Google IO 2025: Logan Kilpatrick and Tulsee Doshi on Gemini 2.5, AI mode in Search, and glasses

Uh, all right. Let's bring in you know what time it is. There they are. Huge day. How you doing? What's going on? Congratulations on today. Uh, what's new? Would you mind introducing yourselves and then just giving a quick breakdown of what's going on today over at Google? Yeah, I can go first. I'm Tulsi.

I lead the product team for Gemini. Cool. And I'm work on some of the Gemini developer tools. What are the top three announcements today? What's the most exciting thing? What is the the one takeaway for the audience? Oo, I'll start with one takeaway.

I think the one takeaway is like Gemini, at least from my perspective, Gemini 2. 5 is just getting better. Um, and it's not just getting better in terms of like the quality of the models themselves, but we're now bringing that capability everywhere.

So AI mode is coming to everyone which is kind of our advanced version of search. Um Gemini is coming to glasses. Um which is going to be amazing. Uh Gemini is coming in even better ways to Gemini app. Um with more weight limits and just more access and better features. I was going to have to glasses.

Is this like a return of Google Glass? Is this a new Yeah. So, so part of the issue for for us today is that we got to watch like 20 minutes of IO and then we started our own show.

So, we still have to catch up, but uh but yeah, and then and then it's good you mentioned rate limiting because uh I was worried I was going to have to spend, you know, this entire conversation moderating John asking for higher rate limits from you guys. V2 need we'll make it happen. Anyways, sorry Tulsi can continue.

No, no. I think it just it's it's like an exciting time for for us because I think the models continue to get better. We released a new version of Gemini 2. 5 Flash today. We introduced um like improved versions of 2. 5 Pro with the deep think kind of deeper reasoning version of the model.

And so I think the models are getting better, but it's also amazing to now see them actually really landing in Google products in in a very real way. I don't know what what are your takeaways or top three are. Yeah, I think my my TLDDR is Google's firing on all cylinders across every vertical, across every dimension.

And like IO is this like weird manifestation of this actually happening where you just like I'm like holy like even as someone who spends all day working with Tulsi, working with our teams and seeing this happen.

like IO is this crazy reminder of like just how much breath Google has in in the in the product areas and just how many people the model innovation touches and it's um yeah I think today's a celebration of all that hard work and and making all that happen. Yeah.

How do you think about the messaging from like the continuum of like consu Google's so big that like consumers will watch a developer focused keynote.

um some of the products that are launched are very consumerf facing but then there's obviously developers that want to build on top of these tools and then there's also stock analysts that are probably watching for how things are going on benchmarks or MAUs or any new data points and so um uh what are you focused on kind of messaging to the different audiences?

Is there a concept of speaking to multiple audiences or do you just try and laser focus it on on dev specifically? No, I think we really do want to speak to multiple audiences. I think you hit the nail on the head in like there are definitely experiences that we talk about here at IO that are very consumer focused.

We also know that developers are trying to build for consumers, right?

So when you showcase amazing consumer experiences, you are also able to speak to devs because you can speak to them about how you actually were able to build that experience and what they can do to build experiences like the ones that you're trying to build, right?

and those and so I think there's an art to try to figure out what are the right demos, what are the right stories that really can land cross audience and then we try to create sections that are maybe more targeted.

So for example, I spoke earlier today, Logan spoke in the developer keynote um really targeted at developers and what developers can do and then there are obviously sections around search or glasses or Gemini app that are much more targeted at consumers. Walk straight off the dev demo stage right here. True.

Yeah, you were you said you could do two and you also said your keynote was starting at 1:30. I was like, are you worried somebody might ask an an extra question? Be in multiple places at once. It's a skill. This is amazing. Um, uh, how important do you think those demos are?

Because, you know, we saw Chat GPT breakthrough with the, uh, the this the studio Giblly moment. Everyone was memeing the Giblies.

Uh I was recently using V2 to generate video of of cars with specific liveries and decals on them driving around and the result was fantastic and but it hasn't broken through in the same way even though I feel like the model and just more importantly the accessibility because I was able to generate a video from from a Gemini chat box and I didn't have to go to a different app for that and I saw in the Apple App Store uh the the Gemini promoted post is basically like V2 is here.

Um, and the output was so good that I showed it to somebody and they they said he was John had generated an image of a Lamborghini that had full TVPN branding on it and I showed it to somebody and they were like, "When did you do that? " And I was like, and they were like, "That's right.

Was this how you made your promo video? It was really just VO behind the scenes. " No, that's our that's our next promo video. But actually, also on that note, you I think you probably haven't seen that part of the keynote yet, but we shipped VO3 today and it's awesome and it has audio. It's It's amazing.

Oh, that's very cool. Yeah, I was thinking about that because it was missing audio and I was like, this is the missing thing. I need to go find another model and then figure out how to inference that or whatever.

Um, but yeah, I mean in terms of in terms of creating these like memeable moments or something that can go viral, is that something that's even thought of or is it more just like turn it loose and hopefully Well, yeah, I think the question is like, you know, you want to distribute the capability and the question is like is that Google's job or is it the developers building on your guys' platform to do that sort of last mile delivery to users?

Yeah. Yeah.

I think the use case bit is is actually really important because I think despite like we're I feel like we're all sort of seeing everything that's happening and we're close to it, but like there's so many people who actually have no idea and it's like every day I think we need like a tracker of like the number of new people entering the AI world every day, but like I imagine it's like some reasonable percentage of all people who are using these products and like you know most products are still a blank empty chat screen somewhere.

So like showing that use case of what's possible is super important. And I feel like a lot of the demos from IO today were like pushing that frontier in a in a really meaningful way. Um but I think also to answer the question of like whose responsibility is it? Like I think Google we get to do both which is fun.

Like we get to go and build a lot of these use cases for developers and make it possible for them to, you know, bring that to their consumers, but then we also get to be like dog food and like really walk the walk and build those consumer experiences ourselves and find a way to make a a memeable moment. Yeah.

H how much of what is announced are more like showcases that could be built into other products long term?

Because I was looking at somebody was saying like Google Assistant has less fivestar ratings than Gemini, but you could imagine Gemini being like a demo for what's coming in Google Assistant, but it seems like Gemini's just gotten so big that it's like bigger than assistant now.

And so you have this weird thing where you're experimenting in this fun sandbox, but if you're really successful, then people just start using that and then you just move off of the other thing.

Is that is that like a deliberate part of the strategy or just like a natural outgrowth of like the modern AI development strategy that you put stuff out and if it goes really big you just let your winners ride. I think it's a it's a good problem to have.

I think if we're in a world where the experiments we're trying or the new kind of surfaces we're building or the new types of use cases we're building actually really resonate with users and then we're in a place where we actually want to build those and scale those out.

I think what we're trying to do, so at the end of the day, what we want to build is products that people love and really want to use, right? And I think there's three ways you can do that. You can try to build new features into your existing products. I think in a lot of cases that makes sense, right?

So, for example, like AI mode on search today. Yeah. Um, but in some cases, if you do too much of that, you bloat the existing product.

it becomes harder to use and understand versus being able to create a new surface where you actually can really rethink the UX, really rethink how users are supposed to think about the product.

And then you can look at that and say, well, does this really make sense as a standalone or does it actually uh merge with other ways that users are using different products, right? And actually fits into their existing workflows in some way.

And so I don't think we want to pigeon holes ourselves into just, hey, we need to consistently build on one product, right? We want to give ourselves the flexibility to create ways for users to build new mental models and then figure out how those things come together.

And I think this search example is actually a great one because you if you saw we announced deep research um in the Gemini app came out I think all the way back in December and it was like pioneered the deep research category and Google launched that in in the Gemini app and people love it.

It's it's a crazy product experience. Um and then like that's now a like I think the the you know the feedback was this is an incredible experience. Not everyone in the world is a Gemini app customer.

Um there's a whole lot of people in search and like now I think it's called deep search is available as part of AI mode inside of just like basic Google search which is going to reach hopefully even a many many millions of more users than are using the couple billion couple billion. Yeah. Um just a just a B.

Talk about MCP. Was there was there uh like a sigh of relief that Google's leaning into MCP? Um, was it ever a question or was it something that was like pretty easily adopted across Google?

I think it's been fascinating to see like how quickly the world has like jump I feel like with standards like people I feel like are often like very slow moving to be like do we adopt this?

There's like a whole you know many year battles with different standards competing and I feel like everyone was just like MCP came out everyone's like okay we'll use MCP. XQCD. It's that XQCD meme. There's 10 standards. We need another ST. We need just need one standard. Now there's 11 standards.

But in this case, we actually got the good ending. It's important though like I think developers are building stuff.

I think in this moment where like you know the world is still trying to figure out how to build robust agentic systems like having people not spend all the time just rewriting the same framework 50 times over like is I think is like actually a net benefit for the world.

So, and I think the the team that built anthrop uh that built MCP at Enthropic did a really really good job and it was a smart move by them to like actually make it an open standard which um they could have easily not done and I think it's it's the right thing for the ecosystem. Yeah.

Puts developers first and I think the right way. Yeah. Um what what is the overall messaging to developers? It seemed like Satcha on stage at at Microsoft's Build was very much like we host every single model. We'll we'll give you DeepSeek if you want. We're partnering with X.

Um Google in my mind has always just been like look, we invented the transformer. We have the biggest like best models that we're we're domining the paro curve. Uh are are you guys seen more as like a one-stop shop or do you think there'll be more uh model agnosticism in the future with regard to GCP?

I think GCP already has this. So GC like we do like obviously you know Tulsi and the team do a great job with Gemini models. We want the world to use Gemini, but GCP does actually have like I think it's over 175 different models that are available through the model garden. You can get anthropics models.

You can get a lot of other models. So I think the and like for customers on cloud like that makes perfect sense. They need that like that's a requirement like the world doesn't want to you know be locked into a single model.

Um but I think like we get to see these like really really interesting experiences that can only be built on into Gemini because of the like stack of of Google's model innovation. Um, so we're going to keep pushing on on Gemini models, too. That's the goal. Yeah.

And then on the uh uh developers that might be working with Google, I imagine that the conversation is a little more nuanced uh if you're actually scaling a growing app on on on Google infrastructure.

like I ran into a rate limit as a random user, but uh if you're if you're actually thinking about scaling a startup, uh unless you're, you know, thinking about building Stargate, uh you're probably able to scale pretty pretty efficiently.

But have you been seeing companies run into um any sort of like scaling or rate limits or challenges as they kind of grow on Google? Email me. My email my email's online. I'm on Twitter 14 hours a day. So, whatever you need, email me. We'll make sure it doesn't happen.

I think in general like this is like as far as like a class of problems that developers have in this moment. I think that like very real like I I hear this and see people all the time. The challenge is like compute is like a it's like literally a currency almost.

So like we have to make sure we get the right compute to the right people who are actually building stuff and like it is a it's not an easy process to do that. But I think if people need stuff please reach out to the team actually reach out to Tulsi too. Uh we'll get yeah spread the spread the love you know.

Don't just send your don't just don't just create CX tickets for Logan. Uh maybe talk talk about a couple of the the other kind of product level announcements you guys had. Jules uh an asynchronous coding agent. You want to speak to that a little bit? I'm curious to get your Yeah, we're really excited about Jules.

So um actually like at jewels. google today you can sign up for the beta.

Um but basically the idea with Jules is we really want to start um thinking through more and more use cases for how we can build essentially um assistance for an agentive assistance for developers right and so you can think about Jules as trying to take aspects of the workload that are frustrating that are timeconuming and can actually operate on them in the background right so the example I gave earlier today is node.

js JS um being able to like update an outdated codebase uh for like the latest version, right?

Um but I think it's really about how can you actually be able to assign tasks and let the model run and complete those tasks and come back and actually then have this like relationship with the developer um where these tasks can get completed can come back can iterate together and and build that kind of collaboration setup and you can run multiple at once.

you can run multiple at once and I think in classic Google fashion you can sign up and use it for free right now which I think is really cool like I think part of this is if you look at what's Google's job in the world right now like I think part of this is like we have this stewardship role of making sure you know developers people who are using search all etc etc like know what's happening with this AI technology and can actually get their hands on it yeah can try it and like know that it's available to them so anyone can sign up I think the queue time for tasks is kind of high right now because everyone is flooding jewels, but hopefully as they scale up capacity, we'll Yeah.

I mean, also and I think like what what's been actually really exciting today is to see and get the feedback on Jules because we've obviously been trying it internally and and using it, but actually getting it in the hands of real developers.

Getting that feedback, that's how we can actually a prioritize what types of activities do we need to prioritize in jewels? What use cases and features are most valuable for developers? How do we build that up? And so, it's actually like a very self-fulfilling prophecy, I think, for for how this will go hopefully. Yeah.

Have either of you had a chance to actually sit and get the project Starline demo. I remember a couple years ago. Uh this is the the 3D How would you describe it? 3D video conferencing. Yeah. The like video conferencing plus. Yeah. Yeah. Plus+.

Uh and and and there were a whole bunch of videos uh from influencers who were there who got the demo and like blown away and they were like, I filmed it with my camera and I you can't really see the effect, but you'll see my reaction.

Um, but it's it's exciting that it sounds like that's going into production in partnership with HP. But, uh, is there anything else to to say there? Have you had a reaction or uh or any any test demos?

I I'll be honest, which I've had I had that same experience that the influencers had, which is I used the the Project Starline room to take us to take a meet call, but I was meeting with someone who wasn't in one of the other rooms. So, I have no idea.

Like maybe I looked different or something like that, but they weren't they was just like they were a normal they were in their home. So, but it looks cool. Like I I I'm hopeful. I'm hopeful. Yeah, it's a very cool setup.

As somebody who works at home too much, um I'm interested so that people can sort of feel my presence hopefully in the future. Well, if you get in your car and drive, you're going to miss all the DMs from developers that are, you know, need Logan to stay home with his Starline set up. Stay locked in. Yeah.

I mean, we've been kind of in like I mean, I I remember I bought a 3D TV like a like a decade ago probably and we've been waiting for like, okay, yeah, 4K, we can't really see anything past 4K, but something like this would be very very cool if it uh Notebook LM number four in the charts, big launch, new mobile app.

uh did uh I'm sure you guys never lost faith internally, but I think people externally were kind of had been frustrated because they liked the product so much and they didn't feel like it wasn't getting kind of the investment. Uh any any color to add there? Lots of Notebook Lon progress.

I think people sort of felt this which is interesting and I've we had a bunch of conversations around why but it was like the team is still cooking like they're still shipping a bunch of stuff.

I think that I think part of it was like the expectations were just like super super high now because a lot of the stuff they shipped was awesome but they had a ton of new features that just landed and it's been awesome.

I think also this team is one if even if you look at the history of notebook LM like we actually shipped notebook LM relatively quietly and then it got a lot of pickup because it was just such an amazing product to use and so I think we've taken the strategy with notebook LM where we've sort of tried to lead with the product leading almost um as opposed to just trying to like cook up a bunch of noise and I think the notebook LM has team has just been silently making the product better and better and better and like we are going to have Tulsi cook up a bunch more noise for notebook alm this is a great way for you to more more tweets is just firing off notebook element updates.

Well, it's good. I mean, the fact that there was like organic real pull from the market and it wasn't just like, hey, we have a bunch of distribution. Let's make everybody aware of this right away. Yeah. Yeah. Yeah. What was the original notebook element?

Because I mean, there's research previews, there's developer previews, now there's alphas, betas. I feel like there's 25 different ways to say like don't hold us accountable to that we launching this product. uh what is the actual flow these days for just getting a product?

Wasn't wasn't Gmail in beta for like a decade or something like that? This is kind of like in in tech culture broadly. Uh but I mean a lot of things people are just like if I can use it I expect it to be a consumer product.

But did did did notebook LM go through uh like a particular pipeline or do you have like a a structure for these launches over there? Yeah, notebook LM came through uh Google Labs.

Google Labs is the Josh I don't know if y'all follow Josh Woodward but Josh Woodward runs Google Labs and the Gemini app now and Gemini app is obviously generally available and not in labs but labs does all the early stage bets.

So actually AI studio the project that I work on um came out of lab same with the Gemini API notebook uh jewels came out of today some things we showcased so we showcased on stage we showcased pinhole um we showcased uh stitch and whis jewels yeah right all of these are products that kind of came out of and it's break those down really quickly those other labs products uh stitch and yeah so I mean you were actually demoing stitch earlier right stitch yeah uh I didn't demoed stitch right before I got on stage But uh Stitch similar sort of vibe coding, vibe designing setup um where you can go in and ask the models to like go and build uh native mobile apps I think is what it's actually building out of the box.

And then the pinhole basically builds off the magic of V3 um and sort of multi multimodal generation.

Um, so I think really what we're trying to do with labs is or Josh is trying to do with labs, what we're trying to do with labs is use it as a way to take these technologies and start testing what new surfaces, what new UX, what new types of products can you build.

Y um, and then like ideally like notebook LM they're well loved and we actually like can hit a moment in the market where we can find new product market fit. Um, and that's really why we're trying to create this incubation engine.

And then we can take those ideas and both bring them to our products as well as actually ship them as standalone products. And principally they're it's they're also like very close to the model teams.

So it's not this like you know they're the the you know they're collaborating with Tulsi's team like PMs are all sitting in the same places. So there is this like innovation at the model level turns into innovation at the product level. And I think those teams are like some some of the first recipients of that stuff.

Yeah. Yeah. It's interesting because like I I feel like Gemini has a voice mode now that you can click over and and talk to the LLM and it's kind of like almost a separate feature path from just dictating and then having it read to you even though that's a very similar auditory experience.

You could imagine notebook LM being like a feature inside the Gemini app, but right now it's kind of going down separate tech trees and it kind of is. So we we brought audio overviews and deep research and like I think the interplay between these things are like all kind of connect.

I think there's like some like very bespoke notebook LM stuff that you'll like probably not have in the Gemini app. I think the stuff that works like there's no reason to not. It's like all it's all under Josh's umbrella. So like he gets to easily be able to bring those. Yeah.

And I think like the underlying technology like Logan said we're trying to make sure that the model teams in these products are very close together.

So, for example, Notebook LM has now native audio, which is what one of the things we talked about today, which is really like Gemini being able to natively generate audio, which makes it more I like multilingual friendly, more natural. And so, that kind of technology a we can use to make Notebook LM better.

We're also bringing that to Gemini Live. We're bringing that to um you know, other parts of the Gemini app over the course of the next, you know, months. And I think that kind of synergy means that you're going to get the best of these experiences depending on your use case in different places.

Yeah, I I it's an interesting product challenge. I've run into weird scenarios with chat GPT deep research. I'll generate some huge research report and then I'll be like listen to like read this to me, but it clearly gets lost once it's reading halfway through and starts hallucinating and stuff.

Like there's some interesting challenge to actually make all of this like if you generate 5,000 words of text actually getting it into the human's brain. It's not enough just to be like here's your research report like you actually have to make it accessible.

Deep research audio overview I think is that flow like it does work. It's like you're giving It's actually awesome because the deep research reports are like 46 pages. I'm like yeah exactly. I don't have time for that. Like there's no 46 full pages. 46 pages.

Like give me a one pager or give me like three minutes of someone like you know having a casual conversation. Give me a YouTube short generated by V3. That's what I need. What are you saying? That makes sense. Eventually you two are just going to be an audio overview with your you know your own spoke tastes. Yeah. Yeah.

Yeah. We're not letting Notebook LM disrupt us. We're coming into the show completely unprepared, but we listened to a 10-minute notebook LM on the thing we're about to talk about, so we have a little bit of context. Uh, anyways, you guys have been shipping like crazy. Uh, what can people expect from you now?

Are you going to rest on your laurels? Are you going to, you know, you know, uh, just get a little overconfident or is it, you know, business as usual? Back to the grind. a little overconfident I think will be good because I think like we've done a lot of there's there's just been so much stuff that the team has done.

So I think like people should be proud of all the work that's gone into it. Um but I think the work just starts now.

I was just I woke up this morning and I was like I'm writing the doc about how we need to go heads down for the next six months and land the next 50 things that need to happen because there's like so much stuff that's still cooking in the pipeline and it's going to be awesome to see all that land. Yeah.

I think also like we're really I really do think we're hitting our stride in so many of these areas. Um it feels like it even even some of the posts like even Sundar's posts like over the last 24 hours I was like he's feeling good going into IO. You can tell you can tell. So the energy is palpable. It's good. Yeah.

And so we want to I think we want to use that energy like I think I think the momentum coming out from today is really about like okay how do we take the excitement? How do we take the energy? how do we take the feedback and like actually really use it to continue to just build awesome things?

Um, so I think we're just at the beginning. Like this actually really is the start of of what's going to be an exciting rest of the year, I think. And we'll make sure that you two are in person for the next uh more than 24 hours heads up if you want to come in person. Yeah. Yeah. Yeah. Yeah.

We need a whole calendar for these things. Yeah. Amazing. Well, thank you both for for making time to come on. It's really great to hear and congratulations on all the progress. We'll do it again soon. We'll talk to you soon. Cheers. All right. Have fun, guys. Yeah. The interesting thing is they are cooking.

They're cooking. I mean, yeah, they they delivered in a bunch of different ways and um I mean obviously like you know what like tech Twitter is not real life in many ways and you see that with the the the Gemini uh mouse and Dows and just five star ratings and stuff.

Um, but I mean you saw it with the the video I generated like like the models are really really good performing really well and I feel like Google has been there's so much chatter on Twitter of people making fun of oh I was in Gmail and I asked it to do this and some of the products are misses for sure. Yeah.

But clearly they're aware of deficiencies and like rapidly shipping on a bunch of different dimensions and uh that you know the the energy they even are bringing to this call and the fact that they're going even just going direct um I think the really interesting thing is like Mark Andre had that quote about like uh chat you uh open AI is becoming Google, Google's becoming open AI.

You remember this?

He said this to us and and I think his his take was that like um you know open AAI is becoming a consumer tech company and Google's uh like most advanced in research and I think that's going back to the Yahoo question we were talking about is like is like if there is a like they invented the transformer right and so if there is a new architecture and it and it does come out of like an academic lab it's totally possible it's still it comes out of deep mind again right like they did invent the transformer there and Ilia just kind of picked up the paper and was like this Amazing.

We need to implement this and go really hard at this. Um and then and then everyone did. Yeah.

But uh it would be very interesting to to see um like like we didn't even get to the diffusion language model, but what other what other foundational innovations they're trying because they've been ahead of the curve on getting bigger context windows, faster inference on that paro frontier and if there's a new paradigm and they implement it quickly like that that that could be an entirely new thing.

Very interesting. Yeah. Overall, Google has been so successful for so long and so dominant that I think that it creates this effect where people almost want them to lose or want them to uh you know get taken down a notch in some way. But they're going to have to go through us because we will take a bullet for big tech.

We will we would we will we would we we won't because we won't have to because you will never challenge the supremacy. But Google is one of the greatest high technology companies in history. and it deserves to be celebrated.

And I need to just put this phrase to bed, but founder mode feels like an individual level, people are, you know, taking, you know, massive autonomy and and really have a lot of momentum. And even I'm I'm on jewels. google. com right now. Yeah. And they're rate limited.

Uh but it looks and feels like the product that a startup would create. It's good. It looks like they just raised like 300 on a billion. They probably did from internally. Internally internally. Um well anyways we should get out of here. I did get some extra context on the Coinbase Oh yeah. issue.

Uh a friend of the show said re Coinbase hack. He was sending it live while we were in that. Uh he said the information that isn't the customer data is sold at a granular level i. e. info on the individual. So, it'll go on a dark web market and be sold individually based on the Bitcoin balance of the individual.

Think of it like a standard three tier. Regular, silver, and gold would be a balance over 50K. Uh, and these accounts are already vulnerable to like SIM swapping and things like that. So, uh, anyways, appreciate the intel. Uh, the Lone Ranger shot that over to us. Very helpful.

But uh last thing, congratulations to Tyler Cowan, a friend of the show who has made the Time magazine list of the 100 most influential people in philanthropy. So congratulations Tyler and we will see you tomorrow day. We'll see you on Wednesday. Fun show. Bye. [Music]