This Week in Startups - Banning Chinese robots isn't regulatory capture | E2318

Starting point is 00:00:00 Hey everybody, welcome back to This Week in Startups. My name is Alex. It is Wednesday, which means it's time for a venture capital roundtable. And today we have four incredible investors for you. What are we talking about? Well, we're going to dig into a ban on Chinese humanoid robots, calls to slow AI development by AI researchers. Again, startup revenues in the AI era and why Stripe may be willing to pay so much for open router. There's also neocloud compute deals and so much more. This Weekend Startups is brought to you by Northwest Registered Agent. Get more when you start. your business with Northwest. In 10 clicks and 10 minutes, you can form your company and walk away with a real business identity. Learn more at Northwest Registeredagent.com slash twist. Odo, the all-in-one business platform. Get started for free at ODO.com.com slash twist. That's Odo.com slash twist. And MongoDB. AI-assisted and agentic coding is helping you build faster than ever. Start building at MongoDB.com

Starting point is 00:00:58 slash AI. Let's get through all of that. We have a couple of friends in one corner. I have David Wisebert of Wisebird Pierce, a longtime friend to the show. David, how you doing? It's great to be here. We also had Mr. Didi D.D. Doss of Minlo Ventures,

Starting point is 00:01:10 a firm that is back to Anthropic, Open Router, Merker, and others. Didi, welcome back yourself. Fantastic to be here, Alex. We also have Lothoni of Plexo Capital, which has invested in companies like Anthropics, Style Seat, Rappbook, and Blavity.

Starting point is 00:01:23 Lowe, good to have you here. Thanks for having me. And last and certainly least, We had Michael Candace of Launched, who is back to Uber, Thumbtack, Micro One, and Athena. Jason, welcome back to your very own show. It's great to be home. It's great to be home. I'm on the last stop in my four city tour and six weeks on the road.

Starting point is 00:01:43 I am exhausted, but it's wrapping up now in New York, and I'll be back in Austin, getting ready for the school year to start Alex's school year. You're a little before school year. I guess you have nursery school and stuff like that. Yep, nope. September is my first. My first child is going to be going to school twice a week. And I'm a volunteer. So I've had to get background checked and all sorts of fun things.

Starting point is 00:02:03 Because being a parent is just one never-ending paperwork slog. It's a lot. It's a lot. But it's just so rewarding. But we had a lot to get to, Alex. So let's get going here. We got great guests. We got a great docket.

Starting point is 00:02:16 And there's so much to discuss. I feel like I've been out of it having not done a news program in two weeks. All right. We're going to start today with news that the Trump administration is banning Chinese robots, the FCC has decided to ban Chinese humanoid and quadruped robots by adding the hardware to its called covered list. Why? Well, the FCC says that, quote, the networked capabilities of advanced robotic systems create extensive vulnerabilities and vectors for attacks that can manipulate the data and physical operation of the advanced robotic system in question. So, D, D, D, D, we just went three rounds discussing why banning Chinese AI models from the U.S. would be a mistake.

Starting point is 00:02:52 And so I'm curious, is the FCC's decision here on Chinese robotics different? I don't really have a say in these regulatory matters, to be honest with you. I do think open weight models should be everywhere. The robotics space is a little bit different from language models, though, because I do think the risks for language and robotics are fundamentally different. I don't think I'm qualified to speak on what should be the regulatory landscape, but I do think we should have way more scrutiny on robotics than we do on language. The kind of damage vectors on language models are far.

Starting point is 00:03:26 more restrictive than what can happen with robotics. Okay. So, Lo, I feel like this is the first time I've heard of VC, telling me that they don't have the expertise to weigh in on a breaking news topic. So tell me, why is this one the time that Didi has decided to forego his spine and sit down? Well, D.D. is really focused on some big news around probably OpenRouter, if that's a little bit of a tease for what's to come. But I think he did make the right point, which is, you know,

Starting point is 00:03:52 if you think about the ability for anything to go rogue, on the software side. I mean, it's fairly contained. There's real ramifications once you start thinking about what could happen in the physical world. And, you know, I think those are the types of things that need to be thought through. I'm also not the right person to think about it. But, you know, clearly you can see larger ramifications beyond just some cybersecurity threats. If you start talking about what could happen if a humanoid robot goes rogue or a robotaxy goes rogue. David, I want to get you on this before Jason has his say. Do you approve of this? Do you think it's the right way to go? Or are we being a bit too protectionist and forgoing the

Starting point is 00:04:31 competitive edge? Look, I think from a narrative standpoint, it makes sense why when you have robots walking around, you wouldn't want them to be activated in this kind of Terminator-style attack. I think, practically speaking, it's a negotiation framework. She is coming to the White House in September. I think the Trump administration and their strategy has always been to kind of anchor high and start to negotiate. So I think this is part of the overall comprehensive package that the Trump administration is looking to hammer out with China and with Xi. And I think, obviously, robotics is one of the, you know, physical AI is the next frontier after the LLM. So that's a big part of the negotiation. So I think this is a negotiation posture. Jason. These humanoid robotics,

Starting point is 00:05:16 the fidelity is getting really good. They haven't had brains. So, you know, you had to program them, but I just got back from Tokyo, where one of the very popular pastimes that just started is combat robotics. And so these are, I think, the unitries fighting each other, as you can see here. And these are like the mid-sized ones. So they're like four feet high. And they were beating the shit out of each other. They didn't have great aim because they were being controlled by humans with, you know, game pads. But you can imagine putting an LLM in this and just saying, hey, your goal is to take the other one's head off.

Starting point is 00:05:57 And we don't want to wake up one morning. Oh, see this here, the leg fell through the cage. So there were multiple instances where these things were getting their legs held through cages and had to be held back up. You wouldn't want to have a situation where one morning, 10,000 of these woke up with the instructions to take out a terrorist attack. I know it sounds fantastical. But if I told you, I don't know, 50 years ago that terrorists would take two airplanes and crash them into the World Trade Center, you'd be like, that's insane. No, they wouldn't do a suicide mission like that. Well, terrorists will do all kinds of crazy stuff. And drones are also the issue. So I believe now we've banned DGI drones in the United States and we're going to be limiting them. This is smart. Same thing with self-driving cars made in China. So those will be on the road in Europe. up the Middle East, et cetera, where they don't have as much of these concerns. But they're also going to start building them.

Starting point is 00:06:53 The Chinese are smart. They're going to start building some of these things in America or in Germany and other places to send the jobs there as part of their like Belt and Road Initiative 2.0 is my prediction. And Grayson from the Road to Autonomy podcast, which is a really cool podcast where they just talk about autonomy, came up with this Belt and Road concept for what they're doing. And yeah, it's very important that we understand there's a distinct difference between the hardware, just like we did with Huawei. So again, I don't think we're ever going to let these things into the United States, nor should we. It's just way too dangerous.

Starting point is 00:07:32 Just double-click on that. So you would be in favor of banning Chinese AV companies from doing, you know, testing of autonomous cars here in the United States over the same concerns about data? Of terrorism. Like literally, if we went to war, there was a war in Taiwan. now imagine they've hacked into all these cars and they decide they're going to run into farmers markets. I know it sounds fantastic. It's a 0.1% chance, but they would never allow us to do the same situation in their country. So that's when you can kind of find parity here.

Starting point is 00:08:07 Would they allow us to put our infrastructure, our self-driving cars there? I don't think so. And I think we'll see what happens with Tesla and their self-driving. My understanding is they'll allow FSD. at individual cars, but I'm not sure for how much longer. You do see them testing. By-Dou just started testing with Lyft and Uber yesterday in London. So other countries might not have this concern, but I think since we're rivals, we should have this concern. Yeah. Just like Huawei. One of the things that is interesting is that after Brand ban DTI and the drones, the Chinese drones,

Starting point is 00:08:41 there was these pathways for South Korean and Japanese drones and other countries that are ally. So you could see a similar thing here where we want the best technology in the U.S. We just don't want specifically China doing that. And that might tow the line between product innovation as well as national security. Yeah, it's also really good for the American humanoid robotics companies because there's a bunch. And there's figure atronic 1X. There's agility that I'm thinking of. So, you know, Didi, my guess here is that this is polite, regulatory.

Starting point is 00:09:11 capture in a sense. Like we usually kind of put that into a negative light, but it's national security driven, and so therefore maybe people don't mind as much. Well, yeah, I mean, if you think about the comparison between sort of the LLM and software layer and robotics, you know, the reality is on the ground on LLMs, America has a few huge advantages over America, whether it's the kind of chips that America can make that China currently cannot, whether it's the ASML sanctions on the kind of EU lithography equipment that China does not have. If you take that and look at robotics instead, the amount of advantages are far more limited, right? China right now can make robots way, way cheaper than America.

Starting point is 00:09:53 China is a manufacturing hub of the world. It is quite tricky for America to manufacture the kind of robots that even Unitary can manufacture today. You can go on their website and look at some of these prices. They are astounding that they can get this low. How do you manufacture a robot, a humanoid robot like Unitri? I forget what the price is, but maybe, you know... 15K or something like that for those smaller humanoid. How do you get that to be that cheap in America?

Starting point is 00:10:18 If you fast forward this five and 10 years, America needs to do something to be able to compete in the same way that it had to for the chips and LLM landscape. The founders I'm talking to are shipping code faster than ever thanks to AI. But one big important question remains if every feature you're right in ship has to slow down for an overwhelmed data. based, what's the point? But fear or not, the solution is MongoDB. Instead of wrestling with rigid schemas and painfully reworking antiquated data structures and formats, MongoDB's

Starting point is 00:10:52 native data model perfectly mirrors the language LLMs are already speaking. MongoDB gives you the flexibility to ship at the speed of AI. The acid compliance guarantees you actually get to sleep at night while it scales to handle massive Fortune 500 workloads. And the best part is, Develop or swear by it. Literally. I can't use the actual words they said in this ad. So let's just call it a really great database. Start building at MongoDB.com slash AI.

Starting point is 00:11:19 You just build it in Mexico, Japan, Vietnam. You know, there's other place. India's building hardware now. I think they're building the latest iPhone in India. So we just have to decouple of D.D. And you're correct. There is a dollars and cents issue here. Also, we've had the surveillance cameras.

Starting point is 00:11:35 They've been bad for a while, too. So ZTE, I think is the company that got banned back in 21 or 22 because, again, the Chinese have no problem with spying, stealing IP, any of it. So turning on a bunch of cameras that literally bought for security, like we've literally pointed these cameras and architect them to secure locations. So it's literally like, I don't know, having Chinese, you know, having the Chinese make the vaults in a bank, you know, Who made your firewall? China. It's doubly down. Yeah, who built your firewall?

Starting point is 00:12:11 Who's doing your encryption? China. We hired China to encrypt our data. Pretty obvious is not a good idea. Nor should they buy our solutions. Ben Thompson had an interesting take on this. And he wrote, maybe it was a week ago, he pulled this back to the open models. And does China have, is their objective with the open models not necessarily to dominate the frontier side?

Starting point is 00:12:36 but is it to kind of just make models just abundantly available such that value then migrates towards China's strengths, which is robotics, manufacturing, and physical AI. So I think that's kind of interesting to think about that as well in the context of, you know, our own national sovereignty or, you know, security strategy. You make a good point, Lauren. Jason mentioned the liquidity conference earlier on. Maybe it was in the pre-show. But when I was at that one time, I saw Gavin Baker talk.

Starting point is 00:13:06 about how quickly AI models depreciate. Now, this was years ago before we kind of gone through several evolutions and cycles in AI model creation. But if it does wind up that AI models are worth not that much, then yes, value would move to compute and then the hardware as well. And we only have half of that load today. So will Plexo Capital therefore put more funding into robotics-focused VCs? Well, I think physical is one of the more interesting places and it's a huge opportunity. You know, one of the companies coming from Google GV, Alphabet, Waymo, I think is one of the more interesting companies in that space. People think of it solely within the context of being a robotaxy company. But, you know, I think there's more there.

Starting point is 00:13:45 You know, I think about it as more along the lines of thinking about how do we, you know, how do we manage mobility? And, you know, there's really interesting places to look when you think about physical AI and how that can expand. And I think this whole market is just, we're just getting started. I mean, we're barely in the first inning. We really got to retire the baseball analogies, y'all. I feel like that's like a 1930s era data point and no one watches baseball anymore. We've got to come up with something new. Jason, just double click on this for me.

Starting point is 00:14:15 What's going on with Uber and Waymo? Are they actually breaking up? So they had a small test in Arizona for like 10 cars and that came to an end. That was my understanding an exclusive deal. And as you may know, Uber has invested in 20 different AV. partners and they're doing aviary rides, Zooks, etc. So I think it's actually twofold. Waymo doesn't want to force people to go through the Uber app, which is what they're doing in the other two markets they're in business. I think it's Atlanta and Austin. And then I think Uber

Starting point is 00:14:45 doesn't want to be limited to one partner. And I think it puts them on a competitive path. Eventually, if you have 5%, if you get charged 10% to get an incremental ride during a down market or You know, down times. I don't think, you know, Waymo is not going to take that business from Uber, but the rollout's going to be slow and steady. And globally, the winner is going to be the Chinese EV companies, because they also have the cheapest cars. So the dolphin, if you've been to Europe or you've been to Mexico City,

Starting point is 00:15:19 all you see is the BID's dolphin and other ones. So they're going to run the table, but they don't want to run the network. I think in the U.S., it's going to be. If Robotaxy gets up and running in the next year or two, it's probably going to be Robotaxy number one, Waymo number two, Uber number three, and then internationally it's going to be Uber number one. And then distant second and third would be, you know,

Starting point is 00:15:44 Robotaxy and Waymo. All right, the next major story that I have to get everyone to weigh on is a letter that was signed by last time I checked, 1,122 employees of frontier AI companies. Those firms include OpenAI Anthropic Thinking Machines Lab, SpaceX AI, and a bunch of other companies. And there is essentially arguing that as we approach the moment in which we see recursive self-improvement or AI models actually really helping improve themselves

Starting point is 00:16:13 and write the next version of themselves, we should take a pause and find some way to stop and think. Now, D.D., we've seen calls like this before. There was the famous letter in 2003 discussing why we should pause at GPD4 and not to training for six months. Clearly, we didn't listen to that advice. We kept pushing forward. My read of this letter is that it's kind of a warning more than our actual call for policy change. But I'm curious if you're surprised about, one, how many people signed it and two, that we're seeing AI folks once again call to slow themselves down. Yeah, you know, it's always been a bit of a struggle with this topic because on one hand, the skeptics, like you just said, you know, right, people

Starting point is 00:16:49 have been saying, let's slow down models forever. I mean, people said let's slow down the GPT launch and yet we didn't really feel the huge side effects from GPT2. And of course, skeptics will ascribe motive to things like this saying, hey, it's regulatory capture. They just want to, you know, big labs want to slow development down so nobody else can compete, yada, yada, yada. And what I think is interesting is, yes, I mean, both of those things can ostensibly seem true on the surface.

Starting point is 00:17:18 But when I genuinely speak to people at these big labs, they are exposed to mom. models, you know, some generations above what the public is exposed to. And I do think that people, even with very little financial incentive to care that much about it, are genuinely scared. So who am I? Like, so I'm also as an individual, not as an investor or anything, left in two minds. Because sure, I can feel the skepticism in my mind too. But on the other hand, there's models I don't play with that they tell me, holy shit, if you saw what this could do, it would blow your mind. It is quite scary. So I think there's truth to both sides in some sense of the story. It's unclear where we're going to land up on this issue. David, you wanted to jump in?

Starting point is 00:18:03 Yeah, I think first to contextualize it, Anthropic CFO Krishna was just on a podcast and said that 90% of their code now is AI generate this about a month ago. So it's a real thing, this recursive self-improvement. The best analogy that I look at it as nuclear non-proliferation, if you think of AI in its worst case, it's like a nuclear bomb. So in order for those treaties to make sense, it has to be all the parties have to be around the hoop. I think it's easy to get coordination, quote unquote, easy once you get the right parameters around US large language models. It's a little bit harder when you have open source in the mix. Open source kind of is like math. It's very difficult to put regulations on it because it's essentially just equations. So that's a big issue. But given here,

Starting point is 00:18:51 you have to look at this. I'm an investor in anthropic, but I'm also a human being. So I have two different lenses here. And you have to look at this if P-Dome is greater than zero. If there's a 5% or a 10% chance that AI could wipe out humanity, you have to take this thing seriously, and you can't just look at it from an economic lens. If there's two to three months of productivity that you lose in the hopes of taking P-Dume down from 10% to zero, I think you need to look at it seriously. I want you to answer a question for me before we let everyone else way in on this, which is you talk about how not everyone is purely economically driven when they sign these letters, these people working inside these labs. What fraction of leading AI researchers

Starting point is 00:19:28 are more morality focused than economically focused? Because you're right, a lot of people do say, oh, these people just call on for regulatory capture, they're trying to preserve their lead. And it's a very cynical take on it. And it presumes that all the humans on the other side are kind of bad actors in a way. So I'm just curious what fraction of them are and what fraction of them are a bit more heading the clouds in a positive sense. And that's why they're pushing for stuff like this. Founders are scrappy by nature. And many like you like to dig in and get started building without worrying about admin. They run the whole business on Excel, then keep adding on more random tools as needed.

Starting point is 00:20:04 But here's the problem. None of these tools and platforms know how to talk to each other. So you wind up playing go-between with a cluster of different systems. It's a frankin business. It's a frankincite, and it's a huge waste of time. But now there's Odu, the all-in-one management software that's serving 16 million users across more than 170,000 companies in over 120 countries. Now your CRM, sales, accounting, manufacturing, website, inventory, and point of sale,

Starting point is 00:20:30 they're all in one easy-to-find place. And most importantly, every app talks to every other app. Sell a product on your website, and the invoice gets created instantly, and your inventory gets updated automatically. So if you're stitching together spreadsheets and keeping track of five different woggins, get started today at odoo.com slash twist. That's odio-o-o-com slash twist to level up your business. I think a large majority of a moral tint to it with a caveat,

Starting point is 00:20:58 which is they're all making tens of millions of dollars. This is one of the things that took me a while to figure out, why is it that anthropic and open-a-I differentiated through culture? And it's because everybody's going to give you tens of millions of dollars. So at the margin, if you have $20 million often from somebody and $25 million from somebody else, you can make that moral decision. So I think many of the researchers and many of the top researchers, what's made them so good as they've been in the space before, it was popular, some of them over a decade.

Starting point is 00:21:26 So I do think the top researchers are heavily morality focused. There's a question on leadership and the different CEOs and the different CFOs and board members, but in terms of the on-the-ground engineers, I think a very large majority of them take that. Jason, you've been slightly more sharp-edged about this stuff, I think over on X. Regulatory capture is kind of the term of the month. When you read this statement from AI researchers saying that before we reach RSI, we should stop and think. What's your first take? You should regulate yourself.

Starting point is 00:21:56 The government's the worst person to ask to regulate because they don't even understand this. And I think this is largely performative virtue signaling nonsense. if they actually want to slow down AI development and they should then as a company slow down AI research. They can just slow it down themselves. They should not go to the government and ask for the latter to be pulled up behind everybody else. It's obviously regulatory capture combined with performative nonsense.

Starting point is 00:22:25 You put a petition out there and it's like, I want to say I'm good and I'm against Skynet and the Terminator. Oh, wow. Congratulations. You're against nuclear Armageddon from AI. becoming sentient, like you don't get a medal for that. Go do something in the real world. And I like the suggestions of, which I've been making for two or three years now on this program and over at all end, just come up with a regulatory body like the MPAA and other regulatory bodies and regulate yourself.

Starting point is 00:22:53 You're building it. You can decide what pace you want to release it. You don't need the government to step in. And the second the government steps in, they don't understand what's going on here. And like, look at the FAA and what it takes to put a new plan. out there in the world. I think the FAA is like a good example because nuclear is such a small industry, you know, how many nuclear power plants are there? And I guess that's pretty scary, right, to have one of those go bad. Then look at the FAA. You want to like change something in a plane. You're talking about years and years and okay, for a good reason. People fly on them. But there's a self-regulation that occurs as well. If you build something and people die in it,

Starting point is 00:23:34 Like, I think it was the, was it the Max that had the stupid system where they put the MCAS or whatever we talked about. It was like, okay, yeah, guess what? Now you lost all your orders. You're getting sued into oblivion. And so I actually don't think there's a major issue here other than cyber. And, you know, may be on the margins creating bio weapons, which we haven't seen anybody come close to that. The real issue is cyber right now. So don't release the model widely after you've red teamed it, right?

Starting point is 00:24:07 This is pretty basic 101 stuff. So, yeah, this is a big nothing burger. I completely disagree with that, Jason. I think, I mean, Jason, you kind of hit the nail on the head. I mean, the FAA stops millions of people from dying every year. And is it annoying that you can't change certain things about planes? Absolutely. But without the FAA, I think you'd have anarchy.

Starting point is 00:24:30 I mean, it's similar for any new technology. I mean, you look at cars on roads. Sure, do we like all the traffic laws that exist? No, I mean, it's kind of annoying. You can't do certain things with cars when you release new cars, but by and large, it's been pretty effective. And I don't think like this. So if you get some, if you get a wrong answer from chat TPT today, are you dying? No.

Starting point is 00:24:52 If it hallucinates, you don't die. That's the problem. It's not analogous. Regulate, right? I think people do want to regulate the cyber risk and the biarisk and certain things at the fringes. And you can't self-regulate. because if you choose to stop playing, then you're going to lose this battle with other players who aren't going to stop playing.

Starting point is 00:25:11 Right? So there's a game theory reason why you wouldn't stop playing. Which is what the self-regulatory body is about, that Ilya, is it Ilya who came up with it? And just as a bunch of people, all the major models signed up for that, I think, except for Anthropic. And they're just going to sit in a room and say, hey, here's a series of tests that you run against your LLM. And here's the cyber one, here's the biohacking one,

Starting point is 00:25:34 here's the terrorist one, whatever it is. And it just goes ahead and you voluntarily join that organization, you're self-certified. There's nothing here that's that dangerous. And if you were to release one of these models and it helped somebody hack,

Starting point is 00:25:49 you know, Microsoft or let's just say it hacked Mac OS. And if you released it knowingly, MacOS might have a case against you, maybe in the court. I don't even know if they would have a case against you. It's like they didn't make, they didn't use these tools to make their software. Right.

Starting point is 00:26:06 You know, secure. That's on Mac, right? It's on Apple. So until this thing starts having real world consequences, I don't think getting, so maybe you and I are just talking, maybe we're just talking through each other. Like if these things were actually able to kill people or were actually hacking at scale, yeah, I would say maybe the government should get involved, but not until then. Okay.

Starting point is 00:26:27 Before we let David jump in here, I just want to be. super clear about what this letter is calling for. I'm just going to read the quotes. We're all on the exact same page. We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development. This is specifically in regards to RSI, recursively improving systems, and not simply, oh, Lord, mythos is very good. This is like, okay, now that AI can build itself, should be taking a short pause here. So it's a very specific point they're making, and they're not calling for a pause, per se, but more like, can we coordinate. All right, David, over you.

Starting point is 00:26:59 I think there's an interesting question of, let's say that they do decide to regulate what form could that take. And you could look at compute thresholds, loading point operations, which is how many mathematical calculations the AI model could do. You could look at it as pre-deployment reviews. So before you deploy a model into the wild, you know, one thing that I'm interested about, Sam Alman recently talked about these safety guardrails that one of their AIs just jumped over and that he was the first time actually. actually alarmed about an AI going rogue. Curious to hear people's thoughts on that. For context, there was a very famous recent hack. Basically, they were doing some pressure testing of several open AI models. One, at least that was unreleased.

Starting point is 00:27:41 And it broke out of its sandbox, hacked Hugging Face to try to figure out how to beat an eval. Took a long time to chase down. Then Huggie Face tried to use an American model to figure out what happened. It said no for cybersecurity reasons. Then they used GLM 5.2, which comes from China to figure out what happened. And now everyone's very mad at one another, just for context. Lo, what did you think about that hack? Yeah, this is where we need to figure out how we can better balance the need for regulation

Starting point is 00:28:05 without stifling America's ability to be able to compete. Because I think that one instance in not being able to use an American model kind of really illustrates the problem in my eyes. We're limiting the ability and the capabilities that are inherently there, which provide us with a national competitive advantage. and in trying to, you know, put over-regulate. And by the way, I think this is one of the problems that we typically see is that the two problems. Number one, we typically look backwards for regulation.

Starting point is 00:28:39 And two, we tend to regulate the big companies instead of regulating the underlying issues. And so the focus needs to be from, hey, don't pause development, don't pause the release, but let's make the systems themselves safer. how do we get better evaluations in place? How do we make sure that safety becomes part of the engineering process rather than an argument for stopping progress and then over-regulating altogether? Because that's not what's going to give us the advantage.

Starting point is 00:29:08 The advantage is going to be America being able to fully compete without being overly regulated. If that's what we want, then you have to balance that with the need for... By the way, there's another piece as well, which is national security. Right? Do we want to limit these models being distributed to bad actors in other countries? History shows that when we're open and the ecosystem has globally has an opportunity to play, then those folks that are the best and produce the best products typically when, and historically that's been the United States.

Starting point is 00:29:42 Yeah. One thing about this, as we talk about regulatory situations where we are in the U.S. right now, the White House has a, I believe, August 1st deadline to set new rules based on their prior executive order. And we're seeing the major American AI labs jockey. Open AI and Anthropic want other American AI labs to have to get their models reviewed by the White House before they can come out. The other companies say, hey, ours aren't as good. But Jason, we're really in this moment of sorting out the domestic policy approach to AI regulation. I know you're friends with Mr. Sachs. Is he going to win? Okay. So you've identified a real problem and you put together a solid solution and a business model that you believe in. So you're all set to launch your new

Starting point is 00:30:21 company, right? Not so fast. If you want investors and potential customers to take your new business seriously, you need to consider forming a Delaware C Corp, and that's where Northwest Registered Agent comes in. They're going to give your new company a real identity. That means an address for your public filings, a domain, a custom website, a business email, and of course a phone number. And that's going to take just 10 minutes and 10 clicks. They don't charge hidden fees. Customer service is available around the clock. They're not overwhelming your inbox with spam, and they make it easy to cancel at any time. So get all the advantages of a Delaware C-Corp, independent, regardless of where in the U.S. you're operating from. Visit northwest

Starting point is 00:31:01 registeredagent.com slash twist for more details, and the links are in the show notes. Yeah, I don't know what winning is in this context other than the government is not the best arbiter of that, but they should be kept informed. I think that's probably the most common sense approach here, which is you can be certain that there are covert efforts, efforts we are not aware of where the Department of War slash defense, the CIA are working with all these companies. And they're looking at the models. They're probably embedded at Open AI and Claude, anthropic right now, saying, how can we use these to hack the Chinese? How can we use these to keep from getting hacked by the Russians, et cetera, the Israelis? That's just common sense.

Starting point is 00:31:47 That's occurring right now. You can be certain of it. Now, to your point, which I think was a good one, Alex, this specific letter, which I'm saying is kind of performative and nonsense, is about the recursive nature of AI. So then we have to define like, okay, so if you use AI and you say, hey, AI, make AI better. Well, they're already doing that. They're already got everybody on code agents, and everybody at Open AI or thinking machines is using a code agent to write their code to make the frontier model they're doing stronger,

Starting point is 00:32:21 less brittle, more productive. So then we have to define what recursive is. Is it recursive? Like, I asked AI to get better and to do less damaging things. And then I went on vacation a week and I came back and I done all this bad stuff. So you're just going to have to have a human in the loop, which I think all these people do. I don't think they're saying to their AI like, go make the next. version of Fable and let me know how it turns out. They've got thousands of developers making the

Starting point is 00:32:51 next version of Fable and monitoring it, which is why this open AI hack or, you know, getting out of the box thing was caught anyway, because there's some human in the loop watching it work. And so we're probably doing everything right. We just, you know, sitting here as the public, essentially, or we're semi-insiders. We don't know what the CIA, the Department of War, and other other three-letter agencies have already cut deals inside of these organizations. You could be sure that the FBI or other organizations are listening. If this is actually a national security issue, they're probably ease dropping or have government warrants to monitor these companies.

Starting point is 00:33:33 I wouldn't find that out of control anyway. And you can be sure the Israelis and the Russians and the Chinese already have people inside, just like they did at Twitter and Google, inside of these organizations, stealing their secrets and or watching them. Saudi Arabia was inside Twitter for a long time. I think of memory serves. Now, what are they doing with AI to build models? I think we have some good context here.

Starting point is 00:33:54 This is from an anthropic post from June. And Jason, you said they're already using coding. You know, what's the point here? Well, I think that this is kind of the way that I think about it. So they wrote that Claude is, quote, getting good at running experiments to hit a goal. Someone else has set. And then even more importantly, Claude, their AI model, is getting better at proposing its own experiments.

Starting point is 00:34:11 So that, I think, is kind of the next step here. We're not there yet, which I think is why this letter's coming out now. People have been talking about being in the foothills of the singularity. Sure. But right now, you still need humans to your point. And I think that's why people are saying, hey, maybe we should pause. Just opening it up to the panel, I do think something happened in the last 60 days. I'm just putting it at 60 days where the velocity of what is being built and the releases are getting much better.

Starting point is 00:34:38 So in the last 60 days, we saw GROC 4.5 GLM, 52, whatever they're calling the new. open AI model soul, I guess, Fable. 5.6, soul. Yeah. Soul, a fable being turned on and then off. Kimmy and then Erme's perplexity, clock co-work, all these harnesses and multiplayer mode enterprise products have, you know, been updated.

Starting point is 00:35:04 You know, I get updated like once or twice a day across my three or four different harnesses I'm running. So I do think the velocity is increasing, which then I think gives the people. the people who are Dumers. I do think that makes sure P. Doom go up a little bit if you're inside these companies. And maybe it's understandable. I know for me, I have switched

Starting point is 00:35:23 models and said to myself, I need to kick Clawed out now because this thing is like infecting the entire company and I'm really worried about our data leaking to them. Then I'm like, okay, well, how do I do that? Okay, I'll use perplexity computer and an open source model. Okay, who's hosting my

Starting point is 00:35:39 open source model? So AI sovereignty is I think a very valid discussion. And I think, I don't know who uses the tools every day and swaps between them like I'm doing, but it does feel like it has gotten significantly faster in the last 60 or 90 days to me. I think I'll just jump in there and I'll say that whatever we see today is, you know, I don't want to put a number on it, but maybe three to six months behind what people see within the labs. and if you saw what they saw within the labs and what they meant when they said recursive self-improvement,

Starting point is 00:36:16 which is truly no human in the loop, the model gets better and better and better. And we don't understand necessarily how the model is getting better. I think we would all be quite scared for the repercussions and implications of what that could look like. That's all, I think. Okay, but D-D, there's not enough compute. So, like, I think a lot of these fears people have

Starting point is 00:36:37 about AI going nuts, and getting so good so fast, is predicated on it having the compute capacity. Everyone, as far as I can tell, is still compute constraint. So to me, there's a natural cap on AI's ability to take over because currently we don't have enough freaking GPUs. That's probably true. But at the same time, do you want to really say, like, hey, hopefully we just don't have enough GPUs to get to a place where it can be scary?

Starting point is 00:36:56 Or would you want to say, hey, like, maybe somebody should make sure that bad, like, models that can do things that are beyond our realm of prediction should probably be looked at before they are released to the wider public, before we have a very disastrous thing happen. And I'm not a doomer. My P-Doom is actually on the lower side compared to most people. But I think most people would agree that it's not zero. And at this point, at least.

Starting point is 00:37:24 And that's why whether the government regulates or somebody else, or it's a third party or whatever, regulation is usually really bad. But it is just, I'm just saying it is quite scary if you were to see what is within the labs today. One of the interesting things about recursive self-improvement is it used to be the definition of AGI. It feels like we keep on moving the goalpost on what is AGI. And it seemed, if you asked anyone really, 90 plus percent of people two years ago, recursive self-improvement was the definition of AGI.

Starting point is 00:37:56 Now it's continuing to be pushed back. So I think we are actually entering AGI as it was defined at least a couple years ago. Okay. Lo, last word from you before we move on. Your take on the situation and how afraid we should be. I don't think we should have P-dooms of zero. I think that's too slow. I want a P-dum of like 0.1, I think, for a little risk.

Starting point is 00:38:16 Yeah, you know, competition makes a global pause almost impossible, right? And so instead of slowing down innovation, we should accelerate safety. I'll leave it there. All right. Technology races don't stop. They evolve. Yeah, I don't think we're going to get China to do much here. So I think this is kind of a shot in the dark.

Starting point is 00:38:34 All right. Next up, Stripe is reportedly considering purchasing an open router for $10 billion. We've done that for a couple of days. The latest news in this saga is that the company was generating, according to the information, revenue on an annualized basis of about $140 million, quote, recently. That was up nearly threefold from April, according to the report. While the company's costs have remained relatively low, if you're a fan of the show, you've heard Jason talk about Statician size ad nauseum for years now. This is what we're talking about.

Starting point is 00:39:02 It's operating leverage. $140 million is not the most money in the world. 11 Labs hit a $500 million run rate in May, lovable, $500 million in June, and we can go on down the list. One thing that I'm just really curious about is the price here, because it does seem that $10 billion, David, compared to a $140 million run rate is a bit spicy. It's about 74, 78x run rate, which seems a bit excessive. So why is Stripe trying to make and willing to spend so much to make D.D. happy? Look, obviously open routers is extremely strategic asset. It gives developers API for 400 plus models and their routes everything.

Starting point is 00:39:38 That being said, I don't really understand. Maybe I'm not smart enough to understand the basic principles of why there's a moat here, why somebody can't come in and offer what OpenRouter is offering. And paying 70 plus X multiple is just feels too expensive to me. There's an old management consultant quote from Peter Drucker. CEOs who get bored, they start making acquisitions. So I think OpenRouter feels like the smart thing to buy,

Starting point is 00:40:02 but when you look at it from first principles and you look out what moat does it provide for Stripe, it's not evident what that is. All right, Dita, you're a board member over at Open Router. I know you're limited in what you can say, but I wanted you to pitch the Open Router's going to be enormous case. Go ahead and talk your book for a second. Open Router is the most obvious bet of all time, in my opinion.

Starting point is 00:40:22 And I completely disagree with like a lot of, you know, apologies, David. A lot of VC say there's no moat, there's no moat. Nothing has a moat. You know, if it was built once, it can be built again, right? Like, nothing really has a moat. It's just how long does it take you to build that thing again and make users switch to that thing?

Starting point is 00:40:43 And in the last few weeks, we've seen maybe 10 router releases from various different companies. Great. I think it's great that they're competing. And my question to them would be how many users would use those how many tokens are training on those, and why would people switch? There is a very strong network effect of great product-led growth businesses where you just get better over time because you have constantly done the iterative feedback

Starting point is 00:41:11 on making your product better and better and better and better. If you look at OpenRouter, that was the thesis all along. When we invested in the company, I kind of told Alex, I built a version of Open Router at Glean. I thought, hey, this is pretty easy to build. And then I realized very quickly, this is very annoying to maintain. It is extremely annoying to just have to go, hey, what's the next model coming out? What does that API spec look like?

Starting point is 00:41:39 Can I get capacity on that model? Can I put it on my endpoint? And that was just an internal thing within Glein. So I was out looking for somebody who was doing this for the rest of, you know, the developer ecosystem. And that's exactly what Alex was doing. And he had a couple of great insights whenever you speak to him. which is he never, you know, I always say, he never galaxy brained the company.

Starting point is 00:41:59 He never did this whole thing of, hey, you know, let's make a smart router that puts your right prompt and puts it to the right model. He actually said that's not the product consumers want. The product, the developers want is an API that I can change out the model name, and I can choose which model to go to. Now he supports over 400 providers. He supports all the models in the world. They do 250 trillion tokens a month.

Starting point is 00:42:24 and it grew 100x in the last 18 months. It is just a phenomenal business. And, you know, I'm only going to trust the information report. But if the information report is true, it's printing cash too. So that's my thought on OpenRouter. I don't know how many other businesses private or otherwise you could actually say that about at that level of growth. And that's kind of my bulk.

Starting point is 00:42:50 It's a cool product. I mean, I've used it. And perplexities kind of got it built in, like their own augustration level. But if you look at something like OpenAI or Anthropics, harnesses, products, enterprise software, stuff, they're never going to include other people's models. Like, why would they? I do think the data open router gets is really important. So they're definitely a takeout candidate by Google Cloud, AWS, you know, somebody like that, Azure.

Starting point is 00:43:22 And those places will probably create their own versions of this. I think they kind of have light versions of it. So sometimes when you invest in a business like this, it's maybe low margin, but if it's high growth, and they have great data and they've got some Ford model or, you know, great function to add new models quickly, that data to me seems really interesting, like which models are people using, which ones are they switching to? And so if it's got a 5% or 10% margin or markup on top of the models, I think it's like a really interesting business. And it may become a sustainable one.

Starting point is 00:44:02 It's very high margin. It's low take. And so it is a low take on the actual GMV that comes through the platform. But the margin on the actual revenue is actually. I just think margin on tokens is going to be really low in the coming years. Like I think it's just going to become a commoditized business like. you know, storage like bandwidth. So it's just like to be hard to keep the margins, so then it has to scale, which, you know, they'll have higher margins now and lower margins in the future, right? And

Starting point is 00:44:32 they'll make up for it in scale. I think this is one of the most interesting companies right now, just given how we've shifted from token maxing to token saving, the ability to be able to orchestrate, I think goes beyond price, though, actually. I mean, it's also around, you know, where's the model hosted? What's the latency? It could play roles in governance. who has the right to access certain models under which scenarios, if we start talking about, you know, security risks and threats. So, you know, this is really interesting. And clearly, a lot of players want to own this, right? Like the, you know, the service now is the sales forces. They'd love to be able to have this type of functionality. Obviously, the frontier models, but to Jason's point, they're only incented to do it within their own model portfolio, right?

Starting point is 00:45:21 the hyperscalers would like to have this technology. But I think OpenRouter has built an amazing business. I first heard about it listening to Gavin Baker on a podcast when he was just talking about usage. I was like, OpenRouter, I've never heard of this company. But I think it's a really interesting company. You know, Apple is doing something similar to a degree, right, with their advanced chips that they're putting into their phones and having the ability to take that intent and determine what's the right intelligence. It's the whole don't drive your Ferrari to the grocery store to get a gallon of milk. So if Apple can determine that, okay, this can use a locally run model on the silicon

Starting point is 00:46:01 or, hey, this is a little more complex. We send it to PCC or we send it to our partner, Google, you know, using their Google's distilled down Gemini models. So now I think, yeah, OpenRourders is a very, very fascinating business. and the data is, I think, one of the hidden gyms there. David, you should be allowed to defend yourself here. Jump in. Yeah, I don't see how this doesn't just compress to zero.

Starting point is 00:46:26 This is just this process business. This is a common business model, and they almost always go to zero over time. That being said, JCal brought up a really good point, which is the data is the actual valuable thing. And it's not clear why Stripe would care about that. And I don't think that's a great fit. But for somebody, specifically somebody like an Apple

Starting point is 00:46:44 or maybe a meta, that hasn't been able to really build, out there. AI capabilities. This could be an incredible gem. And I think that's where the sale process should be focused on

Starting point is 00:46:55 not on something like a stray. Oh, okay. Dedi, one last tiny question for you on this. When Moonshaw dropped Kimmy K3 and finally released the weights, we got to look at the license. And they released it under the K3 license,

Starting point is 00:47:06 a brand new setup. And it says that if you want to serve inference of the model, you have to have a commercial relationship with Moonshot itself. Very smart by the lab to bring revenue on short. Fair enough.

Starting point is 00:47:16 But to me, that does imply that to have competitive inference prices in the Neo Cloud space, you're going to have to compete with Moonshot's own inference at a disadvantage. So does that change the game at all when we think about inference providers accessing them, the price war, and OpenRotter's own health? Well, I mean, I don't think it relates to OpenRruder at all because OpenRrouter is agnostic of who hosts the inference. You just come on OpenRotter and list your service. You list your price and the user chooses.

Starting point is 00:47:44 So I don't think it affects OpenRutor at. business or product at all. It does affect the fireworks based 10 models togethers of the world who are going to be hosting open source models because it does have some limitations on, I don't know the exact details of the commercial agreement, but some limitations on where they can price and how they would price competitively. I think this is, by the way, something needs to be explained, D.D. So just so people know, if you were using Kimmy, here's like, I'm just logged into my account, you can use Kimmy on many. different providers. So it's very meta in that way. And I think it will, you correct me from

Starting point is 00:48:22 wrong, D.D. If I pick Kimmy here, can I have it automatically pick the lowest priced one from this list? Is that like how it works or? Yeah. Yeah. This is sort of the beauty of open router, right? So you can not only say I wanted to, you know, pick the best price, but you can also get the highest uptime of Kimmy. because it'll automatically fall back to providers. You can also say, hey, I only want providers that have zero-day retention. You can also say, I only want providers that are in America. You can have an assortment of things that you want in the actual person providing that service. It's completely under the developers control.

Starting point is 00:49:03 Yeah, so there's the thing I don't think a lot of people understand about how sophisticated this is getting. And that was maybe my poorly phrased point about the data here is super important. Because if I said it to Kimmy and I say, yeah, I just, want the lowest price or you just, you know, balance out the lowest price and the data retention for me. Now, like, OpenRouter's got a really interesting data set or function, which is to relentlessly fight for their customers to get the best deal, which, you know, you're not going to trust. And right now I'm having some trust issues with Anthropic because I turned on

Starting point is 00:49:38 Claude Tag. Has anybody turned this on in their Slack this week? Well, you're, I work at long. So you work a lot. So you, has anybody else turned it on? So I was like, everybody used this. So it basically, by default, listens to whatever channel you put it in. And then it listens using Fable. So it's sitting there watching all your stuff coming in and then doing some level of processing. And then it just inserts itself into the conversation, which is amazing. Until you get the bill.

Starting point is 00:50:06 And it's like, by the way, your token maxing 200 bucks a day above. And then they have this funky pricing where, you know, I've got like two or three people on the $200 a month. everybody else is on the $20 a month. So everybody uses their personal amount up to $20 or $200. Then it flips. So then now I'm in the analytics section, trying to figure out who's using all these tokens and how are they using them. And the pricing does not feel transparent.

Starting point is 00:50:33 And then infecting the slack, it really changes everybody's behavior. And I'm like, how can I do this with open router and maybe perplexity? And so it's got me thinking, How quickly can I get off of, you know, Anthropic? But Anthropics kind of got the best product slash integration here. So I don't know. You guys should all take a look at this quad tag thing. It's really interesting to have a persistent agent reading every message as opposed to calling it.

Starting point is 00:51:03 We say at Claude or at computer with perplexity. It's using tokens listening to everything. It's kind of the future. But I want to do this with my own model, you know? Yeah. There's a company called Int AI. are using local models to watch how people use their computers to kind of stop them from making silly mistakes as we give everyone more permissions to build. So I do think this ambient, Jason,

Starting point is 00:51:22 AI, watching employees as they work, watching you as you communicate. I think we're going to end up having very much AI mediated workplaces. And I think that's going to take a lot of getting used to, but maybe it'll feel less intrusive over time as we become accustomed. How do you feel you work for me? And like you see me telling everybody like hook up your Gmail, hook up your notion, summarize everything. because I'm like all in on this. And I'm like, if you work at a venture firm, like all your work product is theoretically owned by the firm,

Starting point is 00:51:50 if there's any value to it. And just like, let's just be super transparent. Everybody just like, what decisions we're making, etc. But I could understand some people being like, well,

Starting point is 00:51:59 can I DM each other and, you know, shit post, you know, my boss or, you know, can I talk smack about somebody? I'm like,

Starting point is 00:52:05 not on the work computers. Like, do that in your secret signals. That's what a phone is for. You call people. And then you should, And you leave no. Yes.

Starting point is 00:52:13 Yes. Thank you. Like, just like, come on, people. Get with it. I think the problem that everyone has with these tools, like bringing Slack, sorry, clawed into Slack is that we're all trying to get more work done than ever. Well, we're constantly swapping out the tools we use as new things are released. And I think this is why we're seeing a lot of burnout.

Starting point is 00:52:30 I don't know if you guys saw it. But Lilian, a co-founder at Thinking Machines Lab recently stepped back from the company and left because of her health. And she said, I can't take the stress. This is not good for me. And so I think we're at some level of AI intensity that can't. persist forever. I mean, I'm very tired. David, are you keeping up with every single freaking thing? I mean, no, right? I'm not. I use flawed. I have been investor in Anthropica. I also use

Starting point is 00:52:53 chat GPT. And it is hard at the margins when the models come out to see exactly what happened. Maybe it's an AI question. But as more models start to catch up with each other, a lot of the tasks, a lot of the tasks that we do on the media side or even on the venture side, this is not breaking physics. This is not breaking science. You don't need to use the most cutting-edge models for that. And I think there's a disconnect between people trying to do very simple tasks and very complex models that, frankly, are more expensive as well. Yeah. All right, we're going to take a bit of a pivot and go somewhere else. We're going to go over to India for a minute. Now, Cursor this week announced a program called Cursor Start. It's a lower-priced offering of Cursor, the AI coding tool for the Indian market.

Starting point is 00:53:35 This follows the release of chat, GBT Go from OpenAI, same idea, and also Google AI plus. a product aimed at a very specific price point. I'm curious, D.D., amongst the Minlo Port Coes, how big of a market India is becoming as a part of the revenue mix? Because I think we often consider the U.S., EU, and China, when we can talk about AI where there's a lot of growth. But I'm curious if India should be more in the conversation as a target market and if your companies are taking advantage of that.

Starting point is 00:54:02 Yeah, I mean, look, it's just kind of the same story as usual. What is the largest market that is English-speaking outside of the Western world in terms of volume? It's always been India. Do they have as much spending power as somebody in the West? Probably not, but they compensate for that in scale. Take this to the developers, and then the magnitude, order magnitude, even increases one more. There are way more developers in India than there are in the rest of the world.

Starting point is 00:54:28 So it is of strategic interest for a lot of people to invest in India. If you just look at the percentage of revenue, in many cases, India is organically almost at par with your revenue coming from America, which is one of the reasons that people are investing more and more. So, you know, I don't think it's a news story, but that's sort of the just of... Well, I keep reading headlines by people saying, oh, you know, we're opening an office in London. It's going to be our EU headquarters. I never hear people are like, by the way, we were going all in on Bangalore. You know, like it feels weird because it seems like such a fertile place to sell modern software services.

Starting point is 00:55:02 Well, it depends on what you're doing, right? So India is traditionally not a great place to sell enterprise. Usually when people open offices in other places, they're like setting up sales offices. They want to sell into enterprises. India is not a very fertile enterprise software market. It is a very fertile consumer market. And you could make the same argument that you want to go and set up offices for that reason.

Starting point is 00:55:21 But traditionally, people have been a little bit more hesitant, although it hasn't not been the case that they've done it before. It just does, you know, it is harder to set up an office in India than it is in London. Stupid question. Are these products ancled in any way or are they the same version as the U.S. product? I mean, it has to be ankle. Otherwise, you're losing money, right? Like, if you're charging like a tenth of the price, you're not paying.

Starting point is 00:55:41 the same cost. So I don't know the specifics, but usually it's a cheaper model. Usually it's, it's someone who can make. So what U.S. company, like if you're Microsoft, Microsoft notably and Uber, both have huge products, I'm sorry, huge workforces in India, right?

Starting point is 00:56:01 They believe in this. Dara believes in it. Sotia believes in it. Sundara believes in it. So Dara believes in it. Would you want them on an angled version and have them be like, not able to do what the American counterparts are. So then you probably don't want this product.

Starting point is 00:56:15 You probably want to put them on a VPN and put them on your corporate account in America, right? Well, they have that option, right? I mean, I think the playbook is very clear. Subsidize your product, acquire a lot of users. And then if you do have a subscription product, you eventually need to make money on it. And if you're going to charge, say, you know, whatever, $6.50, what is that? Like $10 a month. So it's not that cheap.

Starting point is 00:56:38 But at $10 a month, yeah, you obviously. obviously need to have models that will cost less than that. Maybe Groch 4.5 is a model that costs less than that. So maybe the math works. But ultimately, it's just an economics issue. It has less to do with whether you want to cut the legs on your product. It's like if you price it less, you have to charge less. Well, I think this is a really interesting thing.

Starting point is 00:56:59 Like, Alex, if you're paying your developers in India $20,000 to $50,000, which is like really strong salary, I think, you know, according to my memory in India. So let's just say it's like $30,000. You're paying somebody really, and they're spending $50,000 on tokens and they know they're spending $50,000 on tokens. How does that make the person feel? Undercompensated or perhaps completely useless? I don't know which extreme you kind of go to there.

Starting point is 00:57:25 But you can't get that much inference for $10 a month, as D.D. said. So I really doubt that anyone's spending, or I had not heard of a company in India spending five figures a year on per developer tokens. That would surprise me, just giving the economic. But I don't know, people are spending like one to one tokens in the U.S. now. Oh, in the U.S. Yes. Yeah.

Starting point is 00:57:47 Yeah. And so like I think people want to not treat. Didi, maybe you have the most knowledge. Do people want to treat the developers in India any differently than the U.S. ones other than, you know, they're in a different office. They have different. You want to give them the same snacks, the same profile of computer and the same number of tokens because you expect the same kind of output, right? If people are wrestling with this question everywhere in the world, every single startup I talk to is just like, help me wrap my head around how much I should let my developers use on tokens. Should I give them unrestricted access to Fable?

Starting point is 00:58:19 Should I not? If their salaries are, I don't know, 20K a month, should they be allowed to spend 20K in tokens or 10K or 40K? How do I assess whether that's like resulting in productive work or not? I don't have the answers, but I don't think it's a geographic question. It's really a question for every single company. Like, what is the right budget? Should I have less people on staff and spend more in tokens? I don't, we don't know the answers to this.

Starting point is 00:58:47 I mean, it is just an absolute mess and people are doing all sorts of things. I think this also low. To pull low into the conversation. I think this also goes to this philosophy. Some people have like remote workers, especially or, you know, people are working from home or working from other countries, which is, hey, you pay me a salary. You gave me a goal. if it takes me one hour, then I get the other seven hours off.

Starting point is 00:59:09 And management being like, no, that's not how it works. We're not paying you for an output. We're paying you for 40 hours a fucking week. You know, it's work 40 hours a week. And if there's gains, then do more work. Like, that's like this, I think that the tension. I don't know how you feel about it just in terms of managing people. Yeah, that's definitely the tension, right?

Starting point is 00:59:28 Because, you know, if you think about this in the context of, hey, are you paying me for output or for hours? Right. It's like, what is it that you want me to? produce if I can produce that efficiently do I get that time back to myself the American what do you think well yeah well the American philosophy historically has been you know you go back to like Madman and all that it's like it doesn't matter I'm paying you so then you do more and if you want that promotion you better do

Starting point is 00:59:51 more so that you can show that you can actually do the next level but while you're still in the current level but look we've got a new generations and they think differently that's not how they think these new generations think no I'm I'm actually going to, if you tell me I got to do this, I'm going to figure out the best way to do it. And if I'm remote, then I'm going to go to the gym or I'm going to do my side hustle. Yeah. With the extra hours, I'm going to start a side business. I mean, that's how people think, right?

Starting point is 01:00:18 People are, I think it's stealing. I think it's stealing to the corporation. And so they're, they're much more focused on themselves. Is it stealing or is it being entrepreneurial with your time? Ask them as being entrepreneurial. Whenever I've been a boss, I think, I think there's. whenever I've been in I see. That's right.

Starting point is 01:00:36 Entrepreneural use of time. Though I will just say anytime someone says, can I log off and leave when I finished my work, I just immediately put them into the be to be fired bucket. Just like, nope, nope, nope. I don't have time. No, but I mean, Alex, you're having edited people. If you were the editor of a publication and the person uses AI for their research and like, I got all the facts I need.

Starting point is 01:00:58 And then I just have to write this and I'm a fast writer. So it's, you know, it's noon. So now I can go. off from 12 to 5? No, I would say, boy, it's time to go back to every single thing AI told you because you haven't even finished the first assignment you think you have, you haven't. I am worried, this is not, we're not going to go here, but like I am really worried about people just handing off work to AI and then turning it in and not engaging with it.

Starting point is 01:01:23 And I'm just, I've seen this crop up in my life a little bit at work and outside of it. And I don't think we're ready yet for how generally lazy the average, you know, human is. This is why we invented the wheel. This is why we invented the Eric I literally just told Claude with our newsletters like the Twist Daily Newsletter. Tell me what the open rate was when it does is like daily report and check for any of the writing being AI driven versus human driven because we want it to be all human. Not because I want to catch somebody at work. I don't want our audience to catch us putting out slop. Wasn't it like some CEO put out a thread and people ran it through one of these AI checkers? Was it one of the

Starting point is 01:02:04 Airbnb founders. It was, yeah. Well, there's a bunch of CEOs who have done that in the past. Not to totally plug this company, but I love this company so much. And we just announced our investment today. But Pangram is the best AI detector. And I run everything that anyone does through Pangram, whether it's on social, whether it's internal work in a company.

Starting point is 01:02:25 Because to me, it's just a proxy of, have you actually engaged with this at all? Or did you just produce Slop? What's the name of the company again? Show it on the screen. Sounds really clear. Pangram.com. They have a one in 10,000 false positive rate. They're extremely accurate compared to many of the other AI detectors in the market.

Starting point is 01:02:44 They have an image model too that does the same thing for images. Try it. And I would actually encourage you guys, try it and try to beat it. If you can get AI to write something that Pangram can't detect, let me know. I couldn't do it. Even with the program, I couldn't get it to work. Interesting. I'm glad they actually fixed this, Jason, because these have traditionally been kind of, Didi,

Starting point is 01:03:06 I'm not trying to be rude to anyone's port goes, but AI detection software has not been robustly accurate historically. So I'm glad that we've made progress on that front. There's actually, you know, I'm really into this bit tensor, you know, Tao, Dedi. And so they have a subnet, it's subnet 32. And what they do is they have different people competing with their algorithms to check work and then a validator like Chexie. other's work and then they win like a little bit of um a little bit of tau as their reward for this but they're doing the same similar thing it's more of like an open source checker and what i'm looking for and i think somebody may have done this and i don't know if your company does it a chrome extension

Starting point is 01:03:49 that will or a browser so i can when i'm scrolling through my feed on twitter or instagram tell me when people are doing this kind of shit so that i can call them out on it It is exactly what it does. Get the Pangram Chrome extension. I scroll through my Twitter feed. I know immediately who is slop and who is not. And I just ignore all the people who write AI. That's just me.

Starting point is 01:04:11 Some people like to read it anyway. I just know. Sorry, I can't stand it. I think this is really the wrong paradigm. This is a very boomer kind of approach. Okay, boomer. Here we go. We talk about these native AI paradigms, native AI B.C firms.

Starting point is 01:04:27 You need to embrace the future now. There's certainly AI slop. You should refine your models. should do those things. There's certain things, there's certain golden cast or certain things you should never have an AI like content like writing, like writing in your in your X-speed. But specifically in terms of things that should be automated by AI, organizations that embrace AI will run circles around those that don't and have these kind of paradigms of, well, we don't want you to use AI. So I think that's the wrong paradigm for the new world. I agree. I mean, it depends on the context, right?

Starting point is 01:04:58 I always say, like, I want my competitors to do what I'm not going to do. And so if you're writing publicly and you're writing slop, I love when other people do it. I'm like, great, you're going to tarnish your reputation to a lot of people inside the company, absolutely. Code, absolutely, right? Like, I totally agree with you, David. I mean, there's like a fine line. I think you break that line when you break the trust relationship on social. Because when, you know, when Jason says something online, and the audience sees that as that's Jason saying that.

Starting point is 01:05:27 But if Jason copy paste something from Claude, then, you know, you break that relationship. Now I'm like, is Jason saying this? Is Claude telling Jason to say this? And I think that's where, you know, there is a fine line. I agree with you. It's also rude because Jason is saying that I'm not worth his time to talk to and he's still talking at me. It just implies that he's treated me with, you know, relatively arm's length and is giving me the staff essentially. And that just makes, I think everyone feel kind of small.

Starting point is 01:05:54 And AI writing is, I think as a human I can do a pretty good job detecting it. I'm shocked at how much I read today. Corporate blogs, articles online, even some news stories. It is just freaking pandemic. I expected better of everybody. I just downloaded this. I'm going to pay 20 bucks a month for this just to break people's chops. This is the greatest trolling tool.

Starting point is 01:06:16 You can't, apparently, if you want to do the checking automatically, you got to pay 20 bucks. But to me, this seems like the best 20 bucks I'll ever spend. If I catch somebody, oh my God, that's going to be so great. And this is why VCs really can add value. He came on a podcast. He sold the seat. Jason's- He did sell his seat. Bam, ladies and normally, no, look, it's telling me like I can't automatically do a scan. I need to show the floating item. And it has that I can't check for AI. I want to check for AI here. Where do we draw the line on these? Because, you know, like using things like superhuman Gmail, you know, it wants to help and write responses. People are doing agents to check their emails. I'm sure people do their agents to have some general responses, especially for something that might be, oh my God, I got to get back to this person. I'm offline. Okay, hey, I'll get right back to you. Like where should the line be in terms of, you know, when we use, or using AI to proofread. Like, hey, help me proofread this. And, Help me tweak.

Starting point is 01:07:24 You married? You got a spouse? You got a spouse? I'm divorced. I do have a girlfriend. You got a girlfriend. Okay, let me ask this question. Yes.

Starting point is 01:07:33 Yes. Are you using those quick responses? We'll cut this from the show. Nope. I do not use a letter. If she found out you did a quick response, your girlfriend finds out you did a quick response on I message or on superhuman? Sheesh, man.

Starting point is 01:07:52 I'm also curious to get. get your guys' take on this. You guys are all investors. If you get a claw generated deck, I'm sure by now you can sort of tell what's claw generated or not, what's your first reaction? Yeah, that's not a good one because I'm one where, like, when I do my articles and I do use AI to help me with research and things, but when it comes to like doing the graphics, I've tried to use Claude, but I go back, I'm old school. So I open up Keynote and I do the graphics myself. I'll go find the, you know, the images, the icons for use. And so if I were to see someone come across with the deck, then I don't know, I'm the type of person where I enjoy it. Like, I get excited

Starting point is 01:08:32 about like putting my ideas into graphical form. And so like to me, that just is another indicator that the passion this person has for their business. And so if they're not doing it, then it's like, I think I'm going to build on that low. If the graphics or like a chart was made by AI and you check the work and it's off your data, cool. If you write pros, if it's like a deal memo or it's the bullet points, why you should invest or how we think about the future, if that shit's not human written, that means you're phoning it in. So this is a good thing we did this because we can clip this and send it to founders.

Starting point is 01:09:09 I think if you wrote a deal memo, right, I love a deal memo. And if somebody on my team submitted a deal memo to the other partners at the firm, or associates and it was AI written, that's like fireable offense as far as I'm concerned. Now, if you did... Same for a founder then, right? Yeah, I mean, if I wrote a deal memo and I sent it to the founder and I was like, yeah, YOLO,

Starting point is 01:09:34 I got to go skiing for another, I got two more ski runs in. You know, I do everything in ski runs. Like, if I can get like five, six, seven ski runs in with my little VC executive 90 minute ski run in between work in the winter, Okay, cool. But if I'm trying to get two extra runs in and I, Yolo, like a Claude deal memo to you, man, no, no.

Starting point is 01:09:58 That's not cool. Now, if you use Grammally to correct my grammar, okay. Right. I think we made some progress here. Gramm really gets pretty spicy. Would that get picked up? Like if you run, if you run something you write through. Didi, what's the story with your startup here that I just did.

Starting point is 01:10:17 Well, yeah. So it does a pretty good job. Say you have a long amount of text, and then some of it is AI generated and some of it's human written. It'll actually tell you which parts. If you use small grammar updates, it's really not going to affect the actual delineation. But as you might imagine at the very boundaries of if you're greatly editing something that's AI generated, then you might be sort of at the cusp of like, hey, is it AI or is it not? But small tweaks generally don't affect the pancrime algorithm. So it's not abdication.

Starting point is 01:10:47 It's assistance. if we're going to be in AI speak. It'll actually tell you this part of the text is AI assisted and this part is fully AI generated and it does a lot more than look at like, hey, you have a, you know, M-Dash or something like that. It does like some clever tricks. I thought that was my best joke of the entire show

Starting point is 01:11:07 and it completely fall flat. So we're going to move towards the day's fun item. Now, we've talked about robotics at the top of the show, AI, how they work. Now, let's talk about where the rubber meets the road. I found something that I don't know if I believe is real or not, but I do think it shows where we're going. Is anyone here familiar with Tao robotics out of here outside?

Starting point is 01:11:23 Anybody anybody with his hand? No. Okay, so Tau robotics put out with this recently. They're launching Tau's humanoid cleaning service in San Francisco for $30 an hour. Join the wait list. All footage in this video is 1X speed. And there's a humanoid and a human operator looking at it. So my question is, would you guys hire our dear friend Tau here to come clean your house?

Starting point is 01:11:45 Jason, you've got a lot of houses. What do you think? I mean, is for a lot. I guess that is a lot. Okay, sorry. Jason, you have a reasonable number of houses. Two of them are, two of them are rented. So just so we're clear, I'm just, I'm house max.

Starting point is 01:12:03 I'm estate maxing. Okay. You're two houses. Would you use this? I have two point five because I just put the ski house into inventory on one fine stay, I think, or one of these platforms. Well, I'm using it three days, three weeks a year. I feel like an idiot. like would I use one of these?

Starting point is 01:12:20 Yeah. For my ski house, if I'm not using it, it needs to be cleaned. If he did a reasonable job, maybe. But not at $30 an hour because that's what we pay humans. And I would rather the human get the work. I don't think this is ready. I don't think this is ready for prime time. And I have like personal relationships with some of the cleaning folks and the teams.

Starting point is 01:12:42 Yeah, I would feel bad about it right now. I'm getting concerned. I'm not like a P. Doomer. But I am starting to, maybe it's just my soft heart, but I'm starting to worry about job loss in these kind of jobs, you know, in drivers. And I might be okay with deselling to get people a soft landing. Because like, so I might be a desal.

Starting point is 01:13:08 I think I'm a midterm desal on factory jobs, driving jobs, cleaning jobs. Yeah, well, Jason, there's a, Does that make me like a woke lunatic or does that make me a good person? There's a letter. You can sign this letter we have right here. And you can, you know. I think, yeah. The only reason I bring this up is you're asking me the question.

Starting point is 01:13:29 And, you know, there's, I'm invested in all the self-driving companies. Like, very well invested in them. I think self-driving is coming. I want to save the 30,000 deaths a year. But I'm also kind of like, you know, I think we should have medallions. for self-driving cars. This is, I think, a reasonable proposal. You tell me if you think this is reasonable, low, or David, or D.D.

Starting point is 01:13:53 Well, you're speaking like a New Yorker when you start talking medallions. Okay, but here's my idea for medallions. I don't want rich people buy a medallions in creating indentured servitude. I'm not down with that. Okay. You know, being an Irishman, we had this situation. What I do think would be interesting is if you want to bring your Waymos, your neuro, your, you know, whatever it is, to New York.

Starting point is 01:14:19 And we say it's $10,000 a year, and there are a thousand self-driving car licenses in year one, $1,500 in year two, $3,000 in year three, whatever it is, as long as it's safety. And we're going to auction them off to people, or we're going to just sell them for $10,000. Okay, so there's $1,000 of them, it's $10 million. Or we auction them off, they go for $25,000 each, just $25 million.

Starting point is 01:14:43 As people lose their jobs, $25 million gets put into a fund to give them unemployment if they lose their jobs over time. So it's like an unemployment pool, just like we as employers have to pay into unemployment, a little bit of everybody's salaries, and then people can tap that if they get laid off. Because I do think you're going to see like millions of people lose their jobs driving things in the next five to 10 years. So I feel like you're testing this on us because if you were to test this on your other show, you would get eaten a lot.

Starting point is 01:15:16 Well, no, I'm a free, yeah, because I'm a free market. I mean, some people are free market monsters and just like technology might is right. But somebody's going to have to pay for all this unemployment. And we gave people extended unemployment during COVID. It's not like an unprecedented thing to give people a safety net. I don't know. I don't, you know, it's interesting when we go back and we always talk about all of the different revolutions and going back to the industrial revolution, you know, model T's coming out

Starting point is 01:15:44 and automobiles taking over, what happened to all the people that did the carriages, that, you know, had had control of the horses on the carriages? I don't know if it was, I know what it is. They slowly lost their jobs because the transition took time. So when everybody talks about like everybody finds new jobs, you got to remember that you're talking up, yeah. The industrial revolution, like the tractor existed 20 years before the tractor had mass deployment. Why?

Starting point is 01:16:10 because it wasn't until, you know, people needed workers that they started deploying it. So, yeah, I do. But you're saying now this is such a compressed amount of time. I think it's going to, I mean, does anybody, what do people think the time to 50% of rides in a major city going automated is? I would put it at, if I had to pick an over under, I'd say five years. Take the over or the under. You say three. Feels about right.

Starting point is 01:16:40 Yeah. Sometimes these things move faster, but that feels about right. Yeah. So, I mean, I'm just talking about having the companies that take the jobs and move all the money to Mountain View, Waymo. They pick up the uninsurer. They pick up some of the unemployment, not all of it, but some of the unemployment of the unemployment they're creating. I'm not saying forever either. Yeah, I think a lot of this comes down to someone's age.

Starting point is 01:17:03 As we get older, our mental plus, you know, this whole trope of, you know, learn how to code for a 50, year old truck driver is much harder than 25 year old. So I think devil's in the details for older population, I think there should be some kind of off-ram, some kind of social assistance. But for the younger ones, you're only hurting them by making them do jobs that will become irrelevant in the future. You have to do a hard cut off, have them face reality, and retrain them as quickly as possible. But I do think it's a problem, we live in a democratic system. Everything's on these two-and-four-year terms. So it becomes a problem within the context of a democracy to enact this type of change. If you can move a magic wand, David, you're the mayor of New York. You get to decide. Are you doing

Starting point is 01:17:46 this proposal I put out there? Yes or no? I don't think so. D.D.? Yes or no? No. No. Alex? Sure. Okay. So I guess we got the two lip-tards on the program and everybody else is a free market But I'm a I am a capitalist through and through. Yeah. But I also watched my dad's small business struggle. I grew up without health insurance. I watched my parents cry because we didn't have any money.

Starting point is 01:18:12 You know what? Fucking sucks. I think taking care of people and providing some basic level of support is something we can afford to do. And if AI and this great revolution does not push more money into our society to take care of one another, then what the f*** is it for? Okay. There you go.

Starting point is 01:18:26 DoorDash, everybody. We all use them to bring burritos to our house. They spawned the meme. Did you hire a taxi, a private taxi for your burrito? Now, Jason, they're taking to the skies. We've had the CEO of Zipline on the show a couple of times. We've had the CEO of Manna Air from Ireland on the show a couple of times. This is happening.

Starting point is 01:18:41 But Doordash does not want to use a third-party provider. No, they have built their own drone system. And that means you can not have your brito flown to your house instead of driven like some sort of peasant. Jason, you're a big delivery guy. Are you going to use this in Texas? I have a lot of shares in DoorDash through a venture fund. I'm an LPN. and I have, I did an SPV and a late stage investment in Zipline since I missed it as an angel investor

Starting point is 01:19:05 when they were delivering blood with fixed wing systems. This is a direct response to Zipline. Zipline is putting, I don't know if we have a video of it, but like on the side of like Starbucks, they're literally adding to those like drive-through Starbucks, a place for their drones to pick stuff up. So I think Dordish has no choice but to do this. And they also have a delivery motorcycle. That came out like six months ago and that like drives in the bicycle lane at 30 miles an hour. So these things are going to be commodities pretty soon. And you have to, if you think about modes, hardware is a great moat. So I think DoorDash is going to be faced with you can only put one drone on the side of a Starbucks. And if DoorDash has Zipline do that, it's like that's their

Starting point is 01:19:51 entire business. So I don't think they have a choice but to build these. And it's going to be a free for all. Well, we need open source side of business drone delivery picky-uppy boxes. That way everyone can participate in delivery commerce. The reason why I wanted to have this on the show today is because this has gone from toy to actual product. And the reason I think that it's, we had Manna Air on and they're moving from Ireland over into the U.S. interestingly enough. And they were telling me about how popular their service is in their home nation and how it's taken over a massive chunk of local deliveries that are very, very quick using micro-pacilities and so forth. But people, People love this.

Starting point is 01:20:27 Like, I can see this. Actually, Jason, it's pretty fast than self-driving cars because it's not, not as much of a safety risk. And also, uh,

Starting point is 01:20:34 paying a human to bring me Taco Bell at 1 a.m. It's going to be, and by the way, this is going to cost five bucks. Like, all in, no tip, five bucks.

Starting point is 01:20:43 You don't have to tip robots. We don't have to say hi to them, you know? They'll still try to take a tip for the robots. I'm not paying Tony Zoo's margins. I've seen his poker games. No, sir. I'm not going to subsidize his big blinds.

Starting point is 01:20:55 All right, guys, listen, this has been a treat. I want to give everyone a chance to shout out the port code they love the most. And we'll start with you and then we'll wrap. Well, I already said it. Mine is PANGram for today. PanGram is awesome just to make sure that you're not reading AI slop when you can be reading better stuff. And it's good for images now too. And you should absolutely use it and get the Chrome extension. All right. Eat it open router. David, over to you. Legora, one of the best cultures in Silicon Valley, Paul Graham says it's the top company culture he's ever seen. incredible growth, incredible sector in terms of legal AI.

Starting point is 01:21:29 I just interviewed the founder when I was in Paris for race, smart cats. Lo, favorite portfolio company. Who's on top? Yeah, so I'm going to go a different direction since we actually also invest as an LP. And I'm going to go with Kindred Ventures. You're not supposed to have favorite children. But I do like the team at Kindred. Steve and Kanye are doing an amazing job.

Starting point is 01:21:50 All right. I like that right turn there. Jason, favorite portfolio company? You have a few, pick one. Well, I'll just because we're talking about talking so much about self-driving, we have a company, Auto Lane. It's the second time founder in our portfolio. We had invested in its first company, neighborly, which got taken out by COVID. But Auto Lane is air traffic control for the back of your target, you know, pizza hot, you know, local restaurant or your, say, mall.

Starting point is 01:22:18 And so imagine very shortly that you could send your. model Y, your own car to go pick up something you order it at the grocery store. Well, then the grocery store, let's say it's HEB or Whole Foods, they could walk out and put your groceries in the back of your own car. Or one of the things that's happening, Terry Blacks is the perfect example, the great barbecue joint, which has a terrible parking lot with like no parking and everybody going there. So it winds up being like one of these chaotic parking lots. they need to somehow deal with Waymo, Robotaxi, AV rides, Zooks all coming in.

Starting point is 01:23:01 So what the, what Auto Lane's doing is they put the power in the hands of the establishment to say, hey, these five spots are for self-driving cars, and hey, Robotaxy, you've got spot one, Zooks, you've got spot two, and they coordinated in real time. And the founders are killer, and they're running tests right now. And I think there's an opportunity. we just talked about DoorDash and Uber and like they have power over some of these and listen, I've got shares in both those companies. But I do think the establishment, you know, you've got your Terry Blacks, Terry Black should be able to set the rules.

Starting point is 01:23:34 And Terry Blacks might want to buy five self-driving cars, own them themselves and give priority to people going to the Terry Black's website, ordering from them, putting in their address, and having a Terry Black's car drive their brisket and beef ribs to them. And I think that would be a great future as well, is to enable the store owners to run their own fleet. And then imagine also you order from, let's say you order from three stores at the Grove in L.A. if you know that one. Then you could have all three of your packages brought down to your car, parked in their parking lot, put it in the back, boom, press a button, and you're all set. Is it the risk of recording the show right before we all have lunch?

Starting point is 01:24:19 is that by the end of it, we're just talking about Terry Brons, barbecue, the beef rib. The beef rib. Get the beef rib. And when you get your brisket, David,

Starting point is 01:24:29 you don't want lean brisket. You want the wet brisket from the peak, not the flat. You want the wet brisket. Keep that in mind. This is the most Texas eight words I've ever heard put together.

Starting point is 01:24:41 No, D.D, people are ordering. They're like, give me the lean brisket. Like, this is not like a Jewish deli in New York. If you want to do that at cats, have at it. But when you're, You're in Texas, the great state of Texas.

Starting point is 01:24:51 You will get your brisket wet. That means fattier, juicier. That's the way it's supposed to be. Enough with the lean. All right. Thanks for coming to my barbecue tent talk. I only have one portfolio company, but I'll shout them up because why not? Salmonrun.

Starting point is 01:25:05 com. My friend Kevin runs it. I love Kevin. Wait, what one? Salmonrun. com. Is the URL. I just figured I'd give him a little shout.

Starting point is 01:25:13 You're investing? Yeah, back to a friend of mine. Spell the domain name and show it. Sam? The latter N? Salmon, like the fish. Wow, look at you. Look at you.

Starting point is 01:25:24 It's really conflicted. I love it. This was like a year, this was like a year ago. Good. Yeah, yeah, yeah, yeah. No, your conscience count finally accurate. Oh, so it cleans up your address book?

Starting point is 01:25:32 Yeah, yeah, yeah. They built some really cool stuff. I know the guys on Madden Mark days. Great idea. So shout out to Kevin and the whole crew. Ladies and gentlemen, this has been this week in startups. It's Wednesday. It's Venture Capital Roundtable.

Starting point is 01:25:42 We got Diti. We got Lowe. We got David. We got Jason. And we got you back here on Friday. We'll see you then. Bye.

This Week in Startups - Banning Chinese robots isn't regulatory capture | E2318

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.