Hey Siri, who made you?

0:00Hi, everyone. It's Rana. The future of AI is not built in isolation. It is shaped in rooms where ambitious people challenge each other. Masters of Scale Summit is back October 20th through 22nd in San Francisco, bringing together founders and innovators pushing the edge of what is possible. If you're building what comes next, you should be in this room. Apply now at mastersofscale.com slash pioneers. That's mastersofscale.com slash pioneers. On Pioneers of AI, we often talk about the responsibility that comes with building powerful technology. AI isn't just one thing. It's a force reshaping medicine, education, science, and the stories we tell on screen.

0:48And for that future to work, it needs infrastructure that's built thoughtfully and deliberately. As the essential cloud for AI, CoreWeave empowers pioneers at leading AI labs and enterprises with a purpose-built cloud made specifically to train and run the world's most advanced models. They help accelerate the breakthroughs that matter at scale. The big ideas, the moonshots, the futures we haven't imagined yet. Ready for anything, ready for AI. To learn more about how CoreWeave powers the world's best AI, go to coreweave.com slash ready for anything.

1:46all running on consumer GPUs. With over 3 million downloads on Hugging Face, LTX2 gives you full model weights, training frameworks, and evaluation tools to build real production workflows. Don't settle for a first draft. Get the complete creative engine to reach your idea's full potential. Try LTX2 today at ltx.io slash model.

2:18I love magic as just a creative, inventive, scientific discipline. I've been a magician from age 10 to 14. I did nothing for 35 years, but then my son became a teen, and it was something that we could do together. If you're like Adam Chire and lucky enough to find a way to bond with your teen, you hold on to it. But what started as a fun hobby grew into something a lot bigger. We'd perform at fundraiser parties, at birthday parties. Now I can say I've performed on all the major TV shows. Shows like Penn & Teller's Fool Us. Once you have an idea and a prototype, you need investors to fund your company.

3:10Penn & Teller, would you join me on stage? You can be the ones I'm trying to request. Adam Moonlights is a magician, but what he's really known for is no less magical.

3:25He's one of the masterminds behind Siri. But since creating the now ubiquitous voice assistant, he's gone on to found other companies like Change.org and Game Planner. I see huge parallels between entrepreneurship and magic. because if you think about it, when you're an entrepreneur, you have to have a vision for an impossible future. You're not going to start a new company to create something that exists already. You have to reach far and imagine, what if this could become true? And then you work backwards and you figure out the math and science to make it a reality. I have known Adam for over a decade, and I find all of my conversations with him to be truly inspiring.

4:15And this conversation was no different. Adam reflects on his entrepreneurial journey and reveals his bold prediction on the next big innovation that will reshape the world. But my favorite part of this interview is how he manifests change in his life and how you can too. And definitely stick around until the end of this episode because you might just hear from Siri herself. I'm Rana El-Khalyubi, and this is Pioneers of AI, a podcast taking you behind the scenes of the AI revolution.

4:58Hi, Adam. Welcome to Pioneers of AI. I'm so excited for our conversation. Thank you. Me too. So we go way back. I think we first met almost like a decade and a half ago in Palo Alto because we shared common investors. And I have always been a big fan of you and followed your journey for many years. So, yeah, I'm just excited that we get to do this. Absolutely. It's a small world in the tech world, and there's not that many people who've been working on AI for 15 years, not just the last five. So I think some of the early ones. Yes, we're the OGs. So you started prototyping Siri in the 90s, right?

5:39So just walk us quickly through that history and how that all culminated into you starting Siri. Yeah. So I guess 33 years ago, everyone had a PC on their desk and you would load it up with software from CD-ROMs and floppy disks and you'd put the software on your computer and that's what you use to compute. There was no web yet in 1993, no web browser. And I said, someday there will be content and services spread around computers all around the world, different providers who have content and services to offer, and we'll need a way to discover them and to interact with them. And I never imagined a hyperlinked multimedia document framework as the way to do that.

6:27I thought everyone would have an assistant and you could just say, I want to know this or I want to do that or some combination of the two. And the assistant's job would be to understand you, know where everything was, all the content and services, and it would orchestrate and automate solving that task for you by routing the information, the right servers, gathering it back, presenting to the user, learning from the interactions and helping the user get the job done. So I built it. In 1993, with collaborators, David Martin, Luke Julia, and it was amazing. And that's something I still can't see today.

7:07You know, we're beginning to get it with Agentic, but starting to orchestrate across services that do not know about each other in advance. So that was the original idea. And, you know, I kind of sat on that idea for 15 years until the iPhone came out. And I said, this is going to be huge. Two years from now, Apple will have flipped the game. Every handset manufacturer and telco will be desperate to compete or collaborate with the iPhone. The Siri idea I've been plinking along with working on the side for 15 years, that could be just the thing. To one-up the iPhone, in a sense, to up-level the iPhone or whatever competitive phone it was built.

7:53Yeah. Yeah. So Siri is arguably the OG conversational interface, but I would say it doesn't have agentic capabilities. So first, do you agree with that statement? And B, to your point, we still don't have that incredible chief of staff that you can ask it to do things and delegate stuff to. Why is it such a complex undertaking? And actually, maybe also for our listeners who perhaps don't have that distinction clear in their minds, like what is the difference between an assistant like a Siri and then an agentic assistant? Oh boy, I've got strong views and maybe contrarian views to the entire industry.

8:35So this'll be interesting. I'm curious. So yes, I'm very proud that Siri was the OG conversational assistant. It was the first voice assistant many came after. So there were Google Now, then Google Assistant, Microsoft came out with Cortana, Alexa came three years later. You know, so it was the first. I would say if I have one superpower, it's timing things to market. Because not only did I help create Siri, the first voice assistant, I was a founding member and first developer. Change.org, which today has 570 million members. I was a founder for Sentient, which was the first large-scale compute platform for machine learning.

9:19We had more than 2 million CPUs and GPUs back in around 2009, more than all of Amazon AWS, more than all of Microsoft Cloud as a startup. And later on, Viv Labs, I think, was the first AI that could code on the fly for a user request. So you mentioned agentic and you said, well, Siri wasn't agentic, but today we have agentic systems. I would say somewhat differently that Siri and Viv Labs were more agentic than any system we have today. And we can get into that. They were early and they didn't do everything you'd want. But they did many things that no system does today and that we need today. First of all, I say that AI today is missing three important things.

10:20The first is the right user experience. So chat, today most AI systems come up with primarily a chat interface. And, you know, maybe they put in a little carousel at the bottom that scrolls off the page. I can tell you there's a reason that most websites, most apps are not chat interfaces. They're graphical interfaces. And if you think about most tasks, so like take planning your trip to a holiday, all-inclusive vacation. Chat can help a little bit, but travel is complicated. You want to look at the photos. It's emotional. Totally. So there are so many things missing in today's AI interface of just request response, conversational chat, no real rich GUI, no rich way to deliberate, no way to collaborate, no way.

11:22There's missing things. So the interface is not right. I would argue that Siri and Viv, which eventually sold to Samsung and deployed on half a billion devices of all sorts, from wristwatches, TVs, phones, refrigerators, air purifiers, speakers, it had a better interface because it would seamlessly combine language and GUI. So that's the first thing that does not exist today that's missing is the right interface. The second thing that's missing is the original Siri vision. I want a Siri that can know and do. I want to know this or I want to do that. LLMs are incredible at knowing things. They know everything, it seems, but they can't do much at all today.

12:18And so people say, oh, ChatGPT is so much better than that OG Siri. I'm like, oh, really? I said, with Siri, I say, tell my wife I'm going to be late. And it sends her a message. Does ChatGPT do that? Oh, no, it doesn't send her a message. oh okay with siri i can say set around you know play this music this playlist does chat gpt oh no it doesn't play the playlist but it'll tell you how you can go to spotify and play the playlist it knows things but it can't do things and the architectures for knowing and doing are very different no knowing is about cached knowledge primarily think about google search or chat gpt It's trained on knowledge up to a point, right?

13:07There's a date. There's an expiration date on the knowledge in a large language model, right? And then it can call a little bit of tools. Maybe it calls to web. But doing is not about cash knowledge, search and generation personalization. That's knowing. It's about authentication. It's about transactions. back in Siri days, 2010 when we launched, before it came out on Apple, I could say, book a four-star restaurant in Palo Alto for three people, 7 p.m. Oh, make it 8 p.m. So it was conversational. We weren't the no engine that an LLM is. We don't know a lot of things, but we could do a lot and much better than any LLM today.

13:56And for me, the third thing that's required that's missing. Okay, well, we have a Gentic that can crawl a web page a little and automate it, but I hate that whole approach. We can go into that more. But the problem is there's no business model, there's no ecosystem model that works. Now, if I'm a company, say I'm a travel provider, do I want some unknown agentic agent to scrape my site, automate my site, cut me out basically of the whole process, and the user will never come see me? And that's not going to work. And from the user point of view, they're not happy that I could just say, get me a ride and it's crawling this site or that site.

14:46I want to be able to control my preferences. I like Lyft and not Uber. I like Uber and not Lyft, etc. And so the model for how this works, if I'm a content provider, how do I make money in brand? You're cutting me out of the advertising. So to me, in order to solve these things, it's not just a technical thing. It's not a coding, oh, you need to solve three things together. The right interface, a platform that can actually do transactions, live APIs so I can know if the ticket exists, all of that at a scalable way. You need a business model that supports both the needs of the provider and of the user.

15:34And it all needs to come together because that provider needs to be able to differentiate on the user interface, log in through the transactions. You have to solve it together. And I don't see anyone doing that. And the original Siri did all of that, the Viv Labs, all of that. And so I would argue that those two OG systems are more agentic than anything we have today. You know, I want to add two kind of parameters to your three things. One is this idea of memory as a moat. For these AI systems to work, they really have to have a concept, not just of the short-term contextual information. For example, it just booked a restaurant, so it knows that you're referencing that booking.

16:21But maybe a longer-term horizon of your preferences, it understands your relationships and your networks and travel preferences and all of that. A, did Siri have that or Bixby? And then B, again, like what would it take to implement a memory system that is actually sophisticated? It's not just a dump of everything that it's ever conversed with you. So again, memory is a tricky, big topic, but I'm going to cleave down the middle into knowing and doing. And this is a theme. So today, ChatGPT will start to remember things. Like it remembers that I like buttery, oaky Chardonnays. And in fact, the other day, I was asking a question.

17:09I came along some food and I'm like, I didn't know this type. I'm like, what is this? Tell me about it. And it answered me and it said, and it pairs really well with the buttery, oaky Chardonnays that you like. I'm like, brilliant. Personalized. It made me feel good. It gave me useful information that was just specific to me. So I do think there's a lot that can be done with personalization and memory on the knowing side. Knowing is about cache data and being able to search and generate based on that and to personalize. I put personalization, a lot of it there. However, there's a different form of memory for transactions for doing.

18:00And I I gave the example, if I've booked that reservation through the assistant, it should know about it. And I should be able to say, modify my reservation to APM, and it should know what I mean, right? It should know my credentials for all of it. For memory and personalization, you need to offer transparency and control. That would be awesome, yeah. It was awesome. It existed in 2010. What about empathy? Because as you know, I've spent, you know, 20 years of my life building emotional intelligence into machines. Yeah. What's your view on that? So back in the old days for doing, everything was really about live APIs, structured data.

18:51And we were trying to get this fuzzy natural language and maybe more, right? It would be nice if you knew your emotional state, which affective, you know, when you're interacting, that could give you input that not only am I asking this, I'm perturbed or I'm happy or whatever. Or frustrated or in a rush, right? But largely the doing side was about structured data. And there, APIs are pretty rigid and you kind of have to work with them. Even if you have AI that can code to those APIs, which Viv did, there's not that much room in those technology stacks for empathy. It was hard back then. Now, LLMs come along.

19:41they have learned how people communicate. They've learned subtleties, not only about what we communicate about, but how we do it. It's amazing, right? And there are studies that show that in many cases, LLMs are more empathetic than humans. And there's, in one example, you know, they took people talking to their doctor through a messaging system and the responses from the doctors compared that to what if the LLM wrote the response, and then they scored it not only on accuracy but on empathy. And there, LLMs are great at saying, well, I really feel your pain. Does it really feel its pain? No. Has it ever felt pain?

20:31No. But it knows how to be empathetic, which means to respond in a way that's nuanced and takes into account many subtle cues about that interaction and account for them and tailor responses appropriately to that situation. So LLMs, I think, do have the capacity for the first time. We never had this back in the old days to create, you know, simulations of empathy. We need more ways to give more signal into these systems and to have them actually respond even better. Yeah, absolutely. To be able to see the facial expressions or listen to the prosody's in the voice. Okay. Obviously, Siri is powered and built on machine learning and natural language processing and all of that.

21:20But the voice itself is a human voice. So I want to talk about that for a second. So Siri is, of course, powered by AI and machine learning. But behind the voice of Siri is an actual human voice. It's the voice of Susan Bennett. She's a voice actor. Can you walk us through the process of training, like what it takes to train a voice model like Siri? We're talking really about what I call text-to-speech or voice generation. I've been through many chapters of this. In the early 90s, there was text-to-speech, but it sounded like a machine, and it literally was a machine. You get the Stephen Hawking, like, you know, that type of voice.

22:07No human was involved in the making of Stephen Hawking's voice. But later on, probably mid-90s, something called concatenative speech emerged. The idea was to take real snippets of human voice and string them together and then try to smooth them out into prosody and voice sounding. so you would get voice actors who had to read these some of them very strange wording and the goal was to get snippets of sound of the voice of someone saying ah someone saying buh someone saying cut but in different ways a buh might sound different it might be a little bit different when followed for certain types of words and other types of words.

22:59So you create these nonsensical phrases that a voice actor would spend hours, it was not easy work, going, you know, saying some words, but they didn't make sense. And the goal was to capture all the sounds as possible. And then finally, they would usually record common phrases that maybe an assistant would be. If you're doing a navigation system, you kind of want it to say turn left a lot, right? So you want them to actually say turn left, and you'll have a snippet that'll be a perfect match. So the process of this, the generation of Siri text-to-speech voice was concatenative. If I have some words I want to make sounds for, take the longest span that fits.

23:52If I want to say turn left and I have someone recorded turn left, just use that. But maybe if I want turn left, but I only had them say turn, I'll use turn and then I'll form left by LFT and where can I take them hearing those sounds? So that's the process that text-to-speech used to be, you know, was what I call the second phase of text-to-speech. Today, it's still human-based, but these voice-to-voice models are neural nets that are trained differently, where now you can capture a recording, and it will do this in context not only of the sounds, but of the meaning of the words, of the intent and emotion of the words, and it learns all of that.

24:47Because if I say, look out, for instance, in the old days, look out, it would be, okay, I want the word look, and I want the word out. The latest models, you know that when I go, look out, like that phrase is going to be expressed differently than, you know, hello. And so the process has changed yet again, from kind of the Siri days of text to speech to what I would call version three, where now there's so much data on the web of how people speak different phrases and words that you can produce voice to voice models that learn all of it, not only how you sound, but how to phrase it, how to emote appropriately for that kind of expression, etc.

25:36We'll hear from the voice actor behind Siri, Susan Bennett, at the end of this episode. But first, Adam reveals his blueprint for founding companies. Plus, he shares his not-so-secret sauce to success. That's after a short break. Stay with us.

26:06AI has moved from possibility to practice. The question now is how do today's pioneers build it fast, reliably, and at scale? From medical research and education to scientific discovery and cinematic creativity, AI is already changing how we understand the world. As the essential cloud for AI, CoreWeave is a leader of that transformation. Delivering a purpose-built AI cloud designed specifically for today's most complex AI workloads, it allows researchers, developers, and creators at leading AI labs and enterprises to focus on impact, not limitations, which is how bold ideas become real breakthroughs.

26:49CoreWeave was built for this. Ready for anything, ready for AI. To learn more about how CoreWeave powers the world's best AI, go to coreweave.com slash ready for anything. If you've spent any time building AI products or leading technical teams, you know this. Transformation doesn't fail because of ideas. It fails because teams can't move together. Enter Atlassian's Teamwork Collection. It has planning in Jira, documentation in Confluence, video updates in Loom, and now AI agents in Rovo, which connects the dots across your work so nothing gets lost. It's one AI-powered teamwork platform designed for how modern teams actually build.

27:36Learn more at Atlassian.com slash TeamChanger. That's A-T-L-A-S-S-I-A-N dot com slash teamchanger. When it comes to AI, we spend a lot of time talking about what it takes to move models from research into reality. Enter LTX2 from Lightrix. It's an open-source audio-video foundation model built for synchronized sound and video, native 4K output, expressive motion, and precise multimodal control. all running on consumer GPUs. With over 3 million downloads on Hugging Face, LTX2 gives you full model weights, training frameworks, and evaluation tools to build real production workflows. Don't settle for a first draft.

28:26Get the complete creative engine to reach your idea's full potential. Try LTX2 today at ltx.io slash model.

28:40Okay, you are a serial entrepreneur and you've done this many, many times, all very successfully. And I believe you have a blueprint for how you start and exit companies, which is, again, a little contrarian to the meme that's out there. So I would love for you to share your blueprint. Yeah, look, there are many ways to be successful in companies, many ways to be successful in technologies. I've done it a few times now so I've been associated with five companies three in a major way and two in a little minor way so I have a weird pattern but it's worked for me because all five of those have been successful at some level, right?

29:25Change.org is more than half a billion members, Siri was on 2 billion devices Viv was on half a billion devices you know sentient reached a big valuation etc etc game planner ai my last company we exited to airbnb and you know so it had some success so here are things that i do that i see entrepreneurs doing wrong first of all valuations i think of the game of entrepreneurship as steps right i use the metaphor of a merry-go-round that spins twice as fast every time it goes around, at some point you're holding on, at some point you're going to be flung off. And the game, like a merry-go-round, is to grab the brass ring.

30:13So what does this mean? The brass ring means exit. So either you IPO or you get acquired. You have a positive good exit. Or profitability. You can become profitable where you control your own destiny. So when you raise money at a valuation, say you're worth$50 million, you raise $10 million at$50 million, well, that's great. That buys you time. And the goal is you know you have to be worth double. You have to be worth$100 million. But you know the next round you're going to have to be worth$200, then$400, then a billion. It's an exponential curve. If you've noticed, three of my companies, Siri, Viv, and GamePlanner, all exited at around$200 million.