[AIEWF Preview] Gemini in 2025 and Realtime Voice AI

0:04Hey, I'm Sam Trunington. Welcome to another episode of the TwiMLI podcast. And I'm Swix. This is a special episode of the Latent Space Pod with Twymo at Google I.O. Welcome. Thanks for being here. Thanks for hanging out with us. I'm excited. Logan, you were our first guest. You came back remotely a few months ago. And now you're back. You're sort of a lot of the face of the AI studio, basically, that a lot of people are using. I'm using it. And I think it's a really welcome change for people being more accessible with the rest of the Google. suite. And Shrestha, you've been, I actually don't super know your role.

0:41I just generally have you pegged as PM of the API team with a particular focus on live. Shrestha runs the show behind the scenes. What is public face running the show behind the scenes? Model launches, the live API, generally all the stuff that's happening in the API is Shrestha's hard work. Thank you for that, Logan. But I think everyone knows who really runs the show. There's public evidence there. But yeah, I Yeah, I work with Logan and a few other excellent PMs, but I lead the API side of the house. There's a lot of announcements. I think a lot of people have done their recaps. What are you guys' personal highlights over IO?

1:20I'll break the rule and I'll give two that are not the sort of big, big flashy ones. I think the two that I think developers are going to be super excited about. One, thinking budgets coming to 2.5 Pro. And you'll be able to disable thinking as well. So if you just want 2.5 Pro as like a raw non-reasoning model, we'll have that hopefully in early June. And then thought summaries. So we've had this debate internally about like, do we need to show full thoughts? Do developers want full thoughts? I think developers say they want full thoughts. We have thought summaries right now as a sort of step in that direction.

1:52It'll be really interesting to find out and get the feedback around like, what are things that work with thought summaries? What are the things that don't work with thought summaries? I was reading some threads last night about like thought summaries are now live in cursor as well. And people were sort of reacting to, you know, having summaries versus not full thoughts. So it'll be interesting to see, but I'm excited for both of those things. Thought summaries are live now. Thinking budget for 2.5 Pro will land with the GA model in a couple of weeks. Yeah. And I should say we already do have thinking budgets in 2.5 Flash.

2:23I do think, you know, with all of the features that we are releasing on top of our thinking models, summaries, budgets. I think this is our way of, you know, you have the models, but then we want to give developers as much control as they can on top of models. But coming back to your question about my favorite feature, it's really hard to pick because like all of these features we've been trying to push out for weeks. But I think native audio output is a personal highlight. Right. I actually, Quinn and I have been playing with it together for a bit as well. I think especially with all the obviously the voices sound great.

3:04The fact that it can switch in and out of languages. So Matt Veloso, our boss, actually has a has a demo on Twitter where it actually speaks Klingon, even though that's not an officially supported language. but you know I speak Bengali just being able to for it to switch into and out of Bengali and English that's been special and then if I get to pick another one it's I'd say we released a new tool called URL Context and the idea is that you can use it by yourself or pair it with search to retrieve more in-depth information from web pages in a way that's respectful of our publisher ecosystem, of course.

3:43And I think that this will unlock new use cases, like if people want to build their own version of a research agent, which is something developers ask us for a lot. Yeah. Worth mentioning that just prior to IO, there was a ton of new, interesting new capability, including the update to Gemini 2.5 Pro, as well as the implicit context caching, which I know a lot of folks are waiting for. We made implicit caching happen. I think there was lots of feedback that people are like, explicit caching is nice. Like there's definitely use cases where it makes sense, but people want implicit caching, so I'm happy passing the cost saving on to developers.

4:16You don't have to do anything. It just works right now and you're saving money. It's a great outcome. I don't want to manage that myself. Yeah, there are people like, I think if you, there's so many use cases where you're just doing chat on the same stuff over and over again. And for those use cases, you want to be able to explicitly cache the thing and make sure and guarantee your cache it so that you save money. So I'm happy we have that. Is there any behind the scenes of like what makes caching hard or anything that people don't appreciate about caching as a general concept? So I think this is a very important pricing paradigm that people need to really get behind.

4:50Yeah, that's a good question. I think there's a trade-off between like all of the dimensions of caching, which is around like the sort of latency, because in some cases you're getting latency gains. In other cases, it's like how much, you know, what's the cost for Google? How much stuff do you want to cash all together? So we could have an entire episode and get a bunch of the caching people. It's like a good example of like an infrastructure problem to be solved. And a bunch of the folks who we work with love working on this problem. So we should do a deep dive episode. Yeah. I want to shout out that you've been doing more video stuff.

5:20You have your own podcast as part of your Gemini work. You've also been doing a video with people on the team. It's been fun. We had the long context episode. Exactly. You did the long context one. People loved it. Your reception was very positive about the Long Contest one. So thank you. That was the first time that we did like a more deep technical discussion with folks on the team. And Nikolay is awesome. And we actually just did one with, we did one with Shrestha about the live API, which I'm excited about. We did one with folks on the team about the multimodal capabilities in Gemini. We're going to do a pre-training one, hopefully, which will be really cool.

5:54We've got a bunch of people who are excited to talk about that. So there's a bunch of them in the works and it's fun to make them happen and have those conversations. Yeah. My underrated pick is Gemini Diffusion. Yes. Yeah, yeah, yeah. It's not underrated. I was underrated for all the love it's getting, yeah. So like apart from speed, I wonder like what the potential results of a Diffusion language model could be. Generative UI. Generative UI. This is the way the generative UIs happen is through this experience. The UI bit, just like being able to like say, I want, you know, build the UI on the fly using code based on what a user does.

6:29So like you have no pre-compiled, notion of what your website is. And as a user goes through, as they click buttons, thousand tokens generate, and it just makes that UI for you. Interesting. I think that's going to be possible. I mean, I think there's a lot of work to productionize, make Gemini Diffusion actually a high quality model that meets the bar for us to bring to the world more generally. But I do think that's going to be the killer use case will be this generative UI experience that doesn't exist today because the models just take too long to generate tokens. Yeah. For me, it was really the role that audio and video are taking throughout a bunch of independent product releases from the generative models to the live API to the on-the-fly transcription and translation.

7:15Yeah. It's, I think, kind of foreshadowing the role that that's going to play in a lot of developer applications. Yeah, transcription actually, even before we released native audio, now, of course, you get text and audio interleaved in the output. But transcription used to be one of the biggest use cases we had on the live API. What are you seeing as the challenges for folks getting started with live? Yeah, that's a great question. I think, firstly, awareness, right? Like people knowing that we have a live API. You can do this. That's why we're doing this, talking to you folks. I think some of the areas where, so we were actually the first to market with also video input.

7:59But one of the areas where we've been getting a lot of feedback is in session length. Anybody who's been trying to put this in production, like when we started, you could do like 15 to 20 minutes of audio, I'm sorry, and about five minutes of video. And so we've been putting in a lot of knobs for developers, and we can talk about that more if you guys want, for people to have a sliding window or decide what resolution they want to send video in, but, you know, to basically increase the session length. And then tool calls. That was another area where we used to get a lot of feedback. Again, we were very proud because we introduced tool chaining first.

8:37So you could chain search and code execution, do all kinds of analysis. but then we've had to do a lot of work in improving function calling, improving the performance of search. And anyway, we continue to push on that. I've got a quick one on this too, which is I think the level of commitment you need to make to the model provider in the world of the live API. Like I do think for developers as a higher bar, if you look at like what is chat completions or like what is for us generate content provide from just like a text modality perspective. It's like, it's a pretty lightweight thing. There's a lot of model providers that have that option.

9:10Like I could switch to a different provider if I end up not liking some model provider, which I think is good for the ecosystem. I think if you look at a lot of the live API infrastructure right now, like you really do need to commit that you're like gonna, you know, there's, it's not easily interoperable between different model providers. Like everyone's infrastructure is all bespoke and different. And so it's a different level of commitment that you need to have to really bet your company or your business or your product on the live API, which I do think is a challenge for developers to sort of make that level of commitment in this fast-moving AI world.

9:42But I think hopefully there'll be some level of similarity and you'll get some model-agnostic infrastructure to help make developers feel a little bit easier about being able to move between models potentially. I could go on and on, but if you have, say, more complex workflows, then one of the things is being able to change the system instructions at every step of your workflow. So, yeah, so onboarding some of the more complex use cases with the Live API has been a work in progress as we've released more features. So what kind of complex workflows are we talking about? You know, we have people who are building, say, gaming agents, but which have multi-states, for example, in them.

10:23We have a lot, I mean, this was a famous demo at Next, but we have folks who want to, you know, customer support agents, of course, you know, the sessions can last for hours, right? Then there's a lot of use cases around people showing a certain screen. This is the coolest use case, honestly. Yeah, and I was referring to like the famous demo at Next where Shopify showed how to set up a DNS using Cloudflare, right? So in certain cases, especially the longer your workflow runs, you might have to go from one state to another state and you might want to change the SI. Or if you hand it from one agent to another agent, you might have to change the system instruction.

11:03When you're thinking about building voice-based applications, is speech-to-text and then processing with a standard LLM, would you say that's like a precursor to the live era or are these two distinct paths that are still viable and that you still see being viable going forward? That's a tough question. Right now, we have both out. I do think perhaps eventually for most use cases, as these audio to audio architecture models get better, a lot of use cases will probably transition to that. But, you know, when we talk to our developers, they still very much like those componentized components. So that's why we also put out two new text to speech models at I.O.

11:56So not available through the live API yet, but really high performing, controllable, promptable text to speech models. I have an angle of an answer to this question, which is I talked to Korai this morning, who's our boss's boss, the CTO at DeepMind. And Korai had a really interesting take, which is just around like what makes one of the main things that makes what we're doing at Google with Gemini different than what a lot of the other labs are doing is like, we're here to make one model. And like that model is Gemini. And like, I think, I think you, you do need to, to stress this point, like to make the capabilities work in some cases, like you do need to have these forks that like go off and make that capability and harden it and then find a way to bring it back into the mainline model.

12:40But like, we want to make one model and it's the Gemini model and like not have the sort of splintering of all these different capabilities. And we've done a good job of, I think thinking the reasoning stuff was like the best example of this. We had those, they were separate from the mainline Gemini models so that those teams, the research teams could go and hill climb and make progress and not need to be constrained about like, how do we do this without having there be collateral damage on other capabilities like multimodal or something like that. But the teams went and did that and then they find a way to sort of bring the capabilities together.

13:09And oftentimes what you see is there's tension in bringing them together, but it's the really exciting thing is what happens when you bring the capabilities together. And 2.5 Pro with reasoning is a great example of this where multimodal with video understanding ended up having this huge, it's having this beautiful moment. The model is like soda out of the box because of all the reasoning capabilities that were baked in. It wasn't because they did a bunch of stuff to make video understanding really good. It was just like an artifact of bringing and merging those capabilities together. So I think that as a North Star for Gemini models makes a ton of sense.

13:44I agree with you. and that's what I said, right? Like I think eventually a lot of use cases will end up on Gemini, will end up on natural voice. But I think in order to foster development, like we have these offshoots from time to right. We have our imagined models for image generation, even though now another IO, well, slightly pre-IO announcement, you can do interleave text and image within Gemini also, right? And it unlocks. Yeah, but those are different models, right? Those are different models. One is autoregressive, the other is diffusion. The other is, That's what I'm saying, right? But for a lot of image generation, image editing, high quality photorealistic use cases, developers are still using Imagine, but then, you know, slowly, but surely we're bringing those capabilities into Gemini.

14:29Whoever's watching this, we had a mid IO switch because obviously, you know, there's a lot going on here. There's not AI shapeshifting. I know, I know. But we also have Quinn, actually, who made this podcast happen. But you're a founder, CEO of Daily. Welcome. I'm a big fan of all things voice and audio. So it's fun to be here with you and with Shrestha. Quinn actually runs the voice AI meetup in San Francisco. You are basically consistently the leading community builder. And you're very generous with your time and knowledge. I really appreciate that. And obviously, also recently you started PipeCat, which is this open source framework for voice orchestration.

15:06Which has really great support for all the Gemini models. And you wanted to say something about the relationship with Gemini and Daly? I just wanted to say that it's been a very, very fruitful partnership with Daily. They've been our partners since the launch of the Live API. And, you know, a lot of their feedback that they continuously get has been, you know, instrumental to the success of the Live API. So both Daily and LiveKit were partnered with them. Quinn, I think, you know, we had a little bit of a prep for this. You also wanted to dive into a little bit on the cascade of models in Gemini Live.

15:43I mean, I think Shrest has taken a really interesting approach designing these APIs. So you talked about components a little bit. You talked about how you want to be able to do things both in the live API and in the more traditional chat API. And you've got, originally you designed the live API to have audio in, but then it's a separate text model, the notebook LM models, audio out. What was the sort of driver for that originally? I mean, that at the time was, we wanted to hit a certain quality bar, a certain latency bar. And, you know, Notebook LM was already out. And the TTS models that were powering Notebook LM were very, very good.

16:19But we wanted an aspect of native. So it was native audio in, but TTS out. And we still have that architecture available through the live API. But then now we just released audio to audio architecture. I mean, the infrastructure for this stuff is so interesting because you're always balancing latency, cost, output quality. There's no free lunch. Yeah. And other things like multilinguality. Coming back to your question earlier, Sam, we had a lot of users asking us for, say, better German language support or something, which hopefully now we've delivered on with these models. Yeah. But now you have audio to audio in the live API as well.

16:58In the live API only. is where we have the native audio output models. Continuing to pull on the component versus single model thread a little bit, when I think about voice, I think about it as being an area where to deliver solutions, you need to surround that strong model with a lot of voice-specific infrastructure that is, you know, I'm imagining challenging the scale. So, Shrestha, can you talk a little bit about that? And maybe we can have Quinn talk about that from his perspective. So the first thing that comes to mind is, of course, the voice activity detection models that we have. And we've done a lot of work like finessing that model server side.

17:41But we've also learned that we need to provide some knobs to developers. So now developers can actually tune the sensitivity on our voice activity detection model, as well as, you know, how much of the prefix, like how much of a time duration at the beginning, at the start or stop of saying things. And we also have a mode where now you can disable our voice activity detection and bring your own. But I think the larger point that you're touching on, Sam, that I do want to mention is it is really, really hard to bring all these components together and still get latency down to where it needs to be in the 500 to 700 millisecond range.

18:22It's one of the hardest things we've had to do with the live API. What we see is that the shape of building these real-time voice agents is a different set of developer problems in the shape of non-real-time or text mode things. One of the fun things about partnering with Stress and DeepMind is we work on this open source framework that people use to build these kind of production voice systems. And so we try to solve problems at the framework level like turn detection, like context management. as the models get better, as the use cases get more clear, some of those features migrate from the framework into the APIs, which makes life easier for developers.

18:58The use cases at the same time continue to broaden out. And so there's more things for the framework to do. So we're sort of filling the top of the use cases, building blocks, developer experience funnel, and pushing down as we all get better and we all figure out what this new world looks like. And maybe this is also a good segue into WebSockets versus WebRTC, Quinn. Yeah, you know, there's so much infrastructure. Like for my whole career, I've been building, you know, large scale, low latency network stuff. What we saw from my perspective when we started to see the possibilities of voice AI was you need this packet routing like down underneath the inference layer.

19:32There's like the AI inference stuff, but then there's the just how do you move the audio and increasingly video around the Internet. And so there's a whole new generation of developers who are interested in these networking protocols because voice AI and now real-time video are so interesting, which is super fun for me because I've always thought moving packets around is one of the most fun things you can do on the internet. Yeah. Seven layers of the OSI stack. Exactly. Exactly. Yeah. At pretty demanding real-time latency, this rest is saying, like, human beings expect you to respond in a conversation in 500 milliseconds or so.

20:05And if we're talking to an AI, we don't relax that assumption. We bring our assumptions about human conversation into that experience of interacting with an AI. Yeah, or not respond. That's a great point. Yeah. So like one of the features that we've pushed out, a little more experimental, but we'd love for people to test it is what we're calling proactive audio. And it's available only in the audio-to-audio architecture right now. and what this feature does is it's trained not to respond to irrelevant audio okay so it's like a refusal kind of yeah or you could call it directionally like semantic voice activity detection right um so basically uh yeah like let's say i'm talking to the ai and then quinn comes and asks me a question and i respond to quinn it'll know when not to respond um so yeah that in one of the demos the ai seemed to ignore a background question from someone else and yeah I think there's two threads to pull on there.

21:06One is that's another great example of things that we had to work really hard at the framework level to implement. It's much, much better if it actually migrates down into the model or the API. The other is part of the magic there is this semi-separate feature, but I think they're multiplicative of now your models can actually recognize two different people just based on their voices. You and I replied to that. They have to. This is not officially supported yet. But the model just does it. But just try it, right? Just try it and give us feedback. Is it okay to talk about it? Because it might be my single favorite thing you can do with these models that you previously have not been able to do.

21:45You can talk about what you've observed. I'm just saying it's not official. And what specific models are we talking about? Because speaker identification and diarization has always been really hard for these models. This is all the, it's called, gosh, like model naming now has become, but it's called the native audio dialogue. You'll see it in the live API, but that's the model. And then, you know, to your point again, Sam, about architectures, one thing that we launched on the cascaded architecture that we hope to eventually bring to the native audio as well as asynchronous function calling. So earlier, the way it used to work is if you wanted the model to do a function call, you'd have to wait for the response.

22:27And now you can set a non-blocking parameter and the model can go off and execute the function in the background. I love you so much. Yeah, that's great. We do have to wrap up. So I think one fun thing that we can do to wrap up would be a wish list for like next year's IO. What would be one thing that you would wish? It doesn't have to come true, but you know, wish happens with Gemini. Well, I was hoping for Gemini 3.0 at this IO. So maybe Gemini 5.0 at the next aisle. You want to tell us what you mean by Gemini 5, what you want in Gemini 5.0 then. I'll let Quinn go. I'll just put on my hat as representative of a big community, people building this stuff, more and more languages because AI is global and there are so many communities all over the world that are starting to use this stuff.

23:11Can we do language LORAs? You know, it's hard to stuff everything in one language and it's in one model. Yeah. Okay, but yeah. But they're building one model, as they said. They're building the one universal model. That would be a boring answer, but I think really more and more. No, languages. I mean, wasn't I telling you earlier, like we officially support 24 languages, but you can try talking to the model and cling on and it'll respond to you. So I think we'll get there way before next IO. But I just think more and more capabilities into the main model is what I would say. I'll have to think about this.

23:47Yeah, yeah. It's a fun parlor game, but it also helps people align as to what is possible and what's coming up. Thanks for your time, everyone. This is very hastily organized, but I'm glad that we can make this happen. And it's nice to actually see Sam in person. Same. Yeah. Swix, you think I am totally the PM for the live API, but we did not get to talk about some of all of the other releases as well. We'll save that for your talk at World's Fair. You guys are all speaking and we'll be podcasting as well. Sounds good. Yeah. All right. That's it. Thank you so much.

24:22Thank you.