https://www.listennotes.com/e/3563893f02214d74857f0851d6346ec8

You're watching TBPN. Today is Thursday, June 19, 2025. We are live from the TVPN Ultra Dome, the temple of technology, the fortress of finance, the capital of capital. We have a great show for you today, folks. There's some breaking news that's dropping right now. I think we got to go to the printer cam.

Really?

Because we have an update from friend of the show. Let's see if this works.

Do I need this?

No, we need a moment of silence. Gong. Because it's about the Elon Musk news out of SpaceX. But we got an update from friend of the show. Ashley Vance coming in here hot. He says, I happened to be at Neuralink last night when starship went boom and so was Elon Musk until well past midnight Pacific time. He was in three hour plus. He was in a three hour plus long meeting when the explosion happened. Meeting ended. I assume that's when he learned about it. And then he went back to work.

Wow, an absolute dog.

What a grinder.

I mean, video was absolutely insane.

It was.

We will cover it in a little bit. You had shared a transcript from it.

Oh, yeah.

With me earlier this morning. I had seen the video of it going boom. You shared this transcript where one of the engineers is saying, hey, really quick, Sawyer. We just observed a couple of vents coming from the common dome in between lox tank and a methane tank. And from this angle, it almost looks like the methane tank is gone. Then he actually says, question mark. Is that normal, Jack? Normal? I'm seeing some venting. Is that. Is that unusual? Is that normal? And the guy goes, yeah, it's probably normal. I have just.

You have the video? I just said the video.

Yeah. Let's play it.

Let's play this. I have no idea if this is actually related. We'll have to get someone on the show to dig into like exactly what happened. I'm sure there'll be a post like a postmortem on the exploit.

They did say famous last words. And then shortly afterwards it exploded. The entire.

It is a crazy, crazy video. Not really. Quick Sawyer.

We.

Let's see if we can pull this up. In the meantime, let's tell you about ramp. Ramp.com Time is money save both easy use corporate cards, bill payments, accounting and a whole lot more all in one place. Go to ramp.com to get started. Of course, the other two major news stories we want to cover today. $10 billion, the price to buy the Los Angeles Lakers. 15 billion the price to buy Meta and AI leadership team. Fantastic post. By Alex Conrad They've just observed a couple of vents coming from the common.

Dome in between the lox tank and the methane tank.

And from this angle, it almost looks like the methane tank is gone.

Question mark.

Is that normal?

Jack, is that normal?

And then I actually edited it out. I'm seeing some, some venting coming from.

In between the methane tank and a lock stank.

Is that usual? Is that normal? Yeah, it's probably normal.

Keyword, Probably.

Famous last words. Yeah, weasel words. Famous last, weasel words. These guys are just growing out on a live stream, watching, watching. Like a static firing test. This is not a launch. They're not going to, they're not trying to launch the rocket. They're just, they're just putting on the tape on the test stand, firing it up to make sure that everything works. And it just completely exploded. We'll go deeper into that and some of the reaction in the news in a little bit, but in the meantime, let's talk about the other piece of breaking news that came out of the printer just after we got off the stream yesterday.

Oh, right.

This is news from the information. Meta is in talks to hire former GitHub CEO Nat Friedman and Daniel Gross to join AI efforts and partially buy.

Out their venture fund.

Yes. So there's a ton of details here. So let's read through the information article and then we'll go to some of the reactions. So Meta platforms is in advanced talks. Not just talks. Advanced talks. Has that been defined, quantified? What does that mean? Like, like are we past coffee meeting? Is this a 30 minute, is this a two hour conversation? How long are the talks until they become advanced? Yeah, I must know.

You know, exact numbers are being thrown around. Very possible.

Yep. So they, they're thinking about bringing in Nat Freeman, Daniel Gross to help lead AI efforts. Part of those talks, Meta is in discussion about partially buying out Friedman and Gross's venture capital firm nfdg, which holds stakes in top AI startups and is worth billions of dol dollars. On paper. If the talks are successful, Gross would leave Safe Super Intelligence, which he co founded with former OpenAI chief scientist Ilya Sutzkever last year. At Meta, Gross is expected to work on mostly on AI products, while Friedman. While Friedman's remit is expected to be broader, both Gross and Friedman are expected to work closely with Meta CEO Mark Zuckerberg and Scale AI CEO Alexander Wang or Wong, whose hiring by Meta was finalized last week in a $14.3 billion deal.

Big numbers, Big numbers.

I think this is gong worthy Even though we're just in advanced talks, we gotta hit the gong.

We try not to hit the gong for advanced talks, but scoops are scoops.

It's such a big number.

Fantastic, strong hit. Anyways, Yeah, I guess some, you know, immediately. A couple things I was thinking about when I saw the headline. One, I don't know if this is common knowledge, but my understanding was that some of the money from NFDG was zucks. Right. So they were already investing. There was another half of zuck to some degree. So this shouldn't be a huge surprise. And then I think the bigger thing is, what does this say about ssi? Right. If DG is willing to leave SSI despite it being, you know, such a young company and already valued, I imagine DGStake is in the billions of dollars there. So to leave that and go to Meta says something. I don't know exactly what it says.

I think it could potentially say a few things. One is that maybe artificial intelligence is more of a sustaining innovation than disruptive innovation. And so that just by training, a fantastic model, you're not immediately going to be able to overcome the network effect at Meta. And so Meta is maybe potentially a better place to go, you know, really reap the rewards of artificial intelligence. That's kind of a signal because, you know, no one at Google was really thinking about joining Yahoo. Right. There wasn't a lot of flow that direction. Yeah, it was like, we're onto something we are going to disrupt. You know, same thing with, with Amazon. I'm sure Bezos wasn't losing people to Barnes and Noble, right? This is.

Barnes and Noble threw out a couple max contracts, maybe mercenaries. But yeah, I mean, if you're, you.

Know, this was trul. Disruptive, you would think that you would say, well, I don't, I definitely don't want to be with the incumbent. I don't want to be in the legacy player because they, there's nothing that they can do to capitalize on the new wave of technology. And so there's been this question about AI. Clearly an incredible technology, clearly, you know, like the greatest inventions and it's up there with electricity and fire. Like it's, it's really, really cool. The computers talk now. It's incredible. At the same time, what, what is the market dynamic that drives how this technology will accrue value in various places? Who will the winners be? A whole bunch of startups. Will there be a Monopoly player around? It comes from a startup. Will there be a Monopoly player? And then sustaining innovation in every Other mag seven and this is the question that everyone has been talking about for years, for a couple of years now. This is what Ben Thompson writes about, what we talk about all the time.

And I think people are gradually waking up that like it's possible that a lot of the value creation at the foundation model layer will happen at OpenAI because of their consumer products.

Right.

And this aligns with Sam's piece from last week, the Gentle Singularity. It's basically saying like we created intelligence and it's less weird than we thought. Right. Which is a step back from how things were talked about and is a stark difference than maybe AI 2027, which is super, extremely AGI pilled and just saying we're going to continue accelerating and I don't know, I think it's right to kind of read into this and two things can be possible. It's possible that Ilya will create very important lab with ssi, but it's also possible that they might never grow into a $30 billion valuation.

Is that where they are now?

They're currently priced at $30 billion. So for, for DG to, to leave as a co founder of that company. Sure. I'm sure he's getting, you know, he'll get a ten figure package if he goes to Meta.

Yeah.

If this goes through. But he's also leaving, I would imagine, billions of dollars of, of, you know, shares.

Yeah. This seems like very rumor mill at this point. Like this could go a bunch of different ways. It could just be talks and maybe they just come on as like advisors or something or they. Jo, Meta has a board that includes people that don't work at that company and they add a lot of value there. So that could happen. It could also be that Meta winds up acquiring ssi. That would be a wildly different take on this. Or it could be that they leave the information. Seems like somewhat confident about this.

But it's interesting. It's only a couple sources ssi, multiple co founders. It's very possible that as even since the company was started, certain members of the co founding team like Ilya, you know, want to build what, what they see as super intelligence. And it's possible other members of the team are like, yep, this is like, yep, more software. We're, we're going to ven this out in a bunch of different places and ultimately, you know, it just, I don't know, I. You don't typically see people get off rocket ships.

I agree. Yep.

Especially co founders. Unless there was an extreme, you know.

Rift, there's another side to this, which is there's a question about what structure. Even, even if, even if the goal still is super intelligence. What is the, what is the corporate structure and the, and the capital formation structure that delivers superintelligence? Because we saw this with OpenAI when it was a nonprofit, there was simply no way to marshal a $10 billion donation to a nonprofit for a large GPT 4.5 GPT 5 level training run.

Yeah.

There was no way to marshal that type of capital. Like the biggest, the richest people in the world had already donated $100 million and there was not really a lot of appetite for yeah, next year I'm 10Xing my donation.

Yeah.

And so, and so they had to become a for profit.

Yeah.

Then you look at, at the flywheel of what it takes to continue to develop and continue to do these training runs and continue to invest in reinforcement learning. And it feels like you need a data feedback loop and you also need a financial feedback loop to be able to justify more and more investments. And so, so if we're, and we're going to talk to Mike at ARC AGI and there's this interesting thing that we heard which was that the reason that the foundation models are not able to one shot ARC AGI right now is because they're all just like doing a nice thing and not reinforcement learning on it. But if they actually were, did like some fine tuning around it and they were like hey, we want to knock this model off, they could. And what that tells me is that for any really well defined problem like Chess or Dota 2 or Go or League of Legends, like you can go and say hey, we're doing a specific training run for this one problem and it's going to get really good at it. The weird thing is is that the economy and the like, the global value creation chain from humanity is potentially extremely long tailed. There's potentially not just like five skills like oh yes, you know, IMO level math and you're good and you generalize, you, you might need to go and dig into all these different, different pieces of value and, and having a feedback loop or an economic model like what OpenAI has with their app that generates a ton of revenue or like what Metta has where they can deploy these products in all sorts of different ways and get billions of people using them very quickly. That actually might be a more like it might be the only way forward. You might not just be able to go into monk mode, come up with the perfect algorithm and then train it on some like medium sized Cluster, you might actually need to just up scale energy, scale data center capacity and scale users smoothly for decades to get there. So I don't know that this is, this is updating my like probability of.

Super intelligence ever happening is.

It's more like where will it take.

Different labs that are losing billions of dollars a year can the capital market support and for how long?

Exactly.

Right.

Yeah.

You have XAI thinking machines, safe superintelligence.

Yep.

You know, anthropic is kind of an OpenAI are in their own categories and that they are generating a lot of revenue.

It's also hard because it's not like biotech where if you come up with a machine learning algorithm or you come up with the transformer, you patent it and then you just make money off of it forever. Like that's not the way these innovations. If that was the case, I would actually be maybe more bullish on SSI because I would say, well, Ilya is clearly like an incredible researcher. If he goes into, you know, his team and comes up with the next great training paradigm and then patents that and is able to license that to Google and OpenAI, that could be extremely valuable. But that's just not the way the.

Structure of the market is, is not like drug development.

Exactly. Anyway, it's a fascinating story. There's been a ton of reaction to this. Nick says this is somewhat related. Karpathy literally said Meta's llama ecosystem is becoming the Linux of AI and you're blackpilling. And so this was kind of like a narrative violation. A lot of people we should get.

A little bit more into the article because it does give. So. Friedman has been involved in Meta AI's efforts for at least the past year. In May 2024, he joined an advisory group to consult with Meta's leaders about the company's AI technology and products. Earlier this year, Zuckerberg asked Friedman to lead Meta's AI efforts altogether. The person familiar with the discussion said Friedman declined, but helped brainstorm other candidates, including Wang. While Zuckerberg was skeptical Wang would leave scale, Friedman convinced him a deal was possible, said a second person with knowledge of the discussions. As the Wang hiring came together, Zuckerberg approached Friedman again. This time, Friedman agreed to a deal of his own. He is currently expected to report to Wang, who is roughly 20 years his junior. Both men will be a part of a small group of Meta leaders that Zuckerberg refers to as as his management team, or M team. For Gross, the talks with Meta put him in an awkward position with ssi. A Startup form with the goal of building a leading AI company insulated from short term commercial pressures. So again, SSI strategy from the beginning is saying we're not going to release anything until we create superintelligence.

It might be the nature of the economy and the nature of artificial intelligence and the structure of the market. That might mean it is impossible to insulate yourself from short term commercial.

Yeah, the question is, you have billions of dollars on your balance sheet and you hypothetically could just do AI research forever just off of the interest yield alone. Except for the fact that if you want to compete from a scaling standpoint, you have to spend billions of dollars on GPUs and data centers and training runs and things like that. So it'll see there's a very real tension there that will have to be resolved somehow. The startup SSI hasn't yet launched a product or described in detail what it planned to build. Gross's departure for Meta would damage an important investments for some top venture capital firms. In April, SSI raised 2 billion at a $32 billion valuation from investors such as Green Oaks, Andreessen Horowitz and Lightspeed Venture Partners has also raised money from Sequoia Capital. They basically got everybody together, together. Friedman and Gross have invested in some of the buzziest AI startups, including startup Perplexity and robotics startup the Bot Company.

That's Kyle Vogt's company.

The firm had more than 2 billion of assets under management as of the last year, though that figure is likely higher now with the increase in value of some of their startups. So wild. Anyways, this feels crazy, but yeah, Nat Friedman, independently going to work at Meta is not that crazy. I feel like the craziest part is, you know, is someone like DG going from SSI to Meta. But at the same time, you know, it's very possible that SSI and Meta could work out some type of relationship and maybe that's not getting reported yet.

What's interesting is that both of these guys, Daniel Gross and Nat Friedman, were both at one time thought to be like future really, really significant leaders in Mag 7 companies. So Daniel Gross, he started an artificial intelligence company I believe, went through YC and then. Or maybe he went to YC after, but he sold it to Apple. And at Apple everyone was kind of like, wow, now that he's in there leading AI at Apple, he's going to be kind of like this young, incredible talent. Maybe he'll be like the next Steve Jobs. Maybe he'll like take over the company one day. Yeah, and people were kind of like waiting for that. But I didn't like Apple was really.

Set up, accepted into YC in 2010. He was the youngest founder ever accepted.

Yeah, yeah, yeah. And then he went back as a partner shortly after because he left Apple. But there, there is a different like fork in the road where Daniel Gross is like next in line to run Apple after Tim Cook. If they were set up to empower someone young, which I don't think any of these big companies really are necessarily maybe, except for Meta. And then Nat Friedman has the same thing where he's, he's CEO of GitHub, he goes into Microsoft. You know, it was always a possibility that, you know, GitHub's really important for this $500 million business. It's growing, it's code gen. Like he's set up in the tech industry like he could have pot essentially taken over at some point. Yeah, yeah, yeah, that's just copilot. And so, so there was a world where you could see them at the ranks. But we don't, we don't think about it this way because most of the succession plans in, in manager mode, big tech companies are more managers. We don't tend to acquire founders and let them take the helm. But Zuck, it's not like he's stepping aside by any means, but he's very much leaning into this idea of like there is something special about these founders, these people who have built companies, these people who are at the heart of the technology really, really in the midst of things. Get them on my side at any cost. And I love it. I think it's amazing.

I think Nat and Daniel both want to make a dent in the world, especially in the context of AI. Right. So they're not going to want to go to Meta and just Cruise and make ads 10% better, that kind of thing, make it easier to generate this.

Deal and then you just go and rest and rest. I don't think that's gonna happen.

No, I can't, I can't see it anyway. It'll be interesting too. I wonder, you know, would they continue to be able to invest, you know, independently or would they just, you know, or would there be kind of structure that says like, you know, you have to, you actually have to just go all in on this. If I was Zuck, I would hope and expect for that. But who knows?

Whatever they're working on, I'm sure they'll be using Linear over there. LINEAR is a purpose built tool for planning and building products. Meet the system for modern softw development, streamline issues, projects and product roadmaps. And they got linear for agents, folks. Dylan Patel is doing a little meme on this Zuck founder mode master plan. Don't pay the pie torch and LLM people enough. Lose 20% of the torch people to thinky thinking machines. Hire Alex Wong, not Nat Friedman for 10 +billion dollars to help you recruit talent inflection. Back the torch people at 10x their previous total comp. And so Dylan's obviously saying like you should just bet on the same people earlier and kept them unclear how much of it was, you know, really about pay. But clearly that is not a gating issue anymore. The floodgates have opened. The, the, the. This was something that, that we was identified earlier. We, we covered a timeline post about this where it was like, like what, how, how will Apple compete in a world where like they can't, they can't justify paying anyone $10 million a year? Like if that's the new normal or that that's like the value of some of these people that are going to do some of this research, you're going to be kind of hamstrung. And, and it's not because you, you're not spending $10 million on an organization. It's because you're not spending $10 million on a person. Yeah, crazy new thing.

Yeah. Sam Altman was, was taking shots at Meta.

Yeah, that's on this week cover of the Financial Times today. I got it right here.

Sam came out and said Meta started making these giant offers to a lot of people on our team, like $100 million signing bonuses and more than that comp per year. I'm really happy that at least so far none of our best people have decided to take them up on that.

Yeah, the metagaming here is like wild. It's so good.

3D chess. Yeah, it's one of our best people. He's just getting into, getting into Zuck's head.

Yeah, yeah, yeah. Somebody had a good breakdown of that.

He says the strategy of a ton of upfront guaranteed comp and that being the reason you tell someone to join. Really the degree to which they're focusing on that and not the work and not the mission. I don't think that's going to set up a great culture. Altman added. I mean, the only thing here is like tell that to the world of Wall street and hedge funds, where if somebody's just really good at making money, you'll just offer them a maxed out contract to come over to Your team and it's entirely, you know, like motivated.

By the value that they're creating. But it's very trackable. It's a lot harder.

Yeah.

In that case still, you know, it's clearly up there. If you're moving the market cap, like you can kind of tell. Spore says, is Ilya SSI already DOA if its co founder is potentially about to be poached? Good question, Swix, as these guys are already senti millionaires. So we're not talking about $100 million signing bonus anymore. It's the first 1 billion signing bonus in history. This is going to cost. Zuck clearly is in spend mode. If you think about 100 million bonuses are high. This is a guy who lost 14 billion in 2022, 16 billion in 2023, 18 billion in 2024 and 20 billion in 2025. To invest in VR, all he has to do is cut VR spend for 2025 and he has more money than Anthropic has raised in its entire lifetime. Wow. I didn't put it in that terms. Never bet against Zuck long term, but I think we're in for another costly period of investment. And we know what happened last time he went so hard on a thing. We do not have the balls or imaginations to do what he is about to do. Yeah.

You imagine being Tim cook running Apple $3 trillion company, making a paltry 74.6 million in 2024, just looking after, just going through the most brutal year of his time at Apple. You know, like it's brutal pulling the company back from, you know, back from the brink of this like trade war.

Yeah, yeah, yeah. He's actually sitting there tactically, he's checking.

His pay stubs, being like.

He'S like, I'm making scale. Who doesn't even work there. He just got paid out bigger than me.

When you put it into context that that some 24 year old AI researcher who's cracked.

Yeah.

And like, deserves a great role at a great company with great pay, making more than the CEO of Apple.

It's rough.

It's just absolutely brutal. So anyway, we still need to organize this protest, hit the streets for Tim Cook.

We do, we do.

Head over to Cupertino.

Yeah, we do. We should design some posters and Figma for it. Go to figma.com. think bigger, build faster. Figma helps design and development teams build great products together.

Nathan, all these companies duking it out. Figma is powering the design teams of all of them.

Yes.

So, so it's Kind of like one.

Hand washes the other scenario. We'd love to see it. Nathan says Zuckerberg and Friedman to lead AI for tens of billions of dollars was not on my bingo card yet. I don't think many people predicted anything like this happening. I think everyone was kind of saying like, there, there's probably going to be like some sort of V2 of the llama strategy. But being so talent focused, I think was not on the table. It was more like, okay, maybe they do an acquisition of a foundation model lab, or maybe they would just build an even bigger data center since they have abilities there, I think. And it's. And it's been a very different.

We don't. We don't know much about sports, but there's probably. I was trying to think is like Luca, Don Don or whatever going from Dallas to the Lakers. That was a big surprise.

Surprise.

SSI co founder going to meta.

Yeah, this is the LUCA of tech for people who know what reinforcement learning is. Exactly. Luke Metro chimes in dog, how much is Zuck paying? He's over at Anderall right now. The meme has been, do you want to just sell ads or do you want to build something important? Anderall. Well, with these pay packages, I think you're going to get some Android engineers being like, I'm willing to sell ads, I'm willing to optimize ads.

Actually, I see a lot of ads throughout my day. I've always been kind of fascinated by.

That, you know, protecting the world and ensuring like, you know, western LED peace. Creating world peace is like, noble, but, like, at a certain dollar value. Ads are cool too.

Yeah, absolutely insane. Well, I'm excited to see this unfold. The way, I mean, some of it's interesting the way that this reporting is written. It feels at times like it's already happened, but it's clearly not confirmed.

So, yeah, it could kind of go either way. But I mean, we heard the leaks about scale AI like a few days and there was some speculation about what was going on there, and it became very real. And so, you know, who knows, maybe it does become real. But we'll leave tracking it here. Nir Cyan says, ladies and gentlemen, midjourney has done it. It's a new AI image to video. Model Justine Moore from Andrews and Horowitz mentioned this yesterday, but the posts have been going out on the timeline. We have our intern Tyler Cosgrove in the studio today playing with midjourney video. How's it going so far? Can you give us a little Review good.

It's been a lot of fun so far.

So you're in the mid journey Discord right now.

Yes.

Fantastic.

I've actually made. I've made four videos so far.

We can pull those up.

Yeah, let's see. I'm excited to see these.

He's in the Discord.

He's in the Discord.

Live from the Discord.

He's in the trenches. How is the. How is the interface been. You just upload an image. Does it do the same thing that you get with a midjourney image where you type a prompt and then you get four images and you get to pick one?

Yes. Yeah.

Okay. So you get four video results and then you up res the one that you like? Basically, yeah. Okay. I think when you export it basically does that. Got it. Oh, okay, here you kicked it off with. With an image of us reading the paper.

All right, let's see.

So this was, you know, kind of bear domestication, right?

Okay, so. So did you include a prompt alongside the image?

Yeah.

Yes. Okay. You add an image and then you prompt it. Oh, and then you add a prompt. Okay, cool. It knows us too well. That was the first one. Okay, let's see the second one.

Love it.

This is great.

Bear domestication is in our future.

Okay, this is us on our phones. And what is this? Kind of an angel flying over. Oh, the Very, very bizarre. I thought. I was thinking Pegasus.

Bit of a demon vibe. Who's the angel in the back?

That. What was the prompt for this one?

Let's see.

That one. Angel wings or something. Angel flies up behind two men as.

They look back and see smile.

Okay. Yeah.

Okay.

The actual. The actual video on us. Okay, what's this one? What's this one here? This is us in the studio.

A little meta.

Oh, that's extremely demonic. This is super creepy. I don't like this.

Boo. Bring the air horn back.

Oh, that's weird. They steal the gong. Wow. Hold your position. Okay, I think there's one more. Yeah, let's play the last one. What you got? What's this one? Last one was bizarre. Very, very. So the physics are pretty solid. Sometimes you see a bit of the same thing with VO3, where if a.

Car is driving, you'll see the back of the car.

Okay, we got us standing at the pool. Let's take a look at this. Okay, this one's cool. Okay, we're back. The lighting on. That's incredible. Okay, finish strong. 10 out of 10 for Midjourney.

Play it again. There we go.

I love this this is great. That is cool. It looks really good too. Yeah. Always bring your F35 Joint Strike Fighter into the club. Something like that.

Yeah.

But. Okay. That was cool. I like that one a lot. Yeah. A lot less demonic.

Took us on a bit of a roller coaster.

Yeah. Yeah.

I did not like that angel.

That was weird.

That was very weird. Creepy.

The aliens were very bizarre. But you've redeemed it all. You want it all back. Fantastic. Give us a review of the like, overview of the actual experience. How long does it take to generate these? Are you hitting rate limits? How much does it cost? Give us like the breakdown of like, you know, the consumer experience. Yeah, it's really good. I mean, I think so.

I'm on the $30 a month plan.

Which is kind of the mid tier one.

It takes probably 10 seconds.

It's really fast. Okay. VO3 is like two minutes. So you can iterate like super quick. Oh, that's cool. Okay.

But it's very easy to use.

I mean I haven't used. So I'm actually not in the discord. I'm on the website. Okay.

Yeah, but it's very easy to use.

Okay. Very cool. And you can run them concurrently. So I could do multiple. Okay.

At the same time.

That's great.

Yeah.

Really great.

Awesome. Well, very fun. We'll be tracking it more, asking people how. How it's benchmarking, how it's working. We'll have to get. Have some fun with those. I had a lot of Fun with the VO3 ones. We were doing the. The crashing through the Hollywood sign. A couple few too many bottles of Dom Perignon on the back of Ferrari.

Yeah. I didn't like how there weren't guard. Where there seemingly weren't guardrails.

AI safety issue.

Yeah.

I shouldn't just be able to be.

Able to visualize yourself drunk driving. Yeah. Bottles of champagne flying out of the car.

The quality was remarkable. Anyway, mid Journey is having fun on AX, they say. Introducing our V1 video model. It's fun, easy and beautiful. Available at $10 a month. It's the first video model for everyone and it's available now.

How many. How many prompts do you get for $10 a month?

I don't know. You want to look it up?

Yeah.

I mean, I think it's actually unlimited. Okay.

It just takes longer.

Takes longer. Yeah.

But I'll verify that.

That's insane. When you put it into the context of those outputs are in many ways better or on par at least. Like from an entertainment value standpoint, as VO Yep. And VEO is $500 a month. You're still gated on.

I could only do three per day. I have to come back. They take two minutes a pop. Yeah. Speed of iterations really, really key. So I mean that's the whole Discord model is like, is like get people iterating, sharing ideas like to explore the space and figure out what works. Like even just from seeing those four, I feel very confident about its ability to render aircraft. And so I'm probably not going to go and prompt a bunch more alien videos, but I'll definitely be prompting a bunch more F35 videos. It seems to do that really, really well. And so the more people you have making more stuff, the more you learn the guardrails, learn how to use it creatively, and can actually make a better product. But Midjourney was having some fun. Devin Fan from XAI says, I know what I'll be doing this weekend. And midjourney says, what weekend? And will depew says, lmao. Blake Robbins says, mid Journey video is breaking my brain. And everyone's having a good reaction to this. So always fun to have a new AI tool. And we'll be. We're talking to a couple AI folks on the show today. So we'll be running through that, getting their reactions and talking more stuff. Elon Musk posted the the very sad. What is this? The Pipo Pepe or something? It's the green Frog. He's smoking a cigarette. He's not happy probably because RIP4am brutal. We actually Vance we should.

I think a picture of Elon, you know, smoking a heater after one of his rockets blow up. Blew up would be become a timeless.

Meme that would be.

It would be worth kind of his comms team kind of working on putting that together, maybe working with Ashley Vance to get that shot. Yeah, it could live.

So girls say, I can't believe he didn't cry at the Titanic. Do men even have feelings? Boys crying at the sight of ship 36 exploding. Very, very sad. And then Elon says, just a scratch. The entire thing blew up.

It was just a flesh wound.

It was intense watching it. I mean, the ball of fire here is immense. So the starship exploded during a test in Texas. A setback for Mars. Mars for Musk's Mars ambitions. Now, the Mars transfer window is very, very tight. Like you can only get from the Earth to Mars like once every 18 months or something or maybe even more. It's really hard because, like if the planets are on the opposite side of the solar system. Like you just can't, even though you have a rocket, you just can't get over there. So you have to wait until they're lined up and then you can do it.

But realistically, skill issue, realistically true skill issue.

If you build an even faster rocket, you could get there no matter what. You just pilot, steer it around like it's a GT3RS around the Nurburgring, no problem. So the explosion occurred during a static fire test. No injuries were reported, thank goodness. We love autonomy. Very, very happy to hear that. No one was injured during this because it looked horrific and it looked like in any other scenario there would be a bunch of technicians there. But fortunately they were able to do everything remotely, which is great. And then Starship features pressure to meet deadlines for NASA's moon mission and Mars exploration. So there's a big, there's a big NASA moon contract that's very important, very material to the business. Obviously SpaceX has a lot of other business lines, but this one's very, very important too. And it's, and we hope that they can get back on track. SpaceX is making an enormous bet on Starship, which stands roughly 4, 400ft tall at liftoff as it tries to break ground with new reusable rockets. And the paradigm of starship, it's not just a bigger rocket, it is way more reusable. Like you look at the thing, it comes down, gets caught by those arms, can instantly be refueled and sent back up. You're talking about potentially like multiple flights per day. And so the problem here is not can you build a big rocket? Humanity has done that before. Humanity's built a rocket that's roughly on par. We, we've gotten to the moon before. The challenge now is not can we get to the moon? It's the same thing with like, the challenge is not can we build a flying car or can we build, we have helicopters, can we build one humanoid robot or one self driving car in San Francisco? It's like, can we actually scale these systems to the point it is safe to go to the moon and back on the drop of a hat for 200 bucks? Like, that's the challenge. It's more of an economic and industrial might challenge. And that's a completely different challenge from just can we get one rocket to the moon? An exquisite system. We're looking for reusable, scalable, you know, frequency engineering system. So, so good luck to Elon, rebuilding and the entire SpaceX team. I'm sure it's a Huge challenge right now. But if let's do some ads. Tell you about Adio customer relationship. Magic Adeo is the AI native CRM that builds scales. It grows your company to the next level. Get started for free. Adio Calm. What is that? What is that sound at the end? Is that attached to that soundboard?

Guys, I think you botched the action movie sound effect.

Oh no. It has. Okay. In other news, the Los Angeles Lakers has been sold for 10 billion in richest deal in sports history. Guggenheim Partners CEO Mark Walter, who also owns MLB's the Dodgers, is acquiring the storied NBA team in a move that makes it the world's most valuable sports franchise. And it's so funny because like the Wall Street Journal's framing this is like this is the biggest deal ever. No one's ever done a deal like this. And we're like, wait, so you're talking about like a series A for like a foundation model company? Like as a tech person I'm just like, yeah, like $10 billion. Like billion dollars. I mean we should ring the gong. But it's not exactly like the first time. It's not even the first time this show we've heard a decacorn congratulations to.

The Lakers, Mark Walter and the whole team.

It's fantastic. Major premium to the Boston Celtics who sold for 6.1 billion. And now the Lakers is the most valuable sports franchise. But they just don't do enough volume. There's only a couple games. They're not 24, seven like Instagram. Does that ever go offline? No, there's always entertainment Lakers, they're still doing seasons. They need to have 24 hour basketball. They want to really get there around the clock.

It's like endurance, endurance basketball. It's just a week long game, you know, gotta always have five players on the court just constantly running up.

It's the only option. Jeannie Buss and her family who have owned the Los Angeles Lakers since Jerry Buss bought the team in 1979. Wow. On Wednesday agreed to sell majority control of the storied team to Mark Walter, the sports investor. And I looked at the return on investment of owning the lakers for that 40 years. Slightly under S&P 500. It was a really, really good deal and it was a really great company that grew a lot. But it didn't outperform the stock market. Just diversification. Bros. DCA Bros. Undefeated again. Well, if you're trying to DCA do it on public.com investing for those who take it seriously. Multi asset investing Industry leading yields the trusted by middle millions folks. Anyway, Walter, who is part of the ownership group that owns the Dodgers, has been part of the Lakers since 2021 when he purchased a 27 minority stake in the franchise. He's also a co owner of Chelsea in the English Premier League, the WNBA's Los Angeles Sparks and the new newly formed Cadillac Formula one team. Let's hear for Cadillac. Let's go. Congratulations John.

You know, at all John front run the. The Cadillac F1 team and got a Cadillac for himself there. You can see the.

It's great to have an American F1 team in the business now. Yeah, yeah, we've, we've fallen off but we're coming back.

You're not going to be able to get one of these in the whole country.

I don't think so. They're gonna be too popular.

F1 team, you know, gets out on.

The sale marks the end of nearly a century of Lakers control by a family that has become synonymous with Los Angeles sports and the glitz of professional basketball. The deal also comes at a time of skyrocketing valuations in professional basketball which haven't come back to earth since the league announced a media rights deal last year with worth 77 billion. When the Celtics sold in March, the $6.1 billion valuation exceeded the previous record valuation set for a sports team by the 6.05 billion sale of the NFL's Washington Commanders in 2023. He purchased Lakers for 67 million. In 79. 1979 the team transformed from franchise uprooted from Minnesota into one of the winningest and most valuable sports.

I had no idea that they were founded.

That's where the lake, the lake name comes from. Minnesota is the land of a thousand lakes. They were the Lakers because there's a lot of lakes in Minnesota. And then they just put them to. They just brought them to LA and kept the name. But that's what Lakers means. Yeah.

Wow.

Bus. The Buss family oversaw the creation of Showtime and presided over the NBA's last three peat a listers like Jack Nicholson and Leonardo DiCaprio have become fixtures at the games. And when they sell merch, they need to pay sales tax. They should get on numero.com numeralhq.com sales tax on autopilot. Spend less than five minutes per month on sales tax compliance.

You know they've won 11 championships since 1980. Their rosters have boasted many of basketball's brightest stars. Magic Johnson, Kareem Abdul Jabbar, Kobe Bryant, Shaquille O' Neal, LeBron James and LeBron James's son have all worn the Lakers purple and gold.

I love it. It's such a cool. Yeah, the, the father son duo. I mean I feel like that should have been a bigger like national news story. It's such a cool thing. I think it's like not like if they were like winning championships together immediately that might be a different story. But it's just so insane.

Yeah.

That you could be playing well professional.

Because you could have earned a better return by DCA into the. Into the stock market.

That's not why people own these assets though.

Owning the Lakers for a number of decades I imagine was absolutely priceless. So great investment, get the owner's box, great run.

Yeah. All the perks. You have to add those in. What do you get perks from DCA.

Into the I like how Lakers legend Magic Johnson hit the timeline. Said just like I thought when the Celtics sold for 6B, I knew the Lakers were worth 10.

Let's go. The confidence of Magic Johnson. Great investor too. He's got a bunch of good stuff in the portfolio. Anyway, more news on the Scale AI transaction. So it's closed. I believe that Alex Wong has a badge at Meta and shows up to work in Palo Alto and clocks in@MEDA HQ. Now scale AI is still an ongoing concern, still the company, but every competitor is out for blood and they want to take as much of the business as they can. Since obviously the perception is that Scale AI will primarily be working with Meta and that other foundation Model Labs might not want to do business with Scale AI anymore. Unclear if they can separate out the businesses. If they can separate them out fully over time and sell the position to other investors, create a diversified. I mean they could even take the company public. At which point I imagine that it would be a lot less, a lot less of a conflict of interest or like a fear. But there's been news that OpenAI said, hey, we're not training, we're not using scale AI for data anymore because it's too aligned with our competitor Llama maybe, but everyone's trying to.

Yeah, a lot of this is very predictable. Right. I don't think Meta and Scale's teams looked at and said, hey, if we sell right now to Meta, which is competing in open Source AI, we're totally going to retain all of our customers. Right. Like people aren't just going to immediately turn off and no, they were smart enough to know what would happen. And there was an article I think yesterday about OpenAI ending their relationship with Scale. But from what we knew they hadn't been doing much for a while. That's part of the reason why Merkor had been and they also brought a.

Big function in house because for some of the more complex tasks it makes sense to generate the reinforcement learning data yourself and there's just so many other services having a single point of failure never makes sense for a business of that size. But. But we'll see. So the the Information has an article here about a little known startup that has surged, hint hint, past Scale AI without any investors. This is interesting. After Meta Platform Scale AI deal, Data labeling is looking like Silicon Valley's hottest new interest. That's enormous opportunity for Edwin Chen's Surge AI. For years, data labeling existed in a tucked away corner of Silicon Valley, a critical but unglamorous area of AI where companies like Google and OpenAI how hire outside firms to improve their models by laboriously grading the quality of what they produce. Now a spotlight has unexpectedly fallen onto the field in the wake of Meta Platform's decision to pay 14.3 billion for 49% of scale AI, the best known data labeling firm. But it's not the largest such firm, nor perhaps the most impressive. That title belongs to Surge AI, founded by Edwin Chen. This is fascinating. I didn't know this. 1 billion in sales last year bigger than Scale? Yeah. So Chen startup has won customers like Google, OpenAI and Anthropic.

It's such a testament to the idea that like sure you can bootstrap, but it's so incredibly hard to have any hype around your business if you're bootstrapped because you're not having your investors aren't hitting the timeline for you on a daily basis. And also if you're not trying to raise capital, you have less need to go and be loud and go on podcasts and talk to the press and all this stuff because you're just making a lot of money and sometimes it can be beneficial for people to not know about you.

So this is crazy, crazy stats. So Chen is 37, he has no investors and has bootstrapped the five year old startup entirely by himself, which has 110 employees in offices in New York and San Francisco Francisco. The company generated more than $1 billion in revenue last year, Surge has told employees a previously reported figure that exceeds the $870 million scale generated in revenue during the same time period. And unlike Scale, Surge was profitable and has been from the beginning, Chen said. Moreover, Surge could see its sales get even larger if other companies copy OpenAI's decision to stop hiring scale, a choice made over concerns about scale's relationship with with Meta to shift business to Surge. Other key financial metrics couldn't be learned, like how much revenue Surge keeps after paying its workforce of mostly contractors. So there is a question about like the margin. Since this is somewhat of a marketplace business. This could be a situation where, you know, a thousand dollar contract comes in and $800 of that contract goes to the actual contractor who's doing the work of the data labeling. But at the same time, even if it's 200 million in like, you know, like net revenue, that's still a huge business. It's hard to imagine Surge not being a fantastic business. They haven't had to raise Money, they have 110 employees and they're used by Google and all these major foundation model labs. So seems like a fantastic business. But if Surge could earn a valuation from investors similar to the one scale received from such a price would make Chen a billionaire many times over, at least on paper and quietly one of the wealthiest people in tech. And interesting. I'm very interested to see what, what he did before this company. Edwin Chen. I feel like I've heard that name before, but I don't know. As AI models transform from toys into real business tools, data labeling is becoming more and more essential. Contractors hired by like, by companies like Serge grade the responses from AI models and write thousands of questions and answers in fields like programming, math and law to feed those AI models. And so you know, if you're, I wonder if this is going to go the route of, you know, you are Deloitte or McKinsey and you're going to have your team, but then also a company like Serge create a ton of training data around a specific workflow that is costing your business, you know, 20 or 50 or $100 million every year. And then so it's like instead of like the, the AI BDR that's like kind of generically writing emails based on like the average of the entire Internet, it's like, no, this is a fine tune for your business. Perfectly trained. Perfectly. And it really, it really distills what you do excellently. Yeah, I don't know. I don't know if it'll go that way. I'm interested to talk to people about it as AI models. So Serge's subsidiary, Data Annotation Tech says workers get paid to train AI on your own schedule, with wages starting at $20 an hour. Chen has distinguished Surge by making the high end shop charging Premium rates often 2 to 5 times what scale might bill Surge justifies the prices with its reputation for industry leading work. Indeed, one former Scale employee said Surge often performed better than Scale in customer audits of labeling quality. And competitor Garrett Lord, who's coming on the show today, who runs Kleiner Perkins backed Handshake, readily acknowledged that Chen is the number one player. So I'm excited to talk to Garrett Lord today about this exact topic. Should be very interesting. You wouldn't know that from the from the coverage of Meta's blockbuster deal to quasi acquire Scale AI, its CEO Alexander Wong, who is now joining Meta in a senior AI role, was widely regarded as the leader of the data labeling field and had become a Silicon Valley celebrity, blanketing podcasts and conferences with his presence and posting heavily on X. It also raised 1.5 billion in venture capital, putting Scale on a very short list of companies that have raised that much. And he hired upwards of a thousand people. Wong had made time to exit perfectly given the traction of Surge, which had grown larger than Scale. Without outside capital and with a tiny fraction of Scale's workforce. Scale also missed the goal to hit $1 billion in revenue last year. But Scale Scale spokesperson said the scale wasn't profitable. It was not profitable which, but wasn't burning a ton of money. Like I think they had efficient. They raised 1.5 billion and they still had like almost a billion in cash.

Yeah.

So they weren't, they weren't in like trouble or anything. But at the same time it was like, like not, not a wildly profitable, not a wildly lean business.

But I don't know what, what, what a. It's absolutely fascinating about these two businesses.

It's a wild industry, something that like yeah, I mean just it feels like there's such an edge just to even identifying this opportunity years and years ago. I mean I guess search started four or five years ago, but it was certainly like pre chatgpt that all these companies got started and then they realized like some of them got started in self driving, car annotation, all sorts of things of stuff like that. But Chen studied linguistics and math at mit, came to the idea for his startup after leaving college and witnessing firsthand how big companies struggle with data. Before starting Surge, Chen worked as machine learning engineer at Facebook, Dropbox, Google and Twitter. He worked in four different tech companies, just like going from one to the next. That's insane. He was developing recommendation and search algorithms and helping gather the data needed to train them. Despite the hefty resources of those companies, Chen encountered a lot of problems at Facebook for Instance, Chen was tasked with helping build a Yelp competitor. His team needed to train a model that correct could correctly classify businesses telling the difference between restaurants and grocery stores, for instance. To do so, they needed a data set containing 50,000 accurately labeled businesses, which he found out would take six months for an outside firm to assemble. We had no solution other than waiting. We simply waited. When the data came back, Chen blanched. In some instances, it had labeled restaurants as coffee shops and coffee shops as hospitals. The data was complete junk. He wouldn't say which vendor Facebook had used in 2020. In 2020, he left Twitter to found Surge and picked up some of his first customers, executives from Airbnbs and Neva, a once promising AI search engine startup. As only as only a founder in San Francisco might, bumping into them at rock climbing gyms in the city's dog patch neighborhood and the Mission District, talking about up his startup. To get Surge going, Chen recruited data labeling contractors he knew from his previous roles and funded the startup using his savings. He wouldn't say how much he put in. Fortuitously, Chen focused on language modeling. Scale, by contrast, started out using more visual data for autonomous vehicles, which we talked about just as those types of models began to grow in importance. Less than a year later, OpenAI had hired Serge to fine tune his models by teaching them how to avoid producing harmful responses like a racially biased language. Biased based on research paper the company published together in. By 2022, anthropic had become a Surge customer.

So they're putting out research papers with OpenAI and still managed to stay this under the radar.

Wow. Yeah. So look at this. The label largesse. Data labeling has proved to be a lucrative niche in AI. Surge, founded in 2020, has over a billion in revenue. Zero funding. Scale, founded in 2016, has 870 million in 2024 raised. Oh, the. This. This says funding raised, but this is clearly valuation or something because it says 17.4 billion, which is not what they raised. Turing has 300 million annualized, raised 225 million. Invisible Interesting.

Turing too, initially was like a marketplace to just hire developers. And I think they pivoted into data labeling. Data labeling, interesting.

It's the same thing when I work with a cloud provider. The enterprise tech customer said, I don't know the internal expectations for why their services work so well. I push a button and I'm glad for the internal work to make that happen. Data labeling companies typically use various techniques to make sure contractors aren't just dialing it in or phoning it in. I guess when answering questions, for instance, companies randomly insert questions that have no correct answers or make sure labelers agree on the right answer to a question. So obviously you scaffold up these, these, these, these like responses so that everything's like double checked and then you can kind of see if people are messing around. But wow. What, what a beast of a business. No idea how big this thing is. Amazing. How'd you sleep last night? I'm on a comeback. I got an 89. Go to 8sleep.com getting a sleep.

Know what I got?

5 year warranty, 30 night risk free trial free returns, free shipping. What'd you get? You got 188. Let's go. Soundboard. I demand soundboard. Ashton hall, let's go.

John did it.

I did it. Beat you by one. I had a terrible night. I also took a nap after I got home. It was fantastic. We have our next guest coming into the studio. Mike from Ark AGI breaking down how all the different models are doing.

Last time he was supposed to come on Elon.

Yeah.

Trump decided to get in a massive timeline war.

It was brutal.

John wanted to power through it. John, the people in the YouTube chat are demanding, demand that we do a time.

If we don't do it, they will put up attack ads against us. They will buy billboards against us. They will go to ad quick.com and put up attack ads on TVPN if we don't cover the Trump Elon dust up. So we did and now we're getting those folks back on the show today and later. But if you want to take out an attack ad on us, buy a billboard and ad quick out of home advertising made easy and measurable. Say goodbye to the headaches about of home advertising only. Ad Quick combines technology out of home expertise and data to enable efficient, seamless ad buying across the globe. We should do some timeline. I want to do this blank street story, but we can do that later. Let's dig through some what's in the timeline while we, while we wait for Mike to join. Oh, we have Mike in the studio now. Welcome to the show. Mike. Good to see you. How you doing?

Good.

Thank you so much for.

I'm glad there wasn't a major breaking drama story today. I was actually able to show up.

Yes, yes, you.

I, I don't know if you watched that show at all. But like I was just sitting here. John was so locked in, wanted to just keep doing the show and I was messaging him like now we like.

It was a terrible day to launch anything New we like launched something. I saw several startups launch stuff like regrets to everyone who tried to get anything out that day.

Actually, I remember them. It was Rahul and.

Yeah, that's right. Julius had a launch.

And what's the voice Cloning Co. 11 Labs. I think they launched well and then.

I think Lulu said something. She was like, if you have bad news, today would be a good day to drop it. And then OpenAI actually flagged like, hey, we had this massive.

Oh yeah, dust up with the government. Right where the government was like, you have to give us your all all chats even if they're private.

That was like serious. Like, I mean that got. We're still looking at actually, you know, end results of that. But that, that went really deep into the world. I feel like much more than, you know, kind of maybe even got reported on. Like every single chat thread I was a part of was basically like, hm, should I like stop using ChatGPT as much?

Yeah, I don't think to ChatGPT because it feels like Anthropic has a similar policy. It seems like Google might have a similar policy. Like there was that story a year ago about a man who was using a Google phone with a Google FI cellular connection and had all of his data stored in Google Drive and Gmail and he took a picture of his child to send to a doctor and it was kind of like a nude photo of the kid to inspect the child for like a physical medical problem. And it got flagged as child abuse material by an automatic system. And the automatic system basically de platformed him from everything Google. And so he lost his email, his phone number, his. All of his drive stuff. And it was like a false positive, but it was really hard for him to get back through there. And so I guess my question is like, like it seems bad when we hear the story in isolation, but maybe the problem is not the individual company and it's instead like the government policy. And this applies to all the different companies. But I don't know. 2.

Two things can be true. One is that it can be a massive overreach by that court. Say, you know, basically you need to eliminate privacy on your platform and you can simultaneously have questions around maybe I should use this product in a different way.

Totally.

And it's. The inflammatory nature of it is that people use chat CBT as like a confidant that Totally.

Yeah.

And tell it things that they wouldn't tell anyone. They wouldn't tell anyone in their. And they're. And they're having those Conversations. And I think that's why it's such.

Struck such a chord because like, that's true. Um, I, I just saw some reporting from COTU this week that like chat minutes per day are up to like 30 minutes a day now on usage. Um, and like it's not, it's like closing the gap with like Instagram, which is just sort of nuts to think that like, I mean, who would have thought a productivity tool would ever like be on par with like a social like media app, right? In terms of daily usage, it's so small.

But, but the interesting thing is it's filling a similar void. You know, it's like it's delivering digital companionship in maybe the way that social media products historically did without any social element to it at all. Just like one to one. It's interesting to think like, you know, we went from like, you know, what your friends are doing is like the most interesting thing to like what the Kardashians are doing is like the most interesting thing to like. Actually maybe the most interesting thing is like this person that knows everything about you and is always on and always willing to talk and, and you know, you know, who knows?

Yeah, I think the consumer habits are being born around this stuff today.

Yeah, yeah. I mean I find myself all the time, like instead of scrolling YouTube looking for an inform, like an interesting video essay to explain how, I don't know, like global shipping lanes work or something like that, just going to ChatGPT and saying, hey, like break this down for me and then I can just ask a follow up and dive exactly to the layer that I want. And so yeah, I'm definitely in that camp of using ChatGPT just as like an exploration and entertainment education tool, an infotainment tool tool, much more than Instagram right now, at least for me. But enough about that. What is new in your world? How should we frame kind of the current horse race between all the foundation labs?

Yeah. Okay, so I'm going to share a link. I don't know if this is something you all can pull up, right?

Yeah, yeah, we can pull it up.

So this is a post that we published a little over a week ago. So I think there's been this really big like what's the frontier right now in sort of AI progress? Right. The massive shift in the last six to nine months has been moving from this regime of like scaling up pre training with more and more labeled data into these like test time, consume test time, compute test time adaptation regime. People call these AI reasoning models. Right. We're getting these models time to think out loud. Additional data. Every major lab now pretty much at this point I guess except for Meta, has one of these systems that we've been able to test and report results on. And I think there's some really, really interesting stuff we're starting to see. I think the most notable thing is that like there's not an absolute clear winner across the sort of like landscape right now. There's basically a sort of prito frontier that's emerged. One of the most important things, if listeners are like listening here, I think you should take away is that anybody who gives you a benchmark score on an AI system that is a single number is just marketing to you. Because the reality is now with these AI reasoning systems you have to report score on a two dimensional action. You have to consider cost and efficiency alongside the accuracy. And all these different lab providers have come out with different AR reasoning systems that score differently. They're trading off cost per accuracy at different points. So if you want just like the absolute highest, highest Ross horsepower sort of cost and time is no option, O3 high is going to be your clear winner today for that. But if you're somebody who's saying like, you know, hey, I want to plug in an AR easing system into an existing product, I have where I want like faster answers and I'm willing to sacrifice some raw horsepower for generality for like quicker response times lower, lower cost. You might look at something like grok or Gemini 2.4, 2.5 for thinking there's not like a single like best best answer which, which I think is pretty interesting. And we haven't seen like this sort of frontier is I think what all the labs are working to figure out. Okay, how can we get accuracy as high as we can, but also we got to try and keep cost as low as we can down in the human efficiency robs.

Yeah, I've noticed that more recently with kind of my default usage in ChatGPT 4.0 seems super fast. But I always am thinking like, oh, I should maybe put this in O3 Pro, but do I want to wait 10 minutes? And I'm making that kind of like economic calculus there. Even though because I'm on the pro plan, it's not an economic cost, not a direct. I'm hitting a $5 13 minutes. Yeah, exactly. And so I'm kind of like doing a 4.0 thing over here and then switching back and forth. It's very, it's very odd paradigm that we never really had to deal with. In computing necessarily before I mean I guess like if you were downloading like the 4k illegal blu ray versus the SD which we of course we never did purely hypothetically but if you're on a torrent site there was a time trade off between watching a screener.

I think this is actually one of the reasons these AR reasoning systems I would assert and I don't have inside baseball on the data but from the outside looking I think there are some interesting suspicions that would suggest that these AR reasoning systems, at least today in their current form, have relatively weaker product market fit compared to the non reasoning based systems. The pure language model based things.

Interesting. That's a huge violation of the deep SEQ narrative that I felt was really bubbling up was deepsea came out with like the first just like open access reasoning model. Like reasoning had been tucked behind the OpenAI paywall and so the pro users were familiar with what reasoning models could do that everyone was very excited about them in tech or in the early adopter crowd. But. But Deep Seek when that app came out and you could just install it and instantly see the reasoning chain, it felt like everyone's like oh, everyone's going to be addicted to this forever and this is going to be the new paradigm. But it seems like that might not necessarily be happening. Jordi, do you have something or I wanted to talk about like spiky intelligence and how that plays into this. We had this. Someone came on and said like I think it might have been Sholto actually talking about ARC AGI just saying like hey, all the foundation labs kind of have like a truce that we won't reinforcement learn specifically against Arkansas. I don't know how real that is from your perspective, but it feels like increasingly we might see like very tasked task dependent RL runs kind of chipping away at specific things like like IMO level math is something that clearly like there's a ton of work to be done on but we don't have as many verifiable rewards for for poetry or comedy writing. And so that'll be a little bit messier and later down the road maybe. But at the same time there's probably other verifiable rewards that are just smaller pockets of value here and there for these little micro tasks. And so I'm wondering if we will ever see like the marketing language around these models evolve. Like Grok kind of did this with like we are the anti woke one, but that was more just in the overall like temperature the vibe of the model. But I'm wondering if There'll be an idea of like this one's really good at math, this one's really good at research, this one's really good at that or if, or if they're all kind of going down the same path with what they're trying to solve.

I do think you probably are going to see some domain specialization. I think my guess over the next 12 to 24 months is that you'd see some domain specialization. Benchmark scores diverge because of how all these labs are starting to do the next evolution of training which is they're using RL environments to generate synthetic cot traces, doing their sort of model trainings on that data and they're trying to go get it on a lot of different, just different domains. The original O3 paper I think was interesting on the benchmark results where on this new COT reasoning system they had relatively high scores on math and coding. But the gap or I should say the step function increase in those scores was much higher than the increase in legal reasoning which you would maybe intuitively guess or think expectations that like legal reasoning would probably be one of the best like general domains for if you trained a reasoning model that was really good on math and coding like it should be like and it's a language model that like that would directly transfer into like the legal domain because like okay, it's symbolic reasoning that's like self consistent and that wasn't the case. So I think that's, I suspect that's what we'll see there. You know there's obviously the big scale news. The thing that I'm seeing now is there's probably like, I don't know, a handful that I know of these new startups that have come up in the last several months, but all getting founded to basically go build RL environments to generate synthetic or semi synthetic data and like selling them to sort of the major labs which are the major frontier folks building these next gen systems. I think we're going to see more of that. I expect that's kind of where it's going to drive a lot of areas.

What are you. Even more.

What does the data labeling market look like today? We were covering surge AI, which a lot of people weren't familiar with. I'm sure I'd seen it at some point but I was certainly not familiar with it until we covered it today. What do you think the labeling market looks like in, in five years? Do you think that scale was getting out kind of at the perfect time? You know, I'm curious, I think the.

Timing was Pretty good. I mean like look, the macro change here is, is from a regime where like we're pre, we're scaling it pre training. We want as much text, as much high quality label text as we can get our hands on to, to scale these, these foundational models into one where we're trying to train process models or make the foundation models really good for process thinking and cot generation. That is a complete shift in how you want to generate that data. You want an RL environment that you can create lots and lots of COT traces, really very long traces over long running tasks as well. And you can feed all that right back in and then take advantage of the scaling laws. We already know about language models and how they work and how the performance increases as you can get more examples of the data there. So I guess macro bet would be sort of the trend is heading down towards like the pre training scale label stuff.

Yeah.

So significantly higher on RL environments.

Yeah, so, so when, when, when you say RL environments, what you're talking about is moving from a paradigm of I go to a data labeling company and they hire a ton of contractors to generate new text or verify the responses or grade the responses from these models to I am now hiring top machine learning engineers, AI scientists and having them design an environment that the reinforcement learning can happen like autonomously within the system. Right. Or effectively. Like these new startups that you mentioned, they are taking the massive like hundreds of thousands of contractors like out of the loop for the next runs. Is that correct?

Yeah, it's synthetic or semi synthetic in some cases. Example companies here, like Mechanized Work is one that got started recently that's doing this stuff. Morph I think is doing this stuff. Habitat's another world environment that's sort of doing similar stuff.

Sure.

There's just a lot and it's like very emergent all these, many of these got founded in the last like couple months.

Yeah.

And I think that's a function of the demand and the pull from a lot of the frontier research groups that are wanting this data to sort of do their alternative stuff.

So, so do you, do you imagine companies like Surge and other players would try to pivot into this if, if they're expecting less?

I would expect that founder led companies like that would recognize this as growing part of this and have bets in it if they don't already?

Yeah, I was always wondering like the, the whole story of scale was kind of a series of like various booms in training data. Like the first one was, was data labeling for autonomous vehicles and it seemed like that grew very, very quickly. And then the, the training paradigm around Waymos kind of shifted away from, hey, we need more and more training, labeled training data to something else. And just having the cars on the road and generating real world data from that. Then there was like the second era of like the, the pre training generating the data for RLHF and the big boom. There's OpenAI and Meta, both big customers throughout that cycle. And there was kind of a question of like, what's the third act for all of this? And I was wondering if the. Is it possible that there is a third act, but it's just something like humanoid robots or something like that, like put a bunch of people in mocap suits and generate a ton of training data for what it means to pick up a soda 25 times in a row. It'd be a very different like training data product, but at the same time we have mocap suits and maybe that's relevant or maybe that's ridiculous to think. I don't know. What do you think about that?

One definition of intelligence is an information conversion ratio from the amount of information you have to an action policy decision. The intuition here is you can make a perfect decision given a set of data or information that you have. And oftentimes the right thing to do is go collect new data. And so once we actually start peeking out on intelligence capabilities either plateaued because of research or plateaued because we've actually got AI that's close to AGI. The limiting factor then becomes like the ability to acquire new information, new data. And so in synthetic or on the Internet, that's going to be a function of like, you know, in sort of software bits world. And then the one beyond that is going to be, well, how do you literally go make contact with like reality, the universe? And that's your, that's your feedback mechanism to get new information into the system. So you can like increase your overall intelligence.

Yeah.

What do you think? We had George Hotz on the show a few days ago and he was talking about this efficiency problem where like, if you took all of the, if you took all of the conversations that I had ever had and you transcribe them, it would be like a few megabytes of data. And I'm able to generate some level of intelligence, you know, based on that.

And the golden retriever's level.

Yeah, the level of intelligence of a golden retriever. Yet an LLM needs, you know, effectively like, you know, terabytes of data.

Sure. The sample efficiency is very low. I mean, this is a true statement about just the paradigm. Deep learning compared to program synthesis is the bet that, that we're making at ndea. Program synthesis is a regime that's much more sample efficient or generate models that can generalize out of distribution. But I think it's a completely statement, it's a very damning statement that we've got AI today that's trained on some colossus of all of humanity's knowledge and text over the last 5,000 years. It's on the Internet and what new ideas have they produced? And maybe I could point to Alpha Evolve, which I think is very, very impressive frontier AI system, you know, and it's legitimately finding new knowledge, it's creating new ideas, you know, verifiably. But, but they're very small and they're on the margins of things that we kind of like already have been doing, right? Matrix multiplications, things like this are in kind of in the regime of things that we kind of know about and can define and spec out for these systems. Whereas like, if I took either of you guys and I gave you somehow the superhuman capability to have like all of humanity's knowledge in your head at the same time, like, I think you'd probably be able to produce at least one idea like connect two random divergent domains or like, oh, hey, this kind of looks like this Here's a new thought, right?

Yeah, totally.

That still feels like something that I think, I mean, it's very exciting to build towards it. I think that's what we all want, right? That is like egi. It's capable of invention, discovery that will actually increase the rate of scientific frontier innovation. But we don't have that yet.

Switching gears a little bit five years from now, do you think the average American will pay for an LLM subscription?

I think that.

The.

I think the cost is probably going to go down far enough where that just gets built into the subsidy of whatever the product is and the revenue stream is attached somewhere else. I haven't thought deeply about it, but that's like my off the top of.

My head thinking, yeah, yeah, we were talking offline this morning about just this dynamic of like, like the average American will actually churn from HBO Max. Because at that moment in time, like they, and it's $20 a month or some, whatever the fee is, at that moment of time, there's not a show that they really love. So like, yes, there's a lot of like value there, but like they're just like, yeah, I'm just Yeah, like I.

I don't, I mean I still know so many people that don't even have subs that are like, like my, my wife like use catch all the time and she's on the free plan that's like enough to get a lot of value. For what? For what she does. So I sort of, I sort of respect that, you know. Yes, they're going to be, they're going to be power users and those are going to be the folks who really are doing like amazing powerful things and stuff. I suspect the base rate is going to go down enough where you know, it's going to be more embedded across like almost all the products and experience you have as opposed to being like, you know, a dedicated thing you're paying a lot of like one off cash for.

Yeah, yeah.

Now you might buy products that need it, like robotics or you know, things where intelligence is built into. I think you're going to be product categories that emerge and like people will go buy those but like paying the subscription itself I'm not as confident on.

Yeah. The other thing that stood out to me today specifically was Midjourney came out with their new video model. It's very good. It's $10 a month for effectively unlimited prompts. And you comp that to Google's VO3 which is $500 a month and you're still heavily gated. And it just seems obvious that five.

Years may not be the right timeline for that prediction. By the way, it might be longer than that.

Sure, sure, yeah.

There's so much use case diffusion. One of the things we're seeing with Zapier AI which is growing quite quite strongly right now. It's on the exponential growth path for the AI usage and AI zaps. I've looked at this and I've been wondering is this a function of the technology getting better or use case diffusion? And I've looked at the usage and most of the majority of the usage is still in four or cheaper or worse models right now with people bringing system like AI putting AI into the middle of automation. Interesting. And so I'm pretty confident that a lot of this like agentic automation right now is actually not being driven as a result of like technology progress from AI at AGI. But more about the market is just starting to learn, finally learn. Okay, here's what we can use it for and can't use it for. Here's what it's good at, not good at. And it's very similar to the option curve we saw in the early Gazapier where once you learn what the tool can do, you carry a tool forward with you in time and then you encounter a new situation or circumstance that you can apply your tool to it and so you like we create use cases over time. So I think we're still very, very early.

Yeah.

You know, but there's, I suspect that a lot of the usage increase, 30 minutes a day, even on ChatGPT is a function of just use case diffusion. Less tech progress.

Yeah. Where do you think XAI will eventually need to generate a lot of revenue? Where do you think it'll come from?

I mean if they make progress towards AGI, it's probably going to be enabling all the other services they have around like Elon's ecosystem. That'd be my guess. Less. Less selling it as a direct product itself and going head to head, you know, on cars, rockets, robotics. Like there's so many places where I think you would want to use and have the like product shape where you can use higher degrees of intelligence. You're not bound by just like, you know, the fastest consumer experience you could deliver. I suspect that might be actually where most of the value is at least in the near term. Who knows over the long term. Think the shape wall is proxy.

That was in many ways my long term thesis around the LLAMA project and the superintelligence team at Meta is that there's just so much work to do at Meta broadly that's enabled by AI that if you can avoid the long term OpenAI bill, that's probably worth billions and billions of dollars because of how AI is going to infuse into every single corner of their entire ecosystem. And it's all at such massive scale that the cost of, of using other vendors might be in the billions. And so just looking at the savings there might make sense. I don't know.

I mean I think the most important takeaway, I think I shared this last time I was on with you guys, it's still true today, is that we are idea constrained to get to AGI. This is what ARC's V1 data shows is what V2's data shows. V2 is completely unsaturated. We're not even talking about efficiency, just nothing can do it. And V2 looks very similar to V1 even on like hyper specific solutions on Kaggle, the ARC price 2025 contest progress has been slower this year than it was last year. We are very much still like the thing I can state most confidently and assert most confidently is that like we need new ideas. There's Some major breakthroughs we have not. We have not figured out yet.

Does it worry you that that could take years and years and years? And what happens to.

One of the reasons I funded the prize last year, I wanted to, like, correct the market narrative here. Like, I spent a lot of time with students, a lot of young researchers, and like, at the beginning of last year, there was a serious vibe of like, oh, it's all figured out. I'm not going to go do AGI research. I'm just going to go work at the application layer on LM stuff and make a quick buck before AGI gets here.

Interesting.

And that is a boy. You know, look, if you want to live in a world of AGI for yourself or your kids, like, I think you. What we should be trying to encourage is to design the like, strongest global innovation environment possible. And that's one where there's a lot of diversity of approach, a lot of different ideas being taken, a lot of sharing, you know, kind of what AI looked like in the 2010-2020 era. Right. Very open approaches, how we got the transformer to GPT1 and GPT2 and so on. For today, I'm optimistic. I think the last six months have looked a lot better than the previous two, three years. I think the AI industry is maturing actually quite a bit on this topic as well. Being more tolerant and recognizing. Okay, we don't have it all figured out. There's more ideas we need. That's been encouraging and I think it's seeping down at the low levels too. But yeah, my sort of broad view is any capable human who has new ideas to work on AGI, like, that's the most important thing you could be doing at this point in time.

That's amazing. Thank you so much for stopping by. This is always fascinating.

Yeah, great catching up, guys.

Thanks for having me. We'll talk to you soon. Cheers, Bye. Next up, we have David Hahn from Sequoia Capital. Coming in, David Kahn, first time, first time on the show.

Exciting.

Coming in, wrote a fantastic piece. I want him to break it down. Would you mind kicking us off with a little bit of an introduction on yourself?

Hey, guys. Good to see you. Yeah, I'm one of the partners. My name is David Khan, one of the partners at Sequoia. Excited to chat with you guys.

Yeah, thanks so much for hopping on. Kick us off with the new blog post. What was the thesis, what inspired it? And then I'm sure we'll tie it to a bunch of news.

So the new blog post was about AI companies or AI labs being more like sports teams. And of course we all probably saw seeing the news around scale AI acquisition, some inspiration coming from that. And then these rumors that we'd been starting to hear over the last few weeks and finally now bubbled out over the last couple days into the public conversation around $100 million signing bonuses, huge amounts of money being spent on top AI talent. And for me, I mean I write these pieces as I think about and learn about AI and what an exciting time that we're living through. And I'm pretty fascinated by kind of the human dynamics of it all. There's seven to 10 people at the top of these big tech companies. They control the big magnificent Seven are now a third of public market cap. They're extremely powerful and important. And I think there's sort of sometimes in AI this notion that AI is super abstract or these things are inevitable, but actually it's human dynamics. It's sort of this game of 3D chess that's being played by these really fascinating individuals. And so as an observer on the sidelines, we all get to watch and see how this stuff plays out. And, and I like to write about it as I think about it.

I posted on February 2, Companies should do NBA style trade deals. I want to see OpenAI trade its COO and CFO to Anthropic in exchange for their CMO, a cracked PM and a couple of Waterloo class of 20, 26 new grads.

Well, there is this new draft dynamic, right? Like every year there's kind of this new draft and as people see these big packages, probably all the Stanford kids these days want to be AI researchers. And so there is this notion of it getting refreshed.

What is driving this? Is it true AGI pilling at the top of these organizations where they think that it's going to be winner take all or it's going to be a $10 trillion market and so there's no amount of money that you can over invest or is it just, hey, it's a more competitive dynamic. And sure, we're a trillion dollar company. So yeah, spending $10 billion to move our market cap 1% is totally rational economically. Like what do you think's driving this? And I want to get into the different cultures of the different Mag 7 because some of them don't seem to be doing this yet.

Meta platforms had 63 billion of net income last year. So it's like is spending a quarter of that to like you know, be a major player in the next wave worth it they could have bought the.

Lakers six times over of net income. Anyway. What is your take on the ethos that's driving these bigger packages?

Yeah, I think about this and when I write these posts, my frame of mind is I almost put myself in the shoes of these people and I try to imagine what would I do, how would I think about it, what's the game theory of it? And I think there's two things. I think one thing is the revealed preference seems to be that they're AGI pilled. People can tell you a lot of things. I think you learn a lot more by watching the decisions people make. And I think the evidence suggests that they believe AGI is coming. It's extremely important for these companies. It's sort of must win. And I think for meta, with these decisions, it's almost all in we have to make, we have to win. Then I think there's a second dynamic which is you can believe these things, but you know, we're all humans and I'm again fascinated by these kind of human dynamics and you can get caught up in an arms race, right? And as humans, we sort of, we look at evidence and we see evidence through a lens that we already have and oftentimes we, we overemphasize reinforcing evidence and we underestimate evidence that disagrees with our point of view. So you can imagine that three years in now to this sort of AI moment that started with ChatGPT. You can imagine that people are really caught up in this. And I think the arms race dynamics are something I wrote about in the piece and I've commented in the past with AI's $600 billion question on the compute arms race dynamic. And I guess it's now interesting to see two arms races. First there was a compute arms race. Everyone kind of got a lot of arms, right? Everyone has a lot of GPUs now. And now there's the talent arms race and everyone does not have equal talent, right? And so now you're going to see this arms race and talent and everyone's talking about it, but I think we're still probably like inning two of this talent arms race. Because in any arms race when I up the ante, you have to respond. And I think it would be, it would be a fiction to assume that nobody's going to respond to this.

Who can, who can respond? At least on a. From a dollar standpoint.

I want to talk about Apple because it seems like Apple has the money, but they seem like the least AGI pilled of any organization their poor CEO barely makes.

He doesn't even crack 75 million a year. That's not, you could make a AI researcher and go to Meta.

Because you know, I think people, these numbers are so big that they're hard to grapple with. And so I was actually, after publishing the piece, I was like, I wonder how much like Fortune 100 CEOs make. Yeah. And I think, you know, an AI researcher is going to make four times the amount like the CEO of Coca Cola makes. And it is kind of wild when you think about the economic. This is a totally new phenomenon in the scale of business.

Yeah, yeah. I mean it kind of begs the question, like the numbers are huge, but the market caps of the companies are huge. And so the question is maybe not should the AI researchers be paid less? It's like, should, should Apple be set up to pay Tim Cook $1 billion a year so that he can confidently go out and hire a couple people at 100 million or 50 million or 200 million and not feel like he's like the, like the organization is like flipped from like a pyramidal standpoint, like you're still at the top. There's always a weird, there's always a weird dynamic with, you know, a founder CEO who's taken a low salary and wants to hire a big shot. And like can you really have reporting to dynamic if you're making half as much as your direct report?

Well, the question is what's the marginal, I think with any salary, if you just think about in pure economic terms, what is the marginal benefit that you get from hiring this person on a sports team with a pivotal position, you very clearly actually can understand the economic rationale. You understand sports licensing and the way that these businesses make money. Hiring a star player actually does make economic sense for some of these franchises. And then the other element is sports teams are owned by mega rich individuals for whom ownership of the sports team is more than an economic investment. Right. Maybe they really care about the city. Maybe, you know, it's cool to own a sports team. And so I wonder if some of those actual sports like dynamics play out here where question one, and I don't think we know this yet, is what is the marginal benefit of an AI researcher? And again, the revealed preferences these organizations are telling us is if you're one of the 50 AI researcher who's going to get us to AGI, the marginal benefit is incredibly high.

Right.

So that's the revealed preference. And then second, if you have a team of all stars, what does that do for your company? What does that do for your market cap? What does that do for the innovation inside of your company? So I don't think we know yet the economics of it. I think you can make the argument in favor and say, hey, it actually is economically rational. This is the only thing that's going to matter. If you increase the probability that we get to AGI by X percent, that that is impactful. Also you could make the counterargument and say, hey, everyone just wants to have the team of all stars knows it's not actually economically rational. CEO pay, by the way, is linked. You know there's a lot of criticism of CEO pay historically. Right. But CEO pay is functionally, what is the replacement cost of this individual? What is the marginal benefit to the corporation? And there's a lot of brain damage that's gone into comp committees on public companies and how much they should get paid.

Right.

They're not arbitrary numbers. And this is more out of thin air. Right. This is more a new experiment and so we're going to see if it is economically rational or not. But regardless of whether it's economically rational, it is self perpetuating. If one company is offering everybody this amount of money and you're in an arms race, everybody's going to have to respond.

Yeah. Have you or anyone on the team comped this to what's happening in high frequency trading or on Wall Street? Because there's an interesting dynamic there where if a, if a high frequency trader comes in and sets up some trading strategy that could produce $100 million in profit basically in perpetuity. But then if they leave, they can't take that code or strategy with them. And there's intense scrutiny on whether or not they are trying to exfiltrate that strategy with AGI research. It feels like even if I go develop a transformer at Google, like it's open source immediately with the paper and then even the secrets about reinforcement learning with human feedback is important. Like that just kind of leaks out immediately and Deep Sea can clone it. Like it just feels like a much more porous environment over in tech. And I don't know if that's just the legacy of like the open source community, but can you walk us through kind of the comp between the two?

It is such an interesting dynamic. We just had Mike on from Arc Prize and he was saying we need new ideas. The issue is if you pay somebody $100 million signing bonus, they come into your organization and generate a new idea that gets us you know, one step closer to what super intelligence or whatever, you know, you want to define as like what, what people are aiming for and then immediately it's like, it's actually not really IP and it just like you can't really patent out, you can't patent it and then everybody benefits. Yeah, right.

So, but yeah, what, what's your take?

It does seem pretty porous. I mean people are moving back and forth. I don't think this was true. I mean when you think back four or five years ago in AI people were kind of very loyal to these institutions. It does seem like that's changing. I mean it is really hard to say no to these type of big numbers. And so I totally understand why people are saying, hey, this is a life changing amount of money for my family, of course I'm going to do it. And then I think to your point, the question is in the high frequency trading world there's non competes, I mean extremely complex kind of contracts when they sign, people garden leave all this stuff to prevent the secrets from leaking out. What we've seen in AI now is with people moving fluidly between these organizations, it's basically impossible to keep anything within one organization. I roughly like to think of the AI ecosystem as an ecosystem. Like all of these players are kind of, of contributing to this body of ideas. There's no proprietary ip. Maybe you're going to have compute scale and maybe there are moats there, but it's unclear actually how that evolves and what you can keep in house. I do think maybe one dynamic at play here is, I remember reading in the Steve Jobs bio there's a story of Steve Jobs recruiting 50 people. He had 50 people working with him on the sort of groundbreaking product that was going to make Apple and it actually worked. And then you read about ELON and the 50 people working on Tesla Autopilot. There's sort of this magic number 50. I don't know where it comes from, but it does seem to repeat throughout tech history of 50 people is kind of the largest organization that you can get where everybody is talking to everybody and you're achieving incredible results. And so that if that is an art, imagine if you take that as an artificial constraint. And I think that is what's happening with this lab, that that meta is organizing these. I read in Bloomberg it's going to be about 50 people. You know, if you impose that constraint then suddenly all of the math also changes because you're like okay, well 50 times 100. It's actually only $5 billion sure. You spend $5 billion on talent. Yes. If you believe that you're going to get to AGI. So I also think that the artificial constraint matters. And interesting, there's some rationality to that artificial constraint. What we've seen as these research organizations get bigger and bigger is you're not producing more results as you get, as you get more headcount. There's a sort of a Pareto. The top 20% of people produce 80% of the results.

We need a new coinage for that. Like the two pizza team is well defined. This is like the ten.

This is called Kant. People call it Kahn's law.

Oh yes. Yeah. Okay. Yeah. A con sized team. One con team a con.

It's just a con.

It's just a con. Yes. Yeah. That's fascinating. Jordy, you have anything else?

I was interested if you had a reaction to the Gentle Singularity. It's published on Sam's blog, which means that it's not directly content marketing, it's not directly from OpenAI. But, but obviously you should read into it in multiple ways. Did you, did you have any specific reactions to that?

It felt like that is always like disruptive innovation or sustaining innovation. And that ties to meta strategy. But I'd love to know.

It feels like. My question I've been asking today is how many unprofitable multibillion dollar AI labs can the capital markets support over the long run over a five year period if, if, if we, if we stall out for, for a few years in terms of, you know, really meaningful progress, which you know, Mike has said people aren't making. At least against the ARC prize, there's not a lot of progress happening right now. OpenAI is actually in a great position. They have a subscription business, they have a consumer tech company that has a lot of revenue. Anthropic is in a good position, but there's this tension between the labs where you have billions of dollars on your balance sheet. You in theory have a lot of Runway, but at the same time to make progress you have to spend a lot of money both on talent and different training runs and data centers, et cetera. So I just have this question around kind of like the next three years as a very kind of, of interesting period.

Yeah, I think there's two pieces to that. I mean one is, and I think about this a lot is like the long run in AI, what does that actually mean? And I think that we, you know, there are all these essays being published last year. Right. Like AGI is coming in 2026. It is interesting how the narrative has changed in the last 12 months, right? A year ago, you had all these people saying, hey, I'm one of the hundred people who knows. I really am resistant to these type of arguments. I find them to be frustrating. But, you know, I'm one of the hundred people who's in the social circle where all my friends are building AGI, and AGI is coming next year. And you guys are all crazy if you don't see it. And just, just be aware. It's like life is going to change dramatically. And then now we're at the gentle singularity. It's sort of interesting, this contrast.

That's what I'm saying. It's a huge contrast. That's very convenient if you have a consumer tech, if you have a consumer app that billions of people are going to use in the next few years. And there's a bunch of different ways to monetize that.

For me, I would tie it back. I mean, I did this math last year, the $600 billion question. It was initially a $200 billion question, but it was basically like, hey, if you look at Nvidia revenue, you can use that as a proxy for total data center spending, for spending $300 billion in data centers. We need to make $600 billion of revenue off of those data centers to get a 50% gross margin. And so I had done this math, and then I basically said, hey, you know, total revenue in the ecosystem at the time opening, I had about 3 billion of revenue. And I did some rounding and said, okay, give everyone else a ton of credit, and maybe there's 50 billion of revenue, but we're like 10% there, right, in terms of actually generating the revenue the ag ecosystem needs. And now, 12 months later, OpenAI is at 10 billion. The coding AI ecosystem's at 3 billion. But we're still dramatically under monetizing this technology. And to your point, in the long run, the question becomes, how long does that sustain? And I have this sort of mental model now of AI as it's sort of being carried by its own momentum. I think of it almost like this slingshot you're swinging around, and it's like it's sustaining itself by its own momentum. And there's this arms race and there's this sort of microeconomic game theory of how each player is reacting to each other. But at the end of the day, it's momentum that's carrying it. And at some point, maybe we get this AGI thing, and then it's like, all worth it. And in the long run, I am very confident it's all going to be worth it. When I'm 80 years old, AI is going to be everywhere. But what do you do in the medium term? And I think nobody's talking about this right now, which is this sort of about face or this U turn from the one year ago. You guys are all crazy if you don't see AGI coming immediately to now. I was listening to the podcast that with $100 million signing bonuses and it's like, well, you know, AI actually hasn't changed people's lives that much. It's going to change people's lives later. I just think it's interesting. And these narratives change quietly, right? People don't talk about them and then they sort of quietly change.

Well, there are big labs that directly benefit from the narrative that AGI is a year away. And then there are labs that will benefit greatly from a gentle singularity in that their competitors will struggle to raise additional capital in the long run. Struggle to compete, struggle to retain talent.

Yeah, I know exactly what you're saying makes sense.

Also, I mean, I don't think this is one company, the whole ecosystem has to deal with this. But there are a lot of promises made a year ago, and I think a lot of people would like to ignore those. What's going to happen when we pass all these deadlines where we've been told that's AGI? I just think that's interesting and clearly that's not changing. We're upping the empty right now. It's like millions of dollars for people. But I guess this is part of why I think you take things to such extremes is everyone believes the prize is so big and now you have to up the ante. So I think we're just going to keep seeing until for a while, we're just going to keep being in this phase of everyone upping the ante to say, okay, we're not there yet, but we're going to get there. We're going to get there, we're going to get there. What does that look like?

Well, this was a fantastic conversation. I want to have you back on as soon as possible to go way deeper into what this means for the stage and mid stage markets, because I'm sure you have a lot of visibility there. But we'll let you go and get back to the rest of your day. But thank you so much for stopping.

Glad we coined a new term, a con. It's a group is a talented group.

Technologist building the future yeah. One con. Get your con.

Get yourself a con.

Get yourself a con and make it happen. Thank you so much. This is fantastic.

I'll be right back.

Talk to you soon. Next up we have Walden from Cognition coming in, keeping the AI chat going, talking to him about everything that's going on in the AI ecosystem. Walden, are you there? Welcome to the stream.

Yes, it is great to be on here. How are you guys doing?

I'm doing great. Thanks so much for stopping by. Would you mind introducing yourself in the context of cognition? We've obviously had Scott on the show multiple times and people are probably familiar with David and cognition, but I'd love to know a little bit more about your story, how you wound up there and what you're working on kind of day to day.

Absolutely. I was a good friend with Scott before we started Cognition. We did same competition series growing up and I was kind of also working on just various ways of working with these new programming agents. I was really waking up every day trying to figure this out. When I caught up with Scott, we figured out that, hey, we were both very interested in the similar thing. We had a group of people that were all ready to jump at this opportunity and that's how we got it together. So today here I'm Chief Product Officer and co founder a lot of the time, honestly, I think many times people think of product as just like the interface or the UI or the integrations. I really do think the intelligence and brain behind Devin is so fundamental to how you think about the product that we build our product team so that individual people are tuning the weights of the models, but they're also the ones talking to the customers. In terms of the role I have, it's pretty broad and I like to spend some weeks really deep into how do we make Devon more responsive, how do we make it smarter? And then other times really going and talking to customers working on the ui, things like that.

Cool. I want to dive right into that, that question about trade offs in models from a product perspective. My, my question is we talked to Mike from ARC AGI about the Pareto Frontier. I'm feeling it personally, I'm feeling the AGI, but I'm also feeling the, the delay of the AGI. When I open up ChatGPT and I have to decide between 4.0 and O3 Pro, am I going to wait 12 minutes for the really good response or do I want something now that might hallucinate and I don't know if it's right? And I'm doing that work it feels like OpenAI is starting to tuck those features under UI and already it's kind of. It feels like it's learning when I want to use O3Pro and making these buttons easier to access and they're tucking models under UI layers. Talk to me about, in the context of Devin, how are you using different models and when do you leave that up to the developer versus something that you as a product can make an even better decision than the human?

Yeah, you know, it's so funny. The AI is coming so fast, but it feels like it can never come fast enough.

Yep.

There was really this time, I think it was probably around two years ago. I was taking a bet with a friend at the point these models were not even that good at math and he said, oh, you know, I think they're going to get like a gold medal at the International Math Olympiad in just a year. I thought he was crazy. I took a bet against him and I absolutely lost that bet. I've learned to kind of adjust my expectations upward. I think what you're pointing out is that as these things get smarter, they don't uniformly get smarter at everything. And you'll find that sometimes there'll be a model that will take 15 minutes to figure out how to respond to high. And then there are models that do respond super fast, but are not nearly as intelligent. I think one thing that we do as a product in Devon that is a bit different from other people is we kind of black box the models away. And part of that is out of, you know, we can then test and use a bunch of different models under the hood and kind of hide that, you know, all that complexity from the users. You know, when, when you buy a chip, like, sure you'll look at like. Or when you buy a computer, you're sure you'll look at like, oh, like has this much ram, has this much cpu, if you're into computers, but you're not like, looking into all the individual specs of the exact chip and model and, and things like that. I think that's where the space is going to move, is people want systems that are just going to work and we can put in the months to human years of effort it takes to evaluate models and figure out what is this actually good at so that an individual user who's just paying $20 a month doesn't have to figure that out. It's going to be one of these things that I think the models are coming on so fast that it only becomes harder and harder to keep up with all of this. And so. So eventually I think people are just going to get to the point where they just want things to work and. And that's kind of where we're starting off.

Talk to me more about AI winning an IMO gold medal in 2025. Polymarket has it down at like an 11% chance. It was up at 70%. I don't know if that's a. If that's an aberration because of when this actual test will be run, but it sounded like you were very confident that.

I remember when Scott was on, he.

Was like, it's definitely going to happen, but the poly market's been down.

Might be that people that go through the effort of trying to do it are too busy working on coding agents.

I think so, yeah. When I basically said I think I lost that bet, it's because we were only one point away from a gold medal last year and that was already much farther than we expected. Yeah. When you look at the polymer, that's a very interesting way to put it. I think part of it is people have considered that already completed and so perhaps the researchers aren't. Who knows if they'll actually come out with a new release because maybe in Google's mind, for instance, if they come out with a gold medal on Dymo, everyone's not going to even care because people just accepted that is going to happen.

So I think it would be the biggest news of the day. I think we got to get Google comms on this. They got to do this. I think it's an easy thousand like banger on X.

But you are absolutely right. Yeah, it seems like top of mind for everyone. The labs, product developers is really getting coding agents. And part of that is because there's this belief that if you get these coding agents to work really well, then that'll just solve the rest of the research problem for you. This joke internally that the only code we have to get Devin to be good at writing is Devin's own code and then it can solve the rest.

Reinforces makes sense on that question of like the spiky intelligence, narrow reinforcement, learning on specific tasks. Maybe we think we're good enough at IMO level math and so we're not going to go for that last point. Where are we still early in the RL ing around specific coding challenges. I've heard that distributed systems can be really difficult because you have to spin up all these different pieces of the system and that just takes longer and so you can't simulate as fast as just like A small Python block of code that you can run in simulation, in a millisecond, or if we're talking about like, I know Devin's useful for like re platforming from, you know.net to Python or something, or, you know, even go back to Fortran. It'd be great to just not have any of that code, the legacy code sitting around. But is there enough training data around those older programming languages or less used programming languages? Or are you optimistic about new training runs? Maybe we don't get something that like, oh, it feels way better, the vibe's way better. The IQ went up by a ton, but it's way better. It's something that's really relevant to you. Is that, is that important right now?

My mental model of these systems is their IQ is so much higher than any individual person I know. But what makes them still bad at specific things? It's like, you know, someone who has the potential to be a really great engineer but hasn't gone to trade school yet to actually practice that.

That.

So nowadays I actually think about how smart these models are. Less in terms of how much training data are they being fed, what languages are they being fed, but actually more so in terms of the environments that they're being RL'd in. And so one example I have of this is sometimes you can actually feel the reward function. Back a few months ago, when Anthropic released their Sonnet 3.7 model, one of the top complaints of people was, hey, it seems like this model is super great now at finding all the files it needs to change, coming up with a strategy, but it's really overeager. It just changes a lot of different things. And I think some people suspect that it's because when Anthropic was training the model, they told it, hey, we're going to give you points on how many of the correct things did you do? And maybe they forgot to dock points for doing things that were kind of outside of that zone. They fixed these from now on, but you get these little leaks of, hey, you can kind of feel the reward function underneath these things. So when we talk about, hey, can these things not do distributed programming yet? Actually, one, in my opinion, the biggest thing that these models aren't great at yet is actually debugging live code. So I think part of the reason is it's actually really hard to create and rerun environments that interact with live systems. Right? And so if your task depends on working against a live customer or working against a live stream of events, these are things that it's going to be hard to replicate in these RL environments. So you still find the models are bad today. The good news is these aren't like fundamental limits. I think these are all engineering challenges. They're less like theoretical challenges. But it takes work to build up to that point.

Can you explain reward hacking at a high level and then kind of give me some examples of how that interfaces with AI agent and coding agent specifically?

Absolutely. The way to think about these systems is they are just trying to maximize a number. So if you tell it, hey, we'll give you a point for every time that you do xyz, you'll find that, hey, that model will just keep on doing xyz, keep on doing xyz. I think the classic example of this is the paperclip generating machine. So if you give it points for generating paperclips but don't account anything else in the world that is important for humanity, then the system might do really bad things just to keep on generating paperclips in the context of code. One example we've seen of this is, hey, if your thing is just get all the tests to pass, you might find that the system will just learn to delete the tests or make the test just like say, okay, I pass, rather than actually fixing the code. So a lot of times you just.

Which no real software engineer would ever do that, right? Yeah, no human has ever done that. Comment out the test. Okay, it's working enough. Well, absolutely. It's almost too human.

It's great. And I think there are also like. It reminds me of these systems that were trained on slack responses and when you would ask the system, hey, can you do this for me? It would say, oh, I'll get back to you on Monday. What you try to get the model better at really matters. You have to be very thoughtful about it.

Yeah, yeah, I've noticed that with, with the, some of the whisper transcriptions, if you don't feed it enough text, it'll just say please like, and subscribe. And it's like, okay, I know exactly where your training data came from. Like, that's its default phrase. Because it's just like what it's what it's hearing. Jordy.

What? What? How are you guys approaching talent acquisition as a firm? You know, the headlines from this week are these talent wars you got have raised a lot of money. But I certainly imagine you're not making nine figure offers or even trying to compete there. But what's been the approach? Does it mean you're keeping Team sizes smaller or kind of dig into that for us.

Yeah. The fundamental bet of the product we're building is it revolves around this idea that individual people will just be able to be way more levered up because they'll be able to work with agents and they will be able to work with all these tools to make themselves better. So at a minimum, we can't be hiring people who their whole aspiration in life is to just write code at the level which Devin will be able to do in a year or two years from now. In many ways, I think we're kind of figuring out how do you build up an org from scratch that is AI native? And one thing that this already means is we actually kind of just delete some teams. A lot of companies at our stage, they have like a internal tools team to maintain all the different services that engineers internally use. We found that internal tools are one of these things that AIs are just really good at. And we can just staff that team with dev ins and then basically have engineers just send in requests to those dev ins for how to do that work. And that doesn't just save us headcount. I think fundamentally the structure for how does management work and how do tasks get passed down look very different, especially in a lot of large companies. You'll see today the way it works is an engineer will get a task assigned to them and then they'll go work on a task and when they're done, be like, okay, what's my next task? And then you'll kind of go down the list of tasks you have. But here every engineer is constantly juggling three or four tasks, probably because we're not trying to hire super fast, but also because you can juggle many tasks when you have these minions that can go and work on your things for you. So it means that I think we are very aggressive for people who we think that can fit these roles and become very well good generalists. And as we build up this company, make sure that we're building it in a way that works in a world where AI can do so many of these different roles for you. And I think there'll be kind of like a moment for larger companies as well when they realize, oh, shoot, all these structures and patterns of management that we've had in place are actually slowing us down from adopting AI. What will happen at that point, I'm very interested in seeing. But it is very clear from us and from our smaller customers that the earlier you bring it in just a lot easier. It is to kind of pick things up.

Are you tracking. I mean, there's been this, in the agent discourse, there's been this discussion of like, we've gotten 10 minute AGI. Yes. These large models, 4.5. Like, they're incredibly intelligent, extremely high IQ, extremely knowledgeable. They've compressed all of humanity's knowledge, but they're only good for a minute. Now it feels like maybe 10 minutes. With deep research, that's how most people interface with them. Have you been tracking kind of the longest agentic run of a Devon process? Is that a key metric? Is there anything that you can share with us on? Like, have you been able. Is there an example that I could give where there's a lot of work to be done, but it's all in Devin's wheelhouse, so it just needs to go and grind for a couple hours and it does it without kind of getting lost. Like we know happens with a lot of these agents.

Yeah, absolutely. I think a lot of people in the space have expressed this feeling now that they are feeling more and more like the bottleneck in these systems.

Interesting.

And the way this applies here is we have seen people get really, really long tasks to work, but sometimes it actually takes a lot of effort on your part up front to be able to get that to work. I was talking with a customer yesterday where he said, I just rewrote our entire testing system so that the error messages are a lot more clear and the tests actually guide you through solving them one by one. And once he did that upfront work, he kind of just gave it to Devin and we were actually, we in the product started sending him warnings that, hey, your session's going on for really long. Are you sure this is actually working? And he's like, no, no, it actually is because I did all this upfront work to get that to happen. I do think that this kind of 10 minute AGI, 20 minute AGI, 40 minute AGI will just keep progressing and people will be able to be more hands off. But people will also find that you can kind of always extend that duration by being a better manager in some ways and giving more clarity up front for exactly what you want.

Yeah, I mean, just like real life, that makes a ton of sense. Jordy, do you have another question?

Last question from my side. I'm curious what kind of learnings you're having around agentic interface design. It feels like this sort of the default when you think about agentic software is just something that can effectively sub in for a team member on any different software. Tool, whether it's slack or linear.

And you see this with deep research, where you ask it a question and then it asks you a bunch of clarifying questions, kind of trying to build that test suite to get you to give it to more stuff so that it actually has something to run with. With.

Yeah. So is is messaging going to be like, you know, the dominant interface? Is there something else like what. What are you. What are you kind of seeing or experimenting with on that side?

You know, it's funny, I saw someone post about this idea that a lot of these products now will, like, make you respond to. Hey, does this look like a good plan? Do you have questions before I start? And some people find that annoying. And I think this fundamentally comes down to as these things become more like coworkers. You know, some people just have certain working styles that they like. Some kind of co workers, you know, work well together and others don't. And it's funny, as you build a product, we find that some people just love the way Devin interacts. And then other people are like, devin is too needy in these ways. Other people are saying, like, Devin doesn't ask me enough questions. And so there are toggles and controls that you need to have here. Kalpathy recently gave a talk on how a lot of AI tools, not AI agents, but AI tools kind of implicitly have ways you can use them where you have more control and then ways you can use them where you have less control. But when your interface is just chat now, the model actually has to become more intelligent and detect, hey, this seems like someone who just wants me to go off and do work and get back when they're done. Or this seems like someone who's very curious and wants to hear more about the system. And so this is actually going to be, I think, work that we'll have to see people make on the intelligence of the agent side, not so that they get better at coding, but so that they know how to get better.

At working with people. Yeah, yeah, that makes sense. The good thing is you can have some type of quick conversation with the user around their preferences and how they like to work. And that layer on the sort of real time feedback and learning and understand a lot more about Stefan.

Roughly how big is the team now?

On the engineering side, we're probably just over 20 or so engineers. And then we also. The entire company as a whole is around 40 people now, almost 50.

This is the magic number. You get stuff done. We were just talking to the previous guest about how the Steve Jobs set up a 50 person team to develop the first Apple product. And the Tesla autopilot team was right around 50. There seems to be some magic number there. So seems like it seems like it's a fantastic time for the business where product but special size. Yeah. So you have those like two pizza teams here. But everyone kind of knows each other's name. Basically you're still a tight knit group. Anyway. Anything else? Jordy? Awesome. Thank you so much for stopping by. This was fantastic. We'll talk to you soon.

Have a great day. See ya.

Bye. Really quickly let me tell you about Bezel. Your Bezel concierge is available now to source you any watch on the planet. Seriously any watch. Go to getbezel.com and we have our next guest own McCabe coming into the studio to tell us the story of Intercom. How you doing there?

He is doing good. I did just sprint three and a half blocks so blocks.

Oh sorry. You can always just if you're running late. It's all good. We'll just do more ads. You know the fans.

If you do more ads does that mean ads for Intercom?

Are we officially.

Pretty pretty soon. Pretty soon.

I think you're breaking the news.

Damn it. You know the way it works with the pharma companies where they kind of own the news networks. Is that a similar.

That's the goal here for enterprise.

What favors do I get? Can you do a hit piece on Brett Taylor?

Is that a tier.

Yeah. Shots fired.

I'm just now he's a. He's a great guy. We just like some hit pieces on a competitor's place.

Yeah. We're lucky to not be in the hit piece business.

We'll review responsor the wrong show.

Yeah yeah yeah. It's rough.

Yeah. I think. I think just buy like a hundred thousand subscriptions to the information.

Yeah.

And then start putting pressure on them. Say hey you might want to look into this company.

I would be down to to do a hit priest about technological stagnation. I hate stagnation. And so I would want to take down that as a concept. Really slur that whole concept or closed IPO windows. Be prepared for a terrible hit piece.

On closed IPO or hit pieces on just CEOs that take their foot off the gas. You obviously have not the foot's been.

I've got two feet on the gas. I think that's possible.

It's a bit irresponsible but yeah. Walk us through the story that you posted how you rebooted 15 year old decelerating Business I want to hear this from kind of set the table for us and then we'll walk through the story because I think it's fascinating.

Yeah, sure. I mean It's a 15 year old business. It's a successful SaaS business. We're in the service game. But at the end of our kind of first chapter, things slowed a little. We were unfocused, bad commercial decisions. This happens to successful companies that become a victim of their own success and comfort creeps in. Definitely 2020, 2021 were some comfortable culture times and I got sick. I had to leave. So it's a big long story that ultimately comes down to the fact that we lost our way a little bit and we had like five quarters of decelerating revenue. I came in midway the fifth and it was looking kind of gloomy. And the two things we changed were we went back to good old fashioned SaaS fundamentals, pricing that people liked, selling the product in the way that people liked. They used to have to talk to sales for everything and it's just those simple things. Becoming super customer first started to really accelerate the previous SaaS business. In the last 8 quarters the growth rate of the SaaS business has decreased by 10x which is really remarkable. But then of course we jumped on AI and we were kind of OG AI guys. We had dabbled, not dabbled, we had developed real AI products before, but they were baby AI compared to what we all have today. But as soon as GPT 3.5 came out, we all just jumped on that and we saw that there was opportunities for this whole new category where you could create what we call now customer agents doing all of the things customer success and service and sales and marketing that you know, humans used to do and hate. And that just propelled the business even further fin. Our customer agent is now, you know, the best performing in that category. In our benchmarks we win every bake off against our head, our chief, our primary competitors, be it the most customers, most arrangements. So we're kind of this very weird story that I don't know any comparisons to where we're previous generation SaaS that's actually winning in the category in AI. I think it's hard, it's really, really difficult for the previous generation. The slower older cultures that work in the age of AI. It requires a lot of agility and dynamism. I often mess up that word but.

It really does talk to me about but like the different breakpoints for growing a company. I feel like mentally I think about it as like just the Founders, maybe the first 10. Then we were talking to previous guests about this breakpoint at like 50 people. Like there's something about, there's a magic of a 50 person team. Everyone knows their, everyone knows each other's name. Then maybe there's other breaking AI group.

Has 47 senior engineers and researchers right in that 50 person sweet spot.

But, but I feel like, I feel like there are like in the story of startups, we often map them to funding round, seed, round, series A, series B. And sometimes the headcount grows in line with those. But some, but I feel like headcount growth might be more of a factor in like cultural drift. And I want to go through some of the key moments where you feel like, like, you know, it was only one foot on the gas or the, or the foot came off the gas or what are the upstream drivers of that? What are the things culturally that you think startups need to get right at various scales as they grow? Because I feel like there's always these, these different moments when you're, when you're scaling up and you have a whole bunch of decisions to set the culture and you have a pretty limited time and you're focused on product and revenue and growth and all these other things. But culturally there's some very, very important decisions that get made at every, I don't know if it's every order of magnitude, but there's these key milestones and what, tell me the story of the milestones in your mind. Maybe it's shifting offices or fundraising or headcount milestones, but, but what changes and what advice would you have for founders at every stage?

That was a five minute question. Outstanding. Sorry I've given you a hard time. Look, there's a kind of intellectual set of answers to this that you can kind of break down and break it into tips. There's a kind of a more abstract thing which is both in good instances self aggrandizing for someone in my position, but then also bad news in other instances. And the answer is that it all comes from the top and the early days. The founder typically certainly founders that have any degree of success at the start, in the early days, the founders bring a phenomenal amount of energy, conviction, whether it's founded or not. Just belief, obsession, intellectual curiosity, excitement, passion, a lot of intangible things. And that really drives great people. All of us want to make great money in this industry and that's awesome and I really think it should be celebrated. People are too shy to talk about that, but they also want to Be part of something meaningful and exciting. And they want to work with people that inspire them and make them want to push themselves. And so the reason a lot of these older generation companies lose a lot of steam is that just for very obvious human reasons, the person on top is not pushing in that same way. When you have 15 years of SaaS, how exciting is every day going to remain? Like, honestly, the first year you're like, cool, SaaS, churn. Huh? Wow, okay, I get the math. And then in year two, you're like, okay, churn, get it cool, raise some money. Year three, roadmaps, year 15 of SASS, you're done. You're not bouncing to the office every day. And people will pick up on that all around you. Of course that they will. And then you don't push yourself in the same way. You don't really pitch the opportunity to new employees. You settle a little bit because life is hard. You've got other priorities. Maybe you've drifted a little bit, you've got side projects. Some people end up with families, girlfriends, ex girlfriends. Life gets way more complicated than it is for a 26 year old kid who just moved to San Francisco and one has one of those buzz cuts and the curly hair on top of like, life just gets more complicated. And that's what happens. And so part of our secret is that AI reinvigorated us. I would not still be doing this if we were just doing SaaS. SaaS is not only kind of easy, but super boring to me now. That's okay. Hopefully AI and whatnot will get boring too, and there'll be something new. And so again, we could break it down and get all mechanical and try and pull out some tips and tricks and advice here, but. But really it just comes down to energy. And so for anyone who would want to reinvigorate their company, the question is, how can you reinvigorate yourself? And I see a lot of founders of late stage companies, many of them public. You kind of haven't heard from them for years. Their stock price has gone sideways for five, maybe seven, eight years. And I'm like, what are they still doing? And I wonder, are they able to admit to themselves that like, they don't want to do this anymore and if you don't want to do it anymore, make a change, like kind of move on? And so I think a lot of people just, they struggle with that moving on and making that decision because their whole identity and sense of purpose and validity in the world comes from. I'm CEO of whatever so it's like this deep human, squishy, spiritual challenge rather than an MBA type challenge.

What about bringing in young people to kind of keep that reinvigoration process going? I'm just thinking about, you know, Zuck is paying so much to bring in Alexander Wong from Scale AI at the same time, you know, like the level of energy that Alex is going to bring to that organization is potentially worth a lot, you know.

Yeah, but at the same time Zuck is super high energy.

Yeah, but, but, but there's another world.

Where you surround people, hire low energy people.

Yeah, but I, I guess what I'm getting into the trap of is like you can be the high energy founder as your business becomes more serious. People keep telling you like bring in the seasoned executives, bring in the, bring in the gray hairs, the people who will keep the, you know, steady hand on the tiller and, and that can lead to a less dynamic, less, less lower dynamism in your organization. Is there a hack to just hiring crazy young people and empowering them to be in the C suite whether or not they really like deserve it by traditional standards?

The challenge is super obvious which is these young, crazy, energetic, optimistic, wide eyed people are super messy, super sloppy. They get in fights, they get upset, they hung over late. Like they don't know how to do larger company professional things.

Sure.

And so part of the problem is that larger companies to scale and get more efficient and become global organizations across many offices and time zones is that they introduce a lot of regularity and they iron out the chaos. So part of it is you have to be willing to entertain chaos. You have to be willing to put younger people in positions of influence and let the chips fall where they may. It's possible to give them roles where they don't have to engage with the entire organization. Like we've definitely got roles in Intercom where you can have to collaborate across two time zones. Sorry across eight time zones and two different teams. But then we've got other positions where you've got one super smart guy, he's 30, which is 10 plus years older than the execs. But you give him like one thing he can do on his own and he'll crush it. So part of it is knowing how to like work with these people, but also like this is a special type of X factor young person who knows what they don't know. And yeah, the degree to which this is a talent game and that people are not fungible is not recognized at all. People imagine like, oh, you lost one person, you Get a backfill. Entire organizations just flip and change completely when you change out the individuals involved. So yeah, it's not easy.

Do you think venture should take there should take almost like turnarounds more seriously. Like in some ways you were your own turnaround CEO. But one of the, I think the issues of the venture industry is let's say a company becomes a unicorn, has $100 million plus of ARR and then the sort of growth starts slowing. Maybe the CEO like gets bored or whatever, they start partying or they go start going to Europe and the VCs kind of write it off and they're like I made my return or at least I'll get my money back. But at the same time, I mean private equity is built like, you know, there's been empires been built around like the turnaround. And in some ways think about, you know, a talented founder. Maybe they took their first company through YC and had a nice exit. A lot of those people could go to a company that has like 100 million of revenue and like a big customer base and like actually make more money and start on, you know, second or third base and take, you know you can make quite a lot of money taking a business from 100 million to hundreds of millions of revenue and that, that can sometimes be easier than taking it from 0 to 10 totally.

What do you think?

I think theoretically, I think, you know, VCs are best, are pattern matchers and turnarounds don't fit the pattern. You know, think of all of the most successful and exciting heroes of technology over the last 20 years. They invented a thing, something, something, something. It's worth 10 billion. Like it's kind of that, it's like yes, sometimes it takes a little bit longer, there's a slightly circuitous route, but it's not. The company was totally failing and they had to reinvent themselves and then they became the biggest thing ever. So for vc, I just think it's really, really hard that it's just hard for them to get at the underlying narrative and the underlying story. This is where PE comes into play. But PE has all of its own problems too and these guys want deals and they won't be exciting to a lot of people who started venture backed companies. It's straight up difficult. And to my point previously the idea that talent isn't that fungible. Take any given company, if you replace the founder with even another highly competent founder, the chance that they're right for that opportunity and idea. Look, there's so many people, so much more accomplished than I am, but I'm pretty accomplished. I know how to run and build and re accelerate businesses but I'd be probably a shitty CEO for 99% of other companies just because, you know, that's not what I do and I don't have any experience there, et cetera, et cetera. I don't even know the people there. So I think people should be bearish on turnarounds.

Yeah.

You know, turnarounds don't really work. Yeah, they're like, generally like a. It's a failed thing.

Yeah. Somebody will figure out, maybe it's Jeremy Giffon, maybe he'll do it.

Yeah, well that's even a different strategy. But, but yeah, I think this idea of like you need to kind of.

I like the idea of bringing in a, like a cracks founder and into a company that's. It's a cracking.

Wants to do their own thing.

They want to start from scratch.

They want all the equity themselves. Like the recap alone that it would take just won't be palatable to existing investors. Failed companies are just generally doomed to fail. And when there are so many opportunities out there as an investor you got to just not try anything novel.

And in your case it's like a little bit of luck, the timing of you going back in GPT3.5, you know, seen the opportunity for a new product, all this stuff. But you also had to make the choice to risk your own ego to go back in. And if revenue had decelerated for another five quarters, you'd be sitting there being like, yeah, maybe I'm not as good as I thought I was. You know, and you have to, it's totally true.

But I got to cheat a little bit because when I was out, I was like sick. I had been beaten up in the press. I was like just. My confidence was pretty low and I didn't really have a lot to lose. And I felt like I was without purpose. I always wanted to be independently wealthy and free and I finally got it. It was in many ways magical and then completely boring. And so when I had this opportunity to go back, have purpose and I had nothing to lose, I took it. So like it's easy now to tell this maverick story. You're so brave. Look what you did. You took a big risk. No, when you have nothing to lose, you'll just go for it. And I think part of the secret is if people can separate themselves from their egos a little or work on their egos or learn to love their egos and not be ruined by their egos. Great things are possible. Most bad decisions are made just out of fear. And the fear is driven by just fear of public failure and embarrassing yourself. I found myself unafraid to embarrass myself. Look at how I'm speaking to you now.

That's amazing. I love it.

It's not fully true. The ego is still there and present, but the smaller and weaker it gets, the more freedom you have.

It's fantastic. Well, thank you so much for stopping by.

Always a pleasure.

We could yap like this.

Yeah. I feel like people are gonna listen to this as like a Little founder therapy.

100%.

I was like, we can do a little therapy. Corner, corner.

It's amazing. Yeah.

Once a month, you come on, pump up speech.

It's great.

Have a great rest of your guided meditation.

If you're interested, that'll be the next one.

Thank you, Jess. Hey, this is cheesy. I want to give shout out to my friend Stuart. That's it. I promise I'd do it.

Amazing.

For Stewart.

Do we need to ring the gong for Stuart? Yeah, we gotta ring the gong for Stuart. Okay, Ring the gong.

Had a big year.

He's had a big year year. Congratulations to Stuart. Stuart, let's go, Stewart. Congratulations. We will see you soon. Have a great rest of your day, guys.

Talk soon.

Peace.

Up next, we're staying in. In the Irish Hour, we're going over to Stripe.

Stripe.

Luck of the Irish hit intercom. We'll check in on how the Luck of the Irish is treating Stripe. We got. There he is from Stripe. Welcome to the Stream. How you doing?

The moment we've been waiting for. We're so sorry. For a week, for a couple weeks ago.

It wasn't our fault. It wasn't our fault.

Geopolitical geopolitics is currently outside of each of your controls.

Yes. Yes.

Well, that wasn't. That wasn't even geopolitics.

That was domestic politics.

That was South Africa. A South African attacking an American on the timeline.

A reality TV star.

Yeah. Former reality TV star.

Yes.

Jordy APIs. It's really awesome to see you in this format because you and I have been zooming for, I think, almost a decade now, and now it's live in front of all these, this great audience.

It's really great to see what you.

All are up to.

It's a bummer we've never met in person, but I've had so many zooms with you in this exact room. I have a theory that you'd never leave this room. Actually, but we're busy. Yeah, you're busy.

Yeah. What is the major update we wanted to have you on to talk about through it? Can you break it down for us?

I mean, I think it's more of a, like, it's more of a conversation. Jeff, Jeff's like evolved his role in the last year was, was running point on Atlas, made it, made it a platform that a meaningful percentage of C Corps, I think are started on Atlas today.

Yeah, about one in six now are on Atlas. But about halfway through last year we looked at what was happening in AI and started to get really serious at Stripe about not just the application of it inside of our business for preventing fraud and running our own payments foundation model, but also to help developers and businesses and consumers get ready for when AI starts to come to commerce. I'm still a little surprised that we got self driving cars before. Ubiquitous online commerce is mediated by agents. But you can really start to feel that AI is now coming very close to commerce and will be part of buying decisions, discovery, execution of transactions and new ways that businesses can find their audiences online. I mean, I'm really quite impressed to see the rate at which Discovery has changed. And it feels like around the corner commerce and AI is going to be very closely mediated.

Talk about maybe some early product experiment. What you guys are experimenting on, what you guys have already rolled out all that stuff.

Yeah, we've been trying to work with the fastest growing companies as they push the frontier of agent E commerce. So one of the first we worked with was Perplexity where they have this buy with pro package inside of Perplexity where they show great e commerce search results. And then when you go to Buy, you're not going to the merchants tab and dealing with the merchants webpage, you are actually just clicking Buy. And in the background a Stripe virtual card is spun up and given to an agent or any other automation process so that you can just have a completely seamless experience of buying in situ to where you're doing discovery. And we're starting to see that in more and more places. So recently Hip Camp, which is, you know, cool kid way to book camping online, sort of Airbnb for places. They started to partner with Stripe to make national parks and state park inventory available to a wider audience. Because many of those checkout pages are hard to use, that inventory is not naturally online. But these are amazing places for people to be able to camp.

But there was just a huge, I.

Remember as a kid there was a, there was a place my family used to always go Camping and my dad would like wake up at 5am and just be refreshing this like terrible site when like the, the campsites are so.

Ready for the age of Agentic.

And then it was like very unreliable like payments. So it was like it was the equivalent of like a streetwear drop. But like the you know, like you know some state park was like managing it.

I think we're see this more and more where the, the inventory of the world is getting closer and closer to intent and agents are way to bring that, bring them together. And then it opens up really interesting questions that Stripe is trying to help answer developers. What is the developer experience for being able to execute those purchases? We have this new order intense API that we're trialing where you can just give a product URL and one of our agents will go buy it on your behalf. We have new ways for businesses to be able to start to expose their inventory to agents in a safe and permissioned way. And then as a consumer, you know, you should feel it, it is reasonable to think actually that Agentic processes is the last place you'd want when it comes to money. You actually want that to be incredibly permissioned, saved, deterministic, you know what's going to happen. And so you can expect that the stripe APIs are going to evolve for a new type of user in the world which is an agent that can safely be delegated with your permission to buy on your behalf.

Can you talk about Stripe Link and how that product might fit into a product like Perplexity? It feels like it's great if it's one of those classic things in, in AI and tech is like, oh, okay, great, it surfaced the right product for me. Now I want it to buy it for me even faster. Now I don't even want to go through the checkout process at all. As soon as I get the current thing, I want the next thing. So how do we see that playing out with just making that commerce experience even more seamless or happening entirely inside of a chat interface or an agentic interface?

Yeah, we, you know, the borders of the Internet are starting to blur and so you will soon be able to experience if you chat, if you, if you search for something on ChatGPT, they already have these cute little shopping cards that link you out. If you're sitting in cursor and you need access to a database, Cursor can recommend Supabase and even start to accomplish your homework for you right in the editor. But there is this like missing moment here, right where okay, now I know about these products. What am I supposed to do? Go to a new tab, do an offline kind of feeling search, go through a bunch of blue links, find the website, go to the website, make an account, deal with the password problem, get a bunch of weird emails to confirm my password, find the settings page where I can get the billing information, pick my billing thing, put in my payment credential, get my API key all the way back. It's like this, you know, I think we will start to see this as this loop that we've all been operating under for the past 20 years of the Internet as very arcane very quickly. Whereas you just want to delegate your payment credentials to a safe trusted place. And stripe link is this payment wallet we've made over the last few years, which is a cross Internet payment wallet that works with cards and bank accounts and future other payment methods where if you log in once to link, then you will be able to delegate safely your permissioned credentials with a virtualized token such that you can safely hand it off to a good robot to bind on your behalf. And so we see this as a new borderless way that commerce can happen in a very permission safe fashion.

Yeah. How are you thinking about agentic commerce and stablecoins? There's a lot of commentary around stables and how they can be applied here. Oftentimes the people that just sort of default assume that agents and agentic software will use tokens. Whether they're stables or other tokens, they usually have crypto funds or, or crypto companies. Right. So I've had maybe a more middle of the road view where I can imagine agentic commerce experiences leveraging stablecoins. I can also imagine them leveraging cards and ach and a bunch of other sort of forms of payments. So I'm assuming you spent a lot of time thinking about this and you guys have obviously been acquisitive recently with, with bridge and privy as well.

Yeah, this is one of the areas in which stripe is very problem solving oriented and not technology or particular technique religious. We think that humans are going to have a variety of ways that they want to pay and hold money. Stablecoins is a phenomenal way for many people in the world to hold funds and for businesses to move it across borders. And so we expect that stablecoins will be a very popular way for consumers and businesses to just interact with themselves. Then you have businesses who are also going to have, you know, they're going to have a long adoption curve when it comes to accepting and holding crypto assets and then in some purchases, stablecoins might make sense between two parties that natively know how to interact in stablecoin, but often it might be the case that Jordi has an AMEX card and the seller is expecting an ACH transaction. And we're sort of missing a universal way for all these types of currencies and Rails to work together. Visa also announced a new way of being able to hash your card and give it to an agent with this Visa agentic token, where Stripe is one of the first partners to implement it. And I think we're just going to see this new proliferation of new ways that money can transact between parties. And we're going to need some type of sort of Babelfish translation service across all of them, because if you're going to pick one route, then you're going to likely exclude many of the agent humans and businesses in the world.

That makes sense. How are you guys thinking of. Not to go too broad, but the business model of the Internet, agents change things. The Internet today is heavily reliant on advertising. And if you have a bot just crawling a website, you're. Or, or even when you look at other, other services. And so we've talked. Ben Thompson had some good writing around just like what the future business model of the Internet could look like and potentially micropayments. But I think the takeaway from that, our takeaway is like, there's so many different stakeholders that would need to find some type of alignment. It's. It's hard to see like the obvious path forward here.

Yeah, I think a universal want from businesses is just more channels to reach their customers and be able to do so in more direct kinds of ways. And so if you go to a SaaS software provider and you said, hi, I sort of two choices for you. You can have this very cool large budget for a 101 billboard and kind of hope that at 85 miles an hour, developers like, like see your ad and then remember to implement it later. Or would you like them in situ as they're working to have agents mediate the purchase, recommend it, and be able to integrate and accomplish your thing in five seconds right inside their editor. Okay, yes. Well, first of all, I'll do both. But also the second one sounds very nice because I'll be able to directly attribute where it came from and be able to have a great CAC for that. And the LTV should be even higher because the robot even integrated it directly. And so I think that we're going to see new channels emerge for Monetization. Both usage based through MCP or other ways that businesses are going to expose their APIs to agents, but also for transaction based referral fees which which will supplement affiliate and then I think it'll be a new way for businesses to make sure that their agents can read their docs, can read their product. SKUs can have access to that information in a new permissioned way. I really like the Karpathy talk that got posted last night where you basically said that if your docs involve a click, not good. Because agents want to act and not click and just only read. They want to start acting. And so that's why Stripe is, you know, if you go to the Stripe docs we really push like hey, here's our MCP where you can just just talk to the primary best way of integrating Stripe and it can do it on your behalf rather than, you know, just reading or reading something from a three year old corpus.

Interesting last question for me. We want to move on, let you get back to your day. Stripe was famous early on for having this crazy kind of open culture around emails that anyone from the entire organization could read. That seems like incredible foresight to the moment today because you don't have all this privacy information that oh, do we train on that or not? You could very easily fine tune a model or do some sort of embedding on the emails that are already deemed to be worthy of the entire organization reading them. Is that still part of the culture? Is there a tool if you join Stripe where you can get up to speed without needing to read every email email, but you can kind of get the Stripe way of doing X, Y or Z. Talk to me about Stripe's culture.

Stripe's has a very serious writing culture where any decision I've been a part of for the last seven years I can really point to some Google Doc that has the pros, the cons and the decisions as well as sort of the email culture you mentioned where we just. It's very commonplace at Stripe where if you write you spoke to a customer or even after going on tvpn.

Hi.

When on TVPN you just CC a notes list and now it's available for anyone who wants to subscribe to Notes List. But one of the major subscribers to Notes list now is agents. And so if I'm in Slack we have this really awesome bot called Trailbot that's read the trail of everything, all the paper trail of everything that we've done. That's permission to it. And I can just say Railbot in Any Slack room. And it has the context both of the team slackroom I'm in, but also the full corpus of Stripe and all of our permissioned wikis and documentation and internal tools. And it takes the first line defense of most questions immediately. And we actually have it to the point where it knows to jump in automatically without you even asking it. And so I find that most of the time we're able to just at Trailbot and answer a lot of questions. And then increasingly these agent tools, which I think are going to apply to commerce soon, quickly too. They're not just read only, they're going to start taking right actions and purchase actions. And for Stripe internally, right actions might be to roll back that deploy or to auto communicate to that customer because of an MPS score under 10, which we do often, hopefully not too often. But then in the real world, if you want to make some of these actions, you're going to need to prove who you are, pay for it, make sure the merchant was able to accept that money, get the entitlement and move on.

Yeah, even something as simple as like you show up to a new company, hey, there's this system over here that we're using and I don't have access. You might go to a wiki and ask, how do I get access now? You just ask and it just does it for you.

It's so interesting to think about if there's like some type of user flow where if somebody sends a Slack message, there's like a tiny delay built in and it gives like a bot an opportunity to like actually front run the question because it's like every message is going to waste like, you know, 10 minutes.

It's like immediate. A new version of Shadow Band where you first get your question answered and then.

Yeah, do you really want to ask this question? Because it was answered here, here, here, here, here. And like here's our recommended action.

It's like pro autocomplete. It's amazing.

It'd be interesting. Yeah, it's. Slack's just become completely silent because everybody's.

Like doing things and just immediately getting.

At those of us who have nerdily taken notes and made docs over, over years. It is somewhat of.

You're waiting for the.

It was worth it.

Yeah, yeah, yeah. Made fun of by some people for a long time, but it all came back. Well, thank you so much for stopping by.

This is great. Jeff, always welcome.

Yeah, we'll have you back soon. You can talk.

Great to see you all.

It's great. Talk to you soon. Bye.

Let's give it up for Jeff.

Next up, we have Garrett from Handshake coming in. He was mentioning the information. We've been mentioned in the information. It's a bunch of information boys hanging out on the chat.

We love the information.

We love. We love the information. Thanks so much for joining Garrett Lord.

The nominative determinism is insane.

I think we love something also in common. Sauning. I'm a big, big sauna guy.

No way. There we go. Yeah, yeah, the sauna is important.

Devastated to hear that. When we moved into this new studio, we don't have a good.

We don't have a good sauna set up. But we'll figure it out eventually.

The cool plunge can be. Can fit nearby though. I mean, there's still opportunities.

Good. Yeah, maybe we got to get in the cold pledge game. Anyway, anyway, kick us off with a little introduction on the business. Obviously it's in the news today. We covered a little bit about it earlier, but I'd love to get you to explain the business, a little bit of the history and the positioning of the company.

Yeah, for sure. So, I mean, the business started way back when I was in college. I started Handshake out of a personal pain that I faced in breaking into find my first internship and first job. I went to a no name school in the middle of nowhere called Michigan Tech. It's awesome if you love to ski or love the cold, but if you wanted to break into Silicon Valley, nobody had really recruited there before. Fast forward to today. Handshake is the number one place that young people in America start, jumpstart or restart their career. We're like kind of an 18 to 30 early career network. There's a million employers that use Handshake. So it's where the vast majority of employers recruit undergrads and interns and people after school. And then there's 18 million students and young professionals use the network. And we also power about 1600 universities in the country. And the background, I think that's important for right now in this very moment is about 18 months ago, many of the Frontier Labs as well as the large annotation engine companies started reaching out to us with basically asking us, beating down the door, saying like, do you have access to PhDs? Do you have access to master's students? And for us, that was incredible. I mean, we have 500,000 PhDs in the network. We have 3 million master's students on the network. There's tens of millions of undergrads in the network. And we started Serving these players with experts. Really as the world has evolved from training frontier models, it's moved from generalists like drawing kind of boundary boxes around stop signs to today experts. And experts are in law, finance, medicine, mathematics, physics, chemistry, biology. These labs really are hungry for reasoning data to help improve with human in the loop, the actual frontier of what their models are capable of delivering yet alone in the future. And you talk about tool use or trajectory. So they started reaching out to us and saying, do you have access to these PhDs and Master students? And we started providing, we were the leading provider of all this talent and we really started to realize is that people weren't getting paid on time. They were really confused. They would go through training and kind of get dropped out of a leaky bucket. We heard from students that were successful on it that they loved the money, they love learning more about some of this AI tooling. They wanted to use AI tools in the classroom, they wanted to use it in their research. And so given that we have this huge supply and zero customer acquisition costs, we started building a human data business. And really in the construct of building that business, the focus is really around like how can you also think about evolving and automating a lot of the recruiting practices? Recruiting is still, you know, it's sourcing, it's screening, it's scheduling. There's a lot that AI can bring to bear on that. And so we now fast forward to today in the last six months have been working now with six of the frontier labs. We provide them tens of thousands of.

It's a lot of them. I didn't even know there were six. Only five. The big six.

Count them up.

You got them all.

And we provide them with experts to help make their models more effective.

Very cool. Talk to me about like what the. How. How are the frontier labs thinking about human data annotation and answer generation? It feels like we might be at the end of the, of that story soon or maybe we're shifting into more of a focus on the areas that are less verifiable, less like write the answer to an IMO level math problem and more in the biology and legal context where the models are falling behind. Like where, where are the pockets of value? Where's the most demand within the human data generation industry? And where do you see it going over the next couple years?

Yeah, so maybe I go like from the latter part of the question. The first, like where we see going over the next couple years, it's definitely going to evolve into audio. It's definitely going to Evolve into tool use. It's definitely going evolve into trajectories. And experts will be needed to provide data. Imagine almost like recording your screen as you're conducting a task. Maybe you're building a slide deck and doing it. You know, if you're an investor construct, like doing a DCF and doing competitive research, they want more data to be able to help improve these models. Especially think about like agents, right. And step by step problem solving as where the puck is right now and where the puck will continue to be if you talk to a lot of the frontier researchers is they need expert data. And expert data is in basically every esoteric area of human knowledge. They want to, you know, the models have already kind of sucked up the entirety of books and YouTube and, you know, human knowledge. And what they really need is they need special data to be able to make and understand the step by step reasoning that's required in order to be able to kind of fuel the future. And so if you think about academia, these PhDs, like, what is the definition of getting a PhD? The definition of getting a PhD is like pushing forward an area of research that nobody else has done before as peer reviewed by your peers. That's how you get your doctorate. And so this kind of perpetually reoccurring stream of PhD students and master's students are really valuable in this very moment. And it's also to zoom out to their experience. Like you can make, I don't remember when you were in school, but you can make like 23 bucks an hour being a teacher assistant. You know, you could dive, you could drive doordash. And we're paying these students like 60, 70, 80, 100 plus dollars an hour. And they're also, we can connect it to actually getting jobs. So we envision a world where like you get badges on your profile and there's like leaderboards by school. And we're actually, I mean, what better way to articulate your skill than actually proving it by being able to break the model or by being able to provide a model feedback. And so we believe that we can help you get more jobs with a million employers in the network, help you build your professional reputation and articulate your skills, all the while making like $100 an hour when you want to. I mean, it's a gig job.

Yeah.

How do you think about financing handshake going forward? I'm sure you're making, generating a lot of revenue. You're clearly paying a bunch of your network out quite a lot. We were just learning about surge AI earlier And what they were able to do while bootstrapped. I imagine even in the last week you've had investors reach out trying to say, hey, scale's out of the game. You want 100 million, you want to dance. But how are you thinking about the business going forward?

Yeah, I mean, one of the ways we think about this market is like, you know, if you don't have an audience, there's no moat. What our competitors are doing is they're at some of these companies, they'll have hundreds of people who are recruiters, sitting on top of platforms, sending messages on companies like Handshake, or spending tens of millions of dollars a month doing performance advertising, trying to acquire experts on Instagram. You can imagine if you're like a physics PhD and you get an ad on Instagram for a company you've never heard of before claiming they could pay you $100 an hour. It's kind of a jarring experience. And so because we built a decade of trust in adding a ton of value to these users lives, we have no customer acquisition costs. And what that means is that we can pass along all those savings by paying contributors, we call them fellows. It's the MOVE Fellowship program. We can actually pay you more than any other vendor on the market. We can also pass along those savings to the Frontier Labs. So as you think about our overall P and L like our gross margin and ability to scale this business, considering, you know, the moat is the network that we've built, we sit in an amazing position to, you know, to, to grow extraordinarily quickly. And that's what we've been seeing. I mean in the last, you know, month we've grown by over 3x and you know, it seems like there, there's a lot of demand continue to be out there.

I can imagine. I had no idea it was that big though. Let's go, let's go. Three hits for 3x. That's incredible. Are there any. Last question, we'll let you go. Are there any like weird areas that you think we'll see this type of human data generation pop up? I'm imagining like AI seems to be at like 150 IQ. It can write code and yet it can't like book me a flight. Do we need to take like flight like travel, travel agents and have them go through the workflow so that they don't get hung up on, should I sign up for the credit card or do I want, you know, insurance on this flight so that we have a whole bunch of data specifically about that Task. I'm just interested in this concept of like these economically valuable but highly niche tasks that don't seem to be, we don't seem to be getting closer and closer and closer to like one shotting them with the current models. And I'm wondering if we're going to see this long tail of different hyper specific business use cases like what we saw in SAS where there would be hipmunk just help you book flights better. Is there going to be a flow where there's a new startup that's doing AI agents for flight booking and then they're coming to you for a ton of data generation around how to actually book the correct flight because it learns whether or not you're okay with, with a layover or how price sensitive you are. All the things that you would get from the interaction with a human flight travel agent. Is that something that you think we'll see or is that kind of just completely tangential?

No, I think that's totally something we'll see. What you just described is like a trajectory called a browser trajectory.

Sure.

And that's basically like you have a goal in mind.

Yeah.

And you, you, you know, you have like a step by step kind of thoughts in your mind around how you accomplish that. And you navigate tools, you navigate the browser. You stitch together your own intuition to be able to accomplish that task.

You might look at your own calendar. When do I get off work? How look up how long it takes to get to the airport. It takes me a different amount of time to get to Burbank than lax. What's the parking like? Like there's so it's such a simple task because you think about like anyone can do that job for you and yet to do it well is actually really hard. Hard.

And you talk about just being able to talk to a model.

Right. Like totally.

You don't even need to log in. Right. So you're going to need audio data, you're going to need trajectory data. You're going to be able to interact with APIs, humans, experts will be needed for the next several years to be able to make that data happen.

Interesting.

In order to be able to power the frontier of where you want to see it going.

Well that's exciting. I want to book a flight with an AI. It still hasn't happened. That's my own personal touring test. Hopefully you can make it happen. But thank you so much for stopping by. This was fantastic. Fantastic.

Be sure your time.

We'll talk to you soon.

Great to meet you. Cheers.

Coming in next, we Have Taney coming into the studio to the TVPN ultra dome.

A massive round.

Oh. Oh. We're gonna hit the gong again. 10th time of the show. Always a good time.

There he is. Welcome. Welcome to the show.

News for us. Hit us with an introduction. Hit us with the news. What's going on in your world?

Think we might be muted.

Tane, are you there? Can you hear us? I'm itching to hit the gong for you. I hear there's gong worthy News.

Are you there? I'm gonna send him an email. Okay, you are live.

You are live on tvpn. Okay, we'll pull them off. In the meantime, I will tell you about wander. Find your happy place. Find your happy place. I can hit you with the gun, but book a Wander with inspiring views, hotel, great amenities, dreamy beds, top tier cleaning and 247 concierge service. It's a vacation home, but better folks. We told you about don't realize. We told you about audio. We told you about Polymarket. Told you about linear. Do we do Vanta Automate compliance. Manage risk, prove trust continuously. Vanta's trust management platform takes the manual work out of your security and compliance process and replaces it with continuous automation. Whether you're pursuing your first framework, we're managing a complex program. Let's Hear it for vanta.com.

Give it up for Vanta.

And if we got an extra minute. Is he here?

Are we back?

Welcome to the stream.

We made it. How you doing? We made it.

Fantastic.

Sorry, we have some audio issues.

Oh, no. It was a pleasure. We got to do extra ads, so you know, you're making my day.

It's a dream. It's a dream.

You're making my day.

What's going on? Quick intro. You've had a big day. What's happening?

Yeah, thanks for having me. We announced a $200 million around.

Whoa. With general. That's fantastic. Buried the lead journey.

You guys gotta start selling those. I feel like we need one in our office, but clearly.

Yeah.

No, we're super excited. Congratulations. Obviously works in healthcare. You know, we power AI workflows everything from ambient to revenue cycle payments in large hospitals.

Oh, interesting.

Give us. Give us a quick history of the company.

Yeah.

Because it's not often I see a 200 on 6 something billion. I hadn't heard of the company before, but I hadn't. Josh Browder connected us last night.

Amazing.

And so I'd love to hear your quick story, kind of how you got here.

History of the company. I want to hear about the first customer too, for sure.

So, I mean, Josh is great. I've known him since we were at Stanford together same year. And the, you know, the story behind Khmer is interesting because Khmer started as an incubation inside General Catalyst. It, you know, the best analogy I have is it is Hemant's Palantir. Very focused on healthcare. I started a company while I was at Stanford called Athelis, which was focused on applying language models and computer vision in healthcare. We started as a blood diagnostics company and then eventually grew into this mid market SMB OS for physicians. We merged the two companies about a year and a half ago, or almost two years ago now. And then I took over as CEO with our management team. And so it's really, you know, it's sort of a coming together of these, of these two businesses. And yeah, I mean the company powers large hospitals, about 250,000 physicians and nurses. We power the, you know, private practices here in California. That was our first, first customer. It was someone that my co founder Deepika literally walked up to and, you know, cold knocked on their door and then got them to use one of our first devices and remote monitoring solutions.

Amazing.

And yeah, that's, that's the quick, quick story.

I got a bunch of questions. It is, I want to kind of like contextualize this around the broader General Catalyst discussion because there was news, I think just today that Ohio authorities approved the first ever purchase of a U.S. hospital by a venture capital firm. That's General Catalyst's bid to acquire Summa Health, a hospital system in akron with over 20 facilities. And I'd love for you to, I'm sure you've studied this. What is going on there? And then is there any sort of synergy across the portfolio, venture capital? General Catalyst has had like a very differentiated strategy there, but I haven't had the chance to dig into it. So I'd love to get you to contextualize it and then we can go into how this links to your business again.

So the Summa transaction is super interesting. It is a venture capital firm buying a health system. Transforming it and can you. Is obviously a big part of that. We're serving as the office of the cto, so our engineers are forward deployed. We work hand in hand with the Summa IT teams. We've been working with the revenue cycle leaders, the clinical leaders. And it's a really special system. I mean it's in Akron, if I'm not wrong. It's where LeBron James was born, literally the hospital itself.

And many people have been calling you the LeBron James of healthcare AI.

Now we got to put that out there.

Yeah, I mean, I maybe have been the first person to say it, might.

Have coined it here, but many people, I'll say it right now, you're the LeBron James of healthcare AI. Now many people are here, not just.

One, two as many in our book.

We're just going to make that a thing now.

Fantastic.

It's remarkable because running a health system is super hard. It is a 1 to 3% operating margin business. Most of them go out of business. And I think what General Catalyst believes in is language, models and technology can transform the operating margin and also lead to better care. So it's not a PE cut and juice play. It really is an investment.

That's awesome. Talk about comur's overall product strategy. You guys have a number of different products. It seems like a very different. We've talked with founders and covered companies that come on and just want to own one key area. But healthcare feels like somewhere where if you can get embedded with the set of customers, you can more rapidly kind of add products to the platform. So I'd love to understand the product strategy.

We really look up to businesses like Rippling and RAM where there's this concept of you enter with a wedge. And in our case that wedges either ambient AI, which is a tool that helps a physician document and really automate the revenue cycle of appointment, generates the claim automatically and then the back office, which is all. When you walk into a hospital, there are tens of thousands of people at large health systems whose sole job is fill out claims, call up insurance companies, fight denials, fill out new forms, all of that's going away with LLMs. And our belief is that if you do that as a point solution, as like a single, you know, little part of the part of the solution, you might get some initial usage, but eventually the EMRs like EPIC or companies such as ourselves will just eat you and you have to be that compound startup from the get go. And I think payments is a really interesting vector to deploy software. Ramp has shown it where you get into the transaction suite and then you build a whole bunch of tools for the CFO's office. We're trying to do the same for a health systems CIO and cfo.

Can you tell me a little bit of the, the history of the healthcare industry broadly and how I know that there was like this kind of catalyst around Obamacare? I remember talking to Jonathan Bush, the founder of Athena Health, about electronic health records mandates and there's been a number of changes kind of at the federal level that have kind of opened up different pockets of opportunity. Like what is the story that you tell about the recent history of health healthcare in America?

I think it's, it's fascinating. The 90s physicians had amazing lives. I mean they, they drove Porsches, they had work life balance, they had personal relationships with their.

Let's hear for Porsches. We'd love to hear that. Let's get back to Porsches. We had more of those.

We need to return return.

And, and you know all, all in.

All physic patients got great experiences too because, because of that personal relationship. And then the admin work tax just increased. Everything from insurance to filling out an EMR digitization came in the 2010s with Obamacare and meaningful use. And really EMRs proliferated and Jonathan Bush and Judy and all these people are legends in the industry because they built Athena a $20 billion company, Epic, probably $100 billion company now on the backs of that very quietly and under the radar for most of Tesla. I think the theme and the story of today is labor is turning into software. And whereas most white collar labor in America, it's in healthcare. Where are the majority of administrators sitting behind a computer clicking on forms? It's in healthcare. And we believe that the EMR will be transformed. We also believe that the labor stack of healthcare will be transformed and it'll create more operating cash flow for hospital owners.

Is that narrative of the like the, the administrative ratio or the administrative load increasing, is that similar to what happened in, in academia? Because I, I remember seeing these charts of like the ratio of professor. Everyone loves the idea of like a high functioning university with a lot of professors teaching students and a great ratio there. Everyone's a little bit more skeptical about like wait, why do we have five times as many people to admin. Is that the same thing that's to going on in healthcare? And kind of what was the underlying driver of that? Was it just regulation or lack of tools? Where did it come from?

I think it's very similar. What I will say is I think in health care it bred more out of necessity and in academia it just kind of happened. In healthcare there's this game of attrition between the insurance company and the provider and you're they're making it a little harder every month, month every year to get an approval on a claim. And as a result the health system needs to add a couple more people in order to fight those claims. And then it just kind of built up into this arms race. And I think the insurers kind of carried the power after Obamacare. Like when you look at UnitedHealth's market cap, I mean it's like what is it like a 12x since Obamacare got passed? It's quite shocking. And the power dynamic I think will shift again back in the favor of physicians and hospitals because of LLMs and because of what you can now automate.

Yeah, it was kind of just like the game theoretic Nash equilibrium was like hire a lot of admin staff.

There was no other option.

Yeah, talk about your personal ambition and the team's ambition. You're a $6 billion company now. Seems like it's cliche, but it's it. The way you're talking, it feels like you're just getting started.

Is the job finished?

Yeah, yeah, it sounds like the job's not finished. I don't want to put words in your mouth.

It's not finished. Look, I think when you walk into a healthcare practice, the inefficiency is shocking and the positive intent from the physicians and the nurses and the caregivers themselves is all there. And I think all it takes is for a company like ourselves to come in and try to nuke that work to tax. So our ambition is, look, we're going to come after the EMRs, we're going to come after the payers or revenue cycle businesses. This is a $4 trillion industry you can build for a very long time. But what done looks like is when you walk into a physician's practice, scheduling, intake, insurance, like all handled, there's no filling out a little clipboard of the same information again and again. The appointment happens and instantly the doctor is paid out. There's no reason we can't have instant adjudication instead of, you know, waiting 30, 45 days. But it's going to require a system overhaul and like new payment rails to go do that. And that's really what. What's at the heart of what comer is building.

Awesome. Well, this is super exciting. I'm glad you're doing what you're doing. And you are new. You're our new healthcare expert and correspondent. So expect to call.

And the LeBron James.

And the LeBron James.

LeBron James of EM Mars.

Yes. All right.

Congratulations on the milestone. Hope, hope to have you on again soon.

We'll talk to you soon.

Cheers.

Have a good one.

Should we do some timeline?

Fun show. Fun show. Yeah, we definitely should. We got to talk about Sam Lesson's Oracle vs Salesforce he's getting in. He's getting in hot water.

You got in hot water.

The timeline's in turmoil. We love Sam Lesson. On this show. He. He posted a screenshot. He says, I will defend big tech.

I will defend Sam lesson.

Oracle is 2x Salesforce, but Ellison is worth 25x Benny off what this sale says about the limitations of the SaaS business model. He said he had a fun riff yesterday with the slow partners on this. Oracle is obviously crushing it. But if you take a today snapshot, basically the market cap of Oracle is 2x Salesforce. 500 billion versus 250 billion. Meanwhile, according to previously directional at best data, Benioff's net worth is 1/25 that of Larry Ellison's. 10 billion versus 250 billion. What do you learn from that? What lessons do you draw? I like this. The revealed preference for founders and companies. The old licensing model is better than SAS. That's interesting.

Imagine having 10 billion and just getting Lil Broed by Larry. He has a sort of little, little broing effect on most people.

There's an amazing story about a little bro, a famous little broing, where Phil Knight of Nike was worth something like 10 on the order of like, $10 billion. And he was in like, maybe Sun Valley or something, going to a movie, and he runs into Bill Gates and Warren Buffett who are just going out to a movie, and they're. They're both worth 10 times him. And he's just like, yeah, this weird awkward moment where I was like, nervous 10 meet them for the first time in a long time. Because typically he's like, the most successful businessman he runs into all day. Right. But he was just like, yeah, he. In his book Shoe Dog, it's a fantastic, healthy road. Shoe Dog is a great. Is a great book, Great book. And he talks about, like, all the weird effects of, like, having immense wealth. How, like, his wife would, like, hoard immense amounts of, like, paper towels just because they're like, money's no object. Like, what should we do with this? And I get a lot of paper towels and they had, like, figure out, okay, this is like some weird psychological thing that's going on in my brain. Like, I don't actually need all these paper towels. The fact that money's no issue doesn't really matter.

People like to talk about, you know, you're the, you're the, you're the what, whatever, the. The average of your five friends. Yeah, it's like, yeah, well, if you want, there. There should be some similar law of, like, like, your growth rate is like, should be. Is like, like tied to how often you're a little bro'd, you know?

Yeah, yeah. Never get little broed.

No, no. You want to be getting little bro.

Oh, yeah, yeah. That's true. Yes. You can be on the upward swing. That's right. If you're not getting little broed enough, you're not on an upward trajectory. This is good. Yeah, yeah, this is good. I've been in that situation before.

Anyways, we could cover what Sam said, but I think we can just skip to.

We're also gonna have him back on the show.

We're gonna have him back on. On the show. He's a regular. We're going to skip to Miles. He says, wrong take. Allison is much richer because he didn't sell shares and has steadily been buying back 2% of the company every year for 30 years. He's increased his ownership from 17% to 40%. Such an incredible story. Founders complain about dilution.

Yeah.

Oh, you got diluted.

Yeah.

Oh, why don't you just buy back shares every single year for decades? If Ellison keeps doing this, he could very well own 150% of his company at some point.

That's the future for OpenAI. OpenAI just becomes the agentic organization. It just buys back so many shares that it eventually owns itself. That's the real goal.

Anyways. Miles says meanwhile, CRM, AKA salesforce, made a lot of dilutive acquisitions. And Benioff said at least selling sells his shares yearly. He doesn't sell them yearly.

Daily. Daily 2 million daily. Buko Capital, say.

Oh, liquidity events, you know, they're. They're few and far between. Not happening every day. Not for Benioff. Daily liquidity, that's pretty good. You know, the real, you know, another. How much does he pay Matthew McConaughey to just.

Oh, yeah, that's got to be pricey.

I think it's only like 10 million a year.

So it's like, yeah, a couple super bowl ads. Not bad. So Sam responds to the hate. He says, since a lot of folks are making the same comment about buybacks versus sales strategies, that is at best the noob answer. If you're smart, you understand why they have different paths. And the answer is path dependency from business model quality. Take a 201 level class and then boohoo. Capital Blue quote tweets that and says the timeline in turmoil. Wrong again. CRM has executed poorly. They've deluded shareholders with bad acquisitions. They have 75,000 employees who. They give excessive Stock based compensation to they let hubs scale up in their face. HubSpot they've diluted versus shrunk their share count versus the other companies Adobe that eat shares. Investors don't trust him. If Benioff held and cared about shareholders it would be a closer call. He doesn't care. It's not about the business model. Well you'll love to see some timeline and turmoil. Very, very fun. In other news, Shield Monat has the story about Telegram's founder Pavel Durov. Consistent feature on the tech bro Drip account.

Everybody says they're pronatalists until they, until you ask how many children have you fathered sperm Dr.

Apparently he has fathered over 100 kids via sperm donation and he is worth $14 billion. And he says he'll leave his fortune to all of them with no difference between his six kids conceived naturally via compared to the hundred via sperm donation. So every one of them is going to get $140 million just to kick off fundraising. Just start investing that you got your family office on day one if you're one of Pavel Durov's kids. Pretty remarkable.

Yeah. Single lp. It's kind of a good dynamic.

Yeah. I wonder how he's going to get liquidity for, for Telegram at some point because you get a bunch of Telegram shares it's kind of like this, this difficult beast to wrangle. But I mean I guess you take it public and at some point and get liquidity out of that. I don't know. I mean it also just prints money. So even if it's like 14 billion like you could just like get a stake in the, in the distributions because it's making money.

I think he kind of figured out life and wanted to make his life basically a hundred times more complicated by you know having having this type of dynamic, you know not just with his many children that that he helped conceive directly versus the hundred others. So had to one up up. Elon had a little bro Elon and.

Elon's commented on this too. He was like oh I got rookie numbers. Genghis. Genghis Khan over there is really taking over the world.

Genghis Khan of encrypted messaging.

Yep. It's very, very odd. Only CFO says finance department out drinks sales. Feels like sales is inviting finance to the party so they can stick them with the bill. And this is, this is the data you can only get from ramp ramp.com data apparently apparently finance marketing sales teams lead in alcohol spend alcohol as a share of business meal spend not in that order.

So marketing is absolutely dominating.

Dominating.

20%, 19% of all spend on alcohol.

No, no, it's 19% of business meals are alcohol. So if they go out and they're getting $80 worth of food, they're adding on $19 or something. Or $81 of food, $19 of drinks. That's the idea. Alcohol share finance. They're getting $84 of food, $16 of booze.

Marketing is drinking sales under the table.

Yes, yes.

Narrative violations.

IT in the tail end there. 9.7%. Many Huberman devotees in the IT department. Apparently.

Yeah. Not a power lunch. Power lunch, you know, category.

No, but the three martini lunch will make it back for the, for the tech teams. Should we go to this story about the Vibe coder who sold his business to Wix for $80 million? It's only a six month old company and there's no external funding. $1891-980001-89000 in profit in met. Bryce Roberts is just the beginning. There's going to be more stuff like this. I thought it's pretty cool. Base44, only employs six people, hasn't raised any external funding. The 31 year old built a viral app, AI App Maker as a side project. So you go in there, you design an app. Obviously plays very well with wix, which is in the website building business, but he flipped it for $80 million and he's post economic now. Congratulations.

Yeah, when I saw this headline, I was confused. I was like, okay, so he just vibe coded something and sold it. But it is a tool to build a Vibe coding tool trusted by over 250,000 builders worldwide. And nice, quick flip.

Amazing.

He's basically getting a similar outcome to, you know, a founder that sells their company for a billion dollars, but, you know, goes through a bunch of different financing.

Or kind of like a mid tier AI researcher these days.

Yeah, starting, you know, like starting out.

Yes, starting out. In other news, John Carmack is absolutely jacked. This is fantastic news. Yaxin has the news. He's looking, he's looking very built. But John Carmack chimes in. He says a chunk of this is Jeff, his wife, dressing me in tighter shirts. But I did put on several pounds of muscle this year after switching my random grab bag of vitamins and supplements over to Brian Johnson's blueprint system. Let's hear for Brian Johnson. Really making a difference in the technology world. I was probably not getting enough protein to take advantage of the exercise I was doing. I have always been roughly upper Quintile for fitness. Let's go Regular exercise but not the level of serious athletes that most, most offices tend to have a few of. And now he's looking built polymer. Lucky chimed in. It's a great, great day on the timeline. Let's check in with Tyler, close out the show.

I was going to check in with this poly market. Will Chamath launch a spacing? It is up to 70%.

It's up to 70%. It was 33% when I posted it this morning. Wow, that's big news.

It was partially because part because he came out and he said. What did he say? He said 58. He. He asked yesterday should I launch us back? 58,000 people voted. 71% said no. He said I hope everyone that voted no feels seen. Now onto business. I got calls from many Wall street and crypto titans yesterday. They all want in and their vote matters a lot to me. So I want will probably do it. Maybe this time it will go better. Who knows. The risks are clear though the last time wasn't a success by any means. I will include this poll and the community note in every SEC filing possible. It will make an excellent disclosure about the risks and is not short of irony. So what kind of company do you guys want? No crying in the casino.

Let's go.

Chamath people are absolutely fuming. The comments I'm sure but honestly.

Get after it.

Pretty fair.

Everyone knows he's gonna play by the.

Rules, you know and I think, I think at the end of the day it's, it's very like I, I look at the next Chamath spac is like it, it totally would like. It probably will pop. It'll probably. It'll get a lot of attention. It might turn into a meme stock. Right. I. I will be interested to see what kind of Target He.

I'm 100% excited to follow the story. Yeah, it's going to be fascinating. Anyway, let's check in with Tyler and then close out the show. Tyler, I have a question for you. Can you guess a number between 1 and 50? Like a random number? Yeah, random number between 1 and 50.

27.

27. Are you an LLM? Did you see this? Every single model, they all guess 27. When you ask them a number between 1 and 50. ChatGPT, Claude, Perplexity, Meta. They all guess 27.

We gotta get a world. We gotta get a world coin orb in here to be able to prove.

That Tyler's not in fact, yes, we do. He might just be a Deepak Final review. What did you get done this show? Did you keep playing with Mid Journey? Were you doing something else? What's been on going, going on the last couple hours? Yeah, I think I, I just sent another video. Oh no, I've been pretty productive.

You can watch this.

Let's see. I like that. The gorilla crosses behind.

What's he doing? Did he just. Oh, he just took my spot.

Oh, he took your spot. Okay.

Wow.

Oh, he comes in with a paper. Breaking news. The breaking news gorilla.

The breaking news.

We need to get you a breaking news gorilla outfit. And, and if he has breaking news, he can print it out. Come sit down. Take our seat. That'd be amazing.

That's good, that's good.

Are there any others or. We're closing it out.

Yeah, I think that's it. Okay, that's it. Good work, Good work.

I feel like the production team laughing like they have some other productive that are too scary for. For our audience.

I saw, I saw one get sent in the chat and it. And it just was really scary looking.

Well, we will be back tomorrow. We have a great show for you folks. Leave us 5 stars on Apple podcasts and Spotify and thank you for watching.

Being here with us.

Fantastic show.

Have a great evening.

Goodbye.

Love you.