Google I/O 2025 Recap with Josh Woodward and Tulsee Doshi

0:04Hey folks, Logan here from the Google DeepMind team. Welcome back to a special edition of Release Notes. I'm here with Josh Woodward and Tulsi Doshi. We're sitting in I.O. right now, day two, and excited to talk about all the stuff that we just launched and shipped. And it's been awesome. You two killed it on the stage. It was wonderful to watch. People were, the crowds were cheering. The people online were saying great things. So thanks for sitting down and chatting. Yeah, great to be here. Yeah. I think takeaway from day one for me was very positive. Like people were really excited. We launched a bunch of great products.

0:38I think I pinged someone last night and I said like this year's I.O. feels like Google at its best. Like really like and showing I think this is like the first I.O. to me that felt like AI native and not like us like sort of taking the first step into into the AI moment. So I'm curious to get both of your reactions to everything that happened yesterday. I mean, yesterday was a blast. I think as I was walking around afterwards, someone referred to it as like a Google carnival. and I really do feel like the energy felt that way. It felt like there was a lot of excitement and energy coming from us and then a lot of excitement and energy like being reflected back about what we were sharing and that was really awesome.

1:19And I think for me, it really does feel like this IO is so much of not just launching new products, which is awesome, but also like a momentum story. It actually really feels like we've hit our stride. And speaking from a Gemini perspective, like it really does feel like we're not just telling a story about Gemini as a model being awesome, which we are, but we're also telling a story about Gemini in our products and actually landing with real users in real ways, whether that's like from a developer perspective and the API and Jules, or whether it's from like AI mode perspective, or it's from like a glasses perspective.

1:55Like we're telling the story about how Gemini is at a moment now where we can really bring it to users in the way they want to use it. And I think that's awesome. Yeah, I mean, I think that's kind of the theme. If there was a bumper sticker for yesterday, to me, it would be like, from research to reality. Yeah. It's actually taking a lot of the research. Yeah, that'd be a good one. I could get you a shirt that says from research to reality. It's a Gemini logo. Yeah. Well, I think if you look across the developer products, consumer products, we didn't even talk about any of the enterprise stuff that's happening.

2:23That was Cloud Next a month ago. I think it's a whole bunch of stuff around the company coming together. and IO is always just a crazy time. You're trying to put stuff together sometimes the last minute. Mostly morning. But it's really great. So I think on our side, we had a bunch of stuff coming from Google Labs, the Gemini app. So exciting to get to share some of that and look forward to everyone's feedback too. Yeah, I actually want to start off with another question that I won't take credit for this. Somebody asked last night, I was doing a panel with Yunhan and Biba who are on your team and the panel host asked what we expected IO 2030 to look like.

3:04And it was a really interesting question and they both gave great answers. So the stakes are high for both of you. But I'm curious what your take is of like what IO 2030 will look like. 2030? And maybe I'll give the answer that I gave first. Yeah, yeah, it started up. The comment they made was actually in line with what you said, Tulsi, which was like, I think historically Google has been a company that has had many, many different products and like there hasn't been this like threat. I think the only thread that sort of tied them all together was like your Google account in some ways or like Google infrastructure.

3:38And it feels more and more that like Gemini is this one product thread that brings all of everything that we do at Google together, which is actually really cool. And I was thinking about this in the developer ecosystem. Like we just haven't like different developer communities have historically been like very disconnected. Now like Gemini and Android Studio and Gemini and Chrome and Gemini for web developers and et cetera, et cetera, is like bringing everyone together. And I think we'll see more and more and more of that being the narrative at I.O. of like that being the mechanism to bring Google together and our customers together.

4:08But I'm curious other takes on I.O. 2030. Yeah, do you want to go first? Sure. All right, yeah. I mean, I think it's actually interesting because when you say IO 2030, I feel like there's like three things that come to mind. One is just like, I mean, the technology is going to be wild by the time we get to 2030, right? I mean, we're already seeing like just from 2024 to 2025, like what Gemini can do has been amazing. And so we're going to see this world where like so much of what Demis presented yesterday on this vision of like the universal virtual assistant and this idea of like Gemini as this like proactive part of your life, I think is going to be like actually realized materially.

4:48So I actually think the like way Gemini, the way we're talking about Google products and the way that you were expected to interact with technology and interact with the world is going to be different. And so I think that's going to be a big part of IO. I think the second part of IO to your point is going to be about, I think we're, we're really bringing these pieces together into a more unified experience, right? So it's not like I'm going over here and I'm going over here and I'm going over here and I actually like Gemini is working for me across my surfaces across my devices across like you know And then the third I think is just I expect I O will also look different Yeah in in 2030 Yeah, yeah the internet is gonna look different in 2030 and the way we like consume and engage with content is gonna look different You just send your agent to I O and then it consumes all this stuff Dun dun dun no, but it's just gonna look different I think I think what's also funny is like 2030 feels like so far So far away.

5:41Yeah, so Mark Earth's reaction when you said 2030, I originally was thinking like... Maybe this is 2027, I don't know. Yeah, maybe we can get to Thursday first or something like that. I mean, it's interesting. We have this exercise in Google Labs sometimes we do where we think about like, jump in this kind of like magical flying saucer that shoots you out into the future. You get out, you have a few minutes to look around, write down everything about the future, and then you come back. We've only done that exercise through 2028 right now. So I don't know the actual extra two years. 2030, who knows?

6:10But it does feel like, I mean, one of the things even we're seeing a lot with some of the stuff we launched yesterday at I.O. is this kind of blurring the lines of kind of what's becoming possible. Yeah. And sort of the lowering of the bar when it comes to like how much time it takes or how many skills or even the cost of things. Or who can use things. And who and all that relates to who. So I think when you think about the future of software development or creativity or knowledge or any of these things, yeah, four or five years out, it's going to be really interesting. But I do think it's almost like the arc is towards democratizing a lot of this stuff.

6:45And that's probably when you look at products like Flow, like we launched yesterday or any others, that sort of feels like that'll be a principle. And then, yeah, in terms of the show format and everything, who knows? I mean, to that point, it was also so cool to see how present VO was in the show yesterday. Right? And actually like... Telling the story. Telling the story. And actually creating IO. Yeah. And that's like, I mean, this is like year one of us doing that. Yeah. Like month one. Month one. Yeah. Week one of us. Yeah, yeah, yeah. I think if you just use that as an example of how IO, the literal show is going to change, We actually used like Gemini and VO in so many places to like make the show today or yesterday.

7:31I think that's going to be like super, super wild. Yeah. Yunhan had a great comment on this. Who's on your team and is one of the post trading leads. And she said that her expectation for 2030 was that we hopefully will be able to announce like major scientific breakthroughs that were enabled by Gemini happening. And I thought that was such an interesting answer. where like you could imagine back to this threat of like democratizing access. I think the cherry on top is like, it's a bunch of college students somewhere in the world who are like using Gemini out of the box and like discover some like new scientific principle or like they cure, you know, cure some disease or something like that.

8:07I think that would be like such a, such a crazy and wild story. Yeah. Well, I think that was one of my favorite parts of the show yesterday at the very end, Sundar kind of pulled back and was talking about how AI, like today's research will become the reality of 2028, 2029, 2030, etc. And I think that, think about scientific breakthroughs, whether it's in medicine, the disaster relief stuff with Firesat or Wing delivering, you know, medical supplies or water, Waymo, isomorphic labs. I think there's a whole bunch of stuff there that is just like, it's going to be really interesting to see how these things stack on each other.

8:46It's really cool, I think, to think through also, like, where I think oftentimes when we talk about AI, we often talk about like the sort of more consumer tangible, like opening up the world, the way you can engage with AI on the day to day. But I think the point around scientific breakthrough and the examples you give also speak to like, what is the role we can have in the world and in the concerns we're seeing today, right? Whether that be around like climate change or whether that be around access or healthcare, like how do we take some of those principles and break them down, I think becomes really amazing.

9:20Yeah, I think this is the other thread from yesterday was just like the breadth of Google stuff. And I think the science piece is like, it's awesome that we can, you know, do VO and like have all this like really cool, fun technology and then like also walk the walk and like do real hard science and help the world through that lens. But let's talk about some of the launches. Both of you are responsible and we're announcing a bunch of the stuff yesterday and we can do whoever wants to, whoever wants to go first, Josh, year. Yeah, people want more jewels, people want more flow, people want more Vio, more Gemini apps, so I think that's a crazy amount of demand.

9:55Well, there's a lot. First, it's just a kind of unbelievable team effort across, so anyone that kind of sees people up there, you just know there's a massive amount kind of behind and working on it together. I mean, let's maybe start Vio3. That one was pretty fun to announce. So the big headline, you know, you cannot just animate things now and make pictures move, they can talk. And they have sound effects and background sounds. And I just remember the first time I was watching these come across when we got the checkpoint, just blown away. It's kind of the reaction we've seen in the first 24 hours is it's just really hard to imagine you can describe something in natural language and it just comes to life.

10:35And so that was the model breakthrough. It's state of the art. It's building on a state of the art VO2 model from December. It's a massive amount of creativity. It's incredible. It's really incredible at the core research level. And then we basically, less than 100 days ago, we were like, we need to build a product that's worthy of this model. And so that product's Flow. It's a tool for AI filmmakers. We've co-created it intentionally with folks in the industry. So these are big names you might recognize in sort of Hollywood, all the way down to emerging AI filmmakers who are making their first films using AI natively.

11:15So we've tried to get a whole range of opinions and kind of in a typical sort of Google Labs and sort of Creative Lab and Google DeepMind way, we've kind of co-created. And so the product, the features, the back and forth has literally been tested with just this like great set of creatives. And so when you see things like, oh, you can trim a clip, of course, any product can do that if you're editing video. But when you extend it, there's so much that went into that, even though it looks so simple on the screen, and so much input from folks that are making films of like, oh, if you could do it this way, it'd be amazing.

11:49And so we just tried to build that into the tool. So yeah, it launched yesterday right on stage. The TPUs immediately got very hot. I think we've kind of hopefully kind of added more capacity to meet the demand. But yeah, we're really excited to see where that's going to go. And our vision for that is it's really a tool to empower kind of creative storytellers. It's not a tool to replace. It's a tool to empower. And a lot of the stuff you saw yesterday at Google, that's kind of the theme. And that's, you know, a visual way to do it, I guess. I love that. And for folks who are, and I've seen a bunch, I was watching a bunch of the videos this morning and just being blown away by it.

12:28Amazing. Yeah, they are amazing. How do you actually get access right now? Yeah, yeah, yeah. To access. There's kind of, it's available in two products right now. You can get it in the Gemini app or in this new product called Flow. Right now, because we're just rolling it out, we've put it in this new Google AI Ultra Plan is what it's called. And this Ultra Plan, you can just Google it. It'll come up. It's 50 % off for the first three months here, like Tulsi was saying earlier before we got on the show. And it's going to give kind of access to all the best AI stuff from Google. So we've designed this Ultra Plan to become your VIP access to everything.

13:02So it's got access to VO3. It's got the highest rate limits across everything. YouTube premiums in there, 30 terabytes of storage. We basically tried to create a plan that like, if you wanted all the best there, and that's where you can get it today. And over time, we'll start to bring it into the other tiers and make it more available. I love that. That's the only way to access Flow. Like Flow is like part of the ultra bundle. It's not like a standalone product that you would like pay per video or something like that. That's right. Yeah. Over time, that's actually one of the top requests. Yeah, yeah.

13:36Because people are already like, can I just pay for credits and just pay for use? I saw a bunch of tweets. I was like, Josh, please raise the rate limits. I'm like, this is the one thing I can't raise rate limits on. So you got to, I'm going to start. I'll tag you. Straight to Josh. We heard you. We heard you. We heard you request. We're working on it. Yeah. Josh is actually backstage. No, well, we didn't expect the demand to be as swift as it was. So there's a lot of demand. People are starting to hit some rate limits. Yeah. So we're on it. We're on it. Yeah. And to talk about this like audio plus video innovation, is that like tied to the, we're also telling the like Gemini native audio story.

14:11Yeah, you should talk about that. Are these things, is like innovation level connected or just like similar ideas or like is there any carryover between the two? I think there is similarity, although I do think the actual technical breakthroughs here are distinct. Got it. So in the context of Vio, really what we're trying to do is bring audio and video together, right? And be able to generate these in a way that actually these two things speak together, right? And are unified. In the context of native audio, so what we're also bringing to Gemini is just the ability for Gemini to speak, right? And when we say native, what that really means is instead of Gemini generating text and then that text being converted to speech, Gemini just naturally generates the content that it creates in speech.

14:52And what that allows you to do is really focus on making the conversation feel more natural. So instead of it feeling like I guess you would expect a model to speak in sort of a more monotone kind of notion, it has more emotiveness. It speaks with more tonality differences. You can change the style so you can ask it to speak with a higher volume or lower volume or more dramatically or like a pirate. right like you can you can change the the style of how the model speaks and so when you actually get the outcome it's it's much more it feels much more conversational yeah right um we actually put native audio in notebook lm um and that's been awesome to see kind of how much that's really been kind of a large part of even uh supporting itn right so being able to support multiple languages in notebook lm because native audio actually allows you to go seamlessly between languages You don't have to set explicitly now speak in English or now speak in Hindi.

15:53You can actually have that flexibility. People were talking to me about that. I'm like, I've got nothing to do with Notebook. I was like, I'm so much out. Pretty interesting. Awesome. When you get these new capabilities like Tulsi's describing and the team's inventing, it really changes how you think about building products. So something like Notebook LM and the audio overviews or even in the Gemini app now, kind of out of the box, 45, 55, 70 plus languages just work. and it just kind of opens up access and it doesn't just open up access, it reduced the latency and it makes it way more interactive.

16:26So, I mean, one of my favorite things to do is like you fire up one of these podcasts and you just hit join and just call in and you're talking to the host. You see this similar kind of thing in Gemini Live too, right? And I think these are new types of kind of ways to interact that are really built on these new capabilities. But I think we're just starting to work through what kind of product possibilities is it unlocked? And it's a lot. The things that's kind of cool with native audio that we're still experimenting with, and you can try this in like AI Studio and in the API and like test it out, is this idea of like proactive audio, right?

16:59So the idea that the model can actually tell like when we're speaking to each other versus when we're speaking to the model, right? And so it actually knows when to respond, when to interrupt versus when to hang back. And like, that's awesome, right? And then to your point about like new product possibilities, when you can't do that, you have to design your product around all of these edge cases, right? Because you're like, well, how do I make sure that I have the right limitations in place? But if the model can actually be like intentional about how it responds, you can actually now start creating like much more engaging experiences that are actually part of your day-to-day life, which I think also becomes really cool.

17:37Yeah. A hundred percent. Josh, just a Gemini app question related to all the audio stuff, the live API or the Gemini live mode in the Gemini app feels today like it's this like distinctly different Gemini app experience. It's like you go in and there's all this different visuals than the rest of the experience. How much do you think as sort of all of this, I assume some of this was just like an artifact of the technology being like somewhat distinct from the rest of the stuff, but as it all sort of merges into the main Gemini model, do you think we'll see more of those capabilities just like fused into everything?

18:09I think there was some other announcements. It's available to everyone now to try for free in the Gemini app today. Yeah, that's right. And that's kind of part of our general strategy. We want to bring a lot of this great, cool stuff just available. Fatherless, such a hardcore Gemini Live user. Yeah, yeah, yeah. Love that. Yeah. No, I know. I was using it actually yesterday, just like working through something. I think, yeah, you're right. Right now, you kind of do go into this different mode. There's the like drop animation. It all changes. It feels like you're kind of in a different spot. Which is kind of cool because it is kind of like this magical different experience.

18:40Yeah, yeah, yeah. I don't think that's that. So that was the original thought behind it. I think we're trying to figure out how do we bring some of these elements? Because what's interesting is you get people in that mode and the conversations are like five times longer on average than the text-based conversations because it is this more free-flowing interactive experience. So I think we're interested in, can we take some of those principles and bring it to the rest of the app? Because you think about Gemini, we're trying to make it the most personal, proactive and powerful assistant. and something around proactivity I think has really been missing so far in a lot of the AI products even the ones we've built in Google labs right so I think there's a big opportunity taking a lot of these principles like hey actually this AI assistant of course it'll respond to you if you talk to it but it'll also bring you things and things that are really matter to you are really timely and so I think there's there's probably seeds in the Gemini live experience that can spread in that way it is kind of interesting though because I do feel like I mean I'm curious like it feels like you're building like gemini live does have currently just a different mental model behind it which maybe isn't necessarily a bad thing yeah around this idea of like conversational yeah right like i can go on a a car ride and like have a conversation yeah and have and have that be the but i guess to your point maybe it's about how do you bring the elements of like i can actually take an action for you and what does that look like and merge these things closer together i mean one thing we're already seeing a lot of people do with it is and they even refer to it this way they go live and it's like they want to bring Gemini into whatever they're doing this is why the camera and the screen sharing is such a big announcement yesterday because when you go live with Gemini it's like all right here's my camera now you know help me out Gemini fix this bike or whatever you know what I mean and so I think that's like interesting it's kind of becoming a verb and we didn't even plan for that yeah and so we're going to follow user behavior in that way because there's something to that like oh okay it's like helping me in this moment.

20:31This proactivity piece, I think, is also, it's available on iOS and Android. I think it's also going to be super powerful on desktop. So this is my obligatory 90th thing that I get, I get picked. People want a desktop app. Gemini desktop app, they want it, they're begging for it. Yeah. Can we, can we get a gem? Can you commit right here? On camera. Like this is the approach. I didn't know this was a trap. There's a must-beak request here. Okay, so here's where we're going with the desktop app. What we also announced yesterday is Gemini is coming to Chrome. Nice. That'll be a step one. So that'll work on your desktop.

21:09When you click the sparkle icon, it'll actually give you a floating sort of Gemini you can move around your screen. Awesome. So that starts to give you, if you're a Chrome user, you're going to be able, hopefully, to love that. And we'll add more to that over time. What's the limitation of that? I haven't played around with it yet, but what's the general? or can I do everything that I can do in the Gemini app? Yeah, so where it starts right now, it's got a lot of it actually out of the box. So you can basically bring Gemini in. The cool part is we take the context of the page you're on, and it's automatically just in the context window.

21:38So that helps. Any kind of Q &A with a long site or anything just works. We also built in Gemini Live there. So you can talk to it just like you want. And then of course you can drag it around. So we didn't want to just anchor it like fixed in a certain spot on the screen, because I don't know people like you have multiple screens and then you want to move it around. I only have one extra screen. I don't have that good of a set. I think that's where it starts. Over time I think we want to add a lot more. So the stuff we were showing yesterday around agent mode, some of the project mariner capabilities, wouldn't it be cool if you could just hit that thing or have there is a keyboard shortcut too.

22:13You can just invoke it, send off a task and it just does stuff. So that's where we're going. We've heard the requests loud and clear about the dedicated desktop app, especially on Mac. So there is a team very excited about that. We'll have more to share when it's ready. Another thread from yesterday was just around all of the new Gemini models. So we had DeepThink with Gemini 2.5 Pro. We had all the excitement with Mariner, which is part of the main Gemini model. We had Gemini Diffusion. Tulsi, you announced a bunch of this stuff. I think I was blown away. If folks haven't tried the Gemini Diffusion bit yet, it's crazy.

22:52Give us the... Yeah, you should definitely sign up on the waitlist to try it if you haven't, because it's, I don't know if you've had a chance to try it. It's awesome. So I think what's actually really cool about the Gemini model story we announced yesterday is it ranges the gamut from like us trying to make our models more performance, right? From like a quality standpoint. So DeepThink is really about pushing the frontier of research from a reasoning perspective, right? and trying to say, okay, if we actually enable these models to reason more, and we allow them to actually look at multiple hypotheses, and then identify an answer, can we actually like seriously push coding performance, math performance, reasoning performance, multimodal performance?

23:35And the answer is yes. And we actually like can see that in 2.5 Pro Deep Think, like it's state of the art on the benchmarks. It's also just like super cool to use. And so for us on the Deep Think side now we're like, wow, this is a really a frontier model. It has these capabilities. We need to start thinking through, okay, what is the safety story? How do we make sure we're intentional about how individuals use this model and then roll it out? So we're in that stage. So that's kind of one extreme, which is how do we push the research on the reasoning side? Diffusion is like the other extreme, which is how do we push the research on speed and on efficiency and on new ways of engaging with the model.

24:15So what's cool about diffusion is like the approach we've taken to Gemini so far, it generates left to right. And so it generates like sequentially, if you will. Whereas diffusion doesn't actually have that limitation. It actually like refines kind of step by step through that process. And so if you saw the video we showed at IO, you can actually see it real time modifying the answer. You can barely see it because it's so fast. If you blinked, you missed it, which is the line from I.O. But I think if you saw the slowed down version, you could see it actually going through that process. And it's so fast.

24:54I think that kind of shows where we're hoping for Gemini to go. We're really hoping to, A, push on latency and efficiency and new types of ways of editing and engaging. And we're trying to push on 2.5 Pro and its reasoning capabilities and its performance. And then of course there's our mainstays 2.5 Pro and 2.5 Flash. And what was also really cool yesterday is 2.5 Flash got an update and it's awesome. Yeah, it's great. Like its performance has gotten so much better in just even the last month. And a lot of it is through like just developer feedback and trying to figure out where we're seeing developers use these models, where we want to continue to push them.

25:35It's much, much better at coding, for example. It's multimodal performance is getting better. Like it's just an awesome model. And then 2.5 Pro, we actually pre-released it two weeks ago. So we released the IO edition, yeah, two weeks ago. Two weeks ago. Just so we could get folks starting to build on it. And that's been awesome because like the web app generation is wild, what we're seeing people do. And so that's also been really cool. And GA soon too. And GA soon. Which is super important. And people are, I think a lot of the feedback is, give us a model that we can use for the next year. Yeah, yeah, yeah.