The Exchanges

Q Should we go deeper on sort of the infrastructure and energy point in terms of what it's really going to take to, to get enough capacity or what's most important in that second bullet you were talking about?

A Yeah, I mean, well, so, I mean, there, there are definitely people who are much more knowledgeable about energy than I am and are experts in the space, but here's what I've been able to kind of divine is, um, so first of all, the administration, President Trump has signed multiple executive orders to, uh, to allow for nuclear, to make permitting easier. We've even freed up Federal land, uh, for, for data centers to hopefully to try and help get around some of these state and local restrictions. And obviously the president has made it a lot easier to, um, to stand up new, um, uh, energy projects, uh, power generation, all that kind of stuff. Um, I still think though that we have, um, a growing NIMBY problem at the state and local level, uh, in, in the US that, that is becoming a little bit worrisome. Uh, and, um, And, ah, and if we don't figure out a way to, to address it, then it could really slow down, ah, the build out of this infrastructure. Um, in terms of, of power, so my understanding is that nuclear is going to take five or 10 years. It's just, it's just not something that we're going to be able to do in the next two or three years, so. In the short term, it really means that, like, gas is the way that these data centers are going to get, um, powered. And the, the issue with gas is the shortage there is not, I mean, America has plenty of natural gas and, um, There's, uh,…

AI assessment note: “In the short term, it really means that, like, gas is the way”

Q Let's get into real estate. Let's, let's get into Opendoor. Um, a lot. We recently had Kaz on as well. Why did Opendoor lose its way? What were the mistakes that were made? And let's talk about how we're gonna figure it out.

A Well, there was two. One was predictable. Vinod actually warned Eric and me about this in 2015, that real estate has a cyclical, cyclical nature to it. And you need to make sure that your cost structure will be Acceptable when you hit the lows of that cycle. So he's basically get as many variable costs built into your company culture and as low fixed costs as possible. And then you don't care. Like, so real estate, residential real estate, people's intuition on this is very, very off. At the low market, like when people think real estate's dead, four million homes transact a year. High in the market, six. So you have four to six million transactions a year. You need to build your cost structure that you can be breakeven or not, you know, incredibly unprofitable at four million transactions a year. And Eric and me collectively did not do that from 2000 to 15, 2019. So when the Fed started raising interest rates and it actually did raise interest rates six times in a very compressed period of time, which is the fastest rate of interest rate hikes. The company then went from five and a half million transactions a year as a market as a TAM to four. And immediately started burning money. An extreme example of this that clarifies my point, I think is what happened to Airbnb COVID the first month. Airbnb is a wonderful business. One of the best network effect businesses, maybe the bes…

AI assessment note: “Well, there was two. One was predictable. Vinod actually warned Eric and me”

Q Let's get into real estate. Let's, let's get into Opendoor. Um, a lot. We recently had Kaz on as well. Why did Opendoor lose its way? What were the mistakes that were made? And let's talk about how we're gonna figure it out.

A Well, there was two. One was predictable. Vinod actually warned Eric and me about this in 2015, that real estate has a cyclical, cyclical nature to it. And you need to make sure that your cost structure will be Acceptable when you hit the lows of that cycle. So he's basically get as many variable costs built into your company culture and as low fixed costs as possible. And then you don't care. Like, so real estate, residential real estate, people's intuition on this is very, very off. At the low market, like when people think real estate's dead, four million homes transact a year. High in the market, six. So you have four to six million transactions a year. You need to build your cost structure that you can be breakeven or not, you know, incredibly unprofitable at four million transactions a year. And Eric and me collectively did not do that from 2000 to 15, 2019. So when the Fed started raising interest rates and it actually did raise interest rates six times in a very compressed period of time, which is the fastest rate of interest rate hikes. The company then went from five and a half million transactions a year as a market as a TAM to four. And immediately started burning money. An extreme example of this that clarifies my point, I think is what happened to Airbnb COVID the first month. Airbnb is a wonderful business. One of the best network effect businesses, maybe the bes…

AI assessment note: “Eric and me collectively did not do that from 2000 to 15, 2019.”

Q So they're spiritually close to you, but they're much earlier and they dominate kind of like a company creation, whereas you, you do a lot of seed of course too, but you played all stages. Have you thought about going after that space like pretty hardcore? How have you thought about where you situate in the ecosystem?

A Yeah, you know, it's funny because, uh, we, Paul and, and us started, you know, around the same time. He started a little earlier. Um, and you know, we talked to him quite a bit during that phase when he was running Y Combinator out of his house with Jessica. Um, and you, you know, I have to say we, we never really Thought about kind of being Y Combinator. And I think like a lot of it has to do, you know, my philosophy of business is you have to start with, okay, what can you contribute that's going to be important in the world that nobody can do better than you? And, you know, for us, a big thing that we had done is we had scaled companies built into very large size. That wasn't really kind of Paul's experience. Um, but he had thought super deeply about like the very Initial kind of part of it. Um, so I think that was the right thing for him to do. And we did the right thing for us to do. And I think the world was better with us doing our thing and him doing his thing, but like, he's got a great business.

AI assessment note: “we never really Thought about kind of being Y Combinator.”

Q Why don't you go deeper in terms of what are the unlocks to actually improve housing affordability? Or why don't you give some more context on the problem?

A Well, for housing affordability, housing supply matters most of all. That's the single greatest determinant of housing prices. So one, we know that we need to, we need to build way more units. We're about five million housing units short of what we actually need in the country, and we need to add somewhere between 1.8 to two million units per year just to kind of keep that shortage from getting worse, let alone making up for that deficit. We've only developed about one and a half million units last year. So we actually need to increase our delivery by about. And, um, actually, analysts say that the pipeline is shrinking for twenty-twenty-six and beyond, so, um, they said it's gonna drop by about 50%, so we're headed in the completely wrong direction. So, housing supply matters, but in, in the short term, we can actually get more out of our current supply. So, I'll give you an example, which is almost half of, half of inquiries that go to a rental apartment building Never get responded to. So we, you've all experienced this, right? We've all experienced this, that we send a, you know, we, we want to look at an apartment, but no one, we get ghosted. No one responds to us and that apartment's still sitting there available and you want it, but it's being underutilized just because the process is broken. And so, um, if AI is managing all that demand, we can actually turn vacant apar…

AI assessment note: “for housing affordability, housing supply matters most of all.”

Q Why don't you talk through where we are today in terms of what's, what's been automated, what's not yet automated, and then we can get to what, what is the, the, the full vision? And, uh, in that full vision, what do humans do?

A A lot has been automated. I mean, maintenance, you'd have physical boards with covered in post-it notes with people trying to keep track of what needs to be done. Now that can be automated, um, triaged by AI prioritized based on urgency of those issues routed to the right technician tracked everything automatically. And we see that some operators with this new, these workflows have cut average work order completion times from Four to five days down to under 48 hours, so that's really meaningful for residents. Leasing was pretty, was maybe one of the worst. It was people just spending entire days answering emails, the same 50 questions over and over and over all day long, and, um, just the same basic information, and now AI can obviously handle that and complete that. Uh, those tasks with all of the knowledge of a building or all of the knowledge of a portfolio. Touring was another area of automation. So you'd have to go meet a broker or go meet a leasing agent and be physically escorted to every single showing. But now you can, AI can give you access or you can get access through You know, smart hardware, smart locks, lock boxes, and the AI can still be there to engage and answer all those questions and do the selling. So, um, that gives you a ton of benefits and efficiencies where you're not just having, you're paying a whole human just to unlock a single door and, um, actuall…

AI assessment note: “A lot has been automated. I mean, maintenance, you'd have physical boards”

Q a sort of, uh, business strategy perspective. You know, maybe we saw Meadow with maybe the first big open source push. Um, you know, OpenAI has sort of evolved their tune. I, I've seen even Anthropix seems to be evolving their dialogue, um, a, a little bit. Um, how should we think about open sources as a, as a business strategy in terms of what, what's changed here and why?

A Oh, look, I, I don't think this is, this part is actually, is, is playing out beautifully along the same trend lines of all previous computing infrastructure, databases, analytics, operating systems, like Linux. The, the way it works is the closed source bioneers the frontier of capabilities. It introduces new use cases, and then the enterprises never know how to consume that technology, and when they do figure out eventually that they want cheaper, faster, more control, they need somebody like a Red Hat to then introduce them And, and provide solutions and services and packaging and forward deployed engineering and all of that around it. And which is why the arc generally in enterprise infrastructure has been closed source wins applications and open source tends to do really well in infrastructure, especially in large government customers, regulated industries where there's a bunch of security requirements. Things need to run on prem. The customer needs total control over it. Broadly, you could call that the sovereign AI market right now. Lots of governments and lots of legacy industries are going, wait, this open source thing is really critical to us. So I think Whereas two, three years ago, it was open source was viewed as like this, like, largely philosophical endeavor, which it is. Open source has always been political and philosophical by definition, but now there's an ex…

AI assessment note: “now there's an extraordinary business case for it”

Q Will we see more vertical integration or more horizontal specialization?

A You know, historically, we've seen both, and what's interesting is we're already seeing both now, right? Like Apple, of course, has just been historically vertically integrated. Uh, Microsoft and Intel historically horizontally. Uh, and often, companies will start horizontal and then go vertical. So like, um, you know, Google is horizontal. I mean, it was built on top of normal servers, but then they built their servers, and then, you know, they, you know, They built their own chips. They built their own networking gear, and so I, I think you always get a mix of the two. What's interesting about now is we're actually really seeing both. I mean, I would say that OpenAI is very much a vertically integrated company now with ChatGPT driving a lot of it. I would say Anthropic, a lot of the usage really is more horizontal, and they're doing a great job of that. I think we're seeing this on the model layer too. I mean, a very interesting discussion we haven't had, but it's a very interesting one is like, open source quote unquote really seems to work with these models just because you can't, as a user You know, recreate it. So like, if you look at like BFL, they've done a great job building kind of like a horizontal layer for these models. Um, but then you've got companies like Ideogram, which have built a great kind of vertical experience as well. And so I would say for AI, we've got…

AI assessment note: “for AI, we've got already this early on great examples of both.”

Q When did you realize that this was the way you were going to build a company? That, that, like, how did you intuit, like, hey, distribution is a scarcity, distribution is what matters, and that, because there's a lot of creators out there, but they're not combining it with a tech company. Like, how did you, how and when did you put this together?

A I guess, I guess there was a certain point where I kept going viral that I sort of realized that I know something that ex-LinkedIn people don't know yet, and it is sort of like mastery of the algorithm. Um, Um, and I think, like, the, everything started with the Interview Coder situation. Um, Interview Coder was the earliest prototype of Cluey, and it was a tool to let you cheat on technical interviews. And I used it to cheat my way through an Amazon interview. I made it super public. I posted it everywhere and ended up getting me, like, blacklisted from Big Tech and kicked out of school. Um, and that situation was inherently viral. Like, when's the last time someone got kicked out of an Ivy League and raised five million dollars? Like, this has probably never happened in the history of humanity. Um, so that situation was inherently viral. And at that time, I had no idea that this was, like, a repeatable thing that I could do. Um, but then the launch video happened, and I had my intuitions about the virality of launch video, and I just kept scrolling on Twitter, and I was wondering, like, man, why is nobody doing what Avi Schiffman with friend.com showed the world you could do a year ago? Like, why has nobody done this, done this yet? And it worked. And then I did the 50 interns thing, and it worked, and like, like, I kept doing viral video after viral video, and at a certain p…

AI assessment note: “there was a certain point where I kept going viral that I sort of realized”

Q When did you realize that this was the way you were going to build a company? That, that, like, how did you intuit, like, hey, distribution is a scarcity, distribution is what matters, and that, because there's a lot of creators out there, but they're not combining it with a tech company. Like, how did you, how and when did you put this together?

A I guess, I guess there was a certain point where I kept going viral that I sort of realized that I know something that ex-LinkedIn people don't know yet, and it is sort of like mastery of the algorithm. Um, Um, and I think, like, the, everything started with the Interview Coder situation. Um, Interview Coder was the earliest prototype of Cluey, and it was a tool to let you cheat on technical interviews. And I used it to cheat my way through an Amazon interview. I made it super public. I posted it everywhere and ended up getting me, like, blacklisted from Big Tech and kicked out of school. Um, and that situation was inherently viral. Like, when's the last time someone got kicked out of an Ivy League and raised five million dollars? Like, this has probably never happened in the history of humanity. Um, so that situation was inherently viral. And at that time, I had no idea that this was, like, a repeatable thing that I could do. Um, but then the launch video happened, and I had my intuitions about the virality of launch video, and I just kept scrolling on Twitter, and I was wondering, like, man, why is nobody doing what Avi Schiffman with friend.com showed the world you could do a year ago? Like, why has nobody done this, done this yet? And it worked. And then I did the 50 interns thing, and it worked, and like, like, I kept doing viral video after viral video, and at a certain p…

AI assessment note: “there was a certain point where I kept going viral that I sort of realized”

Q And, and that was celebrated at the time, right? Because we were excited about sort of more integration with China. So that wasn't seen as such a problematic thing at the time, right?

A Oh, it's, it's absolutely, it's even more than that. Like, this was like hailed as like, this is the modern way to do business. Like, uh, like a very An example of this that's just super close to home about how things have changed was that in 1999, the World Trade Organization gathered in Seattle to, um, ratify China as a member of the World Trade Organization. And this is the organization that, like, that sort of navigates tariffs and trade rules between countries. All the cool countries were in it, but not, like, communist, socialist China, you know, without With, you know, a totalitarian government and all that, and so the Clinton administration had really pushed for China to come in, and it actually split the Democratic Party. People today, their heads would explode if you tried this out, and it turns out the Republicans were split too. Yeah. Because half the Republicans were like, free trade, free trade, free trade, and the other half were, were like, we hate communism, we hate communism, and, and, and so the whole world, if you look back in 1998, it look, it's like completely upside down. Yeah. From today. And, and there were huge riots in Seattle. Like, the city was like a mess over this moment, because it was viewed as this basic, like, this, this pro-business, anti-labor kind of thing, because it just meant cheap outsourcing in China. Now, the reason for that was becau…

AI assessment note: “Oh, it's, it's absolutely, it's even more than that.”

Q there was a, there was a breakout starting from Facebook, Twitter, Instagram, Snap, WhatsApp, Tinder, TikTok. Every few years there was this sort of new paradigm, this new breakout, and it feels like at some point that just a few years ago that just stopped. Why did it stop or did it stop? Would you reframe how we should think about that? And, um, where do we go from here?

A I would argue probably ChatGPT was a, was a huge consumer, like, outcome and winner in the past few years. And we've also seen a bunch of other ones in, in various other, like, modalities of AI, um, in, like, image and video and, and audio companies like Midjourney and, and 11 Labs and Blackforce Labs, um, now things like Kling and Veo. Um, mo, like, weirdly though, a lot of them don't have the same, like, social or traditional consumer dynamics that you mentioned. I think because AI is still relatively early, and so much of the new products and innovation has been driven by research teams who are, like, so good at training models, but historically have not been amazing at creating the consumer product layer around them. So I think the optimistic view is that the models are now mature enough, and many are available either open source or via API, um, for people to build great, more traditional consumer products on top of them.

AI assessment note: “I would argue probably ChatGPT was a, was a huge consumer, like, outcome and winner”

Q One thing I heard you guys say is that one surprise that you guys, uh, sort of realized was that enterprises are sometimes adopting these, these products first before consumers, which feels different from, from previous era, or maybe not what we expected. What can we say there?

A Yeah, that has been fascinating, and BK and I saw that a lot with 11 Labs, which, um, we, uh, were relatively early. I think we invested, we did the Series A like a month or so after the initial launch, and I think what we saw was first the, um, early adopter consumers got on board, and they were making memes, they were making fun video and audio, they were cloning their own voices, they were doing game mods, um, but then I would argue it hasn't even gone in many cases to the truth True mainstream consumer. Like, it's not yet like every single person in America or most have 11 Labs on their phone or have a subscription, but the company has these massive enterprise contracts and a ton of huge customers across, like, conversational AI, entertainment, tons of different use cases are using 11. And I think we've seen this across a bunch of AI products, which is like, there's an initial consumer virality moment, and then that actually leads to lead generation in, Enterprise sales in a way that we did not see with the last generation of products. Like, enterprise buyers, there's so much of a mandate to have AI now, an AI strategy, and use AI tools, that they're watching places like Twitter and Reddit and all of the AI newsletters, and they're saying, like, hey, this is some random, looks like a random consumer meme product, but I can actually think of a really cool application of that…

AI assessment note: “an initial consumer virality moment, and then that actually leads to lead generation in, Enterprise sales”

Q And in fact, most healthcare companies today are the opposite of that. And actually some people might describe healthcare today as intimidating, regulated, and complex. And in fact, you two both have referred to it somewhat in that fashion. And so what do you think the tech industry in particular is set to Do in a way that, you know, others haven't in the past to address this unique combination?

A Yeah. I think I'll, I'll address those things head on, which is people often say healthcare is intimidating, which is often a euphemism for healthcare is really big, which is in venture a good thing. It's great to go after big markets. Healthcare, US healthcare is five times the size of the global advertising industry, which is what companies like, you know, Meta and Google make all of their money from. Um, and it's highly regulated, but tons of huge companies have been built in highly regulated markets. Um, you know, if you look at Airbnb or Lyft or even companies that have been built in unregulated markets like Google and Facebook eventually become regulated. So we think regulation is kind of a marker of success, and you run into it regardless, and it's a good thing in healthcare. Um, and healthcare is complex, but every industry is complex when you dig under the surface of things. And we think you really need technologists to solve healthcare problems. Everyone thinks like, you know, healthcare, don't I need a PhD? Um, you know, I don't know how to cure cancer. And the PhDs who are working to cure cancer are doing extremely important work. But the reality is, if you cure all cancer, you extend American's lifespan by three years. And we're still lagging behind other developed countries. And so you need technologists to come and do two things. One is like health care is a logi…

AI assessment note: “health care is a logistics, a data, an operations problem, and technologists are very good”

Q Andres Guido, uh, we want to talk about sovereign AI, uh, AI and geopolitics, and let's start with the news. Uh, our partner Ben is in the Middle East right now. Um, what happened, uh, and, and why is it so important?

A What happened is, um, the kingdom announced that they're going to build their own local hyperscaler or AI platform called Humane. And I think why it's notable is that they are As opposed to the status quo of the cloud era, um, they're viewing the AI era as one where they'd like the vast majority Of AI workloads to run locally. If you can, if you think about the last 20 years, the way the cloud evolved was that the vast majority of cloud infrastructure basically existed in two places, right? China and the US. And the US ended up being the home for the vast majority of cloud providers to the rest of the world. That doesn't seem to be the way AI is playing out. Because we have a number of frontier nations who, who are basically raising their hands and saying, We'd like infrastructure independence. The idea being that we'd like our own infrastructure that runs our own models that can decide where we have the autonomy to build the future of AI independent of any other nation, which is quite a big shift. Um, and I think the headline numbers are somewhere in the range of, um, a hundred to two hundred and fifty billion worth of Cluster build out that they've announced, of which about 500 megawatt seems to be the atomic unit of these clusters that they're building. So what's going on? A number of countries, the, with the kingdom being the one that's most recent, are, have been announcin…

AI assessment note: “the kingdom announced that they're going to build their own local hyperscaler”

Q And is this something the U S had internalized in 2009? Is that, is that why you called the firm injuries and Horowitz when any, every other firm was going?

A No, no, that was a different thing. Um, so what happened then? So when we were raising the money, uh, and it was, you have to remember it's 2009. So it was a difficult year to raise venture capital. Um, in fact, I think there were only two new funds raised that year. There was SARS and, uh, Khosla. Um, so the, the biggest, the number one objection we got on the fund was, You guys are very successful entrepreneurs. What's going to stop you from going out and quitting doing this and just starting a company? And then we're going to be left holding the bag and nobody's going to be investing or watching our money. Um, and we had no plan to do that. So, uh, we got the idea. Well, one easy way around that is just name the firm after ourselves. Then they'll know that we're going to be tied to it forever. Uh, and we did that, and then, um, I had the idea that since nobody could spell Andreessen Horowitz, we should, uh, have this A-sixteenth thing, and that, that was the name of the firm, and of course, immediately, uh, all the competitors, uh, said that we were egomaniacs and, like, narcissistically insane because we named the firm after ourselves, which we, we just ignored. Like, what, what couldn't we do? You know, maybe they have a point.

AI assessment note: “No, no, that was a different thing. Um, so what happened then?”

Q One interesting stat that you guys shared was the percentage of YC companies that are now pursuing AI voice. What are we seeing there in terms of how cohorts have changed and the percentage of, I guess, these new companies on the frontier actually pursuing this field?

A YC founders are typically, you know, young, high hustle, ambitious, and they're like heat Seeking missiles. And so they will pivot until they get into a space that's interesting. So in recent YC cohorts, upwards of 20, 25% of companies are, are building with AI voice, which is really exciting. We're even seeing a lot of companies from past cohorts all the way back to, like, twenty-nineteen, twenty-twenty, are going back now and pivoting into AI voice. The first wave after the infrastructure companies in voice we saw were pretty horizontal platforms that allow Anyone, any business, any consumer to build kind of a broad-based voice agent. Like, I built one that called the DMV for me and scheduled an appointment.

AI assessment note: “upwards of 20, 25% of companies are, are building with AI voice”

Q when they think about technologies replacing humans, and really this idea of augmentation as well. And you mentioned it also in the scenario of, like, let's say you take some company that, you know, only has a receptionist nine to five, what about five to nine or 24 seven? Can you talk a little bit about how you're seeing these AI companies wedge in and kind of start the engines?

A I would say a lot of businesses, I mean, small businesses to enterprise alike are, for their own reasons, like nervous to hand over all of their phone calls and customer interactions to an AI. So we'll often see these voice agents start with a specific wedge that just feels so obvious in terms of ROI to the business, and then as they gain trust, expand from there. So one of the most obvious and easiest ones are these after hours or overflow calls. So if you're a small business, you probably live or die by the ability to kind of get an appointment booked. Having that handled by an AI is kind of a no-brainer. Like at the very least, they can get a phone number and information and call back, but maybe they can actually book a full appointment for you and have like a job, you know, on deck for the next day, which is awesome. But beyond that, there's a lot of other kind of clever things that I think we've seen, uh, companies do. So there's some calls that just don't make sense. To make right now if you're paying human labor. Like, if you're a credit card company, you send out a credit card, uh, and the consumer never activates it, does it actually make sense to call them after one or two or three days and get them to do that? We see, I've seen a couple voice agents that are really successful now with that use case alone. Anything that's back office, it's not client facing, so it's, …

AI assessment note: “we'll often see these voice agents start with a specific wedge that just feels so obvious”

Q Yep. Since we're early days, what's your instinct about moats, right? That's, as you mentioned, like, that's true across the AI ecosystem, not just voice. Yeah. But where do you see moats potentially arising in this sphere?

A I see moats in a couple ways. So one would be integrations, and, and this is, I think, why we're especially excited about these more vertically focused voice agents, It's not gonna make sense for OpenAI to go integrate with every long tail, you know, transportation management software that a fleet company is gonna, or a freight company is gonna be able to need to run their, you know, fleet of, of trucks on a voice agent, uh, product. And similarly, UI. Like, OpenAI and, and other companies have a pretty set, you know, um, system for interaction right now that doesn't work the way that many of these, like, Heavily legacy businesses, uh, want to be able to operate. One of the types of moats that has been the most intriguing for us, I would say, especially for enterprises, is kind of this self-improving data moat. So if you are going to take over calls for, say, a large bank, they have a certain way that they want those to be done, and so you're not going to plug in a voice agent and have A hundred percent NPS on day one. It's gonna take months and months of training calls to make that better. And so you, as a voice agent provider, if you get in early, ah, benefit from having all that special proprietary data that just gives you months of a head start for anyone else who has to come along and go through that entire onboarding and integration and training process. And so I think th…

AI assessment note: “I see moats in a couple ways. So one would be integrations”

Q One interesting stat that you guys shared was the percentage of YC companies that are now pursuing AI voice. What are we seeing there in terms of how cohorts have changed and the percentage of, I guess, these new companies on the frontier actually pursuing this field?

A YC founders are typically, you know, young, high hustle, ambitious, and they're like heat Seeking missiles. And so they will pivot until they get into a space that's interesting. So in recent YC cohorts, upwards of 20, 25% of companies are, are building with AI voice, which is really exciting. We're even seeing a lot of companies from past cohorts all the way back to, like, twenty-nineteen, twenty-twenty, are going back now and pivoting into AI voice. The first wave after the infrastructure companies in voice we saw were pretty horizontal platforms that allow Anyone, any business, any consumer to build kind of a broad-based voice agent. Like, I built one that called the DMV for me and scheduled an appointment.

AI assessment note: “in recent YC cohorts, upwards of 20, 25% of companies are, are building with AI voice”

Q LLMs that are actually helping people vibe code. There are entire Reddit communities behind this. And on the back of that, there are these companies that are being built that help people vibe code. So who's actually participating in this trend? What are the companies that are being used? And how would you also kind of like frame them in terms of like, are there different categories within this trend?

A Yeah. So we've, um, we've seen different companies serve different types of users. So I think there are some, um, more sort of like IDE based companies like cursor that, um, are targeting developers and just making it, uh, a lot easier for them to code by giving an agent a prompt and then it writes, writes code or edits code for you. Uh, and then we've also seen a new emergence, um, both for non-technical and also being used by technical users. Of these like text to web app or text to website companies, um, that are in the browser. And like you literally go, it's, it's a web app. You type it in, you say, Hey, I want an app to track if my dog has been fed, or I want a website for my small local business where people can find information and, and contact me. And then you sort of get this interface where usually they all look quite pretty similar on the left side, you prompt it. And then on the right side, it kind of shows you, um, What it's generating, what the interface looks like, and then you can say, no, I want this button to do this, or no, I want the design to do that. Um, so we've seen a ton of companies emerge in this category. Um, I think the biggest ones we've seen thus far are probably like the Replit agent, um, Lovable, uh, Bolt and V zero from Vercel. Um, but we're seeing more products emerge here every day.

AI assessment note: “we've seen different companies serve different types of users. So I think there are some”

Q Yeah. And give me a sense of scale here, right? Because I think maybe cursor is becoming more of a household name. A lot of people refer to it as one of the fastest growing companies over the last two years. But what about some of these others? Like, are people really using these tools at scale or are we really in the early innings?

A Yeah, it's a great question. Um, not every company has released their metrics. I think Bolt and Lovable both have either tweeted about it or, or talked about it on podcasts and are, are saying things like getting to twenty million in ARR in two months or ten million in ARR in two months, which is an insanely fast ramp. Um, and so I think to us that indicates like there's a ton of latent demand for people who want to use these, um, tools or want to code or make something for the first time and are, are finding A ton of value and being able to do it like in a more accessible way. Obviously, um, there's also a ton of demand from developers to, to code more easily because cursor is growing incredibly fast as well.

AI assessment note: “Bolt and Lovable both... saying things like getting to twenty million in ARR”

Q just never happened. So tell me more about that, because it does seem like, okay, it seems like you could maybe create a static web page, maybe something like a tracker, some of these more early stage, like websites, maybe not full blown apps. Am I getting that sense? Or am I wrong about, like, where these tools are, and where they end, perhaps, and in what you can build?

A Yeah, I think the difference with like the, the vibe coding products versus, um, products we'd seen before to help non-technical people generate websites like a Squarespace or a Wix, um, is that they're actually writing code so they can actually like do more kind of interactive and dynamic things. Like, um, one example is like even me with Lovable made a web app where, um, you input a book that you liked growing up, It does a call to the OpenAI API gets a response about like three books you might like now based on the book you liked growing up. Um, then you can save the book in a database that I like prompted within lovable. Um, and then you can even log in with your Gmail account, which like a, a Google authentication API. Um, if you want to like save books and then rate them. And so that sort of thing, I think is beyond like far beyond just like a basic Static website. It's like truly a dynamic web app. Um, that was a little bit harder than just like pure text prompts. It required me going and, I mean, the LLM would explain it to me. I would say like, hey, I want to add authentication and it would be like, okay, here's the five steps, but you do have to click out to this external link and you do have to be able to like copy in these codes and keys into one places or another. But someone like me who's non-technical was, was able to do it. And so I think our belief is it's just…

AI assessment note: “far beyond just like a basic Static website. It's like truly a dynamic web app.”

Q past. Can we just quickly revisit that idea of the why now? Like, why are these tools so good? Are we a little surprised even that these LLMs are so good with code? Because that, to me, would have not been intuitive coming in. And maybe just tell, tell us a little bit more about like the building blocks that led us to having these tools in the first place.

A Yeah. So I guess I'm on the very lowest level of layer, which is the foundational models. All of these tools are powered by very good coding models out there. So why are the models good? One, obviously it's like transformer architecture and everything, but two, I think it's the placer of the data. The data distribution is there on the internet. Most of the apps nowadays are JavaScript apps. Uh, like if you go on Stack Overflow before AIH, most of the question people Questions people ask are like, why can't this Node.js app work? And then you get answers for that. And now there's a lot more like frameworks that's very well defined, like Next.js, React, right? So you really travel up the abstraction in coding. So as a result, when, you know, like you train a coding model with the data, there's just way more examples of web frameworks, web data. So as a result, it's very good at, you know, building those apps. Another reason I actually think is Web Based developing environment now, it's very mature. So before you may need like a front end, which is like HTML, JavaScript and back end like, um, Rust or maybe back in this sheet C sharp.net or something. But nowadays, like the full stack apps are all JavaScript. Like the most of the apps we see that are being generated are all JavaScript and TypeScript based. So it's in itself a very contained runtime. The agents can automatically ver…

AI assessment note: “So why are the models good? One, obviously it's like transformer architecture”

Q as well in terms of how these work, right? Is it, it might, some people might think that this is just like an API call to a foundational model. Is it that, or is there more built on top? And then adding on to that, how do these different companies add value right past the integration with the foundational model? And how do they differentiate, you know, across the competition?

A Um, these are actually very sophisticated systems when we, you know, lift the covers and look into how they could be implemented. So, um, the consumer interface is always like, like Justin mentioned, there's a generation of the code, there's a preview, there's a prompt window. But then when you kind of lift the cover, what you had to do is that it has to have a execution environment for the agent to actually work in. So it needs to paint the foundational model like, hey, My user have this prompt, like, give me these examples of a code, and then you have to, um, you know, the agent will have to tell the foundational model, like, generate the code only to this, you know, these set of standards. And then once that happens, it runs this code, you know, in the execution environment, which is mostly a browser. Nowadays, it's either something called web container, which is like a very cool technology that leverages your own, you know, Uh, laptop as a computer, and then, you know, spin it up, uh, like locally, or it's, you know, running on the server somewhere, so it, uh, produces that preview, uh, for the user. And then, uh, kind of like what we discussed on the component side, like a lot of apps can't be built without a database. Uh, so you just have to persist the data somewhere that's either in memory in a browser, or it's, you know, more persistent in the actual database, like sup…

AI assessment note: “these are actually very sophisticated systems when we, you know, lift the covers”

Q Kimberly, you wrote an article with a pretty fun title, RIP to RPA. So let's jump into that. But first, what is RPA?

A RPA stands for robotic process automation. Um, and it's a way of basically automating very manual tasks within an organization. So things like data entry or invoice processing that basically every business has to do, but it's nobody's core competency. It's just one of the, like, Dirty, messy internal things within an organization that everyone has to do. So historically it's been done very manually. Like you would just hire a data analyst or you would hire a back office operations person. Um, and there was this like, I would say innovation in the last 20 years where people were like, is it possible to automate these tasks? And so the historical way people have done it is through robotic process automation where you basically build like a little software bot that mimics the actual clicks that somebody would be doing. It's very deterministic, meaning, like, they're literally clicking the different, like, boxes that I would be clicking as a human. But, you know, like, organizations are messy, and the work we actually have to do is not perfectly delineated by a very specific, like, process. So oftentimes, if something veers a little bit off course, like maybe someone misspelled a name, or maybe a website changed where the sign-in box physically is on a page, then historically, that would break the RPA process. And as you can imagine, there's, like, An infinite number of small littl…

AI assessment note: “RPA stands for robotic process automation. Um, and it's a way of basically automating”

Q Normally I ask the question, why now? But I feel like, you know, listeners know that AI is coming, it's here. LLMs are maybe the term that a lot of people use. But is there a deeper why now or specific technological advances within the sphere of LLMs that you can point to that actually make this possible?

A Yeah, I think one thing that we're really excited about is, you know, people use the term AI and they're like, oh, everything's going to change now because of AI. But like, what does that mean? You know, there's a lot of very distinct technological breakthroughs that make different applications possible. And specific to intelligent automation, I think one of the things that makes it much more possible than before Is a lot of the fundamental research coming out of the large labs. So for example, recently, Anthropic announced computer use, which is basically a browser agent that is able to intelligently understand what is happening on the browser level of any sort of desktop and be able to take actions accordingly. So, you know, we talked about how historically RPA basically understood at a pixel level, hey, I should click this thing and then I should click that. But with something like computer use, or I think open AI has something called operator that they're gonna release soon. Agents are gonna be able to browse the internet and browse the web in a much more sophisticated way, which is gonna open up a lot of possibilities for what intelligent agents can do before. So we think a lot of these intelligent automation startups, they're not gonna be doing fundamental research on their own. You know, there's still tech that needs to be done to make a browser agent fully work at scale…

AI assessment note: “Anthropic announced computer use, which is basically a browser agent”

Q What can intelligent automation, or what you refer to as these LLMs in action, what can they do that RPA couldn't?

A Let's use the example of, um, a company that we were actually invested in called Tenor. Um, Tenor does referral management. For healthcare practices. So if I'm a primary physician and I need to refer a patient to a specialist, historically, the way that that would be done is I would literally write something out on a piece of paper. I would fax it to the specialist. The specialist front desk person would take the fax, look at it, look at all the information on it, and then input it into my own database, check, you know, like the insurance policies, check prior history, et cetera, and then decide whether to accept the patient or not. And that was a very manual task that there's just a little bit too much complexity in the way that it's done for RPA to be able to handle. So it had to be some sort of administrative person like human who was going to do it. And with now like intelligent automation, um, tenors come up with a very sleek solution that is basically able to automate that whole process. And it's much more self-serve. Yeah. Because the way that RPA would historically work is you would have to hire like an implementation consultant or something, and they would sit Next to whoever was doing the task, and they would basically just watch, like, what are the clicks that you are doing?

AI assessment note: “too much complexity in the way that it's done for RPA to be able to handle”

Q Yeah, and as we think about all the different jobs that are out there, you see all these charts where people are like, This industry is going to get hit really hard or is going to shift in this way. Do you think about marketing in particular in this Gen AI wave? Why do you think that industry or jurisdiction is uniquely suited to maybe be shaped by this?

A Yes. Marketing almost felt like the most, one of the most obvious use cases of Gen AI, Gen AI when it first came out. You're creating content. So whether using Jasper, Midjourney, any of these platforms, um, and then The natural sort of B to C to B to B evolution is, okay, we're creating an image. Now let's create an ad. Um, and so, you know, LLMs to that point are very good at generating content, but they're also good at gathering and synthesizing sets of data, um, and enabling, you know, hopefully end to end actions. Um, so what does that mean for marketing in particular? One, you know, the most obvious is creating marketing content, whether that's emails and blog posts or ads. Um, Um, and then there's another piece, which is enabling better research and the ability to collect and understand your customer. So that is now being able to be done on a more dynamic and real-time basis versus, you know, sending a bunch of surveys out and then collecting the data later.

AI assessment note: “Marketing almost felt like the most, one of the most obvious use cases of Gen AI”

Q I mean, I think I've seen some data that you, you've all shared on Just how many visits or interactions some people are having? Maybe we just talk about that, because it's, it's kind of shocking in a way, but also not surprising in a way.

A Yeah, absolutely. So on, for the mobile products, as well as the web products, but especially interesting on mobile, because you can really granularly track a user and say, how much time are they spending per day, per month, how often are they coming back? And the insight there that I was, like, shocked by, and I say this as a big fan of the company, but Character AI has the average user Close to 300 sessions per month, which again is like 10 sessions per day. This is like iMessage usage, Snapchat usage, Instagram usage Really, yeah.. So I think we're starting to see like, yes, maybe the product isn't for everyone right now. Maybe, you know, it hasn't reached the true, true mainstream, but it's getting there, and the people who are using it are really using it and clearly finding like deep, deep value from it.

AI assessment note: “Character AI has the average user Close to 300 sessions per month”