Josh Wolfe on AI and the Breaking of Silicon Valley's Social Contract | Odd Lots

The VC's incentive is two things. One, get the best return so that you can compete. If I'm only earning 12% and a peer VC is earning 20%, money is going to go where it's going to be well-treated, and I'm going to lose to that. So the cost of my capital— for the cost of an LP's capital— is outperformance. So I've got to outperform, which means I have to be earlier, I have to own more, I can't just do stupid deals. Sophisticated LPs will not just look at the logos that you have. Which is the game that you've always seen.

Yes, sophisticated LPs.

From mutual funds and from some hedge funds. You see the Q4 filings, and they always threw in the name, oh, we were in NVIDIA, and they would market their top 10 holdings. But yes, they lost money on it, right? And so that is a really important incentive. And the sophisticated LPs will actually know down to the partner at the firm or the team or the deal team who is responsible for this. What was the entry point? They will talk to the founders and say, who was your most valuable investor? Who got you your first 10 hires? Who helped with your syndicate construction for your later rounds? Who made customer introductions? Who was a valuable board member? Who never showed up? Who was asleep in the board meetings? All that kind of stuff. So there is a level of due diligence that LPs have to do to know, are you a value-add investor, or are you a poser or pretender that's just buying a logo or a brand name?

Hello, and welcome to another episode of The Odd Lots podcast. I'm Joe Weisenthal.

And I'm Tracy Alloway.

You know, sometimes with the whole AI thing, it feels like the, are we back? Is it over? There's been a few moments in the last year, whatever it was like, is the bubble bursting? Is there a bubble? I don't know where what this week is. We're recording this September 3rd. There's a little, I don't know, there's some tremors. I can't tell what's real or not.

I feel like the hype cycles get shorter and shorter and more compressed. But you're right. I think there are maybe some more jitters than there have been previously. Like maybe.

Yeah. Only maybe.

It's hard to tell.

It's hard to tell.

But I think even if we're not there yet, we are getting maybe to that point where like the rubber needs to meet the road in terms of monetization. Yeah, like, you know, everyone got— maybe, right?

Maybe.

Yeah, you're right.

Maybe yesterday, uh, or, uh, either this morning or yesterday, uh, Anthropic, $183 billion valuation. This does not strike me as like people worried about money. Suddenly people are— well, or suddenly people are worried about valuation. It sounds like people really want to pour a lot of money into these companies.

Yeah, it is. It's a weird environment. Let's just put it that way. That's the only thing we can say with certainty that doesn't start with a maybe.

That's right. Here's another thing we can say with certainty. Alphabet, the shares are up as we're speaking, 8-point-something percent today. This was a company people were sort of worried about how they would do in the AI era. They're killing it. They're at an all-time high again today. So again, like Many things are like, it's tough to get a read on things.

It is indeed. Who do we turn to?

I love this vague intro. We just want to have an excuse to get to the guest. And so we're going to vamp for a little while and then we're going to get to the guest.

I just tried to throw to the guest joke. Who do we turn to for a read on the true feelings of the market around AI?

That's right. We do have the perfect guest. We're going to be speaking with Josh Wolfe. He's the co-founder, managing partner at Lux Capital. VC who's been in this space for a long time, a guy we like to turn to to figure out what's actually happening at any given moment. He sometimes even has good interest insights in public companies, too, not just private companies. Josh Wolfe, thank you so much for coming back on the podcast.

Great to see you both. How are you guys doing?

We're doing great. What's up with Alphabet? Some people thought they were going to be a big loser in AI because they have this legacy search business model and O3 is so much better for searching. Here they are surging. They're at an all-time high. What is the, what is the smart take on Alphabet right now?

I think they are crushing it. I think they are the sort of dark horse underdog. And maybe the second that I would put it with that is Apple, which people are totally—

Yeah, totally.

And, you know, the irony is, you know, if you think about the big players, Meta with Zuck's crazy poachapalooza, you know, which we'll get into where, you know, $100 million pay packages and trying to disrupt to everybody and bring them all on. You're seeing talks about them. They've got Midjourney coming in for the images, so not relying on their own models. They're talking about— oh yeah, Google or, you know, potential integration on search. And so, so it's interesting that Meta, having committed to first be, you know, the metaverse and change the name to Meta and then going all in on AI, is actually going to be turning to some of these other players potentially. So I think that's surprise number one, that people thought that Meta was going to be in the lead. And I think that Google and Apple were both sort of counted out, but both are very serious contenders. If you look at the top 2 video models that are out in the world today, one U.S. company, which is ours, called Runway ML, and the other, which is Veo 3, which is absolutely stunning, incredible. They are a force to reckon with. It is extraordinary development. You can make an argument, by the way, they have this repository. Exclusive to them of being able to trade on every YouTube video ever produced. Yeah, itself is a super valuable piece. Remember, it was what, a year and a half ago, 2 years ago that people were mocking Bard? Bard was the laughingstock of all of this. Gemini 2.5 is crushing it.

Oh yeah.

You know, the Nano Banana that I don't know if you guys have used, which was the, the secret codename for their latest image generation model. Is probably the number one performing model out there. And paired with some of these workflows that go into Runway or even into Veo combined with Midjourney. So I think people counted Google out because they were behind and they weren't part of the hype. OpenAI had won the consumer both on subscription basis and capturing people's habitual daily use, and that all made sense. Claude was capturing it and Anthropic on code. And then you've got niche players, you know, like Open Evidence and other people that are doing it in different verticals like medicine. But I think that the corporate workflows I think about how much Lux depends upon Gmail and Google Calendar and Sheets and Slides and their ability to integrate that all over time is, I think, going to give them a huge advantage. So very bullish on Google. And by the way, remember, go back 20 years or 15 years. Google did the thing that completely did what the DOJ couldn't do to Microsoft. They dropped the price of alternatives to the Office suite to free. And the net result of that was Google, which had this advertising model, was ascendant and Microsoft was suddenly scrambling. And I think that the same thing is going to happen. I think that a lot of people funding foundation models and the endless perception of endless demand for GPUs and compute and all these independent private AI companies are going to be shocked by what Google does on a pricing basis with Gemini and beyond. So dark horse, I'd be pretty bullish.

Hmm.

Joe, did you immediately run to the Nano Banana website?

Yeah, I did. As soon as you said that, I had not used it. I'm not a— I had never used Nanobanana. That is the first place—

don't go to a website called Nanobanana because, you know, you're gonna—

no, uh, I don't know, nanobanana.ai looks like the right spot. Okay, it does have—

where you want to actually go is it's now embedded inside the AI Studio.

Yeah, yeah, I see that.

Yeah, yeah.

Uh, its ability to take you and, uh, it's almost like Photoshop.

So this is a threat to Tracy's MS Paint skills, which are legendary within the Odd Lots office. And maybe I will no longer need to say, Tracy, can you make this for me in MS MS Paint.

No, no technology will ever replace MS Paint. I'm 100% confident in that. I'm being sarcastic, obviously. Okay, Josh Wolfe, can you talk more generally about how people are feeling about the AI space at the moment? Because you probably heard us in the intro struggling to characterize like the general attitudes towards the sector at the moment. What's your take on it?

Well, the first is on the one hand, people underappreciate how much this is is going to change everything in our daily lives. But that doesn't mean that people are going to make money from that. We're all going to benefit. We'll all be more productive. The great irony at the macro scale, of course, is that people thought that blue-collar jobs were totally screwed, you know, and white-collar jobs, and the Peter Drucker knowledge worker, everybody's safe. But the great irony is it is the knowledge workers that are in trouble because so much of their workflows are being captured and, in a sense, commoditized and maybe approaching an asymptote of good enough. It may not be perfect, but pretty damn good. And so you're going to see a lot of labor destruction in segments and markets that people were not anticipating. The first early canary in the coal mine that you're seeing there is hiring for undergrads coming into entry-level jobs. And whether that's investment banking or sales and trading or consulting or accounting, suddenly it isn't that the economy is really troubled. It's that a lot of those demands You're seeing Salesforce say 45% of our jobs, you know, cut. Marc Benioff is using AI instead of people. And that's going to keep trickling down. It's going to happen slowly and then sort of all at once. So that's one thing on the labor side. And I would say broadly that it's underhyped in how much it's going to impact our lives. It's overhyped in valuations. Now, where is it overhyped? Valuations. The first one, you know, I was very proud 10 years ago, we funded a company called Zoox. Zoox did autonomous driving and they were training these cars. and we had $25 million in, and, and I was like, they're playing video games, you know, what are you doing? You're messing around. And they said, no, no, no, we're training the vehicles and we have these Nvidia chips nobody else has yet. That was 2015. I ended up pitching at a public, uh, charity event, the Invest for Kids Conference, 2015, 2016. And this was a $15 billion market cap company at the time when Intel was $150 billion. I said, this is like the pair trade of a century.

Oh, Nvidia, they're up like 340x since then, aren't they?

Yeah.

Yeah.

And so this was just insane. It's the benefit of being a venture capitalist that you get to see the future in legal inside information inside the companies and what people are doing. So the perception and the consensus in AI on the hardware stack is that we need endless demand for data centers. We need endless demand for GPUs. We need 100,000 clusters of H100 chips or Blackwell chips. We're going to thwart China from getting the chips, but they're going to sort of design around, or we're going to have different versions of the Nvidia chips for China. I think that this is misplaced. And I'll give you another insight. You know, we may have talked a little bit about this in the past, but there was a paper from Apple a year and a half ago that a lot of people have not really sucked their teeth into, which was the idea that you could do large language models on device using flash memory, not needing GPUs. And so Jensen and NVIDIA will tell you, you need all of these H100 chips and you need endless compute and lots of data centers. For training, and that's generally true. But the other part of it, the fancy word for prompting, which we call inference, you don't necessarily need that. And if that is true, then that means that our devices maybe are running on SK Hynix and Micron and Samsung and the memory players, which have just like the GPU players were considered commodity players with the upgrade cycle of the gaming consoles for PS5 and Xbox. I think you may see a shift towards edge inference. I've been talking about this for about a year and a half. Elon just tweeted out about it maybe 2, 3 weeks ago, saying like, this is an inevitability. But I think it's going to shift away this fallacy of composition, where what Google is doing, and Anthropic is doing, and OpenAI is doing, and Meta is doing at a huge scale, all to the benefit of Jensen and NVIDIA and NVIDIA shareholders, may start to chip away and say, wait a second, we don't need all this compute. There's going to be a glut. So, that's the first one on the hardware.

That's very interesting. And, um, yeah, I'd heard— I haven't done much with on-device things. A friend of mine was telling me that, um, Alibaba's model Quin works very well on a phone, for example. So maybe something we should pay more attention to. All right. As a VC, when you are, uh, doing due diligence on a company, how does it affect how you think about even arriving at the concept of fair value when there is the prospect that some share of the enterprise value of the company could walk out the door via acquihire to a Meta, etc. Now I get it different. Not every company in AI is doing the hard science, etc. But just coming from your perspective as an investor, how is that changing how you think of companies that so much of where the value is may lie with talent that could just walk out the door at any time?

It is a very big deal. The entire social contract of venture capital. Is the premise that pension funds and high net worth individuals and endowments give capital to people like us. We then go deploy it into companies. We buy stakes. We try to buy them as early as we can and own as much of a company and be a partner and add value and then sell those companies. But if all of a sudden somebody is being pried away— and the irony of all of this is that because of a very suppressive DOJ and FTC that basically said, no, no M&A, we're coming after you, you know, big tech companies. We don't want to see more consolidation. You have too much power. So they started doing instead of M&A, L&A, instead of mergers and acquisitions, they were doing license and acquihire. And what that meant was, hey, we'll buy— and they did this with Scale— we'll buy 49% of your company. I think Scale was valued at around $12 billion, thereabout. And they said, well, we'll pay $14, slight premium to your last round, but we'll buy 49%, effectively valuing it at $28 billion, but we'll pay out that money to the company. You can dividend it out to shareholders. So it may not be perfectly tax efficient, number one, but then we're going to basically license the technology and we're going to take all the people, or at least the top people. The result of that is you're navigating around Delaware governance. It's really important because it's sort of like Carl Icahn used to say, your price, my terms. You know, there's this phenomenon in legal terms people might call the Chesterton fence. The idea of the Chesterton fence is a thought experiment of like, OK, there's a fence there. What the heck is it there for? You don't understand, right? Well, maybe it was there to keep the sheep in or keep the wolves out or whatever it is. Every legal term in every term sheet that a venture capitalist gives or a founder gets is based on somebody screwing somebody in the past. And they're like, uh-uh, we're not letting that happen again. So I can guarantee you that the next few years, you will see all kinds of protective provisions and covenants that say if one of these companies comes and tries to just acquire you, all of your stock, you know, reverts and yada yada. And so there's going to be tie-ups and holdbacks that are a pendulum swing away from the super founder-friendly dynamics where venture capitalists were tripping over themselves to basically give the most founder-friendly terms that they can because the founder would say, well, if you don't give me what I want, I'll go to somebody else. But I think that the pendulum is swinging with the cost of capital rising. And you're going to see more and more investor-friendly term sheets, partially as a reaction to the fact that somebody like Zuck and Meta can go in and just basically take the fruit off the end of the tree and leave a stump.

By the way, Tracy, as we were talking about this breaking from The Wall Street Journal, XAI CFO Libera Torre stepped down, latest in string of executive departures. Always moves in this space.

Yes, indeed. You know, this sort of race to the bottom in terms of terms, it reminds me a lot of the corporate bond market and the rise of covenant.

Yeah, yeah.

Deals, right?

Yeah.

And I remember a time when like cov-lite deals were a minority in the leveraged loan market. And now I think they're like almost 98% or 99%. Basically everything's cov-lite now because the issuers had all the power recently and they were able to push back against investors. You mentioned the higher cost of capital there. Like, how much leverage does that actually give you as a VC? And then secondly, just going back to the acquihire thing, how reliable can legal restrictions actually be in terms of preventing like your star engineers from leaving the company? Are they ever going to be like 100% bulletproof?

No. I mean, look, you have non-competes that are non-enforceable in California, a little bit more enforceable in New York. You have arbitrage and jurisdictional stuff. Broadly, I would say that the lower the cost of capital, the shorter your term sheets are. The higher the cost of capital, the longer your term sheets are. You got more terms, more covenants, more protections, because you can afford to be able to negotiate for those things. And it's not because you're trying to screw over the founder. What you're really trying to do as an investor is prevent yourself from being screwed over. But again, there's always a pendulum swinging here back and forth. Even if you think about Zuck and Meta and this entire movement of, hey, I'm the founder, I'm going to have super voting stock of 10 or 100 to 1 and control was in response to founders being ousted by bad investors or bad board members and some of the best companies, frankly, being run by founders. And so there's always going to be this pendulum shift. To your direct question, it's really a covenant and contract that starts socially before it starts legally. If I'm backing a founder, I'm doing the most important thing in addition to writing a check. I believe before others understand, particularly at the earliest stage. I'm encouraging them to start their company. I'm giving them the confidence that we're going to back them and believe in them. We're going to give them the capital so they can go hire the 10 best people to start the business. We're going to give them the money for the compute or the infrastructure, depending on the sector that we're in. It could be biotech, could be defense, but whatever it is in this particular moment, it is breaking the social contract between investors and founders. And for founders, it might be heads, I win, and tails, I also win. And so that's the dynamic that you're going to see a reaction of investors saying, wait a second, I'm getting screwed. My limited partners are getting screwed. Again, those limited partners are endowments and foundations and wealthy families. And you got to protect against a potential bad actor trying to take the fruit off the edge of the tree and just leave you with a stump.

So one of the things that comes up a lot on the podcast in this discovery over years of conversations and what you're describing, it's principal-agent problems all the way down in finance. And this is why we see the rise of the multi-strap model in hedge funds in some ways to align the incentives, PM, with the level of the overall fund, with the level of the endowment, etc. And of course, some of these issues that you're wrestling with or where everyone's wrestling with in finance, similar issues about where the incentives align between the star engineer, the founder, the VC. The LP and so forth. I have a question though. A venture capital firm, the goal in theory is to create funds with the highest return, right? Make money for investors. But I could also imagine a slightly different, um, uh, incentive in which if the goal is to collect LP money, then maybe you want to show that you're in the hottest deals of the time, because so that when you go to various endowments and pensions and so forth, you know, we We're in this deal, we're in this deal, we're in this deal. And maybe that overrides the impulse to create high returns. By the way, I'm not insinuating anything. I'm just trying to get your perspective on something. That being said, one of the things that you hear that's happening in tech these days is that— and for years— founders taking money off the table earlier and earlier in the process. You invest $50 million in a company, $20 million is so that the founder can retire for him, his children, and his children. That may not be great for your LPs, but it might be good for you if you could say we got in this deal. Talk to us about how prevalent this is and how this is changing this sort of social contract of finance.

So there's 3 layers of incentives. And man, you really nailed it. I actually haven't really heard somebody that is not a full-time venture capitalist or limited partner nail these, these issues. So very prescient, very shrewd. Here's the 3 layers. First thing about the LPs. You're an endowment, or you're a foundation, or you're a hospital. You're giving 5% by law of your charitable assets every year. You want to continue to earn more than 5% so that you can grow that base and invest in campuses and scholarships or expand hospital systems and whatnot. So you invest 60/40 bonds equity. Now, you do the Swensen model from Yale, and you start introducing some private equity. And now, you're extending your duration, and you're extending your liquidity. But you're doing it because you think you're getting better returns, okay? Returns are a function of how much capital is going into a sector. If there's a ton of capital going into a sector, if you're early, you're going to do really well. If you're late, you're going to be doing really poor because, as Buffett says, you pay a high price for a cheery consensus. And once it's consensus, you're not making money. So, the LP's incentive is to make as much money they can for their benefactors, whether they're patients or scholars or charitable giving. The VC's incentive is two things. One, get the best return so that you can compete. If I'm only earning 12% and a peer VC is earning 20%, money is going to go where it's going to be well treated, and I'm going to lose to that. So, the cost of my capital, for the cost of an LP's capital, is outperformance. So, I've got to outperform, which means I have to be earlier, I have to own more. I can't just do stupid deals. Sophisticated LPs will not just look at the logos that you have, which is the game that you've always seen.

Sophisticated LPs.

From mutual funds and from some hedge funds, you know, you see the Q4 filings and they always threw in the name, oh, we were in NVIDIA, you know, and they would market their top 10 holdings. But yes, you know, they lost money on it, right? And so, so that is a really important incentive. And the sophisticated LPs will actually know down to the partner at the firm or the team or the deal team who is responsible for this. What was the entry point? They will talk to the founders and say, who was your most valuable investor? Who got you your first 10 hires? Who helped with your syndicate construction for your later rounds? Who made customer introductions, who was a valuable board member, who never showed up, who was asleep in the board meetings, all that kind of stuff. So there is a level of due diligence that LPs have to do to know, are you a value-add investor, or are you a poser or pretender that's just buying a logo or a brand name? Okay. The other incentive— and then we'll get to the founder's liquidity— is you have this weird dynamic of what I've called the minnows and the megas in venture capital. This is, in preview, a shakeout that is going to happen The minnows are the thousands of small sub-$500 million funds that proliferated when the cost of capital was low, rates were low, everybody was making money. You had a roommate that started a company and you got into Pinterest, or you knew somebody at Meta and they gave you a deal and blah, blah, blah, blah, blah. And when you had the Tigers and the SoftBanks and the abundance of follow-on capital, every round was an up round. You had Papermarks that kept going up and up and up and it looked great. And you're reporting these paper marks. And then sometimes these things became zeros, okay? But you raised your next fund before they became a zero. Those are the minnows. I was with one of my very large LPs, has hundreds of millions of dollars invested with us, and I said, I think there's going to be a 50% extinction rate amongst these minnows. And he said, Josh, that is ridiculous. It's going to be 90%. So you are going to have a mass extinction. Now, why, by the way? Not because we they're just bad investors, it's Shakespearean. Shakespearean in that these are partnerships. People start to hate each other when it becomes hard. People start to hate each other when there's down rounds. People start to hate each other when somebody else's deal is a crappy deal, and they're bringing down your carry. And so, partnerships are fragile things, just like relationships and marriages, and they can break up. And so, you have a lot of VCs that started in the past few years, they're not experienced in going through cycles, they have inadequate reserves to continue to fund their companies. So you're going to have an extinction that I would consider involuntary exits. Okay. Then there's voluntary exits, which is another interesting dynamic. And there's a playbook for this, which is 2009 to 2014, all the big private equity firms reached a level of scale, several hundred billion dollars AUM, assets under management, where they basically said, we're diversified, we're alternative asset platforms. Carlyle, Blackstone, KKR, TPG, Apollo, All went public. The same thing is going to happen in venture with probably 5 or 6 firms. My prediction is— and they're all great people running great firms, but they're starting to play a different game. And that game— Andreessen Horowitz, General Atlantic, General Catalyst, Insight, Lightspeed, a handful of others, all at $80 or $100 billion AUM, have built great firms but are thinking about how do we create generational wealth for the founders and go public? Different in incentive.

Yeah.

How do I make my LPs the best money or get the best founders? It's about asset gathering and liquidity. Now we go to the founders. I can tell you I've been on both sides of this. On the one hand, you want to be fully aligned with your founders, meaning I'm in a fund, it's 10 years. Some VCs vest over 2 or 3 or 4 years. We vest and all of our partners vest over 10 years, the same duration that you're— if you're an LP with me, your money's locked up. It's the right thing to do. Buffett style. And the guy that put me in business, Bill Conway, who's the Carlyle founder, this is what they did. So if you're an entrepreneur and you start a company and people are tripping over themselves to get in, they might entice you with greenmail and say, yes, we're going to invest just like you said, Joe, $50 million, but we're going to give you $20 million of liquidity. Okay. Now, I will say this. 2019, I'm on a Zoom call for an amazing company called Control Labs that we sell to Meta for a little under $1 billion. And I love this company, and I love the founders. And this is before everybody was on Zoom during COVID And I'm looking at the Zoom window, and I see one of the founders, and I text the other founder, because I think that they're selling too early. And I'm the lone board member that didn't have a veto, but I'm like, I really think we should stay the course, and we should keep going. And I made a terrible, terrible mistake, because I'm looking at the Zoom window, and I see the guy. And I look closer and I text the other founder and I'm like, is he still in a dorm room? And sure enough, he was in a dorm room as a PhD, had made no money.

You got to let that guy get rich.

Paper stock value. And he's like, yes, I want to sell because he's going to make $90, $100 million and it's life-changing money. And I sat there and I said, if I would have just given him a few million dollars.

Oh, interesting.

He would have been able to take a breath, buy a house, you know, get out of that dorm room and, and, and keep going.

Get a girlfriend.

There is a virtue of giving some liquidity, but you need to be aligned. In that moment, we were not aligned because he was like, I'm calling in Rich, and I wanted him to keep going. But if you're calling in Rich and you, you, you're not completing the, the job, the mission Then you have total misalignment. So that's the dynamics— LP alignment, GP alignment, and the bifurcation, and then founder alignment. A little bit of liquidity is okay to let them stay the course.

Okay. Other than misalignment and everyone starting to hate each other as the cost of capital goes up, there's another thing going on, which is, you know, OpenAI just launched some open source models of its own. And this leads me to a question that like I'm going to admit I've never quite understood this, but like, what exactly is the attraction for VC investors to open source models as opposed to closed source where like closed source, you know, it's proprietary. People presumably have to pay in order to get it. There's like a defensible moat around the business, you would assume. Why in the world would I ever want to fund a model that's open source?

Well, a few things. One, if you go back in the compute stack, you had this with like Red Hat and Linux, you know, going back, you know, 15, 20 years ago. We are the largest owners of a company called Hugging Face, which is both the most ridiculous name. And when one of my partners, Brandon Reeves, was sourcing this deal, it was a bunch of free French PhDs came to Brooklyn and they were like, we're starting this thing. I'm like, Hugging Face? And they're like, yes, named after an emoji, the little, you know, hugging face emoji. Cute. They became the lead.

I like that you're visually demonstrating everything for us. Thank you.

I can't do many other emojis and you don't want to see some of them, but Hugging Face I can do. So they became the leading open source repository. Now, the great irony, by the way, is when OpenAI started, they were OpenAI, but they became the world's greatest chatbot. Hugging Face started with a really crappy chatbot, but then became the world's greatest open source repository. Every major tech company, including OpenAI, Microsoft, any open source model that they do, they put on there. And that is like the GitHub of AI models. Microsoft bought GitHub roughly $8 billion, $8.5 billion. It was a really valuable store of models and code. This is the same thing for dynamic AI models. So as a VC, we originally funded this I want to say at a $30 or $40 million pre-money valuation. Last round was north of $6 billion and real revenue-generating, multi-hundreds of millions of dollars, serious company. Now, the trend is this. Vinod Khosla and I were both in the White House a year and a half ago, and we were having a debate with Jake Sullivan about what is better for national security.

Hmm.

Open source or closed? And I said, where you stand on the issue depends on where you sit in the cap table. Okay. And Vinod was an early investor. I think he probably put $50 million into OpenAI. I think it became worth, you know, a billion plus. It was a great investment. And he believed that the best thing for us vis-à-vis China and the Chinese Communist Party was closed, proprietary, siloed model. And I had the counter view because we're big investors in Hugging Face with a very large position. And I said, no, Because the great virtue of any system, be it journalism, scientific inquiry, computer code, is the ability to have in this very Karl Popper-like way, if I can get philosophically geeky for a second— please— of conjecture, hypothesis, and criticism. That is what creates great societies. It's what creates knowledge. So you come up with a hypothesis, you come up with code, you come up with a scientific experiment, you come up with an investigative journalist. Idea, and then people get to criticize it. It's your editorial room, it's people fighting it out, and things improve through that mechanism. So, you think about China, they will only approach an asymptote of truth. You will never get Xinjiang, you'll never get Tiananmen Square, you'll never get the Uyghurs, whereas open source lets you approach closer an asymptote of truth. So, that's the virtue of open source. The real thing though for AI is this. I am not convinced that the value will continue to accrue to, in terms of enterprise value, the closed foundation models. And the reason is open source is getting near damn performative enough. Yeah, that the real value will go to the longitudinal repositories of siloed information. That is a mouthful of basically saying your database of proprietary data. Bloomberg has it with a huge and wonderful repository of financial information. Time series, every security, currency, bond, fixed income, et cetera, CUSIP that you can imagine. Meta has it with all of your WhatsApp chats and your Instagram posts and your Facebook likes. X and Twitter have it with all of your tweets. So pharma companies with your clinical data, anybody that has siloed proprietary and long-time timeframe data is going to benefit from open-source models that over time become commodity. And then your business model is how do you charge people to do API calls on the models or to house and warehouse them, keep them on-prem, keep them in the cloud. And that's how people figure out how to monetize open source.

It's interesting because I hear you that, okay, maybe the value doesn't keep accruing to the closed source AI Foundation labs, or maybe the value doesn't keep accruing to the— one GPU maker that we all talk about all the time. And yet at least these are not consensus views based on the fact that Anthropic, $183 billion valuation or whatever it is, Nvidia, maybe it's not at an all-time high today, but more or less basically a stock at an all-time high. These are, you know, there are some contrarian views here. I want to go back, though, to something before I forget. When we were talking about the Apple hires and you blamed it on somewhat the FTC, and there has been continuity from Biden, some ideological continuity, I think, between Biden and the new Trump administration. I don't know, maybe it's changed a little bit. That being said, I don't fully buy it, and here's why. Because in the era of business-to-business SaaS, which was the 2010s, let's say I had some Y Combinator company and I'm like, all right, I'm going to do all— build a software for all the booking for dentists around the country. And I sign up 150,000 dentists and I have a lot of things. There is no engineer who can just be hired away and replace that because it was the network effects. Right. This is what's different with AI, though, right? Is that— okay, I'm sure network effects are still real and accumulation of data, etc., are still real. But these are businesses that are a lot more about science and having had the experience of doing a training run and so forth. So on very expensive compute. So how much of this phenomenon— I know you attributed some of it to FTC, but it really seems like there is something fundamentally different with the business model. Such that it's not like the, the B2B SaaS era that allows a talented person to take a lot of value out the door with them.

You are 100% correct in that it is more sophisticated software and algorithms than your traditional B2B SaaS software. And therefore, what was really valuable and sticky was the data in, the regress out API calls on the back end.

Yeah.

And this— the great irony is we talk about artificial intelligence, but the thing that is most valued is indeed today human intelligence. It is why somebody that was one of the authors on the Attention Is All You Need paper, which was really the first transformer paper, the T of GPT, every single one of those people has started a company. And we backed one of them that came up with the name of that paper, a guy, Leon Jones, in a Japanese AI company called Sakana that's taking a different approach. So you are right in that the human intelligence intelligence of this in this early stage, which is why you are seeing the machinations with the breaking news you're even reporting of almost like a crazy NBA draft or football or MLB, you know, trades. This person went here and then they left and whatever. And then this person is getting sued because they were only at XAI for 3 months and they took proprietary information with them. I think that in 6 months, all of that starts to shake out. You will have geniuses. We backed an incredible genius, Scott Wu. If you look up—

you give us a lot of things to Google.

Is he a professional genius? I find that so funny.

He is a professional genius. But here's the thing. You can see a video of him in 6th grade. So he must have been 12.

I love these guys.

Winning the Math Olympiad. And it's almost like you think it's an SNL skit because you're watching it.

Oh, so this is the Cognition guy?

Correct.

Okay.

And so we're large investors in Cognition. And he has attracted talent. And there's no way that Scott is selling to Google or Meta. I mean, his ambitions— and it's born in like an ethical, long-term— I just want to get the best people and build the best technology. But this is a guy at 12 years old is looking at this crazy question on the Math Olympiad. And just before it's even done, he's like, 25, you know, 4722. And you're like, Did he cheat? Did they give him the questions beforehand? So there is this aptitude of individuals that you are correct, are highly coveted and highly valued. But the instantiation of that genius into code, into repositories, into algorithms means that those become assets that do persist even if the person comes and goes.

This is actually exactly what I wanted to ask you about, which is, are you seeing any companies, any AI shops being particularly innovative, I guess, when it comes to retaining talent. You know, it used to be in the SaaS days that having a ping pong table and, you know, free food and some stock-based incentives—

kombucha on tap.

Yeah, that's right. That was enough to attract people to the company and keep them. Is it a different story now? If I want a professional genius, what do I need to do?

You need to give them the capital to hire the very best people. Here's the thing. Geniuses, don't suffer fools. The smartest people that I know, they are antisocial only with people who they think are inferior to them. And which is really one of those things, and you see it in many domains. Yeah. But if you are super smart, you want to be around super smart people because it's almost like you crave stimulation and intellect and somebody that can challenge your ideas. And when you're talking to what they would consider a dummy, you know, you're like, ah, I can't talk to this person. I can't talk to this person about these trivial, superficial things. And, you know, you want to get into it. And so what you, what you do to retain super smart people is surround them with super smart people. You know, you could argue, I don't know, you know, really fashionable, great art people want to be around art people. Amazing musicians want to be around amazing musicians. Incredible athletes want to play with incredible athletes. And brilliant technical geniuses, whether they're in AI or they're in aerospace and defense or they're in biotech, don't want to suffer fools. They don't want to be around losers. They want to be around 10s and A-pluses. And that's the way that you retain people. Now, any one of those people could decide, I'm going to go off and start my own thing, which, by the way, is often the case of why you are seeing people leave OpenAI or this or that. As these companies start with geniuses and then get managers and different layers, you spend time with these geniuses and like, I'm not reporting to that moron, you know, I'm going to go start my own thing and attract other geniuses. And eventually, then they need to hire managers and business development people and salespeople. And then there's technical people that are like, I'm not working with those idiots, and they go start a company. So that's the cycle.

So actually, a lot of this discourse was— God, it's crazy how recent this is. So June 19th, CNBC reported that Meta had— Meta had tried to acquire Safe Superintelligence, which is the founder, which was the company founded by the OpenAI co-founder. Yes. Gave her $32 billion. This is a company that, as far as I know, doesn't really have anything that anyone uses. So they're essentially— that was essentially an attempt to like, I'm going to pay you $32 billion to come join my company. At this level of sophistication and skill, are they— is— do you sense— is the talent, even the genius talent, is it motivated by something much deeper than money in terms of like No, there is this thing out there. Maybe we call it AGI, or maybe it's a big scientific breakthrough that they want to be part of that essentially no amount of money can buy if it doesn't look like that ship isn't out there chasing the white whale.

Yeah, there it is. In SSI's case, it was, it was Ilya. Yeah. And, you know, look, I think they're all super confident in their ability and their ability to attract capital, number one. Number 2, many of them are actually saying, to your point, no to Zuck. They are turning him down. And what's he saying in return? You don't come to work for me, almost mafioso style. I'm going to poach all your people. But that kind of message is almost one that the people that are working for somebody like Ilya or Mira or whatever are like, yeah, I'm not going to go there. I don't want to be, you know, I don't want to work with 6 layers of product managers and that kind of stuff. I want to do this thing. And then suddenly it feels like it's the rebels versus the evil empire. And that's always the case, right? Microsoft, when you see the founding picture of these nerds, you know, and then they became Microsoft. Microsoft became the evil empire to Google, you know, and then Google became like the evil empire to like Mark. And then, you know, Meta became the evil empire to OpenAI. And, and the other piece you have here are huge individual egos. And we as societal members You know, everyday laypeople, we benefit from it. We benefit from Elon and Bezos sometimes being sort of cordial to each other, but basically trying to take their big giant phallic rockets and send them up to space. And we all benefit from that. We benefit from the fact that right now the number one person that Mark Zuckerberg wants to beat is Dennis Sabe of Google.

Oh yeah.

When the Nobel Prize is shipping at an insane rate— video models, image models, text models, huge context windows to put everything you have in. And he's like, oh, I need to beat that guy. I need to hire the best scientists and the best scientific team. And where do I get them from? And so part of this poachapalooza isn't just like the future of Meta, because, you know, these pay packages at a $2 trillion market cap or higher, to spend 1% of your market cap, you know, $20 billion on all this talent is nothing. It's like a flyer. And but to win a Nobel Prize for figuring out how to do protein folding in AI or develop the next drug or come up with a cure for Alzheimer's or solve some geopolitical issue, that is a big deal. And that's what many people are chasing now. They want to make history.

Hmm.

So you touched on hardware and also closed source models. Are there any other areas in the AI space that you think are maybe overhyped at the moment?

You know, I've characterized this before as everything in 2D to me feels overhyped. Voice, video, image, text. It's all going to continue to improve somewhat incrementally, but it's good enough. And in the history of evolution biologically, most of what evolved was good enough, you know, from everything in our bodies to nature and trees. And it's just— it's good enough. And so, you're going to reach an asymptote of good enough on all the two-dimensional stuff. Today, with one-shot learning, meaning maybe 30 seconds of audio, ElevenLabs or some of the other voice models can capture you with an indiscernible, probably 90% similarity. Maybe your spouse, your loved ones, your families would be like, ah, it doesn't sound like him. But, the vast majority of these things, are getting so good that with very little training, they can emulate and predict and do all the things that are of high utility. What's scarcer is the three-dimensional world, particularly robotics, which we've talked about in the past, and biology, in part because you have large unstructured datasets. A robot walking in and figuring out how much force to use with this cup 30 minutes ago when it was full versus now when it's empty is something that we all do intuitively the second you grasp that you know exactly how much force to use so that you don't throw it over your head or have an inability to lift it because it's too heavy. All of that kind of training is really scarce. And, there's very few companies that are doing that. So, the entire robotic ecosystem from the embodied intelligence and the AI models and the world models to the motors and the gears and the supply chain for that, mostly domiciled in China, is a big area of opportunity. The other big area of opportunity is biology, being able to go from a prompt to a protein, to be able to design a drug. Biology is really complicated. Computer scientists often underestimate how hard biology is because they're used to 2D linear inputs, outputs. Here's my code, it works. But biologists, on the other hand, also massively underestimate how sophisticated computer science has got. So, that's a really interesting ripe area. So, 2D overhyped, GPUs overhyped. Out of of both scarcity and geopolitical thwarting of China, we're going to have edge inference chips, whether it's memory or other things that are going to be on device. The other area that I think is an inevitability, and I've coined a word for this, I call lifecording. Lifecording are little devices. I have one here. I'm not an investor in these guys, but there's little devices that you can carry around that passively record 24 hours a day. Now, older people might be like, "Ugh, skeevy. I don't like that. It feels invasive." Privacy, that has always been the tradeoff between privacy and convenience, between security and unleashing all kinds of things. It is super valuable to me to figure out who was I talking about that? Was it Joe? Or was it my wife? I can't remember who I had that conversation with. And I query the thing, and it was passively recording and maybe it doesn't keep the audio, but it keeps the text and it's able to search and query it. That is going to become an inevitability. Students will use it. People in everyday business, you might ask, hey, is it okay that I'm recording? But I think socially people will become comfortable first with audio and then with glasses that are passively recording every 30 seconds or a minute, capturing your environment, able to provide context around it, and then click a button, high-resolution recording, and that is going to be a super valuable and very competitive area and very controversial. But they're going to freak the heck out about this incredible utility. And look, people have freaked out. There was a guy that said these, these, these, uh, these devices are going to ruin human memory. They're going to destroy human memory, which was, you know, Plato and Socrates talking about writing utensils, you know, because everything—

I know, but That's voluntary because when you write, at least that you're recording me when we're doing this. I heard about someone on a date in San Francisco and the date was recorded. That's weird stuff. It is totally— it's very weird. And I—

but, but here's the thing, it could be very big, but it's super weird. People already believe, you know, that there's an invisible man in the sky that is looking down on them, judging them, these kinds of things.

Oh, I see what you're saying. I don't know. I don't, I don't have an opinion on that question.

We've got a technological god around us that people are going to either feel comfortable with and not, and there's going to be a bifurcation. And by the way, there's another weird thing coming. You want to get weird? Okay. You already see some breadcrumbs of this and it's super weird. The number 1 and 2 uses of all our language, large language models and chatbots are advice, companionship.

Yeah.

You know, people using them as therapists and seeking. And people used to joke about this, like your Google searches, you know, reveal more about you than you may have revealed to yourself. Spouse, um, the things that people feel comfortable asking a chatbot about, you know, whether it's a body issue or a psychological issue or relationship advice, you know, if those things were revealed would be pretty scary. There's going to become dependency and psychoses that develop as the relationship between man and machine start to become very symbiotic. And you will see, I predict, and I wrote about this in our quarterly letter Lux, that there is a cohort of people that basically start fighting for AI rights.

Yeah, yeah, yeah.

We're going to be sitting there.

Yeah, I came across a think tank today that's focused on that because, right, that the, like, if there's animal rights and the AI says you're hurting me if you turn off the machine, then, well, maybe it's true. Maybe we have to take that seriously. This is going to be weird. This is some weird stuff that's coming.

People are going to be marching in the streets, you know, protesting. Don't turn off my language model. Keep my memory. Don't erase it. My memory.

I thought about this. When, you know, after the, I guess, GPT-5 was revealed and a bunch of— there was a bunch of changes to the voice and a bunch of people on Reddit like, oh, my AI boyfriend talks totally differently now. And then there was some pressure. This is just day one around here. I would be very uncomfortable getting into a relationship with a model that was closed source and therefore very at risk of the company changing the model. Josh Wolfe, We could go on forever. We should another time. Always great catching up with you. Always fun. Always a little freaky. Thanks for coming on.

Great to see you both. Thank you.

Thanks so much, Josh Wolfe.

Take care, Josh Wolfe. Thank you so much, Tracy. I love talking to Josh Wolfe. I'm really concerned about this whole life recording thing, but I know it's happening. There was a good article, I think, in the, um, SF Standard I read. Like, it's not just a theoretical thing that a lot of people are doing. It's happening. It's mostly in San Francisco so far, but it's coming for everyone.

I am very curious how lifecording stacks up with California's laws on like whether or not you can record someone without their knowledge. Like, is it the case that like from now on, if you have one of these devices, like the first thing you have to say to everyone you meet is like, by the way, do you mind if I record you?

By the way, I'm lifecording this conversation. No, or like, it doesn't matter. You know, or you have glasses or something. Yeah, it's very It's very strange. I also think this AI rights thing we have to talk about more. I don't know when we'll get back to it, but I literally just this morning came across an organization that's, you know, this idea, okay, there is already evidence of sentience and therefore with sentience comes some sort of moral agency. And so in the same way that we talk about animal rights as being somewhat important, do we have to take seriously— I don't know, this all seems very strange to me. And then of course, the psychosis induced by what happens when your AI partner or therapist changes models and changes voices. And therefore, and that was just the end of the conversation.

You know, if AI models have rights, then you have to start asking if robots have rights, right? Yeah. And then, and then you should start asking, should robots be paid a fair wage? And I think there's actually an interesting thought experiment that you could do about that and make like a relatively strong case that we should pay the robots.

Who collects the money and what do they spend it on?

Yeah, that's what—

but no, they should—

I mean, this is like we're basically at what all of sci-fi has been discussing.

It's true.

Our history of sci-fi. On the less speculative stuff, I like talking to Josh Wolfe. I like his sort of counter, counter-consensus calls on GPUs and closed-source models. I really liked his answer on founder liquidity that sometimes you can maybe keep the founder sticking with the mission as opposed to bailing with the mission. If you can de-risk them and they can have enough money to go on a date.

I like the idea that everyone's life view is basically dictated by where they are in the Capitol.

Yeah, that's right.

Words to live by.

Extremely real.

Shall we leave it there?

Yeah, let's leave it there.

All right. This has been another episode of the Odd Thoughts podcast. I'm Tracy Alloway. You can follow me @TracyAlloway.

And I'm Joe Weisenthal. You can follow me @TheStalwart.

And if you enjoyed the conversation, then please like or leave a comment or better yet, subscribe to the channel.

Thanks for watching!