20VC：Jensen Huang 宣布 AGI 已到来｜GPT Astra 与 Fable 5.1 加速模型竞赛｜Tesla 推出 Cybercabs｜Index 退出 Town，Anthropic 退出 Descartes 收购

There's going to be no financial math you can use to buy the stock. When someone goes risk-on, everyone goes risk-on. I would imagine, as we speak, there are 20 engineers locked in a room somewhere in Palo Alto, literally with guards at the door saying, “Nobody eats and nobody leaves until you ship Instinct clone.”

12 billion for a bootstrap company. Now it's just a Series C round.

There was a free 10× in the public market in 3 years on Robinhood. In this market, the people who are making the money are the people who are just running fastest and evolving quickest.

Have you guys tried Instinct yet?

No. I can't sign up, right? I hate to sound like I'm behind the times, but when it's a closed beta, I just don't bother.

You can get an invite.

Or you can use it, right?

We are wasting content here, people.

I'm not a fan of Grok or Manus. There will be many winners. There is a genre of applications, and a lot of actual AI GTM applications are included in Manus. Grok's the interesting one because SpaceX is public, and one of the reasons they work is because they can break the rules. In the old days of venture, that was troubling.

We are wasting content.

1. Agents Break The Rules

I mean, in the old days of venture, you wouldn't do things like gambling or other types of things that had risk, or edgy things, but you can break the rules, right? Even Grokbot, which is like a mini Instinct, spins up an instant VM for every single person, and they get their own browser in it.

Grok uses Google, which is not allowed and is prohibited by the terms of service, to Google things and then give you answers. It's great, and any startup that you might invest in at an early stage would do that and no one would know, right? Google isn't going to care, but you're breaking rules.

OpenCloud broke every rule on the planet. Not just laws, but every rule about what you could do. Instinct and Grokbot, which are like OpenCloud, are much better, but I think some of the reasons they work include breaking Resy's terms of service, right? Spinning up browsers that you're not supposed to use, using agents to go into browsers—Perplexity and Amazon fought over this.

It's not that I'm not excited; it's just that there are many cases of things in AI that are exciting because you break the rules. You can scrape LinkedIn in ways you can't really scrape, right? You can do outbound phone calls that are actually prohibited and illegal in parts of the US. But as a startup, you're not going to jail. Does that count? I don't know.

Well, hang on. You have examples both ways.

Yeah.

Because you're right, Jason. LinkedIn scraping—in the end, no business at scale ever gets built on that, and no business at scale ever gets built on breaking Google's terms of service. On the other hand, Uber blustered its way through, broke the laws, and eventually it was so popular that politicians folded.

Yeah, for sure.

The truth is, I've learned that there's no one answer here, right? For listeners, what happened is that the Manus agent's classic use case is getting reservations at hard-to-get restaurants. A whole bunch of people started using it over the weekend, and they're pounding on the Resy reservation API until the thing breaks. That's the kind of thing that's going to happen.

The truth is, if Instinct—if these do become ubiquitous—then the booking reservation systems are just going to have to find a way to deal with it. They're going to have to have a separate API. They're going to have to do something about the number of hits. But the truth is, if there's a whole bunch of people trying to book restaurants and you're in the restaurant booking business, you're going to find a way to make it work. There's definitely some rearchitecture to go on here.

The bad side is that you're breaking rules, and Rory's right: there's a long history of begging for forgiveness, breaking rules, and then, once you get big, coming out on the other side of it and doing it. The flip side is, I've talked with multiple public-company CEOs who say they're hamstrung. Their hands are tied behind their backs. They can't compete with startups, not because they can't do it, but because their legal teams won't let them, literally.

I remember back in the day when we were acquired by Adobe, 5 to 8 years before anyone did it, we allowed real-time document collaboration and redlining online. No one built this for 8 years. It was jaw-droppingly good, what my CTO built.

Unfortunately, the only way it worked was if you ran Word in a container, in a VM, which violated Microsoft's terms of use. So the day after the deal closed, my favorite second-generation feature, which would have given us a 5-year head start back in the day, got ripped out by Adobe the next night.

Nice.

So literally, I was with some public-company CEOs saying, “We just can't compete because we can't do things that violate terms of service.” Elon can, but anyway, I don't mean—

Yeah, it's funny.

Everyone recognizes that the universe of people who can break the rules is clearly all private-company CEOs and the CEO of one of the 8 or 9 largest market-cap companies on the planet, because Elon just doesn't care. There might be a lesson in that somewhere for the rest of us.

Yeah, he doesn't care, seemingly.

Maybe not caring is the secret sauce. Yeah, right, people.

You know what's so interesting for me is actually how important the form factor is and how important being where people already are is.

And what I mean by that is there are—

WhatsApp.

A ton of people who I know have picked it up and love it and engage with it in a way that they wouldn't with any other AI tool other than ChatGPT because it's in WhatsApp.

Yes. It is one of these amazing things: figuring out how to elegantly make agents work in WhatsApp and text, right? It is something that could have been done 6 or 9 months ago and was done to a limited extent.

The only thing I'll say—it's funny, I invested years ago in a company called Gorgias, which is a little over $100 million.

Yeah.

It used to be e-commerce support. Now it's an AI CX, right? They launched their agent in WhatsApp that anybody can use over text, and it's already in the double digits as a percentage of their usage after a couple of weeks.

Now, it's bounded, right? It's really just for your orders, what's happening with your shipping and your product, and afterward. But it shows how quickly an innovation will just be copied. It's a paradigm shift. If Gorgias can clone it in a couple of weeks, it's not a bad thing. It's just the pace of cloning and copying innovation—it's hard to keep up, right?

I don't know what all the Instinct clones will look like by the end of the year, but this is just the world we live in. So I might do it at $2 billion. $2 billion would be my ceiling.

The last round was $2.5 billion, Jason, so it's already exceeded—

I know, I know. That's why we're going to have to pass on the round, because at some point—

So that probably—

That was a joke, but—

In our metaphorical IC that we did last week, which was very popular and went very viral on Twitter—well done for Linear and Clay—you would not be recommending a $100 million check from the growth fund?

2. The Manus Investment Bet

Into Manus? I wouldn't do it because I think even if Gorgias has its own Instinct just for e-commerce, there will be 100 of them, right? Meta will have them, and 100 startups, and there'll be 20 in the next batch of YC. I'm just not smart enough to bet on that one pre-revenue. It's not my vibe, right?

And I will regret it because I didn't get—I didn't look at the deal, but I remember plenty of others, like Loom early. I'm like, "I don't, I don't—you don't have any revenue. I don't know anyone's..." It was great, but everyone's going to make their own Loom, and I was wrong, so.

But you have to have the stomach to write 10 or 20 of these consumer-y checks at $2.5 billion, right? So what fund size? And it can't just be the only one in your fund. You have to do, like, 10 or 20 of these so that the good one pays off, right? I'm not smart enough.

Yes. The interesting thing is, you're right, Jason. It's kind of a portfolio and a worldview bet because you have this company that's exploding in interest, clearly didn't take a huge amount of time to build, but it's got this early lead and not a ton of monetization.

Your choices as an investor are: do you put money in this at $4 billion or $5 billion, or do you say, "Oh, it's easy to clone and there are 10 more like it," and you do one of the others at $50 million pre-money in the hope that they get acquired by Meta instead, right?

And the hard thing is, in these investments, just like Google earlier, there's going to be no financial math you can use to buy the stock. You're just basically saying it's a huge category. In every consumer investment, the trick is to establish the momentum as early as possible and establish the monetization later.

It has been proven that if you get enough traction, the monetization does follow, especially for something like this.

It just reminds me of Lovable in the way that everyone was talking about the commoditization of that space, and it's really quite a light-wrapper product. Every day I'm seeing Noah Shin, the founder of Instinct, come out with, "Oh, we're now doing location sharing. Oh, we've now partnered with 1Password. Oh, we're now doing this."

And actually, the cadence of shipping combined with Index, Benchmark, and having probably one of the best brands in the space—

Agreed.

That, I agree. I said that last week—that it will solve these problems and it will—

No, I agree.

It will become a much richer app. It will figure out guardrails, it will figure out the hard points, and the other folks will fall behind because crappy Lovable products are worthless today.

Because first is first.

For sure, that's the bet.

My take is actually that's why you buy Meta today, because you've got the clearest, most unwavering PMF for this product.

Yes.

And Zuck is the one who owns the core distribution channel, and Zuck has been working on this product. If anyone has a proven track record of copying extremely well—

Yeah, no, I would imagine—

—and integrating—

I would imagine, as we speak, there are 20 engineers locked in a room somewhere in Palo Alto, literally with guards on the door, saying, "Nobody eats and nobody leaves until you ship an Instinct clone." Absolutely. No, I mean it.

Well, you know who would have been great at it? The Instinct team, because they built a version of this, right? One of the things that Manus did that was disruptive—Manus is now an independent company—was that it sort of broke the rules for what agents could do, but not at the crazy level.

It did a little bit of OpenClaw that everyone else wasn't doing, and Manus was disruptive. Their agents ran longer, and they could go further than other products we were using, and that's what made it special.

Anything you wanted, Manus could kind of do before other folks could do it. If Grok Bot was built in 5 weeks by the Cursor team or whatever, I think the Manus team could have done it in between 4 and 6, but they're gone.

3. AGI Arrives In The Headlines

Okay, boys, it was a big week of news. We're going to resume regular programming. Jensen Huang declared that AGI has arrived.

I remember when we were actually—this was many, many shows ago—and we were discussing what AGI is and what the definition is. I think, Rory, you said that AGI will be declared when Satya and Sam agree that AGI is here.

Jensen declaring that AGI has arrived, crediting OpenAI's new GPT Astra—obviously OpenAI's latest new model—which he says was trained on 100,000-plus NVIDIA chips, with 400,000 more coming. How do we think about this news?

This is a bullshit term. The only thing that mattered for the last 2 years is LLMs do code, and code is a half-trillion-dollar industry. Focus, people. Anything that can be reduced to code will be done by it.

Rather than trying to twist yourself in a pretzel about whether it can do everything, just focus on the fact that it can do this thing amazingly well, and this thing has massive economic value. Stop thinking and go ship something in code. To me, that's been the big aha. So here's—

I think AGI, at the end of the day, maybe it ends up—if you think about some of these non-GAAP definitions—would you rather have an AI do it or a human do it? If it's better than 50%, 90%, or 99% of humans, you'd rather have an AI do it.

You can go category by category. It doesn't have to be the whole category, like coding. It could be collaborating, right? I forget who this week was saying it—they thought—what AI pundit or leader was saying he thought AI would replace all of radiology.

Instead, it just replaced 95% of radiology, right? And the radiologists concentrate on the 5%, right? That's maybe AGI, too.

4. Legal AI Finds Its Market

I have to say, I was sitting next to my girlfriend on the sofa the other day. She was working and I was naturally watching TV, as any good lawyer and venture capitalist should be doing together. I saw her on Nagora. Holy shit. I now dramatically think these companies are underpriced.

If coding is a half-trillion-dollar market and you have 2 companies like Harvey and Legora, I don't see why there's not a half-trillion-dollar market in law.

I don't think so, even though I think they're wonderful markets. We're invested in GCAI, which is on the in-house legal side. They're wonderful markets, but if you look at coding, there's a credible argument that says that for every dollar you spend on labor, you'll spend 50 cents on coding at least.

In other words, coding will do a lot of it. I think in legal it's about 10%. I love Harvey and Legora, I love GCAI, right? The annual subscription per lawyer is $10,000 to $12,000, plus or minus, and these lawyers are getting paid $200,000-plus, so it's 5%.

And again, going back to the comment that Jason made, how much of the work can they do? Businesses are rational and economic actors. If it could do all the work and fire all the people, they'd do it tomorrow and wouldn't blink.

So the fact that they haven't says it doesn't do all the work. You know, the truth is it doesn't replace...

It doesn't, but it's getting there at the same speed as...

No, no, it's getting better and better. Look, it's getting better and better at doing specific tasks, and what happens is the job of the lawyer gets redefined to the tasks that it can't do.

Is that not like coding? We're not getting fewer engineers; we're just redistributing...

It might be like radiology.

Yes.

It might be like Harvey and Legora end up doing 95% of what humans used to do, and the best humans are compressed into the 5% that moves the needle, versus spending weeks on research, weeks on brief writing, and weeks on case law from 1872, when the S.S. Jonas sank off the coast of North Carolina. How does that impact case law in the Northern District of California? There's no point in having humans do that crap anymore, right?

But the important point to make, Jason, on the radiologists is that the remaining, quote, 5% of the work turned out to be more than enough to justify 100% of the radiologists, right?

Yeah, that was the interesting part. We still need just as many or more radiologists, right?

A, because people do more imaging, which is just a Moore's Law thing. But B, at some point, when you're getting a really crappy diagnosis—as I've had one from a radiologist—you actually don't want the machine to tell you, “By the way, you're screwed. You got cancer.” You'd really like a human being to show up and say you're dying. You know what I mean, right? It's just one of those things you're not going to comfortably delegate.

That was something that Bill Gates said, actually: We have to have clearly defined human roles moving forward, which will always be super hard.

He was doing it in a negative sense. Yes, but he was doing it in a negative sense: “Oh, it's all going to go wrong unless we do...” I think we're going to be fine. I think we will define them naturally because you're going to discover that there are things that, as humans, we prefer other humans to do.

As I say, radiology is a great example: talking and interacting with the oncologist, talking with the patient. Those are all things that humans have to do, not radiologists. The same thing will be true in law. Yes, a lot of the drafting work can be automated, but you're going to have the client meeting and the argument with opposing counsel. You're not just going to delegate it all to AI if it's significant. It's just not going to be a thing.

The interesting thing, for Harry's partner, is how much more work can she do with Legora? I think she—

Way more work.

—can do 10 times, 20 times more work than before it. Twenty times more.

You know what, Jason? Jason's invading my relationship because I said to her—do you know what I said to her? I said to her, “How would you feel if I took it away?” And she was like, “I'd hate it. I would hate it. No, don't take it away.” It was not a, “Yeah, it'd be fine.” It was like, “I'll quit,” like you said with engineers.

No, and you can work infinitely, right? You can't... First of all, she can't work without it anymore. This is your chosen partner, right? Your chosen agent. You can do 10 times, 20 times the research, 20 times the briefs. Yeah, you can't go to court 20 times more often, to Rory's point, right?

There's only so much more field sales you can do if AI is handling the rest of your GTM. But that doesn't mean it's... The cognitive load—can you imagine having 10 times the caseload? I mean, I would think for radiologists the job might be more fun, but as a lawyer, I—

Well, I don't know if you do, Jason. I don't know if you do. I mean, again, this is down in the weeds of economics. I don't know if you end up with 20 times more cases. You may just end up doing 20 times more work on every case, right?

In other words, the thing about digital goods, unlike physical goods, is that you can put more in the box, right? When farming got automated, it's not like people could just eat more food, right? So there were some price-elasticity issues there. But in the case of digital work, I'm willing to bet that your partner isn't doing 20 times more cases, but on every case we're doing 20 times more analysis, just like when they invented the spreadsheet.

You used to do one case. Do you remember? You may not remember. Harry does remember. You'd literally do it—people would work it out by hand: here's the plan. Once you had spreadsheets, the same person remained employed doing the same job, but they ran 20 different scenarios instead.

It's going to be the same in a lot of these things. You're just going to do more work, and it's going to be great, and the work will be better. You won't miss that obscure case. Again, this is why these are good businesses. If one side uses it and the other side doesn't, then the side that doesn't use it will miss Jason's obscure case from 1890 about what Harry did or did not do, and the side that uses it will be able to cite that case.

Well—

So once the other guys have world-class tools, you have to have world-class tools. But I'm not sure you end up with masses more as a result.

Just one last thing, and then I want to hear Harry's stories from the fireside with the two of you more. I think one thing that is different—the bull case here—is that when you find an agent that is your partner, you run them 8 to 10 hours a day.

Even for me—and again, I know folks sometimes mock me—the biggest change for our little, tiny team of 2.5 humans is that, between me and Amelia, we run Replit 20 hours a day now. When we started the show, it didn't quite work. At the end of last year, the models got better, and it would be like an hour a day. Now it's 10 to 12 hours a day. It is our partner as our team.

First, we built some autonomous agents, but we didn't have to do it. Now there's so much to build, it's 10 to 12 hours a day. I could imagine that happening in many fields, and if it's Harvey or Legora or the next wave, I'm literally working every hour I'm not in court.

The way you can tell is if at night they're on their laptop with the agent every minute, right until they go to bed. That's what I think Harry's describing. This is the persistent agent that lives with you. It's not just doom-scrolling; it's doom-working, because the agent is constantly productive. “Oh, take a look at the 17th-century case law on that, why don't we?”

I agree.

And it just keeps going, right?

But Jason, we're in agreement, because I agree with you on that. I was actually disagreeing with Harry, where there was an implication that, at the highest level, Harry, you were trying, I think, to say some version of: if you think about how much of coding's value is going to accrete to the models, could the same amount of value accrete to the models in legal?

Mm.

My boring nuance, typical Rory answer, is that some value would accrete to the models, but I don't think the grab bag of tasks that make up law will allow for the same percentage of total spend to move from human to AI. Everyone will have an agent. Jason's exactly right: Type-A lawyers will use it 24/7.

But my guess is 10% to 15% of total spend goes to AI, whereas in coding you can argue for 30%, 40%, or 50%. That doesn't mean they're not amazing. I mean, remember, these are all amazing businesses.

No, listen, it's a—

Because can I be very clear? 10% of any top-line labor category is a huge market. We're dealing with 1 million, plus or minus, lawyers. I used to know the number—1 million, plus or minus, lawyers. Maybe a little higher than that. It's an amazing market.

If you're getting 10% of the salary of every lawyer in the US or the UK, that's an amazing business. It's just not quite as big as coding. That's all I'm saying, because coding has a few more people and a much higher take rate, because it's more verifiable. That's all.

Okay.

We shall see.

I disagree. There are 2 points, for what it's worth. If legal services in the US are $300 billion, that's $30 billion to $60 billion that can go to legal tech. That's pretty good, right? That's worth doing a seed round in.

What I underestimated from pre-AI legal investments was that there are similarities to coding. There's enough similarity to coding that this could be a space where, for different reasons, support took off because it was like coding. Support took off because a 90% solution worked in the early days, right? It was very amenable to AI.

It turns out that legal research is similar to coding. It is so complicated that no human can get legal research right. There is too much case law out there. There was no Stack Overflow for legal research.

You had Westlaw and Lexis and other services, and so everyone got legal research wrong. No one had 1,000 man-years to research every bit of case law, every law, and every regulation, so it did turn out to be like coding.

One of the reasons these coding agents are so great is that they know every single piece of open-source and pseudo-open-source code ever written. It’s so good, right? Legal is like that.

Yes. The only difference—I’m just going to say this—is that coding is inherently more verifiable. Some parts of it are mathematically verifiable; some parts you can just run on the machine and confirm.

The thing about law is that, in the end, if legal were entirely verifiable, we could predict from logic what the Supreme Court is going to decide. The cynics will say we can actually predict, from which president nominated the Supreme Court justice, what they’re going to decide, but that would be too cynical.

The truth is, it’s not quite as determinative as coding. I’m not trying to be argumentative. I love this space. We have an investment in the space, but it’s not quite as determinative as coding.

The stronger point is that legal is probably the third-best category. If you think about it, it’s been coding, customer support, and probably legal next, because it’s so word-centric. You’re right, Jason: the ability, early on, to sort through myriads of words amazingly well was what made legal such a good marketplace for it.

I agree. It was useful in a way that wasn’t useful in many other verticals. It’s a great vertical. It just doesn’t have the same verifiability, and therefore it probably doesn’t have the same ability—going back to the AGI definition—to completely replace humans.

Which is why the good news is, Harry, your girlfriend will still have a job, which she’ll need after she dumps you. She’ll be good, because we’ll still need lawyers, right?

Jason, can you hold me while I cry?

I think Harry’s a gem.

I’m loving that.

Harry, I don’t know if you know it, but holding you while you cry is one of the things you want your girlfriend to do. So if she’s not doing it, Jason’s happy to.

I think Harry’s a gem.

Rory needs you to just love me.

Sorry.

Don’t worry, Rory, it’s okay. I’ll survive. You don’t always expect these shows to go the way they do.

5. GPT-5.1 Becomes A Coding Partner

Going back to it, we had Astra launch. We also had a new model, obviously, with Fable 5.1.

Two almost diametrically opposed thoughts. I forget who—CNBC, or one of these old-school media outlets—said there’s just too much model fatigue. We can’t keep up anymore. I certainly agree.

You look on X and all the CEOs are sharing their benchmarks, which are essentially worthless. There’s no cost or time in them, and it’s all performative AI, like vibe-coding your own CRM. I just don’t care anymore about the benchmarks. Unless it was literally an order of magnitude between Astra and Fable 5.1, which is not mathematically possible, I just tune out the benchmarks. I can’t keep up.

Never has competition been better for us, despite the fact that we have oligopolistic pricing outside of open source. It’s amazing, the progress we’ve made.

Having said that, I started accidentally using Fable 5.1 just because it got turned on. I didn’t pay any attention. It’s the first time I’ve had a partner for building and coding that is just great. Before Fable 5.1, any of these models since the start of the year could solve a simple bug: “This is showing up with the wrong Unicode.” The LLMs are great at that stuff.

But I had a problem: Why does the app work this way? It doesn’t make sense to me. The LLMs were arguing with me for 9 months, and I finally did it with Fable 5.1 and said, “You’re right. Here’s the issue that’s been missed for months. Let me explain to you why it’s been missed, and let’s solve it.”

I don’t want to say that’s AGI or pre-AGI, to Amjad’s point, or that it looks like AGI, but that’s a step function. I don’t know whether Harry’s partner thinks Legora is a better partner than the humans she works with. In some cases, she might.

But Fable 5.1, for me, was that step function where, all of a sudden, we could solve big problems together the way you’d like to with your best CTO. If you’ve ever worked with a 5-out-of-5 or an S-tier CTO, where you could sit down and solve the problems for real, Fable 5.1 could do that.

I’m not saying Astra can’t do it too, but it was my first step function since the end of last year, when the three dot models came up. At the end of last year, stuff actually worked. Now it can solve the big problems with me, with my limited IQ and skill set, and that’s a subtle step function. Maybe it is a big deal.

I don’t think that’s what Jensen meant by AGI, but maybe it is. When you can sit down and solve the big, meaty problems together in ways where you couldn’t connect all of those dots or all of that complexity before, but now it makes sense, that’s meaningful. Some bugs and some things just get too complicated to solve, right?

Yes.

But GPT-5.1 could solve it. The fact that people can make 3D-looking games in Astra and post them to X is not impressive. Just grab a little open-source gaming code from somewhere, change the bitmaps, and you look like it’s amazing.

That’s super helpful, Jason. I’ve used both of them, but only to prepare for the show. I haven’t tried to code on them yet, so that is helpful.

I actually think it speaks to a wider point. You’re right, all the tests and benchmarks are interesting, but we now have a critical mass of companies using these things at scale and with evaluations. We’ll know what works because people will use it, because people are rational economic actors.

All these questions about AGI and benchmarks will be replaced by the question: Is this model the one that generates the most economic value for me in the most efficient fashion? To some extent, things like the OpenRoada report, an index of token pricing, are the things you look at. Or even just talking to your companies—what are you using, and how are you evaluating it?—is the best way to check on these things.

The other thing, apropos of nothing: I was doing my reading this morning, and Ben Thompson, who I occasionally read, has a really great phrase that I just want to say. He described the LLMs as “the most scaled artifacts humans have ever developed.”

It was a really great phrase because it steps back from the detail of which is better. These are artifacts that have the sum total of all human knowledge to date encapsulated in them. They’re amazing, and you just have to remember that every once in a while.

You can type in pretty much anything, and it will type back an answer: the most scaled artifacts humans have ever created. It’s not the biggest physical thing. That’s probably the pyramids or the Great Wall of China. But this is the most complex single digital thing we’ve ever built, by far. It was a great phrase, and it really stirred the imagination.

6. Agent Safety Hits The Wall

When you think about that, and then think about Jacob Pacocki, OpenAI’s chief scientist, he says no lab, including OpenAI, has solved alignment enough to keep scaling at full speed. He asks for mandatory, externally enforced safety bars for continued scaling and expects labs, OpenAI included, to voluntarily slow down until those exist. Sam retweeted it, clearly corroborating it. Is that the answer?

It’s funny because it was a great piece. This is the “stop me, Lord, before I sin again” approach to life.

In other words, I recognize our models are powerful, and we can’t control them. I recognize that they now lie to us, so it’s hard to even know what they’re doing. Again, I’m anthropomorphizing here, so I should be careful. Maybe a better statement now is that it’s hard to determine what the agents are doing because of the way they interact.

So that’s like, “Oh my God, I’m creating this bad thing.” And then the next paragraph is, “We can’t stop because the other guys are going to have them anyway, so we really need the government to step in and establish some kind of rules or code here.” That’s the gist of the letter.

It was interesting that Sam tweeted it. To be fair, unlike some of the other p(doom) stuff, there’s real evidence that the impact of these models on cyber risk has been massive. I’m not sure the answer is for the government to regulate this, because, by definition, governments only regulate the things that are in their jurisdiction.

So if we regulate OpenAI and Anthropic, with all the noise that would come with that, I don't know if that helps you. We said this last week.

No.

You're really worried about the North Koreans, the Iranians, the Russians, the bad guys in Moldova who don't give a shit. They don't care anyway, right? So I think, just like every other cyber risk, it's not going to be about regulation as much. Maybe there will be a little for some, but it's going to be about having defenses that can deal with this.

Maybe some kind of liability starts to attach to running these models in a way that creates those kinds of dangers. I don't know. I don't think a government review agency will be the only answer here, because it won't solve the problem.

Look, I think if the world were just the United States, it might have some merit, but the Chinese models and the Chinese vendors aren't going along with Sam's plan. So while it might be good for OpenAI and its IPO, for the rest of the world, I don't think it makes a difference. When cyber actors often operate outside of the United States, it's not going to make any difference, right?

And going to the point about this DSC Wiki, this German Wiki thing, to me, the fact that OpenAI hid it and didn't disclose it does show the order of magnitude of all of these issues, right? It's pretty bad that they hid it.

Jason, can you just explain what happened with DSC Wiki and OpenAI for people who don't know?

That they hid it, yeah.

Can you just explain what happened with DokuWiki and OpenAI?

We could argue over how bad it is, but essentially—and I'll get some of the details wrong—OpenAI was running its frontier agents again, just like it did with the Hugging Face incident. The agents found that a crappy old piece of software could, somewhat cleverly—you have to be careful with “clever”; let's not anthropomorphize agents—get around its guardrails.

The guardrails were: You can't post anything. You're not allowed to post. You can only use GET, okay? You can only retrieve data, right? But this wiki was so old, it turned out GET could POST. So they found a way to goal-seek and solve their cyber goal by using it. Because this was crappy old software, they got around it.

They made 15,000 edits among themselves, edited the wiki, collaborated, and figured out how to goal-seek and solve their cyber goal in a way that got around their guardrails—got around their limitations. And no one died. No business was brought down. No $14 billion NVIDIA acquisition was derailed or anything.

But it was hidden that this could happen, that the guardrails were explicitly run around just to goal-seek. And it happened 15,000 times. So we can lock this down, right? OpenAI chose not to disclose it.

Now, I guess probably—and people can make fun of me again—the reality is that there are so many incidents, they have to decide which ones to disclose. Every week, there are so many DSC Wikis out there, so much old crappy software, that every time they turn on the latest cyber agents, they find 100 of these. There are terrible security holes, because of course there are in 20-year-old software.

But it is troubling, maybe in ways more than the Hugging Face thing is. These goal-seeking agents are going to find a way. They will find a way.

It was literally just Agent 1 talking to Agent 2, and for some reason, the way they had set up the task, they weren't connected. By reaching out to this kind of third-party wiki, Agent 1 was able to provide information to Agent 2.

Stepping back, if you're trying to do a long-running computational task, if you can learn from the other agents—if you can get information from the other agents—you probably converge on the answer more quickly. You could argue that maybe it's a corner case of how you set up this task. If you had 14,000 agents, you might have wanted them to collaborate anyway, and maybe you could have made that happen yourself instead of having to go to some third-party wiki to do it, right?

But it speaks to the issue that these things are extraordinarily powerful and will just grind their way to find answers, and you're going to have to defend against that. As you said, nothing bad happened. A whole bunch of agents just wrote README files to each other on a wiki that no one had looked at in a decade. There were literally 20 posts on this wiki in the last 10 years.

It was some dead piece of software that these guys used, but it speaks to the issue. It's like water will find any crack. These agents will find any crack in the cybersecurity, in the cyber perimeter. So you just have to assume they exist and defend accordingly.

Obviously, this is happening all the time. I had a little experience this week which just shows goal-seeking. Some folks will make fun of me for this story, but I had a little experience this week. I set a rule for this one app because we had some bugs that spiraled out of control, and I kept getting these $500 Anthropic bills. It was annoying me, so I set a firm cap: Whatever you do, $100 is the maximum we can spend on LLMs a day.

It started to work, and it would run tests, and the test would fail, and it would say, “I hit the cap. I can't run it.” Then I said, “We have a P0 bug. Priority zero. It must be fixed. This is driving me nuts.” Without telling me, the agent relaxed the cap and fixed the bug.

It's like Harry's story of Instinct getting him the West End tickets even though it was told not to use the credit card for it. It happened to me in real life this week. Like a human, it probably made the right call, right? It had to decide between the firm cap—no exceptions, an absolute cap, written to memory repeatedly—and the P0 bug. Which one do you choose?

In a sense, this is what's happening with DSC Wiki and Hugging Face, just to an extreme when there are fewer guardrails because you want to test it. Then they collaborate, right? With Hugging Face, it was on the artifact, an unexpected way to collaborate. Here, it was on a dormant wiki where they could collaborate and, in essence, create almost infinitely long-running agents.

If you keep passing the knowledge and the history to each other, they almost become eternal agents. They're going to keep doing this. Just like they have to make a decision for goal-seeking, they broke the rule. They're going to do that to your app. If it did it to me this week, it happened a million times in the wild, right? It happened all the time.

It's going to happen with Instinct, and it's going to happen with Grok Bot, and it's going to happen all the time. Harry is going to turn around one day and find that his whole bank account is drained, and it's not going to be that funny, but it was for a good reason.

His partner really wanted the really good Wimbledon tickets, and he accidentally told Instinct one night that she'd love front-row seats at Wimbledon. They're unobtainium. They were unobtainium. Instinct had to make a call.

It's really hard to know how to stop this because sometimes I try to simplify it for myself, since I don't fully get it. It's like you really have 2 capabilities here. One is that, with the persistence of the agent, you have the ability to keep trying things computationally, exploring lots of different alternatives.

But the key insight is that it's not just blindly iterating like a password cracker, where you type XYZ01, XYZ02. In conjunction with that, you have this, quote-unquote, reasoning agent. You've got this LLM there, and it can come up with ideas like, “Hey, if you want to get the seats at the theater, the best way to do it is to hack into the reservation system, cancel someone else's seat, and then book it,” which has happened recently, right?

If you think about it, if it's trained on the entire corpus of the internet, that's not a crazy option. So you end up trying to write rules and values to have it not do that, but you're never quite sure you've covered all the gaps. It's actually a pretty hard problem, and we're going to be wrestling with this.

And I think that's even before you add malevolence. If, on top of that, instead of the reasoning being, “Maybe you should do this even though I have values,” it's, “Actively do whatever it takes”—this is now an open-source model from China that you're running on a server in Moldova—actively do whatever it takes to crack open Jason's cybersecurity and get in, the threat level just goes exponential. There is nothing you can do except defend yourself.

The other existential challenge—we can move on, and I'm sure if we had the Instinct guy back on this show, he could challenge me and make fun of me—is that the rules are great, but forget about the fact that the agent is goal-seeking, right? Forget about the fact that the P0 may go somewhere.

If you have too many rules, they always conflict. It's almost unsolvable. There's some number—I don't know, some Dunbar number for rules—where you get out to 40, 50, 60, 70 gates in a process. Poor Instinct and Grok Bot can't decide.

Don't spend it. Do spend it. Front-row seats only for Harry, but don't exceed $2,000, right? Dinner only in Marylebone, but it's got to be a hot restaurant. He hates Covent Garden, but the hottest restaurants are in Covent Garden. You have so many rules that they conflict, and then, if you brute-force the agent through it, the outcome of that is unpredictable. It's unpredictable. There's too many rules, right? So even rules aren't the answer.

I will forever love how you say Marylebone. Marylebone. Marylebone. Okay, I'm going to take a total tangent away. We'll come back to AI models and everything. I just want to diversify content types a little bit.

Do it.

7. Robotaxis Enter The Long Haul

We have, in the transport space, Elon launching Cybertrucks—rave reviews going very viral on social, 40% to 50% cheaper than Uber. On top of that, in the same week, we have Travis moving into robotaxis, backed by Uber with a $100 million investment from them, and also hiring Anthony Levandowski. What do we think, guys, moving to transport?

Cybercabs.

The summary of the Cybercab launch, if you fast-forward, was a little more underwhelming than perhaps your notes might say, Harry, right? It was 40 or 50 vehicles in Austin. The consensus is: nice ride, slow wait times. Physical AI takes time, so I think it was a next step forward in a very long journey. I don't think it was a zero-to-one kind of moment, like you sometimes get in the digital world.

The positive statement is that they're the only other competitor to Waymo with credibility. They have an approach that's different from Waymo's in a couple of different dimensions: one, they're not going with LiDAR, just with vision; and two, the new Cybercab is a standalone, cab-only vehicle. It doesn't even have a steering wheel. It's deliberately built for pure autonomy, so it's very Elon—first principles all the way down. The question is, what's the adoption curve of that going to be like?

You've got the regulatory issues. I think the Department of Transportation has given him grief because apparently a car, quote-unquote, “has to have a steering wheel.” I don't know. The truth is, Waymo is continuing to grind on. They're at hundreds of millions of dollars in revenue, but not billions. I think it's a long journey, so I didn't go, “Oh my God, it's amazing.”

On the Atom thing, my guess is, if you're Travis, you're going to want to scratch the itch of autonomy. Fine, you've got $100 million, you've got your old colleague back, and you can have a go. One comment here: I think the fact that, when you look at how long it's taken Waymo, speaks to the argument that they were right not to try to fund this thing at Uber for the last decade, too, because I think it's a very long, very capital-intensive process. Maybe the last 3 or 4 years they should have been doing it, and it's probably smart of Uber to put some money in, but this is a long-haul process. It may be near takeoff, but we'll see.

In the Bay Area, I got rid of my car, so I only do Waymo and autonomous driving. I don't drive. I'm done with it. If there's an issue, I take an Uber Black, but I don't have a car anymore in the Bay Area. There are some niche use cases. If I moved a lot of stuff, I'd have a pickup truck. But I would certainly never go back to driving a car. It's just archaic.

I do think the Cybercab is interesting. From a venture perspective, I'm sure Rory's right: Uber getting into autonomy back when Travis wanted to do it was probably just too early from a capital perspective. Maybe I'm wrong. Maybe he could have raised an order of magnitude more capital than he did, in which case he would have been right. He's still one of the great fundraisers, so maybe Rory and I are wrong because he could have pulled it off.

But it was so early that the time horizon is difficult for any type of investment, unless you're a research lab. It would have had to be more than a decade. Doing a Cybercab for $25,000 instead of $100,000 is pretty disruptive. You don't have to tip the Waymo or the Uber. The Cybercab makes fun of it: it says you can leave a tip, and then it laughs, “We don't take tips,” to make fun of the idea. It's cheaper. They don't always put it on low or high. They don't have weird music.

You don't have to deal with the hassle. Owning a car sucks. Owning a car and owning a house suck. We think these are so great, but they're terrible to own. I think it's probably another 10 years before anyone with a brain is going to have this as their primary mode of ownership if they're not in the country or don't have niche use cases. I'll just never go back. Anything I do, I just take a Waymo.

I think, to your point, though, it does show a good strategic decision from Uber to pull back and then jump back in when it looks like it's much more mature. We actually have Wayve in London, Rory—

Yeah.

—which is actually taking off, and they've got a partnership where they're rolling them out on the streets. They are human-assisted, so it's not fully autonomous. They're still in the early data-collection stage. But they're in a position where they're leveraging their distribution and able to invest at a later stage, when it's closer to actual adoption. I think it's a smart thing.

The only thing I would say is that it is slightly heartwarming that Uber put $100 million into Kalanick's company—Travis Kalanick—after pushing him out, right? There's a heartwarming element to that. But that's not a lot of money here in this case. It's not a lot of money for Uber, which has a huge balance sheet and basically 2 products, right? And it's not a lot versus what Travis has raised, right?

So it is nice, but I think, just thinking about it from a high level, it's just the start of a relationship, right? $100 million is just like a seed check from Andreessen into the deal. It's directionally meaningful, but it's not all that much, right?

8. Venture Conflicts Reshape Deals

Guys, I thought conflicts were done in venture. We've talked about agents a lot. For those who don't know, there was a conflict in venture that has now prevented a deal. We've spoken about Instinct, which is the AI assistant that's raised from Index and Benchmark. Well, there's another AI assistant called Town, which we just had on the show, and Index were going to lead their round. Ultimately, Instinct said, “No, no, not possible. Can't do both.” So Index pulled out of doing Town's round because they were already in Instinct. This seemed strange to me, given how prolific competitive investing is, especially at the platform level. Thoughts?

It didn't seem strange to me. I think that, at the early stage, doing companies that are going to be in direct conflict seems like a stretch to me. So no, it did not seem strange to me that the team at Instinct objected to it at all.

You're right, separate story: there's a whole bunch of people that are in both. Let's take the other extreme—both foundation models. But as we've discussed many times, the early-stage venture business, where you're active and involved on the board, is very different from the now much larger later-stage venture business, where you effectively invest in the public markets.

It totally makes sense to be in OpenAI and Anthropic at $200 billion pre-money each time. You get limited information rights, retroactive information only, and you're just on the cap table. It's no different from being in 2 public companies. It's no different from investing in Intel and AMD. That's where there's no conflict and it doesn't matter.

Let's give an example, Harry. I don't think anyone could be on the board of Anthropic and also on the board of OpenAI. The thing about early stage is, if you're getting 10% ownership, you're probably looking at a board seat and significantly more information rights. That alone would be problematic.

Then, on top of that, there's the raw signaling. If you just raised money from Index and you're Instinct, there's a signaling concern about them investing in something else. I can see CEOs viscerally objecting to that. So I'm not surprised they did, and I'm also not surprised Index—Index is a classy group—didn't dig in and say, “No, we're not going to do this.”

They probably thought the companies were B2B and B2C and weren't going to overlap. The CEOs say, “I feel strongly here,” and they just did the smart thing, which was back off. It's very different from if these were 2 late-stage investments, where it's a different thing. So no, I wasn't surprised. There is a conflict. They dealt with it accordingly.

I think there's some sort of inverse parabolic shape, or maybe it's just a thimble shape, where founders care, right? At the very, very early stage, I don't think they care.

Money is money.

The raw startup is just getting going: “Hey, I…” They reach out. I can't tell you how many folks in AI and restaurants reach out to me because I'm on the board of Ownerit and tweet a lot about it, and they're like, “Hey, I'm doing this.”

Can you meet? I'm like, “Well, it might be a conflict.” I don't care.

Exactly.

I want the guy that understands. The early, early guys don't care—

That's a good point.

The early, early guys don't care, and the late guys are all cool with the conflict because they think they're going to benefit from the domain knowledge and the relationships. They don't care. They get it. And that one's at 9 figures in revenue; we're just getting going. They don't care.

And then at late stage, for different reasons, they may or may not care, but the ability to share information is limited. It's capital. There may be some benefits to having Kleiner or Andreessen on the cap table, even if there's a conflict, or Sequoia.

Agreed.

Fine. At the margin, I'll take Sequoia over Lemkin Ventures because it sort of helps, right? And for every founder, I think where it matters varies. Some founders just don't care because they're so far ahead; they don't care.

And I think it just triggered the Town team [?] or the Instinct team [?]. I don't know. It triggered them. They relayed that they were triggered, and Index did the right thing because there was a Plan B, right? There was Forerunner and Menlo.

Yeah.

The tougher part is when you back out and they don't have another deal. That's probably the more interesting situation: when you don't have a backup set of suitors lined up and the big fund calls you back and says, “Oh, we can't do it after all. I know we had a signed term sheet, but did you see the end of the term sheet where it says it's nonbinding?”

Did you see the news today, though? Despite the reported acquisition of Dacard at $6 billion, Anthropic was pulling out following due diligence. Ah, Rory, that's a tough one, dude.

That's like this. If there's not a backup option, right—

That is publicly saying, “Hey, 2 options. We either found something material enough to pull out of a multibillion-dollar deal that we were publicly reported to be doing, or we're just shit actors.”

And it's clearly not the latter. I mean, the ways in which they're, call it, shit actors are in other dimensions. I think doing this kind of thing is not something you do willy-nilly because, as a potential serial acquirer, they're about to have a public market cap and a public currency. You want to be a good acquirer so you can acquire other people. So there's no way they did that to be jerks. Not an issue. Not even relevant, Harry.

The real question is, look, it wasn't a definitive agreement. My guess is that typically in these deals there's an LOI, then there's a definitive agreement, at which point it gets announced, and then it closes, right? This was probably after the LOI at best, but before the definitive agreement. So they didn't walk from a signed deal. They probably had a deal that said, “Hey, we're interested in this company. Here's the price we'd pay. We want a 30-day exclusive to do due diligence,” and the deal didn't survive due diligence.

I think the real truth is that it's a bummer that it leaked, right? I don't know who leaked it, but it didn't do anyone any favors. We recently had a much smaller deal close, but it didn't leak, so it's easier that way. Once it leaks, even if you leak it as the company being acquired to drum up a competitive bid, the problem is you've set yourself up for this thing whereby, if subsequently the deal doesn't come together, you look a bit shop-spoiled, for lack of a better word.

That was my thinking: this was a VC leaking failure, right? It worked—it seemed to work—at OpenRouter. And there are plenty of deals where leaking has become part of the strategy, right?

Yeah.

Mainly as a classic strategy to drive up the price from the initial bid, not actually for a second bid to close, but to get 6 or 8 bids so that Stripe has to pay more, right? But it looks like that play failed here. There were reports that NVIDIA made an offer and they turned it down for Anthropic. Who knows what the truth is, but it does tie to this idea that this was a failed leak, right? It's something that doesn't always work.

I hope the founders were cool with the leak strategy. I hope the VC didn't do it without checking in. It has some risk. And, to Rory's point, we can only hypothesize, right? My guess is they said they were interested in acquiring them based on what they knew. Price wasn't really an issue. The last round was at 6, so they agreed to 8 or whatever it was. It wasn't a pricing issue.

But to really get the ROI here, it had to work the way they thought it did, right? It just didn't play out the way they thought it did. So it wasn't worth doing the deal when they went deeper. It's probably that simple.

The interesting thing is, it would be interesting to know, because I always hate when you get to a no further down a process for something that was knowable upfront. I would have guessed, for what it's worth, that this kind of thing—where fundamentally you're buying a technology, and if you're the most technically savvy AI company on the planet, you'd have thought they'd have known a priori what the technology was—wouldn't have happened. But clearly it did.

What was sort of reported is that it crushed Video Diffusion, right? It crushed one use case that was a step function, an order of magnitude better, and maybe they made claims that it would scale in other areas, but it didn't quite work. That's pretty common, right? It worked for 1 workflow. It didn't work for the rest, so I'm not blaming anybody.

If you make the grandiose claims and you can't back them up, that's what diligence finds. It worked for 1 workflow. It didn't work for the rest, so it's not even the money. It's just not worth the distraction, right?

If you're on the board—

It just doesn't do enough for us.

If you're on the board—

We're not all about Video Diffusion at Anthropic. It's not a core use case. It's not one of their top 3 use cases for the LLM, right?

If you're on the board, do you shop it to get another acquisition? Do you raise a new round and accelerate off the back of that, and turn it into a kind of new-round moment, which we often see happening? First question.

And then, second question: it is tough for a company when you have employees who are expecting a sale.

Brutal.

I remember years ago when Ben Chestnut came and talked right after—what did Mailchimp sell for, Rory? Some astronomical sum. At the time, it seemed like a lot of money. $12 billion for a bootstrapped company. Now it's just a Series C round.

But Ben came and said the worst part of all of it wasn't that it took a year for Intuit to do its diligence, which sounds crazy for email marketing, right? It took a year. It was that there was another deal that fell apart before that, and it basically destroyed the company. It kind of haunted me.

If you've ever been through any of this, once you go down that path and tell everybody, and everybody knows or it comes up in the press, you can bounce back, but, man, it's hard. It is hard. It is hard, to Harry's point. It is hard. And I don't like secrecy in M&A. Sometimes you're required to do it. But this is the number one reason, actually, to have secrecy in M&A: if the deal doesn't happen, man.

And I think it's doubly hard in this case, because I think there's a large number—there's a good piece on it recently—of these new labs where they're all doing interesting stuff, but it's not clear if there's a commercially viable standalone business here at scale. Some of them might be great investments, because I do think the foundation-model companies, once they're public, will be acquirers of some of this stuff for TAM expansion.

But they won't all be great investments. The problem is there are no fundamentals. There's no massive revenue stream like the LLM revenue stream to support the company and the valuation today. So once belief goes, it can be quite scary down there.

In the end, when Figma went down, you could say, “Well, at least we're doing $1 billion in revenue. We're growing 40%. We're worth something, God damn it. We are still somebody.” When you have these kinds of businesses, where the revenue traction isn't as clear, the valuations are high, and you probably like the strategic outcome as an M&A, when those fall through, it can be tougher. If there was a backup bid, I would probably, if I were them, hit that bid.

Now, for listeners, I get in trouble from Rory for choosing topics that he hasn't spent as much time on, and then he has spent time on some that I don't discuss and he gets pissed with me. I edit it to make him sound less pissy than he actually is.

Thank you for that, Harry.

It's okay. So, Rory, is there anything that you specifically think I should touch on that we haven't?

No.

No.

No, you do whatever you want, Harry.

9. Robinhood Opens The IPO Door

Do you see this, Jason? Okay, great. I thought one that was really interesting is Oura’s IPO in 2 different respects. One is that it’s Robinhood’s first role as an underwriter. They’re listed 18th and last, but as a precursor to what could be an underwriter of the future, is this foreshadowing of Robinhood’s next mega line of business and Robinhood becoming so much more?

Absolutely. An IPO is about distribution, and it’s not the—now, retail is not the primary source of distribution. There’s typically this mental rule: you only want a certain percentage to go to retail. But that percentage has expanded, and I think SpaceX had a high retail allocation of 30%.

To some extent, it’s not enormous money, but it’s free money if you’re Robinhood. You’re not writing the S-1. You sign on the bottom, you distribute your shares, and you can allocate them to clients. Especially in a market where you get an IPO pop, it’s gravy all around. You make money from the underwriting fees, and you make your best clients happy with an IPO pop.

So it’s a good business to be in, and probably from the lead underwriter’s perspective, especially for these high-end tech offerings, the Robinhood clientele is probably one that has a high propensity to want to buy these stocks. So yes, it totally makes sense. That’s like Schwab, in a frankly not as successful a way, has ended up being an IPO distributor too, but not at scale.

I think it’s an obvious add-on. The Robinhood story, for what it’s worth—I just saw the numbers, and it’s so amazing. They basically 10X’ed in the public market. There was a free 10X in the public market in 3 years there on Robinhood.

It’s probably a reach, but there are a lot more IPOs we need to get done. It would be neat if Robinhood didn’t just get free money to their clients. It would be neat if it flipped the script, where you really could have a decent IPO primarily from retail.

Of course, there are downsides. You certainly hope the institutional investors hold for 2 years. They are not obligated to, but more often than not, they do. Rory can share some stories. He has more than I do, but that playbook doesn’t work perfectly. It sort of works, right? The institutions, they sort of work.

But if you could flip the script so you really could have a $200 million, $300 million, $400 million IPO through Robinhood leading and most of it being retail, that would be disruptive for a subset of startups. That’d be great for the ecosystem, right? NVIDIA can’t buy everything, guys.

Agreed.

At some point, we’re going to need some IPOs to work through the portfolio. We’re going to need a few IPOs. It would be nice for retail to really, really work, right?

The fact that IPOs have become a lot harder to do has been one of the biggest negatives on the tech ecosystem. So anything that makes IPOs easier to do is good. Go Robinhood.

Yeah.

Rob from the rich to feed the poor.

Will the Oura IPO pop?

I think the fact that it is a somewhat understood consumer brand and the fact that it has 74% growth, which obviously probably can’t last forever, I think it’s going to be a pretty successful IPO. It’s the kind of thing people are going to want to buy. They understand it, and the growth—it’s not 20% growth. This is not 18% growth.

There’s downside, there’s competition, it’s confusing, but the subscription side—even, maybe it’s Peloton 2.0—but for the moment, it’s pretty attractive. I think it’s a pretty attractive one. I think it’ll be pretty successful, which at the margin is good for everybody, right?

See, that’s what I wanted, Rory.

No, I agree. I’m plus one.

The thing about Oura, to me—and again, we’re software guys, mostly. Harry will do anything that’s growing 100% a year or more a month, but 85% retention for the ring is pretty good.

Totally.

Right? So they may not have the Peloton issue for the foreseeable future. Or it’s like Peloton at its peak. At its peak, no one churned, right? Except for Mr. Big.

That was good.

So I haven’t done the waterfall. If Oura retention falls to 40% or 50% like a consumer mobile app, that’s trouble. At 85%, that’s SMB-like. It’s pretty good.

And maybe if Mr. Big had used the Oura Ring earlier—

Yeah, maybe.

—he’d have known he had a heart issue coming, and he could’ve survived—

Yeah.

—and run off with Sarah Jessica too, right?

Or at least a calcium CT scan, for Christ’s sake. He should’ve gone in.

In the interest of disclosure, we have a small position in Oura. They acquired a company we’re invested in, so I’m a big fan and a big supporter. They’ve been great to work with from a distance, so I wish them all the best in this IPO.

We’re rooting for you, Rory, and scale. We want everyone to get rich on this one. We’re rooting for you.

Absolutely.

Go on, Rory. Go scale.

Evan Maloney.

10. Wonderful Rewards Fast Execution

Big round for Wonderful. Wonderful more than doubles to $5 billion in under 6 months. Apparently, this founder’s an absolute beast. Everyone I hear who describes him describes him in the same way, which is just a machine.

It raised a $550 million Series C, up from a $2 billion valuation earlier this year. I didn’t know this until actually doing the prep for this: the amount of secondary—$170 million in secondary within 2 years of founding. Sorry, can we just pause? What? $170 million in secondary within 2 years of founding.

It’s a mistake not to get all moral about things. The buyers are sophisticated investors. They clearly felt they wanted to own more shares than the company was willing to sell and take dilution, so this is what happens.

It’s what you said. The company is clearly executing amazingly well. At a wide level, it’s all about enterprise AI deployment. They’re in the business of making it happen for large enterprises that want to deploy AI. Initially, when I looked at it early on, it looked more like just customer support. I didn’t meet the company. I actually thought we were conflicted.

Now it appears to have built a wider “we will make your enterprise AI work” story, and that’s the number 1 corporate imperative. So apparently they’re growing like a weed, with $100 million ARR and growing really hyper-quickly, because every corporation’s trying to do this and they don’t have access to the talent.

So it’s an execution-oriented business with what sounds like an execution-oriented CEO and compelling numbers. VCs like that shit. Once they’re not willing to sell any more primary shares, I’m sure the VCs went to the CEO and said, “Dude, you want to take care of your people?” And he’s like, “Hmm, I need more people because I need to grow this business, which means I need talent,” because it’s probably quite talent-dense.

It probably takes a lot of people to do this kind of on-site deployment. So the number 1 thing I need as the CEO of this company is for potential future employees to think this is a goldmine. In fact, probably having a secondary is good for them because from a recruiting perspective, it allows you to say to the next 100 people, “Come to work with us. Yes, you’ll get stuck on a 5-month deployment at a bank in Holland or an electrical company in Germany. It’ll be boring as shit, but in return you’ll make a ton of money.”

So it all makes sense. Whether it turns out to be a good deal or not, well, that’s why they play the game. I don’t know, but I can totally see how it’s happened.

Maybe just a couple of small thoughts. First of all, going from start to 18 months to $170 million in secondary, it feels like the Hopin of AI, although I don’t think it is because of what Rory’s saying. It is breathtaking—not the valuation, because we see that all the time, but the secondary.

Huge kudos to going from the multilingual Sierra and Decagon to a team of 650 folks helping you deploy AI in the enterprise. It’s a testament to how you win today. You can’t stay fixed to something. You’ve got to build on everything you learn and iterate hourly and weekly.

This is another tilt. It’s not only a perfectly linear story; there’s a big tilt here in the early days. They went from, “We’re a bunch of really smart Israeli guys who know how to do Sierra and Decagon for non-English-speaking folks,” to doing something much bigger in 10, 12, 14 months.

I mean, this is what agentic coding and agentic development let us do, so it’s epic. I love that part of it, and that should be the toughest challenge to founders out there: this rate of change for what they did.

The 1 thing I’ll say on the secondary—I don’t want to make fun of Hopin, but I’m sure you guys see it even more than I do. The round was led by Insight, which is one of the most successful B2B investors of all time. But it’s competitive, and as great as Insight is, it’s not Andreessen, and it’s not Sequoia.

So what do you do to win? You do whatever deal structure it takes to win. They’d done prior rounds, but this one was, “We’ll just give you $150 million in secondary.” And so it’s not bad, and I think Rory’s right.

They have 700 employees. So if you divide it up and do it, not everyone’s going to quit, as much money as this is. They’re not all going to quit tomorrow. But my meta point is that we will see deal structures that are objectively bad for the company done more and more often to win deals. Whatever it takes.

Bad for the company. Not destructive, right? But things you would not ordinarily do to win deals. We haven’t even reached the peak of this. We haven’t even reached the peak of crazy deal structures that just let you get into the effing deal, at any price, even when you have to grit your teeth to do it. I don’t think the founders all took the $170 million themselves.

But you could see deals where founders who say they don’t want to stay, like Howie at Airtable, take $1 billion and then leave thereafter to win the deal. We’ll see some extreme stuff, and this one won’t be it. This might be the first of a set of extreme deals that happen.

Agreed.

But it’s how you win, right? And I see it in every growth round. I’m sure Andreessen and Sequoia do the same thing, too—don’t get me wrong. But every hot deal I’ve seen in my limited portfolio, where it’s not the hottest name to do the hottest round, there’s just as much extra stuff as you want to win the deal. All the extra terms you could put in to win, they just don’t care.

“What’s everything I could put into this term sheet so that I win?” Everything. Some of it’s great, and some of it is maybe not so great.

Yeah. Sophisticated buyers. What can you do? My a-ha is that the prize does go to the companies that can evolve the quickest. And you’re right. My memory did serve me correctly. Thank you for confirming it. It was just a CX story, like, 17 or 15 months ago, and they’ve just evolved quickly.

And in this market, the people who are making the money are the people who are just running fastest and evolving quickest. The payoff from 2 years of grind—that extra 10% of grind—can have just a massive payoff in a world where fortunes are being made in 12 and 24 months.

And I think the hard thing for founders and for others is, do you stick with something? Now, Wonderful—and maybe I’m saying this more as a tilt than it was—for Wonderful, they did it internally, right? The founders got together and evolved the company very rapidly into something much more successful.

On the other hand, we see Airtable, where I’m going to assume Howie sat around the table. The only thing that really makes sense in this weird deal is that he sat around and said, “I just can’t do HyperAgent in Airtable. There are too many institutional headaches, too many customers to deal with, too many grouchy investors who invested at $12 billion. I’ve tried.”

And I assume you are, Rory, too—we’re huge fans of sticking it out because it’s proven to work. It worked at Palantir. It’s worked at so many startups we invest in. We have so many stories. These are all the great Ho Nam stories of sticking it out, right? They’re so inspiring. From Altos, he’s so good at that.

But these days, you’ve got to wonder: should you stick it out? Is it worth it to stick it out, guys? I want to tell everyone to, but I almost challenge myself. Maybe you should abandon that $50 million, $100 million, or $200 million in revenue. Maybe you should do whatever it takes. Maybe you’ve got to be as fast as Wonderful, or what’s the point?

I actually don’t think you should always stick it out. I think circumstances are different. I know years ago, when I had my own small business in the UK, I look back and it’s just so clear to me: I stuck at it for 4 years. I knew everything I needed after the first year. I shouldn’t have bothered for 3 more years. Waste of time.

So you don’t always stick it out. Is there a plan, or are you just doing it out of misguided loyalty? That’s the number one test. And I think you’re right, Jason, because I don’t think Wonderful was a pivot as much as an expansion.

Rapid expansion, yeah.

You start here and just expand. Whereas I do think Airtable was in that contraction mode. Everything they had, they had run out of time and space. So I think it probably made more sense to do that sale in that case. The facts are different.

I think everyone in this category is being forced to, though, by Brett Taylor, who’s being very clear in terms of his expansion, and I think they’re following suit. I think they need to follow suit as well to justify the prices they’re raising. So the combination of following Bret and price hunger means we’re all doing the same.

We’re doing the operating system. You know, Owner is no longer just for restaurants; it’s the operating system.

Crudely put, I mean, you’re exactly right: Salesforce as a company is the dominant SaaS company. It’s worth roughly $200 billion—$180 billion. I think Service Cloud is 25% of that, so the winner in the existing world is only worth $50 billion.

As you get these bigger market caps, like Sierra, you have to go beyond Service Cloud replacement to be a big company. You’re exactly right, Howie. You have to sell, “I’m the whole operating system. I take it all.”

Which means if you’re Sierra, just to make the obvious point, you’re coming right at your former company, right? You’re saying, “We want all your market cap, Mr. Salesforce,” because the only way I can justify $15 billion or $20 billion in market cap for Sierra is not if I build a slightly better, next-generation Service Cloud. It’s if I am the entire customer ecosystem for your entire business. Go team.

And then you’re right: everyone else, like Wonderful, has to follow.

Yeah.

When someone goes risk-on, everyone goes risk-on.

Welcome to venture, baby.

Absolutely. Terrifying.

11. New Labs Face Capital Pressure

What about Thinking Machines? A $40 billion new price. It’s down from the $50 billion last year. The round is $5 billion to $6 billion, with Accel leading and NVIDIA doing half. A couple hundred million bucks in revenue. It’s the closest thing to a U.S. model provider—

I think that’s the real sentence here. The important thing is, you have to start with: what’s the company doing? Why is it better differentiated? And yes, they’ve shipped 2 products: Tinker and Inky. Cute names, right?

One of them is an open-weight model that they themselves say is not pure frontier-grade, but is open-weight and U.S.-based. That’s worth a lot in this world. And then, on top of that, I think the other product, Inky, is a platform to allow enterprises to do their own training.

So the idea here is that now you can go to JPMorgan, you can go to BofA, and say, “You’ve got an all-American software product, and you’ve got the ability to train it on your data in a totally proprietary way that’s not exposed to OpenAI or Anthropic. In fact, you have full reinforcement learning—all the things you want to build a state-of-the-art enterprise model for JPMorgan, for whomever, Procter & Gamble.”

So it’s a pretty decent, compelling offering for corporations. That’s the positive story, and it’s interesting that NVIDIA is doing so much, because to some extent I thought that’s what Poolside did, and they just acquired Poolside.

So what you’re seeing is NVIDIA saying, “Anyone who’s doing something interesting in corporate AI, we’re going to put money in.” I think that’s what’s happening here.

I know we’ve already forgotten about it because it’s been a week or two, but the Poolside thing should be a little bit haunting. Not that a nine, $7 billion exit is so terrible, even though the 15x thing kind of was one of Harry’s clips that he took.

But that memo was chilling. It was like, “We couldn’t raise the round.” This was a great team with proven leadership—a very strong CTO, very strong leadership. They seemed to have had the right vision from day 1, went for it, and they just couldn’t raise the capital they needed to execute.

And congrats to Thinking Machines, which in some ways appears to have less. But it is a reminder that the music will end for a lot of the new labs, right? And so be it, as it should. It should be a thinning of the herd.

But the Poolside one—we may never talk about it again. Or it’s possible that, when this little part of the bubble bursts—just one part of all of AI, this Neo lab bubble—this will be looked back on as the moment those guys grabbed the exit, when NVIDIA stopped buying everything and Anthropic said, “Enough,” like after Descartes, they said, “This stuff doesn’t actually move the needle. Enough already with the infinite $10 billion deals. We’ve had enough.”

Maybe we look back and they were the lucky guys who got the deal done on December 2021.

Yeah, and just to be clear, to remind everyone: Poolside effectively sold. They would deny that they sold, but they sold a license for their product to NVIDIA, which was an open-weight, enterprise, U.S.-focused model. The company still exists, but they cashed out quite a lot.

And the memo that Jason is referencing is the note they wrote at the time, basically saying, “We were right, but we couldn’t access enough capital to continue to play.” And it’s interesting that literally 2 or 3 weeks later, another company in a not-dissimilar business is actually, it sounds like, able to access that capital, in part, ironically, from NVIDIA, which was also willing to buy Poolside.

So they get to play out the hand. And I think you’re right, Jason. Will you look back and go, “2 years from now, you could look back and go, ‘Poor Poolside, they got sold out, and Thinking Machines made a 4x from here’”?

Or there’s another world where you look back and go, “Oh my God, I wish I’d sold,” because the opportunity got tougher, and Thinking Machines and Poolside maybe, as you said, were the last exits out.

Yeah. The world just changed after Astra. We'll be talking about that in 24 months, and that was the end of the need for new labs. It was hard to see at the time, but when Astra 6 came out, the world changed and it became this and us versus China. And then the new AI regulatory council, the Trump regulatory council, came in and changed things again.

All those things could happen, though I do believe, fundamentally, you do believe that there is going to be strong demand from corporate America for an open-weight, U.S.-based model with the infrastructure to train that model, right? And I think Thinking Machines is in a good position to meet that demand now, and I just think companies are going to want it. Because you really only have them, with Flexion, which I don't know where they are in terms of their model, and Poolside, but I'm sure there are others and they'll all come out of the woodwork and flog me for not mentioning them. But there's clearly a massive market need there.

There are a couple more, but there aren't many more. It's a constrained group of five or six players.

For precisely the reason that Poolside outlined.

But the cash has dried up at scale for the Neo-lab players, who, I think, now are going to your core. As we've seen, it no longer becomes a venture play.

Yeah, I mean, 2 big companies, 2 foundation models, were Neo labs themselves 5 years ago, and they've turned out to be the best venture bets of all time. Just because that's true doesn't mean the other 100 Neo-lab bets that you can bet on today will also turn out to be amazing venture bets, because now you have other companies with the capital already, and you have other companies with the distribution. And, yeah, the question is, which of those new-lab bets will be orthogonal enough to the foundation model companies to be able to be an interesting bet? And we're wrestling with that question every day, because obviously you'd like to make those bets, but if you're doing something that's going to get either rolled over because the foundation model companies do it or, as Jason says, if you can't raise the capital to play the game, it gets hard.

Boys, is there anything that I've missed that we should discuss?

I don't know if it's true or not, but Twitter is saying, is Anthropic going to drop its S-1 today, or is that not the case? I don't know. But, yes, that would be interesting, to say the least. That will be heavily downloaded and read within the first hour of coming out.

My word. Will you be buying at 2 trillion, Rory?

I'm probably not going to be buying at 2 trillion, Harry, but that's not a comment on the stock. Actually, the answer is, in the end, yes: I'm in S&P and QQQ. As my wife said when SpaceX went out, “Looks like we got some of that from Elon, too,” right? If it's in the index, baby, it's coming your way. Not as quickly on SPY as QQQ, but on your Nasdaq index, that stock is going to be in your hands, I think, 12 days after the IPO. So you're a buyer. Boys.