The Vergecast - Teaching AI to taste and smell could help the future of product design

Episode Date: October 5, 2021

For the next four Tuesdays, Verge senior reporter Ashley Carman will explore how artificial intelligence and machine learning are shaping the future of a variety of industries. In this episode, Ashley explores how AI can be used for product design, but more specifically, for creating fragrances and flavors. Guests include founder of ScentTronix Fredrick Duerinck, computer scientist at Cold Spring Harbor Laboratory Saket Navalkha, and Michael Spranger from Sony AI This podcast was made by producer Liam James, senior audio director Andru Marino, senior reporter James Vincent, and senior reporter Ashley Carman. Learn more about your ad choices. Visit podcastchoices.com/adchoices

Support for the show comes from Retool. Too many companies run critical operations on duct taped spreadsheets, Slack workflows, and whatever else they could cobble together. Not because they want to, but because building internal tools means weeks of waiting on someone else's backlog. That's where Retool comes in. Build custom internal tools just by describing what you need. Prompts something like,

Build me a revenue dashboard on our Salesforce data. And Retool actually builds it on your company's data, in your cloud with enterprise security built in. Go to retool.com slash Verchcast. We all need to retool how we build software. Hey, Vergecast listeners, it's Nilai. For the past few Tuesdays, we've been running a mini-series we made about the many uses of artificial intelligence and machine learning

all over the place. It's been hosted by Verge Senior reporter Ashley Carman. Today is the last episode. Hey, Ashley, how's gone? It's a sad day, last episode day. But like a happy day because our AI future is here. It's true. if we figured it out.

So we've done voice synthesis, AI and video, robots writing books with authors in text. What's this week's episode? Today we are talking about how AI could help us create new flavors and new smells. That's, this is like, it's like over. Now we're like through the looking glass AI taste. What's an AI taste? Well, I haven't tasted an AI taste, Neelai. This is all, you know, in the works.

This is all happening as we speak. I see. So basically like I thought this is. crazy, like AI can't smell. But then I realized I just have to think of like my nose and my little nose hairs as sensors. And then it all made sense. I was like, oh, I'm a computer. Very deep, very existential. And my brain is running an algorithm. This is why you have to start thinking about these things when you think about smells and tastes. It's like, oh, we're just running the algorithm. I like what's happened over the past four episodes here as you've started to consider yourself

more and more as an AI. It's true. I'm replaceable. It's been an interesting month. All right, Check out the rest of the episodes if you haven't listened already. They're all very good. But here is episode four, the Verschess AI series. Enjoy. When I cook a meal, I typically take stock of what's in my fridge and either Google a potential recipe or wing it. I'm definitely not doing anything fancy to figure out what could maximize my meal or flavor.

But I'm a mere human, and if multiple tech companies have their way, that could change in the future with AI. In today's episode, we're talking about AI and how it relates to product design. But more specifically, how people create fragrances and flavors, smells, and tastes. It sounds wild, I know, because, well, at least as far as I know, AI isn't capable of smelling or tasting. But that's what we're going to investigate today. Can AI personalized fragrances, concoct new recipe ideas, and maybe even develop a flavor we've never tasted before? It's extremely complex what we're trying to do.

That's Frederick Dunrick. He's the founder of a company called Centronics, that's part software company, part hardware. Through their algorithmic perfumery platform, he and his team are trying to make the process of developing a personalized scent as easy as answering a few questions and pressing a button. Algorithmic perfumery is on one hand, it's a machine, and it's on the other hand, it's a very big software piece. And how it works is that people go to a website and then they fill in a questionnaire, and the questionnaire can be partly psychology questions or physiological questions. or cultural background questions.

And then based on the input they gave, the answers they gave, a unique perfume formula is created by the AI. And then people get a code and then they can go to the machine and then they fill in the code. And then the miracle of creation is happening. And then the bottle travels and then the unique perfume is created for them. The questions might be as predictable as something like, what sense do you already like? Or more probing.

Like how do you typically dress, what you do for work and even where you grew up? Frederick says this is all essential data to build a scent people like. What is important to understand on scent that it is very cultural sensitive. So it means that the context you are in as a person and what you've smelled very often before is often a thing you're very accustomed to. Like for instance, if you're from a different region, for instance in what we have seen because we have done tests in art festivals in Asia, we see that there are people, for instance, they don't like woody stuff.

They really don't like woody stuff. So whenever the AI concocks something with X amount of percentage or over X amount of percentage of wood or woody ingredients, basically people, they don't like it. As Frederick mentions, he and the team are not only trying to code software and algorithms that can accurately produce a scent you to enjoy, but also pair it with a machine that mixes that smell on demand. They're building their data set from scratch,

meaning that whenever they create a smell, they ask for feedback on it, all in the hopes of improving their software. At this point, you, like me, might be wondering, why do we need AI to do this? Frederick says it's a highly complex field and one with many possible combinations that only machine learning could handle. We have three machines, actually. One machine has 38 accords. An accord consists of between five and 20 ingredients.

I'm going to interrupt Frederick for a second and let you know. that these accords he's referencing are scents made up of several perfume notes or ingredients. They are like mini compositions, mini building blocks where you can work with. And then a perfume typically is built out of a thousand parts. So you have a thousand small drops make a perfume. So then you can start understanding that the amount of possibilities that we have is spectacular. Like it's really gigantic, what we can create and the variety we can create. person receives three cents from the machine, and Frederick admits they're not always a perfect

match. Once every X amount of time, one of the three fragrances that you get is a random perfume. So basically that sets our benchmark, because basically it tells us, okay, if we would give out random sense, how much would people like it then? And how much is contributed by the story or how much is contributed by the setting of the machine or how people interact with it online? you know, we really validate that effect. You know, sometimes it's super bad.

We almost don't beat random. But right now, in the last, I would say, two quarters of a year, it varies between, you know, that we go above 12% difference or sometimes even to 20% difference. So that's significant. And then basically we continue to optimize, like, with the best performing algorithm. So, yeah, no, we are making progress.

But successfully making these on-demand personalized sense is just part one of Centronics's loftier goal. The ultimate goal that we have is we really want to give you something that allows you to feel in a certain way that you want. So for instance, if you would say, hey, you know, I had a bad day and I want to relax a little bit, can you give me something that really physiologically helps you to relax? So that's an extremely complex because then you need to understand, like,

okay, what is your physiological response, but as well, what do you perceive? even if you say, relax, we need to interpret that. So there is basically a lot of parameters to be able to do that one. The other thing is we want to be able to give this to as many people as possible. And we want it to be unique. So I really want to give you something which is uniquely yours. And bless them, there are a lot of companies that you do a questionnaire with,

and then in the end you get a cent, but then those cents are not really customized. They're basically, you know, standard-made sense, and they're not uniquely yours. And so for us, it was really the thing, now I want to give you something which is really for you, uniquely made for you. And this gets extra complicated, Frederick says, because of those cultural associations he mentioned earlier.

Just because Mint is relaxing to me doesn't mean it's relaxing to you. Instead, a key component to measuring how you like a scent is monitoring your physiological response. Of course, Frederick is thinking about that too. We're setting up a research project, which is really about, okay, what physiological response do you get? So then we're going to measure, for instance, amongst others, is your arousal state, like your skin conductivityancy, what is your heart rate doing, what is your respiration pace doing? And then you really can go to the physiological level, because that's the beauty of scent.

Like scent, it's processed in the oldest part of our brain. It's extremely powerful. It can get you in your flight or fight mode without you being able to control it. Like if you would smell a little bit, the smell of burning plastic, but subconsciously just below your awareness level, you could already get, you know, more sweaty hands. Your heart rate could go up. That physiological response is going to be key to getting us to Frederick's dream world. If you go further down the line, we're working on a very small device that basically you can wear it, you can connect to your phone. And basically if your smart watch detects like, hey, you're stressed, it can give you a small puff because basically, you know,

You kind of optimize it in such a way that you're able to control your mood a little bit. So that's the end goal. But to get there, there are so many hurdles that we need to take. And we've only taken, I would say, a very small piece of it. That certainly sounds ambitious and likely a long way off. But once he and the team understand ingredients and the combinations people like, along with why they like them, they can start moving towards this hyper-personalized, AI-driven scent world. This is where I started to wonder if the AI itself could actually perceive a smell like humans do and make a judgment about it.

So that's why I called Socket Navalka. I'm a computer scientist working at the Simon Center for Quantitative Biology at Cold Spring Harbor Laboratory. And I'm interested generally in this field called algorithms in nature, which is this idea that biological systems have evolved. Interesting solutions to solve what are fundamentally engineering and computational. and that there's a lot we can learn from studying these types of systems. Socket studied fruit flies specifically to try and learn how they process smells and use that research to develop new algorithms. So the idea was how do fruit flies perform similarity searches. Okay, so the idea is that a fly, let's say, learns that odor A is something good and odor B is something bad and to be avoided.

But the next time the fly experiences odor A, it's not going to experience. experience it in the exact same way. It's going to experience, let's say, odor A prime, just because of noise and, you know, other things that are happening in the environment. But somehow, when it experiences this odor A prime, it needs to realize that A and A prime are similar enough to each other that it should go and maybe, you know, eat this thing that it's smelling. Okay, so how is it doing that? How is it like finding these similarity relationships between odors? And this is a problem that computer scientists face all the time. To be clear, Socket is not necessarily pursuing this research in order to help computers smell

or to improve the state-of-the-art in AI product design. Really, he's looking for ways to make searching datasets more computationally efficient, to allow systems to quickly sort and compare information, the same way that fruit flies can quickly sort and compare different odors in nature. So, for example, when you go to Amazon, you'll be buying a product, and they'll suggest similar products to that that you might. also be interested in. Or, you know, on YouTube, you'll be watching a video and they'll have this sidebar of similar videos to the one you're watching. So, so this basic idea of, you know, you have

this huge database of things that you've seen in the past, songs that you've heard, videos that you've seen. And then you get this query from the environment in real time and you have to quickly determine, you know, what is similar to this that I've seen before so that I can now know, you know, how to behave in response is really a basic computational problem that comes up in, you know, a lot of cases. This could help create a bridge between the world of organic senses and the digital world of data processing, including AI. If you understand how brains, even fruit flies brains, react to smells, you can maybe

help a computer understand that response. So now you have this smell, how is a person going to respond to that smell? Or how is it going to be perceived? What is it going to be perceived as similar to? because it's not just chemical structure that tells you how similar to odors are. You can add a carbon atom to a molecule and it'll be completely different, a completely different smell. So understanding the relationships between these and predicting the relationships,

I think is an important way to get at this question of how it's going to be perceived. And I think that these kinds of algorithms that we're developing are going to be useful for that kind of classification. These algorithms, Sackett says, could have other use cases, well. For example, they could help pick up on very faint scents in high-stakes scenarios, like in airports, where scent can be used to detect security threats. But software is only one part of the problem. And this is what Frederick was getting at, too. We need to solve for not just creating the scent, but knowing how people will process it, and then

determining whether they'll like it. Maybe the way to do this is by building an electronic nose. I think there's actually two aspects to actually make a successful product like this. So one is getting good sensors by themselves. So that's, you know, having, you know, in our nose, we have receptor neurons that bind to different types of chemicals, and they're really good at sort of sampling from the space of all possible molecules or odor compounds. And so if you don't have a good early representation or a good sensory measurements of what's going on, I mean, no matter what your algorithm is downstream, you're pretty much screwed. So I think there's sort of two

classes of problems that people are thinking about. One is how do we develop good sensors to sort of sample from the olfactory space? And then two is, okay, now that we have good sensors and a good representation, how do we then, you know, do the classification, the novelty detection, the background subtraction, and, you know, all the other more complicated things downstream. So we're sort of more focused on the second problem. The first is, you know, lots of other people are working on that. As Sackett points out, if computers are going to start processing data about the sensory world, they're going to need to have actual sensors that can take in information.

Basically, if you're using AI to design new sense and flavors, it needs to be able to accurately detect how things smell and taste. That's where our next guest from Sony comes in. So my name is Michael Spanga. I work for Sony AI. I'm based here in Tokyo, Japan. My background is in AI research. And one of the key things I think is missing or that we don't understand well enough yet

is this question of human creativity and how we can build AI systems that help in our endeavors for human creativity. The company is indeed working on its electronic nose or e-nose and also eat taste. It's almost like image sensing, right? So you're trying to get digital representations of food. We're related to how the food might taste and smell. But that's, of course, still different than human perception. And so then you get into this issue of essentially trying out different things, measuring how people respond to food.

And there are sort of different initiatives in this direction. So obviously large food companies have a really big commercial interest in understanding how people might react to their products or their potential products that are developing. So there's lots of data in that direction that typically resides in large food companies. But separately from this initiative, Sony is also working on AI that might be able to assist in recipe creation, particularly, for high-level chefs. We had really interesting conversations about their view, the chef's view, on technology and the potential for AI to solve and impact some of their processes and the problems that they might have in recipe creation.

And more often than not, some of the answers we're getting is really in this direction of, oh, I'd really love to know more about, you know, where certain ingredients are coming from, or I'd love to know more about the seasonality of ingredients, or I'd love to know more about the health aspects and sustainability aspect. And I think that's really, that goes to the core of why I think AI is important in this field, is really recipe creation or any kind of creative endeavor, I think it's really a process of constrained optimization. As with the use of AI in fragrances, the idea is that these people are working with extremely

complex datasets, whether for smell or taste, that machine learning sorts through to find hidden relationships. In this case, that might mean finding the connection between the conversexed. composition of a dish and the response of the person who eats it. Do they like this dish? Very, very quickly, the problem becomes very, very complex with different sources of information, continuous measurements of how food impacts human perception, all the way to, you know, the seasonality, these large kind of signals and changes that impact food and the performance of food but also the sustainability aspects. Like, when is it good to how much?

with certain food, I mean, all the way down to how we create food for consumption, more broadly speaking, like, how do we control the agriculture of the food in order to produce it and get the best possible performance, essentially all the food is an optimization problem. Recipe creation in cooking is a situation Michael sees as ripe for AI. A lot of the information that you find in recipes is really implicit. And recipe descriptions are very coarse. But somehow to us, they make this enormous amount of sense. And it's like almost like we can almost visualize the steps that are necessary to take in order to fulfill certain recipes.

But if you look at the actual information that's transported in recipes, like literally what it says, like what are the lines and what are the things that are not being said. And you as a person while you're cooking, you're filling in some of this detail. And so I think that's like a really interesting challenge for computers or more broadly speaking also for AI, of course. None of the Sony technology is in public use yet, so we haven't tried an AI-created dish. unfortunately. But Michael says that they might have more to say about this soon. Ultimately, though, this entire conversation leads us back to a question we asked in prior episodes of this show. What role does AI play? For food, will it design recipes entirely or inspire them? Is AI writing the menu, or is it just the sous chef?

I think AI is going to play a role in prediction. And I think that's maybe more important in your gastronomy space than it is in the image space, because of course in an image space, you can just look at an image. But in the gastronomy space, right, if you have a recipe, what you really have is sort of a texture representation of a potential taste. But that doesn't mean that you, by looking at the recipe, you can feel it on your tongue or you can envision the smell that this thing is going to make. And so you're going to have to actually go through the process of making the dish and then

trying it out. And my vision for the future would be my hope, my aspiration will be that AI can help in some of these processes by predicting what a specific recipe might look like or whether it fits a specific creative vision. That's a lot to take in, and to be honest, I'm hungry now. So we're going to take a break, and when we get back, James Vincent, the Verges AI and machine learning senior reporter, will be here. We're going to chat all things flavor, taste, smell, and AI, plus a little more on product

design. We'll be right back. Support for this show comes from Shopify. Every thriving successful business has to start somewhere. A good place to start is a relatively simple question. What if, given the right tools, I've really put my all into this. One tool that can help grow your sprouting business to new heights is Shopify. Millions of businesses around the world rely on Shopify for e-commerce.

They offer a host of helpful tools you can take advantage of, from payment processing to analytics to website design. Their design studio includes hundreds of templates to help you create the existing. Zach website you've been envisioning for your business. If you're wondering, what if I need help, then no worries, because you're never left to fend for yourself. Shopify's award-winning customer support is available 24-7. It's time to turn those what-ifs into a thriving business with Shopify today.

Sign up for your $1 per month trial today at Shopify.com slash vergecast. Go to Shopify.com slash vergecast. That's Shopify.com slash vergecast. All right. We are back and with us is James Vincent the Verge's senior reporter who specializes in AI and machine learning. Hello, James. Hello, Ashley. It is lovely to be back again and chatting machine learning with you.

How are you doing today? I am great. Good. So you've heard throughout this episode we're talking about product design as it relates to flavors and smells. and as I always say, I need to get your take. What are your thoughts here? What are your impressions? This is a really interesting topic for me.

And I feel like there's something about it, which is it's kind of difficult to talk about, right? I've written about some of these sorts of creative endeavors in the past. And the thing is that they are inherently less, I don't know, measurable, shall we say, right? If we're talking about what text can an AI output that a human can't, you can't just type that in there, you see what it comes up with and you can judge that. But when you're integrating AI into these sorts of product design processes, I'm a little bit more skeptical because I think it's actually very hard to tell what contribution the machine is making and whether that contribution is necessary. Now, I don't want to come across as too skeptical and it's been really

interesting to listen to, you know, these people talk about how they're using it. But I do worry that sometimes AI is used as a bit of a marketing flourish in these areas. And it's saying, oh, we had to use machine learning because it's just so incredibly complex. Whereas I feel like, well, incredible complexity is what humans are used to dealing with when it comes to sensations and tastes and smells and sights and this more aesthetic realm. I wondered, you know, how did you feel about talking to these people? Were you convinced that AI was a necessary part of the equation? You know, it was tricky because in some ways I felt like, okay, I've never made a perfume, for example.

There is probably so much that goes into it that I'm just deeply not aware of. So in that way, I was like, I'm going to have to take your word for it. This does seem complex. But at the same time, like you were kind of mentioning, it was hard for me to fully grasp why we needed AI to do these things. Like, for example, talking about having a puff of a certain smell delivered to us based on the readings happening in our wearable. Like that sounds cool. It sounds cool, no doubt about it. But I also think of like my mom who has a diffuser that I bought her and is like, today I'm in the mood for lavender. So in that way, it is kind of like, I don't know if this is for me a problem

that needs to be solved necessarily. But you're right that it is just harder to like qualify because with our previous episodes, you can see the work being removed from video editing and things like that. You can hear our voice clones. You can read the text and also hear in in this form, hear the text. And you're like, okay, that's really interesting. I cannot believe a computer did this. Yeah. Whereas here you're like, I don't know. Did a computer do this? Yeah. I don't want to, as I say, I don't want to express too much skepticism about this, because I do think that there is potential for AI to be used in these complex data sets. Like, what AI is good at is finding hidden.

relationships, it's finding connections, as you said, within these vast data sets. And I feel that there definitely is a place where that can be used. And I think AI can be really useful at sort of personalising stuff to you and trying to create something that is unique. So I think what AI does, it creates abundance, right? And then it is up to humans to try and choose what is the best of its output. And I think that's sort of like what we talked about in the text episode, right, is that you get these text generation machines and part of the good thing is that they can come out with a ton of content but you do still need a human element to winnow that down one thing that i've sort of tested that is similar to this is an app called endel e n d-el which creates these sorts of infinite loops of music

and they're supposed to be used for working or for working out or for going to sleep or for helping you focus, they're sort of generated on the fly. It's not pre-recorded stuff. And they're supposed to personalise it to your location, the time of day, and maybe if you want to add in some information about how you're feeling, they'll sort of tailor it to that. Now, the degree of tailoring, I think, is probably really simple. I think it's probably just a lot of if this, then that functions, which is not really machine learning. It's just basic programming. But the abundance that the thing can come up with is a factor of machine learning, I think. And I think there is a reason. something really appealing about this idea of having something that is unique and personal to you.

It may not be tailored exactly to your needs. It may not be like this perfect match of your desires, but it is still unique. I think that's kind of cool. Do you think, and this obviously came up a lot in the episode, we talked a lot about just the fact that there's this whole part of trying to personalize things that maybe we can't even access, which is this data around how we physically respond to things, specifically in the sense world, but maybe also in other areas as well. I just wonder if you think that this sort of personalized future is a far ways off until we have Apple watches that can do more intrusive, I guess, monitoring of our like blood levels or anything else. I have no clue. But like, yeah, if you just think that that part of the

technology needs to catch up before we enter sort of the true world where AI can really help personalize everything. Yeah, I think that's a fascinating question because I think that's like, It's sort of a core question of the digital age, which is what can we capture in digital systems? And can we capture every facet of human experience and emotion? Because I think if we're talking about smells and tastes and those sorts of aesthetic experiences, they are just, they're really hard to pin down. They're hard to describe if you're the one experiencing them. And I am a little bit skeptical about this stuff. I think there is knowledge in the world that is hard.

to digitise, to capture, to record, however you want to term it. And I think there's always going to, not always, but I think there is for the foreseeable future going to be a limit in how well computers can interpret this stuff. And this is why I mentioned confidence earlier, because I think it's, a lot of it is about confidence. If you think that the machine is clever enough to read your mind, then you sort of let it. Do you ever feel like this yourself? Like when you're getting something delivered that's tailored to you, I don't know what it might be.

Maybe it's like a workout or something like this or clothes. that are being recommended to you. Do you ever feel that actually, wow, the algorithm saw through to my, what I needed in this moment? Or do you feel that it's just getting lucky? Well, I think the only place I can think of where I really see this is less in physical goods, but more algorithms tailored to me, like my Instagram feed. Oh, of course, yeah. And in some ways, that is, I guess, a physical good because I'm thinking of the shopping feature

where I'm just like, why are these stores not in real life? Everything is so good. And so I guess I'm predictable. I mean, to some extent, that's more of like aesthetic versus sort of the senses and things like that. But similar. It's the same realm, right? It's taste. Yeah.

I can imagine also, though, if I shopped at a specific candle company and I was buying a specific thing all the time, they might be able to be like, you know, we see you really like this. Try this. We think you might like this one too. Yeah. So I feel like it can be in that way. I can be seen. Yeah, I feel that's a really interesting comparison, the shopping and the products, because, like, that is something that, like, is very personal, but it also follows these macro trends that can be mapped in easier way. And for me, that is more legible as data than what I enjoyed at a meal in the restaurant. I feel that comes down to what smell I like. Like, the smells you like are so damn personal. Like, what smell do you associate with, you know, people you love, you know, with your parents, with loved ones in your life? Like, is it because you like the smell?

Is it because you like the person? Like, I don't know. I do feel that there's this realm that I'm skeptical about AI being able to map fully because, yeah, it's based on those interpersonal experiences. We're just complicated beings. We are complicated beings. There's a lot of baggage. We carry around with us.

The AI will never know our baggage. Yeah. Machines aren't always going to be able to help us deal with this stuff. I like this as a place to sort of end on because I feel like that was the point of this series was looking at how AI intersects with these various industries and where we see them in use now. Yeah. And we've seen where they're in use now. But it sounds like to me the broader conclusion we've come to is that one, maybe we're replaceable at some point. But this human touch is kind of really needed right now. And these AI use cases are putting AI more in a position to be an inspirer, an a assistant, things like that, where it's not. fully driving the car, if that makes sense. Yeah, I know. I think that's completely right.

I feel that like we often get quite scared about the, what we see is the intrusion of these non-human systems into what we think of as human domains, whether that is writing, whether that is cooking. And we get scared even when it's only a little bit of our territory that's being impinged on. And it's just novelty, you know. And I don't think that machines are going to end up destroying the ability. of humans to be independent creative entities in the world.

I don't think that's how these things work. I think it just seems like a bigger danger than it is because it's so new and it's unusual to us. I mean, you've been doing all these interviews. Do you feel threatened by the rise of the machines in the creative industries? You don't sound like you, do you sound like you feel maybe reassured? Yeah, I was going to say, wow, I'm so happy we are getting to this point where I'm like, wait, I'm not anxious for once.

But what I will also say is going into this somewhat blind and really learning about AI through this podcast series, I'm coming out of being like, wow, like AI actually does a lot more than I thought, simultaneously does a lot more than I thought, but also a lot less, which I feel like is a really good place to be. Like, okay, I could do some really cool stuff. Yeah, I always get to that point in my reporting where like half of me is thinking, these machines, they're really up to some interesting stuff here. And the other part, I mean, he's thinking,

these machines are so dumb. Yeah, you just, you flip between those two extremes because it's new territory and you're trying to work out what the context is. And it's always very difficult. So, yeah, no, I'm completely with you on that. And it'll be interesting if in three or four years, maybe even less, but I'm thinking three or four years, we did the same series and checked in where things were then.

I imagine it's going to be wildly different. Yeah, yeah, yeah. I feel like a lot of our conversations about, where the future is going to go for this, there has been always loads to talk about and loads of loads of possibilities. So yeah, at three or four years from now, we'll be seeing even more. I have no doubt about it. I'll ping you on Slack for our reunion special. Let's put it in the G-Cal now. Let's get a date. We'll schedule it out now. Exactly. Surely there's an AI assistant that can help us do that. Probably. Awesome. Well, thank you so much, James. This was amazing. I'm beyond happy. We got to

collaborate on this show. You're the best. I hope everyone enjoyed hearing from James. And if you're listening, follow his work on theverge.com, find his Twitter. What's your Twitter, James? My Twitter is JJ Vincent. And thank you so much for having me, Ashley. It's been amazing fun to work with you to talk with you. Really enjoyed it.

Thanks again for listening to this Vergecast AI mini-series. This podcast is made by producer Liam James, senior audio director Andrew Marino, senior reporter James Vincent, and me, senior reporter Ashley Carman. Talk soon.