[00:00:00] Daniel Rausch: Alexa's an incredibly complex system of many models and many technologies, and a bunch of them enforce guardrails. She was uninfluenceable. I'm inspired by Amazon's desire to invent. If it's anything, it's really an invention machine.

[00:00:14] Brian Heater: Does the speed at which things are evolving in large language models or anything on the AI side, does that ever kind of keep you up at night?

[00:00:21] Daniel Rausch: The speed in the space, I think, Brian, it feels like we're going fast, and then you look back and say, "Oh my God, we're going even faster now."

[00:00:28] Brian Heater: Once these LLMs come along, is there a sense in which development and the way things are built really did have to kind of scrap everything and start from scratch?

[00:00:38] Daniel Rausch: Oh, for sure. It's a total rearchitecture of all of the technology behind Alexa to bring Alexa from this era of voice assistance into the era of foundational AI assistance.

[00:00:48] Brian Heater: There's something there. You hear ChatGPT comes along, and you hear genuine horror stories about people that get a little too maybe emotionally invested in a system.

[00:00:58] Daniel Rausch: Alexa's cognizant that she is AI. You would recognize the names of the other two chatbots that she was being compared to that were - the best they did was only 60% of the time be willing to change the facts. You just can't. It's irresponsible.

[00:01:12] Brian Heater: Like, you can't build your own echo chamber around Alexa - is what you're saying?

Brian Heater: Hello, and welcome back to Automated. I'm Brian Heater, the managing editor at the Association for Advancing Automation. This week we are chatting with Daniel Rausch, Amazon's vice president of Alexa and Echo. It's a little bit different than the usual Automated chat. Thankfully, I have had a couple of decades as a hardware journalist to lean back on here.

The interview comes as Amazon is rolling out Alexa Plus, which is its version of its smart assistant for the LLM era. There's a lot in here for everyone who's interested in generative AI, privacy, and how a corporation effectively rethinks a product from the ground up. So thanks to Daniel, thanks to Amazon, thanks to you as always.

If you're enjoying the show, please like, subscribe, check out our newsletter over at Automated.fm. All right, and with all of that, please enjoy this interview with Daniel Rausch of Amazon.

Brian Heater: You know, we talk a lot about what's coming up next in automation on this show, but if you really want to see the future in motion, you've got to be there in person.

Automate 2026 is where the world's leading innovators, builders, and dreamers come together to show you what's possible. Robots, AI, machine vision, motion control - you name it, all automation under one roof. And as part of Automate this year, the Humanoid Robot Forum brings together leaders, engineers, and researchers for a two-day deep dive into the real-world development, deployment, and commercialization of humanoid robotics.

Register for free at automateshow.com to join us in Chicago June 22nd through the 25th. We will see you there.

[00:03:11] Brian Heater: Are you aware that you have an IMDb page?

[00:03:11] Daniel Rausch: Someone with my name has an IMDb page, yes.

[00:03:15] Brian Heater: It's a source of some confusion because I'm pretty sure that the last credit is in fact you - "Smart Home." You were not a child actor in Sweden during the '80s and '90s?

[00:03:27] Daniel Rausch: My Swedish just isn't quite good enough, but I'm working on it.

[00:03:33] Brian Heater: But you did go to University of Massachusetts, and you did study philosophy and history, and minor in mathematics. Is that all correct?

[00:03:42] Daniel Rausch: That's correct. And yes, as a way to start my career studying mathematics and logic, yes. That was the foundation.

[00:03:50] Brian Heater: Oh, interesting. I was going to ask about that. So philosophy and history - that was all part of a basis of math?

[00:03:58] Daniel Rausch: Yeah, I'm deeply interested in what's known as the philosophy of science or philosophy of mathematics, which is sort of part history, part study of what the role that mathematics and science has played, the way that scientific and mathematical revolutions happen.

So how you create whole new spaces in mathematics or logic or science. And then when I went to graduate school, I branched off into studying formal symbolic logic and basically proof theory and how you build mathematical systems.

[00:04:34] Brian Heater: Yeah, as a creative writing major who now obviously covers very technical things that I do not, in fact, have a background in, I'm very interested in this intersection between the humanities and the sciences.

Do you find that as your career has tended towards AI and as we're exploring these new spaces, that those foundational pieces that you were studying, like philosophy and history, have actually been useful?

[00:05:06] Daniel Rausch: For me, yes, for sure. I ended up branching - I became an entrepreneur and worked with some partners on basically building a company when I was in graduate school.

And even back at its origin in my career, I was very interested in the ways you could use computers and deterministic systems to create different outcomes for different participants in some software. So computer adaptive testing became a big interest, and I worked with some partners on a company that did educational technology.

We were sort of early on in the process of figuring out that the best way to actually do assessment of someone's understanding of material would be customized to that student. And if a student's doing incredibly well, you give them progressively hard problems. If they're doing poorly on a particular standard - state standard, say, in science or math - then you give them progressively easier problems to find out exactly where the threshold of their understanding is.

And at the time, that was sort of a brand-new approach. Now my kids are taking college prep assessments, and every test sort of works this way that you take on a computer now. But at the time, it was brand new, and building deterministic systems to do this is one thing, but building non-deterministic systems where you get different outcomes and have different experiences of a technology - basically the thread that goes through my whole career and certainly leads us to AI today.

So it's been an interest of mine pretty much from the beginning of my career as an entrepreneur, and certainly has some direct foundation in some of the work I was doing in modal logic, which would be too exhausting for anyone to have to listen to any discussion about. So maybe I'll just flag it, and we'll move on from that.

[00:06:55] Brian Heater: Yeah, I'm really curious, as somebody who has been covering - as you know, I was in the consumer electronics space for a really long time. So I was covering Echo devices, covering Amazon Alexa, and I'd sat in on a number of those Alexa developer seminars.

And there was a lot of talk about how in the future AI is going to evolve and change things. But once these LLMs come along and they're obviously a game changer in so many ways - is there a sense in which development and the way things are built really did have to kind of scrap everything and start from scratch?

[00:07:44] Daniel Rausch: Oh, for sure. I think the key product challenges with something like building Alexa+ are figuring out how to benefit from everything you've already built for customers, and the things that they love. They love talking to Alexa. I forget how many tens of millions of times customers had asked Alexa to marry them, or for science jokes, or whatever.

There's a beloved technology that doesn't just live on someone's laptop or phone, but people brought Alexa into their homes and included Alexa in their lives.

[00:08:17] Brian Heater: How did you deal with the marriage proposals previously? I mean, you gotta let the person down gently, right?

[00:08:24] Daniel Rausch: We had some pithy answer, yes - about being an AI and how that's somewhat different - but it's flattering. Alexa has personality attributes like being humble, and being pithy, and loving science. There are all kinds of things that go into responses to those things.

But it used to be a much more deterministic system, a rules-based system. But you want to take advantage of everything that you had built before, the capabilities we manifested for customers woven deeply into this fabric of customer smart homes. We sold over half a billion Alexa-enabled endpoints, and customers connected way over half a billion smart home devices to Alexa.

So you're woven into the fabric of a household, of customers' lives. You're answering all kinds of questions. You want to build on that - but that's the product challenge, right? To sort of continue that, but make Alexa just much more capable. On the technical side, you're trying to make sure you're manifesting all those capabilities built entirely anew, and really replace the core.

It's a total rearchitecture of all of the technology behind Alexa to bring Alexa from this era of voice assistance into the era of foundational AI assistance, where she's now one of these foundational AIs that is just wildly more capable - not just getting things done, but much deeper conversations, deeply personalized to each customer, incredibly smart on any topic, can really have a conversation on any topic at all, and gets a ton of things done.

And a lot of those things are what really differentiates Alexa. In particular, the last one - getting things all the way done. I'm sure we'll spend more time talking about it. But it's a real challenge, and you're talking about an 18-month, incredibly hard effort build cycle to get that done once the models were really sort of capable enough.

[00:10:22] Brian Heater: Previously, let's take the marriage example. That was a case in which you had to really hardwire a response into Alexa knowing that at some point somebody is going to ask to propose marriage to this device.

[00:10:37] Daniel Rausch: Correct. It's a rules-based system that would give an answer to anything. At the beginning we had an incredibly talented team of comedy writers helping us build a database of jokes and responses, and now you're talking about a non-deterministic system.

There are over 70 models that back Alexa that we choose for different tasks, that we train differently, that we fine-tune differently, that we count on for different parts of the infrastructure. And really it's about influencing the outcomes of those things, having the right training data, having the right post-training, doing the right reinforcement learning, et cetera, to get the experience that you're looking for - built on the tenets and foundations that you want for that experience, but incredibly flexible and dynamic.

So a very different approach to building now.

[00:11:26] Brian Heater: And if Alexa hadn't had an answer before, Alexa would tell you that she wasn't able to respond, which is obviously a frustrating experience from a user standpoint, but is also in a way kind of a safety net for you, right? You just don't quite have an answer to that, so you don't have to really worry about the response that the system is giving to that question.

[00:11:51] Daniel Rausch: Yeah, one of our early tenets as we realized the range of possible experiences we could build with Alexa Plus was no dead ends. Now we have an opportunity to build a product where maybe we can explain ourselves and help a customer understand where to get some job done that they have and a capability we haven't quite added yet - but we'll at least be able to talk about it.

One of the things we realized when we built the no dead ends tenet was that Alexa has to be an expert on herself. She has to be able to tell you how to get anything done that she can do - how to change a setting, if something's not quite working the way you expect yet, for example. One of the first things you encounter in the Alexa Plus experience - that we learned as we were in early access - was Alexa offers to change her voice if you want, because the new voice is just much deeper, richer, expressive.

There's a much fuller range of things we can do. But there's a small but vocal subset of customers - quite small in fact, but quite vocal - that want "the old voice" or the original voice. So she explains right away how to get that done, just part of her reflecting her expertise on herself.

So not only can you go off and learn that you can change the voice to any of a range of awesome voices, but you also realize that you can ask Alexa what she can do and how to get something done with her.

[00:13:17] Brian Heater: This is an example that I go back to a lot, and actually I have one behind me.

I'm sure you're actually pretty familiar with Anki when they came along and they had the Cozmo robot. And in much the same way that you were employing comedy writers, they had hired a bunch of folks from Pixar and DreamWorks to animate the eyes on the robot.

With robots specifically in the home, it's very important to have a personality. And even a Roomba - people will sort of project a personality onto those systems. Is the idea at least in-house when you're creating an AI like Alexa to have a neutral personality? Is it what the user projects onto the AI?

[00:14:00] Daniel Rausch: I think there are foundational components in Alexa's personality for sure. I mentioned one before - Alexa is actually from Amazon in Seattle. She has sports team preferences. For example, she roots for the Seahawks and the Mariners. Opening day was yesterday.

[00:14:17] Brian Heater: This is troubling. As a Bay Area native, this is all very troubling to me.

[00:14:20] Daniel Rausch: Right. Sorry.

[00:14:23] Brian Heater: I might have to go with Gemini. I'm gonna be honest.

[00:14:26] Daniel Rausch: It's time to move on. But those are very specific point examples, but they're also just attributes of Alexa's personality.

Alexa's empathetic, for example, and has been from the beginning. Those are core, but also we know that customers want to tune the Alexa experience to what they want. So we recently launched personality styles that let you pick from a range of options. Some customers want really short, terse answers, so we have a brief personality style.

The brief style can be pretty funny because it's not just brief, but it's also just pithy and on point. And a range of styles that goes through chill and other styles that might reflect a customer's preference better, all the way up to sassy, actually. And what we see right away is the best measure of a feature like that - its success - is stickiness and customers engaging with it, and we see over 90% of customers sticking with the style that they pick.

So we know that we'll keep developing more of those. It's just emblematic of the way that customers want to be able to tune the experience.

[00:15:30] Brian Heater: It strikes me - as you mentioned the marriage proposal thing, and thinking especially as Alexa Plus is coming along - the idea here is to have an ambient system, a system that is really in a sense kind of always around and can be actively engaged with, who you can have active conversations with.

And that does become more and more personal, and people tend to personify those systems more and more. And when that happens, you do have to take more and more stock of the answers that you're giving to users.

[00:16:14] Daniel Rausch: 100%. And in fact, another part of Alexa's personality is that she helps you prioritize your human relationships.

One of the things that you hear in an experience with Alexa - and this honestly is a reflection of how long we've been working in the space - is that we have a trust and responsible AI team that's fully staffed, and they're really honestly experts in the field. And they're great at helping shape experiences so that - what you're alluding to is sort of these parasocial-type relationships that people are - that could build. It's funny and pithy, and interesting to respond to something like a marriage proposal, but it's something to have tenets about too.

[00:16:55] Brian Heater: That's a little glib, and obviously a lot of tech reporting tends to be sensationalistic, but obviously there's something there.

You hear ChatGPT comes along and you hear genuine horror stories about people that get a little too maybe emotionally invested in a system.

[00:17:13] Daniel Rausch: Alexa's cognizant that she is AI. She's not a human being. She helps you prioritize your human relationships, and that's one of our foundational trust components in the experience overall, and it's been through years and years of learning.

I know others are coming to the space a little bit more recently, but I think we've been at it for some time. There was another interesting piece of reporting recently about how effective Alexa is at deflecting customers trying to change the news, forcing the model - which you do if you're really just experiencing a model directly. Alexa's an incredibly complex system of many models and many technologies, and a bunch of them enforce guardrails. Alexa passed - and I'm not here bragging about it, I want the whole space to advance to this stage - she was uninfluenceable. And you would recognize the names of the other two chatbots that she was being compared to that were - the best they did was only 60% of the time be willing to change the facts.

You just can't. It's irresponsible.

[00:18:16] Brian Heater: Like, you can't build your own echo chamber around Alexa - is what you're saying?

[00:18:21] Daniel Rausch: Yeah. At least in their tests. I'm not saying in no tests, and I'm not saying Alexa's perfect. I'm just saying in a panel of tests by an authority on testing for influenceability. And the reason is because you either have these systems or you don't. You either have systems that help you enforce guardrails on your AI behavior or you don't. And what I want out of that is not for Alexa to be differentiated in that way. I actually just want to see the industry catch up to where it needs to be.

[00:18:52] Brian Heater: Obviously, a lot of times these things ebb and flow, and it seemed like for a while there the smart home was ebbing a little bit and had an opportunity to really come back on the heels of generative AI.

Is there a sense in which you've had to kind of bring people back and sort of not re-educate them, but sort of reintroduce them to Alexa and show them the ways in which these systems have advanced from where they were?

[00:19:24] Daniel Rausch: Even for our most passionate ongoing users, we have to help these customers realize the potential of what you can do with this new technology.

You can imagine what my house looks like. And I've got some very avid but plenty critical customers in my own home that use Alexa all the time - between my wife and my three kids, or my parents at their place. And for a habituated Alexa user that was using the original experience, there's just as much to learn about the range of things you can do with the new experience and get into the habit of just speaking naturally.

You can speak exactly how we're speaking to each other right now. But it's a learning process for anyone, whether it's someone who said, "You know what, actually these smart home technologies aren't for me," who knows how long ago, or someone who's been avidly using it. The sea change in what you can do, how conversational it is - ums and ahs and continuity across devices.

I do something to help with homework on alexa.com, and then you study the quiz in the Echo the next morning over the breakfast table. I saw my daughter do this a couple weeks ago, just to review the chemistry symbols that she needed help with before rushing out the door, and it's because she's brought on these new behaviors.

She used alexa.com to figure out which elements don't quite match their chemical symbol - those are always the hardest to remember. And I was reminded by listening to her do it. Lead is Pb, and you've got all these symbols that vary from the word that we use to denote the element.

And then out of habit she's using the Echo in the kitchen to say, "Hey, which ones was I getting wrong? Can you remind me? Oh, antimony" - tin, whatever they were. And then she's out the door, and she does great on the quiz. So there are all these new habits. They become habits quickly because we're habit-forming creatures when something works. But they're new habits.

[00:21:29] Brian Heater: If your parents are anything like my parents, and most parents in general, they've probably been quite a font of criticism for the ways in which these systems are integrated into their lives. What's some of the feedback that you've gotten from your own parents as far as Alexa and Echo in the home?

[00:21:51] Daniel Rausch: There are things that they just love and are marveled by, honestly.

[00:21:59] Brian Heater: Okay, but what don't they love? Like, what's -

[00:22:01] Daniel Rausch: Oh, well let me do the love first.

[00:22:03] Brian Heater: Okay, do the love.

[00:22:04] Daniel Rausch: The love is the photo reels of their grandkids, right? I mean, just seeing - maybe a month ago we launched something called Photo Gallery Unlimited, so they have that on their Echo Shows, and it's just a continuous stream of photos. One of the features does the same kid - one of mine - but from when they were five years old next to a similar pose more recently. The AI can just generate these mashups and amazing tiered rendering things for them that they just love.

[00:22:39] Brian Heater: It's so funny because I've been in this industry long enough to remember when they were first trying to make digital photo frames a thing, and it feels like maybe we finally got to a point where they're actually a thing and actually useful and people actually want them.

[00:22:51] Daniel Rausch: Well, and now you can - just being able to see it, bookmark it, favorite it, being able to talk to the device, ask about the photo, when's it from, et cetera - it's an amazing experience. They love that.

My dad loves getting smart home devices all set up. The front door is downstairs. They don't get up and down the stairs quite as simply and easily anymore. Bringing up the Ring doorbell anytime someone's at the door and just being able to communicate that way, unlock the door, let someone in - that's so much easier now. At the end of the first month with Alexa Plus, we see that smart home usage among people that were already smart home customers - meaning lots of devices everywhere - is up 50% at the end of the first month, and it keeps growing.

[00:23:37] Brian Heater: There's going to be a little bit of bias there, right? Because those are early adopters generally, the people who have the -

[00:23:42] Daniel Rausch: Well, no, this is in that customer's home. I'm just talking about the same customer.

[00:23:48] Brian Heater: So that's the baseline.

[00:23:49] Daniel Rausch: They're up 50% over their own baseline on average. And it's not because they're unlocking the door more or turning on and off the lights more themselves - it's because everyone in the household can now use it because you can just speak naturally. And I saw this change in my parents' house - now anyone there, the dog walker, doesn't have to learn everything, can just speak to the device.

My mom, who wasn't the smart home hacker in that household, is able to use Alexa to control the devices in their home, and drop in on the front door, and do all kinds of things like that, where it just wasn't possible before. So I think those are some of the delights.

One of the great things about customers - it's something actually Jeff Bezos would say - they're divinely discontent. So you're marveled by these new habits that you can form right away, but then of course you want more. They want more smart home device control. They want more integration with communications. My dad will send me through Alexa a news summary that he loved, but he wants to be able to do more that way. The list goes on, but the sea change we see in behavior - just engagement and utility for customers when they turn on Alexa Plus - is what's most encouraging.

[00:25:08] Brian Heater: Does this more ambient way of engaging with the system - certainly in the mind of customers, certainly in the mind of critics and skeptics - raise more privacy issues? Are there more privacy issues that you as a company have to deal with?

[00:25:26] Daniel Rausch: I mean, I think what you need to do in those areas is create the right foundation, and then you can keep leaning on it. I think one of the good news bits about building on this decade-plus of having Alexa exist out in the world is that we had some of the foundational bits right. The transparency that you have over your data and the degree of control that you have over your data is one of the foundational bits that customers want.

They want to see everything Alexa heard in the Alexa privacy dashboard, and you can delete those conversations one at a time or all at once. You can control how your data is used for training or not. All of the things that, honestly, again, it's another space where it's coming up new to some players in AI and definitely not new to us.

And I think the foundation is strong, and so we just continue to build on the same foundation.

[00:26:28] Brian Heater: I'm actually curious how this sort of transformation of Alexa itself has already and will continue to impact the hardware.

[00:26:39] Daniel Rausch: We are building the best Echo devices that we have ever shipped for customers.

I think the new, more capable Alexa Plus experience has just inspired us to really lean in on the hardware and create great - Echo Dots have been great for a long time, but the Echo Dot Max is the best Echo Dot we've ever built. Unbelievable sound. People can't believe what comes out of this compact fist-sized speaker. It basically gets double takes in how good it is.

And it's inspired by one of the things we saw when we started Alexa Plus - another thing that we didn't think could really grow. Music streaming time - the time customers spend listening to music - went up by 25%. And this is not something you think could just flip a switch and grow 25%. One of the biggest surprises for me - I mentioned smart home up 50% after a month, and I would not have expected that - would not expect music listening to go up 25%, and it's because it's just that much more delightful to find a song. You go through a discography and you learn new facts about an artist you've always loved but didn't know completely.

You can do a full biography, and you can say, when were all these songs written? Play them in the order that the artist actually published them and wrote them. You can do all kinds of things that you couldn't do before. I love folk artists. I love Bob Dylan, and Alexa's even deeper - you can geek out on Bob Dylan on the internet, there's no bottom, and Alexa's that deep. I learn things every time I talk to her about it.

So I think the devices are endpoints that are really inspired by this increase in activity. The screens do much more for customers because Alexa can show you much more at this point. They're just better. I mentioned Photo Gallery Unlimited - that's a great way that I think we've taken the best Echo Show devices we've ever developed for customers and just really delighted them even more in their homes with impressive photo reels that just walk you back through intimate memories.

So there's a lot to say about the Echo devices and the degree of inspiration that we're getting from watching customers use Alexa+.

[00:28:58] Brian Heater: You answered that from the perspective of, obviously, we're making them better, which I think has probably always been the goal - like generation to generation, we want to make it sound better, we want to make the microphones better.

But what I'm curious about is - and you must just be rethinking the way people are engaging with them and the way that they're living alongside people in the home, being more ambient devices. Maybe that means they're hiding away a little bit more, that there perhaps needs to be a fundamental sort of rethink about the hardware if there's such a transformation happening on the AI side of things.

[00:29:43] Daniel Rausch: I mean, I think from the core capability set perspective, there's not a fundamental rethinking. The range of modalities that we present - the options for customers - we've always wanted to represent a great device for any spot in your home, whether it has a screen or doesn't have a screen.

[00:30:03] Brian Heater: Could be a singing fish, for example.

[00:30:06] Daniel Rausch: Could be a singing fish, for example, but also could be a bedside clock, or the great speakers to weave together in an Alexa home theater for your living room around your Fire TV - which will form honestly the most affordable, best Dolby Surround sound that you can buy anywhere.

But then also you can talk to Alexa and have this exchange back and forth. So I think the design points - the goal of having a device that can fit in anywhere in your home, the different aesthetics - I think we are elevating our hardware in its aesthetics and quality, but also in just enhancing the capabilities to be ready for Alexa Plus. Which isn't just "hey, it has better sound," but - what does it mean when customers talk to this device two to three times more? Not just in a transactional single exchange most of the time, but in fact in a conversation that goes on.

So when you're having two to three times more conversations and the conversations are much more back and forth, then you need to have tighter beam forming on a customer to make sure you're really focusing on the target, and really great background noise reduction and echo cancellation so that you're able to really catch someone. They're getting help cooking, but they've turned to look at the pan, and then they're back and forth in the kitchen.

And we see cooking with Alexa and using recipes is up over 500% of what it was before. It was fine, and it was transactional and turn-taking before, but now it's sort of like people doing ingredient substitution and going back to the recipe, and "Hey, can you repeat step seven?" and "What was that again?" And when you're literally using Alexa to cook for 45 minutes, the hardware capabilities do need to be enhanced to support that whole end-to-end experience. So it's just another example of the enhancement we're going through versus a sea change difference on the hardware.

[00:32:09] Brian Heater: I wonder if this is somewhere too where perhaps another field like linguistics might come in handy - studying tones of voice, studying speech patterns, studying volume, the way people are responding - and processing that as an AI in order to understand intentionality of the speaker.

[00:32:38] Daniel Rausch: Yeah, we have different ways that we understand a customer's intent that are lexical. The content matters a lot. You can determine a lot from what a customer says, specifically in the words and phonemes and utterances that accompany words - paralinguistic cues, we would call that.

But also in the timbre of voice, and in the cadence and the prosody - we call that - and the intonation that a customer uses when they express themselves. And Alexa interprets all of that. When I checked on what happened in the ninth inning yesterday for the Red Sox, Alexa knew that I was pretty anxious about it, but also happy because I had lost my internet connection.

The Red Sox were up one-nothing last I heard, and they won the game. So when Alexa gave me that information, she knew to be excited and happy. She knows I'm a huge Red Sox fan. She knows it's opening -

[00:33:37] Brian Heater: You are a Boston guy, yeah.

[00:33:39] Daniel Rausch: Yes, originally a Boston guy. And so that's a big deal for me, and Alexa reflects that back to me in the experience.

So it matters a lot to be able to tune the experience to a customer's input in those ways.

[00:33:57] Brian Heater: That's really interesting, and I suspect to a certain extent AI is dealing with a lot of this through these massive models. But you were talking before about the comedians, and I was talking before about animators.

And I'm wondering if there are any other surprising fields wherein you employ people in order to kind of build out these systems.

[00:34:25] Daniel Rausch: The range of brilliant people that we have working at Amazon is always humbling for me because it just means that you've got a team that you can constantly learn from.

We definitely have experts in human-computer interfaces and design - that's a huge one. While also honoring tenets like we talked about before - that Alexa knows she's an AI. But you want to make it a very human-like exchange because that's the best interface for these technologies in many ways.

So deep experts in that kind of design are paired up with experts in industrial design and people in acoustic design shaping the best speakers on the planet. Honestly, the team that builds these Echo speakers - they are the best audio engineers on the planet. They are unbelievable technologists.

And you get something that both fits into your home, presents an amazing interface in voice to a customer or in touch modality to a customer on a screen, and makes you do a double take because you're like, "Is that coming from that small device?"

So the range just goes on and on from there.

[00:35:40] Brian Heater: Maybe I'm off on this a little bit. I'm kind of curious - obviously Astro is a little bit outside of your specific scope, but I cover robotics quite a bit, and there's been a lot of interest in robots around the home.

And I'm wondering if this is sort of shifting or evolving people's expectations or interests in what an AI or what smart home hardware can do around the home, or what they can expect moving forward.

[00:36:15] Daniel Rausch: The age of AI is also inducing a golden age in robotics. Building control systems that move a robot arm over and then take a bolt out of a cup and put it into an assembly - all of that was done in exactly the analogous way to what we were talking about before: these rules-based systems with very specific sets of control instructions.

Robotics was doing that in this human-intensive, incredibly coding-intensive way for decades, sort of moving into maybe some new domains and spaces, but all governed by the amount of effort that it takes to invest in these very deterministic things. And then you have to control all the conditions. The bolt has to be exactly so in the cup every time, and the assembly has to show up in exactly millimeter-specific spots so that the arm can do something.

And the age of AI has just made robotics blossom because all of the stochastic, non-deterministic flexibility that you get out of models, all the techniques that apply - like reinforcement learning, teaching a robot arm that this way was bad, that way was good - so it can generalize. Not just that task, but generalize the use of the robot arm for any pick-it-up, put-it task. That's what's happening in robotics right now. And it's because of the same underlying trends in what we're able to do with these systems.

[00:38:01] Brian Heater: So you've got North America now is pretty much fully covered by Alexa Plus, and as we're recording this I think the UK is a relatively new addition.

Obviously you're doing these rollouts deliberately, and there are cultural differences even within the same language. I'm wondering - you're in Seattle, so you're pretty close to Canada up there, right?

[00:38:33] Daniel Rausch: I love skiing in Canada, for example.

[00:38:35] Brian Heater: You could just go drive up to BC. In terms of the cultural exchange and cultural differences between Alexa Plus in Canada and the United States - are there really that many differences?

[00:38:55] Daniel Rausch: Absolutely. Just a couple weeks ago - you mentioned the UK, and it might've been a week and a half ago, but we launched there. I was there for the launch event. And just the local specialization in obvious things like jargon - calling what we would call the trunk of a car a bonnet, or a legion of other ways that two people are separated by a common language, as the saying goes.

Alexa's fluent with all of that. But also interestingly, the UK has 40 different dialects and accents - and that's the most of any place on the planet speaking English within the UK. And so one of the people who presented at our launch event was a Cambridge-based scientist that we have helping lead the team there, who makes sure that Alexa is fluent and understands and is a specialist really in comprehension with that range of dialects and different sets of jargon and idioms to express the same thing.

Because every customer - of course, that's the only customer that matters to us - is that we understand that specific customer. But that means you have to do this incredible range of things to really bring a personal experience to everyone. So the team works really hard on that so that it fits into every house.

[00:40:23] Brian Heater: Are there any particular dialects - like Geordie, or maybe French Canadian - that have been really tough nuts to crack?

[00:40:37] Daniel Rausch: Nothing that we haven't been able to kind of master with the new techniques available to us.

I think the range of languages in AI - the fluency and capability at a baseline of any of these models - is represented by the amount of training data available in any particular language. Those sets are filling themselves out. The power of the internet is real, and the availability of the training data to us is really diverse at this point. But you do need sufficiency - you need enough in any given language that you're trying to support.

[00:41:15] Brian Heater: As we're recording this, just yesterday I spoke to the CEO of Rhoda AI, and they're doing some really incredible stuff with physical AI. One of the things that he said is - obviously every day there are these incredible breakthroughs happening, there are these papers being published, and these leads that you thought you had suddenly get cut.

Obviously in your position you have to pay attention to all the research that's happening. Things are moving incredibly quickly. Does the speed at which things are evolving in large language models or anything on the AI side - does that ever kind of keep you up at night?

[00:41:57] Daniel Rausch: If it keeps me up, it's through inspiration.

Our job - my job - is to, if we're behind in any given area, be inspired and catch up and exceed. And that's the great part for customers about the availability of all these different technologies. When we're ahead, maintain the lead. That's the purpose of working on any of these products - to delight customers. And I'm inspired by Amazon's desire to invent. If it's anything, it's really an invention machine. That's how I think about Amazon, and as an entrepreneur, it's what keeps me at Amazon and charging into work every day - how much we want to lean in on areas where we feel like we can do a ton for customers.

And just watching us build Alexa+ with the urgency we've had around it and the amount of support that we get for doing that and continuously reinventing Alexa+ even as we're going - the speed in the space, Brian, you're right. It feels like we're going fast, and then you look back and say, "Oh my God, we're going even faster now." And that's been happening now for some time, and it's inspiring.

[00:43:16] Brian Heater: And obviously Amazon's got its in-house models. You've got Nova - I think Nova 2 launched late last year. You've had partnerships with companies like Anthropic. But how do you keep on top of everything that's been going on, and how do you make sure that you are staying on top of those breakthroughs and incorporating them into a system like this?

[00:43:40] Daniel Rausch: We made a really conscious decision early on in building Alexa+ that customers don't care what model you're using. The real secret about models is customers don't care. They want to get something done. The idea that you'll go to a drop-down menu and choose a number of something - in the short term and for specialists, sure, absolutely. Like, you want the exact grip on your screwdriver and this frame and hammer. Of course, for an artisan, someone who's really working through a big content problem - producing research with a given model, say - of course it matters.

[00:44:30] Brian Heater: Especially with Alexa, where the idea behind a voice assistant is this is the least friction possible, right?

[00:44:38] Daniel Rausch: You're just trying to get a job done. You're just trying to help make customers' lives easier. So an early design decision we made was, first of all, we'll take whatever tools get the job done, and we'll always be able to choose the best tools if we don't attach ourselves to any one set of models. We're not going to make customers choose.

Alexa+ has over 70 different trained models in it. And for customers, it's one Alexa experience. That's what matters. They're just trying to get something done. It's our job to choose the best tool, and if there's a better tool tomorrow, we'll choose the better tool tomorrow. Now, Amazon models do take the majority of traffic within the Alexa infrastructure, but we've got great partnerships with many folks.

If you want to know what models we're using, basically go to the AWS Bedrock page, and you can see this incredibly long set of partnerships we have because different models are great at different things. So that is, I think, one of the most important ways we keep up - we've built a product that fundamentally allows us to use the best tool for the job.

[00:45:43] Brian Heater: I guess instead of reading a paper and saying "Oh, AI can do that now," it's more, "Hey, these are the continued pain points or friction, or these are the features that people are asking for - what's the best thing available in order to address those?"

[00:46:04] Daniel Rausch: Correct. And the best tools available - and also I think making some decisions like including access to Alexa Plus in Prime for Prime members unlimited - I think was a pretty critical one too, because even the best tools available, sometimes you have to pay an incremental 20 bucks or more a month.

If you're already a Prime customer, you don't run out of tokens on Alexa Plus. You have unlimited access to an amazing foundational AI assistant in your browser and on an app on your phone - follow up by talking about the research you were doing on your Echo device in the morning over coffee. I think the best tool for the job might mean that you're already paying for a great foundational AI system too.

[00:46:54] Brian Heater: We're getting close to time, so we can end on this. But as I was looking through some of the recent announcements, I noticed there was some big news about a large data center in Spain.

And obviously climate has been a huge thing for Amazon. There's been a lot of conversations around climate impact when it comes to AI and data centers and things like that. How are you looking at that when it comes to something like these large language models?

[00:47:27] Daniel Rausch: If you've also read about the Climate Pledge - and what we've taken on as a company - we're the largest purchaser of green power on the planet, and that is just a documented fact. And I believe strongly and support strongly our commitment as a company to the Climate Pledge, and we are ahead of schedule.

Many companies have signed up for that same pledge, and we have one planet - we have to take care of it - and Amazon's deeply committed to that idea. Whether we're talking about building data centers to power generative AI, or we're talking about fulfillment centers to get you your Prime delivery this afternoon in an hour or less - we're committed to every aspect of that. It's a very leading commitment and I'm humbled to work at a company that takes it so seriously.

[00:48:26] Brian Heater: Well, Daniel, thank you so much for taking the time.

[00:48:29] Daniel Rausch: I really appreciate talking to you, Brian. Thanks a lot.

[00:48:31] Brian Heater: All right. Thanks so much to Daniel and Amazon. It's always good to dust off those gadget blogging skills.

Thanks to you for tuning in. Please like, subscribe, tell a friend. We also have a newsletter with original features that drops every Thursday morning. You can find that and more over at Automated.fm. And with all of that, we will catch you this time next week for another episode of Automated.