With Logan, Member of Technical Staff — on how they work, more than on the models.
Episode 1 of our podcast. Twenty-nine minutes with Logan, Member of Technical Staff at Google DeepMind, on the way they work — and on what a twenty-five-person company can take from it.
[00:00] Arman: I'll go straight to the point because we don't have much time. Yeah, yeah. What's one of the coolest projects you've seen someone build on AI Studio or using other Gemini technologies?
[00:10] Logan: That's such a good question.
[00:13] Arman: Yourself included if you want.
[00:16] Logan: I think one of the most recent ones that I've seen that just is, is interesting to me because you start to see people do-- actually, the conversation we were just having off-camera about people using AI in the real world, it was somebody doing like 3D CAD stuff, and actually like designing, literally like designing buildings and stuff like that, and using AI Studio to like have a more controllable, playable environment to go and do that. I think it was, it was one example, and I think actually it- Ended up in the, in the intro sequence to the start of the I/O keynote, the countdown was a bunch of people who built those things in AI Studio. and it was, I think that was one of my favorite projects to like see the creativity of the ecosystem come together and, and showcase a bunch of like very, very different implementations of, of that sort of countdown animation. It was, it was, it was very cool. That's amazing. How do you use it yourself? That's a good question. I spent-- So I think the most common use case is I spent a lot of So we, we're sort of like prototyping new features and like, what would it look like if you sort of brought a team of agents inside of the, the AI studio build experience? What would it look like if you, sort of removed all of the UI around the edges of the screen and it was sort of like completely immersive? What would it look like if it was, you know, and so your next favorite crazy idea? And so we, I, that is like most of what I do, outside of work, a bunch of random, random side projects, but that Internally, that's my most common workflow is like cloning AI Studio, making some changes to it, and sort of trying to visualize like what would that product experience really feel like.
[01:51] Arman: So you would try to get the UI kind of cloned and play around with the UI, add features, things like
[01:56] Logan: that. Exactly. Yeah, and we, and we have, internally we actually have like a bunch of templates for teams. So like similar to like if you go into Google Docs, there's like templates for, report and like ten other things. We actually, we have that for teams inside of Google Gemini app and AI Studio and a bunch of these things that you can fork off of so that teams sort of trying to do this really fast prototyping can actually have, not need to wait for the model to clone a bunch of stuff, it can have your design system and a bunch of other stuff just like ready and waiting for you, which is fun.
[02:27] Arman: You worked at Google on the AI Studio and then I think that some kind of merge happened with DeepMind and then you're part of DeepMind. Yeah. Yeah. How do you work? With Google, like how's the collaboration between Google and DeepMind? Because I remember the days I used to be a Googler for seven years, like a long time ago, DeepMind was another entity, you couldn't enter the buildings. How is it today? How do you co-collaborate? What are the processes? How do you share information?
[02:50] Logan: Yeah, they still have building access stuff, which is a, a separate comment, but in general, I think it's, DeepMind has become the engine room of Google. I think actually, like the, I was looking Google, and I think that's how DeepMind is. So the, the engagement model has gotten super, super collaborative, like, our team sort of is, is like one deployment surface, we're like a first party Google DeepMind deployment surface. We take the models, we bring them to the world, but actually the DeepMind team isn't just working with, with us, from a model perspective, they're working with Search and YouTube and Workspace and, you know, ten other product areas all across Google, and my, my good friend Tulsi Doshi Model team, she's got a very tough job, sort of not only building the model for all of our external customers, but working w- across Google with all the different internal product areas. And I think that engagement continues to get tighter and tighter, and now DeepMind is not just building a model, but actually we built a bunch of infrastructure that the rest of Google is built on top of, both model serving infrastructure. So if you wanna, if you're, you know, Google Search and you wanna bring, Gemini to production, you use this model serving infrastructure layer, that, Now the same actually for agents, and so we, we heard a lot about anti-gravity at I/O, but anti-gravity, there's sort of a, an agent harness that's actually powering the agent experiences, not just in DeepMind, but across Google, and it's powering a bunch of stuff, that I think we'll be announcing in the future as well. So it's, it's cool to see more than just the models actually coming out of, of DeepMind these days. It's, it's the harness, it's the infrastructure, which is
[04:29] Arman: really cool Of GPT three, kind of like move the focus a little bit more onto, B2C like chat interfaces. Do you see a change in the focus? Like, do you have to do more stuff for like the B2C people? Or how, how is it structured? How's the balance between keeping, you know, the work on the mission and also satisfying the needs for like direct products? People can put into their hands.
[04:59] Logan: Yeah, it's a great question. I think there's definitely tension around that. I think the DeepMind mission has always been to build AI responsibly and, and make it accessible to the world. And I feel like historically, I think for a lot of teams that started out, and, and actually not just DeepMind, but lots of other research labs who started out in this sort of like research era, it was like they were trying to build AI responsibly, they didn't yet know what sort of making it accessible to the world meant and, and actually how important the flywheel between-- it was like, it, it was sort of in the mission all along, staring you in the face, but like it was less, I think, obvious to folks that it was Truly an important part of the flywheel, and I think now if you look at like, I would, I would posit to say that like it's actually impossible to make that mission true if you don't actually go and like do it in the flywheel where like the models are being deployed to the products and you're getting feedback from real users and that feedback is influencing the direction that the product and the models are going and the different areas that we're exploring from a research perspective. So I think it actually is you need, you need both and I feel DeepMind is actually really well-positioned too, just'cause it, you know, you have so much legacy of, you know, Demis and the team and everyone sort of building, building this technology. then they've done a good job of like having the breadth, Demis says this all the time, like DeepMind has the best breadth and depth of, of talent and bets across the ecosystem, so we're not just doing Gemini. you know, there's, you know, tons of science initiatives, and actually Demis talked about some of those Forecasting hurricanes and, and forest fires and things like that, and that's just like a sliver of the science portfolio, there's so much more happening there. So it is really cool to see the balancing act of building Frontier Gemini models, and also everything else. And, and again, interestingly, it's like the Gemini model improvements actually fuel progress in all those other areas as well. Like we're using Gemini to sort of build the next generation of all these other products and models, so it feels, it continues to feel I feel like more and more and more central to the actual strategy is like the mod-- the main model flywheel has to work for all these other things to work.
[07:12] Arman: Do you have something in mind or can you maybe elaborate on this, on how improvements on Gemini, the Gemini I'm using in my pocket, is influencing something like Gemma Med or I don't remember, but the one, like another model that has nothing to do with, you know, GenAI and yeah, Med, MedGemma, MedGemma.
[07:29] Logan: Yeah, yeah. Well, I mean, and that, that's actually, that's maybe too easy of an example because, the Gemma models are, are fundamentally built on top of the same technology that powers Gemini. So we, we Models, we, we bring it to the open model community with our Gemma models. Gemma four just was released a few months ago, and I think the number is it's already passed like a hundred ten million downloads in the last few months, which is crazy. And the cool thing is for Gemma models, we can do all these fine-tuned variants so you can actually like explore what would it look like if you really specialized an open model on a bunch of medical data and sort of make it first-class with those use cases. So that example is maybe too obvious. I think maybe a, a harder It's a great example, like protein folding. I think the way that Gemini makes those products better is in the, the development of them itself. And so I think there's like two fold to this. It's like there's a product experience that sort of scaffolds the model. It's not just like some model weight living by itself. You need a whole product experience. AlphaFold actually has this. And so you, you know, a better Gemini model, if it's good at coding, if it's good at these things, you can sort of build a better interface around it. but then Greater to go build better versions of the model, and so they have not only like a really great thought partner, but they also have like a system that can go and execute a bunch of experiments on their behalf and kick these things off. and I think we'll see that sort of-- I think this is like the self-improvement flywheel that a lot of people talk about. I think it looks- Much more natural than like scientific, or not scientific, than science fiction in how it's described. It, it just looks like people using AI tools to do their job today. And I think that's, that's sort of how you see the self improvement flywheel actually happening.
[09:12] Arman: Do you see discrepancies in your team? on AI adoption, are there like some users like hardcore using it very heavily, token maxing, and others like a bit less, and how do you kind of balance?
[09:25] Logan: Yeah, it's a great question. I mean, there's, I mean, there's for sure a discrepancy. I think some of it is like what's their job. Like for some, for some roles, it's like much more obvious that there's like low-hanging fruit stuff where like if you're token maxing, there's actually alpha in doing that. For other people on my team, and my team's sort Some of it, like it's actually less obvious that, like, that would actually be really helpful, so it's, it's kind of fine that they're not doing it. I think the important thing is like, are they actually delivering on the stuff that needs to happen? And so if, if AI is a part of that story and can help accelerate that story, awesome. But like, we're not, we're not using AI for the sake of AI. Like, if, if truly it doesn't provide a, like, benefit for them doing their job, I'm like, I, I don Materials to like build a great product that people love and, and do it in a way, you know, with urgency and, and all that stuff. And so I try to keep that in center, center frame, like, and also in the way that we present our products, like the way we're presenting our products isn't like AI for the sake of AI, it's AI to solve the problems that humanity has. And I, I feel like it's, it's a nuanced- Line to walk, and I'm sure we don't always get it right, but like trying to continue to walk that nuance line.
[10:40] Arman: Talking about this, how do you set goals for your teams? Is it like OKRs or do you work on like ninety days or is it yearly plans? And in such a fast-moving ecosystem, how can you set objectives for your teams?
[10:51] Logan: Yeah, this is the hard part. And so I think we, we definitely have like top level Google level OKRs, which, which impact our teams, and our team is represented in a bunch of these top level Google OKRs. That, that sort of has been great for us to think about, like, if nothing changes, you know, here are the set of priorities. And, and actually, it's been most useful to me as sort of a reflection mechanism after that, that half of the year has happened, which is like, going into this six-month stretch, like, what did I think my priorities and the priorities for our team were going to be? And then there's a question of like, did we, as I look back to twenty twenty-four and twenty twenty-five, like, did we actually deliver those things? And you Shift of where the technology changed and, and where we ended up going instead. And so the actual like week-to-week planning process, sometimes like day-to-day planning process is actually becoming more and more important because the technology, everything is just moving so fast and we're, we're trying to look around the corner at the same time that we're trying to react to all the other things that everyone else is discovering. And so it's become, it's become harder and definitely like added tension. But I still really like that like half-year planning process, and we literally, we write out the list of the ten things To, to make progress on or, or fully accomplish in the next six months. And we so we, we do that main exercise like twice a year, and it, it is just like a very-- I love it as a reflection mechanism.
[12:16] Arman: Very interesting. Can you talk about the Demis' involvement in the teams? How does it work? Like when you have such a big company doing so many things, does he come sometimes to unlock certain situations or what's his role and how do you interact with him? And what about your teams? Do they interact and how, how does that work?
[12:33] Logan: Yeah, that A lot of, has a lot of demands on him as a, as a Nobel laureate and Sir Damesh Asabes, so I, I appreciate the, the, the time that we get to spend with him. He's remarkably leaned in as a leader, which I really apprec- I think there's like, lo- He, he has such an important role in, in sort of like telling this massive story of what's happening, like h-- him and the DeepMind team are the sort of like folks, the stewards of this original AGI story, and so I think he's, he's, he's trying to walk the line of like continuing to tell this AGI story and sort of steward the vision of that at the same time that like he's doing like real, useful, important work and like helping, to your point, like unlock a lot of stuff for our teams. I think he's definitely like clued He's super excited about, I think like World Models actually is a good example of this. He's, he's super involved in, in Omni and collaborating with Shlomi and Gabe and, and Dumi and those folks. He, he definitely supports AI Studio and he's leaned in on the Gemini app and sort of helping Josh and the team sort of unlock, and build this like, he's very passionate about this like universal personal assistant story, and I think that's been something that's like top of mind for him for, for many, many years now, and I think the I think he has this like very deep appreciation for the developer ecosystem, and so he's, you know, supporting our team and Anti-Gravity and others in order to help, like, f- from my perspective at least, and I'll, I'll try to get Demis, a direct quote on this, but I think my perspective is, you know, developers are in many ways like the manifestation of the mission. Google's never gonna be able to build products to solve everyone's problems, and so we need to bring our models and, and AI technology to developers so that they can actually, and Communities and, and at the global scale as well. And so I, I feel like it's, it's super important, and he's, he's made it happen for us, time and time again. At,
[14:28] Arman: at your level, what does a week look like? Like, you do a lot of meetings?
[14:32] Logan: A lot of meetings. And I'm trying to get, I'm trying to get out of meetings, and, you know, I'm always balancing this, I love doing IC work. I think like the thing that brings me the most joy in life is like Succeed, and so from that perspective, like working with, with the broader team is a ton of fun and seeing other people deliver, but like I think some of my most useful contributions can actually be doing the thing and delivering. And so on any given week, trying to find the balance of like lots of meetings, lots of planning that's happening, lots of actually like external engagement with the ecosystem,'cause we're, we're building a, a developer ecosystem, and also trying to do great IC work, you know, contributing to the documentation, writing code, planning features, and, and working with users And it gets more and more difficult over time to actually do all this stuff, so it's, it's a hard, it's a hard balancing act in order to, to find the time to do it all.
[15:29] Arman: I wanna get deeper in what happens in your meetings. Is it a lot of- the team comes, exposes like a situation, and then you all think together. When you get like five VPs in the same room, is it about like thinking together, brains together, finding a solution, or is it actually exploring, brainstorming? Just walk me through, like So I can project myself in one of your meetings.
[15:50] Logan: Yeah, yeah. I think there's such a wide spectrum of this. Like, I try to spend a lot of one-on-one time with the team, just like various folks across the team, my team, sort of a bunch of folks across our engineering organization, folks in research, et cetera, just to like stay plugged into, to what's happening and all those things. I think we do a ton of, we're, we're trying to get better about our like review culture, so we do these like informal but also like really All these like API design reviews where we're sort of going in detail about, you know, is this the right design for our developers and agents? And then there's a whole host of like cross-functional stuff around like, you know, working with our marketing teams to figure out like, you know, what's the right story that we should be telling? How do we sort of meet this part of the ecosystem where they are? How do we sort of make sure this other part of the ecosystem that sort of isn't yet thinking about these problems and solutions like start to think about it because we're, we're building Intentional time to do like brainstorming stuff. It happens very ad hoc, late night calls, early morning calls with our team in Europe. So lots more of that that needs to happen, you know, the, the water cooler chats as you sort of walk through the hallway and see someone and they, they make some suggestion. Yeah, there's, it feels like there's so much to do all the time and that only continues to get more, more fast.
[17:09] Arman: My expectation, I, I think your org is structured like a classic technical org, so you have like probably the CEO, then the Teams, managers, and ICs, et cetera. If you project yourself in the next five years, ten years, do you see this org radically evolve thanks to AI and because of what it changes?
[17:28] Logan: Yeah, it's a good question. I mean, I think there's like this potential ownership level that everyone has, which might increase, so you can imagine there's like many CEOs and there's like many things. That, that sort, that notion actually like kind of already exists in some like corporate hierarchies. There's this like GM model where like you're running the team And there's a bunch of, you know, things that are in your mandate, and you basically have like, you know, P&L responsibility for some product. So I think that like mini CEO responsibility sort of exists and, and sort of applies in, in some of our cases with AI Studio and another product that I run, Kaggle. So I think how radically is it going to change in five years, I think will be very interesting to see. I think a lot of this comes back to like comes back to like who the people are to do this stuff. And I think in a lot of cases, like, we are-- I'm, I'm fortunate enough that like our team has these like incredible specialists who, who can sort of like go really deep and solve these problems. You know, is that like the CEO profile? In some cases, maybe, yeah. in other cases, like, probably not. They're like, they have domain expertise in this one very specific area and sort of, when you wear the CEO hat, you wanna sort of have ten other things going on, and so The interesting thing is I have these conversations with folks, is we try to figure out like, how do we set everyone up for success in the next five years? It does feel like there ends up with this like choose your own adventure type of setup, which is very, very interesting. I think it adds lots of tension for people managers, like as I try to figure out like, how do I set my team up for success? You know, choose your own adventure is great in some ways and obviously has a lot of obvious downsides. And so trying to like find the balance of giving people autonomy and You also wanna make it clear like what is actual-- what does success look like, and if you have too much autonomy and not the right like checks and balances in place, like, you know, things, things could go wrong, and so I think, yeah, lots of like Lots of fun organizational stuff to figure out over the next few years. I don't think we have the answers, but we're definitely having a bunch of the conversations right now.
[19:31] Arman: Is there something you find very interesting in the way DeepMind is organized that you think might be unique to DeepMind?
[19:38] Logan: It's a good question I think a-actually, the, the, the one thing that continues to resonate with me, which I don't-- I mean, this is, this is unique to, to the maybe the model labs in general, but the-- and I think maybe to a different degree depending on which model lab it is, but the fact that like research has so much collaboration with the product and engineering teams, I think is fascinating, and I think the-- it's the thing that I've appreciated the most about being at DeepMind is like the last, even actually before we were in DeepMind, when I, I joined And was in, worked super closely with the Google Labs team because they, they were the ones who created AI Studio originally and sort of took Google Labs, or AI Studio out of Google Labs into Google Cloud, spent nine months at Google Cloud, worked with the cloud team to help scale up a bunch of stuff, and then came over to DeepMind early twenty twenty-five. All throughout that process, like having this incredibly deep collaboration with, with DeepMind and with the research teams, and I feel like that, you know, there's, there's so little friction in that. I think it's easy for A super secret thing, and we don't wanna talk to anybody else, and sort of it's, you know, it's all, all this magic, and we can't, you know, we wanna sort of go in our, in our research ivory tower and, and sort of not collaborate with the products. That, that is like couldn't be the opposite of what's happening in DeepMind. It really does feel like the researchers like, they wanna work with the product teams, they wanna bring their, their ideas to life through the products, they sort of wanna share ideas and get feedback early on. And DeepMind, which really does feel unique to, to sort of the organization that, that Demis and Or, Oriol and Jeff and NOME and, and Core i's credit, have, have really fostered and created.
[21:21] Arman: If I give you a blank page and I say, okay, Logan, you have to reimagine the way your team is organized in the AI era, do you have like an idea or can you just think of like how you would do? Would you do the same org or would you do like the mini CEOs or like the GM style thing you
[21:38] Logan: mentioned? Sort of somewhat set up in this way already, like al-almost all the folks who report to me are members of the technical staff now, and sort of I think that, that sort of, it, the, the thing that it means is that the expectation is that you're, you're sort of pushing the frontier of what's possible with AI, in, in the way that you sort of work. And I think it doesn't, it doesn't mean anything more than that today. I think in the future maybe it will as sort of the expectations change, as the tools change, as I wouldn't, I wouldn't make any significant radical changes to how the team is set up. I think it will be interesting though as like the, as the product shape. I have this conversation with, with folks all the time now, which is sort of future-looking, in five years, will Google have, you know, ten thousand products or will we have three products? and I think it's very interesting to think about like, you know, actually that team structure really depends on sort of like which version of the future it is, and it's prob- The hard part is, the answer is probably somewhere in between there, and that, that will make all the other things more difficult. But like, if I was building for the future of Google with ten thousand products, maybe the like single person running a product does make sense. If I'm building for the world where there's actually three products, like actually the structure you want is probably looks a lot different in that world than it does in the world where you have ten thousand products. And I think it's It, it isn't obvious to me right now which of those two is going to be. There's obviously this huge push for super apps, at the same time that the cost to build products continues to go down in general. So it's an interesting, it's an interesting world to live in.
[23:13] Arman: With AI Studio, you're giving people such an amazing tool, they can build anything, it's also so powerful. How do you think about like the, the risks, of like, let's say someone creates a hub where they put all their, health data? Yeah. Do you Security or like the backend, et cetera, how do you integrate those questions into the process of like building the, the tool or building AI, as you
[23:39] Logan: say? Yeah, this, this, I think is the most important problem that, that we as an ecosystem need to solve. I think Google will, will have our sort of own takes on how to solve this problem, but I think people fundamentally don't understand the level of risk that they're taking. and I think part of our responsibility as, as a product sort of helping people build whatever they want and bringing, bringing their Our idea to life is that they need to get calibrated on the level of risk that they're taking, and w-we can make a bunch of like reasonable decisions from a security, authentication, et cetera, side to begin with, you know, maybe those decisions aren't actually the right ones that the user wants to take. They wanna take risk, or they're completely conservative and they don't wanna take any risk. And so I think, we're, we're, we're like very actively having these conversations across the team to make sure that we, we get users calibrated on the right That they're actually willing to take on. It's an interesting thing to try to communicate some of these problems to people who aren't technical. I think that's the real challenge, is like to the person who's like I'm putting my data in here, I don't understand how it would ever become available to anyone, like it's just-- it doesn't make sense to me, it's not obvious, like they're, they're just not calibrated on the risk. So I think we have more work to do, but it's, it's definitely something that's top of mind because we, we know people wanna build things with, with their data, with other people's data even, and so doing it right matters a lot.
[25:02] Arman: What's the biggest challenge or problem you have as a lead that- Cracks your head, you know, like you're like, "I need to crack this, I need to get this right." Or is it just like very peaceful and everything's going well?
[25:15] Logan: I think, and this isn't, unique to my role, I think this is something that a lot of folks who are building products right now are thinking about, but essentially like everything, it feels like we can do anything, you know, especially, you know, Google, we have resources, we have, you know, users, we have a bunch of interesting infrastructure, there's so much we could do About these days is like, what are the things that we can do, as a team that are unique to, to Google, and feel like they really sort of n-not only solve the, the problems that our customers have, but are solving them in a way that only Google could solve those problems. And, and I think there's lots of, you know, vibe coding is a sort of a quintessential example of this. There's so many companies doing vibe coding, and I, I try not to think about the sort of-- obviously, I use a bunch of the competitor products'cause An incredible team's building those products, and, I think there's lots of people doing a good job. But I almost, I, I'm not super worried, and I don't think about the sort of competitors in that sense much, because I think we're solving a fundam-- I think we're solving a problem in, we're solving that problem of the ability to create software and bring it to the world in a way that only Google can solve it. so in, in a sense, it's-- there's this, there's this great Naval quote, which I also Authenticity. And this applies to building products, this applies to like how you operate as an individual in the world. And so I think about this for myself, but I also think about this for AI Studio and the way that we build products, and I want to escape competition through authenticity because we're really building a product that only Google can build. And there's like two recent examples of this that I'll give you, which is the ability to build native Android apps that we just landed in AI Studio and the workspace integration that we have. the way that we went about building those integrations is building Google can build because actually the Firebase team, and because the Android team, and because the Workspace team actually sit with us inside of Google, and so if we can find ways to work with them that like don't exist, and actually hopefully will exist for other products,'cause we'll, we'll sort of blaze the trail to, to make some of these things possible. But I think, I think about that, that intersection of what only we could build inside of Google a lot.
[27:30] Arman: Such a big question, and I think Gemini Spark was a great example of how you can leverage the whole ecosystem that people have been building for so long.
[27:37] Logan: Exactly.
[27:37] Arman: All, all the things happening with search also are fantastic. Twenty seconds left. Yeah. Can you just shoot me things that you find interesting, books, recommendations, movies, whatever? And I'm talking to Logan, I'm not talking to AI, Logan, right? Yeah, that's a good question. You mentioned Naval.
[27:53] Logan: Yeah, I try to listen. I mean, Naval is obviously timeless. He has, he has a new, podcast episode which is on my, which is on my to watch list. I spend a lot of time listening to Lenny's podcast. I think he does an incredible job of interviewing some of the folks like building the most interesting frontier products and companies. So, he would be my other suggestion. I think you cover sort of philosophy with Naval and the way to live your life, and then you cover sort of like And sort of AI product building, which is, very top of mind.
[28:23] Arman: I'm very grateful for your time. Thank you so much. This was awesome.