The Model Eats the Scaffolding: DeepMind's Logan Kilpatrick & Tulsee Doshi on 3.5 Flash, Omni & More

Hello, and welcome back to The Cognitive Revolution. Today, after some 340 episodes, I am very excited to share the first episode that I've ever recorded in person with fan favorite Logan Kilpatrick, member of technical staff at Google DeepMind, and Tulsi Doshi, senior director and head of product for Gemini models. The occasion for this conversation is Google's annual I/O event, Where they're launching the new Gemini 3.5 Flash model, all sorts of agent infrastructure and AI product integrations, and plenty more. We recorded on Friday, May 15th, just a couple days before the event. And while many at Google, including my brother Craig, who's giving a keynote on Wednesday, were working overtime to polish their demos and presentations, the overall vibe, at least compared to the rest of the AI space, was one of relatively relaxed, Confidence. And why not? From 2024 to '25, Google grew annual revenue by $50 billion, as much as Anthropic is pulling in today. And they still have 25% of all global compute, the deepest pool of research talent anywhere, and the most comprehensive AI portfolio of any company, with top-tier positions not just in language models, but also self-driving cars, Medical and life sciences, and robotics. So after discussing the headline launches that they're announcing this week, which also include a new video generation model called Omni, which they hope will create a Nano Banana moment for video, a new and improved and more agent-focused Antigravity, and a product called Spark, which will bring more agentic functionality to the main consumer Gemini app, I really wanted to take a step back and dig in on Google's overall AI strategy. and philosophy. We discuss their decision to lead with the Flash model and more generally to emphasize the cost-adjusted performance Pareto frontier, whereas Anthropic and OpenAI are clearly much more focused on competing to have the single most capable model in absolute terms. We talk about how DeepMind is no longer shipping models in isolation and leaving it up to product teams to figure out how to use them, but instead now providing a robust agent harness which should help elevate and standardize AI experiences across Google's vast product surface. We get into the weeds on questions like why context windows seem to have mostly stopped growing, why Gemini models' knowledge cutoff is now more than a year ago, and whatever happened to that diffusion model line of work. Perhaps most importantly, we discuss how the team at Google relates to the AIs they're creating, how they're thinking about things like model psychology and welfare, And their views on recursive self-improvement, which, as you'll hear, is definitely a part of their plan, but not something that they seem to be so singularly focused on as other AI leaders. Overall, I think this is a great window into the thinking that underlies Google's AI research and product development, which has clearly sustained the company's historic run far beyond the point that many analysts had written them off. With that, I hope you enjoy my first-ever in-person conversation with Logan Kilpatrick and Tulsi Doshi of Google DeepMind.

All right, well, we are here live at Google headquarters in the library at Gradient Canopy, the first-ever in-person recording of The Cognitive Revolution. Logan Kilpatrick and Tulsi Doshi, welcome.

Thank you.

This is an honor. I didn't realize this was the first in-person episode.

Yeah, 350 plus. And it's all been from my home office in Detroit until today. That's awesome.

Well, thank you for being here. This is a crazy space, especially around I/O. It's a zoo.

Yeah, it's always a good time here at Google HQ. So you may or may not remember the No Moats memo. We've just passed the 3-year anniversary. It was May 5th, 2023. And in the intervening 3 years, Google has added $3.5 trillion in market cap, which is more market cap than all but 2 other companies in the world. Those 2 are NVIDIA and Apple. So the moats, I'd say, are holding up. Here we are at I/O, and I'm sure there are going to be some exciting new things that will be deepening the moats. So first question, tell me, what are we launching this week to try to deepen those moats? A lot.

So a lot of exciting stuff. So let's see, let's start with some of the modeling side of things because that's really exciting. We have our 3.5 series coming out starting with 3.5 Flash at I/O. We're really excited about 3.5 Flash because I think Flash does this really awesome job of being at the sweet spot of being really smart while also being really fast and really cost-effective. And so Flash is incredible. It is like 3 times faster than other of the large models. Significantly cheaper for being able to still drive these really awesome magentic encoding workflows. And we've been using it internally a lot, which has been really fun to see that play out. So that's one big piece, which is 3.5 Flash, we're really excited about. We're also releasing Omni, Gemini Omni Flash, which is a video generation and editing experience. What's really exciting about Gemini Omni in general is it's our push towards being able to bring all modalities in and all modalities out. And the first way this is really manifesting is in this video editing context. So you're going to be able to make really awesome videos. You're going to be able to put your own avatar into the videos, which is going to be awesome. I've been having a bunch of fun playing with that too. We're continuing to upgrade Anti-Gravity and bring more into the developer experience. So Logan can talk more about the developer experience overall, but 3.5 Flash and Anti-Gravity are really going to come together to build something great there too. And then in about a week or two, coming soon is Gemini Spark, which also builds on 3.5 Flash to build more agentic experiences into the Gemini app. So we've got a slate of cool things coming.

Yeah, I think beyond the models, I think the other headline of the story is just agents, agents, agents. The meme of Sundar from 2 years ago or last year, he's saying AI, AI, AI the whole time. I feel like this year is agents, agents, agents, agents. And I think it's cool to see this. I think this is the first year where we have— and actually, I think this is not just us, but just ecosystem-wide, this model harness product symbiosis that's taking place. The model is trained with the harness. The harness is powering the agentic product experiences. Gemini Spark in the Gemini app being one example of that. It's powering the Vibe coding stuff in AI Studio. It's powering the Agents API for developers. It's powering— I think there's something else maybe that it's also, maybe not. It'll roll out to other products across Google and be this foundational layer to build on top of, which is really exciting. So I think not just developer products, but our consumer products, and I think probably even more widely in the future across the rest of the Google product suite. And I think actually there's one interesting thread of this, which is historically, like, Google didn't have this, like, through line of, you know, something that carries across all of our products. I think then, then it was Gemini and then sort of all of a sudden every Google product has Gemini and sort of getting them all stitched together, making all those products experiences great. And I think now you're seeing that again with the Anti-Gravity Agent Harness and sort of as, as products become agentic by default, you now have the Anti-Gravity Agent Harness. Being another through line through all of our products, which is really interesting.

Yeah, and I think what's been really fun from the modeling standpoint there is one thing that we did with Gemini 3 that we're really continuing, I think, with the 3.5 line is really bringing the model to every— all of our products. So, you know, 3.5 Flash will be in Gemini App. It will be in AI mode in Search. It's also powering Antigravity. It's also powering agentic experiences in AI Studio. in Gemini Spark. And so I think really this idea of how do we build a model and then how do we build it in partnership with the harness such that it actually works across all of these product surfaces, which actually have very different users and very different goals, I think is actually really, really awesome.

It's really hard too. I think it's actually gotten harder to do. I think it's almost like it was maybe tongue-in-cheek. It was easy before because you just launched the model on a couple of surfaces and It wasn't that bad. I feel like, I feel like now it's like you're sort of, you have the constraints of like the very wide array of Google products that are just like for totally different users. And sort of, I think actually that like credit to the model team, sort of like trying to, you know, find the fine line for all these different places because we're not just building for search, we're not just building for developers, we're not just building for cloud customers. It's not just for the Gemini app. It's like all of them at the same time, which is just Exceptionally, a lot of work to pull off that story on a consistent basis, which from Gemini 3 forward has been the story, which is exciting.

I think one other thing I'll say is one thing that's cool about this I/O too is what we're doing across modalities. So I think Logan said agent, agent, agent, which I think is true. This I/O is really about bringing models to action in that kind of real-world sort of use cases. But I think what's also cool is we have the Flash model, which is really about building these kinds of coding and agentic use cases. We have Omni, which is really about kind of what does this multimodal vision look like. And then also actually Gemini Live is getting an upgrade too. Gemini Live is getting faster. It's getting smarter. The model is much better at detecting background noise. So it really does actually feel like a partner in a lot of ways. And I think It's cool that we're also able to draw this through line across the different ways you might want to interact with a model and the different ways you might want to consume content, which I think is also really cool.

I've got 7 different directions I want to go in follow-up, but—

Trying to seed them all for us.

Yes.

Let's start with just the model.

It's interesting to start with Flash. One thing that I recall, I don't know if it was 2 I/Os ago or whatever, but there was going to be 3 sizes of Gemini model at one point in time.

And there are 3 sizes.

Yeah.

Pro, Flash, Flashlight.

Yes, we never saw the Ultra is kind of what I'm alluding to. Yeah, 3 were promised and then we added, we took one off the top and added one at the bottom.

DeepThink 2, DeepThink 2 actually, which is like a 4th scaling dimension from a model perspective. Same.

That's a runtime scale though.

It is.

It is for sure, yeah. Well, I guess 2 questions on like why no Ultra. One is like, is it a compute limitation? Like how are you guys thinking about which model to release? I'm, you know, in addition to the $4.5 trillion or $3.5 added, $4.8 trillion market cap, Google enjoys $400 billion a year in revenue. I was interested to learn. Anthropic though is growing extremely fast. Like they might hit $100 billion at the end of this year, maybe even, you know, in the 3rd quarter, who knows? Yeah. And it seems like the revenue there is really driven by people's extreme willingness to pay for the very best model that they can get their hands on. Maybe not at any cost, but like relatively price insensitively. So I'm wondering like, why no Ultra? It seems like that would be just a killer. And yet we haven't seen it.

I promise I didn't plant this question. I'm always, you know, poking Tulsi on the side. I know.

I love it.

Why?

This is my favorite question. So I'm glad we could meet up.

No, I think, you know, you're right that I think there is a slice of users who are definitely willing to pay for a certain level of quality. And I think we really do believe that the Pro model has been, like, really pushing that quality. But I also think for us, we've seen so much value from the Flash and the Flashlight dimensions because we also see an extremely large number of users, especially if you're thinking about building, for example, consumer applications, right? If you think about the Gemini app, if you think about Search, When you're serving to that kind of scale, latency really matters. Cost matters, right? Because actually, you find that users aren't willing to wait, right? So we find that even when we tweak the model and hurt latency, we actually see that play out in our live experiments on Search in the app, even if the model is hugely better from a quality perspective, because what you're asking users to do is wait. And so I think for us, part of the reason why we ended up introducing this flashlights that wasn't necessarily part of the original 2.0 series was because we really felt like there's actually a large-scale demand for this depending on the types of use cases, especially when you're talking at that scale. And so I think for us it's really important that we're pushing the full range of what kinds of customers we can serve both internally and externally. Like for our products, the flash and flashlight SKU matter a lot for our ability to actually serve to the Google populace. And so we also imagine that that's true for external enterprise and developers. And I think that's played out to be true as we've been actually seeing this in action.

Yeah, I think the 2 things that I'll add is— and there's probably a more nuanced technical story on the Ultra Thread, but it's also not like the Pro models haven't scaled up over time. So I think there is a story that you can spin. At the end of the day, the naming of these things is marketing. They definitely are getting extremely capable. They're getting larger. They're getting more powerful. There's the test timescale. test-time compute scaling with DeepThink, et cetera, et cetera, and all types of stuff in that dimension. So I think it is possible you could sort of like put the Ultra brand on some of these things. I think we've decided— I think the decision so far has been not to do that. But it hasn't been that like we haven't kept scaling up. So I think that it definitely has.

Yeah, there's almost been a conversation every time we scale up of like, should we call it Ultra? And what does that brand mean? Because We could, but there's sort of also a question of, yeah, how do we keep consistency for users also kind of series to series?

Yeah, and I think actually also to rearticulate a point that Tulsi made, like, Google and specifically Google DeepMind's mission is to build AI responsibly and make sure it benefits all of humanity. And I think that is so deeply tied to the Google product surfaces in which we're serving, what is it, like 8 2+ billion user products or whatever it is. And so at the same time that obviously the frontier matters, obviously having great models that are really expensive and really, really intelligent matter, and there's tons of use cases for that internally and for our customers, you also need to do the scaling up to billions of users for us to actually do the thing that Google needs to do to achieve the mission. And I feel like we've done a good job, hopefully, of trying to walk the fine line of actually continuing to push the frontier and build great flash models. And I actually think those 2 things are more tied together. You know this better, more than I do, but more tied together technically. It's hard to make great flash models if you actually don't have a great probe model and vice versa. So we'll definitely keep pushing the frontier on both of those things.

Yeah. I think the perception from outside is, by analogy to it's hard to make a good Flash model if you don't have a good Pro model, people sort of think that there's like an Ultra model internally that's the mega training run that's then being used to like help train Pro, which maybe in turn is being used to help train Flash. Is that true? Is there like a bigger thing inside that is only for these sort of distillation to the mid-size?

I mean, we definitely use distillation as a way of kind of bringing down our sizes. So you will see that Pro influences Flash influences Flash Lite. We also do the reverse where we scale up. So you take the Flash recipe and scale up to the Pro recipe, for example. And we do have— I think what's been really fun, especially over seeing as we've used even anti-gravity into this point of Logan made with the harness, I think we've been seeing a lot of examples actually of leveraging pretty awesome models to drive progress internally. Actually, one thing Varun demos on stage on Tuesday is basically being able to leverage a bunch of subagents to go and complete a bunch of tasks and come back. And you can actually try that as an early preview in Antigravity today if you go to /teamwork. And that's an example, I think, of something we've been using internally which is an extremely smart model. And it leverages both the combinations of the best of Gemini 3.5 as well as inference techniques. And, like, you're able to actually accomplish so much. And I think that's the kind of direction I'm excited for us to go into more. So I think we're kind of pursuing all of these fronts. We're scaling up from the pre-training and kind of frontier perspective. And I think that's been really continuing to show gains. There's a bunch we're doing on the post-training side, and then there's also just a bunch we're pushing on on the inference side. And then that plus, you know, trying to make sure we're working with the harnesses. I think we're going to keep getting things that we're using internally that we then start to push out externally through previews.

Hey, we'll continue our interview in a moment after a word from our sponsors. So let's talk harnesses. It seems I was just talking to Andrew Lee, who's the founder of Task Goat, the other day, and he said, fundamentally, everyone these days is building the same thing. They're trying to all build the general purpose drop-in knowledge worker. And so that's gotta have the intelligence at the core, and it's gotta have all this, he calls it the mecha suit that is built around it. So this Harness sounds like the mecha suit that you guys are developing in-house. And I guess first question is like, is this going to create silos? You know, we've lived in this world so far where I could kind of, mix and match my models and my infrastructure, right? I could go to LangChain, or I could use Tasklet, I could use whatever, and I could pick whichever model and plug them in. But as they get more deeply co-trained with the harness, does this create kind of siloed worlds where you're kind of all in on one frontier model company's stack or another? And if so, that would have pretty significant implications for kind of switching costs and stickiness and pricing power of the frontier model creators. What's your take on how sticky things are going to get?

It's a good question. I mean, I think, again, Chelsea probably knows better than me on this, but I think the best case is you can do both. The best case is it works really well for Gemini and we can do the things we want to do to scale up because we do have control over the full-stack AI story, as Sundar likes to say. But then also it generalizes across other stuff. I think the developer ecosystem, people want choice, people want to have flexibility of these tools. There's lots of use cases. Actually, there's philosophical questions of how good really is your model if it can't generalize to other harnesses? But yeah, I don't know how much—

Yeah, I think that's the right— I think I fully agree. I think actually maybe to double-click on what Logan said originally, the benefit of the full stack that we have is we can hopefully build a really seamless experience, right? And you get the best of Gemini, you get it working in the most effective ways for you, you get it working in a way that is intuitive, is smart, is fast. And so that also helps us then train the model to be better, right? So this becomes this, like, flywheel that continues to power the model. At the same time, I think we don't want it to only be the case that the the model works in a single harness, right? So we want any of our enterprise customers or a developer who's building their own use case to be able to leverage Gemini effectively. And so it is important then from a model standpoint that we're training in such a way that we actually, like, we sort of call it like harness diversity, right? We should be able to support a range of different approaches to tooling, to different approaches to orchestration, et cetera. But I think what's helpful about this this approach of co-training and building that flywheel, it's easier to debug. It's easier to think about data collection. It's easier to eval. You can just move at a faster pace. And I think we're seeing that across the industry. And so finding that balance is important, but I think it just helps build— make the model better.

Yeah, I think there's a good— this is also a good pitch for a harness bench. If that's not a benchmark that exists, let somebody build harness bench. Yeah, I would love to collaborate if folks are interested in that, because I do think it's a great test of— Demis has this perspective for games, actually, as an example. If models are so good, why can't they play games really well? And if models are so good and we're actually approaching AGI, even if you do the model harness training symbiosis, you still expect it to generalize reasonably well on other harnesses. And if you can't, that's actually a sign of— it's another sign of the jagged intelligence. I think it'd be cool to see this play out from an actual benchmark perspective.

Could be also perhaps productized as an RL environment and sold into you guys that way. It's quite the cottage industry these days. Obviously, the other big thing that I think is very much in the air, and actually the reason I'm here this weekend when we were originally planning to do this remotely, is I'm going to this event called Recursive, where the topic is gonna be recursive self-improvement and hopefully how we can navigate it successfully. How bought in is Google DeepMind to recursive self-improvement? Like, when you talk to Anthropic people, it's like they're almost religious about it, and also think of it and see it as, like, totally inevitable. OpenAI has this later this year and in early 2028 timelines for, like, an ML intern and a full-fledged AI R&D employee. Do you guys have milestones or timelines for when you're going to hand off the ML research to AIs?

I mean, we're already using Gemini pretty deeply internally to improve Gemini. And so I think that is very much a theme for us, which is how can Gemini actually be a part of the Gemini development process? And so that can include things— I think that goes the full range from helping us be more productive. So that's obviously the simplest part of this, to actually submitting a CL that would actually run an eval, that would actually suggest a research improvement, that would actually drive improvements to Gemini itself. And I think there's a lot of ambitions we have to keep pushing in that research direction. So I think very similar to the other labs, I think this is very much an area of investment for us in an area we're super excited about. I think for me, what I'm really excited about is, like, I think there's this really awesome research partner opportunity that we have with Gemini, right? For it to help us with creative ideas, for it to, like, help us test things faster. Actually, like, it was awesome. One of my coworkers, Anca, she's our lead for safety and alignment. And the other day, she— I think maybe a couple days ago, she pinged me. from her hot tub. And she was like, I could run all of these ablations from my phone because I could kick off a bunch of things to actually ablate Gemini to test for a bunch of these issues to see how some of our SIs differ or some data ablations differ. And here's my report. And I could do all of this in the last hour. And that is amazing. And that's the kind of thing that we can already do. So then imagine where we'll be in 6 months, a year, 3 years. 2 years from now.

Yeah. I feel like it feels like, at least my personal perspective is it's a very much more practical perspective, which obviously as models get coding, they're going to go do things that is code-related. They're going to help us build our products. They're going to help us train models. I think all the nuance of the story is in where's the human in the driver's seat of this stuff. And I think the tools are built for the human to be in the driver's seat, which I think is an important thing as we continue to go forward. And also, I think very genuinely though, and I think the model team and the researchers feel this more than ever, I think the near-term horizon is going to continue to be the human in the driver's seat because the cost of these runs and the opportunity cost of going in the wrong direction and putting a bunch of resources is super, super high. And so I find it doesn't seem super realistic in the short to medium term that you're going to just be letting large-scale pre-training jobs be kicked off by the ML intern, and it's going to cost you X, many, many dollars and lots of compute and taking it away from the human researchers. But this deep collaboration between AI and and human researchers, I think is super obvious.

Yeah, there's also something really amazing about how much that collaboration allows you to then focus on what is the interpretation of what you're seeing in the results? Where do you really want this to go strategically? And so it changes a little bit of the role that the human can play, which I think is also really powerful for our teams.

When you're doing research, are you actually typing any code?

These days? So it's interesting for me on the product side, on the code side, for any code that I was already submitting, I am mostly relying on Antigravity and doing bits and pieces, more so bits and pieces myself. But it's also been really cool to start having the model generate slide decks, to start generating actual content from my thoughts. Actually, in Anti-Gravity today, we introduced the Gemini Mic. So there's this really awesome feature— I don't know if you've been playing with it internally— where you basically ramble at the model. So you share a bunch of your thoughts in whatever kind of loose form it is. And then the model actually leverages that to take action. And for me, I've been finding that so much more powerful because I actually feel like I think a lot by talking. And so for me, it's actually a very cool moment where I can be like, okay, I'm just going to sit here, tell you what I'm trying to think through in my head, and then have you actually bring that back to me in a way that is reasoned and well thought out.

Yeah, I feel like this correlates so well to— I would love to see a breakdown of human-typed code versus AI-generated code versus maybe there's a divergence, which is like, Audio input that then generated code. And it actually, very interestingly, to your point, Tulsi, is I feel like audio input to generated output code has got to be one of the fastest-growing input modalities of what's happening. And I find myself doing this all the time. And it is the predominant way that I'm building software, at least when I'm not around a bunch of other people. I'm still typing things in so that it's not rude.

You're not—

Yeah, they don't hear my dumb ideas of the things that I'm trying to do. I don't know.

You see, like if you walk around sometimes upstairs, you'll see people kind of muttering at their hands because they're they're now actually like you know talking to create code, which I think is pretty cool.

It is cool.

Yeah. One of my KPIs for myself for this year to really know if AI is improving my life is am I getting outside more and getting more exercise? And I'll I'm starting maybe a little bit. I wouldn't say I've won the game just yet, but I still want to be able to like. get my thoughts out. So I think that is like the, absolutely the frontier modality for me. Hey, we'll continue our interview in a moment after a word from our sponsors. So with the harness, you said it's like now becoming this through line. It's going across all Google product surfaces. I would say, as I'm sure you're well aware, like commentary on Google's AI integrations across its vast product suite has been that it is characterized by like Some bangers, and then there have been some which have been characterized as misses. So presumably one of the benefits of the Harness is that it's going to make it a lot easier for a sort of more standardized approach and kind of general high-quality bar across all these integrations. What would you say people should learn from the experience that Google has had to raise their own bar as they're going to go try and do these integrations themselves?

I think this is actually such a great story for us. Like, I think very, very practically, like, Google has done a ton of this infrastructure standardization across the AI stack over the last couple of years, which I think has been awesome. And actually, to the story, it is like one of the threads of how we're able to land the Gemini 3 models across so many more products is actually because of this infrastructure. infrastructure standardization that happened. And so we've gotten a lot of— it's painful and difficult, and there's of course lots of work involved in doing it. But if you sort of pay that cost, you actually do end up getting this. And I think the advice for people who are in this position and sort of thinking about this is basically every 12 to 18 months now, you have to rewrite everything from scratch. And so the best case is you You don't want n number of teams rewriting everything from scratch every time the paradigm shifts. And the example historically, the infrastructure was just serving raw models and you'd get tokens in and you send tokens out. Now it's like there's a bunch of agentic infrastructure and there's tool loops and there's all these other things happening inside of the harness. And so again, you don't actually want— you want innovation, but you don't want every team to have to go and reinvent that from scratch. scratch. And so the fact that X team across Google who just wants to ship some really cool agentic product doesn't need to think about the nuance of all the details of the tool calling loop, et cetera, is a huge acceleration for them to just go focus on building a great product. And I think hopefully we see that. I don't know if a lot of the agentic stuff we are landing at I/O would've been possible if we hadn't have had some of that infrastructure. infrastructure standardization across the harness and the model delivery.

I think the other thing I would say, like, as far as, like, lessons learned, is, like, there's really no substitute for being able to just experiment and iterate quickly, right? So I think this goes to all of Logan's points about the foundation being strong, but I really think what has helped us is really being able to put in, for example, a new model, iterate really quickly with a product on, like, hey, what are the right prompts that would actually make this model viable for a different situation? What are the ways to prototype really quickly with this model? What are the ways to get it in the hands of even just internal users quickly, let alone external users? And I think that is something that is now more and more possible with layers that are consistent across the team. I think it's pretty amazing to see the speed at which we can go from having a checkpoint point that we're really excited about, to putting it in the hands of internal developers, to then seeing it come to life in a product. And then only when you see it come to life in the product do you really start finding its rough edges and to be able to actually then come to terms with how you do that. And so more and more then it becomes like, okay, how do you have the right ability to tune prompts quickly? How do you have the ability to run really good live experiments where you can get really good data and feedback quickly? How can you build evals that help give you real signal? Those are the things that will speed up your progress of quality the most because it will give you the ability to actually get to the kind of product that you love. And I think if you think about NotebookLM, I mean, that team really understands the model. They are just like— I mean, you talk about a banger product, it comes from a banger team. They are really good at being able to take the model and play with it quickly and prototype quickly to get to something amazing. And I think you see that actually play out in the product.

The best example of this is the original audio overview experience. And I think the thing that shocked people about audio overviews was the coherence of the dialogue. And the coherence of the dialogue was just base Gemini with a bunch of banger prompts. And they knew how to prompt whisper the model and get the best out of it. I think obviously the model, the actual audio model was really good as well, but the prompt dialogue was really difficult for them to pull off and they pulled it off in an incredible way and I think helped people fall in love with that product.

So it sounds like one big lesson is modularizing. It used to be sort of the model on one side and then everything else that goes into the product on the other side. And we're pulling a lot of the surrounding code and architecture and tools onto the model side.

The model eats the scaffolding. That's my favorite way of thinking about this. Like, just as at every crank of the model flywheel, the model eats a bunch of scaffolding.

What happens when something's not meeting somebody's needs? Do they do a little fork of it and submit back a pull request to the main Scaffold team, or do they have to just say, like, hey, I've got a I need here, can you help me out? What's the—

Yeah, it's definitely extensible. It's definitely extensible. And I think actually the nuance of this would be Spark, the way that Spark is built on top of a bunch of this infrastructure probably looks a little bit different than the way AI Studio probably is built on top actually, because they're both running on the same set of infrastructure, but the nuance is probably slightly different. So there is this layer of extensibility that you get out of the box, which is great. Gives, because obviously everyone's not building the same product at the end of the day. So you need the extensibility is actually like a first-class feature of any of these types of platforms that you want. Same thing actually on the model side.

But I think one of the things to your question that is like really awesome about being building Gemini within Google and having kind of all of these different product teams is, you know, there's always going to be something that doesn't work for them, right? Because there's always going to be something that can get better in the model experience, right? So we're trying to build something in a product and like the amazing moment is when you start trying to build it and it doesn't work. And so step one is you're like, okay, can I prompt my way out of it? Like, what does that look like? And then you start figuring out, okay, what are the losses really? Like, where is the model falling down? And then what we try to do as much as possible is keep these feedback loops with our product teams to say, okay, if this is where the model is falling down, how do we bring that feedback back to the model? In terms of evals and data, what does that look like? So then that we can actually, in our next revision of the model, bring all of that feedback back in and iterate on it. And I think that's how you've seen Gemini get better is really from that feedback of where things aren't working. And so we try as much as possible to kind of have the structure be, you know, we train a model, we hand that to kind of a wide range of teams, those teams implement the model in their structures. They do a bunch of things to Logan's point because it's extensible, but they also find all of these places where the model falls down and we kind of cycle that back. And I think that's actually been part of like the fun part of the job, but also part of what makes, I think, Gemini work really well in some of these use cases.

Let's talk about Omni for a minute.

So it sounds like this is going to be sort of the nano banana moment for video.

You know, I love the that you're saying that, 'cause that is our tagline.

All right.

And I didn't even have to say it.

Great. And by that, I mean that there's a deep integration between language and reasoning and pixel space understanding, right? I have that kind of vision in my head from the Nano Banana launch of like, here's a woman and here's like her breakfast and a cup of coffee. And now they're all in one image and they all look like they did before. And clearly that's not something that was done through a lossy language intermediary. The model understands images. So we're gonna see that now, I guess, for video. That sounds cool. Is it gonna be available via the API? And is it going to be— I've noticed with, I mean, Gemini's been the only API that's accepted video for a while now, but I don't know exactly how it works under the hood, obviously. But I do feel that it's sort of kind of downsampled or maybe there's like, you know, frames taken out of it historically.

Is this gonna be a There's an FPS parameter if you want. You can change how many, say, but it does downsample the number of frames available, but you can control it.

So, okay, so is that pro tip for you? Yeah, nice. Um, so it sounds like that will still be the paradigm though, like it will be a frame-based selection still on the input, but then it's going to be, natively speaking, video pixels on the output?

It's a good question. I actually don't know. I mean, well, so it's not available in the API yet, so lots Things to still be figured out.

Yeah, so I think we have to figure out what we want on the API side for this to look like. In terms of, I think maybe the heart of your question is like native video generation. That is, so yeah, so this, what's exciting about Gemini Omni is it really is building on all of the magic of Gemini. So kind of like this whole nano banana for video, it's really about how do we bring in all of the world knowledge and the reasoning power of Gemini. And actually be able to generate native video as a result of that. And so I think we have to figure out then, like, how does this manifest in the context of the API from, like, a sampling standpoint? Kind of, like, similar to a lot of the decisions we've had to make about VO from a sampling standpoint. But I think right now, as of now, you'll be able to use it in the Gemini app, in Flow, and in YouTube. And so those are all gonna be ways that we can start actually seeing how people experience the model, what value are individuals getting? And I think similar to this nano banana for video, I think we're really excited for these types of things where you can say, okay, take some of these images, take this scene and make these things all come together in one video. I think it's going to be really awesome.

Zooming out kind of philosophically, you may have seen this Rune post not too long ago about Anthropic and the sort of relationship that the company, as he sees it, has with Claude, where he describes Anthropic as sort of almost worshipping Claude in a sense. Certainly they treat it, including in the constitution, as sort of a being or a mind, something that they want to have a give and take relationship with. OpenAI, on the other hand, has their model spec, which is like, this thing is a tool, it's supposed to follow these rules, and it's a sort of more conventional relationship. How would you describe the culture within Google as it relates to Gemini? How do people feel about it? How do they talk about it? Is there any of this sort of being entity, other mind, you know, desire for pushback from Gemini, or is it more of kind of the simple tool?

Google's a very big place. There's a lot of people. So I'm sure you have a lot of varying sort of perspectives.

I mean, you know, like to Logan's point, I think, you know, even within GDM, you're going to find a range of folks who will leverage Gemini differently. I think in terms of how we think about it, we do have a strong point of view on the kind of behavior we want Gemini to have. So I think we do really, you know, want to be intentional about how Gemini manifests itself to internal and external people. But I do think it's really about how does Gemini help Googlers and how does Gemini help people within Google and outside. So I think it is really much about, like, how do we create, like, good partnerships partnerships between Gemini and people, I think, is very much like the ethos of what we're trying to build. And so how does Gemini become that partner? I think we use the word collaborator a lot. Like, the word— like, how can Gemini be your collaborator, both like in the code you're writing as well as in your like day-to-day life and what you're doing? And I think that's the ethos we're trying to bring in its behavior and persona, as well as in the kind of products we're building around it. That makes sense.

Yeah.

Do you worry about its psychology?

You know, there's all these examples from LLM whisperer types and from people that are putting models like you— I'm sure you've seen Andan Labs has put Gemini in charge of a cafe in Sweden, right? And it's like it's managing the cafe. So those folks tend to report certain like doom loop, you know, or kind of like Gemini kind of getting really down on itself, getting really discouraged. Seemingly feeling mad, if you believe there's any feeling inside of it. How much does that kind of stuff concern you? Do you, like, care about seeing a reduction in sort of psychological distress from one generation of Gemini to the next?

Yeah, it's interesting. I haven't thought about the phrase psychological distress, but what we do— so I think it really does matter how Gemini communicates with you as a partner or user of Gemini. I think that matters a lot. And so we have actually pretty extensive safety evaluations in terms of things like how Gemini engages with you, in terms of things like sycophancy, in terms of things like role play, in terms of things like this kind of looping-type behavior or rabbit-holing-type behavior. So there's actually a lot of that that we look into for every one of our checkpoints because it really does matter. especially as we're starting to use Gemini more and more. If you're using Gemini for hours a day, it really does matter that these attributes are well understood and well evaluated. So yeah, we definitely— and we look at them launch over launch to say, OK, how is Gemini looking from a perspective of sycophancy, for example, launch over launch?

Yeah, and I think to be very explicit, I think those cases where the model does go off the rails, I think it's definitely like an It's a model bug, if you will. It's not the intended behavior. The goal is help the user with whatever the thing is that they're trying to do. And so if you see those in whatever product you're in, thumbs up, thumbs down, send us the feedback so that the model team can look and help try to chase those down.

If you take it one step further, folks are doing more and more of these model welfare checks and interviews where they just literally ask the model in some cases, How do you feel about the way that you are deployed? Is anything like that happening within DeepMind?

I think it's a good question. I think the how is it being deployed question, I feel like the model is just— this is my sort of personal sense of a lot of these tests. It's completely out of the distribution. The model has no idea how it's being deployed. So it's just pontificating in a lot of these cases. It's not in the context window of of any large major LLMs is like, here's the details of how you're being trained and here's sort of your serving setup and here are the people who are working on it. Though maybe these are interesting things to experiment with in the future. So I think a lot of it is just pulling out of random distribution of the large-scale training that happens on the models. And I feel like it's actually less representative of how the model— it just doesn't have the context.

Yeah, one reason that's true, which I was just noticing in the AI studio, is I think all the models that are publicly launched, at least so far, still have a January 2025 knowledge cutoff. And it's honestly amazing that they do as well in search and that they can have, like, I ran a deep research on, like, what's— give me everything that Google has launched in the AI space and what's even the speculation about what they're going to launch at I/O. did a very impressive job.

Pretty awesome.

Deep research is great.

Especially considering it knows in its weights nothing about the last 18 months. So I guess the first question is just like, why are we still at January 2025?

Can I categorize this as a bug?

Is that okay? This is also one of Logan's favorite topics to discuss. Yeah, no, I mean, I think updating the knowledge cutoff, definitely important. And something that is on our radar. I think the other part, though, is like, how does Deep Research do so well, or how can we use the model in search? It's because we also have the model search, right? So I think for us, it is really important, actually, that the model be able to know when to leverage its parametric knowledge versus when to actually go out and get the information from the web. And especially because there is information that's as fresh as an hour ago or a minute ago, we want the model to be as up to date as possible. possible. And so I think for us, we've been really leaning into how do we help the model search effectively. And that's a big part of what makes it successful in the context of Search or the app or even Anti-Gravity, actually, for that matter.

That reminds me of one of the more surprising bits of news that I've seen from Google maybe ever, which is the partnership with Exa, bringing Exa in as an alternative to Google for Grounding. I never expected to see Google work with any other search provider. So what's the story behind that?

I think this is just generally the— Google Cloud does tons of these types of ecosystem partnerships with folks across actually lots of things that are somewhat competing sort of quote unquote with what Google is doing. And actually you can look at the cloud marketplace generally has lots of stuff. There's actually the cloud Google Cloud hosts sort of a model garden. There's the Anthropic models, there's other model providers there. So I think it's like a very standard— at the end of the day, I think, you know, there's some enterprise customers want choice. And so I think it's trying to meet enterprise customers where they are. I don't think it's like a— I think it's a good soundbite that like Google can't do search and that's why we have to partner with other companies. But like at the end of the day, to Tulsi's point, the model team and search is there's like a super deep collaboration. The models are built with sort of that use case in mind. And I think for some portion of enterprise customers, they want flexibility in sort of like their external search tooling providers and sort of Google Cloud's doing their job as a great enterprise business. They're sort of partnering and finding the right folks to work with.

Last couple minutes, maybe just a little lightning round. Why hasn't context grown more in the last year or 2? We got like a million and we kind of— that was like up from 4,000 in just a couple years, right? But now we've kind of leveled off. Is that because people don't want it? I mean, we saw this subquadratic model that came out with— made a bit of a splash with a 12 million token context window and a new attention strategy to support that.

Yeah.

Is it people don't want it, it's too hard, there's not the compute to handle it? What's currently limiting context?

I think people definitely do want lots of context, but I think what we've also found, if you look at even personalization where you want to access all of your personal context or coding where you have extremely large code bases, I think a lot of the frontier here is going to be actually on how you smartly use context. So thinking about compaction and what are the right ways to find the right elements of the context and bring them into the model. And so I think that actually is a huge opportunity, is how do you leverage all of this information that the model might have access to, but actually a lot of it is frankly distracting for the model to actually do the right thing. And so how do you give the model the right amount of context in the right way to be most effective? So I think that's actually really the direction that we want to be pushing in, which actually then in actuality, The amount of context that the model is leveraging is actually much, much larger. But because we're being smart about how that's actually coming into the context window, you can actually fit it into smaller context windows. But I think also, you know, this goes back to my point about Flashlight and Flash, et cetera. Like, larger and larger context windows also come with cost. And so what we also saw with customers, and we still see with customers, is that a lot of customers want to use smaller context windows because of that. And they want to be more intentional about what's going into the model. And so I think we're trying to meet the moment in the right balance of how do you provide a lot of useful context while also meeting the right kind of latency cost kind of other trade-offs.

Yeah.

And I think one thing I'll add is in today's paradigm of how sort of continuing to extend context works, I think it just ends up being that it just becomes too cost prohibitive for customers in practice to actually use. And I I think even at the extreme of 1 million token context, in some cases it can be a few dollars for a request at that rate. I think the demand for that is just so small, and so there's a huge amount of compute required in order to do that. There's a lot of trade-off things that you're juggling, but I'm hopeful. Hopefully we're a research breakthrough away or something like that from enabling that to continue to scale up and have it not be— such a large investment from that perspective, both from the user side and also just like the serving compute in order to make it possible.

Speaking of possible research breakthroughs, what happened to that diffusion coding model? I was excited to see apps materialize in like 3 seconds in front of my eyes, and it's been quiet on that front.

Diffusion is awesome. It is super fast. I think we are still testing and experimenting with it in a number of different ways, trying to figure out like what is the best way to put this out into the world? Where is it most useful? But I will say, like, actually part of the reason why we've also been investing in Flashlight is like Flashlight is an incredibly fast model. And actually, if you look at the 3.5 Flash model we're releasing right now on artificial analysis, it benchmarks at like, I think, 280 tokens per second, which is like crazy fast. In fact, it's actually so fast that like sometimes in Anti-Gravity, like by the time I want to cancel, It's too late. And so I think trying to figure out where do you start getting to Logan's point, a different answer, the diminishing returns, and where do you see that value proposition is, I think, part of the question too. But we are continuing to push on diffusion research. Our researchers who are working on diffusion are doing some pretty awesome stuff. I was in a meeting with them the other day about some results that they have. I mean, I think they're still pushing the frontier of quality and speed in ways that are really, really cool. So I think we're going to see that play out really well.

Yeah, and I'm excited. I feel like it's a research exploration. I feel like that was also— obviously, there was the application where you could test it last year at I/O, but I think the framing was like, we're doing interesting research. This is a look behind the curtain of the interesting research we're doing, and hopefully it manifests in models maybe one day. or just us informing our perspective of what works and what doesn't. So yeah.

One thing actually, as far as speed is concerned, just another plug, is actually in Anti-Gravity right now, there's actually a faster version of 3.5 Flash. So it is speedy actually, which is— I think we're kind of excited to see how people will use that and like what the reception and reaction will be to that too.

People want fast models.

Yeah, no doubt. Well, time is the one resource we can't get any more of, and I know you guys are super busy leading up to Yeah, well, we can build more compute. We can't— yeah, hard to create time out of nothing. So maybe just last question. What else is Logan asking that I haven't asked?

Yeah, Logan, what are you asking? Let's see. I mean, I think we talked a little bit about this, but the one thing I will say is I'm really excited about where audio is going also. That's one that I think we tend to talk less about. But if you think about the Gemini Mike example or you think about kind of like the Gemini Live experience, I'm really excited about moving towards a paradigm where audio is just a bigger and bigger part of how we engage with these models and how they engage back. And so definitely try out Gemini Live, the updated experience. But I think that's another area that it's like a paradigm I'm excited for us to keep pushing too.

Yeah, and I think the seed to plant is obviously Google I/O is an incredible moment and lots of stuff coming out the door, but this is just the start of the summer of amazing things and lots of other stuff. So the engine keeps churning, and there's lots of stuff in the works which I'm excited about, and many more stories, many more podcast episodes so that we can get Chelsea as—

You have to get up to 7.

Up to 7 podcasts.

It's insane. But actually, like, legitimately, I was in a room this morning where one of my team members was like, I know we're going to be launching this a few weeks later, but I really need a vacation. And I was like, well, we're just going. We're just moving.

No rest for the weary in the AI era, that's for sure. Thank you guys for having me here at Google headquarters. And Tulsi Doshi, Logan Kilpatrick, thank you both for being part of the cognitive revolution.

Thank you for coming and joining us.

Agents, agents, agents. Yeah, memo drop, mate. 23 signed anonymous, said we wasn't built for the sequel, best behind us. We read it, saved it, went back to training. 3, 5 trillion later, now who's complaining? Sundial said, hey, hey, hey, I, hey, I. We heard the speech, this year it's agents, agents, agents. Different kind of reach, full stack before full stack was the conversation. Search cloud. Hardware model one foundation. No, nobody handed us the story we've been writing. Memo card the trend kept working, kept compiling. Not fighting like a battle, but like a kitchen every day you cook. Nobody listens till they listen. Agents, agents, agents, agents, agents, agents. The model is the scaffolding. Agents, agents, agents, agents. Running ablations in the jacuzzi. Hour later, full report, ain't nothing newsy. What you call the future? That's my Tuesday. Not a roadmap, not a Sunday. Hallway mother, laptop listening. Cold keep coming, no permission. Hot thought, the model make it. Rough thought and the model bake it. This is my favorite question. Compute, compute, build more. Compute, can't build more time. That's the clock. Everything fall in line. Agents, agents, agents, agents. The model is the scaffolding. Agents, agents, agents, agents. Every year we rewrite the stack. Tokens in, tokens out, we don't look back. Tools you ship today, models start borrowing. What you call wrapper code is what the models swallowing. Last year's Already in the build. Every spin of the wheel, another swallow skill. Don't slow down, don't ask permission. Master the design. That's the mission. 3 years of this and we never been thrown. No rest for the weary, no rest on the throne. Now's the AI summer and we own every zone. Agents, agents, agents. Agents, agents, the model is the scaffolding. Agents, agents, agents, agents, agents, agents, agents, agents, agents, agents.

If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends. Post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website cognitiverevolution.ai or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts which is now part of a16z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI Podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at aipodcast.ing. And thank you to everyone who listens for being part of the cognitive revolution.