Neurosymbolic AI outperforms chatbots and product search | E2327

LLMs are not good at this. They regress to the mean and they find the most probable thing, and that's often not what you need personally. When you need to remove hallucination and you need trust, you need to be able to see inside the model. That's where neurosymbolic models can really shine.

You really have built something between Houzz and Pinterest that is more powerful than both of them.

The performance of it is already 2.5x greater than some of the largest search companies in the world. It learns that new information so you don't have to go do another training run. It's literally updating in real time. It's 1/1000th the cost of an average frontier model training run in the US.

This Week in Startups is brought to you by Wyze Security, the on-demand security team for startups. Need enterprise-grade security without hiring a $400,000 CISO? Security gives you 40+ expert engineers matched to exactly what you need by the hour, with your first 6 hours completely free. Go to wyzesecurity.io/twist.

Agree.com.

Stop chasing invoices and automate your entire contract-to-cash stack. Go to agree.com and tell them Jason sent you to get 50% off for life. And Vanta. Compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy for companies to get an SOC 2 report fast. Get $1,000 off for a limited time at www.vanta.com/twist.

We got Zach Cutts joining us. He's the CEO of Onton. They're a search and discovery engine for e-commerce. It's powered, Jason, by a neurosymbolic AI model. What does that mean? You could see it for yourself at ontoon.com. We're gonna find out right now. Zach, thank you so much for joining us.

Thanks for having me.

Pleasure.

Yeah, and you're gonna give us a demo, so let's, let's, uh, you know, show, don't tell.

Yeah.

Oh yeah, yeah, yeah.

What have you built?

Yeah.

We can dive right in.

Let's do it.

Oh, and we're investors in the company, um, just as a full disclosure here.

Yeah. I'll give your audience a chance to go and do some, some basic searches for themselves, but why don't I jump in and do some, uh, more of the mind-blowing searches? Um, and a little bit of, a little bit of background about Onton. We have these things called surfaces within our product. If you've ever, you have, have Lon or Jason, have either of you used Notion before? I'm sure you have.

Plenty of times.

Yeah.

More than once.

So they have these things called blocks there. They let you organize information in different ways. People do this quite often in shopping journeys too, where they create lists, they create mood boards, these types of things. These are surfaces within Onton. And so what I've prepared for you, uh, was I put together, uh, some of your favorite things, Jason, or at least some of the hotels that you've mentioned, some of your The things you mentioned were—

We sent him some cheat codes. Jason, Monocle magazine, a few hotels we know you like. We helped Zach build this.

I went back through some of the podcasts too and grabbed a few others. But we, we put this into a canvas. But what the— what we've realized recently is it wasn't possible to search these without this model. And so I want to use this to demonstrate what Ontology One can do in this, this type, the amount of context these models can take.

Okay.

So I'm just going to go search with it. I'm going to use this just to go, hey, Uh, try to find some stuff.

Okay, so you put images on a mood board of hotels I've stayed in and the design of those hotels. So I saw one of them there look like the room at the Artisan in Singapore.

Yeah.

And I put even more than that. I put actual products in here. I put some text and other contexts in here, and Ontology is able to suck all of that up. So it's, it is a lot like a mood board, but Ontology, one, will look at that whole thing and it will try to understand it. The thing that I want to call out here is that it's learning in real time. We don't have any tags for these things. We don't have any— this wasn't like human labeled. It is going off and learning them, but more importantly, it's storing that understanding in the model. So every next search that's happening with Anton is getting a tiny bit smarter and a little bit faster, which is incredible. We don't have to go back and do another training run with this model to get better accuracy. So it's learning about things like organic burl wood, brass, floral, all the things that it thinks goes into this. And in my opinion, these look pretty accurate or look like they could belong in some of these hotels. I can almost one-shot the room that I'm going for, uh, which is a lot more like how e-commerce search actually works, right? You have a, a chair and you want that chair to match the lamp, or you have the lamp and you want it to match the other parts of your room. Uh, ontology can do multi-category, multi-object searches, but more importantly, it can be honest with you. It can tell you exactly why these results came back definitively. I can look down into the model. It's not a black box like a typical LLM and see why these results came back. So it can be honest and, and explain these things. But I could, I can take this and I could throw this product back into this mood board and I can use this to recursively search. So I've plopped it in here. I could see how it looks.

So you put a Cambridge loveseat here in a 1930s Hollywood style. Only, uh, 16,000 lawns. So this might be right in your zone for—

It's out of my budget, but maybe in y'all's budget. Um, uh, but I can recursively search with this.

Yeah.

So I've now updated ontology's understanding. I could go do another search, but, um, just to show you what, uh, what else this could, could do, I can, uh, I, I've prepared one more for you. I've taken one of your friends, Jason, and, uh, sweater thrown, thrown, thrown. Maybe this meme is already dead. I know news cycles are, uh, pretty fast these days, but I can take this one and I could also search with it. And, um, let's just drag it into the search bar and let's see what comes back. But actually, let's augment it. Let's say couches that fit this vibe of these AI-generated—

And the vibe being Chamath's oversized white turtleneck.

Yeah, Chamath's oversized creamy off-white oversized chunky knit. It's learn— it's learning these concepts in real time, which is incredible. So I think, uh, this is what ontology can do. Yeah, this is what ontology can do, and it's getting smarter with every search. So it's literally learning these concepts. So this is— we're really excited about it. And the performance of it is already 2.5x greater than some of the largest search companies in the world. We have a smaller index size. So I can only imagine where— excuse me. Yeah. I can only imagine where it goes once we start adding some new categories.

Who's your customer for this? Obviously, designers, consumers, everybody wants to find that perfect object, uh, to put in their room, etc. And then also people want to build a new product. So I was looking at this thinking, hey, uh, what if I want to get custom furniture made, or I'm a manufacturer and I say, here's the Blade Runner apartment that I love. I'm a big fan of, uh, Deckard's apartment. So if you— can you get a picture of Deckard's apartment and the Blade Runner set and then make me a Blade Runner Uh, find me Blade Runner-esque, um, yeah, an aesthetic.

So you can definitely do this, Jace. I'm going to share my screen. I did this just to play around with Anton with Twin Peaks, and it's been— it, uh, so you can see I, I just put in a bunch of images from the TV show Twin Peaks, and when you search, it can find anything like rustic. It knows about rustic decor. It knows about, you know, like, oh, you're looking for surreal or weird things. Like, it's actually pretty good. So if I just put like Items for a Twin Peaks living room. It'll—

or a bar, you know, you could make like a restaurant that looks like Twin Peaks.

It's, it's really cool how it sort of understands like, like it's like vibe search or aesthetic search beyond just like items or, or words you're putting in.

This is a problem. Like taste is so beyond what— and you've talked about this endlessly, Lon, just like it can't make jokes. It doesn't have taste. Zach, have you cracked essentially taste here?

I think a lot of people bundle the shopping journey into 2 categories, uh, or in, into one thing. And I think it's actually a lot more nuanced than that. I think there are wants and needs. Needs are like toothpaste, uh, deodorant, the things that you order very quickly. They're often commoditized products, but there are a lot of other products. Broadly speaking, they're above $50 and they're driven by totally different things. Taste is one of them. Self-expression, emotion. imagery. LLMs are not good at this. They regress to the mean and they find the most probable thing, and that's often not what you need personally. So what this model is capable of doing is it's capable of learning about you too. While it's learning from searches today, it can start to update itself to where the search index reflects the things that you really enjoy, uh, in a very different way. So it's, uh, it's not like there's just one output that is the most The most likely it is, it's creating its own separate understanding for, for every search. It's, it's learning that new information.

I've got a question.

And you can go check this out at onton.com. O-N-T-O-N.com. We invested back in '21, '22 during COVID and this was a search engine. This is pre-ChatGPT. The world caught up with you in some ways in terms of AI, and you've now got millions of people doing searches.

Yeah? That's right. Yeah, we're on track for hundreds of millions of searches over the next year, and that's only in one category. So, As we move into others, I would expect that to multiply pretty quickly.

What's the revenue model now for the business? Uh, uh, 5 years in here, 6 years in?

Yeah. Uh, so we do make some money from affiliate commission, but I think our largest revenue opportunities are in front of us. Uh, we're releasing an API of ontology to be plugged into all these stores. And I think, uh, it, there's the reason why there's an API instead of say a Shopify plugin is because we found this model to be generalizable. It can be applied to use cases beyond just e-commerce too, but obviously that's, that's where we're starting. That's our bread and butter. So that's, that's how we're, that's how we're making money. But we have been really focused on, on user growth.

Everything is moving a lot faster in the age of AI, but moving fast often means taking on more risks. Every time your company adds a new tool or a vendor launches a shiny new integration, that's one more opportunity for something catastrophic to go wrong. No one wants to get left behind, but you still need to worry about security. It's kind of job one. That's why you need a partner like Vanta. Vanta is the number number one agentic trust platform, and it's being used by over 16,000 companies. We're talking extremely legit organizations— Ramp, Cursor, Harvey, all those great surging companies. Vanta is making sure these incredible startups are audit ready, and they're ready to do the same for you and your startup, as well as looking out for all the problems that can arise between audits and across all your AI tools, vendors, and your entire work environment. Whether you're a fast-growing startup or a global enterprise, Vanta is here to help you automate your security and compliance and earn and prove trust. So get started today at vanta.com/twist. That's V-A-N-T-A.com/twist. If you do it that way, then anybody could hit this API from their product or service and you just charge them based on consumption, yeah?

That's right.

Yeah. Interesting. Another product category coming?

Apparel is coming next. Um, we see tons of searches for apparel right now that we're not fulfilling. Electronics is probably the next category after that, uh, which you'd be surprised. Um, I think what, what is fascinating about ontology is that as it learns about one category, it makes the next category smarter, right? So as we learn what polyester is in, say, home decor and furniture, the search engine gets better at understanding what polyester means for something like Uh, apparel. Um, and, uh, so it, it makes it a little bit easier for us to move into categories than what's historically been the case with e-commerce, where you have to set up the whole, uh, you have to set up everything. You have to go through, you have to relabel all the products. You, you basically have to build your, your understanding of that category before you move into it.

I was gonna say on This Week in AI this week, we had Izzo Kant on from, from Poolside, and he was talking about how they're pursuing coding models as the route to AGI. Like teaching, if you make a model perfect at coding, eventually it'll become super intelligent and self-aware. I'm curious, do you feel like neurosymbolic models could one day get there too? Because you are kind of teaching the model about our world and sensory experience in ways that a coding model is not.

That's a really great question, Lon. Um, I, I think there's a lot of use cases for large language models. I'm not saying they're not useless, um, but they are great at certain categories of things at a certain scale. They're, they're, in my opinion, they're great at verifiable domains. Coding is one of those. Mathematics, you're starting to see them break into mathematics because they have verifiable, you can use things externally to verify the output of those large language models. But when you need, uh, when you, when you need to remove hallucination and you need trust, you need to be able to see inside the model. That's where neurosymbolic models can really shine. So I'd say the other 3 benefits are, it's less expensive. The cost of this model to outperform those search companies that I mentioned was one, it's, it's 1/1,000th the cost of an average frontier model training run in the US, which is incredible. It learns that new information, so you don't have to go do another training run. It's literally updating in real time. And then the final thing is probably trust. So I think at a certain scale, this model could replace some of those things. It can do things like, I think at a certain scale, all models can actually be pretty great at a lot of tasks. But the question is, is it the most efficient model for the job? And neurosymbolic models are quite a lot more efficient at certain tasks like the E-commerce.

Lon, I just put in like, uh, I got the actual image of the— I think it's the Innes House in, uh, LA. That was, uh, a Frank Lloyd Wright. Yeah. And I almost bought this house, by the way. My wife stopped me. Well, it was only going for like $4 million.

I feel like if you own a Frank Lloyd Wright house, you have to like let people in it to look around, right? Like you kind of—

Well, that was one.

That was one. Well, and it was, um, so there's that. Then you also, Zach, had to restore it, and all these tiles that he made, you would have to go source. And, uh, it turns out, uh, Ron Burkle, also a fan of Blade Runner and real estate, the billionaire, um, bought it. I think he spent like $15 million renovating it.

Mm.

Um, and then I don't think he sold it. I think it's now part of like Um, and, um, like some sort of, um—

Wait, is it trust?

Yeah, I think, I think it— I think you're right. I'll look at it.

Yeah, if I remember the history correctly. So when I was like starting to look at it, it was like, yeah, you, you— this is not going to work, son. You're going to need to spend $10 million for a 2-bedroom house that is totally the opposite of kid-friendly. It's a totally dysfunctional house. But it was— this is like a really powerful, uh, solution that you've come to. Is this Now, a small, like, um, a small language model that you own and is proprietary, or is it like more like a harness when we think about it as a business and the IP here? What is it that you actually own here?

Yeah, this is a fundamentally different type of model. So the way that you can think of a large language model, or actually to oversimplify the brain, you have like 2 halves of the brain, right? Maybe, uh, an intuitive side and a logical side. LLMs are kind of like the intuitive side of the brain, but it's when you apply a logical side to the brain that you really start to unlock the intelligence of it. Some people, psychologists sometimes talk about this as like systems 1 versus systems 2 thinking.

Mm-hmm.

LLMs are great at systems 1. So the, the, uh, this is literally a, a completely different model. It's a union of a couple different things, but we do have a lot of things we've built in-house to make this work, including a graph database that, uh, stores a lot of this information, and it stores information in a very different way. Um, so yeah, it's a, it's, it's a completely new approach. I wouldn't even call it a small language model. You could scale this thing up a lot.

Really super fascinating. And it's great. Like, it's one of the amazing things, Lon, about when you invest in smart people, you know, especially when you're like, we're super early stage. I'm, I'm not sure exactly. You went through the accelerator, then we put a little extra money in. I mean, this was years ago. And I've been like, wow, I wonder what's going on with Anton. And like, just, You guys just grinded and grinded and grinded and figured it out. But you, if you, my investment philosophy is passionate founders who are gritty and figure things out and are passionate about that space. So they're not going to give up. And here we are. This is an incredible product. What's the future of the company, do you think?

Yeah, I think, uh, we're launching that API, um, for a lot of reasons. Like, uh, there's this, this great quote from Sam Altman quoting Paul Graham that says you should always create an API no matter what, good things will happen. And, um, and, uh, the, uh, that was, we can largely thank the Cambrian explosion of AI products that happened, um, from that. It was around GPT-3 that they launched their API and you see all, all the opportunities that developed. So, um, we see a very similar path for this model for our company. Yes, we do still want to be the starting point of every shopping journey. Actually, let me, let me kind of reframe this differently because I I think the, the shape of companies are changing right now. Previously, you would be a product company and you would have a technology that would, uh, really, you would have a technology that would help you reinforce that strength. I think companies are kind of becoming a little bit different now, especially with the AI labs. You have a model that is good at certain tasks and your product is a reflection of what the model is good at. In our case, we have a neurosymbolic model that is great at e-commerce and all these trustworthy searches where you need accuracy. Uh, all the other AI companies that are using large language models, they're, uh, we could take OpenAI, for example. They have a model that was great at conversation. So they launched ChatGPT and now they're starting to see code, uh, the ability to do code. So they launched Codex. So I think the, the, the structure of companies is changing and Anton is no, no different. Amazing.

Um, I'm just like, now I'm totally distracted putting in Tyrell Corporation.

It's fun to play with. Notably, the Ennis House, it is featured in Blade Runner. It's Mayan, Neo-Mayan architecture. It's in Los Feliz. It was memorably in the 1959 horror classic House on Haunted Hill, Jason. That's the—

Oh, yes.

That's the other most iconic movie. And that's one of the reasons it fell into disrepair. That movie came out in the late '50s. It was a sensation, and kids would then go find that house and like scream and carry on outside because it was in this big horror movie. And so people, nobody wanted to live there. And so in the '60s, it fell into Abandoned, disrepair.

Isn't that interesting?

Well, and to show you the to show you the depth of this and and how crazy it is, I was watching this incredible show on I don't know if it's HBO or if it was Apple TV, but I'm watching Seth Rogen's show *The Studio*.

It's Apple TV.

Apple TV. Thank you. And it turns out, and I was like, I said to my wife, "Is that the Ennis house that we wanted to buy?" And the production team behind the studio looked to Wright's Mayan Revival designs for the—

Yes, there it is.

Company's headquarters. And so look at this. In this image, you see the Mayan tiles. They're not exactly the same, but you're like, why is this so iconic? And there it is. That's the Ennis House. That's the studio conference room. And it's just super— Fascinating to me. And this is the interior of the NSA. That's the walkway. And here is, you know, this like famous scene with Ron Howard and Seth Rogen. And they did this. Now, if they had Anton, my Lord, they could have gone even deeper on this. Bring in a faker. Agree is the number one fastest way to go from contract to cash. And right now they're offering an exclusive deal just for TWIST listeners. Go right now to agree.com, sign up, For their all-in-one contract-to-cash workflow. And Agree is going to give you 50% off for life. That's right, 50% off for life. If your company makes money through contracts and invoices, you already know it can be painful, time-wasting with all these steps. You're chasing signatures, waiting to get contracts finalized before you can send your first bill. You're regenerating fresh invoices for every new cycle. You're waiting days for clients to get around. To paying you. And that's if everything goes right the first time. Agree moves all of these tasks into a single platform so you can get everything signed, sealed, delivered, start billing, and finally get paid. If you want to stop chasing invoices all day, go to agree.com. And then if you tell them J. Cowell sent you, they're gonna give you 50% off for life. Do you, do you have like an outreach program to designers to kind of get in here and I think saving the mood boards, you know, you call them something else, but I'll call them mood boards so the audience understands.

That's easier.

The canvases, the mood board canvases, like this feels like it could have a Pinterest-like workflow where, um, you could get designers, uh, Kelly Wurstler, a friend of mine, or any of these folks to start making canvases here. In fact, Ron, can you make sure we get this to Kelly? Sure. Um, and do an intro with Zach and Kelly because she's gone crazy on AI.

Yeah.

You know Kelly Wurstler, I'm sure.

I know someone who knows her very well. Let me put it that way.

That's how you know that person is pretty famous. I know Kelly. Anyway, you know, she has her own studio and, you know, like this whole warehouse and everything, and she's just obsessed with AI. I think you should get some of these, like, iconic folks as advisors to the company, have them start making their own mood boards. And I think taking the next step, which is I want to make this product in the world. That's like another interesting angle. Like, what, what can't I find that I could manufacture? So there are all these interesting people who like to manufacture products. So this— what this could do for furniture, for decor is just unlimited. I— you, you really have built something between Houzz and Pinterest that Is more powerful than both of them. And that would create a sharing. Like if I could share, can you share the canvases so I can share one?

Yeah, they're, they're starting to be public. We, we, uh, we're, we're going down that exact same line of thinking.

Um, what I like about that one, Zach, by the way, is it creates a viral loop. And, you know, this is when you go to market, Lon, one of the tactical, practical things that founders sometimes forget. Which is, what if one person who uses my service introduces it to 1.5 people or 2.6 people and then gets in a search engine? Oh my Lord. Then the site starts growing while you're asleep.

I think one other thing I would say about that though, is while it does help with the viral loop, I, the LLMs in these, these, uh, AI companies are cannibalizing the data that they rely on. for training. Like, uh, you see, think about how many bloggers, Wirecutter, these types of posts that used to exist in the world are disappearing because there's no incentive to write them anymore. They're just going to get sucked up immediately. If you don't build content into your product, if you don't build user-generated content into your product, you have nothing to train on or give these recommendations. Super important for e-commerce. That's why we started doing it now, but it's also super important to think about for a lot of other categories of searches.

Yeah, I just— I'm thinking about the number of pages that Google has for Anton that are— yeah, I wonder like how many pages you actually have in there. And this is super interesting. Do you get a lot of search traffic anymore, or is that game over?

You know, I don't think that— I know people are talking about AEO a lot. We invested a lot into SEO, and it's still continues to drive, uh, users over to Onton. Um, wait, what was the term you used?

AEO.

Answer engine optimization. Other people call this GEO.

GEO is the other one. Generative engine optimization. People go back and forth.

In my opinion, it's largely the same thing. They're, they're using existing search indexes and they're putting LLMs on top of it. So the things that worked for SEO kind of still work for AEO and GEO. It's almost more of a wild— it's almost more of a wild west than it used to be. But, um, you do have to think about how people query these things differently. Like, the queries aren't just keywords anymore. They're, they're much more long-form. So that's the only difference between SEO and AEO, in my opinion, on, on how you're creating content.

It's still all about Wikipedia and Reddit, it turns out. If you, if you rank well on those sites, you do well no matter what.

Exactly.

Um, it does create some signal. I did see somebody create a fake— I saw this on my social media feed. Somebody created a fake brand, a fake deodorant brand, and then they got the LLMs to think it was real.

Believe it.

Within like 3 weeks, they had ChatGPT, uh, singing the praises of a fake deodorant brand.

Yeah.

So AI slop— and what, what is that term they use, Lon, for like the moment the web went from human to—

Enshittification.

Uh, gentrification is one.

It's true, right?

That's like with products deprecating. But there's another one, like the human-created web then giving way to— I'll look at this non-human web, um, where there's a, there's a major problem on Reddit right now with AI slop going in there. And this is where we're gonna just have to make a decision of the human web versus AI slop web. Do you have thoughts on this, Zach, and what, what we need to do?

This is why we built Ontology One. Like, uh, we were— we put LLMs on top of e-commerce data and it would hallucinate hallucinate all the time and come back with products that were completely irrelevant. Uh, e-commerce data is so bad because people, it's been gamed for so long. And, uh, just imagine how that's going to compound. So that's, that's why we built Ontology. I, I really don't envy the position of a lot of the social companies right now. I know Reddit is dealing with bot attacks constantly. We're dealing with bot attacks and AI agents that are going after Onton regularly. It's, it's, it's a, it's a big pain point.

Jason, I think you may be thinking of the dead internet theory, the idea that over time—

Yeah, another one.

Bots will take over and there won't be any human-created content left to find.

I think we already hit it, but it's the, yeah, when automated accounts and AI slop outnumber the traffic and/or content.

I think you, I think you already see this most prominently in comment threads and replies where once there used to be robust conversations after almost every piece of content you could find. And now it's almost always a bunch of bots talking to each other, overwhelming—

Absolutely.

Human participation.

This is why I charge $1 a month to reply to most of my tweets. I have 2,200 people out of my 1.5 million followers, so 1% would have been—

I'm one of them.

15,000.

So I think I have 10 basis points of my follower count who pay $1 a month. I give it to charity, and I'm increasingly thinking I'm just going to give it back to people randomly. Like, I'll just pick one of the I may just pick one of the followers, 10 of the followers.

X money, you could just easily do that. That's right.

And then it's just like, okay, you're— I mean, I might get— then maybe, uh, one of these organizations is gonna come to me and say I'm, I'm creating a gambling multi-level marketing scheme because if you put a dollar in, you have a chance at $2,200 every month.

You're gonna be the Bernie Madoff of X money. Yeah.

Of the people that pay Are there still bots?

No.

Do you see—

No.

Okay.

There are MAGA lunatics who think I'm a libtard, and there are libtards who think I'm a MAGA lunatic. And I told them, you can insult me, you can flame me. If you give the dollar to charity, I'm all with it.

It's totally fine.

Uh, Zach, continued success. Everybody check out ontime.

Yeah.

Thank you.

Very cool product.

Today we're going back to space, my absolute favorite place to build. There are so many startups that are doing interesting things in our local orbit and in our solar system. I love talking to all of them, especially right now in the wake of the SpaceX IPO. I feel like a lot more people are paying attention to space, so it's a great time to sit down with the founder building something I think we're really going to need and not many people know about. So please join me in welcoming Ajeet Desaniyaka to the show. She's the CEO and co-founder over at Ashi, how you doing?

I'm really good. Thank you so much for having me here.

I'm so glad because we get to talk about a pet topic of mine that I really love, which is in-orbit or in-space refueling, which I think is really cool and really matters. But I think a lot of folks don't know what it is and why it does matter. So could you start by just giving people kind of the basics as to why this is a problem that needs to be solved now?

Yeah. Okay. So space industry is growing rapidly and we see so many space companies coming up with like OTVs, own missions, launch vehicles. And all these missions and all these spacecrafts are currently limited by the amount of fuel they can carry, which compromises on the payload, how much payload they can carry, how far, how much further they can go in space. They can't do crazy maneuvers, they can't change orbits. So that's what we are solving because it's like, just think of us like a service station here on Earth, but in space. Because when the highways were built and Like on Earth, you need fuel to go from one corner to the other side of the city, right? Like, oh, from LA to New York, you need, you need fuel. But that's exactly what we need in space as well. So to go from place to place, you need fuel. And right now, all these spacecrafts are launched by the exact amount of fuel they need for the mission. And to replace with fuel, you need to compromise on the payload, or you need to compromise on the range they can travel. So that's where we come in. And we basically give them more options. So now they can travel further, carry more payload.

Here's a nightmare scenario for any unprepared founder. You're about to close a huge enterprise deal and suddenly the legal team pipes in, what's gonna happen if you get hacked? Thankfully you are prepared and you listen to This Week in Startups and you already know about Wyze Security. Wyze Security is over 40 security engineers with elite experience at trusted companies like Apple, Uber, Brex, Robinhood. Hey, I'm an investor in 2 of the 3 of those. And here's the best part. You don't have to hire this amazing team. No, you don't need extra headcount. You can just rent them by the hour. No $400,000 CISO salary, no hefty retainer, just real operators embedded in your Slack where you're already working, joining your standups, and listening in on your actual customer calls. With Wyse Security, your startup will have the enterprise-grade security it needs to close big wins without sacrificing Your runway. Head to ysecurity.io/twist. The first 6 hours of consulting are free. That's ysecurity.io/twist.

The flexibility point is very interesting. I hadn't thought about how the fuel load would be exactly tied to the target mission, but if something goes wrong and they need more fuel, what do spacecraft do today to compensate for that crisis?

Right now, they just have to scrap that whole mission and send a new one. So yeah, it's— and it's not just about missions going wrong. Even the missions that do go well, they're currently being compromised.

So what types of fuels do we need to have in these future orbital gas stations that we're going to have? Because I know there's a variety of propellants used in space, and I'm curious which ones you think are the most important and the ones you're going to target first.

That's a really good question. So I'm going to give a bit of background. So we are doing 2 types of fuels. So we are doing storables as well as cryogenic fuels. So storables are xenon, krypton, hydrazine. And cryogenics are the ones that have to be stored at really cold temperatures, like -190 Celsius or -310 Fahrenheit. So that's, that's really cold. So that's liquid oxygen.

Oh, wow.

Yeah, that's, for an example, liquid oxygen. So if you don't store these fuels at these really cold temperatures, what happens is the temperature increases and you basically have to boil off, like, which means it, it's like water, like it boils off and then you have to vent it out. Otherwise, this— we are going to see an explosion.

So you paid all that money to get it up there, and then you have to slowly release it if you can't keep it at the proper temperatures.

That's terrible. Exactly, exactly. So that's something we cracked during YC. We went through summer of 2024, we went through YC. So me and my co-founder, we cracked that problem. We built zero boil-off technology fuel tanks. So our first idea was, uh, before we even, uh, got more customer demand, we were solving for cryogenic. We were solving to store and transfer cryogenic fuel because that's what the initial demand we got from our customers. So that's why we solved the zero boil-off technology. And then we got reached out by other customers who wanted like water or xenon. And so then we decided, okay, now we have to be more fuel agnostic. So we are gonna do stowables as well as cryogenic fuel. So our initial customers or initial stations will have xenon. And then, but those customers are also moving into cryogenics in the long term. So, so yeah, we are fuel agnostic.

So you're going kind of both ways then, both the fuel you have to keep cold and the fuel you don't, simply because customers want both. Do cryogenic— do fuels that need to be kept cold have a different use than fuels that are less temperature sensitive? Are they used for longer flights or shorter bursts? How does that break down?

Yeah, so it also depends on what our customers, the other spacecraft, their mission is, right? So with cryogenic fuels, you get more energy. So that's the basic idea. So for an example, if you want to go to the moon, they mostly use cryogenic fuel. So we have customers who's asking for like 10 metric tons of cryogenic fuel because of their moon mission.

10 metric tons?

Yeah, so because of their moon missions.

Just to be clear, Rashi, we're talking about bringing all this up from the planet's surface to orbit, and then you're going to provide it, but we're not manufacturing anything in space yet at that quantity?

No, no, no.

Wow, that's a lot of weight.

That is a lot of weight. That is a lot of weight.

Yes.

Okay.

Before we get too far into the weight point, I want to ask about electronic propulsion because I think probably folks listening to this have heard about the use of electronic propulsion to do some in-orbit maneuvers. So why does that fail to meet the needs of customers that are coming to you for cooled or just kind of gaseous liquids?

Yeah. So electric propulsion, like we think electric propulsion can also really benefit from refueling. So it's not like, oh, electric propulsion is not going to work.

It definitely works, but they will benefit from refueling.

So electric, chemical, even nuclear propulsion will benefit from refueling because NASA is building nuclear propulsion and we have spoken to some, like some companies and who thinks they will need refueling even for their nuclear propulsion systems as well. So it's not just one type of propulsion system won't work, it's just they can get more benefit from being able to refuel.

Let's talk about launch costs because I was talking to the folks over at StarCloud, a very cool company. They want to build in-orbit data centers. And the CEO was like, look, our entire company is a bet on launch costs coming down. If Starship and New Glenn don't work, then we don't work because the economics don't pencil out. I'm curious about what needs to happen to make it feasible to bring up those 10 metric tons that you mentioned earlier of fuel, because I think right now, if we did it all with Falcon 9s, it would take a while just to get that one set of material up to orbit.

Yeah.

Do you guys also depend as much as other companies on heavier launch vehicles coming to market?

No, not really, because our initial customers are asking for like 2 metric tons, 5 metric tons. So we have that capability currently to actually send the amount of fuel we need. But yes, when we have to like scale up more, like 10 to 30 metric tons of fuel, yes, then we will have to depend on heavier launch vehicles where we can send more payload to space. But we are not depending on Starship or New Glenn for our business plan to work because right now everybody launches. And where we actually give value to our customers, how much they can do once they go to space, once they're in orbit. So that's when we actually give them value. They can go from LEO to GEO. I watched your session with Tom Mueller where he mentioned like for their Helios, they could, if there was a refueling station, they could get more value of the Helios from GEO to GEO. That's exactly what we are providing. So we don't really depend on the launch vehicles, like launch vehicle cost.

Well, but it does improve your economics. I presume that if we had lower launch costs, then it's a tailwind, not a glass wall.

Exactly. Because at the moment, like, we all pay the same amount for the launch. And also our technology allows us to store for longer periods. to store fuel in space for long periods of time.

And that's your zero boil-off.

Zero boil-off.

Which you invented during YC.

Yeah, exactly. And also we can store, like, for xenon as well, we can store for longer periods as well.

That's fantastic. Okay, so quickly before we get into what you've built and how it's performing up in space, to me this sounds like a piece of work that will apply equally to both commercial and defense applications. One, is that correct? And 2, Where are you seeing the most early demand? You mentioned customers earlier. I'm curious where they kind of fall on the commercial defense continuum.

Yeah, so we see demand from both sides. So we don't just see the demand from the defense side. We have like all our customers right now are commercial customers.

Oh, interesting.

Yeah, because like YC says, build something people want. So we actually wanted to make sure we have people who's going to buy our product. So we iterated so much with what our customers want. So we would build something and then we would always get feedback, and then if they don't like it, we would change it. So, so we got like, over the last 2 years, we got so much feedback in terms of hardware, in terms of type of fuel.

Yeah.

In terms of how our hardware should dock to their spacecraft, or how, like, how much risk they're willing to take.

I'm blown away by how fast your company has moved. You went through YC in— oh, I forget. Was it 2024?

Yeah, summer.

Yeah. So summer of '04, and you've already put something into space. You've already done iterations, invented technology. When I was growing up, space was slow and hard, and it took forever to do anything. And you're thinking about multi-decade NASA contracts and cool stuff. Everyone loves the James Webb Space Telescope, but oh God, did that take a while. So what's changed that allows And I say this with respect, a scrappy startup like Spacium to do so much so quickly?

Yeah. So me and my co-founder, we love space. We are like, we, that's what we live for, for Spacium, to send payloads to space and to build this refilling station. So we built our first payload in 5 months, just the two of us in 5 months. And it was the most accurate robotic actuator ever tested in space. The way we did that was obviously obsession was a big part of it. Like, because we never felt like we are working. It was just every day doing something we love. But, and also we brought all the testing in-house, so our iteration cycle was really quick compared to how space industry usually works. So we brought all the testing in-house. We do 98% of the testing in-house. So, so we build something, we test, and if we test it until it breaks, and then we improve on it. So that's why we were able to do so much with so little time.

When it comes to the fast iteration of building something, testing it, improving it, building it again, we're going to talk about your robotic actuator in a second, but this is a piece of physical technology. So do you guys also have in-house milling machines and all the necessary tools to quickly prototype and iterate?

Yeah, we do. We have T-wax, we have vibration beds, we have—

And you can do all that With 2 people?

We did that with 2 people for the first mission, but then for our next missions coming up, we are growing. Now we are a growing team, so it's not 2 people anymore.

I was going to say, like, I mean, someone has to like empty the trash cans, do the payroll. Like, I mean, there's a lot of work to do just to keep the business from falling off a cliff.

Yeah, yeah.

Let alone do it all.

Yeah, it's not anymore. So we— it was— we had like— we have so many fun stories about what happened during our first mission because it was just the 2 of us and it was like crazy. We have one of our advisors, he was one of the early employees at SpaceX. And at one point, because we went quiet, like we did not reply to him, and then he was so worried something happened to us because it was the two of us building payloads. And he was like, where's Ashi and Reza? Like, we don't know where you guys are. And then, but we were fine, obviously. Glad to hear it. We were fine. But, um, and then after the first mission, we started hiring. Now we have Amazing engineers who's also equally obsessed.

How big is the team now?

We are 7 people. We also like to keep the team lean because it helps us move faster because then we don't have a lot of bureaucracy and we don't have like, everybody owns their piece in the puzzle and then moves forward.

Yeah, no, I've heard that from software entrepreneurs, robotics entrepreneurs, space entrepreneurs. It's like every single person is like, we want to have the minimum viable team accelerated with as much technology as we can. So that makes good sense. Let's talk about your mission though. So, uh, you were on a rocket last December, and then this February you announced that you, quote, flew the most precise robotic actuator ever tested in orbit. So what does that mean and why does it matter for in-space refueling?

So robotic actuator is what goes into our robotic arm. So with that precision of the actuator, and exactly end of the robotic arm will be We'll have a precision within 0.5 millimeters. So with that type of precision, we can dock to the incoming spacecraft so precisely, and it reduces the risk for all our customers. And then we start the refilling process. So that's why that type of precision is so important, because if we can't dock to the incoming spacecraft precisely, uh, it's going to be a huge disaster, basically.

Are the tolerances that fine for docking onto a spacecraft?

Okay. Oh yes.

Okay, why? And this is just pure ignorance leaking through here, but it doesn't make sense to me that it has to be that precise, given that we're talking about a thing and then, you know, like, why can't it just fit once it's attached?

What do you mean, why can't it fit?

Well, you know, go back to Apollo 13, the movie, right? When they're trying to put the thing into the connector port, it kind of missed and kind of hit and then it would slid in. Why isn't there a little more tolerance to the linking of your future space station refueling facility and the spacecraft? Like, what drives that lack of tolerance?

Yeah, so if you can't dock precisely, it really increases the risk for our customer. And that's not a risk our customers are willing to take. So this is another feedback that we got from our customers.

So are spacecraft just less tough and durable than I'm imagining? And therefore you have to treat them very, very carefully and therefore the tolerances are so fine?

That's also a funny part because we build cars to like be really robust, right?

Yeah.

But we build spacecraft like they're like little babies that we can't even like do anything like if something goes wrong. And also in space, we don't have humans on board. So if something goes wrong, we don't have— usually we don't get a second chance to fix that. And so that's why we want to make sure everything goes Smoothly the first time, very less risk, because with risk, like, it's like a chain of events, like more things can go wrong.

Okay, so you got your actuator up in space, you showed that it works. What is next for the company as you work towards building your first actual facility that can accept fuel and disperse it?

So we have a xenon fuel transfer. This is the first time I'm saying this out loud. So yeah, you guys are the first to know. So—

Thank you. We appreciate that.

Yeah, of course. We have a mission coming up really soon, within the next year or so. So we'll be transferring xenon in orbit. And then we have another mission coming up within the next 1.5 to 2 years where we have a free-flying spacecraft and we'll be docking into another spacecraft and we're also transferring xenon in orbit. So those 2 missions are coming up and we are fully funded for those as of right now.

Fantastic. Now, Demand for in-space refueling is not a hypothetical, it sounds like. It sounds like this is a very acute need now, essentially. Okay, cool. Because sometimes you build ahead of the market, sometimes you chase, you know, what's needed. But it sounds like the moment you can get this up and running, you're going to have business.

Exactly. I have heard from like bigger companies, they're waiting until someone actually puts these stations up there. And then they will start building their technology around refueling. But so it's just kind of like chicken and egg problem. And also we do have customers who spring and take that risk and actually get fuel.

I'm really excited once we have this technology. I think it's gonna make everything so much easier and less unfault tolerant. And then hopefully we can build more robust things and do more stuff and get back to the moon. And then I can go to space too. It's gonna be fantastic. So I did talk to a company called OrbitFab a while back, and they were talking about their kind of like standardized enterprise refueling interface, which I think is kind of the connector port. So I'm curious, what's the SpaceTeam approach to connectivity to other spacecraft? Is that a difficult thing, or am I over-indexing on a bit of news that I just happened to remember before we jumped on the chat?

No, that's, that's a really good question. So that's something I would love to talk, but there will be some really groundbreaking news coming up About standardization really soon. So just keep an eye out. And, uh, yeah.

Give me a hint about really soon. Like, should I have my eyes peeled in the next couple of weeks, or is this like a Q4 thing?

Oh no, like within the next month or two.

Okay, okay, cool. Well, we'll have to have you back on to talk about that because I'm really curious. Because I mean, like, um, I hope that there's eventually a USB-C equivalent for in-space refueling so that way it's super simple, super quick, and everyone can use the same thing to avoid I mean, just an unnecessary plethora of dongles in space to do the connections that we need. It seems simpler to have one standard. So I'm hopeful on that front.

Yeah, standardization is really important. And also, it's not just about standardization, it's about how that standardization works and adapts to the current spacecraft or future spacecraft. So yeah, keep your eyes out for it.

We were talking at the top of the show about how once we have the ability to refill spacecraft, they can do more. They can make more adjustments to their course or change their mission or just be more flexible. And right now, if you have to send up a spacecraft with all the fuel it's ever going to have, that impacts design choices. Once you take away that requirement, how quickly will we see spacecraft design change to take advantage of what you guys are building and also just to have more, I don't know, longevity per spacecraft?

Oh yeah, we'll see that change coming within the next 3 to 5 years. Because when we started Spacium, we had to tell space companies, okay, if you refill, this is what you get. But now I don't have to explain to them anymore. Like, they already know, and people are already trying to build their spacecraft around refilling, thinking that, okay, there will be refilling in the future. Okay, if like— because it's such a constraint.

Yeah.

And with all that mass you want to put into space and with all this, uh, Things that we want to do in space without refueling is such a bottleneck. So people have actually started developing their space missions around refueling.

Okay. Now you mentioned that your next couple of missions are already fully funded. I know you guys raised, I think it was a $6.3 million round last year in 4 weeks. That's impressive. Are you guys going to need to access more capital to finish those missions? And have you already kind of apportioned that, or does that initial round take you all the way through your next few missions?

Yeah, we actually raised more money after that seed round. It wasn't— so we have raised close to like $12, $13 million so far. So that covers all our 3 missions, the one that which we already flew and the next 2 missions as well. After that, we are planning to like, um, come into like more profitability soon because we do have commercial contracts, uh, close to $100 million, and we have $2 billion, $2+ billion in LOIs. So all those, once all those commercial contracts convert, We'll be able to be profitable.

Yeah, it's kind of rare that a company comes and says, yeah, we have $2 billion in LOIs for our technology that we're still building. I think that is the definition of building something people want. Now, thinking about the broader venture ecosystem, if you look around technology today, it's kind of all AI, and that's absorbing a whole bunch of capital. Is there still enough investors out there focused on space to keep the space startup industry healthy in the near term?

Yeah, I think so, because space is really hot right now. I have been reached out by multiple investors wanting to talk and better understand our technology. So I think investors are really still interested in space industry as a whole.

And then last question is just, you know, pretty simple. SpaceX went public recently, mentioned this at the top. I'm curious if that has changed the level of excitement and enthusiasm around investing in space-focused companies? Because I think a lot of people that weren't paying much attention and kind of forgotten, you know, the fast launch cadence of Falcon 9s and all that are now once again paying more attention. And I wonder if that's creating positive momentum for the space industry.

It definitely does. Yes, I would say that. Yep.

And how does that manifest?

Because I think a lot of people started giving more attention to the space industry after SpaceX IPO. They're like, oh, space industry. And then they became so interested. Okay. There's actually, there's an industry that we should pay more attention to because, uh, it wasn't— space industry was there, but people weren't really paying attention until SpaceX IPO. And now people are like, okay, space industry is something we should actually talk about. There's so many cool things happening. There's all these new technologies coming up. And end of the day, space industry will benefit life on Earth. So once people understand that, People will pay more attention. But I have, I have been reached out by a lot of people wanting to talk about space industry after SpaceX IPO.

That makes me so happy here. I feel like, uh, it's always good when an IPO can kind of shine light on an entire industry and help everyone at once. Yeah, actually, Ashi, before I let you go, uh, you are building this company in San Francisco, and I tend to hear about companies building things that fly and zip around building in Southern California. So can you just tell me why SF or the Bay Area kind of writ large is the right place to build Spacium?

It's because after YC, we realized SF is where we want to be because the energy is crazy. Everybody wants to build something huge. And that's the— and at Spacium, we are weird. So we want to be in a place where there's a lot of people who want to be like really crazy things. And that really, uh, and so, so we really felt like SF is where we belong. Um, so, so yeah, we made that decision. We want to be in SF.

I, I love San Francisco and I miss my city. But what about, what about you guys is, is quote unquote weird? Because I feel like everything we've talked about is within the, the normal envelope of like technology. So are you weird in a space startup sense, or are you guys a different type of company in startups, uh, as a whole?

I think it's a little bit of both. It's just how we test things, how fast we move. And even like when we interview people, we are never looking for technical knowledge. We are looking for some type of technical knowledge, but we're always looking for that mindset who wants to— who are obsessed, who wants to just work from 5 AM to like 8 PM. So that's when we actually start working. Like every morning we start at like 5 to 6 AM and then, and then everybody just grinding. But it's such a fun work environment that, that's the, that's the cool part. Like sometimes I have to tell people like, please go home. They're like, oh no, I want to finish this. I'll go home after. I'm like, no, please go home. So, so it's not like I'm forcing anybody to stay up late.

No.

Everybody wants to stay late. So, so weird in the sense it's How everything we build is, we always find a way to, we always find a hack to make it cheaper or how to move faster.

I really appreciate that. I think the hours you work sound incredible. I just, I have tiny babies. And so I cannot do that. But I'm always very jealous of people that can devote that much of their time to their core projects. I feel like I get get 60% as much done each day as I need to, and it compounds over time. So now I'm about 8 years behind. Anyways, Ajay, thank you so much for coming on. What's the URL so people can go find out about the company? And are there any roles you're hiring for that you would love to shout out to our audience?

Yeah, we are looking for software engineers. So if anybody who's obsessed and wanna build really cool stuff that can send to space, yes, please reach out.

And it's spaceium.com, S-P-A-C-E-I-U-M dot com.

Yes.

So check it out, everybody.

All right.

And when you announce your next mission or two, we'd love to have you back on, Ashley. But in the meantime, go forth, kick ass, and let's go to the stars.

All right.

Thank you so much, Alex. It was really nice talking to you. Thank you so much for having me.