Gemini 3 Reactions, Cloudflare Outage, The Upsides of Bubbles | Byrne Hobart, Glenn Hutchins, Yogi Goel, Sam Jones, Ali Madani, Amit Jain

Summary

TBPN Podcast Summary

Episode Title: Gemini 3 Reactions, Cloudflare Outage, The Upsides of Bubbles | Byrne Hobart, Glenn Hutchins, Yogi Goel, Sam Jones, Ali Madani, Amit Jain
Date: November 19, 2025
Hosts: John Coogan & Jordi Hays
Guests: Byrne Hobart, Glenn Hutchins, Yogi Goel, Sam Jones, Ali Madani, Amit Jain

---

Overview

This episode of TBPN dives deep into the current state of artificial intelligence, infrastructure, and the financial "bubble" mindset powering the rapid expansion in tech. The hosts and their guests break down the ramifications of Google’s Gemini 3 launch, the Cloudflare outage, how financial bubbles contribute to progress, and announce major news in AI, robotics, cybersecurity, and biotech. The episode features live reactions, sharp analysis, notable stories from industry veterans, and wrestling with big questions: What does the future look like for chips, power, labor, and capitalism in the age of AI?

---

Key Discussion Points & Insights

---

1. The Centrality and Challenges of iMessage in the Age of AI

[01:06–07:28]

• Main Topic: The hosts reflect on how Apple's iMessage has become an “ERP system” for people's personal and work lives. John argues that because of the lack of interoperability, iMessage is a walled garden that AI assistants like Gemini 3 will struggle to integrate with.
• Power Users & AI: Jordi wonders if Apple will ever let users get more value by allowing AI (like Gemini) access iMessage data. Skepticism prevails about any upcoming portability due to Apple’s payment structure with Google and its integration incentives.
• Notable Quote:
  • John: “iMessage has really, really grown to the point where it’s not just like one-on-one text messages, it’s all these group chats, it’s sharing of locations and documents, files, all this stuff.” [03:10]
• Funny Take: John and Jordi joke about whether you'd ever want your AI to draft responses for dozens of unanswered texts, or if a humanoid robot could just “pick up your phone, unlock it, and answer everything.”

---

2. Google Gemini 3 and Model Benchmarks

[11:00–18:56]

• GeoGuessr Benchmark: Gemini 3 Pro beats professional human players at GeoGuessr, surprising the crew but raising questions about overfitting (since the data is all from Google Maps).
• Model Tier Naming: A lively debate breaks out about model naming conventions (e.g., “5 Pro”, “3 Pro”, “DeepThink”), how confusing it is, and how this confusion impacts both power users and the average consumer.
• Notable Quote:
  • “It feels like most of the labs are coming out with three variations on speed right now…I wonder if Gemini will do a model switcher at some point.” – John [15:09]

---

3. Google Antigravity, IP Forking, and the Challenge of Platform Churn

[17:50–20:00]

• IP Forking Drama: Google’s new “Antigravity” agent platform appears to have been built directly on Windsurf/Cascade’s code base (from Cognition), sparking community jokes about missed branding cleanups.
• Industry Practice: The hosts connect this to Twitter’s lingering old branding and the broader challenge of major companies rapidly rebranding and “forking” their own and others’ code.

---

4. TPUs, Nvidia, and the Structure of AI Infrastructure

[20:00–23:52]

• Exclusive Compute Worlds: The Gemini 3 model was trained on Google’s TPUs, not Nvidia chips. This, paired with Nvidia’s earnings, raises questions about true AI chip competition.
• Infrastructure Duality: While TPUs offer a counterweight to Nvidia, the closed nature means other players are left with fewer choices—fueling major industry efforts to build custom chips or tap AMD.
• Demis Hassabis Interview: Demis (Google DeepMind) is focused on “world models,” but points out there’s never enough compute, and balancing business/research returns is an ongoing challenge.

---

5. The Upsides of Bubbles (with Byrne Hobart)

[76:20–109:43]

[Interview spans 90:20–109:43; pre-discussion at 76:20]

• Pro-Bubble Thesis: Byrne Hobart, author of "Boom: Bubbles and the End of Stagnation," argues that bubbles “coordinate different market participants”—they force overbuilding and propel innovation, even if some investors lose out.
• Rich Lose, Public Wins: Byrne points out that in tech bubbles, the rich typically lose more (via concentrated risky bets) while the public benefits from all the cheap, overbuilt infrastructure that’s left (e.g., railways, fiber, GPUs).
• Complementary Innovation: True tech progress relies on “complementary assets” all being built at once—a bubble signals to companies from power generation to chip fabs to models to build for a future that isn’t yet here.
• Historical Perspective: Even after the dot-com crash, the underlying value (like the Nasdaq 100 and housing prices) was higher than at the start of the boom.
• Notable Quote:
  • “There hasn’t been a technological revolution in history that didn’t, at some point, get overhyped. That’s always obvious in retrospect, but less so when we are in the cycle.” – Byrne Hobart [81:20]
• AI Labor Displacement Timelines: Byrne is skeptical that labor-displacement will happen as instantly as some claim; organizations have a harder time adjusting blame/risk models when AI augments—or replaces—jobs. Early adoption will likely manifest in new, AI-native businesses serving smaller, less risk-averse clients and then expanding up-market.
• On Kids Growing Up With AI: Byrne shares personal stories of his children using ChatGPT to settle disputes, cautioning about how using AI as a crutch may hinder the traditional benefits of writing and critical thought.
• Notable Quote:
  • “There’s going to be a cognitive overclass and a cognitive underclass—and you have to decide if you want to be someone who thinks, because you like thinking, or just have a more relaxing time.” [108:33]

---

6. Glenn Hutchins: Financial Evolution, Debt, and the New AI Buildout

[127:12–154:50]

• PE Innovation: Glenn, co-founder of Silver Lake, walks through the evolution of private equity, from capital asset pricing models to Black-Scholes, and how software businesses (like Microsoft) fundamentally shifted finance due to high returns on IP.
• What’s Different About Today: In the current AI era, the scale of required CapEx (for AI datacenters/factories) is unprecedented. Drawing a parallel to Taiwan’s backing of TSMC, Glenn argues that some of today’s US AI infrastructure is only buildable thanks to enormous sovereign or corporate counterparty contracts (e.g., Microsoft).
• Bubble Reflection: Glenn sees today’s investments as more like the Internet buildout, rather than subprime or 19th-century railroads—there is real value and “offtake” behind the infrastructure, even if some ventures will inevitably fail.
• On Debt Deals: He explains how today’s four-to-five-year contracts in the AI space offer a “2x multiple” on invested dollars, and that the existence of solvent off-take agreements (with, e.g., Microsoft) avoids the speculative mistakes of the fiber optic and railroad booms.
• Memorable Moment:
  • Glenn flexing a t-shirt with "70" written in binary for his birthday [152:45]
  • “If an idea is hated intensely, I know it could be really, really good...and when it turns out to be generally accepted, that’s when I sell to them the beachfront property I already purchased.” [132:43]

---

7. Major Announcements: Fundings, Startups & Breakthroughs

[155:34–183:55+ – Lightning Round of Guests]

• Yogi Goel (Maxima) [155:34]
  • $41M Series A to build AI for enterprise accounting, reducing errors and inefficiencies, not aiming to replace humans outright.
• Sam Jones (Method Security) [160:52]
  • $26M seed + Series A from Andreessen and General Catalyst for a cyber resilience platform focused on autonomous, dual-use cyber operations for gov and enterprise.
  • “Our adversaries have better models at home…This is why it’s urgent we focus on resilience.” [163:16]
• Ali Madani (Profluent) [168:09]
  • $106M round (Bezos, Gerstner on cap table) to “make biology programmable,” designing novel proteins via AI, bridging machine learning and wet lab feedback cycles.
  • “We’re moving away from random discovery and relying on nature, to using AI to design bespoke medicines from scratch.” [171:45]
• Amit Jain (Luma AI) [176:05]
  • $900M Series C; also—jointly building a 2 gigawatt compute cluster with Humane in Saudi Arabia, targeting AGI via massive multimodal model training.
  • “Inference, especially for video, will become much larger than training…Generative models give you simulation capability. Simulation is extremely important.” [179:47]
  • Cluster buildout to finish by end of 2027 or early 2028.

---

8. Cloudflare Outage and Platform Fragility

[64:21–67:37]

• Cloudflare’s major outage led to outages of everyday apps/sites; hosts highlight both the existential risk (and inadvertent advertising value) of all cloud infrastructure, and praise Cloudflare for rapid, transparent communication.
• Key quote:
  • CTO Dane Knecht: “I won’t mint words…we failed our customers. The trust our customers place in us is what we value most.” [67:37]

---

9. Additional Colorful Segments

• Nvidia Earnings: Speculation abounds on how Nvidia’s performance telegraphs real AI demand. The hosts joke that even beating earnings might not stop a selloff, given market expectations. [32:00–34:00, 184:03]
• Water Usage in AI: Satirical critique of panic over AI’s environmental/water impact, calling out factual mistakes in "Empire of AI." [53:43–59:50]
• Roomba Robots/R2D2: In a lighter segment, the hosts debate the merits of various robot form factors, and recount a hilarious story about nearly leasing an office with a mysterious “soil cleaning” death machine in a closet. [39:43–52:00]
• Press Release Economy: Observing the event-driven nature of AI and investing, where press releases and partnerships are now as meaningful as actual deals. [62:33–63:55]

---

Notable Quotes & Memorable Moments

• Byrne Hobart:

  • “There hasn’t been a technological revolution in history that didn’t, at some point, get overhyped.” [81:20]
  • “It is so easy to go through life without thinking and it will only get easier. And so you have to decide: do you want to be the kind of person who thinks because you like thinking…or have a more relaxing time?” [108:33]

• John Coogan, on Bubbles:

  • “When the bubble pops, rich people actually get hurt more than Main Street…most people have a diversified asset base; those in the riskiest bets are the ones that get wiped out.” [78:15]

• Glenn Hutchins:

  • “My best idea is the one people dislike. My very best ideas, they hate intensely…by the time it turns out to be generally accepted, that’s when I sell.” [132:43]
  • “Today’s AI buildout is different—every one of these data centers has a solvent counterparty contracted to take all the output. These are built to suit, not ‘if you build it, they will come.’” [145:40]

• Sam Jones (Method Security):

  • “Our adversaries have better models at home…and that’s why we have a national cyber resilience urgency moment on our hands.” [163:16]

---

Timestamps for Major Segments

• iMessage as a personal ERP, and AI integration: 01:06–07:28
• Gemini 3/Benchmarks/Model Discussion: 11:00–18:56
• Google Antigravity/Windsurf Drama: 17:50–20:00
• TPUs, Nvidia, Compute Monopolies: 20:00–23:52
• Byrne Hobart Interview (Bubbles): 90:20–109:43 (discussion from 76:20)
• Glenn Hutchins Interview: 127:12–154:50
• Lightning Round/Startup Announcements: 155:34–183:55
• Cloudflare Outage Recap: 64:21–67:37
• Misc/Comedy/Press Release Economy: Throughout, eg. 31:37, 53:43, 62:33, 39:43

---

Tone and Style

The episode features the hosts’ trademark irreverent, humorous style—blending serious, highly technical analysis with plenty of banter, in-jokes, and “live” reactions to breaking tech news and market moves. The tone remains skeptical yet optimistic, grounded in sharp insight and industry experience.

---

Utility for Non-Listeners

This summary covers:

• The context behind buzzworthy events (Gemini launch, Cloudflare down, Nvidia earnings, major VC rounds)
• Intellectual frameworks for understanding bubbles, risk, and tech infrastructure
• The perspective of iconic investors and operators on what’s changed (and what hasn’t) in the current tech cycle.
• Live reactions and expert analysis from those operating at the center of AI, biotech, security, and finance.

Whether you’re a tech insider or watching from a distance, this episode provides both perspective and practical commentary on where the industry is—and where it’s likely headed next.

Transcript

[A]

You're watching TVPN.

[B]

Today is Wednesday, November 19, 2025. We are live from the TVPN ultradome, the temple of technology, the fortress of finance, the capital of capital. Ramp.com Time is money. Save, save. Both easy use corporate cards, bill payments, accounting and a whole lot more all in one place. Thank you to the good folks over in Australia, Ben Sand. Ben Sands from Strong Compute sent a whole crate of violet crumble. John know this is my favorite piece of candy in the world it comes from. It's their greatest export. It's why we need to defend them at all costs. It's why they belong in Aukus. It's the backbone of geopolitical protection in the Pacific. So Strong Compute for tbpn. Visualize every data center announcement interactive in real time for GPU cluster users see and control GPUs in all clouds. Ben sand sends this from there. Says visualize any cluster. Thank you to the team for sending.

[A]

Very thoughtful.

[B]

Enough violet crumble for a lifetime. What a crazy.

[A]

What do we got today, John? What's your take?

[B]

My take is do you want iMessage in Gemini 3? Do you want iMessage in your AI assistant in your personal superintelligence? After Meta Connect, we left saying, wow, the virtual reality, the Call of Duty heads up display is here. It's arrived. The Meta Ray Ban display. And the technology was really cool. Like the glasses didn't look that crazy. And the heads up display, like the actual HUD was really high quality. Like you could actually read what was going on there. But where we left it was, wow. If it doesn't work with imessage, I can't imagine wearing that because my whole life is imessage. And I was just kind of reflecting on this idea that like imessage has kind of emerged as my personal ERP system. Remember when VCs used to be like, oh, we need a personal CRM? And it was like you're. You've just turned every one of your personal relationships into a business relationship and now you should be using an actual CRM. And many VCs do use actual CRMs. Even if it's like catching up with coffee with a buddy from their MBA program or whatever. Like, people will track that because it makes sense. These are professional relationships, so they should be professionally managed. Maybe in a CRM like Adeo.

[A]

The AI native CRM.

[B]

The AI native CRM. Where is Adeo?

[C]

Here.

[B]

I have a new. I have a new list.

[A]

I'm getting the blood flowing this morning.

[B]

I'm glad I'm enjoying some movement but personal CRMs never took off. And I noticed that like imessage kind of become like my personal data lake, my personal ERP system. Like it's my single pane of glass. Like if it's, it's the source of truth. Yeah, it's like the system of record for my personal life and also we use it for business and stuff. I don't know how unique I am. I feel like a lot of people are stumbling into this world, sleepwalking into this world where they bought the iPhone. They were like, yeah, it's cool, it's got all these apps. Like I could switch to a different phone and like truly you can't if your whole life is an imessage because there's so many different chats, there's so many different like you know, the images and like imessage has really, really grown to the point where it's not just like one on one text messages, it's all these group chats, it's sharing of locations and, and documents, files, all this stuff.

[A]

Files that were shared, you know, PDF that was shared over a year ago.

[B]

Totally, totally. And so, and so my question is like it seems like imessage is important for the heads up displays for the, for the smart glasses. Will it be important for Gemini? And we were debating this like right now imessage when you go in there like the only AI experience you see is like those Apple intelligence summ which are sometimes very funny. I was laughing at it. Summarizing one, is it declared over? Because if someone says it's so over it will just rewrite these. It doesn't get the jokes. Other times it'll just say PNG image shared and sometimes those funny. Sometimes it's a little bit useful. But in general I think that all the Apple intelligence features will get better. With Gemini 3 we saw on the benchmarks we demoed the product. Gemini 3 is definitely a great model, the best model potentially right now. Apple will be able to implement that all over the place and they just won't have to worry about like do we have a good foundation model to build on so they'll be able to stuff it everywhere. But what does the actual flowback look like? Because Google and Apple are famously like walled gardens. Like they can't really just interface with them.

[A]

Some of the best walled gardens of all time.

[B]

Some of the best walled gardens of all time. And I was wondering about if you. If I'm, if I'm using. So the average consumer will just see Apple intelligence and they'll really just see Siri and they'll be like when I ask Siri the history of the Roman Empire, it does a great job giving me the history of the Roman Empire. It doesn't necessarily get confused and hallucinate because it's using Gemini 3 under the hood. But the consumers I don't think will expect if they wander over to Gemini 3 hosted on Google Cloud Platform or Google AI Studio, go to AI, go to Gemini 3 Pro, Google's most intelligent model with state of the art reasoning, next level of coding and deep multimodal reasoning. AI Studio build. That's the URL. The. I think, I think people won't necessarily expect that if they're interfacing with Gemini over in Gemini World, in the Gemini app or in Gmail, they won't expect it to connect to their imessage even though it's the same model that that's powering both of those. And Apple will say that that's for privacy reasons and consumers won't know to ask. But I'm kind of curious about that because that would be an interesting feature and I don't know if you would even want that. Like would you want to be able to go to the Gemini app and have it be able to pull a file that was shared with you in an imessage group chat and then do something with that in the Gemini app? Is that a feature?

[A]

The only thing that I can think is I feel like my entire life runs on imessage and it doesn't feel like Apple is super motivated actually building for power users. And so if there was a way to get more value having that data within Gemini, right, Like hey, draft me text message responses to people that I've texted, you know, more than more than one day that I haven't responded to in the last two weeks and have draft a bunch of messages that I can then just go through and at least like look over and respond to. Yeah, but I don't know. I have zero faith that there will.

[B]

Be portability, any jumping of the wall.

[A]

And the reason for that is Apple's paying Google to white link to effectively white label the model. Leverage Gemini in the next version of Apple Intelligence and they're just going to be focused on integrating it within their ecosystem deeply. And I think if they weren't paying for it, Google would have been able to negotiate for quite a lot more and potentially more interoperability between the products.

[B]

Yeah, I feel like they're. There might be some magic that comes out of a deeper integration between these Two things, it does feel very different than Google Search because the models are actually intelligent. And could I think that the obvious draft a summary, like the example that you gave, draft a response to a text message. I don't know if anyone would even want that. And I do think that Apple Intelligence will be able just to do that out of the box. I'm imagining more of. Of like when I. When I go to an LLM to prompt it for a gift guide. If it has access passively to imessage, it can understand. Oh, like people have been sharing these links with you to things that could be gift. Here's the context around the context. Maybe they shared that link with you. Being like, lol, I would never buy this. Someone for Chris, for someone for Christmas. Or they could have been from a family member saying, you know, this has been like, I would write to Santa for this. And they're like alluding to you actually wanting to buy them for that. So, Tyler, what do you think?

[D]

I think like, when I think of like, AI in, like communications generally, I think it's more like the vision is like, let's say I'm trying to set up a meeting with Jordi. It's like I have an agent. My agent talks to Jordy's agent.

[E]

Yes.

[D]

They sort everything out if we should meet, when we should meet, where we should meet. And then it's kind of like done completely separately from like imessage even.

[B]

Yeah.

[D]

So I think that's more of like my kind of ideal vision of like, what LLMs and messaging look like, where it's basically like, I'm not even doing actual messaging. I'm not sure how important it actually is that it interfaces with imessage. I mean, obviously it's like, good to search through your messages. That's useful.

[B]

But Yeah, I just wonder. The reality of everyone's life is that they use multiple messaging systems. They use email and WhatsApp and Signal and then iMessage and Twitter DMs. And there's never been a successful unification of these. But I was laughing to myself thinking about like a humanoid robot. Because, like a humanoid robot, you could literally just like be like, here's the phone, here's the passcode. Go respond to every message on my phone. And like, it could do that and it would be impossible to like, there's no like, data wall that you can put up at that point. Really?

[A]

Yeah.

[B]

I mean, maybe if you're like world coin scanning constantly to, you know, like eyeball scanning to get into the actual. The actual app. Or something. But it reminded me of like George Hotz was saying that like at a certain point the, the, the full self driving like it's like you don't need to worry about car compatibility because it's just a humanoid that gets in the driver's seat. You want a driver. I thought that was such a funny take because it's like yeah, like right now, Toyota. I believe it's Toyota. But a few of the car makers are basically saying like no third party self driving kits. Like we are encrypting our OBD 2 ports like the actual port where the car. We're not going to let anyone build on top of us because we want to own the self driving stack on top of our vehicles. Y so no, no third party kits. And it's just very funny to imagine like well how are you going to stop a robot from just sitting in the driver's seat and shifting the gears and, and, and, and pushing the pedals anyway. Restream one live stream 30 plus destinations. If you want to multi stream go to restream.com Lisa Lisan Al Gaib has more. Gemini context said Gemini 3 Pro is the first LM to beat professional human players at GeoGuessr.

[F]

Wow.

[B]

We gotta watch. Who's that? Who's that? The amazing geoguessr guy. Does he just go by geoguessr? What's was.

[A]

Oh, I know who you're.

[B]

You know what I'm talking about. The greatest game of GeoGuessr. This is the guy, he's in the thumb.

[A]

Rainbolt.

[B]

Rainbolt, yeah. Geo Rainbolt. I want to see his reaction to that and see how he's doing.

[A]

He's just crying. Crying on stream.

[B]

He's done a few.

[A]

This is one of those things that I think is actually still gonna be wildly entertaining. Even when like chess, right? Like watching him figure out where something is down to a single street is still going to be impressive and probably entertaining.

[B]

It's a pretty cool benchmark. I'm surprised by this but what is this? Oh, so it got a higher score but lower country percentage than a professional player. That's fascinating. I wonder what that says. So it outperformed on score but it underperformed on guessing the country. And I wonder if that's something like it's using different heuristics that are like less intelligible. Because a lot of the heuristics that you'll watch the geoguessers use, the really good professionals is that they will be able to identify like this color of signpost is only used in this country. So even though it looks like it's a tropical, that helps me understand it's this country and not that country. And that might be something that Gemini 3 Pro is not picking up on, but it's still doing a better job of understanding just the references. Also, I mean, this feels like it has to be like, overfit on geoguessing because, like, didn't Google create all the geoguessr, like, data source?

[D]

Yeah, it's all just Google Maps.

[B]

It's Google Maps. And like, it has to be in the training data, like, perfectly. So even if it's like, not intelligently thinking, like, the beauty of watching someone play GeoGuessr is that they're not just doing memorization. They're not just like, oh, I know that street. I know every street because I've memorized every street. They're actually applying a whole bunch of heuristics and patterns and matching.

[D]

Yeah, that's probably true. But also I remember with the. I think it was the GPT5 release.

[B]

Yeah.

[D]

People would. Would like submit just a picture they took like on their phone of like, themselves. It's like, where am I? So that's not like actual. I mean, that's not from Google.

[C]

Yeah.

[B]

It's not overfill. It would still do like. Well, okay.

[G]

Yeah.

[D]

Yeah. Also this is.

[A]

Yeah, they nerfed that pretty quickly because there was so much. There was. It was. Could easily be abused. I remember I uploaded a picture of outside my house and I could. I could tell. I could tell by its response that it knew exactly where it was. Even though there's no street view because it's a private neighborhood. And it was basically like saying where I knew it knew exactly where we. I knew it knew, but it was just wasn't giving like, specifics, but so much like it was within like, at least like a mile.

[B]

Yeah. 2.6m says we should play a round of GeoGuessr on stream. We should. We should get. We should figure out how to actually wire up, like, games. We've done it once before and it was pretty fun.

[D]

I'm also curious where the Deep Think model ends up on this, because this is still just. This is just 3 pro deep think.

[B]

Must be do even better, right?

[A]

Yeah.

[D]

I mean, you would imagine.

[B]

Yeah. So, yeah. How would you benchmark the 3 Pro versus GPT5? Because it seems like 3 Pro is not equivalent to 5 Pro. 5 Pro is more like DeepThink.

[A]

Yeah.

[D]

If you're looking at like, price and like the.

[B]

How long.

[D]

Yeah. How long it takes to Gender and.

[B]

Got it.

[D]

Yeah.

[B]

So 3 Pro is like 5 instant. Or is it like 5 thinking it's.

[D]

5 thinking it's 5 and then 3 flash. If that comes out, that will be instant. Yeah. Like 2.5 light or flash or there's flash light.

[B]

Yeah. So, yeah. Okay.

[D]

That's more the instant model.

[B]

So, yeah, it feels like most of the labs are coming out with like three variations on speed right now. Maybe something along those lines. And then maybe a Deep Research product adds like a fourth to the end. But that's like more of a specific.

[D]

Yeah. Like Anthropic has Sonnet, Haiku and Opus. Those are like the three. And then there's like thinking on all of those. But it's kind of a similar breakdown.

[B]

Yeah. I wonder if Gemini will do a model switcher at some point. Like right now. I mean, I guess like AI Mode has some of that, but maybe they just are. They just don't have to worry about the actual GPU cost at this point. So they're not.

[D]

Needs it. He couldn't figure out how to find the thinking model.

[B]

Oh, yeah. You need the switcher. You need the switcher. It is funny that this happens all the time.

[A]

I asked the model, what model are you? And then it said that it didn't have access to Gemini 3.

[B]

Yeah. That is something that they should hard code in because it is very frustrating. It's happened a number of times where.

[A]

It just makes it feel not intelligent.

[G]

Yeah.

[B]

Where I said, okay, I want to use the latest and greatest. How do you actually do this? They should definitely make that URL or that explanation available in the prompt so that it can answer questions. You need to bake in an faq since you imagine that people will be interacting with the chat directly.

[D]

Yeah. Well, it seems like there's some difference between the naming conventions where the lab DeepMind wants to come out with. It's like a new model. So it's three. It has a number. But then on the product side, you see, it's like numbers are kind of confusing. So they want the consumer to just see, like, faster thinking.

[B]

Yep.

[D]

But then for people who, like, want to use the new model, but they're using the consumer product, it's like, pretty confusing.

[B]

Yeah. I mean, the name scheme is very funny. Right now. There's. I mean, everyone has, like, different models, fast and thinking. But then there's also, like, Deep research, which is deep. And then there's deep thinking and deep research. And that's very hard to communicate the difference between there. Unless you're following this stuff very closely. And then the create videos with vo, but then instead of create images with Nano Banana, it's the banana emoji and then just create images. And so there's like not a lot of symmetry in the way the UI is laid out because I think everyone's moving so fast in this category that it's like, just get it out ship the code word. Oh, the code word leaked. We gotta go with it. There are still people who know Strawberry in the context of OpenAI, which is like a wild thing to be at the level where like no one knows like the code word for the next iteration of the Diet Coke can or whatever. Like, I'm sure that internally there was some project for this, but like there aren't like people following the industry that closely. Maybe there are, but certainly not on the consumer side.

[A]

So yesterday Google announced Google Antigravity, their new agentic development platform. Marvin von Hagen, one of the most powerful names Intact said, which IDE did they use to build Anti gravity Windsurf or Cursor? And Silas over at Cognition said, so Google just forked the Windsurf code base and they even forgot to remove the Cascade branding in some places. Cascade is a part of Windsurf's product, which is obviously now by Cognition. This is funny that they kind of miss this and I think it's fair for the Cognition team to dunk on it. That being said, they, of course people did buy spend however many billions on acquiring the Windsurf ip. So not super surprising.

[B]

But yeah, I mean you'd think like step one is find and replace. You know, just find and replace and just like anywhere in the code base, remove the old branding and put in the new branding.

[A]

Do you have any Rune was moving.

[D]

Quick, kind of hard to do this. I remember like, I mean it should be really easy. But I remember like months after the Twitter X takeover. Yeah, you would still find on docs Twitter branding. I mean that was still like, that's true.

[A]

Months ago.

[D]

I would see Twitter branding.

[G]

That is true.

[B]

That is true.

[A]

Yeah. But less imperative to actually make those changes in my opinion. Right.

[B]

It's like also that's a living, breathing service and like that might be a little bit difficult if it's like, you know, twitter.com is baked into some DNS and if you switch it live, like you're going to have a bunch of downtime or something like that. Like, like this is a new product. Like you can, you can just like the code base is just dead. It's just sitting there like Waiting to run and then you're just about to ship it, you'd think you'd do control.

[A]

Also I think with this was in their launch AI.

[B]

If you got the best AI, you think you'd say hey, go and fix this, go make this change.

[A]

I also think this was a part of. Wasn't this a part of Google's launch? Wasn't it in the launch video? I'm pretty sure the screenshot is from the launch video.

[B]

Oh really?

[A]

Which makes it excellent.

[B]

No, no, no, no. I don't think so. I don't think this is in the launch video. The launch video is like very minimal and this is like clearly has like a streamer in the corner like looking at it. But anyway, whether you are excited and bullish or bearish on Google because of this, head over to public.com investing for those that take it seriously. They got multi asset investing. They're trusted. Kyle Chan says this is the big story here. Google trained Gemini 3 Pro on Google's own TPUs. No mention of Nvidia chips. This is pretty crazy. I mean they've been doing this for a while but Nvidia's announcing earnings today and it's pretty crazy that the biggest store in AI is not really relevant to the biggest company in AI.

[A]

Best model ever created from a benchmark standpoint. Didn't use Nvidia chips which are supposed to be a monopoly. Right?

[B]

Yeah.

[A]

And so yeah, I don't know. This doesn't feel fully priced in yet to either company. But then again, right. It's so hard to predict demand over the next five, 10 years that maybe it doesn't even matter.

[B]

Yeah, I wonder how much because if TPUs are not for sale, Nvidia does have a monopoly like you can. If there's a monopoly on. If Nvidia truly is the only seller in the market because Google is not a seller, then yes, they still extract monopoly power from every other buyer because every other buyer says yeah, I'd love to buy tpus but I can't. So you're the only game in town still. But it's a very weird dynamic where you do have two very clearly performant products that are not actually driving down cost. It must be very frustrating if you're somebody else. But that's why every other all the other labs are working so hard to develop their own chips or bring AMD online. There's a whole bunch of different efforts in this.

[A]

Do you know the background here from this post? It's extremely Google that a flagship consumer product is Named as a reference to inner org drama that happened three years ago.

[B]

Well, there's lots of people saying that. They require context. Let's see if anyone.

[A]

Anti gravity.

[B]

No.

[G]

Oh, oh.

[B]

Is anti gravity the reference the zodiac Gemini refers to twins. Google's Gemini is a reference to two formerly distinct labs, Google Brain and DeepMind that were merged into one lab. Google DeepMind.

[A]

There we go.

[B]

I think that's it. Yeah. And I guess the interorg drama that happened three years ago was just this idea of DeepMind was acquired in. But Google Brain was still running. Isn't this reference to Gemini as in the constellation of the Gemini twins, referring to the consolidation of twin organizations. I like that. That's actually a pretty good name. This original post makes it sound much more dramatic like inter org drama. But in fact it's sort of a way to keep. Keep the lore going basically. Anyway, let me tell you about adquick.com out of home advertising made easy and measurable. Say goodbye to headaches of out of home advertising. Alex. He has a Q and A with demis over at DeepMind. He says world models.

[A]

Alex is on a tear.

[B]

He's on a tear. Sources dominant.

[A]

He's averaging like the challenge when writers go, when journalists go independent. Is actually figuring out a way to get enough scoops to justify a subscription business model as a standalone company. Alex has been. It's basically been at least a scoop a week or like very interesting content.

[B]

Big stuff too.

[A]

Yeah.

[B]

And I don't know, just like it's interesting seeing like there's a bunch of. There's a bunch of interesting things. I mean he did that interview with Mark Zuckerberg that was like seated hour long, you know, in depth interview. There were some big scoops that came out of that. Some funny takes about the. About the bubble. Basically. I think Mark was saying like yeah, we might overspend and that was sort of a viral moment. And then doing some Q and A's also it feels like maybe great timing to just. You look at the stats on this. 57,000views 477 likes Link in the core image. I mean I get it. It's like it's an important scoop, it's an important story. But it feels like a year ago this post would have been buried by the X algorithm. And so people show it to three people. Like really, really great timing on that as well. So just catching different opportunities and capitalizing constantly. So very, very exciting. But, but the actual quote that Alex Heath is sharing from his piece in Sources News, which you can go subscribe to is he says world models are the thing I'm spending most of my research time on. I'd love more TPUs. You look at seed rounds with just nothing. Being tens of billions of dollars is not quite logical to me. Taking shots, shots fired. Do we have the gun? Do we have the gun?

[A]

No, we removed that.

[B]

Oh, we removed it. Okay.

[A]

Might have to add it back.

[B]

I like that we need a taking shots.

[A]

One jumping in. Just a note on tp. Alex says when you talk about the constraints, Google has more computing access with TPUs than most companies. I would think that Google could just go all in on your team's work. But Google also gives TPU access to other startups and even rival AI labs. Do you ever just go give me all the TPUs? And Demis says I love more. But there are business requirements to balance. There's short term and long term revenue and all of these things need to be balanced and smoothed out. It's a huge advantage. We have TPUs in our own stack and we co design the TPUs with the TPU team based on where we know we're going software wise. But yeah, there isn't enough COMPUTE in the world as we all know. For everything that we want to do, there are always competing things. And then there's the question of what is the return on that amount of compute. It can be a research return, a new product investigation return or direct revenue. GENIE is still in the exploratory phase in terms of what we may eventually do with it. So anyways, well, if you're looking to.

[B]

Manage a bunch of TPUs. Get linear, meet the system for modern software development. Linear is a purpose built tool for planning and building products. Greg Brockman looks like he's hanging out in dc. This has to be Washington dc.

[A]

Yeah. This was last night.

[B]

This was last night. He says the future is bright and he's pictured with David Sacks, his wife, of course. Elon Musk Jensen Huang. What a fantastic photo.

[A]

Tyler did a green line analysis on this.

[B]

I think Tyler was getting a little wild with this one. It's barely, barely a photo.

[D]

He's not beating the.

[B]

I don't know, I think he's standing up pretty straight. He's a little leaned over, not full.

[A]

It, but we got to pull.

[B]

Elon is the. Is standing up extremely straight with some wild shoes. People are saying what are those shoes for?

[A]

It's Nvidia earnings day. You got to look for any signal they can possibly get. Yes, pull it up. It's at the bottom of the Timeline squad. Here we go. Here we go. So everyone is pretty dirty.

[D]

That is a very dirty line.

[B]

You did him dirty with that line.

[D]

I drew the line perfectly.

[B]

No, no, no.

[A]

I'm with Tyler. He's. It's all about center of gravity. Right? Center of gravity. Anyways, I think Jensen will put on a show later, right after the show ends, so we will look forward to finding out more.

[B]

Okay. Well, Elon was pictured wearing some very, very crazy shoes. These are his SpaceX shoes. I don't know who made these. Look at these.

[A]

Whoa.

[B]

Would you rock these? Could you pull these off?

[A]

Okay.

[B]

The team likes them. I'm not. I'm not. I don't think.

[A]

Are these. Are these, like, were they made in collaboration with some. Another brand or these just.

[B]

I don't know.

[A]

Is he vertically integrating Drip?

[B]

I don't know. Yeah. Did he make his own shoes? I have no idea.

[A]

These would. These would go for. I have a feeling they'd go for quite a lot.

[B]

Yeah. They seem pretty cool. Let me tell you about Fall to build and deploy AI video and image models trusted by millions to power generative media at scale.

[A]

Bobby says they look like Yeezys.

[B]

They do look like Yeezys. That's right. Quarter. The quarter app has dropped the. Oh, I gotta follow them. Why am I following them? Quarter app has dropped a. An announcement that the Nvidia earnings call will be tonight at 5:00pm Eastern Time. As soon as we log off this stream, you can head over to quarter and start streaming the Nvidia earnings call. Jensen is there. Pictured, all eyes on Huang. And they've done a fantastic job developing this image style. I feel like it's been 2025. The meta on X has been exploding in terms of, like, image macros. We've had a ton of fun with the trading cards. These have done extremely well. Anything where you can bring design and just tell a little bit more of a story, give a little bit more context, texture, something breaks through. And every time they post One of these 3,000 likes, people love them.

[A]

Scroll down if you can, because somebody ran this graphic through midjourney, and it's pretty crazy.

[B]

So bad by comparison.

[A]

I mean, it still goes pretty hard.

[B]

Yeah.

[A]

The arm there is looking a little.

[B]

It's also an interesting testament to, like, I know that the corridor designers use midjourney, they use AI, but they are really, really deep in the SREFs. They obviously have a whole bunch of different stylized prompts. And then it seems like they're also doing a ton of work in post processing layering text on top of it. They've created like a visual style that's distinct. I'm sure people will copy it, but it's definitely create. Created its own sort of style and broken through. And at the same time, like it doesn't feel like, yes, there's AI involved, but it doesn't feel like if you just threw, you know, this prompt at a random person, okay, hey, go make one for Coca Cola next week. I don't know that they could necessarily pull it off even if they had a mid journey subscription. There's still a lot of inspiration to understand what is the texture, what is the style?

[A]

Right now on Polymarket, will Nvidia beat quarterly earnings is sitting at an 87% chance. The real question is how will it trade after the fact?

[B]

Yes, it feels like we're in this weird market where you can beat on earnings and then sell off because nothing is ever good enough for the street right now. But we will see. Nvidia is such a big story now that just the fact that they are going to have earnings is essentially front page news, at least of the business and finance section. Nvidia and jobs data coming reports will provide key signals for investors after a market pullback. The fog masking the direction of the American economy and future of the artificial intelligence boom is starting to lift. After mounting scrutiny of stratospheric tech investments as well as a blackout of federal data during the longest government shutdown in U.S. history. Wall street awaits two that stand to reshape its outlook for the months ahead. AI poster child Nvidia is due to report earnings after the closing bell Wednesday, offering a snapshot of demand for chips that are in that are a linchpin in the tech mania that has lifted markets and helped buoy the economy. Also with the Nvidia news, it's like, how much can you actually read into AI demand based on Nvidia earnings? Because I feel like we're projecting out like these deals five years in advance. We buy the chips, then we install them. Like, are we really seeing that whole rumored decline or deceleration in ChatGPT growth? If that is real and that's happening and ChatGPT usage is starting to plateau from 800 million weekly to, hey, next year it's gonna be at 900 a billion. It's not going to be 5 billion next year if that's happening. Are we expecting that to show up in the Nvidia data this quarter? Like, probably not, right? Because like OpenAI has projected out five years of demand for GPUs. So I don't know, it seems hard to actually read into Nvidia's earnings as a, as a, as a real snapshot of demand. I mean, I guess demand for chips certainly.

[A]

Yeah.

[B]

A sell off in Nvidia has dragged down indexes with Peter Thiel's Macro hedge fund and others dumping shares. Sort of crazy that that's in the Journal.

[A]

Yeah, especially, especially when it's the equivalent of, you know, the average person in tech selling like a $10,000 position in the company. Yeah, it's like not, not like super.

[B]

Notable with no statement either. It's not like it's like oh yeah, he was also on Rogan talking trash. It's like nothing. The tremors extended beyond other AI names into crypto gold and more. Even Warren Buffett's latest big bet on Alphabet hasn't staunched the bleeding. America's richly valued stock market has retreated in similar fashion several times during its years long run up. In every instance, bargain hunters snapped up stocks, tech companies out profits and the economy kept on motoring ahead. The fact that there's. Yeah, we can move on.

[A]

Yeah. I mean the reason there's fixation. Nvidia's currently, it's like 8% of the S&P 500.

[B]

That's crazy.

[A]

So it matters more than any other. This feels like the most important earnings call of the year. Given the sell off in Neo clouds, given the just like pressure and debate around OpenAI given, given Google's investment in progress with the TPU. I mean there's so many different factors. In related news, it got announced this morning Musk's Xai Nvidia to develop a data center in Saudi Arabia. It's a 500 megawatt data center in Saudi. XAI is working with Nvidia and a Saudi Arabian partner to develop a data center in the Kingdom. Musk said Wednesday at an event with with the Crown Prince they're teaming up with Saudi Arabia's AI company Humane. That's going to be 500 megawatts or enough electricity to power several hundred thousand homes for a year. The announcement came at the US Saudi Investment Forum. Of course. The Crown Prince announced a trillion dollars of investment in the U.S. yesterday, President Trump touted Saudi's investment in the U.S. and the partnership between the two companies countries. My question is there's no information here on how this data center is going to be used. Do we expect XAI to be operating and competing as an AI cloud or is this going to be something that they want to have A local version of Grok. Right. And to me it seems much more likely that they're just going to be in the. They just want to be in the data center business.

[B]

Yeah. Oh yeah, that's a very interesting.

[A]

And to me that's always made sense because Elon is clearly fantastic at that. Pretty much best in the world. I mean he was mogging Microsoft for bragging about how many million work hours. 15 million work hours more than. And so clearly very good at large scale physical infrastructure build outs, getting access to energy, doing things on a ridiculous time horizon. And so so in order to support Xai's valuation, I could see them trying to get into that game.

[B]

Yeah. I mean there's also the possibility that if there is strong US inference demand, but latency is not an issue, it might be valuable to actually just co locate the data center next to the oil so because maybe the energy is cheaper. Midjourneys I believe been doing that for a very long time. Doing inference internationally because the data center demand during peak hours in the United States is more expensive than across the world. Let's pull up this video and while we do, let me tell you about graphite.dev. code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. Let's go to Elon Musk saying AI and humanoids will actually eliminate poverty. Eliminate.

[A]

And Tesla won't be the only one that makes them. I think Tesla will pioneer this.

[B]

But humanoid robots.

[E]

But.

[A]

But AI and humanoid robots will actually.

[B]

Eliminate poverty and Tesla won't be the.

[A]

Only one that makes them. I think Tesla will pioneer this, but.

[B]

There will be many other companies that make humanoid robots.

[A]

But there is only basically one way to make everyone wealthy and that is AI and robotics.

[G]

And we can't talk about robotics without AI features.

[B]

What do you think? All.

[A]

All problems in the world solved by one product. I love it.

[B]

I mean it's not the craziest take over a long period of time. You give everyone the ability to sort of marshal anything. I wonder if we'll redefine poverty at that point. Poverty will be not having a beachfront property, a beachfront mansion or something. Something that's truly scarce. That even in a my land thesis, even an army of humanoids can't necessarily.

[A]

Give out their own island. This is, you know, we joke about land a lot. We joke about it being the most undervalued asset by the current generation of investors. But land is the one thing that even with an army of humanoids like you can't as easily like, copy and paste, right? Like, it's just. It's just. It truly is scarce. It's not like land. It's not like land on the blockchain where people were like, no, like, you can buy this plot of land on the blockchain, and that's yours forever. And somebody's like, what if I just make another blockchain?

[B]

Exactly.

[A]

And I can also.

[B]

There will be no.

[A]

I can get from this piece of land on this blockchain to this other piece of land on this blockchain in a second.

[B]

Yeah, it's ridiculous.

[D]

If you have enough humanoid robots, though, then land is actually not that hard to. To get.

[B]

Why?

[A]

Because you're saying, like, you would just put enough dirt in the ocean and like, it's like, oh, or you go to Mars, you had this ocean, or you have a robot army.

[B]

Robot army. Oh, and then you just steal.

[D]

Then you just take.

[B]

But. But I think. I think what Elon's saying is, like, if you assume universal basic humanoid army, everyone gets 10 humanoids. And so the humanoids can cook for you, they can give you shelter, they can clothe you, they can give you health care. So you get everything that, you know, would typically be bucketed in poverty, but there's still scarce resources. There's only going to be one Mona Lisa, and so you got to fight over that.

[A]

Orange is in the chat says the humanoid form factor. Silly. Make it an R2D2. I'm actually surprised. Yeah. So one Matic. I got to give a shout out to Matic. My Matic has been running the reviews. No, no, personally. Been running it at my house daily for months now.

[B]

That's fantastic.

[A]

And it has worked flawlessly.

[B]

Yes.

[A]

Go check it out. I had, like, you know, I feel like everybody's been disappointed by, like, a vacuum robot over the years. And so I didn't have the highest expectations, even though setting it up was, like, fast and it got to work quickly. But I've been shocked at how just well it's worked and haven't had to think about it. You replace the little bag every once in a while and it's great. But R2D2 form factors. Would love to see more of that.

[B]

What do you want out of an R2D2, though? Because the Roomba form factor, the Matic form factor, where it goes and cleans, that's pretty useful. But if you like, Is it an R2D2 that it can fold your laundry, do your dishes? You need to sort of define a few different because clearly we're going to be in the age of spiky intelligence and also spiky humanoid usage. Like they're going to be good at some stuff.

[A]

And so I think in R2D2 form factor, you could reduce the number of motors that you need. Yeah, but what does it mean to carry more weight? Right, Carry.

[B]

So it's going to carry you.

[G]

Are you going to ride it?

[A]

No.

[B]

Explain what it's doing. Because in the classic Star Wars, R2D2 is like basically just like a hard drive that carries like a video. Like that's all he does the whole movie.

[A]

Yeah, but you can imagine he does a little thing. It has a number of different.

[B]

What does it have? What does R2D2 have? Have you seen the movies?

[A]

Doesn't have like a screwdriver.

[B]

Has a screwdriver that's basically a USB cable that comes out and like plugs in when it's like you could just use wireless to hack the network or whatever.

[A]

That's true. No, the other challenge with that form factor is as a screwdriver. So it's great if you have like a one story home. If you have a second floor, R2D2 is kind of cooked. RTD2 starts asking like, totally cooked. He's like, hey, we're going to need some more capex. I need some.

[B]

Can we get an elevator?

[A]

Elevator please. Elevator please.

[D]

I think if you're Robot Wars. I think the. If you want to talk about like Star wars form factors, I think the optimal is General Grievous or whatever. The one that has a bunch of arms.

[E]

Yeah.

[D]

It can walk around like a normal human, except it has more arms.

[B]

No, I completely agree with that. Bunch of arms form factor.

[A]

Yeah. Everybody wants to make a humanoid. Nobody's trying to make the General Grievous.

[B]

General Grievous is six way better lightsabers. You distilling R2D2 as it has a screwdriver, it's like the most just completely mogged. But I mean truly, other than just being like cute. R2D2 can't even speak English. Think about that. R2D2, it just goes beep boop beep boop boop beep again.

[A]

I mean the implication was like the general form factor of something that can roll around and has a lot of different capabilities built in.

[D]

It doesn't have any capability.

[B]

It has no capability. What are the capabilities?

[A]

Can we pull up General Grievous on the screen?

[B]

R2D2 hacked the death Star and saved Luke from the garbage. Disposal, yes. Two things that could have been done with wireless networking. Didn't need to plug in for that. It is a hard drive. I'm getting into battle with the chat right now. Can we tell the story of us risking our lives yesterday? We really should.

[A]

Yeah. This was truly incredible stuff. So we're looking. We're in the ultradome here for at least another year, but we're starting to think about our second the next UltraDome V2. We want to get slightly more space. There's a number of different things that we want.

[B]

There's General Grievous. That is the ideal humanoid form factor. And if you're not building that, it's a zero. What are you thinking? I think Optimus would look way better with six arms. Not scary at all. It is crazy. They do the quadrupeds, but no one's really working on the six legged, six arm. Like the really crazy creepy stuff. There's been a couple humanoid robots that look really scary where they were like, wow, let's put it up on the meat hooks. Remember that one? That was crazy. That was a wild one.

[A]

Was that the video where it started.

[B]

Going, no, no, no, that's a different one. But this was the company that was like, here's our presentation. Like, we're ready to release our humanoid. And they were like hanging it up on meat hooks. And it looks so spooky because it was using muscle fibers basically.

[A]

Nathaniel Smith is very bearish on R2D2. A cell phone can do most of what R2D2 can do.

[B]

Completely agree. R2D2. Cute for sure. Like, definitely, like fun to have around, but more of just like a toy companion. And you know, we looked at that lamp and that lamp. We were kind of like, what is that lamp? And I think there's just like, it's just delightful. Like, it's just nice to have around. It's like this Turbo Puffer here thing. Search every byte, you know, serverless vector and full text search built from first principles and Object storage. Fast 10x cheaper and extremely scalable. Like the Turbo Puffer. I mean, obviously we're sponsored by them, but this is something you might have in your house just because it's cute. People like having cute things. Many people have asked and having an R2D2 in your house would be cute until it runs into a stair and goes tumbling down and smashes into a million pieces.

[A]

Anyways, so we're looking. So we found a space that we love. It's dome. Like we're looking for a space In LA that is fit for the Ultra Dome, there's not a lot of things that qualify. And so we had looked at the space a couple times. I had seen it with Ben. John and I drove by it, and then we went back to look, do another walkthrough. You were excited, and we're getting like. I'm like, extremely.

[B]

You were pitching me.

[A]

Here's where this thing goes. Here's where this goes.

[B]

You made us. On the way to the show in the morning. You make me poke my head through the window.

[A]

I was really.

[B]

Then we go back at the keys.

[A]

We go in, really selling John on it. It's a beautiful space. It's like a few minutes from where we are now. Made a lot of sense.

[B]

Sorry, sorry, sorry. Ethan says R2D2 was the original digital guy.

[G]

True.

[B]

Yeah. Digital guy is incredible, for sure. Sorry. Anyway.

[A]

So anyways, we go for the third time to this space, and I'm just selling John on every inch of the space. I'm like, this is what we're gonna do here. This is what we're here. Here's where the truss is going to go. Here's where the production team's going to go. And we're just walking around, kind of getting a feel for it. And we're basically wrapped up. We're super excited about it. Not necessarily ready to make an offer on it, but certainly we're like, okay, this is by far the best option that we found. We've looked at a bunch of it.

[B]

Checks a bunch of the boxes.

[A]

Checks a lot of boxes. And right as we're about to leave, John looks over and there's a closet door with a key in it. And you just walk over. I just watch you walk over and open it up, and you start looking around. And first I make the joke. I'm like, oh, this is the intern closet. Cause it's like this really long, narrow hallway thing that's just like. It's like the worst room you can imagine. And so the idea of putting Tyler in it was at least entertaining. And then we're like, wait, what's that humming sound? And there's like, this box that's covered up, and it's just like this. Not super loud, but just constant humming sound. And we asked the broker, we say.

[B]

It'S super weird because it was drywall. Like, you walk into this big room, it's a big room. And then within that big room is a massive drywalled box with no entrance. No entrance to the box, but it's drywalled. Like, you don't usually see drywall inside of a room that's not. Doesn't go all the way to the ceiling. And so it was very.

[A]

And so we walk into this room.

[B]

While they were hiding something, basically, and there's no.

[A]

There's no purpose to the room.

[B]

Yeah.

[A]

Other than it just stores the box.

[B]

It stores the box.

[A]

It has no entrance.

[B]

It has no.

[A]

And it's humming.

[B]

Yes. And.

[A]

And we look around, and John's like, what's in the box? And the real estate broker says. The broker says, oh, that's just the machine that cleans the soil. And we were.

[B]

No, no, no. She said, that's just the machine. That's just the machine. And we're like, oh, like, what kind of machine? What kind of machine is in there?

[A]

And she's like, don't worry about it.

[B]

She's like, don't worry about it.

[A]

It's not a big deal.

[B]

Yeah, just like, you know, buildings have machines. Sometimes there's a machine in there. It's a.

[A]

And it's always on, but you don't. It's.

[C]

It.

[A]

We took that out of the square footage, so don't worry.

[B]

Oh, yeah, that was a wild one.

[A]

We're not billing. We wouldn't bill you for it.

[B]

And so we were like, okay, what type of machine is it?

[A]

And then she goes, it's a machine that cleans the soil.

[B]

And we're like, is this on, like, some sort of haunted burial ground or something? Like, what are we doing down there?

[A]

This is a hazardous waste site. And she goes, again, really not a big deal. I would worry about it if you were gonna buy the place, but since you're just planning to lease, don't worry about it. And then we were like, okay, the more you tell me not to worry about it, like, I kind of want to know more. So what's it cleaning up? And she's like, oh, I mean, it's 85% of the way clean. We're like, what. What's getting.

[B]

When did this process start? How long will that go? Has it been going for 100 years? Start an hour ago, and it's just gonna be 15 more minutes. Like, you gave us no context to actually project out what 85% of the way means. And finally she's like, there was a laundromat here ago. And we start piecing it together, and we kind of, like, don't want to press her on it too much, so we leave and start doing some Googling. We figure out that it's not a Super fun site. But Apparen, apparently there was a laundromat there that was using toxic chemicals that.

[A]

No, it's a machine shop.

[B]

Oh, a machine shop. Oh, that's what we figured out. So they said laundromat. And apparently laundromats can give off toxic chemicals that, if they get in the ground, can be very cancerous for a very long time. This was apparently a machine shop, like, almost 100 years ago or something. And they're working on cleaning the soil. But I still don't even understand how you clean all of the soil under a massive building without causing a collapse. Is it like a whole bunch of tunnels that are digging around?

[A]

It's a bunch of R2D2 robots?

[B]

Maybe it's a bunch of R2D2s, honestly.

[A]

Anyway, so she's still saying, yeah, I really wouldn't worry about it. It's just, like, not that big of a deal. It's just a machine. It just runs. You won't even know that it's running. We'll keep the door closed. And granted, the machine would be, like, 10ft from the set. So we'd be sitting here doing the show, and you just have the. The death machine running right there always. So anyways, it was very, very bizarre.

[B]

It was one of the funniest jump scares ever.

[A]

Very good bit. It was such a good bit, too, because I'm far more health conscious, I think, than you. And even you were thinking, there's no way we're gonna lease an ultradome that has a death machine that always needs to run.

[B]

I just. I found it so fascinating that it could sit there and clean the soil for years with a massive machine the size of a giant room. I want to learn more. I want to know what that machine is. I want to know what.

[A]

Invest in that company.

[B]

Exactly. That's what I'm doing.

[A]

We got to figure out how to make money on it.

[B]

You got to have the CEO of whoever makes that machine on the show. I want to get to the bottom of it. We need to do a deep research report. Tyler, can you fire off Gemini 3 Pro? Deep thinking max 24. 7 mode? Where it works for ages, it works for eons.

[A]

So I found two groups. CDE group. Soil washing equipment.

[B]

Okay.

[A]

Our wet processing equipment extracts maximum value from hazardous soil.

[B]

So is it just that corner that has the hazardous soil? What I want to know is, is it going under the building and then over so that underneath us over here, the machines here? Is it digging a tunnel that goes underneath the building and then Washes over here, too. Is there a network of tunnels under that building? I have to know. We have to go back. We have to lease this thing. We have to buy the building just to get to the bottom.

[A]

Just to get to the bottom of it.

[B]

I have to know. I'm ravenous for information.

[A]

Yeah. The cool thing is they use physical and chemical methods to separate heavy metal.

[B]

That's super cool. That's super cool.

[A]

That's exactly what we want.

[B]

That's exactly what we love. Well, in other news related to water.

[A]

I think I found the machine. We gotta pull it up. We can't leave people hanging. Yeah, I dropped one of the makers of these machines.

[B]

It sounds like fracking, right? Yes. If we could frack directly some natural gas out of the soil and then use it to power a natural gas turbine that we use to, you know, run the show and power us. I'm down for that. While you're looking that up, let me tell you about FIN AI, the number one AI agent for customer service. If you want your AI to handle customer support, go to FIN AI in the Water News. Okay, you want to pull up that and then we can go into the Water News. I got to talk about Andy Massley at some point this show.

[A]

I just. I want to see your reaction when you start to see the scale of this. The scale of this contraption and how it pretty much perfectly fits into the box.

[B]

Yes, yes, yes. Fracking with extra steps. Language, please. Was someone swearing? I don't know. Anyway, let's pull that up. Let me also tell you about profound. Get your brand mentioned in ChatGPT. Reach millions of consumers who are using AI to discover new products and brands. Let's see about this water story. Andy Masley is going back and forth with. What's her name, Karen. The AI and the environment. Somewhat related to our own environmental story that we could kind of go through.

[A]

How we doing, boys? There we go. Look at this, John.

[G]

Okay.

[B]

Okay.

[A]

This is from GN Separation Core equipment for contaminated soil washing. And you just look at this machine. This is pretty much exactly what would have been in the, in the, in the room soil washing.

[B]

You have to wash all the soil.

[A]

And there's a graphic. If you scroll down a little bit more.

[B]

I want to know. It's a simple process. It's 85% done. It's 85% done.

[A]

We just need to get the hazardous waste into the decanter centrifuge and then get it into the non acceptable solid second wash. Then take the acceptable solid up to the Coarse screen into the washing fine screen. And then take the washing chemical and bring it up into the washing reaction tank, put it back in the centrifuge, push it down into the soil filter, press dewatering screw, press, and then move it back up through the hazardous waste John. And you're good.

[B]

So Doug is asking if it's behind drywall. Is that because it generates fumes? We have no idea. Maybe it does generate fumes. We have.

[A]

We still don't know how they access.

[D]

It without knocking down all the drywall.

[B]

Yeah, we don't know how they access it. Is the drywall just up? And also, I really want to know, like, was there another entrance that someone could go into? Like, like, what if the machine breaks while we're there? Does someone come by and change out something? Does this machine need to be turned off at night? Does it require. Is it fully automated? Does it just run for years? Would we have never seen a technician come by? What if it gets jammed? Like, is it just the most flawlessly built machine in the world that never breaks? That seems unfathomable. All machines break. All machines need some level of attention from time to time. But maybe it's the most perfect machine possible.

[A]

And the machine is of course made by Hebei GN Solids Control company, which is a China based company.

[B]

Wow. Well, we don't know that this is the actual machine, but who knows? Anyway, let's go over into the environmental impact of artificial intelligence. There was a very funny post from Henry Thunberg who says, whoa, I had no idea. That AI uses 5,329,584 water per year. That's insane. Like, it uses just one water. Yeah, people are all over the place with the water thing. It's so interesting because no one is debating that it uses a lot of energy. Like, you could just have all the same discussions about energy. Like, we're actively burning natural gas for a lot of this AI stuff. Like, like all of the old school. Don't, don't cause global warming by burning fossil fuels. Like, all of those, all of those like, claims apply to AI today. Like, you could just make those claims, but instead everyone seems to have been like, caught up in this water. Oh, the water usage is so bad. And it's like you had right here, which was like, we're burning fossil fuels and that's bad.

[A]

Is it because water feels more scarce to people than electricity? Maybe energy in general?

[B]

It's like, it's like, if I can't drink water, I die. But if I can't access natural gas. Like I can still live maybe.

[G]

Yeah.

[A]

Or the sun beams energy on the earth daily.

[B]

Yeah. And maybe it's easier to spin move out of that being like, well, we're doing nuclear and solar tomorrow. Next year we're doing. We're doing nuclear and solar. So. So like you don't. It's not a gotcha that I'm using natural gas today because tomorrow I'm going to be using nuclear and solar. Maybe, maybe. Whereas the water issue might be like, more like it's not as concise to wrap up in a bow. But anyway, we covered this story yesterday a little bit and I wasn't able to pull up the original post. Andy Masley called me out, he put me in the truth zone. He said, jon, you follow me? How do you not know where the story broke? I broke the story. And yes, Andy Masley, you did break the story. And so we wanted to run through a little bit of this post to actually understand the claim about what he's saying went wrong here. And basically the high level is that he says this is the single most massive factual error in a major book I've ever personally noticed on my own. And I think I'm the first person to notice it. Empire of AI asserts that a data center is using 1000 times as much water as a city. In reality, it's 22% of the city's water. And so the chapter turns to Chile. We talked about this a little bit. It's a unique combination. Look at this line again. So the line says, in other words, the data center could use more than 1,000 times the amount of water consumed by the entire population of Cerrillos, that Chilean city, roughly 80,000 residents over the course of a year. How justifies this number in the notes saying, in other words, the data. The Google environmental impact report to sea stated that the data center could use 169 liters of potable water a second or 5 million. Oh, it's right there. That's the same number. 5 billion liters a year. According to the Water Service Authority in Cerilos, the municipality consumed 5 million liters in all of 2019, the year Google sought to come in. 5 billion liters a year divided by 5 million liters equals 1000. Something isn't adding up here. It doesn't make sense that you could use 1000 times the amount of water used by that city. And so Andy Masley has successfully put this book, Empire of AI in the truth Zone. We thank him for his service let's go back to the timeline. But first let me tell you about numeral.com. let numeral worry about sales tax and VAT. Numeral.com new product from Travis Kalanick. That's exciting. Big tripicnic.com request picnic. Travis says I'll come out of Twitter retirement for this one. Picnic at work lfg.

[A]

Great job with picnic is delivering lunch directly to your office floor with no fees and no tips every day from 50 plus restaurants. Sign up up your office for free.

[B]

Okay. There's only one benchmark for this stuff. We gotta look at the benchmarks. What's the max amount of protein? Is it over 200?

[A]

Are they protein maxing?

[B]

Is it over 200? Because we saw a major, major jump in the amount of protein in a bowl yesterday with sweet green sweet greens at 108. Now this is the most important benchmark in the bowl economy, which I'm a huge fan of. But are we seeing acceleration? Are we seeing a fast takeoff in the amount of protein? I want to be seeing 200 grams of protein, then 1,000, then 10,000, then 100,000. It should be 10x every year. Just 10x that. Yes, exactly.

[A]

Everyone's always talking about fast takeout, but we need to be talking about a fast takeoff.

[B]

Fast casual takeoff.

[A]

No, just a fast takeoff in protein per serving.

[B]

Yes.

[A]

Anyways, I think this is hard is this has to be built on top of cloud kitchens. Cloud kitchens. I wonder if it's a separate company or it's just a subsidiary kind of a front end for cloud kitchens. But either way, I think people just don't like paying delivery fees. And tipping too is still debated.

[B]

Yeah.

[A]

So you get, part of it is like, I feel like a lot of these things, if you just build it into the cost of the food, people feel better about it. But when you're, when people are forced to make the decision around tipping for something they want to do every single day and it's like, well, you know, maybe it's great sometimes, maybe it's not. But you're setting these things oftentimes before. So.

[B]

Yeah, a lot of the tipping stuff, it just, it needs to be like injected in the UI at the right time. And a lot of the apps don't necessarily like, like prompt for the tip at the right time. Like if you ask, if you ask for the tip before the service is rendered, it's hard to use the tip as a, as a quantitative feedback mechanism.

[A]

Exactly, exactly. So, so I will, when I order, I ordered delivery from a grocery store.

[B]

Yeah.

[A]

And I tip up front. Up front.

[B]

Do they see the tip?

[A]

And that's the other thing. I don't know. In theory, I'm like, I'm gonna tip because I want you to not throw the drinks a bunch.

[B]

Exactly, exactly. I do that.

[A]

But then, yeah, getting the fact that we've just normalized, getting an exploded bag of drinks in a bag is just funny.

[B]

Back to the press release Economy. Today's press release is out. Brookfield today announced the launch of a $100 billion global AI infrastructure program in partnership with Nvidia and the Kuwait Investment Authority. There are tons of press releases going out every single day. Danielle Tenrero says running a business is all about partnerships. It's all about announcing partnerships.

[A]

It's not even about. You don't even necessarily need to do the partnership. You just got to announce the partnership.

[B]

I mean, the partnership economy is going crazy right now. The prediction markets are obviously the. The most heinous offenders with a new partnership, like, every single day, it's hard to keep up with. We obviously are partnered with polymarket and we want to celebrate them when they do great things, but there's a lot of these things going on, and so we tend to give you a little bit of a higher level review on.

[A]

The prediction market front. Somebody just leaked a bunch of screenshots from Coinbase's coming prediction market. Everybody's getting into this game. There's different approaches. Some are like, partnering with existing prediction markets, others are building it entirely themselves. I'm more interested to see if, when Coinbase does their prediction markets product, how are they actually running it under the hood? Are they taking on the responsibility of actually managing the markets, being a market maker, what is that actually going to look like?

[B]

They should just hire a guy. On the other side of every trade, where you just go to Coinbase and you say, Look, I want 50 bucks on the Eagle. And the other guy says, like, yeah, I'll take that. I think they're going to lose. And that's how it goes. It should be a guy that you.

[A]

Call almost like a bookie.

[B]

They should acquire my bookie ag.

[A]

Maybe my bookie AG should pivot to not having any digital experience and just lean labeling. No, no, no. Lean into just the guy.

[G]

Oh, the guy.

[B]

Yes. Become a guy. Come back. Well, Cloudflare unfortunately had an outage yesterday. We were not affected, although. Tyler, do you want to take us through how we seem to be dodging exposure to Internet outages lately? What's going on?

[D]

Well, so I mean, to be clear, all my systems were fine, but I recently moved some of the kind of backend processes we use onto like a local machine.

[B]

Yeah, you went on prep.

[D]

Yeah, on Prem, and then I use Cloudflare tunnels to help do API stuff.

[G]

Okay.

[D]

So I was worried for a second that the stuff that I moved off AWS onto on Prem was actually going to go down because of a cloud.

[B]

Yeah, because AWS was down, what, two weeks ago, three weeks ago or something?

[D]

I think it was longer, maybe longer.

[B]

That one was rough. I feel like that day we actually did cancel a bunch of guests. There was a lot of stuff going on. We've had a few rough outages. But let's read a little bit of the. Of the postmortem from the Journal because it did make front page news, obviously. Sending our best to Matthew Prince over at Cloudflare and the team hoping for a swift recovery because we love the Internet and we love them. An outage that knocked swaths of the Internet offline was resolved Tuesday after drowning social media sites, disrupting retail sales, installing transportation networks. Users visiting sites including X ChatGPT, DoorDash, IKEA Metropolitan Transport Authority in New York City were met with error messages related to Cloudflare, a cloud provider used by major companies for security tools that protect from cyber attacks and traffic surges. A spokesperson spokeswoman from Cloudflare said an unusual rise in traffic to one of its services at around 6:20am Eastern time caused traffic passing through the company's network to experience errors. The bug was fully resolved by 9:30, she said in an update. For several hours Tuesday, users were unable to access sites and services from retail and social media to financial services. The outage echoes problems with aws. Cloudflare and AWS services were effectively invisible to users, but their tools underpin many people. So I don't know if there's a full, full breakdown here. Last year, a bug in a. In a tool used by cybersecurity company Crowdstrike upended computer systems across the world. There's just a lot of these going on, but I don't think we have like a full postmortem. I would love to know exactly what happened. It's always interesting we failed at Ms.

[A]

I'm interested to know what happens to the business when they have these outages because on one hand it's a great way to tell the world that the entire world runs on Cloudflare, or at least like a large amount of sort.

[B]

Of a Super bowl ad.

[A]

Yes, very much so, yeah. And then you talk about the stress from the Cloudflare team, where anybody that's built a software product has experienced the product going down and the stress around that. But it's like when your product goes down and then many of the services that people use and love across the country and the world also go down, it's even more stressful. But it also probably brings a ton of traffic to the site and people might start evaluating some features and say, hey, maybe this is a good solution. I'm going to sign up and see how they kind of react to this.

[B]

Well, I mean, the Cloudflare team reacted very well. They got a lot of praise for their response. Dane Knecht here says, or nact. He's the CTO of Cloudflare. He says, I won't mint words. Earlier today, we failed our customers and the broader Internet when a problem in Cloudflare's network impacted large amounts of traffic that rely on us. The sites, businesses and organizations that rely on Cloudflare depend on us being available. And I apologize for the impact of. We caused transparency about what happened and we plan to share breakdown with more details in a few hours. In short, a latent bug in a service underpinning our bot mitigation capability started to crash after a routine configuration change. We made that cascade into a broad degradation to our network and other services. This was not an attack. That issue impacted, caused and time to resolution is unacceptable. Work is already underway to make sure it doesn't happen again. But I know I caused real pain today. Ooh, I know I caused. I know it caused real pain today. The trust our customers place in us is what we value most. Just taking full responsibility here. Lulu says well done response and the comments reflect that. People in the comments are very happy. Mert, of course, always having fun. Okay, thanks, Dane. But have you considered that Blockchain's handling 0.000001% of your load did not go down. Very funny. There was another company has a picture here from Shogun, I think says Cloudflare's comms playbook. I ask permission to commit seppuku because they're just like fully throwing themselves down, just being like, yeah, we're 100% responsible. We won't mince words. Pretty sweet.

[A]

We should get into Mr. Hobart's piece.

[B]

We should. He's joining the show in just a few minutes.

[D]

Tyler, before I. They're breaking news. OpenAI new model.

[B]

New model.

[A]

Yeah.

[D]

GPT 5.1 Codex Max.

[A]

Okay, so they are firing back.

[B]

They're firing back.

[A]

We were debating did they have the juice?

[B]

Well, what's interesting is that Gemini 3, the one benchmark that it didn't outdo Anthropic on it was better in a lot of benchmarks but it wasn't better at Swebench. Correct, Correct. And so and so that was of course a testament to like Anthropic being really, really great at doing just something special in code. Obviously that's aligned with their mission of reaching super intelligence through self replicating code essentially. But a fascinating like you know, durability of their business that Even with this Gemini 3 thing that's so good at all these different things, Anthropic still on top in Swabench but do we know how to OpenAI is faring in this bench. Is there any reaction? What can you tell us about the latest model from OpenAI? Because we got to get to the bottom of it. While you look it up. I'm going to tell everyone about Vanta Automate compliance and security with the leading AI trust management platform. Also, Suno raised $250 million to build the future of music. I'm hit the while Tyler pulls up the reaction.

[A]

Great hit to open up the day. While Tyler gets into that, there is some breaking news. Glue has hit the public markets Christian Tech Group tests investors faith in AI deals on Wall street debut shares in a company backed by formal former intel chief Pat Gelsinger waiver after scaled back ipo I didn't realize that Glue was ipo.

[B]

No, I didn't know.

[A]

Pat's cool. Pat's company shares in a company developing AI software to connect Christian organizations across the US Wavered in its Wall street debut following a scaled back initial public offering.

[B]

That's very cool.

[A]

Glue counts Pat Gelsinger as exec chair and video rental store Blockbuster's former chief operating officer Scott Beck as chief executive rose as much as 5% after it began trading on NASDAQ on Wednesday morning, having raised 73 million from investors. The average share price pop for a U.S. iPO that has raised 25 million or more this year is nearly 25%, according to Renaissance Capital Management. Founded in 2013, Glue hopes to pull the Christian faith into the digital age by using values aligned generative AI to distribute content and sell marketing services to ministries and community outreach groups. There's an imperative to shape technology for good on its own. It isn't good or bad. The question is what it's used for, beck told the Financial Times. And anyways, so wasn't tracking this one but really enjoyed having Pat on the show a while back.

[B]

Yeah, no he was amazing. I'm very excited for him. He's just like, I don't know, just fun to talk to. Let's run through Bern Hobart's Economist piece. But first, Tyler, did you give us any?

[D]

Yes. So previously the Codex sweebench verified was 73.7. And now with, like, the highest reasoning, it's at 77.9.

[B]

Okay.

[D]

And then sonnet four.

[A]

Five is.

[D]

It's always kind of hard to tell, like, what exactly it is because people measure it differently.

[B]

Yeah. Don't they? To take some of the questions out. Sometimes.

[D]

Yeah. Sometimes they do that or. So Sonnet 4.5 is officially 77.2. So that's lower. But then with parallel test time compute, it's at 82%. So it's kind of unclear what that really means, but it's definitely better. So this is a big improvement.

[A]

Isn't parallel test time compute just a real guy who's just kind of sitting there being like, oh, actually, don't do it like that. Do it like this.

[D]

I think the main headline is that they said, yeah, tool use.

[B]

You find a dude who's a bit of a tool and you tell them, hey, I can't solve this. You got to do it human. I got to kick this one out to you. Do this arc puzzle for me.

[A]

Aj, our incredible brokers. In the chat, he's talking about the office debacle. He said, lmao. I still can't believe that happened. Maybe something landlord brokers should disclose before tour.

[B]

He's in the chat watching us.

[A]

AJ's been incredible, finding us every possible possible. Every possible space in the greater Los Angeles area for the next Ultra Dome. Highly recommend if you're in the LA office market.

[B]

Yes. I also recommend figma. Think bigger, build faster. Figma helps design and development teams build great products together. So we have our update. We will keep monitoring the GPT 5.1 Codex Pro Max.

[D]

One more story. There's an interesting headline. They said there were some tasks. They found that the model worked for more than 24 hours, which is like. That's it. You know, if you're. If you're following that. That one meter chart.

[A]

Yeah.

[D]

Where it's the time horizon.

[B]

That's super interesting.

[D]

Definitely good sign.

[B]

Okay.

[A]

Have you ever worked for 24 hours straight?

[B]

Buckle up.

[D]

But it depends on how you do.

[B]

Time to take it for a spin.

[D]

Yeah.

[A]

See that Tyler bench.

[B]

I would love to know what it actually is doing for 24 hours. I want to know the prompt and I want to know the output.

[D]

Yeah, that's what people are asking.

[B]

Yeah.

[D]

Of course, they didn't say that in.

[B]

The press release because it's like just sit there.

[D]

It was working on the easiest problem, just trying to debug it because it's so.

[B]

Yeah. Or I mean, there is a world where it's like, hey, the prompt is like, just go take a crack at Every single open GitHub issue on every repo for as long as you can and work on it. And then you're basically just wrapping another for loop around it. And it's like, is that one continuing? There's obviously a lot of.

[D]

But even if it is, that. That having a task that you can continuously work on is having a kind of plan that you can maintain and you don't get lost is still like a big improvement.

[B]

Yeah. I mean, in general, I would imagine that there's maybe some SaaS productization, but there's also just a ton of value to having agents that sort of roam through your organization continuously and clean up data or look for different errors or just do opportunistic tasks. That seems very valuable.

[A]

Trey says count for 24 hours.

[G]

Yeah.

[A]

Tile.

[B]

Count for 20.

[A]

Count for 24 hours.

[B]

Mr. Beast has literally done that count. There's a YouTube video up. It's 20. It's probably. I think it's 24.

[A]

What about. What about doing one rap of 135 every five minutes for 24 hours? I would probably get. That would probably really hard. Get absolutely brutal by. I don't know. I don't. I don't know if that would think. I think you would. I think most people. It sounds like. Doesn't.

[B]

Yeah, it sounds very easy, but I imagine it'd be very difficult. Difficult.

[A]

Anyways, we have to talk about Bern Hobart, the legend.

[B]

The king of bubbles.

[A]

The king of bubtalk.

[B]

Yeah.

[D]

He wrote the book on bubbles.

[B]

He wrote the book on bubbles. Yes. He's in the Economist.

[A]

He says, how I Learned to Love Financial Bubbles by the author of a book on bubbles.

[B]

So tech stocks have sold off this week over fears of frothiness and artificial intelligence. AI. I love that.

[A]

Economists adding that in.

[B]

Some investors were no doubt surprised by this. This.

[A]

But for the others, some reader out there is just like, thank you. I never put that together that artificial intelligence was the thing that people were talking about.

[B]

They have a very broad audience of the Economist. But I absolutely love the Economist. I've been a subscriber for probably over a decade. The signs of an AI bubble have been there for some time. Cluli whose original product was a tool for using AI to cheat during Zoom job interviews, raised $15 million and then dropped its cheat on everything tagline and pivoted to being a more benign AI meeting assistant. More serious AI labs have been able to raise 10 figure sums at 11 figure valuations, not just pre revenue but pre product. Individual researchers have reportedly been offered nine figure signing bonuses and in the past year the spending commitments made by a single company, OpenAI, total about 1.4 trillion, a sum equal to 1.2% of global economic outcomes output. A frenzy like that is enough to make you long for the relatively sane and responsible days of the pets.com sock puppet or the synthetic CDO Squared. You want to continue reading?

[A]

Yeah, but bubbles are tricky things. The default school of thought is that they're driven by irresponsible speculators who aren't trying to invest in great companies, but to buy something they can flip to someone more gullible. A more benign theory is that that there are wealth transfer from rich investors to everyday consumers. People who bought telecon firms junk bonds in the late 1990s lost their shirts, but the rest of us were blessed with bandwidth cheap enough to support the likes of YouTube and Netflix.

[B]

This is one of my favorite takes of his is that in the bubbles, like when bubbles pop, rich people actually get hurt more than Main street. Which I think is not how it's framed most of the time. Because yes, there are some retail traders that go crazy and put, you know, they have a five figure net worth and they put it all in the most risky NFT and they do lose it. Like there are some anecdotes like that, but in general, most people have a pretty diversified asset base, whether it's their house or their stocks in a retirement fund. And those fluctuate a lot less than someone who's in the most risk on positions. We're Doggler, the first AI native dating and vibe coding platform for dogs. We've raised $150 million as part of.

[A]

Our seed as part of our seed.

[B]

Part of our seed. Dog learning is a good name. It's hilarious.

[A]

There's some truth to this. To this day, America and Britain benefit greatly from rail networks whose construction turned out very badly for the original investors. But there's another way to look at bubbles. The participants in the AI race are all building products that are economic complements to one another. You need the turbines that power the grids that power the chips that run the models that power the products. And you need firms to build their growth and hiring plans around the expectation that ever more of their work will be done by AI, but that every company and every employee will be automating different sets of tasks. If TSMC builds hugely expensive chips chip factories but the big AI labs all decide they've spent as much as they need to. Those factories are a stranded asset. But when asset prices are loudly signaling that the technology is real and the economics will be compelling, it encourages those complementary investments that actually make it happen. There are countless historical examples of this. The car industry's growth implicitly subsidized oil production, and vice versa. Electrification followed a similar path. Appliance manufacturers had to operate on the assumption that utilities would wire up more households, and those utilities had to bet that once power was available, ge, rca, and the like would give people something to plug in. During the heyday of Moore's Law, chip companies raced to build ever more powerful chips, and software companies rushed to ship products that would use them. It's hard for any of this to happen without entrepreneurs getting excited about a business based on its hypothetical future rather than its present profits. And it's impossible for this process to keep going unless investors, too, get excited. Naturally, one side or the other will overshoot. This hasn't been a technological revolution in history. There hasn't been a technological revolution in history that didn't, at some point get overhyped. That's always obvious in retrospect, but less so when we are in the cycle. An investment researcher once circulated an essay called A Home Without Equity is Just a Rental with Debt, warning that house price appreciation was driven by loosening underwriting standards and would inevitably lead to collapse. But it was dated June 2001.

[C]

Wow.

[B]

That's crazy.

[A]

Even at the Post, it's so early.

[B]

Because it's true, it's seven years too early, or six years too early.

[A]

Even at the post crisis low a decade later, the Case Shiller index of American house prices was still 18% above its level when that piece was published.

[B]

Wow.

[A]

Wow.

[B]

I mean, this is like the Bitcoin bubble stuff. Yeah, it's going to crash. And it's like, yeah, crash from 100 down to 90 or whatever. It's.

[A]

Similarly, media coverage of dotcoms described trading as nutty and quoted an investor saying, I don't really know anything about the company. But that article, the Wall Street Journal on the Netscape IPO, was published in the summer of 1995. At its post.com low in 2002, the Nasdaq 100 was still 40% higher than it had been Then signs of a bubble aren't necessarily signs that it's time to sell because they precede the peak of the mania by an unpredictable amount. Anyone who read the quite cogent arguments against buying a house in 2001 or buying tech stocks in 1995 would have benefited financially from completely ignoring them. The famous dictum apocrypha Foley attributed to John Keynes is that markets can remain irrational longer than you can remain solid.

[B]

Keynes didn't say that. That's funny. I always thought that.

[A]

Really?

[B]

Yeah.

[A]

I mean that's, that's, that's like his.

[B]

Well, no, it's not.

[A]

No, I know, but apocryphally, yeah, attributed to him.

[B]

We'll have to get to the bottom of that.

[A]

But this presupposes that everyone has the same information and that irrational traders are simply ignoring it. It's more in the spirit of Keynes to argue that the economic growth is partly a matter of believing that it will happen. Recessions end when people and companies start to spend as if they're over animal spirits. And booms persist when some participants are building the infrastructure that others need to be make that boom happen. When OpenAI announces a splashy new scale up or Meta declares that it has found yet another opportunity to raise its planned capital expenditures, they're signaling to AI users, coders, lawyers, writers, whoever that they'd better be prepared for smarter models. The more people and organizations gear their behavior towards a world in which AI is even more powerful and ubiquitous, the more they're locking in the demand that justifies all of those eye popping expenditures. In the end, a bubble functions like an industry cluster that exists in time rather than space. If you want to be a movie star, you move to Los Angeles. If you want to start a hedge fund, you move to New York. And if you want to be. If you want a part of being first to something in AI, first to build, first to use, first to profit from asset prices are insisting that now.

[B]

Is the time to act. I love it. Fantastic.

[A]

Someone in the chat was saying earlier they were expecting Nvidia to beat and then trade down 5 to 10%, which I feel like is the consensus view now. Which means that I think we might see something else happen.

[B]

Something else happen? Who knows? It's all been very unpredictable. Well, I'm very excited for Burn to join the show. I'm also excited to tell you about Julius AI, the AI data analyst that works for you. Join millions who use Julius to connect their data, ask questions and get insights sites in seconds.

[A]

It really is like a UAV for your business. It is.

[B]

Although that's night vision. That's not a UAV sound. Play the uav. UAV online. That's the UAV for your business, baby. Crazy.

[A]

While I was reading, Ben got me this. What do we got here, Ben?

[B]

That's bub talk, baby. Wow. That's actually a lot of bubbles. I like that.

[G]

Whoa.

[A]

All right, we will be ready to go when Byrne joins the show. Hopefully those don't get on the camera.

[B]

There's some pretty crazy news. The founder of an ADHD startup is found guilty of conspiracy in an Adderall case. What a crazy story. Ruthie Ahee.

[A]

Gotta give credit to Will for, like, predicting this, like, years ago at this point, he was just saying, like, all this stuff is. That seems deeply.

[B]

So back in 2021, Will Menidis said on Oct. 3, 2021. So four years ago, he said telemedicine psychiatry startups have driven an unprecedented wave of amphetamine abuse. So he was worried he was sounding the alarm bells four years ago about ADHD medications being overly prescribed, too easy to prescribe. He said after tweeting this, an executiveelloahead.com DM'd me from an anonymous account details of my care history with them, asking that I delete the tweet or caveat that they are not bad. This is an unimaginable violation of patient privacy. And it's like a threat.

[A]

Just the worst person to.

[B]

It's also insane because he didn't specify. He didn't call anyone in particular out, but then he got a threatening message from one person in particular, so that was very rare.

[A]

Just say you're responsible.

[B]

Yeah. And so Will has followed it up and said, worth Remembering that in 2021, 2022, many major healthcare venture investors funded a cabal of Internet pill mills that operated with mafia tactics to silence regulators and drive an unprecedented wave of amphetamine dependence in the United States. Well, today there has been some justice done, I suppose, for these ADHD startups. And so a jury found Ruthia he guilty of conspiring to distribute controlled substances after her startup, Dunn Global, became a ready source of Adderall prescriptions for more than 100,000 patients. The jury found he and Dunn's former top doctor, David Brody, guilty on two conspiracy counts and four counts of distributing controlled substances. The former CEO was found guilty of conspiring to obstruct justice. So the company was the subject of a series of articles in the Wall street journal from 2022 to 2024. Maybe they got Maybe they're reading.

[A]

It's interesting. Tech and Venture basically decided, like, doctors were a bug, not a feature. Yeah, it's like, yeah, why waste time talking to a doctor just to get the medication that you want and that, you know, you need? It's like, oh, actually like having. Having somebody that is like, pro, you know, even if it's slower, like, having somebody that's there and actually understanding the patient and having some personal connection with the patient feels very much more and more like a feature.

[B]

And also having the economic incentive of the doctor being like, they get paid a lot of money and live a great life just to give great advice and follow the Hippocratic oath and be.

[A]

Like a pennyware to give great care.

[B]

Of their society, not to increase conversion rate. Exactly. As opposed to like piecemeal. How many scripts did you write? That's the pill mill model. And so defense lawyers argue that, that he. So during a seven week trial, prosecutors argued that he sought to enrich herself by making it easy to get Adderall and other stimulants. While the government classifies this as a controlled substance with high potential for abuse. The startup collected more than $100 million in revenue. And the defense lawyers argued that they just wanted to make it easier to get the drugs when there was a shortage of providers. Quote, I think the goal we want to optimize is to help patients manage their ADHD in a convenient way. And there's some good reasons for that. Not everything. Sometimes you actually can't get to a doctor. There were good arguments on both sides, but in this case, it does seem like they pushed it way too far. There's some crazy, crazy quotes in here. Whoever is the first person to get arrested, I'll buy you a Tesla. Ruthie told concerned staffers, saying, like, don't worry, you know, bend laws.

[A]

Wait, don't if you get arrested.

[B]

A former company executive.

[A]

I'll buy you a car.

[B]

Testified that the CEO encouraged staff to, quote, unquote, bend laws. Okay, I'm encouraging everyone here who works for TVPN never bend the law at all.

[A]

Operate within the laws.

[B]

Operate within the law, Tyler. 100% looking at you, Tyler.

[A]

Operate within the laws of physics.

[B]

Yeah, don't bend the laws of physics.

[A]

You're studying physics. Make sure to operate within the laws of physics. Always.

[G]

Yeah.

[B]

Whoever is the first person to get arrested, I'll buy you a Tesla. That's a crazy thing to say.

[A]

Was that in writing or was that just like a quote from one of the employees?

[B]

No, that was the former executive Testified on the stand that this was said by the CEO, which is pretty, pretty crazy. Anyway, fortunately, the bubble in ADHD medication is winding down as the justice is being served. We have Bern Hobart in the restream waiting room. Let's bring him in to talk about other bubbles. More positive bubbles, more beneficial bubbles.

[A]

Welcome. Great to have you here. We brought the bubbles.

[H]

Awesome. You brought bubbles. Great to be here.

[A]

The bubble king.

[B]

It really goes everywhere. For those who don't know you, please can you kick us off with a little bit of an introduction on yourself? And thank you so much for taking the time to be here.

[H]

Yeah, absolutely. So, hey everyone, I'm Bern. I am probably best known for writing the newsletter the Diff, which you can check out at the Diff Company covering topics in tech finance, everything adjacent to them, everything in between. Also a partner at Anomaly, an early Stage Frontier tech venture capital firm. Also co authored with an Anomaly partner, co authored the book Boom, Bubbles in the End of Stagnation, published late last year by Stripe Press. So yeah, I'm the robot, the bubble.

[A]

Boy on a roll.

[B]

How can we talk? How should we set the table? Do you want to talk about just. It feels like you've been sort of defined as like pro bubble. So I feel like asking you the question, are we in a bubble? Is a little bit irrelevant, but would you agree that we're in the bubble? In a bubble? Maybe at the start of a bubble, maybe at the end of the bubble, but it feels like we're in a bubble. And it's safe to say it now.

[H]

Yeah, totally. It's great. Yeah. That is like. That is the curse of co authoring a book that is trying to rehabilitate the image of bubbles is that every time the NASDAQ hits a new high, people start calling you and asking you if that's good. And yeah, my obligation here and the model advanced in the book is to say that, yeah, it is pretty good. Not to say that stocks will always go up forever. Not to say that everybody's options or their weird quasi equity participation units will all be valued at their present prices. But yeah. So the general argument we advance in the book, I guess you can rewind a little bit and say you can go through these different ways of talking about bubbles. One is to just say they're stupid. It's when people are, they just get overexcited about some new technology or in the case of like housing, they get excited about some very old technology that is newly easy to finance and they just lose their heads. And by the end, everybody knows they're overpaying. Everybody assumes someone else will overpay more in the future. And just when you run out of stupid people, prices collapse. So that's like the bubbles are really stupid vein of thought. And then there's one which is a little bit more nuanced, which is, hey, yeah, they're stupid, but they're actually a wealth transfer from hedge fund people, venture capitalists, etc. To everyday consumers. So I lived in San Francisco in 2015. I remember I long for the incredibly cheap, you know, universal basic Uber where every, you know, you could get anywhere for like $8. It was amazing. And that was, yeah, it was great. It was like, you know, wonderful wealth transfer from the Saudi Arabian sovereign wealth fund to me and I really appreciate it. Thanks guys.

[A]

Well, it's also notable that they've done fairly well. Even though there was like it was a bubble, it wasn't necessarily sustainable, but it created a enduring business or at least one out of the two.

[H]

Yeah, and that's, that's often what we get.

[G]

Now.

[H]

Uber is a little bit of an abstract case. What we often get from bubbles is we build too much infrastructure, but that means we have the infrastructure and we have enough infrastructure to build the next thing. And that was the case in the 19th century with railroads. That was the case in the late 90s with telco infrastructure. And you know that that could end up being the case today with GPUs. But there is the pro bubble argument, which is that what bubbles really do is they coordinate different market participants, founders, employees, investors, regulators, customers, suppliers. They convince everyone that this is happening, it's happening right now, and that if you build something, if you overbuild for today's demand, you will still have underbuilt for future demand. And if everybody's doing that at every layer in the supply chain, then you actually do build enough to satisfy future demand. And so like making it more concrete, if, if TSMC does not buy into the idea that AI is a really big deal, they're not going to build enough fabs, Nvidia will not be able to ship enough chips, the next models will not be quite as good, or we'll have to do more of a trade off between training and inference and the whole thing slows down. But if everyone is wildly optimistic, then they do build all of that infrastructure, all the way from the power generation to the end use cases, they're building all of that and it's kind of like just in time, manufacturing of the future and the prices like the crazy prices they are this signal that, that this is the time, like it's happening now. If you build something on the assumption that OpenAI is going to keep shipping better models and that they will need a lot more compute and that will need a lot more power to make those models work, if you operate on that assumption, you're making the right call.

[B]

How have you been processing some of Ben Thompson's maybe jitters around the idea that the infrastructure that gets left behind in this particular bubble might not be something that's with us for 100 years? He's been advocating for, hey, let's do energy, let's do nuclear, solar, let's build out a lot of energy. But if we're just like, yes, we way overbuilt, like a bunch of H1 hundreds and then, you know, years later we're like, those days actually aren't that valuable. They don't. They maybe depreciate over a few years. That's sort of how he's articulated some of the fear of the overbuild not being as durable. Does that resonate with you or how have you been processing that?

[H]

Yeah, I think, I think you can look at these different lags and you also look at how generalizable the use case is are for these products. And so it is true that GPUs depreciate, but depreciation, it is this economic concept that's tied or it's an accounting concept that's tied to an economic concept. And it actually ties in a couple different things, one of which is just if you use a machine, there is friction, it can break, it can overheat, whatever. And so eventually it's trash. And then the other piece, more on the economic side and much more relevant to GPUs is if the GPU is producing fewer tokens per watt, and that just relative to newer ones, it can be economically worthless, even though it can still actually do something useful. And so if the GPU build out slows down, that actually decelerates the depreciation for all of the world's existing GPU fleet. And power is harder to slow down. The just the lags are a lot longer. It takes a long, long time to build a new power plant and to build all the equipment that goes into it. So you could have this case where bubble pops OpenAI has to do some weird recap at a much lower valuation. Nasdaq's down by half, or probably not down by half, but down by a lot. And you know, a lot of a Lot of people who felt very smart as of like a month ago look pretty stupid, myself included. And you know, we could have that but there will still be this increase in power generation capacity because that stuff is locked in. And the like the gas turbine companies, they, they can really lock their customers in because there are just not that many places you can go to buy one. And so if you're going to buy one and they say okay, but you have to actually guarantee that you'll pay for it. Even if you really regret this, this, that's pretty much what you'll have to do. So you could have this case where what actually happens in the aftermath is power costs decline. And so these GPUs actually become higher margin when they're doing inference. And so you get really, really cheap, abundant intelligence at today's model capabilities. And that that is still.

[A]

Yeah, the GPU is fully depreciated and power costs have come down which would make them, you know the concern around depreciation as you just of kind get a chip that is so much better that it's just non economical to run one of these old GPUs. But the scenario that you're pointing out, they could potentially be valuable for much much longer.

[H]

Yeah. So if you look at a company like Core Weave, where their business model is, we stack a bunch of GPUs in a data center, we lease them out to various people in varying terms. That is actually kind of a bet on this narrow slice where AI is not a complete flop. We don't find out that it was actually just Sam Altman typing those answers really fast all along. But it's also doesn't completely revolutionize things because that like if there's another generation or two of GPUs or if TPU's take more share of inference then maybe those GPUs end up being economically stranded. But if there's a world where we're not building many more GPUs but we are using, we do find all the use cases for the ones we have that might be a world where actually Core Weave is you know, reporting pretty nice GAAP profits and, and their investors are happy. And when you look at Corey, one of the interesting things about them is on their cap table. So they, one of their big backers is this hedge fund Magnatar which does incredible. Has done incredible things with structuring various bets like if you control f their, their prospectus. Magnatar is actually mentioned more often than Nvidia and Magnitar likes to make interesting bets on like the relative volatility of different things or the time, relative timing of things. And so you could see this as them making this kind of esoteric bet. That AI is actually both a really big deal and somewhat overhyped now. It's always, always hard to read into. Like what you're reading into is Magnatar is involved in this. They have done a bunch of really interesting deals throughout the core we've cap structure. But it is, it is still striking that they are very sophisticated about this exact kind of trade.

[B]

That's interesting.

[A]

What is like, what is the right way to view a bubble? Is it this monolithic structure in which there. Or are you viewing it as like. Because in my view it's like we just read your article or essay in the Economist right before you.

[B]

Congratulations, by the way.

[A]

It's great. Thank you. Great kind of summary of everything, but.

[B]

High honor to be in there.

[A]

But you're saying like there's pieces of a bubble that are feeding into each other and making the possibility for durable value creation to be higher and higher because you have all these parts combining. I feel like another view maybe is similar, but slightly different is like you have these rolling bubbles that are all kind of like building up and right now there's like, like we have a private credit bubble, maybe there's a Neo cloud bubble, maybe there's an LLM bubble. Right. Like do we, it's unclear if we need, you know, the, the 50th closed source LLM. Right? I don't know. Maybe we do. Right. It's hard to predict. But what, what is like the right kind of like way to even just visualize the bubble, the AI bubble broadly.

[H]

So like a lot of other bubbles, it starts out as this really differentiated, unique thing where most people do not know, you know, like five years ago, most people did not know or care that much about AI. It was kind of this thing where you would listen to the quarterly call from Google or from the company then known as Facebook and they would talk about how they're an AI company and you'd think, okay, like I'm glad you have your science project nerds, but I really care about more ad clicks and more dollars per ad click. So good luck with whatever, whatever robot experiments you're doing. And then when it starts growing, what it starts doing is actually connecting with the rest of the economy. So now like the marginal dollar of AI Capex is increasingly to general purpose power generation infrastructure. So. And meanwhile AI is getting much more broadly distributed. Like initial use cases were one. It was A really good autocomplete for coding. And two, if you needed to create original content in order to spam people, or if you were replacing like the lowest valued bloggers, you could do it and it was cheaper. But then it became this thing where it's like a lot of. There are just a lot of things where you wish you could apply a little intelligence to it. It's really not worth your time. But if you can get the right answer easily, then you should do it. And just like a lot of cases where you'd want it. Like, I use it a lot when I'm writing as a research tool where it's like, I want examples of this.

[A]

Phenomenon or it's a research tool, not a writing tool.

[B]

Yeah, right. Are you even using it as a first draft or is it more like you have a bunch of facts here and then you are actually typing out.

[A]

The sentences that you were like bloodhounds for AI content. And we did not. And no, no alarm bells went off when we were reading.

[B]

Oh yeah, the Economist article was like.

[A]

I've just been, I've been surprised.

[B]

The most clearly human written.

[A]

It's so, it's, it's extremely notable that like using AI for writing has become the most low status thing that you can do on the Internet.

[B]

It's like, just feels like disrespectful.

[A]

It is lower status than making just like sloppy memes.

[B]

Like, there are status things you could do.

[A]

Yeah, okay. Like adult content. Adult content.

[B]

Or like use my coupon code to sign up for prize picks. Here's my parlay. Who's riding?

[A]

But still, it just communicates. People feel disrespected by it because it's like, hey, you put this out. You wanted me to read this. And if it's completely obvious that a computer generated it, it's like, well, well, was this even worth my time? Right? Like if you couldn't have said it in your own words.

[H]

So there is this dynamic where they're just sometimes when there's increasing efficiency with something, we find out that some of the effort was load bearing. And that doesn't mean the technology is bad. It means we do have to adapt. And so in the case of writing, one of the things that used to be the social norm was if you can produce a grammatically correct, lengthy document about some topic, that is an indication that you probably know what you're talking about. And to get into a position where you can do that, you have to read a lot so you acquire knowledge. And if you want to write something Persuasive. You probably have to talk to a lot of people and find out what's persuasive to them, what's persuasive to you, etc. And if you can just, just ask a model to admit that, then you can basically write at a level that is much higher quality than your ability to think. You can write well beyond your wisdom. It's kind of like when people use some peptides and steroids, they end up getting weird injuries because they're just like, mechanically, their body is not actually suited to lift the weights that their muscles can move. So they do get serious injuries unless they train pretty rigorously. So I have a 9 year old who has in the past used ChatGPT to write emails to me explaining her side of a fight that she had with one of her siblings. And the email is very clear, very articulate, lots of EM dashes, lots of it's not X, it's Y. And it's, you know that I think if someone that age sat down and write this coherent letter explaining their side of an argument, that would actually be impressive. Like you'd say, okay, this person's actually thinking seriously about what happened. But in this case it's like she can write two sentences in ChatGPT and answer some follow up questions from it and then produce this nice coherent looking document.

[A]

So do you, how much do you, how much do you worry about a new. We have kids younger than that, but how much do you worry about potentially a generation of young people? Never, like, you know, maybe in a classroom setting teachers can be like, put your phones in this box and you guys are all going to write a paper on this. And it's possible that writing will become like highly supervised because the only way to prevent somebody from just generating the.

[B]

Written word, it literally already is at many schools.

[A]

But yeah, yeah, yeah, but, but, but even, even then it's like when I think about growing up and being forced to think deeply about topics, oftentimes it was because I was assigned to write an essay on something and I didn't have the world's best autocomplete tool and I just had to sit there and kind of wrestle with an idea and actually learn about it. And I had to read a book or read a bunch of essays and really put it together. And I think it's possible that just a lot of time spent deeply thinking is just fully lost forever.

[H]

Yeah, I think it comes back to that load bearing effort question. So I do tell my kids that there is just a qualitative Difference. And also that, that when they get an assignment at school, it's not because the teacher has this burning desire to read an essay about, you know, whatever about Charlotte's Web or something like. It's not like the teacher is absolutely, you know, they have been pining for this. It's like the point is the effort, and the point is the way that you think about things and that writing is actually just a very useful way to think something through. I don't really understand why that is. Like, I don't know why it is that. That if you just try to talk to yourself for 20 minutes straight about a topic, you won't get to the same level of clarity that you do if you type it out. Even though the typing it out process is just really similar. It is one word after another and then a little bit of editing sometimes. So I think some of what the education system has to do, which different schools do to different degrees, is actually explain to kids what the purpose of what they're doing is so they understand what that purpose is. And then we also do have to make this adjustment of sometimes there are things that it is. It used to be necessary for basically every adult to be able to do no longer as necessary, and fewer people will be able to do it. And maybe the ones who do it will still take a lot of pride in their craft, but they won't strictly have to. So think of it as like, I don't know, things like manual labor and I don't know, wilderness survival skills, things like that. That there was a time when being physically strong and knowing how to, like, being able to navigate in space and figure out which way is north if you're lost was actually a pretty important skill that a lot of people had to have. And there's actually.

[A]

You'd be mocked if you could.

[H]

Right, right. And so then you go through this generation where there is a lot of mocking, there is a lot of bullying, but the nerds are actually probably right that this thing is not so important. And then the next generation is only the hobbyists who do this. Thomas Sowell had this argument about, I think it's in his book on knowledge and decisions where he's talking about how if you live in a really, if you live in a subsistence level tribe somewhere, you actually have to have this incredible breadth of knowledge. Like you've got to know all the landmarks, how to get from one place to another, all the signs of danger, everything you can eat, everything you shouldn't eat, and you know, which. Which local tribes are Friendly, which ones aren't. And you just don't need nearly that level of knowledge to survive in a modern city today, there are all kinds of things about where your food comes from and what is safe to do and not to do that you simply don't have to know because you're not exposed to any of the risks. And so we, we actually have just a much lower knowledge requirement in, in more advanced societies, on the other hand, we have much higher returns from having that, having unique kinds of knowledge because now that whatever value you can create can be amortized over a much larger number of people and there's just more stuff to go around. So the rewards from being really, really smart are a lot higher. And, you know, I hope that when I talk to my kids about this stuff and I basically say, like, there's going to be a cognitive overclass and a cognitive underclass you can opt into and it's super easy, you tell your.

[B]

Kids that it's amazing.

[A]

You must escape the cognitive underclass.

[H]

It is true. Like, it is so easy to go through life without thinking and it will only get easier. And so you, you have to decide knowing that the thinking part is increasingly optional in a larger and larger number of domains. Do you want to be the kind of person who thinks because you like thinking and you like creating and discovering new things, or do you want to be the kind of person who has just a much easier, more relaxing time? Because they don't.

[A]

We could continue on this conversation for a long time, but I wanted to ask you about what scares you about this current bubble. Like, things that are not necessarily, like, bad today, but could get bad. To me, like tech indulging in leverage for the first time, maybe as an industry or as like a lot of the leadership has not. They weren't in, they weren't participating in the telecom bubble. They didn't get blown up. Maybe they've never gotten blown up by leverage, and maybe that's a concern. But I'm curious how you think about it.

[H]

Yeah, I mean, I'm less concerned about that. I think the current generation of tech leaders, there's a lot more tech history that they can know about and they just seem more interested in tech history. You can actually go back and see that. But the people who were more obsessed with tech history tended to do better. Like Steve Jobs was obsessed with the story of Polaroid. It's this beautiful consumer device, changes everybody's behavior. Really simple tool. You look at it, you know exactly how to use it, you know what it does, and it does what. It's what it looks like it's supposed to do. Jeff Bezos gave a TED Talk when that was a much cooler thing to do, I think right after the dot com bubble had rolled over. Where he's talking about the early days of electrification and how the Internet is like that, partly in the sense that we did not know how to use it. We didn't know all the applications. And he, I think he, he said, I think that's where I heard that the original appliances, like if you bought an iron originally, it would actually plug into a light socket. Like you'd unscrew a light bulb, screw in the iron, and then iron your clothes in the dark and then screw in the light bulb again. Or I guess you'd iron your clothes during the day. But anyway, like, we. It was very janky. And so you could have looked at it at that time and said, like, this is just a clown show. Like, okay, sure, electric lighting, I get it, but what are you doing with all these other weird gadgets?

[C]

And who needs that?

[H]

Like, we already had irons. They were fine. So I think that a lot of tech people are actually pretty keenly aware of history, and a lot of them are just. They're way more obsessed than you would think with the prospect of their company becoming irrelevant in six months and a total failure in two years. So I think we're, you know, it is riskier to borrow than not to borrow, but we're probably safe on that front. I think one thing that could go wrong is some combination of corporate behavioral norms and regulatory norms and investor assumptions where we decide that this stuff is really dangerous, we should not touch it. It will blow a giant hole in somebody's balance sheet because we know it happened and it'll happen again. And it just becomes untouchable for a while, which did kind of happen in the dot com space. And I think people underestimate that when they look at things like Mark Zuckerberg start social network in 2004. Is that. That was. You could have looked at that as really, like, now it looks really forward thinking. At the time, it kind of looked dated. It kind of like. The example I use is like, if you. If in 1999 you moved to Seattle to start a grunge band, like, you missed it. You were, you were way out of date. And that's what it looks like. So it was still a kind of contrarian thing to do. And it was still a company that was started in the aftermath of this dot com bus when people were cautious so but with AI, the capital requirements are so high that it is actually a really big deal. If investors decide that the space is uninvestable, progress actually stops. Whereas you just don't need a lot of capital if you're in your dorm room on your laptop, just slinging php.

[B]

Yeah. Or you can at least monetize much, much earlier. And you see that with like the Google earnings like pre IPO is like a massively profitable business. Undeniable, didn't need any permission. What do you think about sovereign AI? How bubbles spread internationally? I was listening to Tyler Cowen talk about one of the weird side effects of tariffs is that other countries might copy America's tariffs just for sort of mimetic reasons. And America might be in such a powerful position that tariffs might not actually wind up hurting America because of its position in the global economy. But if another country says, oh, let's copy that, they might be hurt more. More. I'm wondering about how bubbles propagate. At the same time a lot of the telecom magnates in foreign countries that just kind of copied our telecom build out, well, they're the richest people in those countries now. So how are you thinking about the bubble spreading internationally?

[H]

I think it is a really cool toy for petro states and some of them have actually done some really impressive work. So I don't really begrudge them. I'm not sure how many, how many general purpose models the world needs. I suspect what the world needs is lots and lots of special purpose models. And that could be the level of okay, this model just knows rust, but it is insanely good at rust and it has not polluted its mind with any bad habits from C or C or anything else. It's just pure rust. And then you could also have even more narrowly scoped models where it's like this model is this one person and it will give you the best approximation it can of what this one person does. And, and if you have a lot of different models and people who interact with models interact through a router where the first thing the router does is figure out which sub model to send things to and it can do many iterations of that and eventually might be sending some things to, you know, maybe delegating some things to an agent that ends up talking to an agent at some third party service. So like I'm thinking of things like if you are planning, I mean everyone says if you're planning a trip, let's say you're planning a really complicated tax sensitive global MA transaction. So maybe you need like the French tax law bot to interact with the US tax law bot and they both need to make sure that the economics of your weird tax thing also makes sense. In that world, you could actually have this great diversity of models with a great diversity of model use cases. But for the general purpose stuff like I don't, I don't think there, I, I think that there is enough room for customization at the user level that we probably don't need, need 50 different models that are close to the frontier.

[B]

Yeah.

[A]

What are your labor displacement timelines? Because every CEO over the last year has used AI as the reason behind layoffs. And I think everyone has been calling BS on a lot of that. It's just like they need a good reason to do around a layoffs for other more real reasons. And everyone I think has seen the chart by now of job openings versus when ChatGPT was released. And at the same time, if you've used these tools, a lot of people.

[B]

It doesn't feel like a drop in replacement.

[H]

Yeah.

[A]

Meanwhile, you have engineers, LLMs are incredible at coding and you have engineer. If you're a talented engineer or even a high agency engineer, you probably have more opportunities, opportunity than ever. So I'm curious about how you're thinking about timelines.

[H]

Yeah, this stuff takes a surprisingly long time to deploy because one of the load bearing inefficiencies is that if something required intelligence, there's at least one human being whose judgment is implicitly tied to the output of that product. And it's really hard to go from there is some specific person to blame, like if a mistake was made, someone made it to. If you scale up your work by 100x and now 95% of the time you do just fine and 5% of the time you mess up, is that your fault? Is that Claude's fault? We don't want to blame Claude. Claude's so nice to us. We don't know. So we actually have to rethink how people get judged. There's this sense in which everyone becomes a kind of engineering manager who everyone in software becomes this engineering manager who is describing what needs to get done and vetting what has been done, but is writing less code themselves. On the other hand, and LLMs are actually pretty good at doing the opposite. Where you are the junior coder, you are doing the grunt work and what it's doing is looking at your overall architecture and telling you what things you missed and what design mistakes you have made that are just a lot easier to fix upfront. But a lot of organizations, they don't want the risk of their workers are massively more productive, but they're also producing some mistaken things. And that's actually going to be a big hit to the company's reputation. So you'll probably see. What I think you'll see is that a lot of AI deployment is that there will be a legacy version of something. There will be an AI native version of that thing. The native version will sell to smaller customers, those customers will grow faster than legacy companies and then the AI native product gets sold to all the legacy companies. So this is kind of the stripe model where they started out doing payments for early stage companies that had pretty simple requirements and had some tolerance for error. And then they, as long as they stay good enough to maintain whatever their biggest customer is, they are necessarily building out the feature set for other companies the size of that biggest customer. So you get some deployment that way. But it has this, it actually takes a while because the big companies, they just, they want to be somewhat cautious on this. And you sometimes have this case where there's a top down mandate at a big company saying everybody's got to use AI and there's also this bottom up insurgency of I can use AI and it makes my job more effective.

[A]

Well, also there's going to be a dynamic, there's a dynamic too where we will see scenarios where employees say, well, I don't want to adopt this AI, this one's a little too good. I'd be worried about losing my job. Right. And so I think we're going to see like more friction between even with tools that actually can replace labor, like truly not just being like a copilot and the friction to adoption because the people that would be adopting them, and that's probably years out, certain investors have been underwriting early stage private market bets to they're saying like labor is the tam. Like how do you view that framework? Is that it feels overly simplistic to just say like any dollar that is spent that goes out through any type of payroll system today is up for grabs. But there seems to be some element of truth to it.

[H]

Yeah, there's a little truth to that. But I think it's in the same way that Netflix says that time is the TAM and their biggest competitor is sleep. It is broadly true.

[A]

It's marketing.

[H]

Marketing that, yeah, like it is marketing, but it's also a way to frame the scope of the opportunity. So what I would say is that when the labor in question is mostly delivering value by producing a Sequence of tokens, whether that is writing a document or building an Excel model, or writing some code, that that is the addressable market for an AI tool. But the real world just has enough complexity that models have to develop a really good world model. And one way to think of it is in software they have a really good world model because that is their world. Like their world world is this abstract world that is defined by whoever wrote the compiler. And to a lesser extent that world has some complexities if you're actually working with real world physical systems where someone can trip over a cord and unplug one of the servers in your distributed system. And that is just not contemplated in the purely software world model. But as soon as you move out of pure software that is running on one machine for one user, you start to get some real world complexity. And then when you're trying to automate something like building a financial model, you need pretty tight feedback between what assumption works in the real world, what actually maps economic realities, and then what assumption is the, the most probable token in this cell that needs a token. And I think that will like, as AI gets deployed in messier parts of the world, what you'll actually see is that more of the world will get structured in an AI friendly way and that more, more of GDP will be in that world that is already pre structured for AI. But then you still have the rest of the real world where just really hard to get ego onboarded. And you can actually see that with things like when, when the company then noticed Facebook was growing internationally, one of the obstacles they ran into was in many places almost nobody has a computer. So that's one of the reasons that they went into mobile early. But they also realized they could market themselves through Internet cafes. And that apparently for a while in the developing world, if you went to any Internet cafe, half of the unused computers would have the Facebook logout screen. And that was actually a huge source of user acquisition for them in developed market, in developing markets. And then once smartphones came out, those people migrated onto smartphones and then Meta was able to keep them and continue to sell them.

[A]

So somebody would be on Facebook, they would use a computer in an Internet cafe, they would get up and leave and then somebody would sit down and they'd be like, what is, what is Facebook? And then they would just create.

[H]

Exactly.

[C]

Yeah.

[H]

Yes, there's, there's a great story in Chaos Monkeys about this and about how they wanted to have an ad on the logout page because they're like, this is otherwise just wasted Real estate. And it turned out the logout page is this mission critical thing in all the, in many of the non US markets. So there was a big internal fight on that. That and they did end up doing ads on the logout page only in developed markets where growth had slowed down enough and they were already a dominant market share. But like they, they needed the outside infrastructure to catch up with the product. And once it did, the product was already there kind of waiting for that infrastructure and saturated it really quickly. But this, this is another thing that happens with bubbles in general. Technology bubbles in general. It's like you, you don't consider the podcast an electricity company. Like you don't think of yourselves as that business. But the business doesn't really function if you can't plug something into an outlet or use a. And actually get power from.

[B]

Nothing can stop us from podcasting. Let's be clear.

[A]

You could yell yes, do without microphones, without cameras.

[B]

We'll just find a crowd of people.

[A]

And scream at the megaphone. But yeah, on the rooftops.

[H]

But like so in one sense, the 1920s bulls who were like, I'm all in on electricity. This is the future. They were absolutely right. But if you transport a trader, a stock trader from 1925 to 2025 and you're like, okay, go buy all the electricity stocks is like, well that's everything. Like every company uses this. So there is no real way to make a direct bet on it. And to the extent that there is, the direct bet is now a totally different bet now actually that, that particular time traveler, if he arrives in, in 2023 instead of 2025. Actually his, his 1925 thesis of just buy levered power generating companies that put all your money to that, it's actually brilliant. So these things, you know, the cycle does repeat itself a little bit. But yeah, they as it, as it disperses, you've got a little bit of AI and everything. And it's, you know, Internet is the same way. Like you don't consider Target an Internet retailer. They do a lot of E Comm. All the physical, basically all the physical retailers do a lot of online sales. The, the fast food restaurants do a lot of their sales through apps and through kiosks. So there is just this convergence where by the time the bet is such a big scary bet that you're like the whole economy is dependent on this. You're also like, well, it's just mixed in with the whole economy. Like you can't actually take the AI part out of the US economy. And the US growth story without completely breaking things. And at that point it kind of converges. It settles.

[B]

Yeah, makes a lot of sense.

[A]

I have one more question that is probably worthy of like a 10 minute answer, but we'll see. We only have a few minutes. Is it the CEO's job to disconnect the stock price from reality?

[H]

Well, it's partly the market's job to tell employees where their equity, where they should go if they want to max out the value of their equity comp. And this is something that I, I used to not really believe. And what happened with Metta in 2022 kind of converted me this view where Mark Zuckerberg did not actually have to care that his stock was under $100 a share. Like he, it's not like the board is going to vote him out. Even if, even if he didn't have voting control, people, they're just not going to kick him out. But it did mean that it was harder to recruit people. And so if your dream is we're all going to live in the Metaverse, we're going to have this legless utopia. You can only hire the people who make that possible if they think your stock is going to go up. Otherwise you have to pay them entirely in cash. And then your stock goes down even more. And suddenly you're in this position of making really hard decisions that you don't want to make. So sometimes you, you take a foot off the gas pedal in terms of massive capex for something investors are skeptical of. And as long as you're still in the lead, and this is what like investors would send hedge fund people. There was a great hedge fund letter that was plaintive. It's like Mark, even if you cut your metaverse spending in half, you'd still be spending the majority of the world's metaverse money. Like, you know, you're still a winner, you're still, you still get the trophy. But please, just give us some free cash flow. Buy back some stock. It's cheap.

[A]

Yeah, it feels like such like disconnecting your stock price from reality, at least to the positive. Positive can be a massive advantage. And you can see different. Palantir is a good example of this, or Tesla is a good example of this. And if you can keep it going, it's tremendously effective because investors want to be in companies where the stock price is not necessarily always going to be tied to fundamentals and even employees can benefit. And maybe the opposite side of that is Dylan Field with Figma. I feel like he just he wants to be valued fairly and accurately and just wants to make the business better and better and better every day. But of course it's a double edged sword and it's great when it's disconnected to the higher end. But anyways this was super fun. Thank you. Thank you so much for joining. Would love to have you back on again soon.

[B]

Let's do it again soon. This is absolutely pleasure. Have a great rest of your day. We'll talk to you soon. Baldur, you too. Before we hop on with our next, let me tell you about Privy. Privy makes it easy to build on crypto rails securely, spin up white label wallets, sign transactions and integrate on chain infrastructure all through one simple API. Our next guest is Glenn Hutchins. He is the co founder of Silver Lake Partners, a legend and the chairman of North Island North Island Ventures. I believe he's in the restream waiting room. We will bring him into the TVPN Ultra dome. We were keeping him waiting just a few minutes.

[A]

We'll keep the Bub talk going available.

[B]

How are we doing if he is on the line? Glenn, good to see you. Sorry for keeping you waiting. Welcome to the show. How are you doing?

[A]

We don't have audio. Can we check how we doing team?

[B]

Mute button. Check on that and see if we are getting audio through the call. I will give some more context. He is the chairman of North Island, North Island Ventures, the co founder of Silver Lake, the vice chairman lead independent director of Cander.

[A]

There we go.

[B]

He's also the lead independent director of Cor Weave and he's here on the show. Welcome to the show. Thank you so much for taking the time.

[G]

I just want to say I'm a big fan of your show.

[B]

Oh that's amazing.

[G]

So it's so much fun to be here. Really a real pleasure to meet you guys almost in person.

[B]

Yeah. Well next time you're in Los Angeles, please feel free to stop by the TMP Ultradome here in Los Angeles.

[A]

We are so excited to have you on. So much, so much to talk about.

[D]

Yeah.

[B]

Why don't we start with just a little bit of your career arc. I know we're going to want to talk about the dot com bubble, the dot com boom, your experience there. But walk me through your career up to co founding Silver Lake in I believe it was 1999. Right?

[G]

That's right. Just really, really briefly please. Basically when I got by Silver Lake was my third and now I'm on my fourth. Essentially startup.

[B]

Yeah.

[G]

In and around investing largely called private equity the first one, I was a junior partner to a guy named Tom Lee, founding what people look back on now say as one of the first private equity firms. Then I took some time off, worked in the White House for Bill Clinton and then was recruited to come to a young little firm called Blackstone.

[D]

That.

[G]

Was getting into the private equity business and wanted to build a private equity platform. About five years later the Blackstone guys helped me start Silver Lake which was the. They invested in it, which was the first large scale organization to combine private equity type of investing with technology. And now I'm on my fourth, which is my platform called north island, which has one very difficult limited partner, which is me. So it's my fourth startup in investing and maybe we can get into this a little bit later, but the origins of private equity might be something worth talking about at some point if you'd like. But we'll come back to that. What's your next question?

[A]

Let's start there. I'd love your view on it and it really is quite funny to think that you couldn't have maybe picked better stepping stones, stones across the whole journey must have been. You must have had some good intuition.

[G]

Yeah, you know, it's better to be lucky than smart. But so, you know, the one thing I would say is can I speaking. Can I get a little geeky for you with you guys for a moment? We can come back to the more personal dimension. But there are four or five real advances in largely quantitative approach to finance that enable the creation of kind of what I've done over the years years, especially in the early 80s when we started thinking about private equity. And the first was the capital asset pricing model which allowed us to really do very good in depth valuation of equities, which has not been done before. The second was Black Scholes option pricing model which allowed us to value options and really understand what embedded options, how to value embedded options inside of equity securities. Oftentimes when you bought a private equity company, you, you paid for the company and then you identified something inside the company had a real upside and how to value that and how to pay for that was the question. Second or maybe third was understanding fixed income. A fellow named Marty Lieboitz came up with something called inside the yield curve would let us really value fixed income. And then Mike Milken really understood really good work on understanding the risk reward associated with high yield securities which became a tool that we were able to use use to build these companies. Michael Porter at Harvard Business School did a bunch of research using standard economic Analysis about the five forces that you could use to extract value from companies, which wasn't being done in a very systematic way in those days. And then finally, modern portfolio theory with sharp ratios and efficient frontiers were adopted by places like Harvard and Yale. And a key part of that was having an allocation of private equity. And as that model was rolled out across first pension funds and then sovereign wealth funds, a huge amount of money flowed behind us.

[B]

That makes a ton of sense.

[G]

We figured out how to value companies, we figured out how to use debt, we figured out how to extract value from the companies, and then we had a big flow of money coming in to back us doing it.

[B]

Fascinating.

[G]

So that's kind of one way to think about what happened over the last two years.

[A]

How quickly did those ideas and methods actually disperse? And tying that to the present. It feels like with AI labs today, there'll be some sort of innovation that gets discovered, and then one of those people gets immediately poached to another lab, and suddenly another lab is developing the same type of system or approach.

[G]

That's a really good question. Now, someone quote. You could look it up.

[E]

Someone quoted.

[G]

I read a quote recently, said the future is already here. It's just not evenly distributed. And so, you know, innovators try to. I've always tried to find the next way to be successful in investing and stay ahead of what people do who copy me. One of the things I say is that my very. My. My best idea is the ones that people dislike. And my very best ideas are ones they hate intensely. Because I know if someone really hates something, I'm thinking about it, and I know it's like, could be really, really good. And then by the time it turns out to be generally accepted, that's when I sell what I bought before to.

[B]

Them.

[G]

If you know what I mean.

[A]

Yeah. So you're basically like, if they. If they. If they think an idea is dumb or silly, that gives you a window of opportunity to signal it's gonna be good. Yeah. Well, and it gives you a window to, like, you know, get as much value out of that idea before it becomes common knowledge or an accepted approach.

[G]

Occup. Occupy that territory before they get there. When they come there, then you sell to them the beachfront property that you've already purchased.

[A]

That's right.

[G]

And then. But to go back to it, the half life of innovation on Wall Street AI is a little bit different, but the half life of innovation on Wall street is the time it takes someone to read a prospectus and then copy what you did. And so like, you know, someone does a SPAC and everybody does SPACs, someone does a digital asset treasury thing and then everybody wants to do a dat, you know, I mean. Yep, it's like on, in, on Park Avenue, New York City. As soon as it starts raining, the guys with the umbrellas come out. It's almost like they knew it was raining and then blocks on either side of Park Avenue. Everybody's with umbrellas and got to figure out something else to sell because the umbrellas are already there. But the. So you know, when we first started doing in private equity was a way of exploiting value that was latent inside companies because you didn't have the financing to be able to purchase these companies and you didn't have the toolkit to extract value from them. Those are the issues that we resolved with what I talked about earlier. Especially when Mike Milken untapped this high yield market that we could borrow from to finance these companies. Then people rushed in and in part the Blackstone idea you'd have to talk to Steve Forsman was to build a platform that you could take to scale scale where you could raise an amount of money that people who would come into the space couldn't match you. And so you could mine a different vertical layer of companies that were immune yet to private equity disciplines by getting to scale in the enterprise. You see what I mean?

[B]

Very controversial in Silicon Valley to call that out. Nowadays there's a big discussion over platform funds and funds that might be doing exactly that.

[G]

Yeah, okay, we'll come back to that. And, and then I decided that another path that the technology industry had reached a point where there were scale companies where you could use debt and more private equity style skill sets to buy the companies. First big one we did was something called Seagate where we borrowed a bunch of money to buy a big tech company. But if you look at companies like. So when I was coming up, people looked at companies like Microsoft and they said, oh, these are big, very risky company. Steve Ballmer and Bill Gates were college classmates of mine. By the way.

[A]

You'Re a lucky guy.

[G]

Class of 77 at Harvard. Steve and I graduated, built the nut, but he did better.

[B]

On the financial innovation. It feels like a lot of what you identified, Black Scholes, modern portfolio theory, Sharpe ratios, all of that, that's all pre1999. I'm interested in understanding what was the key unlock to bringing private equity to technology. Specifically, were you thinking about Metcalfe's law Network effects, zero marginal cost. Were you looking at Businesses that fundamentally differed from the traditional widgets business or industrials business and had different structures or what else was going on there?

[G]

Really good question. So at that point technology was in an important transition. This is like the future is here. It's just not evenly distributed. Yeah, the. It was thought to be an area. Two things. One, it was thought to be an area of expertise where you. And it was true. You really had to have specialized expertise to understand the companies to invest in them successfully. You couldn't just wander off of out of Wall street with your pinstripe suit and sort of think you could figure, go to a couple conferences and think you could figure out how to buy, you know, a technology company. Because the process of evolution was so rapid.

[E]

Rapid.

[G]

And then secondly to that point people did not understand how technology companies had evolved that point. Technology companies were big. They consumed huge amounts of cash in investing in R and D to build the products and they had very volatile earning streams as a consequence of being pioneers in a space that came very quickly. And so people looked at that from. There was a venture capital gig. But look at that from a private equity perspective and say you can't do it. But at that point, Microsoft, for instance, that's why I was talking about Microsoft, got to a level of scale where it was one of the greatest economic enterprises in world history, where you make this piece of software that comes out of someone's brain, has almost no capital expenditures associated with it, no kind of fixed cost and sell it a billion times. Right. I mean that's. And it just this massive flywheel of cash comes into that company. And we, I, you know.

[A]

And when was that? When was that like light bulb for you? No, but for you personally.

[G]

I observed it in the 90s. I had the benefit then of living in Boston and the venture capital business was pretty, pretty vibrant there. Then it moved later primarily to Silicon Valley. But there was a big footprint in Boston that days because you remember Data General and Digital Equipment were kind of there, the micro companies, the mini computer companies. And so you can watch it happen. And you say, you know, that's a better way to make money than just trying to extract value from rationalizing legacy industrial companies that have been poorly managed.

[A]

Yeah, widgets business.

[G]

Right. And then the other thing is, remember is that people will, in thinking about exiting business, the market will pay you more. For companies that have good. This was an insight those days. It's not now, by the way, in those days to borrow money you had to have assets to back it with, you know, like you Know, inventory, working.

[A]

Capital, you couldn't use the, you couldn't use the company that, and, and then yeah, you couldn't use the cash flows in this, in the software business didn't you know, they maybe had a lease for an office, but not a lot of like servers.

[G]

So what we had to do was teach the markets to lend against cash flow, actually lend against assets. So cash flow lending became this kind of new thing that we had to teach people how to do. And when then once you got that, then you realize that if you had a rapidly growing company like a Microsoft that had an extraordinary cash flow engine, huge barriers to entry. At that point people, when I came in the investment business, people said tobacco is the best business to invest in because I'm serious, because it was very stable, it had stable cash flow, stable pricing and it didn't vary with recessions. I said, let me get this straight. The businesses that addicts and sickens its customers is better than Microsoft? No, I'm sorry, I don't agree with that.

[B]

Right, right.

[G]

You got to look at the modern world and understand that these cash flows are sustainable and these businesses are extraordinary because it doesn't, because the product comes from someone head. You don't have to build a factory to build the thing.

[B]

Fascinating.

[G]

And so we just built this business that got, that had that set of insights and as a consequence we were able to build. So the other idea there was to build a strategic competitive advantage of commanding heights that you could occupy that made it very hard for anybody to compete with you.

[B]

So how does this bridge to some of the debt financing that's going on today? I think that there are a lot of, of folks in the tech community that are very used to a bunch of 20% dilution equity rounds, maybe a growth equity round. And the idea of bringing on a partner like Blue Owl for some massive deal, it just doesn't map to the traditional like tech startup like Path. And yet folks who are trying to understand where AI is going and where the big hyperscalers are working start have to grappling with, with debt and how debt is coming into this generation of this technology.

[A]

Sam Altman has said we maybe need new. He said we need new ways like financial innovation. We need financial innovation, not just technological innovation. A lot of people have kind of shunned him for suggesting that. But I think based on what you've been describing of what enabled this wave of value creation and unlocking the values of these private companies and the value of their cash flows, it can be done. Responsibly it can be done responsibly. Yeah.

[B]

Yeah.

[G]

Wow. That's a really, really good question. And I know maybe we'll have to do a second show just on that because this is a complicated topic.

[B]

Right.

[G]

But it is. Technology is very. I like technology. You know, gotten into it full time for now. You know, 25 years ago ago. I still feel like I'm new to, like I'm still new to golf even though I've been playing about the same period of time. Yeah, the. But one of the things great about is it's constantly changing and you have to constantly adapt your thinking and develop new modes of sort of how of investing. And so this AI thing is come brings us back to the future. That which is. It's technology. It's a technology, software driven, driven LLMs technology enterprise that requires a scale of capital investment that we've never seen before. And that's a really unique kind of challenge. It's one of the things that's drawn me into the kind of investments that I've made there. He reminds me a bit historically, reminds me a bit of when the semiconductor companies went fabless about 25 years ago and the industry split into companies that design semiconductors and basically tsmc. Right. And TSMC succeeded largely because the country of Taiwan was willing to essentially lend them the credit rating, the scale of capital necessary to build a fab that could design these, that could manufacture these wafers with a nanometer scale that they had at prices that were cost competitive that could continue to drive adoption. Adoption of technologies based upon semiconductors was only approachable by a national credit rating.

[B]

TSMC had a backstop, basically had the.

[G]

Backing of the government of Taiwan to go get this done. There's a reason why it's in Taiwan. You couldn't do that in those days that the capital wasn't available to do something like that in those days at that scale for that kind of enterprise. Very similar to this, which is the scale of financing that's. That's required to do to, to build all these fabs. Not fabs, I'm sorry, factories, data centers. I call them factories because I think they're factories manufacturing data now. And what I say to my people, what I say to my friends is America is now the leader in the world in advanced manufacturing because we're building these data centers to manufacture data.

[A]

Yep.

[G]

Right. And that's kind of what it is. It's a massive factory manufacturing LLMs and applications for both training and inference. That's kind of one when the second point would be that people compare this to. So the question is, are we in a bubble that's kind of underlying your thing? You raised right. What kind of bubble is it? And the question you have to make a decision whether or not this is more like subprimes in 08 or more like the Internet in 1999 2000, you know, whether the subprime is just kind of something that's not real, it's going to collapse. And when you're left, you're just left with a bunch of debt and no value there because the home values all went down. I am more in the Internet camp which means that of course there will be companies that will be formed that won't be successful. Of course there will be investors who put capital in bad places and lose money. Of course there will be some number of scoundrels and shysters who come in because money gets moved around and they get attracted to this. Right. But there are one major. You mentioned Blue Owl. The one major difference today between the build up. So what was happening simultaneous with the Internet was being when the dot com companies were being built could say they're the LLM equivalent to today. The fiber optic networks were getting constructed all around the country. The CLECs and those all went to zero and people lost their money on it. The major difference between that and people use that as analogy today and maybe the railroads is another analogy. But the major difference between that and today is every one of these data centers, almost all of them has a counterparty solvent counterparty that is contracted to take all the out output. They're built to suit. Not if you build it. They will come.

[B]

Yeah.

[A]

Yep.

[G]

Okay. Microsoft has I think the world's best credit rating. If you sign a deal with Microsoft to take the off put for your data center.

[A]

Satya is good for it.

[G]

He's good for it.

[B]

Yeah.

[G]

And by the way Microsoft's going to survive if that has a collapse at some point before it comes back again.

[B]

That's a good point.

[G]

It's a very different kind of financing structure. And the last point I would make in this thing finish this is that each of these deals so far as I understand it is done in a way that essentially generates in. In the four to five year period of the deal generates about a two times multiple of money on the cost of buying the GPUs and standing up the data centers.

[B]

Oh, interesting.

[G]

Right. So the contra and they're about four to five year contracts.

[B]

Yeah, yeah, yeah.

[G]

And the output has a now and and then and talking about okay, embedded options and how you've added value those. Right. And the, and then the owner, the of the, the date the GPUs in the data center has an embedded option on the value of the used GPUs which will be worth something. I mean your 5 year old iPhone is still worth something of course, even though people are buying the new ones. Right?

[B]

Yep.

[G]

And so the, each of the model, each of the business, each of the contracts and builds right now has an as a. Has a commercial proposition in it. And when done well, these companies that are doing this, like corporations, we are putting one of building a wall with one of those bricks on top of the other. Yeah. Do you see what I mean? So it's not, it's not analogous at all to the Sealex where they put a bunch of money in the ground and then went to get the customers and the customer for there.

[B]

Sure.

[G]

That's a very different thing.

[B]

That's a really good point. I hadn't considered that. That makes a ton of sense. That's great. Yeah. I feel like a lot of people in tech are, are just struggling to. There's been this narrative for a while that ChatGPT is the new Google. And then you look at how capital consumptive OpenAI will be before profit comes or cash flow comes versus what happened with Google where they were throwing off millions of dollars in cash well before ipo. And the prospectus just looked so clean. This super high margin business, very fresh out of the gate and it's just a very different, different world that we're in where we're delivering is something similar. It feels just like a weapon.

[A]

Partly because. Because OpenAI has to compete with Google.

[B]

Yeah, maybe, maybe. But it's just a different, it's just a, it's a capital consumption. So the model's changing a little bit. Yeah.

[G]

Each wave of technological innovation, companies are created that don't obsolete the company that went before them. They do something completely different.

[B]

Yes.

[G]

Right. And they're sometimes very, very different. Like you know, so you've got, you know, the soft, the Microsoft software was unlike that was the operating system, the applications was unlike anything we had before because we didn't have the PC. And then Amazon was not anywhere near like Microsoft. It was a whole different kind of innovation that was based upon the Internet that was built. And then Google was something new and Facebook was something entirely new. So these aren't companies that say I I'm taking your business, your thing away from you. Yeah, right. Each one is a very different kind of unique unicorn type of business that occupies a niche itself and eventually obsoletes the other businesses because they stop growing. Yeah, right. Like Facebook might stop growing if consumers go to open AI, but it's not because they're going to open AI because it's a new social network work. It's because as a different use case, valuable them Today.

[B]

What's been the biggest learning surprise, sort of update to your mental model from working with coreweave?

[G]

That's a really good question. The pace of change the scale we talked about, I mean the thing that just amazes me is the scale at which this thing is growing and. The rapidity that you have to have in order to act at to be successful at this kind of scale, with this kind of growth. It's unlike anything I've seen before. You saw the adoption curves of, you've seen the adoption curves of OpenAI versus Google versus other things. Right. And it's just like this asymptotic thing going to 700 million customers right overnight, all the infrastructure to support that is like unlike anything we've ever seen before.

[A]

Yeah, yeah. It's still under discussed how much bigger and faster the outcomes can be when you have the Internet as a distribution. As a distribution and engine. So like during, when like, you know, you founded silver lake in 1999, I'm sure you've looked at a bunch of companies that had a lot of potential that if there was already billions of people using the Internet, they would have done very well. And the challenge at that point is there maybe wasn't enough Internet users to support even ideas that were maybe like structurally good ideas just missing enough of a user base. How much time do you spend finding and meeting and backing new managers? I feel like every new technology cycle, the hottest hedge fund of the year is situational awareness or at least on X. And that feels like I imagine there will be more of those. And so I'm curious how much new fund formation you're seeing and what you're most excited about on the GP side.

[G]

Yeah, so I have investments. I don't do venture capital investing per se, so I've got investments in some of the major venture capital funds, you know, in my investment platform and I know all the people and watch what they do. But my business model outside of that is to find a small number of companies where I can put a fair amount of capital and be engaged, helping to create the outcome. See, that's kind of where I spend my time. So I not doing the whole build up massive portfolio thing, I'm picking my spots. And you mentioned the European bank that I'm the lead independent director of. You know, we've got that stock up 3.5x in the last three years since I invested.

[A]

Hit the gong.

[B]

We have a gong here we'd love to hit.

[G]

Right. So you know, there you go. Oh, that's great. Thank you. Thank you.

[A]

I don't.

[G]

But I gotta tell you, I didn't get founder mode, guys. I don't know. I don't know what's going on here.

[B]

Here we go.

[G]

There you go. They didn't get founder mode.

[B]

Come on, you got to do founder mode. For you. The goat is also very important.

[G]

I love it.

[A]

You know.

[G]

You know, I've got children about your guys age. I just love your generation. I love hanging out with them. It's a lot of fun. So by the way, you see this logo here on me, my shirt?

[D]

Yes.

[B]

What is that?

[G]

That's binary code. See that? You know what that's binary code for? 1, 0, 0, 0, 1, 1, 0. What's that? 1, 0, 1, 1, zero.

[C]

Zero.

[G]

It's binary code for 70.

[A]

70, 70.

[B]

What was my.

[G]

It's my 70th birthday logo.

[B]

Oh, very cool.

[A]

There we go.

[B]

Happy birthday.

[G]

Thank you.

[A]

Incredible.

[G]

So as I say, I've got. I've got kid jewelry age and I really love hanging out with your generation. It's been a great pleasure for me.

[B]

Yeah.

[G]

Thank you.

[A]

So we love hanging out with you too.

[G]

You asked me another question that we got distracted from it.

[E]

Oh.

[G]

So what I'm trying to do is find a small number of enterprises in which I can engage, get involved with them at a senior level. In both cases Cor Weave and Santander. I'm lead independent director. It's another term for non executive chairman. There's usually an executive chairman and I'm not non executive. And then really work with the enterprises to build value. Yeah, that's kind of how I think about it. And then I. I let venture capitalists who I invest with and I still invest with Silver Lake be on the rock face every day building these portfolios.

[E]

Yeah.

[B]

What I try to do analogy. I like that.

[F]

Right.

[G]

Well, you're the mountain climber, right?

[B]

Yeah.

[G]

Yeah.

[A]

Right.

[G]

So although you're this your height of a basketball player, I think probably the wrong for.

[A]

I think you're referring to like one of our early episodes where I was joking, remember about you climbing those like.

[B]

Oh, yeah, yeah, yeah, yeah.

[G]

That's what I heard.

[F]

I heard that.

[G]

Yeah. I thought that was just. You were just kidding. Him. Right.

[A]

We were messing around just because. Just the idea of John, a 6, 8 guy scaling.

[G]

Okay. Yeah, I agree.

[B]

Rock climb.

[A]

You could do it. I believe in you. But I'd be. I think that would, that would probably violate. I wouldn't want to be a blonde. Insurance.

[B]

Yeah, I know.

[G]

I definitely want to be above him on the wall.

[H]

Anyway.

[G]

So, you know, so I'm trying to pick my spots and really add some value.

[B]

Yeah. Well, thank you so much for coming on the show. We have to have you back soon. This is fantastic. We could talk all day long.

[A]

Yeah, there's, there's so many, so many more questions I want to ask.

[G]

Okay. Congratulations on the success of the show, guys.

[B]

Thank you so much.

[A]

You're welcome.

[G]

On, welcome on any day when you're on the east coast. Come see me.

[B]

Fantastic. Yeah, we will.

[A]

Thank you so much. I'll play him off with great hanging, Glenn.

[B]

Thank you so much, Glenn for taking the time. We'll talk to you soon. Let me tell you about getbezel.com shop over 26,500 luxury watches fully authenticated in house by Bezel's team of experts. We're going to our lightning round. We got Yogi Goel from Maxima announcing a massive round. Let's bring him in to the TVP and welcome the show. What is your T shirt? Introduce yourself. What do you do? Give us the news. What's the latest?

[F]

Absolutely. So nice to meet you both. Jordy and John.

[A]

Great to meet you.

[F]

Yogi here from Maxima. We are an enterprise accounting platform focusing on waging the war on the month end close process.

[B]

Let's go.

[F]

We are in a short focus. Our AI agents are co writers writing the monthly financial package and preparing the data for the accounting team. We've been around for now 5, 6/4 and helping incredible companies like Scale AI Rippling, SpotOn Press juice. So a lot of companies in both tech and non tech world to make accounting sexy again.

[A]

Very, very on brand to count the time you've been in business by quarters.

[G]

Exactly, exactly.

[B]

And what's the news today? Break it down for us.

[F]

Yeah. So we just raised $41 million in C Series A.

[A]

All right, there we go.

[B]

Congratulations. Very exciting. Explain to me how this plugs in. Obviously there's a lot of folks that have an accounting layer of record, a single pane of glass, an erp, an accounting suite. Do you want to just put plug into that? Do you want to rip and replace that? There's so many different folks eating around the edges creating different solutions. I Don't think anyone knows exactly how the market will play out. But what have you built?

[F]

Yep. So we built a system of action and system of intelligence, which works with any system of record. Okay, so when you go to an enterprise company, asking them to replace the ERP is like asking them to do a brain surgery. I was a rubric and I would not agree to that. And we say that, hey, your system of record is where your data should eventually sit. We help with automating the human work of grabbing the data from upstream systems, doing the manual, doing our agents do the automated work, and then eventually finding anomalies and errors. I don't know if you're following, but last year was the maximum number of companies in the US which had material misstatements and up to 40% stock drops.

[B]

Stock price drops, most mistakes from accounting, specifically last year. That's not good. Hopefully we can fix that. I have one last question and then we will let you go. I want to know, give me some examples of where the current crop of AI models really excels in finding these types of problems. And then where do you want to still leave the human in the loop? Where do you want the human? What's the really intractable problem that maybe we'll solve in a few years with the AI, but for now you'd leave it with the human?

[F]

Yeah, you gotta have somebody to fire.

[B]

Who's the last guy in the accounting office, I guess is the question. But you know, I'd love some examples of problems that really excel for AI and problems that are maybe more intractable.

[F]

Yeah, look, I'll just start with saying we are not going after the human labor salary. We are going after errors, inefficiency and pain that I personally face both as an auditor and as an accountant for 20 years. There are not enough accountants in the world that you can truly hire for the amount of work that's there. So in terms of where AIs are very good at today, they are very good at taking a defined set of instructions and following things over and over again for a variety of transactions. Provide avoided. You give them deterministic operators which we have built, that they will only use those tools and then come up with the right answer. So we are using this hybrid approach where agents follow maxima tools to come up with the exact same answer. And so when Deloitte and EY comes knocking, looking at the work that maxima produced, they will do two plus five. And the answer will always be seven. It will not just be 15. So that's one thing we determined really well. Second is really good at finding anomalous behaviors and errors that might happen because it is looking at millions of transactions over time within the company. It can just see that, hey, your legal bill used to be $50,000. Suddenly it's $500,000. Turns out Jim had a late night and he had one extra zero and that's why it went up.

[B]

Yeah. And I mean, artificial intelligence has been used in fraud detection for years and years and years.

[D]

And.

[B]

And so applying that sort of heuristic based, stochastic based, more, less deterministic computing, more probabilistic computing makes a ton of sense there.

[A]

I love the positioning around pain and errors. You've got to talk to the venture capitalists who are yelling loudly to anyone that will hear, we're going to replace all labor. Give me more money to replace labor. It's like you can show the sort of optimistic, like, positive.

[B]

Well, thank you so much for taking the time. Congratulations on the massive round. We will talk soon.

[A]

Yeah, great to meet you. I'm sure you'll be back on soon.

[B]

And have a great rest of your day. Thank you.

[A]

Great show too.

[B]

Let me tell you about eightsleep.com exceptional sleep without exception. Fall asleep faster, sleep deeper. Wake up energized. Our next guest is Sam Jones from Method. We will bring him in from the Restream waiting room into the TVPM Ultra dump. Sam Jones, how are you doing?

[A]

Welcome.

[B]

Good to meet you guys.

[I]

Great to see you again.

[B]

That sound effect kind of just. I don't know, it doesn't have enough of a crescendo for me. We need to work on that one anyway.

[A]

We're working on our drum roll.

[B]

Thank you so much for taking the time to hop on the show. Please introduce yourself, introduce the company, tell us what the news is today.

[A]

All right.

[I]

I'm Sam Jones, the CEO and co founder of Method Security. Our mission is to deliver cyber resilience to the US government and critical enterprises. Think of what we do do is building the command and control layer for autonomous cyber operations across defense and offense. And the news today is that we are announcing our 26 million combined seed and series A investment from Andreessen Horowitz and General Catalyst.

[A]

Incredible.

[B]

Very good.

[A]

Preemptively last night.

[B]

Happens but we got you. Happens in the Ultra.

[I]

You'll have to raise more money to pay you back for that.

[A]

I'm sure. I'm sure you.

[B]

I would love for you to get me up to speed on how you're thinking about that story in the wall Street Journal about anthropic. I'm sure you know the one.

[A]

Is this AI on AI violence?

[B]

Exactly. Is this relevant to your business? Are you building a solution to that or do you even have a comment on it or anything? Can you just get me up to speed?

[I]

Highly relevant. And that's kind of the moment that we've been, we've been building for, for a couple years now. Like we've known this is going to happen. AI is effectively, you know, taking at the limit, taking cyber offense to infinity and taking the cost to zero. And this is bad news for good guys, bad news for the defenders, as our adversaries are essentially eliminating their requirement or limitation on human headcount. So what we do is essentially allow organizations to safely become the threat, to test their own defenses before some adversary does. And the best offense is the best defense has always been a notion in security. But AI is really the unlock to do it at scale. The hard part is you need to do so safely, ethically, legally. And that is the infrastructure that is like needed to do and that's what we build specifically. So like in that report, it's almost like no news to anyone in the security trenches. Like obviously this has been happening, happening. Obviously that wasn't the first.

[A]

Yeah. What are some other. Without, without naming names or any details, like on, on the individual companies that were attacked. Like, like I'm assuming when you see, anytime I see a report like that, I'm like, okay, this must be happening like a ton. And just a lot of it just never, never hits like headlines. But what are some of the most like kind of common strategies that bad actors are using today in the context of, context of AI to carry out whatever their goals are?

[I]

If you think about pre AI malware, it was already autonomous, but it was basically reliant on like if, then decision making. What AI basically allows it to do is like a broader non deterministic path planning that allows it to harness a multitude of tools, thus do a lot more damage. That's what's different now. And I guarantee you the most sophisticated actors are not using vanilla Claude code to run their operations. That's ludicrous. Our adversaries have better models at home that they make themselves, that they're using, that we have no telemetry on. And so they're essentially using it to scale and speed up their operations. Which for us and why we have a national cyber resilience urgency moment on our hands, is that all of these exposures that we've left out on the Internet and In our, in our enterprises are now easy takings for these types of attacks. And that's essentially why it's so urgent that we focus on resilience.

[B]

Can you talk to us a little bit about traction? What unlocked this $26 million fundraise across these two rounds? Are you doing? Yeah, just walk me through how you actually show progress in when you're building a product, but you're also trying to do deals with the government. That can be very difficult. What does progress look like?

[I]

So we are deployed in production with a number of organizations to include the Department of War, US federal government and Fortune 500 organizations. That's probably the biggest hallmark of traction. And we're doing so across defensive and offensive use cases that get to the heart of resiliency. So, so that's I think the unlock that we were able to do that. And we've had this hypothesis and mission from the beginning that in order to secure what matters, you need to be dual use. And we set out to basically pick what is the most intense, hardest government customer you could go after from the beginning and what's the commercial equivalent. Those were our first two customers. And then basically just continuing to build on that from the adversary's perspective. They do not just discriminate between public and private, and neither do we. And that's why I think the ultimate game changer solutions will come from dual use companies like ourselves.

[B]

Now I'm thinking about the hardest to hack Fortune 100.

[I]

And it's not necessarily always the hardest to hack. A lot of times it's the hardest to sell to, hardest to deliver for.

[B]

Okay, yeah.

[I]

Think about the government done accreditation, deployability, like interoperability, huge technical challenges. That's why startups would never dare their touch there. But when you think about what matters, they are what matters and that's what we've built this company to serve.

[B]

Yeah.

[A]

What were you doing before this again?

[I]

So I started my career actually seeing this problem firsthand at the U.S. air Force. So I was a cyber operator in many ways. We're building the tools that I wish I always had. I joined palantir about 11 and a half years ago, you know, pre product and building out both their Cyber, Commercial and DoD business. And then I was also at Shield AI pretty early. And so you can kind of think of this company as we were the users. My CTO and co founder also started his career at nsa. His last name is Hacker by the way, if you want.

[B]

Oh, there we go. Get that. And then also I want an overnight success for overnight for being in this industry for 15 years, he had no.

[I]

Choice other than to work at nsa. But we met at Palantir and did great work together. But we're combining our knowledge of like we were the users, we know how to build hardcore, hardcore software and dual use businesses and then we built AI before you know, it's become a meme and certainly in no fail scenarios which I would group security in for sure.

[B]

Yeah, this is very cool.

[A]

Very bullish.

[B]

Yeah, extremely bullish. Thank you so much for taking the time to come chat to get the update.

[I]

Sorry my, my background wasn't as good as Glenn's mahogany.

[A]

I showed this startup you brought a.

[B]

Wood which is, which is good. Yeah.

[A]

Thank you for bringing. What's the biggest fish you've ever caught?

[B]

Yeah, yeah.

[I]

Probably a nice, nice walleye, I'd say.

[A]

Okay, there we go. Good answer, good answer.

[I]

Midwest shout out.

[B]

Thank you so much.

[A]

I cannot wait for the bees. Congrats on all the progress.

[B]

Yeah. Have a great day.

[A]

Cheers.

[B]

Bye. Let me tell you about wander.com, book a wander with inspiring views, hotel grade amenities, dreamy beds, top tier cleaning and 247 concierge service. Our next guest is already in the restroom waiter room we have Ali Madani.

[A]

Welcome to the show.

[B]

How you doing? Thank you so much for taking the time to come chat with us. Please introduce yourself, introduce the company, tell us what the news is today.

[C]

Sure, absolutely. My name is Ali. I have a PhD in machine learning from UC Berkeley. I've been working in the space of biology and AI for almost a decade now. Previous to this, I led a moonshot at Salesforce pioneering language models for biology. And what started out as a purely scientific endeavor to develop, develop transformer models for sequence generation has led into profluent specifically where our mission is to make biology programmable. And I'm happy to kind of break that down.

[B]

Yeah, yeah. I've seen obviously there's a ton of just like momentum in the space AI. Curing cancer is like a buzzword that a lot of people are throwing around. How are you thinking about concretizing what you're actually, actually how you're trying to fit in? Are you a tool? Are you a drug maker? Is it uncertain? Like who are your customers? How much are you in science project world? You could be a nonprofit in another era versus you're ready to commercialize, you're going to market. Not that there's one path that's wrong or the other, but I'd love to know how you're thinking about the Business right now.

[C]

Totally, yeah. I think there's a lot to unpack there. I think the meme that came to mind specifically, I don't know if it's a south parking or otherwise where it's, it starts with like build something and then there's a dot, dot, dot, question mark and make profit.

[B]

Yeah. Step one, Step one, you know, make biology programmable. Step two, bro, down with your boys. The step three, profit.

[C]

And I think a lot of folks, you know right now there's an incredible amount of excitement around AI and it's kind of like step one is make a chat bot, for example, and then it's question mark and then solve disease or cure cancer specifically. Whereas what we're actually trying to build here is actually tackle on the disease head on specifically. So what we do is we build language models. So the same language models that have enabled GPT 2, 3 and 4 and chat GPD specifically, these incredible models and algorithms that can learn on sequences, what we can feed instead of words in a sentence is actually amino acids that are strung together to form a protein. And why that's actually important, why making biology programmable. Maybe to take a step back like people usually shut off their brains when it comes to biology. And when it comes to like rockets landing on a platform in an ocean, we're amazed, right? And that makes sense, right? Like it's, these are man made machines, we can see it, they're incredible. But honestly, biology is not that much different. There are these molecular machines called proteins that enable us to breathe and see they're responsible for everything in human health and disease, disease. And also they sustain the environment involved in daily products like even our detergents to begin with. And how let's actually stick to drug discovery in particular, how we've gone about finding these solutions, these molecular machines that we utilize day in, day out has actually been through random discovery. So you know that middle school example of Alexander Fleming coming across penicillin, right? He had a petri dish, they molded for example, and then he found the advent of antibiotics. And now, now after you get a cut on your skin, for example, where bacterial infection happens, it's no longer a death sentence, right? That actually is not the exception, it's the rule in which we've gone about finding life saving medicines. Even fast forwarding to today, CRISPR Cas9 was actually found in a Denisco yogurt facility where people found these interesting bacteria, doing these interesting characteristics, having these interesting characteristics and we've taken a molecule, plucked it from nature and then crammed it within human therapeutic applications to actually save lives. And honestly like to put this really in rudimentary terms, that's kind of absurd.

[B]

It's almost caveman.

[C]

Like in terms of our technique, our techniques that we have and methods that we have available for us for drug discovery. And what we're trying to do is actually move away from random discovery and finding a needle in the haystack and relying on nature altogether and using AI to design bespoke medicines from scratch. And that's like, that's our mission to really gain control and mastery over biology and perform bespoke design. So in terms of your question of like where are we with respect to, you know, is this just a science project or how's the commercialization looking specifically, I would still say we're in early days, like the equivalent of GPT eras of like maybe GPT1 or GPT2, but we've already seen an incredible amount of traction. So we had this project called Open CRISPR specifically where we took is we took these, these language models, trained on gene editing proteins specifically and generated a novel protein from scratch called Open CRISPR1 that thousands of people use now in pharma, large pharma, small biotechs, academics, industry users and scientists as well. And over thousands of people use this over worldwide, worldwide today. And I think that's like, it's amazing to actually see us solving problems today that have lead to commercial traction and that we have partners, partners both from therapeutics to diagnostics to biomanufacturing, even agriculture that are utilizing today.

[A]

Can you talk about like how you create feedback loops as a company? Because you know, there's no shortage of people in AI that talk about the opportunity of like curing various diseases. Many of them aren't saying that from the standing in an actual lab, you are standing in a lab that makes me more excited about what you're doing because you're not just kind of, you know, like you're not just saying like oh like the next version of the model, we'll just do this like don't worry about it. It's like no, we're going to run a lot of experiments. But yeah, talking about, yeah like you know, using AI to learn and generate potential approaches but then actually bringing it into a lab setting.

[C]

Absolutely, yeah. We operate within a pre training and post training paradigm within proteins similar to NLP and natural language processing as well. So the pre training step really involves similar to how we have all of the Internet that we can scrape from and can learn these underlying principles and Grammar and semantics as to what makes human generated texts. We've actually collected a tremendous amount of data of proteins that have naturally evolved through nature for selective reasons. Selective pressure is an evolutionary, evolutionary kind of pressures that have shaped those proteins specifically to make a functional protein. And just to put that into context, AlphaFold 3 was trained around, it was exposed to around 2 to 3 billion proteins. What we've actually trained to date so far at proflant is over 100 billion proteins. And to put that into tokens, that's over 20 trillion tokens. Exactly. So there's an incredible amount of data for pre training purposes that utilize. And then what you see behind me as well is the data that we're doing, the assay labels, labeled examples, meaning actually taking protein sequences and then measuring their function, not just in vitro and test tubes and petri dishes, but in human cells and relevant cellular contexts and seeing how well they actually perform. And we could feed that back into our models. So I think that's, you know, the future is really an integrated future where you're building frontier AI models and having the, the closed loop specifically with respect to the wet lab, which is what's behind me today, to actually test these and feed them back into our models to get better and better over time.

[B]

Well, congratulations, I want to ring the gong for you.

[A]

And by the way, Gerstner and Bezos, I mean, how much? Potentially the coolest cap table, the cap table of the year.

[B]

How much was the deal?

[C]

Yeah, absolutely. Yeah, it's $106 million. And I think what's more important, what's more important than number are these legendary investors that we have. I mean, Jeff Bezos is a legend. He's transformed industries. And I think what's exciting for him and for us as well is that biology is the next frontier for AI specifically that will have tremendous impact and really honestly is the most important quest in our lifetime.

[B]

Thank you so much for coming on.

[A]

I'm sure you will be back on.

[B]

Very soon and congratulations on the progress.

[A]

Great, great to meet you.

[B]

We'll talk to you soon.

[A]

Talk soon.

[B]

Have a good one. Our next guest is already in the Restream waiting room. We have a hard stop at 2. We gotta run, we gotta hop on with New York. So let's bring in Amit from Luma AI with some massive news. How are you doing? It's been too long. Great to see you again. Welcome to the show.

[A]

What's happening?

[B]

Give us the news. What happened today? What happened? Break it down for us.

[E]

Yeah, so we did two massive things. One Luma raised a 900 million Series C.

[B]

Okay, what was the second one?

[A]

I'm sorry, I'm sorry. What? Like was there not another 100 million lying around? You couldn't, like, you couldn't. You gotta. You're gonna make us wait for the Luma one billion dollar round. Come on, dude.

[E]

I'm very happy to accept friends and family checks.

[B]

Okay? Oh yeah, if you got 100 million and for your fort from your friends and family to round out, that'd be fantastic. And then what? Yeah, what is the second thing?

[E]

And yeah, the second thing is basically along with humane, which is a, you know, this AI company being built in Saudi Arabia, we are building a 2 gigawatt compute cluster that we are going to use to train, you know, multimodal AGI.

[B]

This is the big news. This is actually the big news. This is much bigger. This is much more important. This is. That's right. We actually need the compute.

[E]

So the TLDR of what happened here is basically, so far LLMs and LLM Labs have had the right resources and multimodality world simulation. These problems actually were side projects for most companies. Now there is a lab and there is a company in the world that has this level of resources and is going right after AGI that can help us in the physical world, AGI that can help us simulate and generate the universe. So I think that's actually what happened.

[C]

Basically.

[B]

Amazing.

[A]

How do you think about, how do you explain the scale of 2 gigawatts? Because it sounds like 2 is not a big number.

[B]

And is it 2 gigawatts because you're expecting 2 gigawatts worth of inference? Or do you need a particularly big cluster for some sort of pre training run that you're planning on doing?

[E]

So it's both inference and training. But inference is actually so, you know, majority of the workloads as we go forward, right. Like you know, as AI deployment goes forward, as we mature from just text only models to models that are able to like, you know, generate videos, models that are able to explain things to us in video, what's going to happen is most of the workload and tokens will move to video understanding and video generation. And video tends to be computationally much more intense than language. So we need this level of compute to be able to deploy this technology and to be able to train. But this is mostly inference. Honestly, even today, Luma's inference to training compute ratio is 2, 1 already. And we are seeing that ramp actually growing further and further and further. While we do deep research and train some of the largest models in our space, space inference is the one that is actually taking off.

[B]

Okay, react to this take I got from someone who's also building a world model, generative world model. He told me that he believes that it's more likely that AGI something fully paradigm shifting emerges from world simulation than merely scaling up. Next token prediction. GPT. 5, 4, 5, 6, 7, 8, 9. Getting away from text is actually somehow foundationally important to the next major breakthrough in AI as we know it as a whole.

[E]

I think getting away from text is a mistake. We need to build models that combine audio, video, language and image. So we need to build things that operate like human brain. If you remove text, you remove the entire interpretation of the human logic and reasoning and those kind of things. So we need the physics that comes from video, we need the causality that comes from video, and we need the text which actually makes all of this interpretable and logically connected across the world. So no, I think what we need to do is build these joint unified models. But on the simulation side I agree and I think that's really, really important because does think about robots or think about systems, how they would operate. They need to be able to understand the world. So this is world understanding, which is where world models are going to be very, very powerful and multimodal models are going to be very powerful. And second is simulation being able to run the process or idea in your head and drawing out conclusions. What If I go 20 meters this way, would I fall? This is a simple question, but as robots become more general purpose and day to day in our lives, we need this level of simulation capability in their heads. So generative models give you simulation capability. Simulation is extremely important. Second thing is LLMs are really good at things that can be represented more or less fully in text, code analysis, these kind of things. But when we think of the physical world especially acts like designing, manufacturing these kind of topics. One of the things we think a lot about at Luma is, is manufacturing of a jet engine, right? Or manufacturing of a rocket engine. These are one of the most complex things humans do and it takes a decade to build one. Imagine having models that are able to run these physical simulations and get to an answer. It's not about the visuals, it's about getting to the right answer. People do that in cad, people do that in software today, but it's very inaccurate. But if you're able to build models that can accelerate building of these complex systems, humanity has a chance at building Better and better. Better things for ourselves, for our planet. So that is why simulation is really important and that's why multimodality is really important. That text is just the first step to. Text is like 1990s Internet. Then we got images on the Internet, then we got videos on the Internet and today videos is the Internet for humans at least AI would not be any different.

[B]

Yeah, last question from my side. What is the actual timeline for building a 2 gigawatt cluster?

[A]

Yeah. And where, where will the majority of the infrastructure be?

[B]

When can I see it? When can I go inside? I can be trusted.

[E]

So some of it already exists. So by the way, we are building this in partnership with Humane in Saudi Arabia and today it was announced here. So we are in D.C. right now for the U.S. saudi Investment Forum.

[B]

Oh no way.

[E]

It was announced by President Trump and conference Mohammed bin Salman. So the data center is going to be built in Saudi Arabia. Quite a lot of capacity is actually already available and Luma is actually an active customer and using that today. But the deployment of 2 gigawatt is going to take time. That's an absolutely colossal amount of power and infrastructure that needs to be built starting with 2026. And currently we believe that by end of 27 or early 28 we will have majority of the capacity at hand and we'll go from there.

[B]

Fantastic. Well, thank you so much. Incredible progress while you're traveling to come chat with us and break down what's going on. Congratulations on the amazing news and good luck with the next phase. I'm sure there's a lot going on.

[A]

Next time you call in, come call in from Saudi.

[B]

That'd be amazing.

[A]

From the desert, that'd be amazing.

[E]

The first time actually I was on tppn, I was in Saudi.

[B]

Oh, no way.

[A]

There we go. We already did it. We'll check that box next time I want one of those four by fours that you know, call in from the desert from Humane. I want to be live on the ground with you. Amazing. Great to see you again. Congrats on the progress.

[B]

Thank you so much for jumping on. We have to hop on with New York first.

[A]

We have one post we gotta pull up and it's a post that I made right when I saw that Nvidia beat earnings.

[B]

They have traded up. The Stock is up 3.3.91%. Massive.

[A]

It is at the very bottom.

[B]

There were signs. This is your prediction, one of your, one of your many predictions.

[A]

But this is all the only data.

[B]

This is the only data you need to know. You know. You know, you said this. I think he's going to beat earnings because he's drinking beers. And Ev was like, yeah, you belong in a pod shop. And he was saying it, like, sarcastically, like, you know, to be in a real hedge fund, like, you have to be much more quantitative than that. Turns out you don't. Turns out the vodka analysis, you gotta.

[A]

Take it all in.

[B]

Gotta take it all in. Thank you to everyone for tuning in and watching our show. Leave us five stars on Apple podcasts and Spotify, and we will see you tomorrow.

[A]

Global economy continues.

[B]

Continues. The party continues, folks. White suits tomorrow.

[A]

Gabe in the chat. Gabe's getting drunk. Drunk responsibly. See ya.