Marissa Mayer: AI to Make the Mundane Magical

Thank you very much for having me.

I've been fascinated by AI for a long time, and there's obviously people who have longer tenure in it than I do. What I will say is, when I was at Stanford there was this interesting major that ended up being my major called Symbolic Systems, which combined cognitive psychology, philosophy, linguistics, and computer science. The idea was: cognitive psychology is how people learn, philosophy is how people reason, linguistics is how they express themselves, and computer science: can you create something that can learn, reason, and express itself? For years in AI there was a lot of focus on the learning and a lot of focus on the reasoning in the models, inferences, and things like that. For me it's an oversimplification, but I have to say that I often was somewhat dismissive of the linguistics piece of it, being able to express itself. But it was really fascinating to watch as OpenAI released ChatGPT and people could actually chat back and forth and watch the artificial intelligence actually express itself in language that sounded human. I think that was really interesting to see that you really have to work on all three of those pillars as you're building AI, but it was really the expressiveness that helped capture people's imagination and really show people the potential of what generative AI could do.

I think it's been able to do it for a while because to me reasoning is just the abstraction of learning. So yes, I think that now we certainly see today that these models are reasoning quite well, but I think they've been reasoning quite well for a while. If you look back at other examples that were in many ways some of the underpinnings of what we see happening in AI today, translation: pulling in all the language from the web and understanding this webpage in English, this is in Spanish, and therefore beginning to make inferences if someone types a new, wholly unseen English phrase, how would you translate that into Spanish? Similarly with facial recognition, the fact that it can have me be tagged in one photo, but then in another photo where the angle, the lighting, the shading on my face are different, yet it still recognizes me. That's in my view these models really reasoning. It takes a set of learnings, very specific learnings, and abstracts them so you can actually reason and say, 'Well that was Michael, so that must be Michael as well.'

I think so, and I also just think that what you're seeing is a scale law playing out. So what looked like groundbreaking work in 2010 in translation, when you play that 15 years forward and you've got something that's rapidly evolving and scaling, it starts to look a lot more sophisticated. It's not just doing rote translation from one language into another anymore; it's doing complicated reasoning across hundreds of thousands of web pages on a topic and putting together a synthesized answer. So yeah, there's very complicated chains of reasoning that are at play now.

I don't know that we're actually approaching a scaling law, and I think you have to be working pretty directly with these models to really observe whether or not that's what's happening. One of the things I do suspect though is we've gotten really good at collecting information. I think that a lot of what you're seeing with a lot of the LLMs in particular is that the world of words online is vast, but it is in fact finite. So once you actually crawl all that information and you ingest it, yes there will be differences between the different engines in terms of how they process and use that information, but they'll probably be more alike than they are different. I think that's one of the things you're seeing: that consumption of new information to ingest and synthesize is in many ways getting to the 80/20 rule, or maybe even the 90/10. So that element: do we have new creative thought, new creative content for these LLMs to ingest and learn from that wasn't created by another LLM, which has its own difficulties? I think it's clear that they can consume information faster than humans can generate it, in large part. So I think we might be seeing some element of a plateau there, which I also think will lead to some amount of commoditization across LLMs where in many cases they may be more alike than different. When we get there, the value is probably going to shift to other aspects, like distribution or capability or specialization of one model versus another.

I think that what's happening on a product level with AI is that the most powerful technologies change end user behavior and they change end user expectations. We saw that in the early days of search. All the search engines were kind of the same, they could all kind of get things done, they weren't incredibly efficient. Then Google came along and had just better answers, causing people to search more on Google, causing them to search way more. There's way more internet search being done today than there was before Google existed. It basically caused people to ultimately change their behavior in terms of how often they searched, how they searched, what they searched for, and also change their expectation in terms of what they got. I think that what you're seeing now across many forms of generative AI, whether it's chatting to get a better understanding of a problem, getting editing help, getting visual help in visualizing something and creating an image or a graphic, the way that you can do that, the speed with which you can do it, and the ease with which you can do it is going to cause users' behaviors to change and their expectations to change. So I think that ultimately the groups that will be valued most in organizations as a result of this change will be the ones that can adapt, be most insightful about those user behavior changes and expectations, and adapt most quickly in response. I don't necessarily know that it will be a functional revaluing across companies, but I definitely think the companies that respond to these user behavior and expectation changes are the ones that are going to win.

Not surprisingly, I think that technology-rich industries are ones that are going to adopt this most quickly and readily. So I wouldn't be surprised if we see consulting actually do well in terms of adopting it, because they generally do use technology effectively and study its impact on HR and office workflows as part of what they do. But I think that anywhere where there's a tension between the new way of doing things and the new way your customer wants to do things, and a tension with your existing business model, those are the companies that have really hard choices to make. They can probably cling to that existing business model, ala your consulting example, and make more money in the short term, but in the long term they're really mortgaging their future. So I think that right now most companies need to be thinking about, and most people working in those companies need to be thinking about, how can AI make me better and faster at my job? There are all kinds of people who are doing studies now in terms of who does better: the consultant, the doctor, the lawyer, or the AI. If you go to the AI for suggestions, there certainly are some professions and some cases where the AI is outperforming the people, and if the people attempt to actually edit what it's suggesting, things only get worse. But it's early days and there are anecdotes right now that cut both ways.

Yeah, I've seen that study and I've seen some in a few other industries as well. Those industries where the AI solo can outperform, and the value that the people are adding could be negative, which is kind of the point that makes. I don't necessarily think that it was a big enough sample study to say that conclusively, but it definitely is thought-provoking. It does make you realize that where you should be spending your time and where we can add value versus where the AI should be adding the value is something that companies are going to have to be strategic about.

I think that advertising still will have value because I do think one of the critical signals that happens in search is that a result's ability to pay does affect its quality. The classic example is concert tickets. If I type in 'Taylor Swift concert tickets right now', there are lots of articles talking about the fact that the Eras Tour has been such a phenomenal seller, but if I type that in, I'm probably looking to buy tickets. There are a lot of vendors who are very economically motivated to get in front of you to ultimately achieve that sale, and I want them to get in front of me because that's what I'm looking for. It's counter to what a lot of people say, because they say if they have an economic motivation to be in the search results and they're willing to pay, then there's not a quality factor. But I think that Google has really shown over time that they do have a really good handle on good quality ads and formats, really using that ability to pay as a quality signal, not exclusively because you can't use it exclusively to understand if it's a good match to the search, but really including it. The other key thing that I think Google got right about advertising in this classic model is the search results and the ads are congruent. They're clearly marked as sponsored, but it's easy to absorb the information because it's in the same format as the search results around it. If search results fundamentally change their form, they're more synthesized, they are paragraphs that summarize content, people are going to want that. It means that when I'm doing that concert ticket search, I don't want to have to go through four different ticket vendors looking for the best price and the best seat. I'm going to want that brought together and be given the best available seats in the section or the best available seats for this price, and have the same type of synthesis applied to the paid results as you would have to the organic results. So I think that one of the things is that notion of what's native or what's twinning with the content, where you've got organic content and paid content but they live in parallel to each other. If the format is not 10 blue links, then it can't be 10 ads each with their own link; that has to change too. If we're starting to look at things that are more paragraphs, as I said, synthesized information, I think the ad has to follow that format. That's something that we know works well for advertising. So it may be quite disruptive to the type of format of what people are expecting.

Exactly. 'You want these tickets, and these are the best ones that are available, most reliable, and most likely to actually materialize from verified resellers.' That element could come in quite strongly. So I think that just like AI answers right now, those AI summaries at the top are in fact summarizations. Right now you have an ad summarization that's there as well that summarizes: 'Yes, you were searching for a TV. In summary, we have TVs that range in this size to this size and from this price to this price, and ultimately let you play with those parameters.' The company that will be most successful is probably the one that can accommodate the right format.

We do have a portfolio strategy; we knew it would always be that. We started the company to take mundane tasks and make them magical and easy, and apply AI where we could to everyday tasks. Today, what we're very focused on is taking photo sharing and fixing it using AI, helping apply facial recognition so we share our photos with the right people or the people we forgot to share them with, understanding how to cluster photos so we can say, 'You and your friends each took the same photo a bunch of times. This is the best one: most eyes open, best lighting, the most share-worthy,' a term we tend to throw around a lot. We're very focused on how to take photo sharing, make it easier, make it better, once you make it powered by AI. How does the AI make it better? We think that right now sharing photos works reasonably well when you text person to person. We also think photo sharing works really well on things like Instagram where you're trying to get your one amazing photo in front of thousands of people. But we think things are pretty broken when you're dealing with small groups, say five or more up to a few hundred people. Maybe it's an event, maybe it's a conference, maybe it's a soccer team, maybe it's a party. We have started working in some events basically because we know when someone's about to have an event, they're about to take a lot of photos that matter to them, and people want to generally share them at those types of events. But we think that type of photo sharing is really broken. So trying to understand groups and relationships among groups, the types of photos that you want to post to that group, and who should be in the group, those are all things that we're working on. Understanding people's photos and helping them share them more efficiently and effectively is just a really hard, fun, and interesting problem.

I've always worked in consumer tech and I love consumer technology. I've generally worked on consumer tech that people use every day in really small increments: search, email, Maps, stock quotes, sports scores. What I really love when I build a product is thinking about how people use it and how can we really simplify it and really kind of grind it down so they have a simple, beautiful experience where there could be very sophisticated technology working behind it, but yet it presents itself as extremely intuitive and straightforward to the user. That's really how I like to approach things. Coming at it from another angle, as you mentioned in the beginning, I've been fascinated by AI since my college days. When I look at the landscape, I think one of the biggest problems that we can work on today isn't necessarily the building of these models. That's very complicated, there's a huge barrier to entry. I do think to some extent there will be some commoditization across it. When you're doing a startup, there's almost nothing worse than saying, 'Look, there's a large barrier to entry that could be in the billions of dollars, and when you're done building it you might end up with a commodity product.' That's very scary and daunting. But at the same time, to me, I think that humanizing AI, making it useful every day for people, and having them understand how it can help them in these types of everyday tasks is one of the most important things that we have to get right in the field of AI. So AI isn't necessarily this abstract thing or this thing that I kind of play with when I want to chat or make pictures, but it actually helps me every day in something that I want to do. I think that's really important. I heard that a physics professor likes to ask people three questions: 'What are the biggest problems in your field today?', 'What are you working on?', and the third one is 'Why are those different?' For me today, I'm very excited about where AI is going, but I have real concerns. I grew up in Wisconsin. If I went to the center of the country and asked people, 'What do you think of AI?', they're really concerned. 'Is it going to take their jobs? Are we going to lose control of it? Will it hurt us?' There's a lot of fear in terms of what it is, and it feels like something you can't touch or use or understand. So I think creating applications that can use AI in a way that makes it really obvious how helpful it can be for people is something that is ultimately really empowering and can move the whole field of AI forward in a positive way.

What we found is we started building on contacts mostly because for what we want to do, contacts is foundational. If you're trying to understand groups and relationships, how people like to share, how they like to communicate, you really need to make sure you've got the right phone number, the right email address, the right methods of communication, social media handles for a particular individual. If you send someone an email or an invitation or a photo and they don't get it because it went to the wrong place, it doesn't matter. So we started off in contacts. But it's counterintuitive because contacts are of course in theory everyone knows, but contacts are inherently not that viral and not that social. You've never said to your friend, 'Hey, there's this great contact manager, you've got to try it.' That's true. I'm very proud of Sunshine Contacts. It is the highest rated and was App of the Day in the App Store. It's a great product, but it didn't have the type of social and viral growth that we ultimately want. So that's one of the things that made us think, 'Okay, how will some of these grow, especially that they are in spaces where there are really fearsome competitors?' So we thought, 'What really has to happen is we've got to work on spaces that are inherently social,' which is why we moved to photos and also implicitly events. We feel like the event space has not been disrupted much, and there's just a lot of friction and organizational overhead that we think we can make easier both with technology and with AI in particular.

Right now what we have is the Shine app on the phone, which organizes photos and it can do event photos and has great suggestions as you pointed out. You can also do things like sharing over time. We introduced this new feature, Streams, where you can define a family or a soccer team or a classroom and share photos over a few hours or over years. I think of it almost like WhatsApp for photos. Right now everyone dumps photos into WhatsApp, but WhatsApp is just not a great photo viewing experience. So what we try to do is build something for groups that really helps you share photos in that same type of conversational flow. So we have that app, and then on the web we actually have Shine Events at shine.sunshine.com. There we allow hosts, it works a lot like Evite or Paperless Post or Party Folds. It really seamlessly goes back and forth between web if you want to do email-based invites or link-based invites on mobile. It does both, but it uses generative AI to actually come up with really cool, witty, stunning, clever invites. We were doing a pizza party and we asked it just to create an interesting and witty pizza party invite, and it recreated the picture of the Last Supper with the disciples all eating pizza. You're like, 'That was a much more fun way to invite people over to your apartment for pizza.' My twin daughters were obsessed with dragons and they were turning eight, and it created this amazing intertwined two dragon bodies that made a figure eight as the invitation. So we see it doing really clever and interesting things that you can't really get outside of generative AI, where it's customized to your event, it's thoughtful, and it has a wit to it. We feel like it's a fun way to use AI again to humanize it and bring it into people's everyday lives as they're organizing their events. Then we can streamline that because we know if you're organizing an event, you're probably going to want to share photos of what you did there. When you RSVP to the event, you're immediately joined to the stream for that event, so your photos that you take there get joined with everyone else's.

For me, my company now is about the same size as Google was when I started. So it's been exhilarating to be back in that very early phase where everyone knows each other and we all know what each of us is working on and what we're getting done on a daily basis. I love that close-knit team. We have a tremendous team, I really love our team, and I love the work that we get to do every day. I love thinking about what users want, how we can make this better, how we can make this grow faster. I really love building product, so it's been really fun to be refocused on that. I will say that I miss scale. So now my goal is really to try and take Sunshine and build it into a scale that really marries the two, because I love working at scale and I love working on building product. To have something where I've had a foundational role and it operates at scale is really the hope and the dream, and that's what we're working very hard to do here.

I think it's maybe been the opposite because I always like to lead by example and I don't like to ask people to do things that I wouldn't do. So I will say that when I first got here, I was designing the business cards, building the desks, ordering the refrigerator, figuring how to get it installed. At some point you also have to say, 'Wait, if I'm spending time doing all of those things, what are all the other things that are more important in terms of the long-term success of the company that I need to be focused on instead?' So as exhilarating and fun as it was to be in service to the company and the team at those early days, and setting that example, you also have to strike the balance of saying, 'Okay, we've got to make sure that we're focused on the right things and that's where we're spending time as a team and that's where I'm spending my time.' I would say it's less of the big company piece because I just think by definition it's just different. There's an element of just, 'Look, I don't have to worry about how this is going to work for 20,000 people. I just have to figure out how this is going to work for 20.' I've always been a big fan of thinking about management in terms of orders of magnitude. I think that each order of magnitude things change pretty profoundly. So you go from tens to hundreds, or hundreds to thousands, or thousands to tens of thousands. Those types of scale events happen gradually over time, but that's when processes break and need to be redefined. Right now we're still at a stage where we haven't necessarily crossed those types of scaling barriers from a company size standpoint. So we haven't necessarily had to redefine processes; we've just tried to find what process works really well for us.

Right now Shine is really focused on events and sharing photos, but a lot of these are tasks that are requested by the user and they're somewhat mechanical: moving this photo here to there. Yes, we can do some magic with it in terms of grouping your duplicates and identifying people, creating a great event invite. But over time, I really hope that AI can become much more suggestive. I don't think that everyone will let all AIs in, but I think you'll selectively decide to let different applications that use AI into your life, allowing it to suggest, 'Here's the event we think you should host or the one we think you should go to,' based on what we've seen in your photos. We've noticed that you always like to ski with this person or you always like to play soccer with that person, and it seems like it's a nice Saturday in October or November, so probably you should reach out to that person and see if they want to do this. I think there's so much that can be analyzed and learned, particularly in the world of photos, where we could actually really help enrich people's lives and their interactions with each other by being able to go that extra step and not just necessarily do what was requested of us, but also analyzing and making suggestions around how people should spend their time.

You can do sunshine.com, or check us out on the App Store: Shine Photo Streams for Groups. And if you're throwing a holiday party, definitely check out shine.sunshine.com where you can see Shine Events.

Thank you.