Dazzle CEO Marissa Mayer Talks AI Assistant for Photos

SPEAKER_1 (0:02)

Bloomberg Audio Studios, Podcasts, Radio, News.

Ed Ludlow (0:09)

Most AI assistants start with a prompt. Marissa Mayer is betting yours should start with your camera roll. The former Yahoo CEO and longtime Google exec is launching Dazzle, a personal AI assistant that uses your photos and screenshots to learn what matters to you. Marissa Mayer joins us now here in SF. This is so interesting.

Again, the basics, we're used to saying, do this for me or through the voice, do this for me and the agent acts. This is live real-time monitoring of camera roll.

Marissa Mayer (0:40)

Yes. So what Dazzle does is a personal AI assistant that learns from your camera roll, and it does two things. One, it looks over your history of your photos, and it tries to understand you, who's important in your life, what are your hobbies, where do you like to go, how do you like to spend your time? And then we also have an AI agent that monitors your camera roll. So when you take a new photo or a new screenshot, it monitors and says, wait, why did Ed take this picture today? What can I do to help him?

Ed Ludlow (1:08)

Give me a case study. I have one for you, but you give me one first. A case study that helps the audience understand why this is a good idea, like why it works.

Marissa Mayer (1:15)

Sure, well, a couple of different case studies. So one is, the idea is like if I see something I like, I'm shopping online, I'm busy, and so I can take a screenshot, or if I saw something, I could take a picture of it. And Dazzle actually comes back to me and says, did you want to buy this?

And it can overall make it a seamless purchase. And the other thing-

Ed Ludlow (1:35)

We're showing that kind of case study on the screen right now.

Marissa Mayer (1:37)

Yes, exactly. And the other thing that it can do is, it really understands at a very deep level, things that you're into, things that your kids are into. My kids are off of school on Friday. I asked Dazzle, what should they do on Friday? And he said, you know, Maryelle's been really into magic lately.

Why doesn't she and her friend go to a magic shop, and they picked a specific one here in San Francisco, Misperceptions Magic Shop, where they can go buy tricks, learn tricks, and spend the afternoon. And so the fact that it could actually understand that my daughter is going through a phase where she's really entranced with magic is pretty fun and pretty interesting. And same thing, I asked to tell my mother to buy skis. And it said, well, because she's 76 years old, and because this is where she skis, and roughly her stature, and she loves the Rossignol brand, why doesn't she buy this pair of skis?

Ed Ludlow (2:30)

Based on photos.

Marissa Mayer (2:31)

Based on photos, because it's got pictures of her skiing.

Ed Ludlow (2:33)

Of those experiences. Let's talk about technology. How did you build this? How did you get to what we're showing is Dazzle AI today?

And tell me a little bit about the underlying technologies to support the agent.

Marissa Mayer (2:49)

Sure. Well, we obviously are using some frontier models to handle users' requests and chats. But what we're doing is we're also building a deep context layer. So we try to understand who you are and add that context as relevant to the queries that we see. So when you're chatting with Dazzle, we're using that context to augment and basically make it a more personal form of AI.

But in addition, we're also using an agent to screen your photos, new photos that you're taking, to understand how can we help, how can we turn these into tasks that aren't necessarily a to-do list for you, they're a to-do list for our agent because then our agents can kick in and say, how can we buy it? How can we book this repair? How can we add this to your account?

Ed Ludlow (3:31)

So that goes to the definition of an agent. Is that agent genuinely autonomous? It's able to make a transaction or an action that's real. In Dazzle's AI case, it is?

Marissa Mayer (3:40)

That's right. So our agents do go ahead and they are able to do most of those tasks. We also think it's important, because some tasks are either more efficient or are high value enough that it makes sense to have humans in the loop. And that's another thing that we do at Dazzle. So most of the work is done by agents. But for the more complicated and intricate tasks, we actually have humans in the loop that make sure that we've got high accuracy and high fidelity.