The Exchanges

Q Amazing. Where does physical AI show up first in a way that's really economically real? Are you seeing that already?

A We conjectured that, ah, that robotics was going to come along and decided that the first application of robotics that has both a large enough market Um, relatively standardized technology so that we could scale and get the flywheel going, um, and has real economic value was, uh, self-driving cars. And so, uh, inside Waymo, uh, our chips from NVIDIA, uh, at, at Tesla, we were in the car, uh, now we're in the data center. Um, uh, Mercedes, we're in the data center, we're in the car, we're the software stack. Uh, we, uh, worked on Alpamayo, and we open sourced it, and the reason why we open sourced the self-driving car stack is because you need it for agriculture, you need it for mail delivery, you need it for warehouse AMRs. There's so many different ways that you could apply, um, uh, autonomous navigation, uh, and none of those markets are big enough to be a self-driving car market, and we thought it was sufficiently diverse that we would create the whole stack for it. And so we're working with autonomous vehicles in all kinds of different places. Our robotics business, autonomous vehicle business, basically physical AI business is probably almost, it's like ten billion dollars, so it's really, really big already. Um, likely this will be one of the largest industries in the world, and, um, uh, it'll take longer than a couple, two, three years. It'll take less than 10, and so th…

AI assessment note: “first application of robotics that... has real economic value was, uh, self-driving cars.”

Q fetal DNA that naturally exist in the mother's blood, and that this would someday be universally adopted. This was a radical idea at the time. Before Billion to One, most genetic abnormalities could only be detected via amniocentesis, an invasive procedure that is only used in high-risk pregnancies. How is the key insight that enabled you guys to do this when no one else was able to do it before?

A We have realized that DNA that is coming from the fetus and the tumor is both very dilute and rare, right? So you might only have a few molecules among billions of other molecules. So every molecular diagnostics approach here requires In the lab using a process called PCR to amplify this DNA billions of volt. The problem is that this DNA amplification process can add tremendous noise so that the small signal that you have can be lost. So what we have done is to add a synthetic DNA into the patient sample that we get before any amplification happens. These synthetic DNA Allow us to know how much amplification happened at different genomic locations. You know, what are the errors that are being introduced by the amplification process? So then we can remove those errors from the sequencing data, the data that we get at the end, so that we know what was in the sample to begin with. That converts a difficult biology problem to almost a simple mathematical problem.

AI assessment note: “what we have done is to add a synthetic DNA into the patient sample”

Q I'm always curious. I mean, you were mentioning second sight sort of, you know, flashes of light, and yet, you know, here, you know, how did you figure out the API? I mean, if I was, you know, trying to reverse engineer it, I guess I would, like, try to measure the signals. Is it similar with, you know, biology?

A It's just, it's difficult to measure the signals. So, brain-computer interface research and development is limited by your ability to record and stimulate these signals. The neuroscience, comparatively, is actually pretty simple. As soon as you can record the signals, we've very quickly figured out what, we talk about neural representations, what they are. Second site's instructive. So in the retina, there's three layers of cells that matter. There's a hundred fifty million rods and cones. This connects to a hundred million bipolar cells, bipolar because they've got two ends, and that connects the rods and cones to 1.5 million optic nerve cells. We call them retinal ganglion cells. Ganglion is like a fancy word for, like, reaches a far distance and connects to somewhere. We stimulate the hundred million bipolar cells. Second sight stimulated the 1.5 million ganglion cells. And so they were trying to get the signal into the brain past that 100 X compression. And the retina was doing a lot of computation there. The eyes of camera light shines in from the front. It hits the rods and cones like that. The representation in the rods and cones is a bit mapped image. It's just like you take the image, you tile it across the rods and cones. That's what it is now. And the The 1.5 million optic nerve cells. It's not like that. Like if you just project an image onto them, you get a bunch o…

AI assessment note: “As soon as you can record the signals, we've very quickly figured out”

Q Maybe let's start way back. I mean, when you were, this is a room full of 18 to 25 year olds, um, that skews younger because the founder set is younger and younger. Uh, can you put yourself back into their shoes when, you know, you were 1819, you know, learning to code, uh, even coming up with a first idea for ZIP-II. What was that like for you?

A Yeah, back in 95, I, I was faced with a choice of either do, uh, you know, grad studies, PhD at Stanford, uh, in, in, uh, material science, actually working on ultracapacitors for potential use in electric vehicles, essentially trying to solve the range problem for electric vehicles, uh, or Uh, try to do something in this thing that most people had never heard of called the internet. And, um, I talked to my professor who was, uh, Bill Nix in the material science department and, uh, said like, um, can I defer for the quarter? Uh, because this will probably fail and then I'll need to come back to college. And, um, and then he said, this is probably the last conversation we'll have. Uh, and he was right. Um, so, but I, I thought things would most likely fail, not that they would most likely succeed. Um, and, um, and then in 95, I wrote, uh, Basically, I think the first or close to the first maps directions, uh, internet white pages and yellow pages on the internet. Um, I just wrote, I just wrote that personally and didn't even use web server. I just read the port directly because I, um, couldn't afford, uh, and I couldn't afford a T one, uh, The original office was on Sherman Avenue in Palo Alto. Uh, there was like an ISP on the floor below. So I drilled a hole through the floor and just, uh, ran a land cable directly to the ISP. Um, and, um, yeah, uh, my brother joined me and ano…

AI assessment note: “back in 95, I, I was faced with a choice”

Q Can you give some examples of these challenging Taser capabilities that people should explore that it can do now that it couldn't six months ago?

A Yeah. So, okay. One example is the model can now rewrite essentially any code base from one language to a different language. It's, Just sort of crazy. Like it's this work that would have taken just like a very long time as an engineer, and now the model's like quite fast at it. So, so one example of this is, um, Cloud Code is built on the BUN JavaScript Runtime. It's a open source JavaScript Runtime. Um, it's an alternative to Node.js. It's kind of a faster node. BUN was written in Zig. Zig is a systems programming language. It's kind of like C, it's, it's very low level. One of the problems with C, with, uh, with Zig is you have to manually manage memory, and so it's quite easy to run into situations where there's like memory leaks and, you know, other memory management issues. And so one thing that the Bund team was doing is they were having Claude fuzz the code base and try to simulate and trigger memory leaks, and they were doing this for, you know, for a long period of time. They were able to find a lot of memory leaks. It was sort of like a case at a time, and that was kind of the capability of the model at the time was doing this fuzzing. And then at some point, Jared on the team was like, okay, let's just like rewrite it. Maybe the model can do this. And I, I think this is like one of these test problems that he kind of threw at the model with every new model generatio…

AI assessment note: “One example is the model can now rewrite essentially any code base from one language”

Q Is that part of the thrust behind NVIDIA being so involved?

A I think, well, first of all, I, I need to understand agents because agents is the new software, and how is this new software processed matters a lot to computer architecture. And the, the more intimate we are about, um, the nature of agents and how it's different than, than, um, uh, chatbots, which is how different than, than, um, maybe inference in the very beginning, however we think about these processing layers, the more intimate we are about the nature of the processing, the better we could design systems. We, we kind of have to live in the future five to 10 years because it takes three or so years just to build a system. It takes a couple of years to ramp it up and you're dealing and you would like them to be able to use the computer for 10 years after. And so you kind of have to live in the future for a while. And so agentic systems for us at the first principles is just what is the workload? What's the algorithm? How is it going to evolve? Where are the bottlenecks? You know, where are the MDOS laws problems? And, um, uh, how does it scale? Uh, what happens to concurrency? How do you deal with sandboxes? Um, how do you deal with MCP? How do you deal with, you know, working memory, long-term memory? How do you have all these autonomous systems, asynchronous systems working all the time? And so what kind of design architecture makes perfect sense for that? And so we have …

AI assessment note: “I need to understand agents because agents is the new software”

Q Lawson at Twilio built that company from nothing. To four billion dollars in, you know, actual revenue, like stock up 390% since IPO. I mean, by all accounts, you know, smash rip roaring success. And then his super voting shares, uh, expired after 199 days and he's out. And so less than half a percent of shareholders did that. You know, how did that happen? Like, What, what's going on?

A It's like, it's honestly unbelievable to me. Well, here, let me make the case for why he needed to be fired. And then you'll see if this makes sense to you. So what happened was, so he took the company public. He agreed as part of the IPO prep process, as a lot of founders do, that he would have dual class control, founder control. He'd be the mission guardian. The protections would sunset after seven years. Man, you're taking a company public. Seven years sounds like a long time. But man, in the public market, seven years is just, that's barely a beginning. It's like a handful of quarters. Anyway, that was the deal. He made the deal. His advisors and everybody told him, don't worry about it. You can always extend it. You know, you could, it's always too early until it's too late. It's kind of the idea from the book, like, okay, whatever. So seven years come and go. Now it happens to be Those seven years include the pandemic years, as you well remember, the run-up in telecom and tech stocks we had. It was like Golio stock was just up an insane amount. That bubble burst and the stock came way down. So at the time he was fired, the stock was down like 80% from the peak. And it's like, oh, well case closed. But if you measure from the IPO or even from the pandemic peak, revenue was up. It's like, did, did the business go down? Was revenue down? Was there some kind of problem? No. …

AI assessment note: “The protections would sunset after seven years... the stock was down like 80%”

Q a very unique Technical insight that in the past robotics folks would have kind of gasped and be shocked because robots need to run in real time. A lot of times all of the compute runs in on device, but you guys have done something very different. Can you tell us more about that? So that this works in, in, in real time with large models and, and really well.

A So the context here is that, you know, we talked to many companies that would like to deploy robots and one of the First question we get is what compute units should we get on the robot? You know, it's expensive, it's going to increase the bomb cost, and they're worried that it's going to go out in fashion very quickly because the model changes, the model gets bigger, how do I make sure that the hardware that I'm going to commit to today is going to be viable for, you know, a couple of years? It's a very difficult question. People are often really surprised when I tell them that almost all of the robot evaluation that we run at Pi today, including the really Complicated demo that we have shown making coffee, folding laundry, mobile robots navigating around. The model actually hosted in the cloud. Um, and you know, this is not like a cloud as in a server in the office. It's a real cloud. The model is hosted in a data center somewhere, and within this high frequency control loop that, um, is controlling the robot, the robot is actually querying an API endpoint that hosts the model, sending it images and language command and getting back action that then Executed directly on the robot. And this is surprising because of precisely the reason that you mentioned, you know, how do you actually make it work? This is why it's really important for Pi to couple system, hardware, and model …

AI assessment note: “The model actually hosted in the cloud... you can actually bury the inference time”

Q used the product themselves, they have used many products that under the hood are actually powered by what you built. And so indirectly, they're actually like using stuff built on top of Variance, like a lot of like software infrastructure kind of products. Can you maybe tell people about a Like a specific product that they've used that's, or have probably used, that's powered by Variance and how it works?

A One of the customers, GoFundMe, is a platform that you can use in order to build your own fundraisers. And GoFundMe is effectively a payments platform. And GoFundMe actually has some very strict compliance requirements because they are liable for facilitating payments to, let's say, an organization that they shouldn't have done so. So Variance is used to verify, for instance, if a fundraiser is going to be built for A military operation, for instance, or if you're building a fundraiser for, um, a crisis that you were not part of, which would be fraudulent, well, GoFundMe sort of has the responsibility to be able to detect this. They're using AI agents, so variance AI agents, to conduct those investigations, so make sure that the money is going to the person that they say they are, and also make sure that the money is not going to be funneled to sanctioned countries, for instance, Or be going to, um, anything that could be a compliance risk for them.

AI assessment note: “One of the customers, GoFundMe, is a platform that you can use”

Q What are other like products that people use that also are powered by variants and they don't even realize it?

A So the scale has been, um, impressive. The number of use cases that our agents can be used for. So We have been running complex identity reviews for marketplaces, gig economy platforms, so when you sign up to actually be, for instance, a delivery driver, your identity needs to be verified based on selfies, based on your driver's license. All of that data is then going to be reasoned on by our AI agents, and then it's also going to be validated based on the company's standard operating procedures. That's a good example of how Variance is used. We're also used for complex What we call KYB verifications. So if you sign up to do any sort of business online, whether it's for a marketplace or also to, with a financial institution, they have the compliance requirements to verify that you are actually linked to the business you say you, you own. So a good example is, for example, I sign up to say I'm going to be doing business with variants. Well, The legal name of my company is Decoy Technologies, and it is tied to Korean Malata. That's a simple example, but building that graph at scale is really hard, and oftentimes you're going to see company have multiple shell companies, be tied to multiple different agents and different other identities, and within that really large graph, you expand the area of risk for the company where one of these nodes could be in a sanctioned country. One o…

AI assessment note: “when you sign up to actually be, for instance, a delivery driver”

Q Unstructured, isn't it? Don't you just, like, pull it, like, right, do you pull, like, out of their, like, relational database?

A I, I wish it was this, this simple, but usually, I mean, I'll give you a very concrete example. So, for instance, if you need to verify a fundraiser, I'll pick the fundraiser example, usually the data is going to be scattered around the user identity data, so you need to have information on the user, you need to have information on All of their login behaviors, the devices that they've had, the PII that they've onboarded at the beginning. You need to also be able to pull in information about the business, and then you need to have also all of the information on the fundraiser itself and all of the history of that fundraiser. What has been hard is that oftentimes, whether you're a financial institution or you're a marketplace, that data is going to be scattered across five to 10 different systems. It's going to be into different data stores. And one thing that's been really interesting in that is that sometimes that data is going to be hidden behind a UI. So the only way that the Variance AI agents are able to sort of scoop up that data and reason over it is to be able to directly scrape from a UI that was built for a human. So the data piece and being able to scoop up all that data and bringing it to Variance was one of the hardest technical challenges, really an onboarding one.

AI assessment note: “I wish it was this simple, but usually”

Q Well, that's awesome. Maybe let's change gears here and talk a little bit about the, the origin story. How did you and Michael end up starting this company? How did you end up working on this problem?

A So Michael and I both met, we were co-workers at Apple. Um, we were both engineers on the fraud engineering team, and I was a data engineer. Michael was a machine learning engineer, and it was really interesting because the team in and of itself was sort of the Fraud engineering as a service team of Apple. We were providing our services to the iMessage team, the iCloud team, and we were this centralized fraud team, and Michael's machine learning decisions were then dispatched to the rest of the organization through my own streaming jobs. So we had a very sort of symbiotic relationship, uh, from the get-go. We knew that we worked really well together. I remember telling Michael, oh, well, What if the right vehicle for this product was a company? And I told Michael, oh, we should apply to my Combinator. That was sort of the origin story. I think it really started from the product. We really, really wanted to just see this product exist. Um, and we wanted to see that problem that we were solving at Apple be solved in a much more efficient way, in a much more self-healing and resilient way.

AI assessment note: “Michael and I both met, we were co-workers at Apple.”

Q Okay, so the, the, the first big battle was, like, getting the first customer. Took eight months to land IAC. You really did it the hard way because you went enterprise from the, from the very beginning. Have there been any other hard, like, challenges of building this company?

A I think starting a company tests you in a lot of different ways, um, and we have a really interesting story. You know, after IAC, we got to onboard a lot of great customers in the trust and safety space, um, Medium, of course, GoFundMe, Redbubble, And around July, 20, 24 was one of the times where the company was growing rapidly. We were onboarding more and more enterprise use cases. I think during that month, our revenue was doubling within the month and then doubling the month after. It was really exciting, very step function because it's enterprises. And we had just wrapped up one of the largest trust and safety conferences, uh, called TrustCon in, in San Francisco. And I think a couple, I want to say the day after we had worked so hard for this conference, 12 hours a day, we were super tired. Um, I was going back to the office on a Sunday afternoon, um, and I was on the bike lane, and a truck hit me. That was a really, really crazy experience to, to go through as a founder. I mean, the company was doing well, but at the end of the day, we were A 10 people team, and the CEO just gets hit by a truck.

AI assessment note: “I was on the bike lane, and a truck hit me.”

Q I guess for V two, what, what caused the, so one was clearly reasoning, two, a benchmark doesn't care how you solve it. I guess embedded in what you said, like, were people using code gen to then solve?

A That's right. So not, not necessarily code gen, uh, per se, but, uh, the Frontier Labs has been targeting ARC V two and, uh, the progress you saw on ARC V two is actually a result. Uh, this very, very large scale targeting. So what you can do to solve RG-II is you ask your reasoning model to make more tasks like those in the benchmark. Uh, and then you try to solve them using, let's say, let's say program induction, for instance, uh, uh, still using your reasoning model. Then you verify the solution. Again, it's very viable. So you can, you can trust, uh, the answer. Um, and then you fine tune the model on the successful reasoning chains. And then you keep repeating, like, you generate new tasks, you solve them, you verify the solution, you fine tune the model on the reasoning chains, and, um, you can keep doing this millions of times, right? Like, you just need to spend more money.

AI assessment note: “not necessarily code gen, per se, but, the Frontier Labs has been targeting ARC V two”

Q You're one of the best examples of someone who came from a pure software world and then went into hard tech and now is actually doing real breakthrough type of research and work that is also commercializable. The people watching, they might be on a similar track. Knowing what you know now, like, what would you tell to the sort of 2016 version of yourself?

A So I think there's two things. The first is, um, like the thing that I did, and then there's the thing I didn't do. The thing that I did, I think, was I had, I had a, A clear sense of what I wanted, and then I was very high agency towards that. When I was in college, I knew that I wanted to work in brain computer interfaces. There was a great lab that was doing that work at Duke where I went. And I was pretty persistent in figuring out how to like place myself into that lab. It was in the medical center. They didn't usually take undergrad. They were like, it took me a little while to get in there. I eventually figured out that I could sneak in by taking an independent study in the chemistry department that would like be a backdoor into this like primate neuroscience group. But then really most of my education in college happened in that lab. So yeah, I, I grew up programming in my, my deepest hard skill is software, but I, I've been doing primate brain computer interface, like closely neural decoding stuff since And so that was just like, you had to be pretty high agency and, and like, um, persistent in trying to like, if like follow through on that, but that only works if you have a sense of where you want to go. And so the first is like, figure out what you want. The thing I didn't do was my, so after college, I started a company, um, called Transcriptic. That was a, the, it …

AI assessment note: “So I think there's two things. The first is... figure out what you want.”

Q And then compounded with the agents who then are sort of the oracle telling you what the best tool is. And we are actually seeing some of those trends with the growth of YC companies, dev tool companies. That are doing really well because of this, of all these trends. Maybe we should talk about those, and why is that?

A I mean, one that springs to mind, it's like a, a stat, actually, a friend of ours, Yuri Sagalov, had mentioned to me a while ago. It's just like, then, if you look at, like, the number of databases being created over, like, simple database, Postgres databases over, like, the last 12 months, and the numbers just exploded, um, and that's because it's all people vibe coding and building apps, and the agents going out choosing, like, A database tool and like a knock on effect for that for YC company is Superbase has just seen like an explosion at the demand for databases, right? And the age was interesting is the agents are choosing Superbase as a default tool to like set up and host their Postgres database. Like, because if you like go out and read the documentation online, like Superbase has the best documentation is reasonable for the agents to assume that that's like the best tool to use.

AI assessment note: “a knock on effect for that for YC company is Superbase”

Q What do you think are some of the tips for anyone that wants to build a coding agent since you've done it a lot? What are, what are some now lessons that you learned that you want to share?

A I mean, I think the number one thing, uh, is managing context. Well, basically we kind of had like a checkpoint. For, uh, I think it was O three, like one of the reasoning models. And then we did a bunch of fine tuning on it, um, and reinforcement learning where it's like, oh, you're given a bunch of questions to like solve these coding problems or like fixed tests or whatever, implement a feature. Um, and then the model was RL to respond to those. And so I think most people are not going to be doing that. Right. But the things that you can do are figure out like, Hey, what context should I be supplying to this agent to get the best possible result? And so for cloud code, if you watch it working, it's like, oh, I'm going to like spawn a bunch of these explore sub agents. They will like search for different patterns in the file system. They will come back. Uh, they will have this context. They'll summarize it for me. And then I'll have someplace to go. It's interesting watching like different agents structure this context. Uh, like I think cursor takes an approach where they actually do semantic search, where they embed everything and figure out like, Hey, what query is closest to this? If you look at a codex or a cloud code, uh, they actually just use like grep. Uh, and I think that works because well, yeah, it works very well because code is very context dense. Um, like if you…

AI assessment note: “I think the number one thing, uh, is managing context.”

Q What about internally? Like, what was the vibe internally at Amplitude on the team? Like, was there anyone internally asking, hey, like, should we be building AI products or doing something with AI?

A There were a few, uh, people that were kind of testing out ideas, but I think we were so focused on our regular motion. Like, we were coming out with a lot of other products outside of analytics. We were doing You know, we launched experimentation. We were building session replay internally. We were building this thing, uh, activation, which targets your users based on their behavior, and so there was a lot to do. That was just kind of right in front of us. Um, that was clear, like, okay, hey, we can be much more competitive with, uh, you know, and, and there's this revenue that's just sitting here if we go build these things. You know, there were a few folks that were messing around with it, and there, there were some people that had used cursor and other things, but the, the, the organization as a whole did not have, it's not like it was conscious and aware of this change, this, this, this massive change that was about to happen. I give a lot of credit to Wade, and I give a lot of credit to the command team of being the tip of the spear in terms of showing people what is possible, um, and then getting the organization to, to embrace. By, uh, early twenty-twenty-five, I was convinced, and I was like, all right, we have to be very aggressive on this. Um, so the first order is to train the organization of what the capabilities of AI are, specifically engineering.

AI assessment note: “There were a few, uh, people that were kind of testing out ideas”

Q And so, I guess, to get the most out of a tool like Juicebox, how detailed should my search be? Like, it's almost like I'm prompting Juicebox. So, like, am I trying to write, like, a full system prompt here? Or, like, give it, give us, like, a sense of what I should be doing?

A I think the easiest way to learn is kind of by starting with what feels natural. That might be a couple sentences about the role, and then Juicebox gives you a few different ways to filter further. And so we have a feature called Autopilot, which is personally my favorite feature in the platform. It lets you define criteria, and then you'll get a check mark for every profile who matches that criteria. Criteria can be super specific, like has published a paper in at least two peer-reviewed journals, something that doesn't exist as a filter otherwise, and you can write your own. You'll get check marks for everyone who matches that, and then continue customizing from there. And so I'd start with like a fairly broad initial prompt, maybe a couple sentences, and then as you fine tune in and get more narrow, adding additional criteria.

AI assessment note: “I'd start with like a fairly broad initial prompt, maybe a couple sentences”

Q When it comes to sort of like the channels sort of outreach, we have different channels, um, Should I be trying to contact the same candidate through every channel, or should I be focusing on one channel, um, and do certain channels tend to have better response rates than others?

A Ideally, yes, we'd reach out to every candidate on every channel. Um, practically it becomes difficult and quite time consuming. I'd say the default should always be email because it's what you can automate and get real data on. What are your reply rates, open rates, and more. If you're just sending a single message, your best response rate is probably going to be on I guess apart from like a Twitter, it's probably going to be on LinkedIn, but the advantage of email is that you're not just sending a single message. Instead, you're orchestrating this multi-step campaign, and so with those multiple steps, you're going to get a better response rate on email than you would with a single message on a LinkedIn. Um, and so I think the easiest way to get started is just email automation, um, improving your response rates by adding in a LinkedIn step, and then if you want to get creative beyond that is when you can start layering into other channels.

AI assessment note: “Ideally, yes, we'd reach out to every candidate on every channel.”

Q I have a question for you, which is that, you know, you've talked to a lot of companies, I assume, through the interview process that are doing crypto. What are the, the crypto companies that have stood out to you as, like, exciting for YC, and what are their characteristics?

A Yeah, you know, I mentioned a couple of them, like, Dollar App, um, Aspora, um, Sort of both in sort of like the, I would say like neobank built around building a new experience for their users with stablecoins in particular, but just more generally a better experience for users in those countries. I think a really interesting one that I've personally been involved with that I'm excited about is called Courtyard. And they started out as a sort of a collectibles marketplace. And what they, I think they were doing was, um, interesting and innovative is that they wanted to like actually verify that the collectibles were like, um, authentic. And so they would, they would actually manually verify That, like, you, the, the baseball card was what I said it was, or the collectible toy was what it said it was, and then put that onto sort of like the, the blockchain is just like, you know, immutable proof of the authenticity, and then sell that on the marketplace.

AI assessment note: “Dollar App, um, Aspora... I'm excited about is called Courtyard.”

Q What are some of the other, like, applications of stablecoins that you're personally really excited about, or do you think builders might not recognize as great opportunities?

A Well, this is one where I, I think that there's an opportunity that runs counter to the narrative. Because there's a ton of excitement about dollar stablecoins, and I'm really excited about dollar stablecoins. I think there's so much more that we can do with them. But at the same time, you know, as I've talked to builders and as I've talked to regulators and businesses in countries around the world, so many of them have excitement about dollar stablecoins because they solve an immediate need. You know, if I have inflation in my country, I can now save in dollars and that that helps me, you know, have a more stable savings account or more stable business account. At the same time, when I talk to those folks, I think that there is an intuitive feeling of Oh man, we have our own local economy and we don't want that economy to be dollarized, right? Like if I'm in Kenya or Brazil or Nigeria or, you know, Indonesia, we have excitement about building our own local economy on the Brazilian real or the Nigerian Naira or, you know, that, you know, kind of excitement about investing in one's local economy. And then the, the, I think kind of intuitive feeling of, Ooh, there's this tension between getting the value from dollars, but also not having dollars dominate in a way where they crowd out the local economy. I think to me, that is one of the most exciting opportunities to kind of lean …

AI assessment note: “entrepreneurs in all those countries who are creating stable coins for their currencies”

Q How about for like US-centric builders, um, How should they think about the opportunities to build in stablecoins right now? And maybe a bit of a segue here, but I know that Coinbase just launched like a commerce payments protocol with Shopify. How does that tie into this and how does that create opportunities for founders in the space?

A Now that we have the fast chains, we have the, the scaled stablecoins, we have the mature regulatory environment, we have the easy to use wallets. I think the opportunity is basically looking at every single part of the financial system that exists today and figuring out how do we turn that from Legacy systems that in many cases are 50 to a hundred years old into programmable smart contracts that live on chain. And you can look at every part of the world and be like, wow, look at this system that is written in cobalt that, you know, has millions of lines of code that has all of this cruft that has massive fees. Can we write that into a smart contract that's 500 lines of code? And the answer is yes. It's literally yes. And this is exactly what we saw with Shopify. You're talking about, you know, this commerce payments protocol that we built. That lets, um, any Shopify store in the world accept USDC on base from anyone else in the world. And the thing that we did to unlock that was we embedded with the Shopify team and, you know, shout out to Toby and Shopify. They came to us and said, we think the moment is now for us to look at all of our existing acceptance systems and figure out how can they be rewritten in smart contracts. And what we did over a nine month period was exactly that. We went into their systems, which are literally millions of lines of code where they're doing t…

AI assessment note: “I think the opportunity is basically looking at every single part of the financial system”

Q So there was some wisdom to figure out which bucket it fit. Is this just for this vertical or it could be generalized? So could you give us like an example of what that looked like in terms of the products and verticals and what fit in one bucket versus the other one?

A Yeah, I mean, probably the, the sort of like most basic example here is, is sort of the invention of the Palantir ontology itself. And so when we first started talking about working with, you know, the U S government and specifically working intelligence, you know, should we have a database table for people and a different database table for money and a different database table for this. And this is super obvious. I think at this point, if you, if you go down that route and you try to deploy to multiple people, your, your database doesn't make any sense. And so, you know, the, the change here would say, well, we need to pull this up to a higher level of generalization. And instead of thinking about specific types of objects, Um, we should allow that to be defined per customer by the forward deployed engineering team. And so that's the sort of origin story of where Palantir famously got its ontology.

AI assessment note: “most basic example here is, is sort of the invention of the Palantir ontology”

Q Yeah, but you wouldn't, don't try this at home. Mindset, right? Like now everyone's sort of, it's become, like Diana Dana said, it's become very commonplace. Has that, one, has that surprised you? And then two, like, why do you think that's happened?

A This was absolutely a surprise to me. That, you know, my first, second, and third pieces of advice to people who are thinking about trying an FDU strategy is like, don't, Don't do this at home. If you can avoid it, like it's, it's probably bad for you. Probably you're going to end up doing services. And then only if you really try hard not to do it and fail, then, well then maybe actually it's a moat for you. If it's the only thing that can possibly work in your market. So what's special about this market, right? Why does the AI agents market, uh, work this way? Maybe the, the starting place is why did Palantir have to adopt this? The Palantir market is not One coherent market, right? So we were working with national intelligence agencies, with national law enforcement, with the military. All of these organizations had some similar projects, right? But even, you know, the difference between a counter-proliferation workflow and a counter-terrorism workflow, one you're trying to figure out, you know, who's building bombs and the other one, well, who's building nuclear bombs and who's building IEDs. And those are actually quite different in terms of how they work. And so there's this incredible heterogeneity, and the market, you should really think of the market as different segments. Inside each segment, you can build something, and, you know, the crossing chasm story a little bi…

AI assessment note: “This was absolutely a surprise to me.”

Q the curve, who are in your circle. Like, you know, How should people approach that? Like, what has worked for you, and what would you pass on, like, you know, to, to X, to your children? Like, you know, what do you tell them when you're like, you need to make your way in this world? You know, here's how to construct a reality that is predictive from first principles.

A Well, the, the tools of physics are incredibly helpful, uh, to, to, um, understand and make progress in any field. Um, first principles means just, obviously just means, you know, break things down to the fundamental axiomatic elements that are most likely to be true, and then reason up from there as cogently as possible, as opposed to reasoning by analysis or metaphor. Um, And then it just, simple things like, like thinking in the limit, like if you extrapolate, you know, minimize this thing or maximize that thing, thinking in the limit is, is very, very helpful. Um, I'd use all the tools of physics. Um, they, they, they apply to any field. Um, this is like a superpower actually. Um, so you can take, say, take for example, like rockets, you can say, well, how, how much should a rocket, rocket cost? Um, the typical approach to how, to, that people would take how much rocket should cost is they would look historically at what the cost of rockets are. And assume that any new rocket must be somewhat similar to the prior cost of rocket. A first principles approach would be you, you look at the materials that the rocket is comprised of. So if that's aluminum, copper, carbon fiber, uh, steel, whatever the case may be, um, and say what, what, how much does that rocket weigh and, and what are the constituent elements and how much do they weigh? What is the material price per kilogram o…

AI assessment note: “break things down to the fundamental axiomatic elements that are most likely to be true”

Q We haven't talked about Neuralink at all yet, but I'm curious, you know, you're working on closing the input and output gap between humans and machines. How critical is that to AGI, ASI, and you know, once that link is made, can we not only read, but also write?

A The Neuralink is not necessary to solve digital superintelligence. That'll happen before Neuralink is at scale, but Uh, what, what Neuralink can effectively do is solve the, um, the input output bandwidth constraints, especially our output bandwidth is very low. The, the out, the, the sustained output of a human over the course of a day is less than one bit per second. So it's, you know, 86,400 seconds in a day, um, and it's extremely rare for a human to output more than that number of symbols per day. So, um, certainly for several days in a row. Uh, so You, you really, um, with, with a, with a neural link interface, you can massively increase your output bandwidth and your input bandwidth, um, input being right to, you have to do right operations to the brain. Um, we, um, we, we have now five humans who have received the Uh, the, the kind of the read, uh, where it's reading signals, and you've got people with, with ALS who, um, really have no, they're, they're tetraplegics, but they, they can now communicate at, with, um, Similar bandwidth to a human with a fully functioning body, um, and control their computer and phone, uh, which is pretty cool. And then, um, I think in the next six to 12 months, we'll be doing our first implants for vision, where even if somebody's completely blind, um, uh, we can, we can write directly to, um, the, uh, the visual cortex. Um, and, and we've…

AI assessment note: “The Neuralink is not necessary to solve digital superintelligence.”

Q So some people would say that is what we have today. You sort of describe what you want and out it comes. What would you say to that? Like, are we there yet? You know, what are the steps to where you really want to go?

A We're seeing the first signs of things really changing. Um, I think you guys are probably on the forefront of it with YC. Because I think that in smaller code bases with smaller groups of people working on a piece of software, that's where you feel the change the most. Already there, we see people kind of stepping up above the code to a higher level of abstraction and just asking essentially agents and AIs to make all the changes for them. In the professional world, I think there's still a ways to go. I think that the whole idea of kind of vibe coding or coding without really looking at the code and understanding it, it doesn't really work. There are lots of end order effects. You know, if you're dealing with millions of lines of code and dozens or hundreds of people working on something over the course of many years, right now, you can't really just avoid thinking about the code. Our primary focus is to help professional programmers, to help people who build software for a living. In those environments, people are more and more using AI to code. You know, on average, we see about people using, you know, having AI write 40%, 50% of the lines of code produced within Cursor, but it's still a process of, you know, reading everything that comes out of the AI. And so an important chasm for us to cross as a product will be getting to a place where, uh, you know, we become less of a p…

AI assessment note: “In the professional world, I think there's still a ways to go.”

Q I guess historically, do you have a sense, like, when did it happen? How did this happen? You know, there was definitely a moment where The Democratic Party did not have this problem, and yet it sort of emerged maybe over the course of 20 years.

A I like to tell this piece of history, and I, and I think it's true, that we've really had two very different regimes of building in this country, um, in the last 100 years. Between, say, the 19 thirties and 19 sixties, we changed the physical environment of this country dramatically. We built so many homes In the 19 fifties, we built so many bridges. We built so many dams and connected so many houses with rural electrification, like the Tennessee Valley Authority. We built so much in this period right after the Great Recession up until the 19 sixties that there was a backlash. There was an enormous backlash, and part of that backlash was earned. The air was disgusting, and the water was disgusting. That's the world of the 19 forties, 19 fifties that was reacted to. And it was reacted to along several dimensions. One, you had this explosion of environmental legislation, the Clean Air Act, Clean Water Act, NEPA, Endangered Species Act, which made it harder to change the physical environment, which imposed strictures on the physical environment. At the same time, you had Ralph Nader, um, who created this revolution of adversarial legalism, where without simplifying it too much, he just made it cool. For liberals of a certain age to sue the government and sue businesses to stop things from happening. In fact, in the 2000, when Jim Lehrer asked Ralph Nader why he deserved to be pres…

AI assessment note: “explosion of environmental legislation, the Clean Air Act, Clean Water Act, NEPA”

Q And you also met your co-founders, um, uh, partially through Stanford, was it?

A Yeah, the GSB was, was, was an interesting, you know, Time for me. I mean, I, um, it, I, I applied on the recommendation of, uh, one of my mentors at, um, at eBay at the time, who was the CEO, John Donahoe at the time. And, um, he thought that it'd be great for my personal development. So that was really, that was kind of the thesis. Um, and, and so I didn't go into Stanford, in other words, trying to think about starting a company or, or even looking for co-founders. That was very fortuitous. Um, meeting Andy Stanley and Evan, Um, Evan turned out to be a roommate, and, and, and that was just purely from a social circumstance. Andy and Stanley were roommates at the undergrad. We happened to meet. We worked on different projects together. It wasn't intended to necessarily over time become a company or anything like this, but that's really how we got started.

AI assessment note: “Yeah, the GSB was, was, was an interesting, you know, Time for me.”