Marissa Mayer Is Back with Dazzle, the New AI Assistant That Learns From Your Camera Roll

1. 0:00 – 0:54

   Trailer

   1. CVCarlos González de Villaumbrosia0:00

      You're able to get information at a very granular level, right? So I'm curious to know how you're able to get the taste of a person based on their camera roll.

   2. MMMarissa Mayer0:07

      The photos I took of fun ski trips, it could figure out my mother's favorite ski brand, her ski style, where she was, and probably the appropriate ski length.

   3. CVCarlos González de Villaumbrosia0:15

      We're seeing the examples of Amazon shutting down access from agents, but we're also seeing Walmart, and you're a board member there, how they are creating an agentic protocol to allow agents to just sign up as an agent instead of a human. So curious to know your take on this agentic commerce and-

   4. MMMarissa Mayer0:27

      So the problem is that they're kind of like people, but they have this always-on superhuman productivity level that just means that they are, in fact, slightly different. And so I think having a protocol for them to interact with commerce websites so they're not constantly crawling and scraping and trying to get prices, and they can do this in a way that's, that's more metered and less wasteful of resources, really.

   5. CVCarlos González de Villaumbrosia0:46

      How would you define success now as a startup founder for you?

2. 0:54 – 3:36

   Intro: Carlos welcomes Marissa Mayer, founder of Dazzle

   1. CVCarlos González de Villaumbrosia0:54

      Hey, this is Carlos, CEO at Product School and your host on The Product Podcast. My guest today is Marissa Mayer, Google's first woman engineer, employee #20, then VP of Search, then CEO of Yahoo for five years. She also started the APM program at Google, which is the reason Product School exists. Today, she's a founder again. Her company, Dazzle, just launched, and it's a personal assistant that learns everything about you from your camera roll. Here are the things we'll cover: 40% of your photos are not memories, they're tasks. How it picked her mother's ski length from vacation photos. First task done in 45 seconds. No connectors, no setup. Why her agents never impersonate you. They act on your behalf. Ten engineers shipping like 100. Let's get into it. AI helps us generate ideas, draft plans, and spin up work in minutes. But faster doesn't necessarily mean better. When everyone has their own AI tools, their own outputs, and their own context, you get fragmentation, fractured version control, scattered review processes, decisions happening in parallel, not in concert. Rebuilding context between people and platforms costs you time and momentum. That's where Mural comes in. Mural gives teams a shared visual canvas to bring work together. You can generate structured starting points right on the canvas instead of moving outputs between tools. Keep all of your context in one place. And compatible AI agents can contribute alongside the team, bringing in context from the tools they already know. The future of work is still human. Learn more at mural.co/mural-ai. Welcome to The Product Podcast, Marissa.

   2. MMMarissa Mayer2:44

      Thank you very much for having me. I'm excited to be here.

   3. CVCarlos González de Villaumbrosia2:48

      You know, I thought you didn't need an intro, but then I talked to a couple of, uh, Gen Z employees at my company. I told them, "Hey, guys, Marissa Mayer is coming on the pod." They're like, "Who?" Like, "You don't understand." You know? Like, you are the main source of inspiration behind this, my company. Like, you started the Associate Product Manager program at, at Google, and in a way, that inspired me to start Product School as the first associate product manager program outside of Google.

   4. MMMarissa Mayer3:13

      So well, yeah, the Associate Product Manager program is, is amazing. So you know, and it's what really is defines it is the people in it. But yeah, I'm excited to hear that it inspired the work you're doing 'cause I just think product thinking and product management is so important, and it's also evolving so much with AI.

   5. CVCarlos González de Villaumbrosia3:30

      The rest of the Gen Z-ers who might not know who you are, obviously, you joined Google as one of the first 20 employees as a software

3. 3:36 – 4:22

   From Google’s first woman engineer to Yahoo CEO

   1. CVCarlos González de Villaumbrosia3:36

      engineer.

   2. MMMarissa Mayer3:36

      I was Google's first woman software engineer, and I was in the, the first 20 or so. I interviewed when there were just seven people there and started in the first 20. And then I stayed there as a software engineer, product manager, then became director product of consumer products, then became the vice president of search and user experience, and then in the end, was the vice president of maps, local, and contextual searches, which eventually became Google Assistant. And then I went to Yahoo. I was CEO there for five years, and since then I've been doing startups. So I had a startup, Sunshine, uh, and now my second startup, Dazzle, which bought my first Sun- m- my first startup, Sunshine, is, has just recently launched. So hearkening back to days of Google Now and Google Assistant, we're building a personal assistant

4. 4:22 – 5:20

   Why go back to founding a startup after big tech

   1. MMMarissa Mayer4:22

      using AI.

   2. CVCarlos González de Villaumbrosia4:22

      So what we are going to talk about all things with your new startup, but the first question I have is what made you consider becoming a startup founder after having such an incredible career as a big tech executive?

   3. MMMarissa Mayer4:33

      Well, I've discovered that I just love all phases of companies. I like small. I like single digits. Every order of magnitude, single digits, tens, hundreds, thousands, tens of thousands. I like being an IC. I like being a manager. And I love all phases of companies, hyper growth, turnarounds. And so I really wanted... was inspired to try and build something. In my career, I've been lucky to be surrounded by amazing founders, Larry, Sergey, Filo and Yang at Yahoo, obviously, and many others of companies we've acquired, you know, teams inside of companies. And so I really wanted to start something, and I'm very passionate about consumer tech. Most of my career has been in consumer technology, so I wanted to start a company to see if I could have an impact on the world the same way I had, had, had the opportunity to at Google and at

5. 5:20 – 6:21

   What Dazzle is: a personal assistant that learns from your photos

   1. MMMarissa Mayer5:20

      Yahoo.

   2. CVCarlos González de Villaumbrosia5:20

      So tell me more about what is Dazzle?

   3. MMMarissa Mayer5:23

      Dazzle is a personal assistant that learns from your photos. So we use your camera roll. We ask for permission. You give us full select or no per- no photos at all. And the more photos we have, the better we work. And what we do is we look at your historical photos, your past photos, to try and s- understand your tastes, your lifestyles, important people and places, and we do that at a pretty granular level. I'll come back to that later. And then we-- the other thing we do is we look at the photos you took today to try and make your camera button a get it done button. If you need to buy something, you need to fix something, book something, add it to your calendar, take notes that was written off a whiteboard, you just snap a picture of it, and Dazzle figures out what your intention is. We ask the question, "What, why did Carlos take this picture today?" And then we try and help you get it done. So we basically form a to-do list, but we call it a get-it-done list because it's a to-do list where it does itself with AI agents.

6. 6:21 – 7:26

   The key insight: 40% of camera roll photos are tasks, not memories

   1. CVCarlos González de Villaumbrosia6:21

      One of the conversations we had before this, you mentioned something around 40% of the camera roll photos are actually tasks and not memories.

   2. MMMarissa Mayer6:29

      Yeah, that's the transition point from Sunshine to Dazzle. At Sunshine, we were doing various utility applications, one of which was photo sharing, and we had built a really excellent shareworthy model. When was it likely that a photo was worthy of being shared? When we were looking at that, we were looking at precision and recall, and we realized that only about 60% of the camera roll, even after you remove blurry photos or du- or near duplicates, was actually m- passing our shareworthy model. So then that begs the question, well, what's the other 40%? And the other 40% is generally things you need to get done, things you need to remember, things you need to coordinate with someone else. And so we thought, well, if we have an AI agent that's on the camera roll waiting for you to take a picture and then says, "Okay, if it's just a happy memory, I'll let it go. I might try and glean some information from it for future, but if it's something that you might need help with to get done, it can swing into action and offer to do those actions."

7. 7:26 – 10:05

   Why photos are the richest, lowest-friction personal data

   1. CVCarlos González de Villaumbrosia7:26

      You mentioned before that you can, you're able to get information at a very granular level, right? So I'm curious to know how you're able to get the taste of a person based on their camera roll.

   2. MMMarissa Mayer7:36

      Photos are the richest source of personal information available with the lowest friction. So some examples of things that we've done. Just a few weeks ago, we had a user, we were doing a session here to understand how people use Dazzle, and, and she said she was a prolific reader, but she had a list of books she needed for book club that she hadn't bought because she feels like she had some already. So she said, "So to use Dazzle, should I just go and take pictures of all my bookshelves?" And I was like, "Maybe, but let's do something else first. Why don't you just give Dazzle the reading list and see what it does?" And she's like, "Well, how will that work?" And I was like, "Let's just see if it does." And sure enough, she gave it to Dazzle, and Dazzle came back and said, "Actually, you do have this book, and you don't have these other two books, but you do have books by the same author." And I said, because basically the principle is, like, unless your bookshelves are in the closet, when you take a picture in your living room of family or friends, of cute kids, right, we're also getting all that information in the background. Dazzle knows, for example, all the appliances in my house, makes, model numbers, everything, 'cause it's all in the background of my photos, just like it knew this user's book- what was on her bookshelves in a really granular level. Another example, early in our prototyping phase, my mother wanted some new skis last ski season. It was right after we had d- gotten Dazzle kind of up and running. And so I said, "You know, my mom would like new skis." And Dazzle came back and said, "Well, because Margaret," 'cause it knows Margaret is my mother, but I hadn't told it that, "Because Margaret is 76 years old and she often skis in Jackson Hole, Wyoming, and she usually wears Rossignol skis, these are the skis we would recommend for her." And then I even, uh, got a recommendation on length, and I said, "You know, she would like-- She's thinking about the 164s versus the 157s." And it came back and it said, "Well, because Jackson Hole is very steep and because she's actually, you know, a little bit on the older side, shorter skis will be easier to maneuver." And I said, "My mother's pushing back because she says her friends all have skis that are in the 160-centimeter length." And it said, "Yes, but in the photos, your mother is usually shorter and slighter than her friends and therefore needs a shorter ski." [laughs] Like, which is, like-

   3. CVCarlos González de Villaumbrosia9:48

      [laughs]

   4. MMMarissa Mayer9:48

      ... really, I mean, just, like, incredible that it, like, from the photos I took of fun ski trips, it could figure out my mother's favorite ski brand, her ski style, where she was, and even the, probably the appropriate ski length. So there, it's greatly amazing things

8. 10:05 – 11:18

   Is there a moat? IP, patents, and moving fast in AI

   1. MMMarissa Mayer10:05

      like that.

   2. CVCarlos González de Villaumbrosia10:05

      It's an amazing angle, and one of the things that come to mind as, as I think about product is we also have examples of how companies saw something happen and then copied that, right? In social media, you have Instagram copying stories from Snapchat. Now we're seeing the same with AI meeting note takers, Granola and Whisperflow copying each other all the time. Like, how are you thinking about keeping your moat?

   3. MMMarissa Mayer10:26

      Well, I mean, I think that in the fast-moving world of AI, there isn't much of a moat. We do have a lot of intellectual property. I do think that we were, you know, we found-- we've created some really amazing technology here, and we have patents pending on what I think will be foundational patents in terms of how to use images to help LLMs and AI understand context for a person. We were thoughtful about really wanting to protect that intellectual property 'cause we think it is unique, but I think that there are also things that are obvious to those skilled in the art. Now there's a whole crop of personal AI assistants that have, have launched, and it was interesting is, like, many of them launched invite-only, as we did. Many of them launched on iMessage, which we've been working on since the spring. So it's interesting to see some of the ideas went, as we were building the product, validated now by other competitors, though the photos piece does

9. 11:18 – 13:00

   Why Mayer is betting big on consumer AI over enterprise

   1. MMMarissa Mayer11:18

      remain unique.

   2. CVCarlos González de Villaumbrosia11:19

      Look, your product is focused on consumer applications, right? It's not in a business context.

   3. MMMarissa Mayer11:23

      That's right, and we think that, you know, I think consumer AI is going to be much larger. I've been fascinated by AI since my college days. I've always worked in consumer. I think consumer AI is going to be huge. And when we were pitching Dazzle last year to VCs, one of the things I learned was that basically today, only about 6% of investment in AI goes to consumer, yet the consumer market is expected to be twice the size of the enterprise market by 2030. So it j- basically just points to the fact that there's a big opportunity there. And it also means that I love to build for consumers because I think I'm-- it's easier for me to adopt that consumer mindset, that user mindset, to really build something they want and that works for them. But it presents unique challenges. So one of the things that we've been most fixated on at Dazzle is how quickly can we get someone who isn't tech savvy up and doing their first task with an AI agent? Because when you look at things like OpenClaw and many of the others, you have to set, do so much setup, connectors, everything else. You really have to understand the technology to do it. And it's a point of pride that at Dazzle, from the moment you first open the app to actually doing your first task on your get-it-done list, is usually 45 seconds or less. And I think that's remarkable. I have friends You know, I grew up in the Midwest, and so, you know, I have friends I've, I've tested the product with, and they just can't believe. They're like, "It doesn't even really necessarily-- It doesn't feel like there's a lot of setup. It doesn't feel like you're using an AI agent." It really just feels much more fluid and natural, and it's really fast, and I think that's really a key element for consumers.

10. 13:00 – 13:39

    The AI assistant supercycle: no new app to download

    1. CVCarlos González de Villaumbrosia13:00

       I agree. I think the reason why they-- we're leading through this super cycle with AI assistance now is because finally regular consumers can use these capabilities within their current apps, like WhatsApp, iMessage. You do not need to download anything if you don't want to. You don't need to worry about connectors. At the same time, Marissa, I see that you raised around $8 million. At least that's what was publicly reported, right? You have-

    2. MMMarissa Mayer13:22

       Yes

    3. CVCarlos González de Villaumbrosia13:23

       ... a, a nimble, a nimble team, and you are up against Muse, Instinct, GrokBot. Now I'm sure OpenAI and Claude are also going to come up with their own solution for this. So how are you thinking about staying competitive given that they can, they can subsidize this product for a very

11. 13:39 – 14:46

    Dazzle is free today: focusing on early users and engagement

    1. CVCarlos González de Villaumbrosia13:39

       long time?

    2. MMMarissa Mayer13:39

       Yes. Well, Dazzle is currently free as well, so we also are subsidizing. But I think that, you know, for us, we're really focused on our early users, what do they like about the product, and the feedback's been very strong. The engagement's been re-really incredible, as has the retention. So these products, when, when they meet users' needs, are very sticky. So we need more distribution, so we're working on, on that, and we also need to keep evolving for new features 'cause it's a really fa-quickly evolving space. But we're really focused right now on the users and on building the best product while keeping an eye on how do we get this in the hands of as many people as possible and keep it really simple, keep it really easy to set up and use. Because in my view, there's a s-shockingly small number of people who are really using AI at all, let alone to its fullest capabilities, and the gap in terms of what people know to be, th-that AI can do is just getting ever wider between the people who use AI and the people who don't. And that's really one of the things that Dazzle's designed to address, is how do we take people who, uh, are maybe skeptical of AI, hesitant to use AI, in a way that really helps them in their personal lives outside of the work context. It's something that's just purely a help. It's trying to give you the gift of time.

12. 14:46 – 17:09

    Agentic commerce with a human concierge for high-stakes buys

    1. CVCarlos González de Villaumbrosia14:46

       And so speaking of monetization, you mentioned the example with your mom, that she was recommended a pair of skis, and I saw that you also have a human concierge team, right? That in a way is acting as a human in the loop, depending on what products are being recommended. So how are you thinking about the agent e-commerce application on top of your own platform?

    2. MMMarissa Mayer15:03

       Yeah. We do use agent e-commerce, and the vast majority of transactions that Dazzle does go through agents. But if something is particularly expensive, particularly complicated, we route them to our human concierges for a final check. In the end, this might end up being something that's done where we have agents checking agents. But right now, we know one of our investors is Sam Lessin. He previously ran a personal assistant company called Fin, and he commented to me, "You know, you have no idea how quickly you lose a user when they're expecting red shoes and green shoes show up." And I was like, "You believe it," because, you know, people are picky, and when they decide to splurge or use a personal assistant to get something they really want, it needs to be right. And some of these soft goods in particular are really complicated. Last week, I was trying to buy something that came in clear, mirrored clear, and mirrored smoke. And I mean, you know, the agent understandably, understandably got confused. We saw a low confidence signal, and we intervened, right? And to just make sure, like, that we could actually get it right. But there's also, commerce itself is really complicated. For Halloween, my daughter, uh, was doing a group costume, and they decided on a T-shirt that came in a two-pack. And we went to order the two-pack of T-shirts, and it was sold out. Not available till November 17th. Consternation, like, really upset about it, and it took me a minute to think about it, and then I was like, "Oh, maybe the two-pack, maybe the black T-shirt we need is available in a different two-pack." And sure enough, while black and red was sold out till November 17th, black and light pink was, was available. So we bought it, got the T-shirt, no problem. But that's hard at this point. Maybe in 2027, agents are gonna be able to figure that out, but that, you sometimes need that human intuition of being able to see, or human insight to be like, wait, given how this is packaged, I wonder if I can get the same product packaged a different way. And sure enough, you could. So there are things like that where in personal assistant tasks, if you can figure out a way to make it happen, that's what you have to do, and that's what our, our concierges are focused on.

13. 17:09 – 17:34

    The web is built for humans: Amazon vs. Walmart on agents

    1. CVCarlos González de Villaumbrosia17:09

       All righty. I think the other challenge we are seeing with agents trying to do commerce is that the web is still designed for humans, right? So we're seeing the examples of Amazon shutting down access from agents, but we're also seeing Walmart, and you're a board member there, how they are creating an agentic protocol to allow agents to just sign up as an agent instead of a human. So curious to know your take on, on this agent e-commerce and the agents are going to be able to, uh, identify themselves as an agent and not just as a fake human.

14. 17:34 – 20:06

    Why proper agent-commerce protocols fix so much waste

    1. MMMarissa Mayer17:34

       Well, I think agent e-commerce is a really good, uh, approach to it, and I think having proper protocols for agents just eliminates a lot of problems. A few months ago, I remember I was reading... Because basically, these agents, they come in and they look like humans, but then they behave almost like a denial-of-service attack. I remember reading about one story where someone was trying to find kiss cam footage of themselves and their boyfriend, and they sent their agent out to find it, and I think it emailed a production assistant at a local TV station 1,800 times in nine hours. Like, you know, so the problem is that they're kind of like people, but they have this always-on and, like, superhuman productivity level that just means that they are, in fact, slightly different. And so I think having a protocol for them to interact with commerce websites so they're not constantly crawling and scraping and trying to get prices, and they can do this in a way that's, that's more metered and less wasteful of resources really, I think is, is actually a good and prudent idea. Here at Dazzle, in the meantime, though, we have a slightly different view in how to approach this, which is we like to do things on your behalf as opposed to as you. So, what that means is our agents, both AI and human, have their own email addresses, their own phone numbers, their own credit cards, their own accounts at most of the major vendors, and therefore, they can go and they're not acting as Carlos when they go to do a purchase. 'Cause we've also seen instances where when AI can impersonate you and it does the wrong thing, it can cause big problems. We've heard stories about canceled flights with other assistant apps, and you'd be like, "Well, how can that even happen?" But the AI is calling the airlines pretending to be you, and they have all the credentials to make that happen. I have a personal assistant, and I understand that it's a point of privilege and it's a, and it's a real luxury, but she never has to impersonate me, right? She always calls as herself on my behalf, not as me, and she's really effective in that way, and we think that that's, at least at this current moment of AI, how we should approach it. When you go and buy something, it's either the AI agent buying it as the AI agent, and they have names and accounts as such, or it's the human, and then we ultimately have you reimburse us for the cost, and we obviously get the item shipped to you. But the way we behave in this agentic commerce model and agentic inquiry model, when we're look at, doing things like booking home repairs and things like that, is quite different.

15. 20:06 – 20:37

    Behind the scenes: Mayer’s real involvement as CEO

    1. CVCarlos González de Villaumbrosia20:06

       Got it. So I want to switch gears for a second and talk about the magic behind the scenes. Because you have a nimble team, and you are working hard, right? So first of all, start with what is your real involvement as CEO of the company?

    2. MMMarissa Mayer20:19

       You know, I'm here five, six days a week in person, and I'm working every day in person, but I love it. I have a great team. I think we have a great idea. It's been really amazing to see how our early users respond to it. It's been really positive, and so it's a lot of fun, but this is my full-time

16. 20:37 – 21:25

    The secret sauce: how ten engineers compete with giants

    1. MMMarissa Mayer20:37

       gig.

    2. CVCarlos González de Villaumbrosia20:37

       What's the, kind of the secret sauce, in a way? How can a team of ten engineers and a few others, like, really compete with those giants?

    3. MMMarissa Mayer20:45

       I think one is focus. As a smaller company, we have to. It's not a luxury, it's a necessity that we have to decide this is what we're gonna focus on. We're gonna focus on how do we take photos and make them incredibly useful? How do we work on this task? So a part of it is just making sure you're not spread too thin. But I also have to say, with the power of agentic coding, you know, we have a team of 10 to 12 engineers, and they code more like a team of 100, right? Because each of them is running so many different agents, right? We generally have most people running five, ten, sometimes even more agents at any given time to get tasks done. So we're able to produce in a way that usually would require, and, you know, just a few years ago, would've required many more people.

17. 21:25 – 22:23

    A model-agnostic architecture running 3-4 frontier models

    1. CVCarlos González de Villaumbrosia21:25

       In terms of, uh, model usage and, and, like, how you're architecting the, the company behind the scenes so you can focus on your competitive advantage, like, what can you tell me about it?

    2. MMMarissa Mayer21:34

       Sure. Well, the harness itself that's part of the Dazzle product is model-agnostic. So whenever there's a new model that comes out, we assess which tasks is it good for. And by the way, at any given time, we usually have 3 or 4 frontier models in play, and we also have a few smaller models, you know, some classifiers and things on device and some models on device, as well as, as some others that work on server side. So we're constantly looking at how can we swap out to the best technology? How can we use whatever is highest quality and most efficient? 'Cause pe- speed also matters for people, too, in the product. And here at the startup, we all use AI a lot, right? Both for coding, for strategy planning, for trying to analyze the data that we have coming off of our early users. It's been an invaluable tool.

18. 22:23 – 24:00

    Returning to the startup grind after big-tech success

    1. CVCarlos González de Villaumbrosia22:23

       So Marissa, you came out of stealth recently, and the launch seems to be going well. At the same time, you've achieved such a level of success, kind of in your previous life as a big tech executive, that there's a big gap to fill, right? So as you think about the future, when you look back, like, how would you define success now as a startup founder for you?

    2. MMMarissa Mayer22:44

       Yeah. I mean, I think that we wanna have a meaningful, positive impact on a large number of users. And so right now, I think that I can see the green shoots where our users really love what we're doing, and they really value the gift of time and how easy we're making it to do things. And I know myself, I find it incredibly useful. And so, you know, that is fulfilling for me to see that we're really helping people. Now we need to do it at scale. And to do it at scale, that means that, you know, we have to get more distribution. Also means we need to keep evolving the products. We need to win more and new users with new features and new capabilities. So for, for me, this really comes down to the positive impact on individuals' lives and then hopefully a large number of individuals.

Episode duration: 23:35