The Exchanges

Q And then what was that demo day experience like?

A Well, we felt great internally as a team where we answered our own questions that those three questions, because, you know, to us, the most important thing was that was this project worth continuing forgetting about, you know, you know, the circumstance of demo day or, or how awesome of an opportunity it was to get in front of investors. Can we prove to ourselves that this is worth continuing? The answer was yes. Demo day, however, was not successful. I don't know exactly where we stacked ranked That day, but we certainly weren't amongst the favorites. Um, and, and it was really tough. I mean, the company almost went out of business because, uh, you know, we didn't, we didn't raise a lot of money from Y Combinator, uh, with the initial grant, and so we were looking for seed financing, and we were, we were a couple weeks away from going to zero. I think at the time, people, um, just weren't sure if this would be a business. Yes, we had great, um, metrics. Yes, we had Positive evidence, you know, to each side of our marketplace in terms of thinking about whether or not they would participate. Um, but it was 10 weeks of data. And so at the end of the day, if you were an investor looking at DoorDash on demo day or anywhere near demo day, it would be a conviction bet. It'd be a conviction bet on the team on that there would be a potential market. You know, I think one of the things …

AI assessment note: “Demo day, however, was not successful. I don't know exactly where we stacked ranked”

Q Um, and so you have all these funny anecdotes. Um, you guys did such a great job of doing things that don't scale. Can you just share some of the anecdotes for how you got your first customers and how you got it off the ground?

A We actually had a bunch of crazy stories. So I remember we, we were in the winter, 2013 YC batch and we hadn't launched, uh, uh, yet. Uh, and so this is like, I remember doing office hours distinctly with PG, like, Two weeks into the batch, uh, and PG is like, oh, so like, how many customers do you have? And I remember we were like, well, we haven't launched, so we have no customers. Uh, and he was like, well, why don't you, why haven't you launched? And we're like, well, we don't have a parking lot to park people's cars. And he's like, well, that doesn't, like, that sounds like a thing you guys can figure out. Uh, like, you need to launch, like, tomorrow. Uh, and we really took it to heart. We were like, we need to launch, um, and, and looking back, like, that was the right advice, you know, for, for us, uh, because that's how you get feedback. And So we were like, okay, well, where would we park the cars? Uh, and we were like, well, there's a BART, which is the, the local Bay Area, uh, like, transit system. There's like a huge, like, basically subway parking lot at BART, which is like five minutes from the SFO airport. We can just park there. And so we literally launched, like, within a couple of days, and we, we basically would meet our customer at BART, get in the car, drop them off at the airport, and then come back, and then we told them we would park the car, but we woul…

AI assessment note: “we basically would meet our customer at BART, get in the car, drop them off”

Q And then you guys were old enough to rent a car, were you?

A And yeah, we were all, I think, 18, and so, you know, like, a lot of car rental agencies do not rent to, like, eighteen-year-old kids, for good reason, uh, and so we found this, uh, company called, like, Super Cheap Car Rental in San Jose or something, and this dude was like, yeah, I'll rent you guys, like, 3040, you know, Crappy base model Corollas. Uh, and so we're like, great. Uh, and so we actually just, me and my, one of my co-founders, we literally took the Caltrain down to San Jose and drove up a car one at a time, like, 15 times each, which was terrible, and then parked it at the same BART lot, and that's what we initially started renting out. That's how we got supply. Uh, and so, you know, you, you realize, like, doing such an operationally intensive business, like, like, flight car, when you have to, like, Do something scrappy. It's like at a, it's like almost at a comical extreme, you know, compared to like running a B to B SaaS company today.

AI assessment note: “we were all, I think, 18, and so, you know, like, a lot”

Q Yeah, we sat in group office hours together. We saw all these companies at the most nascent stages, and then, you know, I guess you're a bit of a masochist, uh, and you, I remember you saying, oh, you know what, I'm just, I got another company and be, I'm gonna go, I'm gonna do a do-over, um, maybe, yeah, what was your thinking on deciding to do another startup?

A Yeah, no, I mean, it wasn't a decision I feel like I had to make, you know, I knew I was, after Flycar, I knew I was gonna do it again, uh, and, you know, but this time, uh, we'll get to that, but I was, You know, we were gonna be much more thoughtful about it, and, and I had spent a lot of time, uh, working with Lou, and getting to know Lou, uh, when my co-founder, when we were at Airbnb, um, uh, he was my engineering counterpart at the company, and, and so, you know, we were working on, like, we had been noodling on ideas and stuff for, like, pretty much two years, really, until we, we ended up starting Zip, so just kind of on the side, and, and we, we both committed to both quitting, basically, and, and kind of going full-time, um, Literally March 31st, uh, 20, twenty-twenty. So that was right after our batch. Yeah. It's like a week.

AI assessment note: “it wasn't a decision I feel like I had to make... I knew I was gonna do it again”

Q And then, uh, did you know that you were gonna start nothing, or, you know, on day one, or, you know, do you have a moment in there where you're trying to figure, figure out what the next thing would be?

A Yeah, for the longest time, I thought I was gonna travel the world for half a year, at least, because at OnePlus, I was so busy. I was working at least six days a week, and just no real vacations for seven years. But after 10 days on that holiday, it got kind of boring. I felt like I was burning away my life, and I was still young, and I was staying in fancy hotels, and enjoying the, the drinks, and the dinners, and whatnot, but I felt a lot of anxiety that I wasn't contributing anything, so. I quickly wrapped up the, the, the trip after 10 days, uh, went back to Stockholm, where I grew up, and started reaching out to entrepreneurs there, because I had never raised money before. So I started reaching out to all the entrepreneurs I knew, and they kept introducing me to people. So basically, in like three weeks in Sweden, I got coached by all the Swedish entrepreneurs on how to raise money. Uh, it was a super welcoming community. I'm sure it's the same here. I haven't spent as much time here.

AI assessment note: “for the longest time, I thought I was gonna travel the world for half a year”

Q Next up, Crystal at Siteful. Along our journey, we've done surveys, interviews, and had many conversations with users. Once the MVP is live and we invite users to use and test our tool, what are the top methods you recommend for capturing and measuring feedback?

A I often think that startups try to scale this effort way too fast. Um, the things that I've seen work really well are you call the users, yourself, personally. You invite the users to your office. You host a dinner for your users. Um, you put your initial users in a WhatsApp group that you can all talk with together, but like things that absolutely don't scale. Um, another famous one was Airbnb where customer service was one of the co-founders Joe's cell phone number. Like, this is the perfect area where like you want there to be so much information flow from the customers to you that like you have to do things that don't scale. You shouldn't be thinking about, Oh, how do I get my, like, ideal intercom setup and my Zendesk setup? In the other days, you do not need that stuff.

AI assessment note: “call the users, yourself, personally. You invite the users to your office.”

Q So, I thought it was interesting out of the seven examples, something that was on the AI side. Is that, any ideas of examples there?

A Yeah, sure. So the, the question was, um, of the examples, I didn't mention any AI companies, and that's a great point. I probably should have included one. There's a really famous YC company that's an AI company called Cruise. You guys heard of Cruise? Cruise built self-driving cars, and they got acquired by GM for a billion dollars. Um, and yeah, Cruise is a great example of a hard tech company. Um, the original Cruise car, um, Was built in less than three months during YC. Kyle basically just, like, was in a garage building this car and writing code for, like, three months solid, and by the end of YC he had an MVP that he could use to drive on the highway to show that, like, basically he could build a self-driving car.

AI assessment note: “There's a really famous YC company that's an AI company called Cruise.”

Q What was that process like? How'd you find the right, right folks?

A So in one sense, we, we had the good timing that ARKit just got launched at that time. I mean, that's another funny story when it got launched. So there was a lot of positive enthusiasm for the space around that. That was one. The other one, we were lucky to really get connected with investors that really believe in this future and with us that it would take a long time. So we got, um, backed up, our lead, Investor for our seed run was, uh, Jeff Clavier from, um, Uncork Capital, who really was with us in really believing where all these pieces were coming about and, and really taking that bet on us. And besides that, it was also Founders Fund, also seeing that vision with us that, yes, this makes sense. So it's kind of finding those, those kind of people.

AI assessment note: “our lead, Investor for our seed run was, uh, Jeff Clavier from, um, Uncork Capital”

Q work on a small team of four, um, where everyone kind of knows everything, um, and everyone, everything is in everyone's heads versus a team of 30 or 50 or for some of us, you know, hundreds, um, that is very different. So, uh, you kind of constantly need to be assessing is this style of working or this methodology, um, is this right for this stage of the company?

A Yeah, I think the point of evolving the process is something we've been, ah, doing a lot. So back when we were at Escher, we were only with four engineers and me building the product. So back then, ah, it was very messy. We're just trying to move as fast as possible, really, so zero documentation. Everything was kind of whiteboards, and everyone kind of could, ah, handle everything in their heads and, and build it out. And in terms of, we didn't really do sprints, because I think that some engineers didn't quite like it as much. And there were just too many big tasks, and we just trusted different individuals to go and tackle kind of one area, and then we would integrate it. But now, things are very different. As we hired, I think the team has tripled already in the past eight months for, for, for the AR platform product. When you have a bigger team, you definitely need a lot more process to be able to be efficient and communicate, because you don't want to duplicate, and not everyone knows everything, and The system grows in a lot of complexity. So we went from zero documentation to a lot more. We went from, we used to have a little CI, but now is, um, very much of a CI that actually builds to all the architectural flavors to Linux, Mac OS, Android, XSX, Android Arm, iOS, et cetera, and actually runs all the tests in all of them, and actually catches a lot of the bugs, and we …

AI assessment note: “the point of evolving the process is something we've been, ah, doing a lot”

Q And so Jared, you have experience with this directly. Having hired quite a bit at Scribd before, and you were also part of the Work at a Startup, before it was known as Work at a Startup, conference in 2010, right?

A Yeah, so it's actually not a new idea, per se, for YC to try to help its companies hire. Hacker News, from the very beginning, when we first created it back in, I think, 2009, um, Part of the goal was to help companies hire, and so it's had a job section from day one, and when I was at my startup, I used the Hacker News job section a lot, and it was incredibly helpful for hiring. I think probably half the engineering team that we hired came through Hacker News, and that's not atypical for YC companies. So we have sort of a long history of this. Um, we've also long had an internal forum for YC founders, where they share Candidates that aren't right for them, but could be great for somebody else. And, um, that's also worked extremely well. And so it was the, the origins of this project were kind of like, we saw that this was working and that there was tremendous demand for it and everybody wanted us to do more. And so the work of the startup product is trying to build just a better, more efficient, More sophisticated version of the Hacker News jobs board and the internal form that we've had for a long time.

AI assessment note: “when I was at my startup, I used the Hacker News job section a lot”

Q And so what do you guys suspect are like the, the main issues right now that if addressed, MBA students would be more successful disproportionately?

A I, it's, it's interesting cause I'd, I'd love Jeff to, to talk through like the causes. This is what I see as kind of the, the red flags. So the first one is a lack of a technical co-founder, um, and the general willingness to outsource the tech part of a tech business. Um, the second one is commitment. It's, you know, can you get your entire team committed to working on this startup first and foremost? As opposed to, we will do this startup if we receive funding, if we get into YC, if a variety of other criteria are met. Um, and then I think the last one is, um, what we, what I like to call traction, which is just basically this idea of how long you've been working on this and what have you done? And that speaks to Jeff's point around MVP versus research. I think research isn't, is extremely important, but I think that being able to do research with a live product Is far more valuable than being able to do research with a general survey. And so I think like those three things like time and time and time again, and don't get me wrong, those are the central problems that most YC applicants have. But once again, like I would imagine that like within this population, these should not be the things taking these folks down.

AI assessment note: “The first one is a lack of a technical co-founder... second one is commitment”

Q to you and what encompasses the role can be quite impactful. And the only other advice I'd have is that oftentimes kind of goes back to the selling thing. People forget to sell the company in the job description. And so that should be at least 30%, maybe more of the job description is why is this a great place to work and, um, what would that involve for you?

A Often companies when they're describing themselves in their job description or even on their website, it's almost like they want to cast as wide of a net as possible. They want to be attractive to everyone. And as a result, it becomes across as very bland. It's like just the same set of values on every single job description. Um, I wish companies would actually be more opinionated at times, like, say what they will trade off. Like, hey, we value collaboration, which means you might not, like, we value that more than just autonomy, which is like one example. Um, what do you think? Is it, is it better to cast a wide net or, or should you try and be opinionated in how much?

AI assessment note: “I wish companies would actually be more opinionated at times, like, say what they will trade off.”

Q incredible amount of access to GPUs and data, but for context engineering, everyone here could do it. You have, you just need the API to something like Gemini, and then work on your own setup for your own retrieval, your own tool calls, and et cetera, et cetera. So, How does, what are some tips for everyone here? How does everyone get better at and become exceptional at context engineering?

A Yeah, I mean, I think, uh, a really good way to do it is to use these models and, and sort of harnesses and tools and so on to try to solve problems. And then some, sometimes you can actually see where the models are failing. And often you can actually make the model work better and succeed at that kind of problem by not just adjusting the model parameters, which is hard to do from the outside, but from, you know, creating better guidelines for the model, you know, writing skills for the model to know how to use different tools that would be incredibly useful for solving this particular class of problem. And I think as you do that, you end up on this kind of Improving, self-improving of the setup that you're trying to use to, to solve things. Uh, and you know, that, that's a really good way to get better at understanding what, what additional information the model would want in order to become more capable.

AI assessment note: “use these models and, and sort of harnesses and tools and so on to try to solve problems”

Q How do you find those? I mean, are those things effectively, uh, out of distribution from the training set, and what exactly is the problem shape that fits that?

A Yeah, I mean, I think, uh, sometimes it's Uh, a product that you build that might have access to particular kind of data that the underlying model might not, the general model. So it might be you're building something to help users organize all their own personal information, and the model won't necessarily have access to that. And so there you can have a big advantage because all of a sudden your model has visibility or your product has visibility into important data. Um, It could be some incredibly hard problem where if you get the right training data and you can train a more specific model than a general purpose one, you can actually do that in a very affordable way. Maybe it doesn't take that much compute to train a niche model for this particular problem, but you can get something that's highly accurate. That can sometimes be a really good building block for, for solving a important problem that is maybe not Handled very well by the general model.

AI assessment note: “a product that you build that might have access to particular kind of data”

Q How are some ways you implemented this particular workflow for your agents internally?

A Yeah, I mean, we have, ah, you know, harnesses, and then we have a whole set of skills, ah, particularly in the internal Google development environment, we have skills so that the agents can know how to use lots of our internal tooling for coding, or for code reviews, or for, you know, measuring performance, or, you know, fetching log files, and, um, those are just skills that you can add to make the base model more capable, even though it hasn't necessarily been trained on exactly the way that You know, Google internal, ah, engineers would fetch log files from our, you know, proprietary system with the right kind of skill, ah, definition, you can actually get it to work. And that, that improves the usefulness of the agents.

AI assessment note: “we have harnesses, and then we have a whole set of skills”

Q Now let's assume now every founder gets good at running hundreds of agents at the same time, and all the code is written for them by the agents. What becomes the scare skill?

A Yeah, I mean, I think it's really having incredibly good taste in what you ask your agents to work on, right? That is the, the crux of, you know, from my background, ah, a research problem. You know, a researcher can have all the tools and all the techniques, but often most of the battle is what problem are you gonna spend your time on? And if you pick the problem well and you succeed in, in, in solving it, That's way better than if you, you know, ah, delightfully execute a research investigation into a rather boring problem. And so that high level wisdom of what to work on, I think is incredibly important. I think models are not necessarily going to be that good at it. So you're going to have people steering Uh, a lot of AI assisted computation in order to accomplish great things and more quickly. Um, but that essence of, of what it is you want your models to do is the, the key thing you should focus on.

AI assessment note: “having incredibly good taste in what you ask your agents to work on”

Q a bit about more about that second path of working with people that you really like in a small team. You've been able to be an incredible mentor and manager to many, many engineers, and you've been able to build huge systems. Um, what are some, some of the lessons for everyone here on how to get the most and how to work with smart people or find smart people?

A Yeah. I mean, You always want to find people who have really good skills in some, some area that's needed in, you know, a team you're trying to form, whether that's inside a company or starting a company. Um, but you also want to find people that are People you delight being around, right? Because you're going to spend a lot of time around people working on really hard problems, and you want people who are low ego, that are team players, that, you know, have complementary skills to your own, perhaps. Um, I always find working in a small team where people know things that I don't know, and where maybe I have some skills that other people don't have as much of, You know, it's super fun because you're collectively building something or working on something that none of you could maybe do individually, but in the process of working on that, you actually gain a lot of new knowledge and new skills, ah, for yourself, and so do they. And you, you kind of want to view your engineering or research career as you have an amazing tool belt of techniques, and you always want to be adding new tools. To that tool belt, because you never know when you might come across a problem where you need these four specialized tools rather than these three. And adding more tools makes it more likely that the problems you, you encounter in the future will be solvable by you.

AI assessment note: “You always want to find people who have really good skills in some, some area”

Q which is kind of wild, but, uh, not only that, now you have, uh, a ideal, a personal AI that's gonna tell you, you know, hey, these are some ways to do it. Um, do you think that would have helped you, like, accelerate even faster? Like, you know, talk to, what do you think it's like to start a company today with, with AI in the age of AI?

A Yeah. I mean, I really think, I think we're at this, like, amazing moment in the world where the bottleneck is not the progress of the AI models. The bottleneck is diffusing that through the rest of the world and, and helping the world adapt to this amazing technology that already exists. Like, I think if the models didn't improve at all from today, there would still be, like, decades and decades of, like, total upheaval and change in the economy and how the world operates and, and everything around us. And, Um, so I think it's, as a result, it's like one of the most incredible, it's probably a, like, once in a civilization opportunity to be a dreamer, and to have a vision, and to have ambition, and to impose a view of how the future world should look by building something amazing. Um, you know, one of the things that we were, we were chatting about, um, uh, you know, backstage is, you know, when When I started scale, or, you know, 10 years ago, if you start a company, you had to be, um, you know, it was like David versus Goliath, and you had to be clever, and you had to find, like, an angle into the market, and you had to sort of, like, you know, figure out, um, a way to compete even though you had much fewer resources. And now I actually think with the power of agents, um, and AI broadly speaking, it's much closer to Goliath versus Goliath. Like, I think, but maybe the startu…

AI assessment note: “when I started scale... now I actually think with the power of agents”

Q what would you say to people in this audience right now? Like, this is sort of a real question that people are sort of facing. Like, should they become more word cell and less shape rotator? Like, what, you know, what's the move? And, you know, has that changed, um, The kind of people you're looking to hire and, you know, how you manage your teams right now at Meta.

A I think systematic and rigorous thinking are still incredibly important because, you know, the abstraction layer, I mean, I didn't used to believe that this is how this is going to play out, but it really has, like, the abstraction layer just keeps changing. So, you know, when I started a company back in my day, we wrote code. Um, and, and now, you know, I'm sure nobody here writes code anymore. That's ridiculous. But, um, but now it's about how do you orchestrate the agents together? And then it's like, How do you develop these organizations of agents? Like, how do you get, like, a million agents to work together well? And then it'll be, how do you get, like, a trillion agents to work together well? Like, I think that there's going to be this continued, um, uh, need to figure out how you structure, uh, workflows at the abstraction layer that we're going to be operating at. And that form of, like, rigorous systematic thinking, I mean, traditionally the way this would work, like, in my era of starting companies is you would start by writing code, And then you would have organizations of humans, and you'd figure out how you, how to organize those humans. Um, and that required systems thinking. And now maybe it's like much more, much closer to first you, you orchestrate the agent, then you figure out how to orchestrate like these armies of agents. But, um, but I think systems thin…

AI assessment note: “I think it's definitely a mistake to go all in on word cell.”

Q I believe this thing, but I haven't found my people yet. Like, is that delusion? You know, what, is that actually maybe even a gate that you would propose people have? It's like, well, you need to find, you know, five people who might believe that, or even one, like a co-founder, you know, people, maybe that's one of the explanations for why we like co-founders at YC so much.

A Yeah, I think if you can't find anybody else that shares your belief, you should pay attention to that. Um, and it's also very hard to do a startup on your own and very lonely, but I don't think you need to find a lot of people, and in fact, if everybody believes it, that's also like a bad, a bad sign. Um, the question, I don't know if this is still what feels like the limiter on more startups, but five or 10 years ago what felt like Was limiting the number of good startups YC could fund is figuring out how to solve the co-founder matching problem. Like, a lot of really talented people, and they just couldn't find their tribe, they couldn't find their people. Um, you know, it used to be very heretical that YC told people they had to move to San Francisco. And looking back, it was clearly right. Like, I think, I don't know if this is still gonna be true for the next 10 years. I suspect it will be, but the best thing you could do if you wanted to, like, find your people to do a startup with was to move to the Bay Area. You just, like, magic happened. You got a lot of, like, lucky chances and collisions. Um, I still think that's probably good advice, although I feel like I have less of an intuition for it now. Um, I also think that I can say this. Gary probably can't. I, I think the premium on doing YC is bigger now than it's like ever been before. Uh, the distance between like YC…

AI assessment note: “if you can't find anybody else that shares your belief, you should pay attention”

Q Let's talk a bit about how then you build this new prompt when there's a new model release, like for everyone in the room, everyone will want to try Opus five and they're going to press delete on their system prompt. How do they go about building, rebuilding the system prompt? How do you set up your environment?

A So you do, you do it kind of piece by piece. So the first step is you delete. The next step is you use it. And you don't want to guess what's the instruction that the model needs because you might not predict it correctly. The thing that you want to do is you want to run it. And if it's like a custom agentic product that you're building, you want to kind of run the products. Uh, you want to see where it fails with the model. You want to see what it does well. If you're using quad code, you want to see where it does well with your code base, or maybe where it stumbles over, you know, the architecture or stumbles over something else. And only when you see it repeatedly stumble on the same thing, That's when you add it back. But you don't want to do it too early. Because remember, like, the model is going to read this instruction every single time you use it. So you really want to make sure that the model needs this instruction. I, I think this is sort of the crazy thing about building on models. It's just so different than all the engineering that I've ever done. Like, in the past, when you built on systems, you built these, like, big, beautiful systems, and you really think about the system design up front. You have, like, a big suite of unit tests. You think about everything, and you know, like a re-architecture is a big project, and sometimes it takes months. I've worked on re…

AI assessment note: “The first step is you delete. The next step is you use it.”

Q How do people get better at that and effectively how do people get better at Prompt engineering. Do people still need to do a lot of prompt engineering, or is that changing as well? Tell us about where this is going.

A Yeah, I remember, like, a year ago, one of the most popular job openings was prompt engineer, um, and then it kind of changed, and then I think it became, like, context engineer. So there's these kind of waves of it. I think these will kind of, like, come and go. I, I think the skill nowadays is less about prompt engineering and more about figuring out how do you give Cloud a hard task that seems a little bit too hard? And then how do you make it possible for Cloud to verify its work along the way? And the verification, I think, is probably the single most important thing that people do not get right, largely. Um, one example of this is people were, uh, you know, we have this desktop app for Cloud. And it's built using Electron. We've made it quite fast, so now it's like a pretty awesome experience. Six months ago, it was like sluggish, and it wasn't very reliable. Now it's pretty awesome, and, you know, it's the thing that most of the team uses. As an experiment, though, I wanted to see, like, what would it feel like if it was native? And so what I did is I, I started a quad tag session, and quad tag is just, you know, it's a, it's a new product we have. It's just quad running in Slack. My first question was, hey, tag, do you have access to a macOS runner on GitHub? And, uh, it said no, and then I, I hooked up a runner, so it was able to start a Mac virtual machine, uh, using,…

AI assessment note: “the skill nowadays is less about prompt engineering and more about figuring out how”

Q And that then created this amazing product just that keeps going. So let's talk about, um, What are some areas and how should future founders here think about on hobbling Claude and fixing this product overhang?

A So there's a couple of things that I will think about. One is you should give the model slightly harder tasks than what you think you can do. I think a really common mistake that I see is people are using quad code, they're using quad, and they, they just give it, like, way overly specific instructions. They're like, I want you to do this, but I want you to do it in this way, this way, this way. You must do, like, one, then two, then three, then four. And for modern models, that's actually really not the way to do it. You want to go a little bit higher level. You want to describe the task. You want to describe the guardrails. You want to describe, like, the exit criteria, and then just go with the model cook and come back in a little bit. And I think it'll, it'll surprise you. Like, and again, like, this is just not something that would have worked six months ago, but it does work today.

AI assessment note: “One is you should give the model slightly harder tasks than what you think”

Q we, I just really resonate with your story. I think that everyone here, I mean, would love the wisdom of, you know, your journey coming here. I mean, what should a young person learn now, given all the things that you're seeing, all the algorithms that are going to take hold in society? Um, what should a young person learn now that will still matter based on what you're seeing?

A Well, some of the things that I saw today and some of the starters I met today was really, really quite, quite, um, encouraging. And, and, and the thing that, that, um, the big takeaway is, of course, the simple stuff is going to get automated away. And when I say simple stuff, I mean software, you know, coding. And the idea that you would, you would do a, you would solve a problem by sitting in front of a computer and you're, you're, you're actually writing, you know, writing code, that concept is obviously going to get automated away. Um, you know, in my generation, when I was, when I was growing up, we had to do long division. I mean, for God's sakes, who has to learn long division, you know? And so that got coded away. That got automated away. And so I think the simple stuff is going to get automated away, but the hard problems, the hard sciences, um, physics, chemistry, biology, uh, you know, computer science, uh, computer engineering, systems thinking, Uh, you know, all, and, and, and particularly the domains that are intersecting, uh, those hard problems will never go away. And so AI is just an incredible tool that helps us become even more ambitious, even more, um, impatient about solving these extraordinarily large and incredibly hard problems, uh, than before. And so, you know, if you, if you look at my generation, when I first graduated, A chip designer would design …

AI assessment note: “the hard problems, the hard sciences, um, physics, chemistry, biology”

Q Quan, because this task is less familiar than laundry folding, do you want to explain what the robot is doing here and what Ultra is, like, doing as a company?

A Ultra is a company that wanna makes it really easy to adapt Robot to, you know, new tasks, um, and right now they're focusing on logistics space, which is really important because, you know, there's lots of labor shortage in logistics, and the task that we focus on together here is, you know, if you order an item from Amazon, you sometimes get this soft pouch that item gets shipped from, and the task here is you have a tray of these items here, and the robot is supposed to pick one of them at the time, and place it inside this pouch. The machine will then close it, And then pick up the pouch and put it, um, on the left here to be ready for shipping. Now, this heart is hard because there are many different types of object that can be in this tray, and the opening here is actually very narrow, so you see this interesting example of the robot kind of nudging the item to go into the pouch, and that's, that's really hard. Like, that requires a very good understanding of the scene and, like, very Precise motion to nudge the object into the pouch. Um, the other thing that's hard about this task is the level of autonomy that's required. Like this is running for an entire day. There is still human intervention. I want to say in, um, this like full day operation. Um, but the level of intervention is actually quite minimal.

AI assessment note: “Ultra is a company that wanna makes it really easy to adapt Robot to, you know, new tasks”

Q Could you help us understand where we are now in terms of, like, what's working and how well it's working? Like, we're not quite at the ChatGBT moment yet. Like, where are we? And I think you brought some videos that you were going to show us to, like, help everybody visualize what the current state of the art actually looks like.

A I think where we are is I think if you have a test where it's okay for the robot to make a mistake, um, and it's possible for you to set up a mixed autonomy system where You have a person that takes over when the robot make a mistake and provide corrections. It is possible to get to a level of performance where it starts to make sense to think about scaling robot deployment. And the example that I specifically want to highlight here is this blog post that we did with Weave and Ultra. And you know, it's great that these are both YC company. I want to provide a little bit of context here first. The context is that Pi Is a primarily research organization. We want to focus on building the best model. Um, but we also want to not be tunnel vision. We want to make sure that the model that we built actually going to be useful and actually perform tasks that people in society cares about. And one of the really good way for us to do so is to partner really closely with company that want to get robot out there today. And the way that these relationships work is that we treat each other like we're on the same team. Very free flow of information, um, and we design a system that try to get the best possible performance for the tasks that this company care about. So, let me talk about we first. What you're seeing in this video is a system that we built together, folding really diverse item of…

AI assessment note: “What you're seeing in this video is a system that we built together”

Q thousand companies like Ultra going after, you know, every like menial job in the economy and like getting a deep understanding of the customer, building a robot that can solve that problem, doing that like mixed human machine deployment until it like can run fully autonomously and Building a company in, in every sector. Is that, is, is that the future that you see people building on top of Pi?

A It's funny that you mentioned Cambrian explosion, because when we wrote this blog post, there was that term that was very kind of like hotly debated. We are, I think, academics at Hurt, and we want to be kind of very major when we communicate. But, you know, myself personally, I believe there's going to be a Cambrian explosion of, um, robotic company across the entire world and across many, many different vertical. Um, just because it's just so much cheaper to build, and it doesn't require, um, you know, someone with 20 years of experience in robotics to start anymore. You know, it requires someone that is really scrappy, that can move really quickly, um, can do the system integration, um, can understand customer what they want, to start the deployment,

AI assessment note: “I believe there's going to be a Cambrian explosion of, um, robotic company”

Q Do you have an example of something like that with one of your customers where the system was able to do something that just like no human team of fraud analysts could ever have been able to do?

A Yeah, one customer specifically, uh, and I think that fraud pattern came to be during the elections. Um, we had one customer that is processing a lot of content, and they're also, they're a fortune 500, they're hosting large communities, and they're also fairly politically exposed. And throughout the elections, because our AI agents had access to the context of entities in relation to other entities, so How does this user fit into all of the other users that we're looking at? We were able to detect really complex fraud rings of especially state sponsored actors that were pushing one narrative over, and I don't think this would have been possible if you had one classifier in isolation that was looking at one piece of content after the other. But because AI agents are able to directly query our data stores, they're able to materialize features on the fly, and they're also able to Use one step to reason over what should be the next step and the next tool call that they make. We were able to detect much more sophisticated fraud rings than you would have been able to do before.

AI assessment note: “Yeah, one customer specifically, uh, and I think that fraud pattern came to be”

Q like, a million questions, honestly. I mean, one of the things that I'm super curious about is, like, well, what is the qualia of the person who has Prima, and what is, you know, I'd be curious, like, with the bio-hybrid approach, like, what does it feel like? And, you know, is it like having a second screen? Like, you know, is there an input or output? I'm very curious.

A Yeah, so for Prima, actually, on the topic of plasticity, In the time that the patients are blind, the brain, the brain wants to see. Like again, you, the thing you experience is this world model constructed by the brain, and that is this, is this generative model that is conjuring your reality. And so when it's not getting input from the, from the optic nerve, it is still trying to see things. So it kind of turns up the noise. And so, um, blind patients often report like hallucinations and these like internally generated percepts. When you first turn on the implant in these patients, like you hit it with the laser, Um, they'll, they'll say, oh, I see a flash, but then you can do a thing where you'll, you'll turn on the laser, they'll see a flash, and you'll play a tone, and you do this a couple times, and then you, like, don't turn on the laser, but you play the tone, and they're like, I see the flash. And so for the first couple hours of rehab, they kind of just have to, like, learn to, like, dissociate the real percepts from the phantom percepts, because the brain is, like, so, it is, like, so turned up the gain, like, turned up Turn down the noise for that. Um, just like getting, learning how to discriminate real information coming in from the optic nerve takes a little bit of rehab. The qualia of Prima is, is normal sight. Um, it's black and white. It's only a, it's a smal…

AI assessment note: “The qualia of Prima is, is normal sight. Um, it's black and white.”

Q I mean, like anything, you sort of bootstrap with the thing that works, which I think, you know, what, what you have is a clear breakthrough as it is. And then if you look at like the PC revolution, for instance, it's like, could you believe that all of this that we have today started with like a little blue box, like in Altair?

A It still takes some suspension of disbelief because I think biotech has just been so incremental. Like it's been so, like there's, there's been big advances, but at the same time, these time constants historically, I mean, you could easily spend 10 years on something that feels very incremental. And I think that One of the things that's so exciting about what's happening now is that no longer really feels so incremental to me. To me, it feels like we're firmly in like the takeoff era now, like something new has happened on earth. But I think it's also important to remember that this didn't start in like, 2019 or 1999. This started in the late 1800 with the industrial revolution. Just a few years before the industrial revolution really kicked off, I mean, life was more or less unchanged in a fundamental sense for several thousand years. And they, Didn't really even have like a concept of progress in many ways. And I don't think there's any way they could have imagined like the way that their life would have changed over the course of the like first 1015 years of the steam engine. And that is how I feel like looking at the next 15 years right now.

AI assessment note: “It still takes some suspension of disbelief because I think biotech has just been so incremental.”