Sundar Pichai: Path To Recursive Self-Improvement, Compute Allocation Dilemma, Future AI Interfaces
Publicado no IA Dots ·
Agentes de IAAGIAI Signaldot-artificial_intelligencedot-googledot-llmdot-openaiDotsidioma-enInglêsMuseyt-F3ItSTZNtT0
Vídeo em Inglês, publicado pelo canal Alex Kantrowitz.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
Alphabet CEO Sundar Pachai discusses Google's approach to artificial intelligence, the development and release of the Gemini Argon model, the balance between frontier AI research and cloud infrastructure, and the future of agentic workflows and interfaces.
Key points
- Google launched the Gemini Argon model, achieving strong benchmark results across coding, math, science, and knowledge work.
- Pachai explains the internal balance of allocating computing resources between the cloud business and cutting-edge AI research.
- The video highlights the growing importance of agentic workflows and multi-agent frameworks for both consumers and enterprises.
- Recursive self-improvement and defense-in-depth safety measures are discussed as critical elements in current and future AI model development.
- Google is pursuing both large language models and world models to improve physical understanding and robotics applications.
Topics
Artificial IntelligenceGemini ArgonCompute AllocationAgentic WorkflowsRecursive Self-ImprovementWorld ModelsCloud Infrastructure
Video transcript
Transcript from available captions; it may contain errors.
We are at early stages of the most consequential technology ever. This is not like fiber. It's not dark. We are using it. AI researchers for them it's like the bigger the computer, the better the AI. You give the researchers the biggest possible computer even if it is to the detriment of today's profits like the cloud business. So the compute we have today is a function of decisions I made in 2024. Today obviously when you sit and look at anything everyone is short not just us other companies and everyone wishes you had 2x of what you would planned for where's Google's muse or it's dot's product users both in consumer and enterprise will have access to kind of some multi- aent framework for them working all the time doing things and so you know it feels like a moment of infinite possibilities Alphabet CEO Sundar Pachai joins us to talk about Google's plan to compete on the frontier of AI, the balance between cutting edge AI research and the company's cloud business and what the future of the interface looks like. And Sundar is here with us in studio in Mountain View today. Sundar, great to see you. Welcome to the show. >> Thanks, Alex. Great to be here. >> So Google is back on the frontier with the Gemini 4 Argon. In plain speak, that means it is a worldleading model. It's benchmark breaking. It beats open AI and anthropics uh public models on a variety of benchmarks including on coding, math, science and knowledge work. Now the conventional wisdom had been that Google had basically given up on the frontier that it was going to make these fast models that were good enough and use them in its products and it wasn't pursuing frontier any AI anymore. But that was wrong. >> I mean very wrong. Look, I I I think we are so committed to being at the frontier. I think it's in our uh DNA. Uh we wouldn't do it any other way. Uh we've been thinking about the AI frontier for uh well over a decade now. We definitely uh the frontier is very dynamic. It's very uh competitive. Uh it's very exciting. Uh I couldn't be more pleased with uh how uh Argon has come together. We took our time with it. Uh you know we we made sure we tested it safely and deployed it uh uh carefully. Uh we took a bit longer than our other models given given its capabilities. But I think across the board be it for agentic work, knowledge work, coding, uh creative tasks, uh I think Argon is uh you know one one of the uh really good frontier models and you know in the first 72 hours since we launched it's you know almost up to 25 trillion tokens used which is about 3x uh what Gemini uh 3.0 had in the first 72 hours. So uh and we are capacity constrained as we are serving it. So uh excited uh but there are always areas you know both uh the frontier is just last uh it's very dynamic and you have to keep evolving the models and there are areas which we can see we need to improve as well. So we'll be at it at work on all of that as well. >> Which areas? Okay. Um, on any given model, you know, Argon will be across the board better at everything if you just look even two months back. But the space is so evolving. Um, but even on coding for example, there are there are areas pockets where we we know we need to do better uh and you know there's clear feedback and so we'll we'll make it better. Let me at least make the argument for why standard models might be good enough for Google. Yeah. I'm going to channel what people say about this. Yeah. >> Um and that is that the frontier of AI today um is is getting so good uh and the baseline of AI is getting so good that the frontier is really only valuable for a specific set of tasks. And if you have a standard model, you could actually build great products with a standard model. Take Metamuse for example. metab muse on um a good model but not the best models. >> You know if you need to do like biological research um maybe the frontier is really necessary but the people who would say Google doesn't really need the frontier they would say you just need those standard models and you can put them in products like maps and Gmail and and search and um you know again take that meta example that would be sufficient. What do you think about that? Look, I it's definitely not how I have thought about uh thought about it at all. Uh I think we are deeply committed to driving the state-of-the-art and actually push what this technology can go where it can go uh all the way from transformers. That's what that's the journey we've been on. So you will see us continue to do that. I think when you when you're at the frontier, you get an earlier look into the kind of uh emerging capabilities earlier than others, which helps you better prepare for the future. We are fortunate to have a wide set of opportunities across consumer and enterprise. Some of our cloud customers for sure want the model at the frontier. I think the frontier models will be better at developing the next set of models in a in a in an era of uh recursive self-improvement. So I would argue it matters there. Internally we have definitely seen a steep point of inflection by deploying argon to our uh developers. It definitely tackles problems which our previous models could not do. And so being at that frontier puts you at the innovation frontier and and uh helps you drive that as well. There is room for models of all sizes and capabilities. We've always that's why we did flash flash light flash etc. And I agree with you that you can develop very good product experiences with uh with not necessarily the the most frontier model because they tend to be slower and more expensive too. So I think I think but I we care about the parade of frontier and and uh so we are committed to doing all of that. Well >> you said that developing the frontier gives you a look into what's coming. Yeah. >> What do you see on the horizon? Uh you know the exciting what makes this uh moment so exciting is you you you you build something which you think is a culmination of like months of work and you put it out you know you already see clear headroom ahead right I think I mean the frontier is going to we we've all been working on these agentic long horizon capabilities but clearly the frontier is going to have multiple of these agents working together to do things right. And and so that's an example of an exciting uh paradigm uh underway. And and it's also clear to me that users both in consumer and enterprise will have access to kind of some multi- aent framework for them working all the time doing tanks. And so you know it feels like a moment of infinite possibilities. You mentioned uh recursive self-improvement. This is the models effectively making themselves better. >> Yeah. >> Maybe without humans in the loop. Um people call it like an intelligence explosion or expect to hear this sound like foom and then all of a sudden AI is just like this superhuman uh capability. And in fact, uh I think Sergey was um doing this round table with researchers at Deep Mind and he talked about with Argon um that the core uh one of the core innovations for Argon was relying more on the model to help develop the model, right? Um we've seen what's gone on with AI models recently. They are not they are well aligned, but they're not perfectly aligned and the imperfections are a big problem. Um, obviously the tech world and now the world more broadly knows about the hugging face incident where open AI's bots uh broke containment and hacked into hugging face and did some crazy stuff. Google uh has also had a scenario where its bots uh got access. I think they were guessing passwords and they got access to a site they shouldn't have. They stopped. But these models, there are there's enough evidence that they're not doing exactly what we want them to be doing that this idea of pursuing this recursive self-improvement or even um making the models better much faster uh is it's sort of freaky. It it makes me scared. Um, don't you think uh there's reasonable worry here and there should be something done to um have the tech industry chill out about this? >> Well, maybe maybe three parts to your question. Let's first let's talk about RSI itself and you know you're right Sergey mentioned um uh you know RSI being you know critical contributor as we are developing argon and but to be very clear today you know in a continuum of we're pretty far from uh you know the end state of RSI or something like that right so today you have the the right way to describe it is you have RSI loops along with AI researchers in various aspects of model development, right? So, you know, I would say you almost have agents helping AI researchers uh in in some automated way uh work together to build these models. So, I think I think it's good to at least understand where we are today and it's super exciting and I expect that continuum to keep making progress as we go into next year. Uh the second question is around safety. Look, I the way we are I mean it's a safet security comes from defense and depth. You're talking about alignment. Obviously, you know, we all are doing a lot of work to make sure the models are uh as aligned as possible, but it's an active area of research and there are clear gaps, but that's not the only thing we rely on, right? So, you have the work you do on alignment, but beyond that, you know, we we put a lot of effort into uh real time monitoring as these models are working. So making sure you have good telemetry into what's happening and so that you can intervene in real time if they if you see unexpected behaviors and how you do that uh the architecture with which you do that how do you make sure it's all sandboxed and isolated in a way that it can continue doing its job during real-time monitoring. These are all areas of investment we are doing uh uh doing and obviously when we deploy these models giving customers enterprises control both with real-time monitoring but their ability to decide where they deploy what type of data these models ever have access to etc. So all that work is very very important and we are doing a lot of that with uh with our all. So I think as an industry we need to constantly up our game and take the time to do all this well including dedicating the compute required to to make sure you can safely deploy these models. The third part is you know is it all enough and you know so I think I think this is why I think you know there's a lot of conversations in industry and I think it's reasonable to be worried but you know I won't right way to approach uh something you worry about it is early on thinking about what more we can do and I think all of us are you you know I think as an industry I think you're definitely uh seeing an increased levels of work on alignment uh testing for safety etc. As we deploy these models I think the next step would be developing better standards so there are crossindustry standards of how we do these things including ability to uh evaluate using independent evaluators. You know if you look at areas like financial controls we have like accounting standards we have independent auditors etc. Right? So those are all the next natural steps for us as an industry to work towards this and regulation will end up playing a role as we go through all of that as well. Okay. Technology question for you. Um Google is somewhat unique in that you research the language or the LLM side of the equation which started with you know the transformer language. Uh and then you also pursue world models and you know there are some AI research efforts out there like open AI that have basically determined that the LLM is the path to AGI and they've actually um subset part of their world model research which is like the AI model that tries to understand the physical world. Uh but you're pursuing both. So talk a little bit about why you make why you find the world model side to be important. you know, I mean, uh, as humans, you obviously perceive the world around you and, you know, you're taking all those inputs. Um, you know, making sure these models are able to understand the laws of physics, uh, simulate the world and high fidelity, all that I think are key enabling blocks for intelligence. And, and I think as you're working, we working across all possible applications of this. uh you know take an area like robotics it's not an accident argon if you look at uh you know embodied reasoning benchmarks argon is state-of-the-art right I think literally sought on many of these benchmarks >> and that all comes from you know our experience building world models as well so >> inherently I would expect you know there will be architectures which will converge all of this in the future but I think it's for us we view this is today um you know in this jagged frontier that is intelligence we want to be at the frontier of all the all the possible approaches and I think that will uh help us both drive better research better insights and develop a much better uh you know AGI model uh you know is how we think about it. Yeah, I looked at your your Twitter after the Argon uh announcement and it's all just like real world applications. It's not like text or uh you know a model generating text. Okay. I want to talk to you about the way that you balance uh allocating uh compute which is your most precious resource between cloud and AI research. This is a big topic about Google. Um, I've had people suggest to me that Google should uh get out of the AI research game and um, you know, take all of its cloud computing resources and its chips and just sell that. And you look at the world of business right now, Nvidia, which has like doing that effectively uh, $5 trillion company. Um there's a push and pull between cloud that wants to you know sell these the this computing resource at a high markup uh and AI research that needs a very big computer to make the best AI. Now you've said a big part of your job is figuring out how to allocate compute. Um how do you do this and how do you balance these two? Look, I think we are um look the ground zero for us is making sure we are investing in the compute that's needed to be at the AI frontier. So that is you know in 2015 I set the company to be AI first. So AI is that key horizontal ingredient across everything we do all the way from search to cloud to way more to everything we do. So I think that is ground zero for us. Uh making sure both for research and developing the models that we have enough compute needed. Beyond that, look, we have these independent businesses which each have their own long range plans and so they they're using those long-range plans to project forward and we plan for compute for all of them. And so the cloud team, you know, they have a business and they have a projection and they have demand signals. They've committed contracts. So with that help, we can plan on a long-term compute trajectory for them. Um, I think we're fortunate to have this wide set of opportunities. I think there's no company in the world which isn't making trade-offs at any given moment in time. But the first principles we operate on is to making sure we have enough compute to invest in the frontier. And that's non-negotiable and everything else works uh works on top of that, >> right? >> I think what you're seeing in these conversations is if you sit now remember we are building our own compute, our own data centers. So there's long lead times. So the compute we have today is a function of decisions I made in 2024. Even if you think back to a year from today, last year the agent recording takeoff hadn't happened. So today obviously when you sit and look at anything, everyone is short, not just us, other companies and everyone wishes you had 2x of what you had planned for. I think that comes with the territory, right? And like you know at any given moment in time we are planning for what we would need 2 to 3 years out in the future. So I think I think you're going to find some decisions you make may look suboptimal 2 years out right depending on where the state of the art is. So so that you know I don't spend too much time thinking about it. The way I think about comput allocation is making sure Google deep mind both for research and frontier model development is at is cutting edge in terms of compute access right and after that we plan for all our businesses I think it's great we can provide our compute to other people too like we have we have done it allows us to scale that up get real test for it in the widest application set possible you know test it in the real world and make sure we are at the cutting edge of techn technology at every layer of our AI stack as well. >> Um, you know, briefly there was a a podcast where u Ben Horowitz relayed that Andy Jasse had said something like computer sold out till 2028. Yeah. Are you seeing the same thing? >> Um, you know, I think it's a yeah, it's a reasonable like you know if if today a customer is coming out and asking us for compute in uh 27 that is uh incredibly difficult to do. um you know there there's uh part of the reason even last year we went out and leased some short-term compute was to accommodate some customers and help them out uh but yes yeah we I think we are all dealing with similar similar timelines >> um one bubble question >> uh there's talk of how every technology shift has seen overbuilding >> um railroads fiber >> um do you think there's overbuilding here >> it depends on the time frames you're looking at. uh are we you know this level of build you know if I take a long enough time frame it makes complete sense to me we are at early stages of the most consequential technology ever and this is not like fiber it's not dark we are using it we are capacity constraint on a daily basis um the applications the underlying technology is improving rapidly right but running a company I the way I work through the moment is you look at your demand signals. You look at your uh like committed revenue opportunities that you're getting in the forms of contracts. You balance that with like the underlying technical capability are internally eval how we are using it internally the model trajectories and based on all that you take a disciplined approach a confident aggressive but you know a disciplined approach to plan out plan out compute. But when you talk about the broader economy as a whole, you're you're h having people operate with very various levels of uh uh risk and and so you know so it's tough for me to I focus on what I can control as a company and and plan through it. If your question is is the can there be an or build? Yes, because you know you can have some constrained components and remember what matters is a square set of everything you need to build out that compute. >> Memory, electricity, electricians. >> Yeah, like >> there are like 2,000 compliments or more, right? And so so you could be you could get constrained at some point. There could be things which cause bottlenecks and you know and then you could have uh over bills. There could be external shocks right to the economy etc. So there are macro risks uh all all all that is possible at any given time but you know you am I confident that the compute we are planning for we will have uses for it given enough time absolutely yes >> okay one more on this um you know if you're going to do both uh there's been a line of thinking out there that uh AI researchers for them it's like the bigger the computer the better the AI And if you know you're someone I was in the room when you said that AI would be as profound as fire and electricity and uh we all laughed at the time and I kind of regret that laugh now. Um but if you're someone that actually believes that then the thinking would be um you give the researchers the biggest possible computer even if it is to the detriment of today's profits like the cloud business. Your thoughts? We're all doing some form of that, right? I think uh our cloud business is at any given moment in time foregoing short-term revenues because we are making compute tradeoffs to support frontier AI development. So I think what you're saying is is is true already, right? And and uh but you you know you want to be thoughtful, right? even our AI researchers um you know some constraints are also good that's what inspires creativity you know that's what helps you come with a more efficient algorithmic breakthrough maybe to use less memory etc so you know so uh what we are committed is uh to invest what's needed to stay at the frontier right that's how I think about it >> I wonder what you think the future of the interface looks like the way that we use computing in the future as AI starts to roll out. I was speaking with Greg Brockman, the president of OpenAI about this a couple months ago. He had a very interesting quote. Uh he said, "What is the interface that you want? What is the product that you want?" The answer is that you want no interface. You want no product. You want this to be what's the interface between you and me? Basically saying just talking. that if AI is that smart, there shouldn't be like a, you know, maybe a screen or just so many buttons and things that you have to touch. Um, what is your view of what the interface looks like in the age of AI? >> The way I think about it is I do feel, you know, we computing historically has been humans adapting to computing. With AI, it'll be the other way about the technology will adapt seamlessly to how humans live and go about their day-to-day lives. So, there's an aspect of it. Um, you know, you want that AI to be able to speak to it when you want it to. It has context. It can, uh, be persistently running somewhere, give you continuity, you're not dealing with a lot of complexity. It's natural. Uh, all that makes sense to me. That doesn't mean it won't have an interface, however, to me. Um look today for example I mean nowadays I just even in my phone like you know the Gemini voice models are so good I can just press press the microphone button I can just ramble that's how we call it rambler right I can ramble it kind of intelligently corrects with my personal context so even my friends names like weird names it kind of gets it right even if it first guesses Is it wrong? It's amazing. It's adapt. I'm not going and backspace editing once once I finish saying it, right? It's remarkable how good it's gotten. But there are going to be moments you will need an interface. You know, you may want to be entertain yourself. You may want to watch a video like, you know, so you may ask your AI to create something and you'll want to watch it, consume it, right? So I think the way I think about it is you want it to be as seamless and as natural as possible and that you are not doing the hard work adapting to whatever the interface is. The technology is adapting to how you're working at any given moment in time. >> Yeah. I'm just going to um play the OpenAI hits. Yeah. Because um you know last year you sent them into a code red. Yeah. >> I'm sure you know um Gemini uh was doing very well in OpenAI batten down the hatches. Um, and I was speaking with Sam Alman about the challenge from Google. What he said was he thought OpenAI would fend you off because they're building AI native and I don't know if he said it specifically like this, but he implied Google is bolting on AI to existing products. What is your view of whether um, you know, a bolt-on, I wonder if you would accept that framing uh, can compete with an AI native product? Look, first of all, uh I mean it's a very dynamic moment. We are all pushing each other. I think OpenAI has pushed us hard as a company too. So this is not just uh you know we we are doing something to them alone etc. far from it. >> Yeah, you went into code red too. I mean I know it's disputed about whether it was but I mean it's been a the AI moment all of us had to respond with varying levels of urgency at any given moment in time. Look, I um I think you know we're constantly building new products as well. I mean Gemini was built from the ground up like like as a product the Gemini app I'm talking about >> or you know tomorrow as all of us are adapting to agents are you bolting agents onto chat GPD or are they building it natively? How would you think about that? Right. So you can always flip these questions back. >> Right. >> Right. So I don't think I don't think of it that way. I mean I think of it as you want to build worldclass product experiences uh that are AI native and they are winning and uh and some of it is adapting product some of it is rethinking new experiences so they will all be part of how you approach this moment talk a little bit about the way that you see the business of this AI interface in the future so um for instance if I'm asking my agent to go out and do something for me and um it suggests a product should it I I mean maybe even if disclosed does it get a cut of the sale if I make the sale or when you think about the way that you're going to make money from these products what's the plan >> I think you know obviously I'm going to put enterprise aside I think yeah consumer question so you know if it's consumer >> look I think there are consumers are going to increasingly value this technology they're going to meaningfully I mean take software developers today I think developers are willing to pay for these coding models right and because it really helps makes them more productive nobody would want to code the way they were coding 3 years ago or even like same time last year so I think when you're delivering value there's value to be had I think history of consumer product shows if you build a consumer product I've never seen examples where consumer products really scale to billions of users where people are using it intensely and hasn't had monetization to go along with it. If you're providing value to people, there is inherent value to be had. >> Totally. >> And so the nature of it may change, right? I think you're asking a scenario in which agents are doing stuff for you. At the end of the day, you will have commercial needs. The agents need to fulfill those commercial needs. the providers on the other side are going to see value in that and there's going to be some mechanism by which it gets fleshed out. That's what an efficient uh system does. But I think what matters is that you're building those world-class consumer experiences, right? And so um am I answering your question? I just want to make sure. >> Yeah, definitely. >> Yeah. >> I promised myself one product feature request and one only. >> Yes. >> Here it goes. Um, is it possible to make search in Gmail powered by Gemini as good as search in Gmail that you access when you plug Gmail into your chat GPT interface? What's going on here? Um I think you know we we have one of the exciting things we announced recently is in a unified Gemini agent uh for enterprise which we we have announced and we're bringing that to consumers right so these are uh persistent agents which run 24/7 in the cloud like you're spinning up a microVM effectively for every query Um, and you know from the runtime, the harness, the model, it's all a headless service, right? And it'll work everywhere and it it has your context with you. >> Mhm. >> You can address this agent whether you're in workspace, in search, you know, in Gemini, all of it. So I think agents are going to drive this unification from a customer experience which which people are looking for. >> Yeah. And you are right when we when you have a broad product portfolio and you're moving forward fast there are always areas where you're moving forward faster than some other areas that that's that's a valid feedback right and uh but we are putting a lot of time and effort into making sure as we build out our agentic experience that unification and that capability works well across all these surfaces. >> Okay, that's a good answer to the interface question too. >> Yes, it is. Yeah. Uh when people have thought about where Google uh plays next, a question comes up. You know, where's Google's Muse or its DOT's product? Uh those have kind of come out of nowhere recently and uh they've taken off. You have a version of that? Um do you do you anticipate creating a new product that competes with let's just say Muse for instance? Um look uh we already today so the way we have envisioned is in fact we we we describe Gemini our goal with Gemini app to be a personal proactive and powerful assistant right that's what we are working towards >> and you saw us describe the early versions of it at Google IO um uh Google IO this year with Spark >> but effectively already you know if you open the Gemini app as a or an ultra user, you see chat and task. >> Mhm. >> And and and and they start giving you that agentic experience with Argon. Now that capability is dramatically accrated and what we announced recently in our enterprise event, we are bringing that to consumers as well. But over time so you will soon in Gemini uh you will be able to ask something and that will it's a continuum you don't need to think about whether you're asking a chat or a uh agentic task it'll just kind of automatically uh do the right thing and it can work persistently on the background securely and get things done for you already empirically when I look at what's coming into Gemini in our app about last year you would have tagged them and most of them would have been chats but you know a steadily growing percentage is already agentic tasks in Gemini and so >> so yes we are going to bring um you know personal agents to our consumers it'll work in the Gemini app it'll work in Google search it'll work in Gmail you can you can address it call it etc but we will you know we we want to make sure we're getting it right it'll be powered by Argon and so we very excited to uh bring it to our users. >> I've got that pro subscription I've been tasked. So, I'm looking forward to using it more. U all right, couple of random questions. Uh Google's committed and invest invested and has committed um tens of billions of dollars uh to Anthropic. Uh what do you see in Daario that makes you want to have such a deep commitment and investment to him? Very early on when we looked at the founding team of Anthropic, it was very clear to us um you know this was a world-class team which was involved in some of the earliest GPD models and so we were quickly able to assess that you know they they would be at the frontier and so we've always tried to identify uh and partner with companies which would be at the cutting edge of technology. We've done that with SpaceX. We've done that with Anthropic. And uh you know that's what we saw and uh we had a chance to partner some of the partnerships turned in the construct of our uh TPUs as well. So I I so I think that that's how we approached it. >> Okay. With with Apple, you had a partnership with them to license Gemini to them uh to help improve Siri. Is that partnership expanding, contracting, standing still? I mean we are uh uh look both for both us and Apple that partnership was about translating the capabilities of AI into better experiences for uh for for users and uh you know I thought Apple has done a a phenomenal job of incorporating they customize these models and build it into uh Siri and uh it's been great to see the reception and uh so we are committed to doing more. you're going to do more with them. >> There's nothing specific to comment, but you know, I mean, it's a it's a partnership that's working well. >> U data centers in space. Uh you've announced that you have uh you're doing more there. What have you learned so far about the feasibility of that project? >> I mean, it's a it's a very exciting project. You know, projects like this remind me, I still recall that moment in uh in the desert uh for the DARPA Grand Challenge, watching the Stanley car uh you know, compete in it, which ended up being the basis for Whimo. Uh right. And so you're always with all these technologies, you're you're at a moment in time where it looks a bit like sci-fi, but you know, 10 to 20 years down the line, it's actually working uh in a meaningful way. I feel the same with space. You know, we've always believed in betting big on these technologies. Quantum has been a great example of it. Uh we probably 10 years into the journey. Way more we are about almost 20 years into the journey space. We are two to three years into the journey, right? And uh it's it's clear to me that it it's an approach which will make a lot of sense in the future. uh being in sun-synchronous orbits with uh univers 24/7 access to the most abundant form of energy we have access to will make a lot of sense for this exponential increase in compute that we are going to need but having said that I mean this is a challenging project the team has disciplined milestones and so it was a super exciting milestone to send the TPUs up to space and sometime as we are speaking we are expecting a hello world moment uh from them and so I look forward to uh you know to seeing that happen and we'll continue investing here too. >> You've you recently was it maybe last year you said that 25% of code uh written within Google was written by AI. What's the number today? I mean we have far I think we have almost that number in our world today is like meaningless in the sense that every part of code we now touch is uh influenced by AI deeply what's now what we are ex we are tracking with argon for example we are now able to see it's been a point of inflection people are using argon to really go tackle complex projects right and including projects which affects millions of lines of code. So we have definitely in the four weeks since we've launched Argon we've seen a 30% increase in uh sales and and and more importantly uh almost a similar increase you know one-third uh one-/ird decrease in like roll backs right and so I think people are writing code more productively and better code with it uh code with it so I think uh I think so we are tracking it with newer metrics now >> um you know, talking about the opposition to data centers and how do you feel that? Um, and if you can't build data centers in the US the way that you want to, where else do you go? We're, you know, we're very committed to building our data centers very responsibly. I think if when I look at our largest data centers in some of the communities where we have been the longest uh you know I I visit them I talk to people >> you there are people who who would come and tell us their kids have gone to college because of our investments in these communities and so and we've grown in those communities and so I think to me this is about doing it right you know being engaging with the communities early doing it transparently we are committed to making where we are bringing incremental power and we cover 100% the cost of our power needs uh for these projects. It's understanding the impact on the on neighborhoods and making sure you're working early to address them etc. So I think done rightly uh you know I think I think you can continue making progress but you know we this is obviously we are constantly taking feedback into account and and and so on. I do worry not everyone operates uh the same way. >> Yeah, people have been surprised by data centers showing up in their backyard. >> So I think I think those are all important to uh as an industry we should continue to do better. >> But part of the reason for in a 20-y year time frame we are thinking about things like data centers in space is making sure we you know we are able to put these data centers where you know where maybe the trade-offs will be different. Is there an unproven discipline in AI right now that Google's working hardest to crack? So people have talked about continuous learning, getting these AIs to continue to update after they get released as opposed to forget through every session. Look, there are multiple parallel bets. I mean, you saw with Argon, we have a million output tokens, right? You know, you know, with our Gemini models early on, we had a 1 million context window on the input side, right? These are all uh examples of breakthroughs we are constantly working on. You know, how do you think about synthetic data? How do you think about continual learning? There are uh how do you think about more efficient architectures? So part of we do dedicate a meaningful percentage of our compute to these research bets too. And so it's always uh >> uh you know it's always goes hand in hand including our world models etc. right? So those are all bets we are uh uh placing on sometimes as part of these bets there are ideas that emerge which just make their way into the mainline models and so maybe you know uh you don't uh uh uh you know fully fully uh talk about it but you know I I think we have multiple bits in these areas. >> Okay. Okay, I want to end with this. Um, to be involved in the tech industry the way that you are, there's a there's a certain inherent optimism >> to it, right? A belief that uh if you build, you will improve the world. >> And this is, you know, something deep within the tech industry. Now, there are sides to that, right? If you it requires betting oftenimes like betting a lot in hoping that the thing that you're working on is going to come to fruition uh in the way that you want it to be. Uh but it doesn't always work that way. I'll give you an example. Um I was just visiting El Salvador for the first well I I went in a 5-year bracket and I saw um in this one town uh just a a amazing amount of uh enthusiasm for Bitcoin and the blockchain and there was a construction happening and it was booming and then it didn't happen the way that that folks expected. Uh and this visit was basically a ghost town with a lot of unfinished construction projects. Um, I think that that to to do what you do there, it requires, you know, a type of optimism that says it's worth going for, you know, even if um the bets don't pan out and uh and there could be consequences there. So, I just love to hear your perspective on you spoke about this at your Stanford commencement address about the need for optimism. Um, but it's a nuanced form of optimism, right? So I'd love to hear your perspective on um sort of how you maintain that and and is it like a pure optimism or is it a a couched optimism in a way? >> Look, I think uh uh I genuinely have a deep conviction that uh you know that you know improving technology and improving access to technology makes people's lives better. It's been my personal experience in my life. So out of which comes that deep conviction optimism but with that I've always felt you know humanity's journey is always about working hard to harness that technology to make it work well for society. So the optimism is not an optimism that you just build whatever you want whenever you know no you you work hard to make sure you're building technology in the service of humanity to to make uh the human condition better and and that's got to be the defining goal as we are approaching it and with that framework I'm very very optimistic and I think we are in very early stages of the kind of progress that we are going to see on many areas you know we all talking about areas like healthcare and stuff but you know I think a lot of the progress is ahead and the ability to make things in a fundamentally cheaper way on on many many things where things are things have been getting expensive for a long long time discovering new materials etc which really bend the curve I think I think it's tough to step today and understand the impact of all that but they're going to be very profound So I'm very optimistic. Uh but you know we have to be rightfully concerned about all the ways this technology can go wrong and that's how you make sure you you you solve those concerns. So I think I think they both go hand in hand. >> Well Sundar uh I know you don't do many of these. I really appreciate you spending the time with me today and it was great to get a chance to see you and speak with you again. >> Thanks Alex. I appreciate it. Thanks. All right everybody. Thanks so much for listening and watching and we'll see you next time on Big Technology Podcast.
Assistir ao vídeo original no YouTube
Assuntos: Agentes de IA, Dots, Muse, AGI. Publicado no YouTube em 2026-10-20.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.