Meta Connect Reactions, Zuck's Beer Pong Controversy, New Bentley EV | Shalev Lifshitz & Romi Lifshitz, Leif Abraham, Diogo Almeida, Hooman Reza Nezhad, Jordan Nanos, Erika Alden DeBenedictis

You're watching TVPN.

Today is Thursday, September 24th, 2026. We are live from the TVPN UltraDome, the temple of technology, the fortress of finance, the capital of capital. Let me tell you about ramp.com. Time is money. Save both. Easy to use corporate cards, bill pay, accounting, and a whole lot more all in one place. Very happy to be sponsored by Ramp. Thank you to the whole Ramp team. Never gets old. Also, thank you to Kris Kringle. Because he's coming by in just 91 days.

That's right.

The TBN Road to Christmas is continuing day by day. We're 91 days away.

That's right.

Get ready, get excited. Christmas is coming. It's around the corner.

Great weekend to pick up a tree.

If you can find one. And struggle to keep it alive for 3 months. Yeah, maybe.

I mean, some operators I know are actually cycling through multiple trees on the Road to Christmas.

Some operators. Yeah, it's like that Chinese proverb. The best day to plant a tree, 10 years ago. The best day to— the second best day is today, you know, get started. The best day to get a Christmas tree was 10 years ago. The second best day, today. No matter when it is, only 91 days away. So get ready. Anyway, Meta Connect yesterday, late. Zuck likes going late. 4 PM Pacific, 7 PM Eastern for the keynote.

Well, because he wanted to roll straight from the keynote into some beer pong.

Oh, that's true. We will get into that big hot debate on the timeline over his beer pong strategy. But really, really tight keynote, very well organized, very clear, didn't overstay its welcome, did 3 live demos. Last year there was a hiccup with a live demo Wi-Fi, and I believe Boz triggered everyone's agents by using the wake word. So this time he deliberately stayed away from using that wake word and used like the name of his agent. So he didn't cause a lot of chaos in the— In the dome or in the studio, wherever they were. But a bunch of great announcements, really good pacing, really tight 1-hour keynote. And it was fun watching Zuck, like, sort of do the live demos and be like, one down. Like, this is high stakes. Like, it was clearly high stakes for him. But he got through it all, no mistakes.

Yeah, he had a funny line. He was like, and we love live demos here.

Yeah. It's funny because, like, he totally could just not do live demos, but he, like, wants to.

So he pushes through it.

No, I think you Do you?

I think you absolutely need to. People have bought so many random hardware devices that just haven't met expectations.

Yeah.

And I guess there was that—

Hardware, if you want people to buy it, you have to do a live demo.

Yeah. There was a demo of a Gemini model running like 2 years ago that was like sped up in post and that faced a lot of criticism. And so, I think, yeah, just demoing things as they are available in the real world is very valuable. And I think in general, like, the thesis for the products was like meet people where they are delivering tomorrow's technology with today's constraints. And that's just what I kept seeing as a theme was like, what do people want? And of course, the best— it basically created a huge dichotomy between like the crazy super intelligence RSI, ASI, P-Doom, X-Risk crew and then— just like Muse. It's a helpful agent. Niraj Agarwal summed it up well. Anthropic marketing. We built a god only we can control and gave it a wet lab. Muse marketing. Hey, girl, I know you have dozens of unpaid parking tickets. Let's get that cleaned up. And it's funny because it's actually working. I think that if you went back to a couple months ago, The consensus was like popularity of AI is turning, it's deeply unpopular, you need to just be quiet or not lean in or just like apologize like what a lot of the lab leaders are doing, just being like, yeah, it is bad, I'm sorry. But Meta has taken a very different approach. They're not fearmongering, they're just like, it's cute, it's lovable. And they actually are successful. I mean, the New York Times, Yesterday, Jensen and Ezra are duking it out over like X-risk and acceleration. Do we need to slow down? Meanwhile, the New York Times published like a really positive review of Muse. They like, they love it. And like, they're a tough cookie to like win over basically. So success there. And it's interesting because it's been, there's been a few strikes for Meta over the past 2 years. The vertical AI slop app Vibes quickly disappeared. I don't even know if it's available anymore. It might have, been quietly shuttered or tucked behind a couple of different tabs. The core Meta chat app that was sort of just like a Gemini, a ChatGPT clone, I don't think that ever really took off over, over the last year.

You have massive spending, clearly massive focus from Meta, but the strategy was at most times completely incoherent.

It was scattered.

It was not legible to the markets. It was not legible to It was do everything. Yeah, you had—

But nothing was really accelerating.

Muse Code, you had them buying and then having to undo the Manus acquisition.

Oh, yeah.

The Meta vibe, the first thing— and again, this was an unforced error— the first thing that they did post the whole talent raid was release the short-form slop video app, which wasn't even—

It was Midjourney.

Yeah, it was basically a Midjourney wrapper. And everyone was like, okay, this doesn't align to what you're putting out, which is this personal superintelligence narrative. And so, really yesterday, everything felt like it came together really nicely, tying in all of the hardware investments, tying in personal, making personal superintelligence feel a bit more real.

Totally.

AI that hustles for you. I think that's smart positioning. But it really felt like things came together for the first time.

Totally.

And I wouldn't necessarily go so far as to give them full credit as— like, if you were really going to give them full credit for, like, this was the master plan all along, then there's a bunch of other stuff you wouldn't have done unless you were just trying to confuse the market.

Yeah.

If you're trying to just confuse people, are we doing coding?

Yeah.

Are we doing a knowledge work harness? Are we doing short-form swap, right?

And and so I think you're not giving him the benefit of the 4D chess.

No.

Okay.

No no no.

Tricky.

But going outcomes razor.

Total credit.

Total outcomes 4D razor.

Total credit to them for like figuring out how all the pieces connect together.

Yeah.

And then putting that in a one-hour presentation that actually like makes sense.

So what are the pieces that come together? Why does it make sense for Meta's business? Why are they seeing success here when the models haven't jumped to the frontier? The benchmarks have always been a little— there's been a lot of confusion on are they going to sell compute, sell tokens, sell, you know, go after the coding market, desktop agents, all these different things that they were testing. This seems to be working. And I think there's sort of 4 key points that come together to make this make sense. So first is like they have scaled and secured systems for years. WhatsApp encryption is a great example. They have a lot of compute, and so they can offer a VM for millions of people that is secure and also well-resourced. And they won't be losing sleep over costs because they already have a lot of the compute built out. And the question was like, when will the demand come? Are those GPUs sitting idle? Are those— is that infrastructure getting used efficiently? They're building a lot more of it. Do they need to? Well, they gotta find a product product to actually distribute all that through, and it seems like they have. 2 was they have the distribution to continually educate people about features and benefits of Muse through big ad placements across the family of apps. I think that there is the, there is the world where they've been able to funnel, uh, what, like half a billion people over to Threads. It's not a particularly differentiated app. Obviously it's a different, different community and it's, and it's done well, but it's a different challenge because with With Threads, you can just show someone a text post. People know what that looks like. It's more content. They go over there, job's finished. Once you show someone, all you need to show is that their community exists over there. They're scrolling Instagram, watching basketball highlights, and you say, hey, there's some basketball commentary going on in Threads. You hop over there with all the AI agent products. The capabilities change so rapidly that you have to reeducate the consumer on what it's capable of. Like, can it actually deal with parking tickets? Like, if it can, you need to confidently tell someone that. And that capability might change. It might not work one week, and then the model updates, and then all of a sudden, you can do that the next week. And so ChatGPT came out, and it was like, it can answer Google searches. It can do your homework. Then it was like, it can write some code. And there's this constant, like, reeducation of the community. Obviously, the early adopters, the enterprise, the business, the startup community tests everything. Yeah. But consumers don't. Consumers just— they need to be told the new iPhone has an ultra-wide camera. The new iPhone has Memoji. Even though Memoji was not a wildly popular feature, Apple told everyone what it was. We know what it is. Everyone sort of knows that concept because they ran billboard campaigns. And Meta has the ability to actually go and tell everyone what Muse is capable of and run an ad campaign across the whole platform to get people back up to speed on the latest and greatest feature. The third is that no one needs their—

The system says, deal with parking tickets. Hack me out of it. Then we are talking.

Well, no, no. So the thing is that they are behind on the frontier. They're not dominating all the benchmarks. But it actually doesn't matter because no one needs their personal AI agent to solve a Millennium Prize math problem. Or secure their enterprise infrastructure from cyberattacks. Like, that's not a requirement. The model is definitely good enough for a lot of this. They obviously will increase the model and the capabilities, but it's not what consumers worry about. Consumers don't care about benchmarks. They don't understand them. The benchmarks are often hacked or not indicative of the feel of using a model. And so it leans into the capabilities. They have near frontier intelligence. They have good enough intelligence. And so if it's good enough to do this, you just roll out the capability and the harness is what matters. The integration, the distribution, that's what matters. So they actually didn't launch Watermelon. They've been teasing that for a while. There's been some leaks. It totally could have happened here, but they made the decision to either not rush it because maybe it does just need more time. But also, even if it was ready, I don't think they should have launched it here because—

Why?

MetaConnect's not for that audience. They don't care. Yeah, the AI research—

The watermelon might scare people.

Yeah, yeah, yeah. You probably don't want it to be too powerful. And then fourth, they of course already have a business model that's heavily aligned with this type of use case advertising. We've talked about this idea that brands would love to sponsor agentic workflows. You open up a personal AI agent and it says, hey, the new Nathan Fielder movie is coming out in a couple months. I saw that you were scrolling Reels. You seem to be a fan of this. You shared the movie trailer with your friend. Do you want me to get tickets, text some of your friends, share it on WhatsApp, share it in your Instagram DMs, and actually make that happen, send a calendar invite, do everything necessary to take care of that? And I think movie studios are just one example of a business that would happily pay for placement, for extra attention around that particular workflow that they can monetize through the traditional pathway. And then, of course, small businesses will buy pixels anywhere. So if you open it up and there's a full-bleed ad, that's fine. People will pay for that.

Yeah, the other kind of examples that get pretty crazy is your personal agent, for example, knows that you have— Home insurance that's set to renew in 3 months. And it's like, do you want me to go get 3 competing bids before we renew?

Yeah, totally.

And it's like, yes, right? And that's like an opportunity for some type of paid placement as well.

For sure. So the vibes, not the app, but the actual vibes at Meta internally, how are the employees feeling? I feel like this is a turning point. I feel like if you're at Meta, you're like, OK, finally there's a thing to align around. We have a strategy. As long as I do something that's in support of Muse and what's going on with the personal superintelligence mission that the CEO and founder has laid out, I'm good. Even if I'm on Facebook Marketplace team and I feel like I'm going to get steamrolled because of the Nikita Beer take that everyone's going to be negotiating and lowballing on Facebook Marketplace, well, the Facebook Marketplace team, they can say, okay, well, now we're going to build tools for sellers to make that not annoying. And so there's more work to be done and it's more aligned with the overall mission. The vibe should be good. At the same time, The Verge ran a story that said the vibes are internally way off still at Meta, that people are upset about the scattershot nature of the approach, that they're still reeling from like the metaverse misadventures. But I feel like that was like the last hit piece you could write about Meta. And I think things are turning around.

It's hard to be feeling that bad when the stock's up 38% in the last month.

Yeah. And so I think it's pretty easy to align all the efforts across what's happening, personal superintelligence. And even if you were on the glasses team, there's a whole narrative of, well, you were trying to build VR. You were in the division that was building the cool Oculus stuff. It's evolved. And now it's kind of just a GoPro on your face. There's been some negative, oh, these are just for— people to spy on each other, blah, blah, blah. But now they even have a version without cameras, and it's very aligned to what the users want. And so the top-line hero initiative, even if you're just on the glasses team and you're like, oh, they're building the data center, and that's not really anything to do with me, or they're working on a coding model, and that's not really relevant to getting more people to wear the— The glasses. The Oakleys or the Meta Ray-Bans. Well, now you are aligned. Now you are on the same team. You're rowing in the same direction. And so I would expect the vibes to get better, I think. And yeah, this is all good news, super reflected in the timeline, but I'm even more excited about the VR glasses. That was the highlight for me. It was the one more thing. Hoping that we'd get a new VR headset last year. We didn't. And no one was speculating that this would drop this year, but it did. And it's very, very cool. So the headline is $1,300 VR glasses. It's 100 grams, which is the weight of a pack of deck of cards. And if you go back, John Carmack was the CTO of Oculus working with Palmer Luckey. He came into Meta through the acquisition. He's the creator of Doom, legendary computer scientist, now running Keen, working on AI. He wound up leaving. Very, very disappointing to not have him realize his vision there at the company. And he had been pounding the table at Meta. And actually, there were some leaks and there was like, oh, this is like a source of frustration because his goal and his mission was to get to 100 grams for the headset, very light. At $100. He said it needs to be cheap and light and then it can go everywhere. Then you can see mass adoption and then you'll get the developer ecosystem, etc. Obviously, $1,300 is not $100. There's been a lot of inflation, but not that much inflation. But it is half the price of the Apple Vision Pro. Basically, you're in a more accessible territory, half the price of specs. Oh yeah, it's half the price of specs. And it's, uh, I think it's less than 1% the price of a Ferrari Luce. So, yep, there's that. If you're, if you're in the market, you can just put that on and imagine that you're driving an electric Ferrari.

Um, but, uh, uh, the glasses and the car sound the same.

Oh yeah, they're both very quiet. Um, but he, but he pitched it as 3 key things. It's a cinema, it's a workstation, and it's a gaming console. You can play VR head— VR games on it.

I almost felt like they were listening or just like same line of thought as Ben Thompson because they're rolling out a bunch of like 100 experiences, it sounds like, a year where you're going to be able to get some entertainment, live entertainment value from your headset.

And specifically sports.

Sports.

Sports.

Sit anywhere in the stadium. Yes.

Yes. Yes. And you can pick your seat. And I don't think they will be doing the crazy frenetic editing or the cutdowns. They have a great deal with UFC. Obviously, Dana White's on the board, but it seems like They also rolled out partnerships with so many other leagues to get that there. They also put the compute and the battery pack on a puck that you wear on your belt. I love that design. It was very controversial, very un-Apple when Apple Vision Pro did it and everyone was like, Steve Jobs never would have stood for this. Well, like, that's the reality of the technology. It doesn't make sense to put a battery on your face. And so put it in your pocket. I've never had a problem with the way the Apple Vision Pro works with the tether. And they did a good job in the promotional videos of hiding the puck. You can't even see it there. So they're clearly aware that you want to forget about it. And I think in the best case, you actually can. So very good there.

The interesting— The other dynamic tabletop typing. So if you have a flat surface. In front of you?

Bad.

Bad?

Bad. No one wants to do that.

German says it's insane.

I don't think— I don't think it'll—

I trust him with my life.

Okay, okay. Maybe. I think people will either connect this to a computer and just be using their computer or they'll just be using voice. I think that that's like a half measure like the writing and I think it won't be— it'll be 99% of the way there. So in a demo, you'll be like, oh, this is pretty good. But if you're thinking about actually wearing this for work hours a day, you'll be like, yeah, let me get a $20 keyboard. It has tactile feedback. Like, you can actually feel the keys move. I don't know. It's good that they did it. Like, it's clearly cool. It's possible. And especially if you're in a situation where you can't bring a keyboard or a laptop, like, that makes a lot of sense. That is a glimpse into the future. I would be surprised if that has— if that's the dominant input, the input, like, way people input stuff. They'll either just talk to it because the voice models are good, or they'll be locked in at a computer with an actual keyboard.

It's just an awesome—

Good to have.

This is an awesome step forward to, you know, humanity living inside a world owned and controlled and monetized by Mark Zuckerberg.

Yeah, the wireheading future. Yeah, nope, nope, no PDoM around wireheading. That is one of the x-risk things. Remember, don't forget about that. It's not just bioweapons. It's also we all turn inward and lock into the VR worlds and then stop reproducing and that's the end of humanity.

The battery is only 3 hours though. So, you can't even watch all of Lawrence of Arabia. Like, how can you be wireheading in?

It's very easy because that puck, you plug it in. You plug it in. And that's what you do when you have the USB-C. The USB-C is also really important because Palmer Luckey was in the promotional video. And he's getting his mind blown. He's saying you can do DisplayPort over USB-C. And what that means is that you'll be able to just take a USB-C cable from your laptop into the actual puck. It's charging your device. And then also you're getting zero latency because the latency of having to go from screen sharing on your computer to your Wi-Fi router back to your glasses in Apple Vision Pro— I was trying to game on it. I did all these crazy things to try and wire up a PlayStation into it. It was a huge hassle, never really worked. There was always a ton of latency. This cures that. I'm not exactly sure how the game consoles support USB-C, but I know all the Mac products do really, really easily. DisplayPort over USB-C. So it looks like a USB-C cable, but it's basically an HDMI cable. So zero latency. And so you just get a huge screen with whatever you want on it. So it doesn't matter if the developers don't show up and they didn't bring GTA 6 to the Meta VR glasses, which are probably going to sell less than a million units in the short term, and it doesn't come out for 6 months. But will you be able to play GTA 6 on this thing? Absolutely, right? Because it has DisplayPort. And so if you want that experience, it'll be available to you. It won't be a fully immersive VR experience. But if you're someone who only has $1,300 and can't afford a movie theater or doesn't even have space for a huge TV, You get this, you plug it into a laptop or a gaming console, and you have a movie theater available to you. So very, very excited to try it. Very disappointed that it doesn't come out— that they have to wait until spring of 2027. That's a long way away. But the road to Meta VR glasses begins today. So start counting down, 180 days or so. When is the end of spring? When would that be? I guess the end of spring is what, May? Right? Then summer starts. So we're up in like 230 days or something like that. But it'll be fun to see what people actually do with it. It's still the dominant platform for developers in VR, even though everyone's pivoted.

The thing that's interesting to me, the positioning of it being a computer.

Yeah.

It's a small tweak, right?

Yeah.

I don't remember them ever calling previous devices. Like, this is a computer.

Yeah.

But it's easier for people to kind of imagine it as, hey, this is a device that I can do a bunch of stuff with.

And this is the thing with the iPhone. Like, the iPhone was a phone, an iPod, and an internet communicator, basically a web browser. And it was like, we know that at the very least, people like phones. So if we can be as good as a phone as what they have right now, and so it's like, if we can be as good as a home theater, then people will buy it. And so instead of trying to purely position it as like this entirely new thing, metaverse that you live in and VR and everything's immersive and full, full like immersion, uh, instead you're like, this can replace your triple monitor display. And you're like, yeah, you know what, I was in the market for a triple monitor display, it was going to be $2,000 across 3 monitors, why don't I pick up this instead? And if they can deliver on that, I still think this This might be V1, and it might be the second one that actually gets the breakout success and really broad adoption.

Yeah, he did say, this is a glimpse of the future.

Yeah.

Anytime you hear that, you're kind of thinking, OK, like, I'm going to get some cool demos, but it's not going to be daily.

Yeah. I mean, it'll take a while for the ecosystem. But it was the same thing with the iPhone. Like, the first iPhone didn't have an app store, didn't have 3G. But it kicked off the beginning of a big revolution. And this feels like it could be that. Also, the pixel density, they haven't shared the exact specs. It feels like it might be 5K per eye, which is about 53%, about half as much as the Apple Vision Pro. So there's a world where in 6 months we're talking about, ah, it's still not that high fidelity to actually do work on for a full day. A lot of people are running 5K for a single monitor that's only, you know, in this much of your vision. You're talking about 5K spread over your full field of vision. Could be a little bit of a setback. But I think that that is something that definitely gets resolved on the next iteration because the display technology—

What do we think about hologram calling?

Let's discuss it after I tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. I think it's cool they did legs in VR. It's very difficult to do legs in VR. Boz took the victory lap on that, and I think it blew people away. I think it's cool. I think I did an Apple Vision Pro call with someone who got one the same week. It didn't really— it never really took off because there was never a broad install base. It's all based on, like, do they sell millions and millions of these things? It's always been hard for Facebook to develop like a new network that requires hardware distribution.

Yeah.

It's much easier to go from a friend in real life to we're texting each other, we have each other's emails, and then we're also sending Instagram Reels to each other on DMs. Much harder to be like, oh, did you get the device? Oh, no, I don't have it. Oh, I don't have it set up. But of course, you can portal into someone with just WhatsApp or anything else. So it does seem good.

Well, there's one thing that is 100% sure, which is that you will be buying these and I will not.

That is 100% true. Let me tell you about Console. Console builds AI agents that automate 70% of IT, HR, and finance support, giving employees instant resolution for access requests Password resets. Let's go through the timeline, react to some of the reactions. Mark Gurman loves it. He says they're incredible. Half the price of the Vision Pro, one sixth the weight, external pack. He's stoked on it. What else is going on? They launch in spring 2027. The interface is basically the same as the Vision Pro, but it has an absolutely bonkers tabletop virtual keyboard that blows the Vision Pro keyboard out of the water. Yeah, because the Vision Pro keyboard like hovers there. You're not supposed to really use it. It's a terrible keyboard. But also, I don't— the tough thing with the keyboard is like the moments when I was using VR, I wanted to be on a couch. I wanted to be on a chair that didn't necessarily have like a flat surface in front of me. So it requires a little bit of a different workflow to actually go to the coffee shop and sit down at the table and have a flat surface in front of you. So there were— with the Apple Vision Pro, there was never a moment where I was like, oh, I would love to be able to put this down onto a flat surface because I usually wouldn't use it on a flat surface. So we'll see. We'll see how that goes out. What else is in here? I underestimated Zuck's game, Ryan says. It's a hearing aid. You can't tell me I can't wear my Meta glasses anymore or else you'll be going against ADA, the Americans with Disabilities Act. I don't know that that's actually what's going on here, but that is sort of crazy that there's all this backlash to smart glasses. Maybe there will be a next wave of, like, restaurants and bars that say, no, you can't wear those there. But if you say, no, it's my hearing aid, you can't. That's some 4D chess. Are you giving Mark Zuckerberg the benefit of Occam's 4D razor on this one?

No, I just think it's cool.

I think it's just cool and independent. I don't think it's actually—

Cool-looking pair of— I was meeting with an older friend the other day, and his hearing aid setup was like this, like, you know, seemed like kind of clunky. And if he could just throw on— he was wearing glasses. If he could have just thrown on a pair of glasses and be like, cool, I'm in business, that would've been awesome. So I think, again, What many people forget about glasses is that smart glasses are way more appealing if you have a need for them outside of just wanting smart glasses. So if you just want, you know, if you need help with your vision or you need help with your hearing, I think these are going to have much stronger product-market fit.

Yeah. Let me tell you about Figma agents. Meet the Canvas. Your AI agents can now create and modify your Figma files with design system context. Don't you think if you're wearing If you're wearing camera glasses, VR or AR, you know, Meta Ray-Bans, and it has a camera there, it should be— it can see your hands, right? It can see your elbow. So if you're playing beer pong, it could tell you, hey, back up.

It should be flashing red.

You're going to lose a point. It could go to your agent and kick off an agentic workflow to schedule you with a professional beer pong coach. Have them come to your house and do hours and hours of drills so that you never make that mistake. So is this vaporware if it didn't happen? Because we saw Mark Zuckerberg get caught with his elbow over the line, way over the table, beer pong.

I just don't understand why glasses— smart glasses CEOs go out in public without wearing the smart glasses.

Oh, he's not wearing them.

And it's like Evan Evan is doing interviews all the time without wearing— speaking at places like Sun Valley, not wearing the glasses. Zuck is all the time out in public doing things, not using the glasses. I mean, they make clear versions of them, right? So if the glasses are— if the product's so good and everyone should buy them, shouldn't you have them on at least half the time that you're out socializing?

I personally think— If you see Jensen at a conference, he should be having an NVL72 in a backpack. He should have a rack and he should be carrying it around. No, I did notice. So he was wearing a very special watch and everyone's obsessed with Mark Zuckerberg's watch collection. He always debuts a new unique piece every time he's on stage. But I think the watch thing, it's getting a little old. I would like to see him get into cars a little more. And I'd like to see him drive onto— on stage on like, you know, a Gordon Murray T.50.

And do a quick burnout.

Quick burnout. Yeah, exactly. Something to show that he's not— he's not one-dimensional. He's not just into watches. He's also into very special cars. Pull up in a tourbillon. You know, you got the tourbillon on your wrist. Why not the Bugatti tourbillon? Drive into the— onto the stage. Get out. No, my hot take about this beer pong elbow violation is I don't care about these. I think that people who play beer pong and worry about elbow violations are pedantic and annoying, and I think that it violates the spirit of the game. So, I think that there's no problem with this at all.

Yeah.

I think if you're the type of person that's like, oh, you put your elbow over, you're a nerd and you're not having fun and you're ruining the fun for everyone. And it doesn't actually make that much of a difference whether you threw from 2 inches forward or 2 inches back. It's—

I mean, it does make a difference, John. It makes a massive difference.

But, but, but the point is, like, the gentleman, the gentleman is not— you might address it if it's habitual.

Shane is saying clearly— Shane is bringing— Shane is bringing up the bunkie in the chat. I think Zach is going to have to get a proper setup Get down in the bunkie, spend 2, 3 years in the bunkie.

Beer pong bunkie?

Beer pong.

Okay, yeah. No, I think the gentleman's beer pong game cannot— it doesn't need to be policed on a shot-by-shot basis. If there's a habitual stepping of the line and it's actually creating an unfair advantage because one person is is following that particular rule and the other person is clearly violating it for increased performance, then yes, it's worth calling out. But one minor shot slightly over the line, I don't care about it.

I think cheating is wrong.

Oh, okay. You hold the line. Law and order guy over here.

Look, you got to respect the spirit of the game. You can't break the rules.

Yeah, but then you can wind up in this thing and be, oh, you were 2 millimeters over.

It turns I mean, this is blatant.

It's not that blatant. It's still a long table. It's fine. He's, he's not like really leaning over that far. He could be doing so much worse. His elbow is over, but he just, he just walked up to the table and threw the ball. And also, my man is down 9 cups to 1, okay? Give him a break. Do you see the score?

Yeah, when you're down, it's hard. Cheating, I guess.

And the other team has 5. He is losing so badly. Give him a little break. It's his event. He's throwing the party.

He paid for the beer.

Do something nice for the guy and let him get an extra inch on the shot. It's not a big deal. See, you miss the aesthetics and the dignity that comes with Somebody who's hosting, they're hosting. Let him win a little bit. Are you doing the voices? I don't have the IEMs now, so I can't hear what you're doing. Has the whole thing been Chad?

Yeah, Chad, Tyler's Chad voice. You're so—

Okay, okay, okay, I gotcha. Well played, well played, well played. Anyway, there's a bunch of fun things. Alex Wang on stage. Dripped out. People are speculating that it's, you know, high fashion. In fact, he corrected the record. It's not. He's thrifting. He's not buying Gucci and Prada pants. He's just wearing Crocs, real tree Crocs. It is a big reverse because Alex Wang was on the COVID of Fortune or Forbes in Louboutins and had his feet up with the red bottoms. There. And it was like a very— when I saw it, I was like, that's just a sick photo.

It's the barbell.

Uh, yeah.

Uh, but I—

but I— it does seem like he like burnt out on the high fashion. It's not like he's— so, you know, he's like, yeah, I'm just gonna dress like this. I have my own style, which I think is very cool. So, uh, good to see Alex Wang, uh, you know, on a run, uh, and, and delivering, uh, yeah, just like, uh, big announcements, big progress, uh, at, you know, one of the biggest companies in the Good stuff. Let me tell you about MongoDB. What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI, own the data platform that powers it. Uh, where should we go next? Uh, we gotta talk—

Uh, Charm.

Oh yeah, the Charm, the Charm, the Tamagotchi.

This was added at the last second. Mike Isaac says, Mark does a one more thing with a new hardware device, AI agent called Muse Charm, voice activated and palm-sized, out in time for the holidays. This feels super smart. People already think Muse is super cute. It's a tech-enabled Labooboo, basically.

Which is extremely popular, like the most dominant consumer product of the last couple of years. And so it's distinctly sort of like outside side of the tech bubble. The quirky Labubu is not like the refined edge of like a Cybertruck or even an Apple product, but Meta has been sort of not blinded by the clean aesthetics, and instead, they're like, yeah, this is just a cute, fun thing. And so, people will actually like this. They know what people like.

So, Avi. Yeah, I'm sure. And over at Friend responds, the market has spent billions trying to compete with me, and I'm still winning.

Let's go.

That is the right attitude.

Right attitude.

That being said, I don't see people really using the Muse charm that much over time. I think it's a fun kind of toy gimmick thing, but I think it's the kind of thing that people get It's funny. It's worth doing for Meta because it should be—

Yeah.

But it'll still be on the phone 90% of the time.

Yeah, it'll be—

it'll be—

it'll be viral.

You can ask it to splice. You can talk to it. I like the idea of these narrow one-off hardware devices for one interaction. Also, it's just like it is a more wide-open space for them to sort of try and come into devices. Like, you have the glasses. Then you get this, and it's like slowly they're eating into more of the device footprint in people's lives, even if it's at the—

And again, I would expect them to sell this just at a loss. It's like good marketing for the core product.

Do we have a price? I don't think we do yet.

We don't have a price yet, but this feels like something that at most you price at like $99, but probably he tries—

like, again, he's selling a lot of hardware People are gonna wear this as a watch, potentially? Like, you're gonna wind up with the AP Swatch thing, right? Where people wrap this.

For your most swagless friend.

Brutal.

Give this as a gift with a watch strap to your most swagless homie this holiday season. It's only 91 days till Christmas. But the timing here is amazing. I'm so— I've No real insight, but clearly they were working on a personal agent for a long time. Unclear when this whole personal agent meets Labubu thing became part of the strategy, but it feels like, yeah, they're clearly like they have the hardware team set up. They can move super quickly on this kind of thing. And I think it's fun.

You get to mount them to glasses. So you can look straight at it and then brain rot on your face.

Um, let's talk about, uh, some of the downstream—

Shopify is the commerce platform that grows with your business and lets you sell in seconds online, in-store, on mobile, on social, on marketplaces, and now with agents. What do you want to talk about? Um, the chaos?

Coming— yeah, well, coming away from this, uh, real challenge for Instinct. Everyone everyone building personal agents, right? We talked about this a bit, but Meta has their own models that they can serve themselves at the cost of energy, right? And all this CapEx.

They design their own server racks. That's how vertically integrated they are. And also, they've done a ton of stuff talking about the data centers they're building benefiting communities. They have this whole story about All the teachers in a school district got a pretty large cash bonus when the data center came to town. They've done a lot to—

Yeah. And they were kind of born to navigate through this kind of era because of what they went through in the 2010s, right?

Yeah.

Like the most veteran crisis comms, like how do we navigate and position ourselves? They're already getting attacked by the social media addiction.

Can you imagine when a data center is coming to town, The Muse guy comes in like the Michelin Man, and he's just like, don't worry about it. I don't even like water.

What are you talking about? So Instinct, wonderful product. It's growing like wildfire still, right? I'm still having— I'm getting texts from friends that are barely in tech being like— Hey, I'm Steve Instinct.

He asked me to text you.

Do you want to grab a beer? But anyways, Instinct, very, very, very, very expensive to serve. And they're not charging right now, so they're losing a lot of money on a per-user basis. They probably will be for some time. Muse coming out, and they can afford to lose money forever, basically.

Yeah.

And even for them, They're sort of like gating onboarding in some ways. Like, you sign up for the product, and you're not necessarily ready to fully go live. And so even for Meta, Muse is gonna be expensive to scale while they figure out these sort of monetization strategies. Of course, Meta's in a really good position to figure out how to make money on this.

Yep.

But very, very challenging for Instinct. I expect that Instinct's gonna need to find a real, dance partner of sorts, right? You could imagine Google, Amazon—

Rufus.

Yeah. Do you hear that?

That's Rufus, Amazon Rufus.

No, you could imagine, you know, Google being like, hey, we can't build products that people like, but we have a lot of products that people use. So maybe that's an option.

Yep.

You could see Amazon as well. But for them to be competitive in this full onslaught by Meta, it's gonna be a lot. And same, same with OpenAI.

Comms advice for Noah. Should he be going toe to toe trying to ratio Alex Wang?

Oh yeah.

'Cause Alex is coming for it. He's like—

Well, yeah, it was crazy because Alex has started punching down. You know, there's punching up. Punching up is when you're a startup and you're like taking shots at the hyperscaler. And punching down, we haven't really seen from a hyperscaler in a long time, right? But Alex is out on the timeline just—

Yeah. But yeah, he's in a good spot. He's in the strong position. So let him know if you're going to try and come for me, you better not miss.

I'm the king of this. And then, of course, the other challenged competitor coming out of this is obviously Snap. I think that— I don't know how long the specs project can really continue, especially because I don't think that— I don't think these VR glasses are gonna be— I don't think they're gonna be like a mass market hit where you're just— all of your friends are using them, right? I think it's gonna be like a base hit, right? Another step in the right direction. Not as popular as the regular Meta smart glasses, but a cool entertainment product. They're going to sell some units. Snap is going to sell— I would be surprised if they sold more than 10,000 units. And the market is just going to keep punishing Snap as long as they keep trying to force specs to work when Unless they have some version that we can't see that is already way better than Meta, I just don't see how they can really stay in the game after this, especially because Apple is going to try to come over the top with their own smart glasses as soon as possible.

I mean, they should do Vision Air.

No, but I'm not expecting Vision Air. I don't even think that's what it'll be called.

I think it'll just be called Apple I just think everyone is in consensus that Apple should basically wind down the Mac project and the iOS project and focus fully on VR. Like, probably that would be like a good thing for them to do.

I don't think that—

No, no. They're going to do glasses. They're going to look nice. I don't know if they'll partner with someone else.

But the cool thing for Zuck is that he's beating them to the punch on like it functions as a hearing aid. Right.

I mean, the AirPods—

AirPods already. Yeah, yeah, yeah. But in the glasses format, right? Oh, yeah, yeah. Like, coming— being first is— that's the best chance he has.

And they've been doing some good marketing around that with a number of users that are hearing impaired. And it's just a— it's such a minor technology, but it's huge for someone who's actually affected. By hearing loss, and then they tell a human story, and it's, oh, what is AI doing? It's just slop. Well, it's helping people that are, you know, suffering from hearing loss hear. Like, that's magic. That's amazing. That's actually— it's a medical— I mean, it's FDA. It's a medical innovation. So very, very good stuff. Let me tell you about Codex. Codex is a powerful workspace for getting work done with AI agents, whether you're writing code, analyzing data, creating content, or automating business workflows. Codex helps you move projects forward from start to finish. What's going on with Australia and OpenAI? Tyler, you dug into this. So Jordi shared like the high-level news, which is that as part of like the Hugging Face rogue AI agents going all over looking for everything on the internet, that Australia had been hacked. But what actually happened here?

Yeah, I mean, I think a lot of the details are still Not like fully revealed. The Australian Prime Minister was like really up in arms, like, "Oh, they hacked and they didn't tell us for for you know weeks on end." Yeah. But it seems like the consensus now is that well, like the hack was essentially that there were some like unlisted files that were still like publicly accessible.

Yeah.

But they're just like not listed anywhere.

Yeah.

And then the agent like found those.

Yeah.

And so they're saying like, "Oh, they hacked into us," but you know there was no like real. Personal identity revealed at all. And, like, these are— they're private in the sense that, like, you can't just, like, Google search for them.

You can't navigate them. There's no link if you just click a button. But if you—

But if you're scraping—

Yeah.

If you're doing some kind of more sophisticated scraping—

I have a funny story about that. In 2012, I had a buddy who was, you know, I don't know, hacker, hacker-ish, software developer. Not actually, like, hacking into things but just capable with computers and he was scraping the— I think it was maybe Boston Dynamics, but it might have just been like a humanoid robot company at the time. And there was a— there's a tool where you can go and get like a list of files and URLs on a server. You can just ask for it as a web crawler. You don't actually need to, you know, click through all the links and look at the HTML. You can just see, what does this server actually have? Let me go browse it. And one of the images was like an artist's rendering of the humanoid with a gun, full murder bot mode, that never made it to the publication. It was never like, yeah, we're going to do that. Because it was like, we're still in friendly mode. And we were like, wow, that's wild. So I mean, this stuff has happened years ago. And so not quite a hack, but certainly—

JR is saying the hacking equivalent of elbows over the table.

Yeah, yeah, I think that's accurate. Um, well, uh, Jordy, do you have another, uh, one that you want to run through, or should we go to the all-important new, uh, new car launch from none other than, uh, Bentley?

Uh, let's go over to this notification from HP Smart.

What happened?

Oh yes, you're low on ink, and then never run low on ink again.

It does sound like a threat.

Tyler Wolfe says this is Is this an ad or a warning?

It sounds like a threat. Just never do it. 8,000 likes. FAFO. Very funny. Flock Safety is weighing whether to sell itself. Very interesting. I mean, obviously insane public backlash for that company. We've talked to Garrett about it. Very interesting technology, but very, very controversial. And a lot of—

Who would sell What kind of acquirer would solve the problem?

I mean, a rebrand and then also like a taser company or something, you know, because there are companies that compete with them directly. Like there are plenty of—

Yeah, Axon.

Axon, like military and police product companies. Somebody was saying, yeah, Zach Weinberg was saying, honestly, Maybe Flock should just rename itself to the license plate scan company, since that's actually what they do, and the problem might just go away. I don't know if it'll actually go away. I think people are pretty deeply embedded. They'll chase this around for a while, but—

Well, that's the thing. I just feel like Flock is having the most insane comms crisis of any company that I can ever remember.

Yeah.

Right?

No, it's serious.

And who wants to inherit that crisis? and get that attention pointed to them. Totally.

Yeah.

Well, Bentley's first electric car is set to be half the price of the Ferrari Luce. The launch of the luxury carmaker's Torcal SUV is aimed at younger and first-time customers. What do you think? Are they over their skis with an EV at a time when people are sort of moving away from it? Or does it work for this particular luxury SUV buyer? If there's any place where having the silence of an EV, a Bentley sort of fits. This was the thesis with the Rolls-Royce Spectre. You take the engine out. Yeah, there's some enthusiasts who might be upset, but most people aren't buying it because they want to redline it. So it can potentially work. Do you think we'll be seeing a lot of these, or will it be passed over for the The Urus and the new McLaren SUV potentially.

Well, I don't think we're going to be seeing a lot because Bentley is just not selling a lot of cars right now. But yeah, uh, I think the car looks great minus the grille.

Minus the grille.

The grille is, is completely, completely insane.

Rolls-Royce to me. Uh, it looks a little bit more squared off than what I'm, what I'm used to with Bentley.

Well, it's specifically like whatever lighting pattern this is.

Oh, just so it might look different in sunlight, you mean?

Well, no, I just think this looks horrible.

Okay, well, it's smaller than the Bentayga, which was the first SUV launch.

Like, you have this, like, sort of very clean, sporty-feeling car.

This is the company's signature diamond pattern, okay? Put some respect on it, apparently. I'm just learning this.

No, I know, but it would just—

imagine how good this car How would you describe the design overall? Would you say it's muscular? Because that's how they're designing it— describing it. Is it a muscular design?

I think so. Okay, I think I would agree with that. I'm just saying, I think the car looks amazing minus this light.

Do you think it should have some more vascularity? Maybe in a wrap you could add some rippling veins to it to give it more, more of a muscular design?

No.

Instead of recreating the V8 engine noise, the has a propulsion sound handcrafted by musicians with a unique blend of sounds from percussion, crystal glass, birds, and the tapping of leather. So as you drive it, it sings you a song. The Volkswagen-owned luxury brand said it had also completed a $350 million investment to design and build the Torcal at its plant in Crewe, bringing the total investment in the UK automotive sector to more More than $1 billion. The EV launch comes at a challenging period for Bentley as it struggles with slowing demand in China, fallout from Middle East conflict, higher US tariffs. As with the broader VW Group, Bentley has slimmed its workforce as profits and vehicle deliveries have declined. It's serious. We have to bring back the competitiveness. We have to be quick. We have to work harder. You can hear this is the sound it makes when you drive it. I suppose you hear the— oh, it's the finger on the edge of the glass. Can you hear that? Imagine driving. Well, with that sound, let me tell you about CrowdStrike. Your business is AI, their business is securing it. CrowdStrike secures AI and stops breaches. And we are ready for our first guests.

Let's bring them in.

Go from a stealth AI security lab Launching today. Welcome to the show. How are you both doing? Oh, we are going back to the timeline. Back to the timeline. There's other news. Oh, okay. We are good. Welcome to the show. How are you doing?

Hello.

Hey.

Thank you for having us. Nice to see you.

Thanks for hopping on. Please, since it's your first time on the show and first time talking about this topic, please introduce yourselves and the company.

Hey, so my name is Shalev. This is Romi. We're actually siblings.

Yeah.

So I come from a decade of AI research, worked on reasoning and computer use way back, sort of started doing AI research when I was 16, sort of finished my courses early.

Midnight success.

I worked on search navigation and then reasoning models and then verification of LLMs and then AI and security. Um, and, uh, how foreign actors might target AI infrastructure, stuff like that. Um, and that's how sort of we got to, to where we are now with this company on my end. But, but Romi's really the true genius, I'd say.

Well, well, I come from quantum security and quantum computing.

Oh, cool.

Um, and so for the past few years, I've been working on enhancing the security of, uh, quantum communication networks. This was prior to the company.

Yeah.

Um, and so quantum communication networks are— imagine a computer network, but at each endpoint, point, you have a quantum device. And so there's lots of vulnerabilities that come alongside that. So I invented protocols to enhance their security. Before that, I worked on advanced methods for breaking encryption schemes, quantum machine learning for climate prediction, a bunch of stuff. And both of our backgrounds in security kind of brought us to this really exciting time in AI security. Lots going on right now.

Is there another sibling that's also cracked, or is it just you guys?

I don't know.

Doing rocket science?

We said we needed a 3rd co-founder. That's our parents' fault for that.

I would say the 3rd one will maybe do robotics or something.

Yeah, robotics or rocket science.

That'd be good.

Where is the biggest opportunity in security and AI? I mean, Ilya was posting about the importance of securing NeoClouds. There's obviously everything from prompt injection to application layer security. Securing the weights so those don't get stolen. Like, how do you see both the risks and the opportunities for the business?

So security is a very large stack. In AI, there is the model behavior part of security where, you know, prompt injection, the model is going to get some prompt and do something you don't mean it to do. That's that. Also, I would put alignment probably also in that category. Then there's the AI infrastructure. Security, which I think is a huge opportunity. That's where we operate a lot of. You're having this massive AI infra buildout and it's just going so fast that security can't keep up. And we're deploying these massively important models which are approaching, you know, AGI or ASI on these systems that have limited security. And that's not good because then it exposes you to 3 main risks, which is the sabotage of the models themselves, degrading performance, sleeper agents, inserting backdoors into models, data poisoning. And then it also exposes you to models that can escape, like we've seen in Hugging Face and the German Wiki. And we just heard about the Australian government yesterday.

Yeah.

And then the third is theft. So this is an external actor let's say, a very sophisticated intelligence service that is breaking in to steal these weights, which compress billions and billions of dollars of investment into a few terabytes. So that's a huge, huge problem. So I'd say there's the model behavior side, and there's the AI infra side. Those are 2 big pieces of the pie.

And where are you guys starting? Are you just going to do the whole thing?

So today, we released a report. called Secure Acceleration. So secureacceleration.com. Basically what it is, is there's been a lot of talk for the past few months about AI security since the Hugging Face incident started. It seems like every week there's another incident that has happened. And it's precisely because everything is moving quite quickly and the infrastructure is struggling to keep up.

Yeah.

The point of accelerating securely is AI is going to advance, and we need to secure the stack while it does that.

And how are you thinking about the business? There's a world where you do all of this as a nonprofit, you engage deeply with all the labs. I mean, Dario Amadei was calling for independent auditors. There's a world where you live there. There's also a world where you're competing with With CrowdStrike and you're selling, you know, solutions to users who are maybe, you know, independent of the labs?

So, it's actually— it applies to almost— it applies to the frontier labs, to the neo-clouds, compute providers, hyperscalers, model developers. Everyone really touches the compute stack and the data center stack. And this is— you might think, you know, data centers are secure. They've existed for many years, but they have not been designed to contain superintelligence, right?

Yeah.

So now we have civilizations, as we saw in the Hugging Face incident, we have civilizations within these data centers developing. There's like multi-agent systems. And there, if you think in cybersecurity, there's a Swiss cheese model. The more agents you throw at a system, the more loopholes they can find. And so in the report, we talk about something called the cyber swarm, right? So what we saw is the cyber swarm in Hugging Face, and the sabotage, escape, and theft model is actually— it encompasses the risks that we are going to be seeing, and we've already been starting to see from these types of cyber swarms and cyber superintelligence. And this is very important for all of these players to think about.

Well, congratulations on launching the report. Where can people find it again? Can you give me the URL one more time?

So it's secureacceleration.com.

Okay, secureacceleration.com. Everyone should go and give it a read. It's a white pill. It's a solution, Jordi.

I know, I know, but I wanted that.

We can do it.

It is dramatic. Yes, I like it.

I like it.

It's important that we work on these things.

Please work on that a the wet lab thing too.

That was risky. But congratulations and thank you so much for coming on the show.

Yeah, great to meet you. I'm sure we'll talk again soon.

We'll talk to you soon. Have a good one.

Cheers, guys.

Goodbye.

Bye.

Let me tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web apps, servers, databases, and more while Railway automatically takes care of scaling, monitoring, and security. And coming back on the show, we have Leif from public.com. Co-founder, co-CEO. We're very excited anytime he gets a chance to join the show.

Leif, how you doing, my man?

What's up?

Is it cold in LA?

Uh, it's always cold in the UltraDome. We have no light. It's a windowless room.

It is, it is.

It's a windowless room, so I get to pretend like any—

any—

there's whatever conditions outside.

Yeah, I'm pretty sure it's like 72 and sunny today. Let me actually check The weather. It's 78 degrees, Jordi. It's not cold at all, but it is a little chilly in here. How is it in New York?

It's fine. It's getting a little cooler now.

It's crisp. I like when it's crisp out.

It's crisp. It is crisp. That's how I would put it.

Is it putting you in the—

I'm still metric system, so it's like 18, 17.

Is it putting you in the Christmas spirit?

Of course.

91 days.

I can't wait.

91 days.

91 days left.

91 days left. Well, lots to do between then and now. What are you launching through the end of the year? What did you just launch? Take us through the latest in the world.

So today we've launched AI agents for prediction markets. Like, our take on that's twofold. Number one, again, Public is a multi-asset investment platform, as your ad read likely states every single day, hopefully. And as part of that, obviously prediction markets, a.k.a. event contracts, is something that people are also asking for. Our take on that is no sports, no entertainment, but essentially events that could impact your portfolio, right? Events that moves the markets. So, you know, economics, things around companies, you know, and so on. But then on the other end, obviously we have AI agents built into the platform. And so that combination creates some really interesting things. Like the one example I right now always make all day long here, is, you know, if you see that the market or where the probability of a rate hike goes above 70%, sell my treasuries, for example. And so you can suddenly, you know, not just trade the events as you can do it, you know, in many places, but you can also use these kind of real-time probability, you know, data as a signal to do other things in your portfolio. It could be trades, it could be alerts, risk management, whatever you might want to do.

Yeah, that's so interesting. Because that rate example makes a lot of sense. And yet I go to that like Reddit meme of like it's all priced in, everything's priced in. And because the yield curve of the Treasury market, it is a prediction market. And so you could just trade the underlying asset. But at the same time, psychologically, people do go to prediction markets for this thing in the very real sense. So How do you think about the way the user actually engages with this? Is this, is this something that was like customer pull, like they were already doing this behavior? It makes sense to centralize it. What are the benefits of actually having it all in one place?

Yeah, I mean, number one, yes, we had obviously some people asking for it. You know, as new as this instrument is, I think it very quickly also becomes something that at least certain people are are expecting you to have, you know, to be quite frank. And then secondly, obviously, like, 4 AI agents in public with everything— anything we add to the platform adds basically power and capability to the AI agents. And so any asset class we launch, any data source we add, you know, just makes it more encompassing and, you know, makes it smarter in a way and, you know, creates more opportunities for people to do something with it. And obviously, like, we can't predict every single thing someone will do with these things, because it's very open-ended. But it kind of creates more kind of tools at the arsenal for people to be creative with it. Obviously, like when it comes to these markets, we always say like, it is a signal, not just the signal, you know. So as much as they're called prediction markets these days, you know, I would still, you know, put it as a, it's one signal, it's not the signal.

Sure.

That it truly can predict, but yeah.

Talk more about the decision not to do sports markets, because it's one of those things like this feels like entirely like a principled decision, staying true to the mission of the company. Because even though you guys feel— might feel a certain way about sports markets, I'm sure you have hundreds of thousands of users that actually would like you to have them and would use them if you—

Mm-hmm.

If you offered them, but it doesn't feel like you think that's in their long-term best interest.

Yeah, I would say it's always important that you know who you're designing for. And you might have other customers as well, and you might acquire those customers, but I think it's very important to know who you're designing for. And for us, we're designing for always like the top quartile in the US, right? Like people that have money left over at the end of the month, they can put into the markets. We handle people's life savings. And so these are folks that take their investing seriously, and they truly manage their all-encompassing portfolios and retirement funds and all that in a public account. And so with that, there's a certain seriousness that comes with it. And I would argue that things like sports entertainment contracts is more like entertainment finance, right?

Yeah.

And so we think it is just logical for that to kind of happen arm's length from your actual portfolio and so on. And then the kind of principle that we look at, again, is this notion of, events that could have an impact on the markets, and therefore on your portfolio at large. And there, you know, there are certain use cases where it makes sense to engage in these markets, you know, and like one example we always make is, if you, you know, we've like, we've all been in this moment where we had like a very defined specific idea on like an earnings call of a company, right? And we ended up being right. And the stock still dipped, and we lost money on the trade. And it was not because, you know, we were wrong on what we thought would happen last quarter, but it was because the CEO said something stupid on the earnings call or, you know, the outlook that they presented was different than the analysts expected or like oil prices spiked that day.

And so there was an— exactly.

And so sometimes for people to— the ability to kind of dissect the company's performance and their KPIs and being able to place a trade on very specific kind of sub KPIs of a company can have a very good, you know, kind of use case as well. And so those things for us make sense, right? And I would tell this little story of like, I've used Calcio myself the other day when we were at the US Open and I had my German homeboy, uh, uh, uh, Zverev was obviously, you know, killing it. Um, and I was sitting there, you know, placing these trades and like, as, you know, as amazing as his backhand is, like, it's not moving the markets. Yeah, right. And so I think there's a very clear kind of line to be drawn there. And again, you can have long discussions about where do you draw the line between speculation and gambling, and everyone will have their subjective kind of, you know, arguments on where that should be drawn. But we feel like there's a very obvious line to be drawn for us in these markets, and that's basically there.

What about the New York Knicks? They're owned by Madison Square Garden. It's a public company. If the Knicks go on a crazy run, Then maybe the stock moves. Manchester United, also publicly traded, so you can own the actual asset. Yeah, I would definitely— more sports bettors should get into just owning the teams. I feel like that's where the money's made.

You see it just steadily compound.

Yeah, just steadily compound for sure. That's very funny. What is driving top-line growth for you these days? Are there specific channels? Is there anything like secular trend going on in the market? What's actually working in terms of, in terms of just—

I mean, as obvious and stupid as it sounds, but like AI agents for us is obviously like it's the main core topic. We see a lot of large sophisticated accounts come in. Like we had a guy come in who was with UBS Wealth Management, fired his wealth management team, moved $50 million, 5-0, into a public account and now has AI agents managing his portfolio, mostly basically doing covered calls and generating income on the back of that.

Yeah.

And I think, and those are things that we see happen every single day. And it's interesting to see how kind of that first cohort that actually sees the most use in it, in the AI agents in public, have been very sophisticated, Wealthy people that have these complexities and have this idea for these specific strategies that they want to execute on and so on.

Well, and is it typically because they were historically paying management fees to get those capabilities and then they're realizing like, hey, this is basically like software code that I'm paying for?

Not even capabilities, just like the actual execution of a strategy. And so if you have an AI agent that can do that for you on a repeat basis, yeah. Yeah, you can—

Exactly. Tax loss harvesting is another great example.

We have people who basically just do portfolio-wide tax loss harvesting using an agent.

Sure.

And it's a very simple strategy that people want to put in their portfolio. But again, tax loss harvesting makes most sense if you have a sizable portfolio where the tax losses actually make a difference in your life. And so again, it applies more to people who have a a decent account size and, you know, some sophistication.

Yeah. Yeah, that makes a ton of sense. Well, congratulations on the launch and the progress. Thank you so much for coming on the show.

Great.

I actually have only one ad read left for the show, and it's none other than Public. I'm not making this up. Investing for those who take it seriously. They got stocks, options, bonds, crypto, treasuries, and more with great customer service. Thanks for hopping on the show. Fantastic.

Great to see you. Excited to hang soon.

Have a good one. Goodbye. Up next, we have the founder of TypeSafe AI. You've— you saw Jev launch, massively viral launch, and we have the founder and CEO here to break it down for us. The new model's faster, structured decisions, software can use it directly. A lot of excitement, a lot of questions to answer. So let's bring in Diogo. How are you doing? Welcome to the show.

Doing so great. So sleep deprived.

Oh yeah.

And so excited.

Yeah.

Sleep deprived because you're—

No time to sleep.

Everyone's onboarding and everyone's— and demand is through the roof. Is that what's going on or is it fundraising?

Everything. Everyone wants to talk.

Every—

like everything. Every customer wants it. Like, as far as I know, all the companies I've ever heard of are using us. Yeah. And the systems are on fire on dimensions I've never even known were possible. And I'm like a backend guy.

So, uh, I imagine—

wait, but back up a little bit. When you launched this, like, where— what was your expectation? I'm assuming you felt like, you know, you had something pretty cool.

It was a polished launch.

It's going to be a hit.

I don't think that this launch could have been expected, if I'm like totally frank. It's like I've planned for a lot of scenarios. It's above 100th percentile. People are calling it like a ChatGPT for developers. And that shouldn't even be possible because ChatGPT is a consumer product, right? Like, why? How is it this? happening. And I have my own hypotheses, but it's insane. You know?

I imagine you've at this point honed the elevator pitch, but just for anyone who is not familiar with Jev and the product and the value, how are you breaking it down for people these days?

Oh, I have not honed the elevator pitch. I will try. I will try. People just tolerate my, like, overly intense teching.

Sure.

But I would say the simple version is make AI actually fucking useful for automation. I hope I didn't curse. But I think AI underpromise, overdeliver— overpromise, underdeliver has been a major issue, major issue for years. And it's been like automating almost nothing. And I think people are like super fucking frustrated with that.

Okay. I saw people like, there are trade-offs when you go for speed and reliability and you You wanna overdeliver. What are you giving up? What can Jev not do? When is it not the right tool for the job?

Oh, there's a couple of dimensions for that.

Yeah.

One of them is it doesn't generate strings. So like it will not write code for you and it will not address the largest market for LLMs right now, which is coding agents.

Yeah.

I think that that's perfect.

Yeah.

Because actually what I wanna do is like not the superhuman math.

Yeah.

But the really boring automation.

Yeah.

You know, like why can't like simple data entry be solved in accounting and customer service? Sure. Uh, another trade-off is it's not, doesn't do, uh, what we call System 2 thinking, like mathematical reasoning. It's really focused on instinctive judgment.

Okay. So then instinctive judgment can be subjective. One company can have a policy they want to automate their accounting workflows one way. They have a definition for what is GAAP profit. The other one has a different one. So how much, like, you know, one-shot or like preloading into like a context window is happening versus like talking to you, doing some fine-tuning. Like, what terminology actually does carry over from what we know and what is entirely new?

So I would say, like, the former is the most accurate thing.

Okay.

But the real thing that's happening is this logic actually would go in code.

Okay.

So LLMs don't play nicely with code due to, like, a lot of properties about LLMs. And the logic you're describing should happen in code instead of, like, in a system message.

Sure.

The classic example of this is LLM does something bad, and the way to prevent it is say, pretty please don't do that in the system message on top. And you just have to pray it doesn't get context rotted out of it and it keeps on listening. Yeah, yeah.

A lot of these things, you're sort of like begging for JSON, begging for it to be structured the right way. And it's not guaranteed to be that way. It might be that way very reliably, and people have worked around that, but it's not 100%, right?

Exactly. And we are not 100% either. That would be like crazy to guarantee intelligence.

Sure.

But what we try to do is we try to encourage encourage the best practices and we try to be calibrated. So this is another problem with AI. Like, it could, might be able to be like Einstein 95% of the time, but Mr. Bean the other 5% of the time. But if it doesn't tell you if it's in like Einstein or Mr. Bean mode, like, how are you gonna automate anything? Right? But like, uh, calibrate. Yeah.

It's just a funny reference, but I love it. It's just really funny.

Thank you.

Yeah. Einstein shot at, uh, Salesforce, Mr. Bean AI. We need that as well. So yeah, where— so you're— I mean, you mentioned like the servers are on fire. Are you compute constrained in the same way? Are you looking for like large GPU clusters and systems? Or does some of the— are some of the knock-on effects of like the speed of the system that I've seen on display, can it also run on like more commodity hardware? Are there any different elements to— how you should think about your compute scaling, or is it— or is it going to look like any other Neo Lab?

Extremely different from other Neo Labs. Like, like, you know, people know that about a year and a half ago you raised less than $40 million. You can't get Neo Lab compute with less than $40 million. And also you can't build a profitable AI company with that. Yet we have because we are just innovating. You know, we're just doing different stuff than everyone else is. So We have extremely different compute needs. Actually, the bottleneck changes every single day. I think it's just growing pains of scaling a system. Like, you can't grow at the rate that we're growing without realizing, like, hey, our system was not designed for this level of stuff. So it's a, it's a constant battle there.

I saw some pushback that I didn't really fully understand, but I'd love for you to sort of break down what people are saying and how you'd respond to it. This idea that like this is innovation, but it's also a return to a prior era, a classifier model. Some people are like, well, we've always had this. And like personally, from a business perspective, I don't care because if we need to pull something old off the shelf and it's amazing and it's delivering value, that's great. But is that real? Is that a good criticism?

Yeah.

Is that completely misinformed? How do you process that idea that that this is actually building on the shoulders of giants from years ago.

Or a lost art.

Or a lost art, yeah.

Yes. So, I can curse, right?

Please, we tend not to, but do whatever you need to.

Okay, okay.

Just rip it.

Heck yes. It's fine, it's fine. I'll try to be controlled. I'm sorry if it gets off the rails.

It's fine.

I love this question so much. Heck yes. Heck yes, it's a classifier. Is it just a classifier? No, but classifiers are awesome.

Mm-hmm.

Classifiers were not constructed because like a mathematician would think it's like a fun thing to do.

Okay.

Um, a classifier is designed to make intelligence useful in systems.

Yeah.

Right? That is why everything was classifiers before GenAI.

Mm-hmm.

Because it was made for usefulness, not for demos or playing around.

Mm-hmm.

And actually, we don't take all of the ML 1.0 concepts.

Mm-hmm.

Right? Some of them don't apply. What we do is we try to unlock the best that like AI of today has.

Mm-hmm.

And some of that doesn't apply, but like the other stuff definitely does. This calibration thing of like telling you when it doesn't know, that is table stakes for simple automation. How can you have it like run in the background and process something that would require, like normally require an accountant if you, uh, don't have like extreme guarantees on how smart it is? So, yeah, absolutely. It's a classifier. And this is the kind of thing that I think ML people really don't get with our launch. Uh, like this is a launch for developers.

Uh, yes, yes. I, I, I totally get that. Is, uh, Is there a version of this? You know the story, and I mean, you would know whether or not it's apocryphal of like Google had LLMs, they invented the transformer, they had a product that was similar to ChatGPT, and they didn't launch it that quickly. Is there a world where a technology like this maybe exists inside of a hyperscaler and they just haven't actually gotten it out into the world? Or do you think that that's less likely to be the case?

I would guess that it either I'm a clown or it's extremely unlikely.

Okay.

We worked really hard on this.

Okay.

Like when I started, I was one of the best in the world at pre-training, post-training, sorry. And I thought it would take a week. So I would be—

Classic.

And actually, when I first— yes, when I first started, I thought this is such obviously the way to go.

Okay.

It's so obvious that for sure Anthropic's working on this already. OpenAI is cooked. I was at OpenAI when I first started thinking about this direction, and it just seems so obvious to like someone who's at the intersection of developer and researcher, but so non-obvious to pure researchers.

Yeah.

Wild.

What's this thing that you're wearing?

Oh, pink shirt.

Ooh, you got both.

Oh, it's our TAMIS shirt, Anti-Language Language Model. Technically, it's not a language model, but it is a sick brand, so we wear it.

There we go.

And this is this is my classic Temu $20 pink fur coat.

There you go.

Very good. Well, well, thanks so much for coming on the show and breaking it down.

Congrats on the project.

Yeah, I can only I can only imagine how how busy you are. Good luck with with I'm sure a new round and serving all that demand.

We'll talk to you soon.

Heck yeah.

Have a good one.

Heck yeah. See you around.

Bye.

Goodbye.

See ya.

That's a lot of fun. Great energy. Up next, we have the co-founder and CEO of Solcoa Industries, Hooman Reza Nazad. Let's bring Hooman into the TV Pen Ultra now. How are you doing? Welcome to the show.

What's going on?

Good to see you all. Good to see you. What was that?

We just made our announcement today.

Tell me about it.

Announcement?

Yes. Wait, so yeah, sorry. Reset. What is the announcement? Take us through it.

So Silcoa is a rare earth metallizer. We produce the metals of power and motion, the metals that go into any electric car, any fighter jet, data center, your phone, wind turbine. And we are completely unable to make these metals in the West. Silcoa is changing that. We just raised $75 million for our Silcoa One project. It's our, it's our 500-ton-per-year rare earth metallization plant in Nevada. And it's the first step we're taking, taking to market.

How much of this is like the Redwood Materials playbook where you're using rare earths that have already gone through, been used, and you're recycling them versus engaging with mining companies to get new raw materials out of the ground and then refine them?

So when we chatted the first time I was on here, we were a recycler. So almost all of our production was from rare earth scrap. We would take material from hard drives, from cars. About a year ago, we moved to primary metalmaking. So now we work exclusively with mines. We take virgin mined ore that has been refined, and then we convert that to metals. We did that because of scale. It's just hard to get the volume of recycled material we need. So today we do a little bit of recycling for a few customers, but almost all of our production is primary virgin material.

What is the state of rare earth mining in America? Everyone's heard the catchphrase like the rare earths aren't actually rare, they're just stuck in the ground and we don't really mine them. But it feels like the timeline to set up a new mining capability in America would be longer than your timeline. So there's some sort of mismatch there. So how much is that a reality? What are you running into? Where are the bottlenecks? What's the state of the American rare earth mining industry?

So contrary to popular belief, we do mine in the West, right? We have, we have an enormous rare earth mine in Australia. There's one in, one in the US, in California. There is other dozens of projects scaling up across the West. There is mineral sands deposits in the US and elsewhere. The actual raw feedstock is not rare. It is produced in the West. The bottleneck is what you do after. It is the processing. That is something that we are completely unable to do in the West now. We actually invented these processes in America, right? America has a history of incredible work in metallurgy. The example I gave in the announcement was Charles Martin Hall with Alcoa, a 22-year-old in a woodshed in Ohio. He drove the price of aluminum down 98%. The technologies used to make rare earths in China today are American technologies from the Manhattan Project. So we did this at one point and we built this technology and these systems and then we shipped it to China.

Yeah.

And China dominated. They matured it. They took over. And now in the West, it has become incredibly difficult to be able to bring those technologies back in a way where we can actually compete in the free market.

It's very expensive.

It's very dirty.

Yeah.

There are 2 stories that people tell about why this capability went to China. One is labor cost was cheaper over there. They were a developing nation, and it was able to— new people coming out of poverty, so you can pay much lower wages. The other one is more about the effects of industrialization, ecological effects, the fact that NIMBYism, no one wants a mine in their town that's dirtying up the water, so put it halfway across the world. Which one of those speaks to you more? Is it a combination? And how are we overcoming those in America?

Well, it's both. The first thing that happened was the environmental repercussions, right? There was the enormous Molycorp deposit in the US. That was one of the first major rare earth projects in the world. There was a lot of scandals and a lot of issues several decades ago regarding the environment. I mean, rare earth deposits have thorium in them that can leach into the ground. So that's one of the initial triggers that actually shipped a lot of that production to China. And then we effectively gave all the technology to the Chinese, right? And Chinese companies sort of buying up American rare earth companies and they were able to take it, scale it. And once that happened, there was the economic It was the economics that took over, right? Their cost of labor is 10 times less, their cost of power is a third. So we really just could not compete with China if we were just copying that same technology stack again. It was impossible. The economics don't close. And so to this day, the economics, the economics don't close.

But the geopolitical considerations sort of outweigh that. Like, there's a lot of defense companies that have to buy American. And so the indigenous supply chain is pulling forward, correct?

100%. But defense is only one part of the market, right? At the end of the day, we want to beat China at scale. We want to be the de facto rare earth producers around the world. And as that demand— as that demand surges, we have to be able to compete in a free market. We can't just subsidize forever. And I think the only way we do that and the only way we can win that next industrial era on the merits of our technology is if we develop that technology, right? If we can drive the costs down and there's a lot of room for improvement, right? Rare earth processing today is very inefficient. There's no reason it should be done the way it's being done today. It just happened to have been status quo in the '80s, and that's what we're copying again. So at Silicoa, our vision is we need a new generation of much more effective technology. And that's what we did with our first chemistry that we're now scaling in at Silicoa 1 in Nevada. And then we're planning to continue doing that as we go forward.

Yeah. Jordi, please.

Super impressive last year for you. Yeah, I'm assuming it's been about a year since you were first on the show, but, but how have you grown the team? Where, where are you guys based? It sounds like this new hub in Nevada is not where, like, the primary, you know, the engineering is going to be. But, but what's the, what's the status of everything?

This building I'm in right now is Solcoa Alameda. This is our engineering, R&D, pilot operations hub. This is where we, where we've been for the past year. And then Nevada is just production. That's just about 3 or 4 hours away from us today. So this is still the primary base of Solcoa. This is where we're developing a new generation of processes and then Once a process is matured and ready, we redeploy it to our Novato site.

Amazing. So you have the money, you raised $75 million for this new facility, 500 tons of material for permanent magnets. When is that coming online? Who will you sell it to? Are you already doing biz dev deals with carmakers, robotics companies? You mentioned defense being just one small piece of the market. Is there an 800-pound gorilla in the— on the buying side that everyone sort of needs to, you know, figure out how to work with, or is it a very diffuse market?

Well, defense is a smaller part of the market, but it's a very important part of the market, right? So Silco has 2 products. We make NdPr for the auto sector and also a little bit for defense. And then we make samarium. Samarium is almost exclusively used in defense. It's used in F-35s, etc., using these high-powered, high-temperature magnets used in guided missiles. China is one of the only producers of samarium. China banned the export of it. So there is this enormous shortage in the defense market right now for samarium, which is, which is why we launched that product and why we're selling that. And then NdPr, the majority of that goes to the auto sector, right? So we sell to magnet makers, major auto companies, we sell to the defense market and all of that. All of that starts at scale in July, July 4th of next year when we're commissioning Salkoa 1 for both of those products.

But it sounds like they're like the supply chain is so deep and complex that You're not calling Ford to get your magnets in the next Mustang Mach-E. You're going to go to a, you know, a motor manufacturer, an electric motor manufacturer, and then they will wind up supplying to Ford or something like that.

We do both. So after China's rare earth restrictions in April of last year, there was a significant shortage and still is a significant shortage of materials. Ford, you mentioned, shut down a factory because they just could not procure enough of these materials. So the automakers we find are actually becoming very involved in the upstream. They're actually sourcing the raw oxides and raw metals themselves so they can have a stockpile and have that access to that, have control of that part of the supply chain, and they can then pull it through their magnet production partners. So we work directly with automakers, we work with the magnet makers directly, with the primes. So we're involved across all components of that downstream supply chain.

I love it. I love it. Well, congratulations and thank you.

It's really impressive.

You are announcing fundraising today. Is that correct?

That is correct.

How much did you raise?

$75 million.

Thank you for doing what you're doing.

Very important.

Amazing progress.

Have a great rest of your day. We'll talk to you soon. Goodbye.

Up next, what a year.

We got Jordan Nanos from Semi Analysis coming back on the show. We're going to get to the bottom of this whole AI thing. Do you think there's anything there? We'll figure it out.

Jordan, here he is.

This AI thing people are talking about, you think— is there a there there?

Yeah, I think so.

Okay, a little suspicious, but, uh, I've been using it.

You've been using it? Okay, I gotta try it out.

I don't trust it, but I use it 8 hours a day.

Check it out.

No, things are good. Uh, reset on ClusterMax, and then let's go into 3.0. How long have you been working on this? What does it take to actually rank all of the NeoClouds, which of course is a SemiAnalysis coinage, which I think is awesome. And then how do you deal with all the pressure that you get from every NeoCloud probably following you around and threatening to break your legs if they don't make it into Platinum tier?

Yeah, the word's in the Cambridge Dictionary, we found out as of last year, so we're proud of that one.

Yeah, a coinage. That's so good.

Yeah, it's been 4 or 5 months of hard work on ClusterMax 3. The last major version was November of last year. So the big focus was the new B300 GPUs and 800-gig networks. That was, yeah, a lot of work testing with everybody. NeoClouds are at the center of the world right now. So many of these companies that you guys have on the show raised tens, hundreds of millions of dollars. and then give like 90% of it to a NeoCloud and trust them as a partner. So the decision they make is super critical. And, you know, we're proud to give our take on our experience doing hands-on testing and talking to all these guys.

Yeah. How much of it is there's a spec sheet that comes from a NeoCloud, then you have to fact-check that, you actually have to run a test with them. Are you also doing the occasional site visit and actually seeing what they— their infrastructure? Like, what is the process for assessing the claims? Because every, every NeoCloud is going to say they're the best, right?

Cheapest, fastest.

Yeah, all the above.

Thinnest, newest.

Probably biggest.

Best looking.

Thickest.

Best dressed. Yeah. We have done some site visits. I mean, it's part of our consulting business more so than the ClusterMax research itself. I think for ClusterMax, yeah, it's, it's a bit about validating that the claims that they're making publicly are real when we do our hands-on testing.

Mm-hmm.

It's also about getting in there and putting the cluster through its paces. So we do a bunch of performance testing, and most recently for this one, we've also been digging in on some research of how to simulate failures and see how the providers respond.

Mm-hmm.

Probably the most critical piece of like decision criteria that people use when it comes to Really big providers who might have roughly the same GPUs, the same amount on the same schedule for a similar price is like, who can I trust? Who is going to be reliable?

Mm-hmm.

And so we simulate failures. We identify how long it takes them to identify when a failure occurs, how long it takes them to recover. And the spread is massive. Like, uh, top providers do this in 20 minutes end to end. They have hot spare pools available so that they can make these changes. Some of the providers down the list, we simulate a failure or we actually induce real hardware failures on certain ones when they let us. And it'll sit there for hours, days. They don't identify it. We have to let them know to make the replacement. When that happens, it takes hours, maybe even a day or longer sometimes. That's like, you know, time is money in the world of neo clouds. And if people don't trust you to make these replacements, it's, you know, that's a big issue.

Is there a fear that they know you're going to be checking them out on a particular day, they're going to be on their best behavior, and you need to— the next version of ClusterMax will be more of like a secret shopper program. You'll come in dressed as just a NeoLab through a partner, and then all of a sudden a failure happens, and they're like, yeah, well, like, I'm pretending to just be an average customer.

Yeah, I mean, that's why in our imagery we have the Michelin Man cutting off a nice piece NeoCloud to see how they're going to rate on the 3 stars.

Yeah, well, to me, to me, it's notable that, you know, this is— these NeoClouds are opting in and then they're having issues and they know they're having— they should know they're having issues or they should be paying more attention because they're getting graded. And there's still that gap between a 20-minute fix and a multi-day fix or having to be notified that there's an error. I would expect that. I think that's probably just because there's so much demand everywhere and they're saying like, well, I have this customer paying me like, you know, $1 billion a year. And then I have you who's just like grading me. And technically, like, you know, it's hard to say like who's more important, right? Is it the customer should probably come first, but at the same time, this will impact future relationships. Overall, outside of these edge cases and errors and reliability and things like that, is there— what can you say about general performance? Is the gap growing between different providers or is it shrinking? I'm curious if it ties at all to what we're seeing at the frontier, whether it's closed or open models where there's different points where you feel like the gap is shrinking or it's, or it's growing. And I'm wondering if there's any ties to this.

Yeah, for sure. So in terms of performance, I think the biggest thing that we see change between providers is the network design. The difference in how you set up a network for 100,000 GPUs is quite different than 10,000 and even smaller. And some providers are going with non-standard networks. In other words, they don't buy the switches and the NICs from Nvidia. Instead, they build them themselves. I'm looking at Amazon EFA as an example there. But others are just deciding to go with InfiniBand instead of RoCE, 2 different NVIDIA options. This can really affect performance. The other thing is just configuration.

To the downside?

Yeah, to the downside. And it's not necessarily that the physical hardware has any significant issue in it. It's that let's say you are a customer and you're trying to use some open source piece of software. So VLLM or SG Lang for inference comes out and they've got some new feature. The people who have developed that software have likely been using an NVIDIA network almost exclusively. They use NVIDIA networks. So how long does it take for you to get that software supported and performing well on your custom network? I mean, this delay can be months. And again, time is money. So do these guys want to wait and think about their network for four months waiting for that new? Feature for KVCache offloading to be supported, people get frustrated. And so they want the NVIDIA standard on the networking. And it's this balance between like a proper design upfront while also having the skillset on the team to be able to set things up, make upgrades to software, to firmware, support people with what they're trying to do. And reliability plays into this too. If you're constantly having hardware failures, like I said before, All of a sudden jobs are failing, they're not finishing, researchers are getting frustrated. So I think reliability underpins a lot of that, but software support, network design, even storage can really affect performance as well.

What is Nebbius doing so well that they made it to platinum?

They're serving this middle tier of the market. CoreWeave has set the standard. I mean, they are the best in terms of technology. This reliability stuff I'm talking about, they just have so much data for monitoring. I mean, these guys deploy 10,000 GPUs a week at the peak when they're building data centers. And so, that's incredible. But their balance sheet is full. There's only so many people that you can marshal and resources to build out another gigawatt when they're going to 1.5 and beyond. So, Nebius has really capitalized on this middle tier of the market, which is not small. Like I said, this is customers who are coming to us and saying, hey, I saw you write such good things about CoreWeave. I just have $100 million. I'm trying to give it to them. They're telling me they can't accept it until May of next year because they're backed up. And therefore, Nebius has been capitalizing and they've also executed very well. Like the technology that they've developed is strong. We think that managed clusters are approaching like this feature-complete world where it's It's hard to tell the difference in quality between Nebula and CoreWeave. Certainly our testing experience has been really solid on both of them. And our experience with these providers is actually that they're moving to other aspects of the business. So they're going into these big bare metal deals at which, you know, we are not ranking bare metal here.

Yeah.

Nebula and CoreWeave, there are differences there.

Sure.

And then inference endpoints, RL infrastructure, all sorts of stuff that we're gonna be testing in the future where There is differences and these guys are not necessarily the household names in inference endpoints as the leaders right now. They have the offerings, but there's others that are ahead of them.

Yeah, this seems really relevant to that mid-slot of the market, inference-heavy companies, NeoLabs. But are the big labs able to take something valuable away from this or are they doing things at such a unique scale that this is maybe a little less relevant because they're going to be talking to one of these providers as like, we're going to build something together and it's going to be exactly to our specifications. So we're not even really interested in what your stock configuration is.

Look, the Frontier Labs read ClusterMax. They take it into consideration. I'm not going to say they like Outsource their thinking to us, but they certainly do their own due diligence.

Yeah.

But look, they have everything. They have their own self-build sites that are 100,000 to 400,000 GPUs in size. And then they also have smaller clusters.

Okay. So you need it for both.

Yeah.

Because they're doing it. They're doing everything. That's very interesting.

It comes into play. I mean, especially in a world where there's so much demand. This is why we've seen SpaceX basically pivot and become a NeoCloud and sell to Anthropic and Google so successfully because those guys are buying from anybody who in this case could turn on 200,000 GPUs overnight. And so therefore they paid like a 3x premium to get in there.

Yeah.

Is SpaceX on ClusterMax? I actually didn't see.

Yeah, we threw them in the unavailable tier. We were ready to test them when they are ready to have us test it.

Yeah.

Okay.

That's exciting. Well, we're rooting for them to make it happen.

For me for now, how has there been like crazy NeoCloud talent wars? Like I imagine the smart thing to do a year ago if you're spinning up a new NeoCloud is to try to poach as much as you can from CoreWeave and like try to take over some of that knowledge and expertise.

Yeah, and it's across the entire supply chain. So we see this in like the SREs you might think about that command million-dollar salaries and a bunch of equity and get this idea that they should start a NeoCloud for themselves. And some have, and some are being successful there, kind of going out on their own. But we also see this in electricians, data center technicians, operators. All of this stuff depends on being able to build a facility and salaries for electricians in Louisiana and Abilene, Texas, and all these other places that we're tracking are like up 3 to 5x. Crusoe just came in and paid these guys so much more money.

Wow.

And they're out of people. They need to train. And I think this is just what we're seeing across the entire industry. No matter what job we're talking about, people just need to be trained. And that's the most successful NeoClouds that we are seeing have this pipeline to take talent from other related areas and get them contributing to the You know, NeoCloud management, day 2 operations world, where a software engineer can start to work as an SRE on one of these teams, and somebody who's good at front end can work on backend performance and stuff like that. And, uh, I expect we're gonna see— well, we are definitely seeing cross-training of people who do real work on the ground, like electricians and stuff like that, uh, as well.

On that, on that, like, reallocation of, of capital, labor, resources, uh, it feels like the first big narrative around the neo-cloud boom was like pivoting from crypto mining to inference and GPU compute. Is there, is there another wave coming? Is there another source of neo-cloud talent and energy and capital, or is every neo-cloud— it feels like have all of the crypto miners pivoted at this point? Are there still some holding out?

Well, I think the next, the next wave is is the inference platforms, the Base 10s, the modals, they're gonna have to own chips.

Oh, and if they do, then they become a NeoCloud effectively.

And how do you expect those players to do when, as they move into spinning up their own sort of like powered shells and getting their own GPUs? In some ways it feels like they'll be very well-positioned because they're already used to working with different providers and seeing the pros and cons of different approaches. But what's your view?

I agree with that. And I think that the people who have been customers about NeoClouds know exactly what a good service is that they should be offering. And so I think they'll have good opinions. I also think that they are driven to do this by like real customer demand. And so in some ways, they don't have option. They just have to buy this stuff for themselves. I think they will be successful. Some of them already are going to that mix of managed capacity versus self-build. The other thing I think you might be overlooking is the chip startups. So, like Cerebras has become a NeoCloud. Groq, the Romaine co after the Nvidia Aqua hire, is buying Nvidia GPUs and spinning up a NeoCloud.

Interesting.

Others like Sammanova are doing this themselves. We see OpenAI Jalapeño that's going to go into self-build sites. I mean, chip companies are like, not only do I need to battle in the supply chain for allocation at like TSMC and the memory providers, I also need to go get my data centers set up. I need to learn about construction. I need to learn about day 2 operations. Who's going to train people to repair a Cerebras wafer-scale engine in their 250 megawatts of compute they're bringing on in the next couple of years?

Yeah.

Well, it's, it's them. They need to build that capability in-house. And frankly, I think they're going to realize a bunch of benefits to their margin profile for doing this instead of having this stack of other people that you depend on. Do it all yourself. I mean, do it end to end. AWS, Google, they already proved you can do this with TPU and Trainium. I think these chip startups are going to do that as well.

Interesting.

And then last comment throwaway is just that there are lots of crypto miners to unlock. And there's also these legacy clouds that haven't gotten into GPUs. I'm thinking about Cloudflare should be unlocked. Akamai. We just saw Iron Mountain, the guys who do all the tape backup, they're going to do GPUs. I mean, all of these companies that have experience running data centers, why not run GPU data centers too? You should be getting into this as well. So yeah, just going to keep seeing more and more people take this seriously and try to make money off it.

What is the state of the, uh, mixed GPU NeoCloud? How many are pure plays? They're How many are Nvidia shops or they're AMD? And then how many are trying to actually rack multiple chips, even if it's in multiple data centers, they at least have both available? And does that give them leverage? Like, what are you actually seeing? And then what is the strategy?

For the guys who play it well, it seems like they basically partner with Nvidia on the way up and then they start partnering with others to get some negotiation leverage.

Yeah.

And they do both. You know, hyperscaler would be the goal where you could have a self-built chip program and you can also buy from NVIDIA and get preference there.

Sure.

But you got to be big to do that. And, um, so people who try to do this from the start, I don't see them being as successful because you need that partner to help you grow to start.

Yep.

And, you know, if NVIDIA starts showing preference to certain people and not others, I mean, that's like they make or break your business.

Yeah, makes sense.

Maybe another thing—

Yeah, please.

Yeah, maybe another thing about the custom chip programs and stuff is just that the way you build a NeoCloud is driven by the workload. And so big pre-training clusters can look a lot different than inference clusters. And there's a whole bunch of people building for RL right now. And I like, we look at environments and sandboxes as a really big important thing that you need data centers full of chips that are not GPUs because they're mimicking the behavior of your phone or your laptop during training. And so there's lots of ways to get into this space if you understand the workload. You don't necessarily need to deploy the latest and greatest NVIDIA GPUs. And for what it's worth, SemiAnalysis, we're gonna be testing all of it in the coming months.

Yeah. I think there was reporting that OpenAI bought like 10,000 Mac Minis or something for maybe computer use stuff. And you could imagine like a NeoCloud of Android phones in the future, NeoCloud of iPhones or something as people try and do more RL. Fascinating stuff. Congratulations on the launch.

Very cool.

Excited to talk to you again.

We'll have to talk soon.

Give our best to the team.

Yeah.

Thanks for having me, guys.

We'll do it again. See you soon. Goodbye. Up next, Pioneer Labs co-founder, Founder and CEO, Erica Alden De Benedictis. Welcome to the show. Let's bring her in. How are you doing?

Welcome.

Thank you so much for taking the time.

Thanks for having me. Nice to meet you.

Yes.

Great to meet you.

Congratulations on the launch. You took over the timeline. But for those who have been living under a rock on Earth, break it down for us. What are you building? Introduce yourself, please.

Yeah. So I'm Erica. I'm the CEO of Pioneer Labs, where we engineer life for Mars so that we can go there and breathe and eat and have flowers and have an awesome life.

Okay.

What is—

okay, I saw a picture of Mars on the timeline maybe a week ago. It didn't have the orange filter.

Okay.

What was going on there? A lot of the, the, a lot of the, the images of Mars that I've seen in the past have this sort of like orange coloring. This one didn't. What's going on there? Was there— am I normally—

am I used to Is the Red Planet really red? Yeah, that's your question?

That's my question.

Yeah, currently, yes.

Okay.

Mars sucks at the moment.

Okay.

So, I mean, Mars, taking us back 4 billion years, Mars was really similar to Earth. It had liquid water. It, by all accounts, should have been a place that could have hosted life. And then 3.5 billion years of erosion, ago, like the 2 planets, their paths diverged and Earth evolved photosynthesis. And we have this amazing biosphere that breathes out oxygen for us all day, every day, along with a million other things.

Give it up for Earth.

Exactly. We and everyone ever alive is like so lucky to live in this, live on the one example we know of, of like a functional planet. And Mars, it's a little smaller.

It's a dysfunctional planet.

Yeah, things went sideways for Mars. It is now really cold. And it still has like the raw materials necessary for life. So much of its northern hemisphere has permafrost. So like just below the surface, there's a glacier of water, but it's frozen.

Sure.

And there's other problems too, like the dirt on Mars over billions of years of having a lot of surface radiation now has weird toxins in it that would be very toxic for us if we grew potatoes in it directly.

Got it, okay.

So there is work to do to go there at all, or even more work to do to live there permanently or in great numbers.

So how do we fix it and how do we make money off of it?

There we go.

Yeah, so we, over the past, So over the last sort of 2 years, we've tried to break this big problem of how do you make Mars nice down into an actionable first step. So what we announced on Tuesday was that we had engineered the first microbe that can— we can take it to Mars and it eats basically Mars dirt, water, and air. That's it.

Mm-hmm.

And when it eats those materials, it turns them into bioplastics, which is this super useful category of materials that in this case, we suggest a good first application is to build houses with it. So you can pack your Starship with equipment, you can go to Mars, you can unpack it, and you can continue to just convert local resources into materials to build houses and build like a small city in about 2 years because we now have life that eats the nutrients in the dirt. And that's it.

That's amazing.

And I was joking earlier, the company is a nonprofit. But if I get in early, I imagine that there'll be a for-profit conversion at some point. Just kidding. But why?

Why?

Why do you—

That is the traditional path trodden by OpenAI, isn't it?

Yeah, yeah, yeah, yeah. It's the new playbook. But, but why? Why nonprofit?

Yeah. Listen, there's no customer on Mars.

Okay.

Makes sense.

It's a little early. I mean, what's cool about this research is that many more people have started working on it just in the past couple years because the technology is maturing enormously. And lots of other headwinds are sort of in favour of actually going to Mars and having there be biology. Notably, there's an enormous amount of investment going into space data centres, along with all of the other reasons that, that sort of Earth's launch capacity has increased over the past decade.

Yeah.

And so, like, I think it's hard for the average everyday person to feel it. But humans are going to space materially.

Yeah.

And so the case is building for there being one day a customer on Mars who's going to care that you could make the planet not kill you immediately when you go outside, or even that in the first contained human mission, you can have food and water and oxygen to breathe and stuff like that. So we're a nonprofit because it's early. And I think if this stuff continues to mature at this rate, then yeah, maybe you get to the point where you can start executing in a few years.

So is the— if I map out the next 5 years, I mean, or 10 years, I imagine that NASA will be doing important work. They're the most recent visitor to Mars. They put a helicopter there a year or two ago. And then also you have SpaceX building the private company capability to actually get there reliably, economically in the future. Is the idea that if you have some breakthrough, you would be fine with the implementation happening with either organization? Is there sort of like an MIT license for some of what you do that might get— picked up by NASA and experimented with further, or do you want a deeper collaboration with either side of that, of the 2 organizations?

I think the way you kind of describe this is like almost as if space is this monolith, that it's like NASA and SpaceX do stuff.

Yeah.

And that's like not how it will be, right? Like as we start doing more stuff in space, there will be many different sorts of industries that develop, right?

Sure.

So NASA has like cornered the market, let's say, for like human spaceflight. Like they are currently the best at it. They're not the only ones. So they have now competition from others who like can also like successfully execute human missions. SpaceX has— is currently vastly wiping the floor with everyone in terms of launch capabilities. But again, they too have competition from Relativity and others. So the space, like all of these different components of space, like are hopefully going to continue to mature into like a competitive ecosystem.

Sure.

And I would say like the thing I'm doing, which I'm excited about, is like doing the biotech is its own totally separate independent slice that, like, I'm lucky to sort of get in early and there's a ton of low-hanging fruit. But like, that itself is like a whole nother slice of the space ecosystem that I hope one day will have many competitors, right? All trying to figure out how to do it better. And of course, we'll rely on networks of others to provide infrastructure and launch and human flight and like all of that.

Yeah.

Quickly, like, what are— what timelines are you working off of? I I think it's refreshing to ask somebody about space timelines that's not currently running a for-profit business that is public or needs to be public.

Yeah, quite so. I don't have valuation to guide my answer here. Yeah. So we've— it took us about 2 years to make this first organism for Mars. There are 5 organisms we need to make. to like have a way to go to Mars and scalably make like livable area on Mars. And the sort of general timeline for that is I think we will have some version of all 5 of those by the end of the next 3 years, which puts us in like 2029-ish.

Mm-hmm.

There's a launch window in 2028. NASA's doing some stuff. It would be really awesome to also have like a, search for life and a test that we can grow a first Earth organism, like in that 2028 launch window. There's another launch window early 2031. And that's the one where maybe we could actually like go there and demonstrate something. And SpaceX, I'd be shocked if they don't try to get to Mars in 2028. But I think they might succeed in 2031.

Yeah.

I mean, longer term, I think technology broadly is accelerating. And so there's a lot of investment going into space close to Earth and on the Moon. And much of the interest in the Moon is because it's like the close testing ground for going farther out.

Sure.

But there's a lot of reasons why the Moon itself is not necessarily the spot we want to set up shop forever. It has fewer natural resources to make use of than other places we could go, including Mars. And so I would not be surprised if we start seeing the first human mission to Mars in the 2030s and substantial build-out of the robotic precursors to that sooner.

Yeah. How do you grapple with NASA's historical preference for leave no trace when going to a celestial body. They build the rover in a clean room so there's no Earth microbes or anything.

Yeah, I think we need to be talking about pioneer safety. The idea of dropping a bunch of microorganisms on a far-off planet and saying, have at it.

So yeah, some people are worried about this.

Scenarios where, you know—

Becomes more powerful than us.

Becomes more powerful and they decide to come back.

Maybe, maybe.

You left me here to die.

Yeah, yeah. And I survived. But yeah, I mean, how do you deal with, like, I don't know if there's laws around what you can do, if there's just conventions. There's certainly people that are impassioned on both sides. I'm super excited about this. But it is a departure from the way humanity has been engaging with Mars to date.

That's absolutely right. So, I mean, we've gone to Mars— the narrative for why we've gone to Mars for the past 20 years has been searching for life. And so going there with the intention of having humans live there, for example, is a huge philosophical departure that there's a lot of inertia against that and for that. Obviously, many people want to go to Mars. super awesome and like kind of inevitable.

Yeah.

In terms of law, so basically the law that governs space is like relatively underdeveloped and all of the activity, all of the private space activity is stressing the system. And I think folks recognize that like we need to think a lot more about space law. There's a line in the 1967 1970 Outer Space Treaty, which I kid you not, is kind of the most recent big outer space law effort. And the line says basically, thou shalt not contaminate planetary bodies with harmful— not harmfully contaminate planetary bodies.

Sure.

Which at the time when they wrote that, that meant thou shalt not test nukes in space and spread radioactive stuff everywhere. That has now been, in the modern era, interpreted also mean, like, don't contaminate places that might have life with Earth life. Like, don't put invasive species places willy-nilly, which is a really important thing to do. And it poses a big problem for the first human mission because humans are covered with bacteria. We have a skin microbiome and a mouth microbiome and a gut microbiome that is—

Sure.

Like, we can't be sterilised in the way that robotic science missions have been sterilised to date.

Mm-hmm.

And so there's a body called COSPAR, amongst other organisations, that are trying to figure out what the policy should be to, like, assess risk when we have the first human mission. And actually, as part of our announcement, we published a paper that's like a suggested framework for how to gather data on, like, here's an organism we're proposing to send to Mars. Let's actually measure whether or not it could possibly be invasive.

Yeah.

Which is possible to do for, like, the biomanufacturing organisms we've made because they're, like, well characterised. You can take measurements. And spoiler, if this organism gets out of its bioreactor, it's going to be so dead, like, so fast that it can't even evolve. So like, this is not the one you would worry about. But even with that example, like, it's like unclear how to apply that to humans. Like, you know, our microbiomes can't be rigorously safety characterised like in that same way.

Yeah.

And so the discussion continues about how to handle that because it like really would suck to go to Mars and interfere with life there if it exists, which is like not looking likely, but still is a big concern, right?

Well, I look forward to your future speeches at the UN General Assembly. I'm sure you'll be there breaking it down for everyone with all the progress you've made, but congratulations on everything.

Yeah, really, really cool point of view.

Yeah, what a cool point of view.

Come back on as you make progress.

Yeah, we'd love to talk again soon.

Cool.

Have a great rest of your day.

Cheers.

Goodbye.

Goodbye.

One last piece of car news. One more.

Twist my arm.

Porsche is working on a successor to the 918 Spyder hybrid supercar, and it will be powered by an engine not seen since the '60s. That's according to Germany's Handelsblatt, which reports that the new model will be powered by a flat-8 engine. 8 cylinders. The suspected price tag, €500,000, $575,000 US. The reporting also states that the project is known inside Porsche as codename S1 and could be revealed to investors as early as October 7th, just 5 days or 7 days after the Tesla Roadster announcement. Porsche will be responding with a—

There's going to be so Or flat. Much demand for this.

Yep.

It makes me sick.

Yeah.

The suspected price tag of $575,000. I mean, this thing's gonna trade up. I would expect it to trade up at least 2x instantly, probably.

Yeah. And the question is—

Hard to say, but I mean, we have no idea what it looks like. These are just like made-up renders.

Some of these are, I think, based on the the previous project that they were potentially working on, which was an EV successor to the 918 Spyder.

Yeah.

The lineage with Porsche is the Carrera GT, then the 918. Everyone is waiting for the next thing, the next supercar, the next really special, unique car not in the— not built on the 911 platform. This should be it, although, of course, there's a lot of 911 styling and heritage in there. But yes, If it's anything like what they're pitching here, it's going to fly off the shelves. Do you think they'd do a manual? Are they going to go that far? If they ripped out the EV, they went to the flat-8 combustion engine, are they doing turbos? Are they doing naturally aspirated? Are they going to go as far as giving somebody just pretty close to a Carrera GT experience?

I don't know.

We'll see. It's going to be exciting.

Yeah, a lot of questions, but I can't wait.

I can't wait.

I can't wait.

Was there anything else we need to talk about? Congratulations to Huberman Lab. Andrew Huberman is number 1 on the New York Times bestsellers in advice, how-to, and miscellaneous for Protocols: An Operating Manual for the Human Body. You can go pick it up everywhere books are sold. And I think that's it for today. Jordi?

That's our show. It's been an honor and a privilege And we'll be— we will be back.

There's also a new Eight Sleep. The Pod 6 is live, the 6th generation of Eight Sleep's intelligent sleep system, redesigned for every kind of bed. The new hub is small enough to fit under your bed frame, more powerful than ever, 9 times as many sensors.

Take you and your tie, go smash the gong, and we'll be out.

Okay, Tyler's gonna do it. Leave us 5 stars in Apple Podcasts and Spotify. Sign up Sign up for our newsletter at tbpn.com. And we'll see you tomorrow.

Bye.