Gemini 是怎么被“逼”出来的？谷歌创始人谢尔盖·布林（Sergey Brin）亲述：我为什么回到 Google？

I think that, well, first of all, we've definitely flopped on a bunch of things. We don't need to get into all of them right now. But we've had a long list of failures at the same time. So, part of it is just trying. I think that because of the kind of academic roots, maybe we were more inclined to try hard things. And I think coming into this last decade or so, especially the hard things have become more and more valuable. I guess if you look at AI, which is obviously a huge trend, the amount of computing that has to go into that, the amount of deep math that has to go into that, those are all technically deep and challenging problems. And I guess it's just a twist of fate that that turns out to be important at this stage in the world. I mean, there was a while where you could do pets.com, you remember, you can put anything on .com. It wasn't really that technically deep. You need marginal understanding of the web and you can do whatever .com. Fortunately, we were doing search, which did require some deeper technical skills. But the technical sophistication level has only gone up. In fact, now the people we hire are much more qualified than I was at the time. I was kind of a mathy computer science major because during college I did both math and computer science, which was somewhat unusual in my class. But nowadays, as we hire people out of Stanford and all the other top programs, these people are pretty sharp mathematically and computer science wise, and a bunch of them are physicists because physicists have to do the hard math and a lot of the stuff they do is very computationally limited, so they need to have some degree of computation skills. So I just think somehow it has happened to be the case that some of the deep hard tech has become increasingly important, and I think we just kind of lucked out on having set the bit early on in that direction.

Okay, let me think how to answer that without just pounding my own chest. Yes, it's a huge amount of investment for sure. I guess I would say in some ways we for sure messed up and underinvested and didn't take it as seriously as we should have, say, eight years ago when we published the transformer paper. We actually didn't take it all that seriously and didn't necessarily invest in scaling the compute, and also we were too scared to bring it to people because chatbots say dumb things. And OpenAI ran with it, which good for them, it was a super smart insight, and there's also our people like Ilya who went there to do that. But I do think we still benefited from that long history. We had a lot of the research and development of neural networks going back to Google Brain. That was also kind of lucky. It wasn't luck that we hired Jeff Dean. I mean, we were lucky to get him, but we were in this mindset that deep technical things mattered, and so we hired him. We hired a lot of people from DEC honestly because they had the top research labs at the time. He was passionate about neural networks, and it stemmed from his college experiment. He was like, whatever, curing third world disease and figuring out neural networks when he was 16. He's done crazy things. He was passionate about it, he built up a whole effort. In my division at the time in Google X, we had him, but I was like, okay, Jeff, you do whatever you want. He's like, oh, we can tell cats from dogs. I'm like, oh, okay, cool. But you also trust your technical people. Soon enough, they were developing all these algorithms, these neural nets that were doing some of our search, and then Noam came up with a transformer. We were able to do more and more. So we had the underpinnings, we had the R&D, we did underinvest for a number of years and didn't take it as seriously as we should have, but we also at the time had developed the chips for it, like the TPUs go back 12 years or something. Initially we were using GPUs, we were probably among the earliest to use GPUs, and then we used FPGAs, and then we tried to develop our own chips, which have now evolved through a bazillion generations. So I guess it was the trust in going after the deep tech, getting the more computation out, developing the algorithms, and in parallel we were big investors in compute for a long time. So we've had the data centers for a long time on a scale that I don't think many have. Amazon AWS also has very sizable data centers, but very few have that scale of data center, have their own semiconductors, have the algorithmic learning algorithms and so forth to be able to perform at the forefront of modern AI.

It is absolutely amazing, the rate of innovation, and it's hugely competitive now, as all of you see, between the top US companies and the top Chinese companies. If you skip the news and AI for a month, you're way behind. So where is it going to go? I think we just don't know. Is there a ceiling to intelligence? In addition to the question you raised, can it do anything a person can do, there's the question of what things can it do that a person cannot do. That's sort of a superintelligence question. I think that's just not known. How smart can a thing be? We've had however many hundreds of thousands of years of human evolution and whatever millions of primate, but that's a pretty slow process compared to what's going on with AI.

Are we ready for the speed the technology is advancing? So far, I think people are getting great use out of technology. Even though there are doom and gloom forecasts here and there, everybody's pretty well empowered. The AIs, truth be told, are periodically dumb enough that you're always supervising them anyway. But occasionally they're brilliant and give you a great idea. Especially as a non-expert, if I want to figure out how to create a new AI chip, I could talk to our expert designers, but as a base case, I can whip out my phone and talk to an AI about it. It'll probably give me an 80 to 90% decent overview and understand it, or whatever my health questions or whatnot. I do think it makes individuals very empowered because generally you don't have experts in XYZ all around you all the time. I think that empowerment can create a lot of potential, whether it's career, enterprise, health, or living well. I don't think I have all the answers. I just think it has a huge potential to improve individual capability.

I think it's super hard to predict exactly what will happen. From the advent of the web to cell phones and so forth, those have transformed our society profoundly, have transformed the kinds of jobs and careers and studies people do for sure. AI will 100% change that. But I think it's very hard right now in a rapidly shifting landscape to say exactly what, and also the AI we have today is very different from the AI we had five years ago or the AI we are going to have in five years. So I don't know, I think it's tough to really forecast. I would for sure use AI to your benefit. There are just so many things you can do. Just myself as an individual, whether it's choosing a gift for my friends or family, brainstorming new ideas for products, or for art, I just turn to AI all the time now. It doesn't do it for me because I'll ask, give me five ideas, and probably three of them are going to be junk in some way that I'll be able to tell, but two will have some grain of brilliance, or possibly put it in perspective for me so I can refine and think through my ideas.

I chose computer science because I had a passion for it. It was a no-brainer for me. I guess you could say I was also lucky because it was such a transformative field. I wouldn't not choose computer science just because AI can be decent at coding nowadays. AI is pretty decent at a lot of things. Coding just happens to have a lot of market value, which is why a lot of people pursue it. Furthermore, better coding makes for better AI, so a lot of the companies like our own that work on it care a lot about it. We use it a lot for our own coding and even for our algorithmic ideas. But that's because it's such an important thing. So I wouldn't switch to comparative literature because you think AI is good at coding. The AI is probably even better at comparative literature, to be perfectly honest. I don't mean to disrespect comparative literature majors, but when the AI writes code, sometimes it doesn't work. It'll make a mistake that's pretty significant. Getting a sentence wrong in your essay about comparative literature isn't going to have that consequence. So it's honestly easier for AI to do some of the creative things.

Wow, okay. That's a big responsibility, like planning the dean job will come.

I guess I would just rethink what it means to have a university. I know that sounds kind of annoying. That's the kind of thing Larry would say and I would be really annoyed with him. But we have this geographically concentrated thing with buildings and fancy lecture halls. But realistically, information spreads very quickly. Many universities have gone online, including Stanford, but MIT with open courseware early on, and all these startups like Coursera, Udacity, you name it. Teaching is getting spread, and anybody can go online and learn. You can talk to an AI or take one of these classes and watch YouTube videos. So what does it mean to have a university? Are you trying to maximize impact? In that case, limiting it geographically is not going to be so effective. To be fair, the Bay Area is a special place, but I don't know that for the coming century the idea of a School of Engineering and university is going to mean the same thing as it used to. People move around, work remotely, collaborate across. It's a little at odds because we're trying to get people into the office, and I think they do work better in person together, but that's at a certain scale. If you have a hundred people together, they don't have to be at the same place as these other hundred people. Increasingly, I see individuals who create new things regardless of degree. We've hired a lot of academic stars, but we've also hired tons of people who don't have a bachelor's degree or anything like that, and they just figure things out on their own in some weird corner. I think it's a really hard question. I don't feel like I'm going to magically deliver the new recipe, but I just don't think this format is likely to be the one for the next hundred years.

No, it was great. Actually, it was a bit deeper.

Wow, it's a great question. Is the academia to industry pipeline crucial? I'm going to give you an I don't know on that. When I was a grad student, the time from a new idea to it being commercially valuable was many decades. If that compresses, then that no longer makes as much sense. In academia, you have freedom to think about it for a while, apply for grants, and you can spend a couple decades thinking about it, and then it percolates, and eventually maybe there's some big company or your startup pursues it. The question is, does that make sense if that timeline shrinks a lot? I think there are certain things that for sure make sense. Even with AI, I periodically keep up with Stanford research and other universities, and occasionally we hire those folks and collaborate with them. But I don't know that they needed to have that period of time. They were trying some new attention thing, spent a couple years experimenting, and then took it to industry. Obviously, industry is also doing all those things, so probably not a huge argument for that. Radical new architectures and things, maybe, but the time industry will scale it will be much faster. Quantum computing comes to mind. It was first brainstormed by Feynman in the 80s, postulating the idea of quantum computing, and now there are a bunch of companies doing it. There are also university labs trying new ways to do it. That's kind of on the fence. I would say if you have some completely new idea, like you're not doing superconducting qubits or trapped ions, but you have some new way, maybe you need to let it marinate in university for some number of years. Those things are kind of hard. It could make sense, but at some point if you decide it's really compelling, you're probably going to take it commercial. I want to give you a clear-cut answer, but the top companies now invest in much more fundamental research, and with AI starting to pay off, those investments are paying off. So it would shift the proportion of endeavors you would do, but I do think there are still some things that take a decade of more pure research that companies might be reluctant to pursue because that's too long a time to market.

Oh, what mindset should you adopt to avoid repeating earlier mistakes? When you have your cool new wearable device idea, really fully bake it before you have a cool stunt involving skydiving on airships. That's one to both of you. No, I actually like what we were doing back in the day with Google Glass as an example of prior mistakes. I tried to commercialize it too quickly before we could make it as cost-effectively and as polished as needed from a consumer standpoint. I jumped the gun. I thought, oh, I'm the next Steve Jobs, I can make this thing. That's probably one. If I encapsulate, everybody thinks they're the next Steve Jobs. I've definitely made that mistake, but he was a pretty unique guy. So I would say, make sure you've baked your idea long enough and developed it to a far enough point before you get on a treadmill where expectations increase, expenses increase, and you have to deliver by a certain time. You might not be able to do everything you need in that amount of time. You get a snowball of expectations and you don't give yourself all the time you need. That's the mistake I would have tried to avoid.

From my point of view, all the things you listed have already been bigger factors than scaling compute and scaling data. I think people notice the scaling because you're building data centers and buying chips, and there were publications from OpenAI and Anthropic about different kinds of scaling laws. So that attracts a lot of attention. But if you carefully line things up, you'll see that algorithmic progress has outpaced even the scaling over the last decade. A while ago, when I was in grad school, I saw a plot for the n-body problem. There was huge Moore's law increase in compute since people started worrying about that in the 50s, but the algorithms to do the n-body problem far outpaced that compute scale-up. So I think companies like ours are never going to turn down being at the frontier of compute, but that's just the dessert after the main course and veggies of actually having done your algorithmic work.

Okay. What emerging technology is being seriously underestimated? Wow. I obviously can't say AI because it's hard to argue, but it could be underestimated.

A lot of people wonder about quantum computing and what it will bring. It's probably not what I would hang my hat on to answer that question, although I definitely support our efforts in quantum computing. But there are many unknowns. Technically, we don't even know if P is not equal to NP. On the computation front, there are so many unanswered questions, and quantum algorithms are specific for particular structured problems. That said, I'm a big proponent, but it's hard to put my finger on that. Perhaps the applications of both AI and quantum computing to material science, because what could we do with different kinds of materials that are better in a whole host of ways? The sky's the limit.

Yeah, I was going to say exactly the same thing. I watch the spotlight move around, and the spotlight is very large on AI right now, but it was shining on biology and it shouldn't stop shining on it. There are all kinds of things going on in synthetic biology, very exciting things. So I think we need to broaden that spotlight a little bit.

Huh. Limiting beliefs. I guess I had a very... my life expanded pretty dramatically at a bunch of stages. I was born in Moscow in the Soviet Union, and it's very different, very poor. Everybody was very poor. I lived in a little 400 square foot apartment with my parents and my grandmother and had to walk up five flights of stairs. I didn't really think about the world outside. I was lucky that my father got a hint of the world outside. He went to some conference in Poland where they told him what the western world was like, and he decided to move us, which was very controversial at the time in the family. Eventually we got to the US, and we were still very poor and had to make our way out of having nothing. I had to learn a new language, make all new friends. It was a challenging transition but awakening. When I came to grad school at Stanford, it was similar. I had all this freedom, the professors entrusted me, and something about California was very freeing and liberating in thought, given the tradition of the state. One that we're a little bit getting away from in California, if I'm being honest, but I'm not going to complain about that. I guess it's this experience. I guess I'm answering the question backwards, not really a limiting belief. I had the experience of expanding my world in ways that seemed very painful at the time but later paid off because of my personal history. Those challenging transitions can pay off.

Thank you. What is the definition of a good life? Well, I guess it's being able to enjoy your life, whatever you build. I like to have family. I have one of my kids here, my girlfriend is here. I feel grateful to be able to spend quality time with them. I do feel quite grateful to be able to be intellectually challenged at this stage. I actually retired a month before COVID hit, and it was the worst decision. I had this vision that I was going to sit in cafes and study physics, which was my passion at the time. That didn't work because there were no more cafes. I was just kind of stewing and felt myself spiraling, not being sharp. Then I was like, I got to get back to the office, which at the time was closed, but after a number of months, we started to have some folks going to the office, and I started to do that occasionally. Then I started spending more and more time on what later became called Gemini, which is super exciting. To be able to have that technical creative outlet is very rewarding. If I had stayed retired, I think that would have been a big mistake.

Okay, I'm going to share it and then I'm going to try to undo it. I realize that sometimes I'm embarrassed to ask things I don't know, but I will go ahead. Wait, what is management science and engineering like? Is it like a, I'm going to manage? How does that work?

It's a class.

Wait, this class department is management science.

But what do you study? Like what are the classes?

All right. Well, I guess I didn't really know that. So that's my embarrassing truth. But I'm glad I asked.

Okay, I'm going to try to do this without advertising. The thing I like to do, but you shouldn't do it now because we have a way better version coming, is I talk to Gemini live in the car often and I ask... but the publicly available version right now is not our good version. So you shouldn't do it today, but give me a few weeks to actually ship what I have access to, because we have an ancient model behind it in the publicly released version right now. It's a little embarrassing. But I do like to ask it, whatever I want to develop a data center, how many hundreds of megawatts of this kind of power, that kind of power, how much it's going to cost, and I just talk to it about stuff on my drive. That does seem kind of self-advertising with Gemini. I do periodically listen to a whole bunch of podcasts. I like the All-In guys, they're actually one of my favorites. They're great hosts. We just visited Ben Shapiro, another broadcaster down in Florida, got to see his studio. A bunch of these podcasters are actually pretty fun to meet in person. But that's not how you're going to learn about it. I do just listen to them, see what's up. But I prefer to have an interactive discussion on my drive, so that's why I talk to the AI, as embarrassing as that sounds.

Sort of a glimpse of the future, I think. Actually, that's a good way to end. We'll probably all be doing it.

Thank you.