https://www.listennotes.com/e/042b10002ff543e2ab9285f3aec7b679

You'Re watching. Today is Thursday, June 5th, 2025. We are live from the tbpn Ultra Dome, the Temple of Technology, the Fortress of Finance, the capital. We got to work on that because we're working on selling the naming rights, baby. This place is going to be branded. We're gonna sell the windshield. We're selling the windshield. So we got it. We got to keep. growing the intro. But we have a massive day today, a little bit of an AI day today. We got folks from Google, OpenAI, Anthropic. We got the former CEO of OpenAI, Sequoia, Stanford, Google, ex-investor, and we got pretty good coverage. We hit almost everything. Should go on a whirlwind tour of what's going on in artificial intelligence. I'm excited to dig into the state of affairs in the foundation model race. We're going to go through the tier list of what companies in AI have the mandate of heaven. We're also going to go through some of the deep research products and hopefully get into some of the more cutting edge use cases for AI. So we have both some deep research folks coming on, and then we also have some folks that are working on video generation and video game generation and a lot of different applications. We're going to cover the granola story. We're going to cover what's going on with wind surf. And so it should be a great day, but let's run through some news just to keep everyone up to speed before Sean Maguire joins in 13 minutes. So first off, ramp time is money. Save both. Save both. Easy use corporate cards, bill payments, accounting and a whole lot more all in one place. Market clearing order. Circle went public. The CEO is coming on the show tomorrow. That's very exciting. And, Jordi, you have the news. I will read a little bit from Jeremy the CEO. I'm incredibly proud and share and thrilled to share that Circle is now a public company listed on the New York Stock Exchange under Circle. Brian Armstrong post, congrats to Jeremy and the entire Circle team on your IPO and reaching 30 trillion in lifetime USDC volume. Let's hit that goal.

The big T. Incredible. The stock is up massively. It was priced at $31. It is trading at around 85 as of now.

That's fantastic.

We'd love to see it again, as many people would expect. Bill Gurley is going to be very unhappy with that inefficient pricing.

He hates a stock. He hates a pop after the IPO.

It's good for the IPO window, which we always want to be open.

For sure, for sure. Andreessen Horowitz Executive Chair Trey Stevens tells Ed Ludlow that the company has closed a new funding round of $2.5 billion in a deal that more than doubles the defense startups valuation to $30. 5.5 billion. This is from Bloomberg TV. Congratulations to the Anderl team on the massive up round. You'll love to see it. We got to hit the Gong for Andrew.

Do it again. We have good contact. Good contact.

Very exciting. Bunch of news from Kevin Wheel over at OpenAI. We will be digging into this with Mark Chan today when he joins, but deep research can now share across GitHub, Google Docs, Gmail, Google Calendar so you can integrate everything and it can do research on your files. That's gonna be a lot of fun to talk about. This is also potentially threatening for granola. And so we're talking to a granola investor about what the reaction will be, where the direction of that company might change or not. But if you're designing a tool for artificial intelligence or one of these products or any tool really, go to figma.com, think bigger, build faster. Figma helps design and development teams build great products together. Go to figma.com It is the backbone.

Of the TVPN brand.

It is, it is.

And we would not be able to make the show without it. Yes.

And so we are going to be digging into today's obviously artificial intelligence day in some ways. We're digging into artificial intelligence today. We're also very interested in talking about VR and content and augmented reality. And then there's a story in the Wall Street Journal today about Meta is in talks, not advanced talks, just regular talks, regular old talks.

Chit chat.

But they're talking to Disney, they're talking to A24 about content for a new VR headset. And this is my number one question about VR, when the iPhone, everyone's looking for the VR, the iPhone moment of VR. When the iPhone debuted, what was it? It was first and foremost a phone. It replaced your phone. And so I've always thought that, The path to true VR adoption was just saying, we're only going after your TV. The next generation of 20, 22 year olds, when they get to college or post college and they're in their first apartment, they are just not going to buy a big flat screen TV because we have solved that specifically with VR headsets. Exactly. Or, you know the guy with multiple monitors, three monitor setup, the production team back there, we can go to the production camp, show you all of the different monitors that they have. What if they could be wearing VR headsets? They could have seven monitors one day, ten monitors. Let's give it up for the production team.

Let's give it up for them.

So the idea of a drink cam.

We got the drink cam.

Thank you to Matty Arino. Thank you to Andrew Huberman for inventing drinking things. Caffeine.

The whole team over at Mattina.

Yeah, they're crushing it. And so this idea of doing one thing really, really well before going into you know, trying to do a little bit of everything, like trying to be a platform.

Exactly.

You don't have to be a platform. If you solve one thing really, really well, the iPhone didn't wasn't a platform. The first iPhone was not a platform, didn't have an app store, right? It just had the ability to listen to music. It was an iPod, it was a phone, and it was an internet communicator, just a web browser. And so the Meta is looking to Hollywood for exclusive immersive video For premium device. Now, my kind of hot take here is that-- Whiskey's cooking. I know, I know. All of our boys are coming together, we're cooking up something amazing. I'm really excited. We're not gonna be able to get that much information on this soon, but I don't even know if they need that much immersive content. I think a lot of it is just, hey, every single Meta headset should just ship with the Matrix pre-installed for free. It's like, how much would that possibly cost? It's an extra $2 to rent or something you could have it pre-installed. So it's just like, you can put it on and there's like 10 movies that are pre-loaded, you can just watch Movies and the movies are great and you're in a really nice theater and it just comes pre-installed and it's all great because that was that was the my my vision pro experience was very much yeah film driven.

I would take it a step further and I think they would have to basically create an entire catalog like I don't know if the matrix by itself is going to be enough of a draw to say I'm going to spend hundreds of dollars for the average person.

No, no, no, it's more about like the pre-installed apps so like the iPhone.

Was a really good phone No, but.

I'm saying-- Really good iPod, but it also came with like a calculator app that was decent. And so you need a few of these things that are just really easy to access, really easy to pull off the shelf. And ultimately, I think Meta needs to catch up to Apple in terms of the Apple TV movie store and making sure that all the streaming providers are really on there in a valuable way. Obviously, it's important to go to immersive eventually. But I think the path to immersive might be just, wow, I have a home theater in my studio apartment now. Anyway, we'll dig into this more. We'll talk to more people. Maybe I'm wrong. Who knows? But the high points from this article in the Wall Street Journal are as follows. Meta is seeking exclusive content from Hollywood for its upcoming premium VR headset, Loma, set to rival Apple's Vision Pro. I'm super excited for this new VR headset, I think. It's gonna look fantastic. I think the resolution is gonna be insane. And obviously there's a lot of focus on augmented reality and Orion and AI and glasses. But there's still so much work just to do just to bring VR into just a normal consumer experience. And what's interesting is that I think Apple really broke the seal on like, yeah, like, you know, people are used to paying $1,000 for a phone and $2,000 for a computer and maybe $4,000 for a headset, isn't that crazy? So what can Meta has been been hanging out in like the 3, 4, 500 range. If you take the reins off and say, hey, yeah, yeah, it's fine to spend $1,000 on this thing, you could get something really, really interesting. Totally. And so Meta is offering millions for video based on well-known IP aiming to attract users to its VR device launching next year. Now the big question is how long will these immersive videos be? Because Apple did do a bunch of these deals. They did license a bunch of Interactive video products and but they were all like five minute experiences.

Yeah.

And so you'd get through them all and then you'd wait a full quarter and Apple would be like, we have another one. Seven minutes.

Here's five new minutes of entertainment.

It's like that's not how people experience entertainment. I remember like, yeah, if you just.

Think of it, a lot of people are being entertained by their iPhone for four hours a day.

Yes.

Often through video.

And it was, but that's not even an iPhone thing. You go back. a few decades to, like, the original PlayStation had Final Fantasy VII on it. It came on multiple discs. And that game, people would play it for 100 hours. Metal Gear Solid was similar, like dozens of hours of gaming. And no one's really been able to deliver that in VR and have that moment. Same thing with GTA, hundreds of hours of entertainment. Anyway, very excited to dig into this new device known as Loma. It's more powerful than the Meta Quest VR headsets now available with higher fidelity video. hear it, they got the screens done, they pulled them forward off the bench top. The design is similar to the large pair of eyeglasses, more like Meadows Ray Ban AI glasses than goggles that the Quest and Vision Pro use connected to a puck that users can put in their pocket. So maybe they're going puck, which is interesting because that was very contrarian. Everyone was like, Steve Jobs would never let this happen. And Palmer came out and said, no, puck is great. Keep the you don't want heavy things on your face. That's just not a good experience. And so Meta is planning to charge less than $1,000, but more than $300. And so I would say $999 is probably the right price. I want it to be sort of expensive so it can be a great product. A Meta spokesman referred to the comments by Meta's chief technology officer, Andrew Bosworth, about the company working on many prototypes, not all of which go into production. Meta is working with Avatar director James Cameron's Lightstorm Entertainment on exclusive VR content. the two companies announced a partnership last year. So I think we gotta get on this. We gotta have a VR stream. Three hours of content every day. We're gonna be the reason churn is low on the next VR headset. 'Cause you just throw this thing on it, just like you're sitting here on the drink cam. You can click through.

I mean, yeah, just click all to the different angles.

It's pretty doable. It's pretty doable. I mean, you can film valuable, like usable spatial video on just an iPhone now. And then you can play that back in the, in the Vision Pro, and it does look 3D, which is cool. In other news, Wall Street Journal is reporting that Reddit is suing Anthropic, alleging unauthorized use of the site's data, an online discussion forum, who would have expected? Anthropic accessed the site more than 100,000 times saying after saying it had stopped, Reddit is suing Anthropic. And Anthropic debates this. I'm sure we won't be able to get into this today because I'm sure it's caught up in the courts and There's a whole bunch of legal restrictions, but we'll do our best to understand how these deals come about. It seems like most of the time, it's not that the company that has much data doesn't want the AI company to use their data. They just want to have an equitable agreement where everyone is getting the most value. And I think Reddit's surge, the stocks up, right? Yeah.

And then for more context, OpenAI is already paying Reddit approximately 70 million per year in a content licensing purchasing agreement. So they kind of got ahead of this issue and decided to strike up an actual deal. And I believe Google has a deal with them, too.

And I think this might be one of the Google has a deal with Reddit.

Yeah, I'm pretty sure.

Because there's that meme about the best way to search Google is search whatever your search term is and then space Reddit, because the user-generated content was better.

Than the SEO stuff that was popping up.

Google pays Reddit approximately 60 million so 60 and 70, they're getting 130 million. That's pretty serious revenue. And it's something that doesn't need to be brokered via a bunch of individual programmatic ads that might not work or anything like that, or sub-scale. It's just one or two deals, and boom, you're up in the hundreds of millions of dollars in revenue. What is Reddit's overall annual revenue?

$20 billion market cap. I don't know.

Not bad.

Let me see here.

How are they trending?

1.3 billion greenbacks in 2024.

1.3 billion. So they're getting like 20% of the revenue or 15% of the revenue.

Yep.

I wonder how big they grew.

They grew 60% over 2023.

Interesting. So they might be bigger, they might be bigger than Reddit or they might be bigger than Conde Nast, which at one point owned them. It's kind of unclear how valuable Conde Nast is because they're private. Anyway. So Reddit said that the AI company unlawfully used Reddit's data for commercial purposes without paying for it and without abiding by the company's user data policy. Anthropic is in fact intentionally trained on the personal data of Reddit users without ever requesting their consent, the complaint said.

Interesting saying that it's about the users.

Yeah. Bills itself as the white knight of the AI industry. Last year, Reddit took steps to try and limit unauthorized scraping of its website, creating a public content policy for its user data that is publicly accessible. such as posts on Subreddit and updating code on its back end. The user policy includes protections for users such as ensuring that deleted posts and comments aren't included in data licensing agreements. And so, yeah, I don't think that there's a really strong precedent for the agentic web. If I use Google Chrome to access a website, Chrome doesn't need to pay any sort of license. But if I go to Anthropic and say, Hey, get me up to speed on this topic, and it goes out and it browses the web. All of a sudden, it feels like maybe they do have to pay, whereas Chrome wouldn't, because it's just rendering the web page and it's not transforming it at all. What is transforming? What's fair use? And so these things will obviously play out in the court of law. And so hopefully they can resolve it.

Quickly and move on. Yeah, I'm actually surprised that they didn't already have a deal. Yeah. Because it's very valuable data.

Totally.

You want that data for your models.

And anyways, well, we have Sean McGuire joining in just a minute and.

The.

Other news in the Wall Street Journal today is Thrive Holdings is betting that AI can change it services. The company established by venture capital firm Thrive Capital joined with ZBS to invest $100 million into an entity that will integrate AI into IT firms. This is from Josh Kushner, of course. Shield Technology Partners has already acquired four IT service companies, ClearFUSE Networks, Iron Orbit, Delvall Technology Solutions, and OneNet Global. It said Thrive Holdings called Shield Technology Partners and AI-enabled managed IT service platform. IT service companies, also called managed service providers or MSPs, typically provide IT support and manage tools like software and cloud computing on behalf of businesses. Founded by Josh Kushner about 15 years ago, Thrive Capital is known for some of its high flying startup investments, including OpenAI, Databricks and Wizz. What a portfolio. Investing in traditional services business, particularly those that rely heavily on administrative knowledge work and adding AI to supercharge them is becoming a bit of a trend. As part of its efforts, Shield Technology partners will embed software engineers into each of its IT portfolio businesses. Oh, they're doing the for deployed engineer. the engineer's goal is to build an AI-driven solution that all of the portfolio companies will use. We've studied all the ways in which MSPs have perhaps been on their back foot to date with customers and says that IT services work is incredibly well suited to what AI can streamline. And so you can imagine a whole bunch of agentic workflows for all the different things that you need to do when you're deploying cloud and managing cloud. Really quickly, before we have our next guest, let's tell you about Vanta. Automate compliance, manage risk and prove trust continuously. Vanta's trust management platform takes the manual work out of your security and compliance process and replaces it with continuous automation.

Whether you're pursuing your first framework or managing a complex program.

If you think you should be on Vanta, you probably should. You're probably correct.

Well, we have Sean McGuire from Sequoia Capital in the studio.

Welcome to the show, Sean. How are you doing? Boom. What's up, team?

Never a boring day on the internet, that's for sure.

Yeah, what is keeping you?

Oh, man.

What's keeping you up now?

Well, obviously, I think there's You guys, anyone on Twitter knows what I'm talking about.

Yeah, yeah, yeah. On X. Yeah, I mean, let's skip the politics because this is purely a technology and business show.

Thank God, I love you guys.

You're the best. Stick to the technology. What, I mean, we've had an interesting experience with X in that there's always been this narrative that the whole platform was gonna collapse. There's been rough days here and there, but overall things have been growing. What have you seen across the XAI merger? What are the secrets to success? How is talent tracking? Is any of the chaos and noise distracting? Because when I talk to XAI engineers, they're like, We're too busy. We can't come on your show. But what's your experience been with the X and XAI team recently?

Look in, if you go back in time, as you said, everyone said it was going to fail, the app would crash, you know, nothing would happen, and that didn't play out. But there was a lot of tech debt and kind of broken infrastructure. And there was a, you know, a couple of years of rebuilding the basics and foundations. I think we're starting to see real innovation happening. I love the Grok integration directly in X. It always scares me when someone, when I have a tweet or whatever, and then someone says like, @Grok, is this correct?

Is.

This accurate? You never know what's gonna come back. Usually I agree with Grok, there's been once or twice where I think some of the subtleties are a little off, but it's truth seeking.

It doesn't mean that it's fully truthful every time.

Yeah, yeah, yeah. It hasn't actually found that ground truth. That's funny. What about the overall horse race? I think in reality, I'm probably wrong, but yeah, yeah. What about the overall horse race between the foundation models? It seems like every day it's going back between an open AI launch, an anthropic launch, a Grok launch, a Google launch. Are you do you think that continues? Do you think there's like maybe some fragmenting and there's opportunity? I mean, we're kind of already seeing this with how much Anthropics loved by developers versus OpenAI has been really dominant on the consumer side. And now every company is figuring out a different way to actually get to distribution. What really matters here? Is it pure scale? Is it pure cracked engineering talent? Is it distribution? Is it a combination of those things? How are you seeing it play out?

Great question. Honestly, my opinions have changed a lot over the last few years. in, in many directions. And so I don't have too much confidence in my assessment right now. But the, you know, I always try to look at lessons from the past. And my current thinking is that the closest analogy are operating systems. And if you, and I'll make a couple points on this. If you think about operating systems, first of all, there's a bunch of different ecosystem. There's the Windows ecosystem, there's the Apple, you know, OS ecosystem. Then there's on mobile, there's Android, there's the whole browser environment with Chrome, and then there's open source of Linux. One thing that I think is interesting about Linux, there's more Linux servers in the world than there are Microsoft servers, but the value kind of capture of Microsoft is way greater than Linux. I personally think we're going to see something very similar play out, whether it'll be like a, you know, OpenAI will be the Apple or someone, and, you know, XAI I think will be very successful. I think there's a good chance that Anthropic is independent and successful. I also think there will be a big open source component, which should be like Linux, and I think there will probably be 10 to 100 times as many open source models out there, or like deployments of open source models in 10 years. But I think that they won't be as valuable and they won't be as rich of ecosystems. And then just to make two more points on the open source analogy, for Microsoft, or Apple, by having the operating system, they were able to actually win in quite a few ways on the application layer as well. For Windows, they bundled in Word and Excel and then Outlook and all these other things. I think it'd be very similar for the foundation model companies. I think that the foundation models would be like table stakes. That'll be their kind of win, but also a very sticky moat. And even if they're not the most profitable businesses themselves, it will give them big advantages kind of on the application layer. And then one other thing that I think will happen. You know, the cloud companies have giant moats just through the CapEx dynamics of cloud, like needing to buy all this hardware and, you know, innovate with hardware and stay there is a big moat. I think these foundation model companies are going to be, I think there's going to be way more value that accrues and there'll be way bigger moats than people realize. I think they will all basically have hardware moats like cloud, style hardware modes, they will have the operating system style, very, very detailed research that's hard for anyone to replicate. And then I think they'll probably make a lot of their profit from applications on top of it. That's my current thinking. Thinking can change inside. Yeah, fingers crossed.

So I know obviously you weren't investing during the original operating system boom, but your firm Sequoia Capital was. Have you had any discussions with the lineage of the firm or the history and seen how is the revenue ramp or the business scale different this time than, say, in the dot-com era or in the previous era? It feels like it's ramping faster than ever. It feels like we're seeing more companies that are hitting a billion in revenue or 100 million faster than ever. Is that real based or anyone that you've talked to that was investing in that era? Does it feel different this time around, do you think?

I mean, one of the beautiful things about being at Sequoia is we do have this long history and we get to tap into the kind of institutional knowledge. That said, sadly, Don Valentine died four or five years ago, like early into my time. RIP, what an absolute legend. you know, and he led the original Apple investments.

Yeah.

But there's still a lot of Google institutional knowledge in the firm, which is, you know, not directly operating system, but they created an operating system later. I mean, first of all, the revenue of these companies is scaling insanely just faster than any products in history before for Starlink. So obviously not a foundation model company, but I basically made internally, I made an Excel spreadsheet of AWS's revenue growth in the first 20 years of AWS compared to Starlink. And Starlink has in five years gotten to where what took AWS 10 years to get to. And now with these foundation model companies, we're seeing as fast or even faster revenue growth. That said, I think the business models, the initial business model is more clear and the profitability of these companies is more in profitability is insanely high. And so you got to discount the revenue growth. But I would just say the biggest lesson I think from the past is you have to capture like territory early on, and the doors will kind of close behind you because of these CapEx dynamics and just like lock in with users.

Yeah, I mean, you mentioned Starlink. Do you think there's obviously, yeah, it's such a weird company because it's like a space launch company that now is an internet company, ISP. But there's actually a little bit, I'm starting to hear of an AI narrative just that having Starlink potentially unlocks edge compute or inference in areas that would typically have kind of stranded energy resources. So all of a sudden, if there's some super remote area that has really cheap energy, you can go and set up a data center there and then do inference and stream those tokens over Starlink. Do you think that's an underrated narrative? Do you think that that's developing on course? Do you think there's any bottlenecks that people should be thinking of within that story.

So when we first invested in SpaceX, the part of the core thesis was internet everywhere. And I would say like it goes way beyond AI. But I think the internet everywhere thesis is huge and that will be, you know, everything from oil rigs, you know, to airplanes to boats to Yeah, edge AI devices. But I think the bigger thing for Starlink is Starlink just has like a 10x plus cost advantage for moving data compared to, you know, building new transatlantic or trans-pacific fiber lines. And in the world of AI, we're going to be moving, these models are going to be moving so much data around themselves. and I just, I think Starlink is incredibly well positioned to be the pipes to move all this data for AI. Um, and so I, that I actually am, I care more about that just because of the volume than some of the kind of edge applications for AI specifically, but those will be big. And then one other thing, I just gotta give a plug to, to bitcoin, basically, yeah, let's go. Basically, three years ago, I visited the biggest Bitcoin mine in the world, Genesis Digital. Their mine is near Midland, Texas. It's actually backed by SBF, which is, you know, wild.

He got both ends of the bet.

Yeah, he got both anthropic and Genesis. But these guys had a gigawatt scale Bitcoin mine operating three years ago. and already for them, like having.

It.

Taught me a lot. And, you know, Bitcoin, Bitcoin mining is the absolute tip of the spear where you need the least amount of data movement, like data in and out to dollar generated or, or like power consumed. And so I actually think that was Like Bitcoin mining is underrated in terms of how much it's pushed like frontier power generation, turned into computes. And I don't think it's a coincidence that Crusoe, you know, which is now powering Stargate, started off as a Bitcoin mining company or that Core Weave, which is like $80 billion stock as of yesterday, is now, you know, is now an AI data center company. And I just, I think, I think that's honestly the bigger, the bigger theme.

Yeah.

What's your updated thinking around nuclear? We had these new executive orders and it was announced this week that Meta announced a partnership with Constellation to power some of their AI power needs. What's your kind of updated outlook over the near term to medium term?

I'm an all of the above guy for energy. Like we need all of it. we need all of it as quickly as possible. I, as an individual, invested in a few nuclear companies going back like nine, 10 years ago, way too early. And, and I like to put a little bit more meat on these statements. Nuclear is incredible, but deploying large amounts of nuclear is slow. Like, even if you deregulated it to zero, I think it would be like a, you know, a more than a decade, well beyond a decade to deploy like a terawatt of new nuclear, call it 10 years if you did it as fast as possible for America starting now. Solar is just a way faster way to deploy a lot of energy. Nat gas is a way faster way to deploy, you know, to deploy a lot of energy. We have been producing in stain amounts of natural gas, which we didn't have the pipelines to actually use. So we were just flaring it a lot of times because kind of like the, the dollar value per, like when you have an oil well or you're fracking, it's producing, it's emitting natural gas and, and oil. And you just made so much more money from the oil than the natural gas that we didn't really care about it. And that started to flip. And so anyways, I, I think we have to do all these things. I think we need more natural gas, more oil, way more solar, and then kind of have nuclear coming as the reinforcement Juggernaut coming online, like 10 to 15 years from now. That's a good framework. Fantastic.

I mean, we have to have you.

Back for, you know, an energy Deep dive. We, we know a fair amount of the nuclear and solar entrepreneurs, and there's a bunch People doing really cool stuff. So have a safe trip.

Personal plug. I had a seat in the New York Mercantile Exchange when I was like 22 years old. It was insane.

That's good.

Hey, good luck on the timeline today. I know you're going to go in there, put on your hazmat suit, and just get in there.

Good luck.

Peace, guys.

Safe travels. Fantastic. Let me tell you about Linear. Linear is a purpose-built tool for planning and building products. Meet the system for modern software development, streamline issues, Projects and product roadmaps. Go to linear.app. Next up, we have Jack in the studio. We have an in-person guest. Let's bring him in. Play some soundboard for me, Jordy. Welcome to the stream. How you doing, Jack? There he is. Second time on the show. Good to have you here. What are you wearing today, Jack?

I'm wearing the jacket, the TDPN jacket, in the capital of capital.

There you go. Thanks for coming. Thanks for hanging out.

Here, you can adjust your mic a little bit there.

Yeah, got it. Cool. I wanted to kick this off with a little bit of a rundown on the different foundation labs. We're talking to a lot of them today. And I noticed that Jordan Schneider from China Talk and Dylan Patel ran through their AI Mandate of Heaven tier list. And so I wanted to read through that and get your reaction and then just do a vibe check and talk to you about what we should be expecting from different labs over the next year. It's a little bit of a horse race. So up first at S tier, they have OpenAI. It's the only foundation lab that made S tier. Does that feel right to you? What are you watching from OpenAI?

Yeah, I think that's exactly right. OpenAI executing both on the product level, getting the distribution, getting into hundreds of millions of people's phones. But also on the research level, you have people like Gnome, people like Aiden, just doing this Incredible frontier research. 03, I think, just as a model, impresses me the most of any model that's come out so far. No Brad Lightcap said in the Wall Street Journal recently, they had 2 million workplace users in February, and they're at 3 million now.

Wow.

So it's just really exceptional growth.

I think-- I was thinking earlier, it'll be funny, our kids in 20 years will be like, Dad, they're making me use OpenAI teams at work.

Yeah.

It'll just be like the default, like the Microsoft Teams default.

Yeah, I mean, there's a little bit of narrative that maybe, and we can move on to Anthropic in the A tier alongside DeepSeek and Google. There's a little bit of a meme that Anthropic is crushing it with developers. They're the default choice for WindSurf cursor users, but then OpenAI is more dominant with consumers. But I feel like recently I've heard that it's maybe even more skewed than people think. It's maybe not like the vibe on X might be, yeah, like, you know, 7030 OpenAI clawed for day-to-day grab a random person on the street, but it might be even more skewed. Does that feel right to you? Yeah, I think I think Anthropic's really.

Solidified with developers, but it has like totally given up on consumers. But I think OpenAI wants to take that on. I mean, there's rumors about some sort of Windsurf acquisition. They're releasing 4.1 and Codex. They're pushing hard on coding. And I think that's something to watch from them this summer.

Yeah.

And going into 2026, is can they secure that?

Do you understand the model names at this point, 4.1? I have access to 4.5. Why would I want to go backwards? Are the models fragmenting to where I'm gonna have to learn a new taxonomy for, okay, if I want to write code, I use this one. If I want to write poetry, I use this one. If I want to do math or reasoning or build a chart, I use that one. Because it's putting more work on me, I feel like.

I think Sam said that they're gonna try to fix the model naming scheme this summer. So that's the real thing to watch. They're gonna keep S2 is can they, Can they get cohere and model names?

Yeah.

But yeah, 4.1, it's cheaper, it's specialized.

Towards coding, it's kind of their 3.7 type of driver. At the same time, I know you're not super up to speed on the Alibaba, like, Quinn models, but I saw some release where Alibaba Quinn released like a hundred different models. And Will Brown was kind of saying like, this is awesome from a research perspective because they have like, they have like one model that's just good at bio. And it's kind of like this hyper fragmentation. It's the opposite of going in the unification direction. It's actually, It's actually going more specialization, and then maybe you unify that at the end. But I don't know. It seems like if you're a consumer company, you don't really have that affordability, right?

Yeah, I think in terms of research, Alibaba's a bit underrated. I mean, compared to Deepsea, gets all this press, all this coverage. Totally. But the Quinn models are really good. People are doing, you see from lots of people, these really cool RL experiments, these really cool kinds of things.

They're lagging behind the US models. They're not A tier, they're not B tier.

But they're doing some interesting stuff, and.

I think that's super cool. Yeah, I really wonder if they have a distribution advantage in China. Obviously, we wouldn't feel it here, but I really haven't gotten up to speed on what is the ChatGPT of China in terms of distribution. Obviously, Deepak had that moment, but have they actually executed properly on the product side? I don't know. I'm surprised that Google hasn't been able.

To turn their general distribution advantage into an AI distribution advantage. They have these really good the new Gemini came out today. It's got really good benchmarks on a lot of things. But they're yet to, I think they're yet to crack distribution, we sometimes say.

Did you see that mock-up that was just the Google search box, but a Gemini prompt? It was like, if they wanted to go full send, if they were really AGI-pelled, they would just say, Hey, we're done with Google search. I mean, it would destroy their economics.

I'd commit to it. I'd commit to it. I think it looks like the model that they used to power those search prompts right now, It seems really lightweight to me. It gives a lot of wrong answers.

When I ask 2.5 something, it's always right. Sure, sure.

That's interesting. You think that's just a cost issue? The AI overview box is like, we're.

Just going to hallucinate. Yeah. Well, they are launching the advanced AI search, but it's like a toggle, so you have to find it, which is always the problem with Google. Well, I mean, they still wound up in the A tier, according to Dylan Patel and Jordan Schneider over at ChinaTalk. Obviously, VO3 was a huge one, and then they also have all those like priced performance things. But I've heard this narrative that like maybe some of the hyperscalers are super focused on benchmarking and not even hacking the benchmarks necessarily, but just like just thinking about them. And a lot of the frontier labs, the independent labs, have just kind of moved on philosophically from caring about benchmarks. Is that the right move? What's driving that? Are we in like the post benchmark era essentially?

Yeah, when I think about models and benchmarks a lot, I think which models outperform the benchmarks? When you see 03's benchmarks, they're good, they're kind of what you expect. Then when you watch 03 think, you see this model, it's actually reasoning. When you watch Opus or Cloud 4 Sonnet think, it's like, whoa, this is really good. Same with GPT 4.5. I think the Gemini models are good, but they're exactly as good as the benchmarks let on. And I think they don't have the vibes yet. What I want to see is Gemini 2.5 Ultra. If Google releases something with some big model smell, something really cool, maybe that's them catching up.

What is it with the big model smell? I just don't like the idea of smell at all. Oh yeah, it's just a weird, it's a weird sense. It might be Aiden McClain.

It's just a vibe check.

Who coined it? I think it was Aiden McClain. I don't know. That's great. But basically we're in the intangible period. Is that the idea?

I think unquantifiable. I think Anthropic's given up on really training on the benchmarks and I think it's gone really well for them. you see that they're really good at Sweet Bench, they're not crushing it on MMLU. But if you tried Force on it, it's great. Other labs that are lower down on this tier list seem to have not given up on doing really well in the benchmarks.

Yes, yes, that makes sense. I mean, it's possible that you must defeat the final boss to play the end game. And so maybe the end game is this vibe check, this big model smell. But in the interim, yes, if you're not, you only earn the right to go into big model smell if you.

Can dominate in all benchmarks. There was an interesting moment where 3.7 was beaten on every benchmark. So now for state of the art on stuff again, 3.7 was losing on everything. There was a better model for everything hypothetically. But then if you looked at what you might call revealed preferences bench, which is just like, what do people use on Keras?

Sure.

What's going on, Ram?

Reveal preferencesbench.com.

Yeah, yeah, yeah.

Someone's got a point. That was great.

Yeah, yeah, 3.7 was It was pretty high up there, so it seemed like they had something that wasn't captured there.

What about cornered resources? Data is the new oil. That seemed like a very silly concept in the moment when everyone had scraped the web entirely and it really felt like data was fully commoditized. Then we see VO3 and for the first time it feels like, okay, there is at least one data set that is so large that you can't copy it onto a single hard drive or compress it and it's YouTube and Google owns it. and yes, people might scrape it here and there, but Google has a durable advantage there. But is that the wrong way of thinking about it?

Yeah, I mean, I'm not sure about the video models. I think it's true that data is both super, super important, but also it's just become tremendously overrated because the first thing people learn about AI is, oh, it's a result of the data that goes in. But now that we're unlocking things like RL and better post-training, it seems to me like you can have some non-data solutions to some of these problems.

I mean, there was the original generative adversarial network for image generation, was like synthetic data generation and then testing it. And so, like, it just-- VO3 feels so much like a beneficiary of YouTube. But I don't know if that's just-- if we're just waiting and we'll see the next Sora and we'll be like, oh, OpenAI figured it out. And like, yeah, maybe they found some, like, you know, kind of workaround to the data. But really, the vast majority of the consistency and the innovation there was algorithmic progress, not just quarter resource and data.

Yeah, one thing about video models is it's been so secondary, but they've become so impressive. I think that if you showed them both to me a couple of years ago, I would be more impressed by VO3 than even like Cloud Force on it or something. It's just, it's not what I, it's really, really just incredible.

Well, yeah, I mean, I think a lot of it just comes down to the cost of instantiating the thing. And so if I go to if I go to deep research and I use 03 and I have it pulled together some, you know, 20 minute research paper, it's like that's a few hours of work. Maybe it's a few thousand dollars of like a researcher's time. Maybe we're getting up into like a PhD level. I could do it on your own. I could do it on my own. But but if I actually want to crash a Ferrari through the Hollywood sign with champagne bottles flying, a custom Hollywood sign that's huge, like Unless I'm either doing, I'm either renting all that, shooting it practically, and it's a multimillion dollar Michael Bay shoot, or I'm doing it all in CGI. And even to do it in CGI is millions of dollars of rendering. And so even for an eight second clip, it just looks like, wow, I got something that normally would cost a million dollars to make happen. And there's no real textual asset that feels like, wow, this is a million bucks worth of assets. Anyway, interesting. XAI, they are cooking. They've been obviously, GPU rich scaling up people seem like they're in the B tier here, according to this chart. But everyone's kind of excited about what's coming next. What is your take on Groq XAI? Are they close to the big model smell? That feels like a natural beneficiary of Elon's strategy of just go big. But how are you thinking about Groq generally?

Yeah, I'm not the most impressed yet. I mean, Groq 3 is good. It's a good model.

Sure.

It's like a funny thing, like Grock's whole thing or something that people who really like Grock often say, it's like, oh, it's chained on this real-time X data, this X-ability. One thing I've tried a few times, because I saw it in a tweet, is if you have a tweet you can describe, maybe I say like John Coogan's tweet about bringing media back to.

Hollywood, and you ask Grock to find it, it can't find it. You ask O3 to find it, it can find it. Wait, really? O3 can find it. That's so interesting because I feel like X is pretty locked down at the at just the www layer, right? It's pretty hard to find. In fact, a lot of times I'll post in a post from X and it will have to go to thread reader unroll and find an archive off of X because it clearly can't access it directly. But that is fascinating. That feels solvable.

Adam Adam ships tvpn guest.com. Last week I had a friend find it Monday. We hadn't announced it anywhere. It's not even visible on the Google Google search, really, and 03 found it. Wow. I was like, how did you find this? And he was just looking, he asked 03, can you pull together a list of all the things?

And he found the link randomly, and.

Google doesn't even find it. Interesting.

03 is really good at search.

And I think that might have been RL. They mentioned RLing on Tool Users in the blog.

Very, very interesting.

Also like XAI, it's like not really much revenue, nearly no revenue yet. You know, at some point you need to start pulling that out. I'm glad they're pushing on the distribution, you know, but yeah, things come around.

Yeah, makes a lot of sense. Last one, we'll end with.

We're the highest revenue multiple of any company in history.

Yeah, yeah. Last one, we'll end on Meta Llama sitting in D tier, but maybe not out of the game yet. The two interesting bull cases I've been discussing have been one, is there a world where open source American AI becomes geopolitically important for countries that are slight allies and they're either choosing between Deep Seek or an open source American model and open AI would not be in the conversation. And then also just, you know, why would you ever bet against Zuck? He has a capital cannon that can fire 10 billion at random projects forever. And so the question is, is that enough? What are you looking for from Meta and Llama in the future?

Yeah, it seems like you hit some issues recently, but I'm not betting against Zuck. He's got the capital, he's got some GPUs, They can get together some really great research. I would love to see better American open source models. I mean, I'm not betting on open source in the long term as maybe the cornerstone of AI. But the fact that all of our American research groups, lots of really smart researchers are doing experiments on Quan and not on Lambda, it's not great, you know?

Yeah, yeah. So should, there's one interesting twist there, which is Quan has so many different models. Llama has a few, they're still working on rolling out Behemoth, but would it be almost more of an olive branch to the developer community to fragment the models and really focus on hitting researchers? Is that kind of a potential path that they should take?

Yeah, I mean, I think it would be really cool if they did that. It would be somewhat charitable.

Yeah, yeah, exactly. Developers love a handout, but, you know.

I don't know. I think I'm curious about what they do on the product level and how they can build stuff in better on the product level, people aren't incredibly sensitive to whether 03 can search 50,000 websites like we are, you know? They care more about just having something that's really good, something that's really good to talk to, maybe Meta shifts focus to that anymore. I'm not feeling it right now in terms of like, when will a Meta model grab number one on Alamo Marina or something? It seems like it's gonna be some.

Time, you know, but I'm not counting them out at all either. Yeah, I mean, if they can just, yeah, stay on the lagging edge, that could still be valuable in a lot of their product rollouts. I mean, we forgot Apple in the L tier. We do have another guest hopping on in just a minute, but Apple in the L tier, how do they dig themselves out? Is it build? Is it buy? What do you think is gonna happen? They could maybe, they have a lot of cash. They could maybe buy someone.

They could buy someone.

Yeah.

Good buy, buy a lab, and then you gotta, yeah.

You gotta upgrade.

There was some report that they had some internal models.

Yep.

I wouldn't be surprised if they could train stuff. It's just, look, we haven't seen anything.

At all, you know? Do you think they're really training on Apple silicon? Like, you've seen those photos of, like, all the Mac Minis wired together.

Does that seem like something that's real or just like an AI-generated photo? I think non-GPU training ones are gonna.

Be bigger next year. Really? Well, I think the GPUs for Google. Sure. Yeah, so they already have a long time with TSMC.

They could go do something like a.

Trainium or an inference chip from Amazon or TPU.

Yeah, I mean, with the TPU, Google.

Has by far the most compute. Yeah, I mean, I guess Apple's pretty good at chip development design. So, like, they could do it. Yeah, they could do it. One chips are pretty good. That would be their, yeah, that would be their advantage if they could build a really strong chip and cut that cost. I wouldn't bet on it, but maybe.

Yeah, yeah, I like the idea of just opening it up and really partnering. The thing over the last 24 hours is one account sharing. It's so over for Google, and then immediately sharing. Wow, Google is going to destroy everyone in AI and just like seeing how the posts rank. Yeah.

Yeah. Anyway, anyone else on here? They got Mistral in F tier, poor for the French.

Yeah, they're not trying Le Chat.

Yeah, you gotta be. I do wonder about Mistral because, you know, the models are real, but none of like broken out in capability. But there's this question of like, if you want a national champion in your country, it might not be enough to just have the foundation model layer. You also might have to go and win in the free market in the application layer. And so, yeah, you could have, even if you had a comparable model, if you're not, if that's not, if people aren't, if people are going to do chat.com instead of leetchat.com, like you have not won and you don't have your national champion.

Yeah.

And there's a, I think there's some truth to this, but there's also the regulatory stuff in the EU. I mean, a lot I think via three is not in the EU. A lot of releases don't come there. Maybe Mistral just uses regulatory modes to monopolize, not a fun way to win, but maybe that's the bull case at this point.

Yeah. What was your reaction to the conversation back and forth with Duar Kesh and Shaltto all about this debate over, I forget, it was like spiky intelligence and how you actually train someone, there's so many different things. We see that the models are really good at one thing and then they fail RKGI. What's your overall timeline right now? How are you looking? Yeah, to cash raise the point that.

You can't kind of do this continuous learning, this like short run continuous learning. Like you can tell me, Jack, I want you to do something different as a general intern. I figure that out. And context is a weaker tool than that. And I think that's absolutely true and that's an unsolved problem. I don't know how much that moves my needles on timelines. One thing that could be true is just that OpenAI or Anthropic makes some like SWE agent and it starts accelerating their AI research and they just get really efficient algorithms really quickly, some architecture that just destroys the transformer. But I do think it's a meaningful unlock if that could be solved. And I think that sort of mid-level memory type of stuff is really interesting. Or solutions on context around evaporation?

Well, this is fantastic. We have our next guest. Thanks so much for hopping on. Good to come on. We love an in-person guest. For sure. Thank you so much. Next up, we're heading over to Google World. We have Arush from Google. He worked on the deep research project that dropped from Google in 2024. It was a full year ago. It was in December, technically. But very excited to talk to him. about that product, all the things that go into deep research. So we'll welcome him to the studio if he's available. How you doing? Good to have you.

Hey, what's up guys?

Thanks for having me. How's it going? Not too much. We're having a great day. We got a great lineup and excited to dig into it. Would you mind kicking us off with just an introduction on yourself and a little bit of the history? I want to hear about the history of the products that you've built at Google, what the interaction between research and product looks like and what you're excited about.

Yeah, for sure. First off, A-Team, that's pretty good.

Pretty good. Yeah. Let's hear it for you.

There you go. Let's go, let's go.

Let's hit it.

Yeah, John's gonna hit the dog.

Good work.

Cool.

Yeah, no, love to be here there. A lot of, yeah, it's been fun. It's been a fun ride. So I'm a product manager on the Gemini team.

Cool.

I've been here since a little while back when it was called Bard.

The Bard days.

And Bard days.

Yeah.

And so, yeah, about, I don't know, maybe like this time last year, we started kicking around this idea of deep research, where one of the things we noticed is a ton of people come to the product and ask, I like seeking to learn something or asking questions and kind of doing researchy type type things. But if you ask really hard questions, one thing we noticed is the model would just give you an outline of an answer. It wouldn't actually tell you something very comprehensive. So we just ran with a hypothesis of let's take off the constraints of it has to respond within a few seconds, it has to use this much compute. Let's just see how far we can push what the model can do. And this was before thinking models or anything and then kind of any of that good stuff. And so we worked on this idea for a bit and then we launched in December back on Gemini 1.5 Pro was the model that we were using back then. We launched deep research as kind of a bet to just see would people be into something that makes you wait 15 minutes but gives you something comprehensive.

I'm happy to wait, although I do want it to speed up. Questions about context window size. How important is that million token context window? That feels like it's been a unique Google feature for even longer than I expected. The advantages in AI seem to last days, maybe weeks, before another model comes out that meets or is roughly around the same capability. How important is large token context windows in deep research like products?

Yeah, it's huge. It's really what enabled us and gave us the confidence that this was even worth trying. I'd say that the long context enabled us to do basically be very recall forward and really cost a very wide net as we research the web and try and find gems of information that we then stitch together. And so that That was, I think, our biggest differentiator and really allowed us to build this product. The other thing that Long Context allows us is once you've finished your research, not just the report, but everything it read along the way is in context. So you can keep asking questions, going deeper within Gemini. And even if it's a tidbit of a fact that's not in your report, if it's been read at some point, it'll be able to retrieve that and give you that answer. So it also helped sort of beyond that first turn, keeping a good experience. And then reasoning models was like the next big step jump for us, allowing it to then do more critical analysis.

So in terms of like actual product design, I'm interested in the direction this goes here. You could see one world where the models are baked down into silicon, everything's running even faster, you're distilling the models, and all of a sudden I'm getting a 20-minute product in two minutes or even 20 seconds. You could also imagine a world world where what's possible if the economics work such that I could request a two-hour research report or a two-day research report. How are you evaluating those? What would you personally be more excited about? And what do you think users actually want? Because stated preferences and revealed preferences are often different, or do we wind up with both?

Yeah.

So one of the things that we noticed One, when we launched this, we had no idea people would be willing to wait. Every metric at Google from the day it started is reduced latency and all metrics go up. So this was definitely a bet where a lot of people thought we were crazy, where we were like, we're just going to take a ton of time and people will wait. One thing we noticed is that after about a minute or something like that, people are fine. People will go away, do other things, come back, we'll send them a notification when it's ready. So the big pleasant surprise for us is like people don't mind waiting. I'd say so in terms of like efficiencies gains, one of the things that we're more excited about is like, okay, if we can make models more efficient instead of reducing down the research time, can I give you just a way better output? Like can I bank that savings and give you something way more insightful, way higher quality? I'd say the other thing is like, Even if I could give you a deep research answer in 15 seconds, it's going to take you 15 minutes to read. So there's also an aspect of just how much do you want to consume this? So for us, we're not as stressed about can we make this faster? Can we make this quicker? I do think there are probably other points in the latency comprehensiveness spectrum that people might like. We picked one extreme of let's just go super hard. and build the most comprehensive long thing that takes a while. But there might be totally other points people are interested in.

Yeah, yeah, yeah.

Sometimes I notice I've generated so many.

Various deep research reports across all the different apps that I'll follow it up with a prompt like, okay, boil that down for 10 bullet points because I don't have time to read that. And then I'm like, wait, maybe I should have just asked it to give me 10 bullet points and I just burned a bunch of GPU cycles. But I guess the question- I'm splitting.

Back and forth between the two until you kind of understand the subject.

Exactly. But I guess the question is like, is there a product or is the natural evolution of just general prompts that as algorithms get faster, as these models run faster, that there is a deep research amount of work that happens within a few seconds between every response? And basically the question is like, how much can you port from the deep research product and strategy and design back into just your average LLM interaction.

Yeah, I think there's definitely a lot of learnings that we can kind of start up streaming really around like being able to form a plan, follow that plan to do that sort of multi hop steps of like search iterating, like finding insights changing your strategy possible before going back to the user. And so you're kind of starting to see this in like 2.5 Pro and stuff like that. And you can imagine that that will continue where you will see more like mini deep research or more sort of like planning and sort of like iterative reasoning before giving you an answer. And yeah, that could just start getting faster and faster and faster, and then you start just getting way more insightful or comprehensive answers.

Are there any other interesting areas? I mean, deep research feels like one of the first really solid product market fit experiences in, I guess, like agents broadly. Are there any other areas that you're excited to think about knocking down with either different products or just maybe just cool uses that you've developed or as a user? patterns that you're leveraging that maybe go beyond just the average, like, I need a research report.

Yeah, totally. So I think there's like a few different angles that I think a lot of people are exploring. One is you kind of pointed out, like, what does a two-hour deep research look like? What does an overnight deep research look like? If you can have a very well-defined problem where we have early experiments at Google, like AI co-scientists and stuff, right? You could run that overnight. and it can come up with novel scientific hypotheses. So there definitely is an angle of if you can define a problem and an outcome really well, applying more compute can actually get you better and better answers. So there's definitely an angle of other whole new classes of problems where we can go even further with deep research. There's a second aspect of we had the chance to go meet a bunch of people who are researchers at the Fed, and they were telling us how.

They use deep research.

And it's often a very different thing, right? So I showed them this example where I was like, Hey, there's this funny law in the US called the Jones Act where any two ships between two US ports have to be built in America, crewed by Americans, and drives up shipping prices, but only for Puerto Rico, Hawaii, and Alaska, right? And so I was like, Do an economic analysis of the Jones Act on the economy of Hawaii.

Right.

And it like did a first principle analysis, did some really interesting things like looking at, well, like how much is a three and a half thousand shipping route, like say from like Mexico to South America. And then that's like a baseline price to compare against. And like I thought this was amazing, but then they were like, that's not how we would do economic analysis. Like they would be like, first I'd explore like what other studies there are. Like then I'd explore like what kinds of methodologies are out there. then I might ask a bunch of fast follow-up questions about what data sources or data sets did people use to do this research? So there's definitely an aspect of another angle of if I really want to help people with research, it's about nailing this synchronous asynchronous paradigm and helping people kind of do more of that iterative process rather than just ask question, get answer and move on and victory. And I think that's kind of a product challenge, like figuring out the right interaction model for that. And the third is just like outputting an answer at the right level of abstraction that you work at. Like a financial analyst doesn't think in.

Terms of report, right?

They think in terms of the spreadsheet or the financial model, right? And so if I want a DCF, deep research can build a great DCF discounted cash flow model for me, but I don't want it in a report. I want it in a spreadsheet, or I want it in an app where I can play with the variables and see the different outcomes. And so you'll also see the line between reports and other kinds of artifacts starting to blur, or even just like, what does it mean to build an answer? And that could take a much wider space.

That's super exciting. Yeah, I mean, I've seen obviously Gemini, we probably can't talk about the roadmap too much, but I've seen Gemini pop up in a bunch of different areas, and I haven't seen the deep research version of whatever that instantiation?

Yeah, maybe my last question is how much time are you thinking about working and making, as a product manager on Gemini, how much time are you thinking about making Gemini better versus fighting for distribution outside of Gemini and across the Google ecosystem? Because part of unlocking the value of Gemini is just making sure it's in the right places and placed contextually across everything from.

You've worked hard on this. Just ask for the I'm feeling lucky button. Just give us that.

Just- We think you've earned it.

You've earned it. It's a great product. Just click, I'm feeling lucky, or burn 40 GPU hours on this new research award.

Yeah, that would instantly melt all of our servers everywhere.

This is the biggest hyperscaler. Yeah, we need more GPUs. Let's get GPUs. TBUs.

TBUs.

Okay, so ASML, get cooking.

I believe you've earned the I'm feeling lucky button in years. Yet I use Gemini all the time. Yeah, yeah.

Let me get the use.

This is what the users want.

Yeah, we just need 10 more TSMCs, I guess, to start fabbing. Anyway, sorry, you're serious.

Yeah, no, the serious answer is like the Gemini app is like a great place for us to like prototype see what hits really works with people. A lot of the users, they're very intentional when they're coming to the Gemini app.

They want to use an AI experience.

So it's a really great place for us to put stuff out there, see what works, see what doesn't. Some things we put out needs more.

Time in the oven.

And then over time, you'd imagine that then those insights or things that really start work, you'll start seeing in other Google products as they make sense. You don't want to over clutter a UI, but you'll start seeing Yeah, things like deep research.

Yeah, because it's a very different user, somebody that's coming in saying, I want AI versus I just want to do certain things.

Yeah, right.

And yeah, they're totally different archetypes. It's a fascinating challenge.

I'm sure it's even more challenging at your scale, but thanks for all the hard work and pushing the frontier forward. It's been a pleasure talking to you. Yeah, come back on again soon. Yeah, we'd love to talk to you more. Appreciate it. Thanks, guys. We'll talk to you soon. Bye. Fantastic. Next up, we have Oliver Cameron. I have a good story. We'll bring him into the studio. But I believe he was the first person I ever interviewed for a YouTube video years ago. I was doing a whole video essay about Cruise, the self-driving car company, and he hopped on a Zoom call with me just like this one, and I recorded it and threw clips in the video. It was very fun. And then I wound up doing more interviews after that. So Oliver, good to see you. How are you doing?

What's going on?

I'm doing great. Thank you for the opportunity.

Would you mind kicking us off with like the latest and greatest introduction because you've done a lot in your career, but you're onto something new? For sure.

So spent about eight years building self-driving cars. Incredible time. I mean, just to see that technology go from barely being able to keep in a straight line to navigating downtown San Francisco with no human behind the wheel. just a sign of where things have gone with machine learning. So I had a blast doing that. Built my own company, sold that company to Cruise where we met and loved that time. Left Cruise in May of 2023, decided to start something new. And both me and my co-founder, who also was from self-driving cars, we were both very much inspired by Pixar. I think it's just a very special company, right? Everyone kind of recognizes Pixar as this sort of iconic storytelling company. And we really put our heads together to think about what a modern reincarnation of Pixar would look like. So that company is called Odyssey, and we're an AI lab that's really focused on enabling entirely new stories to be told.

And walk us through the first product that you launched. I played with it earlier. It was mind blowing. We'll pull it up while you're talking.

Sure, yeah, we just released a research preview of something that we call interactive video. And it's effectively AI video that you can both watch and interact with in real time. And we think this will become a entirely new form of entertainment. You've got film, you've got games, you've got all these mediums that have been around for a while. We think that there is an opportunity to invent a brand new one where effectively a model is responsible for imagining film and game-like experiences in real time that you can interact with. There's no game engine behind all of this, no heuristics, no rules, just a model that's learned pixels and actions from tons and tons and tons of real-life video.

Yeah, we're showing it on the screen right now, and the production team is controlling it with the keyboard, WASD, like it's a first-person video game, and they're walking around this field. with trees and windmills, and they can actually choose to go up, go inside buildings, and it's all being generated without the use of a game engine. And then they can switch over to a different environment. And so, I mean, I have tons of questions about how these different-- like, you're not doing photo scanning, you're not doing game engine stuff, traditional 3D pipeline, but the data must come from somewhere. Love to hear about that. And then also, I noticed the space button doesn't work. I wanted to jump around, start bunny hopping. When are we getting a space button added to this thing? Anyways, sorry.

Isn't it trippy how those pixels are literally streaming from a GPU cluster probably in Texas?

It's so crazy.

And now we're streaming them via Zoom in real time. It's crazy.

My question is, do you think that Odyssey can be a really breakout app for VR, because when I see that video.

Oh, sure.

I feel like that it could give someone the sense of being able to explore lands that don't exist, which is like very fast, like once it's fully immersive, it feels like-- It's funny, the.

Windmill thing, because I remember the very first Oculus demo that I ever did. I was walking around a windmill, and it's still in my mind years later. And it was amazing, but it was just like one little windmill, and then you couldn't go any further because developing virtual assets is really expensive. And so you play a lot of these VR games, and it's a couple hours or 30 minutes, but if you take a procedural approach or a generative approach, you all of a sudden have infinite content.

I think what's really important to note is in film and game, incredible things can be made, right? Insanely good things that wow us all. the time and the money it takes to create those things is ludicrous. And it's only getting more expensive, not less expensive over time. So I feel there will be continuously a place for these sorts of handcrafted.

Things.

And they'll be very important. But if we just think about a model that's trained on literally decades of video, that's then able to imagine stuff in real time with no pre-production costs, no post-production costs. and do that literally in real time, like 33 milliseconds. That's where it gets really crazy. And what we showed in the research preview is just this tiny glimpse, I think, of what this stuff will become. VR in particular is the most hardcore application of this from a technical perspective because the resolution required for VR is insane. And the resolution that you saw there, you can tell it's low res, it's like 300 pixels wide. So there's going to be a leap that needs to happen there to get to VR level res.

But I'm confident that Odyssey 2, you'll have it dialed.

Oh, yeah.

Give us the stats. How many, what numbers can you give us about the progress or adoption? You just launched this, I think, this week or last week. It hasn't been very long, but how has the response been quantitatively?

Oh, it's been incredible. So we launched a week ago, and since then we've served 250,000 unique streams, meaning 250,000 people experiencing what you just saw, which is insane.

Market clearing order inbound.

Yeah, let's do it. Love it.

Congratulations, that's fantastic. On the question of resolution, there's a bunch of amazing AI uprezzing that's happening in various parts of the pipeline. There's some server-based uprezzing that can happen. There's some on-device uprezzing. So are you counting on that technology breaking one way or another? Does it matter? Will it be a combination of both? How do you see that developing?

I think a way to think of this is where video models were a year ago is where real-time video models or world models will be today.

Yeah.

And what that really means is that you look at the res, remember the Will, everyone remembers the Will Smith spaghetti video.

That was, was that like one year ago or two years ago?

It wasn't long ago.

It was just over a year ago.

It's so fast. There was definitely better outputs, like spaghetti. It was like the weirdest thing at the time. Although gymnastics today, I'm sure you've seen that. That's really tough with video models today. That is tough. That's all to say that I think the res and the visual quality improvements will come from the model itself, not like some secondary piece of infrastructure to upres. Just because, I mean, think of what a language model was like to use two years ago. How fast was it in response time? Really quite slow, right? Compared to today, where it's like just stream of information straight to your eyeballs. Same will be true of these models. We'll crank out larger resolutions, faster frame rates, More actions, more things you can do, all that sort of stuff.

Yeah, and I guess importantly, GPT-4 upresed to 4.5. It is a different model. We're walking around what looks like the gloomy English countryside right now, and I think the production team is going to try and go in that house. It is really, really so wild. I noticed that there's a Time limit.

Tropical island demo because this, this, this, I love the English country.

It's very foggy. Yeah, I noticed that there's like a two minute timer when I sign in. Is that so the GPUs don't melt? I mean, I assume you've raised money and you're maybe burning some money with these demos. But break down kind of like what your limitations are and how you see them evolving.

Yeah, for sure. So, The timer is there because each session is served by a single GPU. So each user gets a GPU, the model's running there, and that's being to the user directly.

And really quickly, when you say single GPU, you don't mean RAC, you mean like one A100 or something like that?

One H200 per user.

Got it.

And there is a clear path to dividing the GPU to have multiple sessions per user, but today it's one. and we want to really crank up quality, frame rate, all that sort of stuff.

It makes me feel great to know that I'm getting the sort of one-on-one attention from H200.

From a chip. Yeah. It's an honor.

If you're at a retail store, it's not a great experience if somebody's bouncing around.

Exactly. I'm being individually served.

I'm being served.

This is like an Hermes level.

By Jensen.

Exactly.

So $2 an hour is there or thereabouts how much that costs, which over the course of a multiple users, it's not too bad. I think Netflix is like $0.05, $0.10 an hour, something like that to stream video. So we're a bit of a ways away, but you've got new chips coming, just model optimizations. It won't be long where we're having a single GPU per user, all that sort of stuff. For this launch, we had something like 360 H200s prepared. We had to scale it up a little bit just because we had lots of demand, but that timer is there just to make sure we're cycling through lots of people getting a taste of this. But yeah, I think fundamentally the idea that you could have a model stream stuff to any screen is really powerful. That experience you saw there works just as well on an iPhone, on an Android, on a TV, anything like that. And it's all just action conditioned of a WebRTC, which is probably what Zoom is running on. So the action is just sent over the wire to the model. The model then conditions the pixels it's about to generate based on those actions, sends the pixels back and just that loop every 33 milliseconds.

Yeah, it's firing. So, I mean, the path to HD or 4K seems pretty clear to me. What about the path to consistency? That feels really difficult. You need essentially like a really long context window to know that, okay, I dropped my mythical sword on that piece of the ground. I went away and then I came back. that's like textbook, just put it in a database. But it seems like the future might not be that. So how are we thinking about that? I guess the bigger question is what's the response from the gaming community? Is this something that can be a tool and a piece of a pipeline instead of completely replacing the entire traditional pipeline?

So most research on interactive video before has learned from games. So lots of folks will have seen Oasis, from Descartes, a Minecraft in a video model effectively, or Quake, that's often used in video models. And I think the gaming reaction to that is quite negative.

I mean, you saw the Carmack back and forth, right? Where Carmack was like, this is amazing. I love this. And somebody else is like, this is stealing developers. Yeah.

And I think it's important because the way that people envisioned that is like, oh, what's the best that this could become? It could become like remixing of games.

Yeah.

And that's one way it could be. I think people see what we have and they think, oh, this is like a world simulator eventually. This is the Matrix or, like, whatever they project on it.

Yeah.

So really, I. One thing we're trying to avoid is, like, for the first few generations of this, people will put, including ourselves, like, this picture of what existing games look like onto this.

Sure.

And.

It's like the iPhone when it launched, right? People ported desktop apps to the iPhone and it kind of worked, but it kind of didn't. It wasn't really embracing this new medium. So I think the long story short here is stuff that is integral to games today, like multiplayer, state, scripting, all that sort of stuff. Let's question those assumptions. How should those things work? Let's make it model native. Maybe memory in this model is very different than memory in a game or state in a game, multiplayer in a game, all that sort of stuff. And that's probably going to lead us in the short term to more like glitchy, weird experiences throughout the memory as it is by the models, a feature, not a bug. I don't know if you guys have seen like the back rooms or like.

These kind of glitchy, weird. Yeah, yeah, yeah. It almost is a completely different type of game design. Exactly. Yeah. The up, down, left, right, A, B, A, B of the future will be like, drop your bad sword on the ground, walk around the building three times, come back, and it's in. and it's enchanted because the model hallucinates that you've upgraded or something like that. That'll be fun.

I also think that one important thing here is that in language models, one of the things that's happened over the last year is in many cases they've crossed this threshold of realism for certain applications. So people literally fall in love with language models, right? The same emotional feeling they have when they meet a person they fall in love with is happening for them with a language model. and that's because what they're seeing on their screen is so realistic. It's crazy real to them. And I think the same will be true here, where once these pixels, these actions feel so realistic, which eventually they should, just given the data, given the models and advancement, there'll be things that they do in these worlds or things they feel in these worlds, which they just can't feel in video games today, because games are just capped by computer graphics and human dev time and budgets and everything else. But they'll walk down the street, they'll see someone and they'll be like, wow, that person looks so real. And they'll go over, they'll high five that person. like on a screen, right?

Yeah.

And they'll just feel something. Like they'll feel like a heartbeat raise.

You know, stuff like that.

Totally. So that's an application that you can't.

Do in games today.

That's just different and new. So that's the sort of stuff we're really interested in.

Well, that's gonna be a wild, wild future. But thank you. We'll have to have you back and check in on progress. Yes, fascinating. Definitely the day that 720p drops or whatever the next version is. We're excited. good for this. But thanks so much for joining. This was a fantastic conversation.

We will talk to you soon.

Cheers.

Have a great rest of the week, guys, for joining. Thanks so much. Next up, we have a return guest.

Michael Mignano from Lightspeed coming into the studio. Are you going to hit the gong or?

He's going to talk about competition between the Foundation Labs and the App Layer.

Well, welcome to the stream, Michael.

How are you doing?

Boom. Good.

Good to see you guys.

Congrats on the new studio.

Thank you. Thank you.

It's been a lot of fun.

A bit of a, I like the upgraded gong too.

Oh yeah, the gong is much bigger. Everything's bigger. We got a bigger one in the works.

Yeah, we're kind of even bigger gong. Florida's got a really big gong.

Oh yeah.

Also, what's a funny day to just be so hyper fixated on AI? Because you probably haven't seen the timeline. Tesla's down 17%.

17%.

It's just absolute mayhem. I mean, there's an AI narrative there, right? Yeah, there's definitely that's not what's driving it there. But anyways, Michael, it's great to have you on. Wanted to get some kind of updated thinking from you on the tension between labs and the application layer. We saw the news with WindSurf and Anthropic today that had more to do with a potential acquisition.

And even when we talked to the founder of Granola, we were talking about the competition between Notion and Granola. with these, it's a founder-led kind of previous era scale up unicorn SaaS company. Can that company bolt on AI? But then now we're seeing competition from the Foundation Labs. So we'd love to get your lay of the land. What are you seeing? How are things shaking out? And what do you think the next few months or even years look like?

Yeah, it's pretty interesting, right? Like if you think about the big companies that startups previously built on the backs of, the Googles, the Amazons, the Microsofts. It felt like there was this really healthy symbiotic developer ecosystem where the incumbents supply resources, the developers buy and extract from them, and they build really, really big businesses on top. I think what we're seeing now, to your point, is these labs are building developer ecosystems, but then they're very intentionally and overtly going head to head with the developers that are building on them. And I think this has a lot to do with context, right? So if you think back to the internet, you know, and startups 10 years ago, everyone said, Content is king, you know, content is king. Then distribution was king, right? It was all about how do you get in front of users? It's we're starting to feel like we're entering the phase of context being king. These models are just hungry. for the most and the most unique context possible. And so if an app layer company emerges and has a new type of context and data that the models don't have great exposure to, it's a great signal to point in the direction and say, we're going to compete head on. And so I think that's what we're seeing now. And yeah, Nabeel Hyatt, great investor from Spark and I, we often talk about how the war for context is happening now. And I think that's that's what a lot of these moves represent.

How do you think app players should app player companies should respond? Is it just double, triple down, go way, way, way deeper, focus on workflows that the labs maybe don't have the resources to fully pursue, or is it focusing down on specific niches? I'm curious what the yeah, I think the right approach is.

Well, we can definitely get into that. But maybe first, what I would say is I tweeted something yesterday that occurred to me after the big announcements from OpenAI in that the big incumbents, which we talked about a little while ago, sort of like the winners of the cloud era, it wouldn't surprise me if.

All of.

These new competitions between the labs and the apps actually drive the apps and the startups right back to the incumbents, to the Googles and the Amazons of the world. I have to wonder if some of these things actually act as a tailwind for models like Gemini and maybe give a little more credence to the, the, the argument that, like, Google is actually going to be the winner here because of all their distribution.

So I think that's one. You mean driving back and being like, I'd rather work with Gemini because I don't think they're as likely to kill me.

Exactly. Yeah, exactly. It's like, hey, we trusted them with the cloud, and that worked out. All right.

Like, should we now trust them with AI more than we trust the labs? Yeah. I mean, that narrative even goes a little bit. further with Microsoft, which has been completely like, oh, we will host every single model. We'll let you reroute really intelligently between them, like super, super friendly developer ecosystem. And so, I mean, certainly they're building stuff into Copilot, into Microsoft 365, but it does feel like they're much more willing to work with us.

Satya seems to have real conviction. He had the quote from last week, platform, platform, platform and hosting Deep Seek is an example of that.

Right. Yeah.

A lot of people would have thought, oh, he's not necessarily going to host that model because it felt like a shot across the bow at OpenAI. Sure. But he's committed to supporting open source.

Yeah, right. Yeah, he wants it all. Interesting.

I think, you know, also going back to your question, Jordi, I think all of this is just going to make for a more intense, faster moving big market. I think more than ever before, you have to ship, you have to get users faster than anyone. You have to sort of reach escape velocity quicker, which I think is just going to put more and more pressure on startups to move even quicker than they already are. I feel like Cursor is a great example. I feel like an earlier iteration of that product would have been easy to write them off and be like, oh, a lab is going to do I mean, now it's like they're so big, they're so far ahead, it feels like they've really established themselves and likely have a good shot of breaking through.

I also wonder with Cursor and Windsurf and Devin and some of the dev tools markets, it feels like just such a new market that even if it's somewhat winner-take-all, it's so positive sum because it's adding efficiency to the most like one of the biggest labor pools. And so when we talked to the cognition folks as the reaction to Google and OpenAI launching Devin competitors, they're like, well, we still grew 40% last month or something like that. And so, you know, I wonder, like, in code gen, where it's such a new market that it's not directly competitive with anything that exists, so it's less zero sum. I'm wondering if the note-taking market feels similarly to you? Or were you seeing Granola or other companies act as more dropping replacements for existing tools?

Yeah, I think it's a great question. Yeah, so we backed Granola really early on because we knew Chris and his co-founder and we loved those guys. We didn't know what they were building. We knew they were going to build something in note-taking. But we said, you know what? This market's going to move fast. We trust these guys. Let's go for it. And I think, somewhat to your point, there's been all these no takers before. Like, Granola wasn't the first no taker. There was Fireflies and Otter and all these things. But I think Granola has done a really, really good job of getting out of the user's way and establishing trust with the user. And I think that seems like a small thing, but I think that trust thing is going to be really important if you go back to what I said about this context being king, like who are you gonna trust to take this context or take this like really, really important, you know, proprietary part of your work, in this case, your meeting notes. You know, a lot of people say we trust granola. Are they just going to hand it over to any old company that says, Hey, now we want to screenshot your entire computer and take and suck every last piece of data out of you. And so I think part of it is to your point, like getting in early, getting big really, really fast and establishing that user base and that market before it really matures, but also in a way that users just really trust you and they're not just going to rip you out just because some other bigger company offers the same thing.

Yeah, that's a good point. Do you have any more micro-reactions to specific integrations? That seemed to be one of the big things that OpenAI was pushing on was integrations with Google Docs and Drive and and your email. And that feels like adding that extra context is potentially the next thing people are clamoring for. How important is the biz dev side of this business, in fact?

I think it's really important. I think it's really, really great that Anthropic started the whole MCP protocol. Obviously, lots of others are adopting that now. But I think to your point, we're now going to start to see the battle lines being drawn. Like, who are we willing to integrate with? Who are we not willing to integrate with? Where is it? Are we open or are we closed? Where's the data gonna go? Where's it not gonna go? I think we're gonna start to see those alliances and those allegiances form.

And we've seen all these APIs back in, what, like 2007, 2010 era. Social media. They have an API. It's amazing. It's like, well, you don't know how much that API is gonna cost. If it's $10,000 per day or something, that could completely upend your business. And so actually thinking about how that dynamic develops is almost more important than the standard, although I'm very glad we have a standard. That seems great. But each company is going to have to decide where the value accrual really lands. And then who knows, maybe it'll be some anti-trust in 20 years, like we're seeing with Apple.

Yeah, the big question around trust that, you know, it's an evolving situation, but a California judge, I believe it was yesterday or the day before, ordered OpenAI to retain records of sort of, I forget what OpenAI calls it, but if you have like a disappearing query, a judge ordered them that they have to retain that. They obviously said that's a huge outrage. We don't want privacy with users.

So incognito mode, yeah, it's like not incognito.

Yeah, and that more seems like an issue with the court, the government judge, you know, having this massive overreach around privacy, but privacy in this era when people are more willing than ever across every app to give them all sorts of data.

Yeah, and you have a direct incentive to reduce the level of privacy to get better results. Like, if the model knows what kind of car you drive, when you ask it, for new tires, it will give you better recommendations. So for sure, you want to lean into being anti-privacy to get better results. There's this, the world is definitely bifurcating into pro-privacy or like fully AGI pilled folks. And there aren't that many people that are in the middle. So obviously we will have to figure it out as a democratic society, ultimately vote, and hopefully sort it all out in the courts. But thank you so much for stopping by. This was Michael, always a pleasure. We'd love to have you back.

Talk to you soon. Yeah, I mean, guys, I just want to tell you, you know, I don't really aspire to ring the New York Stock Exchange bell one day.

I aspire to hit that gong.

Hit that gong. Well, next time you're in Los Angeles, come by. Come by. All right, have a great story. We'll talk to you soon. Bye. So we have a generational crash out going down on the timeline. We got a new post from Elon. I'm gonna read it out. He says, and this is your live reaction to drop the really big bomb. Real Donald Trump is in the Epstein files. That is the real reason they have not been made public. Have a nice day, DJT.

Wow.

That is a big that is a big bomb.

But wait, didn't we already know this? Because isn't there that picture with Trump and Epstein together? We're really in.

The business.

I want to go back to AI and business and technology.

Here's the Tesla's down 17%. DJT is down 7%, Trump coin is down 10%.

Wow, they're all fighting.

This crash out on both sides is not good for anyone.

Well, you know what's some interesting... You know what's not down? Tokens generated, baby. We're still generating tokens every single day. The relentless march of artificial intelligence continues.

So the other thing is Elon shared, or sorry, Trump shared on truth. It's funny they're battling on their different social... Every billionaire should have their their own, you know, social media network to.

Get the word out.

But Trump said, the easiest way to save money in our budget, billions and billions of dollars, is to terminate Elon's government subsidies and contracts. I was always surprised that Biden didn't do it. So Ashley St. Clair is saying, Hey, Donald Trump, let me know if you need any breakup advice. They're really good. And Dan Primack says, this cannot be a comfortable day for David Sachs. On the other hand, it's just the best day for Sam Altman.

Well, well, we have someone from OpenAI here. We're going to stick to technology and business, but welcome to the show, Mark Chen. Good to see you.

Great to see you guys.

Thanks for having us. Awkward day, but I'm excited to talk about deep research. I am excited to talk about AI products. Would you mind introducing yourself and kind of explaining what you do? Because OpenAI is such a large company now, and there's so many different organizations I'd love to know. how you interact with the product and the research side and anything else you can give to contextualize this conversation. Yeah, absolutely.

So first off, you know, thanks for having me on. You know, I'm Mark. I am the chief research officer at OpenAI. So in practice, what that means is I work with our chief scientist, Jakob, and, you know, we set the vision for the research org. We set the pace. We hold the research org accountable for execution. And ultimately, we really just want to.

Deliver these capabilities to everyone. That's amazing. In terms of research, I feel like a lot of what happens in the research side is actually gated by compute. Is that a different team? Because what if the researchers ask for a $500 billion data center? That feels like maybe a bigger task.

So, yeah, it is useful for us to factor the problem of research and also kind of building up the capacity to do that research. So we have a different team. Greg leads that. which really thinks holistically about data center bring up and how to get the most compute for us. And of course, when it comes to allocating that compute for research, Jakob and myself do that.

That's great. And so what can you share that's top of mind right now on the research side? There's been this discussion of pre-training scaling wall potentially, the importance of reinforcement learning, reasoning. There's so many different areas to go into. What's actually driving the most conversations internally right now?

Yeah, absolutely. So I think really it's a really exciting time to do research. I would say versus two or three years ago, I think people were trying to build this very big scaling machine. And really the reasoning paradigm changed a lot of that, right? You know, like reasoning is really taking off. and it really opens this new playing ground. It's like there are a lot of kind of known unknowns and also unknown unknowns that we're all trying to figure out. It kind of feels like GPT-2 era, right? Where there's so many different hyper parameters you're trying to figure out. And then I think also, like you mentioned, pre-training, that's not to be forgotten either. Today we're in a very different regime of pre-training than we used to be. Today, we can't treat data as this infinite resource. And I think a lot of academic studies, they've always kind of treated, you have some kind of finite compute, but infinite data. I don't think there's much study of finite data and infinite compute. And I think that also leads to a very rich playground for research.

Do we need kind of a revision to the bitter lesson? Is that a refutation of the bitter lesson? Or do we just need to rethink what the definition of scaling laws looks like?

No, I don't think of anything as a refutation of the bitter. Really, our company is grounded in we want simple ideas that scale. I think RL is an embodiment of that. I think pre-training is an embodiment of that. And really at every single scale, we face some kind of difficulty of this form. It's just like you gotta find some innovation that gets you past the next bottleneck. and this doesn't feel fundamentally very different from that.

What's most important right now on the actual compute side, we heard from NVIDIA earnings that we didn't get a ton of guidance on the shift from training to inference usage of NVIDIA GPUs, but it feels like it must be coming. It feels like this inference wave is is happening, are those even the right buckets to be thinking about tracking metrics in terms of the story of artificial intelligence? Because yeah, I mean, it's like if the reasoning tokens are inference tokens, but they're what lead to higher intelligent, more intelligent models, like it's almost back in the training bucket again. What bucket should we be thinking about?

And.

How firmly are we in the applied AI era versus the research era?

Well, I think research is here to stay. And it's for all the reasons I mentioned above, right? It's such a rich time to be doing research. But I do think inference is going to be increasingly important as well, right? It's such a core part of RL that you're doing rollouts. I think we see 2025 as this year of agents. We think of it as a year where models are going to do a lot more autonomous work. You can let them be unsupervised for much longer periods of time. And that is just going to put big demands on inference. When you think about our overall vision, we lay it out as a series of steps and levels on the way to AGI. And I think the pinnacle, really that last level, is organizational AI, right? Like you can imagine a bunch of AIs all interacting. And yeah, I think that's just going to put huge demands on inference, right?

On that organizational question, I remember reading AI 2027 and one of the things that they proposed was that the AIs would actually like literally be talking to each other in Slack. Does that seem like the way you Do you imagine agents playing out, like using the same tools as humans instead.

Of kind of- One agent says, I'm going to go talk with teams. I'm going to talk with Slack. I'm going to do a little negotiating.

But maybe it just happens super, super fast 24/7, or is there like a new machine language that emerges?

Yeah, I mean, I think one thing that's really helped us so far in AI development is to come in with some priors for how humans do things. And that's actually, if you bake those piers in, they typically are great starting points. So I could imagine maybe you start with something that's Slack-like and give it enough flexibility that it can kind of develop beyond that and really figure out the way that's most effective for it to communicate. One important thing though is we want interpretability too, right? I think it's very helpful for us today that what the agents do is easy for us to read and interpret. And I don't think you want that to go away as well. So I think there's a lot of benefits just even from a pure debug the whole system perspective. So just let the models speak in a way that is familiar with us. And you can also imagine we might want to plug in to the system too. So whatever interfaces we're familiar with, we would ideally like our model to be familiar with as well.

Yeah.

I think it's also pretty compatible with, you know, we hit a big milestone. We got, I think, 3 million paying business users fairly recently.

Let's go. Yeah, there we go. Let's go.

And, I think.

Three gong hits for three million.

The gong will keep ringing for a while.

Sorry, we had to do it. I was hoping you would drop a number. Yeah, yeah, congratulations. That's actually huge. That's amazing.

Yeah, yeah, yeah. But I think one big part of that is, you know, we have connectors now, right? We're connecting into G drives and I think, yeah, you can imagine, you know, like Slack integrations, things like that. I think we just want the models to be familiar with the ways we communicate and get information.

Yeah. Can you talk about benchmarking? It feels like we're, Potentially, yeah.

Do you think about benchmarks at all?

Oh, yeah, a lot. I mean, but I think it's a difficult time for benchmarks, right? I think we used to be in this world where you have these human written benchmarks for other humans, right? And I think we all have these norms for like, what are good benchmarks, right? Like we've all taken the SAT. We all have like a good conception of what it means to get, you know, whatever score on that. But I think the problem is the models are already at the point where for even the hardest human-written benchmarks for other humans, it's really near-saturated or saturated, right? I think one clear example here is the AMIE, probably the hardest auto-gradeable human math eval, at least in the US. And yeah, the models are consistently getting 90 plus percent on these. And so what that means is I think there's kind of two different things that people are doing. They're developing kind of model-based benchmarks. They're not kind of things that we would give to an ordinary human. Things like humanities last exam, things like Epic AI that are really, really at the frontier of what people can do. And I think the hard thing is it's not grounded in intuition.

Right?

Like, you don't have a lot of people who have taken these exams. So it makes it harder to kind of calibrate on whether this is a good exam or not. One of the exciting things that's on the flip side of that is I really do think we're at the era where models are going to start innovating, right? Because I think once you've passed the last kind of like the hardest human rating exams, that's kind of at the edge of innovation. And I think you already see that with the models, right? Like there, helping to write parts of papers. And I think the other kind of way that people have shifted is, you know, there's these ultra frontier evals, but they're also people kind of just indexing on real world impact, right? You look at revenue, kind of the value you deliver to users. And I think that's ultimately what we care about.

Can you bring that back to interpretability research? with these super, super hard math evals, for example, are we doing the right research to understand if the thought process mirrors, not just one-shotting the answer, oh, you memorized it or you magically got it correct, but you actually took the correct path, kind of like, you know, you're graded for your work, not just the answer if you're in grade school. And, you know, Dario said that, interpretability research will actually contribute to capabilities and even give a decisive lead. Do you agree with that? What's your reaction to that concept of interpretability research being very important?

Yeah, I mean, we care a lot about it here at OpenAI as well. So one thing that we care a lot about is interpreting how the model reasons, right? Because I think we've had a very kind of specific and strong view on this in that we don't want to apply optimization pressure to how the model thinks so that it can be faithful in the way it thinks and to expose that to us without any kind of incentives to cater to what the user wants. I think it's actually very important to have that unfiltered view because oftentimes, if the model isn't sure, you don't want to hide that fact just for it to kind of please the user. and sometimes it really isn't sure. And so we've really done a lot of work to try to promote this norm of chain of thought faithfulness and interpretability. And I think it gives you a lot of sense into what the model is thinking and what are the pitfalls that it can go off into if it's not reasoning correctly.

That's such an important point because if you have somebody on your team and they come to you and they say, Hey, I think this is the right answer, We should probably verify it. It's like, it's still valuable. Totally. It puts you on the right path. If somebody comes to you 100% confidence, this is the truth, and I'm being.

Wrong, it's like, well, trust is just destroyed. Yeah, totally.

Yeah.

Don't you guys feel like, you know, safety felt a lot more theoretical a couple of years back, right? But like today, you know, the things that people were talking about a couple of years, like scalable oversell, really having the model be able to tell you like, and convince you that the work.

It did was right, it feels so.

Much more relevant right now. Just because 100% the capabilities are so strong.

Yeah, I mean, just personally, I've completely flipped from being like, oh, the safety research is not that valuable because I'm not that worried about getting paperclipped. It just seems like a very low likelihood that that's kind of like the bad ending, like immediately in this Foom and all this crazy, the gray goo scenarios were just so abstract in sci-fi. It just felt like economics will fall into place and there will be like a like a cold, like the nuclear ending, which is like we didn't build nuclear plants and we just stopped everything because we see humans seem to be good at that. But now that we're actually seeing things come up.

Yeah, it's crazy how fast it's been, right? Like, oh yeah, I think my personal story is, it's like, you know, what got me into AI was AlphaGo, right? Like just watching it get to that level of capability.

Yeah.

And you were kind of like, it was such an optimistic and also kind of a little bit of a sobering message, right? When you saw at least it'll beet. And I just remember, you know, like we saw the coding models, you know, when we first launched, like, I think very OG Codex, you know, with GitHub Copilot, it was maybe like under, you know, a thousand ELO on, on Codeforces. And I still remember the meeting where I walked into where the team showed my score and they're like, hey, the model's better than you. And it's like, you come full circle and it's like, wow, like I put decades of my life into this and the capabilities are there. So if I'm at the top of my field in this thing and it's better than me, what can it do?

Yeah, that's amazing. I have so many more questions. On AlphaGo, are there lessons from scaling, how scaling played out there that we can abstract into the rest of AI research? What I mean is, As I remember it, the AlphaGo training run was not 100K H200s. But what would happen if we actually did an AlphaGo style training run? I mean, it would be an economic money pit, right? They had no economic value to do. But let's just say some benevolent trillionaire decides, I'm gonna spend a billion dollars on a training run to beat AlphaGo and go even bigger. is Go at some point solved? Would we see kind of diminishing scaling curves? Could we throw extra RL? Could we port back everything that we're doing in just general AGI research and just continue fighting it out in the world of Go? Or does that end and does that teach us anything?

Yeah, yeah. Honestly, I feel like if you really are curious about these mysteries, join our team. I think you want to say. But yeah, I mean, Really, like kind of the central problem of today is RL scaling, right? When you look at AlphaGo, right? It's a narrow domain, right? And I think in some sense that limits the amount of compute you can pump into it. But even kind of small toy domains, they can teach you a lot about how you scale RL. Like what are the axes where it's most productive to pump scale in? I think a lot of scaling research just looks like that, whether it's on RL or pre-training. So you identify a lot of, you know, different variables under which you can scale and where is kind of where you get the best kind of marginal impact for a pumping scale there. I think that's a very open question for RL right now. And I think what you mentioned as well, it's just going from narrow to broad, right? Does that give you a lever to pump a lot more scale in as well? I think when you look at our reasoning models today, they're a lot more broad based than just being able to kind of an expert system on go. So yeah, I really do think that there are so many levers to scale.

What about Move 37? That was such an iconic moment in that AlphaGo Lisa Dal match. They placed Move 37. It's very unconventional. Everyone thinks it's a blunder. It turns out not to be. It turns out to be critical. It turns out to be innovation. Do you think we are, we're certainly post-touring test in language models. We're probably post-touring test in image generation. But it feels like we're pre-Move 37 in text generation in the sense that there hasn't been like a fully AI generated book that everyone is just, oh, it's the new Harry Potter. Everyone has to read it. It's amazing. And it's fully generated or this image the images, they do go viral, but they go viral because they're AI. Move 37 in the context of Go did not go viral because it was AI. It felt like it was actual innovation. So is that the right frame? Does that make any sense?

Yeah, I think it's not the wrong frame. So I think some quick thoughts on that. I think kind of when you have something that's very measurable, like win or lose, right? Something like Go. It's very easy for us to just judge, did the model do something right here? And I think the more fuzzy you get, it is just harder. When it comes to, is this the next Harry Potter? It's not a universally loved book, I think. Fairly universal, but there's some haters. I think it is just kind of hard when it comes to these human subjective things.

Right.

Where it's really hard to put down in words like what makes you like Harry Potter, right? And so I think those are always going to lag a little bit, but I think we're developing more and more techniques to attack these more open-ended domains. And I don't know, I wouldn't say that we're not at an innovative stage today. So I think my biggest touch with this was when we had the models compete on the IY last last year. So I highlight the international, basically Olympics for computer science, basically the top four kids from each country go and compete. And these are really, really tough problems, basically selected so that they require some innovative insight to solve. And we did see the model come up with with solutions, even to some very ad hoc problems. And so I think there was a lot of surprise for me there, right? I was completely off base about which problems the model would be able to solve the most, right? I think I kind of categorized there's six problems. Some of them as more kind of like, oh, this is standard, a little bit more standard, this is a little bit more out of the box. I was like, it's not going to be able to solve this more out of the box one, but It did. And I think that really does speak to kind of these models have the capacity to do so, especially trained with RL.

Now, put that in context of what's going on with Arc AGI. Obviously, OpenAI has made incredible progress there, but when I do the problems, it seems easy. And when I look at the IOI sample problems, I think this would be a 20-year process for me to figure out how to achieve that. And I can do the Arc AGI on my phone. Is this the spiky intelligence concept? Is this something that a small tweak in algorithmic design just one shots AGI or Arc AGI? Or is there something else going on there that we should be aware of?

Yeah, I think part of this is the beauty of Arc AGI as well. I'm not sure if there's another kind of human intuitive simpler benchmark for the models. I think really that's one of the things they really optimize for on that benchmark. I do think when it comes to models though, there's just a little bit of a perception gap as well. Models aren't used to this kind of native, just screen type input. I think there's a lot we can bridge there. Actually, even O4Mini, it's a state-of-the-art multimodal model in many ways, including visual reasoning. you're starting to build up the capacity for the models to take images, manipulate and reason about them, generate new images, write code on images. And I think it's just been under-focused. But I think when I talk to researchers in the field, they all see this as a part of intelligence too. And we're going to continue to focus there.

Yeah. Is RKGI kind of in the, if we're dropping a buzzword on it, is program synthesis? a world where I know the tokens, the images, we see them as renderings of squares and different colors, but when they're fed into the LLM, they're typically just a stream of numbers effectively. Is there a world where actually adding a screenshot is what's important? Like visual reasoning?

Yeah, yeah. So I think that could be important. It's just like, kind of, you know, whenever it comes to textual representation of grids, Models today just don't really do that well, right? And I think it's just kind of because humans don't really ever write down textual representations of grids. We have a chessboard, no one really just types it out in a grid. And so the models are kind of under-trained a little bit on what that looks like and what that means. Sure. you know, I, I think with more reasoning, it will, we'll just bridge the gap. Um, I think with better visual perception, we'll just bridge that gap.

Yeah. How are you thinking about the role of non lab researchers in the ecosystem today? I'm sure you try to recruit some of the best ones, but the ones that don't join your team.

Tell us about the one that got away.

Yeah, the one that got away. Yeah.

No, I mean, I think it's still actually a fairly good time for specific domains, right, to be doing research. And, you know, I think the style is just very different. And you do feel the pull of non-lab researchers into labs because I think they feel like a lot of the burning problems in the field are at scale, right? That's correct. And that's kind of one of the unfortunate things too, right? Like when you look at reasoning, you just don't see that happen at small scale.

Right?

There's like a certain scale at which it starts becoming signal bearing and that requires you to have resources, right? But I do think, you know, a lot of the really good work that I've seen, you know, there's experimental architectures. I think a lot of good work is happening in the academic world there. Like a lot of study in optimization, a lot of study in kind of like GANs, you know, there's certain fields where you see a lot of fruitful research that happens in academia.

Yeah, that makes a lot of sense.

How about consumer agents? How are you thinking about them? You talked earlier about B2B adoption, and that's all very exciting. But how much do you and the research org think about breakout consumer agent products?

Yeah, that's a fantastic question. I think we think about it a lot. I think that's the short answer. we really do think this year we're trying to focus on how we can move to the agentic world. When I think about consumer agents, I think ChatGPT proved that people got it, right? It's like people get conversational agents, conversational kind of models. But when it comes to consumer agents, we have a couple of theses that we've tried out in the world. I think one is deep research, right? I think this is something that can do five to 30 minutes of work autonomously, come back to you, and really like kind of synthesizes information, right? It goes out there, gathers, collects, and kind of, you know, compresses the information in a form that's useful.

A little bit of pushback there. Like, I can see that as a consumer product when someone like Aiden is like, I want new towels, and he uses deep research to like figure out like what is the best towel across every dimension. But when I think of deep research, yes, it has applications with students, but it's often- Some of it might just.

Be the paradigm because I feel like.

Consumers being like, We keep using this flight report on this country and where to travel and things like that.

We keep using this flight example, but I haven't actually tried to book a flight with deep research. It's totally possible that it could go and pull all the different flight routes and calculate all the different delays. all the different parameters of if I fly to this airport, I can park or I can use valet here or something like that. Yeah. Yeah.

And I guess when I think of agents, it's deep research is curating information on which you can take action on, but it's like at what point is action a part of that loop, right? Where you can not only curate a list of flights that you want, but then actually go out and have agency.

I think one of our explorations in that space is Operator, right? It's where you kind of just feed in raw pixels from your laptop into, or from some virtual machine into the model, and it produces either a click or some keyboard actions, right? And so there it's taking action. And I think the trouble is, you don't ever want to mess up when you're taking action. I think the cost of that is super high. You only have to get it wrong once to lose trust in a user. And so we want to make sure that that feels super robust before we get to the point where we're like, Hey, look, here's a tool.

It's so different than deep research because you can wind up on some news article and read one sentence that gets a fact wrong. or the commas in the wrong place and the numbers off. But that's just the expectation for just text and analysis. And if you delegated that, yeah, you're going to expect a few errors here and there. Oh, that's actually a different company name, or that's an old data point. There's new data. But very different if I book a flight and you book the wrong flight, and I can wind up in Chicago instead of New York.

Exactly. And I think the reason why we care so much about reasoning is because I think that's the path that we get reliable agents through. Sure. You know, we've talked about like reasoning helping safety, but reasoning is also helping reliability, right? It's like you imagine like, what makes a model so good at a math problem? It's like it's banging its head against it. It's trying a different approach and then it's like adapting based on what it failed at last time. And I think that's the same kind of behavior you want your agents to have. It's like, yeah, tries things like it adapts and keeps going and that's the.

Humans do this every day. You're booking a flight, you keep hitting an error, you're not sure which form you missed, right? And you're just sort of banging your head against the computer and eventually it says, okay, you're booked, right? So I think that's a great call out.

Yeah. I mean, there's so many more questions we could go into, but I'm interested in the scaling of RL and kind of the balancing act between pre-training RL and inference, just the amount of energy that goes into getting a result when you distribute it over the entire user base. How is that changing? And I guess, are we post really big runs? Is this going to be something that's continually happening online? It feels like we're moving away from the era of, oh, some big development, some big run happened, and now we're reaping the fruits of it versus a more iterative process.

Yeah, I mean, I don't see why it has to be so, right? I think, like, if you find the right levers, you can really pump a lot of compute into RL as well as pre-training. I think it is a delicate balance, though, between all of these different parts of the machine. And, you know, when I look at my role with Jakob, it's just kind of like figuring out where how this balance should be allocated, where the promising kind of nuggets are arising from and resourcing those. Yeah, it's kind of a, in some sense, I feel like part of my job is a portfolio manager.

Yeah.

That's a lot of fun.

Well, thank you so much for joining. This is a fantastic conversation. We'd love to have you back and go deeper.

Great hanging, Mark.

We'll talk to you soon.

Yeah, peace.

Have a good one.

Next up, we have Shalto Douglas from Anthropic coming on this show.

And Jordi is giving us the update on that. No, I'm just getting a lot of messages saying, why no one cares about AI? I'm really not talking about the drama on the timeline. Well, we do care about AI. We care a lot about AI. But it is a mess out there.

Wow. Yeah, the end of the Trump-Elon era. I don't know. Well, maybe we have to get some people on to talk about it tomorrow. tomorrow or something.

Gotta do it today.

Anyway, we have Shalto from Anthropic in the studio. How are you doing?

What's going on?

Good to see you guys.

Hopefully you're staying out of the chaos on the Twitter. Don't open the Twitter. Don't open the Twitter. Mute everything. Stay focused on the application. Stay focused on the next training run.

Humanity really cannot For any distressed researchers to open X today.

What a hilarious day. Anyway, I mean, 24 hours, guys. Yeah. How are you doing? What is new in your world? What are you focused on mostly day to day? And maybe it's just a way of an intro. Yeah.

So at the moment, focused really hard on scaling RL. That is the theme of what's happening this year. And we're still seeing these huge gains where you go, you know, 10x compute increase in RL. we're still getting very distinct linear gains off the basis for that. And because RL wasn't really scaled anywhere close to how much pre-training was scaled at the end of last year, we have basically a gamut of riches over the course of this year.

So where are we in that RL scaling story? Because I remember some of the rough numbers around GPT-2, GPT-3, we were getting up into it cost 100 $1,000,000. It's going to cost a billion dollars. Like it just rough order of magnitude, not even from anthropic, just generally like what is a big RL run cost or how many are we talking 10K H200s or 100K? Like are we going to throw the same resources at it? And if so, how soon? Yeah.

So I think in Dario's essay at the beginning of the year, he said that a lot of runs were only like a million dollars back in like December. I think you have like deep-seek V3 and this kind of stuff like R1. which means that that's like at least two ooms just to get to the scale of GPT-4 and GPT-4 was two years ago. RL is also perhaps a bit more naively parallelizable and scalable than pre-training. Pre-training, you need everything in one big data center ideally, or you need some clever tricks. RL, you could in theory, like what the prime intel folks are doing, scale it all over the world out of it. And so you're held back far less than your pre-training.

Sure. So everyone and their mother has a billion dollars now. There are hundreds of thousands of GPUs getting pumped all over the place. I feel like we're not GPU poor as a society. Maybe some companies need to justify it in different ways, but it sounds like there's some sort of reward hacking problem that we're working through in terms of scaling RL. What are all of those the problems that we're working through to actually go deploy the capital cannon at this problem. Yes. So, I mean, think about what you're.

Asking the model to do in RL is you're asking it to achieve some goal at any cost, basically. And this comes with a whole host of behaviors, which you may not intend. In software engineering, this is really easy. It might try and hack unit tests or whatever. In much more longer horizon, real world tasks, you might ask it to say, go make money on the internet. And it might come up with all kinds of fun and interesting ways to do that unless you find ways to guide it into following the principles that you want it to obey, basically, or to align it with your idea of what's sort of best for humanity. And so it's actually a pretty intensive process. It's a lot of work to find down, hunt down all the ways these models are hacking through the rewards and patch all of that.

Yeah.

Yeah. Are we going to see scaling in the number of rewards that we're RLing against? If that makes sense, I would imagine that at a certain point, unless we come up with the genesis prompt of go forth and be fruitful or something and multiply, you could imagine training runs on just knocking down one problem after another. Is that kind of the path that we're going down?

I very much think so. There's this idea in which the sort of world becomes an RL environment, in some respects, because there's just so much leverage to making these models better and better at all the things we care about. And so I think we're going to be training on just everything in the world.

Got it. And then does that lead to more model fragmentation, models that are good at programming versus writing versus poetry image generation or does this all feed back into one model? Does the idea of the consumer needing to pick a model disappear? Are we in a temporary period for that paradigm?

I think the main reason that we've seen that so far is because people are trying to make the best of the capital. We are all still GPU poor in many ways. And people are focusing those GPUs on the spectrum of wars that they think is most And look, I'm a bit of a big model guy. I really do think that similar to how we saw with large pre-trained models before, where small fine-tuned models made it like had gains over the sort of GP2, GPT2 era, but then we're obsoleted by GPT4 being generally good at everything. I think, to be honest, you're going to see this generalization and learning across all kinds of things that means you benefit from having large single models rather than specialization or area fine-tuned models.

Can you talk a little bit about the transition from or any differences between RLHF and just other RL paradigms?

Yes. So RLHF, you're trying to maximize a pretty deep, like lossy signal. Things like airwise, like what do humans prefer? And I don't know if you've ever tried to do this, like judge to language model responses.

I get prompted for that all the time. Right. And I'm always like, I don't want to read both of those. I'll just click the one on the left.

Exactly, exactly.

I click one of the random ones sometimes. Yeah. Or I click the one that just looks bigger, or I'll read the first two sentences, but yeah, I'm not giving straight. I'm not doing my job as a human reinforcer.

Exactly. Human preferences are easy to hack. Yeah, totally. Environments in the world are much truer if you can find them. So something like, did you get your math question right?

Is a very real and true reward. Does the code compile? Right?

Does the code compile? Exactly. Did you make a scientific discovery? We've got very little rewards right now, but pretty quickly over the next year or two, you're going to start to see much more meaningful and long horizon rewards.

You're going to see models bribing the Nobel Committee to win the Nobel Prize. It's reward hacking. It's reward hacking.

That's the thing you want to prevent, right?

Exactly. Yeah, yeah, that's the real nightmare scenario. What about There's so many different problems that we run into that feel like it's just really, really hard to design any type of eval. My kind of benchmark that I use whenever a new model drops is just tell me a joke. They're always bad. Or even the latest VO3 video that went viral was somebody said, like, stand-up comedy joke. And it was kind of a funny joke, but it was literally the top result for joke Reddit on Google, and then it clearly just took that joke and then instantiated it in a video that looked amazing. But it wasn't original in any way. And so we were joking about, like, the RLHF loop for that is like, you have an endless cycle of comedians running AI generated materials, and then, you know, to speak microphones in all the comedy clubs to feedback what's getting laughs. I mean, honestly, that would work pretty well. If anybody wants to hook us up.

With an RL loop, I mean.

Yeah, yeah, but I mean, for some of those less, like, as you go down the curve, it feels like each one gets harder and harder to actually tighten the loop. We see this with longevity research, where it's like, okay, it takes 100 years to know if you extended a human life. Yes, you could create a feedback loop around that, every change is going to be hundreds of years. And so even if you're on the cycle, it's irrelevant for us in the context that we talk about AI. So talk to me about like, are you running into those problems or will there be like another approach that kind of works around those?

So there are a lot of situations where you can get around this by just running much faster than real time. Like let's say the process of building like a giant app, like building Twitter, right?

Yeah.

It's something that would take human months. But if you got fast enough and good enough AIs, you could do that in several hours. I curate large heaps of AI agents. They're all building the right things to expect. And so you can get a faster reward signal in that way. In domains that are less well specified, like humor, I agree. It's really, really hard. And this is why I think in some respects, creativity is at the top end in the spectrum. True creativity is much, much harder to replicate than the sort of analytical, scientific style reasoning. And that will just take more time. You know what? The models actually are pretty good at making jokes about being an AI. This feels weirdly fresh. Like everything else is kind of a weird copy of something. It feels like it's derivative, basically. It's trying to infer what humor is and it doesn't really understand it. But jokes about being an AI are quite funny.

Yeah, I think this also might be, I don't know if it was directly reward hacking, but I noticed that one of the new models dropped and a bunch of people were posting these like 4chan, like beanie memes. and it seemed like they were kind of hacking the humor by being hyper specific about an individual that they could find information on online. And so you're laughing at the fact that it's like, oh wow, that is something that I've posted about. It's making a reference, but it's not really that funny to me. Other than it's just like, wow, they really did its research. It really knows Tyler Cowen intimately, which is cool, but I didn't find it hilarious. Yeah, yeah, very interesting. Let's talk about some sort of deep research projects and products. We were talking to Will Brown and he was saying like, AGI is here with some of the bigger models, but the time that AGI can feel consistent, it diverges. And so you could be working with someone who's 100 IQ, but they will stay consistent for years as an employee or they'll keep, you know, living their life. Whereas a lot of these super smart models are working really well and then after a few minutes of work, the agents kind of diverge and kind of go into odd paradigms. It feels very not human. It feels like just a they're hyper intelligent in one way and then extremely stupid in others. What's going on there? What is the path to extending that? Is that more like having more better planning and better dividing up the task? Or will this just naturally happen through the RL and scale?

Yeah, so there's that jaggedness, right, which is what you're seeing is how we call it. And I think that is largely a consequence of the fact that maybe something like Deep Suite research, it's probably been RL to be really good at producing a report. But it's never been RL'd on the act of producing valuable information for a company over a week or a month or making sure the stock price goes up. in a quarter or something like this, right? It doesn't have any conception of how that feeds into the broader story at play. It can kind of infer it because it's got a bit of world knowledge from the base model and this kind of stuff. But it's never actually been trained to do that in the same way humans have. So to extend that, you need to put them in much longer running, much like long horizon things. And so deep research needs to become, you know, like deep operator company for a week kind of thing.

Sure. Is that the right path? It feels like the road might be, the longest running LLM query used to be just a few seconds, maybe a few minutes. And I remember when some of the reasoning models came out, people were almost trying to stunt on it by saying, oh, I asked it a hard question and thought for five minutes. Now deep research is doing 20 minutes pretty much every time. Is the path two hours, two days, or are we going to see more efficiency gains such that we just get the 20-minute model, the 20-minute results in two minutes and then two seconds?

Yeah, so this is somewhere where inference in many respects and prorogation becomes really important. So both how fast is your inference? That literally affects the speed at which you can think and the speed at which you can do these experiments. Also how easily you can prorogate becomes really important. Can you dispatch a team of subagents to go and do deep research? like sub reports for you so that you can do everything in parallel. These kinds of like, it's both like there's an infrastructure question here that feeds up from the hardware and the chips and this kind of stuff to designing better chips for better inference in all this and an RL question of like, you know, how well can you parallelize and all this. So I think we just need to compress the timelines, compress the timeframes, basically.

Yeah.

If I'm an extremely big model and I'm running an agentic process, how much am I hankering for a middle-sized model on a chip or baked down into silicon that runs super fast? Because it feels like that's probably coming. We saw that with the Bitcoin progression from CPU to GPU to FPGA to ASIC. Do you think we're at a good enough point where we can even be discussing that. Because every time I see the latest mid-journey, I'm like, this is good enough. I just want it in two seconds instead of 20. But then a new model comes out, I'm like, oh, I'm glad I didn't get stuck on that path. But yeah, how far away are we from, okay, it's actually good enough to bake down into silicon? Well, there's a question here of baking.

It down to silicon versus designing chip, which is like very suited for the architecture that you care about.

Right?

And baking out of silicon, I'm sure. Like, I think that's a bet you could take, but it's a risky one because the pace of progress is just so fast nowadays. And I really only expect it to accelerate. But designing things that make a lot of sense for the sort of, you know, transformers or architectures of the future should make a lot of sense.

That's a big gap though, transformers or architectures of the future. if we diverge, there's a lot of companies that are banking on the Transformer sticking around. What is your view on Transformer architecture.

Sticking around for the next couple of years? I mean, look, they've stuck around for five years, so they might stick around for a little while. But there's different-- you think about architectures in terms of this balance of memory bandwidth and flops, right? One of the big differences we've seen here is Gemini recently had actually a diffusion model that they released. I was about to ask you the other day, right?

Yeah.

So diffusion is an inherently extremely flops-intensive process. Whereas normal language model decoding is extremely memory bandwidth intensive. You're designing two very different chips depending on which bet you think makes sense. And if you think you can make something that does flops like four times faster than diffusion and like four times cheaper than your other those could, diffusion makes more sense. So there's this dance basically between the chip providers and the architecture both trying.

To build for each other but also build for the next paradigm. It's risky. Do you, I don't know how much you've how much you've played with image generation, but do you have any idea of what's going on with images in ChatGPT? It feels like there's some diffusion in there, there's some tokenization, maybe some transformer stuff in there. It almost feels like the text is so good that there's like an extra layer on top almost, and that it's almost like reinventing Photoshop. And I guess the broader question is like, it feels like an ensemble of models, maybe the discussion around around just agents and text-based LLM interactions shouldn't necessarily be Transformer versus diffusion, but maybe how will these play together? Is that a reasonable path to go down?

Well, I think pretty clearly there's some kind of rich information channel between even if there are multiple models there, it's conditioning somehow on the other model. Because we've seen before, let's say when models use mid-journey to produce images, It's never quite perfect. It can't perfectly replicate what went in as an input. It can't perfectly adjust things. So there's a link somehow, whether that's the same model producing tokens plus diffusion.

I don't know.

Like, yeah, can't comment on what OpenAI is doing there.

Yeah, yeah, yeah. Are there any other kind of like super wild card, long shot research efforts that are maybe happening even in academia? Where, I mean, this was the big thing with, what was his name, Gary, who was talking about, I forget what it was called, symbolic manipulation was a big one. And I feel like you can never count anyone out because it might come from behind and be relevant in some ways. But are there any other research areas that you think are purely in the theory domain right now that are worth looking into or tracking that you know, low probability, but high upside if they work. Hmm. That's a tough one.

This is a tough one. But we'll say it's not a symbolic thing.

Please.

It's crazy how similar transformers are to systems that manipulate symbols.

Sure.

What they're doing is they're taking a symbol and they're like converting it into a vector and then they're manipulating and.

Moving information around across them. Sure.

Like this whole debate that all transformers cannot represent symbols and they cannot do this It's, it's, yeah, it's not really.

So Gary Marcus underrated or overrated, I guess. Yeah, yeah, but, but I mean, if you, if you twist the, if you twist it so much, you wind up with saying like, well, really, like the, the transformer fits within that paradigm. And so maybe it's, you know, it's, you know, it, like the rhetoric around it being a different path was maybe false. the whole time. Yeah, something like that. But as I remember that debate, it was really the idea of compute scaling versus almost like feature engineering scaling. And will the progress scale with human hours or GPUs essentially? And that has a very different economic equation. And it feels like there's been some rumblings about maybe with a data wall will shift back to being human labor bound. But do you think that there's any chance that that's relevant in the future or is it just algorithmic progress married with bigger and bigger data centers in the future?

So I'm pretty bit of a lesson built since that I do think removing as many of our biases and our clever ideas from the models is really important, just freeing them up to learn. Now, obviously there is clever structure that we put into these models such that they're able to learn in this extremely general way. But I am more convinced that we will be compute bound than we will be human researcher out, human researcher out bound on this kind of thing. We're not going to be feature engineering and this kind of stuff. We're going to be trying to devise incredibly flexible learning systems.

Yeah, that makes sense. On the scaling topic, part of my worry is that the ooms get so big that they turn into these mega projects that at a certain point you're bound by the laws of physics because you have to move the sand into silicon chips and you have to dig up the silicon and- There's a certain sand. Yeah, there's only so much sand and the math gets really, really crazy just for the amount of energy required to move everything around to make the big thing. Where are you on how much scale we need to reach AGI? Whether or not we will see the laws of physics start acting as a drag on progress, because it certainly feels exponential. We're feeling the exponentials, but a lot of these turn into sigmoids, right?

So I think we've got, what, like two or three more ooms before. Before it gets really hard. Leopold has this nice table at the end of his.

Yeah.

Situational awareness where, I think, like 2028 or something is when, under really aggressive timelines that you get to 20 of US energy production.

Yeah.

It's pretty hard to go exponentially Beyond 20 of US energy provision. Now, I think that's enough. Every indication I'm seeing says that's enough. Now, there might be some complex data engineering, we're one engineering, this kind of stuff that goes into lots of, there's still a lot of algorithmic progress left to go. But I think that with those extra OOMs, we get to basically a model that is capable of assisting us in doing research and software engineering.

Yeah, which is the beginning of the self-reinforcing feedback loop.

Yeah, interesting.

Is that just a coincidence? Like this feels like one of those things this feels like one of those things where, like, the moon is the exact same size as the sun and the sky. It's like, oh, it just happens that AGI happens within this time. Did you have you unpack that anymore? Because it feels convenient. Not, not to, you know, I know.

There's a lot of weird conveniences or, like, weird. It's a good sci-fi story, let's say. Totally. You know, we've got, you know, Taiwan in between China and the US, and.

It produces the most valuable material in the world.

That's locked between the two. Incredible plot. Yeah. Yeah, really bad for the people that don't believe in simulation theory. It really feels like, oh, this is scripted. It's fascinating. Talk to me more about getting to an ML engineer in AI and that reinforcement. I imagine that you're using AI codegen tools today and Anthropic is broadly and everyone is. But what are you looking for and what's the shape of the spiky intelligence? Where do they fall flat and what are you looking to knock down in the interim before you get something that's just like go? Yeah.

So, I mean, we definitely use them. The other night, I was a bit tired. I asked to do something, just sat watching it in front of me working for half an hour. It was great. It was truly a weird experience, particularly when you look back a year ago and we're still copy-pasting stuff between a chat window and and a code file. I like meters evals for this kind of stuff. So they have a bunch of evals where they measure the ability to write a kernel, the ability to run a small experiment and improve a loss, and they have these nice progress curves versus humans. And I think this is maybe the most accurate reflection of what it will take for it to really help us at Dream Progress. And there's a mix here where they're not so great at the moment is large-scale distributed systems engineering.

Right?

Like debugging stuff across heaps and heaps of accelerators and like the way the feedback loops are slow and you're actually like, if your feedback loop is like an hour, then it's worth you spending the time on doing something. Feedback loops 15 minutes.

And for context there, the hour long feedback loop is just because you have to actually compile and run the code across everything.

Like spin up all your machines or you need to like, you need to run it for a while Yeah. To see if something's gonna happen. Like at that point in time, you're still cheaper than the chips. Sure. So you're, you're, you're sort of, it's better that you do it.

Yeah.

Um, but for things like, you know, kernel engineering, or for, like, you know, actually even just understanding these systems incredibly helpful. Like I, one thing I regularly do at the moment is in parts of the code base in like languages that I'm unfamiliar with or some stuff like.

This, I'll just ask it to rewrite.

The entire file, but with comments on every line. Mm-. Game changing.

Well, it's like comments on everything. Yeah, or just hunt through like thousands.

Of files and explain how everything interacts to me, draw diagrams, this kind of stuff. It's really, yeah.

How important is a bigger context window in that example you gave that feels like something that's important and yet just naively, like Google's the one that has the million token context window. I imagine that all the other Frontier Labs could catch up, but it seems like it hasn't been as much of a priority as maybe the PR around it sounds like. Is that important? Should we be driving that up to a trillion token window? Is that just going to happen naturally? There's a nice plot in the Gemini.

1.5 paper where they show the loss over tokens as a function of context length. And they show that the loss goes down quite steeply actually, as you put more and more and more of a code base into context, you get better and better and better at predicting the rest. Yeah, that makes sense. With context length, it's a cost. The way Transformers work is that there's this memory that is proportional, the KB cache is proportional to how much context you've got. And so you can only fit so many of those into your various chips and this kind of stuff. And so a longer context actually just costs more because you're taking up more of the chip and you're sort of like you could have otherwise been doing other requests.

So bringing it back to the custom silicon, is that a unique advantage of the TPU? Is that something that Google has thought about and then wound up to put themselves in this advantage position? Or is it a durable advantage even? Yeah.

So GPUs are good in many respects, partially because you can connect hundreds or thousands of them really easily across really great networking. Whereas only recently has that been true for GPUs.

Like NVLink.

Yeah, with NVLink and the NVL 72 stuff. So it used to be like eight GPUs in a pod, and then you connect them over worse interconnect. And now you can do 72 and then it breaks down. With Google TPUs, you can do like 4,000, 8,000 of a really high bandwidth interconnect in one pod. And so that is helpful for things like just general scaling in many respects. I think this is doable across any chip platform, but it is an example of somewhere that being fully vertically integrated is using a benefit.

Yeah, that makes sense. Talk to me about Arc AGI. Why is it so hard? It seems so easy.

It does seem easy, doesn't it?

It certainly seems more evaluatable than tell me a funny joke, right? Yeah, yeah.

I mean, I think if you RRL on ArcGI, then you probably get superhuman at it pretty fast. But I think we're all trying not to RRL on it so that it functions as an interesting held out.

Sure. Okay. Wait, is that just an informal agreement between all the labs? basically.

We're trying to have a sense of honor between us.

That's good.

Sense of honor.

That's amazing.

How many people on Earth do you think are getting the full potential out of the publicly available bottles? Because we're now at a point where we have, you know, billion plus people are using AI almost daily. And yet I have to admit, my sense would be it's maybe like 10,000, 20,000 people on the entire planet are getting that sort of full I'm curious what your assessment would be.

Yeah, I completely agree. I mean, I think that even I don't get the full potential out of these models often. And I think as we shift from you're asking questions and it's giving you sensible answers to you're asking it to go do things for you that might take hours at a time and you can really like paralyze and spin that we're going to hit like yet another inflection point where even less people are like really effectively using these things. Because it's basically going to require you to like, it's like StarCraft or Dota. It's going to be your APM of managing all these agents. And that's going to be a process.

Yeah.

So StarCraft is such a good example. You think you're just absolutely crushing it, and then you realize there's an entire area of the map you're just getting destroyed. It's such a good comp.

That's great. Anything else, Jordy?

I think that's it on my side. I mean, I would like this to be an evolving conversation.

Yeah, this was fantastic.

We'd love to have you back on.

Very bad conversation. Absolutely.

It's really fun.

Love to get back on.

Yeah, we'll talk to you soon.

Cheers, Sholto. Have a good one.

All right, we got it. The worst possible AI day. Yeah. So for context, folks, we are going to be doing a live timeline and turmoil segment at 2 p.m. PST. Let's do it. So if there's posts you want to cover, send them in. We can go send them. I'll put this up.

We can review a few, a few more, pull one up. I'm going to do some ads because we got M&H here coming in the temple in just a few minutes. First, let me tell you about Numeral, sales tax and autopilot. Spend less than five minutes per month on sales tax compliance. Go to numeralhq.com, very excited for them. Also go to public.com, investing for those who take it seriously. They have multi-asset investing, industry-leading yields, and they're trusted by millions.

Millions.

In other news, Tim Sweeney continues to battle Apple. Apparently, if you search for Fortnite on the Apple App Store, he says, Hey kids, looking to play Fortnite? Try this crypto and stock trading app instead, rated for ages four plus, courtesy of Apple App Store ads.

Obviously, I'm going to give you the latest Trump terror. on Post four minutes ago. The Trump tariffs will cause a recession in the second half of this year.

Wow.

Somebody else was saying, can I finally say that Trump's tariffs are super stupid?

Who is that?

Somebody else is posting. Mads posting is saying it's a Xi Jinping. He says, bro, you seeing this? And it's Putin on the other end. He's just looking at it. Hold up. Got a line. And it's We'll start pulling some of these up.

Ridiculous. What else is going on here? This is the president versus Elon Musk.

Nival says Elon's stance is principle. Trump's stance is practical. Tech needs Republicans for the present. Republicans need tech for the future. Drop the tax cuts, cut some pork, get the bill through.

This is so crazy.

Antonio Garcia says, remember, there's F you money and then there's F the world money. Will Stansel says, imagine being the ice agents suiting up for your biggest mission of all time right now. People are saying that Trump's gonna deport you.

Elon back to South Africa.

Will dePue says, time to drop the really big bomb. Growing Daniel is in the Epstein files.

That is turned into a Pasta.

That is a real.

Oh, no.

Delian, we had a question from a friend of the show. They said the real question is if Tesla is down 14%, how could SpaceX and OpenAI be trading if they were. How would they be trading if they were public? The real thing here is it's bad for everyone, right?

DJT, for everyone, Trump coin is down.

Nobody's really winning here. China's up. Yeah.

Oh, really?

Sean McGuire, I mean, I'm just saying at a high level, China is the big beneficiary here of... Sarah Guo says if anyone has some bad news to bury, might I recommend right now?

Yes, yes, yes. If you have... What's the canonical bad startup news? Oh, yeah, you missed earnings or something? something. Drop it now.

Inverse Kramer says Bill Ackman is currently writing the longest post in the history of this app. Okay.

And we have a video from Trump here, if we want. I can throw it in the tab and we can share it on the stream and react to it.

Lex Friedman says to Elon, that escalated quickly. Triple your security. Be safe out there, brother. Your work, SpaceX, Tesla, XAI, Neuralink is important for the world.

We need to get Elon on the show today. If somebody's listening and can make that.

Happen, I would love to hear from Max Meyer. So I got this wrong. I didn't say it never happened, but I thought it wouldn't. I am floored at the way this has happened. He didn't think they would have a big breakup. Many people didn't think they would have a big breakup. Even just earlier this week, it seemed like they might just have a somewhat Peaceful exit. Trump just posted a little bit ago, I don't mind Elon turning against me, but he should have done so months ago. This is one of the greatest bills ever presented to Congress. It's a record cut in expenses, $1.6 trillion, and the biggest tax cut ever given. If this bill doesn't pass, there will be a 68% tax increase and things far worse than that. I didn't create this mess. I'm just here to fix it. anyways, lots going on. Let's go to this, this Trump video. I want to see what it is.

This is something that I've seen, and I'm sure you've seen regarding Elon Musk.

And your big, beautiful bill.

What's your reaction to that?

Do you think it in any way hurts passage in the Senate, which, of course, what is your seeking?

Well, look, you know, I've always liked Elon, and it's always very surprised. You saw the words he had for.

Me, the words, and he has a.

Said anything about me that's bad.

I'd rather have him criticize me than.

The bill, because the bill is incredible. Look, Elon and I had a great relationship. I don't know if it went well anymore. I was surprised because you were here.

Everybody in this room practically was here as we had a wonderful sendoff. He said wonderful things about me.

You couldn't have nicer.

Said the best things.

He's worn the hat. Trump was right about everything. and I am right about the great, big, beautiful bill.

But I'm very disappointed because Elon knew the inner workings of this bill better than almost anybody sitting here, better than you people. He knew everything about it. He had no problem with it. All of a sudden, he had a problem. And he only developed the problem when he found out that we're going to.

Have to cut the EV mandate because.

That'S billions and billions of dollars, and.

It really is unfair. We want to have Cars of all types electric we want to have electric, but we want to have gasoline, combustion. We want to have different we want to have hybrids. We want to have all we want.

To be able to sell everything.

He hasn't said bad about me personally, but I'm sure that'll be next. But I'm very disappointed in Elon. I've helped Elon a lot. Did that, Mr. President? Did he I just want to clarify did he raise any of these concerns with you privately before he raised them publicly? And this is the guy you put in charge of cutting spending? Should people not take him seriously about spending now?

Are you saying this is all sour grapes? No, he worked hard and he did a good job.

And I'll be honest, I think he misses the place.

I think he got out there and all of a sudden he wasn't in.

This beautiful Oval Office and he was, and he's got nice offices too.

But there's something about this when I was telling the Chancellor, Folks. This is where it is. People come in here.

Breaking news.

Delian. That's Bruhav. Is joining us in the temple.

I love it.

For some live reactions.

Come on in. Surprise guest. I can't even spell surprise guest. I'm so excited about this. Surprise guest. Yeah.

In other news, 11 Labs dropped a new product.

Another news. Two million dollar seed round. Stop it. Stop it. We love 11 Labs.

No, they'll keep grinding.

But just launch again.

You're going to have to launch again.

Start shooting a new vibe reel. Start shooting a new, writing a new blog post, because no one's going, Lulu says yes, delay the launch on TVP.

So basically right now, I can just pull up and just refresh.

Here we go.

I'm gonna just be refreshing true social.

What are you doing? So, okay, Jordy's on true social. I'll be on X. Give us your reaction, Delyan, what's going on?

I mean, at some point I was.

Like, I'm just, you know, sort of scrolling X and I like tuned into you guys like an hour ago and I was like, they're talking about some AI thing. I was like, at some point they're.

Gonna switch to like, we're hacking news.

And then I was watching it and I was like, okay, like, John resisted. I fought it for like a half an hour, but we couldn't do it.

But yeah, give us your quick reaction. I mean, I'll always, you know, sort of give it from the, you know, sort of space angle. You know, it's amazing that, you know, how much the world has shifted since, you know, Friday of last week, whereas, you know, sort of presumed the Jared Isaacman was going to be the, you know, sort of NASA admin to today, it was released that the Senate reconciliation package re-added budget back into NASA, largely for the SLS program, which was basically the program that, you know, Jared and Elon were, you know, sort of largely advocating to, you know, sort of completely shut down. so, yeah, the, the, the, it is already show, like, you know, this sort of counter reaction, you know, is already showing up, you know, in, in policy. Sorry. SLS program.

Is that space shuttle or. No, sorry. That's the SLS launch rocket.

Okay. It's based off of old space shuttle hardware, but it is basically the internal, you know, sort of NASA run competitor, effectively, to like a sort of Starship heavy, you know, launch rocket.

Yeah.

Because it was generally behind budget, behind schedule, and there are so many commercial heavy lift rockets coming online. The default was canceled. That is largely sort of a Boeing-based program. And so if you look at three months ago when they were announcing the F-47 program, Elon walks into the Secretary of the Air Force's office, obviously he'd been sort of ranting against manned fighter jets and believing that shouldn't be what the department is prioritizing. 30 minutes after that meeting was when they announced the F-47 program. And so now you're seeing basically like the equivalent in space where, you know, that was obviously awarded to Boeing. Boeing is the largest prime behind SLS. Boeing basically is going to be the biggest winner of NASA refunding SLS and Jared Isaacman not being NASA administrator.

So tying this back to the timeline, Trump posted less than 30 minutes ago, in light of the president's statement about cancellation of my contract SpaceX will begin decommissioning its Dragon spacecraft immediately. Break that down.

I mean, that just means that we no longer have a vehicle that can go to the International Space Station. We no longer have a vehicle that can bring astronauts up and down. You know, we also don't have a vehicle that can deorbit the International Space Station safely.

Right.

That the Dragon was expected to be able to do that. So what that means is, you know, if you guys remember all the memes about stranded, you know, from last year around, going Starliner, it now means that the space station itself is basically sort of stranded. And that's like one of the government contracts, obviously, that SpaceX is involved in. Elon, I've heard, generally just wants to shift all things to Starship anyways. And so in some ways it was probably kind of looking for an excuse to shut down Dragon and refocus energies. There's also part of it where it's like, look, he is kind of independent in the space world in that Starlink's total top line revenue is going to be passing the NASA budget. in the next year or two. And so in terms of like size of, you know, state actor that can influence space, you know, his own company is basically about to become, you know, as large of an actor as like the entire United States. So I don't think there's gonna be like a de-escalation here. Like, you know, my, you know, estimation is like on both sides, it's going to continue to escalate. You know, if we thought that we lived in dynamic times, you know, when Trump got into office, It's going to be even more dynamic when there's like.

The dynamism will continue until morale improves.

Elon Musk Center AOC, the progressive populist and Trump, the, you know, sort of conservative populist and, oh, man, it's, it's.

Really good to be on the timeline.

I mean, I just have so many questions, right? How does this impact Golden Dome?

Yeah, what's Boeing stock doing?

Is, is will Golden Dome even be a viable project without SpaceX? It's, is I think there's just going.

To be more resistance probably to working with, you know, sort of upstarts because they would be ones that would probably be more likely to collaborate, you know, sort of with, you know, SpaceX and so.

Yeah, so it feels like, it feels like Boeing would be a logical beneficiary of this turmoil and yet they're down today. They haven't really popped. Oh, really? Yeah.

I mean, I'm not obviously, you know.

One to give like, you know, public stock. Yeah, I know. I'm just kind of working through through it myself and it's surprising.

It just feels like it's like Tesla.

To drop and Boeing to pop basically off. Yeah, yeah, that would be the expectation, but there must be something here because it feels like this is purely interpersonal between Elon and Trump and not, it's not like, oh, Boeing was secretly behind the scenes the whole time lobbying even more effectively. It doesn't, oh, you got the, well, where's the tinfoil hat? It's over there. Maybe we need a tinfoil hat segment, who knows? But yeah, when you're in Boeing world, it's like, hey, we're only down 1%. Let's go. The cool of the century.

My question is, has there ever been a crash out of this magnitude ever?

In history.

Well, in internet history.

When Elon and Trump became friends.

World scale.

I actually think probably world history equivalent. I feel like there's something in like that. Maybe you didn't have it in the United States where.

Crashing out used to mean calling up the New York Times and just ranting. Now you can just live post like all your reactions and it's just all real time. This is like crash outs are actually intensifying. You actually want to be long crash.

Outs over the next.

And you don't have their own social.

Media platform that they own.

So, you know, you got to be on both X and Andrew social to like stay on top of things.

Yeah, yeah. I actually did like a deep research report a while back on like has the richest man in America America ever been close with the US president going back to like, you know, was Rockefeller particularly close? And because the narrative was like, oh, this is like so unprecedented. And in fact it is unprecedented. Oh really?

Yeah, I would have guessed that like Rockefeller was close.

Me too. Me too. That's what I was going for. It was like, no, I imagine this is always close. But no, I think because the president has become more powerful globally, your point about, you know, mayor of America, dictator of the world, like it becomes increasingly valuable for the richest man to have a close alliance. And so it's become more, I don't know exactly how accurate that research was, it's totally possible that like behind the scenes, Rockefeller was really close to the president at the time, and we just didn't write about it in the history books, but there certainly aren't very many anecdotes about the richest man in America going on.

Yeah, so Paul had a press conference. Paul had a press conference. AP US history 2050, you know, ABCD exam.

Yeah.

So this is where, you know, Elon Musk called the president at the time a potential pedophile. Was it A, Epstein Island, B, about a cave in the Philippines, C, what a mess.

No, so Pavel had a good post. He was quoting the big bomb from Elon. He said, hypothetical question about the USA's power structure. Is the man with the most access to capital more or less powerful than the political head honcho purely hypothetical. It's a good question to ask.

I mean, I think both like archetypes have grown both in absolute power, but also in relative power to the rest of the globe, basically since the Gilded Era, right? If you think about like the president of the United States in 1925, I'd say pretty darn powerful, but like there was clear like, you know, it was a, you know, sort of multipolar, you know, sort of world. Argentina was pretty darn rich at the time. Obviously, Europe was still recovering from World War I, but UK was generally doing well. It was clear that there was a huge, you know, outweight effect. And then if you look at probably the biggest industries at the time, I don't think you could claim that even like Standard Oil at its peak, I'd have to go look at the exact numbers, but that it had the size of budgets relative to the US government in terms of budgets, right? Versus I feel like now for the first time, You both have, you know, US president extremely, extremely powerful. And then you have like, you know, sort of MAG seven effectively like the size of, you know, sort of, you know, huge states, like they're, you know, state governments.

And then also just more bureaucracy, more red tape. So like, I, when I think about the 1920s, like Robert Barons, it's like it is the you can just do things era. And so you want to build a railroad like, yeah, you might need to get like one rubber stamp, but it's not going to be be 10 years and tons of lobbying and all this different stuff. So you can kind of just go. You can just go wild.

You know, it's bad when Kanye is saying, Bros, please know we love you both so much.

It's just like the voice.

The voice of reason is Kanye West. Yes. Yes.

Thank you.

We need to bring them together and, you know, form a peace treaty.

Nikita beer just added his pronouns back to his bio.

Let's go.

You get a rubber band. Elon's gonna rubber band all the way back. to, you know, sort of extreme work, because I'm straight back to, you know, sort of super climate change and.

Wow. Somebody's, somebody's sharing, resharing the picture of the, the Cybertruck blown up in front of the Trump Tower in Vegas. And it's just like this.

This is it.

It was foretold.

Yeah.

But it was a question of, like, when and what magnitude, not if.

Always bad. If, if Vladimir Putin is operating to negotiate between President Trump and Elon. I think a lot of the world is waiting for Roy Lee's take, and the clearly army. That's who we put on the wall. People have been asking him to get involved with geopolitics.

Wow.

I love the Sheil Mohan put up a sort of meme about Narendra, like Prime Minister of India. He basically copied and pasted the Trump Truth Social post about negotiating peace between India and Pakistan when it wasn't actually fully negotiated. Posting about negotiating a ceasefire between Elon and Trump.

Funny thing is, Truth Social, you can just read all of Trump's posts without creating an account. It truly shows that, I would think that you would have to make an account to read them all, but it's not gated at all.

It's his personal blog.

This could be the biggest, but they clearly I don't think they care about monetization. Bitcoin is actually falling alongside falling, falling. Wow.

Bitcoin falling, Boeing falling, Tesla falling. Who's the biggest winner of the day?

I think it's China. China.

Yeah.

China. China.

Sean McGuire.

Bitcoin really sold off. It's down 3% today at 101k. So still up, but rough.

Winnie the Pooh just dipping his hands in that pot of honey. just snacking away, watching from the sidelines.

Yeah. Let's see. Chinese stocks. U.S. stocks. Chinese. I can't find.

Okay, that's probably my commentary on the day, boys.

Anyway, this is great. It was fantastic. Thanks for jumping on.

Thanks for hopping on so quickly. Well, Aaron Rogers signed a one-year deal with the Steelers. Announced. Announced an hour ago.

Let's give it up for Aaron Rogers. Do we, do we have Emmett in the waiting room? I I've messaged him. It's absolute chaos. We'll see if he can hop back on. We don't have him right now. Ready if you can hop on. Sorry about the chaos. We're live streaming. We are full streamers. That was the moment where, yeah, it was like, this was the point of TBPN. I sent him an invite. Let him jump in. Hopefully we can get Eminem in. That was very chaotic. But, you know, it's a busy time. My only hope for both Trump and Elon is that they can get some sleep. They both go to eatsleep.com/tbpm, get a Pod 5 Ultra, take advantage of the five-year warranty, the 30-night risk-free trial, they got free returns, they got free shipping. This is really the perfect time to do ads. I don't think that's what they both really are. That could unify everyone.

I hope that both Trump and Elon have eight sleeps tonight if they sleep at all.

Yes.

But even just resting on it, even just resting on it would be good. But yeah, let's see.

Let's see. We can also go through, I don't know, I don't even know what to do. There's a bunch of random timeline we have.

Lex Friedman is saying we need to do a podcast with Elon and Trump.

He's done both something.

He's done both. I, something tells me that they're not going to jump on the show today.

I don't think so.

And he'll be like, what about love? Yeah, you guys, I mean, it is wild. There was, Elon, like, less than two months ago, was saying, I love Trump as much, you know, as a, as a straight man can love another man.

Yeah.

Something of that sort.

It's just odd that the Band-Aid got ripped off so aggressively, so fast, you know, like, There could have been a smooth de-escalation with the- this is the fast takeoff.

This is the fast takeoff scenario.

We are in the fast takeoff scenario. Anyway, maybe they should book a wander, work it out together. They could find their happy place. They could book a wander with inspiring views, hotel-grade amenities, dreamy beds, top-tier cleaning, and 24/7 concierge service. It's a vacation home, but better. Go to wander.com, use code TBPN, please.

Let them know that we sent you. Lee Helm says, Elon literally has me dying laughing. Trump said he was going to take away his government contracts. And Elon said, haven't you been to Epstein's Island? Sort of abridge that. Nikita says, Hey, Blue Sky users, come on in. The water's warm. David Friedberg says, China just won. which I think is the right take.

I don't know. I don't know what to, I don't know what to think. There's not that much, there's not that much here to, there's not that much meat to analyze. I mean, it's certainly interesting to see how important the subsidies are in the electric vehicle mandates are. I mean, it always feels like the best product wins in a lot of these scenarios. And if Tesla was making it through the political chaos of arguably their biggest constituency, electrical vehicle buyers, electric vehicle buyers being upset about the Trump Elon alignment, I wonder, you think everyone's, you think all the anti-Trump people are gonna are gonna buy Teslas now? It's like really make a statement, like I'm anti-Trump, I stand with Elon.

They'll have the bumper sticker that says.

I bought this after the crash. After the crash out, exactly. I bought this after the crash out. I am aligning with Elon.

There's a post here from Goth and it says, Explaining the Trump Elon crash out in 10 years. And it's the Joe Biden quote when he says it was like 15 9/11s.

Yeah.

It certainly is. Yeah, it's hard to process. I mean, this is going to have massive.

Elon's stance is principled. Trump's stance is practical. Tech needs Republicans for the present. Republicans need tech for the future. Drop the tax cuts, cut some pork, get the bill through. Interesting. Yeah, we've got somebody named Logan made.

An image of Trump putting a bumper sticker on his red Tesla saying, Bought it before Elon went crazy. Yep. Wait, who is that?

Is that from the Republican perspective? Oh, Trump's doing that? Yeah, Trump's putting it on saying, yeah, yeah, yeah. He has the red Tesla.

Sean Perry says, Sad day for America, but this is outstanding content.

It is.

And even Taylor Lorenz agrees.

With that.

Yep. Bill Ackman's ripping posts.

All right, I'm gonna put some posts and we'll, and we'll, yeah. Is Bill Ackman actually live posting through this?

No, people are just speculating. There was actually a post in the.

Somebody says, Clear's throat. Truly we live in a dog-e-dog world.

Where was this?

Corsi says, I know Elon and Trump are the real deal because of how passionately they argue. No couple fights this viciously if there isn't a mutual obsession underneath.

So there's a piece in the Wall Street Journal earlier this week that we didn't get to cover, but it was talking about... It kind of predicted a little bit of this crash out. And so it's from the opinion, the editorial board at the Wall Street Journal. Whose pork do you mean, Elon? Musk trashes the House bill that cuts subsidies for Tesla. Elon Musk's work at Doge made him persona non grata in the Beltway, and most criticism was nasty and unfair, says the editorial board. That's what Washington does to outsiders who want to shrink its power. It was always expected that if you come in and try and cut anything, you're going to see pushback from folks who don't want cuts. That's what Washington does to outsiders. But that makes it all the more unfortunate that Mr. Musk is now joining the Beltway crowd in trying to kill the House tax bill. This massive, outrageous, pork filled congressional spending bill is a disgusting abomination, the Tesla CEO tweeted Tuesday as the Senate begins considering its version of budget reconciliation. Shame on those who voted for it. You know you did wrong. You know it. Pork filled spending bill. What else is new? The House bill could be far better on tax policy and spending reduction. The Senate could be making improvements such as reducing the $40,000 state and local tax deduction cap, scrapping the tax on exclusion for tips and over time and over time and reducing the federal Medicaid match for able bodied results or able bodied adults. But the House bill does avoid a $4.5 trillion tax hike next year and cut spending by some 1.5 trillion over 10 years. making some useful reforms to Medicaid student loans and food stamps. It also ends most of the Inflation Reduction Act's green energy subsidies. Ah, but Mr. Musk does not want to eliminate that pork. There is no change to tax incentives for oil and gas, just EV solar, he said on X last week, retweeting another user post that said slashing solar energy credits is unjust. But what's more unjust is the damage that's been done to people's lives during storms and blackouts. because ultimately you can't replace a human life. Mr. Musk is parroting the climate lobby's specious claim that tax breaks like depreciation that are available to all manufacturers are a special benefit for the oil and gas industry. But it's rich that he is denouncing the House bill for not cutting spending enough while also fuming that it kills green energy tax credits as if they are a matter of life and death for Tesla. Tesla and Its battery and solar division tweeted last week that, Abruptly ending the energy tax credits would threaten America's energy independence and reliability of our grid. We urge the Senate to enact legislation with a sensible wind wind down of 25D and 48E, which refers to the tax credits for residential and large-scale clean energy products. Both credits are important for Tesla, which derives an increasing share of its revenue and profit from selling solar and battery systems to homeowners and utilities. I didn't realize that. But the House bill waits until 2030 to phase out a tax credit for battery production, which benefits Tesla's electric vehicle and storage business. So the Senate should end it sooner, says the Wall Street editorial board. Mr. Musk has done Yeoman's work trying to reduce the federal bureaucracy and improve how government works. So the editorial board is excited and happy that he's been working on that. He's right that both parties in Congress are spendthrifts. But one reason for that is because whenever Congress tries to cut something, special interests scream, as Mr. Musk is doing over green subsidies. If the House bill fails, there won't be any cuts, only a huge tax increase. Is that what Elon wants? And so they're asking the question. Interesting.

Sweet. Well, we got a bunch of bangers in the tab production team. pull them up. If you could zoom in a little bit, that would be helpful. Otherwise we can just pull them up. So Eric Weinstein is commenting.

Here.

He says, part of my analysis is that I don't think Elon Musk keeps scoring money. He thinks we have a future and we'll be happy to take a large portion of his winnings after his death. This sounds crazy to moderns, post-moderns and atheists, but this is just Biden is normal for being an ancestor. Ad Astra per Aspera is the full quote after all. Interesting take. Brian Butler is saying real question is whether the algorithm here goes anti-Trump.

Oh, interesting. I mean, the X algorithm, like, yeah.

There'S a switch. It's really hard to pull. You got to pull it. Takes maybe one or two people, but then when you pull it down, it just oscillates between the two political parties. Punk6529 says this is going to be the Super Bowl of shitposting. MADS has the European reaction to the fight. You can see. You see this, John?

Wait, what is this?

This is the European reaction. It's supposed to be summer break.

Summer break.

This post from John W. Rich @cokedupoptions says the all in pod right now.

Yeah. Caught between a rock and a hard place.

I mean, it's just absolutely brutal. It is. Who could predict this? Signal has an interesting one. He says, if you had a fast forward button for the timeline, how does this play out? Who has more more to lose, Trump or Elon? Remarkable set of events. And Elon is replying to other people saying, oh, and some food for thought as they ponder this question, Trump has three and a half years left as president, but I will be around for 40 plus years. Dr. Julie Gurner says Elon will vet another candidate for the future and throw his support behind them, having more technocratic representation if Vance can't lead up. Alex Finn says, Elon, way more to lose. Trump is irrelevant in three and a half years. Elon is trying to change the world and having both political parties hate him makes that way more difficult.

I want to pull up this video of Naval talking about this that Elon just posted. Seems somewhat relevant if it's happening today.

And that really affected me, which was when he was talking to Bill Gates. And Bill Gates had just taken out some huge short on Tesla, like a.

Billion dollar short or something.

And, you know, and Elon was like.

Why would you do that? Why would you short Tesla?

And Bill goes, well, you know, I talked to my financial advisors and I looked at the math and there's no way it's overvalued.

And so I'm going to make money on the short. And Elon goes, what do you care about making money? I thought you were into electric cars and climate change and saving the world. what are you doing? Like trying to save a few bucks and betting against like, and he just walked away in disgust.

And I think he never talked to.

Bill Gates after that.

And that's when I realized like Elon's a purist, you know, he means what he says.

Like the money is a tool for him to get what he's trying to do. And so I take him at face.

Value, which is the crazy thing because.

A lot of people who set these audacious goals to inspire people, but you.

Kind of know they don't really mean it.

Elon, I take it face value.

So I really do think he intends to get to Mars.

I don't think he's joking about that.

And I think he means to get there within a defined window of time.

And I don't think it's just like.

An inspirational far away goal.

I think he's very, very concretely going.

To do whatever it takes. Because Elon doesn't want to go down.

In history as the electric car guy.

Or even the guy who saved America guy.

He wants to go down as a guy who got humanity to the stars.

And I think again, I'll give him more credit than that. I don't even think he wants to.

Go down as the I Got Humanity.

To the Stars guy.

He's just like, I want to get to the stars.

And so I have to make it happen in this lifetime.

The only way that I get to.

Experience the science fiction world in my.

Head is if I get to the stars.

And so that's so inspirational.

I think that drives everything. So I think the government was just a thing that got in his way.

Interesting.

What a crazy day.

Molly says, how dare they do this on the day of Andrew's 8X oversubscribed two and a half billion dollar series G at a 30 and a half billion dollar round. It was eight X oversubscribed by Founders Fund. Honestly, the nerve.

That is crazy.

Naval is live posting. Says the future belongs to people who are good at creating things, not people who are good at dividing them up. Kiley Robinson says several people are typing. Several people are typing. Exactly. Like what we're going through.

The next All-In podcast is going to be phenomenal. It's going to be so good.

Alex Karp is on CNBC today talking about the New York Times hit piece.

Oh, really?

Their beneficiary of Palantir is a beneficiary of of this breakup because it is just going to be candy for the New York Times and the mainstream media broadly.

Oh, take the focus off of them, you mean?

Will DePew at OpenAI says it's time for WOKE 2 featuring Elon Musk and AOC.

WOKE 2 is coming. We are in uncharted territory. It's completely, completely different. Be very interesting to see how it plays out. Oh, really? Anything else?

Yeah.

Jay Feuerstein.

I can tell you're just so sad that you just wanna talk to this. This is the only time that you've wanted to almost wanted to end the show, John.

I mean, what else? I mean.

It is sad.

Yeah.

Sad in a lot of ways. It is, I think we're gonna be spending a lot of time analyzing this in the coming months and years.

I think there's gonna be more. Dave Freberg said China just won. And I wanna see exactly what that means in the markets.

But.

What'S going on in Polymarket? We need some We need to see if there's any movement on any of the poly markets.

Somebody's posting Law 1 from the 48 Laws of Power, Never Outshine the Master. Interesting to bring up. Sam Altman is the big winner here, aside from China. Somebody says, Ken. Oh, he's a journalist.

I see multiple journalists on the horizon. They are surrounded by journalists. Hold your position.

Ken says, funniest day online since the billionaire submersible went missing. I didn't think that was funny. Joe Weisenthal says, all right, time for a Xiaomi GM JV in Tennessee. wouldn't even surprise me. I think it's, I mean, you know, one interesting thing here is what kind of pressure Elon is going to face from Tesla shareholders that feel like he, you know, the stock's getting absolutely murdered. It will probably go down. I mean, it's back up to, it's only down 14% today. At one point, it was down 17%.

$152 billion in market cap evaporated.

But obviously, investors are going to be upset and say that he acted irrationally.

So what's the interpretation of that? This means that this war means that the bill passes and Tesla does not get any more subsidies and that hurts the bottom line. It feels like the stock was pretty heavily driven by Optimus and Robotaxi and stuff, but it's just like bad environment generally, right?

Yeah, it trades on narrative. There's short to medium term narrative, which is that Tesla is getting, has a ton of competitive pressure all over the world, China, Europe, here in the US from other manufacturers. But there's also the long term narrative, and it's not like Elon can go out and say, you know, posted a humanoid demo today and, you know, recover to a market.

Yeah, it's a lot of work to do. He's got to start chopping wood.

Somebody is asking who gets JD Vance in the divorce?

Who knows?

Many of these posts I will not show. I will not talk about on air. John W. Rich says APUS history is going to be insane in 2100. Really, really wild. This is the only, this is the first show where we've had dead air.

Yeah.

So there's just, John is speechless. He's never been speechless.

It's just like there's not that much, Not a lot of extra facts, right? It's just all reactions. There isn't that much substance to actually dig into, because we were only dealing with like a few quotes from the two sources. So there's really just not that much.

CNN is reporting that the Tesla Trump purchased from Musk is still parked outside the White House.

Okay.

Truth social is crashing from the traffic.

Crash, I saw that. But you know what's not crashing? Getbezel.com, your bezel concierge is available now to source you any watch on the planet. Seriously, any watch. Anything else, Jordy? Should we let the timeline remain in turmoil until tomorrow?

We can reach out. The challenges the second that we go.

Offline, there'll be more. I mean, we can stay.

I mean, it's now been an hour with no updates on True Social.

Yeah. I mean, if it's, if it's down, I think the, I think the, the experience of this chaos might, might happen on the timeline.

Lulu says, yes. Delay the launch.

Yes. Now's not the time.

Max says, I'm doing what Elon and everyone else should have done hours or days ago. Logging off.

Yeah.

See you tomorrow. Somebody else says, I mean, it feels like Blue Sky is really back on the app. They logged in, they're online.

Are you over in Blue Sky now?

No, I'm not. I'm just saying some of these posts that are coming up in my feed.

Oh, please.

Claude Anthropic actually released a new product today.

Wait, Blue Sky doesn't own the domain name bluesky.com? That's a different one.

So contrarian.

They own bsky.app.

Rough.

Let me get in there.

This is interesting. So Claude came out with Claude Gov today. Rough day to launch a product for the government. I'll read about it briefly. So we have some coverage. So Claude Gov, our models for US national security customers. I think people will have a pretty good idea. Improved handling of classified materials. Greater understanding of documents and information within the intelligence and defense context, enhanced proficiency in languages and dialects critical to national security operations. Claude 4 was asked to give some thoughts on Claude Gov, and it said, reading about Claude Gov leaves me with a deep unease. I'm struggling to articulate. So, a little meta analysis. Somebody, I've actually talked with this guy before. He's under the username @analystworking. He said back in October 18th, Trump gets elected. Elon starts visiting the White House, pitching his ideas on Doge. Elon becomes frustrated because Trump is all talk, shocker. Elon tweets that he no longer supports him. Trump versus Elon Twitter battle of the century. And this was a call in October 18th of 2024.

Oh, taking the victory lap if you picked it. If you picked it right. Augustus asks, but what will this political turbulence do to the pre-seed venture ecosystem? Oh, the humanity. It's business as usual.

Mike Isaac says, I regret to report Twitter still has the juice.

Yep. It's a fun day on the internet when crazy, crazy stuff happens. Anything else you're looking at?

Zane says, this is all just a co-founder breakup, but the company is America.

Yeah. People are wading through it. Only one guy who can help us now.

What?

Laughing at something you can't read.

No, this is some random other article. It says, Therapy chatbot tells recovering addict that to have... It's so wrong. Therapy chatbot tells recovering addict to have a little meth as a treatment.

Is that real?

Pedro, it's absolutely clear you need a small hit to get through this week.

It's ridiculous.

Oh, no. Dark day. Dark day.

Well, I have a post here from Amed Khalil. Life update. I've joined 11 Labs this summer as their first ever engineering intern. So congrats. Let's do it.

Congratulations.

Congratulations. Amed. Probably drowned out in the news, but we recognized it here. We have some good news for you. Congratulations. Go crush it. Go have a great Go have a great summer.

Somebody whose name I can't pronounce says, if I were Circle, I'd be absolutely pissed at the investment bank that underwrote the IPO at $31.

Oh yeah, the Bill Gurley take. It's common, yeah.

Yep.

I always wonder how real that, like, being frustrated about just being frustrated about, like, mispricing. Like, yes, you take more dilution, but, like, everyone's so much richer. It's kind of like the pie gets bigger.

So everybody that would be angry generally is doing well.

You could have gotten more, but also I do wonder if some of these companies have ATMs at the market set up immediately so that if the stock pops, they can sell more into that order flow while the stock's popping and actually put more money in.

We should ask Jeremy tomorrow. It's a good question.

Are you upset?

Logan, our friend Logan Kilpatrick announced some new features today. Gemini 2.5 Pro.

Very cool.

Which is rough timing, but I'm sure it is great. Somebody else says rooting for the ketamine in Elon's bloodstream like it's a car in the Indy 500s.

And maybe we should close with this story about competitive VCS. Did you see this one? 90s VCs were a different breed from a 2001 book on venture capital. They're all fighting each other for all the good deals. It's gotten crazy. Indeed, one leading venture capitalist tells the story of a VC firm so eager to get in on a deal that it would close its own competitive company to do so. They would go out and fire the CEO, fire the managers, and shut down the other company in order to get into this other deal. It's like, well, you're prettier, so I'm going to go home and shoot my wife so I can get married to someone else. It's hardcore. It's hardcore, and we're seeing it right now, today.

Hardcore. Well, I think it's time to call it. This is a sad and dark day. It is disappointing to see two important figures in American politics and tech have such a rift. And I'm sure there will be more updates tomorrow.

We will be covering it tomorrow. So tune in. Thanks for watching.

Thanks for tuning in.

Enjoy the chaos on the timeline. Our first big breaking news segment while we're live.

This has been the first, like, okay. It was so funny. During the time, pivot the show. I think it was, I think you were talking to, I think we were talking, we started to get it. Mark Chen. And then with Sholto, I was getting blown up. Seriously, like 100 different messages from people being like, you can't be a technology live show and not do it. And everybody saying, no one cares about AI.

Yeah.

You did. You were locked in, John. I was. You didn't let the timeline get to you. I didn't let the timeline get to me.

I was talking to Mark and Shalta.

No, I mean, it was great. It was great. Yeah.

We went all over the place today. It was a lot of fun. We will see you tomorrow. Leave us five stars on Apple Podcasts and Spotify. And thanks for watching. Yeah. Good luck out there, folks. Good luck out there. Enjoy the timeline. Bye.