YC Demo Day, Spoon = Bent (Again), iPhone Duo Reactions | Garry Tan, Om Buddhdev, Karen Serfaty, Antonio Li, Edward Ge, Alfredo Gonzalez

30:28You get your early adopters there. And I noticed that this was happening a little bit in the gundo where you would see one guy's running a machine shop. One guy's running a cloud seeding. One guy's running a drone company. and they'll run across the street and say, I need a new drill or hammer or whatever they need. But I'm wondering in the hardware sense, are there YC companies that are actually able to get products in the hardware space that bring value to other startups or is it more of like camaraderie around we're in the hardware world, we'll communicate and share notes, but we're not actually selling to each other on day one?

31:03I mean, you don't really have to necessarily even sell to each other, though that's happening. Like the space ecosystem's coming together, like StarCloud is a great buyer of lots of other space tech. That's been happening. I mean, we got defense tech coming together and people trading notes. Like we're putting together a Krata so that we can actually help all YC defense companies get in theater with real Department of War departments right off the bat. There's just a lot that we can do. Next week, we actually have a DC defense tech conference in DC. We're bringing in a bunch of our top defense companies and they're meeting the policymakers directly.

31:41They're meeting the people from the Pentagon directly in their own backwoods. And that's actually stuff that YC can do now that, you know, it just increases the chance and likelihood of people actually succeeding. I mean, defense tech is the wildest thing that continues to happen in this administration. is that I think, you know, huge credit to the people in the current administration because they're realizing like, hey, we got a drone war happening. We have an AI war and America needs to be, you know, can't just pay, you know, cost plus to these defense, you know, the sort of like industrial complex that exists today.

32:19Like we actually need, and they're willing to give seven figure contracts to teams of two or four people who suddenly become teams of 20 or 30 people plus hundreds of agents. Yeah, it's been remarkable how many opportunities that have been there, even in the post-Palantir and Anderil world. There was a time when I think people thought that there might be more of a winner-take-all dynamic. It might be more like social, but there's been so many teams that have had impacts in various ways. It's been exciting. So 25 % of the batch is hardware up 40x. What about the other 75 %? What are the trends that you're seeing on the software side, dev tools?

33:01I think part of why some people are driven into hardware is they're looking at how easy it is to make software now and how fast it is and thinking, well, I don't want to go, if I can build this in two days, someone else can build it in two days. So maybe I want a hardware component. But at the same time, we're still seeing a bunch of early stage software companies rip. What trends are you seeing across the board with agents and infrastructure and all those things? Yeah, this is, I mean, honestly, this is the golden age of software as well. And it's interesting because, I mean, my prediction would be you look at the public companies of today, any company that's got thousands of engineers who are doing it the old way.

33:41Like they're going to be replaced by a YC startup or a company like that, that has maybe 20 people working there. And those 20 people each will do the work of like 100 engineers themselves, right? So you got like a 2 ,000 person workforce and you're just like blowing them out of the water. And so that's going to be the biggest shift. Like you're going to have AI native startup companies that are just brand new that have started in the last two or three years. They don't have any of the bad habits that were developed over the ZERP era. And they're just going to come in and out-execute every single SaaS company that doesn't do it that way.

34:18And it's going to be pretty epic. It's going to be a turnover of epic proportions in the world of software. And that's sort of the rest of the story. Are you seeing pitches for NeoClouds, people that want to go through YC? Yeah, I mean, Wafers is killing it. Shadeform is killing it. I mean, you know, the amount of inference in the world feels like it's going to go up like 10 ,000 to 100 ,000 X. Like people are just not really prepared for, you know, it's actually kind of the funniest thing. Like you can work in SAS where, you know, things are going to grow 10 X. And then or if you look at inference, you look at data centers, you look at semiconductors, you look at GPU and even CPU demand.

34:57It's going to go up like a thousand X. Right. So that's even it's even bigger an opportunity. community. And I think people are just not, it's not priced in. Personally, I think it's not priced in at all. Are you still using the term little tech? I always like that. Oh, definitely. I mean, yeah. Where does little tech stand on the AI slowdown debate, regulatory capture risk? When I look at some of the companies that we're going to have on the show, I feel like they might not be affected positively or negatively by the discussions in Washington, but how are you seeing it broadly? Because typically regulatory capture can, you know, harden the position of incumbents and that's bad for little tech, but how have you been dealing with the current conversation?

35:47Yeah. I mean, I, you know, I think we should be talking less about this Jacob Coxon guy. We need to be talking a lot more about what is actually happening with hugging face, right? Like, are agent swarms going to come and take over infrastructure en masse? And then what are we actually doing about that? Like, I don't want to hear about some guy who worked at Anthropoc for two months. And I don't really care that there's a, you know, there's a coordinated effort to try to influence politicians to sort of do a knee-jerk response. And, like, the thing is, like, we shouldn't be, that's a smokescreen, you shouldn't be paying attention to that.

36:18We need to be paying attention to what is the actual thing that we can do to prevent, you know, agent swarms taking over entire data centers. Right. Like what's our what's our shutdown strategy? Right. Like how do we ensure provenance? Where is this agent actually located? Like, you know, what can what software can we build and what cybersecurity defenses can we build today? And like, you know, that that's the level of discourse that I think we need. And we just don't have that. Right. People are just saying, you know, well, discourse is discourse is one thing. Are there or is there a meaningful number of YC companies now that are actually focused on that set of problems?

36:52because one way to think about AI safety is to go on podcasts and talk about it and write essays and write blog posts. And another way is to actually build cybersecurity products, specifically with, like, you know, agent swarms in mind. And, you know, thinking of Ilya's post about neocloud security, it's like, well, I hope there's five companies that saw that post. And we're like, you know, I'm sure that Palo Alto Networks and CrowdStrike and all these companies are thinking about that opportunity as well. But it just feels like anywhere there's fear, there's also an opportunity for commercial solutions to the problem that won't fix everything.

37:30Yeah, we got that too, right? We got a ton of company. Like Silmaril is able to do 100 milliseconds in request loop, prevent prompt injection. Sure. Right? So you got an agent swarm. It wants to take over your AI systems. And it's in production now. It basically prevents prompt injection, and it can do it at a level that it sort of doesn't matter how much smarter the underlying AI becomes and how much more powerful it is at cybersecurity. If you have a kill switch on an entire request because, hey, that's a prompt injection, that's sort of the stuff that we need today. So I don't really care about science fiction, man.

38:08I saw Terminator 2 also. We're not here to talk about that. We need to actually talk about what's really happening with the servers, what's actually happening with the agent swarms, and how do we actually prevent that. And I think that Little Tech has a role in actually providing the software, the services, the hardware, the things that actually solve these things. And then when it comes to regulation, it's like, let's actually pass regulation that prevents the thing we actually care about instead of whatever a socialist says in the New York Times. I don't care about that. what uh what are you telling yc startups not to do right now i'm thinking there's opportunities and like selling data right now if you just wanted to generate as much revenue as possible before demo day maybe brokering data could be a way to do that but maybe it's not the best way to build you know a super enduring business but maybe that's one example but but what i mean i think like mercore are very very uh you know they're they're body shops at some level and then we're seeing this new wave of data companies like Afterquery or DataCurve that basically beat them head to head.

39:14And they don't do it by being body shops. Like you don't need thousands and thousands of people. You need, you know, dozens of people. I mean, it's sort of the same deal. Like you look at, you know, a workday, it's full of thousands of engineers who are not, you know, using agentic systems. And like, I would argue that like a lot of those, you know, first wave data companies don't. They're basically kind of like glorified labor marketplaces. But the new wave of data, we have more than 10 companies making more than$10 million a year in data. And then the top two are making hundreds of millions on the way to billions.

39:51And so there's a new wave. And those new wave companies do not have 1 ,000 employees and won't ever have thousands of employees. They will have 20 or 50, right? And how are they doing it? They're actually doing it in a proper agentic way. And so that's like the same thing in software, same thing in hardware, same thing in data. There's a new way to build these things. And either you're building it that way or you're doing it the old way. People can still raise money. They can still go to Sand Hill and get that crazy growth around. And people are just going to do it the old way still. But the new way is going to come, and it's going to wipe out the old way.

40:28Makes sense. It's gaming. A lot of people have seen Astro demos from the last week, people just building GTA 6 because they're impatient. They get something that is who knows how many hours of play it'll actually get, but it's an open world game that has multiplayer components, things like that. Do you think that we'll see a gaming boom in the next batch of just small teams that are like, Like, hey, I can kind of build anything that I want, and I can build the kind of thing that would have taken hundreds of people and years historically, and I can build it in three months. Yeah, I just funded a company called Summer Engine, for instance.

41:07It's growing really, really fast. It's just like Roblox, actually. And I think that that's the way to think about it, right? Like CodeGen is opening up all this stuff. And then actually, like, shouldn't people, you know, a lot of the value is actually going to be in your guild, your people network, like getting your friends to come on and play, knowing when they're playing. Like, you know, designing all of that is actually the ultimate value. And then I think it's the same thing all over again. Like, if you look at, you know, a company like Roblox, like, I don't think that they're building around, like, this idea that you could vibe code anything.

41:42Like the future is already here. People just act like it's not. Right. And, you know, anytime that happens, just like infinity, just opportunity. And, you know, the startups are finding it. They're going to go for it. And, you know, that's the future. I'm psyched for that. Yeah, I would almost disagree. I feel like Roblox is going to come out with an amazing tool for. Yeah, no, but even seeing like Astra being like, hey, we can make this even better. but there's still plenty of opportunities for early stage companies to just adopt these tools even more aggressively and come up with entirely new.

42:16Yeah, Roblox is the one that's like on the edge of the network effect. But we can debate it more later. Thanks so much for hopping on, Gary. Yeah, great to see you, Gary. Congratulations on another fantastic job. Excited to meet a handful of these founders and yeah, keep it up. Yeah. Always great to see you guys. Have a good one. Cheers. Goodbye. Let me tell you about Cisco. critical infrastructure for the AI era, unlocks seamless real-time experiences and new value with Cisco. What do you think of the iPhone Duo animation? Marques Brownlee shared it on X. Do you love this? What do you think? So when you open it, it fades as if it's looking through the glass.

43:00And when it closes, it shows you what it's about to show you. It could just show you the actual outside, but instead it animates so that it looks like you're looking through glass and it appears three-dimensional. What do you think? It's so incredible. I mean, it really pushed personal computing into an entirely new dimension. I think it's amazing. No, it's beautiful. It's beautiful. Well done. Well done to the team. And I think this is what makes Apple so special is that the folding phones exist. Somehow no one thought of this. how did they create an environment, maintain an environment where you can still go and experiment and think of a little touch like that that makes something that's been around for a while, Gary has one, a folding phone, still make it special.

43:49And now, you know, you got everyone talking about this. So good stuff. Anyway, our next guest is in the waiting room. Let's bring in Om from Olam Labs, the co-founder and CEO. How are you doing, Om? Hey, you got my name right. Congrats on the pronunciation. How do people mispronounce it? No way. People say Om. They say Om a lot. I used to watch your YouTube videos, all the company documentaries and founder stuff a couple of years ago. It was great. Thank you, man. Well, thanks for hopping on the show. He was eight years old that time, John. Stop making me feel old. Introduce yourself at the company and then we'll chat about it.

44:29Of course, yeah. So we do our environments and training data. We work with labs to evaluate behavioral things in models. What models lie more? Who's more deceptive? Who collaborates better? All those kinds of things. So typical RL environments are a lot more about how well can you complete this coding task? How well can you solve this math problem? That kind of thing. We're testing a lot more of the real... We're aiming to create real-world environments. Companies talking in Slack to each other. You have a PM. You have a CEO. You're playing diplomacy and negotiating. And pretty much grading all of those things that are more qualitative.

45:00and more difficult to verify, but also a lot more valuable if we can start to measure it well. Yeah. So what's the first step in building the company? Like when you started this, was it like, I'm going to go build the environment and then I'm going to go to labs and say, hey, I have this thing. Do you want to buy it? Or are you just going to them first and saying like, look, I'll build anything. I'm going to build. Like what is the flow? Because the demand is so clear and high. I imagine that it's working very well. But what direction do things flow right now? It's a lot more of the former of what you said.

45:33So the latter, going to a lab and saying, I'll build anything you want, that only works if you have relationships and a lot of trust with them or you used to work there. When you're selling data, the best mental model is that you are selling a capability to a lab. You are selling them something to get their models better at. So if you have a benchmark or you have a data set on, here's how you can get better at coding, here's how you can get better at, I don't know, like collaboration, like just anything, then you can sell it to them. That's why all these companies, like any data companies, even posting their evals on Twitter and trying to get to go viral.

46:03Because that's proof to the labs that, hey, I can make great data and package it well. So that's why. So the answer is the former. Everyone's trying to show that we can make great data and then use that to sell to labs. So do you have evals? Do you have benchmarks? How much time do you spend thinking about benchmarks or building new benchmarks? Yeah, so I mean, there's two things. So making great benchmarks and publishing them is great for just the broader industry and like especially in our case where like we're measuring a lot a lot of things that relate to like safety and all like the opening eye hugging face stuff going on the things about like how agents actually behave in these like multi-agent environments that's more our domain so we want to we are publishing stuff we've released evals on how models all compete relative to each other how much they lie against each other how well they collaborate we have a bunch more coming pretty much but yeah it's like our we've had some tailwinds in recent times on like the kinds of things that people are looking for is pretty much exactly what we bet our entire company on four months ago.

47:00So it is working out well in that regard. What, if any, are some historical comps for your business? Like, what was your business 20 years ago? Oh, God. Like, I mean, it didn't exist 20 years ago. Well, no, no, no, no, no. That's what I'm asking. I'm trying to understand, like, Like, was there companies that were helping Amazon in the way that you were helping, like the labs, for example, where it was like this helping them deliver core product quality increases that were scaling this way? And part of why I'm curious is because I'm trying to understand what your business looks like in five years.

47:50Yeah, five years, 10 years, and how you think about that? I think a comp that I would give is all of the internet infrastructure companies, everyone who is doing fiber and internet and CPUs and RAM, I would say that is the closest comparable thing to data companies nowadays. You're building the foundations. All the data company really is, like Mercor, all those guys, they just go out and capture stuff from the real world and then give it to a lab. which isn't, you know, like if you're buying CPUs, you're just constructing stuff in the real world and giving it to like a large cloud to then like service their customers.

48:25So I think that would be the closest like comparable thing that I can think of. There has been a question about the viability of these businesses, but I will say is that I think the viability of these businesses would be the exact same as the viability of like the, let's say NVIDIA. I think for AGI, for ASI to come, there are only two bottlenecks right now. We have no bottleneck with architecture or compute or agents or anything. It's data and compute. Compute is the biggest bottleneck, but now we're going to have 10x more compute coming online in the next two years. So then compute even more becomes data.

48:58If you're scaling law pill, and if you're pill in all the ways that these AI labs are, then data is the actual bottleneck. The new DeepSeq model that came out last night, the reason that's a crazy dominating model is because they spent all the time on amazing data. That's it. Like most of the time was on curating amazing data. That's pretty much all they did. But yeah.

49:23Are you compute constrained? Yes. How so? What is the nature of it? So I think the best data companies are research companies. So one thing that we do is like when we create like post-training tasks that we want to sell to a lab, we test them ourselves on GPUs by like post-training like open source models. Okay. So for that to source GPUs and do that as we grow and as we scale and if we want to do even more and more, I'd say at our stage, it's very easy to source the amount of GPUs that we need. I think especially, I see a lot of other companies I know in this space where it's doing research broadly, even Neolabs, they're all compute constrained.

49:58I think that is where the labs have their biggest advantage. They have the highest margin on every megawatt that they make. So they will continue to gobble up more and more compute faster than anyone else. So I think that if I was to name a risk, that would be the main risk for our business and other businesses like it. Andrew Ross has a comp, LIDAR Flyover Company Selling to Google Earth. Oh, yeah. That's a good comp. The only thing is Google Earth always was sort of... I guess it is constantly the intelligence of Google Earth is always scaling, right? Because you can always get higher resolution.

50:34I was also thinking about like, uh, like this is a weird example, but like movie production studios that sell content to Netflix. Like you want the Netflix library to keep growing. So you constantly have to be making new capabilities, new content for the network. Uh, I mean, there's also just like in the Amazon example, it's like the sellers, like, like there's a company that like makes the knives that are sold on Amazon and makes the appliances and stuff. So, uh, yeah, all, all a huge ecosystem. fascinating stuff. Hugging face, very scary. Are you optimistic? Are you a techno optimist? Do you think it's solvable or are you just spinning your wheels before the singularity?

51:15Well, I am optimistic. I've been optimistic my entire life, but I will say it is getting scarier and scarier. We were clapping for optimism. Oh, okay. Okay. The part that is very scary is that normal RL environment companies that are selling coding data, even they're now having to do safety grading. Because they're seeing... When you're having Claude do a normal coding task, they're trying to reward hack and trying to escape the sandbox into normal environments. So it is getting... Everyone's becoming, whether they want it or not, a safety company. Just because it is becoming a critical issue for everyone.

51:52And I think that is only going to increase and continue, not kind of randomly go away and get solved. Sure. Sure. Yeah, that makes sense. What numbers from your demo day pitch can you share with us? You want to hit the gong? Oh, yeah. I can't. There are some numbers I can't. How many employees you got? It's only me and my co-founder. There we go.

52:19But needless to say, you're ripping. That's kind of the vibe I'm getting. Seems like you're doing well. Yeah. Most of our metrics are public if you go through our stuff. Yeah. Okay. Yeah. Well, have a great day. Thank you so much for hopping on. Great to hang. Come back on soon. We'll talk to you soon. Cheers. We have Karen Cerfati from Agent Card, the founder and CEO, hopping on the show next. First, let me tell you about MongoDB. What's the only thing faster than the AI? Market your business on MongoDB. Don't just build AI. I own the data platform that powers it. And let's bring in Karen from Agent Card, talking about giving AI agents credit cards.

53:01Why not? Amazing shirt. What is, let's start with the shirt. Let's start with the shirt. What's on the shirt? That is a great question. So this is Braxter. It's an AI character that we came up with. Okay. You know, this way, C events are like an ocean of black t-shirts with very small logos. Yep. And it was very important to distinguish ourselves. So yeah. Okay. Smart. Tell us about the character. It looks like a real person of sorts. Yes. So it's like a bunker from the 80s. He does all our marketing. If you've been to our website, you'll see that he's there promoting agent card. Yeah. What is the elevator pitch for why not just use a standard card?

53:45Why not give your personal agent just your personal card? You can do a chargeback if it messes up, maybe. But why do you need a new card? Actually, we think that you should be using your real card. Agent card is not a new card that you give to your agent. We actually help you use your Chase card or whatever it is that you have. So the agent can use it in a way that's safe. Because I fully agree with you. I think that when you ask consumers, people actually prefer to use their real cards. Because you get points. You get disputes. Like all the things that you hear about. So that is actually what we do.

54:19So then at what point do you plug in to a company that wants to enable agent of commerce, to an agent platform that wants to provide more secure payment interactions? Who are you selling to? Exactly. So we work with companies building personal assistants. So I'm sure you've seen online companies like Instinct, Orchid. These are the type of companies that are trying to enable authentic payments for consumers. And what we do is basically we're the payment platform that they can use. So their users can use their cards safely. Interesting. And then do you charge a fee that's based on the transaction size?

54:58Or is there just a SaaS fee associated with using the service? Yeah. So we charge a monthly fee to companies. Okay. And they get access to our main product, which is called Vault. Vault allows companies to store customers' cards. And we have a second product that's called a purchase API, which is basically a single API that these agents can use to buy from any merchant online, which is a very hard problem to solve. If you want your agent to be able to buy something, maybe you've tried it. You go to Codex and you say, hey, buy me a table from Amazon. And what the agent's going to do is open a browser and start navigating, which is very painful to watch.

55:36I got banned from a website for doing this exact thing. It was like, you're a bot. I was like, I'm just trying to buy an expensive shirt. Give me a break here. They lost my business for maybe ever, but definitely a few days while I was on the cool down from using an agent. What do you expect e-commerce merchants to do differently? Because when my experience was having an agentic interface to their e-commerce website was going to accelerate my purchasing. I was going to spend more, more confidently. I was going to become a more regular customer. It felt very accretive for them. And yet, clearly, their current defenses and their posture was, we don't want that business.

56:22Yeah. In my experience, merchants do want agents to be able to buy. They're just not sure about how to implement that in a way that is not a risk for them, which makes sense. We think that the way to enable authentic payments and authentic purchases is to solve all this flow. So the first part is like, how do we give agents a card that they can use? Second part is how we help agents navigate these websites. And the third part, which is actually the hardest part, is how do we help agents complete a checkout page? Because like the problem that we see now is that like we've been working on this, like Clubflare has been working for ages now on preventing bots from completing checkout pages.

57:04And now we need agents to be able to do it. So that is the problem that we solve. It's not only the card, it's the whole thing. How do we help merchants sell to agents? How much of that is a technology problem? Because anytime I see a CAPTCHA, I think, okay, you could throw a lot of compute at this and beat it. Or you could just meet with the CEO of that company and say, hey, let's figure this out. We're going to do a deal that's mutually beneficial. How much of it is on the phone and on meetings versus technology? I think that a big part of a company like this is the partnerships. but what I think too is that we cannot wait for the merchants to be ready.

57:43There are a lot of companies, this generation of companies that's building this next consumer app, they need this right now. And they are educating consumers so they can buy things with agents. So we need to solve this right now. In the meantime, we need to be talking to the merchants and we need to be talking to the PSPs and to everybody that's involved so they can enable payments for agents. But in the meantime, we need to solve it. How are you sort of tracking the growth of this market? It feels really primed to be something that 100X is over the next year, if it's done, which means it's a great category to be building in as a startup.

58:21But how are you tracking it today? What are you looking at? Is it the instincts of the world driving most actual volume of payments right now? Now I know that Muse seems very interested in actually driving a bunch of conversions and purchases with their agent. But how are you looking at it? Yeah, I think it's impossible to tell. I think that we have two intuitions. The first one is that we've been using agents to buy things for months now. And we know it's only a matter of time until someone figures out what's the right interface so consumers can do that too. I buy things from my terminal. I've been buying flights from my terminal with my agent for months.

59:01But I know that my mother is not going to do the same. She needs something that's better. And I think that that's where companies like Instant come in. They are building on top of iMessage, which is a platform that's a lot more familiar to consumers. We have no idea the size of this market. Our guess is that this is going to be larger than when e-commerce started. Because agents are faster. They can make more payments. They're more efficient. and it's just, in my experience, it's just a better experience to buy with agents than to buy from a website. Thank you so much for coming on the show. Yeah, great to meet you.

59:38Congrats on a great day. We'll talk to you soon. Thank you. Have a good one. Goodbye. Talk soon. Let me tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. Maybe you start at YC. Maybe you make it to the NYSC. Up next, we have Antonio Lee from Nori Robotics. not going to spoil it too much but building humanoid robots and the price will shock you shockingly cheap humanoid robots too cheap too good to be true let's bring him in and ask him antonio we're going to bring him in in just a second uh tennis matches are finishing at 3 33 a.m we'll get to that story later for now we're staying with humanoid robots because we are joined by antonio lee from nori robots whoa there he is now i understand how it's so cheap i can see the cost reduction how much does that thing cost i want one uh 16.88 under 2k 16.88 16.88 no cheaper than an iphone duo than an iphone duo that's crazy hogging turnus and it has not one hinge but multiple hinges how'd you build this thing what's the idea what is it actually capable of why should people buy one for 1688 great question um so we built it entirely in San Francisco.

1:00:55We put a bunch of motors together. Majority of it is 3D printed plastic, and then we have an aluminum base below it. It's able to do almost anything. Well, obviously, it's a humanoid robot, and you should buy it if you want to do development on American humanoids. Shotgun bench. Dev kits. Cannon, shotgun, that ramp, cold brew on the desk. Do anything within the API. I imagine you're not doing Boston Dynamics level demos juggling and dancing yet but uh it's an opportunity for someone to come in and solve a particular problem do you see this as more uh consumer focused or industrial what do you uh are there any forks in the road that you're designing around obviously price is a really key differentiator but where do you see your design your design decisions today leading to in the future yeah so we see it being used within the whole mainly for now right now it's able to handle to 80 to 90 % of the majority of home tasks.

1:01:56But in the future, we obviously want to be able to deploy into hotels, grocery stores, nursing homes, anywhere that can be of help with the sort of labor shortage that we're seeing right now. Sure. Yeah. Talk about the importance of just low costs in the sort of takeoff of humanoids. Because I think one of the, like, I know so many people that would buy an iPhone-priced humanoid just to mess around with it and have some fun. Honestly, I think a lot of venture capitalists will buy it just purely because they want to understand the category and the progress being made. But a lot of people are – when you're starting to see a humanoid that's priced more like a motorcycle or a car, it just gets into a different – like people are going to have such high expectations where they're like, hey, I spent – I could have got a Model 3.

1:02:44The robot dog, you can actually buy that, but it's like$50 ,000, and I still haven't really seen that many people pick them up for fun or for experimentation because they're too expensive. Yeah. Yeah. Please. Yeah, so obviously I think when it comes down to early adopters, you need to have an initial low cost as a wedge into the market. Right now, humanoids are a little bit stupid. It's because the AI behind it is not very well developed. so you need to have it more be like a larger toy in some ways and be within that price point to be able to appeal to a larger consumer base what's the weight limit for what it can pick up we got a question from the chat it's around 1.5 kilograms per arm per arm okay so yeah a plate a cup a beer a four loca it can it can maybe get a monster with the monster energy uh what what tasks do you guys use them for around the office and how do you actually control what are the best ways to control it as a consumer are you just giving it you know verbal commands yeah asking nicely of course so some like the easier task we have it grab a spear so at the end of the day and also grab us waters so some of the simpler stuff right now the majority of the controls come directly from llm so far so we've seen actually an incredible improvement in the ability of like llm to control robots under lower level so chat 35 astro has been absolutely amazing for us and And it's been able to one shot a lot of tasks really well.

1:04:05Yeah. How important is the development of agentic coding systems and these harnesses? Because it feels like, you know, a decade ago, people would grab like a Raspberry Pi, but they'd have to write all this code. They'd spend days just wiring up like, you know, some gadget that would turn on their lights, you know, around Christmas or something like that. this feels like even if you're interfacing with the robot at a pretty low level via an API, there's a lot more flexibility on the dev side to accelerate these demos and these experiments. Yeah, of course. I would say coding agents have been so incredibly important for robotics because it actually allows you to use a generalist robot in a generalized way.

1:04:49Obviously, if you take 10 hours for it to just code it to grab a beer, then it's not going to be very useful in the house. Whereas if it can take under like a second for the robot to write a thousand lines of code and deploy it to the robot, it kind of is a game changer and how useful the robot can be. Yeah. So are you seeing, like, I can imagine I get this piece of hardware and I want to write a script that has it go from my living room to the fridge in the kitchen. And I can use an agent to write that code, but I could also be feeding the actual data live via the API to the actual model. Are people doing that?

1:05:26Are you successful there or is it too slow at this point? Like at what point do you stop actually writing any glue code between the model and the robot? So I think in the near future, you still need a lot of that glue code. The main point is to make it as easy as possible for the LLM to use the robot. So for example, let's say if you want to move the robot from the kitchen to your bedroom, you can simply call on the onboard SLAM, which is a 2D LIDAR that has a navigation stack. And that makes it, it's like two lines of code. It just goes from one place to the other very easily. It's all about simplifying that and abstracting it away for the LLM.

1:05:59Got it. What is your capability? You've built one. Can you build 10 this year? Can you build 100? I imagine there's going to be a manufacturing facility. There's a lot of different steps to actually scale up to any meaningful volume. But what does that road look like for you? So right now we're making around two to three per day. And we're mainly deploying to our earliest photo research partnerships. We are manufacturing them all entirely in San Francisco, or right down 10 months from here from tons of points. That's crazy. And our goal is to basically, by the end of six months from now, to ramp up to around 400 robots per month manufactured here in San Francisco.

1:06:34Yeah, that's not bad. So already doing 1 ,000 a year. Next year, 10 ,000 at least, something like that. What do you, so how do you think that, like, where will the intelligence, like, specifically with the long term, like, I can imagine, like, somebody just has, like, a Codex or Claude, and Claude just has fully access to the robot. And so you, as the manufacturer, Nori, don't even necessarily need to provide the intelligence layer as long as, like, the robot has, like, the right sort of control mechanisms. Is that one potential scenario? because I can also see the scenario where you guys end up building effectively your own end-to-end agent and you have local models and cloud models that help people accomplish tasks.

1:07:22I think the way that we see it is it comes down to how fast LLMs progress in the next one year or so and they have been progressing very, very fast for the past year. I think somehow, I'm not sure how they did at OpenAI, they managed to give Astra the ability to understand 3D spaces very well, which has been a game-changer for robotics control. This was, I thought, to be impossible even a year ago. I think for us, it ultimately comes down to, I think we don't quite have control over how fast these big models or these big labs progress. So we're just going to be using what's on the market to get the best use cases and the best consumer product to our users.

1:07:54Last question from the chat, and we'll let you get out of here. Do these robots have lifelike hands? Will the hands seem a little shy below the table? What do we got? What are we working with? Is it a claw? It's a claw. Okay. Looks good. Obviously, the cost is a pretty large consideration for us. We tested out a bit more dexterous hands. And from the results that we've seen, you're able to do 90 % of what a hand can do with a claw for now just because the control policies are not quite there to fully leverage the dexterous nature of a hand like this. Yeah, no, that makes a ton of sense. Trade-offs and everything, and I think you made the right one.

1:08:29Well, congrats and thank you so much. Very cool. We have a drum set here at the office. We got to get a Nori to just play the drums. So when one of us has a good line, he can go... That'll be good. Have a great rest of your day. Great to meet you. We'll talk to you soon. Good to meet you guys. Great stuff. Cheers. Let me tell you about console.com. Console builds AI agents that automate 70 % of ITHR and finance support, giving employees instant resolution for access requests and password resets. We're going to get to the tennis story, but we are joined by Edward G from Exosat, founder and CEO building sovereign satellite infrastructure to compete with Starlink going up against the easiest man to compete with Elon Musk.

1:09:11What is the plan? That word doesn't look worried at all. Not worried. What is the plan? How many satellites do you need to launch? What cadence do you want to get these up? What is the key pitch for Exosat to differentiate you in a very competitive market? I mean, we filed for 11 ,000 satellites. We actually have a published ITU API and CR spectrum filing for 11 ,000 satellites. I think it would be the biggest satellite network outside of the United States, possibly even China, too. The idea is this, right? A lot of countries don't want to rely on American defense contractors or Chinese state corporations for critical network infrastructure.

1:09:53And we are coming in and we're offering them something in the middle that's cost competitive that can be launched on any launch option in the world, be it an American rocket, an Indian rocket, a Chinese rocket. And that just ensures no single nation can block them from getting into orbit. And yeah, going up against Elon Musk is a huge proposition. But think about it this way. If we are even a mild annoyance to one$2 trillion company, how much would we be worth? Yeah, sure. You're in San Francisco, probably a Delaware C Corp. Feels like an American company. How are you actually pitching this? We are not a Delaware C Corp.

1:10:35We are a Singaporean parent company. We're one of the only single. Yeah. So I think if we want to see batches around like two or three weird companies, like for So they're either in crypto or some weird reason like us. They are not at the Delaware C Corp. And we are one of these companies. So all of our engineering is done outside of the U.S. We know we don't use U.S. vendors. We don't use U.S. supply chains due to ITAR. And that means we can launch, build satellites that can be exported to and launched from any nation in the world. Interesting. Interesting. So are you basically going after like the long tail of countries?

1:11:10You mentioned it could be launched from a Chinese rocket or an American rocket, but it seems like you'll be selling to India, Japan, loosely Western countries. But is everyone? Lots of global South countries like Kazakhstan, nations in Africa, Indonesia and Malaysia. We have a huge, huge interest there. A lot of it is we found there's a lot of countries out there that want independence from U.S. supply chains, want independence from Chinese government owned infrastructure. and that's what we're really going after. The world is getting more multipolar. We're starting to see a lot of countries start to essentially become, like, what was the word again when it comes to supply chain independence?

1:11:54Sovereignty. Sovereignty, yeah, but sovereignty goes both ways, right? Sovereignty affects both moving away from both the Chinese sphere but also the U.S. sphere, and we are the company in the middle that's giving them the satellites I do want. How do you approach launch? How do you get these things up there? Well, we work with Indian, Chinese, and American launch providers in the U.S., right? We can use, like, SpaceX. We can use Stokespace Nova Rockets in China. We have the LAND Space EQ3 as well as the LAND March 10B. In India, we have the Vikram 1 and the Vikram 2 from Skyru and the PSLV launch vehicle.

1:12:31And in New Zealand as well, we have the Rocket Lab vehicles. So we can launch anywhere. Interesting. What was your background? So I'm a second-time founder building the second space company. Before this, I dropped out of school in 2022 to build my first space company called Aethro. At Aethro, we were doing radiation-hardened satellite edge computers and orbital data centers. And we grew the company quite a bit. We actually got to millions of dollars in revenue. And we actually put two satellites into orbit successfully. But the big thing was we didn't—I think the biggest—so two things. I'm a huge believer in sovereign technology, right?

1:13:09So I think a future where just SpaceX and China dominate all of humanity's space infrastructure is not a good future to be in. It's like AI. You don't want to just open AI and throw it to rule all of humanity's AI. That's a pretty bad outcome. So I think the other 80 % of humanity should have a say in that future. And the second thing also is I think the biggest space companies in the U.S. and China have already been founded. I think there's so much competition going after these big anchor customers like the U.S. military or the PLA in China's variant. but like space is a field where incubant advantage matters so much and that the best move to do might be just going after everyone else going after all of these countries that traditionally haven't been prioritized by space companies and that's what we're doing amazing well congratulations and thanks for coming on and breaking it down yeah very cool great to meet you have a great rest of your demo day goodbye cheers let me tell you about shopify shopify is the commerce platform that grows with your business and lets you sell in seconds online, in-store, on mobile, on social, on marketplaces, and now with AI agents.

1:14:13And I'm also going to tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web app servers, databases, and more, while Railway automatically takes care of scaling, monitoring, and security. We have our next guest, Alfredo Gonzalez from Momonta, coming on in just a minute. Just a minute. There are so many good iPhone Duo reactions. The toughest thing for Apple's marketing team, the sequel, they're going to call it the Apple Duo 2. Is it not the iPhone Duo? It is the iPhone Duo, right? It's not the Apple Duo. It's not a separate. It's the iPhone Duo.

1:14:51iPhone Duo 2, iPhone Duo 3. That's going to be a mouthful. They never did. They don't have iPhone Air 2 yet, do they? It's just iPhone Air, and then they just stopped it. They haven't refreshed it yet. Air Max. iPhone Air Max 95. They'll get there in 94 years. Something like that. Well, anyway, we have our next guest in the waiting room. Let's bring in Alfredo from Omanta, building a personal research lab around individual patients. Alfredo, how are you doing? I like where this is going. I like the one-liner. Thank you so much. Thank you for having us. Thanks for hopping on the show. Congrats on a great demo day.

1:15:31Introduce yourself in the company. Yeah, so a little bit about me. Our team is essentially biomedical researchers that have been dedicated to just improving patients' health outcomes for the last decade. And we realized that there's a huge gap right now with the resources that patients and clinicians have as far as being able to go really deep into their data and ultimately understand the biology of their tumor and their diseases. and that's what Amanta does. We're laying down the infrastructure to be able to essentially have a whole team of LLM and AI agents that are supporting patients and clinicians with their care, helping decoding their specific disease, their biology, and finding the, not just the ideal therapeutic path, but additional therapies that would have been missed otherwise.

1:16:19So that's kind of our trick. So do you sell through doctors? Like if I have a tumor, I go to a doctor, they biopsy it, send it off to a lab, There's some raw data. The doctor might be interpreting that for me, having a conversation for me, but they might also give me a report from you or access to an app from you. How do I see the results here? Fantastic question. I appreciate you did your homework there. That's definitely much more knowledge than most people have. So as of right now, the way that the medical system works is they do take your sample and they do these very limited sequencing panels, typically about 50 to 500 genes.

1:16:56Our whole gene space is about 20 ,000 plus genes. So it's trying to debug and not having access to the whole code base. And to be able to fully do this, it's knowledge-intensive work. But now in the era of LLM, some protein models, that's where instead of having one single person that could do all this, you have a whole team of agents that can do that deep dive. So there are labs out there that can do the full sequence. you might advise a doctor to pick one of those labs because it's not just going to be more data, it's going to be more data that's actually actionable thanks to your platform. Yeah.

1:17:34So our platform is that missing link that helps the patients, the clinicians go deeper. Because as of right now, what they have is these very static reports. Sure. Right. So they're not dynamic enough to be able to keep up with the pace of discovery generally or do discovery in real time for the patients. Do you have a pretty typical business model, like insurance reimbursements or like the clinic will pay for this as a service? Or is there something else going on? Great question. So right now is direct to patient, direct to consumer, or direct to a sponsor. Ultimately, though, there has been so much research showing out that the earlier you do the sequencing and the informatics, you actually can significantly improve the median life expectancy of patients, but also reduce the cost on insurance payers.

1:18:25So it's definitely something where we are already seeing the signs of insurance starting to cover. Actually, insurance covers sequencing for recurrences and advanced tumors in most developed countries, right? And you are starting to see already how insurance, the telewins are there. Insurance will eventually cover this because ultimately it is cheaper for the payer insurance. And you're also improving life expectancy. Yeah. What is the biggest bottleneck right now to AI progress in cancer detection in medicine? Honestly, I think it's the infrastructure. So it's a fairly huge data science problem.

1:19:03I think we have these legacy systems that are built around megabytes of patient data outside of imaging. but the genomic sequencing you're talking about hundreds, like 200 gigabytes, 30 to 200 gigabytes, like it's a significant data science problem. So the infrastructure is not set up yet for these systems to fully take advantage of it. And that's in the sense what we're also setting up. Very cool. What kind of traction are you sharing at demo day? Oh, so we have launched almost a month now ago, ago and we are at 319k MRR with over 150 patients that have a PMI. Crazy. Bang. Congratulations. You got a lead with the ARR number.

1:19:52Yeah. Pick the bigger number. I thought you were going to say 319 ARR. You had me in the first half. It was good. Well, congratulations. I'm sure it'll be a successful demo day. Thank you so much. Yeah, and it's good to be able to help patients and their families and help improve the health outcomes for everyone. So thank you so much for the support. Yeah, great to meet you, Alfredo. Have a great rest of your day. We'll talk soon. Goodbye. Let me tell you about Figma. Agents meet the canvas. Your AI agents can now create and modify your Figma files with design system context. Okay. We got through demo day.

1:20:28We got to go over to tennis. Tennis, tennis, tennis. There was a match that ended at 3.30 a.m. People are not happy about it. The Wall Street Journal's headline is, No, no, no, no, no, no, no, no. There's nothing cool about a tennis match that finishes at 3.30 a.m. It's ridiculous. Sorry, but tennis at 3.30 a.m. isn't cool. It's ridiculous. What happened? Bill Ackman loved it. He said, one of the greatest tennis matches ever, and very few got to see it live. This needs to be fixed now. It's unfair to the players, fans, and sponsors in the sports. What is the actual proposal? What do they want to change?

1:21:09Start the game earlier or cut it off entirely if it keeps going? Yeah, so I'm trying to figure out why this ever is or was a thing. Hopefully the journal article can break it down. But I watched a little bit of this game. You did. I went to bed. Yeah, I think I went to bed. It was actually a later night. I went to bed around 11. But the game was going at well past three. Ben Shelton, who actually won the match, it was an upset, was hitting 146 mile an hour serves at Carlos. Carlos at one point was shooing off the camera crew because they were filming him, and he was actively going to throw up.

1:21:57He was pushing it so hard that he was going to throw up into a towel. So really, really crazy. Anyways, Wall Street Journal says, Sometimes this glorious sport can't get out of its own way. Everyone is expected to pretend night tennis is a fabulous experience. We're supposed to deploy the word epic and act like we didn't brush our teeth and hit the hay after Shelton took the second set. I do actually have a family member that did exactly that, like blowout second set. They were like, all right, we're good. We're good. And it ended up being this incredible game. The U.S. Open has long cultivated the idea that the wee hours are part of its identity as much a part of the formula as the$23 Honey Deuce cocktails and the plume of cannabis smoke wafting in from a nearby park.

1:22:45Even ESPN feigns enthusiasm for the Opry Midnight when it would never schedule a tentpole like the Iron Bowl for such an uncommon hour. It's a city that never sleeps. Why should tennis? Okay. How many 3.30 a.m. finishes have you attended at Yankee Stadium? I appreciate the romance. New York remains a restless town where you can get a steak dinner and a tattoo on your back long after quitting time. The subway is 24-7 theater. You can crawl out of a bar at dawn. There are amateur bike races that start in the dark at 5.45 a.m. It's zany. But Alcaraz versus Shelton isn't a corned beef sandwich at an all-night diner.

1:23:27This sounds like AI slop. It's a marquee showcase for a sport. I've been reading this, and I think it's just like fun turns and reasons. No, it's good. It's a marquee showcase for a sport that needs every bit of U.S. attention it can get, and it got buried like a Tom Snyder sit down with Buddy Hackett. Even people in Hawaii were like, this Alcaraz-Shelton match is running pretty late. This isn't a routine to celebrate. Okay, I'm sorry. A lot of, a lot of, a lot of, we're going to have to run this through Pangram. Parallelism. Yeah, this isn't a routine to celebrate. This is when someone needs to step in and scream, what are we doing?

1:24:04You know, I found two new AI phrases. I don't remember both of them. One of them is in our increasingly online world, when someone's writing about online culture or something that happened on the Internet or like with the boom of social media, like there's two phrases that were popularized because journalists who needed to write about things that happened online kept using this phrase. that got baked into the LLM and then now it's getting spit back at people. And so a lot of people are turning away from the phrase in our increasingly online world. So avoid that one too. Sash Shane says nighttime.

1:24:44You wouldn't understand that. It's a New York thing. Yeah. Yeah. It's funny because I feel like you can maintain the mystique of the U.S. Open by just having games wrap up around midnight. It's still very much nighttime. Totally. Like it's dark out. You know. people are sleeping. Anyways, definitely seems just incredibly rough and unfair to the players. This has happened in UFC a few times where in order to hit a U.S. time zone for the pay-per-view, the actual fighters have to be fighting at the equivalent of like 6 a.m. Oh, yeah, internationally, yeah. Which is just, imagine waking, imagine one going to sleep at like 10, knowing that you have to wake up and get in a death match at 6 a.m.

1:25:34Sounds pretty rough. The alternative is just to stay up and, like, you know, just try to be ready. Because it starts at 6, so you've got to be warmed up. Like, the whole event is starting at probably midnight. Well, I guess the move would probably be just don't adjust for jet lag. Like, get there on your schedule and just treat it like it's a 6 p.m. fight for whatever your local time zone is, I suppose. I don't know. What do you think about this? Chairman Burb Bernanke. You can eat the same thing for breakfast every single day for like 20 years in a row and no one will say anything. It's normal.

1:26:08You eat the same thing for dinner for a month straight and people call you psychotic. I think it's fine to eat the same dinner every night for a month. What's wrong with being a psycho? I used to do chicken and rice every night for months, for years. Yeah. That's a good point. Let me tell you about public.com. Investing for those that take it seriously. They got stocks, options, bonds, crypto, treasuries with great customer service. And they got more. Anything else in the timeline before we get out of here? We do have to cut the show a little bit short today. We have a little bit of family stuff going on.

1:26:42But we will be back tomorrow at 11 a.m. sharp. We have a great show. Alex Heath is joining us. He has some big news. He's becoming a venture capitalist, putting some money on the line, finally risking it all in the markets every day. putting, putting down his capital J journalist hat, hanging it up. Now he's keeping the hat. He's keeping the hat. He's keeping the hat. He's keeping the capital J journalist hat, but he's also going to be right. Ripping some checks. Okay. Ripping, taking some flyers. And, uh, yeah, we got Sager and Jetty coming on tomorrow. I'm sure we'll talk about that. Yeah. Safety stuff because, uh, he's been a friend who's been evolving.

1:27:20And hopefully we get an update on Sager bet. Yeah. Features. Definitely. I've been enjoying seeing the, uh, Sager bet, uh, references in the chat on and off, but leave us five stars on Apple podcasts and Spotify. Sign up for our newsletter, tbpn.com and we will see you tomorrow at 11 a.m. Pacific sharp. We love you. Goodbye. Have a great rest of your day. The best Thursday afternoon of your entire life. Goodbye.