SPEAKER_00: All right, guys, welcome back to Twist. This is Alex. AI is absolutely everywhere. We have been talking about it in terms of new models, faster reasoning, who has the best inference engine, the list goes on. But AI is not just about the core technology. When we think about AI, we have to think about how we're going to power the future and how AI is going to change business today. So we have two interviews today with amazing Twist 500 companies. The first one is with Zap Energy. Now, in this case, CEO Benj Conway and I talk about why they're pursuing Z-Pinch technology in their approach to fusion and how that's going to get us all the way to commercialization. I'm very excited about fusion technology, not just in the AI context. And it's chats like this, they really give me the confidence to say, hey, we are going to figure this out. Then we're talking to Poly AI and its CEO, Nikola Mirstic. Now we break down how AI powered voice assistants are taking on the customer service game and are changing it very, very quickly. So if you care about how AI actually interacts with the real world, how it impacts jobs, how it impacts workflows, that's the chat for you. We're going to start with our Zap Energy interview. So here's my chat with Benj Conway, all about fusion and why it's coming

SPEAKER_01: sooner than you think. This week in startups is brought to you by Oracle. Oracle Cloud Infrastructure or OCI is a single platform for your infrastructure, database, application development and AI needs. Save up to 50% on your cloud bill at oracle.com slash twist. Squarespace. Turn your idea into a new website. Go to squarespace.com slash twist for a free trial. When you're ready to launch, use offer code twist to save 10% off your first purchase of a website or domain. And Atlassian. From MVP to IPO, Atlassian for startups provides your team the right tools to plan, track and collaborate on work. Head to Atlassian.com slash software slash startups to see if you qualify

SPEAKER_02: for 50 free seats for 12 months. If you've been listening to twist for the last couple of months,

SPEAKER_03: you've heard more about AI than I'm probably sure you ever wanted to. And one theme in that conversation has been enormous power demands. Because AI compute requires so much power, people are talking about massive solar arrays, demothballing nuclear reactors. There's a lot of fun approaches out there. But one that I'm personally most excited about is the progress being made by commercial fusion companies. We have a couple of these on our twist 500 list. But one that I am most excited about is a company called Zap Energy. Now I am not a physicist, have some in my family, but I was the family idiot. So I have brought their CEO and co-founder Benj Conway onto the show today to help explain to me how their approach to fusion works and how close we are, hopefully very close to seeing commercially viable fusion power. So please welcome Benj to the show.

SPEAKER_05: Benj, how are you? Well, how are you? Nice to see you. Thanks for having me on.

SPEAKER_03: I'm good because I'm excited because prepping for this chat, I learned more about the different approaches to fusion than I ever thought I was going to be able to. And frankly, I think what Zap is up to is fantastic and very exciting. But for folks out there who are a little bit less read in Benj, I was hoping we could start with just something basic. How does, from a very high level,

SPEAKER_11: the toroidalnaya kamramagnetimi katushkami, which is a toroidal chamber surrounded by magnetic coils, the idea being that you have big magnets to confine and compress plasma. And then bucket number two, which is build a big laser. And your listeners may be familiar with the NIF experiment, the National Ignition Facility Experiment at Lawrence Livermore that achieved a scientific energy break even in late November 2022. So we use, we don't occupy either of those two buckets. We use a twist on a well-known bit of physics called the Z-pinch. So Z-pinches were understood long before nuclear fusion was understood. A couple of scientists in 1905, Pollock and Barraclough in Australia, looked at a lightning rod that had been struck by lightning in some tin mine in Southern Australia and saw that that lightning rod had been crushed down. It looked like it had been crushed down the length of that lightning rod. And that's because if you remember from high school, your right-hand grip rule, a current going through a conductor creates a magnetic field that curls like your fingers. And if you squeeze your fist, that's the force that gets exerted on that lightning rod. You sometimes see it in a science museum. You put a current or an empty can of coat, and that can of coat goes crunch under its own self-generated magnetic field. So this Z-pinch was actually the earliest approach to fusion. And then what was the secret 1950s fusion program called Project Sherwood? It was called Project Sherwood because the guy running it was called Dr. Tuck. The Z-pinch was actually the first attempt. So the idea was they would take some plasma, they would put a big electric current through it, it would create this magnetic field, and it would compress and crush just like that lightning rod or that empty can of coat. The problem was it was like taking a water balloon in your hands and doing that. You know, it hit the walls in a nanosecond. And so these Z-pinch plasmas became unstable very, very quickly. And they then abandoned this idea and built magnets and lasers for the next 70 years. What we've worked out how to do is to stabilize that water balloon. And we do that with something called shear flow. So imagine that water balloon is not static, but is now flowing. And it's flowing in a way that it's flowing faster on the outside than on the inside. When you're in fast moving traffic, we all know it's very hard to move, change lanes into fast moving traffic. So if you have this column of plasma that's flowing faster on the outside than on the inside and ever sort of concentric rings going towards the center, and then you put electric current through it, it compresses just like that lightning rod or like that empty can of coat. But when it tries to destabilize, it can't change lanes because the traffic's moving really fast. In fact, the traffic's moving really fast on all sides of it. And it compresses a bit further. And the traffic tries to change lanes, but the traffic's moving really fast. And so you end up now with a really stable compression that lasts for 10,000 times longer than that normal instability. So for now, for several microseconds, this thing is stable. And that gets to the point where you have fusion conditions. So on these two-meter devices that cost single-digit millions and that we can build really, really quickly, we're now doing fusion in a way that rivals some of the biggest experiments in the world. In fact, in the last 12 months, we published an electron temperature experiment result that really only a handful of fusion topologies in the last 70 years have achieved. We did it on something that you could literally fit in the trunk of your car.

SPEAKER_16: So going back in time, we've known about Z-Pinches for a long time. They were an initial approach to

SPEAKER_18: thinking about fusion. They went away for 70 years. You guys have brought them back and you think you are on, from what I can tell, the cusp of making them viable in a net positive power

SPEAKER_11: generation environment. So what's exciting about our approach is that the scaling is really significant. The relationship between the amount of electric current, so the amount of lightning bolt that we put through that shear flow stabilized Z-Pinches, and the fusion reaction rate is an 11th power relationship, meaning if you double the current, you 2 to the power of 11, so 2 times, 2 times, 2 times, 2,000 times the fusion reaction rate. So it's a really strong lever. So what we've been doing really successfully over the last few years is driving more current through our shear flow stabilized Z-Pinches, not by building ever-increasing complex and enormous machines that cost billions of dollars, but on that same small device where we can iterate really fast, driving more current through a shear flow stabilized Z-Pinches and increasing plasma parameters, for example, neutron yield and other things that demonstrate we're getting to a hot, dense plasma.

SPEAKER_17: So I'm going to play a clip here from you guys that shows what we're talking about, I think in a much more illustrative format than just words, but when you talk about your comparison to systems that cost

SPEAKER_18: billions of dollars, to me, what ZAP is building is kind of like the anti-ITER, the major fusion reaction over in Europe, because it just seems so much simpler, smaller, modular, and easier to tweak

SPEAKER_23: because you don't have to spend 20 years building a magnetic array that can bend space and time,

SPEAKER_03: you know? Right. Why hasn't this approach been more popular, Benj? Because it seems very logical to me to approach it in this way because there's so many advantages from my layman's perspective.

SPEAKER_11: I believe that the reason that we don't have fusion energy has got nothing to do with the science, and I can hear my 160 scientists next door suck air through their teeth as I say it, but the reason that we don't have fusion, I believe, has nothing to do with the science. It's because we've been building these billion-dollar experiments that take several years to design, several years to build, several years to commission, several years to do science. So on these 10-year time scales, we've been building these enormous devices where it's just impossible to rapidly iterate. So, you know, imagine spending a billion dollars on an iPhone prototype and building one iPhone prototype every 10 years. You know, you would never, ever achieve a commercial product. You know, to go from Windows 3 to an iPhone in 15 years, that kind of iteration just isn't possible. So it's one of the key differentiating factors of Zapp. It's our superpower. The ability to build devices with single-digit millions, so orders of magnitude cheaper, order of magnitude faster. So we can build a new device in a year, less than a year. We spend, you know, single-digit millions, not hundreds of millions or billions, which allows us to iterate really fast. The fusion electricity that we produce is going to be competitive. And I think that's one of the things that the fusion community has largely ignored over the last 70 years, which is, how much is this going to cost? And, you know, if fusion electricity or fusion heat can't compete, there's going to be one fusion power plant in the world, and kids are going to go look at it on their school field trips, and they're going to say, this is a fusion power plant. It's not going

David Friedberg: to scale. Okay. Even if you think it's a bit overhyped, AI is everywhere from self-driving cars, medicine, and just business efficiency, right? We're all using it all day long. And if it's not in your industry yet, it's obviously coming and it's coming fast. You need to get ahead of this giant wave of computing and innovation. But as you know, you're going to need a lot of compute power. You're going to need a lot of speed. So how are you going to compete without your costs spiraling out of control? Well, it's time to upgrade to the next generation of the cloud. That's Oracle Cloud Infrastructure or OCI. OCI is a blazing fast and secure platform for your infrastructure, database, application development, plus all your AI and machine learning workloads. That's key. So save up to 50% on compute and up to 80% less on networking, according to our friends at OCI. So you're going to save a pile of money. Thousands of businesses have already upgraded to OCI, including Vodafone, Thomson Reuters, and Suno AI. Right now, Oracle is offering to cut your current cloud bill in half if you move to OCI for new US customers with a minimum financial commitment. The offer ends March 31st. See if your company qualifies for this very special offer at oracle.com slash twist.

SPEAKER_11: That's oracle.com slash twist. You're about to show an image of one of our devices.

SPEAKER_17: Yeah. So this is actually one of the things that helped me really understand the difference in the zap approach. If you're on the audio version, I'll play the audio for you. If you're watching

SPEAKER_34: the video of twist today, you're going to get a better show for the next 43 seconds.

SPEAKER_36: Let's look at how zap energy creates Z-pinch fusion within an experimental core. First, a puff of gas is

SPEAKER_38: injected into a vacuum chamber. As the gas expands, an intense pulse of power ionizes the gas into a plasma. Its currents and magnetic fields cause it to accelerate down the chamber. As it comes past the inner nose cone, the plasma collapses into a thin column. The Z-pinch magnetic field powerfully compresses the plasma for a few fractions of a second, producing extreme temperatures and densities. As this happens, a wave of fusion reactions produces highly energized particles, which can be harvested to make heat and electricity. Soon after the pinch disappears, the cycle will begin again.

SPEAKER_03: So I want to double click on the, the cycle beginning again, because I know that you guys

SPEAKER_18: are working on a century, which we'll get to in a second, which is all about doing more and more of these pulses. Um, and you said they last for a microsecond. So how many pulses do you guys expect to have eventually in a minute timeframe? I'm not quite sure what the correct question there to ask is, but how frequently do you spin the cycle?

SPEAKER_11: Yeah. So what, what, what you showed there is our, our, our, our fusion core. It's what, what will power the, the, the center of our, of our power plant. But that, that's not enough. You need to surround that with something that then turns neutrons into, uh, electricity ultimately. So what, what, what you see there is, is what our R and D team are working on day in, day out, what we've done in century, which hopefully we're about to show, um, is really that first, um, uh, integration of power plant relevant technologies that you need in addition to your hot, dense plasmas. So century includes, you know, repetitive pulse power, this liquid blanket, um, uh, uh, durable electrodes, um, all integrated into something that is about the size of a double decker bus. It's the size that ultimately it'll be in the power plant, which shows you that this fusion module, uh, you know, is, is, is incredibly compact. Um, but I think century, uh, is, is probably the most important demonstration of some of these key enabling technologies around fusion that's ever been achieved by a fusion company. Sure. Um, and, and it demonstrates, I think, you know, also for that, uh, our technology is highly differentiated, but our strategy of doing plasma physics alongside systems engineering is also highly differentiated as a when I set up that I, I, and again, I'm not a, I'm, I'm one of that sort of rare, uh, fusion CEOs. He's not a nuclear scientist. I'm not a nuclear engineer. And it seemed barking mad to me that the focus for 70 years had been on plasma physics only and all of this other technology that you need to make fusion happen had basically been ignored. Um, and I think people were hoping that Siemens or Hitachi would, would, would deliver it, uh, invented it, uh, at some point at zap. We've got almost as many people focused on these key enabling technologies, um, as we have on the plasma physics side. Um, and century is a really

SPEAKER_03: incredible demonstration of, of, of those. So century is much more than just the fusion core, right? Is it

SPEAKER_18: essentially a, a working, like, for example, if you had, uh, perfected the fusion core technology,

SPEAKER_23: would the century system you're currently building function as a power plant, or is it still a, a, a portion of the pieces you eventually need to build this fusion plant you're discussing?

SPEAKER_11: So it's, it's, it's the first generation of all the pieces that we're going to need in order to do, in order to do a fusion plant. So, you know, the repetitive pulse power is century. We pulse every 10 seconds, uh, in a power plant, it needs to pulse 10 times a second. Okay. Um, and everything is, every, everything will need to be scaled up, but it really is the first, actually it's not even the first generation. It's, it's, it's sometimes the second or third generation of the pieces that are needed to put together for, for an eventual power plant. So century was showing off last October

SPEAKER_18: when you guys also announced $130 million in new capital. It's a lot of money. It's been a couple of months. Uh, how has progress been on the century project? Have you reached any new kind of key milestones that provide extra confidence in your approach? Yeah, we have, we, we, we, we hit a key

SPEAKER_11: milestone, um, that we'll be announcing soon. Um, and that meshes with, uh, uh, a DOE milestone, uh, that I don't want to get too, too far over my skis. Oh, that is such a tease. That's brutal. Okay, fine. Um, uh, and, and also on the R&D side, we, we bring on screen some really new,

SPEAKER_49: new, new, uh, exciting, uh, configurations that we think we're going to push plasma physics pretty,

SPEAKER_11: pretty, uh, successfully this year as well. So yeah, no, a lot, a lot of pro pro progress. When, when you're iterating this fast, um, it, it, it, it's really amazing how, how, how quickly you do, um, progress in all of these different programs, whether it's on the R&D side or whether it's on

SPEAKER_08: the systems engineering side. Now, $130 million, it sounds like a lot of money to folks out there

SPEAKER_18: who have $130, but in the realm of big science projects, you've mentioned the cost of some fusion installations runs into the billions. So what I don't have a good idea, Benj, is how much money is 130 million for you guys? Um, I know you've raised a little over 300 total, but does 130 get you all the way to commercial viability? Does it get you just to the end of the century project? I'm just

SPEAKER_08: not sure about scale of capital versus work to be done. We're probably the most capital efficient

SPEAKER_11: fusion company out there. Um, so $130 million is, you know, is, is, is, is worth a lot to us versus perhaps other fusion approaches. Um, are we going to need much more capital to get to a commercial product? Yeah, we're going to need, we need several billion dollars. I would have thought between now and launching a commercial product, not because the fusion component of our power plants cost billions of dollars, but because zap is, um, unique in that we'll be building our modules in, in a factory. So we'll get the sort of economies of scale that you get when you build a fusion module in a factory, a power module in a factory, some of the tritium cycle modules in a factory to then, uh, um, uh, install them, uh, on site. Whereas other topologies you're building almost like a conventional nuclear power station on, on site with zap, we are going to be building up manufacturing capabilities in order to be able to deploy and scale. And everyone knows setting up large factories

SPEAKER_18: for complex machines is incredibly easy. It doesn't cost any money at all. And you can do it overnight. Right? Yeah, exactly. So on the competition point though, there's, there's a lot of people out there who are doing cool stuff. I mean, I, I, I've been familiar with Tocomax accelerators for a while. The zap approach was new to me. I know, um, helium energy is also doing a kind of a different approach to it. Do we eventually reach a point in which someone gets there first and then kind of owns the market or are we going to see eventually several different approaches to fusion

Chamath Palihapitiya: power generation become kind of the norm around the globe? Just like we have different types of nuclear reactors. Predicting fusion is really hard. I mean, I don't, I, I, I, I can't imagine. I, I,

SPEAKER_11: I don't know another sector where so many smart people have got predictions wrong over the years. So I, I, I predict, I predict humbly, um, the, the, the fusion, the fusion future, I believe there is room for multiple players. Um, it would be like saying, well, you know, one person, uh, uh, you know, one, one group launches, uh, uh, uh, a reasoning LLM and that that's it. We just need one. Um, uh, there there's going to be, there's, there's going to be, there's going to be more. I do think that the idea that there's going to be when it comes to commercial fusion, multiple approaches, I, I, I, I, I, I suspect not, I, I suspect there's going to be one way of doing fusion in the most economical way. And that will be the version of fusion that scales. I don't think we'll be in a world where we have Tokamak fusion power plants and Shiflow Stabilize Z-Pinch power plants and stellarator power plants and laser power plants. I think there'll be, there'll be one way where this is the, the cheapest. Yeah. And that will be the technology, uh, the, the, the, the scales,

SPEAKER_03: but it's very difficult to know. So why do you have circulating liquid metal walls

SPEAKER_17: inside of the technology? Because that to me sounds like science fiction in the best possible sense as a big sci-fi guy. Um, but explain liquid metal walls to me and, and why they're important. I'm just,

SPEAKER_08: I'm just curious. Yeah. I'd love you to come and see it. Uh, it's, it's, it's really incredible.

SPEAKER_11: Um, and I think zap is, I think one of the only fusion companies has actually ever done these, these, these liquid blankets. You need, you need something in order to, um, interface with your neutrons. So, um, uh, uh, a, a liquid wall, uh, replenishes, um, you can circulate heat out of it quite easily. Uh, and so it's the blank, it's the blanket, um, and the, the material which interfaces with your neutrons, neutron output. And rather, rather than being solid, um, it's a, it's a liquid, um, which, uh, you know, has all the properties which enable that to, uh, you know, as I say, transfer heat and with complicated sort of thermal management that allows that to, to, to, to get hot

SPEAKER_49: and to, to breathe the tritium that you need in order to produce one of your fuels.

SPEAKER_64: All right, founders, let's talk about your website. Yes, I know it's embarrassing. Yes,

David Friedberg: I know you're too busy to upgrade it. Well, if you're going to launch something new or you need to give your brand a refresh, which I know you do, you need Squarespace. It's the all-in-one platform that makes building a stunning professional website ridiculously easy. Whether you're selling products, offering services, or showcasing your portfolio, like I'm going to do with one of my kids. They want to go and do artistic things in the world. So I'm going to make them a nice, beautiful Squarespace site when they apply to college. Squarespace gives you everything you need to grow. And you know what? Obviously you've heard me talk about how beautiful Squarespace's templates are, but now you can get a fully customized website in minutes with their AI tool called Blueprint. Here's how it works. You just answer a few questions and then you get a beautiful bespoke website in minutes, personalized layouts, visuals that are going to stun your customers. And voila, you're done. Want more control? Choose from award-winning templates and use intuitive drag and drop tools so you can make it your own. Easy peasy, lemon squeezy. You get to do it the way you want to do it. So here's your call to action. Just fall in love with Squarespace like I did. Squarespace.com slash twist for a free trial. And when you're ready to launch, go to squarespace.com slash twist to get 10% off your first website or domain purchase. That's squarespace.com slash twist, the longest running partner for this week in startups, because they were a startup and they still build their product with the enthusiasm and the cutting edge technology of a startup founder. We love you, Squarespace.

SPEAKER_17: And given that my knowledge of thermal management is opening my gaming PC's case to let more air in, I'm going to leave the physics to you. But just to make sure I understand this,

SPEAKER_18: when the neutrinos are fired out from the Z-pinch moment of fusion, they're collected by the liquid metal, which then heats up and then you can extract that heat from the flowing liquid metal. And then

Chamath Palihapitiya: essentially, do you take that out to do thermal power generation? Right, exactly. So neutrons, not

SPEAKER_11: neutrinos. Just an important correction. But yeah, exactly. Well, they're just different things. Exactly. No, no. Sorry. But you're exactly right. So you need to somehow get the heat out and you need to turn heat into electricity. And humans are very, very good at boiling water and spinning turbines. They're so good. And so yeah, everything from boiling water onwards looks very much like a conventional power plant. We've been doing a feasibility study on a decommissioned coal fire plant, not far from where we are in Seattle. Obviously, you burn coal, boil water, turn turbines. Fusion does fusion, boils water effectively, turns turbines. Everything from boiling water is very, very similar. So we think there may be some efficiencies in terms of scaling commercially by retrofitting some legacy energy infrastructure like the old coal fire power stations.

SPEAKER_17: Well, if all power generation is just making a tea kettle blow,

SPEAKER_23: then why not just take your fusion power and attach it to an existing tea kettle? I mean, that makes perfect sense to me. That's the rationale.

SPEAKER_18: Okay. So Benj, I have to let you go, but I'm curious, when should we sync back up with you to see the next milestones you're talking about that you can't quite share yet? Just, what's the timeline for the next big piece of Zap news?

SPEAKER_11: We'll be publishing later this year. I mean, we really believe in the peer reviewed process. We'll be doing that more this year. But I would really encourage you to come visit. We'd love you to see what we're doing in person. It'll fulfill all of your preconceptions of what a fusion startup looks like. So look, we'd love you to come visit later in the year and we'll be announcing some progress soon.

SPEAKER_08: Not an impossibility given that I grew up in Oregon and I have family on the West Coast and I got to

SPEAKER_18: go see the student reactor at Oregon State University when I was a kiddo. So I've gotten to see some of the related technologies. Benj, thank you so much. Zap Energy on the Twist 500, one of the

SPEAKER_17: coolest companies in the United States today and at the forefront of fusion energy production. Thanks, Alex. And that, my friends, is why I consider nuclear energy very important,

SPEAKER_00: but merely a stopgap on our path to fusion. I can't wait. It's going to be amazing to have unlimited clean, free energy. All right. Next up, Poly AI and its CEO, Nikola Miršić, were talking about AI, but pay a special attention to the difference between conversational AI and generative AI, how the CEO thinks about it and what that means for the market. All right, let's go.

SPEAKER_76: Hey, everybody. Welcome back to Twist. My name is Alex and we have another Twist 500 interview for

SPEAKER_18: you today. Now on the podcast, we spend a lot of time talking about models, new models, faster models, reasoning models, how new models are built, what they cost. But what might matter more is how AI is being used today in a business context. And for that reason, I'm really excited to have Poly AI on the show today. And we're going to bring up CEO and co-founder Nikola Miršić. Nikola, hey, how you doing?

SPEAKER_77: I'm doing great. Thank you for having me. And thank you for the one of the best pronunciations

SPEAKER_76: of my last name that I've ever heard. I did much better before we hit record and I kind of butchered

SPEAKER_77: the second time, but hey. No, you did great. You did a great honor of Serbian citizenship to you.

SPEAKER_79: I'll take it. My country's going through some weird times. So never bad to have one of those.

SPEAKER_77: We've always been in weird times. So yeah. What a strange time we live in. Putting aside

SPEAKER_18: geopolitical jokes though, you are in London. Thank you for being up so late to record this.

SPEAKER_80: I am. No problem at all.

SPEAKER_18: I'm excited about Poly AI because I'm a big fan of I think where AI is going, which is going to be me talking to it in both a personal and a business setting. But I thought to start before we get into exactly what your company does and its history is to define conversational AI because you guys

SPEAKER_81: differentiate that from generative AI in a very useful way. So Nicola, if you don't mind,

SPEAKER_83: can we just start there? Yeah, look, I mean, I think conversational AI

SPEAKER_77: is to do with using technology, AI or not, but really it all is AI to build conversational systems. So to build technology that will allow us to speak to machines. And then, you know, well, generative AI is just, you know, are using generative models to apply and applying

SPEAKER_84: them in different fields of work, right? So it's kind of the question there is like, are you using electricity, right? So yeah.

SPEAKER_17: So then I guess it's simpler than I thought, but I was going to ask when I do use, for example,

SPEAKER_18: just chat GPT in the consumer context, and I'm using it in a me speaking, it's speaking back to me thing, I'm using conversational AI, which is just undergirded or supported by generative AI technologies.

SPEAKER_77: Yeah, I mean, look, I would basically think of it as one is like a foundational model layer, you know, kind of like if you're thinking of computer networking, one is like a cable, right? It passes packets through it, right? And the other is more of an application layer, things or generative AI for chat GPT or for our systems, etc. is used to build a conversational

SPEAKER_83: application, right? So by and large, conversational AI is about applications and it lives in the application layer. Yes.

SPEAKER_90: And for folks who don't know, I know this is a simple question, but can you define

SPEAKER_46: application layer for folks? We use that term a lot, but I think it's good to hear from the expert exactly how you and your company define it.

SPEAKER_83: Well, I mean, I think the application layer really boils down to like,

SPEAKER_77: are you doing something useful for people, right? You know, one is just the technology, and the other one is like something that you use as an application, right? Like the same way to use an app on your smartphone or on your computer, right? So similarly, these are maybe applications used by enterprises to give them a voice on the phone that allows them to speak to

SPEAKER_95: their customers anytime of the night and to do a really good job, right? To open them up to the world and to contact.

SPEAKER_17: Well, I think anyone who's been stuck in a call waiting line to talk to a human, there's like three of them, it seems that every single major credit card company can understand

SPEAKER_18: why this might be faster and simpler and better. But let's go back and head to 2017 when you founded the company. I think a lot of people didn't think about AI in this type of context until Chad CPT came out, which was 2022 if memory serves. You guys founded the company the same year as the attention is all you need paper that kind of launched transformers and LLMs into the technology context. So I'm curious when you founded the company, did the technology exist that you needed to actually get to where you are today? Or did you found it almost in anticipation of technology having a couple of breakthroughs that would make your vision feasible in the market?

SPEAKER_77: Yeah, so I think like it's a moving target, right? Technology is advancing and we're doing increasingly impressive things. I had incredible fortune in my life path where I ended up doing a PhD at Cambridge starting in 2014 with a professor called Steve Young, one of the most cited guys in speech recognition and a believer in deep learning. So deep learning really starts kicking off around 2012 when people figure out how to, you know, Jeff Hinton, his students and others, figure out how to basically pre-train deep neural networks, right? At that point, we start getting much more powerful machine learning technology that allow you to generalize, to understand people, you know, without using one of three words to say, you know, balance check. You might be able to say, look, I want to check my balance or I want to know how much money I have in my account, right? My PhD was largely about that, how you move away from like exact matching of sentences and patterns to like what someone said into like mathematical representation of a sentence where a neural network looks at it and goes,

SPEAKER_104: well, is this the right response for this or not, right? So we pioneered a lot of that stuff.

SPEAKER_106: And now we measured by distance in vector space.

SPEAKER_104: Yeah, yeah. I mean, like similarity, right? So yeah, cosign similarities typically used to kind of like say,

SPEAKER_77: these two things are similar, and then you have different neural nets that learn to look at two things that might be in different spaces. So it might not be distance, it might be like learning a mapping between them and there's a lot in there, right? It's gotten pretty complicated in terms of how it's implemented and these things get bigger and more powerful and it's getting harder and harder to

SPEAKER_84: really see what they're doing behind the scenes, right? It's not like they're writing code, they're learning to have an intuitive understanding of, you know, is this sentence the right answer for this, right? Or once you say that sentence, what do I say next, right? Because that's really how they're

SPEAKER_109: trained to think, to reason, right? All right. If you're shipping a product or rolling out an update, you're building a company, you need to be organized, right? We know that. Atlassian has

SPEAKER_111: exactly what you need to streamline your work and smash your goals. And the Atlassian for Startups program is packed with all the tools you need, like Jira, where you can track every task, sprint, and bug. That's the industry standard. Confluence, another industry standard for team collaboration and documentation. And of course, Loom for quick video explainer creation. Now, Loom is really brilliant. My team started using Loom on their own. They started paying for it on their own. Why? They wanted to get credit on the investment team for communicating to me, the general partner of the firm, why they wanted to invest in a company. So they would do a Loom where they recorded over a recording of an interview they did with a founder or visiting their website and going through why they want to invest in a company. And this was so great for me. I would be skiing in Japan. I would be on a flight to New York to see my parents. And all of a sudden I get a notification from one of my team members. Hey, watch this Loom. And I get the link for the Loom. I click it, and then I can put comments at any time. So it's like doing a conference call, but on my time asynchronously. And I can communicate right there on the video. Also included in Atlassian for Startups is Compass, Jira Product Discovery, Bitbucket, so much more, all powered by Atlassian Intelligence. That's their built-in AI. Atlassian software helps companies like Canva, Cloudflare, and Rivian keep growing and keep innovating. Whether you're brainstorming on sticky notes or scaling to the big leads, Atlassian is here to accelerate your startup's growth. Check out Atlassian for Startups, where eligible startups get up to 50 seats for free for one full year. That is absurdly generous. Why can they be so generous at Atlassian? Because they're the standard. Atlassian is the standard, and they are generous to startups because they were once a startup. I remember meeting them 20 years ago in Australia. What a great company. Head to Atlassian.com slash startups slash twist for complete details.

SPEAKER_18: This is one of the things that impressed me about your company, because when I was learning more about the space and what you guys are up to and just how to go about it, the number of pieces that you had to put together to build your in-market product were interesting. So I'm going to attempt here a kind of breakdown of how this works in your context. And then I would love for you to correct me and tell me where I got it wrong if you're up for it. Absolutely. So ASR is automatic speech recognition. And essentially what that does is takes me talking and converts it into text that an LLM or some other kind of technology system can ingest. From there, we need natural language understanding, which takes those now text-based words and converts them into meaning. And I believe that you use an LLM for that. And then after you have an output,

SPEAKER_34: you use speech synthesis to turn that back into spoken words. That would come back to me, the customer. How'd I do?

SPEAKER_118: For kind of like the pre-LLM era, pretty well, right? Because that meaning, like whether you're

SPEAKER_77: breaking down that meaning in the middle and then writing logic to say, you asked for an Italian restaurant in, I don't know, Northwest London, like I'll go to a database and figure out what they are.

SPEAKER_104: Like, yeah, that's kind of how it worked. Kind of like when we started the company, right? Now already there, it's always been about adding data to make the whole thing work with less logic written and with more just data passing through, right? So the more data you see, and like the culmination of that really is things like modern large language models,

SPEAKER_84: where basically you don't write any logic. It's just, it flows. It predicts the next word and through that it actually in the background, it does a lot of reasoning. It's able to call APIs

SPEAKER_77: and you can kind of like train it to do a lot of these things. So increasingly now, like you're able

SPEAKER_104: to avoid a lot of that ontology-based kind of like specific use case-based encoding and you get to the point where you just have a single model. So at this point, like the end-to-end voice really stands for models that take voice in and produce voice out, subsuming all the components in one. And deep learning has always been about just doing that, being focused on creating intermediate representations that don't have to have like humans saying for this application, pull out this and then do this. And instead of just having a single thing that from data learns how to do it all,

SPEAKER_84: right? So we have, you know, my co-founders, my work during our PhDs has always been about having

SPEAKER_104: fewer components, right? So some versions of that dialogue stack have like five components, right? And really you don't have to have all that anymore, right? Now you can kind of like prompt an LN and say, hey, you're a restaurant concierge. Like if you need to find a list of restaurants, here's this function. And the function is give it a location and it'll return like the list. And when you see that list, well, you know, keep talking to the person, then you might see how it talks,

SPEAKER_18: right? So as LLMs have gotten better, you've been able to do more instead of one specific model, say. And I know you guys have the, you have a convert model. I think that's the name for it.

SPEAKER_104: So convert was our, um, kind of like previous transformer model. It was, uh, is 600 megabytes. So puny in comparison. That's insane. Yeah. I mean, like a lot of people these days are like small and small, large language models. It's like, what does that even mean? Right. Like, uh, but, um, everyone was just writing code that was kind of saying like, let's just figure out how to, you know, have these intense entities through database searches. And that stuff was a nightmare. That's how that first generation of voice IDR is built. Right. So like the one step forward was not exact matching of words and stuff like that. That happened around like 2014, 15, 16. Right. Then you got a bit more expressive, but it was still like, you know, those loops where you get stuck. You're like, no, do you want a credit card, debit card? You're like, no, I'm calling you because of my mortgage. Ah, was it debit or credit? You're like, no, no.

SPEAKER_127: That's when I take my phone and I, I throw it across the room.

SPEAKER_104: That's exactly where violence happens. Exactly. Yes. And, um, you know, really like we've, we've kind of gotten further and further away from that. Now with convert, um, and we no longer use it much. Right. Uh, you were able to just say like, look, Hey, you system, the user will say something and you can say one of these 500 things, pick the best one. Right. So it's kind of like an LLM except an LLM just generates a sentence all out flat out. Right. That's where you have hallucinations, all those things. So this was like, you know, a bit less expressive than an LLM, but very safe for the enterprise. It's a good approach. And these days, you know, we use, uh, LLMs, we tune our own, right. Uh, we have a lot of data. We have

SPEAKER_77: many customers that take, you know, millions and millions of calls with us. So we have a serious data

SPEAKER_104: mode that allows us not to be better than open AI or deep seek or who else right at like general LLMs. And we don't want to be that. Right. But it allows us to tune open source LLMs so that they're better at customer service over the fall. Similarly, that speech recognizer, we still treat separately. A lot of people hope they will be able to kind of like do these end to end models. And, you know, we're excited about it. That's ideologically where we've always been going. We're experimenting with it, but for the enterprises there, you know, we still kind of keep two separate things that we tune for them. And that's really needed to drive that performance higher. Right. And the truth is it's better than it's ever been. It's not solved. Do you know this intuitively from the fact that if you tell Alexa to set a timer for 15 minutes, I don't use Alexa as the

SPEAKER_130: state of the art for anything because Alexa is dumber than my dogs.

SPEAKER_104: Well, we can talk about that. It is. It is. But the one thing which is, I think, a good illustration is you say set a timer for 15 minutes and you're just as likely to get 50 as an outcome. You know that everyone who uses it is aware of this. Right. And that's just like stuff that has to do with speech technology. It's not at the cusp of working 100% of the time. And that's something that you must build into how you build the systems. And that's what I think

SPEAKER_84: the market's mostly these newer people working on it. They don't really get it because they just, you know, they started working on it three months ago and they think that that jump happened in three

SPEAKER_79: months. It did not. Versus it taking seven years or eight years since the Transformers paper came out.

SPEAKER_123: Transformers are not the beginning of deep learning history, right? Transformers are like, yeah.

SPEAKER_135: I'm just going back to 2017. When you founded the company and that paper came out, going back to 2012 is 12 years.

SPEAKER_77: Yeah. Yeah. Yeah. And look, I mean, that was just like, you know, there are people who would tell you that it was 90s. And then there was one winter and another, and we might yet slow down,

SPEAKER_104: but I think society as a whole has not put so many great minds and resources into AI that I think

SPEAKER_137: we're, we're kind of like all in now. So it's going to take a lot for us to slow down.

SPEAKER_34: Well, what matters in the business context is that probably AI has been in the market now for some

SPEAKER_18: time. And you guys raised a series C last year that was quite large. I think it was a $50 million round. And I know that a coast load was in there and I think Nvidia's venture capital arm was in there. Two questions about the state of things. One, how fast is your technology improving? We talked about the market itself overall, but I'm curious how much better you guys are getting at the kind of in-market task of handling customer support calls for your customers. And two, how quickly is the business itself growing since you raised that hefty sizeable series C last year?

SPEAKER_103: The technology is improving really fast, right? And that's to do like getting the opportunity to

SPEAKER_77: work on really important big columns for very large companies, right? If someone trusts us with, you know, 20, 30, 40 million of their annual calls, like we take it pretty seriously. And, you know, we've been building up to the challenge for years, and there's a lot to handling that, right? With some of

SPEAKER_104: these companies, we're able to do 90% of the calls in a fully automated way with like a customer satisfaction score at comparable and sometimes even better levels than humans, right?

SPEAKER_76: Nicola, that 90% number, what was that three years ago for Polyae?

SPEAKER_104: Yeah, it's hard to talk about it generally, like, but you know, we've had three years ago, we had like maybe the highest one might have been in the low 70s.

SPEAKER_34: Okay, so you've gone from pretty darn good to almost all of them. That's, that's probably harder than going from 0% to 20%.

SPEAKER_77: Yeah, but I'll tell you something else about the business context that I think anyone listening and thinking of applying this needs to know, right? Like, if like,

SPEAKER_104: the LLMs right now are outstanding, they're really good, they might improve still, right? Yeah. But the reason that we don't see like a 20% annual GDP growth of all countries that have access to it, is because we need to connect it to the world in meaningful ways. You need to find a conduit for it to excel at what it does and produce value from it, right? So for us with enterprises, it's really about like helping them navigate their transformation of their processes and to work with people of their contact center, so they can reap the benefits. It's really hard for a company to structure their processes in ways that they're comfortable with enough to then entrust it to AI. That's a journey that we help them navigate. That's as difficult as implementing the AI itself, right? Because it involves their IT teams, their people leaders, their contact center leaders,

SPEAKER_103: their chief operating officer, risk officer, brand, everything, right? Companies doing this are still, I think, early adopters.

SPEAKER_18: But one thing I've also seen is that there are other companies working in the broader voice AI or conversational AI space that are also raising a lot of money and making a lot of noise. 11 labs just raised a bunch of money, play AI as well. And to me, I read that as investors noticing the momentum that you're describing in your business and the market itself and just placing multiple bets, because it's going to be a market big enough for multiple winners. But I'm just curious from your perspective as the leader, how much competitive pressure are you seeing in the market when you go out to try to land an account? Do these other names show up? Is it still kind of a greenfield opportunity?

SPEAKER_77: You see everyone, right? I think there's like the recent style was that 30 to 40% of all white combinator companies are working on voice agents. So this is the new gold rush, right? And I feel extremely privileged with kind of like when we started how we started everything we've done, right? Because unlike a lot of those companies, we have road stroke clients, we have a bunch of people in every department, right? Yeah. And just the cashier and the data that allows us to have,

SPEAKER_104: you know, we have full model strategic autonomy, we're not a rapper, right? So do we see competition? Hell yeah, more than ever. That market map came out. It's like 20 different verticals like, you know, healthcare, restaurants, hospitality, travel and logistics, financial services, obviously. In all of those, I think we probably have more revenue than most of these new challengers times probably like five to 10, right? That's a formidable advantage. But it's also like

SPEAKER_77: something to, you know, I think, you know, you think of all the companies that have gone the way of old favorites, you know, your Netscape sales, others, right? You know, I don't think we can get arrogant here, right? I think there's a lot to do. We fight Google left and right, we fight like hyperscalers, there's Seacast companies deploying things, it's a gold rush,

SPEAKER_104: because everyone understands that, you know, between North America and Europe, there's a trillion dollars a year of labor costs in this. And yeah, listen to this, people don't want to do those jobs.

SPEAKER_76: That that's, that's why in my notes, I don't have Nicola, what are you going to do with all the people you put out of business because calls into jobs? I don't think I've ever put anyone out

SPEAKER_18: of business. Oh, I'm sorry, that's not what I meant. Um, what are you going to do to people that you displace who are doing these rote jobs? But I just don't think we need to have humans answering these questions for other humans. It just it's a waste of human potential in my view.

SPEAKER_151: Yeah, no, no, no, no, I heard you. I mean, like the, the real thing is, I don't think that we ever

SPEAKER_104: had like large scale deployment where we implemented this, where someone was like happy with what they're spending on the contact center when they had enough people. Like that is a fictional noun at this point, right? Right. And there isn't really like this, you know, proverbial private equity backed CFO who's like, haha, I'm gonna fire 1000 people. That doesn't happen. Right? Because people who don't have a problem don't tend to go and fire 1000 people. Right? Yeah. And if they have to fire 1000 people, that company has already let those people go, and their service levels have gone to hell. Right? So now they need technology to dig themselves out of the grave. Right? So that's really the business we're in. The only people we hope to put out of business are all our competitors, right? But yeah, but yeah, no, it's, it's a really fun time. But the other thing that's really important to say is that we're creating a whole new generation of highly paid knowledge workers in the contact center, right? And many ways we think about it, like we're turning the contact center into a command center, right? Because where you previously had, you know, this department that is there as a bandaid for failure management, and you do something wrong as a business, you roll out a product that needs to be recalled, people call the contact center, your product doesn't work, contact center, you overcharge someone in your back office, they call the contact center, and everyone points a finger at them. And they're like, hey, it's your fault. You didn't pick up the phone, you're not working hard enough. Oh, trust me, I've spent a lot of time contact centers, the number of empathetic people that are doing their best, while one after another, callers are being incredibly bad to them. They're terrible. Like the,

SPEAKER_84: I don't know why people think that being mean will get them something better. They're, they're think it's going to get them a refund or whatever. People behave terribly on average. Like, I don't think the contact center leadership talks enough about what their people are going through, right?

SPEAKER_18: Well, I don't think they want, they don't want to admit it, they just want to put that onto someone else and pay them, you know, 14 bucks an hour. This is by the way, if you're listening to this, here's a hack. The next time you talk to a person on the phone, who's helping you with your

SPEAKER_17: credit card miles or whatever, be nice. Because, because one selfishly, they'll go to bat for you. But two, they didn't do it. Whatever your problem is, they didn't cause it. Be nice to them.

SPEAKER_18: One last tiny question. And then I promise I'll let you go. There's a lot of money sloshing around the world of AI. And you mentioned all the YC companies that are out there trying to kind of follow in your footsteps. Um, how often do someone try to acquire poly AI? Yeah, we've said no many times. Is it more frequent that people show up and try recently? I'm trying to get an idea essentially for what's the state of startup M&A in the realm of applied AI

SPEAKER_102: today? Uh, well, I think like post-election stuff, the inbounds and stuff have increased.

SPEAKER_77: Okay. Um, and I think that in general, the state of the markets is such that I think a lot more is expected, right? I don't like, I mean, the previous period was pretty quiet, right? I think when we started, it was left and right, several companies were really, really keen. And that was like the

SPEAKER_104: first golden age of like deep learning acquisires. And we are, have a pretty strong team. So that was no surprise. Right. Then kind of like when COVID hit and like the whole thing, like kind of like afterwards slowed down, but we were busy building. Like I didn't even like respond to those emails. Right. And I still don't, right. Cause I think that what we're doing is like really,

SPEAKER_84: really, really, really exciting. Right. I would all these companies like don't want to buy you,

SPEAKER_17: you know, you want to park there. Let's let's do a deal. Well, okay. Um, whatever your next ARR

SPEAKER_18: milestone is 25, 50, a hundred, whatever it is, come back on the show, tell me about it. And I want to ask that question again to see if you've turned the tables and are you now going to buy other companies, but in the meantime, Nicola, thank you for coming by. We appreciate it. And, um, just hit us up with all your news. Cause twist is always here for more startup stuff. Appreciate it.

SPEAKER_00: Yeah. Thanks for having me. It was really fun. I love talking to founders and that's why I'm so glad we're doing this twist 500 project. If you don't know the twist 500, which is at twist500.com is a list of the 500 most important and potentially most lucrative private market companies out there. Essentially the goal is we want to find the top 1% of startups and then talk to them. So these interviews are going to keep coming as we add companies to the list. We're in the back half of building out the twist 500 now to expect a lot more to come. In the meantime, twist does go live Monday, Wednesday, Friday at about noon central, 1 PM Eastern over on YouTube, LinkedIn, and every other social media application you can name. We're also out there on podcast platforms wide and far. And if you want even more from me, well, I write over at cautiousoptimism.news. But in the meantime, I'll see you on Monday. Bye everybody.