Hello and welcome to the Vergecast, the flagship podcast of David and Nilay both being mysteriously unavailable at the same time, so we've got to fill in. I'm Jake Kastrenakes, executive editor of The Verge.
I'm Hayden Field, senior AI reporter at The Verge. And we've got a lot to talk about today. Google I/O just wrapped up 15 minutes ago. It was as long as ever, a nice, relaxing two-hour presentation. And I think they talked about AI for, I don't know, about an hour and 45 of it.
Yep, and the word agents. It was non-stop, like every 10 seconds. Agents, agents is the sizzle real word of the year. Yes, yes, yes, and we have a lot of that to talk about. David unfortunately under the weather. Nilay, if you're following along on the site, he was live vlogging. He is out in Mountain View. So expect more from him. He's out there for some fun stuff that we'll have for you, I believe, next week, so stay tuned. But for now, we're here to talk about Google I/O and all the big announcements. If you have thoughts, questions, hit us up in the comments and we'll be checking them out. So there was a lot that happened. It is hard to even know where to start. Tons of Gemini announcements, some huge announcements for search. There was a big aside on shopping, but I do think agents is the buzzword of the year. Everything kind of laddered up to agents, including, you know, what the shopping stuff, right? Ultimately, that was agents. The search stuff, ultimately, that's agents. Everything seems to be about agents. I'm curious, Hayden, you spoke to folks over there, and you came away with a bunch of thoughts about, you know, what they are thinking about what 2026 looks like in terms of what the purpose of AI is. What do you think they are trying to accomplish here?
Yeah, that's a great question. I feel like I've been covering agents for years, like since, you know, late 2022 when generative AI was, you know, officially a thing basically with ChatGPT's release. And I feel like every single year that I've covered them, there's been a different like milestone in each year. So, it's like the first year was the year of ideation in 2023. And then the next year was the year of deployment. Then we had the year of experimentation. And these are all terms that are coming from like either both users and the leaders. So, it's like this is kind of the vibes that everyone's getting on both sides. And this year I'm really feeling like from all the Google announcements and from like interviewing the CTO of DeepMind and stuff, it's going to be the year of usefulness, like actual usefulness, we hope. We hope they're not out yet. Yeah, they're not out yet. So, we really don't know. But I will say that, you know, we've had promises of these things actually being useful and being able to do what they say they're going to do for a really long time. And I'm always really skeptical and it usually happens besides like coding, of course. That's where they really shined. But I do think that if anyone's going to make them useful, it's Google because of all the integrations they can seamlessly have with like Docs, Sheets, Gmail, Calendar, everything else. And it looks like they're kind of learning from their mistakes this year because a lot of their announcements they made about agents, they were, you know, not to say like copying, but kind of copying and like learning from OpenAI, ChatGPT Pulse, Codex, Claude Cowork. Like they were really taking a lot of the best features of all these competitors and just like kind of integrating them into their own agent plans. And I think that's going to work pretty well for them because they're Google and they can easily seamlessly integrate all this stuff into their own tools that we all use every day, like Photos, Gmail, etc. And their new two agents that they're talking about the most, the Daily Brief and Gemini Spark, I think are going to be pretty... I think they're finally listening to like what people actually want with AI agents, which has been a problem in the past. Like, it's like, do you really need it to like scour 50 websites to find the perfect pair of shoes, or do you just want it to like give you a daily briefing on what your day looks like and like how long it's going to take you to get from point A to point B, and then maybe help plan an event that you're planning, like a wedding or a block party. We saw examples like that in the stream. So, and in the report or briefing called the day before. So, that's kind of what I saw. I think they're at least hoping to make this the actual year of usefulness for AI agents, but I can't confirm that until we actually test them.
I mean, this is what's so complicated, right? They are starting to have these abilities, but I actually think that finding ways to put them to use, figuring out things that I do want to offload to AI, and finding things that I trust AI to handle correctly, like, that's actually a really big gap that, you know, things like Daily Brief are like baby's first AI thing. Like, you can maybe trust it. It's really, really low stakes. If it tells you the wrong thing, like, you're probably going to know. It's not necessarily giving it full access to your computer like you've done with OpenClaw. On the other hand, they're also launching their own version of that, which is Gemini Spark, which, and you don't have to go out and buy a Mac mini like with OpenClaw. It'll run in the cloud, but they're also going to give it access to your Mac if you have the Mac app. And so, it'll be able to do those OpenClaw things and run wild with your stuff. Maybe not as well. They said that you have to like specifically choose which files it has access to each time, but it's really interesting. Everything that Google is trying to do is to make this more useful, but they're making it more useful with agents, but they are trying to find the specific spots, like with Daily Brief and like with search, where it will just kind of automatically do these AI things for you, so you don't have to invent them.
Right. And there also something I thought was interesting about Spark is that it, you know, like one of the memes that came out of like all the OpenClaw craze was like, you know, leaving your laptop open 24/7 and like letting agents work for you 24/7 and ever a lot of people in tech were like, 'Oh, I can't even sleep. I can't even move away because I'm getting so much, you know, done.' Even though that was kind of like a productivity, we don't know if it was like, you know, productivity for productivity's sake or if, you know, people were actually accomplishing that much. It kind of depended on who you were. But something interesting about Spark is that it's cloud-based, it can run 24/7 without keeping your laptop open and it can sync across like the web, Android, and iOS. So, I think it's helpful that Google has so many resources here, so they can just like eat the cost of this and just be like, you know, 'Yeah, go ahead and go crazy. Go wild. Run these 24/7. We'll handle it.' And, you know, obviously, you know, they have subscription plans and everything, but they're actually cutting the cost of two of their... or they're introducing a new lower-cost subscription plan for Ultra and then they're cutting the cost of their top-tier Ultra subscription. So, it's like I think they're really trying to compete with everyone here by being like, 'Okay, we're Google. We have more resources. We can kind of, you know, temporarily eat some of these costs.' And the biggest complaint a lot of people have had lately about agent stuff is just the speed and the cost. And Google's kind of trying to undercut its competitors in both areas right now.
This is true. And I maybe I'm getting ahead of myself here, but they also had Gemini 3.5 Flash, a new model, and they mostly emphasized that it is faster, it is a little bit more efficient, it's lower cost in certain applications. And I think they're seeing this is the area where it impacts them number one, but also the developers, right? If you're paying for this stuff, you know, you've got a budget. You got a budget. You can't be burning through... they said there's like a... I don't even know. What was it? Over a quadrillion tokens were being, you know, processed. Like it's...
And over... and a ton of companies were using more than a trillion per... Was it a month? A trillion daily tokens. As a lot of companies...
Honestly. It's... And they said that like, I don't know, a lot of the startups and like, you know, smaller company leaders I've been talking to lately, their biggest complaint has been having to do all these crazy calculations on their side, trying to figure out the recipe for which models they should use. Um, you know, triaging some evidence or certain aspects of like plaintiff lawyers' jobs. I don't know enough about plaintiff lawyers' jobs, but the more complicated aspects. And then they're using the cheap models for the simpler tasks and like refining that ratio all the time. And so Sundar was saying, you know, Gemini 3.5 Flash specifically, it's a big cost saver. It can be half or a third the cost of some of its comparable competitors. And that if you shifted like 80% of your workload as a company to Gemini 3.5 Flash, then you could... and you were using a trillion daily tokens, you could save more than $1 billion annually. So that's obviously kind of a cherry-picked example, but still shows, you know, they're listening to people being like, 'Hey, we're spending a ton of our time just figuring out how possibly to save some costs.' It also shows they know where the interest and the money is, which is in coding right now. This is the biggest use case. It is the biggest way that they are going to get corporate customers. And it is not clear to me that Google's solution has been anybody's favorite. Right, it's Claude, it's Cursor. I've seen people in the chat saying that Antigravity, their harness for agents, is not very good. So this is a problem, right? They're trying to focus on the stuff with the new Flash model. They're trying to speed things up, they're trying to make things more efficient, but coding has not been the area where Google has excelled so far, right? They've been focused on, I think, these much more tangible applications because they're Google. They're focused on can they make a model that works really, really, really fast when you search for something. And I generally think that they're quite good at that compared to, I think, the alternatives. But as for coding stuff, I, you know, I tried... I don't even know. I think it was AI Studio. I shortly after it originally came out and got absolutely nowhere. Whereas, you know, I paid Anthropic $20 for Claude Code access and I cannot stop making apps with it. And it's like... I don't need these apps. Like, I need to stop. It's not a good use of my time. It's not a good use of tokens. But it's so easy and so capable. And I think one of the questions here is can they get there? Can they, right? They're improving agents, they're improving their harnesses. Can they make it easier for anyone to do this stuff and especially for developers to do this stuff so that they will switch to Google's platform.
Yeah, one of the things they announced was like they keep calling it like Antigravity 2.0. So like expanding Antigravity so that it can just be a like full-on agentic development platform beyond just coding. So they're introducing like a new standalone desktop app which they want to call like a central hub for agent interaction. They're just like really going all in on the marketing here. And, you know, just trying to kind of court developers and get them to switch. And so it's interesting like I think the pricing will help there, but, you know, like people said like the harness hasn't been that good. So we'll see. And also I remember like... I don't know. And when Gemini 3 came out, everyone had temporarily, of course, as they always do for like a week, was like Google won the race. Like it's over. And then like a week later or a couple weeks later, you know, Claude kind of stole the thunder with coding prowess. And then, you know, it held on to that for months in terms of the hype. And so we'll see if the next few models that come out in the 3.5 suite kind of take some of that hype away or not.
So we should talk about those. We should talk about the models. We should talk about some of the Gemini-specific updates. So I think the first thing from a consumer perspective is there's like a new design in the Gemini app. It's glowy. It's dark mode. Like it's very nice. It's very lovely. I don't think this will improve the AI experience dramatically. They're making a big deal out of it. They're calling the design language neural expressive, which I do not think sounds cool.
Yeah. It's like a little... It kind of sounds like something you'd have to like get medicated for. I don't know. [laughter] They... Yeah, they should have worked on that.
But, you know, behind that is I think a zillion and a half different Gemini updates and new models. So we already talked about 3.5 Flash. 3.5 Pro is coming, what, this summer or something like that. They're throwing... They now have different Gemini live conversation modes that they're just shoving into every single product. They're making it more accessible within the Gemini app. They're allowing you to have a live conversation with your Gmail account, which is one of the like... I don't... I get it. I get that you have the technology, you're like, why not? I don't foresee using this.
Here's the thing. This Gmail search is so horrible that maybe I would use it. Like honestly, like it's...
Yeah, yeah. It's the type of thing I would never use if the Gmail search wasn't so bad. But like I can literally be quote searching the title of an email and it won't come up. And yet like I'll go find it myself and it's there.
This is when you've hit rock bottom at an airport looking for like a reservation number. You're like, I guess I will talk to it. I believe you actually have to pay. I... [snorts] I might have it wrong. I'm not... I forget which one it is. They're adding live talking to like Google Docs as well. And for some of them you have to pay for like a hundred bucks a month to get that feature. And it's like, I don't know if that's the one you needed to pay for. I like...
Yeah. There's not a lot of them that are...
And I'm like... I was a little surprised by like what's free and what's not. It is... I think for Google in particular, it's like Anthropic, I'm like, yeah, they need money. Like I understand. With Google, it's like, oh, if you guys have to charge a hundred dollars for these features I don't want, like you must be losing a lot.
Yeah. But then it's like it's funny because some of the more like complex ones are just free, like information agents. At part of their search overhaul is offering like information agents to basically just like, you know, which are supposed to roll out I think to everyone this summer and they're supposed to do research on your behalf or like alert you kind of like a personal assistant would anytime something happens. So like if you're like watching a certain part of the stock market or you're like trying to track the weather or even like flights maybe. They didn't mention flights, but it seems like from what they talked about, like maybe even like flight price dropping. Like, you know, things like that it'll monitor for you and just email you when it happens or, you know, it pings you somehow when it happens. And so that's free. And then also, yeah, the custom experiences and the... They were really into the haptic stuff in the app. So...
Cool. Yeah, they're into it. I don't know... All going... matters. If maybe if the app is cool enough, the vibe will shift on people's reaction to non-stop AI. Yeah. So the other big model they announced, and it... the other big model they announced is Gemini Omni, which they are calling a model that... I believe I'm going to misquote this, but they're saying it's a model that can accept any input and generate any output. And this sounds like a very, very big deal. And yet, despite how big of a deal it is, they are doing it... they're giving very little explanation to what this is or why it is or what comes next for it. But right now, it's basically just a video generation model. And you might think to yourself, 'Hey, don't they already have one called VEO?' And yes, Google has now created a second video generation model, and it is not clear why. But this new one can accept any kind of inputs. So, you can throw a PDF and some photos and a song at it, and it'll use all of that to generate a video for you. Whereas VEO is just text to video. And I guess their long-term perspective is that they want this model to be, I guess, much more multimodal to a degree that other models are not.
Yeah, they want to like combine Gemini's text intelligence with its video and image models. Basically, they're trying to combine everything they already have into one. And also, they just really want to go all in on world models, which is kind of, if you remember, like when OpenAI canceled Sora, they didn't say they were going to never use it again or like just scrap the whole thing. They said they were going to instead use the research behind the model behind it to develop world models or do world model research. So, this is kind of the new big thing that has been happening for a little while, and Google's leaning all into it, too, with world models, because they think... and the CTO of DeepMind said this, too, that this is going to be the next big step to AGI. So, Google DeepMind is one of the only companies I've talked to recently that isn't kind of trying to get away from the AGI term. They're like all in. They're like, 'Yeah, we got to get to AGI and this is how: world models.' So, yeah, Omni is that for them. They, you know, gave a lot of examples like giving it a simple prompt like give me a, you know, claymation explainer of protein folding and then getting a video back and, you know, yeah, just covering all the inputs you provide like text, image, video, voice, whatever. And the good news is it includes SynthID, so hopefully the videos you output will... you'll be able to tell if they're AI-generated, but as we know, people can get around those.
It's really interesting because they introduced Omni with this sort of big hype up about world models. And the way I've seen world models used so far is these sort of interactive video generation experiences where you kind of just like... it's like you move the arrow keys and the world generates around you as though you're playing a video game and then the world immediately forgets what is around you because they're not working yet. And it's very cool. And they set this up as though, 'Okay, this is world models of the future, so we've created Omni.' Is Omni a world model? That's not clear to me. They did not specify that.
They told me it was and they said it's critical for the AGI journey. The AGI journey, and from two points of view. One is like understanding the world and then reasoning about it as a big part of intelligence. And then also the fact that like our intelligence is built in this 3D world. So, they think of it as like a way to make their models much smarter because they're learning how to interact like physically with the world. And eventually, like Google hasn't said this, but a lot of times when a company is working on world models, sometimes their like unsaid eventual goal is to basically give their AI model a body in the world. Like, you know, it's a way to, you know, create smart robotics systems. So, we'll see. They forgot to mention that.
Actually, okay, actually Sean Hollister on our team got an exclusive demo. He went by Mountain View a day or so ago and got a look at... so Google Beam, which is the, I think, current name for Starline, which is their like fancy like 3D video chat service, which... it cost like 25 grand. There are probably a single-digit number of them in existence. They're really cool demos and I don't know why anyone would buy them. But they showed him a demo where they're making a like 3D fake woman who is, you know, powered by AI. I don't know why. I don't know why, but this is... There's like... it's also like... It's happening. With Omni, you can apparently like create a custom avatar of yourself and drop yourself into like the videos you create, but again, it kind of reminds me of Sora and like why, you know, but... Actually, Travis is telling me we can throw to the video for this.
Yeah, let's play a clip. First journalist to test out this preview demo with me. Pretty groundbreaking stuff. This is Google Beam, the $25,000 teleconferencing system. It makes it feel like the other person is in the same room. Google wanted to create an AI agent that isn't just text on your screen or a voice in your ear. It's something you can walk up and talk to, too. What am I holding up right now?
Ah, I see now. In your left hand, you have what looks like a transparent phone case with a magnetic ring built in, probably for MagSafe. And in your right, a small circular accessory. Is that a magnetic ring stand or a grip that attaches to it?
Yeah, that's right. I'll be honest, it's all over the place right now. It does Google things fairly well, pulling up maps, speaking in multiple languages. Because it can see its surroundings, it can answer questions about a book, even recognize a sign language signer, too. Here's another hand sign. Tell me about this one.
Oh, I know that one. You're making the sign for 'I love you.' Yeah, see, this is just... I don't entirely understand what the end goal is here. Like this... this is... they didn't mention this on stage, and I feel like this is where this starts to get pretty creepy.
Yeah, and the fact that like it takes so long to even respond, you know? I mean... Yeah, there... it's not working that well. I mean, it's impressive for what it is, I guess, but I mean it's kind of creepy just thinking about what people are going to do. Like people already have a lot of really, you know, emotional connections to models, and I feel like this is... and then further that. You know, fortunately, you're going to need $25,000 and maybe need to break into their prototype labs in order to get this experience, but it... yeah, I think to your point, like this is on their mind. Once you create this technology, it will get there. But I do think, just to bring it back to the world models, it sounds like this is maybe the next big... we're seeing a very small sliver of this right now. But from what you're hearing in the world, right, the world models are the next phase. Yeah, some of the like biggest AI pioneers launched World Labs. When was that? I think 2024? No, 2025. Like in November last year, and they left where they were, which was a ton of top AI labs, to start and go all in on just world models, and the company's called World Labs, and it was kind of like you were describing, like right now it's basically like using arrows to more... [clears throat] like navigate around an AI-generated world, but, you know, all the like investors I've been talking to, all the AI leaders I've been talking to in the past year have been all in on World Labs saying this is the way to make models smarter with spatial awareness and then eventually robotics, which is what most AI companies that have models eventually want to do in some way, shape or form. It is going to be really interesting. Right now, it is just videos from Google. But I will say like one of the, you know, I think a very fun thing for Google. It's there... they're saying all this stuff is supposed to be available right away. So, we will find out very quickly exactly what it is capable of. So, I look forward to being able to test this stuff out. All right, we got to throw to a quick break, but when we come back, we're going to be talking about search and a lot of the fun stuff they announced that you're able to just immediately try out and use.
Support for this show comes from Superhuman. AI tools are everywhere right now, but a lot of them end up creating more work instead of less. You bounce between tabs, rewrite the same thing five different ways, and spend more time prompting the AI than actually getting things done. Superhuman wants to change that. Superhuman is an AI productivity suite designed to work where you already work. It brings together the intelligence of Grammarly, Coda, and Superhuman Mail into one connected system that helps you write more clearly, move faster, and stay focused on what actually matters. Think of it like an AI-powered team working alongside you throughout the day. Instead of opening a separate chatbot window every time you need help, Superhuman is built directly into your workflow, whether responding to emails, organizing ideas, collaborating on documents, or managing projects. It's AI that works where you work, however you work. You can unleash your superhuman potential with AI that meets you where you work. Learn more at superhuman.com. That's superhuman.com.
Support for this show comes from Shopify. Starting something new isn't just hard, it can be really scary, too. So much work goes into this thing that you're not entirely sure will even work. But here's a better thought. What if it did all work? What if your instincts were actually right all along? Shopify wants to help you get there. They're the commerce platform behind millions of businesses worldwide and nearly 10% of all e-commerce in the US. From established brands like Allbirds and Heinz to companies just getting started. Their design tools make it simple to create the exact online presence you're envisioning with hundreds of ready-to-use templates available. And with built-in marketing tools, you can launch full email and social campaigns in just a few clicks. So you can connect with customers wherever they are. It's time to turn those what-ifs into... with Shopify today. You can sign up for your $1 per month trial today at shopify.com/vergecast. You can go to shopify.com/vergecast. That's shopify.com/vergecast.
Support for the show comes from LinkedIn. If you're a small business owner, you know that every hire counts. The time and resources are limited. Finding, connecting with, and screening the right candidates takes up valuable time you could be giving to your customers. That's where LinkedIn Hiring Pro comes in. It's built to be your hiring partner, helping you find the right candidates faster. That way you can hire with confidence without turning it into another full-time job. Hiring Pro streamlines the entire process from drafting your job to shortlisting candidates and conducting AI-powered interviews for initial screenings. Its updated conversational interface lets you describe what you need in plain language. Nearly 60% of hires find a candidate to interview within a week. With Hiring Pro, you spend less time searching and more time connecting with the right talent. And instead of getting buried in resumes, you get a focused short list that actually moves your hiring forward. Join the 2.7 million small businesses using LinkedIn to hire. Get started by posting your job for free at linkedin.com/track. Terms and conditions apply.
All right, we're back. It's time to talk search. Time to talk vibe coding. And yeah, there's a lot... there's a lot. So it's very interesting. Google search is sort of a big thing for them and search ended up getting like smushed in the middle of this presentation because as Sundar Pichai told us, apparently they decided 10 years ago that they're an AI-first company. I missed that announcement. I don't know. Feels like more of a slightly more recent thing than that.
A big deal out of it. He was like, 'It was 10 years ago that we pivoted our entire company to be AI-first and we were the first to do it or something.' And I was like, 'Okay, yeah, but you're still a search company, too, though.' Yeah, yeah, yeah. [laughter]
And so all of the stuff that they are building, though, does end up in search. And they kind of came out today to try to make a very big splash with search saying, you know, 'We're kind of completely overhauling it.' And that includes just like infusing AI into every corner of search. And it's very interesting because I feel like a lot of what they're announcing may already exist today. Very recently I've been doing searches, I would say, particularly on my phone and I get... I start reading stuff and I'm like, there are no links here. Like I am just reading an AI result now. And it keeps forcing me to have a conversation. You know, and I'm like, I have to exit out. I'm like, no, I'm like trying to scroll. But yeah, I mean there's a lot... They said it was also the biggest overhaul to their search bar or like search box in 25 years, and it was a little like they basically said they were going to make it bigger and allow for longer queries and then automatically take you into AI mode, which is kind of what I feel like is already happening when we use it, but I guess it's going to be for everyone instead of just some. Yeah, it's unclear how much they're just sort of...
Saying like, no, this is official now. We want you to think of us as being in this different format. And I think that is probably part of what's happening here. It's really interesting. Google, for many years now, has been going out there saying like, you know, younger users and people in certain countries, they actually like they don't use properly formatted search queries. They don't just say like Indian food. They say, 'All right, I want a place that's open during these hours, that's great for my parents, and that serves, you know, Indian cuisine, and that's near.' And it's like I'm like, who does that? Who on earth makes a search query like that? Like that's nuts. And I will acknowledge like the more that they have infused AI into stuff, and the more that I have gotten comfortable using ChatGPT and Gemini and all these other things, the more that like I feel like I'm losing the ability to make a succinct search query, and I'm like I'm going to search for something crazy and specific because it will more often than not get me a slightly better result. And so this is what they're saying. They're like, 'Now the search bar, you can somehow you can add Chrome tabs to it. You can throw in files and photos.' And yeah, put it in a giant sentence. And you can like also get back. This was what I thought was really different. You can get back custom widgets and visualizations of your search. So like if you're searching for how a black hole works, obviously reading about it, it's a little hard to parse, but it can make you a custom visualization or custom widget or like, you know, so that's really interesting. Like it will automatically decide when I think when you need something like that and then deliver it to you and it'll be custom made for your specific query. It won't be like every time anyone searches for a black hole, that's what's going to happen. This is like both fascinating and terrifying.
