Cerebral Valley: Marissa Mayer (Sunshine) in conversation with Max Child

Well, I just love building things, and I've been fascinated by AI since I was in college, and AI meant something very different then. So I'm glad that I started Sunshine and Shine, and I really saw an opportunity to try and apply really advanced technologies to everyday practical problems. But I couldn't have predicted what's happened in terms of the moment that AI is having now. So what we're doing at Sunshine, it's funny because one of the other names we considered that I think would probably have been better for this crowd was Mundane AI. We thought about calling it that because we were like, we're not building models; what we're trying to do is take AI and apply it to everyday obvious things that it can do really well. So right now we're very focused on photos and events. We're using generative AI to make events much more interesting, more beautiful invitations, more interesting descriptions, and we're applying AI on photos to basically try and solve small group photo sharing and basically photo messaging. How can we do facial recognition, near duplicate detection, and help understand who you should be sharing this photo with? And we're looping the two together because we know when people spend time at events, they often have photos they want to share afterwards.

Sure. Well, I think one of the things I've been surprised by, and I guess I probably shouldn't have been, but I do think events. I think people are just so visual, and I've seen it throughout my career around things like image search and later things like Flickr. Photos and images just say so much to people, but people are really electrified with playing with generative AI as well. So I think it's been really exciting to see how people like to use this tool that can really enhance their creativity. So I think that's been really fun and exciting. When we started the company, the very first place we started was contacts because we knew that eventually we wanted to work on groups and events and photos, but we also know that everyone's contacts are a mess, and if you try and share something or communicate with someone and they don't actually receive that, that's a problem. So we built a really great contacts app. It does a lot in terms of person disambiguation, finding publicly available information, figuring out their LinkedIn and all their social media handles, analyzing your email and signatures, and really fixing your contacts. But it's kind of counterintuitive because your contacts are all the people you in theory know, but contacts is inherently not viral. Just because people don't share contacts that much with each other, it doesn't ignite imagination the way that, for example, DALL-E does or being able to do generative AI images.

Sure. Well, I think that for me, the photo space really breaks into two areas. There's kind of archival beautiful photos, which is what in my view Flickr, Google Photos, Apple Photos has really built a name around. And then there's photos as communication, which is like Snapchat, Instagram, where the photo isn't necessarily memorialized over long periods of time, but it expresses something in the moment. And so we're much more interested in the latter, photos as communication. And one of the things we think is, if you are just going to text a photo person to person, that works, just go ahead and do it. If you want to take a great photo, put it out to express what you just saw or what you heard today or something that inspired you, send it out on Instagram. But if you're trying to share a photo with 10 to a few hundred people, that experience really doesn't work. And being smart about which photos get shared, finding the best photo of like five or six that a few different people took, things like that, AI can actually really help with the behind-the-scenes work that makes that experience much better. So we really think of the app as what WhatsApp would be if it were photos-based as opposed to chat-based. Because if you think about it, right now the experience of trying to share photos even in WhatsApp, which is a great application overall, is pretty clunky. Someone drops in a big set of photos, you scroll through them, you can't go to the other set of photos someone else threw in there. So this is basically a group messaging app that is much more photo-centric.

Well, we just put out the most recent version of Shine, which is the app that's in both the Android store and the iOS store for this photo messaging, 10 days ago. So it's pretty new, and we're still going through the post-launch different fixes, things that we learned, things that weren't obvious before we launched. So we're working our way through that. We are excited about different things that we think can overall ease that photo sharing experience, in terms of whether it's on the web, whether it's an app, things like that. That's kind of what's up next for the photo piece. And we also just want to continue to make it work better and better for people, particularly in the world of photo sharing across different events and across different sectors, and really trying to understand relationships, how people like to relate to each other, what context they relate to each other in.

Well, I think from a product perspective, one of the things when I think about products that I like to look at is changes in consumer behavior and changes in expectations. In the early days of Google, for example, I saw in my own behavior that I was searching a lot more as search got better. You searched a lot more as your expectations in terms of how high quality the results needed to be went up. And lately, what I've been seeing both in my own behavior and people around me is that there's a lot of queries that would typically have been keyword search that are now generative AI conversations. So I think that, and I do think the AI summaries that Google is largely incorporating into a lot of different search results are a very good first step. But right now they're still kind of bolted on to that original search experience. The old search experience is still there, but there's this new search experience, an experiment from Google's perspective, there on the top of the page. And I think those are actually quite good, but I think they give you an idea of what the future of search looks like, where people start to expect much more synthesized information. They don't want to have to sift through 10 links themselves. And that also fundamentally changes the business model, because while I think advertising will continue to exist, I think it will continue to be really lucrative for both the advertiser and for the search engine and also for the consumer in search results, I think the format of that advertising is going to change pretty dramatically. And to that end, I think it's a real rethink of the business model, and that can be not necessarily threatening or scary, but just disruptive.

Well, I think the obvious thing that comes to mind, and Google's definitely taking steps in this direction with things like their shopping site, is that you actually need to have a much better understanding of what it is the advertiser has to sell. They can't just say, 'Here's 140 characters that describe what I'm selling.' You actually need to know how big is it, what color is it, what varieties does it come in, how does the pricing work, all kinds of different aspects of it. Because one of the classic examples we used to talk about how ads make search better was concert tickets. So when people search for concert tickets, the fact that there's an advertiser there that has tickets to sell you and they're willing to pay to be in your search results is actually a sign of quality, and it's also somewhere where the searcher actually is happy. They don't want just articles about the concert; they want to see tickets to purchase. So there's a nice meeting of expectation on both the advertiser and the searcher side. I think now in the modern day, when someone says concert tickets and asks for a specific concert, they actually want to see exactly what seats are available, where they are in the stadium, the pricing. They want information synthesized much the way they see it synthesized with generative AI. So I think it means that advertisers are going to have to partner even more closely with Google and other search engines to really allow their wares to be showcased and synthesized into the answer. It's not necessarily comfortable, and I also think that Google's going to have to start to look at who are partners that can fulfill some of these inquiries that are best filled with advertisements.

I think that's why the tension has been there all along, and that's why to date the model has remained undisrupted in some ways. But I think it's pretty clear if you look at where search ads were 10 years ago versus where they are today, and certainly like where Google Shopping is, there are a lot more advertisers that are giving full information of their inventory and a lot of different aspects and facets of the data. So I think that trend is ultimately going to continue.

I've tried many of them. I find that when I look at my own behavior, I do find myself going to ChatGPT sometimes instead of Google search. I will say I often, obviously I'm very allegiant to Google, I really like the product and was involved with it for a long time. But I actually do think the AI summaries, I know that they've been sometimes controversial, but I actually think they're quite good and they are getting better over time, which is what you would hope to see.

I think that Google has seen challenges like this in the past. We used to see things like link farms. When we came up with AdSense, we had a moment where we had a real pause because we were making it possible to monetize content, and we saw people creating a lot of content in an automated way. So we had to figure out how we could look for duplication, things that paragraphs were cut and paste across hundreds if not thousands of pages just for the sake of getting more page views. We ultimately came up with algorithms to detect that and weighted appropriately. I think that obviously the challenge now is that sometimes it's very hard to tell what's been created by generative AI. I think the benefit that Google has is that they do have the index of the web running over the last 25 plus years. So in theory, they know which pages existed before November 2022 and what came afterwards. So in a crude way, they could probably do some weighting in terms of how long this piece of information has been around on the internet and therefore know that it could not have been generated by this technology. But I think they're also going to have to look for much more sophisticated ways that will work for a little while, but there's obviously new topics coming up every day, some of it written by humans, some written by AI. Sometimes actually the AI is something you would want to index and actually rely on. So they're going to have to come up with more sophisticated ways, but I think that longitudinal access that they have over the web and the information contained therein is something that they can actually leverage quite effectively both in their search engine and in the LLMs as they move forward for understanding what to train on, etc.

I haven't really thought that through, but I will say I have played around with it. For example, I know a bunch of different coding languages, but I don't know iOS at all. So I actually have been able to kind of hack together very simple programs using some of the generative tools, and it's definitely thought-provoking in terms of how people are going to program in the future. But I do remember back when I was teaching computer science at Stanford, one of the interesting things that came out which was counterintuitive to me at the time is that programming ability is actually more highly correlated to verbal SAT scores than to math SAT scores. A lot of people think computer science should be math-based, but a lot of times actually the best programmers are really strong verbally and in terms of being able to express really clearly what they want to do. And I think you're going to see that with these generative tools play out to an even greater degree, because if you can really clearly express what you want the program to do, that's going to be much more useful than the mathematical excellence of the zeros and ones that are passing around.

Yeah. I think it's just so clear these tools give people such a huge productivity boost and such a huge creativity boost that I would want all my employees using it for sure. And you can see a tension among some corporate leaders about how it should be used and when it should be used, and you especially see it in education, which to me is a little bit crazy because if it helps someone learn something, why wouldn't you let them use these tools? At the same time, you also understand what they understand without the tools, so I do understand the tension. But I think that someone made the point to me last week that all of the people in your companies are using these tools whether you expect them to and whether they're admitting it or not, at least the smart ones are. So I think you need to set up work streams and productivity loops that actually reward people who are using these tools and using them well.