Vibe Coding Renaissance of Gemini 3 | Logan Kilpatrick from Google Deepmind

I screenshot the AI studio. I drop it into AI studio. I say, clone this. He literally just types in makeup, blah, blah, blah. And all of a sudden, Gemini just goes nuts coding. In AI studio, there's literally an idea of everybody. If you want the inspiration, you have no idea what's possible with Vibe coding, you click one button, everyone is actually instinctively creative. He's teaching his 40-year-old some kind of crazy hack with Vibe coding. I think the world and the tools shackle you in a way that makes it so that you don't feel great. The creation of software and the distribution of software is like one of the most economically impactful things in human history. Hundreds of millions of more people being brought into that. We're going to see this run us on today. I'll be right back. Welcome to the AI Applied Podcast. Today, we're missing our friend Jaden, but we have been graced with the presence of friend of the show and great friend overall, the amazing Logan Kilpatrick. Logan and I now go back probably about two and a half years now, but it's amazing, Logan, to watch everything that's happened with Gemini and when you and I get to catch up just to hear your enthusiasm for Google, for the team. I know you moved from sort of a big competitor a couple of years ago. To me, it always sounds like you just couldn't be happier where you are right now. So congratulations. And you guys are just 2025 was the year that Google sort of exerted a little bit of dominance on that. So for our audience, I know you're a friend of the show. Is it your third time back? It might be my fourth time back. People are going to get this lobbying guy off. It's great to have you. Thank you for taking the time. I know We have crazy schedules with the Gemini 3 blowing up. Hey, listen, just to kind of kick everybody off, can you kind of explain your role inside Gemini, especially for like a non-developer audience, like what you do and how it impacts people's experience out in the real world? Yeah, no, that's a great question. So my role is helping lead our product team and actually our developer relations team as well inside of DeepMind, specifically for our developer products. So my team works on the Gemini AVI, AI Studio. and AI Studio is sort of going through this renaissance of now enabling Vibe coding, which is this really, really exciting experience to see this come to life. So I spent a lot of my time Vibe coding these days and sort of the transformative moment of what it means to be a developer changing. And now like everyone has the opportunity to be a developer, not just people who started computer science and sort of know how to code. It's been really cool. I think Gemini 3 actually brings that narrative to life in a way that I think like we sort of all in an abstract sense saw that coming. If you're sort of tracking the progress and the models from a coding capability standpoint. I think Gemini 3 delivers and brings this to life. So this is my only product plug. But if you haven't tried Gemini 3, ai.studio slash build, five code something, bring whatever your idea is to life. Be ambitious with this. I was talking to Cunery yesterday and he was telling me about something he wanted to make. And he was like, I've got these 30 requirements that I want for this thing. Give me five. And I was like, give me the 30 because really Gemini 3 is that capable of a model. I think we sort of all have tried to not push the models too much because we're worried about, is it going to get it right? And I think now Gemini 3 gets a lot of those things, right? So you have to be more ambitious with what your author models. You know, let's actually just stay right there for a second because so as somebody from a non-technical background like myself and then you who actually do, I think, a really good job at sort of like trying to balance those things. It's almost your entire world. You know, to give you an example of sort of like how I remember you and I were, I think, out there in Chicago and you and I were texting our mutual friend, Ali K. Miller there, and you're like, oh, Connor isn't vibe coding. I was like, you got to vibe. This is why I just vibe coded the last 10 minutes or something. I'm like, all right, so I'll get on it. So I got my guy, Robert Haslam, who I work with. And I'm like, Robert, talk to me, right? Because he's a huge Gemini user. And he's like, well, what do you want to do? And by the way, and Logan, you know me, right? Like, this is my world. I go out to teach companies. Like, this should be my world. And yet I don't really vibe code because I'm falling into the trap. I always try to warn others against switches. No, just try it. But even to me, it felt intimidating, right? So like this whole idea of like vibe coding, you hear it and you're like, no, I know I don't have to know anything, but still, I can't just ask it. And so Robert sort of took our little AI mindset website. He's like, look, what do you want to do? I'm like, I don't even know. He's like, okay, I have an idea. Let's make an AI mindset branded Wordle with all our kind of key terms. And I'm like, okay, how do you do that? He's like, Conard, idiot. And he literally just types in makeup, blah, blah, blah. And all of a sudden, Gemini just goes nuts coding. And I say all that to say for our audience, this is what it really is, right? Logan, it's sort of like you literally just type in a prompt and say, because when people say oh i one-shotted this you know angry birds thing or something like that it really is that and you can't break it right like what's been your experience when people have come to you been like well what do you mean yeah i think one of the i had a really interesting a b test of this which is i was at an event with some sort of like leaders across google in arizona and amar and who's on my team works on ai studio uh and i were sort of like spending a bunch of time demoing a studio and we're sort of like i obviously there's a huge range of breadth obviously not And everyone at Google is software engineers, those people in sales and marketing, all these disciplines. A range of technical audiences as well. And we were sort of A-B testing at real time as people were walking by and sort of seeing these demos, like what resonated with them. And this is a long-winded story of saying sometimes you, what I experience is for people who have never VOD coded before, the thing that gives them this aha moment is this thing that they're familiar with. I think like your website is a great example of this. Like if you have this thing, this digital interface that you interact with on a regular basis or a non-digital interface that you interact with, trying to bring that thing to life using Vibe Coding. And literally like one of my most common flows is I work on AI Studio as a product. So I screenshot the AI Studio. I drop it into AI Studio. I say clone this. And then I suggest clone this UI, comma. And then I say the thing that I want, like as a speaker, move this button here. And then I can sort of do this product iteration in real time. I think if you can find the example that's like relevant to you and is like in your life, it really does bring it to life in a way that I think in like the abstract, you know, I can type something in and get something out. It solves this blank prompt problem, which is also another thing we spent a bunch of time in the studio. Like, how do we get actually inspired by some of our conversations or just like the blank canvas is this really powerful tool, but it's also super intimidating. and one of the sort of things Google had had for a really long time is this I'm feeling lucky button I said go in AI studio there's literally an I'm feeling lucky button you just click it like if you want the inspiration you have no idea what possible with vibe coding you click one button and you can automatically start to see how capable and how powerful it is And like we instantly removing that friction of like what could be possible It something that been super tough So I love that. It also just reminds me that the people who I love, you know, we probably follow a lot of the same folks on X and you get followed a lot on X. Again, our friend Allie had that. I don't know if you saw it on X. She kind of blew up with this, you know, Gemini 3 doing a sign language recognition thing. And she was just like literally doing it. So people have seen this. It really has blown up. Blown up. And I guess it's because you don't know what's going to blow up. And part of this reminds, because we do, you know, we think a lot, you guys must think a lot too about like, what does this do for labor? Right. I mean, like we get asked this about this all the time in AI. And one of the reasons I'm so bullish on what this does for labor, it's obviously going to take a lot of tasks or obviously going to like have a huge disruption in the labor market. But, you know, there's a lot of like, this is going to steal creativity. But, you know, you and I are both fans of Ethan Mollick and of Ali and others. And I am of what you do as well. And what I see all the time is it still takes a creative mind to just come up with stuff, but it just allows you to do that. So like people who are creative, I feel like can lean into this in a way that's not lost. Right. I mean, that's like to your point about the stuff, like things just I'll take it one step further, which is like my personal journey. But my feeling that everyone is actually instinctively creative. And I think that is sort of the world and the tools sort of like shackle you in a way that makes it so that you don't feel creative. And as somebody who's like not artistically inclined, I'm not super talented in any way. As I spent the last year and a half using a bunch of these image models, I think it really has given me like more. It's expanded my mind of what was possible and it has given me confidence to actually go and try some of these things and help bring some of my ideas to life. Because it feels like I have a thought partner. I feel like I have a creative partner. And the coding example, this is one of the most exciting through lines across all of these capabilities as people are having that partnership, that AI human collaboration that I think hopefully gives them the confidence to actually try this stuff and like bring their ideas to life in ways that weren't otherwise possible. coding, I think, is that example. Nano Banana Pro is that example with like really creative visual creation. I think VO is that example of her video. So like we're seeing this across all these trends where people now have the floor is bringing it up for everyone. And it actually increases the value of creativity in that process, which is really, really cool. Yeah, I totally agree. And I think it's why Apple got a little crushed, no pun intended, when they had that advertisement where I crushed all the instruments into like the iPad and people like, oh, like the whole point is the creative process. And I think that, you know, all this is going to do is augment how you think. But you're right. It opens up new things for all this. And I think there's a very human thing to not want to try things because you're going to feel stupid. Like if somebody said, hey, draw this thing, I wouldn't want to do it because I don't want to look stupid. But maybe that sort of leads into another question around. You mentioned Nano Banana, which is just absolutely unbelievable. And now Nano Banana Pro, which I only sort of just some of my other friends from Google were sending me this new trend, which I didn't realize was a trend until like yesterday. But, you know, taking like a LinkedIn post and making it into an infographic and things like that. It's amazing. but it's also practical, right? Like, so I've done this thing a million times where I'm like, change the color of my shirt. Now change it again, now change it again. But one of the things that I see in companies anyway is a lot of drive for people to use tools is just people discovering quote, quote, cool stuff, right? And we haven't had that many moments of cool things, but Gemini has owned a few of them. And so when I think about like Notebook LM, which was one of my three big jaw-dropping, like, oh my gosh, what is this thing, right? Nano Banana, which is like, holy cow, this is like just different. Talk to me a little bit or talk to our audience a little bit about like, These are like, and vibe coding is the perfect example, right? You can, again, vibe code, you know, I think you were talking about vibe coding like Flappy Bird or something like before. I just thought it was super cool. But like, these all seem like really cool individual things. And I know my daughter, Lucy, 14, she uses Notebook LM to learn biology. It's incredible. And other people, how does this fit into the sort of the thesis of how Google is as a company? Like, they all seem like really disparate things. Like, is this all driving underneath towards something else or what's happening here? That's a great question. You know, one of the blessings and sort of curses of building products inside of Google is the sort of breadth and the range of what Google does. I think in many cases, it's jaw dropping to just see how much reach there is. There's like six or eight billion plus user products, many products, which many hundred users. It's incredible to see that. And in a lot of cases, very different user personas of who uses these products. And that's actually, there's this interesting thread now, which I think I had this realization at IO this year, which is Gemini is now becoming this through line across all of our product suites. Historically, your Google sign-in was the only thing that was connecting you across these things. But Gemini is now this cross-Google-wide effort, and it brings these models to all these different product services. I think that will, over time, help fuse a lot more of the product experiences, which Obviously, Gemini is the through line, but some of the product services are still desperate across the board. I think that will help to be the glue that's bringing more of those things together. I think it's also just an acknowledgement of this moment that we're in, which is like you need to build a new product. Like in some cases, you can sort of upscale existing products to sort of, you know, benefit from the new AI form factor. But in a lot of cases, the sort of blessing of the curse is that you have lots of users and you sort of don't want to disrupt your existing users too quickly, but you still want to be able to innovate. So in a lot of cases, you have to sort of go and build new products. And that's what places inside of Google, like Google Labs do. And they can sort of get many shots on goal of what would an ideal form factor be for audio or for video or for something flow or notebook lab. So I think that piece is critically important. And again, there's a little bit of a cost of doing that, which is like for the in the know users like us. There's a little bit of cognitive overhead of like, well, Google has lots of products now in the AI space. But I think it is important. I think it's better that we do that than we sort of don't show up and we don't page and we don't do all the things. people will do as well so there is like a very real threat of we need to make it more cohesive so that customers sort of have clarity and i think we're constantly working on that but i think overall it's on the margin hopefully we're finding the right balance and if we're not please continue to send us the drive no but but no i love that and i think that you know there's not a ton of models that come out that get i don't say universal applause because i would see some people like oh what's the big deal but gemini 3 even sort of a few days before i guess people discovered it or something like that. And people are like, this is just kind of state of the art. And I think that's why people are like, oh my gosh, this is the moment. You live in a bubble. I live in a bubble. Like we live in a little bit of an AI bubble And when you kind of get out into the world as I know you do too because I know you have a life outside this But you know there a lot of people that don use AI right And it feels to me like Gemini 3 is this underlying engine that just going to make everything better. And I think that for the most part, still most of the world, they just want like a better experience on their Gmail and on their calendar and things like that. And that's what I think has been cool for me. So, for example, as Ali is, you know, creating that sign language thing. And I'm seeing all this insane stuff, you know, all these like demos and stuff like that. For me, I had the moment, this is going to sound kind of like dumb, but you'll get it right away, which is like, I'm literally typing an email and, and, you know, Google just auto fills the rest of the sentence in a way that I hadn't seen before. Usually I'm like, Oh, come on. I was like, I actually even posted a screenshot of it on LinkedIn, which I never do, but I'm like, okay, this feels like something different. And I think it was like the day before like the release or something like that. So maybe it was already integrated. I don't know. But the point was that that feels like something where people can actually use it in their everyday life. Like where because I think we're in this bubble where like this is amazing. But for a lot of people, they just want their life better. And I feel like Google or maybe let me ask you this. Do you guys have that sense of responsibility that Google has to be sort of like the adult in the room because you're holding so many eyeballs and brains and ways of doing life? Yeah, no, that is definitely a top of my thread. And I think there's two dimensions to this, maybe even three. On one hand, Google has this responsibility to do right by the customers in the products that we own to make sure they have a great experience and can get the things that I think Google has this responsibility even more broadly with search of the stewardship position that search has as the front door to the Internet in a lot of contexts and a responsibility to make sure that we do that in a way that the Internet economy continues to be the backroom place that it is. And I think now increasingly, we also have this responsibility on the frontier model capabilities and make sure that we're doing in partnership with all the right folks across the world, the right thing from a model safety perspective and a testing perspective. There's this entanglement of the increased amount of risks that models have as their capabilities increase. And specifically for code, actually, it's one of those for the cyber risks, which various model labs and others have been reporting incidents of continues to increase as the models get better and better at making code. so there is very top of mind to make sure that that happened and like how do you sort of balance the innovation with making sure that we're also doing right by all these various stakeholders that we have so we didn't say yeah it's a top balancing yeah i'd say so and one of the reasons that it's fun to sort of encourage and why i like these moments where you know a notebook lm and nano banana and stuff like that is that it drives people to the chatbot itself right which i actually i kind of dig because google has colossal model cloud and all these products and enterprise suites and stuff like that. But the cool thing about the chatbot is now that it's really integrating into everything I do. It's just making life. Like yesterday, we had parent-teacher conferences and I have a spreadsheet of the list. And I literally just took the heart. I wonder if this would work. I literally just put in calendar invites and said to my calendar, I was like, okay, that's pretty awesome. And so when it does that, I'm kind of thinking of Gemini as a chatbot, also like an integration. But then it sort of begs the question of if Google has so many eyeballs and if more and more people are using the underlying model of Gemini. How do you guys think about the personality of it? Because you're working with a ton of developers who are going to be building Gemini into these tools that everybody's going to be using, right? Like, I mean, this is a very positive way, but like a virus sort of like throughout the world, like, you know, making everything better with AI. We have seen so much in headlines and I haven't seen it in Google. This is my question, but you know, Chachapiti goes from four to five and everybody's like, where was our friendly four? Do you know what I mean? Like in five, it's still cold. And then you have, you know, Grok and And obviously, Grok is sort of like a little off the rails. And then you have Mustafa, who came over from Inflection and Pi, which was meant to be empathetic. So that's an interesting, you know, Microsoft co-pilot is intended to feel empathetic. I haven't either read or really necessarily felt Gemini with a personality that you guys have had to keep correcting one way or another. Is that an internal conversation? How do you guys think about that? Yeah, well, I sort of have an interesting personal perspective and I can sort of share some of the thought philosophy. I think it partially speaks to this moment we're in, which actually sort of as you continue to enter into the model, it is not easy to have that persistence. So I think from like from purely a product perspective, like obviously some of these products are like getting some uplift because they have a personality that folks are engaging with. But like the practical reality of our product is it is impossible to persist that behavior from what it is like fundamentally a different model checkpoint. As you put the parameters back through the blender, you sort of get a different personality. That's just the reality of making these models. So I think there is actually a really interesting, and I haven't been in any of these debates, but there is actually an interesting product debate of like, do you even want to try to like lean too much into personality right now when you know that you won't be able to have consistency? And I think maybe the 4.0 to 5 transition for opening, I sort of showed some of that. I think our, like the challenge for the Gemini models and the cool opportunity as well at the same time is that we're making this model across Google. It's made available across Google. It's available through our APIs for developers and millions of products around the world. So we can't be, you know, you don't want to be too opinionated because like what the personality or the vibe of our model would be in Google surfaces is maybe the completely different vibe than what you want. So I think having this like really standard base set of A capabilities, but also B model persona and then really giving sort of each individual product service and developer the ability to flex, like customize. So like the Gemini app, for example, as sort of a deployment service and a product service that has the models available, should be able to go in and say, hey, here's how we want the model to respond. It should be helpful and it's a friendly assistant and it's trying to do these things versus, you know, maybe for us on the developer side in the API, we sort of want it to be a blank slate. We want to sort of have this base capability and obviously hopefully trained on sort of core human principles and it sort of let the developer control this themselves for their different applications. So I think that that approach from like a model training perspective, I think, makes sense. And then I think each product service needs to go and figure out sort of what's the right lane for them to go and how deep do they want to go. And but it is interesting that it is it feels like it's a differentiator for some of these products. Like I think even if the model's not smarter, just like how interactive it is clearly like keeps users around and engaging. Yeah it really does And I like the way you said that just because it really has to be all things to all people because Google is all things to all people right Like there not sort of like oh now we have developers and you have this like you just dealing with humanity and it has to be the same chap It must be an unbelievable challenge A couple last questions I'll get you out of here. I know that you have flights and all that kind of stuff, but you know, we don't usually think about benchmarks because it kind of feels like, oh, you know, now you have like 12 PhDs in your pocket, but are there benchmarks for this that you get excited about? I know sort of like on the humanity last exam, I maybe can just sort of say two words about what that is. Like, does that excite you? Do you sort of be like, ah, but like, I mean, to me that actually feels like where we're going with AI actually could have some significance here. Yeah, it's a good question. My personal feeling about benchmarks continues to evolve just because it's such an emerging space. And I think as you, well, I think the headline of the story for folks that didn't see is like Gemini is a very state of the art across many, if not most benchmarks or a vast majority of benchmarks across the ecosystem. And I think the cooler part of the story is even things like there are lots of folks who are skeptical about benchmarks and they're like, oh, obviously the model teams just go and find the benchmark and they're like, let's go and make the model really good at those things. But the reality is there's many, many hundreds of benchmarks around the world. And even if you tried to optimize for most of the known ones, every time a new model drops, 25 new people come to woodwork with this brand new benchmark. And obviously that happened with Gemini. There are lots of benchmarks that I'd never seen before. And Gemini's state of the art on those as well. So I think the capability transfer is really happening very broadly across the spectrum. I think the challenge is, as you sort of look under the hood of a lot of these benchmarks the question i always have is like how representative are these of what people really do with the models how representative are they of that use case that you described where you sort of take its data out of a sheet and then go in and create calendar invites and like maybe that is represented in some of these like function calling tool use benchmarks maybe it's not like i actually have no idea like you actually have to go and look and obviously the model team do some of this of like trying to understand what these benchmarks are actually measuring. But I think product surfaces as this like scaffolding to understand how people use models, I think is really, really like ultimately the best measure of this is like, do people get value out of your products and do they create value in the world? I think we'll see more and more of this. And I think we actually already kind of are seeing more and more of this where like the product is actually becoming much more a part of the benchmark story than just the model itself. And now you sort of take, for example, people are doing this with coding specifically is like the most hobbyist example of this. But like you take like OpenAI's Todex and, you know, Google's new anti-gravity developer environment and Gemini CLI and Cloud Code and a bunch of these other products. And they actually run those with those providers models as well on a bunch of these coding batch marks. You can see the model out of the box. And then the model with the tool and product scores much better with the tool and product because it has all these extra capabilities and the scaffolding and they're trained to work well together. So I think that it's a really interesting thread between those two things. And part of my reaction as well, more broadly, is that I think this becomes like another scaling dimension. These models is like actually the product scaffolding is another dimension by which you can scale and improve the quality of these models and sort of get more out of them. So it'll be interesting to see how that happens. And I'll make one last comment, which is the reaction that people have to the models across these products is also very different because it depends what the scaffolding is that they have. So if you go to one product and you sort of try the model and if it's not set up in the same way or it's not optimized in the right way, you might get a different experience than if you were to use the Gemini model of some other products. So I think as the model become more capable, maybe it becomes less of a problem. Maybe it becomes more of a problem. Actually, I don't know. I think there's an interesting thread of just like how that plays out. Yeah, it's a great way to say it because I really think I love what you said there about it's ultimately like how people use it every day. And I think that's the goal. Like, is this good? Because Google, as a company, changed the way that we work. And I love the idea that this is doing two things at once. It's actually keeping up with those folks and making their experience better, but also on the frontier. So maybe just to sort of leave off with this for you, what are the exciting things when you look at a model like this or what's coming? I mean, obviously, if you have agents and agents base and anti-gravity and all the other things that we could spend hours on, but anything we didn't get to or anything like that. But what for you are you like, man, that is just cool? Yeah, I think the Vitoating piece was sort of my personal aha moment. Like, as somebody who was actually formerly trained as a software engineer, I spent time writing code professionally. Now I write much less, I think, maybe by volume. And I probably write a lot less code than I used to. And I feel it's this feeling of empowerment that just gets me so excited. And, like, I think it gets me excited as somebody who, like, probably could go and build something if I really put my mind to it. Because I spent years learning and all this stuff. And I think there's hundreds of millions of people who want to build something who can't because they haven't spent that time yet or they don't have the means to spend the time or whatever it is. And I think that sort of brings those people into the fold and lets them start to build the things that they have. And I think it's exciting because not only does it let them be creative and expressive, but also like the creation of software and the distribution of software is like one of the most economically impactful things in human history. and hundreds of millions of more people being brought into that, I think is what we're going to see this renaissance. I think 2026 has this year for like software and probably more broad than software, changing the way that it's built and distributed and all that stuff is going to be, it's going to be incredible to see. So I think it gets me excited. Yeah, it's, that's the sort of the challenge, right? That we sort of like say, which is just, and I hate to say, just try it, but it's go on Gemini. One quick, like I'm feeling lucky and you have built the thing and your mind is blown. But that's the thing. Like, I'm feeling lucky because, again, sometimes it's hard to even conceive of what it can do. And, like, that's where those demos, that's where AI Studio, where can they find that if they were looking for that on the Internet? Yeah. AI.studio slash build. Click one button and have your first vibe coded out. I was watching it. It's unbelievable. So I just encourage everybody to go try that. Logan Kilpatrick, I know you're probably missing a flight on our behalf right now. You're working harder than anyone I know right now. Listen, man, thanks for coming on. Thanks for doing it. I'm excited. hopefully he'll be back in Chicago and we get dinner again soon looking forward to it alright thanks everybody I know poor Jaden is so bummed that he missed everybody he's teaching like his four year old like sort of like some kind of like crazy hack with vibe coding Logan you're the best anytime you want to come back love having you guys thanks for coming on AI Applied and to everybody else thanks for listening and we'll see you all next time