are like, oh, as you know, we've been thinking about deploying Neo clouds, but your material on this, this is technically better and teaches me more than Google's own material. Right. So it's like this is helpful to people on multiple factors.

A (66:27)

Yeah. What about the software side? Google's built their own internal stack to compete with Cuda. How much of that are they going to actually give to their customers who are buying tpu? Because that feels like you, it feels like potentially you could over rotate on. Oh well, Gemini 3 is really good. But why is it good? Is it just because of the hardware or is it also Google's incredible prowess, multi data center training, all this fancy stuff that they have that they won't be giving you when they sell you the tpu.

F (66:58)

Yeah. So that's the interesting thing is some of the software will remain closed source, but you can still use it.

A (67:03)

Okay.

F (67:03)

Right. And then some of the software they are trying to open source aggressively and then some of the software they're just never going to give out there anywhere.

G (67:10)

Right.

F (67:10)

So it sits in three kind of buckets.

B (67:11)

Right.

F (67:12)

The interesting, I guess, newer thing that we did in the piece was we looked across all these different open source AI software, whether it's Pytorch, whether it's Vllm, whether, you know, all these different open source libraries. And we calculated and counted up how many Google commits there were, right? And you can see there's a chart in the article where the number of commits that Google's doing on TPOs has exploded over the last handful of months, right, as they've decided to, to shift their strategy, sell TP's externally. They also recognize software has to be open for this, right? Only the giga brains that Anthropic can figure out how to do everything themselves. It's those people outside of anthropic types that need a bunch of open source software that builds on top of it. And what's interesting is when you look at, hey, Nvidia, the biggest argument that Nvidia doesn't really make for GPUs, but they should is that about 40% of the software that's open source sourced is actually just from China, right, on Cuda. And that's the Cuda mode, right? It's like 40% of the software is just like open source stuff, whether it's people committing to VLM or Pytorch or all these other libraries, right? Bytedance, open sourcing stuff, deep seq, open sourcing stuff. And Google, you know, they don't have people open. Anthropic is not going to open source software. So Google needs to catch up, not just by hey, here's all the software we have internally, let's open source it. They also need the ecosystem to build a ton of software on top of tpu. And so that's the real big challenge there. And there's an element of software there that Nvidia is happy to open source and customers of Nvidia are happy to open source that Google will never open source because it's, you know, Google Cloud is selling the tpu, Gemini is the one actually using it and developing a lot of the software. And these two groups are not always going to be aligned.

A (68:52)

Yeah, isn't that like, I mean, what are the other kind of just problems with Google becoming an actual seller of tpu? It feels like there's obviously an opportunity because Nvidia has high margins, there's demand, it's a great chip. But culturally, structurally, Google tries a lot of different things. They have a lot of advantages, but occasionally they fall flat on their face with just like they can't even get an RSS reader out or something like that. So are there other risks to the TPU not really finding its footing for reasons that, that aren't just the laws of physics.

F (69:30)

Yeah. So the biggest challenge I see with them is it's everything is non standard, right. Google, for years, they developed liquid cooling first, right?

A (69:37)

Sure.

F (69:38)

For AI computing they deployed rack scale architectures. Everyone's talking about GB200 rack scale architecture. Google did it first with GPUs, right. But when they did all of this stuff, they didn't give a crap about, hey, this has to go in 5,000 thousand different people's data centers. Right. This has to go in my data centers that I designed myself. So everything is super vertical. The entire liquid cooling supply chain is super vertical. Entire. The racks aren't even the standard width. Right. So when I look at like a data center, it's like the door, the loading bays because they're so much wider, the Google racks are like three times as wide. It's like maybe it might not even fit into the data center like physically like through the doors. So there's like all sorts of random, like. I wouldn't say random. It's Google from first principles, design stuff.

A (70:19)

Totally. Yeah, yeah, yeah. But if you're neo cloud and you're like, like the hot thing's gonna be TPU next year or the year after and I wanna be able to sell into that market, it's not just flip a switch, drop in, replace with tpu, you have to maybe build a whole new building. Like it might be that significant.

F (70:37)

Right. Or like knocked on some walls and then like, you know, I need to go get liquid cooling. Not from Dell and Supermicro and HPE who I've, who service me already. I need to go get it from some random supplier who's only ever sold to Google. Usually they're sitting across the table from like some giga brain engineer who has a team of 20 people working on liquid cooling instead of like you know, my one guy who does liquid cooling procurement and negotiations and like also does procurement of like network stuff.

A (71:03)

Yep, yep.

B (71:04)

There was, there was a, there was a tinfoil hat theory floating around that Meta leaked their tpu, Meta leaked their TPU interest to try to gain some sort of leverage over maybe some negotiations with Nvidia. I don't know if you see any possibility in that, but how do you think those conversations are going? Jensen doesn't want to discount and compress his margins, but at the same time he can't do this kind of like equity rebate thing. If he took a big position in.

A (71:34)

Meta, it'd be very suspicious I totally get the OpenAI investment. That seems like it makes much, much more sense than saying, hey, we're going long. Meta as a $4 trillion company.

B (71:47)

Yeah.

F (71:47)

@ the end of the day, right, like TPUs have like a set of maybe 10 customers. Right. Because you have to be super.

A (71:52)

Sophisticated.

F (71:53)

Yeah. And so what really is challenging here is, is, you know, Meta. Meta looks at the numbers. You know, it's like, okay, open AI, I'm getting 30 off because they're paying, they're investing in me as a result. Obviously they get equity, but they're investing in me. And I get 30 off on these GPUs as a result. Right. Meta, you can't do that. So, so Meta. I don't think that they're just negotiating, Right. Are they just negotiating with Nvidia when they buy amd? No. There any engineers, they're developing all the software they're actually deploying. Llama 405B was exclusively on AMD for a number of months for inference. Right. So when you look across, hey, is Meta just like playing around trying to negotiate? It's like, no, they're looking out for what is best. Right. And META is power constrained and TPUs are currently way more power efficient. META is compute constrained. And TPUs are potentially higher performance per watt and higher performance per dollar. Right. At least that's what we believe for TPUV7 it is. So they'd be dumb not to look at it, right. And they have the time, they have the people, they have the team. Now Nvidia at the same time has to play the game of chicken, right? Yeah, sure. They could discount the pricing somewhat. And because what's funny is Nvidia is more vertically integrated than Google is when selling hardware. Right. Google has to pay Broadcom, who pays tsmc, whereas Nvidia gets to pay TSMC directly. Right. There's this vertical integration challenge where Nvidia could drop the price a little bit and they'll be fine. But they don't want to. Right. You know, the whole point is you charge the highest price possible. And then the last thing is they've got this like, you know, they've got this view about antitrust, right. You, you don't want to cut deals for specific customers because that looks.

A (73:28)

Bad.

F (73:28)

Right. Instead you want, you know, right now Dell pays the same price for GPU as gigabyte as Meta. Now the networking hardware, there's different pricing because there's a lot more competition and Nvidia can cut a lot more there. But on the GPUs themselves. Nvidia's pricing is very fair, Right? Fair in the sense that they're making a shitload of money off of everyone, you.

A (73:48)

Know?

B (73:50)

Yeah. Talk about kind of Jensen's leverage that he has around Rubin allocations as some of these customers start to at least consider.

F (74:02)

TPUs. Yeah. So as far as, like, next year's TPU deployments, it's pretty set in stone for the vast majority of the volume. Right. Anthropic's got a bunch and then there's some sprinkled else elsewhere. But as we go into 2028, where Google can actually ramp, you know, the flip side is Ruben is also ramping. And at least based on our research, looking throughout the supply chain, you know, over a year ago, when OpenAI started their chip team, they poached like 15 Google people overnight, right? In one week. Like someone I knew, I heard Gup was like, oh yeah, I'm joining OpenAI. Then I text like, another three people I know and they're like, oh yeah. I'm also like, what the fuck? So Google's a lot of their best TPU engineers have left, right? They also have a ton left. And so that what that's done is, you know, chip timelines are so long. That didn't affect TPV7, that's affecting TPV8. At the same time, Google's trying to diversify their supply chain, get from not just broadcom, but also MediaTek Tech. And so Google's got a real challenge on TPV8 in that it's good, it's an improvement. But then when you go look at what Nvidia is doing with Ruben, Rubin is so much better because Nvidia is just pedal to the floor paranoid as we have to be the best and we have to be way, way, way better than everything because how much better I am than everyone else is my margin. Right. And so Nvidia. Nvidia has the sort of, like, at least currently, we think Nvidia is going to be so much better that they'll be fine and they'll be able to maintain margins. Right now, things can happen. Ruben can delay or TPUs can delay and the position looks better or worse. Right. There's a lot of unknowns to go through, but as far as like, what is Jensen's leverage is, look, I'm going to make the best hardware and plus my software advantages and I'll be able to continue to be dominant and dominate the market. Right? There's curveballs that could go which is like, oh, Google Software, they could open source enough software that actually their software ecosystem is not far behind Nvidia. Maybe they don't want to. Right, right. Or hey, they could execute everything and Nvidia has a 3, 6 month delay now all of a sudden they're a lot more competitive. Right. And so all these things are still open questions. But Nvidia can play the allocation game as well, of course. Right? Hey, I'm going to give all of the GPUs initially to companies that probably could buy TPUs, but that ends up being all the AI labs and hyperscalers, right. At least like meta, right. And by people that would actually be willing to buy TPUs. And then you end up with this weird situation where okay, well that's like 75% of the GPU market anyways. When I look at the AI labs through the neoclouds, right? When there's Nebius and Iris Energy and all these other core weave and all these folks are deploying for OpenAI anyways, right? This sort of ends up being like, well sure, I could stiff like some people in the allocation, but at the end of the day, everyone who is a potential customer for TPUs is sophisticated enough to be where they were going to be on the beginning of the allocation anyways.

A (76:49)

Right? How are you framing clustermax these days? Is it for customers who want to buy services from Neo Clouds? Is that the primary goal of clustermax? Because I feel like some people look at it and they're like, this is a buy rating, this is a sell rating on the.

F (77:07)

Stock. So the funniest thing is Clustermax V1, the title of it was Clustermax how to Rent a GPU. Because we discussed all of that. And then in Clustermax V1, I believe we put Iris Energy and underperform. Right. At the same time, the research side of the business, we explicitly were like, dude, they've got these data centers, it doesn't matter if they suck at running GPUs, they've got these data centers, they've got this power. If you just value them on a watts per, you know how much money they can make, it's a long. At the same time as like Jordan, right? Jordan, he's running Cluster Max is like Iris kind of sucks. And it was other people in the technical team before him, you know, it's like, it's like Jeremy who's running the data center side and I think he's been on TVPN is like, dude, Iris Energy is a stock, right? So it's like, it's kind of like, you know, it's like what the technical side of the house does versus what the, the research side of the house does. Yes. They talk to each.

A (78:00)

Other.

F (78:00)

Right. Jeremy did ask the team, like, hey, what do you think of Iris Energy? I think it's a log. And the team working on Cluster Max is like, I don't know, like, you know, it's a bad cloud and it's like, that doesn't matter. So clustermax has nothing to do with the stock right. Now obviously there's going to be some correlation with how good is a stock versus, you know, who's going to want to rent from them. Yeah. But at the end of the day, right, like, Cluster Max is the goal, purpose, sole purpose. And what we explicitly say in there, there is it's for people renting anywhere from like, you know, hundreds of GPUs to, you know, right below the AI lab scale. Right, the AI lab scale. There's different considerations, but in that range, tens of thousands of GPUs, all the way down to hundreds of GPUs. That's who we're targeting. Plus we're saying we're giving a bunch of feedback for people to make the cloud ecosystem.

A (78:44)

Better.

F (78:44)

Yeah. The unsung hero between cluster max v1 and v2 is that we moved the bar up. Right. You know, what it required to be in gold, like was, was much more. What it required to be in silver was much more because everyone improved so much. Right. And as, as we continue to like increase the requirements, make it harder.

B (79:02)

And harder, you got to move, keep moving the goalposts.

F (79:05)

Right? People keep improving the ecosystem. And actually, you know, this is, this is the funny thing. It's like clustermax is evil. It's like when I, when we look at the quotes and we've got hundreds of quotes on clustermax AI, all these companies are like, dude, I love this. This one specific bug that this NEO cloud has bad. They fixed it as soon as you wrote about it. Right. Or like, hey, it helped me understand the reliability, helped me understand this or that. People are like, love clusterbacks. And you know, altruistically, like, I think we're generating billions of dollars in value just from, hey, like all these clouds are more efficient and there's less failures and it's easier to get your workload running on any random GPU cloud and the market is more efficient. Now I'm not making any money off of that. How am I making money off of Cluster Max? I'll be very clear. If is people who hire us to do due diligence, right? So people who want to acquire a Neo cloud, people who want to sign a massive, massive deal that's not just like thousands of GPUs, but tens of thousands of GPUs. And then lastly it's people who want to, you know, invest in an E cloud. Those are the three areas where we're making money off of quote unquote, cluster max. But not really. We're not selling ratings. We're not, you know, you know, we're in fact like a customer will do a consulting project with us or want to, want to buy some research from us and I'll explicitly put in our Slack shirts, send an email to the CEO like, dude, just so you know, the people working on this are not the people who are doing cluster max rating, right? You know, the people who are buying, you know, the research on like these data centers are there and this is the power ramp or here's the accelerators or here's the tco. That's not the people doing cluster max, right? And I don't care about, you know, whether you buy it or not. I, you know, at the end of the day, Google and Amazon and Microsoft are way bigger customers and Flip Stack and those kind of companies. And yet some of those are ranked in silver and some of those are ranked in platinum and gold. And that's because technically what matters not. Hey, obviously when we talk about who buys our research, the biggest companies in the world are going to pay me more than the mid sized companies in the.

A (80:55)

World. Okay, question from the.

F (80:57)

Chat. And the price is discriminated based on.

A (80:58)

That. Would you change the rating of a Neo cloud if Sholto promised to do the dishes for two weeks straight?

F (81:05)

Right. You know, there was an argument I saw someone was like, who does the chores? And it's like, brother, we, we live together by choice. You know, we, we pay someone to come once a week if you cook something, you do your own dishes. But like, you know, frankly we're, we're working so much and I think like, you know, I think, I think Dorkash has ordered pizza from the same spot three nights in the row.

A (81:27)

Before.

F (81:27)

Right? Like it's, it's.

B (81:29)

It'S. Wait, so, so question is, is, is being an adult man with roommates.

F (81:33)

Underrated? Did so. I haven't lived with people in years and then when I moved to.

B (81:39)

Sf. So you came back.

A (81:40)

Here? This is.

F (81:41)

Crazy. I moved with, I moved to SF this year. You know, I'm like, oh, you know, I should live with friends just so it's more fun. And the first house kind of fell apart, so I moved into this house with these guys and we've been talking about it for months. I love it. Right? It's like, look, we, we, we, we have, you know, if you think about, oh, what if we all rented our own places that were good and then we pulled that budget together. We have a nice.

A (82:02)

Place. Yeah.

F (82:02)

Right? And then, and in that place, we have plenty of space for ourselves. We pay for someone to come and clean once a.

A (82:08)

Week.

F (82:08)

Right? So at the end of the day, what is the negative here is like, well, we're living with our.

B (82:12)

Friends. It's just guys being.

A (82:14)

Dudes. And the beauty is if you do bunk beds, you have more room for.

E (82:20)

Activities.

A (82:21)

Exactly. Anyway. No, no, sorry, sorry. Actual question from the chat, when is TPU going on Inference Max? We got to.

F (82:28)

Know. So we're working on it, right? We're working with Google technical folks. You know, funnily enough, actually, we triggered a security warning for this Google Engineer. Kimbo went to a JAX conference, right? Jax is the opposite. Is like pytorch, but for TPUs is the most simple. It's Google's own internal thing, right. That people do use externally. He went to this Pytorch or this JAX conference. A Google engineer presented something. He's like, can I get the slides? They send it to him and then Google security alert, like locks him out of his computer because he sent us like a. Some technical like, information. And like for three days, the guy can't work and he's freaking the fuck out. And I'm like, I emailed Jeff Dean. I'm like, bro, this is like, do not fire this guy. He sent me stuff that you presented at a public conference. He's like, oh, okay, yeah, yeah, I'll get that fixed. But anyways, like, we're working with, we're trying to, you know, implement it. We have access to some TPUs. The software stack is different.

A (83:20)

Right? Yeah, you know, just so you have to, so you basically have to rewrite or re implement Inference Max. Like the code that actually.

F (83:26)

Has. I won't say it's that much work as much as completely redoing Inference Max. But there's a ton of work, right? So we're moving as fast as we can. Internal target is this.

A (83:35)

Year. Don't load us to it then. The obvious question is, I feel like Inference Max is my North Star for TCO relative in AMD versus Nvidia Land. There was a bar chart of TCO for GPU versus Nvidia. It looked like, like it looked like TPU was doing really well on that chart. The bars were very low. Where did those numbers come from? Do you have confidence in those numbers or do you think the numbers will change once you actually get TPU on inference.

F (84:04)

Max? Yeah. So inference max shows performance tco.

A (84:08)

Right?

F (84:09)

Okay. You know, it's great, great. Like you know, like guess what, like you know, TCO of like a Raspberry PI is incredible. It's like five bucks.

A (84:15)

Right?

F (84:16)

Sure. You know, versus a GPU is $50,000. Performance divided by TCO is what matters. So that bar chart is saying, Look, TPUs are cheaper and at least on quoted specs. Now let's make some assumptions around utilization. And in the article we explicitly said, look, we don't know what the utilization is. It's going to change customer to customer. Here's a range. Worst case, it's a little bit worse than GPUs. Best case, it's way better than GPUs. Right. And so inference max will tell us what the actual performance is in inference because we don't know yet. Right. Currently the open source software for TPU's is not good enough for us to just take the open source software and say that's the performance. Right. Because that's obviously like not real. Right. Anyone who like is actually buying TPUs is going to spend engineering hours to work on it. And so we're trying to work with Google to get a real performance number that is achievable by people you know and will be upstreamed into the open source software. Because this is an in progress thing. Right. No one cares what TPV7 can do today. It's about what it does in six months. Months. And so you know, obviously we don't want to be, you know, today. TPUs if you're using VLLM, are worse performance TCO than GPUs. Without a doubt. But the target is moving very fast and you know, there's a ton of like low hanging fruit for us to implement before we actually put a number out there. Right. And so where does Google sit there? We'll see. I personally believe the TCO side of things. The total cost of ownership is based on what we know on supply chain. Right. How much do you, how much do the chips cost, how much do the racks cost, how much does the liquid cooling cost, how much does the memory cost, how much do the cables cost, et cetera, et cetera, et cetera. Right. That's based on our estimates up and down So I think the TCO side of things, we're pretty confident it's the performance side of things where we don't know, right. There is a wide range and that's what we sort of tried to state in the article. Right? Performance is a wide.

B (85:58)

Range. Can you explain more about Google and Broadcom's relationship? Max Hodak from Neuralink and Science was asking on the timeline last week, why Broadcom as a middleman, couldn't Google do the design and place the orders from TSMC themselves? But what's your read on that relationship and how durable it.

F (86:22)

Is? Yeah, so when you think about chip design, there's a few different stages, right? There's defining the architecture and then there's actually like implementing that architecture onto a process technology. There's laying out that architecture into gates on the chip chip and then there's like the whole supply chain side of things, right. Negotiating contracts, getting allocations, et.

A (86:42)

Cetera. That takes like 18 months, right? Isn't that like an 18 month.

F (86:46)

Process? Basically, yeah, 18 months or more. Right. I would say actually like Nvidia is faster side and Google's on the slower side just because, you know, Nvidia's been doing it for longer. They have a bigger team. Right. But at the same time, intel has the biggest chip design team and they move even slower than that, right? They take like four years at least. That's what they did a year or two ago. We'll see what the new CEO can get into the, you know, you know, reorgan. Right. But as far as like Google, you know, when they first started the tpu, it was a very few people and they relied heavily, heavily, heavily on Broadcom to do everything right. They just defined the top level architecture. And Broadcom did everything I said below, right. Negotiating with supply chain, figuring out pro, figuring out how to lay out the gates, everything. Right. As time has moved forward, Google has taken on more and more of this right now. They use, you know, they've talked a lot about alpha chip where they use AI to help floor plan the chip, Right. Once you have the architecture, how do I physically lay it out onto the chip? Right. They've done more and more and more there. They haven't taken over everything yet, but that's, that's sort of the point. But Google, Broadcom has this like super big advantage, right? Nvidia, they acquired Mellanox, you know, call it five, six, seven years ago. Huge acquisition. Who's the biggest networking company in the.

G (87:56)

World?

F (87:56)

Broadco. Right. Broadcom is the biggest networking company in the world. And you know, when you talk about AI, it's, it's, it's the architecture of the actual processing elements. It's memory which you're buying from, you know, you know, the memory companies, right? Hynix and Samsung and Micron. And then it's networking, right? When you try and boil it down to the most simple things in software, right? The networking side of things is so important and the, let's say, technical competence of everyone around the world besides Broadcom and Nvidia and networking is so low, or rather it's just not as good as them. They're actually good. But it's like Broadcom and Nvidia are just so good and Broadcom is better than Nvidia in many ways at networking that, you know, when you think about what is Google doing, yes, they're defining how the network topology is, but when you're talking about the physical network serdes, you know, how packets get transferred, all these different things. Broadcom has heavy, heavy influence there. So to this day, right, Broadcom is still charging margins like they did did three, four years ago, even though Google has taken up more and more of the work. But at the same time, Google can't leave until they figure out how to do the networking and supply chain themselves or with a partner. And so what are they doing on tpuva that is potentially a distraction that's slowing down their execution is they're working with MediaTek, right? MediaTek at times has helped Cisco with their network chips. MediaTek has a lot of work on some of this networking stuff. They're nowhere close to Broadcom, right. On revenue, right? For that's one metric on technical competence, you know, that's another metric. I think MediaTek is good, right? But like they're just nowhere close to Broadcom. So now Google is having to work with, you know, I don't want to say subpar vendors, but inferior vendors to.

B (89:30)

Broadcom. And that's just to increase their margin on.

F (89:33)

TPU8. I would even say their angle when they started this project was never we're going to sell TPU's externally. It was, dude, we're paying, you know, a 3x markup to Broadcom. And half the cost of this chip is memory. Like what the fuck are we doing, right? You know, at the same time it's like, well, sure, physically the cost for the networking is not that much, but what value does the networking bring? As you know, sort of Broadcom and then Broadcom's also doing the like game theory, not science of like well, you can't really leave us, so we're going to charge you what we think is fair or what we think we can charge. And Google's like, oh, no, we're stuck to you. Right. So MediaTek is taking way, way, way less margin. They're not passing the memory through.

A (90:10)

That.

F (90:11)

Right. And so, you know, this, this ends up being like, hey, that's a huge advantage for them. Flip side is like, well, they've, they've, they've got to engineer all this work that Broadcom was doing. Instead of working on a way better architecture, they've got to work with a worse vendor. Right. Objectively worse. Although MediaTek, like I said, is very good to try and implement TPUs more directly with TSMC with less Broadcom sort of in the.

B (90:35)

Middle. Very.

F (90:37)

Helpful. And Google, you know, because it's risky, is going down both paths. Right. They're continuing to work with Broadcom on TPV8 and then separate TPV8 project. They're working with MediaTek, right, because they can't risk, you know, find whatever 30 points of margin, 40 points of margin, 50 points of margin. I can't risk the TPU being late because ADS runs on that, Gemini runs on.

B (90:55)

That. Yeah. Can you, can you give any takes on the Nvidia's $2 billion investment in Synopsys that got announced this morning? I don't know if you, you saw it. I'm assuming you.

F (91:05)

Did. Yeah. So in a time where, you know, let's say the two biggest chip makers, Broadcom and Nvidia, are making more money than ever and everyone else in the supply chain and all the hyperscaler are trying to design more and more chips. Everyone's, everyone's sort of working on that. You've got, you've got the, the EDA vendors are at the lowest possible valuations or lowest valuations that they've had. They're still very expensive, but lowest valuations they've had on a earnings multiple basis for a long time time there. And, and, and this is on the eve of, hey, like, objectively, are there going to be more chip designs or less chip designs in five years? A lot, lot more. Right now, the flip side is AI chip design is comp coming. There's 20 plus companies doing AI chip design. It's a, we've got a really long article coming on that soon that'll sort of explain the landscape. But AI chip design is going to shake up.

B (91:50)

Everything. AI. This is AI. AI chip design. Correct.

F (91:55)

Correct. AI helping chip design, whether it's For AI chips or for like power.

A (92:00)

Chips. Okay, yeah, got.

F (92:02)

It. And so, so the question is like, you know, Nvidia, Nvidia has a lot of tools internally, right? The, the dirty. The thing about EDA is that there's three companies that own 95% of the revenue. But at the same time, Google and Nvidia and Broadcom. And all these guys also design a lot of their tooling internally, although they are massive customers of all three vendors. Right? So it's kind of like an oligopoly where the customers also contribute a of lot lot. And so Nvidia's whole goal here is like, how do I get every EDA flow working on GPUs? Because today a lot of it is running on, on FPGA, a lot of it's running on CPUs and AI. AI chip design is going to get a lot more AI influenced. How do I get everything working on GPUs in terms of like the operation of it, even if it's helping people design, not GPUs, right? And, and I don't have enough engineers to work on all the software. They've open sourced a lot of software, right? Like Kulitho, it's software for lithography, right? And they've got all this software up and down the chain all the way from lithography to laying out chips and all this other things. They just want to make it all run on GPUs. And so that's what their goal here is, right? And now they've given Synopsys a huge, huge. They're buying Synopsys at the lowest valuations that Synopsys has ever had with all this cash that they were going to give away in dividends or buybacks anyways. And they're getting Synopsys to, to now make GPUs first class, right? And so I think this is a win win for Synopsys and.

A (93:22)

Nvidia. Well, we could go way longer, but I know what's on your calendar. You got to hit the gym. Thank you so much for coming by and chatting with us. This is really helpful. Have a great rest of your day. Enjoy the holidays with your family. Great catching up. We'll talk to you.

B (93:39)

Soon.

A (93:39)

Cheers.

F (93:40)

Goodbye. See you.

A (93:40)

Guys. See you. Let me tell you about public com investing. For those who take it seriously, they got more multi asset investing and they're trusted by millions. We have Ro Khanna in the Restream waiting room. Let's bring him in to the TVPN ultradome. Ro, good to meet You. Welcome to the show. How are you.

G (93:58)

Doing? I'm doing well. You guys have become quite the celebrities in my district. Everyone is tuning into your.

A (94:05)

Podcast. I'm glad to hear it. I'm glad to hear it. And we're. And we're happy to have you on the show. Thank you so much for. What was.

G (94:11)

That? Are they all in here guys? Are the all in guys jealous? Or do they. Do they respect.

A (94:14)

You? I think that they have left. They have. They're in the stratosphere. They have left the, you know, the.

B (94:21)

Low. The.

A (94:21)

Scraps. Yeah, we're picking up the scraps compared to.

G (94:23)

Them. I think like any great Silicon Valley startup, you're nipping at their. At their.

A (94:29)

Heels. Well, it's funny, we had. We had a New York Times piece about us. It was a very nice, you know, just like, here's what TVPN is doing, just kind of an explainer piece. David Sacks, of course, got a little bit more of the investigative journalism treatment. Got five reporters. We only got one. And so I think that tells you about the relative importance of the shows. But anyway, I'm sure we'll get into that. I would love for you just to kind of.

B (94:51)

Set. Normally when our guest join and they're wearing a suit, we say, thank.

A (94:56)

You. Thank you for.

B (94:57)

Showing. But I think this is. I'm assuming it's one of your daily.

G (95:00)

Drivers. Yeah, well, you know, I'm not back in my district. It'd be. The only way I'd lose my seat is if I started to show up with a suit to, like, the place is there, but in D.C.

A (95:13)

Uniform. That's the uniform in D.C. but I was hoping you could sort of take us through a little bit of the prehistory, since it's the first time on the show. Just explain how you wound up in this position, a little bit of your backstory. And then obviously, there's so many hot topics that I want to talk about in artificial intelligence and tech broadly, and I want your opinion on everything that's going on. But I'd love to kick it off with a little bit of, like, how you wound up in.

B (95:39)

Congress.

G (95:41)

Sure. Well, I am the son of immigrants. My parents came from India in the late 1960s. My grandfather spent four years in jail alongside Gandhi as part of the Indian independence movement. And that really inspired my love of public service. When I came out to Silicon Valley, I had a professor, Larry Lessig. He said, if you care about policy, go out to. To Silicon Valley. That's where the interesting things are happening. That's where the big things are happening. So I went out and I ran when I was 27 against the Iraq war and I got killed. I got crushed 71 to 19, but came to the attention of folks as someone willing to stand up for against the war. And then I worked as a tech lawyer. I supported President Obama. I got to go work for President Obama. And then I wrote a book about what we needed to do to build new manufacturing across this country in 2012, what we needed to do to really have the modern economy in different parts of the.

B (96:44)

Country. You're one of the first American beginning of the American dynamism movement. Would you.

G (96:50)

Say? Yeah, I was gonna say to President Trump, he stole all my ideas in terms of manufacturing. But you.

D (96:56)

Know. Well, that's.

B (96:57)

Good. That's good though. That's good though.

G (96:58)

Then.

B (96:59)

Right? Right.

G (97:00)

Yeah. No, look, I support the American dynamism movement. I'm a fan of sort of what Mark Andreessen wrote in a Wall street op ed about like, how do we not make masks in America? How do we not make basic things in America? When my parents came to this country in the 60s, we were the place to be. We were humming, we were brimming with confidence. Kennedy said to go to the moon. And my first book was about why manufacturing still matters. I think it was a colossal mistake to let China eat our lunch on so many key industries, especially now with rare earth metals and magnets. I mean, we should have a Manhattan Project to do that in the United States or New Zealand, Australia, Chile. But, you know, so I, after my time in the Obama administration, after I wrote this book, I said, technology is going to shape so much of the future of this country. I have a vision of how we can make sure that it helps everyone in my district and around the country. And maybe I have something to offer to Congress. I ran against an incumbent again, lost again. California is a machine dominated state. It's very hard to break in. And I persisted and won on my third.

B (98:04)

Try. There you go. Third time's a.

A (98:06)

Charm. And so for this year, in 2025, how would you frame your top priorities? There's this weird disconnect between that we've been tracking on how relevant is AI. It's so dominant in tech and yet, yet if you talk to somebody at Apple, they'll be like, we didn't want to focus on AI this year at all. We wanted to focus on battery life because that's what helped us sell phones. And AI was actually not a driver of iPhone sales, for example. It's a deeply pervasive discussion point. And yet it's not necessarily. And yet it's widely used, but also widely hated. It's such a unique technology. But just in terms of political priorities, what's been on the top of the stack for you this year?

G (98:53)

Year. I want to answer your question on AI. Obviously in the last few months, what's been a highest priority is getting these Epstein files released. I mean, Thomas Massey and I passed the Epstein Transparency Act. It was My bill passed 427 to 1, 100 to 0 in the Senate, and Donald Trump signed it. Most urgently, it's about justice for these underage girls or a thousand victims who were raped at Epstein's Island. But it's also about this kind of idea of elite impunity that these rich and powerful people, I call them the Epstein class, don't play by the rules, which you and I have to play by. And people are tired of it. And it also is a story of how in the world you get some things done in Washington. How did a Bay Area progressive congressperson end up getting Donald Trump to sign his bill and getting 427 people in the House to vote for it and 100 senators? So that has been the immediate priority. But what I say to.

A (99:48)

Folks. So are you optimistic that the American people will ever get a truly cohesive narrative on the Epstein story, or will it be our generation's JFK.

G (100:00)

Assassination? I'm confident we're gonna get far more than we've had so far. The release is now mandated by law December 19th or December 20th. I think more names are gonna fall. You've already had some high profile names fall because of their affiliation with Epstein, covering up for him or being inappropriate. There are going to be other names that come out now. Do I think that it's going to satisfy everyone? No. There's always going to be some sense that we didn't get a full justice. But it's going to be much better than these women who were denied justice for decades, which was not partisan. I mean, they were shafted by a justice system that didn't work. And there are a lot of rich and powerful people who got away with with it. But look, I, what I tell people is that AI is going to matter even more than anything. And, and to your point about Apple, it's not AI literally as just AI, as Grok or ChatGPT or a technology that detects patterns and can predict the future based on patterns. It's more that AI has become a symbol for a technology revolution. That people know is changing everything about their way of life and the economy and where they feel like they don't have control, that they don't have a full say in what that's going to mean. They don't have a full say in what that's going to mean for their kids in terms of having good paying jobs. And they're unsure if their kids are going to have as good a life as their parents had. They don't know what that's going to mean culturally for them as citizens, are they going to have the same sense? Or they're just going to be manipulated by algorithms and they don't know what that means culturally as their kids are on phones in school and, and becoming sort of creatures with machines. And so this whole concept of how technology is going to be something that empowers people and that people feel comfortable about as opposed to fearful of, is the challenge in my view of our time. And I've gotten attacked from some people in Silicon Valley saying, oh, he's kind of a Luddite. And I was like, no, I'm not a Luddite. And of course I believe I can do a lot of great things in medicine in coming up with new disease and lowering costs. But I don't think we can be oblivious to people's concerns about keeping jobs and keeping social cohesion and making sure their kids are going to have a good economic future. And so I've tried to be thoughtful about how we adopt AI, how we adopt technology in a way that keeps the American dream alive and benefits.

A (102:40)

Folks. And I mean, that's such a wide remit how we adopt AI technology, because you can see it implemented from a chatbot that, you know, some random person uses or, you know, kids are using AI all the way down to, you know, deeper in some, the bowels of some enterprise software product that, you know, no human was ever interacting with to begin with. And then it's just streamlined a little bit with some AI dropped in the middle of some big system. How are you thinking about creating some sort of taxonomy around AI? Do you like a divide between generative AI and more traditional machine learning workloads? Do you see a divide between consumer and B2B applications Self driving versus is what happens in a chatbot? Like how are you thinking about actually breaking apart that problem? Because there's so much there and we say AI?

G (103:36)

Yeah. I would say that the key distinction is is AI going to enhance human capability or eliminate human beings? That is the distinction and that we need to figure out as A society how we get more AI that is enhancing human beings as opposed to just eliminating them. Let me share two thoughts on this both of people who influenced me. Steve Jobs described a computer as a bicycle for the mind. He didn't say computers would eliminate the mind, he just said it would make the mind go really faster and better. And my view is how does AI do the that? And then Darren A. Smoggle, who won the Nobel Prize at MIT has this idea of total factor.

A (104:28)

Productivity.

G (104:28)

Sure. Let me try to explain it simply. If you just had AI replacing human beings and those human beings then becoming not productive, not only would you have frustration in our society. Right. I mean, who wants to just get a check without contributing people at Pride, but you also wouldn't actually maximize total production because you have all these people who could be doing things who are not productive and are not being able to earn a living and spend money. And so what he says is that there is some savings of that for consumers and for shareholders if a technology just eliminates labor. But the best technologies like electricity, like automobiles don't just eliminate people. What they actually do is they increase people workers ability to produce that they are technologies that increase human capability. And so you have the benefit of the synthesis of the technology and the worker. And that is actually what transforms lives. And he calls it total factor productivity. And so my ideas around this has been how do we do that? How do we make sure we just don't eliminate 4 million commercial drivers? How do we make sure that the adoption of things is actually making us more productive and that it's being done with respect to workers and.

B (105:48)

Capability. Okay, but so let's make it more specific because I agree at a high level with a lot of that. But let's talk about a specific role or job like truck driving. You've generally come out against or have concerns around AI based job displacement with long haul trucking and truck drivers. On the other side of that, if I'm running a trucking company and I want to deliver the best possible service for my customers, it's possible that AI would be able to support that. What kind of policy do you think is right in order to create? You want some guardrails around the industry how AI should be used in trucking? I'd love to kind of understand.

G (106:45)

More. Yeah, I would say have a human in the loop. And so what does that mean when I on a plane? You know, a lot of it is automated, but we still have a pilot there. And I'm glad we have a pilot. I wouldn't want to just fly in an automated plane. And so does this mean that a truck driver's job may become more appealing? Because right now, as you know, we have a shortage actually of truck drivers. More demand. But if they have a assist from a technology that maybe allows them to rest more, that's less taxing, they're there for the edge cases if something is possibly going wrong. They're there to deal with maintenance. They're there to make sure that, that you have loading and unloading happening. We can reimagine what the role of a truck driver is going to be. And we can certainly have a temporary view that for the next five years that you should have the driver there. Now, that doesn't mean that at some point there may not be jobs or certain parts of things that don't require a driver. But it doesn't seem unreasonable for five years to say you want a driver in the.

A (107:54)

Loop.

G (107:54)

Loop. And let's rethink the types of jobs that that will be. And if we need the government to be helping invest in the developing of this technology, fine, but do it in a way that's going to be complementary with.

A (108:09)

Drivers. Yeah, this is kind of happening already with Waymo, where there is a human in the loop and it's. But you know, the ratio of tele operators to cars on the road is, Is potentially higher than 1 to 1 right now, according to some reports. But over time, I think the Waymo team expects there to be fewer and fewer humans in the loop over time. The question is, how fast does that happen? And you're sort of proposing maybe try and make that as gradual as a process as possible. Because, I mean, you go back to the elevator operator used to be a human. Now, now we use buttons. And no one's really missing those jobs they phased out over time. I think the main thing is everyone is concerned about rapid job displacement. Not necessarily the. If I told you your grandson can't be a truck driver, you'd say, oh, he'll find a different job. But if it's like every truck driver out of the job next year, that's obviously much more dis. Disengaging to the US Economy. Is that how you think about it in terms of just timelines more than strict rules.

G (109:20)

Forever? I think that's thoughtful. There's a famous economist who once said in a gender time jobs for the father, not for the son. And by that he meant, look, we've got to make sure that people in their 30s, 40s, 50s, 60s have jobs that doesn't mean that that's exactly what their kids are going to do or their grandkids are going to do. For a lot of these human in the loop legislation, we're talking about five years, we're not talking about 15 years. And we're talking about roles evolving. Right. I mean, it may be that these evolve and then there's less of a need to hire folks down the line. And you have a natural transition of folks, but you're taking people who are workers and making sure that they're productive and they have have a good life. Let me explain why I think this matters. Phone operators, which people often give an example, and Alexis, who's at Chicago, had a great point about this. That was 2% of the workforce. Commercial drivers are 10% of the workforce. You already have an anger in the country of so many people displaced by globalization, displaced by the concentration of wealth in some areas. And you really want to throw into this mix a rapid mass job loss, displacement, and then what, just compensate them and have people stay at home and just get a check? Like, is that the society that we think is going to be productive or do we rather figure out how they have some role and some say and the transition be managed in a way that is, that considers their interests as well. And that's, you know, and I get that it's a good natured debate and people will say, okay, kind of you're adding some cost to the issue. And if all you cared about was shareholder profits and minimizing consumer costs is your only holy grail and you didn't care about jobs and you didn't care about communities, then people have a legitimate critique of me. But I would argue that that was the mentality during globalization and it's what's led to the. So much of the polarization, not just of our politics in the United States, but in the Western world, led to things like Brexit, led to anti immigrant sentiment, and maybe we should consider jobs and communities not as dispositive, but as a factor, just like we consider consumer costs and shareholder.

A (111:47)

Profits. Yeah. What's your, what's your look back on how the Uber story played out? Because that was a weird moment where there was a big pushback from the taxicab drivers. Those jobs still exist, but they're just way less profitable because the medallion system has kind of been undone. But if you want to make, if you want to make money driving someone, you can, but it's. You're making less money. Like, do you think that we should have handled that differently if we could run back the time or do you think it happened slowly enough that it was actually okay and delivered enough value to the consumer? Because Uber is one of those weird examples where the amount of, you know, taxicab like activity, ride sharing activity, it 10x'd and more people take these rides than ever before in the taxi era. And yet it did have remarkable impact on the market structure of that.

G (112:47)

Industry. Well, I'd be hypocritical for saying I'm against Uber. I take Ubers all the.

A (112:53)

Time.

G (112:53)

Right. I'm the same way expose next time I do an Uber. But I'll say this. We should have done more for the medallion owners. Right? I tried actually, in New York. This is before Zoran Mandani became Zoran Mandani, when he was an assembly member. He was really focused on a lot of these taxi drivers who had lost their medallion value and were underwater and what could we do to compensate them? And I'd actually reached out to Jamie Dahmen, who tried to do something to his credit through JP Morgan and it ended up not working out. But we should have, as a government done more to help those folks who had medallions who lost all their value. That's an example of something where we could have been more proactive. And then there's a huge debate about Uber drivers and whether they're getting enough value and have enough say over their, their lives. I, I would argue that, that we need that and I'd argue we need national health insurance. This is the biggest, biggest area where if you're not going to be employed at a, as a traditional employee, it would really help if people didn't have to buy healthcare on an exchange that has soaring premiums. So there are better things we need to be doing to help that Uber driver. But do I, am I glad that there is a technology like Uber? Yes, I am. I think it has created jobs and it has made life easier for many.

B (114:21)

People. You brought up Mamdani. He made a post, I think it was yesterday or the day before that a bunch of Silicon Valley types were agreeing with, which was. You don't see, you haven't seen that very often. It was around basically around SMB deregulation, making it easier to get a small business off the ground. Should that be a more important conversation in every state and region? I feel like growing up in California, I've seen so many businesses try to get off the ground and you end up seeing like a finished restaurant that just has its door closed because they're waiting on Some, some permit or something like that. And it's obviously hard enough to start a restaurant. And it seems like oftentimes local governments can get in the way. Do you think that needs to be just a bigger part of the conversation? As you know, given that starting a business is a great way to insulate yourself from at least some job displacement risk with.

G (115:25)

AI? Yes, it does. And look, Zoran became famous in part with his halal video where he was basically saying it takes too much regulation to have a halal stall and we need to streamline that. And so I believe, yes, we need to make it easier for people to start a small business to be their own business owner. That's not just making the permitting easier, it's also making sure people have access to capital. A lot of times that's a barrier. But I'll tell you one thing that I think is often a blind spot for folks in my district. District. I love small businesses. I love entrepreneurs. I think that there's a lot of people who want to build.

A (116:03)

Wealth. Completely agree. Completely agree. We love small businesses here.

G (116:09)

Too. But. Here's the but, okay, Most Americans, most Americans are not going to go just start a small business like this. This idea that every person in Bucks County, Pennsylvania, where I grew up or Western Pennsylvania should start a startup or build a business. Like my dad never did that. He had a middle class life. He worked for the same company for 30 years. And there are a lot of people who just want a decent job and they just want a job that can support a family. And there's nothing wrong with that if they want to be in manufacturing or they want to be a nurse or they want to be a child care provider. And so sometimes our rhetoric becomes like, why can't everyone become an entrepreneur? It's like, why can't everybody become a politician now maybe is a better life, but like, a lot of people just don't want to do that and they still want to have the American dream. And so all I'm saying is let's think about how to help small business owners, but let's also think about the 4 million people who are drivers and like, what is their life going to look like? And it's important to have that.

A (117:10)

Balance. Give me some lessons from the recent trip to China. I'm fascinated by how they're dealing with AI. Are they doing anything right? Are they moving even faster? Do they have a solution to the job displacement problems? Is there anything good or maybe risky that you found going out there? What were your.

G (117:30)

Takeaways? 3 takeaways 1. 1/3 of the AI talent is in China. What does that mean? That means it would be totally counterproductive to ban Chinese students from coming to the United States or Chinese entrepreneurs coming to the United States. We want to have that talent come to the United States because we still have a better ecosystem for capital and for investment. Second, we need to make sure that we're developing the talent in, in. In AI here in the United States and investing in STEM and making sure that we're encouraging the, the local development of that. But third, and this is the most important, important guess how much youth unemployment is in.

A (118:17)

China. It's nearly 20%. It's really.

G (118:19)

High. You guys are too smart on.

A (118:21)

This. It's really.

G (118:22)

High.

A (118:23)

20%. But that's crazy. How is that possible? I feel like, can't they just go build more bridges and create more jobs? I thought, I thought it was a command and control economy. I don't know. It's.

B (118:32)

Always. They have enough to empty sky.

A (118:34)

Rises. I think.

F (118:36)

Maybe. I don't.

A (118:36)

Know. Yeah. What was your takeaway from.

G (118:38)

It? As I describe it to people, you can't build dating apps in China. Right. Like so. So, you know, the people who have.

A (118:44)

These. Is it.

G (118:46)

Banned? Yeah, I mean, it's such a directed economy. They want everyone to, like, make stuff, manufacture stuff, not do things that they would consider.

A (118:55)

Frivolous. Sure.

G (118:56)

Sure. Sports app, a music.

A (118:58)

App.

G (118:58)

Interesting. All the cultural stuff that we do that improves consumer life or thinks about consumer needs and, you know, so you're someone who gets the same fancy education in college, and then they're like, okay, go work at a factory. And just like, we've undervalued people who want to work at factories in America. We should be having more trade schools and more respect for factory workers. They've undervalued people who don't want to work at a factory. And the reality is, like, you should have both choices. So these people, they, they, they're there and they don't want to go necessarily to build a bridge or necessarily to build the next factory of robotics. And it was hilarious because I would talk to the Premier, Li Chang or others and they'd say, well, it's a voluntary unemployment problem. These are just folks they should be getting doing these jobs. But what if in America we said, okay, you know, is one of the newsrooms, when they were being laid off, said to someone, go become an electrician? Well, that's as offensive as telling a steelworker to become a coder. Like, like, you know, people do things and they want to do what they aspire to do. And China is a command directed economy that is overvalued. Manufacturing doesn't have that diversity we do. Our problem has been the opposite that we undervalued making things, we undervalued the trades. And so what we need is sort of a balance for America to have manufacturing but also this incredible ecosystem of the service economy which can employ people where China, China can't. And that's ultimately why I bet on America. I'm also one pointer sick of this argument that let's just go be like China where they're going to eat our lunch. Really the Chinese model is crony communism. Like okay, GPING gets rich and a bunch of people who are running these companies get rich and the rest. And then you have 20% unemployment and you have consumer welfare declining and look at how most people live. They don't live live in nice houses with you know, two cars. So like I don't want China as a model and I'm not going to compromise every American having economic security just because we're chasing China. China is not the model. America needs to be more like America of how we built America in the 1940s.

A (121:15)

50S. I completely.

B (121:17)

Agree. Want your takes on a couple things. Housing affordability. I think a lot of people agree right now that housing affordability is sort of like upstream of a lot of the problems that we're facing as a country. What's your current stance on how we can improve affordability at kind of the local level and at the federal.

G (121:41)

Level? I'm a yimby, I'm an abundance guy on housing. We got to build far more housing in California. I don't endorse people or sort of zero housing people in my district we got to realize that esthetics matter but economic equality of opportunity matters more. And you can't have 5 trillion dollar companies in my district and expect to live like where the valley of the hearts delight. Like if you got that many companies you got to have housing near transit and dense housing to make sure that people can live there and that it's not just a, a place where wealthy people can live and that the working and middle class is getting shafted. We also need to stop private equity from buying up single family homes. People say oh this is a red herring. No it's not a red herring. In some places they have bought up too much single family homes. So pro building, pro streamlining, making it easier to build and having zoning reform and stop private equity from buying up these Single family.

A (122:42)

Homes. What about international.

B (122:44)

Policy? Will we make progress at least in California on those issues in the next 10.

G (122:51)

Years? Yes, because I think people realize we didn't make enough progress over the last 10 years, that this is a failure of California policy. And whoever is elected the next governor, I can't imagine it won't be on a abundance agenda when it comes to housing and it's not going to be okay, let me do it at the last year try to do something of an eight year term. It's going to be day one. How do we start to do things that it's going to build more housing. So I think it's been a wake up call for.

B (123:20)

California. That makes sense. Any quick comments on the current state versus federal AI regulation? We didn't get to touch on that earlier. And you had some Comments recently on SB10, 10:47, the bill in California, but what's your updated view on where regulation should be.

G (123:45)

Happening? Well, look, ultimately we need a federal regulatory framework. But the way you get good federal legislation is having legislation in the states. That's federalism. And I don't understand how you would have a moratorium on having state legislation when federal legislation right now looks bleak. The prospects of it are bleak. It is such an unpopular position even among Republicans. So my view is build a consensus that you can have thoughtful regulations at the federal level and work on that. Don't stop states from regulating. And this idea that, okay, you're going to stop all the growth. I mean my district is $18 trillion of value. We've got five companies, over a trillion dollars east of the Mississippi. There's not a.

A (124:35)

Single, you know, California's undefeated. It's so.

G (124:42)

Good. You talk to folks in like Bucks County, Pennsylvania, where I grew up, and they're like, come on, come on. Producing more wealth than ever before. Like what we want to know is how is, how are our kids going to fit into.

A (124:55)

This?

G (124:55)

Yeah. And I just think that, that I wish more tech leaders, you know, who sometimes gets it as a. Jensen Huang is talking about this. Like, yeah, I mean about how do we create economic development opportunities in places that have been left out. How do we make sure that everyone comes along on the revolution. I just think it's in tech companies interest to embrace this in a similar way as the economic royalists embrace the New Deal eventually. I mean you can't have just a capitalism that is only working for some with large chunks of the country suspicious and left out.

A (125:34)

Yeah. I just worry that we don't know the shape of what we're regulating yet, like the unintended consequences of social media took five, 10 years to develop. I mean, two years ago we were reflecting on this. People were worried about AI killing everyone and creating the Terminator. And then what wound up happening? Well, it wasn't really political misconduct information. It was much more people chatting with it for a really long time, going crazy. You know, maybe overbuilding, maybe risk in the debt markets. Like, the risks were very hard to predict. There were risks, but it wasn't exactly what we thought. And so I'm always, I'm a little bit like hesitant about, like, you know, maybe there should be regulations. But how, when will we be confident that we know how to regulate? Is it right? Is now the right time? Do we have clarity? Because a lot of the stuff it stands on, you know, we already have fair use, we already have copyright protections, and so a lot of it can be enforced through the courts, I would imagine. But of course, if new problems come up, they need to be resolved. And that's the way we resolve them in a democratic.

G (126:40)

Society. I think that's fair. The places I focus on are jobs and American.

A (126:47)

Citizenship. And I agree with you on the jobs part, but. But it just feels like the jobs, we haven't seen a collapse. And even people building the AI technology are like, this is gonna put everyone out of jobs and that's good. And then the people that hate the technology are saying it's gonna put everyone out of a job and that's bad. And it's kind of crazy because they all agree that the jobs are going away. And yet what do you get when you actually look at the jobs figures, it seems like we still have jobs. It seems like we actually can't delegate to the AI And I can't change. Just say, hey trucker, I want the AI to handle this one. The technology is not there yet. And will it be a year? Will it be 5 years, 10 years, 100 years? There's a whole bunch of incentives to say it's coming right now. And it's hard to get a read on. And predicting when things will happen is. Fortunes are won and lost on.

G (127:35)

That alone, totally agree with you. John Maynard Keynes said we'd all be working 15 hour work weeks and he was.

A (127:41)

On.

G (127:41)

Exactly. You knew more about economics than any of us. So, you know, it's hard to predict. But I think what we can do is when you look at Darren, a smuggler, who says, well, why don't we have a neutral tax code so we're not incentivizing depreciation of investment in technology and automation over hiring people. I mean, there are things we can do that make it that we, we prioritize having people in the loop. And then there are things we can do in our, our social media environment that protect us as citizens and kids. Two things are like, let's eliminate bots. Right. Elon Musk talked about doing this on X and there's still a ton of bots, but a lot of the bots that use AI are, in my view, hurting our democracy. And then let's protect kids from some of the harms on social media. Yeah, you know, so, yeah, you know, I love, I love sparring with folks and I appreciate sort of the criticism I've gotten from some of the tech folks where they. The tweets on, on AI and drivers. But I guess what I would hope for tech people listening to this is don't resist every form of regulation and sort of dismiss people's anxieties instead be part of how we get smart regulations and how we answer people's concerns. Because if 70% of the American people believe the American dream is dead and have a concern about AI, like, the answer to that for anyone who's been like in a relationship is not to dismiss it and say they're dumb. It's to say, okay, how do I address that anxiety so that we can move forward? And I guess my hope would be that there'll be more tech leaders like that. Victor Peng is one who's the former leader at amd. I mean, there's some people who are thinking in that way, and I think it's in Silicon Valley's interest to have that kind.

A (129:25)

Of. No, that makes.

B (129:26)

Sense. I think you really, you really freaked people out with there should be a tax on mass job.

A (129:31)

Displacement. Well, there is a tax on the profits, right? Like we tax profits. So, yeah, I mean, there's a question of like, maybe we adjust that. But it's, it's all. These are all dials that already exist. We're just discussing how we turn them, I would imagine. Yeah, I don't.

B (129:46)

Know. Well, thank.

G (129:47)

You. A lot of times, you know, this is one thing different for me than other politicians. I is I toss ideas out there. If I think there's good pushback, then I adjust my views and I'm like a politician like this. I just talk. Like I talk to someone over a drink over at a bar. You know, I, and everyone else is like so scripted. Oh, you can't put out an idea because, you know, maybe it'll come back 10, two years later on Face the Nation. I just don't think that's what our politics are. It's like, put your ideas out there. Like, what human being doesn't have some ideas that are dumb? Like maybe, maybe Einstein didn't or something? Most of us, yeah, we put up good ideas, we put up bad.

A (130:22)

Ideas. I love.

G (130:22)

It. You think.

E (130:23)

Experiment. I love.

B (130:24)

That. I love that approach. And it certainly sparks a conversation.

A (130:28)

And it certainly fits with what we've done here today. This was really fun. We really appreciate you coming on the show and just like going all over the place and just talking through all this stuff. It's fascinating. I'm learning a ton. And we really appreciate you taking the time to come talk to.

B (130:40)

Us. Thank you so much for coming.

G (130:41)

On. You guys are doing great. Seriously. You're elevating the conversation in Silicon Valley and it's an honor to be on and I look forward to.

A (130:49)

Being. Yeah, yeah. We'd love to have you back on the show and go way deeper on all of these. And I'm sure by the time, the next time you're on, all of the data points will be different and we'll be looking at and we'll be staring at new problems and they will require new solutions and new discussions. And so thank you so much for taking the time to come talk to.

B (131:05)

Us. I appreciate your.

G (131:06)

Approach. Thank.

A (131:06)

You. Have a great day. We'll talk to you soon. Bye. Before we bring in our next guest, let me tell you about numeral.com. let Numeral worry about sales tax and VAT compliance. Compliance handled so you can focus on growth. Our next guest is Jonathan. Jonathan Swordland from Function. Hey. Sorry, we were in a political quagmire. We were in the.

B (131:32)

Swamp. We were in the.

A (131:33)

Swamp. We went to the swamp. We don't normally go to the swamp. Normally we talk about series B's. We talk about large series, baby. He did get us going, though. He was telling us how much value has been created in his district. It's in the trillions. It's in the tens of trillions. And we were just foaming at the mouth about the market caps. And then we said a bunch of other stuff. But thank you so much for coming on the show. For those who aren't familiar, introduce yourself, introduce the business, tell us what's going.

I (131:58)

On. Absolutely great to be here. Now you're climbing out of the swamp. We're going to talk about something a little less.

A (132:04)

Swampy. Thank you. Talk about.

I (132:06)

Health. It's great to see you.

B (132:07)

Guys. Well, I mean, health is. Health is like, honestly more political than politics. It can be, but, but this conversation won't.

A (132:16)

Be. It's funny, we say, we actually.

I (132:18)

Say that biology is bipartisan, though. And I like this idea of everybody can agree that nobody likes to.

A (132:26)

Suffer.

I (132:27)

Yeah. You know, and, and everybody can agree that preventable death shouldn't.

A (132:33)

Happen.

I (132:34)

Yeah. So it comes down. But of course, the nuance of how you get there can become political because who's going to pay for it?

A (132:40)

Right. But.

B (132:42)

Also. Well, that or the. Well, this.

A (132:45)

Diet. This diet's right wing. That diet. Oh, working out. That's a right wing thing. Or like, oh, this is left wing. And like, you know, different ingredients became politically charged over the last few.

B (132:55)

Years. My powder is better than your powder powder for.

A (132:57)

Sure. And sometimes there's political influencers on the right and the left who actually have the same supplier and then they put different branding on top of it and they sell that. That's a fascinating rabbit hole to go down. But anyway, we're not here to sell supplements. Let's talk about the.

I (133:11)

Business. You know, it's funny, it's like, I'm not left wing, I'm not right wing. I'm a whole bird. Otherwise you fly around in circles is kind of the.

B (133:19)

Idea. I love it, I love it, I like.

A (133:21)

It. The whole bird, that's.

B (133:22)

Great. The whole bird. The whole.

A (133:22)

Turkey. So, yeah, take us through the shape of the business these days. What, what's the value prop to consumers? What's the progress been? How big is the company kind of set the table for.

I (133:32)

Us? Okay, so simple value prop is get on top of your health. It's time you open your health. So what does that start with? It starts with a new platform. It's $1 per day to join and the platform includes twice a year comprehensive lab testing at over 2,200 locations. Any quest diagnostics around the country. You go and you test everything, heart, hormones, liver, kidney, thyroid, cancer signals, you name it, up and down. All of that data goes into a platform, into an app that explains you what's actually happening inside your body. And these are the things that you would not get in a physical. This is like a true, true deep look. What function is created is this entirely new standard for your health that every year for the rest of your life, you know that you're well on top of whatever's happening inside your body. You're seeing how it's changing over time. You're making sure that you're getting well ahead of disease and you're doing everything you can to feel your best. So that's, that's the. The value proposition, and that's what function delivers right now. We started with lab.

A (134:28)

Testing.

I (134:29)

Yeah. Because that's, that's like. That's the most impactful data. 70% of medical decisions are based on lab testing. And recently we acquired a company. You might have heard about this. It's called Ezra. And Ezra is an imaging business. And so Ezra does. And what has been amazing for us, we've gotten FDA cleared AIs that have reduced the time that it takes for somebody to get an mri. Okay, so why does that matter? One, nobody wants to be sighted. An MRI machine, typically. Two, it also massively reduces the cost and it picks up the.

A (135:07)

Efficiency.

I (135:09)

Yeah. And so you. What we've actually done is we've introduced lab testing, became one of the largest, most powerful lab testers in the country. And then we went into imaging on the imaging side, bringing down the cost. And what you're seeing actually emerges this new standard for health. We took the most impactful parts of the health system for capturing your data, and we, we packaged it up into something that's really simple to understand and really affordable.

B (135:32)

For. For many, many people talk about how. Talk about how MRIs were used historically or these things that, that get done when, like, you're act. You have like, acute pain or you have an issue and then you're doing.

A (135:47)

It. Terrier.

B (135:48)

Acl. Yeah. Or this feels like kind of flipping it and saying, using it as, like, preventative care. Is that the right.

I (135:54)

Read? That's right, Reed. And not just preventive. I would just say responsible, because this idea of preventive is great, but it's also what might be happening right now that you don't even know about.

G (136:05)

About.

I (136:06)

Right. And so the word preventive and the word early are a little tricky for me because the word early, it's like, why is it early detection? We just call it detection. Can't we just get rid of the word early? What MRI does is traditionally it allows somebody to look inside the body. But to do that, it's been really, really expensive. You get an mri, you know, Terrier, acl, something like that. Like, you basically, you have to send thousands of dollars to look inside your body. And that's the way insurance is set up, and that's the way MRI set up. But what MRI can do is it can look at every single organ and look for tumors that are 0.2 centimeters, 2 millimeters. It can look for stroke risk, aneurysm risk. Endometriosis, hernias, tears, everything. So if you actually want to understand what's happening inside the body, an MRI is an incredible way to do it. But it's been so arcane and so difficult. It's never actually been architected and set up to look at the body and get well ahead of things. And it's usually been, oh, you go in a hospital, you broke something, you have an issue, you go look at this one particular area in functions case. You can actually look at most of the body through an mri and you can detect cancers early, detect aneurysm, stroke risk, et cetera. And you can do it for $499, and you can do it across almost 200 locations by the end of this year. There's never been anything like this. This is the first time in history this has been possible. It is the first time in history it's been possible geographically, from a cost perspective, technologically and culturally, it's changing. People are realizing this. What I was alluding to before, it's a really important point, is a new standard of health is emerging. And that standard includes twice a year, comprehensive lab testing. It takes a 10, 15 minutes each time you go in, you get your whole body testing, you find out what's actually happening inside. And the second thing is now a quick MRI every year. And if you do it, what you're doing is you're actually creating a baseline for your whole health and you're seeing how things are changing over time. You're catching velocity, you're seeing bad trend lines, and you're also just flagging critical issues as well as finding out what you can optimize and what can be better in your life. And what's crazy to me is the current standard, like the status quo. We've all done this, We've all gone into the doctor's office, office. They test you for like 20 things, get a phone call in three weeks, you're good to go. John, Jordy, see you in six months, a year, two years, whatever, and you move on with your day. And that's just this episodic, once in a while, very narrow perspective on your health. That's gone. But they miss. They're not looking at cancer, and they're really not even looking at heart disease. The two leading causes of death, let alone metabolic dysfunction, hormonal issues, thyroid issues, and functional looks at all that. That I'll give you a crazy stat. New study just came out. 45% of people that were hospitalized for their first heart attack did not have what is considered high risk cholesterol. That should be terrifying. Why? Because if you go to a doctor's office today, a regular old physician's office for a checkup, you get your LDL checked, right? You guys have done.

A (139:10)

This? Yeah.

I (139:10)

Yeah. Okay. That marker was born in the 1950s. It's older than my.

B (139:17)

Father. Vintage. It's a vintage.

A (139:20)

Marker. Some people would say Lindy. Some people would say, that's Lindy. Okay, let's just steel man it for a minute. They might say it's.

I (139:29)

Lindy. So look, there is no world where any top cardiologists say, I'm just going to rely on LDL cholesterol. Basically, when every top cardiologist tell you, let's look at apob, let's look at lp, let's look at lipid particle center. For most people, they don't. Those words are foreign to them, but they should be. I mean, it's this avant garde stuff. Right, But. So there are way better ways to look at the heart. But we're relying on something that's back in the 1950s as status quo. And what function's done is for hundreds of thousands of people now we've actually delivered a new standard of health that includes twice a year testing of everything that looks at your heart, that's looking at your kidneys, your thyroid, your hormones, everything. Women don't have to go to their doctor and ask for a horn panel and get chased around. Instead, they can actually get a look at what's going on with their hormones. And then on the cancer side, real quick, cancer side 400, you're. You're four times more likely to survive cancer if you catch it early. But right now, status quo is you have to wait till you have symptoms to catch.

B (140:35)

Cancer. Yeah, it's.

I (140:37)

Crazy. There's no way that's okay. I don't want that for my family. So. And now there's technology where we can actually, with an MRI as well as with a grail test, test that we test for many, many, many thousands of people. We can actually get way ahead of these things. So I can talk.

B (140:51)

About. Yeah, yeah. So I'm sold on the product. I think it's hard. I think it's hard not to be. It's the best kind of like value offering, I think, in health, like, period. And I was sold obviously, when you were raising pre seed back in the day, two and a half years ago or something like that. Feels, feels like forever. Can you give us like a. I'd love to get your view on an Update of like the market structure. A lot of companies have seen you guys weren't function wasn't the first lab, you know, testing and health platform like this to exist, but your guys execution and the growth. I think you're one of the at least growing faster than a lot of the fastest growing AI companies that we're seeing out of the last year. Give us an update on like the shape of the market, how you see the market evolving. Because like I was saying a lot of people are trying to like ride, ride your, your coattails. But I'm curious for, for an update.

I (141:58)

There. You know, we, the, the market is realizing that the word consumer health has been this like dirty word for 20 years or something and it's not, it's, it's. What it is is it's premise on the most primary thing that we experience as human beings is our biology. It's our life experience. And what's the LTV of your health? Right? You'd be willing to pay anything for health. It's the most valuable thing in the world for you. And so we're finally in a place where we can actually see technology and products broadly applied to health. And so it's, it's, you're looking at a TAM that conservatively is $7 trillion and some has give it up for $7.

A (142:44)

Trillion.

B (142:45)

TAMS. John, hit the, hit the, hit the size gong for a 7 trillion dollar TAM. Had to, had to. Anyways, continue. I love.

I (142:56)

It. I love.

G (142:57)

It. No, so, so, so look this is, this.

I (143:00)

Is. People have been saying spending absurd amounts of money on their health through these massive service platforms like insurance companies, big health systems. And finally people are saying you know what health happens outside of the doctor's office. And I'm taking it into my own hands. And what we're doing is we're bringing scientific and medical rigor directly into a platform that people, they themselves can sign up for, they themselves can manage and so they can make decisions for themselves and that gets them way ahead of diseases. You know, as you were saying before, like this is, this is not a, it's not a trivial space. I think it's, it is the best is the most anticipated service for AI. It is the best application of AI in the world is to our health because that is the major experience. And so of we're, we're, we're surprised that the category and all the, all the competitors aren't, aren't. It's not that it's not bigger that more people are jumping into this. Like we, we know that people are going to try to ride these coattails. But, but we just, we are, our head is down and our focus is on how can we deliver as much value per dollar for each one of our members. We have hundreds of thousands of members soon millions of members. We have been growing really fast because we're at, we're actually delivering something to somebody that has real, real, real substantial.

H (144:19)

Value.

I (144:19)

Value. And at a time when a lot of technology can do a lot, it's like, what are we really paying.

A (144:25)

For? And it's like, where can people get started? You mentioned it's a dollar a.

I (144:29)

Day.

A (144:30)

Correct. Does it take me through like the customer flow? Is it just a website? And then I go to the lab, explain how people can get, can get.

I (144:38)

Going. Okay, so it used to be $999 when we started. It was manual and it was per.

A (144:43)

Day. Thousand dollars per.

I (144:45)

Day. No, thousand dollars per.

A (144:47)

Year. Sign me.

B (144:47)

Up. Thousand dollars, he said per year. Don't, don't worry. John's just messing with.

I (144:52)

You. So it started at a thousand bucks per year. Then we worked really hard to bring up the efficiencies and tech got it down to 499. And a couple weeks ago we announced it's now 365. It's like when actually $365 per day because health, health is an everyday thing and it's an understandable price. And when has healthcare actually been.

A (145:12)

Deflationary? Yeah. And then bliss. So go to Equest Labs probably twice a.

I (145:17)

Year. You go to functionhealth.com.

A (145:18)

Yep. Functionhealth.com functionhealth.com.

I (145:20)

Simple. You just sign right up right there in the scheduler. You sign up for your lab.

A (145:25)

Appointment. You show up at the lab.

I (145:27)

Your blood drawn, urine collected. You walk out 10, 15 minutes later. In 24 hours, results are pouring in. Now your app is live and all the data is coming in. It's making sense of it. And you test every six months, under six.

B (145:41)

Months. I think people. One of the reasons people underestimated this kind of category as it was emerging is how many people got burned on like DNA testing. Open DNA. DNA like the 23andMe's. You test it once and then there's like zero incentive to retest. Right. You did 23andMe. You have the.

A (145:59)

Data? Well, it depends. Are you working on your DNA or not? Have you been modifying your DNA? If you modify your DNA regularly, you should probably be testing your DNA regularly. You never.

B (146:08)

Know. You never.

A (146:09)

Know. I might have rewritten my entire DNA, all of it from Start to finish, every base pair, it's different now. Sign me up again. I'm ready to.

G (146:18)

Go. We have to study.

I (146:20)

You. If that's the case, we're gonna have to bring you.

B (146:22)

In. John needs to be.

A (146:24)

Studied. Honestly, Generally, what does 25,000 Diet Cokes do to the human body? We're gonna find out what is 500 Diet Cokes a year, a dollar.

B (146:37)

A day on function, and four diet. And at least four Diet Cokes a day for John. We're actually running a split test. We have the exact same lifestyle. We show up at the gym every morning. We work out, we prep the show, we do the show. We hang out with our family. We're gonna do that forever. But John drinks Diet Coke and.

A (146:56)

I drink Matayina Yerba Mate. Podcast. And a can from Andrew Huberman, of course. And we're going to find out. Yeah, we're going to find out. Well, thank you so much for taking the time to come chat with us. We have a small bit of breaking news. I want to get to you before our next guest, so we will be seeing you soon. Oh, one last thing. Give us the numbers in the last fundraising round. I want to ring the gong for.

I (147:17)

Real. Yeah. Let's do it. Series B, 298 million raised. $2.5 billion valuation. The thing to think about here is that's basically a dollar for every American.

A (147:30)

Adult. There we.

I (147:31)

Go. And so what. What that is, is that's a. That's a vote on your health. That's not.

A (147:35)

Just. I love it. Well, thank you so much for taking the time to stop by. We will talk to you.

B (147:40)

Soon. For. For the next one for the Sea coming in person. We got a seat here for you. Be honored to have you in.

A (147:49)

Person. Be.

I (147:50)

Great. Let's go.

A (147:51)

Brother. We'll talk to you.

B (147:52)

Soon. Great to see.

A (147:52)

You. Goodbye. Let me tell you about Vanta Automate Compliance and Security leading. Vanta is the leading AI trust management platform. Also, if you're running a NEO cloud, you got to get on Vanta because that's one of the criteria for Clustermax. I'm not kidding, not making this up. SOC 2 compliance is a big factor in actually making it up the tier rankings for clustermax, because, of course, if you're training on customer data, you need SOC2 compliance. You need the whole process. Anyway, the breaking news that I wanted to get to really quickly is Josh Kushner is partnering with OpenAI. OpenAI. He says. We are excited to announce a strategic partnership between OpenAI and Thrive holdings through our partnership OpenAI will become an equity holder in holdings and collectively we will set out to deliver frontier technology to our customers. For decades, technology has transformed the world's largest industries from the outside in. We believe the AI paradigm will be different in that some of the most profound transformations will now occur from the inside out. We view the businesses that we own and operate as the right reward system to build, test and improve industry specific products and models. So the race is on. Is it inside out or outside in transformation? What's going to happen? These are the new fast takeoff, short timeline, long timeline. Are you an inside out guy or an outside in guy? This is going to be the defining debate over the next couple.

B (149:22)

Days. Well.

A (149:22)

Said. Get ready to lock in. We'll be covering it here. We'll probably have some people on who are digging into this, investing in this, have long takes, short takes, who knows. But I want to get to the bottom of what this outside in versus inside out transformation will look like. Like we've been digging in a little bit, talking to some folks who are building companies, buying.

B (149:43)

Companies. Taylor says deal guy.

A (149:45)

Yuga. This is the deal guy Yuga. It's happening. It's happening. Well, before we bring in our next guest, let me tell you about Figma. Think bigger, build faster. Figma helps design and development teams build great products together. We have Cristobal Valenzuela from Runway in the Restream waiting room. Let's bring him in. How are you doing? Good to see you again. Thank you so much much for taking the time to come talk to us on such a big day. Kick us off with a reintroduction on where the company is today and then the news. I'd love to know about the.

D (150:17)

News.

H (150:18)

Yeah. Thank you for having me.

B (150:19)

Again. It's been a.

A (150:20)

While.

H (150:20)

Yeah. So big news. We just released our latest Frontier model. Runway Gen 4.5 is some model we've been working on for quite some time. It's the best video model right now in the world, which is a pretty remarkable fit. So I think it's, it's pretty.

B (150:40)

Good. It's pretty fun to play with. Basically that's my audio. I'm at.

H (150:44)

It. But perfect.

A (150:47)

Timing. But let's play some of the video. I want to see the demo videos that you put out, the examples and I want to ask you a bunch of questions about it because it's an extraordinary claim. Google is a serious company. They have a very serious asset in YouTube and I'm fascinated by. So first, give me the news Video arena leaderboard. That's the ranking that you're using, how is that scored? How does that actually.

H (151:17)

Work? So it's kind of like a way of crowdsourcing performance. You basically ask people in the Internet to vote against two videos and it's anonymous, so you vote left or right. And then as you keep on voting, you accumulate more votes. Once you vote, you can see who you voted for, but beforehand you don't know. And so over the last couple of months, we've been working for this entirely new way of, I would say, training both Vita models and image models in such a way that hopefully we thought it would all compete others in the arena. And we got results a couple days ago. And yes, we managed to basically outcompete all other video models, including both Google and OpenAI, which is a very remarkable feat if you think about the scale of resources. Like, I think it's the era of. Ilya was saying this is the era of research again. And I agree. But it's also the era of efficiency. Like really good, really focused teams with highly efficient mandates can get really.

A (152:13)

Far. Yeah. Tell me about what you optimized for here, because Sora seems. It's an incredible model and it was for like a minute, like, whoa, really mind blowing. Then I feel like I kind of developed an immune system for it. And I can clock a Sora video. And it feels like Sora was very much trained on TikTok almost or vertical social media video. And so what have been the breakout Sora videos? It's been a lot of dash cam footage and doorbell nest camera footage and facing.

B (152:51)

Videos. They upgraded the model dramatically, have.

A (152:53)

Degraded the model a lot. Whereas VO3, it felt like it had. It had a little bit of the Hollywood polish, but it was more like Michael Bay. When I looked at it, it looked very saturated. It was cool, it looked good. But what you went for, it feels a little bit more, I want to say cinematic, even though that's kind of an overused term. But talk to me about what your goal was or even if you have a goal when you go in into a training run like this, it.

H (153:22)

Does. So I think there's an explicit goal and an implicit goal. I think in a way, all models, specifically video models that are more visually clear or perceptible, have some sort of personality behind it. And I think that personality reflects a little bit both the point of view of the company and the way you want to train the models in the first place. To your point, if you want to make consumer slop and quick, shareable stuff, you're going to train the models just from the ground up very differently for the stuff that we're trying to do, which is a much more professional, high quality, very controllable sort of tools. And so a lot of what you're basically outlining is, I would say, the personality of the models and somehow also reflects the personality of the companies. If you're trying to sell ads, you're going to do a very different model. If you're trying to make creative tools. So I don't think there's one single recipe or one single ingredient. It's more of a just like, taste. Like, I think that word gets thrown a lot in research this stage. Just taste. And I think taste is both the research. Like, what do you want to work on? Like, having vision, like having. Okay, I want to pick these specific problems I want to work on and this is how we're going to solve them. And this is what we've learned over time. That's one form of taste. And the other one, more aesthetically is like, what things look good.

B (154:37)

Like. And that's like beavers on a construction.

A (154:39)

Site. This is actually very pure taste. That's pure taste. Yeah. Look at this. Hilarious.

H (154:45)

Right? Look at the motion of the donkey moving. Like the camera, the angles, like, the amount of data creation our team of artists and like filmmakers and like people have spent. It's not like trivial, to be honest. I think that's also the taste component. Like shots like.

A (155:00)

This. It's like some of this is horrifying. I mean, I guess that's the point.

B (155:05)

Really. Had to summon the demon on this one. Have you been inspired by Anthropic at all? It feels like somebody could put you in the anthropic for video bucket and that. Like, they're just like extreme focus on code and ignoring everything else. And meanwhile, your competitors are putting a lot of resources towards this, but they're not betting their entire business on it in the way that you.

H (155:29)

Are. Yeah, I think it's like a mercenary versus, like visionary type of like, I would say bet. It's like you want to have people who. Who feel very committed to the vision long term. And the way you do that is you're very focused on the culture and that culture eventually shines in the product. I think Entropic has also that you can tell who works there and how they think. And it's also very cohesive in a way. I think we spent somehow a similar amount of time doing that in a way. And I hope you can tell via the models themselves that Personality comes across nicely as well. Yeah, and I agree that at the end will be perhaps the most defining part of the companies that stay in the long run. I think if you just throw money at the problem, you're not going to get too far, to be.

A (156:17)

Honest. What went into the actual training run? Are you at a scale now where it's a meaningful capital investment to build a model like this? We saw the scaling paradigm change from maybe it's $100 million to do a big frontier language model run. Then we were talking about billion dollar training runs, bigger and bigger training runs. The results are remarkable. But has it been a remarkable amount of investment to get here or are there more efficient ways to actually get to a frontier result without spending frontier.

H (156:56)

Money? Yeah, I mean, it's definitely not cheap. Like, this is not like traditional SaaS. So you definitely have to spend more money, more resources. But. But I think we've proven that we are not spending tens of billions of dollars to get there and to overcome the challenges. And look, to be honest, the model is not perfect. There's a lot of things we're going to improve and we're going to fix and we're going to do larger training runs and we do more over time. But it's kind of, I would say the most expensive thing is the natural intuition the team builds around what kind of works and what doesn't work. Let's go back to the idea of research stage. So like, you can't throw money at it. You just have to spend enough time. We've been working on RHINO for almost a decade. And so there's a lot you've learned over time about what works and what doesn't that informs a lot of the efficiencies on training. And yes, like, expensive models will, like, you'll need more money to train larger and bigger models. Like, if this is the worst the models will be, imagine them in like two years. Like you're going to get there by training larger models for sure, but also knowing how to train them in the first place. And that's the part that I think is hard, hard to quantify per se. And what I'm really excited about is not only what the models can do, but also the efficiencies are not only on training, but on inference. This is a price point that's very comparable to our previous models. So it's actually very usable and hopefully you'll be using it in real time very soon. And so that level of, I would say efficiency at inference level, we haven't yet Seen it and I think we're going to get there very.

A (158:25)

Soon. Yeah, fascinating. I mean, some of those videos are very distracting. They're pretty.

B (158:31)

Remarkable. Your unlimited plan includes 2,250 credits monthly. How much video can one actually generate with.

H (158:43)

That? Well, technically.

B (158:44)

Unlimited. Okay, I was confused because it said there's still like a credit.

H (158:50)

System. No. So we have a queue. We have compute and there's a queue and you get into the queue and you generate a stake. The queue becomes a level. If you just want to generate like fast, you pay for credits, but depending on how anxious you are with your generations, it's a measurement of how fast you want it. But eventually you can just literally generate unlimited. By the way, I think no one else has a plan like that. It's a pretty good.

B (159:14)

Deal. What are the length of generations that are most commonly being done today? And is that a metric that you trust? Like, are people consistently? Is it like a 20 second scene? That's the most common today. And are you trying to get to two minutes or two hours? Like, how do you think about.

H (159:36)

So. Well, technically you can do like arbitrary durations if you want it, but like the average scene duration in like a short film or a movie is like actually two to three seconds long at the most. And that's actually been trending down like the scene, right? The scene itself, the cut is like 2 to 3 seconds long on average. So when you actually, when people mean like I want to drink a 45 minute like long thing, you don't want 45 minutes of like one camera, like fixed. You want like scene cuts and world. And you want the character like a shot, a medium shot, a long shot. And you have, you know, that's a different problem from like creating one continuous long sequence. So the one continuous long sequence for me is less interesting than the multi shot approach where you can create much more compelling narrative work. And I think we're not that far away from that being a reality where you can generate consistent narrative work. Really good visuals, really good stories with the level of quality of the videos that we're seeing right now here. But they're all tied together in a way that just makes it feel cohesive with each other. So that's a different problem, I.

B (160:43)

Would say altogether there was some debate on the why doesn't the cursor for video exist yet? Do you have any thoughts.

H (160:52)

There? What's the cursor for.

A (160:54)

Video? Basically a nonlinear editor, like a Premiere Pro, a DaVinci Resolve, an Adobe after effects for video cursor for video, like replacing the actual bones of the software that the editor that the video creator uses. There's been a couple apps that have spun up Runway. Originally the reason I was using it back in the day was for Green Screen, for Chroma Key, basically. It was fantastic for that. And it feels like that building a canvas, building an nle that feels like one potential pathway to victory. It's also very difficult because you can't just fork VS code. There are no leading open source NLEs. On the flip side, if you wanted to play nice with Adobe, you could be a vendor a la the way nanobanana is now vended into Photoshop and that could be a solution. There's a variety of ways to win. I'm interested in hearing your.

H (161:54)

Approach. Yeah, definitely an interesting question. And by the way, shout out for you for being an OG on Runway since.

A (162:01)

2019.

B (162:02)

Yeah. Something.

A (162:03)

Crazy. I love.

H (162:04)

You. Yeah. My two thoughts are first, the art of NLE and editing and film. It's an art and there's a lot of pacing and details that are very nuanced and specific. It's about granular details and it's hard for, I would say model or system to automate that level of decisions. That's on a purely NLE.

D (162:27)

Side.

H (162:27)

Right. But I would say, at least for us, more interestingly is the question of do we need an NLE in the first.

G (162:35)

Place?

H (162:35)

Right. Like do we actually need these primitives? If you think about non linear editing, this idea that you're like stacking frames of video against each other and like you're cutting them. Before it was with physical racers and now we have detail racers, you're cutting things together. My bet is that you probably won't need like NLEs. Like that whole paradigm will feel like a fax machine like in a few more.

A (162:57)

Years. And so I feel that's somewhat what's happening with like the, the Devins and the clogged codes and the codexes of video. I do wonder if there's going to be an intermediate step or maybe it'll just be absorbed by the current NLEs. I mean, I'm sure that's what your customers are using.

H (163:16)

Right? Yeah, I don't know. We'll see it play. But I'm not too fond of pushing better versions of NLEs out there. I think there's something around how you make videos and how you interact with this AI systems that just naturally actually allows itself with different primitives. And if you think also about the fact that very soon you'll start to See this happen in real time when you make real time narrative work or videos or experiences. How do you want to call them? You don't need to edit things async because you're generating on the fly and you have people interact with them. And so it changes. That's what I'm saying. It changes the nature of those things in the first place. And there's a transitional period where we're seeing like NLE is being augmented with AI, but I think it's that transitory. I don't think it's going to play out like in the long.

A (164:03)

Run. Yeah, yeah, no, I.

B (164:04)

Think. Has Hollywood capitulated yet? What's going on there? It's funny, I've been hearing more and more about Suno from not just guests and friends of the show, but just like random people out in the world. It sounds like every single musical artist now is like using it in some degree even if they're not, not willing to talk about it. What is the case in traditional Hollywood and entertainment? You can't exactly hide that you're using AI video. It's basically out in the open immediately and there's just so much negative energy that gets focused on it specifically from people that are within the.

H (164:45)

Industry. I think the negative energy is like the water problem with the AI. You know, like it's kind of this, this unrealistic like and very noisy, not representative sample of what's actually happening within the industry. If you go to la, you speak with the agencies, with the Thailand, with the filmmakers, with the studios, with the production teams, they're on board on AI like years ago, like months ago. Like their fans, they're using it, they understand it. Of course there's pockets of people who are like, like more advanced than others. But I would say that the narrative publicly hasn't yet caught up with that because some people might not want to speak about it. It's much more interesting to say all the negative things, to say the positive things. I would say Hollywood already has overcome that and they're pretty much on board. I would say gaming companies are now where Hollywood companies were a year and a half ago or two years ago. So that's, I would say an industry who's now catching up more to what AI can help them and how they can use it. So yeah, I would say some of those narratives are a bit fake, to be.

B (165:51)

Honest.

A (165:52)

Yeah. Well, thank you so much for taking the time to come on the show. Congrats to the busy day. We appreciate it and I can't wait to Play around with the new model. We have a benchmark here, bezel bench, where we try and recreate a very complicated shot that we shot practically with with a bunch of different watches with our intern or gap semester Tyler Cosgrove. And it has a very. The shot's very long, it pulls out, it twists around. It's a pretty complex shot. And that's our current benchmark. And we'll be testing and we'll let everyone know how it goes. But thank you so much for taking the time to come chat with.

H (166:28)

Us. We'll talk.

A (166:30)

Update. Goodbye. Let me tell you about Julius AI, the AI data analytics that works for you. Join millions who use Julius to connect their data, ask questions and get insights in seconds. We have Vincent from Prime Intellect in the Restream waiting room. How are you doing? Great to see you. It's been too long since last.

D (166:51)

Weekend. Thanks for having.

A (166:52)

Me. Congratulations master of finding the one day that we're not live to launch your new new Tell us what happened on Wednesday, the one day that we were off of.

D (167:06)

Streaming. Yes. So excited to give you a rundown. So basically for the broader context with primitive, kind of like our broader goals really creating open frontier models and infrastructure for everyone to create them. And last week we released Intellect 3 which is basically really like a scale up towards scaling RL and post training and creating like a sort of model especially for like more agentic tasks. So basically what we did is we took GLM and did a whole SFT stage and RL stage to create kind of like a state of the ART 100 billion parameter and model and really kind of like that whole infrastructure is kind of quite a challenge like from like the RL environments to the broader like code sandboxes and the whole stack to do post training that's basically what we built over the last half year. I think Will Brown came on this show show to unpack some of it on the verifiers and environment side. So basically that's kind of like what we released last week and really proved that kind of like we got performance at 100 billion scale that thus far in the open source only 300 to 600 billion parameter models like Deep Seq Aron for example achieved before. So basically getting to better performance actually at a much smaller scale. And I think in general it showcases that like like open models are starting to catch up. Obviously I think quite interesting in general seeing the trend that not just with our model but also more broadly with other releases like Deep Seek today and over the weekend that actually they're also on par with the close models now. And I think really our goal is so what's almost like a preview release but already sort of is we basically released like our early checkpoint and we actually scaling it much further also on more like agentic capabilities, but basically really making it sort of across a range of tasks. And really I think the foundation of this, which was quite interesting, is that we created this environment hub where anyone in the world can create one of these RL environments which we ultimately then included in a training run. So basically different people in the open source contributed actually to the RL environments that we trained on for those.

A (169:22)

Models. So yeah, give me a concrete example of this shift of businesses that need to buy a model that has been trained in a specific RL environment. We've heard the example of someone's creating a clone of doordash and they're figuring out how to do doordash orders agentically. But what else are you seeing? What are some other good examples of when a business would pull this off the shelf from all the different opportunities, from all the different APIs that are out there and create something I guess semi custom for a specific business use case? Like what are you seeing out.

D (170:00)

There? Yeah, so I think what's interesting is there's I think two buckets, basically there's a bunch of these people like creating RL environments for the labs, like the doordash clones, et cetera. So basically to push really capabilities. So I think we're in this paradigm right now obviously where ultimately like scaling RL is the main way on how these, like we've seen it with opus or with GPT5 and Gemini, like that was mainly like I think a scale up in rl. But basically what we are seeing are two things. It's like on the one side it's like there's a lot of demand for these RL environments. But then the other side, RL is very sample fission. So you can take an old model and then really create an RL environment for the specific use case you care about and scale capabilities for that. So I think good example of this was for example cursor with components that was what's widely believed or known to be a scale up of an open source model. And RL environment was cursor. They basically just gave it the tools and the things within the harness and application of cursor itself. So they trained basically that model on getting really good at using cursor and I think we'll see the same play out across all the applications. Where basically the broader theory is, is like every application, every company will be an AI company or AI native and will have an opportunity to really post train and use RL to make the models work specifically on the application. So even if you take examples of like say figma, right? Like if they want to make their platform agentic, really they need to create an RL environment around figma and post train on that environment to be able to serve that with a. And Figma like kind of like out of the box. Like the closed models won't be perfect at like renavigating and making those applications agentic. So I think that that's the broader theory. I think really it's also, it's like, it's so like the capital requirements are much, much lower than I think the big labs want to believe you, like in the sense where it's like you can for like hundreds of thousands of dollars like post train a model, right? It's like to be much better on your application and then also you are able to like serve the model.

A (172:10)

Cheaper. One weird trick, post train a model for 100k and create a better. So I mean that's basically what you're saying is that if I'm Figma as an example and I could use a frontier model that's really expensive and beefy and it knows everything about. It knows some stuff about figma, but it also knows about the Roman Empire, I can go in RL on just my particular application and have a smaller model that's fine tuned on open source. Exactly, Open source model and get better performance than with the big beefy do everything omni model. Is that.

D (172:45)

Right? Exactly. And I think really you get better performance but also at a lower price point potentially because you can really specialize the model to be extremely good for your use case. So I think you could see this with cognition posturing their own model with first up posturing their own model composer and composer is also much cheaper to serve. It's much faster. Same for the model Cognition was building. I think what we're seeing and we've started to work with dozens of customers on helping them basically do pulling in rl. I think we're basically starting to see a huge pool in terms of enterprises realizing that if they want to get a specific capability, RL is the way to get it and ultimately enables them quite capital efficiently to train those models and serve those models and then really get to a point where even in deployment all the interactions from the user help improve the model. So I think with the cursor example. Example, for example cursor tap interaction. Every yes and no that a user gives to the model is updating the model every two hours. That's what Dwarkesh talks a lot about. Two hours online rl. Yeah, they're basically continuously training the model in a two hour interval and pushing updates every two hours to Cursor tab. So basically every user using cursor for the last two hours is being pulled trained on so to speak with kind of like an online RL loop. I think that's something which we'll see more and more that basically applications will do their own rl, their own post training actually then and since like really how we un. Hobble basically towards AGI it's like the question is like why haven't we say automated like specific valuable knowledge work yet? And I think the answer that also like SH was speaking about for example on with the example of like autom automating Texas and accounting formula, right? So like no one has really created RL environments, post trained on them and then serve the model in the application where the end user is and then ultimately the end user's interaction with the agent can improve the model further. Right. So I think that's really the paradigm that we see play out, which I think is really a paradigm of like thousands of models or like millions of models that like basically continuously improve and where actually the applications win to some extent through distribution. Like ultimately they own the end customer interaction, right? Where they like even the cursors and permissions have like an advantage there over folks who are basically just model providers and who don't interact with like millions of developers. And I think we'll see the same play out like across all the different applications and it's something like talk about also in the context of like Copilot and Microsoft, they own distribution, they can create the cursor for Excel, for PowerPoint or other.

E (175:24)

Things.

D (175:24)

Right. And then whole strain on all those interactions. So I think we'll see this play out I think across all the different verticals and I think it's like a broader trend of just like every company needs to become AI native, right? And also to keep owning the distribution, they don't want to give all of it up to the big.

A (175:40)

AGLMs. Yep, that makes.

B (175:42)

Sense. We got a question from our intern Tyler, if we can shoot over.

D (175:47)

There. Yeah, I guess I saw you.

G (175:49)

Guys talk about this a little bit.

B (175:50)

On but is there any point of you guys training your own base.

H (175:55)

Model?

D (175:57)

Yeah. So basically I think one interesting release in this Context was today we actually released, we supported RCI in their base model release which is kind of like catching up to the Chinese base models. So basically we supported them in training a small moe based model which achieves pretty sort of results, results. So we released that I think like an hour ago with them and and we're actually now like ramping up with them to towards like a much bigger base model so fully like pre trained from scratch. So we actually just had like 2000 B3 hundreds going live I think yesterday to ramp up like towards like a much bigger pre trainer. And I think like it's really, I think like the broader pattern is like thin kind of like Llama had some reorgs and changes and Misrol became sort of like a forward deployed European enterprise play or something. I think there's really no one left outside of China right now to go end to end in the most stack. I think others like Reflection I think are trying to also pick that up but I think there's very few players I think outside of China. So I think that's our broader goal is really serving like the world more global globally but also like the west and the us US with like an end to end.

B (177:14)

Pipeline.

D (177:14)

Right. It's like from data to pre training to training to post training like the full stack and making that accessible to like enterprises and people who are like trained on models. I think that there's a huge I think pool where a lot of enterprises or even like, like sovereign like nation states etc. Like you can't train on Chinese open models but they also, they can't rely on closed models. So I think there's a huge gap in the market right now that we're trying to fill of really like serving kind of like that whole.

A (177:41)

Segment. Do you have anything else.

B (177:43)

Jordy? No, this is.

A (177:44)

Great. I want to know one last question about, you know, what will the market structure look like in maybe a year or two around like implementing these RL environments for companies? Because when I, when I see you know you say every, every company is an AI company, I believe that's somewhat true and I believe every tech company, maybe every founder led tech company under 10 years might be able to say okay yes, we're going to go and train, fine tune a model and turn our application into an RL environment. But if I'm the Coca Cola company, I might not be at that level of going and building RL environments for every business process. I'm probably more of a buyer of this army AI as SaaS almost. So how do you see that kind of breaking out? How do you see a truly legacy non tech company adopting a fine tuned LLM or an RLED model?

D (178:49)

Totally. No, I think there's early adopters and later adopters. I think Coca Cola might be more like later doctor and might not need to adopt it early on. But I think they are even are adopting it just like in less obvious places. It's like ultimately I think they're initially just like using the AI tools that use us for example in a sense where it's like say customer service. Right? It's like is a perfect example of where you get a lot of gains out of post training. And then they might put basically the AI native customer service platforms might use us to post chain using Coca Cola data to serve them a better model. So I think what we'll see play out I think is really just like making a lot of that so accessible to your point that it feels more like using SaaS where I think one element of it is we are launching also our whole RFT platform basically and offering to make it extremely easy and plug and play. But then there's also forward deployed element where you can outsource a lot of that stuff to our team. And I think the other element is really we're walking, walk in terms of making our own thing kind of agentic and autonomous that you could basically just use like an autonomous AI researcher to do all of it for you. That you basically just plug it into your system and the AI even creates AI for you. I think that's the next paradigm is really making in general training models like fine tuning models, post training models as accessible as bytecoding is today in a sense where it's like I think with vibe coding literally every human on earth is able now to code some stuff and I think we'll see the same play out with AI over the next 12 months. And that's one of the big things that we're playing into. We're kind of like pushing towards autonomous AI research where AI can do most of it for.

A (180:31)

You. Well, thank you so much for taking the time to come and talk to us on the show. Congratulations on all the.

B (180:37)

Projects. Thanks for having me.

A (180:38)

Very. And we will talk to you.

B (180:39)

Soon. Great to see you.

A (180:40)

Vincent. Goodbye. Have a good one. Let me tell you about Privy. Privy makes it easy to build on crypto rail, securely spin up white label wallets, sign transactions and integrate on chain infrastructure all through one simple API. And I'm also going to tell you about adquick.com out of home advertising made easy and measurable plan. Buy and measure out of home with precision. Our last guest of the show is Ben Hylak. Did he do the Jaguar rebrand? That's.

G (181:05)

Ben.

B (181:05)

Welcome. And we'll follow him.

A (181:06)

Forever. Grab a seat, hang out. Good to see. Oh, you brought hats. Fantastic. Fantastic. Thank you. Please grab a seat. Introduce yourself, introduce the company. What's the.

E (181:19)

News? Yes. So my name is Ben Hyluck.

B (181:22)

Yes. Let's take a second for the.

A (181:23)

Flow. Fantastic. Thank.

E (181:25)

You. Thank.

B (181:25)

You. It's kind of like a vintage Silicon Valley flow. Somewhat of a lost.

E (181:30)

Art. I appreciate it. You guys have great hair as well. You know, I felt a lot of pressure. You'll notice I'm not wearing a hat today, because I did notice, actually. I kind of. I discovered a blow dryer, I think, around nine months ago, 10 months.

B (181:43)

Ago. So that was a big blow dryer in your.

E (181:45)

Life. It's been. Never been the same since. But, yeah. My name is Ben Hylak. As you guys know, I'm the CTO of a company called Raindrop. So, really, simply put, we monitor agents in production. So we were building a product ourselves probably around two years ago now, which was like a coding agent, and we realized that there was just this huge gap of, like, if you're using Sentry, if you're using traditional analytics, you know, they're covering, like, the things the users are clicking, and almost everything that's happening in your product, if you're making an agent is just not covered. So you just have no idea what's going.

B (182:20)

On. These agents are going absolutely.

E (182:21)

Wild. They're going.

B (182:22)

Crazy. They're going.

E (182:23)

Haywire. You know what's been insane? I think one of the things that's been, like, really kind of critical to our growth in the last couple months has been realizing that as agents get better, this problem gets worse. So that was not necessarily intuitive to us in the beginning. You think, oh, well, agents are going to get better. Maybe this problem becomes less important. But it's like, actually, as they become more capable, they can use more tools, more valuable. Exactly. So, for example, if you take a company like Replit, it's like maybe a year ago or two years ago or when they first launched, you couldn't quite get as far. Maybe you could just get a personal website or something. And so if it messes up, at that point, it kind of gets stuck. It's like, okay, maybe it's not the end of the world, but now with repl.it you're able to build just like real applications, like people are building real production applications. So now if you get to a point where it gets stuck, something goes wrong, suddenly it's like it's a real issue. So that was not intuitive.

A (183:15)

Before. So agent's a pretty overloaded term at this point. I think of, you know, when I fire off a deep research report in ChatGPT, that's an agentic workflow to some customer service agent that's happening completely behind the scenes and the customer might not even know that they're dealing with an agent. And then there's coding agents, There's a few that you mentioned. Are you dividing the market and trying to focus on an early landing zone first, or do you want to do all of.

B (183:45)

Those?

E (183:45)

Yes. So we focus on essentially, and I will say I agree, the word agent is overloaded. We're very hesitant to use it for a really long time, and then we realize it actually matters, of course. So we focus on products that have some sort of user input and some sort of assistant output eventually. So that's sort of our focus. So what we're not focused on is, for example, like, we're not going to focus on like specific, like ML pipelines or things like maybe like translating text or like summarizing text even. It's like we want to see like the user, the user is sort of like, has some sort of request, the assistant is responding to that request. And we do. You map essentially everything that happens in between that initial user input and to what the assistant actually.

A (184:28)

Responds. And then what's the go to market for.

E (184:31)

You? I mean, it's been a little crazy actually. We've had a lot of inbounds, so some of our biggest customers have been inbound. A lot of it has been like when we first launched, I think, I guess this was like six months ago or seven months ago now. Agents weren't as big of a deal. And so I think in the first month or two we had a lot of customers that were like, okay, like, I have evals, we'll need this. Yeah, but it didn't really make sense for them. And a lot of them came back in the last like, like a month or two after that. And we're like, holy shit. Okay, now I get it, we need you. Like, so it's actually been a ton of inbound. We do what we don't really pay for advertising, anything like that. You know, if we see a really crazy failure in the news, we'll reach out to that company, obviously, and be like, hey, this is something we can Help.

A (185:12)

With. Sure, sure, sure. How are you thinking about the target? Like the best type of customer? Are you segmenting it by size? Do you want to go enterprise upfront because they're implementing agents at scale? Or are you more likely to see immediate results of the startup that just kind of gets it and they can hop on really quickly? Like how are you thinking about prioritizing if you are at.

E (185:34)

All? Yeah, it's a really good question. I think that, that we really look at the entire range and I think that we see and have always seen startups as being a really core part of keeping our company.

A (185:44)

Healthy.

E (185:45)

Sure. I heard a while ago that posthog has this metric where they look at what percentage of YC companies in every batch are using.

A (185:52)

Them. Sure, sure.

E (185:52)

Sure. And so that's why we started with startups. They're able to move faster. So for example, when a new model comes out, just to give actually a very specific example. So GPT5 introduced intermediate reasoning. Right. It was kind of one of the first models to do this where like it's going to make tool calls, it's going to look at the results of those tool calls, Think about it and then make more tool calls. Take that, think about it. You know, more tool calls. It sounds small or subtle, but actually it kind of means that you know, if you architected these, your system, your pipelines in the wrong way, you just couldn't use that like and it really helped. So where startups will just like just throw everything out the next day. Right. And they'll ship a whole new thing in a week. You don't see like, you know, like if you look at like the biggest enterprises, they're not going to do.

A (186:40)

That.

E (186:40)

Sure. So you can learn really fast by with startups. That being said, on the flip side, I think that the problem we're solving is actually most painful for.

D (186:48)

Enterprises.

E (186:49)

Right. It's like the most critical high stakes environments are where like failures cost the most in every single.

B (186:57)

Sense. Yeah. How much of categories of agents that you're excited about that are maybe under hyped today? Coding agents. Coding agents are like sufficiently hyped. I think coding agents.

E (187:09)

Are. And for good.

B (187:09)

Reason. For good reason. But like, and maybe they're deserving of more hype. Yeah, but, but what, what other category? You know, I think, I think people have been saying sold on the AI bdr. Yes. Haven't exactly. Maybe companies are getting a ton of value from it and they're getting so much value they don't want to come on TVPN and talk about it because they don't want their competitors to know. And then obviously, like CX feels sufficiently hyped. But what else are you.

E (187:40)

Seeing? Man, there's so many different things. Like I think your speak, for example, language learning. I think the better, better, as models get better, that experience just actually starts to become really, really, really viable. So like, that's an example of something where it's like, yeah, it existed a year ago, it existed two years ago. But like, as voice models get better, as the models themselves get better, it's actually not just like, you know, if you try to use ChatGPT, for example, to learn a language, you sort of can. But if you ask it to like critique you, for example, it just never will. Like, if you say something wrong, it just isn't going to stop. Stop and be like, hey.

F (188:13)

Look. Actually like still glazing even if you're absolutely.

E (188:16)

Right. Yeah.

A (188:16)

Exactly. Donde esta la bibliotheca Is the most complicated Spanish.

E (188:21)

Sentence. It will, it will, right? You're fluent. Exactly, you're fluent. It's like, yeah, you're pretty much good to go. And even if you can get it to the point where like, if you can really, really, like prompt it into critiquing you, it'll just like start critiquing everything you know, which is also not what you want as like, you're learning a language. So like it turns out, and I think we see this with a lot of products that like getting some something right is actually a lot of details and really, really understanding that domain. So I think we're seeing that in literally every domain. Like whether it's like marketing, whether it's like even just like the idea of having a personal assistant, like, notably we don't have that yet, which is crazy, right? Like, we have these assistant models, but then none of us actually have an assistant. We can just chat and be like, hey, send this email. Yeah, right, I don't. And so I think. But I think we're starting to see products actually like nail that like small are.

A (189:04)

Mostly. But how are you thinking about. Just. I don't know if, I don't know if like, if you're sentry for AI agents. Does sentry actually handle this? But just types of AI failures that happen for more infrastructural reasons. So just the GPUs are on fire or like there's just not enough GPUs in this particular cloud and you just see a spike in demand and so you just can't provision more like those types of more, more tactical errors. Do you help with.

E (189:29)

That? Sort of would be the answer. So I think it's actually really interesting is that one thing we realized about evals is that they don't catch those sort of issues. Like, you know, you're kind of testing just like the model, what is the model responding. But then there's all of these things that happen in between. Like I remember really, really early on when we launched, one of the issues that a customer caught was like their file upload was broken. So a bunch of users all started complaining about like, oh, like the file upload's taking too long. It's like, okay, well that's not like an AI.

D (189:56)

Problem. But it.

A (189:56)

Is.

E (189:56)

Yeah. And so we see that with like tool calls. We saw one of our customers had an issue, sort of what you're saying, which is that like they started having like, they have their own GPUs. They started having like an infrastructure error and it was mixing up responses between users. And so users all started complaining like, hey, that's not what I. Like, what are you talking about? That's not my sense. It was like an increase in that, in.

A (190:17)

Like. I don't know if you're talking about meta, but I think that happened in.

E (190:19)

Meta. It wasn't meta. They're not one of our customers yet. But, but there was a situation.

F (190:23)

Right, that's people could.

A (190:24)

Share. It was not that bad, but it was something like I could share my chat with you, but if I shared it with you and I didn't know that I was sharing it, it would go out everywhere. And so, yeah, stuff like that.

F (190:32)

Happens.

E (190:33)

Totally. There's all these sorts of things so you can actually catch those sort of problems. It's actually one of the, one of the things is like that ground truth is actually really, really important because if you just see like a few errors, like let's say you have tool, like your agent calls tools like, yeah, that's going to error once in a while, right? Like, that's might not be the biggest deal, but especially once you, if you can see when it actually starts to affect users, like that's really, that's really.

A (190:54)

Powerful. Yeah, yeah, that makes sense. What about degradation of models under the hood? I feel like people. I don't know if it's just a meme. I've noticed it here and there. I'm not, I'm not benchmarking everything every night, but it does feel like that sometimes, right? It feels like sometimes I'm like, wait a minute, I used to respond in this tokens. Now it responds this way. It used to look hd. Now it looks standard.

E (191:18)

Definition. Like, I agree with you. I think. I think it's real. I know that I can't say too much. I know that at least on one occasion that I. I think people were led to believe that there wasn't a thing. I know that there was. Okay, so that's it. You know what I mean? I can't say who. It's a big company, and because I noticed this and I thought it.

A (191:38)

Was like, whose hands.

G (191:39)

Were. I can't say which.

A (191:40)

One. I caught some red.

E (191:41)

Hands. Yeah, exactly. And, like, it was like. I thought it was a cursor problem. It was like, some really absurd behavior. And then I went into ChatGPT and it was doing the same. Oh, I just said. But anyway, yeah, like, I think that the reality is that, like, every single one of these providers are, like, having these sort of problems, and they're trying to optimize costs. They're trying to, like, make changes. So I think it's.

A (192:00)

Natural. And some of them. I understand where I'm like, oh, okay, well, yeah, realistically, I haven't used that in a long time. I came back. I kind of churned. I don't really mind that you put me on the lower.

E (192:08)

Tier. Yeah, yeah.

A (192:08)

Yeah. I just hope that for the people that actually, like, went and built businesses around this that are using at the API level, that are hopefully paying for the service at a high gross margin to you, you're not degrading the service behind their backs 100%. Right. So anyway, who did the deal? Anybody we.

B (192:25)

Know? You want to hit the.

A (192:27)

Gong? You want to hit the.

E (192:27)

Gong? Oh, let's do it. Yeah.

A (192:28)

Yeah. Hit the gong. Tell us how much you raised. How much did you raise? How much did you.

E (192:35)

Raise? We raised. So we raised $15 million in.

A (192:39)

Total. From.

E (192:39)

Who?

A (192:40)

Lightspeed. Who did the.

D (192:41)

Deal?

I (192:42)

Bucky.

A (192:42)

Yeah. Let's go. Let's hit it.

B (192:45)

Again. This one's for.

A (192:47)

You. Yeah. We're big fans of Bucky over here, so it just wants to get him a shout.

E (192:53)

Out. Us.

B (192:53)

Two. Us.

D (192:54)

Two. Us.

E (192:54)

Too. I think the moment we met him, we're like, okay, like, he mashed our energy, like, great.

A (192:59)

Vibe. Yeah, yeah, he's.

B (192:59)

Doing. How's building the team.

E (193:01)

Going? It's going. It's going. I think we're really, really picky. We've realized, and so it's really hard. And I think hiring in San Francisco is really hard. We have a great team. It's Honestly, really, really small.

A (193:14)

Still. Well, if you want to get out of San Francisco, you could book a wander with inspiring views, Hotel, gritty menus, dreamy beds, top tier cleaning, 20% concierge service. It's a vacation home, but better. You could do an off site.

E (193:25)

There. We could do our off.

A (193:26)

Site. That's beautiful. Do your.

B (193:27)

Offsite. I once used a team off site as a recruiting tactic. I said, we are going on an off site in two.

G (193:33)

Weeks.

B (193:33)

Okay. Posted a.

A (193:34)

Picture. Oh, yeah. You want to.

B (193:35)

Come? We got an amazing.

E (193:37)

Tire. We'll do.

A (193:38)

It. I'll do.

B (193:39)

It. Create some.

E (193:39)

Urgency. We're doing it. So if you're watching right now, we'll. I'll post the picture soon of. Of the.

A (193:44)

House. Okay. Fantastic.

E (193:45)

Fantastic. But we have an.

B (193:47)

Amazing. Yeah, no, yeah. I figure if you're. If you're picky and you're in San Francisco, it's like the most ruthless, like, talent war.

E (193:54)

Constant. You know, the other thing is that, that I think when you hire amazing people, they have zero tolerance for working with people that are not amazing. And so I think you can't even fool yourself as a founder if you start like, you know, whether you're work telling whatever it's like if they're just, you know, if it doesn't fit. Like everybody, everybody knows and feels that.

A (194:12)

And. Have you had to bring anyone soup? Are you familiar with.

E (194:14)

This? I'm not familiar with.

A (194:15)

This. Okay, so apparently the AI this is from Ashley Vance. This is a soup scoop just dropped on Core Memory on the podcast. So he had Mark Chen, OpenAI's research chief, on the show as part of a post Gemini 3 sit down to get the update from OpenAI. And he said, I knew the AI talent wars were rough, but not this rough. Zuck is out there apparently delivering handmade.

G (194:37)

Soup.

A (194:37)

Wow. And OpenAI has soup counters. And so I.

B (194:42)

Guess. Wait, they count how.

E (194:44)

Many. Wait, sorry, soup counter. What are you talking.

A (194:46)

About? I don't even know what this means.

E (194:48)

But. Oh, I.

A (194:48)

See. No, no, I think it's like a.

B (194:51)

Count. Like a. Just like a place where you have.

E (194:53)

Soup.

A (194:54)

Right. Aggressive. What exactly does this tit for tat? We can. We, We. We can play this on the show later. But. But.

E (195:01)

Yes. I mean, no, no, My, My. My partner has come up with mules for someone, you know, creative sort of thing works. You know, we, we do typewritten. I'll write a note on a typewriter. You know, when we do our offer letter like that. So that add something a.

B (195:15)

Little. Are you messing with.

E (195:17)

That? I Love. No, I'm serious. I love.

A (195:18)

Typewriters. No, I, I, I like that. It's just a.

B (195:20)

Way this message.

A (195:23)

All the text is AI generated, I'm.

E (195:25)

Sure. Of course. Yeah. I'm just copying from.

F (195:27)

Chat. I think it's like a little bit of.

A (195:30)

A. You're not just the newest hire, you're a.

B (195:33)

Revelation. This is a.

F (195:34)

Statement. Yeah.

A (195:34)

Yeah. Having fun. Well, that's great. Congratulations on all the progress. Us Very excited. I'm sure you'll be back on the show soon, giving us plenty more updates. And it's been fun because, I mean, I believe that we started tracking your journey via your viral joke post about doing the drag warp or something. But we've always had fun.

B (195:56)

Featuring. Great to have you here live in.

A (195:58)

Person. Live in.

E (195:59)

Person. One year ago today, I remember roughly one year ago, I was sitting in a parking lot and I was listening. It was the first time I ever heard of you guys. You were reading, like, one of my tweets. It was just so surreal that, like, people from the Internet are reading my tweets. Like, one of our customers sent it to us, actually. You printed.

D (196:15)

Out.

E (196:15)

Yeah. So I called my mom today. I was, like, telling her. I was like, hey, I'm gonna be on this. I was like, you're not, you're.

A (196:19)

Not gonna know what it.

E (196:19)

Is. But remember that those guys that were talking about that tweet, this was.

A (196:22)

The whole, this was the whole shtick, was like little love letters to Silicon Valley folks. Just like little messages of just, hey, we found something that you did. Fun because anyone can like, anyone can read post, you know, it's easy to send a small thing. It's very hard to actually print it out. Sit down, talk about it. But we appreciate your post and we appreciate you coming on the show and hanging out today. So thanks so.

G (196:43)

Much. Thank.

A (196:44)

You. Talk to you in just a second. While he's walking off, let me tell you about getbezel.com shop. Over 26,500 luxury watches. Super intelligent, dedicated in house by Bezel's team of experts. I also need to tell you about eightsleep.com exceptional sleep without exception. Fall asleep faster, sleep deeper. Wake up energized. I had a rough night. Kids have been all over the place. But I still got 92, 98, 98, 98. That is.

B (197:16)

Remarkable. Well, is there, There are a bunch of. Yeah, we'll.

A (197:21)

See. If you want to go through some breaking news. Buco Capital bloke is on the timeline. You can feel the panic behind the urgency and intensity with which people are defending Nvidia. It feels visceral and quite intense. You can tell how much is riding on this. It makes a lot of sense. What. What else did you want.

B (197:40)

To. I thought it was notable. Pager duty has fallen to a $1.1 billion market cap at 500 million of ARR. So trading. They're not growing anymore. They're trading at 2.2.1x ARR. It's profitable according to Jason Lemkin over at Saster. So, yeah, rough time out there if you're not growing regardless of the revenue scale. Two days ago we shared that Enron back. November 29, 2001. Nvidia replaced Enron in the S&P 500. I saw this post go out from our incredible team and I immediately googled to.

A (198:19)

Fact. I was like, there's no way. Someone has made a terrible mistake on our team and we are doing fake news unironically now. We used to have some fun, but apparently this is.

B (198:30)

Real. It's real, it's real. Hamilton was like, I'll take that.

A (198:33)

Spot. November 29th. Obviously that's not how it works. It is much more mathematical than that. I believe Standard and Poor's picks the luck largest companies and after certain ebbs and flows of the market, they swap folks in and out. But this went pretty viral. 5,000 likes. But what is really interesting is of course the Nvidia Enron like comparisons are just so silly to me. Obviously it's like, you know, the discussion is like, will it go from being the best business in the entire history of the world to being like, like, you know, somewhat competitive and have to deal with like minor competition from other people. It does not seem like it's some ridiculous Enron situation that's like so, so insane. People are just having fun with that headline. But what is incredible is this, this branded shirt he's wearing. Look at this.

B (199:26)

Thing.

A (199:27)

Fantastic. So awesome. I love.

B (199:29)

It. Not enough people trying to go snipe vintage Nvidia.

A (199:32)

Merch. It's a great shirt. It's a great look and I feel like it's gotta make a comeback. The button down. This is the pre Silicon Valley. I'm just in a T shirt era. But it's post suits. You know, it's like, we're not suits. We're working in technology. We're still gonna throw on a collar, but we're gonna dress it down a little bit. No.

B (199:51)

Tie. Guys, scroll up. Scroll up on this for a second. Yeah, I'll keep going. Keep going. Oh, who's not following? Tyler, you gotta follow the.

A (200:00)

Account. No, this is not my account. I think this is a. This is more of like a burner account.

B (200:04)

Situation. Oh, it's a scraper that we use to.

A (200:06)

For. It is. It.

B (200:08)

Is. You got to correct that. Tyler. Come.

A (200:10)

On.

B (200:11)

That's Gorkham over at Fall had an absolute banger. This was a chart showing ASML sells fewer than 500 units per year and generates 37 billion in revenue. Is there any company in the world with a wider moat? And Gorkham says series A pitch meeting. Sorry to cut you off. Off. But what happened in December 2024 since there's like a slight dip in the.

A (200:34)

Chart. Yeah, what. What did happen? Why did their revenue drop in 2024? I actually don't know. Is it just so much pull forward from 2023 or.

D (200:47)

Something? I don't.

B (200:49)

Know. Maybe they. They were developing some hubris. They decided to get.

A (200:53)

Complacent. Yes, I mean, I certainly understand. Understand the concept. Okay. According to the CEO, customers in Taiwan had delays and weren't ready to take delivery yet. And orders got pushed back. At the same time, China raised to get as many machines as possible before export controls tightened. Okay, that makes.

B (201:07)

Sense. Let's hope that Sash Zatz says Oxford Dictionary didn't get the memo. Apparently Rage bait named. I think. I think.

F (201:18)

It. No, no.

A (201:19)

No. It's appropriate that it would be the word of the year. But it is so funny that you. You posted this and then. And then Oxford.

B (201:27)

Dictionary. Yeah. So this is true. Rage. According to the BBC. Rage bait named Oxford word of the year 2025. It certainly feels that way on the.

A (201:36)

Timeline. Your post. 1 million views on this. 3.6 thousand likes people. Really? This really set the agenda for a little bit. Wow. Congratulations on what a banger essay. Should TVPN do a word of the year? I like that. Or.

B (201:53)

Motion.

A (201:54)

Motion. Motion might be our word of the.

B (201:56)

Year. Word of the.

A (201:58)

Year. Motion's a pretty good word of the year. Motion named named word of the year 2025 by.

B (202:05)

TPP. If you have it. You'll.

A (202:06)

Know. You'll know. We'll call.

B (202:08)

You. Tyler has.

A (202:10)

Motion. In other. In other breaking news, Keith Raboy is taking shots at Airwall X. Airwall X is now on the other side of a billion dollars in ARR. What I love about this chart is that isn't that we hit a.

B (202:22)

Billion. This is the founder.

A (202:23)

Jack. It's how fast the business is accelerating. Took us more than 6 years to 100 million. Ar. What does Airwallex do.

B (202:31)

Exactly. I think they provide payment rails for a bunch of American fintechs. Handle Inter.

A (202:37)

International. Okay, okay. And. And so Keith Raboy been on the show multiple times, says cool growth chart. Have you disclosed to US customers like Rippling, Bill.com, brax, Navon, that you're quietly sending their customers to data to China? Air Wallace has become a Chinese backdoor into sensitive American data, like from AI labs and defense contractors. You must already know this, but you're China. Based on OPS infrastructure, and investors create legal obligations to assist with CCP espionage. Upon request, through Airwallocks, Beijing can assess supplier payments for AI labs so they could know who's using what models. Payroll data for defense contractors, personal data for employees abroad. That's obviously not good. Obviously, many companies do business in China and that's not inherently a bad thing. But you, your company has become a guaranteed vector for data transfer to the Chinese government, and that's a different thing entirely. You have multiple points of vulnerability. People, legal structure, cap table. What's happening? You route global payments for US Companies in critical sectors without disclosing that you're under a Chinese jurisdiction. You moved your HQ to Singapore. Well, that seems like a step in the right direction. Maybe, but your largest operational footprint is in China. Okay, no. So maybe one step back. And one step back and hundreds of your engineers in mainland China touch production payment systems. You are subject to Chinese law that requires Air Wallix employees to support CCP intelligence requests and quietly hand over data when asked. You hid this from your customers, but you are well aware of your obligations to China, and that's why you insist on protection of Chinese data access to your contract. Thanks to you, the Chinese government now has direct, direct, covert, legally enforceable access to sensitive financial information. This is a big story. This is a crazy scoop from Keith Raboy, and I will be interested to see where this goes, how quickly they can remedy this. This popped up a couple years ago during the Clubhouse era. The Clubhouse backend, I believe was. Was at one point, you know, was working with a Chinese company. Or maybe. Maybe it was that there was a company that did like, peer to peer audio streaming that was based in China. And so if you were building a competitor, you might use that.

B (205:04)

Company. Yeah, and so I became familiar with airwallocks through the 20 VC episode that Harry did with Jack the.

A (205:13)

Founder. Is it ripping? I mean, it seems like the business is doing really.

B (205:16)

Well. Yeah, yeah, yeah. Oh, well.

A (205:20)

Anyways. Well, what else to.

B (205:22)

Say? We got to get on with. With Menlo.

A (205:26)

Park. Okay, well, thank you so much.

B (205:29)

For listening hanging out with us.

A (205:30)

Today. We will see you tomorrow. Please leave us five stars on.

B (205:33)

Podcast and Spotify the break the Thanksgiving break was absolutely brutal for us. I would say every single.

A (205:39)

Day. But hopefully you had a great.

B (205:40)

Thanksgiving. Wake up and just twiddle my my my thumbs wishing we were podcasting. It's great to be back. Hope you had an amazing break or or a little holiday and we will see you.

A (205:50)

Tomorrow. See you.

B (205:51)

Tomorrow.

A (205:51)

Cheers. Goodbye.