Pichai Long Interview: Google Prioritizes Compute for Gemini R&D, Cloud Business Gives Up Some Near-Term Revenue

How much of Google's code is written by AI? Last year, CEO Sundar Pichai gave a clear number: 25%. But when the host asked about it again, his answer changed: "That percentage no longer means much."

Because AI is already deeply involved in code development inside Google. In the four weeks after Gemini 4 Argon was released, Google's code change submissions increased by about 30%, and the number of rollbacks fell by about one-third. Developers began using Gemini 4 Argon to handle complex projects involving millions of lines of code. Compared with the share of code written by AI, Google now pays more attention to metrics such as code change volume and rollback situations. In Pichai's view, these changes better reflect AI's impact on development efficiency and quality.

Recently, Pichai appeared on the Big Technology Podcast and spoke with host Alex Kantrowitz about Google's investments and trade-offs in the AI competition. In the first 72 hours after the newly launched Gemini 4 Argon went live, usage approached 25 trillion tokens, about three times that of Gemini 3.0 in the same period, while compute supply could not keep up with demand. When the outside world once believed that Google might give up competing at the AI frontier and focus on integrating models that were good enough into existing products, Pichai's response was very clear: "That judgment is wildly wrong." Pichai revealed that to safeguard frontier model research and development, Google has prioritized allocating compute to AI research teams, and the cloud business has therefore given up some short-term revenue.

In addition to model development and the scramble for compute, he also talked about AI participating in model development, continuously running personal agents, world models, and how Google is responding to competition from OpenAI. The following is an organized transcript of the interview.

TL;DR:

Alex: The outside world once thought that Google was no longer pursuing the most frontier AI and that models just needed to be good enough. But the launch of Argon shows that this is not the case?

Pichai: That judgment is wildly wrong. Pursuing the technological frontier has always been Google's core goal. In the first 72 hours after Argon went live, usage approached 25 trillion tokens, about three times that of Gemini 3.0 in the same period. We are very satisfied with its performance, and we also see areas that need improvement. We will continue to enhance its capabilities.

Alex: Some people believe that ordinary models are enough for Maps, Gmail, and Search, and that only specific tasks such as biology research require frontier models. Why does Google still pursue the most frontier capabilities?

Pichai: We want to continue driving technological progress. Some cloud customers need frontier models, and stronger models can also help develop the next generation of models. Internally, Argon has already solved some problems that past models could not handle. Of course, models of different sizes and capabilities all have applicable scenarios, and we will continue to invest in them.

Alex: You said that being at the technological frontier lets you see the future earlier. So what upcoming changes have you already seen?

Pichai: We have been researching how to let agents continuously perform long-duration tasks, and next we will move toward multiple agents collaborating with one another. In the future, both individual users and enterprise users may have multi-agent systems that continuously work for them.

Alex: Models still cannot fully act according to human wishes, yet we are pursuing recursive self-improvement to let them improve their capabilities faster. I feel worried. Should measures be taken to make the tech industry cool down on this matter?

Pichai: The concern is reasonable, but first we need to distinguish the current stage of development. At present, it is agents assisting researchers in developing models, with humans still involved, and we are still far from fully autonomous recursive self-improvement. Safety also cannot rely only on alignment; it also requires real-time monitoring, sandbox isolation, intervention in anomalies, and customer control. Next, the industry needs more complete standards, independent evaluations, and regulatory participation.

Alex: Some institutions regard large language models as the main path to AGI, while Google is also advancing world model research. Why do you think this route is also important?

Pichai: Understanding the laws of physics and accurately simulating the real world are important foundations for developing intelligence. Argon's performance in embodied reasoning also benefits from experience accumulated in world model research. In the future, these capabilities may be integrated through new architectures, and we hope to stay ahead on different technological paths.

Alex: The cloud business hopes to earn profits by selling compute externally, while AI research needs large-scale computing resources to develop the best models. How do you allocate compute between these two sides?

Pichai: First, ensure frontier AI research and model development, then allocate compute according to the long-term plans, demand forecasts, and contractual arrangements of each business. The resources needed for frontier R&D must be guaranteed. This is our first principle.

Alex: If you truly believe that AI's impact will be as profound as fire and electricity, then even at the cost of near-term profits from businesses such as cloud, shouldn't researchers be given as much computing resources as possible?

Pichai: We have already done so to varying degrees, and the cloud business has therefore given up some short-term revenue. However, appropriate constraints can also stimulate creativity and drive breakthroughs in algorithm efficiency. Our principle is to ensure sufficient resources are invested to maintain a frontier position.

Alex: Greg Brockman believes that sufficiently intelligent AI should let people express their needs directly through conversation, reducing reliance on screens and buttons. What do you think the interface in the AI era will look like?

Pichai: Technology should adapt to people's habits and make interaction as natural and smooth as possible. But interfaces still have a role, for example watching videos or enjoying content created by AI. The key is to reduce the effort users have to expend in order to use technology.

Alex: Does Google have products similar to Muse or Dots, or is it preparing to develop a new product to compete with them?

Pichai: Our goal with Gemini is to make it a powerful assistant that understands individual needs and can proactively offer help. Users only need to state their needs; they don't have to first distinguish whether they are chatting or assigning a task. It will take the appropriate approach, keep working in the background, and provide a coherent experience across different products.

Alex: You said before that 25% of the code inside Google is written by AI. What is that proportion now?

Pichai: That proportion is no longer very meaningful, because AI is deeply involved in development work. We focus more on efficiency and quality: in the four weeks after Argon's release, code change submissions increased by about 30%, rollbacks fell by about one-third, and developers also began using it to handle complex projects involving millions of lines of code.

Alex: Is there any AI research direction that has not yet achieved a breakthrough that Google is focusing on? For example, enabling models to continue learning and updating after release?

Pichai: We are exploring synthetic data, continual learning, more efficient architectures, and world models in parallel, and investing a considerable portion of our compute in them. Some of these research results will be applied directly to our main models.

1 Is Google giving up the AI frontier race? Pichai: That judgment is wildly wrong

Alex: Today, we have invited Alphabet CEO Sundar Pichai to talk about how Google is participating in the AI frontier race, how it balances resource investment between frontier AI research and the cloud business, and what future interfaces will look like. Sundar is here today at our studio in Mountain View. Sundar, great to see you, and welcome to the show.

Pichai: Thank you, Alex. Glad to be here.

Alex: With the launch of Gemini 4 Argon, Google has returned to the AI frontier. In plain terms, this is a model at the world-leading level, setting new records on multiple benchmarks and surpassing the publicly released models from OpenAI and Anthropic in tests including coding, mathematics, science, and knowledge work. Previously, it was widely believed that Google had essentially given up on the frontier race and planned to develop fast, capable-enough models and apply them to its own products, no longer pursuing the most advanced AI. But reality has proven that judgment wrong.

Pichai: And wildly wrong at that. We have always been very firmly committed to standing at the technological frontier; pursuing the frontier is, you could say, in our bones, and that is what we have always done. For more than a decade, we have been thinking about the development of the AI frontier. This field changes quickly, competition is intense, and it is exciting.

I am very satisfied with how Argon ultimately turned out. During development, we did not rush; instead, we made sure to conduct thorough safety testing and deploy carefully. Given its capabilities, we spent more time on it than on other models. But overall, whether for agentic tasks, knowledge work, coding, or creative tasks, I think Argon is one of the very outstanding frontier models.

In the first 72 hours after launch, usage had already approached 25 trillion tokens, roughly three times the first 72 hours after Gemini 3.0 launched. At present, we are still capacity-constrained when providing the service. We are excited about this, but there are always areas that need improvement. The AI frontier changes quickly, and we need to keep improving our models. We also see where Argon still does not do well enough, and we will continue to work on those areas.

Alex: Which areas specifically?

Pichai: Take Argon: compared with the model level two months ago, it has improved in every respect. But this field changes very quickly. Even in coding, some areas still need to be better. We have also received clear feedback and will continue to improve these issues.

Alex: Let me approach this from another angle and explain why some people think ordinary models may already be enough for Google. Let me restate their reasoning.

Pichai: Okay.

Alex: Many ordinary models are now strong enough; Meta's Muse is one example. Some people argue that Google could also use these models to do a good job with Maps, Gmail, and Search, while reserving the most frontier capabilities for specific fields such as biology research. Why does Google still insist on pursuing the strongest models?

Pichai: I have never looked at it that way. We have always been committed to continuously raising the level of technology and exploring just how far this technology can go. Since Transformer, we have been working in this direction, and we will continue to do so.

We are fortunate that, whether for individual users or enterprises, there are many opportunities to apply these technologies. Some cloud customers do indeed need the most frontier models. Moreover, in the era of recursive self-improvement, frontier models can better help develop the next generation of models, so maintaining frontier capabilities is also important in this respect. Internally, after making Argon available to developers, we did feel a clear change: it can handle some problems that past models could not solve. So being at the technological frontier also puts you at the frontier of innovation and helps you continue to drive innovation.

Models of different sizes and capabilities each have their own suitable scenarios. We have always believed this, which is why we developed models such as Flash and Flash-Lite. I also agree that it is not necessarily true that you must use the most frontier models to create a good product experience; after all, these frontier models are usually slower and more expensive. However, we also attach great importance to advances in frontier technology, so we will continue to invest in all these directions.

Alex: You just said that developing frontier models lets you see earlier where the future is heading. So what upcoming changes have you seen?

Pichai: What makes this so exciting now is that after months of effort, you finally build and release the results, and at that point you can already clearly see which directions you can continue to explore next.

We have been working to enable agents to continuously carry out longer-duration tasks. Obviously, the next frontier area of focus is enabling multiple such agents to collaborate with one another to complete tasks together. This is an emerging and exciting new paradigm. I also clearly see that in the future, whether individual users or enterprise users, everyone will have some kind of multi-agent framework that lets these agents continuously run and handle things for them. So you feel that this is a moment full of infinite possibilities.

2 Letting models participate in R&D and also letting models understand the world

Alex: You mentioned recursive self-improvement, that is, letting models improve themselves. This process may not even require human participation. Some people call it an "intelligence explosion," as if with a "bang," AI suddenly acquires capabilities beyond those of humans. I remember that Sergey, in a roundtable discussion with DeepMind researchers, said that one of Argon's core innovations is relying more on models to help develop models.

But we have all seen what has happened recently with AI models. They are already doing well on alignment, but deviations still exist, and these deviations can cause big problems. The tech world has already spread the news of the Hugging Face intrusion, and more and more of the public are gradually learning that OpenAI's AI agent broke through isolation limits, intruded into Hugging Face, and did some very outrageous things. Similar situations have also occurred at Google: your AI agents gained access they should not have had, and later they stopped these behaviors.

However, there is already enough evidence to show that these models do not always act entirely as we want them to. Under these circumstances, continuing to pursue recursive self-improvement, or even just letting models advance at a faster pace, is an idea that makes me uneasy, even afraid. Don't you think it is reasonable for people to have such concerns? Should some measures be taken to get the tech industry to cool down a bit on this matter?

Pichai: Your question can be answered in three parts. First, RSI, that is, recursive self-improvement itself. You are right that Sergey did mention that RSI played an important role in Argon's development. But one point needs to be made clear: if we view the development of RSI as a gradually advancing process, we are still very far from its final form. More accurately, some parts of current model development already use AI to help improve models, but this process still involves researchers. You can understand it as agents assisting AI researchers through a certain degree of automation to complete model development together. I think that, at the very least, we should first clarify exactly what stage we are currently at. I expect that by next year, RSI will continue to make progress.

The second issue is safety. Safety needs to rely on multiple layers of protection, that is, defense in depth. You just talked about alignment; we are all doing a great deal of work to ensure as much as possible that models remain consistent with our intentions. But alignment is still an issue that requires ongoing research, and there are indeed obvious shortcomings at present. However, we do not rely only on alignment to ensure safety.

Beyond that, we also put a great deal of effort into monitoring the model's runtime behavior in real time, using telemetry data to fully understand what is happening and to ensure that we can intervene promptly once anomalous behavior is detected. Exactly how to implement this, what kind of architecture to adopt, and how to use sandboxing and isolation mechanisms to let the model continue completing tasks while being monitored in real time are all directions we are investing in. How to achieve this, what architecture to adopt, and how to ensure that the relevant components run in a sandbox and are isolated, so that the monitoring system can continue to fulfill its own responsibilities during real-time monitoring—these are all directions we are investing in.

Of course, when deploying models, we also give enterprise customers the corresponding control, including real-time monitoring of the model and deciding for themselves where the model is deployed and what types of data it can access. This work is all very important, and we are pushing it forward vigorously. So the entire industry needs to continuously raise its safety level, spend enough time doing this work well, and also invest the necessary compute to ensure that models can be deployed safely.

The third question is whether doing all this is actually enough. This is also a question that is constantly discussed in the industry. I think the concerns are reasonable, but the right response to these concerns is to think early about what else we can do. You can indeed see that the industry is doing more and more work on alignment, safety testing, and so on. As these models are put into use, I think the next step should be to develop better standards, so that the entire industry has norms that everyone follows, including bringing in independent evaluators to conduct assessments. For example, in the field of financial management, we have accounting standards and independent audit institutions. For the AI industry, these are all things that should be advanced next. In this process, regulation will ultimately play a role as well.

Alex: Next, a technical question. One thing that is somewhat special about Google: you are both researching large language models along the direction opened up by the Transformer and advancing world model research. Some AI research organizations, such as OpenAI, have basically concluded that large language models are the path to AGI and have scaled back some world model research, that is, research on AI models that try to understand the physical world. But you are advancing both directions at the same time. Can you talk about why you think world model research is important?

Pichai: People perceive the world around them and obtain all kinds of information from it. I think enabling models to understand physical laws and accurately simulate the real world are both foundational capabilities necessary for developing intelligence. We are also exploring various possible applications of this technology. Take robotics: Argon's leading performance in embodied reasoning tests is not accidental. It has reached the current state of the art in many of those tests, and this also benefits from the experience we have accumulated in developing world models. I expect that in the future, architectures will emerge that can integrate these capabilities. But at present, the frontier of intelligence is developing unevenly in different directions, and we want to maintain leadership across all possible technological paths. I think this will help us conduct better research, deepen our understanding of intelligence, and develop better AGI models. That is our thinking.

Alex: Yes, after Argon was released, I looked at your tweets, and they were almost all about real-world applications, not just examples of text or models generating text.

3 Who gets compute first? Google is willing to give up some cloud business revenue for frontier AI

Alex: Compute is your most valuable resource. Someone told me that Google should exit AI research and just sell its cloud computing resources and chips directly; Nvidia is an example. On the one hand, the cloud business wants to earn relatively high profits by selling compute externally; on the other hand, AI research needs computing systems large enough in scale to develop the best AI. There is tension between the two in terms of compute allocation. You have said that an important part of your job is deciding how to allocate compute. So how exactly do you do it, and how do you balance the needs of these two sides?

Pichai: For us, the first thing is to ensure that we invest enough compute to maintain our position at the frontier of AI. In 2015, I set the company's direction as "AI first." Therefore, from Search to Cloud to Waymo, AI is a key foundational capability running through all our businesses.

So we first need to ensure that there is enough compute for both conducting research and developing models. On that basis, we then arrange compute for each business. Each business has its own long-term plan. They forecast future demand based on these plans. For example, the Cloud team has business forecasts, customer demand, and signed contracts. All of this information helps us plan long-term compute investment for the cloud business.

We are very fortunate to have so many businesses we can develop. But I think there is no company in the world that does not need to make trade-offs at all. The first principle we follow is to ensure there is enough compute to support frontier AI research and development. This must be guaranteed, and arrangements for other businesses are based on it. We also need to take into account that we are building our own computing infrastructure and data centers, and these facilities have long construction cycles and require advance planning. Therefore, how much compute we have today depends on decisions I made in 2024. Even just looking back a year ago, agentic coding had not yet exploded. Looking at it today, it is obvious that everyone does not have enough compute, not just us but other companies as well. Everyone wishes the compute scale planned back then could be doubled.

I think this is a problem we inevitably have to face in this industry. We always need to plan our compute needs two to three years ahead, but two years later, as technology advances and we look back, we may find that some of the decisions we made at the time were not optimal. I don't spend too much time agonizing over these things. My approach to allocating compute is to first ensure that Google DeepMind can get the most advanced compute support when conducting research and developing frontier models, and then make arrangements for the various businesses. At the same time, providing compute to other companies as we do now is a very good thing. It helps us scale, and it also allows the relevant technologies to be tested in as wide a range of real-world applications as possible, ensuring we stay ahead at every layer of the AI stack.

Alex: In one podcast episode, Ben Horowitz relayed something Andy Jassy said, to the effect that compute has already been sold out through 2028. Is that what you're seeing as well?

Pichai: I think that's basically consistent with the actual situation. If a customer came to us today hoping to reserve compute for 2027, it would be hard for us to meet that demand. Last year, we even rented some short-term compute specifically, partly to meet the needs of some customers and help them solve their immediate problems. So I think everyone is facing roughly the same compute supply cycle.

Alex: Some people say that every technological transformation brings overbuilding, such as railways and fiber optics. Is there a similar problem with this round of AI infrastructure construction?

Pichai: That depends on the time horizon you use to judge. From a longer-term perspective, the current scale of construction seems entirely reasonable to me. We are in the early stages of the development of one of the most far-reaching technologies in history. And unlike fiber optics back then, the compute facilities we have built are being actively used, and we face capacity shortages every day. Applications keep evolving, and the underlying technology is also advancing rapidly. However, as the operator of a company, when I make decisions, I look at actual demand, at the revenue corresponding to contracts already signed, and then combine that with underlying technical capabilities, internal evaluations, our own usage, and trends in model capability development to plan compute construction. We need to be confident and invest aggressively, while also remaining prudent and arranging things sensibly.

But if you look at the overall economic environment, different players bear very different risks, and it is hard for me to judge everyone's situation. I focus on what the company can control and plan accordingly. If you ask whether overbuilding is possible, the answer is yes. Because compute construction requires various conditions to fit together, certain components may become constraints, and to actually build these compute facilities, all the necessary conditions must be in place.

Alex: Memory, power, and electricians...

Pichai: Right, there may be two thousand or even more components involved. So a certain link may become a bottleneck, causing other parts that have already been built to be underutilized, which in turn leads to overbuilding. In addition, the economy may also be affected by external shocks, and macroeconomic risks always exist. These situations can happen at any time. But if you ask whether I believe that, given enough time, all the compute we are planning now will ultimately be put to use, my answer is: absolutely.

Alex: Let me follow up on this. Since you are doing both cloud business and AI research at the same time, there is a view from the outside: for AI researchers, the larger the computing system, the better the AI they can produce. When you said back then that AI's impact would be as profound as fire and electricity, I was there, and we all laughed. Thinking about it now, I somewhat regret laughing at the time. If you truly believe AI will have such an impact, then even at the cost of near-term profits, such as profits from the cloud business, you should provide researchers with as large a computing system as possible. What do you think?

Pichai: We are actually already doing this to varying degrees. To support frontier AI research and development, we have always been making trade-offs in compute allocation, and the cloud business has given up some short-term revenue as a result. So what you describe is in fact already happening. However, this also requires careful balancing. Even for AI researchers, appropriate constraints are beneficial. They can stimulate creativity and push researchers to make breakthroughs in algorithmic efficiency, such as finding ways to reduce memory usage. So we will ensure we invest enough resources to keep us at the forefront of AI technology. That is the basic consideration in how I allocate compute.

4 Interfaces in the AI era must adapt to human habits

Alex: I'd like to know what you think future interaction interfaces will look like. As AI becomes more widespread, how will we use computers going forward? A few months ago, I talked about this with OpenAI President Greg Brockman, and he said something very interesting: "What kind of interface do you want? What kind of product do you want?" His answer was that you actually don't need an interface or a product; you want to communicate directly like you and I are doing now, with no interface between us, basically just conversation. What he means is that if AI is already smart enough, there shouldn't still need to be a screen, or so many buttons and other things that users have to operate manually. What do you think of the interface in the AI era?

Pichai: My view is that in the past, humans have always adapted to computers; with AI, it will be the other way around, with technology adapting to people's lifestyles and daily habits. You want to be able to talk to AI whenever you need to; it understands context, runs continuously, and lets interactions flow naturally from one to the next without you having to deal with all kinds of complicated operations. I agree with all of that. But in my view, that doesn't mean interfaces will no longer be needed.

Take my experience using my phone now: Gemini's voice model is already very good. I just press the microphone button and can ramble on for a while—we call it "rambler." It combines what it knows about my personal context to automatically correct the recognition results. It even gets my friends' more unusual names right, and even if it guesses wrong at first, it corrects itself afterward. The way it adapts to me is truly remarkable. Now, after I finish speaking, I no longer have to manually delete or edit the transcribed text. That it has improved to this degree is genuinely surprising.

But sometimes you still need an interface. For example, you might want some entertainment or want to watch a video, so you might ask AI to create some content and then watch and enjoy it. So in my view, the key is to make interaction as smooth and natural as possible. Users shouldn't have to work hard to adapt to a particular interface; technology should adapt to the way users are doing things at that moment.

Alex: Yes. I want to talk about a few more topics related to OpenAI. Because last year, you put them into a "code red" state. You surely know that Gemini was performing very well at the time and OpenAI was on high alert. When I talked with Sam Altman about the challenge from Google, he said he believed OpenAI could handle your competition because they are building AI-native products. I'm not sure if he used exactly those words, but his implication was that Google is just adding AI features to existing products. What do you think of that characterization? Of course, I also want to ask first whether you accept that description. Can adding AI features to existing products compete with AI-native products?

Pichai: First, this field is changing very fast right now, and we are all pushing each other forward. OpenAI has also brought a lot of competitive pressure on us—it's far from being only us putting pressure on them.

Alex: Yes, you've also been in a "code red" state. Of course, I know there's some dispute over whether that state was actually entered.

Pichai: I mean, at this stage of AI development, all of us need to respond with varying degrees of urgency depending on the situation at the time. And we've also been building new products all along. Take Gemini—I mean the Gemini app—it was built from scratch. And next, as everyone shifts toward agents, would you say OpenAI is adding agent features to ChatGPT, or building native agent products? How would you judge that? The same question can absolutely be asked of them in reverse.

So I wouldn't frame the question that way. In my view, what matters is building world-class, AI-native product experiences that can win users. Sometimes that requires transforming existing products; sometimes it requires rethinking and designing entirely new experiences. These are all part of the work we need to do in responding to this round of change.

Alex: Let's talk about the business model. What do you think the business model around this kind of AI interaction will look like in the future? For example, I send my agent out to do something for me, and it recommends a product to me. If I end up buying it, should the agent take a commission from that transaction? Even if users are clearly informed in advance, is that appropriate? Or, to put it another way, how do you plan to make money from these products?

Pichai: Let's set aside enterprise business for now.

Alex: Right, I'm asking about products for individual users.

Pichai: For individual users, I think they will increasingly recognize the value of this technology. Take today's software developers—they are willing to pay for coding models because those models genuinely improve their work efficiency. No one wants to go back to the way programming was done three years ago, or even the way it was done at this time last year. So as long as a product creates value for users, there is an opportunity to earn corresponding returns. Looking back at the development of consumer products, I haven't seen any product that truly reached a scale of billions of users, with users still using it frequently and deeply, yet had no corresponding way to monetize. As long as a product truly creates value for users, there is an opportunity for commercial returns.

Alex: Completely agree.

Pichai: Of course, the specific form of monetization may change. What you just described is a scenario where an agent does things on your behalf. Users will ultimately have commercial needs such as buying goods or services, and agents need to help users fulfill those needs, while the goods or service providers on the other end also gain value from it, and ultimately a corresponding commercial mechanism will form. That's how an efficient system works. But I think the most important thing is still to build a world-class consumer product experience.

Alex: I want to raise a product feature request. After connecting Gmail to ChatGPT, searching emails within ChatGPT works very well. Can Gmail's own Gemini-powered search also reach the same level?

Pichai: We recently announced a very exciting development, which is a unified Gemini agent for enterprises, and we will also bring it to individual users. These agents can run continuously in the cloud around the clock, essentially launching a micro virtual machine, or microVM, for each query. From the runtime and the agent execution framework, or harness, to the model, the entire system runs as a service that does not depend on a specific front-end interface. Therefore, it can work across different products and carry over your context information.

Whether you are in Workspace, Search, or Gemini, you can invoke this agent. I think agents will connect the usage experience of these products and deliver the unified experience users have long expected. You are also right that we have many products and are moving quickly, so inevitably some products progress faster and some slower. The problem you raised does exist. However, in building the agent experience, we are investing a great deal of time and effort to ensure that users can get a coherent experience across different products and that the relevant capabilities all work properly.

Alex: Okay, that actually also answers the earlier question about the interaction interface very well.

Pichai: Yes, indeed.

Alex: Speaking of where Google will focus next, people also ask: where are Google's own Muse or Dots? These products seem to have suddenly appeared recently and become popular very quickly. Do you have similar products? Or do you plan to develop a new product to compete with products like Muse?

Pichai: Actually, our vision for the Gemini app has always been clear: to make it a powerful assistant that understands users' personal needs and can proactively offer help. That is the direction we are working toward. At this year's Google I/O, we already showed an early version of it through Spark. In fact, some subscribers can now see two entry points, "Chat" and "Tasks," when they open the Gemini app.

These features are already beginning to let users experience the capabilities of agents. With Argon, these capabilities have improved substantially. The capabilities we recently announced at our enterprise-focused event will also be opened to individual users. Soon, you will simply tell Gemini what you need without first distinguishing whether you are chatting with it or asking an agent to perform a task. It will automatically take the appropriate approach, run continuously in the background safely, and get things done for you.

Looking at actual usage of the Gemini app, if you categorize what users sent last year, the vast majority was chat. But now, more and more users are asking Gemini to perform tasks, and the share of such requests is steadily rising. So yes, we will provide individual users with their own dedicated agent. Whether in the Gemini app, Google Search, or Gmail, users can invoke it. We want to get this experience right first. It will be powered by Argon, and we are very much looking forward to letting users use these capabilities.

5 From Anthropic to space data centers, how is Google betting on the future?

Alex: Next, a few questions on other areas. Counting both invested funds and committed investments, Google's investment in Anthropic has reached tens of billions of dollars. What did you see in Dario that made you willing to make such a large investment commitment?

Pichai: We paid attention to Anthropic's founding team very early on. It was clear at the time that this was a world-class team, and they had participated in the development of some of the earliest GPT models. Therefore, we quickly judged that they had the ability to stand at the technological frontier. We have always wanted to find companies with the potential to lead technological development and work with them. That is true of our cooperation with SpaceX, and also true of our cooperation with Anthropic. That is what we valued at the time. Later, we had the opportunity to cooperate with them, and some of that cooperation also centered on our TPUs. That is the consideration behind these decisions.

Alex: Now about Apple. You reached a cooperation with Apple, licensing them to use Gemini to help improve Siri. Is this cooperation now expanding, shrinking, or staying the same?

Pichai: For both us and Apple, the goal of this cooperation is to turn AI capabilities into a better user experience. I think Apple has done an excellent job in this regard. They customized the model and then integrated these capabilities into Siri. We are also pleased to see positive user feedback, and we hope to deepen cooperation on this basis.

Alex: Will you carry out more cooperation with them?

Pichai: There is no specific information I can disclose at the moment, but this cooperation is progressing very well.

Alex: Next, let's talk about space data centers. You've announced that you'll be doing more work in this direction. So far, what have you learned about the feasibility of this project?

Pichai: It's a very exciting project. Projects like this always remind me of an experience: years ago, in the desert, I watched Stanley, the autonomous car, compete in the DARPA Grand Challenge, and that scene is still vivid in my memory. That work later laid the foundation for Waymo.

These technologies often go through a stage: at first they look like science fiction, but ten or twenty years later, they can really be put into use and play a practical role. I feel the same way about space technology. We have always believed in making bold bets on these technologies and investing in them on a long-term, large-scale basis. Quantum computing is a good example—we've been investing in that direction for about ten years; Waymo has been at it for nearly twenty years; and we've been exploring the space direction for two to three years.

Pichai: In my view, this is a path that will be very valuable in the future. In sun-synchronous orbit, getting around the clock the most abundant energy available to us will be very meaningful for meeting the exponentially growing computing demand of the future. Of course, this is still a challenging project, and the team has set strict milestone goals for it and is advancing in a planned way. Sending TPUs into space is a very exciting milestone. Right now, as we speak, we're waiting for them to send back a "Hello World," and I'm looking forward to seeing the first successful run and signal sent back, and we'll continue to invest here.

Alex: Maybe last year, you said that 25% of the code written inside Google was written by AI. What is that proportion now?

Pichai: We've now gone further. In a way, that proportion no longer means much to us, because almost all the code we handle now has deep AI involvement. So we've started looking at other metrics. Take Argon, for example—we clearly feel that the way it's used has changed: developers have started using it for genuinely complex projects, including projects involving millions of lines of code. It's been four weeks since we launched Argon. The number of code change lists (CLs) inside Google has increased by about 30%, and code rollbacks have decreased by about one-third. This shows that with it, people can not only write code more efficiently, but also improve code quality. So we now use these new metrics to measure progress.

Alex: What do you think about the voices opposing data center construction? If you can't build data centers in the US the way you'd like, where else would you consider going?

Pichai: We take building data centers responsibly very seriously. I'll visit the large data centers that have been operating the longest and talk with residents of the local communities. Some people tell us that it's precisely because of our local investment that their children have the chance to go to college. Over the years, we've also grown together with local communities. So I think the key is to do things well: communicate with communities early and proceed with construction in a transparent way. We commit to adding power supply for these projects and bearing the full electricity cost of the projects. At the same time, we also need to understand what impact data centers will have on surrounding communities and address related issues early. I think that as long as this work is done well, construction can continue to move forward. Of course, we've also been listening to feedback and adjusting accordingly. However, I do worry that not all participants will take the same approach.

Alex: Yes, some residents suddenly find out that a data center is going to be built near their homes, and they feel surprised.

Pichai: So these issues are all important, and the whole industry should continue to improve. Part of the reason we're looking twenty years ahead and considering options like space data centers is also to explore other deployment locations. Building data centers in different environments may also involve different trade-offs.

Alex: At present, is there any research direction in AI that has not yet achieved a breakthrough that Google is focusing on tackling? For example, people keep talking about continual learning—hoping that AI can keep learning and updating after release, rather than forgetting previous content after each session ends.

Pichai: We are exploring multiple directions at the same time. For example, Argon can already output one million tokens, and earlier Gemini models had already achieved a one-million-token context window on the input side. These are all examples of how we keep seeking breakthroughs. We are also researching how to use synthetic data, how to achieve continual learning, and how to design more efficient model architectures. We do devote a considerable portion of our compute to these explorations, and research such as world models is also advancing in parallel. Sometimes, new ideas that emerge from this research are directly applied to the development of our main models. We may not publicly describe every piece of work in detail, but in these areas, we are indeed exploring multiple possibilities at the same time.

Alex: Finally, I want to talk about this topic. To devote yourself to the tech industry as you have, you need a kind of deep-seated optimism—a belief that what you create can make the world better. That has also been a long-standing conviction in the tech industry. But believing that often means you have to make a commitment, sometimes a very large one, expecting that what you are doing will eventually become reality as you envision it. Yet the results do not always turn out as hoped. Let me give an example: I recently visited El Salvador, about five years after my previous visit. Last time I went, I saw a small town where people were full of enthusiasm for Bitcoin and blockchain; there was a lot of construction and a sense of prosperity. But later developments did not meet everyone's expectations. When I went again this time, it was basically an empty town, with many unfinished construction projects left behind.

I think doing the things you are doing requires this kind of optimism: even knowing that a commitment may fail and may bring adverse consequences, you still believe it is worth trying. So I would really like to hear your view. In your Stanford commencement speech, you also talked about the importance of staying optimistic. But that optimism should also have its preconditions, right? How do you maintain that mindset? Are you always optimistic, or do you also hold something back?

Pichai: I have always firmly believed in advancing technology, enabling more people to access technology, and making people's lives better. This is something I have experienced personally, and it is also the source of my conviction and optimism. But at the same time, I have always believed that in the process of human development, we must always work to harness technology so that it better serves society. So optimism does not mean developing whatever we want whenever we want. We must make an effort to ensure that the technology we develop serves humanity and improves people's lives. That should be the most fundamental goal when we develop technology. On that premise, I am very optimistic. I think progress in many fields has only just begun. Everyone is talking about applications in areas such as healthcare, but a great deal of progress is still ahead.

Pichai: The costs of many things have been rising for a long time, and technology has the potential to significantly reduce those costs. Breakthroughs such as the discovery of new materials may also change existing trends. Standing here today, it is hard for us to fully understand how much impact these advances will ultimately have, but that impact will be very profound. So I am very optimistic. However, we must also take seriously the various problems this technology may cause. Only by facing these problems squarely can we begin to solve them. I think we need to uphold both optimism about the promise of technology and a serious approach to the problems it may bring.

Alex: Thank you very much for taking the time to talk with me today, and I am very glad to have the opportunity to meet and talk with you again.

Pichai: Thank you for the invitation, Alex.

Podcast link:

This article is from the WeChat public account "InfoQ" (ID: infoqchina), translated by Lin Qibei, edited by Cai Fangfang, and published by 36Kr with authorization.