If I had to be your work assistant for a week, I'd want access to your email. If I had to be your personal assistant — booking dinners, figuring out what to do this weekend, buying gifts, buying clothes — I want to know what your life looks like. Because if I can see your camera roll, I can figure out important people in your life, what you like things to look like, how you spend your time, what your hobbies are. But for us, our base level context is the camera roll.
Hello everyone and welcome to MTS. Today we have an exciting guest joining us. We have Marissa Mayer, who is the founder and CEO of Dazzle. This is the new personal AI assistant that works from your photos. And Marissa was also formerly the CEO of Yahoo, and I just realized, the first woman engineer at Google, and then you worked your way up from an IC to being one of the early executives. What a crazy career trajectory. So we're super honored to have you on MTS today.
Thank you so much for having me. I'm excited to be here.
Thank you for coming. I did want to talk to you about Dazzle. This is this new AI assistant, and it's really based off of photos. It's not based off your inbox or your email or the traditional way that we would typically work with personal AI assistants. And so I wanted to just understand a little bit from your background. You've worked in consumer products for a really long time. You understand the space from a very big bird's-eye macro level. How did you guys end up developing Dazzle?
Well, I've been really passionate about consumer all along, and my degrees from Stanford are both in AI — symbolic systems and then computer science with a specialization. And so I've been really focused on consumer throughout my whole career, and I knew consumer AI was about to have its moment, and it feels like at least the first of what I think are going to be many moments is here.
And I've always been obsessed with personal assistants. In fact, actually, my thesis back at Stanford was about a personal travel assistant that you could operate by voice, 1997.
Wow. So you're way ahead of everyone else. Like, they didn't even — in 1997 people didn't even know to major in AI. That was just so niche.
Yeah. So anyway, it was a long time ago. But it's really quite funny because, you know, people have been hypothesizing about AI personal assistants for a long time. And we think at Dazzle the next frontier is really around user context. Why aren't LLMs as personalized as you want them to be? Why aren't they as proactive and aren't able to anticipate what you need? And part of that is because they just don't have your context. So sometimes a lot of different assistants are doing cold start, where it doesn't have any context on you. We've seen some people veer towards things like email or calendar. We probably will go those directions too, and in fact Dazzle does interact with your calendar. But for us, our base level context is the camera roll.
I've often said if I had to be your work assistant for a week, I'd want access to your email. If I had to be your personal assistant, booking dinners, figuring out what to do this weekend, buying gifts, buying clothes, I want to know what your life looks like, because if I can see your camera roll, I can figure out important people in your life, what you like things to look like, how you spend your time, what your hobbies are. And we really have seen that as we've developed Dazzle.
Previously, we were at my former startup working on photo sharing, and we were building a model to try and understand when do people want to share photos — and it doesn't mean just publicly post them, but also share them with friends and family. And what we found was that basically around 60% of the camera roll is shareworthy,
which then prompted the question, like, well, what's the other 40%? And that's when we realized it's all the stuff you need to do, all the stuff you need to remember, all the nuts and bolts of your daily life. And so it's the picture of the appointment card as you leave the doctor's office. It's the sweater you were looking at online and then realized you didn't have time to buy right then, and you screenshotted. Those are all the things that end up in the camera roll. So in addition to creating this context layer for our users that can help create a much more personalized AI, we can also create a task list — what we call a get-it-done list, because it's not a to-do list for you. We can create the to-do list, but then our agents can do it for you. We can actually go ahead and buy it for you, get a repair booked, book hotels, reservations, all those types of things, and basically make things a lot more seamless in people's everyday lives.
And they're already doing something — they're already doing what gives us the fuel to really fuel those insights, which is take photos. Five billion photos will be taken today, 98% of which are on smartphones. And that's all information that can help these assistants be really efficient, that's just already there.
Yeah. No, it's pretty interesting because there's many different personal AI assistants, especially in the last few months. I feel like we've been on a renaissance with Muse and Instinct and Grokbot, and now we have Dots and all these different ones. But the same thing where I have this assistant, it starts off knowing nothing. It's like a spond. It doesn't know anything about me, and then I have to individually pick which permissions to give it access to. It knows a lot about my work, which is nice because it has context to my calendar and my email, but doesn't know anything about me personally. And as you're describing that, that does sound a lot like my behavior. I screenshot the sweater that I want to see. I screenshot everything. Actually, screenshots are a huge part of my workflow. I'm curious what some of the hardest engineering problems are, because not only now do you have all this rich context, but you have to now take that context and then put it in a way that the agent understands what that context is for. So tell me from like the engineering standpoint, how does this work?
Sure. What we're doing is we build this context layer. It's essentially, you know, it started off as a RAG model. It's probably a little bit more sophisticated than that now. But when you take a new photo, we essentially have an agent that's sitting on your camera roll, sees that you took a new photo and says, "Wait, why did Sophia take this picture today?"
Taking context around time, location, right? Because it gets all the meta information off of the photo, but also gets all the contents in the photo. And from that moment, we can be like, "Wait, we think because she's already at home, she already owns this. Maybe it's something she needs to fix, or if it's something that needs to be replenished, maybe it's something she needs to buy again. If you're traveling, right, it means something different," and so on and so forth.
So we basically are looking for that new photo. So we've got one agent that's on your camera roll, basically looking for something new, really trying to pull together that get-it-done list: what are things that you need help with right now?
We have an agent that's answering your chats, but it's answering your chats using that context layer, saying, "Okay, if this is your question..." One of the questions I asked it was, "What skis should I get my mom?" And you know, where lots of other LLMs would come back with more questions than answers, Dazzle came back and said, "Well, since your mother is 76 years old, this is where she skis, you know, and she prefers Rosynill. These are the skis you should get her." And so it can come back with incredibly personalized insights.
And then we also have other agents that, one, get tasks done, and we have agents that are running to also look at proactive ideas because we know you and your personality and your likes and dislikes so well. We can actually suggest different things that you do. So we have a get-it-done section in the app, we have an ideas section, and we have chat. And there's different agents running on all of those. So yes, we're taking the context layer, which is really your history of your camera roll synthesized in a way to be useful for an LLM, and then we also have the agents that are working on task management and task execution.
I want to shout out our sponsor, Lovable. You know, the app you've been meaning to build, the internal tool, side project, or product you'd otherwise lose a weekend to. With Lovable, you get software you can ship today. That includes an editable codebase you can inspect and change, plus two-way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, O storage, hosting, payments, and more. So you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now, on to the episode.
It was pretty interesting. There was this tweet that went a little bit viral saying everyone's building personalized agents like a husband, where you give it a specific task and then it does that one task. But saying, I want a personal assistant that's like a wife, where it takes that mental load. It has all of this context about me and knows what I need before I actually need it, and it just does these things with all the hidden work without me asking. If I have to have the mental load of needing to know exactly what I need to do and spoon feed it, and that's the only way you get something done, it's not that helpful to my actual daily tasks. And as you're talking about this, I'm kind of getting the — oh well, this almost feels like the AI assistant wife, where it has all this context around.
Also, it's more personal, because I will say, you know, we love email as well, and we will probably go on to add email to Dazzle at some point. But we wanted to start with photos because email naturally tends towards work and it naturally tends towards tasks, where photos show you, you know, this is your favorite color over time, this is a hobby that you've enjoyed, these are places that you like to go. And actually, right before I came here, Neil Kosla just posted and said he had ordered flowers through an AI personal assistant and had shown up, and he actually told his fiancée it was ordered through AI, and she's like, "Don't use AI to order flowers." And I was like, "Actually, go ahead and use Dazzle to order flowers, because by the way, we know what you think looks beautiful."
So we can actually do a good job making sure the flowers look the way you want them to look based on your camera roll. And the other thing that we haven't talked a lot about is we also have humans in the loop. Right now we're using agentic browsing. There's a lot that we can do with agents, but there are limits to what those agents can do with fidelity. And especially once you start to get to high-value or very sensitive tasks like sending someone flowers, you actually do want a human to look it over and make sure that it's right, the same way a human personal assistant would. And we think over time we're going to push the limits more and more towards actual agents as opposed to humans. But that's something right now where last week we had someone actually replace their roof
using Dazzle. They took a picture of it. We sent out, you know, bidders. They chose the bidders. And when you're doing something that important and that high cost, you actually do want a human in the loop just to make sure that everything's really coming together the way that you want it to. Will that always be the case? Probably not. But today, we think the state of AI is it's sometimes good to still have a human hand on the wheel.
Yeah. No, that's really interesting that you bring that up about where humans actually need to come into play, what can be fully autonomized, but it's not everything. Not everything needs to go fully towards agents.
