The Exchanges

Q People thought the model was gonna take over, right?

A Yeah, look, the model sounds really, um, um, um, reasonable, and I think logical on face value, where if you can, you know, on that spectrum of restaurants where you have, you know, high-end service on one end and hospitality only to perhaps delivery only, or more of a manufacturing concept on the other end, It seems reasonable that you can actually just borrow a small square footage of space, um, not incur a lot of, not just the fixed costs, but also the labor costs of actually running your restaurant, in quotes, and then selling through a delivery, you know, platform, or acquiring your own customers, or, or doing something like both. It just turns out it's extraordinarily difficult, however, unless you're a large brand, or, you know, um, you know, a, a house of brands, someone like DoorDash, To be able to attract enough customers to make that math work.

AI assessment note: “It just turns out it's extraordinarily difficult, however, unless you're a large brand”

Q That's super cool. Um, what is the typical adoption? Are people using it for, you know, email chat support, because that's the easiest modality. Do they, Adopt us for everything, including phone and stuff like that.

A It's changed a lot over the past two years, but I'll say the median customer and they'll describe some like interesting outliers, which I hope are sort of glimpses of the future. So most will start with one channel and a few use cases. Um, so, you know, at a lot of healthcare companies, phone remains the dominant channel. So say, hey, for a few types of phone calls, let's have the AI agent take them and see how it does. Do people like it? Are people comfortable with it? Does it lower our cost? Does it raise whatever, you know, metrics? Usually it's customer satisfaction. Um, and does it work more effectively? So for example, uh, like for a car insurance company, it'll be like first notice of loss. You know, I got an offender bender, you know, and that would be the typical way you start. Um, for a lot of more digitally native companies, they'll start with chat. Um, and it, uh, and kind of similar. Um, but almost all of our clients will do both. Um, so SiriusXM, if you call them on the phone, their AI agent, Harmony, which I love that name for SiriusXM, will pick up the phone, and if you go to their homepage and you see the chat, that's also the same agent. So the neat part is, I think it's pretty neat, because you have, like, literally, you have all of your, I'll say, customer experience team, or, you know, whatever you might call it at your company, they can spend all their tim…

AI assessment note: “most will start with one channel and a few use cases”

Q Where do you think has passed sensible rules?

A Um, sensible. Um, I think Virginia's bill last year was pretty good. Um, it defined, it did a few things well, um, and one thing I don't agree with. What it did well is it defined, uh, a modest data retention period of 21 days. I think that's fine. I like 30, but tomato, tomato, it's fine. It wasn't, I think the ACLU was lobbying for three minutes. It's a little tough. It's like hard to swallow. Uh, I think, you know, seven, 14, twenty-something days is like enough. There's a trade-off there. Um, they mandated, uh, formal auditing, which I think is great. Enough of our, not enough of our customers audit themselves on a regular basis. We can build software to make that easier, but we need to be pushed to do that. It was like, customers don't want it. They need to be told to do it. So I think that was good. Um, it also validated that this can only be used for criminal investigations, which I think is really good. Well, that's, That's obvious. It's helpful to write it in law. I think the only thing that I disagree with is they did say, you know, effectively, there's no participation with the federal government. And I think that's just, it's their choice. I think it's their choice. And that's the beauty of the country is like, Virginia should do what feels right for Virginia. But I worry about the types of cases that you, you don't want to read about on the news that tend to get so…

AI assessment note: “I think Virginia's bill last year was pretty good.”

Q That's very interesting. Last question, you know, so you guys have grown with cameras out there, uh, in cities now getting into drones, uh, building the software OS. To help law enforcement agencies and others kind of synthesize all the information they have. Just what comes next? What future product ideas are you playing with? We're doing it now.

A So I think about it, um, we talked about this earlier, um, failure for flock is prison population goes up. It's actually like really bad. Um, and we look at, you know, the products today are very much focused in the middle of a crime. A crime has already happened and therefore we should solve it. And that's really good. And I think we're, We're definitely not done, but we've, we've done a lot of work in that category. I get pretty interested in expanding that and going, well, what about, what can we be doing from a product perspective to prevent crime from happening? Um, and that actually doesn't necessarily look like software. It's like one of the interesting things that we started last year is what we call our Thriving Cities Fund. It's probably an analogy similar to your, like, Stripe Press, which is like, it's never going to be the core of your business, but like, you feel really good that it's a part of your business. And so when we go in places like Greenville, Mississippi, we also commit to deploy capital as, as growth partners to those businesses. Because if we want to convince that sixteen-year-old to not be a criminal, there does need to be jobs. Jobs that like a sixteen-year-old can get. And so we deploy capital in, you know, restaurants, nail salons, like, pick your business that you can be 16 and work at easily. And like, we want more of those to exist. Um, you kno…

AI assessment note: “what we call our Thriving Cities Fund.”

Q And so in 2018, the reason Netflix was contrarian was, it was obviously a great product that people loved, but they were burning a lot of money, and so it was just not clear, was this like a classic tech company, or would they ever make money? Like, that's why it might have been contrarian without piece.

A Yeah, I think the, the issue is like, there's very few tech companies that are massively capital intensive, right? Almost every tech company loses money for a period of time. But the typical software company you're looking at, like, you know, there's operating losses, you leverage it, and then people are very used to that business model. Up until the LLMs are very few tech companies where it's like a huge fixed investment. And then the incremental, ah, margins on the sales are extremely high. And so, what that meant was that Netflix was investing heavily in content, which is a fixed cost, and then they were selling that to consumers. The next year, they were investing more in content and selling it to more consumers. But you're constantly investing more and more in content.

AI assessment note: “there's very few tech companies that are massively capital intensive, right?”

Q Do you think you guys were more focused on retention than others?

A I don't know if we're more focused. I mean, I could tell you though, you know, one of the things that was happening, especially when you see a competitive fight, is you see everybody race towards it, right? Everybody is going to try to make offers, you know, to customers, try to give discounts, try to give coupons, you know, free this, free that. One of the things that we had looking backwards is we actually did not have a large budget. In fact, you know, between 2016, 1718, we barely were able to raise a dollar, you know, relative to our peers. As a result of that, Um, that made it a constraint. One of the constraints is, okay, you can grow, um, but you cannot spend in order to do it. So in order to do that, you effectively have to actually come up with ideas in the product to actually stand out and make a difference and have organic growth, um, you know, carry you. And then once we were able to, you know, demonstrate to ourselves first, um, that we had a product with higher retention, you know, than other people and higher frequency, And then we were able to raise capital. Then we actually made the decision to pedal to the metal and actually go and acquire customers because we had an unfair advantage.

AI assessment note: “I don't know if we're more focused... we had a product with higher retention”

Q is a technology company that happens to be in payments. It's not a payments company. And You know, you, you, you've always challenged, I remember, you know, um, you know, your customers to think about, don't just think about accepting payments and, you know, there's more that you can do. So I am curious, what more can we do? Should we be doing with things like crypto or stable coins?

A Two topics that are on the minds of all our customers, uh, today are, yes, stable coins and then, uh, AI specifically with respect to changing buying, you know, agent of commerce. On stable coins, I think maybe DoorDash is not where I would start. Two places that we're seeing tons of interesting stuff is one, any kind of cross-border use case, where if you want to send money to a hundred countries, it's already the case that stablecoin's a much better way to do that than anything else. And that's where we're seeing a ton of adoption, where just that long tail coverage, and they, they really work for that. Um, the second, and this is where Tempo, our new, uh, initiative, That, um, you guys were nice enough to work with us on is, uh, coming in is very scalable, uh, crypto payments, and obviously agents is what we have in mind as we're, uh, developing Tempo, where, you know, if you want to pay for your API consumption, or if your agent wants to be able to pay for things around the internet, you actually need a really good scalable blockchain for that. Ok, so that's all the, the crypto side of things. Other thing that we're thinking about a lot is just how does commerce change now that people's expectations are increasingly, they ask their AI for stuff, and so it feels like you should be able to ask your AI, whether that's, you know, ChatGPT or Gemini or whether that's Siri or, lik…

AI assessment note: “Two places that we're seeing tons of interesting stuff is one, any kind of cross-border”

Q Hmm. So there's a next best use kind of math. Which restaurants have actually invented something that's hard to copy?

A Well, I actually think any restaurant that's been around for, let's say, two plus decades probably has very interesting IP. And obviously, you know, some of this information is not public. So, you know, um, but, but you can imagine, you know, when you go into At McDonald's around the world. That french fry, you know, perhaps they don't sell the same exact items in every single store in every country, every city, but the french fry almost always tastes the same. That is an extraordinarily difficult feat to accomplish. There's a lot that goes behind the scenes, just like there's a lot that goes behind the scenes at DoorDash in terms of getting you, you know, one order on time to make that sentence true. And The same can be said about a lot of other businesses that have been around for very long periods of time. That's on the, you know, big brand, you know, QSR side where a lot of the innovation, if you will, is in the process innovation, and then also how do you run, um, large groups of people And have very high and consistent standards of service. Extremely difficult, extremely difficult, and that's the IP, I would argue, for a lot of these large brands. Now on the other side, there are small restaurants, you know, some of whom have been around for almost a century, actually. There aren't that many of them, but, but when you look at what makes them special, it tends to be the sa…

AI assessment note: “when you go into At McDonald's around the world. That french fry”

Q just doesn't work, and so, like, sometimes I'll be using my phone, and I'll use the IOS keyboard transcription to type in the field and then, like, say a bunch of stuff and then send it off. But this suggests to me that consumers really want voice mode that works, and yet it's just not working yet for the major LLM apps or for anyone. Why is it working yet?

A It is pretty hard to do, because you want, you want two things. You want, uh, you want to be able to say things that you want, but you want sometimes for it to execute it, sometimes to wait for you to, like, finish and add something in the sentence. Sometimes you want it to be interactive, so it asks you questions back to clarify and get some of the additional detail, and all of that is actually pretty hard. Like, that's where kind of the, the magical, like, ideal version of a voice agent for us comes through, where you need the speech-to-text element, you need the transcription side, unique You need then the kind of the turn taking mechanism. So like, when do you finish sentence? When, when is it likely based on silence, based ways likely on the context? And then sometimes you want it to speak back and clarify, or at least give you the text back to clarify, and then maybe execute set of instructions. So that problem is still very hard research. So I agree with the claim that like this orchestration side has not like passed a true conversational agent Turing test. Where it like behaves as you would expect from another person where you can say.

AI assessment note: “It is pretty hard to do, because you want, you want two things.”

Q So when you say can bring convert conversational agents is the biggest priority. Is this for customer service type use cases? Like what are the most popular use cases for conversational agents?

A Yeah, like we want to be a partner for like full interactions between business and businesses and their customers or their audience. Um, I'm saying that the audience because that will apply in support. Support is the easiest one because that's where it's most ready, but like, and that's maybe the big difference to how we see ourselves to some of the other companies in the space is this can also apply to sales. You can, you can have the proactive side of reaching back. You can have AISDR versions of that. Yes. Um, and then you can have all the way to the marketing use cases, where we are your partner for, for working on, on, on, on, even outside of like the, the conversational agent space of how you create a great marketing campaign.

AI assessment note: “Support is the easiest one because that's where it's most ready... also apply to sales”

Q It feels like a big part of the magic of 11 was your voices were much more human sounding. How did you accomplish that?

A Kind of give you a, um, a quick, quick synopsis of what, how we think about the models on the text to speech side today. In any model you need, you need the architecture, you need compute, you need data. So architecture innovations were one thing. The data part was the second big thing. With audio, you will have, um, you will have a lot of audio data available, but frequently you will not have it annotated in the right way. You won't have which speaker is speaking when, um, some of the what Uh, is annotated, but the how isn't. So, like, as we are speaking now, what's the emotions that we use? What are the actions that we use? So we would invest a lot internally on effectively creating our own data labelers, our own team, to be able to create those data sets that will be better. And that was a combination of, of course, like, semi-automatic techniques, and then, and then, and the manual techniques. And actually, a lot of the models that we did afterwards actually spun out from a lot of that research, too. So speech-to-text model, Initially it was a model we did for ourselves because the models on the market just weren't good to annotate that data. And then another brilliant researcher on our team was kind of being able to construct it so we could span it out as a model that we brought to the customers.

AI assessment note: “we would invest a lot internally on effectively creating our own data labelers”

Q stuff for Stripe or any other company to do. And then the AI productivity story in Other roles is just a bit less clear because as we've discussed, AI is kind of uniquely, um, uh, well suited to coding. And so what do you make of just how does the AI productivity show up? I feel like every company in Silicon Valley is trying to figure this out right now.

A Well, first, I think I'll go back to my why I believe in applied AI. I think the atomic unit of productivity in AI is a process, not a person. I don't think AI Uh, I don't know if you have an assistant, but if you do, he or she might help you prepare for a podcast, might help you prepare for a meeting, he or she might also get you a cup of coffee. AI will be really good at the first two, but quite poor at the last one. So no matter of AGI, short of robotics, will get you a cup of coffee. So I think it's wrong to think about AI as like, sort of replacing people, uh, in addition to being inhumane. It's just sort of nonsensical because AI sort of operates in the world of digital technologies. And I think if you go to, like, an example of even a mundane process in your business, like onboarding a new supplier, think about all the departments and people involved in that. There's a legal department to do a contract. There's some, uh, finance department procurement to negotiate the relationship. You probably have IT that's involved to sort of onboard them into your core systems. And then there's usually a business that's sort of sponsoring it. Fairly mundane happens all the time. If, Let's just say you tracked what is the median amount of time it takes to onboard a new supplier, and it was, um, 17 days, just for argument's sake. I bet you could say, as a CEO of a company, I want to us…

AI assessment note: “I think the atomic unit of productivity in AI is a process, not a person.”

Q Yeah. Yeah. Now we get to it, which is you described building stuff that you know you're going to throw away because the model capabilities will get there. And you're like, occasionally they are developing capabilities that you developed yourself. Isn't Sierra itself kind of short AGI? Sorry, I said I couldn't resist.

A No, it's, it's the right question. Uh, you know, the short answer is I don't know. I mean, the fog of war in the software industry is pretty thick right now. I really believe in the applied AI market though. Uh, I, I think, I think most companies don't want to buy models or buy software. They want to buy solutions to their problem. And if you just go back to, um, the cloud industry, why, why doesn't Amazon and Microsoft do everything for everyone? There's not really like a sort of by somewhat similar logic, like why should any software as a service company exist when you have Bigger scale, all this technology. In theory, they could just develop all the software, and actually, many of them have tried. There's actually competitors to Salesforce and almost all of the above. I think there's so much nuance in how these companies align themselves with different departments at these companies, solve their very unique problems in very specific ways, that is a mix of product, not technology, but product, go to market. Um, it's, It's an ecosystem around it, and I think a lot of that still exists because I'm not sure, like, coding the software was necessarily the hard part, and then similarly, I, I actually think, especially in enterprise software, how you engage with your clients really matters, and, uh, you know, I think, uh, it turns out that, you know, um, GPT-V and, and, you know, Cl…

AI assessment note: “the short answer is I don't know. I mean, the fog of war”

Q You have four big decisions a year that are kind of grounded in science. How quantitative versus qualitative do these end up being? Are you Rick Rubin where it's all taste based and you just like the feel of this direction? Or are you Billy Bean where it's a Moneyball type, you know, with the ROI pencils and.

A I think the system does a lot of the Billy Bean. I think that's a change at Lilly that's made us more successful. I think we've actually put together a decision process that's quite a bit more rigorous than it used to be. And that leads to fewer bad decisions. That's good. So that's sort of like the, the bumpers on the bowling alley that you put up. So, but then within that, whether it's a strike or a single pin, that's a little bit of the judgment and taste. And there though, you know, wisdom of crowds, I think we have a great leadership team and we all come with equal voice and sort of debate. We actually have a rule to like never decide in one meeting. So you're asking about the day, but we, We, like, come back to it, think about what others said, and kind of push it again.

AI assessment note: “I think the system does a lot of the Billy Bean.”

Q is correct for the rest of it. But where I was going with this is, if you were to listen to this podcast, I think you'd maybe come away thinking, wow, pharma is hard. Like, good god, you know, there's so many things, and you know, things roll off patent, and we have the Chinese competitors, and things like that. So, so what is it that investors have confidence in?

A Well, I think the track record of success. We've been on a growth curve for 12 years or so. It's certainly gone a little more hyperbolic lately, but, um, I think that builds confidence. I would hope some management piece, but also, you know, the, the ability to, um, predict where to move. And I think if you say, okay, what's your recipe? It's an R&D business. The, everything else is around the edges. So you have to create something better for people that improves their health. If you can do that, you're going to win policy, this and that, the commercial strategies, that's the 20, the 80 is this. And I think we do three things better than others. One we talked about already, which is cycle time. It's a basic concept, but if you can make software faster than someone else, you're going to win. And the same in the drug business. The second is prediction of where to tack the investment and allocating a meaningful part to ideas that may not be obvious today, but actually are big problems without markets. And we're drawn to those. That is the third box.

AI assessment note: “Well, I think the track record of success. We've been on a growth curve”

Q If you could wave a magic wand to speed up U.S. solar deployment, not manufacturing, just deployment, what would you, what would you do?

A The, the single biggest intervention you could allow is, um, by right development of solar panels on private land, um, and ideally even like BLM land would be great. Right, BLM land has been set up for ranching, Um, and you can run your cows on it with very little friction. Um, and, and I think sections of that should be set aside for like, it's just electron wrenching. You can put your solar panels there and you have, you have something like California's fire, um, uh, insurance scheme, right? Uh, so that if, if, you know, one batch of solar panels goes bad or after 50 years we have fusion and we want to take them out, there's money there that we can take them out and put them, put them away. It goes back to being desert. Um, but right now to permit a solar array, even on private land for the sort of applications we want to use, it's often Just as tiresome and expensive and difficult as permitting a new chemical plant or, you know, something that's much dirtier. Much dirtier and much worse for the, actually, I think solar overalls a net benefit for the environment. We should be encouraging it. It should be like, you get credits for deploying more solar on government-owned land. Yes. 90% of Nevada is, is BLM, is like federal land. Yeah. And that's the best place. The reason it's federal land is because they couldn't give it away in the day, right? Um, because it's so inhospitabl…

AI assessment note: “The, the single biggest intervention you could allow is, um, by right development”

Q And you're judging this poor analyst, you know, and their job on the mock portfolio, which you're saying is like updated weekly. And so aren't you not actually looking at companies on a three year time horizon because they have to like perfectly perform at every step along the way and they can't be misunderstood for any short period?

A Yeah, the way I think about it is, any given time, we are planting seeds, and then we're harvesting. Like, there's some companies that we invested in a year ago, and our thesis is starting to play out, and you're making money, hopefully, on that idea, and, um, you know, maybe you're selling that idea at that point, and then you're putting in new ideas in the portfolio that might take another year to play out. And so, it's not like, if we started out with a bunch of ideas that all had three-year targets all at once, like, yes, but like, The portfolio is a living thing that, like, you have a range of positions, some you've had for a year or two, and you think the, you know, the business is going to turn the stocks into work, you know, and then there's other companies that you're buying now because they're really depressed and really out of favor, and you know they're not going to go up in the next, you don't think they're going to go up in the next three or six months, but over any medium term retirement they will. So it's like always.

AI assessment note: “The portfolio is a living thing that, like, you have a range of positions”

Q Okay, so the thing I want to get into you guys, um, About, ah, or spend a lot of time on is the history of the valley. One interesting place to start might be, can you tell when you're in a bubble?

A So my experience is no, ah, and, and, and, and the nuance that I would put on that, ah, I'll describe two. The nuance that number one is there's, there's an old line with respect to economists, economists that also applies, I think, to investors and entrepreneurs, which is economists have predicted nine of the last two bubbles, ah, or, or nine of the last two crashes. Um, and so it, it is extremely common. It, it's a difficult question because it's extremely common for people to call a bubble. When they're correct, they will then go around for years claiming that they're the one who called it. What, what you, what you, what you find with those people generally is they were calling it continuously for the, for the, for the, for the, for the, for the years earlier.

AI assessment note: “So my experience is no, ah, and, and, and, and the nuance”

Q So, so tell me more about your users. I, you know, one of the hardest things that I thought about building products at Retool was that it's such a widely applicable tool when you're building developer tools. And how do you decide who your target user is and who to build for?

A Uh, so from, from the beginning, a part of the vision was like, okay, there's going to be a rapid explosion of new startups with, like, one person creating a lot with just AI, and we want to build for those founders. And as, like, we started building out the product and the AI wasn't, like, as good as it is today, of course, then we looked around and thought, like, similarly to Retool, actually, is this a tool for, Um, building internal applications, and, um, we thought that that might be the case because we spoke to a lot of users that liked it for this application, but then once we just launched and, um, really saw what people were using it for, uh, it became clear that We don't need to be very, very specific about what you should use it for. Uh, today it's used, um, a lot for, as you said, like creating this first version of your startup by, both by teenagers or kids that are like, I want to make money and create something. And serial entrepreneurs that raised like fifty million dollars, but now they were completely alone in Lovable. Uh, and now increasingly inside of companies, like you said, even you had a hackathon at Stripe, right? Where you could, uh, where it's a tool for designing or ideating exactly what type of things should, should you as a company add to your offering or, uh, you, uh, use as a tool to, uh, make yourself more productive. So, so it's, it's used all,…

AI assessment note: “we don't need to be very, very specific about what you should use it for”

Q Yeah. I mean, it's incredible. It's like, like you said, you're just democratizing, Democratizing access to, to development, which, you know, most of us just have, have never had. Um, how much education do you have to do with users on this just completely new paradigm, paradigm of building?

A Yeah, this is a great question. I, I think. The wow moment for like, oh, this would have taken months to build this beautiful websites in the past. That, that happens instantly almost. And that hooks people and they're like, they realize that in itself educates people that I can build something now. This wasn't possible before. Now I can build something. I have to become a bit better at how to prompt and so on. Um, but as you start building something more complex when it requires, um, like setting up payments to Stripe is quite simple, but like if you have many features, Then, uh, you do need a bit of education to understand what are the different components here. So there's some things running in my browser, and then there's some things that need to, um, for logging, logging in, for example, you connect to our default integration to Supabase. Uh, and that takes a bit of figuring out or watching a YouTube video, for example. And, uh, yeah, we're increasingly letting the AI do more and more of the education as it talks to the, to the user.

AI assessment note: “you do need a bit of education to understand what are the different components here”

Q size that you guys are at, it's in itself sort of pretty unprecedented. Um, I, I think at Stripe we, ah, we're, we, we found we're shipping 30% more code. Um, and so, and who knows if it's better code, but it's, we are shipping more. We're seeing some, some, some traction there. Um, what models are you doing? Have you, are you using? Have you built your own models?

A We haven't done that. We looked, we have looked at fine tuning our models and what we're seeing is that the foundation model labs, they, they come with out with new models and we, the hardest part to get right is this inter interaction between like the AI, the complex agentic systems of like how the models are used and the user experience. And we want to spend as much part on the last two as possible. Like how, how are the, how are the AI reasoning and so on used, and they use, how does this translate to a user experience for, um, what the human actually experiences. And if we're spending time on the AI, it's time we could have spent on making the other two parts better. Um, there's, there are other opportunities in making the response times faster by creating our own model. So we might look at it in the future, but it's not the lowest hanging fruit.

AI assessment note: “We haven't done that. We looked, we have looked at fine tuning our models”

Q talent density problem? You just need to get everyone together. Like everyone's, you know, everyone's kind of spread out. There isn't sort of a clear like Silicon Valley hub. Is it like a funding problem? Is it a, like, you know, if you were to sort of like solve Europe, solve Europe's, uh, yeah, having to build a startup in hard mode problem, what would you, what would you say?

A Yeah. Yeah. Yeah. So it's, um, it's like the, the talent is, I guess the same and it's more available in Europe. The, The seniority on different things is lower, ah, like the specific, um, skill sets and functions, um, but the, um, ambition level is the biggest problem, I think. Like, there's fewer people that are, like, super ambitious, like, okay, this is a unique opportunity in the history of mankind, and we're here to build the best way to build software applications with AI. Instilling that mindset is easier in the US. Uh, and if there was, if it was more of this high, super high ambition in Europe, then we would see much more successful companies from here.

AI assessment note: “ambition level is the biggest problem, I think.”

Q long time horizon, but if you're looking at a short time horizon of an insurer or an employer, they don't necessarily. And so, that's great, this challenge for reimbursement where, you know, not as many people reimburse GLPs for weight loss as you think would be rational. Yep. That just will always be the case with prevention, and so how do you actually develop drugs that are commercializable and reimbursable?

A Yeah. Well, in the obesity case, I'll, I'll take a little bit issue with your first assertion and then add two other problems. The, the data actually is becoming more clear that within actually a two-year timeframe, and I hope it's Stripe you reimburse these medicines for your patients or for your, your employees. Within two years, you can, you can break even based on total medical costs. So there's this group called ICER, which is funded by someone who hates our industry and the insurance companies, and they analyze all new drugs. And usually, Seeking to prove that they're not worth it. That's sort of their mission in life. They just analyzed our medicines, trisepatide and semaglutide, and they said, actually, they're both cost effective at current pricing. In fact, Zepbound or Trisepatide was, the threshold they have is to save a 100,000 dollars per person per year in downstream health costs. And it was twice as effective as that at the current pricing. And the current pricing isn't going to stay. Let's be honest. There'll be more competition. The government wants to lower our prices. So, you know, we're, I think we're in a good place there. Now, the two other problems are there's sort of this incumbency problem in healthcare, like many things, but particularly in healthcare, where the last thing in, Is scrutinized the most and the base stack of services and products we use i…

AI assessment note: “The, the data actually is becoming more clear that within actually a two-year timeframe”

Q has my, uh, sort of, you know, lots of my history, and ask, is that the kind of product that I would appreciate? It would, like, it would just return yes, right? Because it's like, it solves the problem in its entirety. It's like a travel adapter, which, like, for any country, to any plug, to any other, um, all sorts of stuff, Good, good product, right? I want it.

A Okay, so we're, we're, we're getting down the road here. I mean, it feels like step one is, that's going to happen with agentic commerce is, I feel like existing aggregators, if just aggregation is their raison d'etre, then they're going to have a lot of, ah, questions that they need to answer, because ChatGPT and the AI apps end up as a new aggregation point. And so, It seems to very much benefit the tail. It benefits all the Shopify merchants, just like, you know, we had a new set of content creators emerge on YouTube that effectively compete with kind of TV. They're both just entertainment, and the tail becomes much more powerful when you have over-the-top distribution through YouTube. Similarly, it feels like once you have over-the-top distribution through AI apps, it becomes, you have a much more powerful position as a tail merchant because, you know, Momax is a tiny brand, Uh, but they can just get recommended as the best product. So that's step one that I think happens.

AI assessment note: “you have a much more powerful position as a tail merchant”

Q Yeah. Did you try to monetize projects before you started building them?

A No, because, so what happened, I was, I was making music, um, and I was DJing, I was organizing, like, nightclub nights in Holland and UK and stuff, um, drum and bass music. The drum, you're Irish, you know, drum and bass. Like, drum and bass, most people don't really know, but, British people know, Irish people know. Um, so when YouTube started, um, monetizing videos, um, I was uploading my music already. So I was one of the first YouTube monetizing people, and I was like number one or number two in Holland for a while in the top channels. But the thing is, I started making money like 1000 dollars and 2000 dollars and 8000 dollars per month, a lot of money, and I was in university.

AI assessment note: “No, because, so what happened, I was, I was making music”

Q I think we could be, but, but I think we definitely could be. But, but again, like back to the question of, uh, of asking What problem does this solve, right? That the human is pretty good with their eyes, you know, at some of this stuff. So the bar is pretty high, you know, in order to do some of the things you're talking about.

A Yeah. As I just kind of think through AI applications for DoorDash, one thing I'm struck by is that the LLMs are pretty good recommenders just by tossing things into context. You don't need to train a custom model, but I don't know if you've ever tried for like book recommendations or TV recommendations. You just like tell it a bunch of the stuff you like already. Restaurant recommendations. It's like, here are 15 restaurants we like to go to. Give us other recommendations and it'll do a really nice job. And yet within products, you know, if I open DoorDash, it's kind of the same categories and things like that. And it's less personalized to me than if I took my DoorDash history and put it into, um, put it into an LLM. Isn't that, like, shouldn't we be somehow using the fact that LLMs are pretty good recommenders? Within products. It's not just DoorDash. It's every product I use. It feels like those recommender capabilities are underutilized.

AI assessment note: “one thing I'm struck by is that the LLMs are pretty good recommenders”

Q It was like awful dubbing, you know, happening in Poland previously. So that's like one example of the, you know, the second order effects. What are the other second order effects you're seeing of ubiquitous good text to speech, speech to text? It seems like across a broad array of languages, because whatever about in English, just this didn't exist in Polish or Irish or, you know, pick your language.

A One, like, breaking down the language barrier, we, you know, the, kind of, the inspiration came from, from the movie side, but it also applies in any, in any communication setup, like, could in the future, could I travel to another country and speak, speak Polish or speak English, and that, that language isn't being understood in the local native language. Like, from Hitchhiker's Guide to Galaxy, this version of the Babelfast, exactly, that you can, like, actually understand the world. And voice, of course, will be an interaction layer, but similarly, all of us will have our own, Kind of extension and voice agents that can help on, on our behalf. And there is like very clear and, and, and, and great examples of that of people that lost their voice and can get it for the first time, for the first time back. We see that everywhere, whether that's people that lost it due to ALS or throat cancer that can get it back. Uh, just recently there was an example of a patient that had Neuralink and worked with them to bring the voice that that person could speak with their own voice back to the, Back with the, with the family around, we worked with, with, um, with, with the lady that lost her voice before, before she got married, and, and then finally technology became possible. We, we were able to recreate that voice, and for the first time she could replicate the, the marriage ceremony a…

AI assessment note: “One, like, breaking down the language barrier... people that lost their voice and can get it”

Q Was there something that demonstrated to the outside world that, oh, they got this? Was it Gemini three that changed people's minds or like, I don't follow all the timelines.

A I think, I think the real model, uh, probably where People, uh, saw it was maybe Gemini at 2.5. Um, and, you know, I'm getting to the frontier on Particularly around multimodality. We made a bunch of, I mean, credit to the Google DeepMind teams, right? They, we, I think we paid a bit more of a fixed cost upfront, but we designed the Gemini models to be very multimodal from day one. And so there were, there were areas, uh, I think, I think, um, we started, the strength started showing. Nano Banana was an example of it, right? So you were able to see it all together. But look, it's an amazing, amazingly dynamic frontier. I think there are two to three labs who are pushing each other pretty vigorously. You know, at any given month, we feel like, oh great, we've done this well. Oh shit, there's like a couple things we're behind, right? But I think the picture will again be dynamic in a few months. So I think the frontier is intense as you would expect it to be. So that's how I think about it.

AI assessment note: “the real model, uh, probably where People, uh, saw it was maybe Gemini at 2.5”

Q executive has severe AI psychosis right now, and is spending, you know, a huge amount of time writing code and talk to AI and things like that. I thought that was a funny take and not without Any truth to it? And I'm curious, what were your feeling the AGI moments along the way of the recent, or, you know, to what extent do you have AI psychosis these days?

A My first feeling the AGI moment was, uh, uh, when Jeff Dean demoed the earliest version of Google Brain. This is when the neural networks recognized a cat, right? So that was in I went with Larry to the DARPA challenge. It might have been in the year, I think I need to be exact about when we went there, seeing the cars drive there. Demest demoing the earliest versions of the models, having what we would call as imagination. So there have been many moments like that so it was obvious the technology is progressing. In terms of living now and kind of having a visceral feel for it, I think the closest I would say is if you're coding and you give it a complex task and you never open the IDE and you're in some agent manager world and you see it kind of do it, You know, and how powerful it is. So, you know, if you can, you know, call it field AGI. So there are moments like that.

AI assessment note: “My first feeling the AGI moment was, uh, uh, when Jeff Dean demoed”

Q Yeah, it's a good way of putting it, that every company is, uh, different, uh, in some new way, and so yeah, this could be yours. Um, what was the basis for the, um, decision where you won the election last year? Like, was there any interesting policy angle?

A Right, so the whole point is that the government cannot stop any type of contract unless it makes a finding that's against public interest, and it has to fall within certain categories of war, terrorism, assassination. And the CFTC was taking the stance that they were trying to fit elections into any of these things. They're like, oh, elections might be illegal under state law, and, you know, because betting on elections, there's this one state that in bucket shop law, they try to find something. To stop it. And we knew we were very, very clear on the law, like elections have economic impact. If the elections have economic impact, they need to be allowed to trade on a futures exchange or derivatives exchange. Um, and it was. Basically, I think what the, what the lawsuit did is it told the CFTC that they couldn't just do whatever they wanted. And that kind of like.

AI assessment note: “the government cannot stop any type of contract unless it makes a finding”