Google I/O 2026 Recap with Logan Kilpatrick, Josh Woodward and Tulsee Doshi

0:05Welcome back to Release Notes. My name is Logan Kilpatrick. I'm on the Google DeepMind team. Today I'm joined by Josh and Tulsee, the dynamic duo. We are back. Maybe we're the dynamic trio. I like it. I like it. Yeah. At Google I.O. 2026. So many announcements yesterday. So many fun launches, models, products, everything in between. Maybe we can start talking about models and then we can talk about products. Sounds good. Sounds good. Yeah. Let's do it. Tulsa, you want to kick us off? Yeah. And actually, I think the cool thing is this year, really, you can't really talk about the models without the products.

0:35So I think they're going to be a really nice tie in. And maybe that's the first thing is that, like, I feel like this year, the phrase we're using to talk about Gemini 3.5, which is one of our big releases, is intelligence with action. And I think that really speaks to what we're trying to bring through all of our launches this year with the models is that we're really trying to focus on how do we enter the agentic era even more seriously than we have in the past with models that have amazing tool use capabilities that are really strong for coding, that can handle really complex, like long workflows.

1:08And then how do we bring that magic to life in the product? So Gemini 3.5 Flash is our big model release. Really excited about that model because I actually think it's a great combination of performance and speed. And it does this like really good job, I think, of bringing a lot of that magic to life in a way that you can actually quickly iterate with. We also released Omni, our multimodal model. It really actually brings out like the nano banana for video vibes. So really actually allows you to edit videos, build kind of amazing experiences. And the world knowledge from Gemini that you're bringing into the model makes it even stronger.

1:46It's super fun. We did an episode with the Omni team and they edited this sort of preview at the beginning. So let's take a look at it. There's so many different kinds of edits. Okay, there's the parrot. No, one of the best parts, they're all the yellow birds, right? Yeah, yeah. But there's one where... There's the slot. I still don't know why am I holding a plant? What was the... This is... The cat. I know. There's nothing there. It's just me. Yeah, it's just Logan straight up. A few hyenas wandering across the set. I think you could just have an endless possibilities here, right? You can do anything.

2:21That video, I showed that to my girlfriend and she was like, I get it now. I think there's something interesting about how we showcase what it's capable of. So I really like that format of taking something existing that you know and people doing something normal and then sort of changing it in a way that like doesn't change the coherence of what they're saying but then like just makes it feel magical a little bit yeah like magic i was trying this yesterday with some dance videos of mine and like was having it like every time my arms move in a certain direction like sparkles would come out or like petals would fly or things like that and again like the coherence of the video doesn't change the actual dance has not changed yeah but you're just adding something to it that makes it like memorable or remarkable in a way that you didn't think about right the other cool thing with Omni, I would say it's incredible at camera angles.

3:09So I showed on stage 360 degree view, but you can do drone footage. You can take shots. One of the most amazing things I think we showed was Google flow took one shot, 16 different frames from 16 different angles. And it contains the same sort of scene consistency, character consistency. It's pretty fun. That's really cool. Yeah. Yeah. People are going to go wild with us, um, in the Gemini app in YouTube and in flow, So they can make all those dynamic videos, which will be really exciting. One more thing is like 3.5 Flash is our most powerful model to date. It is the sort of like most capable model that we've shipped.

3:45And I feel like it's, you know, Flash, I think people see as this workhorse model and sort of it's like a little bit less intelligent. And I feel like we keep doing this thing where like Flash ends up actually being really, really capable. In a conversation yesterday with Oriole and Jeff sort of just like continuing to hit home on like the distillation capabilities we have are just so good. And Jeff was like, yeah, actually, it's pretty we're pretty much doing the same thing that we had in the original distillation paper. And like it just keeps working. So maybe there's like something magic about like how we keep landing the pro intelligence into a flash model.

4:19Yeah, I think there's two things. Right. So, I mean, maybe to like hit your point home harder, like this model literally exceeds 3.1 pro. on most benchmarks. And I actually think this is a cycle we've been seeing, right? So the previous Flash exceeds it. I think there's two things there. One is, I do think Flash is still a great workhorse model, but the kinds of work that you can do with it is now changing, right? So I think when we thought about workhorse model, you know, like two years ago. This is a great tweet for you. I love this. I'm just like, because I think you say that and I'm like, I get it.

4:53That makes complete sense. Yeah, like I think it's just like the kinds of things you want to do in the Gemini app today versus what you could maybe do a year ago or two years ago are meaningfully different. And Flash is still the model powering it, but you need a more powerful Flash to be able to actually like do the kinds of things you can do in Gemini Spark, for example, which, you know, you should talk about. But like, I think that's pretty awesome. And then to your point, like this model is really an amazing showcase of post-training excellence, right? because it takes Gemini 3 and actually builds on it.

5:26And so what we've done from a perspective of RL, what we've done from a perspective of taking the feedback actually that we're seeing in JetSki and actually bringing the model and the harness closer together to actually train the model more effectively, I think that's a really great example of what that can do for model performance, especially when you think about just like real world usage, which is really awesome. Well, and I would say we're using it all over the place internally to build. I mean, Sundar talked a little bit about that on stage. He's showing our token count internally. I was like, this is cool.

5:58Yeah, yeah. Token count. It's up in the T's. It's Harrison using all the tokens. That's what the problem is. Half of the tokens is just Harrison on whatever side quest he's working on. That's what side quest encouraged. No, but I do think it's unlocking all kinds of things we didn't think were possible. And so it does come through in features like Gemini Spark. it comes through even in things you wouldn't expect, just the depth, the quality of the answer, so much better. Yeah. Gemini 3 feels like the era of product, model, harness, symbiosis, and then this sort of like flywheel, like you actually can't have one without the other to your point from earlier.

6:34And maybe Spark is like the, maybe the perfect example of that, I don't know. Yeah. And so for folks who've missed the news, maybe what is Spark and - Yeah, yeah, yeah. Well, Spark's are 24 seven, always on agent, works in the background. I think the way to think about it is when you use it, it almost feels like you're just throwing tasks over your shoulder and Spark catches them and just runs with them. So you can say, just brain dump a bunch of stuff into it. It'll break it down into individual tasks. It'll give you this great dashboard. You can see what it's doing. And then you come back to it and a lot of stuff's just done.

7:07So it's kind of this amazing step of how do we move Gemini from something you're just chatting with to like Tulsi said, something that's acting on your behalf. And so it's a big shift. We're very excited about it. It'll be rolling out next week, starting in the US for Google AI ultra subscribers. And the plan is to kind of basically see how people use it. We have no idea also how many tokens it's going to consume. So we want to start kind of small and then scale it up very quickly. We're very excited about it. A lot of us have been using it. Yeah, it's a daily active thing. Once you get into the swing of it, you realize there's so many tasks you can just hand off and come back to.

7:47And so I think that's one of the things we're really excited about. Yeah, Josh, one of the things you mentioned is sort of you throw things over the wall to it and then some of the tasks are done. For the ones that aren't, I think something you mentioned, if I remember from yesterday, is just like it actually does come back to you when it needs sort of like a human in the loop in order to take some action. If there's something like, you know, you're about to spend a thousand dollars. Maybe you should check in with me before you go and order me a bunch of crazy stuff. Yeah. Which I think is a really interesting paradigm.

8:17And I'm actually curious, like, is it if that's also one of the things that we're like, it's like still a research frontier, like how how yeah, how good it is at knowing when it should come to you versus not. Yeah, absolutely. It is very deliberate sort of product design choices right now. We're erring on the side of being a bit more conservative because one of the main pain points we've seen when people set up their clause or anything else is they'll like, sure, I'll YOLO this. And then they realize like, no, no, I don't want that to happen. So we're going to start a little bit more conservative.

8:47We'll check in. It'll say it needs input. Usually you go in, you say yes or no, and it keeps going. And then we're going to gradually allow you to kind of set, like always remember this choice. So we want to make sure we build up that trust and that comfort and get it right. So today it might ask you if you want to book a calendar meeting, and then you can be like, yes or no, and then you can kind of remember that. Anything with payments, anything with accepting consents, these are all areas where we want to stop it, have it come back to you. But I'm very excited. One of the other announcements Vidya did actually yesterday was some of the stuff we're doing with some of our agent payments protocol.

9:23This is very interesting. It's very Google in the sense that it hooks up to Google Wallet, and you'll be able to set budgets for your agents and even per merchant sort of constraints. So I'm okay to spend this much money with this merchant. And it's really interesting as we're testing this, it feels a lot like you're giving almost like a teenager their first debit card. It's an allowance. Exactly. So it's kind of constrained, but I think that's a whole other area by people using all the cards they already have in Google Wallet to be able to sort of spin that way with their agents later this summer and into the fall.

9:58Yeah, and I think actually this is where we're going to start seeing how different users have different boundaries of comfort. Yeah. Right. And I think we want to be able to provide everyone the flexibility. Right. I think there's going to be users who are going to be a lot more comfortable saying, yeah, I scheduled this meeting. Remember it. I'm good. And also based on the constraints of maybe their job or their personal life or who they're even setting up meetings with. And then there's going to be others who maybe want to keep that level of feedback loop. And you should be able to have the flexibility to decide how you want to engage and what level of kind of control you feel comfortable with, I think.

10:31Yeah. And I think actually for both of you on this, one of the things that we announced yesterday was that the Gemini app sort of passed, I think, 900 million monthly active users, which is crazy. Josh, you're on the war path to get, you know, what's the slide where we show the billion user products? I'm like, I showed, I saw yesterday we put it up. It was like, what, it was like 12 products or something crazy like that. So hopefully, hopefully Gemini apps. Josh, is this like a big party? Yeah, that would be, you know, no, no, no wait till I own next year, but we should definitely have a party when that happens.

11:02I feel this tension for both of you sort of as we try to make a model that works for everyone, as we try to make a product that works for everyone, like obviously the like, I'm still showing up like the next hundred million users to come to the Gemini app, like maybe it is the first time they've ever like used these product experiences. And then the other end of the spectrum, there's a person who's like, give me a 24 seven, always on agent that like, I'm just throwing things over the wall too. And I've got my payments set up and I can sort of give it a different budget. And I'm curious, like it feels like the frontier continues to push forward.

11:33And yet we still have that capability overhang. We still have people who are sort of just coming into these new products. And for both of you, like the tension of trying to build in that world, I'm curious, like thoughts or reactions. Yeah. When I first joined Google, I worked on search and I worked on YouTube. And when I joined, those were already massive products that had really started to already build the muscle of how do you build for a very diverse set of users, right? How do you think about running live experiments in ways that you can get insight from how various different users use the product?

12:04How do you really lean into user research with different segments of users? And I think it's awesome actually that now we're at the point where we're having this conversation about how do we do this with our Gemini models and our product. And so I think for us, actually, what started becoming even more important on the model side is actually this idea of like feedback loops with every single product surface and not just with the surface, but actually going deep, like a level deeper to say, OK, what are the feedback loops with users who are using maybe the flashlight version of the app versus Flash versus Pro?

12:36How is that usage differing? Right. How does that usage differ in different regions? How does that usage differ for like power users versus users who are coming in like once a month? And actually like being able to run those live experiments, get that feedback and actually like take that and say, OK, what are our priorities then for the model? We're seeing that even just actually as we are starting to use the models more and more for these kinds of always on tasks internally, it's been really helpful to get that feedback. But that does look different than what external feedback looks like. Like even running a live experiment on our internal version of anti-gravity versus our external version of anti-gravity give us different results.

13:10And so like now it's really about how do we scale that feedback loop, which I think has been really cool. And it's kind of awesome. Actually, we're like now pulling in some of the best tools that we've been using across Google for years to actually like even evaluate our results because we have that infrastructure in place and we have so many people who have done this for so long. How do we like bring that into our model development process? Yeah. Well, and I think that on the modeling side, it's all of that. And on the product side, it's an incredible design challenge. How do you design an interface that can scale to someone's first time they've ever touched AI?

13:44Which there's still billions of people in the world who haven't actually fully adopted or tried AI. To people doing the 24-7 agent, setting up payment, getting very sophisticated. So I think we're trying on the Gemini side, at least, to think about how do we design it in a way where it kind of scales with your level of comfort and skill. And so you'll see things like even how we launched Gemini Spark. First, it's for the paid users to start. They tend to be a bit more sophisticated. They want to open up the hood, upload their own skills, set up heartbeats, schedules, triggers. So it allows for all of that.

14:16But then you also saw us do something like the Daily Brief, which is a very accessible daily morning digest that's personalized for you. It's all running the same 3.5 model under the hood, but it kind of how we package it and present it is different. And I think maybe that's actually a really important point, which is like as the models get better and as we tie to your point about this whole symbiosis of like model harness product, I think actually a big need for the model is to be steerable. And then a big need for us between model and product is to actually start designing these things more hand in hand so that we actually build in that notion of like what is good for different types of users in how we're actually like building this out.

14:57So like, okay, if we want to be able to support that level of flexibility, what needs to be true about the model? And then how do we actually design that end to end? What I hear, Tulsi, is that your job just becomes more difficult over time. It's like, no, don't just make a model actually also work with every product. I mean, you're already doing this today, but work with more products or cost-cruple. The amount of time we spent actually, like Josh and I were talking about this earlier. Like one thing that I think has gotten a lot better over time with Gemini is like how our teams work together to actually land the plane.

15:26It's not like we build the model and then throw it over the fence and we're like, good luck. Enjoy, right? It's actually like PMs from my team are sitting with folks from Josh's team. We're literally iterating on system instructions together for different features at the same time, then running a live experiment and being like, okay, system instruction maybe went a little too far. How do we actually kind of recalibrate? And like that process, I think has been working even better and better every launch, which has kind of been kind of fun. It's beautiful to see. I'm curious, like from last year to this year, anything that's like surprising as you look at the set of announcements and things that we've landed over the last couple of days?

16:03Okay, I have one. Last night we got together with some of our power users from across a lot of our projects, Gemini and Google Labs, AI Studio people too. And one of the questions in the audience was, of all the things you announced at IO this year, how many of them were on your one-year roadmap? And I thought for about, I don't know, a second, zero. So I actually think that is probably the biggest shift is the things we were doing on stage at IO this year, a lot of them weren't even in our minds yet. And I think it's a testament to sort of how fast the capabilities are coming and how much we're trying to invent new things.

16:43And I eventually settled, I thought about it to try to give a little bit more satisfying answer to the guy. I was like, actually, maybe we have like a 90 to 120 day roadmap. That may be all we have. And then I thought about it. I was like, maybe for the last five years, that's kind of been how we've been operating ever since it all got started. I think to give you guys more credit, though, I do think it's like you, I feel like you sort of know directionally you want to do like, you know, it always on assistant. He's like sort of like direction, but it's like, you don't know how to make it happen.

17:13And then sort of the capability lands. And then it's like, all of a sudden you have the how maybe. But I still think, I think you're right, which is like the actual product implementation. It's not like we were working on Spark. We were loosely working on Spark maybe for the last year, but not in the way that like people would think. Yeah. I mean, we had something in Gemini agent mode last year. There were kind of early glimpses of it. But then one day you wake up and it's like, it's here. It works. I think it's like a really important thing for people building out there. And how we try to think about it is like, you want to be right in that like almost possible zone.

17:46Because then it'll tip. one fine day. And then it's like, here we go. Hopefully with 3.5 flash. That's right. Well, I will say a lot of things tip for us. And I mean, we haven't talked a lot yet, but there's a lot of labs projects that suddenly with 3.5 flash have just become possible. What are some examples? Well, I mean, when we showed on stage yesterday, we talked about Google flow a little bit. A whole big part of flow is the Omni model. But another part is this assistant director. You can now talk to flow almost like you're talking on a set like this and you'll be able to be like, change this, change that, change that.

18:18It's like, just does it. So that's incredible. Everything with flow music is also leveraging the same 3.5 sort of goodness that showed up. There's some very exciting notebook LM stuff we have coming soon. So watch out for that. But I think it kind of completely rewrites kind of things we thought were impossible are now possible. And so that's like, oh, you always want to be on the lookout. So now we're like having to reset our expectations of like what's going to be almost possible for the next rev. This is very fun. Yeah, this is my advice to everyone about just like you have to be I framed it before as like you have to be reset your ambition because I feel like the models humble you so much to be like, you know, that will literally just not work.

19:02And so you're like, OK, I'm going to try something else, try a different version of this. And then you sort of sort of set your bar for ambition. And then every time the model crank turns, you have to like reset it and go in with a fresh mind, which is so as a, as a human building with the technology, it's so hard to do, but like it is, I think it is where the alpha is when you're building this stuff. It also is like, there's something really amazing about people who can play with the models and actually find the magic in them too. Because I think some of it also is like, yes, we are, we are intentional about the capabilities we want to build into the models, but they also continue to surprise us.

19:33Right. And it's not really until you start playing with the model and trying something, like trying to build an assistant director that do you realize like, Hey, the model is falling over in places we didn't even realize it was falling over. Or B, there's like something magical that you can eke out of the model that you didn't realize. And then once you find that, then it goes like really quickly from there. And then you also see that kind of industry-wide because then you're able to like, someone builds on your idea, you build on someone else's idea. And there's actually this kind of like cross industry development magic that's happening right now too, that I think is kind of cool and very unique because everybody is sort of learning from each other, whether you're in the same organization and building the same product or actually building kind of cross products.

20:13Yeah. Oh, can we talk about one other? Please, please. Voice. Yeah. I think voice is having this like right on the cusp. It's a moment. It's a moment. I mean, we can talk a little bit. I know you all teased like AI studio, the mobile pre-registration. I showed some stuff with the Gemini Mac app using voice. There's all kinds of stuff in Gemini live. I think anti-gravity is going to have voice. There's just so much that's going to be possible there. I'm really proud of the, the, where the audio models are getting to, I think like, um, you know, there's two parts to like the voice part that we're talking about here.

20:48One is this whole idea of like, just being able to throw out your ideas, like being able to almost like ramble your, your way at a dock or at a, uh, at actually you can already do this in anti-gravity. Like you can just talk and, and see what comes out of the other side. And it's actually so good. That's actually been my number one workflow change right now is I'm just talking at my laptop. You can do what you do best. I can do what I do best. I just talk. Just ramble it. I ramble away. It's great. It's awesome. No, it's actually though. It's kind of weird. And then also the dialogue models are getting really good.

21:19And so actually like the, for those who haven't been building with them yet, even you can already build with them in AI Studio. You can already build with them in the API. the naturalness of the voice really does make a huge difference in the experience. It's good. The expressiveness, the dialects we showed, pretty fun. So it's going to be, I think voice is going to be a really interesting next few months. And also, it's selfishly for all of us, it makes doing a lot of these demos way easier because if you're like, you historically have to like, this is my least favorite thing is like trying to like talk while I type things.

21:52And I know Tulsi, you, you cannot multitask. So this is perfect for you because you just talk right into it. Actually, Tulsi, you need to answer the question. Something surprising from this IO. Oh, so I think to the point about like what we didn't predict, I think one thing that's interesting is last year we talked about how the through line of IO was really Gemini and Gemini coming into all of our products. I think this year one through line is actually anti-gravity and the harness. And I think it was interesting as we started putting the package for IO together really realizing how much that actual, like, the harness consistency across our products was starting to come through, right?

22:30So, like, behind Gemini Spark, behind things we were showing in AI mode, obviously behind Anti-Gravity, the product itself, AI Studio, we're really starting to see that actually not just having a consistent model, but actually building on a consistent harness speeds up our productivity. It makes the products more consistent. It allows us to push the model capabilities as well. I think that also was something that I like didn't expect to to see be as like consistently through. Yeah, that's a great one. Predictions. I owe 2027, which I was saying to someone yesterday sounds like a made up year. Sounds like 2027.