Inside Google's AI turnaround: AI Mode, AI Overviews, and vision for AI-powered search | Robby Stein

It feels like something has changed internally at Google. Just last week, Google Gemini hit number 1 in the App Store. I feel like nobody saw this coming.

Google's mission around having any information be universally accessible is a very enduring, very motivating thing, and it feels like with the AI moment, we can actually achieve that more than ever before. What I'm feeling now is just an incredible sense of focus and urgency. Things have hit a tipping point where these models are now truly able to deliver for consumers.

As ChatGPT emerged over the past couple of years, as Perplexity emerged, a lot of people were just like, “Google is dead. Nobody wants to sit through search results and click links.”

The core Google Search isn't really changing, in my opinion. We're not seeing that. People come to Search for a ridiculously wide set of things. They want a specific phone number, they want a price for something, they want to get directions. I think the vastness of that is underappreciated by many people. AI is expansionary. There's actually just more and more questions being asked, and curiosity that can be fulfilled now with AI.

You've built a lot of very successful products. You use this phrase, “embodying relentless improvement.”

You need to be the physical manifestation of 2 pieces of things. One is just relentlessness: complete effort that's always exerted in a direction of positive productivity. Then the second is to make things better. You have to always make things better. You're never content.

You built and launched Stories at Instagram. Back in the day, it was quite controversial because it basically took what Snapchat was doing really well and then said, “Hey, let's bring it to Instagram.”

Not every great thing is going to be invented by you. Facebook probably created the modern feed, but there's a feed for every single product. At the end of the day, you're just robbing your user base of an opportunity to have a better product.

Today my guest is Robbie Stein. Robbie is VP of Product for Google Search, and is responsible for essentially the entire Google Search experience, including the new AI overviews, AI mode, multimodal AI experiences like Google Lens, the ranking algorithm, and a lot more. He's at the forefront of one of the biggest shifts in Google's history, and has already made a massive dent in Google's trajectory. He's also made a massive dent in the trajectory of Instagram, where he was head of product and led the launch of Instagram Stories and Reels and Close Friends, and through that, grew Instagram to half a billion daily active users. He's also on the founding team of Artifact with Mike Krieger and Kevin Systrom, started two companies of his own. Very few people have had this level of impact on two global consumer products at this scale, and Robbie shares all of the biggest lessons that he's learned about building great and successful consumer products, along with a bunch of insights into where Google is headed in the world of AI. A huge thank you to Bart Stein for suggesting topics for this conversation. If you enjoy this podcast, don't forget to subscribe and follow it in your favorite podcasting app or YouTube. It helps tremendously. And if you become an annual subscriber of my newsletter, you get a year free of 15 incredible products, including Lovable, Replit, Bolt, n8n, Linear, Superhuman, Descript, Whisperflow, Gamma, Perplexity, Warp, Granola, Magic Patterns, Raycast, ChatPRD, and Mobbin'. Head on over to lennysnewsletter.com and click Product Pass. With that, I bring you Robbie Stein. My podcast guests and I love talking about craft and taste and agency and product market fit. You know what we don't love talking about? SOC 2. That's where Vanta comes in. Vanta helps companies of all sizes get compliant fast, and stay that way with industry-leading AI, automation, and continuous monitoring. Whether you're a startup tackling your first SOC 2 or ISO 27001, or an enterprise managing vendor risk, Vanta's trust management platform makes it quicker, easier, and more scalable. Vanta also helps you complete security questionnaires up to five times faster so that you can win bigger deals sooner. The result, according to a recent IDC study, Vanta customers slashed over $500,000 a year and are three times more productive. Establishing trust isn't optional. Vanta makes it automatic. Get $1,000 off at vanta.com/lenny. This episode is brought to you by Jira Product Discovery. The hardest part of building products isn't actually building products, it's everything else. It's proving that the work matters, managing stakeholders, trying to plan ahead. Most teams spend more time reacting than learning, chasing updates, justifying roadmaps, and constantly unblocking work to keep things moving. Jira Product Discovery puts you back in control. With Jira Product Discovery, you can capture insights and prioritize high-impact ideas. It's flexible, so it adapts to the way your team works, and helps you build a roadmap that drives alignment, not questions. And because it's built on Jira, you can track ideas from strategy to delivery, all in one place. Less chasing, more time to think, learn, and build the right thing. Get Jira Product Discovery for free at atlassian.com/lenny. That's atlassian.com/lenny. Robbie, thank you so much for being here, and welcome to the podcast.

Thanks so much for having me.

1. Google Consumer AI Surges

This is such a cool week to be recording this podcast. Just last week, Gemini—Google Gemini—hit number 1 in the App Store. I have it right here. It's still number 1 in the App Store. It's above ChatGPT. I feel like nobody saw this coming. I feel like everyone's always saying, “Google, what have you guys been doing? You build all this amazing tech, and where is it? Why didn't you have anything working in consumer? Why is ChatGPT winning? Why are all these amazing companies doing better than Google?”

First of all, let me just say congrats. I know this isn't all you, but I imagine you had some part in this, so just congrats.

Many, many more people, yes.

It feels like something has changed internally at Google. Things are starting to really work, especially on the consumer AI side. In terms of the growth, is Nano Banana a source of a lot of this recent growth, or is there something else going on?

People are really excited about—

Okay.

—Nano Banana, to be clear, very much so. But I think people are also recognizing that there are just so many cool things you can do across the Google set of products, and they've become quite powerful. I'm always shocked, even for things in Search, because we think they're very obvious—they sit right in the core Search experience. Then on X I'll go look and see, “Oh, I just found out about this AI thing,” and it seems very obvious, but I think a lot of people are just discovering how powerful these tools are now.

To go 1 level deeper, to your point, there's been all this incredible tech. You guys wrote the original Transformers paper that has powered so much of the innovation, and then it's just like, “Where has Google been? And actually, why aren't you building the thing that's winning?”

What has changed? Is it just, “Okay, we need a…”? Have there been major reorganizations? Have new leaders been put in place? Is there just a new philosophy in the past couple of years that has led to this moment where Gemini is now the top app in the world?

Yeah, I've been at Google now—this is my 2nd time at Google. I started at Google in 2007, did a bunch of things in between, and I've been back at Google now, so I can't speak to that whole period, for many, many years back to today. But what I can tell you about what I'm feeling now is just an incredible sense of focus and urgency to deliver great products quickly.

I think that is in part leadership, for sure. We work very closely with our partners at Google DeepMind and, obviously, across the organization. It's an incredible group of people, and also an incredible group of researchers and technical thinkers who've been thinking about this for a while.

When you have that energy, and when the product teams and the tech and research groups are working very closely together, we're able to move, and we're getting a lot done. I don't think there's any one thing that has happened. A lot of times, people ascribe a lot of momentum to a one-time change or a single person, right?

I find a lot of this is actually a compounding effect. Every month, you're ruthlessly improving the product or the models and getting better every day. Then it just hits this tipping point where people like it, use it more, and enjoy it. That's more of the feeling that I've had: we've had the right investment and focus, and then it just hit a moment where people are seeing the effects of that now.

2. Search Expands With AI

As ChatGPT emerged over the past couple of years, as Perplexity emerged, and with all these other chatbots, a lot of people were just like, “Google is dead. Nobody wants to sit through search results and click links. Why not just get your answer right there?”

It feels like that's not actually happening. It feels like you guys are doing just fine. What can you share about the state of Google Search specifically? Then we'll talk about AI Mode. How is traffic going? How is Search going, considering all these things are out there? What are you seeing in the data since the launch of, say, ChatGPT?

Yeah.

What’s interesting is that people come to Search for a ridiculously wide set of things—all kinds of things. They want a specific phone number, a price for something, directions, or a payment webpage for their taxes. Every possible thing you can imagine. I think the vastness of that is underappreciated by many people.

What we see is that it’s not changing. AI hasn’t really changed those foundational needs in many ways. I think what we’re finding is that AI is expansionary, and so there are actually more and more questions being asked and more curiosity that can be fulfilled now with AI.

That’s where you get the growth. The core Google Search isn’t really changing, in my opinion. We’re not seeing that. But you’re getting this expansion moment.

What we’re seeing, as a few examples, is that you can now take a picture of something and ask about anything you see. Google Lens, one of the fastest-growing products out there, is growing 70% year over year in visual searches, which is already at a massive scale. It’s billions and billions and billions of searches in that way.

You can take a picture of your shoes and say, “Where can I buy this?” Or take a picture of homework and say, “Hey, I’m stuck on question 2.” You can take a picture of your bookshelf and say, “What are the books I should get based on these books?” AI can help you with those things now. That’s an example of why I think there’s so much growth left and why we’re so excited.

Okay. So you’re not seeing the death of Search.

No.

3. AI Mode Rebuilds Search

Along the same lines, you guys recently launched AI Mode, which I don’t think enough people are talking about. I think you get there at google.com/ai. Is that the right URL?

Yep.

Is that the right URL?

Yep.

Okay, cool.

Yep.

I’ve been playing with it as we were prepping for this conversation. It’s really incredible. I asked it, “What is the best newsletter on product and growth?” It’s very smart. It said Lenny’s newsletter, so that’s my evaluation of how well this thing does.

Fantastic. One out of one. Perfect evaluation.

Perfect. Also, if you go to it, there are these recommendations for things to ask it that are just like, “Wait, how did you know I care about this stuff?” One of them is “Help me switch to product management,” right on the front page. I’m like, “How did you know?” It tells you that it’s based on your Google activity.

Talk about what people should know about AI Mode, and maybe what they don’t really understand about the power of this thing.

I can tell you there are 3 big components to how we can think about AI Search and the next generation of Search experiences. One is obviously AI Overviews, which are the quick and fast AI answers you get at the top of the page, and many people have seen them. That’s been growing very, very quickly. When you ask a natural question and put it into Google, you get this AI answer now. It’s really helpful for people.

The second is multimodal search. This is visual search and Lens. That’s the other big piece. You go to the camera in the Google app, and that’s seeing a bunch of growth.

Then, with AI Mode, it really brings it all together. It creates an end-to-end, frontier Search experience on state-of-the-art models to truly let you ask anything of Google Search. You can go back and forth, have a conversation, and it taps into and is specially designed for Search.

What does that mean? One of the cool things that I think it does is that it’s able to understand all of this incredibly rich information within Google. There are 50 billion products in the Google Shopping graph, for instance. They’re updated 2 billion times an hour by merchants with live prices.

You have 250 million places in Maps. You have all of the finance information. And then, not to mention, you have the entire context of the web and how to connect to it so that you can get context and then go deeper.

You put all of that into this brain, which is effectively this way to talk to Google and get at this knowledge. That’s really what you can do now. You can ask anything on your mind, and it’ll use all of this information to hopefully give you super-high-quality and informed information as best as we can.

You can use it directly at google.com/ai, but it’s also been integrated into our core experiences. We announced that you can get to it really easily. You can ask follow-up questions from AI Overviews right into AI Mode now.

Mm-hmm.

The same goes for the Lens stuff. You take a picture, and it takes you to AI Mode, so you can have this back-and-forth conversation. You can ask follow-up questions and go there, too. It’s increasingly an integrated experience within the core part of the product.

I imagine much of this is “wait and see” as you learn how people use it, but what’s the vision for how all these things connect? Is the idea to continue having AI Mode on the side, AI Overviews at the top, and this multimodal experience? Or is there a vision of somehow pushing these together even more over time?

I think there’s an opportunity for these to come closer together. I think that’s what AI Mode represents, at least for the core AI experiences. But I think of them as very complementary to the core Search product.

You should be able to not have to think about where you’re asking a question. You just go to Google, and today, if you put in whatever you want, we’re actually starting to use much of the power behind AI Mode right in AI Overviews. You can ask really hard questions. You could put a 5-sentence question right into Google Search. Try it.

It should trigger AI at the top as a preview, and then you can go deeper into AI Mode and have this back-and-forth. That’s how these things connect.

The same goes for your camera. If you take a picture of something and ask, “What’s this plant?” or “How do I buy these shoes?” it should take you to an AI preview. If you go deeper, again, it’s powered by AI Mode, and you can have that back-and-forth.

You shouldn’t have to think about that. It should ultimately feel like a consistent, simple product experience. Obviously, this is a new thing for us, so we wanted to start it in a way that people could use and give us feedback, with a direct entry point like google.com/ai.

I recently had Brian Balfour on the podcast, and he shared this quote that’s really stuck with me. It’s something I think about as you talk about all this. It was by Alex Rampell: The idea that startups are a game of getting distribution before incumbents can innovate fast enough.

It feels like you guys are finally there. It’s like, “Oh, man, now here comes Google.” I don’t know if I have a question here, but it just feels like there’s been all this time for people to find distribution, and now it’s like, “Okay, now Google is coming.”

What we found is that people are asking these questions in Google. They’re trying to get this out of Google. If you can have an AI that’s powerful enough to answer a really hard calculation someone is trying to figure out, or take a picture of a multiple-choice homework question for a chemistry problem, people are doing this.

Now that you have this really sophisticated AI that’s based on our frontier models, we can handle increasingly more and more stuff for people. Hopefully, that’s the more natural on-ramp here.

We just need to make it easy enough for people to use, because these are new products and people are used to using Google in a specific way. They type in keywords, which we sometimes call keywordese.

But you can actually use natural language in Google. That’s the biggest shift we’re seeing: people asking really long, hard, complex questions. You just don’t think, “I can go to Google and type in, ‘What’s a great place for a date night? I already went to these 4 restaurants. I’m looking for outdoor dining, and my friend has this allergy.’”

You could put that into Google, and I think that’s the kind of thing we’re excited to continue making easy for people.

It’s interesting how we’ve come around. Back in the day, there was Ask Jeeves, which was this whole idea of just asking a question as if you’re asking a human.

Yeah. Ask Jeeves was surprisingly prescient on that. They had something way before its time that I think folks have rallied around now.

Too early.

Yeah.

What’s your take on this whole rise of AEO and GEO, this evolution of SEO? I’m guessing your answer is going to be, “Just create awesome stuff and don’t worry about it,” but there’s a whole skill involved in getting content to show up in these answers. What should people be thinking about here?

Sure. I can give you a little bit of an explanation of how this stuff works under the hood, because I do think that helps people understand what to do.

But when our AI constructs a response, it does something called query fanout, where the model uses Google Search as a tool to find and do other querying. Maybe you're asking about specific shoes. It'll add all of these other queries—maybe dozens of queries—and start searching in the background. It'll make requests to our data backend, too, so if it needs real-time information, it'll go do that.

At the end of the day, something's searching. It's not a person, but there are searches happening, and then each search is paired with content. For a given search, if your webpage is designed to be extremely helpful, you can look up Google's human rater guidelines and read this very long document that's been thoughtfully crafted for decades around what makes great information. This is something Google has studied more than anyone: Do you satisfy the user's intent? Do you have sources? Do you cite your information? Is it original, or is it repeating things that have been repeated 500 times?

There are best practices that I think still largely apply because it's ultimately going to come down to an AI doing research and finding information. A lot of the core signals—whether this is a good piece of information for the question—are still valid. They're still extremely valid and useful, and that will produce a response where you're more likely to show up in those experiences.

The only advice I would give is to think about what people are using AI for. I mentioned that this is an expansionary moment, right? It seems that people are asking a lot more questions now, particularly around things like advice, how-to questions, or more complex needs versus more simple things. If I were a creator, I would be thinking, “What kind of content is someone using AI for, and how can my content be the best for that given set of needs?” I think that's a really tangible way of thinking about it.

It's interesting, your point about how it goes into searches. When you use it, it's like searching a thousand pages or something like that. Is that just a different core mechanic from how other popular chatbots work? The others don't go search a bunch of websites as you're asking.

Yeah. This is something that we've done uniquely for our AI.

Mm.

It obviously has the ability to use parametric memory, thinking and reasoning, and all the things a model does. One of the things that makes it unique is that we designed it specifically for informational tasks. We want it to be the best at informational needs, right? That's what Google is all about.

How does it find information? How does it know if information is right? How does it check its work? These are all things that we built into the model, and so it has unique access to Google. It's obviously part of Google Search, so it has Google Search signals, everything from spam—what content could be spam that we probably don't want to use in a response—all the way to, “Wow, this is the most authoritative, helpful piece of information. We're going to link to it, and we're going to explain, ‘According to this website, check out that information,’” and you're probably going to go see that yourself. That's how we've thought about designing this.

4. AI Learns Human Conversation

You've worked on a lot of AI products at this point. It wasn't just Google or Artifact. At Instagram, you did a lot of AI stuff. What's something you've learned about building AI products that you find people maybe don't truly understand? Maybe something that's surprised you about building successful AI products?

I think the most recent lesson—and this is true even of something from the last week or 2—is how obvious it is that the interface is becoming more human-like, with how you can communicate with and steer AI. I think it used to be, even just months ago, that you had to do a lot of work to get the AI to do the thing you were trying to get it to do. You had to use these incantations. You had to prompt in a really specific way. People would have all these hacks like, “Hey, act like you're a coach and you do these things,” and you had to really push it.

To use a tool, more on the technical side, you had to do post-training. You had to take this foundational model, show it data, train it, and actually update its weights to do more sophisticated things. You'd tell it, “Here's documentation for an API if you ever have a problem. Ping this API. Here's the documentation,” as if it were an engineer you could talk to, and it would have no idea what to do with that, or it would have some idea but wouldn't really do it.

Increasingly, you can just use language. If you were to write up an order, you could say, “Here's my new startup. Here's my data internally. Here are the APIs to it. Here's the schema and the URL. Here's when to use it. By the way, make sure that if you get this kind of question, you really get it right.” That'll end up doing a lot in the model.

The model has now been encoded to be able to say, “Okay, I'm going to use more reasoning or thinking budget for that kind of question,” or, “I'm going to use tools or code execution in order to connect to this API I'm told about.” That's a relatively new thing, so I think it's going to open up a lot of this democratization of accessing these models and building incredible things. You don't even need to do a lot to get the most sophisticated outcomes. Increasingly, I don't think you need to do a lot of this heavy-duty fine-tuning.

It makes me think about a recent guest, Nasreen Shankel, on the podcast. She was a PM at Google, and she worked on Google Meet. She was a delight PM, working on making products more delightful. She talked about the reason Google Meet did so well and now feels like it's killing Zoom: They compared the experience of Google Meet to a human meeting rather than making it the best possible video conference. It was, “Let's make this as good as a human experience.”

That's interesting, what you're talking about—how that's almost the goal here with AI. Just make it feel like you're talking to a person.

Exactly.

It might be obvious, but think about that.

Yeah.

5. Relentless Improvement Drives Products

Okay. Let me zoom out, and let's talk about broader lessons you've learned over the course of your career.

Mm-hmm.

You've built a lot of very successful products, which I shared in the intro at this point. I—

Many, many of them are also on the other side of the spectrum. We've got the whole portfolio.

Okay, perfect. We'll talk about some of that. I asked you, as we were getting ready for this conversation, what's one thing you wanted to get across? What's something you think would be really helpful for product builders to hear to help them build more successful products? You used this phrase, “embodying relentless improvement.” Can you talk about that? What does that mean? Why is this so important?

Of course. I think that you need to be the physical manifestation of 2 things. One is relentlessness: complete effort that is always exerted in a direction of positive productivity. The second is to make things better. You have to always make things better. You're never content.

I think this actually came out of a funny story. I was at Instagram at the time, doing a big all-team meeting—one of my first—and they had this icebreaker. It was, “What's 1 word to describe yourself?” In the backstage area, I texted my wife really quickly: “Hey, 1 word to describe me. The first thing that comes to your mind.” She wrote back, “Dissatisfied.”

I was chuckling in the back room because I was, first of all, kind of offended. I thought, “It's not loving, caring, or something good.” Then I saw her little bubble, and she wrote, “Okay, there's more.” Then she wrote me this really thoughtful thing: “It's not that you're just unhappy. It's that you want the world to be better. You're driven by a deep desire. It's that you feel this sense of dissatisfaction with what the world gives you. You want to make it better, and you're pushed and motivated to do that.”

I thought about that afterward, and it wasn't until we built a bunch of products—some that didn't do well, some that have had a lot of really large success, and now billions of people use them—that it felt like one of the big differences. Obviously, a lot of it is just the conditions of the product and a little bit of luck here and there, too. But for the things that went well, there was always this spirit of, “We're going to get it eventually if we just make 2 more moves to make it better.”

Eventually, as I talked about earlier in our conversation, you get to this tipping point where it just kind of tips over into being net useful to people because of the amount of compounding effort that you put into something. You're always the harshest critic and the most dissatisfied person in the room about your own work, basically.

And I think that's really meaningful. There's this other incredible story that Tony Fadell told on a TED Talk about 10 years ago. You can look it up. I think it's something around “Think Younger,” as a title.

He talks about what it means that, as we grow up and age and become grown-ups—I have 2 little kids, so that's something I think about a lot—we habituate to everything. We accept and tolerate what the world gives us everywhere, and we just go, “Oh, that kind of sucks. Oh, well,” and shrug our shoulders and move on.

But if you don't do that and you ask why—“This sucks. Why am I tolerating this, and how do I make it better?”—he has this incredible story about going grocery shopping. He goes on for almost 10 minutes about this story, where he talks about getting a piece of fruit, like a plum or a peach, and how it has that sticker on it. It's got that sticker, and you think, “Who put that sticker there?”

Then, when you get home, you take your fruit out of your bag. You're ready to eat it, and you're all excited. You stick your thumb under the sticker, and it punctures the flesh. He goes into incredible detail about how it punctures the flesh of the fruit. The sticker comes off, and now the fruit's bleeding. Then you flick the sticker, and it misses the garbage. You bend over and pick it up, and put the sticker back in.

I was like, wow, that is embodying this mentality of, “Why is this here? How can this be better?” I think the best product people, the best thinkers in the space, that's how they think, in my opinion.

I imagine there are many examples of you doing this in the many products you've worked on. Is there one that comes to mind as a good example of this in action—of this actually working really well and delivering something really huge?

Honestly, a big thing is working on AI Mode. I think a lot of it was that we saw in AI Overviews that people were trying to ask harder questions, and we weren't able to answer a bunch of them.

Mm-hmm.

Or AI Overviews just didn't show up. A bunch of us sat around and said, “Why can't you just do this for everything?” Instead of saying, “We don't need to solve for that,” or, “That's not something that's in the most addressable next thing,” we actually saw people in the query stream putting the words “AI” at the end of their queries because they were trying to get the AI to do the thing.

We would look at that and just be like, “This is ridiculous. We need to build something here.” That was one of the big motivations: identifying that user problem and being very disgruntled on behalf of the user. We were failing the user every day. We were not helping them actually get their thing and get it better understood, and we were going to go build a whole thing because of it.

That's hard to do, by the way—to build all of that—but it was so obvious that that's what we needed to do.

There's kind of 2 buckets of people, let's say. Hypothetically, one bucket is, “Just make things better, make amazing experiences, and you're going to do great.” There's another bucket that's, “Drive metrics, drive goals, hit our KPIs.”

I know what you're not saying is just, “Work on things. Just make things better. Relentlessly make things better.” How do you think about that overlap of, “Okay, make things better, but here's what we really need. Here's the strategy. Here's the vision”? How do you think about this divide?

Yeah, I don't think of them as an either-or. I think they have to be intersected, because basically, the way to think of it—the way I think about it—is you actually start with a problem, or the inverse of that, which is a vision. But they're connected.

Most great companies and most great products come out of a problem. Out of the problem comes a better way: “What if, instead of this crappy thing, or way of living, or thing that we all tolerate and accept, some entrepreneur comes up and says, ‘What if we did this other thing?’”

It comes out of this dissatisfaction and this sense of better—that you need to make things better. But then you're going to build, and at the end of the day, you need your instrumentation to know if you're on the right track.

That's where you bring in tools. You build your first version of the product. Do people like it? Each product goes through its journey. The way you understand whether people like it is you scrutinize it. Typically, you talk to people, but you also add some analytical tools there, and you might look at something like a J-curve.

This is the retention—the percentage of people still using the product on day 7, day 30, and day 90. Does it flatten, or do people just drip out of there over time? It's just not exciting people, and that would go to zero. If, on a long enough timeline, no one's going to use it, you don't get past that, you're toast, right?

Then, okay, some people are doing it. Great. We need more people to do it, and it needs to be good enough that people talk about it, and then it grows. That's another gate. Then there's another one, which is, “How big can this get, actually? Is it a small thing? Is it a medium thing?”

I think most companies have an aspiration of being big, but you can't start big. Everyone has to go through that journey. No product has started big. Even ones that get big really quickly—even within a week—they had something, and even internally, they started small. They started small with 100 to 200 people.

You have to be metrics-focused, I think, in order to know if you're doing the right thing. On the other side of the spectrum, you're running a big thing. There, you need metrics to be your guide.

If your product is, let's say, “Our core metric's down 5% this week,” it's like, “What's going on?” You need to be really close to root-cause analysis and say, “It turns out that there's an issue. Is it in a region? Is it on a device? Is it in a demographic? Is it in a use case? Where is my problem? Why?”

When you get to it and understand the problem, this improvement thing comes back. It's like, “Okay, I'm going to fix that thing. What's the treatment for that disease?” Then you're back to growth again.

You need this, and you're always looking at, “What's the system that I'm working on, and what are my instruments, as a pilot, to know if this thing is going and flying correctly?” But it doesn't tell you exactly what to do. You have to think for yourself about how to make it better. It can just show you a little bit of the way.

I love that you just gave a master class on how to prioritize and pick what to work on. I want to go on a quick tangent. Speaking of products that have done really well and become really big, Stories—you built and launched Stories at Instagram. It's quite an infamous product launch back in the day. It was quite controversial because it basically took what Snapchat was doing really well and brought it to Instagram. It was not great for Snapchat.

Now that it was so long ago and it's so far in the past, I'm so curious to hear about that time, reflecting on that decision—what you guys talked about, how you decided to go ahead with that, and anything you think about looking back at that.

6. Instagram Makes Stories Its Own

I think there are a couple of really important lessons from that launch. We went on afterward to launch Reels, a bunch of updates to direct messaging, and feed ranking. There was just a huge era between 2016 and 2021 or so when I was there, where so many products got built.

An interesting lesson in all of those, and particularly in Stories, was that you have to really understand why someone uses your product and know when something is actually an existential question because there's just a better format or a different way of doing something that has worked and works. You need to figure out what that might mean for you.

Not every great thing is going to be invented by you. But a lot of these things can become formats that you can make your own, and you need to learn from the world and what's happening out there in order for your product to always give the best thing to its users.

For Stories, we looked at Instagram and asked, “What's the point of Instagram?” It is sharing your life and connecting with people, ultimately. If there's a way to do that that lowers the pressure because it doesn't have likes, or it's an ephemeral format, and it's optimized well for mobile because it's this full-screen experience, it's a really great format. Kudos to Snapchat for inventing it.

We didn't think of that as a deterrent—that we had to go make “Instagram photo clock.” Actually, there were early versions of this idea where you try to take the core Instagram feed and make it ephemeral.

Whenever you try to mix a core product that's very cemented in someone's mind and physically looks a specific way, and you're trying to contort it to do something new, it's usually a bad recipe.

And so we knew we needed to do something new, and it was so clearly critical to the core essence of what the product could do. It could fit in naturally. But the question was: How do we make it our own, and how do we build on this? If you think about it, there were a bunch of things we did that made it Instagram. For example, it had different creative tools, including neon drawing and really sophisticated filters that people loved.

We also looked at this through the lens of being dissatisfied. A lot of times, people want their main camera to take a picture of something, and then they want to upload it to Instagram because they want to save it and have it be a very high-quality, high-resolution photo, because it’s a memory. Snapchat at the time didn’t allow you to upload photos; you had to use a Snap camera. So we made a bunch of decisions like that: Why don’t you just let people upload their photo? This is back to the dissatisfied point. That’s frustrating.

There was another example where you couldn’t pause if you were consuming a story. You couldn’t pause it. It would just go through and be done.

Mm-hmm.

That was because it was this ephemeral thing, and you wanted to create safety. Why can’t you just pause it? It goes by too fast. So we added this pause. It’s such a small thing, but now you put your finger down to pause the story. There was a whole set of those things that made Stories feel like Instagram; it wasn’t like you just had some other thing.

And it turned out that worked incredibly well. So much so that someone on the team mentioned that, at the time, they didn’t realize it, but it was almost like Instagram was missing the story-sized holes at the top of the page, and that completed the product in some weird way for them. I think that was an important lesson.

Instagram definitely got a lot of hate for that moment, especially from a lot of founders who were just like, “Hey, you guys just stole this idea, and that sucks.” How did you guys deal with that internally? Was it just, “This is something we have to do. We have to focus on our shareholders and grow this thing, and that’s how it goes sometimes”?

I think it’s more that we were focused on our users and the people who loved Instagram. We were denying them the opportunity to have an easy way to just share a photo and have the thing go away. Ultimately, that’s what we were trying to add. At the end of the day, that is a format that people adopt.

In the same way, think about feeds. Facebook probably created the modern feed, but there’s a feed for every single product, right? There’s a LinkedIn feed, and there’s a feed for DoorDash. These things become core primitives and formats quickly, and at the end of the day, you’re just robbing your user base of the opportunity to have a better product if you’re not making the best possible product for your use cases.

For Instagram, it was used differently. People used Instagram differently than they used other products. It turned out there were these experiences in WhatsApp, Messenger, and many other social products over time, and they all were used differently, actually, which is fascinating.

7. Mature Products Find New Growth

Something else I want to talk about is that you came into 2 products that were already doing really well: Instagram and Google. On the Instagram side, there was transformative growth and improvement. With Google, it’s happening—we’re in the middle of the improvement and growth you’re driving.

Not a lot of people get to do this, where they go into an existing product and make it grow significantly. A lot of people want to do this. They have a product that’s been around for a long time, and they’re asking, “How do we make this grow and be more successful?” Is there anything specifically that you’ve learned about coming into an existing product, figuring out where the big opportunities are, and then hockey-sticking growth? This is what everyone wants to do.

There are a couple of lessons here. I think the first lesson, by the way, is to always be humble, because it’s incredible to be able to work on products that have such an impact on people. I view product like golf: You’re always 1 stroke away from shanking, and as soon as you think you’re good, you’re not. You don’t know anything; the world changes quickly. You always have to be a servant to your user base and the people who are out there, and learn from them.

The first thing I always do and think about is getting in touch with why people are using this product and where the areas of growth are. Usually, even in a big, mature, complex system, there’s a part of it that’s growing, a part that’s mature, and there could be a part that’s declining or isn’t growing as much. Certainly in Instagram, there has been a big shift over the years from sharing into public, very large broadcast posts and Feed to these more lightweight formats like Stories and DMs—actually, private sharing as well. You have to observe that, because every month and every year, the world changes and people’s needs change.

So the first thing you do is get a sense of what people want out of this product and what its true essence is. I think a lot about the jobs-to-be-done framework, which is one of the things that I’m a big fan of. Clayton Christensen’s book Competing Against Luck is one of my favorite books on this topic. You have to really be a student of causation: Why is someone using this product? What are they doing with it, and what are they trying to get done with it? That usually leads you to bigger next-stage ideas, and it removes this belief that you need to solve the problem with the current tools.

In the Instagram example, it was like: You have to make a square photo do more for people, right? That would be how you increment the product. In Google’s example, there’s something very specific with the core search experience that needs to change, like a subtle tweak. You have to think, “What’s the big thing someone is trying to do?” Someone’s trying to ask Google a really hard question, so what’s the best way to do that for them? It makes you think more from first principles, and that’s the first basis of this.

And then, from first principles, you’re like, “Oh, this newer thing—it could be a shift or a new format.” In many ways, the AI version of Google, Stories, and Reels are all kind of similar in that they’re new formats in the world that people are expecting and wanting more of. By adding them, it becomes complementary, not a replacement. In both cases, Stories didn’t replace Instagram; it expanded it, in the same way we’re seeing with AI.

What’s interesting is that then you think, “How do I bring that into my world?” You have this big, mature product, and the best way I’ve seen is by making it complementary, having it be a core part of the experience but clearly defined as a distinctive thing that has its own attributes associated with it, because people think spatially. If you have a Feed and you have holes with pictures, people expect those holes to do things. If you make one of those holes with a little clock, and that one goes away the next day, or you can’t like it, or it operates differently than the other parts of your Feed, it’s going to be super confusing for people. It sucks.

So you have to add product carefully, but it needs to feel coherent and different. Stories have a similar aesthetic. They obviously use your camera roll in the same way; they work, you can share them in DMs, and they work in the ecosystem, but they have a different primitive. In the same way, AI Mode is a full-page experience that you can pop out now. You can have a follow-up conversation with it, right? People have a set of expectations you need to snap to for those use cases.

And then you’re constantly learning how to best make these new products work within your world. You never just want to snap in something that’s working; you have to make it work for your users, your expectations, and what people are trying to do with your product. It’s actually one of the things I see people fail on the most: They assume something working for one system will work in your world, but someone else’s system is totally different—the types of users they have and the consumer expectations of that product are a totally different set of expectations.

So you have to respect that and say, “What can we learn from that and bring here?” That was, I guess, if you were to talk about the method that I’ve seen now twice, how these products have developed.

I love this topic. It makes me think about this balance people always try to find between optimizing something they already have versus trying to take a big bet on something. You’ve had so many examples where you’ve taken a big bet on something totally new, and it’s worked out incredibly well. Do you have a heuristic for how you structure teams and prioritize across, “Okay, we have an amazing Google Search experience today; what percentage of resources go into improving that versus trying something totally new?”

That’s one where I actually do feel like the more analytical, systematic thinking helps a lot. Because if you’re trying to produce value in the world, you want to quantify it in some way.

And so, if you're seeing this growth curve and you're trying to understand, “Wow, people are using it more and more. They're liking this product,” products are young, they grow, and then eventually things mature. You can break out product suites and different features of products in the same way—certain features that are growing fast and other features that are not.

And you get to these points of diminishing marginal return in every system, where it feels like you could put 50 people on this project and it's just not going to dramatically move the needle. Part of it is this bottoms-up thing, with your own team being really thoughtful about what the expected value of that investment is and knowing when it's starting to approach zero—diminishing marginal return.

And then, when that happens, these are moments that usually coincide with something fundamental changing: either people's expectations externally or market saturation. There's something happening where you need to adjust, and you then find your next growth driver or set of drivers. That's where you need to go more first-principles and try these new things more.

Then, when you land a new thing, that creates this new little growth engine, and then you put people on it and optimize it. Each change is a 10% win, a 20% win, a 4% win, and it clearly still has so much value and headroom to make it better for people. You can see that in the data.

And so, as I talk about this instrumentation, it becomes your guide for knowing if you're making good calls. Otherwise, if you don't know where you're headed and you don't have a quantitative goal for what you're trying to do, it's really hard to know if the thing you're doing is mattering to anyone, because you'll just think, “I made the product better, but is anyone using it? Does anyone care? Or are we just congratulating ourselves?” Ultimately, you want to have an impact on people, and that's what matters.

So it's essentially tracking S-curves on every product and understanding—

Yeah.

—if you're in the plateau and if it's time to invest heavily somewhere else.

Yes.

This episode is brought to you by Orkes, the company behind Open Source Conductor, the orchestration platform powering modern enterprise apps and agentic workflows. Legacy automation tools can't keep pace. Siloed low-code platforms, outdated process management, and disconnected API tooling fall short in today's event-driven, AI-powered agentic landscape. Orkes changes this. With Orkes Conductor, you gain an agentic orchestration layer that seamlessly connects humans, AI agents, APIs, microservices, and data pipelines in real time at enterprise scale. Visual and code-first development, built-in compliance, observability, and rock-solid reliability ensure workflows evolve dynamically with your needs. It's not just about automating tasks, it's orchestrating autonomous agents and complex workflows to deliver smarter outcomes faster. Whether modernizing legacy systems or scaling next-gen AI-driven apps, Orkes accelerates your journey from idea to production. Learn more and start building at orkes.io/lenny. That's O-R-K-E-S.io/lenny.

Maybe it would be helpful to talk about the journey of AI Mode—how it emerged and the steps that you took to make it such a big part of the Google Search experience. When did this start? How did you decide this was worth betting on? And what were the steps to get it further and further rolled out?

8. AI Mode Ships In A Year

I think it probably started earlier on with AI Overviews, actually, which was the first way we brought generative AI to Search. In that world, we noticed that people were asking these questions, and many people were trying to put natural-language questions into Search. How can you provide helpful context, links to go deeper, and an AI that made sense for Google? That was our first version of these models that could do this for people.

Then, by building into that and seeing this observation around people wanting more of it and direct access to it, and then being able to ask follow-up questions, you need a new modality. It's going to be really hard to build all of that within the construct of the core Search experience.

That led us to form a small team of folks: a few technical leaders and a couple of designers. It was very small, just to prove out, “What if there was an almost blank screen?” Delete everything—make a fresh document with a blinking cursor. What if there was a new page where you could ask the question, ask whatever you wanted of it, and tap right into the AI that was originally powering the top of the experience in Search?

But we invested in making it much more powerful in the ways I described before. It could search for you, it had reasoning as part of its model capability, and it had multi-turn context. So if you had a conversation with it, it could keep track of that context. It had some unique pieces to it. What would happen if we tried that quickly?

We got, I mean, probably five to 10 people originally.

And how long ago was this team formed?

This was probably over the last year, last summer—

Oh, wow.

—basically into the fall.

Oh, wow. So about a year ago.

Into the fall. Yeah, maybe it started about a year ago. We were really plugging away on it, and then we saw this little version of it emerge that wasn't very good, but it had this moment of brilliance.

It's actually, again, kind of like golf, where you hit the perfect shot and you're like, “Oh my God.” You get that feeling where everything worked. I asked it a question about—I forget. I was doing something with my daughter, and I was planning an experience, and it found all this incredibly useful information about parks. It had links to go to the site and confirm a bunch of things. It had Google Maps information showing that, for my daughter, we could walk there. It was walkable.

There were early examples like this where it just blew me away—what it could do, what it could find, and how helpful it was. It gave us the conviction that we should go further.

Obviously, there are lots of people involved in this type of decision, with tons of support from leaders across the organization. But it was just a little working team. You have to build something, and then you have to feel it yourself. It's very entrepreneurial in that way.

Then, when you see it tangibly, you're like, “What's a version of that that's good and could work?” That gave you hope. So we basically built it out and built the first version that launched in Labs.

So the first big milestone was, “This is working.” It was just a qualitative experience of, “Oh, wow, there's magic here.”

Yes, it's working. And then we brought it, before Labs, to a trusted tester group. There were maybe 500 people externally that we added to it, and we had conversations with them. Some of them were friends and family.

We tried to treat it a little more like a startup because we feel like you have to have people test it who will tell you the truth and tell you when it sucks, because it probably does. Then they message you. I had a friend who was loving it but also hating it for lots of good reasons, and would just message me all the time with screenshots: “This broke. This broke. This makes no sense.”

We had that for a while, and then we got to a point where it was feeling good. The trusted testers were liking it and reporting good stuff. Then we brought it to this Labs moment where anyone could turn it on.

We used that to make it better with real query data. We could actually see what people were using it for at greater scale, and that allowed us to tune it and make it better. Then we launched it out to everyone, or at least in the US.

Now we've been on this journey to expand it to all countries and languages and have more people be able to access it.

It's incredible that Google went, in roughly a year, from an idea to a significant change to the Search experience that's AI-powered. I think this is not what people imagine Google is like, and it feels like things are different and things have changed in how you operate. What has allowed this to happen so quickly? What's changed? Is it just top-down leadership—“We need to get shit done”—or is there something more?

No, I think it's interesting how organizations change. When you feel like there is a moment in time that's clearly critical to deliver for people—people are trying to get information from Google, and we're not able to answer certain things or help people in certain ways—and there's this technology that can do it, that creates urgency.

Obviously, there are lots of people building lots of things, the market's crazy, and lots of things are shipping all the time. So there's a really exciting and healthy moment for us to build and build quickly.

And I think it’s just exciting to be able to capture that opportunity, because I think people believe—and I certainly believe—that the next year or so of product is going to establish how people use the next wave of products for many years. At least, I can only speak for myself: I feel this obligation to our users to give them the best version of Google powered by AI, one that gives them access to the full knowledge of everything Google knows about the world and makes information accessible to people with AI. That’s driving a lot of the excitement.

Yeah. It’s such a good point that people are building their new habits. It’s wild how many people just now rely on ChatGPT and how quickly that happened. I could see Google being worried that everyone’s changing their habit from searching Google to searching ChatGPT. And now Gemini is number 1.

I was actually looking at the list of top apps. In the top 15 apps, Google has, I think, 5 of them—a third. It’s out of control. Killing it. When people look at AI Mode versus ChatGPT, Claude, or Perplexity, what’s the way you think about the positioning of AI Mode versus these other tools? Is it trying to be a direct competitor, or is it actually pretty different, and here’s what it’s for?

Yeah, AI Mode is a way to ask Search anything you want. It’s designed specifically for information, and it should give incredible, helpful responses for the things that people come to Google for. Think about planning a trip, trying to buy something, or working through a question for a research project. It needs information, and that’s really what it’s focused on.

It’s less focused on things like creativity, although there are things it can do that are nice there. It can help you rewrite something; it’ll do that, like any core AI product. But we’re less focused on creativity and productivity—uploading a spreadsheet and outputting graphs for you, for example. We’re really focused on what people use Google for and making an AI for that, so you can come to Google, ask whatever you want, and get effortless information about it, in context, with links to verify, dig in, and ultimately go to the authoritative sources that people want and that we hear from people.

Those end up becoming the distinct qualities of this product versus something more like a chatbot. Maybe you would talk to it and even have a bit of a “Hey, how are you doing today?” conversation with that chatbot. We have some of that, but people are usually coming for information. They’re trying to learn something, and we’ve focused our product on that.

Got it. Okay, AI Mode is not your therapist. Maybe zooming out again a little bit and reflecting on all the amazing products you’ve worked on and all the places you’ve worked, if you had to pick 2 or 3 core product principles or philosophies that have helped you build such amazing and successful products, what would those be? What comes to mind?

9. Three Principles Build Better Products

I mean, there are typically 3 things I think about. If I were to write a book about how to build great products, there’d be 3 chapters. I mean, there’d probably be more than that, but 3 chapters.

I love that. I love how short that would be. I think that’s the ideal book.

The first—I’ve thought about these 3 areas for a while, and they’re always consistently the 3 things.

Mm-hmm.

The first is: deeply understand people. I think we talked about this a little bit with the jobs-to-be-done point and Clayton Christensen’s book, Competing Against Luck, which I loved. It really helps you be a student of why someone ends up, in his words, hiring a product. Don’t think of users as using your product; think of users as hiring you to do something for them.

There’s this famous quote—I think it was Theodore Levitt—“People don’t want a quarter-inch drill; they want a quarter-inch hole.” So what is someone trying to do? You have to understand that deeply, and then you can build an amazing product. Also, why is someone not using your product? It focuses on these techniques to extract causation.

He talks a lot about the interview. He calls it an interrogation, where you talk to a user: “Hey, why’d you use my product? Where were you? Were you in bed? Were you at work? What were you doing?” “Oh, I was talking to my wife in the morning.” “Okay, well, what brought it up?” “Well, I guess I was reading the newspaper.” “Okay, well, why?” And then you have this aha moment. When they first decide to use your product, he calls it the big hire. The information that you obtain ends up becoming the most critical because that is what caused someone to use your product. If you can study that and understand it, you’ll be much more on your way than just building things that sound cool.

So that’s the first chapter: deeply understand people. The second is really around analytical rigor and understanding your problems. You have to understand your problems, and this is a little bit of what we were talking about with root-cause analysis and understanding: The metrics are dropping—why? If someone’s not using your product, why? You have to be able to dissect that to get to the true root causes.

It’s like, “Well, they went all the way to the end and then bailed.” You talk to them, and then you understand, “Oh, it turns out that…” We actually learned about this with Close Friends at Instagram, where it totally failed at first when we shipped it in a bunch of markets. It turned out that people were only adding 1 close friend to their list because it was mistranslated as “best friend” in many markets. So people just put 1 person on the list, and then the probability that person saw it and wrote back to you was like 0. It’s a product that’s broken. You’ve got to understand your problems.

And then the third one is around really designing for clarity instead of cleverness. A lot of people are like, “Oh, we’re going to differentiate the design.” We talked about this a little bit with Stories: “We’re going to make a new version of something.” But if something’s a standard and people understand it, if you lean into it, you’re going to get so much more leverage than if you reinvent it. You have to be really thoughtful about when you reinvent and when you don’t.

I think on this one, there’s this great book—Don Norman’s The Design of Everyday Things is a big one—but he has this incredible chapter in there about doors. Why is it that, after all these years, you can walk up to a door and, based on how it’s designed, still not know whether you should pull or push it? If you try to build the most beautiful, symmetrical glass door with 2 handles, 1 on each side, it doesn’t communicate any information to you.

I’ve seen this all the time: We’ve designed new icons when we could have used global icons. “Wouldn’t it be so cool if we used a camera that’s kind of a camera, but is mostly an AI-looking thing, and then has the dots in it that connect it to this other product?” People just—it’s a camera. Just put the camera in. Maybe you could add a little thing to it, and that’s how you get people to use your products.

If you do those 3 things, I think you typically can do well. And then, sorry, the fourth one, which may be more of a coda, is: be humble. Constantly and always question yourself, listen to others, listen to users, and be open to being wrong.

I love these. On that third point, I feel like AI Mode as the name is such a good example of clarity. What is this? This is AI Mode. So we talked about it.

We talked about it internally. If you look at it in the tab, everyone knows. You see it, and you’ll know what it is.

Right.

Or we could call it something random, but then what is that? Now you’re working against yourself.

So if I were to reflect back these 3 pieces—basically, this is the book you would write to help people build more successful products—it’s: understand the problem you’re solving for people deeply. What’s the job they are hiring you to do? I love the lowercase “jobs to be done.” It’s not like the—

Mm-hmm.

—the rigorous whole thing that—

Exactly. No.

—you know, everyone. Yeah.

Lowercase for sure.

Okay. This is just: Why are people hiring your product to solve a problem for them? What problem are they solving? So it’s basically: figure out what problem they’re having. Then, through data, understand the problem and whether you are solving it. And then just keep it really simple: clarity over cleverness, essentially.

Exactly, yes. And be humble.

Is there…

An example that we haven't talked about that shows this in action is: Here's the problem we found, here's how we figured out this is the solution and whether we're succeeding, and then here's a very simple way of solving it?

Honestly, this Close Friends example from my Instagram days was really wild. It took 2 or 3 years to get Close Friends to work, and I think it totally failed originally. This is the product that lets you add a private list of people, and then you can post to your Story and only those people see it. It's this very exclusive, private space, so you can feel really comfortable sharing maybe more—

Green circles.

Green circles, yes. It was—at least when I was there—one of the most popular features of Stories and did really well, but it totally failed.

I think what we found was that we actually used a bunch of these techniques. One was that we first thought about it as an overall system problem, and you could add a Close Friends post for anything. You could do a Feed post or a Stories post, and you also had a Close Friends profile. So if Lenny went to Robby's page and we were Close Friends, you would see extra stuff from me on my profile too.

We shipped it and thought it would be great. This is the “be humble” part: it wasn't great. It was super confusing. You would see this really beautiful photo, and then in the Feed right after it, this blurry, very vulnerable moment someone was trying to share with their friends. It felt so out of place and weird for the reason people use Feed.

It was also confusing because it had an extra little green thing on it, but the Stories one didn't. If you opened the Story, it had a green thing inside the Story, and people were just so confused. There was this other issue with the list, where the list didn't work because it was mistranslated and people didn't get it. I think it was actually called Favorites originally, I want to say, and that encouraged people to put only 2 people on it.

This gets to the framework, I guess: deeply understand people. What are people trying to do with this? What they're trying to do is share something vulnerable and say, “Hey, I'm lonely. Hey, what's going on? Are people up?” It feels very much like a friend-group thing. If you only have 2 people on it, the job we're doing is actually connecting you to your friends. If you don't get a DM back, it's broken.

What we're really doing is getting you a DM and getting you connection—a sense of being connected to your close friends. That is the job. It's actually the other thing Clay Christensen talked about in the book: There are utility jobs and emotional jobs. People usually discount the emotional ones a lot. This was really an emotional thing as much as it was a utility one.

The product was broken, right? People didn't even know that it was a Close Friends Story. They just saw the little head, because you had to click on it to see the thing. People stopped using it.

We went through and did these revisions where we simplified it, updated it, and went through this change list: Take this out, take this out, change the name here. What we saw was that it was working really well for people who added 20 to 30 people to their list. You would put 30 people on your list, and then 2 of them would write back to you in DMs. Now you've closed the loop and you feel connected to those people. It's a winning thing.

We designed the whole system around that, and it also only worked in Stories. We were looking at the data and trying to understand where it was working and where it was failing. We updated the name to Close Friends, so it didn't feel like Favorites. It wasn't 3 people; it was more like 20.

In the list, we built this list builder where we recommended a set of people based on some data and a cool algorithm created by an engineer. Then we updated the design to put the green ring on the outside of the Story. This was the design for clarity.

We were being cute. I think at the time we thought it was like a secret Story or something, and if you opened it, you would see it. It just wasn't clear to people. We put the green ring on the outside so users would see it in the tray and think, “What's that little green guy?” Then they'd click on it and say, “Oh, this is a private Story for me.”

That system worked and did incredibly well. That was the process we followed, from a total flop to something that was very successful.

That is an awesome example. This process took 2 or 3 years, you said?

Yeah, it took a while. That was actually one of the longest projects we worked on. The reason we did it was that we asked people, to deeply understand them, “Why aren't you posting to your Stories? What's preventing you from doing it?”

Everyone had some version of, “My ex is on it. I have a teacher on it. A friend who's judgy is on it.” There was this commonality around audience problems. Someone had an issue with people watching them. That gave us conviction to go this hard at it for so long, because we knew that was a core problem with the product.

Was this also connected to the Finsta-Rinsta trend?

It was. Actually, I think that informed us. Everyone had a Finsta, and there was a Binsta, right? It was like the best—

What is a Binsta?

Best friend.

Best.

Binsta. It's like this layering of people having 20 Finstas down to a partner Binsta. I made that up. I don't know if it's true, but I'm sure there are Binstas out there somewhere.

Okay.

We were like, “Wow, people clearly are trying to hack Instagram to create these private, smaller-group settings, so we should just make a product.” Yeah.

How did you actually do this testing? Was it rolled out to some percentage? Was it rolled out in New Zealand or something as sort of a—

Yeah, we rolled it out in a few other countries. Exactly.

Okay, got it.

We had a basket of countries that we tried it in, and then we would do research. I think Australia was one of the first ones for that one.

I was going to ask if you can share the countries. So, Australia.

I think that was one of the earlier ones, yeah.

Okay.

But it's always—every time you ship something, there's a slightly different reason why.

Oh, interesting. So it's not always Australia that gets all the new stuff.

No.

Okay.

Although sometimes it is. Australia and Canada get a lot of stuff, just because it's easier for the teams to see feedback from them.

Yeah, they speak English.

Yeah, exactly.

Awesome. Let me go in a different direction and talk about something that you have a hot take on. There's a lot of talk these days about lean teams, small teams, creating with limited resources, and not hiring at all. You have an opposite perspective: You actually need a lot of resources to build really big breakthroughs. Talk about your experience there.

10. Big Breakthroughs Need Big Teams

Yeah, I think it obviously depends on what you're trying to build, and there have famously been small teams building high-impact products. But I think there's this cult of lean, scrappy, fast, throw away your product quickly, keep moving. At some level, it's true for internal conviction.

But to build a product that works for a lot of people and is based on a technological breakthrough, a lot of times I see teams just give up too early or underinvest in the product. Obviously, the space matters. If you're building a single product that's a way to do something with a digital app that's fairly straightforward, that's going to be different from building a robotics company. What you're building does change.

But even for software, for really hard technical problems, think about the amount of time and effort it took for teams to build a foundational model, and how many years and hundreds and hundreds of people were needed for that to happen. You think about these large companies that have had huge impacts on people, and I think particularly for bigger companies, something I've seen internally is that it's almost too scrappy because it never gets enough momentum.

The product never gets good enough internally, and then it just dies on the vine. Whereas if you put more people on it, you have to be careful not to put too many on too soon. But I see the opposite as more true: People hold on to small teams too long, and then you...

Either it takes forever to get to the thing you’re looking for, like this close-friends example I mentioned. This was actually a small team. One of the reasons it took us forever was we kept the team so small and scrappy; the loop cycle was so short, and by startup age, you’d be dead probably. So you can maybe do that in a bigger company, but as a startup, I don’t know if you have that luxury.

And so I think you need to actually think: What is the group I need to build a version that’s great? From first principles, really think about it instead of just embracing blindly, “Okay, we’re gonna be the 2 of us until this thing has escape velocity and market fit,” which isn’t always true.

This is definitely counter to the narrative we see on Twitter. Anything you can share about the heuristic you use to decide, here’s when and how long to keep it small? I know there’s not going to be a do-step-one, two, three. But what I’m hearing is: start small to prove out the concept—designer, PM, engineer, maybe. When do you find that it makes sense to go big?

Yeah, I think it’s mostly when you’ve hit the conviction moment. I think there are 2 big milestones. There’s internal conviction: for yourself, do you believe in it? And you believe in it because there’s some external validation. You put 20 friends on it.

By the way, I found out very quickly, building startups, that if you put 20 friends on something, they’re not going to do you that many favors. They’re not going to use a product every single day just because they’re your friends. Thirty days in, 60 days in, 90 days in, they’re not using your product unless you’re doing something that’s useful to them.

And so you get all this feedback, and you’re seeing people really enjoy it. You get to that moment, and then I think that’s not a product that would win externally, because if you were to ship it, it’s broken and doesn’t work great. Then you need to invest enough to make the best version of it, or as good a version as you can, to get it out the door and ship it.

I think that you want to build the right product eventually—that’s the mentality—and you can only really do that with the right group.

I’m going to take us to a recurring segment on the podcast that I call AI Corner.

Okay.

What’s some way that you’ve found a use for AI in your work or your life that is really interesting, really helpful, and maybe other people can be inspired by?

11. AI Expands Visual Inspiration

I think one of the coolest trends ever is how AI is affecting multimodal visual and inspirational needs for people. We’re early in this, and I think this is something that I’m actually working on as a project as well.

But right now, if you think about what AI has done, in large part, it was born and grew up in this text modality. It was chat. For a long time, if you were to ask it to help you redecorate your bookshelf behind you, it would describe that to you in text, because that’s what it knows.

But increasingly, AI is going to be liberated to help in every possible modality. We’ve seen a lot of this with the explosive use of Google Lens, our image search, and image features, and with this deep understanding.

What I’m actually starting to use internally—and some things that we’re excited about, with more coming up that we actually announced at I/O that we’re going to be building more of—is how AI can help with inspiration, how AI can help with shopping, and helping you really get things done that are more in the inspiring bucket of needs versus these core utilities, like code, math, and homework, kind of side of things.

I’m really excited for things that are coming where you can ask it for inspirational tasks, and it’s starting to do really fascinating things in terms of what I’m seeing. Hopefully, we’ll share more on that soon. But I think the one thing I can share is there’s a visual version of AI Mode that we talked about at I/O. You can reference some of those keynotes, but that’s in the process of being rolled out.

Hmm. Mysterious.

You’re going to be able to ask, “What’s a beautiful, mid-century-modern office design with dark themes?” It’ll be able to produce this image board that’s inspirational, and you can do multiple turns with it.

You’ll be able to say, “Actually, I want more of a light theme, more creamy, more California, more coastal vibe.” It’ll do that, and it’ll understand that. It’ll actually see the images and be able to talk with you in the way that text works, which is going to be really cool.

I think that’s going to be one of the more exciting things that will be new to AI soon.

What I’m hearing is Nano Banana integrated into AI Mode. Recipe for success.

Well, it’s a little different than Nano Banana, because Nano Banana—

Oh, okay.

— is an image editor.

Mm-hmm.

This is more like helping you find images on the web. So it’s a little bit more like AI inspiration, AI image search, and allowing you to then talk with it to effectively get visual responses with natural language.

That’s going to be a little bit different than, “Edit this photo so that it changes it.” Although, potentially, an interesting idea too would be to have the ability to take a picture of your living room. I think AI will help with that too, ultimately.

Pinterest is in trouble. It feels like this is what people use Pinterest for: here’s all the inspiration. Now it’s just AI doing it all.

By the way, Nano Banana—where did this name come from?

I don’t actually—I forget the story. There’s a story somewhere. I forget it now, honestly. But the team—

Good.

—I think it came from a scrappy, fun group of people building this, and they wanted to go for something fun for folks to—

It feels like that’s part of the reason things have started to work. There’s just more fun and delight and random, crazy stuff coming out.

It feels a little more like when I was at Google the first time through, where you kind of just have so much stuff and this fun curiosity happening, where people want to try things and ship things. Hopefully that continues.

Yeah. It feels like Veo 3 would be even more successful if it had a wacky name. And I like that this is the opposite of your advice on clarity. I don’t know what Nano Banana is, but it works.

Yeah, that’s the other thing. No advice is right universally, right? But yeah, Nano Banana.

Robby, is there anything else that you wanted to share, anything else you want to leave listeners with as a final nugget of wisdom before we get to our very exciting lightning round?

This concept: be curious. I think of embodying everything as really being about curiosity. It’s about wanting to know why everything is the way it is. Why is someone doing something? Why does someone have a different opinion than I do? Why might this not be working?

The people who really have that level of intense curiosity, and who chase things down until they know, I think, are well-served by that. So that would be my only parting thought.

Let me follow that thread, actually, because it’s maybe the most trending term on the podcast over the past few months: curiosity. It comes up a lot when I ask people, “What are you teaching your kids and then embracing with the rise of AI?” Curiosity comes up all the time.

Is there anything that helps you? Is it just, like, “I’m good at this, and I’m innately curious, and this is valuable”? Is there anything you can share that helps you or others around you embody that and actually be curious?

AI is obviously the ultimate curiosity engine, and that’s what’s so cool: you can now ask anything and just get information. I find that people underappreciate just how much they can learn about whatever they want.

But also, I think that a lot of this comes down to studying what you want to know about and knowing where the branches of knowledge live. A lot of times, I’ll read old papers and PDFs that are free online on a statistics topic if I want to learn about that. I think people underappreciate those as analog, old-school, great learning.

AI can help you discover them. I’m using AI, particularly at Google, to help discover all these cool links and things to read. But I find that it’s an interesting hybrid where it’s not just AI, but really going to original sources more.

I find that these books I mentioned in the chat here—I find that you need a blend of all of those things to ultimately really get to the bottom of things.

Like actually reading the thing, not just reading the summary of the thing.

Yes.

Let me actually ask you this question I’ve been asking all these people who are at the cutting edge of AI. You have kids. Is there anything you’re thinking about and leaning into to help them learn and develop as AI emerges and becomes a big part of the world?

I have younger kids, so the biggest thing I’m doing is they’re using live versions of AI that they just talk to now much more.

Funny enough, we actually just launched Search Live out of Labs this week. You can talk to Search in a live AI setting, which is conversational voice mode. When you're driving, you can just talk about all the knowledge I talked about—what you can do with Google—in a normal conversation with your voice, and I found that to be incredibly accessible for kids.

My kids come home and ask, “Can I talk to Google about something?” I'm like, “What do you need? What do you need to say?” Then they go to my app, hit the Live button, and start talking to it. They want to know about animals, certain history things, or something they learned about in school. It's so natural to learn in that way that I think it's helping them become much more AI-native than any other thing I'm doing.

Life as a parent is gonna be way too easy now. Whenever your kids have questions, just go talk to the AI.

I know.

But I don't think that's bad.

Yeah.

So this is within the Google Search app. There's a Live button. How do you access this?

Yeah, exactly. You go to the Google app.

Okay.

It's one of the apps in the App Store you mentioned. You open Google, and there's a button now that's labeled Live, right on the home screen.

Mm-hmm.

If you tap on it, it's a live version of AI Mode that you can just talk to. It's a full-screen experience, and it will say, “Start talking.”

In the show notes, I'm gonna link to this project that Eric Antonow built, which I love. It basically shows you how to put a little speaker into a little stuffed animal, and you connect the speaker to— it could be Google Live or it could be ChatGPT, whatever you like—in voice mode.

You put it on your shoulder, get a little magnet that attaches, and your kids could talk to this parrot, for example. You could tell it, “Talk in a pirate voice,” so they're talking as a parrot.

Yeah.

Oh, that's really funny.

Okay. That's really cute.

It takes about 15 minutes. You just get an X-Acto knife and sew it and stuff, and it's kind of fun. I made one for my nephew, and he was looking for treasure with this parrot.

That's really adorable. I'm definitely gonna look into that.

Robby, with that we've reached our very exciting lightning round. I've got 5 questions for you. Are you ready?

All right, I'm ready.

What are 2 or 3 books that you find yourself recommending most to other people?

Definitely the 2 I mentioned: Clayton Christensen's Competing Against Luck and Don Norman's The Design of Everyday Things. But I also really love Aurora, a fiction book David Koepp wrote. It's about an electromagnetic pulse from the sun that knocks out—it's fiction just for fun, and it was a really fun beach read. Apparently, it's gonna be made into a Netflix show. It didn't work out. I don't know. I was sad to see that fall apart, but it's a really fun book.

There's a book along those lines that I love. They're making a movie of it right now called Hail Mary.

Oh, I'm in the middle of reading that right now.

Okay, awesome.

Yes.

We're of the same mind. Yeah, they're making a movie of it. How about that?

I'm in the middle of reading it. It's getting a little wacky where I am right now.

Yeah.

I'm excited to see where it goes.

It only gets wackier. The ending is extra wacky.

Oh, really? Okay.

Prepare yourself.

Okay.

What is a recent movie or TV show you've really enjoyed?

I love The Bear. I think that's just an absolutely awesome show. Dune, of course. And I thought the new Top Gun is a little old now, but I think the new Top Gun was so fun and awesome.

Is there a product you've recently discovered that you really love? It cannot be AI Mode.

I'm gonna use a nondigital product.

Perfect.

I'm super into this new pillow that I got called Purple Pillow. I've been recommending it to everyone at work. We're on a pillow chat now. It's a thing. We talk about what pillows we're getting, but it's this really cool thing where it's got this new technology of a honeycomb polymer inside. It supports you, and it has these little micro-holes, so it doesn't get hot. It's really cool. Big fan. Strongly recommend Purple Pillow.

Wow.

Yeah.

I have never heard of this thing. I am excited. I recently got an Avocado Pillow, focusing on low toxins.

Oh, those are good. I've heard good things about those too.

Yeah.

Yeah.

Okay, I gotta join this pillow talk. It's a great name for it, by the way.

I didn't know you were into pillows too. That's great.

Huge.

I love bedding.

No, I'm just joking.

Yeah, great.

But I did upgrade my pillow. This is not Mr. Pillow or whatever that guy's product is, right? The controversial pillow guy.

No, no, no.

Okay. Purple Pillow. I'm gonna ask AI Mode about this.

Yeah, you should. Definitely.

Next question. Do you have a favorite life motto that you find yourself coming back to in work or in life?

I think it's “Be curious.” I almost named a company Curious. In terms of getting things done and understanding the world, people, your kids, and your family, you always just wanna know more and question things outside yourself, not feel like you have all the answers. I think it's really important.

I love that. Final question. Speaking of startups, you started a company called Stamped back in the day. It got acquired by Yahoo. I heard there's a story where you got Justin Bieber on your app, and that was a big deal and a big inflection point in the success of the app. Can you just tell that story?

Yeah. It's kind of a wild story. Just to set the scene a little bit, I was 25, right after Google, where I had been an IC PM, in New York with some Google friends building this company. So we were very early on and, maybe in a good way, had no idea what we were doing.

Basically, we decided that the concept of Stamped was to put your stamp on your favorite things and get recommendations from friends and people you trust. Think of a Twitter feed, but it's all stuff that people think is cool. It's like books—

Like products.

Restaurants, maybe food. Products, exactly.

Pillows, possibly.

Pillows could be on there. I would totally stamp this pillow, and then you could discover it. One of the cold-start problems was obviously that you want a group of people who are on it and already using it, who could be tastemaker-type folks.

We had a bunch of people who were chefs, and we had people who were kind of literary folks. Then we wanted to get a couple of people who were more musicians, artists, and these kinds of influential folks. My co-founder and I basically got the contact for Scooter Braun, who was Justin Bieber's manager, and we sent him an email. We were in New York, and we said, “Hey, we're gonna be in LA tomorrow.” I think we said something—I don't remember all the details—but it was something like that.

And you were not gonna be in LA tomorrow.

No, no.

Okay.

“Do you happen to be there?” He just wrote back some one-line thing like, “Meet me at this hotel for breakfast.” We were like, “Okay.” We immediately went to the airport. I just remember going straight to the airport, flying to LA, and meeting with him.

We gave him the whole pitch, showed him the product, and he was like, “Okay, I think this would be super cool. We can help be involved, and maybe you can be an advisor.” We ended up going back and meeting with Justin Bieber, showing him the product, and even filming some little clips with him. It was actually really funny.

It was a really fun moment, and he was using it to stamp his favorite stuff. People were like, “Oh, Justin's into this song,” or, “He's into this stuff,” and would post that. It was one of the ways that we got lots of people to try out and see what we were doing.

So that's a little extra scrappy moment in time. But I think it embodies a good lesson: do it now, be scrappy, be immediate. Intense urgency usually wins over thinking about it for a long time, and that certainly proved to be true in that case.

Incredible story. Thank you for sharing that. So many lessons to take away. 2 final questions. Where can folks find you online if they wanna reach out, maybe learn more about what you're doing? And how can listeners be useful to you?

Yeah.

I think on X, @rmstein is probably the best single place. And then, to be helpful, send me feedback. DM me, just mention me, ping me, let me know problems with Google products, with AI in general, but also just anything. As I said before, you have to always listen to people and understand their experiences. So ping me ideas and feedback. That's the best way to be helpful.

Wow, what an onslaught of feedback you're about to receive on the search experience.

No problem. Yes, please do.

Robby, why is this link second? Why is my site not at the top? I can only imagine the kind of stuff people complain about. Robby, thank you so much for being here.

Thank you. It was great.

It was great. Bye, everyone.

Take care.