Noah Shinn - Building Instinct: The Personal Agent - [Invest Like the Best, EP.493]

27:47I don't know the numbers on email and what proportion of users connecting email, but you can imagine full access to an email inbox and into a calendar. So there's a lot of service area here. There are systems that we've put in place that are detached from instinct, the agent architecture itself, that are put in place to be proactive about these things and to sort of like decouple the risk, if I were to say. So for example, any piece of content, any piece of text, anything, any form of media that comes in that might be consumed by instinct goes through what we call these firewalls which can intercept, that can reject, that can block malicious pieces of content coming in and hitting instinct and trying to convince instinct to do something.

28:36There is also for every action that instinct might take or for every thought that it might have, that is being actively monitored by a system that is decoupled from instinct itself, which is able to pause it, intercept it, to approve or disapprove of what might happen next before it takes the action. I mean, those are just two pieces put in place, but there's just so much more. There's so much more under the hood that is put in place to enable the agent to be as capable as possible, but also safe and trustworthy. I have this question around, I don't have a better word than alignment, and I know alignment is a very loaded word in AI.

29:12I don't mean humanity scale alignment. I mean an agent aligned with me personally. If I think about other agents I've hired, like employees, I pay the money and therefore I trust them to have my interest at heart. It won't be perfect, but they don't have some other alternative incentive stream that guides their behavior. They're guided by their employment. How do you think about the business model vis-a-vis alignment? Like you could, well, you'll walk us through how you've thought about it, but many approaches will be taken. Some will be paid. Some will be free. The free ones will monetize in different ways.

29:46Can you walk us through this decision tree of what the business model is or will be and how you arrived at that as the ideal conclusion for the agent? We don't want instinct to influence the user's behavior in a way that is not aligned with what the user wants. That sounds very good on the surface level, but I want to call out how important that is because imagine a world in which instinct is generally smarter than the user. I'm talking more socially intelligent, more socially aware, like textbooks smarter as well, right? I think it would be a very dangerous world if instinct were influencing the user's behavior to purchase something that they don't want to purchase or to subscribe to something that they don't want, right?

30:31And using its intelligence to be able to convince that, Yeah, so I think that, you know, you look at most of the major consumer businesses today, who's, you know, are able to influence users' behavior against what they might want to do. You know, I'm talking the major platforms, whether that's Google or TikTok or Instagram or Snapchat or so many others, where you have this platform and users, you know, it's free for the users. But then there's so many instances in which, you know, paid ads are pushed to the user and they try to convince the user to then purchase. And by the very nature that they do purchase, it now becomes this game of these brands paying to convince users to purchase items that they may or may not want.

31:15This is the idea that if you're not paying, you're the product. Exactly, exactly. As a programmer, as a technologist, I just don't want to build that reality. I think that's a very dangerous reality. And so we instinct should act on behalf of what the user wants. We take this very seriously when we're building product, when we're through the various research projects that we have, which are Instinct is not, unlike any other AI product, Instinct is not a task accomplisher. So what I mean by that is, you know, with most other AI products, you write a prompt and then it does the task and then it tells you what happened, right?

31:56That seems good in theory, but what happens is that if the user is now asking for something, you know, that might not be well-intentioned. Well, if you have a task accomplisher, it's just going to do that and listen to the user. Instinct follows higher-level objectives. So instinct will follow, you know, learn to build trust with the user, learn to make the user genuinely feel safer with you, learn to watch over the user and have their back when things might be dropped or other things like that. And when the user asks it to do something, if it is well-intentioned and well-meaning. One way to communicate safety and trust is just to do the tasks.

32:37And so we end up just doing a superset of what most other AI products are able to do. But I think focusing on higher level objectives is very important here because it enables Instinct to be more robust to these edge cases. So we don't want to influence users' behavior in a way that is not aligned with what the user truly wants. Even if you look at the business model side of this too, there's over a billion dollars flowing through the platform now every year. and we're just getting started. And honestly, we're going at 10 % day over day. So you imagine the transaction volume is also compounding at 10 % day over day, right?

33:10So that's not just a billion dollars flowing through the platform. Now that's 1.1 tomorrow. And then that's like 1.2 something the next day and 1.3 something the next day, right? It's still very early. But when I see transaction volume that is so high that flows through the platform, what I see is very basic case. It's similar to like an Apple Pay or it's similar to like an Amex or any other platform that provides distribution to underlying services and a great user experience for the user. Apple Pay is a great experience, right? You can go anywhere and you can scan your card and the user doesn't have to pay for it, right?

33:44The user is getting a free, great experience and the merchants on the other end who are benefiting from the business are paying to be a part of that platform. So I see a blanket transaction take rate being enforced across the platform, which is just us exchanging distribution for being able to serve products on behalf of merchants. Can you say a little bit more about how that runs into the existing world? So I can imagine the layers being you could be a card issuer. You could be something like Stripe. You could be Visa MasterCard. There are sort of rails that have been built in the payments world that create convenience and reduce friction and take a vig as a percent of the transaction.

34:25and those are some amazing businesses, to be sure. How do you think about which of those are partners, which of those are potential things you would displace? How does that vision of a small take rate on the transaction volume on Instinct, because it's a free product, how does that slot into the existing world, do you think? Just to put that into context, those are very great businesses and there's so many people along the side. You make one digital transaction and there are like 40 people along the line that make money on that. And just to take a step back, those 40 people making money at each line of the stack are really sharing what two and a half percent two percent it depends on where the transaction is coming from so it's a very actually small piece of the pie and um we're not primarily interested in whatever like doing what amex does best and have the network to do or what uh even even like what what stripe does uh on the internet or or you know some of the underlying payment infrastructure do because it's it's such a small piece of the pie and i think that they're providing real value where i see the majority of the value just like thinking on the business side is, you know, if you look at most of the other, you know, major platforms in the world and what take rates that they're able to achieve by, you know, again, it's free for the user.

35:34It's a great free experience for the user. And the merchant, you know, is now recognizing the distribution source. You have Shopify that I believe is like between two, two and a half and 3 % that are providing their services and taking, you know, some take rate from those businesses. You know, I'm not sure where Stripe is. I think it's maybe on the lower end of that. And then you have Amazon that has a great platform and taking upwards of 10%, right? And then of course, the premier, you have Apple where you make any in-app purchase, which is 30%, right? So I'm not saying that we're going to be 30%.

36:05I think it's very unrealistic for a lot of businesses. But what I'm saying is that we still don't know where along this curve of distribution power versus take rate that we're going to be. I just want to call this out again, because I don't want it to be misunderstood of it's a free experience for these or it's a free grade experience. For the service that we're providing to all of the underlying providers, I'm not focused on that to finding like 30 bips on the 2.5 with some partnership with some payment provider. I'm looking at the, can we provide so much value that we're on the upper end scale of this?

36:39The reason why I think that this is possible is very practically 50 % of the transaction volume flowing through the platform is travel alone. And we all know the travel industry and some of the rates that we've kind of... The OTA rates are high, yeah. It's very high for flights. Maybe it's on the lower scale, but how many single digit percentages can you take off of a flight? Or for hotels, some of these boutique hotels, they're offering to pay up to 30 % for every transaction that you're able to deliver for them. And I'm not saying, again, we're going to be at 30%, but you can see the range. These are existing business models that we can boot shop off of in the early days, but then what would it be like to take that and extend it to effectively every major industry?

37:25This is only the case. We only have the ability to do this. If it is true that most digital behavior then moves to these new types of interfaces, which is, well, again, like no interface, right? But it is only the case if there's significant distribution power. So I think it's actually a quite intellectual or interesting intellectual question here about how this evolves over time. It seems to me like there's going to be a serious corporate agent war that you've already seen this with Amazon and Muse. You've got all these established players with tremendous vested interests in relationship with customers that this could, both instinct and other agents, could really disrupt in a major way.

38:11What are you thinking about that? Like, how do you interface with the other great services out there? I'll pick a random one. You know, I use Uber Eats a lot. Like, I order from Uber Eats all the time. It's kind of a pain in the ass to click through the thing. I can imagine a lot better experience on a snappy WhatsApp connection with instinct or something saying, hey, I want my usual from this restaurant. And that's it. Like, there's nothing else. And that it's got its computer and it goes on to Uber Eats and it orders or whatever. At what point does Uber Eats not like that anymore? At what point do these great businesses that have been built up start to be adversarial against agents, do you think?

38:44What do you think are the most likely sources of conflict and reconciliation? It's going to be really interesting to watch. So I'm serious when I say, I mean, you're alluding to this too, that I think most digital services or industries are going to be not disrupted, but just changed and transformed. And so I was running through an exercise the other day of pulling effective like every digital service or business or application from various different verticals and industries and sort of plotting it along the line of saying what proportion is like the user's attention, you know, experience on the product as a proportion of revenue?

39:20and then what proportion of it is delivering the underlying service? Meaning if the user didn't use the app, what proportion of that transaction would be, or what proportion of their revenue is due to the underlying good being provided or service being provided? And so I think that your question is of that upper end of the scale where let's say the majority or some significant amount of revenue is due to attention on the application or advertisement or things like that. So yes, that's Uber. I actually don't know what Hooper's number is. That's a lot of the restaurant or the food delivery services.

39:56That is the travel agencies. That's that's even Amazon. Amazon themselves with with the upselling that they do on their platform. OK, there's one simple way to look at this, which is that, OK, well, it's 70 percent ads, 30 percent good. And so therefore, you're going to slash the 70 percent and they're going to be a 30 percent business moving forward. There's another way to look at it, which is I would, you know, take any business. Let's take Uber Eats or DoorDash as an example. They don't know the numbers. I don't know the numbers, but they don't know the numbers of as they reduce the number of clicks needed to check out to order food to your house or to order an Uber.

40:35As they reduce that time to check out, the transaction volume increases because it's reduced friction to get the same underlying good. and you know we're taking something like instinct and and making that friction to do anything almost zero it is literally zero in the case that there is proactive behavior for uber let's say for ride-sharing right if instinct has access to your calendar and is it you know owns your calendar and is booking you know all of these events and knows where you need to be in person here and in person there you know what if instinct just always had a car lined up for you at you know every time where you need to be so that it's out of, you know, out of your mind to know, oh, like, you know, in five minutes, I need to order this Uber because I need to be in this place in 45 minutes.

41:19And there might be traffic. So let me check the app to see how much traffic there is to see when I need to order, you know, order the ride. Instead, what if it was just, you know, proactive behavior, a car will always be lined up, and you will never be late because it's going to calculate traffic, it's going to find these things. What would that do to Uber's business or what any red share business, if now the default, you know, for existing, you know, riders that love, love Uber or love Lyft or any other ride share business, what would it be like to have proactivity now make the friction to, to, you know, experience that, that, or have access to that good or service to be nearly zero?

41:53I think that that will result in that 30 % looking a lot, lot bigger. You know, you talk food delivery service, maybe you're coming home late off of an airplane and you need to be at your house and you haven't eaten yet and it just texts you and it says, you know, are you hungry? Do you want the same thing that ordered yesterday or five days ago? I can send it to your house if you'd really like it. I know you're, you know, because it ordered the Uber too. I know you're in the Uber and you'll get here at this time. And then the user just says, yeah, go, that's great. Or thank you. Yeah, please order that.

42:22So with very little friction now, it's a, you know, what will that do to the total transaction volume or the amount of times that a the user interacts with the business, I actually think it'll go up. So it's this interesting game. I think it's this interesting transition period between users spending a lot of time on apps, like painfully spending a lot of time on apps, and that being a monetizable surface because that's where the user's attention is, to users actually interacting with the business even more, which is counterintuitive, actually interacting with the business more because the friction to do so is actually just much less.

42:56So I think it'll be this interesting game. And all these various industries are going to move in a different way. And we hope on our end, I think that the way to approach any big change like this is not to come in hot. You know, we're still, you know, we're a new company. We're just getting started. Not to come in hot and immediately just start disrupting, you know, certain businesses. But go to them and just say, hey, this is what we think, you know, this is what we think your business looks like. You know, you guys certainly know what your business looks like. This is how users on Instinct are interacting with your business now already.

43:28What can we do in collaboration to make that a better experience for both sides where it makes sense for your business? It makes sense for our business. And certainly it's just a much better experience for the end user. That is something that we're exploring and that we're learning across so many different industries right now. If you think about the things that traditional companies could be doing now to prepare themselves for an agent-rich world, let's just pretend half of Americans or something have an agent that's doing all the stuff you just described, which sounds incredible and magical and very democratizing.

43:59I think that's like maybe something to highlight is this is going to bring to everyone capabilities that have been rare or expensive. And I think that's just like a really cool feature of agents in general. But we can come back to that. What kinds of businesses are going to thrive in that world? What should businesses think about doing to prepare for that world to be successful in it? Do you think? I think it comes again back to that breakdown. That's why I was doing that exercise the other day, which is just, you know, where is your revenue coming from? What proportion of that is the user spending time in your application?

44:33What proportion of that is the user, you know, having access to the underlying service? And it's just very clear if your business benefits from more transaction volume, not at the cost or with even at the cost of less time on the application, then you're going to be in a really great spot because Instinct is going to make it, you know, 100 times easier to do that. and if you're in a spot where nearly 100 % of your revenue is due to the user's attention, I think that even in a lot of cases against the will of the user and what they want to do, and there's so many games that, or so many, I would say, malicious product building almost of trying to convince the user against their will to use the application more.

45:16We're thinking about a lot of these, I would say, a lot of these social media companies that are, you know, the user doesn't want, they don't feel happy when they're on the app, right? They're unwillingly giving their time to the app and they can't get off and they keep scrolling, right? But it's because that underlying business is benefiting from the user's attention. It's almost like liberating for the user to be able to, you know, again, this is why it's so important for Instinct to act on behalf of what's best for the user, because it's able to deliver experiences like this where you can actually liberate the user from, you know, being sucked into these, you know, these, you know, these scroll, infinite scrolling moments like that.

45:52I would say any blanket advice is probably just like not well thought out. I think it's a case by case basis. And it's certainly different within different industries. I would say very practically, we're finding where, you know, there are early partners with very innovative, you know, CEOs or other executives that are really thinking ahead. They're really thinking ahead and they're willing to be early partners. And what we're discovering is almost like a playbook for how every business, no matter where they exist along that risk curve, can kind of discover, honestly, what are the risks so that they have a little bit of data to be able to work with.

46:24And then we can work together on finding ways where we can land in a happier spot for both sides. One of those things is you don't have to go all in, right? Like you don't have to say, let's just turn it on, you know, for let's just go and now you see like 70 % of your revenue going to zero and then now you're stuck in this odd place. You can mitigate the risk, right? You can scale down the experiment. You can run A-B tests to figure out, okay, if we enable this certain thing across 1 % of users or something like that, and we find how they interact with the business. And honestly, does the user like it more?

46:59Is it a more enjoyable experience for the brand side too? Does the transaction volume go up? Does the willingness to buy the product or access the product itself? There's a lot of work in product discovery too, right? There's a lot of times the user doesn't know. They want to buy something, but they can't find it, right? And so does that actually increase the user being able to find exactly what they want? So I think doing scaled experiments here is actually a really great playbook to run because you can scale the risk accordingly. At the end of it, you get the data. It's proportional data, so you don't have to run it across your entire user base.

47:33And that's what we're finding is really working so far. Again, it's still very early, so we're still discovering this in real time. And before I ask more questions about the world reordering nature of personal agents, I'd love to take a little side quest in the conversation and talk about what it takes to do all this, to provide all this. I want to hear about what's been hard about building this, the technology itself. I want to hear about compute. I think you said to me at some point that you spend, I don't know, a big chunk of your time just thinking about compute right now. And maybe that's different five years from now, but certainly in the moment when you're growing fast, this is a really important thing.

48:07I'd love to hear your thoughts on that. Talk us through what it's been like to build instinct itself and the key and hard things to do so. Maybe before we do that, just because I'm remembering all of our conversations, it would be helpful for you first to frame up what you want it to feel like and why it's so important to you that it sort of has this distinctive quality and performance before we talk about then how you deliver those things. So maybe first just say a quick word on that, like what you want instinct to feel like as a product. I could talk all day about this because I think that is just so important.

48:42And honestly, as a product builder, it's, I think, a new muscle to flex and to build. But really thinking beyond capabilities is something I really want to push here, which is, I think over the last three years, we've seen, you know, all these different product launches and new products and saying, AI can now do this, right? I can now do that. And did you know that it can do this thing because of this like small technical thing that happened under the hood? And, you know, I think the consumer is first like fatigued by all of this, like, you know, they don't know how to access it. And then second, I think we're missing the point.

49:12So early on, actually, one principle that we held was let's not focus on capability. Let's only focus on understandability. So how much does a user understand about what's happening? How much what is their ability to predict what will happen when they ask this or when they do this or when they interact with it in this way? Honestly, I think that that is one of the major angles or factors that has led to, you know, engagement numbers that are completely off the charts or the viral word-of-mouth growth that is happening at 10 % a day. It's understandable. It just feels, if I were to describe it, for lack of a better word, it should just feel good in some way, right?

49:53In the way that it communicates to you, the purpose of communicating or sending a text message is not just the meaning of the text itself, right? It's down to underlying, you know, even what is the shape of the text message itself, right? And how will the user feel when they see that, right? If you see a big blob of text, that requires the user to scroll a little bit to find, you know, what the next message is, versus maybe you front load some of the information so that the user really gets it in the first 30 % and then can optionally read the rest. You know, thinking about how the user might read or even, you know, like read patterns where they use, I don't know if you know about this.

50:25No, what's that? The way that most people scan big chunks of text, they might read 80 % of the first line and then maybe like 50 % of the next line and then it tapers off so it looks like a flag. That's a consideration that should be made. If instinct, it has the ability to think about things like this, right? Okay, the user is going to spend most of their time reading the first two lines and then certainly the first part of each line too, right? So how does it craft its message to deliver in the lowest fatigue way or with the least amount of cognitive load placed onto the user? So there's a lot of consideration that's put into this in terms of product building.

51:01That is, there's quite a bit of work that we do on the infrastructure side to make it fast and make it affordable to serve. But this is another area that I think is just so important and is going to be a differentiator. How much of that is your personal and the team's taste versus them being the result of a quantitative type process, like iterative quantitative process of, okay, we did ABC testing of all, you know, emoji reaction versus short versus long. And this is the thing that is optimal. Like how much of it is an optimization exercise that's data driven versus your own sensibility and your team sensibility?

51:36Honestly, it's just all of the above because, you know, when you're thinking about building evals or, you know, evaluation or other testing frameworks to be able to test for these like very soft qualities or these actually very long term qualities, right? Two or three weeks in, does the user trust instinct, right? How do you measure that? How do you run, you know, simulated evaluations to test if the user is going to feel trust in two or three weeks from now, right? And actually a lot of it is just staged rollouts over time. So I might come in and build a slightly different experience and then I'll release it to myself and I'll play around with it for a little bit and see how I feel about it.

52:12And, you know, I'm just like very opinionated about these types of things. And then, you know, then if I feel comfortable with it, I'll send it out to the team. I'll say, hey, you guys should try this. We'll see how they feel about it. And then they'll send it out to, you know, our, you know, smaller early access group and then they'll play around with it and see how it feels. And then if we're confident there, if there's any tweaks that we need to make, we'll do that. And then we'll eventually roll it out. to the general public. So I think this is very important because I think that Instinct is quite capable and is one of the most capable products out there.

52:40But I'm not saying that over the next couple months or years that others are going to come and deliver the same, seemingly same experience. But I think that this deep focus and priority on how the user feels and how to make it the most enjoyable experience beyond the words that it's saying, just the way that it feels is just is just so important. One of the great things about the history of technology is this race between incumbents getting quality and innovation versus upstarts like you getting distribution. Obviously, you've built an incredible product, the feel of it. Like you said, it's the worst it'll ever be.

53:15How do you think about that challenge? Like your speed of scaling, what your ambition is for how to get big really, really quickly. Do you think this is a winner-take-most-take-all market? Like what will the market shape of agents be? I'm just really curious how you're thinking about, okay, you've got this foothold, you're growing 10 % day over day. You do that math, it gets really big really quickly. You need a lot of compute. It's a free product. Like there's all the, it just, you're a very chill guy. It just seems like a stressful situation to be in, facing down how big this could get as quickly as it could get.

53:47Talk us through that many-headed monster of a problem. Well, I think you just described everything all in once in about 20 seconds there. And I think maybe if we think about the growth story so far, just to describe like where we're coming from, we are, I guess, like famously or infamously serving an invite only product, which is honestly not meant to be an exclusive thing. Although some users are treating it like that, but that's really not the intention. And the goal here is to be able to, you know, I'm ambitious. I want to scale this thing as fast as possible. But I also want to do it in a responsible way that enables us to, you know, not wake up one morning and have 10 times the number of users.

54:32And then 80 % of them actually can't talk to it because there's not enough compute. The interesting thing, both the benefit and also the detriment is, well, we first started out this program. We gave it to like 200 people. It was like, you know, close friends and family members. And we just said, like, go try it out. and then like the next day like say like five people came onto the platform because they just referred it to somebody oh just to just to describe a little bit more um it's an invite only platform um and every user um will have five invites to be able to get so just five five invites um and so so we rolled it out 200 people next day it was like 205 and the next day it's like 210 so it's like okay cool like you know a couple people are sharing it to like one person right But then that just started to accelerate.

55:15It was very odd. It was like, okay, it was like not 1%, 2%, then it started to be like 3%, 4%. And then once we hit, you know, a couple thousand users, some people just started to share it online too. Just like natively share like, you know, some cool use case that they had with it. And then that accelerated the growth. It actually started to turn into like 6%, 7%, 8%, 9%. And now I believe we're like 10 or 11 % day over day. And just to call that out, it's not that we're doing some sort of creative marketing event. It's like every, you know, we spent zero dollars on marketing so far. It's not that we're doing something every single day to be able to support this growth.

55:48Every day, about 10 % of the audience, or slightly less because you can refer multiple people, are making a decision to give up one of their five valuable invites to somebody else. But that is happening every single day at a 10 % rate. So I think there's something very significant there. When I talk about strength of word of mouth, I think this is, it's honestly a little bit surprising. it's it's one of the strongest uh cases of word of mouth growth of you know i i find all these stories of people saying um hey actually like they'll ask me and they'll actually be ashamed they'll like email me and say like hey can i please get an invite i i i think i have a friend that has it but i don't know if i make it into his five friends right and then there are also people who are saying who are bragging like oh i got these like i got three invites left right and I'm just like holding onto it right now.

56:38So there's so much happening of people that want access to the product, but there's like these odd social games happening of people using their invites. I was seeing the other day, there were some invites that were selling on eBay too. I don't know if you saw this. It's like 300 bucks. People were buying these invites on eBay and it's just to control the growth. So then, you know, the big question comes, the thing that most people are asking right now is, yeah, I mean, you have much bigger players that are able to distribute to, you know, a billion or two billion people on the planet, you know, immediately.

57:12And maybe they might not be compounding naturally as fast, but they have such a great top funnel distribution. And then you have us, you know, compounding at, you know, a very, very fast clip every day. But then, but, you know, we don't own a major, you know, service that has you know two or three billion people and able to distribute immediately that day so it's this interesting question like where does the curve like line up and where's the inflection point and and um there's another problem that comes with this too which is that it's an interesting scaling problem this is what is i think the core of the problem that i i spent like 40 percent of my time uh uh just um worrying about which is it's not like the traditional other consumer products that grew very fast where it's you know to double the number of users on let's say, Instagram or Facebook or something, it would be, you know, this many number of other requests going through the platform.

58:04And like, yes, there's a scaling story there. And it's certainly hard infrastructure work. We also have that to be fair. But what do you do when the underlying compute also needs to, you know, 10 % day every day, we've been doing this for several, several weeks now. What does it mean when the amount of compute that you need access to is now doubling effectively every week. Do we buy compute, you know, 2x of what we have right now? Well, we're going to consume that in a week, right? So then do you buy 5x? Well, we're going to consume that in, you know, less than three weeks, right? Do you buy 10x?

58:38So now it's like your 10x leverage, right? It's if you can even stomach what it's like to buy 10x ahead. But then you're going to consume that in, you know, a couple weeks from now, right? So that is the hard problem. It's thinking about how far ahead how far ahead do you buy it's going at a faster rate than cloud code or codex or some of those other applications where they also had to you know reason about other similar exponential type of problems the other subtlety here is that it's not like a certain business where when you double the number of users you can buy like 2x more resources in order to power it it's that the resource has a lead time of several months right so you can't just go out tomorrow and start you know buy compute because honestly you get you get taxed like three or four x what it is if you're wrong you're wrong by three or four x but then now so two or three months at at 10 let's say it slows down right let's say we don't actually do 10 you know for a very long sustainable period let's just say it's like five to eight percent right but five to eight percent compounding day over day for three four months which is the lead time to get you know bring compute online and that's even aggressive itself.

59:45That's 100 million users, right? So then do you buy compute for 100 million users? So those are the types of questions I'm wrestling with, which is just like, if you're wrong, you're very wrong. You get charged three or four X. What about if you zoom in on the individual user and the cost to serve them on a per day basis? Do you have a set or per week, I don't know what the right metric is, like how much me using my instinct costs in inference per day or something like this? Do you have a sense of like that scale? What is that scale? One thing that I think is to our advantage is we've figured out how to serve the product, which is when we run, no matter how you evaluate, whether it's A-B tests, whether it's internal evaluations or it's tracking engagement across users that might be on one model or the other, we're able to deliver the same performance as, honestly, Opus 5, which is now, I guess, we're dating ourselves.

1:00:40Opus 5 is, you know, like frontier level intelligence. We're able to serve, it's the same engagement rate, the same, you know, A-B test performance. It's the same internal evaluation performance, but at a cost that is very, very low. It is actually very affordable. You know, we're running this program. Every user has the product for free. And our goal is really to deliver this product at an affordable, I'm not going to commit to free for a lifetime for now, but that is my personal goal, to be able to deliver this product for free for everyone for a lifetime. It's just hard infrastructure work to, I don't know how deep we want to get on this, but just if I give one example, if you use Frontier APIs from some of the main providers, you are taking a blanket cost on a certain request.

1:01:31And for all of the requests that might be needed to power that product for that month. But there's a lot of different work that's happening through proactivity that is happening throughout the day that doesn't actually need to finish in hundreds of milliseconds. It needs to finish in, you know, minutes or even hours. There is a batch work that is consuming a lot of content that can be served with, you know, deployment shapes that are 3x, 5x, 8x more efficient with the same underlying compute. So I think that when you customize these inference deployments to be able to perfectly shape the data and the workloads that you're serving, you're able to find these 30 % here, 5x there, 6x there, 10 % here, and all of those compound to a rate where we're able to serve it at a very low cost.

1:02:22How do you think about solving the bigger problem of how far ahead to buy? And if I think about this at true scale, like you get to scale of a billion users or something like this, how much new compute demand do you think this will represent. I mean, it seems CodeGen's obviously been an enormous amount, but ground us in some sense of scale of like, you know, whether it's per user or some way of understanding just like how much new compute this will require. I'm just saying it's going to be a lot because if you just think about, let's just ignore compute and just think about how many tokens are flowing through the platform.

1:02:54The products that are, you know, I would say breakout products of a couple months ago, let's say in the code generation space where, you know, it requires the user to prompt it and then it goes and runs for something and then it comes back to the user and then asks for something else and then the user sends something again. Sure, a lot of that is background work and so there's quite a bit of tokens that are consumed there, but here it is, Instinct has the ability to wake up and to sleep at any moment in time during the day. You know, that might sound a little odd, but its architecture is enabling it to do that so that it can really think about, if you have a meeting that you're running late to and you need to order an Uber or if they ordered the Uber and the user's not showing up, be able to help the user through those moments.

1:03:34So proactivity, I think, is going to continue to expand over time. There's this big build out, this big compute build out with the earlier breakout products in the AI space that are just scratching the surface of proactivity or of background work. Here now, we have something that is almost natively proactive. It is a smaller subset that is actually interactive, right? So I think that the amount of compute that is going to be needed is going to be, honestly, orders of magnitude more than what we thought that we needed. It's just with more productivity comes more care for the user, more time to think about certain things that could go wrong or not go wrong.

1:04:19There's just so much that's happening under the hood. But maybe Instinct wakes up at 6 a.m. because it knows that you wake up at 7 a.m. And then it goes and scans everything, makes sure that everything's ready for the day for you. And then it realizes, hey, actually now is not a great time. It should just go to sleep or do this one thing in the background and not notify the user. And then it realizes, oh, at 4 p.m. there's something that's coming up. There's value that could be provided. Maybe the user doesn't even know how to interact with Instinct in that way. And so but instinct thinks that it's a it's it's like well-meaning and it's something valuable to your user.

1:04:54So it might then just wake up at 4 p.m. and then do the task and contact the user. And then the user will, you know, could lean into it and do that task. You see how much work is going on in the background. That is, you know, with, you know, I would say these coding products, you don't really have that much, you know, background or productivity work happening. So, yeah, I don't mean to quote a number here. I'm just saying that the shape of the product and of the workload that's going to be run is, I think we're just scratching the surface in terms of how many tokens we need. What do you think of Muse?

1:05:29What do you think of the product? I think it's a great product. I was playing around with it for a little bit. And I think it's interesting. It's a different take, right? Because sure, you can say some of the underlying architecture might be similar, but I think it's fundamentally different. I think that instinct is meant to be a simple and very easily accessible. And, you know, of the soft qualities that we were talking about earlier, of, you know, really thinking from the user standpoint about what's important and what's not important and how to make this task easier and easier to read and other things like that versus like a new application and a new interface.

1:06:04And sure, maybe over time we have an application that also delivers on a different set of tasks, but I think it's a great product. I just spend very little of my time thinking about competition and other players in the space because I think, you know, you walk outside, you go to the nearby cafe and you think about how many people within that cafe are actually using AI in the way that they imagined that they would or the way that they want to be using it. And I would say very little. Right. And that's and that's that's here. Right. You know, you go over to other countries or other cities and and it's certainly lesser of the case.

1:06:42So I think I think it's it's still an open space. It's still an exciting time. You know, I think it's an interesting game that's going to be played and rolled out over the over the next coming months and years. So I'm just I'm just focused on building the best product experience. What have been the blunder so far? Like what if what's gone wrong? What have you done about it? I'm sure more things will go. I mean, this is just going to be an explosion of emergent properties and mistakes and the end is going to be really high. How do you think about things to guard against proactively, things to be reactive against?

1:07:14Like, yeah, talk us through the darker side of building this or the harder side of building this. I think it's important to never be reactive and to always be proactive, to look ahead for what new surface areas are being introduced and what new risks might come. And I would say this is the reason why we ran this early access invite program from the start, which was, yes, an early version of the product did have, you know, it didn't have firewalls in place. It didn't have these active monitors in place and so many other pieces of infrastructure that were meant to get ahead and be proactive about being able to provide that much control to the agent and existing system to be able to secure.

1:07:57So yes, there's an early version of the product that had some of these qualities or some of these mistakes. And we addressed it. We went above and beyond and didn't just patch the problem. We built a different system to systematically solve these types of problems. So I think the key here is, again, I said this earlier, the user should always be in control. The user should always be in control of their data. The user can share as much as they'd like or as little as they'd like. And if they ever change their mind at any point in time, they can always, you know, retract to access these certain services.

1:08:32Makes me wonder, what's the future of security? Even if you're the best in the world at this, which maybe you'll have to become, you're going to have so much information and context on so many people. And security is a big problem, I think, across the world. I mean, all these great hacking examples that we've studied now, it's wild what these things can do. The capabilities are going to get stronger and so on. Do you have like a general philosophy? I'm just curious for you to riff on the future of security and safety and guarding against, I mean, early in technology revolutions of the past. Always, there's these enormous hacks, there's enormous data breaches and so on.

1:09:07It's hard to imagine this won't happen in this technology revolution too. How do you think about this and the responsibility of providing safety, given how much you'll know about people? it's the most important problem. I think it comes to building security and safety and a mindset towards the, I don't know, some of the core values of the company and the people and the way that you build product. You know, this is topical because anytime, if you look in the past, the last, you know, 20, 30 years, when any new sort of like breakout or new consumer experience has been has been revealed, there's always been, you know, immediate backlash of like, whoa, this is, this is different and confusing different with being unsafe.

1:09:51And, and there's, there's always this retrack the last 20, 30 years, seeing some of that now. But also, I think just staying core to the principles that you hold, you know, the users, the users always in control of their data, they should never feel out of control. And then being proactive about the systems that you put in place to be able to get ahead of these types of things. If I call out one example, like a hallucination case that existed. First, language models hallucinate all the time, but with a product like this, you don't want a language model to be hallucinating. So we put in place a more systematic solution that will detect before an action is taken, before a thinking trace is executed as a tool call before any action might be taken, it is validated and scrutinized by something that is decoupled from the same incentive system as the underlying agent.

1:10:44It's like a watchdog, yeah. Filter that's able to find, hey, this proper noun was just generated out of thin air due to some sampling error in the underlying model. And so it's very easy in hindsight to be able to capture these types of mistakes. And these are hardened. So we have world-class security teams that are constantly working proactively to find harder and harder adversarial cases, to try to find edge cases here and there. And it's becoming rarer and rarer over time. One subtle part about building a platform like this is that it's getting better over time. As we continue to do more adversarial testing, as we continue to be more creative about certain edge cases, these models are just getting better and better over time towards being robust to these types of attacks.

1:11:30On the other side of the ledger, everyone always shows that beautiful visual of each generation of the iPhone. You can see it getting better over time and more and more refined. What's that arc for you? It's so interesting because it's not an application. It's not a device. It's an interaction through existing communication channels, WhatsApp and iMessage, et cetera. What are the things that you're adding and envision adding over time that will make the the platform and the product more powerful than it is today? You know, the product experience right now is very simple. Very, very simple. Simplicity is one key that we focus on.

1:12:05But I actually think that it becomes even more simple over time, which is we may, you know, roll out an application in the near future, but I actually think that we're going to trend more towards a simpler interface, which is do you even need to open the iMessage application and type in a certain you know a piece of content and send the message out and then look at what the response is after when it's done. We have a certain there's a certain subset of users that interact with Instinct only through voice actually. Like more than 90 % of the interaction of the messages that they send to Instinct are primarily through voice.

1:12:42Where I even have this like action button you know the action button on your phone and it's paired to like I can I click on the action button and I can say like, you know, hey, say hi to Patrick and in two hours from now, I think you have his email address. You can go send him an email and then I can just send and then it goes and sends. And I don't even now need to unlock my phone and then open, you know, the iMessage application and type the message in. And, you know, you can even imagine with real-time voice, with voice recognition that knows what your voice sounds like and has the discretion to know when you're addressing it and not addressing it.

1:13:20Where you could have, let's say, like an AirPod, you know, not a new type of AirPod, just like an AirPod because it's good enough. And it's just on. And it's just on and you might be going on a hike or on a bike ride or on a walk, on a run or something like that. and you're just like catching up like, hey, this contract needs your review. Okay, I need to review that. Or hey, this news article just came in and here's the headline and here's the takeaway. This new project was released and you should take a look at this. And you're just like, and you're like, okay, got it, got it. Put that on my calendar.

1:13:50That's 15 minutes. Okay, yeah, that's not important. So go clear that part out. Oh, this person needs my help. I know the answer to that. So just tell them that it's this, right? It's going to be much, much easier over time. So I think we'll see. I think that there are long-term and short-term considerations here, where I think in the long-term, I think that the interface has become, yeah, again, much, much simpler. But in the short and medium term, you know, maybe there are more expressive interfaces to be able to share. You know, we have like a files feature that enables Instinct to be able to send, you know, effectively entire sites, like full web applications to, you know, show just much more, whether it's like a trip itinerary or, I don't know, like a wedding plan or something like that.

1:14:31I don't know, something that just requires much more and more creativity and more surface area. It can generate those on the fly and be able to show that to you. And I think that there's something interesting there where you look at the last 20 years of application building or product building, where anytime that you needed a new interface to be able to showcase some piece of information, you needed to, well, build an application for it. And it's this long software development lifecycle to be able to produce the application and ship it out to people. And then you find some feedback like, hey, I wish it looked like this.

1:15:00and then you take it and you make the improvement and you ship out a new version, right? And this is on the order of like months, right? And then over time or the last 20 years, there have been all these applications built for so many different purposes, so many different things. And now the consumer is just fatigued, right? It's like, every time I need something, like, is there an app that does that? Like, I don't need, I think that all of that collapses down in the future. I think that all of software is going to collapse down into honestly a single, very, very easy to use interface. Don't get that confused with capability being limited.

1:15:31I think capability is going to expand. What's the craziest thing you can imagine capabilities-wise? When we move away from more of these daily tasks, like I want to do this, can you do it? And that friction is going to go down to zero. I think that it will start towards pursuing higher level objectives that are aligned with Teaser, where the user is able to describe not just, um hey can you track this can you track this workout for me i did this and i and i performed here and this is how i did um but more about hey in in three or four months from now can you can or over the next three or four months can you work with me to make sure that i hit these certain goals and actually a lot of people are are starting to experience that that or are starting to discover that today of you know you want to um you know gain this much uh you know this this many pounds, you want to lose this many pounds, or you want to hit this certain mile time, being able to specify objectives and goals and then be able to work with it to then achieve those goals and objectives.

1:16:28I see small businesses being run now that are native to instinct itself, meaning the entire business is now run on instinct, where the entire back office is now completely functioning on top of instinct. And even parts of it are fully autonomous too, where, you know, talk about high level objectives. It's in this certain area, make sure that this inventory level doesn't drop below this and doesn't go above this. And so now it's not saying, hey, please order this thing. It's not static. Yeah. It's, you know, pursue this higher level objective and use the tools and devices that you have available to be able to accomplish that.

1:17:05So I think that's, I think that's a future. And I'm not like, you know, there are so many cases of objectives that you could, state and have it pursue for an unbound number of amount of time that we could brainstorm here. But I think that's the meta level picture is it's the moving away from individual use cases over to higher level objectives is going to be where I think interaction will progress. Do you want it to have a personality that's distinctive for each person? Like I'm thinking now the movie Her, where there's this relationship that forms between this omniscient, omni-capable agent and the user.

1:17:46What do you think about that aspect? I think of instinct as this very seamless, almost quiet, extremely capable, reliable thing, but not as having a cheeky personality or something. How do you think about that component of the product? Well, you mentioned Her, I guess, comically. It's certainly relationship building or anything in that area is not something that we want instinct to do or to pursue or to develop with users. If I describe more about what it should act like and feel like and represent within the person's life, it's almost like that socially aware operator that knows no matter what room that they're in, what is the best for different people and what is the best interaction pattern for, and it learns that over time.

1:18:38So I think it's possibly one of the most customized, if you even call this an app, it's like one of the most customizable apps ever because it's able to not just, you know, we're not just able to serve a different version of it, it's able to evolve over time. That's why we put so much care into like the software pieces, right? Pick up on when the user didn't like it being said in this certain way, or if there's a lower response rate in this certain way because it's just too much text or it's like, you know, they don't want to look into a huge file to be able to know what you're talking about. It's going to learn those over time and just become easier and more delightful for the user to use.

1:19:12So I think we're not trying to impose certain experiences onto users. I think we want to solve this more from a higher level. It's just ability to adapt to exactly what is the best communication style and task execution style. Why is it called Instinct? Instinct I liked as a name for so many different reasons. I think the main thing is that instinct should not be this like playful thing that you might like bully once in a while and that you like think downwards on and that it's like this thing that only takes the dirty work off your plate. And I think it's this new creative, exciting, you know, competent actor that there's like mutual respect and trust and they feel safe and they trust that it has their back, that it's intelligent, that it's competent, that it's socially aware to know how to act in certain situations.

1:19:57and it's a bet, but I like that instinct is not named like, I don't know, someone's name or something like that to try to personify it into like a human. It's more of, I don't know, I don't even know what an instinct is. I kind of develop my feel for what it is and for what the brand is through just using it. And again, towards really thinking about this feeling of what it should be like, I think that it's a benefit that every user is able to come into it with a fresh slate. There's no prior in their mind about what it looks like or what it's named or anything like that. It's purely through the product experience itself.

1:20:32Who are your enemies and allies? Like I can imagine in WhatsApp or something, you could get shut off. That's a Facebook product and it's a key channel for you. iMessage is an Apple-controlled product. Like who are your friends? Who are your enemies? How do you think about just some inevitable competitive realities here? I would focus less on like Instinct is an iMessage app or Instinct is a WhatsApp app or something like that. And more about just very grounded from first principles. What are the interfaces that users trust today that they're familiar with interacting with today? And how can we deliver the experience of Instinct through the channels that they already interact with every day?

1:21:11And just primarily thinking like we're not pinned to iMessage. Actually, more than 50 % of our traffic doesn't actually run on iMessage itself. But certainly users that are comfortable with iMessage are very familiar with it. I think it's just more about just thinking about what is most practical for the user. And we're going to shapeshift as we see towards just delivering, again, that very simple, easily accessible, sort of like zero friction experience that it is today. I know your latest round is something like a billion dollars at roughly a$10 billion valuation, some of the best investors in the world, Sequoia, Benchmark, KOTU as the leaders of the round.

1:21:52How do you think about the future capital needs of the business alongside this and how did you pick the partners that you picked? I'm just very lucky to be able to work with some of the most supportive partners around the table who have been through different but similarly shaped technology transformations in the past. And I would focus less on the numbers and more about just on the demand in the space, the demand for a product like this or the demand for an experience like this. It's because it really is. It's delivering on that AI experience or that product experience that I think that we've all really been waiting for.

1:22:25And so it's also capital intensive, right? We're a new company, right? We're trying to create something that's going to hopefully be distributed to billions of people on the planet every day and at a very affordable cost. When I say affordable, I mean, you know, compute costs are high and we're trying to deliver it in a very affordable way to our users. So it's just very capital intensive. There has to be some bootstrapping, right? There has to be some bootstrapping, especially if you think about the business model that we're chasing after. Because we could very easily just say, hey, it's going to be a certain subscription and everybody on the platform needs to pay$100 a month, right?

1:23:04And there are shorter term rewards that we can chase after, or we could raise a little bit of capital, use that towards, you know, what is venture capital meant for? To be able to, you know, take certain calculated risks towards, you know, if there are small speed bumps that require a little bit of capital to get ahead of, to be able to prove, you know, transaction volume or to prove, you know, experience across certain industries. That's how you escape these local optimum of, you know, the subscription plan or other things like that. When I do these, I ask the same traditional closing question of everyone.

1:23:34What's the kindest thing that anyone's ever done for you? I've been able to surround myself and be able to build such great relationships with so many different people, whether it's in the industry or outside the industry and where I work or in my personal life, that I'm very grateful to be around. And hopefully, certainly, I can also extend the same level of kindness and thoughtfulness towards them as well. So I would say, I know I'm not answering your question. It's more on a higher level. It's just the qualities of things that I noticed that I really appreciate. Noah, you're building a fascinating company, maybe the most fascinating company of today's AI ecosystem in the world.

1:24:16Thanks so much for your time. Thank you so much.

1:24:22If you enjoyed this episode, visit Colossus.com. You'll find every episode of this podcast complete with hand edited transcripts. You can also subscribe to Colossus, our quarterly print, digital, and private audio publication featuring in-depth profiles of the founders, investors, and companies that we admire most. Learn more at colossus.com slash subscribe. Patrick O'Shaughnessy is the CEO of Positive Sum. All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sum. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions.

1:24:58Clients of Positive Sum may maintain positions in the securities discussed in this podcast. To learn more, visit psum.vc.

1:25:25you know how small advantages compound over time that's true in investing and just as true in how you run your company your spending system is your capital allocation strategy ramp makes it smarter by default. Better data, better decisions, better economics over time. See how at ramp.com slash invest. As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at vanta.com slash invest. The best AI and software companies from open AI to cursor to perplexity use WorkOS to become enterprise ready overnight, not in months.

1:25:57Visit workos.com to skip the unglamorous infrastructure work and focus on your product. Ridgeline is redefining asset management technology as a true partner, not just a software vendor. They've helped firms 5x in scale, enabling faster growth, smarter operations, and a competitive edge. Visit ridgelineapps.com to see what they can unlock for your firm. Every investment firm is unique and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your data, understanding your process, and producing real outputs. Check them out at rogo.ai slash invest.

1:26:31Thank you.