Did OpenAI Steal the Navier-Stokes Solution? | E2335

NYU mathematician Tristan Buckmaster put out a statement saying OpenAI only started prompting its model on this problem after word of his own work reached the company.

Not surprised at all. You look at the YouTubes and the Dropboxes and the like, right? I mean, it's a common trend.

Trust these companies with your data at your own peril to some degree.

And in the scientific community, right, people are very conscious about what they share with who because they want to control who builds on top of their work. And if you start doing that work in ChatGPT Claude, there is a possibility that it makes the next researcher Find that direction faster. Agents are scraping the entire web looking for partial solutions and can build on it. So the importance of like locking in your research becomes increasingly stronger.

I would not trust them with anything that's important or proprietary. This idea that you're going to give Instinct or Zuckerberg your Gmail, then your Notion, then your docs, then your databases means they're going to steal your IP, train it, and then give it to your competitors.

This Week in Startups is brought to you by Vanta. Compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy for companies to get a SOC two report fast. Get one thousand dollars off for a limited time at www.vanta.com/twist and NetSuite. For the first time ever, you can try NetSuite Next for free. If your revenues are at least in the seven figures, go to netsuite.ai/twist. Built for every industry, ready for every boardroom. netsuite.ai/twist. Northwest registered agent. Got a new business idea? Northwest Registered Agent helps you bring it to life. Get a free domain, email, phone number, and more with no purchase required. Learn more at www.northwestregisteredagent.com/twistdomain.

All right everybody, welcome back. It's This Week in Startups. We do our This Week in VC roundtable every Wednesday. What do we do on this roundtable? Well, Lon reads the news, I do a bit of moderation there, and we have 3 great guests who are active investors in startup companies, and we've got quite a lineup today. Lon, introduce our guests, please.

Yeah, well, first up we've got Yohi Nakajima. He is the general partner at Untapped Capital, and he also runs Agent Fund, a 2025 AngelList rolling fund for autonomous agent startups. Plus Ben Lierer. Oh, he's the co-founder of Lierer.

That's a major firm. What a get.

See, firm he started in 2021. Finally, we've got Rebecca Lynn. She's the co-founder of Canvas Prime and a 5-time Ford Midas List, uh, entrant.

There you go. There you go.

Welcome, everybody.

Jason had no idea that I was on. He's just like, oh my God, who? Like, wow.

I was hamming it up a bit.

I mean, come on.

A little hamming.

Then showed up. Amazing.

I mean, you know, whatever.

Listen, uh, we have got a crazy docket. And, Lan, I guess we have to talk about the end of the world and how we're all going to manage that.

Sure.

Because, you know, typically our funds last 10 to 15 years. But right now I'm trying to figure out how to get DPI if in fact the world is ending in 10 years or 2030, according to Anthropic. So let's discuss this for the 50th time this year. Anthropic people quitting. And OpenAI folks have done it as well. But let's discuss people quitting because they saw into the black AI box and decided it's the end of days and we're all going to die.

But he's just going to start a hedge fund or a venture fund, right, Jason? I mean, isn't that what's coming next? I expect that to be the next announcement.

That would be.

That would be. Laad, tee it up for us. Who is this lunatic? And then his coworker lunatic who decided he would retweet it.

So this is Jacob Coxon. He's a 27-year-old pre-training researcher, spent 3 years at OpenAI, credited as a core contributor to GPT-4o before moving over to Anthropic in July. He joined Anthropic specifically because of its safety reputation, and about 2 months later, he quit the company and the industry entirely in a 7-post thread on X that already has tens of millions of views. He told The Wall Street Journal that researchers inside these labs have started using words like crunch time and endgame, and that by the end of next year, some of the most aggressive scenarios could already be out of control. Then it got stranger because 2 current Anthropic researchers publicly agreed with him, including alignment scientist lead Evan Hubinger, who said that he personally puts the odds of AI killing all humans in the next decade around 10%, give or take. 10% chance we all die. And all this is landing while Anthropic, of course, reportedly heading toward an IPO at something like a $2 trillion valuation, with its safety-first reputation as a central part of the pitch to investors. Uh, Yohi, what are you making of this latest instance of doomerism?

I mean, it's gonna continue, right? We're gonna continue to see incidents like the Hugging Face. There are gonna be things where we see a hint of AI doing something bad and we're gonna see what, what if this happens at an exponential scale? The conversation is not going to end, but the conversation is important. The conversation itself is going to hopefully prevent any likelihood of that happening because we're going to put things in place. So I would think that part of the reason that you see this really publicized is Anthropic talking about this a lot and researchers talking about this with the hope that this triggers conversations that changes, that makes sure that people are building in a safe way. I think some of the doomer people might actually be speaking more doomish on purpose just to get people to push in that safety direction more than if they had honestly said their practical, like, oh, you know, it's going to be a, you know, 100 years or something. But if you say 10 years, people are going to lock down on safety more. So that's my kind of take on it.

I'm going to go last because I have spoken on this so many times. Ben, what does this mean for our portfolio?

Well, first of all, one thing I think is pretty funny—

Do you believe this stuff is so powerful that it's going to kill us all?

I actually think it's pretty funny that there's 92 million views on this. Clearly Elon is like, like pinning this to the top of X for all users in all countries. I mean, it's like, look, I think that every month I become more astounded by the speed that these models get better and what you can do with them, um, in ways that I'm excited by as an investor. But like, if we keep ending up with these half-lives getting tighter and tighter, Yeah, you know, it's like, it is creepy to, to think about how powerful they're becoming. And it's obvious that we do not have, you know, sort of a government that knows how to deal with any of this stuff. You know, I think it's, it's on the one hand, I think it's hysterical. On the other hand, I do worry about, you know, it making its way into national security and just like the state of the world a little bit. It freaks me out, to be honest. This one doesn't mean anything, although I do think you've had a lot of people complaining about OpenAI and leaving. I think there's been fewer people who have sort of raised the red flag on Anthropic, and they've done a great job, like, with the messaging around that. And, you know, Dario got very out in front of it very early. But, like, we're dealing with some pretty unbelievable technology that's moving much faster than I think Like the average person should be comfortable with.

I personally look at this and it, it, it's, you know, to me it's a lot, it's clickbait. I mean, this stuff is, um, AI is incredible. Like what it can do. When we first saw, you know, GPT-4, I saw it in August of '22, right? It was incredible. And just what it can do now, you know, versus then. And I think there's just a natural tendency in society to be afraid. And what we should be, I think, is more excited and, and yes, having this conversation, but Yeah, I did nuclear engineering in my old life. And, you know, if you recall when we were growing up, and Ben, I'll put you in the we bucket, right? And Jason, you know, the Russians were gonna end the world and we were all gonna have, you know, it was nuclear was like the fear. And then if you read back, and we were gonna talk about AV, autonomous vehicles later, but it's really, I always love to read historical fiction and like look back at what people were thinking, you know, at different points in time. And when, you know, cars first came out, much less, you know, ones that drive themselves, people really thought they were gonna kill everyone. And so I think this, I mean, I'm super pragmatic, but I think it's a natural reaction to change that we have. And what we should be doing is having the, the positive conversation about, you know, how it can help us live longer and, and lead happier lives. And, and I think AI can ex— can do exactly that. And so it's, it, for me, it should be incredibly exciting.

Yeah. This is like, it's super interesting to be here now because in 2012, '13, and '14, I had 2 neighbors who I was independently friends with, and we would go have dinner in Brentwood. One was a philosopher neuroscientist named Sam Harris, who hadn't started his blog yet, but asked me, like, how do I start a blog? And I was like, well, you get a microphone and a guest. He said, what else? I said, that's basically it. And then Elon Musk, who lived in Bel Air— I lived in Brentwood— we'd go to Popone's, this great Italian restaurant. We'd sit there and I'd watch these 2 geniuses talk. And I'd ask some questions about it. And the one thing we talked about over and over and over again, uh, was AI. And this was, like I said, '22, '23, 2014. And, um, the— what was the book, um, that came out right along that time that got every, uh, everybody super hand-wringing about the superintelligence? Maybe, um, Nicholas— whatever.

The Bostrom.

Yeah, Nick Bostrom's book had kind of been making things. And then there was this crazy a conference in Puerto Rico by the Future of Life Institute conference. You can pull up the picture of, um, everybody at this conference, but this is where it all coalesced, uh, just for a brief history lesson here. And there's like a really famous photo of everybody there. Somebody needs to go back here and just draw the lines on it because this was the famous conference where Elon— and I have a point here— Elon, you know, after conversations with Sam Harris, decided he would put like $10 million towards this, funded OpenAI, Credit to the folks at this event. They saw that there was some downside here, and they started mobilizing, and they did exactly what the industry needed. Hey, is there a nonprofit who could work on this and maybe keep an eye out for it? That nonprofit that was supposed to keep an eye on this issue was OpenAI. Obviously, famously, they went to become a for-profit company, and we know how that story went. Um, but these folks are now the worst possible spokespeople for AI. 80% of Americans hate this. And if this keeps up— and I, I don't even think you can reverse it now. I don't think there's a way to reverse the negative sentiment in America. They're going to block every AI data center. They're going to burn every Waymo in the streets. And they're going to ban AI in schools. Like, you know, New York is— I wouldn't exactly call it a ban, but they're You know, kind of banning it or limiting it. And I have some, I think there's some positive aspects to limiting it to make sure it's being used properly, not to cheat or whatever.

Regular listeners already know that if you've got a great idea for a new business, our friends at Northwest Registered Agent wanna help you bring it to life. They're gonna be the most amazing partner you've ever had. Even if you're not ready to form an LLC, you still need to take care of some basics. So Northwest Registered Agent is now offering free identity services. That means a free domain name, open-source website hosting, a business email and a phone number, and everything else that's going to make your new startup look and feel like a professional company, all with no purchase required. And you know you can rely on Northwest because they've been helping people like you start businesses for nearly 30 years. If you have an idea you can't get out of your head, or even if you're already building something amazing, you already got started, Northwest Registered Agent is the Best way to establish your new company. Learn more at northwestregisteredagent.com/twistdomain.

Congratulations to the team at Anthropic and OpenAI. You have succeeded in scaring the bejesus out of Americans. They're going to burn Waymos in the street. They're going to burn data centers, and there's no way to stop it now. Congratulations! You're going to have a huge IPO because who in their right mind would not buy the company that's you know got a 90% chance. Of being the last company in the world, you're kind of obligated to buy it if it's— there's a chance that it's the truth, and, uh, a 10% chance of the world exploding, um, well then money has no purpose. So congratulations on a PR strategy that is either brilliant or insane, and you're all suffering from psychosis. I literally cannot figure these people out.

Well, this is an interesting one because is it a conspiracy theory? Because this guy says he's quitting, so So he wouldn't be doing this to boost Anthropic stock unless there's some kind of deal they worked out with him. So is this— that where we're going with this? Like, are we actually going through the looking glass?

This is the most liquid stock in the world, Anthropic, right now. Like, it's more liquid than like buying Microsoft or Amazon on Robinhood. You can find it literally in the 100 milliseconds it takes to trade a stock on Robinhood. I think in 10 milliseconds you can find a buyer for Anthropic. I'm being facetious, but Ben's shaking his head in agreement. It's like super liquid. This person probably cleared their position ahead of the chaos.

Jason, I think you made a good point bringing up the—

Or not. I mean, or yeah, I mean, it doesn't—

Or he left a double up, like maybe it's going to triple.

Yeah, effect— it has no effect on like— it's, it's a, it's a runaway train. It's, it's unstoppable. I mean, it really is. I think you're right also that the negative sentiment is, is unstoppable, but, but that's— there's like a bunch of different downsides here. You know, when I talk about it ruining the world, it's like, are we going to give, you know, our defense systems over to AI? That's where, like, you actually have, like, real fear for the end of the world. My hope is that that is not—

You understand that Trump has, like, a little red button that he can just push, right?

I think there's some people around him who may—

And, like, what do you mean?

A special power that, like—

Oh, I hope not.

No, they literally in the first administration—

In the first administration, yes.

Yeah, they made like a—

the adults in the room like, disconnect his button.

They were like, literally, just— yeah, they literally pulled the red wire off the button and they're like, yeah, let him click it 3 times and tell him it's got a 60-second delay, but it's, it's definitely gonna fire off the things.

Uh, it turned on the gorilla channel on TV.

Yeah, the reason— look, the reason that people— I mean, I think like the, you know, the masses don't like AI is because realistically it is going to take a lot of their jobs. And we don't have— by the way, I believe it's going to create tons of new jobs also. We invest in AI all day long. I'm not like a doomer here, but it's going to get worse before it gets better. There's going to be a ton of industries that get wiped out. You see this right now with kids graduating from college. There are no jobs. There's very few entry-level jobs. The wealth divide is going to grow. If you were in the right companies and you have access to the private markets, then you're going to create some kind of wealth. And if not, Like, you're going to have— you're going to be on the wrong side of that divide. Like, we're— and, you know, we don't have particularly stable leadership worldwide for a bunch of reasons. And, like, it's a tough backdrop. And, you know, I don't think AI is going to destroy the world, but it's going to make a lot of people's lives much harder before, like, we get the full benefit of, like, you know, the largesse that we're all promised.

Yes, the abundance is not here yet. Yohi, What do you think here? You, you, you, you and you, you and Ben raced in with their—

No, no, I, I was just gonna say like the, it's, it's, you brought up the 2015 doomer thing. I was gonna say like, it's important to understand the context that this doomer thing is not new, right? They've, they've been around forever. They're just getting attention finally because AI's getting attention and negativity just gets more attention, right? I was actually think talking about AI with a whole bunch of people and I asked a whole bunch of people, was social media net positive or net negative? And overwhelmingly people immediately jumped to the negative. People aren't connecting with each other. But then as soon as you start bringing up some of the positives, like this show runs on social media, there's been people who didn't have a voice who— and then once you start talking about those, people are like, oh yeah, there's been positive as well. And I think to some extent the same thing's happening in AI where, you know, we are talking about people not having jobs and whatnot, but at the same time, I've also talked to a lot of people who I would think wouldn't hate AI that, that love GrokBot. They just can't stop talking about Super Grok. And there is AI impacting people in positive ways today as well.

I can't agree more. I mean, AI is helping people grow their businesses, right? In, in incredible ways. Even, even like the offline businesses. And so yes, there'll be change, right? But I think people fundamentally, like this doomsday headline is what sells. It's what gets clicks, right? It's what gets everyone engaged. But when you look at even the advances in, in healthcare that people are, are experiencing, even on the micro level of being able to go and research your own conditions and really untangle what might be going on. Uh, you know, AI is not taking away the need for doctors. If anything, it's pushing more people in because they want to go and, and discuss these things with their doctors, for example. Right. Um, I, I, I see it as the net negative. And yes, some jobs are going to go away, but I think a lot of people are reporting, you know, busier than ever, largely due to how people are, the consumer is leveraging AI.

That's certainly been my discovery. I was very concerned about office workers. Uh, and in my organization, I scared the bejesus outta everybody. I just said, if you don't use AI as your first attempt to solve problems, and when you have a question, if you don't ask 2 or 3 different AIs how to solve that problem, if you're not doing that, you can't work at the company.

Yeah.

And I saw like us go from 20% of people logging into Claude to like 50%, then 60%, then 70%. And then we have like a monthly bonus system inside of our venture firm. And the same people kept winning each month. Like, in fact, they won 10 out of 12 and 11 out of 12 months, the top 2 people. They were all using AI, and their impact became 3, 5, 10 times that of the people not using AI. And I think what's started to happen here is people, if your organization Can show people, my Lord, your coworkers are now ten times more effective than you are. Then that should be the kick in the ass that people need to embrace these tools. You're not the famous quote: You're not going to be replaced by AI. You're going to be replaced by somebody using AI is exactly correct. Jensen nailed it because a customer support person, literally a person who just picks up the phone and does customer support, can use these tools to build a system to make themselves ten x. And they don't need any computer training. They just need to know how to talk to Grok, Bot, Claude, Cowork, pick your poison. You could just tell it, how do I get 10 times better at my job? And then here's my job, here's my email box, here's my Slack.

It will do it.

It will literally make you a plan. And I think that's what happened in corporate America. And when that happens, what I'm seeing is more opportunities emerge.

Yeah.

If everybody's 10 times more effective, okay, what else can we do? Can we start another podcast? Can we, you know, run another incubator session? Can we find better companies? Can we help our companies find, you know, downstream funding and introduce them to VCs faster? Can we sort through applications for funding faster? Like, all of those things are occurring. So I, I, the only area I'm actually really concerned about in the short to midterm is Drivers, DoorDashers. Like, I think those people are gonna be—

Yeah.

The first wave where we see millions of people lose their job. If you're a founder who's not using AI to make yourself more efficient and productive, trust me, you've been left behind. But fear not, there is NetSuite Next. You already know NetSuite. They're bringing together your financials, your inventory, HR, CRM, and every other system of record that you need. They're already trusted by 44,000 companies. Now they have NetSuite Next, and that offers the same amazing platform, but with AI running alongside everything. Your guide on the side. Just ask your question in plain English, obviously, into Ask Oracle, and you'll get a real answer pulled from your company's live data. Plus, they've got an interactive canvas for automating your workflows, monitoring your real-time data, and managing your AI agents. Now, for the first time ever, you can try NetSuite Next for free. If your revenues are at least in the 7 figures, go to netsuite.ai/twist. It's built for every industry, ready for every boardroom. netsuite.ai/twist. Thanks, Dave.

Well, and I think, I think drivers, like truck drivers, I think are the first ones actually. But, but to your point about making people more efficient, I mean, this is exactly our thesis. And to plug a company, sorry, Jason, but Savvy Wealth announced—

No, you're not plugging anything too loud.

Okay, good. Savvy Wealth announced, you know, their $100 million round today. And what they're doing for advisors is exactly that. People thought, well, wealth managers are gonna go away. I've never believed that. I think people crave that personal one-on-one attention, but these wealth managers are made, you know, 10 times better. And, you know, the, the numbers they're giving us that we have, you know, have checked is they're getting 19 hours of time back in their week every week to do what they want to do most, which is, you know, serve clients. And so by doing that, they're actually increasing their AUM 3 times faster, the assets under management that they have, because they're able to take those 19 hours that they're getting back by using AI in a really smart way. And put it back into their, their relationships actually with their clients. So that's exactly the thesis.

I think this, that, like you said, Jason, 100% of people, and they, that means they can charge, they could take their 1.5 that they're charging and bring it down to 1 or 50 bps. If they have 3 times as many customers, then they compete in the marketplace and say, well, I have 3 times as much assets under management.

Yeah.

I don't need to take 1.5 of your 4.5 returns. I'll make it 1%, which is what Wealthfront did. They used technology to undercut the advisors and bring it down. And, and they had great success with that. So that ultimately benefits the consumer and those financial advisors. They both do their job better. Yes.

They both do.

And they have more, they have more time back. The same thing with lawyers. We did case tax, right? Same thing with lawyers. Everyone's worried lawyers are going to go away, but you know, lawyers actually spend so much time on these mundane tasks when they really wanna spend time with a client with strategy and very different, very different job for them.

The, the ma— the makeup of jobs that people do has changed significantly over the years, right? If you look at today versus 10 or 20 years ago, and, and I think it's obvious that it's going to change over the next 10 or 20 years. The question is how, and I, I don't have an answer.

I'm just throwing, throwing things out there. Well, drivers I think are one, right? You know, Jason, I did, uh, Luminar years ago, uh, and it's since went public and, and all of that. But I thought when I did Luminar, my son who turned 16 this year wouldn't be driving. And so it's actually been slower than what I thought it would be in terms of the uptake of autonomous driving. And I really truly told my son when I made that investment that he'd never have to get a driver's license. And, uh, he has a driver's license, but, you know, also likes Waymo. But it, it's taken a bit longer actually than I thought it would take.

Yeah.

I mean, 10 years longer than I think everybody anticipated. People thought we would be there 10 years ago.

Elon.

Yeah.

Waymo, et cetera.

Yeah.

Well, you know, there's, I think the reason for that is the expectation, Ben. Like, people's expectation is if you're going to have self-driving cars, you take out the steering wheel, nobody can die. That is the public's— it has to be nobody goes to the hospital, no broken bones, right? No deaths. That's a crazy benchmark for a couple of thousand pound vehicle going 65 miles an hour now that they're on the highways. That's the reason this is a slow rollout. They can only be perfect.

That's why, by the way, we're early in Zipline. And I'm sure you saw last week national news that an Amazon drone dropped a package in a swimming pool. And it's like, you know, there are tens of thousands of deliveries happening every day now, and one package is dropped in a pool, and it's like, oh my God, drones.

The end of the world.

Silly joke. And it's like, you know, we're focused on On the wrong things. Oh yeah. But that's—

So then I have to ask you, because I wanted to do this series, it was either the A or B in Zipline, but did you invest when it was a toy company?

We invested in the pre-seed when it was—

It was a toy company.

It was, it was, it was, we invested in Keller as like person building robotics and his first, the first product was a toy. Yep.

That's so cool. Yeah, it was like a little robot or whatever. It wasn't even a drone. That's what I had seen.

It was called, the first product was called Remotive.

Yeah, I'd seen it when they were doing the drops over for medical and things like that and loved the technology of not having to land. And so I'll tell this, it's a funny story. So it was an investor who shall not be named. I was actually talking with the Zipline crew about potentially letting them fly the drone on my ranch because they needed a big area of flight to fly. And I knew a very early, early investor. And so I was having, um, breakfast with this, uh, not to be named investor and said, hey, I love this company, Zipline. It was in his portfolio. And he is like, oh, you know, we invested in something totally different. It was a toy company. We'd never do the deal today. And it was fairly negative on it. And I walked away just deflated, right? Because I really loved the deal. Uh, said investor then led the round that was announced 2 weeks later entirely, right? So I learned a very hard lesson early in my venture career from that experience, but really love Zipline and all the innovation they've managed to do there.

I too missed it. I met Keller when he was delivering blood on fixed wings. And I said, hardware is really hard. I had him on this very podcast. I said, I would love to donate to this, but I don't think I can put my LP's money in it. Then like last year, he's like, you know, you missed the investment. I was like, yeah, I know. I feel like an idiot. I didn't know you were going to deliver burritos. He's like, no, neither did I.

That was not— they were like, they had like a UN contract or something to deliver like medical aid in Africa, right?

Which is literally in my notes. I was like, slow sales cycle. Like the UN is not the best company.

It's okay. I sat with Kalanick for 3 hours one day as he was discussing taking over the CEO spot. bought an Uber. And I was like, and I didn't do that either. So that's like—

Funny story. When he told me somebody else was going to be CEO, I said, if you do that, I'm going to not invest. You have to be the CEO. And he was like, do you think? And I had this really long conversation with him. And to this day, Ryan, who was going to be the CEO, is like, you dissed me so hard. And I was like, I didn't.

You're wise.

I saved you. I helped you.

You're Steve Jobs.

We're talking about Steve Jobs here, or like Elon Musk or Bill Gates. Like, that was in all of our best interests. I shouldn't be on the management team. I shouldn't be the CEO. I would be amazing. So anyway, long story short, uh, I have now a backstop if I miss an investment like Zipline. I have thesyndicate.com. I've got 80 family offices in our tight little circle.

Nice.

And I say, hey, I found this incredible company, Zipline. I missed the investment, and I would love to have you join me in putting in $5, $10, $20 million. And I have a personal in with the CEO. We're friends, and he's been on the pod, and my CFO knows their CFO, and we boom, just— yeah, this is when I had him.

Nice.

I literally talked to him at this Launch Festival about making up for my missed investment. But anyway, we don't have to play this. By the way, guys, if If you didn't know, here's Steve Jobs if he worked at Anthropic and launched the iPhone as an Anthropic employee, as you can see here. The iPhone increases anxiety, disrupts your sleep, kills democracy, creates misinformation, eating disorders for children.

It turns out it wasn't the iPhone, it was actually just social media that did all of those things.

Yes.

Anyway.

Fair enough.

Access to information.

The App Store did it. Yeah, I should make it the launch of the App Store. We're now gonna partner with people who are gonna destroy our lives. I mean, it's crazy.

But he had to do it before someone else did, Jason. If, if we imagine if somebody else had the iPhone had fallen into their hands.

Yes, it, it is.

That's the argument for the nuclear bomb, isn't it?

Here's a lesson for all my founders out there. Trust closes deals. You can have an amazing, beautiful product. It solves every problem your customers have, but If those same customers who love your product don't believe it's totally secure, they're out. And that's why you need a partner like Vanta. They're the number one security and trust platform in the world, and that's going to get you compliant fast for all the major frameworks from SOC 2 to HIPAA and beyond. Plus, Vanta is going to make sure you stay compliant by continuously monitoring your controls. And now you can interact with Vanta's agents wherever you're working. Even if your team is knee-deep in Claude or Cursor, it doesn't matter. Having Vanta as your partner means your deals keep moving and your engineers can focus on building that next great feature, not tedious documentation. That's why Vanta is already trusted by more than 16,000 companies, including Ramp, Ryder, Harvey. You've heard of them. So prove to your customers that you're ready for business. Learn more at vanta.com/twist and get $1,000 off. We literally had this discussion the other day. You know, Yeah. There was an internal debate of if they should have done a demonstration, a public demonstration to the people of Japan, you know, like drop a bomb, you know, I don't know, 10 miles offshore, or, hey, clear out this city.

20, 30.

Whatever it is.

Just get it over with.

Clear out this island, atoll, you know, and we're going to drop a bomb and you're going to have to really rethink your, you know, commitment to this World War II kind of thing. And they got over, uh, they got vetoed because, um, they felt it wouldn't work. And I think that's like maybe a major error in the history of, uh, computing, the history of technology, like in technologists here. But anyway, let's keep going.

Glad, glad Jobs did it first for the iPhone.

Yes.

Thank goodness. Thank goodness he got there first. Before we go on, Jason, we should mention I, I'm wearing my Plaud pin right here. You can see it on my jacket.

Oh, oh, cause for the cause.

We do pause for the cause. We do applaud Plaud, our friends over there. If your work depends on conversations or interviews or meetings or calls, we're recording this show right now, by the way, on Plaud. We will upload the Plaud transcript and summary afterwards in the description so you can read it yourself. Uh, a great way to keep up with everything that you're doing, everything that you're listening to, all your conversations, just get yourself a Plaud pen. This is the Plaud Note S.

I have it too. And here's the thing. It has a little red light. So if you're gonna use it, you don't have to worry about privacy, whatever. You just ask people, hey, can I record this meeting? And I'll send you the notes. It does great notes. I take it when I go on hikes. All right, let's keep going.

So check out Plaid at plaid.ai/twist. Use the code TWIST for 10% off.

All right, there we go.

Move on with the show.

Okay.

Sorry. What's up next?

Yesterday, Meta shipped an agent called Muse that they've been testing for teeing up. Since Zuckerberg's infamous 6,000-word manifesto last month. Muse is an AI agent tailored toward the everyday person. You message it just like you'd message another person in its own app or in WhatsApp, and it goes and does work for you— booking travel, filling out forms, turning a recipe into a grocery list. You get it, the usual stuff. Stripe has partnered up with Meta to develop the payment feature for the Muse agent, and it's free up to 100 million tokens a week. with $20 and $100 tiers above that. No advertising in Muse at all so far. The whole thing rests on trust. This is, of course, the company that just agreed to an $18 billion settlement over social media harms less than 2 weeks ago. They're now asking people to hand an agent their inbox, their calendar, their credit card. So we'll go to you, Ben. Is Meta going to be the company that finally gets everyday consumers to trust an AI agent with every aspect of their lives?

It's a funny question. On the one hand, Meta already has all of my data. It couldn't have more. And so I guess what's the harm of throwing in another thing? Yeah, well, I gave up years ago. I think that Meta's got a big trust problem with— I'm actually not sure how deep the trust problem goes society-wide. I know within sort of tech circles, there's lots of snickering about Meta being very behind in 100 ways, but obviously there's 2 billion plus people using their products every day. Distribution is— that's a huge distribution advantage. I'm sure they're going to get a lot of people. I mean, by the way, you know how many people are using Threads?

It's close to the amount of X, I think, right?

I mean, Meta can take any crap and jam it down the throat of hundreds of millions of people. I haven't used Muse yet, and obviously there's Instinct in town and a lot of companies like playing around this space right now and that are, you know, there's going to be a lot of capital thrown at it. But, you know, it's hard to— on the one hand, it's sort of hard to bet against Meta getting something real going. On the other hand, traditionally they've had to buy all their big successes. And so, you know, it's, it's, it's sort of hard for me to imagine them like winning this with a homegrown product because they, they haven't done that since like the blue app.

Yeah, they, they're good at Stealing. I'll go with Ben on this. They're very good at stealing people's innovations, like features, and putting them into their core product set. But Instagram, WhatsApp were purchases. Obviously, Facebook was the original thought. This is feels like they're just ripping off OpenClaw, Hermes, and most of all GrokBot in making the most simple, easiest way for you to communicate with an agent, and then giving it to. our cousins, our moms, our uncles who maybe aren't super tech savvy. And in that regard, I think it's gonna do very well if it's integrated into Instagram and they give prompts. So the reason Threads, in terms of the ram and jam down your throat concept, even if 4 out of 5 products don't work, if they ram Threads into the main feed over and over and over again, which is what they do, Uh, and then when you're on Instagram, I— my total— the total amount of time I spend on Threads is when I'm on Instagram and I type a snarky funny comment on somebody's Instagram post, there's a share on Threads, which is like completely stupid, but I do it just where there's something on my Threads, I guess.

Um, were those Threads stats that you just put up?

Yeah, we— the— this is thanks to our, our pals at harmonic.ai.

$11 billion of revenue. This year in Threads.

Wow.

These are your Threads stats. 500 million monthly active users in Threads.

Like, I don't know a single person who, since the week that Threads launched, has mentioned Threads.

I've not.

No.

No one.

No.

In our circle, though.

And they're doing, like, the same revenue as all— Threads is doing more revenue than the entire consumer AI category, with the exception of OpenAI.

I will, I will tell you, sometime on Instagram they'll show you like some of your friends who are active on Threads. And I do, there are a few people. I think it's a lot of people who are still active on Facebook. They get sucked in a lot to Threads through some sort of integration. So—

But Zuckerberg admitted on, when he announced Muse, he was on somebody's podcast and they were confronting him about like, oh, you're active on X. And he's like, All the AI people are on X, so I have to meet them where they are if I'm going to talk about AI stuff. So I post my AI stuff to X as well as to Threads. Um, so he kind of made that admission.

But the funny thing is it doesn't really matter because all that revenue is coming from targeting, targeted advertising, right? And the AI people are not the ones buying all that shit. It's like, you know, 50-year-old housewives in the Midwest, right? And so He— I think the joke's kind of on, on Grok in a way, because it's like, you know, he's like, yeah, sure, I'll go talk AI there, but I'm gonna sell all my shit over on Threads, right?

The higher your IQ— Yeah, the higher your IQ and your income, the less likely you are to click on ads. That's the great conundrum, you know, for X. You have high IQ people who are like, that's an ad, I'm being manipulated, scroll, or I'm gonna pay for the premium version so I don't see ads. It's always been their challenge since Evan launched the company.

Uh, Yohi, what are your thoughts here on, uh, well, I think Meta did try to buy their way into this, right? That was probably what Manus was supposed to be until that got shut down. And then I think I heard a rumor that they tried to make an offer on Instinct. I have no idea. I just saw, I just read it this morning. But so I think that Meta was trying to buy their way into it, but you know, none of those fell in and then there was enough open source stuff out there that they could replicate it.

Explain what happened with Manus, Yohi, because I haven't been keeping up on that, but that was the Benchmark-backed Frontier Labs based in China that tried to move to Singapore.

Yeah, it started in China. They had moved it already to Singapore, then grew, then Facebook tried to buy it. But because it was technically started in China, China was able to effectively quote unquote block it. But by that time, the transaction has gone through. So then they had to like try to buy, find buyers for Meta's shares somewhere else. And I think that's where I kind of lost the thread. I don't know if anybody has.

Yeah, China forced Meta to unwind the deal.

Yeah.

Wow.

So they had to go look for someone else to buy that equity that Meta had bought.

And, and they, they were already like, they'd already started integrating the team. So Manus people were already working at Meta and they had to like, sorry, you have to leave now and go back to your old job.

Give us the badge back. We're gonna need that laptop.

Yeah.

Really? It's like, that's very dramatic to lose your job that way.

This is why, you know, I was talking to somebody who is in that region, um, and they were like, If you're a founder in China, you have to move your family immediately to Singapore, like before you get traction, because they will just sanction you and you'll wind up in a reeducation program like Alibaba's CEO, Jack Ma, who is now painting.

Muse is good though. I tried it. I feel like every generation of these agents that comes out takes all the learnings from the last generation and just embeds it. So, I mean, even just playing around with it, like simple stuff like connecting to Gmail on, on bot, on GrokBot and Muse, you can just click a link and it just opens up and you OAuth into it. With ChatGPT in Claude, you still have to like connect your MCP, you have to find the connector. There's the, the chat doesn't just give you a link to click and OAuth into. So there's small things like that that just get like—

And where, and speaking of Gmail, like where the hell is Google in this whole thing? Right?

That's the question.

That's the question. Like where aren't they just gonna come in and just take the table at the end of the day? Right? And just kind of watch all these things happen. I mean, talk about a company that's shitty at building product and having to buy every product they've ever launched, right? Except for maybe the very, very first one. But, um, but that's what they do. And so all these startups are popping up, they're proving the use case, right? That consumers want these things. Aren't they just gonna come take it at the end of the day?

No, I mean, I think they're gonna buy it. Buy it.

Well, they'll buy it. They'll buy something. Now we're in the US.

They'll buy one thing, right?

And, uh, I don't—

I mean, given the market caps of the Mag 7 and that Lina Khan is now, uh, working in New York to destroy my hometown, like, I think there— we're now in a position where, like, well, no, she can apply her incredible, uh, lack of experience, uh, in economics outside of, like, whatever Ivy League school she went to. But she turned off M&A for 4 years.

Yeah.

Yep.

Biden hired her to turn off M&A for 4 years. Now Trump's in. You can think what you want about Trump. He's, he's all about our biggest companies, our national champions getting bigger. And that is extraordinary. Cursor got sold for $60 billion. Andrew Wang's company, the one—

Scale AI.

Scale AI got bought for $10, $11 billion. If this—

$15, $15, $14, $15, whatever.

$14, $15. I mean, these are extraordinary. And if this company, what is it called? Instinct is the new—

Yeah, they're raising $2.5 billion. Yeah.

I mean, you know, it's like Instinct's raising $2.5 billion right now. And it's still like invite-only, pre-monetization.

It's so like, that's just crazy. That's just like a crazy acquisition strategy. Like, must get, must get invited. And then, oh, you're off the waitlist. I mean, it's just such a crazy acquisition. It's just a consumer like mindfuck essentially.

But here's the thing. If you think about it from first principles, um, these companies are worth $2 trillion, $3 trillion. $4 trillion.

Okay.

So 1% of, let's just pick a $3 trillion company, 1% is $30 billion, correct?

Yeah.

So 1% is $30 billion. If you make 10 of these acquisitions and you spend $300 billion on those 10 acquisitions, or if you do, let's say you do 30 for $10 billion each. If one works out, it's going to be worth more than $300 billion, which is exactly the math that Sergey and Larry did when they bought YouTube and Android, and when, uh, Zuck bought Instagram, WhatsApp. It's just an— it doesn't make sense to the business community maybe watching it, or to people who watch CNBC, or, you know, normies. Like, a normie can't understand this, but I think we all can. It's like, it's kind of worth it if only 1 in 10 or 1 in 20 work out.

Like, here we go.

Let me get Yogi in here. I was asking myself, has there been— I would consider this a consumer-ish category startup that was king-made successfully, as in like a lot of money helped a consumer-facing startup really grow? I feel like King— Uber. Oh, okay.

That's that.

Yeah, Uber. But no, but the reason I know it so well is I was an investor in it, which I guess you guys didn't know.

But, um, wait, hang on.

Yeah, it's breaking news story.

It leaked. I had no idea.

It leaked.

It leaked.

Um, no idea, Jason.

Somebody just, I think Wall Street Journal just did an exposé. I was the 3rd or 4th. Um, they were the first to use capital as a weapon, but nobody else has really done it. Yeah, maybe OpenAI and Anthropic are doing it now by buying 90% of compute. So we might have 3 examples that, off the top of my head, where they're using capital as a weapon. They tried to do it with WeWork. That didn't work.

That was—

There's lots of examples of of kingmaking capital. So capital basically flexing that everybody else should stay out of our way and we are the champion of the category.

Examples. What comes to mind?

But we're asking if it works.

Harvey today is using capital.

Harvey is still— I think the jury's still out on Harvey.

By the way, I was in Times Square today. I drove through Times Square and the side of a building 40 stories tall was a Harvey ad. And the ad said Harvey, $15.5 billion valuation, and then listed Sequoia blank blank and listed 20 VCs. And that was the ad.

But doesn't that remind me of San Francisco? What's going on here? Isn't that reminiscent of living in SF in the dot-com? I mean, same type of thing. Money doesn't mean that you've like won. It'll be interesting to see, I think, at the end of the day where it goes, right?

B2B makes sense to me because if you have money, then I can try— I can trust a company with more money because they have more backstop. Consumers, I don't think care. You have the— was it Kiwi or Q-U-I-B-I? There's Clubhouse.

Remember Name back in the day that was quote king-made social network, right? Way long ago.

Here's the thing. If the product category— if you look at all 3 of those, Clubhouse was a bad product category. Ning, bad product category, like groups, who cares? Like it's small. And Clubhouse, small groups, it was also kind of bad. If you look at Quibi, it's actually the right idea, maybe at the wrong time. These shorts in China, and there's a name for it, Lan, you would know. What's the name of these dramatic shorts? There's a category in Hollywood.

Microdramas.

Microdramas, thank you. Microdramas are She printing money. So I would say Jeff Katzenberg got it right, but was 3 years too early. I think you have to look at each of these on a case-by-case basis, because if you kingmake a category and that category sucks, or you're too early, what does it matter?

Right?

Yeah.

You, you actually nailed it on that. When Snapchat was like having its big moment, Hollywood got obsessed with this microdrama idea, and a lot of companies spun up trying to do like, oh, we're gonna do like soap operas, but it'll play on Snapchat. And it was just too early. But now that's the time for, for it.

So, but also the cost to produce. I mean, the difference with, with Quibi was Hollywood talent, Hollywood quality, $100,000 for a 3-minute video. And, you know, these microdramas are made for, you know, $1,000 for an hour or whatever it is. I mean, it's a totally different math equation.

That is very true. All right, let's keep going. Yohi, did you have something you wanted to add there?

I saw you. No, no, no, no. I just think it's a fascinating— I agree. I'm fascinated. how these, how these consumer ones do.

In terms of venture capital, this is about more of these, these king-made consumer companies, because I mean, it's happening all day long in B2B, but it is interesting.

Oh, here's Magic Leap is another one.

Magic Leap.

I just remember Jawbone, Webvan. I remember being so jealous of my friends who got friends and family in Webvan.

And Cosmo and Urban Fetch in New York were the big ones.

Come on, like, Ducero, that was I— that would blow my mind.

I mean, I think there's like spectacular failures, but there are like Webvan, Cosmo, and Urban Fetch were examples of pouring too much capital onto management teams, and they just spent it too fast. If those 3 companies— and Urban Fetch and Cosmo, you probably remember, Ben, because we lived through it, I think.

Yep.

Um, they were selling everything at half price in New York to get Yeah. Market share. And so people in my office— and they had no minimums.

No. We would order a single stick of gum. And they didn't charge you. Like one stick of gum at Cosmo.

And that was the key to Cosmo.

Yes.

And you didn't have any delivery fees.

No.

So they would spend $10 delivering you a pint of ice cream for $2.50 instead of $5. So they would lose $12.50. Then when they went out of business, people were like— and then right as Cosmo was going out of business, I remember talking to the investors because I tried to buy the asset from JPMorgan who owned it and restart it. And they became profitable in the top 3 cities because they added a $10 delivery fee and they stopped discounting. So they literally had figured out DoorDash, but they ran— and when the dot-com bust happened, they ran out of capital. I wonder if you're looking at something like Harvey today, spending on the side of a building, if they have the right business, the right team, but maybe they blow through the capital and they're like, wow, we spent $10 million on advertising that we could have just put into— The product and sales team.

This was Casetext. I mean, Casetext was, you know, launched in 20— launched, you know, with OpenAI and is powering Co-Counsel now for Thomson Reuters. Right.

Mm-hmm.

And Casetext was sitting at like $30, $35 million of ARR and hitting profitability, you know, 3 years ago and not having to spend all that money getting clients. Like their CAC was, was very low. And so your point, Jason, you think is just like managing your acquisition cost and are you really, you know, kingmaking or not? And I'd argue you're not if you don't control your own destiny to some extent. And if you're just like torching—

Managing large amounts of capital, Yohi, is something young people are not good at, inexperienced people. I watched this in New York when I was doing Silicon Alley Reporter, my first magazine, my second magazine. And DoubleClick, Kevin Ryan and Kevin O'Connor raised a second secondary, like after they went public, they cashed up for like $2 billion. market crashed, but they were deploying it very slowly. And then I watched other folks who had raised an equally large amount of money, and they were just plowing through it. And I saw this with Robinhood and Uber up close.

Yeah.

Robinhood was giving away a free stock if you, if you signed up your friend. You got a free stock, they got a free stock. Here's what they knew, because I remember asking Vlad about this, and this is like months before pre-launch. He had told me they were going to do a give-to-get like Uber and Dropbox ad. I said, explain it to me. He goes, well, we're going to give a free share, whatever. I was like, because it's $300 to $500 for E-Trade to get a new customer buying ads on CNBC, right? He's like, yeah, how did you know that? I was like, oh, I worked at AOL. I had sold the company to AOL and we watched that happen up close and personal where they were spending $100, $200, $300 to get somebody to put an AOL CD in their computer. And they just decided they would spend $15 on either side. So they're spending $30. They're spending 10% But they figured out a device that was just super appealing to consumers. If you told the consumers they got $25 each at $50, they wouldn't have done it. But if you told them you get a free share of a company, yeah, or a fraction of a company, they'll do it for $5 to $10 in the free share. It was crazy.

It was, it was, it's really smart CAC. I mean, I've heard Harvey's on the order of $400,000 acquisition cost for every new account. Right. And giving away—

So that's a law firm with an annual spend of what, I guess, is what you're thinking.

But, but giving away a free year and it, it depends, like the law firms Wait, just to touch on the Robinhood, that's also genius because of the retention thing too, right? Because if you get $25, you have no reason to come back to the app. But if you get a stock, you're going to come back to check it.

You're going to come back and check it. I just wanted to touch on that.

That's pretty smart.

Oh yeah, I didn't consider that. You're right.

It actually is a reason.

Your 7-day retention is going to be way higher for somebody who has one share.

But Yoshi, to your point, you have to actually probably set up the account and give them all your information to actually get that one piece of data. stock, right? So that's gold.

We got to talk about this NYU mathematician. He's accusing OpenAI of fighting dirty on the Navier-Stokes solution. Yesterday we covered OpenAI on This Week in AI. We covered OpenAI claiming its internal model had cracked Navier-Stokes, one of the 7 Millennium Prize problems. Today, though, we're getting a different side of the story. NYU mathematician Tristan Buckmaster put out a statement saying OpenAI only started prompting its model on this problem. after word of his own work reached the company, and that a whole team and an insane amount of compute went into it. Buckmaster had spent most of the year on this with Levant Alpoge, a mathematician who works at Anthropic, and he ran the report's drafts through OpenAI's Codex the whole time. So when they asked whether the model had been trained on their sessions, they, they never got a clear answer from OpenAI. Buckmaster also argues that OpenAI twice pushed him to drop Alpoge from the paper, because he works for their competitor. He then threatened to go public and was asked, why would you ruin your career? OpenAI's Sébastien Bubeck calls all of this false and inflammatory and has since apologized for that line and said he retracted it on the call, though OpenAI's write-up does confirm its effort only began on September— September 1st after they heard the rumor this is being worked on. So, uh, Rebecca, we'll go to you. Does a story like this change what you're willing to put into a frontier model? Are you Surprised OpenAI may have stolen or borrowed some of these ideas from a researcher?

Not, not surprised at all. Not surprised at all. But, but impressed with what, you know, the models are able to, uh, to come up with here and how they're pushing, uh, going far beyond, you know, reading your email and summarizing it to actually solving these mathematical equations. But not at all surprised. Um, I mean, and if you look back, I mean, if you look back at like very large companies. And, you know, in terms of, you know, stealing data to get started, right, or to, to build, you know, you look at the YouTubes and the Dropboxes and the like, right? I mean, it's, it's a common trend, I would say. So no, not at all surprised.

Ben, do you think— I mean, if you were an academic or a researcher, are you gonna keep using Codex for this, or are you gonna be like, well, I don't want them to swoop in at the zero hour and publish right at the same time as me and take credit for my work.

Well, we're actually in a stealth company that would solve this problem potentially. So maybe I'll come back on at some point and we can discuss this.

Yeah, there you go.

But look, I— we use— we're a Claude shop here. We have a bunch of information living in Claude, you know, hopefully in secure ways. I, you know, I'd like to think that what they claim about the data privacy and what they're training on and all that is true. You know, we're also in a world where, like, these companies are so big and so powerful, I don't even know how they get policed. I mean, like, Facebook pays $18 billion. It's like, okay, you're just like, as we just talked about, Jason, just feeding tickets.

Facebook can go—

it's feeding ticket. Facebook can go buy 10 companies for $30 billion apiece. as a way to take a, you know, an attempt to see if they can get 1 out of 10 right and win a category. I, you know, these companies are at this point, like, in some ways too big to police. And so, no, the reality is I think you, like, you know, trust these companies with your data at your own peril to some degree. And, you know, there's some kind of work where, like, you'd prefer your data not to go out, and there's others where your entire life work is, like, working on a math problem that if somebody else solves it before you, you're, like, out of luck. Probably shouldn't. You know that that's not the kind of work that I would do.

Ben, I'm dying to hear about your stealth company.

Yeah.

Ping me later.

It's it's a good one.

Next set.

Oh, save a slice for J. Cow. Yo, he what do you think? I got you.

I got you. I got you guys.

It's a complicated issue. What do you think of?

I went deep on this. It's a complicated issue. The stuff that we deal with—that's private information, right? And you know, making sure private information doesn't get out—that's very different from the discovery piece. Like you said, finding a new solution, discovering a new drug, and for those, even if you. strip out the PII out of it and you turn that into synthetic data, that might— let's put it— makes sense that that might still be enough hint for the model to help discover what they were trying to discover. And so there's this issue of like in the scientific community, right, where people are very conscious about what they share with who because they want to control who builds on top of their work. And if you start doing that work in ChatGPT and Claude, there is a possibility that it makes the next researcher find that direction faster. And I do think that is something that's going to be brought up a lot more in academia, especially as these things happen. There was actually another paper recently where a professor unknowingly had a preprint available online, and somebody else's AI found the preprint and then built on that research. And then the student proposed a full solution to something the professor was working on, and he had never meant to share. that 6 pages, but now agents are scraping the entire web looking for partial solutions and can build on it. So the importance of locking in your research becomes increasingly stronger. I think this OpenAI thing is really nuanced. There's a lot more— I think it's pretty complicated. I think no one was trying to hurt anybody, but it just fell through in a bad way.

Well, I actually helped you solve the problem, right?

Yeah.

I'm going to take the other side here. I've been saying for a long time that I don't trust the frontier model companies not to look at the results. not to look at the prompts. They have plenty of ways to get around this in the terms of service. And I would not trust them with anything that's important or proprietary. And this idea that you're going to give Instinct or, you know, Zuckerberg your Gmail, then your Notion, then your docs, then your databases means they're going to steal your IP, train it, and then give it to your competitors. Now, if it's for your, like, ordering your groceries, who cares? It's probably for the benefit of everybody. But if you're trying to solve one of 7 remaining mathematical problems in the world, I have to say, like, why on earth, if you look at the history of these companies, would you trust them? And I always try to explain this to founders. There's no free in the world. No free beer, no free pizza. There is no free vacation. You still have to sit through that seminar webinar for your timeshare. and get the high-pressure sale for 3 hours. There's nothing free in the world. So when they offer you free credits for your startup, OpenAI offered like millions of dollars in free credits. Anthropic is giving people discounts, whatever. And then what does Anthropic do to Figma? They launch design, Claude Design. What do they do to Cursor? They launch Claude Code. They lied, from what I'm told, and what's been discussed publicly, about their intent with the Cursor competitor. Cursor was their biggest customer. And so Lovable, ElevenLabs, Harvey— none of these companies should trust the Frontier models. And it's nothing personal, it's that they're under so much pressure with $100 billion, $250 billion, $1 trillion build-out of data centers, um, to make maximum money. And I don't think they can make max money from tokens. I think they have to win the application layer.

Yeah.

So they're lying and they're going to go after the application level layer. And when they do, your startup and everything you've built will be sucked into their new product, period, full stop. It's a trap. It's a trap. Put in the it's a trap Admiral Adbok here from Star Wars.

So are you guys running everything on open source models?

That's what Harvey said they're doing now. Harvey released, I think, just 2 or 3 weeks ago that Um, they had gone open source. There it is, it's a trap. Uh, they had released— uh, they were now doing open source for their model.

Yeah.

Then they were doing contained models for their customers. So, Ben, if you were concerned about, like, your Sherman Sterling, which I think is still the largest law firm in, Wilson Sonsini, Orrick, you know, Latham, if you think you have some proprietary data and you don't want them to go from one law firm to the other, because you use Harvey, Harvey's now addressed that. And they're saying you can have your own verticalized harness, small language model, whatever it is, they're going to fork open source. And so—

Is it on-prem? I mean, like—

It should be. And I have, like, one of the big successes in— this will probably wind up being the biggest success we've ever had out of our accelerator. We've had a billion-dollar company before, Grin, and multiple $300, $500 million companies. This one, Go.ai, formerly known as Abacus. They do on-prem, uh, and they basically make this Go One in a box, and you order this for your company, and now you're going. And they build assistants for you, a studio, um, and you're now compliant if you're healthcare, finance, any of the regulated industries that can't put their data into these services. Yeah, now you're boom, you're on-prem.

I love that.

with a server. So you don't have to wait to build a data center. You could just pop this into like some IT person's office and just start, you know, experimenting. And so everybody check out go.ai.

I have a plug here that's really relevant.

Plugs are— just to be clear, VCs saying what they invested in is not a plug. It is a description of conviction. It is a description of conviction. We love that here.

I have a company called Covenant Labs that's figured out how to encrypt open-source LLMs. So if you wanted to put it on a cloud, You can just take an open-source model, they'll encrypt it, give you a secret key. You encrypt your input, you get jumbled output back, you just decrypt it. So the GPU provider inference host never sees unscrambled data.

Okay, let me repeat that back to you. I think I understand it. I want to send a job to a cluster, but I don't want it intercepted because I'm renting the cluster. Uh, so I can send a job up encrypted, it gets processed. If, like, if I was building a frontier model or a vertical model, sends me my data back, sends me my model back, whatever it happens to be. And the service provider, if it was Crusoe Cloud, if it was AWS, never even knows what was processed.

Yes, the input is jumbled, the output is jumbled. You encrypt it with your secret keys beforehand, you decrypt output with your secret keys afterwards, but it operates— the model works exactly the same way as if you've never— as it was never encrypted.

I mean, only the paranoid survive. Right, Rebecca? Only the right things survive.

Yeah. And we have a SkyFlip, and it's the data security layer for AI. And they have signed some massive deals lately where they're basically— containerizing is not the right word, but protecting the data that are going to the AI models. So these big enterprises like Walmart is a big customer of theirs. can control and knows where their customer data is going into the model and when it's coming out. And so Skyflow has created— they were in existence prior to, you know, all of the huge step forward in AI, and they just were the right place, right time to sort of answer this question about protecting really sensitive data and allowing companies to use AI in that way. So Super interesting company.

Yeah, love it.

Yeah, really incredible founding team. I mean, the team's—

That's the one thing that I think has changed. You know, we were talking about this being analogous to, um, the dot-com bubble recently, and I think if you look at this, the, the real difference is these are real companies spending large amounts of money and getting very early wins from their spend. Does anybody here believe that the AI spend is not resulting in massively positive results, uh, for corporations and individuals? Like, does anybody— I think if anything, these AI tools are underpriced and the spend is underpriced.

They're running the math on the acquisition costs and the upsell and sales and all of that, and it's working, right?

So, well, I do think there's cases where you have people where, you know, everything is so new and I know of examples where there's major law firms that are, you know, let's go back to the Harvey example, using Legora and Harvey, trying both, figuring out what they want to be their sort of system of record, and they're going to pick one and drop one.

Or building their own. Orrick built their own, right? Orrick rolled their own.

So, yeah, I just think you're going to see, like, I don't think that all the revenue is going to be sticky, but I do think that the budgets are going to be only continuing to grow.

Yes, the budgets are going to continue to grow. I mean, the fact that somebody would buy 2— buy the top product and the second product— shows that they're willing to be inefficient and not do it sequentially, right? Because you could say like, okay, let's try Lotus 1-2-3 for 6 months, then we'll try, you know, Excel for 6 months, and then we'll have a bake-off. It's like, no, this is so powerful. Half of you get Excel, half of you get Lotus 1-2-3. The top 10% have both. It doesn't matter if we double spend. In our company, we have Claude running full steam ahead. We have Grok running full steam ahead, Perplexity, and we have local models running on Mac Studios. Like, we're doing it all and then just spreading it out across everything.

Yep.

And now I think I have to resubscribe to ChatGPT, which we kicked out early on 'cause I was like, eh, I don't know if I trust these guys totally with our data, et cetera. But now with their new model, I'm like, Can we afford not to have access to this?

But I'd be a little cautious. I mean, what we're seeing is that a lot of these enterprise companies that are, that are doing this, I mean, they have massive R&D budgets that we've talked about, right? And they're seeing the spend as a paid pilot, essentially. And so they, they will make a choice. It's not going to continue for forever, but they're, they're seeing it as a paid pilot. And, you know, a lot of these— I talked to a medical practice the other day who was using a company. There's a lot of these right now that go out and do Prior auths and all this kind of back office work and patient enrollment. And they had signed up a pretty major institution. And I called the institution and I was like, hey, excited to hear what you like about this company. And the answer was, you know what? It was free. They just let us sign up for a year. So we didn't really do any work on it. And I'm afraid that's a lot of what we're seeing in some of these kind of quote fast growing, you know, B2B company or B2C companies is, no, B2B, sorry. Companies is that, you know, they're, they're kind of trying everything. And I think, and I think it will shake out. I think they will, they will pick a horse. And so I, I don't know that their R&D budget will necessarily shrink or their spend will shrink on AI, but I think they'll pick a horse.

I think I'm spending like $5K a month in our firm across 20 people.

Yeah.

And I was totally reasonable. And I'm like, that's real, $60K. Uh, I'm like, and they're like, can we spend $400K on this a month. And I'm like, I think you could double the spend and we're going to get the value from it. Like an extra $60,000, $70,000 a year. Okay. That's like one, you know, entry-level person. I think entry-level person makes $60,000, $70,000, $80,000. So do they beat the— does that spend beat the incremental entry-level employee? And then the answer in every case, Yohi, is—

We're never going to have an analyst like at this point.

Which, by the way, coming back to the first point of the day. There's a reason that a bunch of people are pretty freaked out about AI. No one's hiring analysts again. But, uh, yeah.

I mean, that is— that's interesting you guys are seeing it that way. I'm hiring 5 researchers at a time, hoping one becomes an associate. We have an associate-in-training program. But now when we hire them, we came up with a trick. We give them a project to do on their own. We pay them $500 to build something with AI, like a dashboard, an agent to solve a problem, like write a deal memo, research companies, whatever it is. And, uh, they can't use something as amazing as Harmonic, right, Lon? Because that's paid. Uh, Harmonic AI, which is a partner for our program here. Uh, they have to use like whatever they have access to. So if they want to spend $20 on Lovable, you know, while they do this project. And we had 15 people who were like good candidates. Half of them wouldn't do the project. And I'm like, but we're paying you $500 for 10 hours to do this project. And I realized, oh, those half are not capable of using AI. They're intimidated for it. So the 5 we wind up hiring, none of them were intimidated by AI. And that's the key. Like this stuff, people are explicitly— there's a group of people, it's probably 50%, who are explicitly not using the tool. And that's like a major red flag. If you just are too scared to use it, like you just don't exist in the world anymore. That'd be like not using the internet or not using Microsoft Office or whatever, Google Suite.

I started doing intro to AI sessions, not for portfolio companies, but telling my CEOs they can invite all their employees to just demystify where to get started. The idea that like if I can level up all the employees across all my portfolios, even a little bit, just making them more comfortable, I think that has a good impact.

Yeah.

We started doing that. It's been pretty cool. I get a lot of good feedback on it. I'm hearing that employees who are hesitant to use AI felt much more comfortable starting to play with it.

Well, and then what's brilliant about what you're doing is if you were— if you used AI in January, March, or June, and you have impressions of it, your impressions are completely wrong because so much changes every 3 months that like, You would have missed like 4 or 5 new product features from GrokBot to, you know, the new OpenAI model, the 6x models, Claude Fable. You would've missed all those. Claude Tag, you just wouldn't understand the state of the art. So people think, I tried, you know, I tried AI, it didn't work. And I'm like, okay, when did you try it? They're like, that was like, I think it was like May or June. And you're like, doesn't— what?

Yeah.

That's like 1990. You used the internet in 1990 and It didn't work for you? Okay, it's changed a bit.

So it makes it so hard to invest, right, Jason? Because it's, it's, you know, people start developing these applications and you look at it and you're like, well, in 6 months, are the models just going to be—

Or it makes it easier because the only thing you're investing in in year 0 and year 1 is the team. The team, the team, and the team. Yeah. Who are these 2 or 3 co-founders and can they deal with this pace? Will they—

By the way, more and more of these exits are, I mean, like Cursor, Pure pivot. Yes. Like, I mean, all, you know, like you're looking at like the big exits of today. Most of them started, my under— my, I'm actually not sure. I was told that Instinct is a pivot as well. I don't know if that's true.

That is the classic case for outlier success. Even SpaceX, they have a launch business. They add Cursor. Okay. They've got a, you know, $3 billion in whatever Cursor's revenue is, or maybe it's $4 billion now. And then they're like, we have extra compute. Now we have Elon Web Services inside of it and it's $30 billion or something insane. It's like dwarfing the other categories. So, so much opportunity space. It just requires a team that is just absolutely relentless. And we had a company, Micro One. They were a $12 million company. We invested $500K. We did a little syndicate. We just fell in love with the founder, Ali. And it turns out, like, he made this incredible— he's a data lab now. Hundreds of millions in revenue, $4 billion valuation. Again, we invested in when it was $12 million. So this will be a fund returner, like times 5 to 10 if, um, you know, things keep going the way it goes. Um, he started with just trying to figure out— he made an AI test to figure out who were the best developers. So he built an AI tool just to figure out who's the 1% of AI developers. Then he saw Scale AI and all these other labs and was like, wait a second, I could just find the top 1% of legal accountants, whatever, you know, we need training data for. Okay, let's just be a training data company. They just see it. That's what I always tell founders, peripheral vision, super important. You will see something in your peripheral vision that is bigger than what your target was, and then you just turn and you change the target. You just move the crosshairs. It's not that big of a deal.

But you have momentum, and you can't start from zero. Lyft was that. Lyft was Zimride, right? And they saw what Uber was doing and they just turned a little bit, right?

Yeah. I mean, they just— But they were up against a generational CEO. Yes. Goes back to my core premise. Like, just bet the jockey, man.

It's just, it's always the jockey. Always the jockey.

All right. Listen, great episode, everybody. We could be here for another hour. So much to talk about, but, uh, let's end with, hey, what's your, uh, what's the fund returner or the most fun investment? You can pick whatever you like in, you know, this year. Last year, whatever it is. Just a company you want to give a shout out to that's doing exceptionally well in your portfolio. High performers. Let's let's make that the category. Yogi, you got a high performer you want to share with us? Yeah.

I mean, we've all talked about sandboxes, but E2B was one of we we did one of the earliest rounds of E2B that was right after Baby AGI, and I remember Baby AGI being in E2B's deck. So it's been really cool to see see them just the whole category just blow up and them them staying at the lead on it.

Ah oh so these are the machines. Yep. that things like Muse and Cursor pop up and run virtually.

Exactly. So Perplexity, I think, uh, you were using, uh, Manners, GenSpark, those guys were. So when they write, write code and run it, they wanna run it in a safe sandboxed computer so it doesn't touch anything outside of it. So that's, uh, that's what these guys offer.

Talked about, you know, Savvy and, and, uh, and, uh, Skyflow. But, you know, one of my most recent deals, I absolutely love this team. When you talk about founders, it's a team of 3, like MIT, Eng, and they're doing—it's called Daptic—and they're doing compliance monitoring, which sounds super boring, but it's so cool. Google Data Centers just signed on. Honda is a huge customer. Komatsu—all these really, you know, sort of old-line industry players—and what they help companies do is ensure they are regulatory compliant around the world. And what I loved about this. company is when I, when I called some of their customers, they, they talked about the ROI behind the contract they just signed and how they— one company in particular said, we just signed 4 times more, you know, more pieces of equipment in this country than we would have signed without Daptiq because Daptiq could show our customer that we were in— we checked every regulatory box they have.

Awesome.

And what they're doing— yeah, super cool. What they're doing now is they're going upstream into new product development. And so if you want to design a drone, for example, and know that you can fly that drone in every geography out there that you want to be in, it'll tell you exactly how to build the drone.

Wow. That'll save a lot of time.

Yeah.

Amazing.

It'll save a ton of time.

Sounds incredibly boring and like a money printing machine.

Exactly. That's what I love. Incredibly boring and money printing machines. Yes.

Uh, Ben, tell us about Augmodo. Ben, tell us about Augmodo.

Yeah. Well, you know what I like about this company? What, what, like, The reason I wanted to highlight it is it's, it's in a space that is sort of just a graveyard. Like going and selling into big retailers and supermarkets is, you know, one of the— it's like an absolute no-no for venture. It's like, we've tried this 50 times. Like, we understand it's a trillion-dollar category, but like sales cycles are too long. Like, no thanks.

And margins are too thin.

Well, there's like 1,000 reasons that like you're basically told that. absolutely under no circumstances should you engage in such things. And you guys have just sort of like really bucked that trend and gotten like unbelievable uptake with retailers who, you know, shouldn't, shouldn't have trusted a company of this size in the early days. And, you know, what they, what they figured out is how to, you know, essentially solve the what's on the shelf inventory pricing problem. In real time without the human labor going and doing checks and going and, you know, like the armies of people that are, that are running around, you know, counting inventory on shelves and passively solving that with a piece of hardware that store workers are using as their sort of name tag. And over time, it'll become like the physical hub for how they use AI in their day job. But that also passively monitors everything happening in the store and ties back into the inventory system. And just like, this is a really, really— yeah, it's a really cool company.

I actually looked at this early and passed, unfortunately, and because I had seen all the robots and things like that in prior things. But I mean, this solves so many more kind of problems that retail stores are having, including the shrinkage issues and things like that that they that they see in their stores.

And then, Jason, this comes back to your idea about like founder. Ross had sold his last company to Niantic and was really like, you know, like very, very, very, you know, top 0.01% AR thinker and applied, applied sort of like that giant weird brain to this category. And by the way, had no background selling into these kind of retailers. but just was a very special person. And, you know, we're getting rewarded for trusting our instinct on the person when we were really like, like all, every smart person in the world would say just like, don't touch the category. And I like, I like when we're right when, when like we're being brave, not just when we're right when we do the obvious thing and take the money.

Yes, it's the non-consensus bet. And this is so of the moment because AI, spatial computing, And you look at the badge, right? There's this whole movement to like Amazon tracks employees. We have companies that do this for fast food workers with the standard cameras that are there just to make sure, you know, things are operating optimally. And, you know, you run the risk of it being creepy. Here, what I like is it's not a creepy thing. It's just the person's walking through the store and if it's got incorrect pricing, low stock, it just figures that out from the badge. And then you've now turned the average retail employee at a supermarket or wherever else they could put this into like a superhuman robot, right? Like, they couldn't possibly process this much data. So the store is going to be more efficient. And this goes back to like what I was saying earlier. You're not going to be replaced by AI. You're going to be replaced by somebody using AI. This is the perfect example of it. Like, this person is 10 times more valuable now, which means they could have a higher salary, which means the store could be more profitable or more sustainable, or they could pass on savings. So this is The promise of AI here, it's going to make the store more profitable, which means the store can raise wages, have less turnover. It's just absolutely awesome. And yeah, to your point about being a VC, like this, the startup graveyard, I feel like is the perfect place to look for new, uh, versions of old ideas. There was an Uber and a Lyft.

Yeah, yeah, totally.

It was called GoTaxi or Taxi Magic. That was on Palm Pilots and by text. So there's this company called Taxi Magic you could call, but it was pre-GPS, pre-iPhone. They were like— Taxi Magic was like, yeah, there it is. This is—

It's like Cosmo and DoorDash, right?

Cosmo and Urban Fetch. The 2 led to DoorDash. Yeah. And so look in the startup graveyard, folks. Like, we get pitched all the time on like a travel plan. Because of AI.

Yes.

Like the whole idea of like a travel planner was like group travel planner. Well, people go on a group trip like 5 times in their life. It's not like a common thing, but we had an AI travel planner. It was doing so good for us. It was called Roam, I think. And then the founders, and they were like during the ChatGPT 2.5 when it was only an API era. And they were like, yeah, ChatGPT 3.5 is coming out. You can just do this directly with ChatGPT. And I was like, no, no, no, you have $500,000 in the bank. Keep iterating. ChatGPT is never gonna build an interface. They're never gonna have a logo. They're never gonna have a concierge. They'll never have a $500 a month product. Like, it'll do 80% of what you're ever gonna do. And I was like, please don't quit. This is a different one.

Oh, wow.

This is Roam Around, but we had one called Roam. Unless that, yeah, no, I don't think that was it. But it's an obvious idea of like travel planning. And now I've got a bot named specifically for Tokyo and specifically for, um, the Middle East, where I take— this is my new workflow— I just take a screenshot of whatever's interesting that I see, and then I send it to the Grok bot, and I say, put this— when I send you a screenshot, add it intelligently to my database itinerary for my next trip. So I sent it like the world's 50 best restaurants. I said, tell me which of the restaurants on the top 50 list are from Japan.

In a city that I'm go— yeah, I mean—

In Tokyo. And it was like, bing, and it made the list. And then I'm like, okay, now go make me a Google Map. And it's like, sure, I can do that. I can make a Google Map for you. I'm like, okay, that's just— why do I even need a travel planner? It's like incredible. All right, everybody, another amazing episode of This Week in Venture Capital. Thanks, Ben, Rebecca, Yohi. Great to see everybody.

Excellent job.

Thanks to our amazing news reader, news reader Lon Harris.

Well done, sir.

Well done. We'll see you all next time. Bye-bye.

Thank you.